{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 10 | April 2025 | 526\u2013534\n526\nnature energy\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nFeasibility of meeting future battery demand \nvia domestic cell production in Europe\n \nSteffen Link\u2009\n\u200a\u20091,2\u2009\n, Lara Schneider\u2009\n\u200a\u20091, Annegret Stephan\u2009\n\u200a\u20091,  \nLukas Weymann\u2009\n\u200a\u20091 & Patrick Pl\u00f6tz\u2009\n\u200a\u20091\nBatteries are critical to mitigate global warming, with battery electric \nvehicles as the backbone of low-carbon transport and the main driver \nof advances and demand for battery technology. However, the future \ndemand and production of batteries remain uncertain, while the \nambition to strengthen national capabilities and self-sufficiency is gaining \nmomentum. In this study, leveraging probabilistic modelling, we assessed \nEurope\u2019s capability to meet its future demand for high-energy batteries \nvia domestic cell production. We found that demand in Europe is likely to \nexceed 1.0\u2009TWh\u2009yr\u22121 by 2030 and thereby outpace domestic production, \nwith production required to grow at highly ambitious growth rates of \n31\u201368%\u2009yr\u22121. European production is very likely to cover at least 50\u201360% of \nthe domestic demand by 2030, while 90% self-sufficiency seems feasible but \nfar from certain. Thus, domestic production shortfalls are more likely than \nnot. To support Europe\u2019s battery prospects, stakeholders must accelerate \nthe materialization of production capacities and reckon with demand \ngrowth post-2030, with reliable industrial policies supporting Europe\u2019s \ncompetitiveness.\nBatteries are of critical importance for the rapid reduction of green-\nhouse gas (GHG) emissions to mitigate global warming and meet the \n1.5\u2009\u00b0C target of the Paris Agreement by enabling the widespread use of \nrenewable electricity1,2. This requires transformative changes in the \ntransportation sector2\u20135 as transport accounts for approximately 20% \nor 7\u2009GtCO2-equivalent (7\u2009GtCO2e) of global annual GHG emissions6, with \nEuropean transport emitting ~800\u2009MtCO2e (ref. 7).\nWhile some studies have emphasized the difficulties involved \nin decarbonizing transport4,8, there is robust evidence that battery \nelectric vehicles (BEVs) will form the backbone of future low-carbon \nroad transport2,5. Accordingly, BEVs prevail in the future portfolios of \ncar manufacturers9,10 and several European countries will enforce 100% \nzero-emission vehicle (ZEV) sales for cars by at least 20359\u201311, banning \nlarge-scale sales of conventional vehicles as sufficient quantities of \nsustainable fuels are unlikely11. Other key markets, such as the United \nStates and China, have also set ambitious ZEV targets from the 2030s9,11.\nRegarding battery demand, electrified transport is widely \nrecognized as the key driver12 and catalyst for battery advances13. \nHowever, any projection of battery demand is highly uncertain and \nscenario-dependent, generally based on concealed models and sub-\nject to unclear assumptions. Hence, global demand projections for \n2030 cover a wide spectrum, typically 3\u20136\u2009TWh\u2009yr\u22121 and up to almost \n9\u2009TWh\u2009yr\u22121, with European projections at around 0.7\u20131.4\u2009TWh\u2009yr\u22121 (for \nan overview, see Supplementary Table 2).\nRegarding battery production, electrification requires indus-\ntrial transformation and the establishment of new battery ecosys-\ntems alongside the entire value chain from raw material extraction \nto end-of-life, and including concepts of circularity, such as second \nuse or recycling14,15. In addition, the COVID-19 pandemic and other \ngeopolitical tensions have created awareness of vulnerable economic \ndependencies and spurred the development of modern concepts such \nas technological sovereignty16,17 and resilience14,15. In response, the \nEuropean Union (EU) recently finalized the Net-Zero Industry Act18, \nintending to ensure ample capacity for strategic net-zero technologies \nby 2030, including a target to satisfy at least 90% of its battery demand \nfrom domestic cell production18. Global battery production capacity is \nReceived: 20 May 2024\nAccepted: 24 January 2025\nPublished online: 6 March 2025\n Check for updates\n1Fraunhofer Institute for Systems and Innovation Research ISI, Karlsruhe, Germany. 2Institute of Electrical Engineering (ETI), Karlsruhe Institute of \nTechnology (KIT), Karlsruhe, Germany. \n\u2009e-mail: steffen.link@isi.fraunhofer.de\n\nNature Energy | Volume 10 | April 2025 | 526\u2013534\n527\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nin Europe, which theoretically add up to 2.55\u2009TWh\u2009yr\u22121 by 2030. We \nimputed an S-shaped production ramp-up for each announcement to \nderive its effective, materialized annual capacity using a logistic growth \nmodel (LGM) and probabilistic parameterization of influencing factors \n(that is, utilization, scrap, delays and materialization probabilities). \nThese characteristics result from two typical ramp-up phases following \nan earlier planning period: time-to-market and time-to-volume30,31,27. \nThe time-to-market stage typically comprises pre-series and pilot  \nproduction and ends with the start of production and is characterized \nby low utilization yet high scrap30. The time-to-volume stage repre-\nsents the progression towards established processes with lower scrap,  \nhigher utilization and high output30,27. Finally, we determined the  \neffective European production capacity per year by accumulating the \nindividual announcements (Methods).\nSecond, we considered high-energy batteries for BEVs to be the \nmain demand driver with lithium-ion batteries as the prevalent cur-\nrent technology. We collected historical automotive data, such as \ntotal sales and segment splits, as well as historical BEV-specific data, \nsuch as model-specific sales, installed battery capacities, segment \nsplits and energy consumption. We combined these data with pro-\njections of future European battery demand based on probabilistic \nparameterization of future BEV shares, segment-specific battery \ncapacities, segment splits, total sales and additional demand from \nother mobile or stationary applications. To project BEV shares, we fit-\nted an LGM to historical BEV shares and potential future shares. Here, \nwe assumed the potential BEV share by 2035 to follow the announced \nZEV targets, and further varied the inflection point to refine the \nS-curve and capture an extended feasibility space for battery demand \n(Methods).\nEuropean demand and production approach \n1\u2009TWh by 2030\nFigure 1 shows European battery demand and its domestic produc-\ntion capacity up to 2035. Fuelled by substantial BEV diffusion up to \n2035, European battery demand is likely to surpass 1.0\u2009TWh\u2009yr\u22121 by \n2030 (in 69% of all scenarios). The interquartile range (IQR) in 2030 is \n0.97\u20131.2\u2009TWh\u2009yr\u22121. Some high-demand scenarios may exceed the 1\u2009TWh \nthreshold as early as 2026 and even approach 1.6\u2009TWh\u2009yr\u22121 by 2030,  \nwith the top 10% exceeding 1.30\u2009TWh\u2009yr\u22121. High demand may emerge \nfrom the superposition of high total sales, high BEV shares, higher seg-\nment shares for larger vehicles, larger battery capacities per segment \nexpected to approach 7.0\u2009TWh\u2009yr\u22121 by 2030, with European capacity at \naround 0.8\u20131.6\u2009TWh\u2009yr\u22121 (for an overview, see Supplementary Table 1). \nHowever, it is unclear what proportion of the announced capacities will \nmaterialize or whether production facilities can expand fast enough \nto meet growing demand.\nHence, in this study, we addressed the following research ques-\ntion: how likely is it that Europe can cover its future battery demand \nvia domestic production? Leveraging probabilistic modelling, we \nfound that European demand is likely to exceed 1.0\u2009TWh\u2009yr\u22121 by 2030 \nand thereby outpace domestic production, with production capacity \nrequired to grow at almost exceptionally high growth rates (31\u201368%\u2009yr\u22121) \nover the long term with the latter\u2019s momentum increasing after 2025. \nHowever, it is very likely that Europe can cover at least 50\u201360% of its \ndemand via domestic production by 2030. Even 90% self-sufficiency \nseems within reach, yet far from certain, so there is an increased risk \nof domestic production shortfalls. In contrast, if lower production \ncapacity materializes and domestic production remains limited, this \nwill likely pose high economic risks for Europe and imply less European \nbattery sovereignty and setbacks for rapid climate change mitigation.\nModelling future battery demand and production\nProbabilistic modelling is developing as an accurate and flexible tool \nto assess the feasibility of future climate change mitigation pathways19 \nand modelling technology diffusion20. We adopted this approach to \nindependently project future battery demand and domestic produc-\ntion in Europe and to evaluate Europe\u2019s pathway towards battery \nself-sufficiency via feasibility spaces and probabilistic statements. \nAccordingly, we extended the recent advances by Odenweller et al.21, \nwhich were based on Roger\u2019s concept of technology adoption22 and \ntypical S-shaped diffusion curves23.\nFirst, we estimated domestic production using existing and \nannounced cell production facilities and their stated capacities. There \nare different motivations for publicly announced production facilities \nand taking them at face value can easily lead to the overestimation of \nactual capacity and high ad hoc availability. However, uncertainties \nin multiyear projects are inevitable as they invariably involve delays \nrelated to construction, permits or equipment readiness, cancella-\ntions or postponements, step-by-step expansion plans, evolutionary \nadvances in production technologies or quality issues15,24\u201329. Therefore, \nannounced capacities usually decrease to an effective, materialized \noutput. In our model we used N\u2009=\u2009144 facilities between 2020 and 2030 \n2,000\n1,750\n1,500\n1,250\n1,000\n750\n500\n250\n0\nBattery capacity (GWh yr\u20131)\n2025\n2030\n2035\nYear\nProbability density\nfor year 2025\nProbability density\nfor year 2030\nEfective European battery production \nTotal European battery demand\nFig. 1 | European battery demand and its domestic production capacity up to \n2035. Probabilistic feasibility space for different scenarios (N\u2009=\u20091,000) of total \nbattery demand and battery production capacity up to 2035 for Europe. The \nshaded areas show all potential model results, the dashed lines indicate the 5th \nand 95th percentiles, and the solid line indicates the median. Probability density \nplots for 2025 and 2030 are shown on the right. Reference values for production: \nAround 150\u2009GWh\u2009yr\u22121 of stated capacity was available at the beginning of 2024.\n\nNature Energy | Volume 10 | April 2025 | 526\u2013534\n528\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nand increased demand by other applications, and vice versa for low \ndemand. Low-demand scenarios barely fall below 0.85\u2009TWh\u2009yr\u22121 by 2030 \n(5th percentile). Concerning achievable BEV sales shares (Extended \nData Fig. 1), the feasibility space is deemed to narrow towards 2035 \ndue to the European ZEV regulations. Accordingly, we will most likely \nobserve around 30% BEV sales by 2025 (median, IQR\u2009=\u200925\u201334%) and \n70% by 2030 (median, IQR\u2009=\u200965\u201380%).\nIn contrast to the demand scenarios, the feasibility spaces  \nfor domestic production have more distinct peaks (Fig. 1). Domestic pro-\nduction is unlikely to surpass 1.0\u2009TWh\u2009yr\u22121 by 2030 (39%) and the IQR is \n0.93\u20131.04\u2009TWh\u2009yr\u22121. A few high-production scenarios reach 1.2\u2009TWh\u2009yr\u22121, \nwith the top 5% exceeding 1.10\u2009TWh\u2009yr\u22121. Low-production scenarios \nbarely fall below 0.86\u2009TWh\u2009yr\u22121 (5th percentile). Low production  \ncapacity may emerge from the superposition of a low materialization \nrate, low utilization, extended delays and high scrap, and vice versa \nfor high production.\nThe resulting average growth rates for demand (Extended Data \nTable 1) are ambitious at 31\u201343%\u2009yr\u22121, but they have been witnessed \nfor other technologies. Historically, wind and solar capacity growth \nrates have been at least 15%\u2009yr\u22121 and often between 39 and 50%\u2009yr\u22121 \n(ref. 21). General technology adoption growth rates have typically \nbeen below 13\u201314%\u2009yr\u22121, but occasionally have exceeded 30\u201340%\u2009yr\u22121 \n(ref. 32). In contrast, the calculated growth rates for production are \nmore ambitious, ranging from 55 to 68%\u2009yr\u22121, and close to the levels \nwitnessed historically in times of emergency when unconventional \nrates of growth were observed, as indicated by Odenweller et al.21. \nAccordingly, if such exceptional growth rates were not realized, we \nwould observe substantially lower European production capacity by \n2030, affirming that materialized capacity expands more slowly than \nannouncements might suggest.\nSelf-sufficiency of 90% in 2030 is feasible but far \nfrom certain\nFigure 2 shows the individual scenarios as trajectories (left) and density \nplots with cumulative density (right) for production versus demand \nfor 2025 and 2030. It is very likely that domestic production can meet \nat least 60% of demand in 2025 (90.1%) and 2030 (99.5%), although the \nresults for 2025 reveal a slightly elevated risk of short-term domes-\ntic production shortfalls. For 2030, the 90% self-sufficiency target \nseems feasible as this corresponds to the mean and median value \n(IQR\u2009=\u200980\u2013100%), but is far from certain as nearly half of the scenarios \n(49%) do not reach the 90% self-sufficiency target. Production capacity  \nexceeds domestic demand by more than 10% in a minor share of sce-\nnarios (11%).\nOur results are stable, even if we replace the LGM with Gompertz \nor Bass diffusion models to capture asymmetric growth (Extended \nData Fig. 3a,b). The results from the Gompertz model show narrower \nfeasibility spaces at lower growth rates (20\u201327%\u2009yr\u22121 for demand and \n31\u201348%\u2009yr\u22121 for production, Extended Data Table 1) and an aggravated \nshortfall of domestic production capacity, particularly around 2025. \nMoreover, if we limited the growth rates (15\u201339%\u2009yr\u22121) for the total pro-\nduction capacity using the Gompertz model (Extended Data Fig. 3c), \nwe would observe ~0.89\u2009TWh\u2009yr\u22121 by 2030 (IQR\u2009=\u20090.84\u20130.96\u2009TWh\u2009yr\u22121) \nand exceeding 1\u2009TWh would be more unlikely (15%). Accordingly, the \n90% self-sufficiency target would be even more at risk (25%), while at \nleast 50% self-sufficiency would still be very likely (98.1%, median\u2009=\u200983%, \nIQR\u2009=\u200975-90%).\nGrowing engagement of European companies\nBeyond mere domestic production capacity and self-sufficiency, the \ncompany\u2019s origin is relevant in the context of accessibility and techno\u00ad\nlogy sovereignty. While the corporate landscape was nearly 100% Asian \nin the early 2020s, the share of European companies is projected to \nincrease substantially. In 2025, around two-thirds of the materialized \nproduction capacity is likely to result from Asian-affiliated compa-\nnies and more than one-third from European companies (Extended  \nData Fig. 2). By 2030, European companies are projected to hold  \nthe largest share (45\u201355%), while the share of Asian companies  \nis expected to decline (40\u201350%) with US companies anticipated to \ncapture modest shares (3\u20138%).\nEurope\u2019s position on battery raw materials is \nimproving\nThe complexity of battery value chains13,33 implies that production \ncapacity should be assessed alongside raw material sourcing, where \ndomestic availability is a decisive factor for Europe and also for other \ncountries34,35. Expressing the European battery demand in terms of \nrequired raw material quantities reveals that the cumulative demand \nfor key materials, namely, nickel, cobalt, graphite, lithium and man-\nganese, is projected to increase substantially by 2035, with expected \n9-fold (cobalt) and 12\u201315-fold (nickel, manganese, graphite and lithium) \nincreases relative to the quantities required in 2025 (Table 1). While \nEurope will rely on raw material imports until 2030\u20132035, three factors \nindicate a strengthening position. First, and in relation to expected \ndemand, substantial domestic reserves of manganese and natural \ngraphite are available, with possibly lower prospects for lithium and \nnickel, but primary cobalt is scarce. Second, existing self-sufficiency \nassessments (Table 1, rightmost column) indicate progress in building \n2,000\n0.030\n0.025\n0.020\n0.015\n0.010\n0.005\n0\n2025\n2030\n1,750\n1,500\n1,250\n1,000\n750\n500\n250\n250\n500\n750\n1,000\n1,250\n1,500\n1,750\n2,000\n40\n60\n80\n100\n120\n140\n0\n0.2\n0.4\n0.6\n0.8\n1.0\n0\n0\nBattery demand (GWh yr\u20131)\nBattery production (GWh yr\u20131)\nCoverage of total demand by domestic production (%)\nProbability density\nCumulative density function\n90% European\nself-suficiency\n2025\n2030\nFig. 2 | Comparison of anticipated European battery demand and supply in \n2025 and 2030. Left: comparison of results of different scenarios (N\u2009=\u20091,000) \nfor domestic production capacity versus demand in Europe. The trajectories \nfor 2025\u20132030 are included as grey lines. The black line corresponds to supply \nmatching demand (100%). Right: feasibility space (N\u2009=\u20091,000) for the relative \ncoverage of total European demand. Probability density (histograms, left y axis) \nand cumulative density (right y axis) are shown for 2025 and 2030. The black \ndashed line marks the 90% self-sufficiency level.\n\nNature Energy | Volume 10 | April 2025 | 526\u2013534\n529\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nEuropean value chains, however, ramp-ups must be extremely quick. \nWhile cobalt and nickel imports (all grades) are likely to remain neces-\nsary for domestic processing, it is likely that major shares of lithium \nand most of the manganese can be sourced and refined domestically. \nNatural graphite (all grades) is likely to require both local sourcing and \nrefining as well as imports. However, global supply diversification is \nanticipated to also lower general dependency risks36,37. Third, emphasiz-\ning the circular economy and recycling, as proposed in the EU\u2019s Critical  \nRaw Materials Act38 or incentivized by the US Inflation Reduction  \nAct35, is likely to reduce dependency and further improve sustainability \nwithin a comprehensive battery ecosystem13,39\u201341, also securing mate-\nrial availability even beyond 205042. Yet, current estimates of battery \nreturns36,43,44 strongly indicate that recycling or second use/repurposing \nof batteries will be less relevant in the early 2030s, but will sharply grow \nin importance thereafter, in particular, unlocking potential for nickel37,43 \nand cobalt41. In addition, battery technologies such as sodium-ion \nbatteries are an intended alternative to lithium-ion batteries; these \nare currently at the beginning of large-scale commercialization24,45.\nUncertainties and policy implications\nOur results indicate that the course has been set for transitioning from \nbattery research and development (R&D) to production and establish-\ning value chains in Europe. While substantial capacity additions can be \nexpected until 2030, reaching the adopted EU 90% self-sufficiency tar-\nget by 2030 is highly challenging as production capacity would require \na highly ambitious growth rate. Nonetheless, our results imply that the \ndomestic production of batteries is unlikely to be a bottleneck for BEV \nmarket diffusion in Europe and will conform to the transport-specific \nemission budgets required to comply with the 1.5\u2009\u00b0C global warming \ntarget (that is, ~70% ZEV sales by 2030 and 100% by no later than 20335).\nHowever, our analysis has limitations related to probabilistic \nparameter choice and parameter assumptions, independent model-\nling with missing feedback mechanisms, data availability or neglected \nvehicle trading. Nevertheless, our individual projections align with \nother studies (Supplementary Tables 1 and 2), and furthermore, we \nhave demonstrated robustness by varying the diffusion models and \ncertain parameters.\nFirst, parameters were sampled from independent programme \nevaluation review technique (PERT) distributions (Methods). Specific \nminimum, most likely and maximum values were primarily based on \nliterature values, but some were in-house assumptions and projections \nbased on recent developments. While this implies representative fea-\nsibility spaces per parameter, real-world distributions are unknown, \nparameter dependencies were disregarded and potential trend breaks \nmay have been missed, such as declining popularity in sport utility \nvehicles (SUVs) or shrinking battery sizes. Future studies could further \ndecompose parameters such as demand from other applications.\nSecond, capturing dependencies and feedback mechanisms \nbetween supply and demand might constrain the joint feasibility space \nas substantially decoupled scenarios, such as high production with low \ndemand or vice versa, may be deemed improbable. Similarly, demand \nscenarios where BEV ranges remain low but their sales shares rise high \nmay be considered improbable, given that perceived range anxiety \nremains a significant barrier for many potential buyers46,47. Future \nstudies could use other approaches, such as system dynamics.\nThird, production capacities were derived from individual \nannouncements (bottom-up, cut-off January 2024) rather than \nmodel-based or demand-driven outlooks (top-down), confining our \ndataset and results. On the one hand, we emphasize that the lack \nof standardization makes it difficult to distinguish what is or is not \naccounted for in the announced capacity. If capacity utilization were \nincluded, we would obtain higher production capacities by 2030 \n(IQR\u2009=\u20091.15\u20131.30\u2009TWh\u2009yr\u22121) with unrestricted growth rates and an \nalmost parallel trend slightly below demand with limited growth rates \n(Extended Data Fig. 4a\u2013c). On the other hand, as there are no explicit \nannouncements beyond 2030, supply curves flatten towards 2030 \neven though our model accounts for delayed realization, potentially \nleading to underestimated production capacities between 2030 and \n2035. Of course, additional capacities may emerge after 2030. However, \nno announcements have been made so far, even though there are only \n5\u2009years left before this date and not every factory is likely to expand at \nshort notice.\nFourth, we linked domestic battery demand to domestic sales, \nassuming that European sales equal production figures and neglecting \nany foreign trade, as with the batteries themselves. While total Euro-\npean vehicle sales and production figures are a good match (\u00b110%), \nexports dominated imports in the trade balances from 2015 to 2022, \nwith clear distinctions between imported (smaller and volume-type) \nand exported (larger and more premium-type) vehicle segments.\nIn our analysis, we disregarded upstream and downstream parts \nof the value chain that might constrain Europe\u2019s prospects too. Most \nnotably, capacity for the refinement of raw materials as well as produc-\ners of further processed cathode and anode active materials (CAMs \nand AAMs) and cell components will be required to supply factories \ndomestically. In this respect, there have been several announcements \nin Europe, with more progress for CAM processing27 than for AAMs48. \nNevertheless, most of the announced battery cell plants are likely to \nhave secured long-term supply contracts, mitigating the risk of material \nTable 1 | Overview of European demand, production, reserve (primary) and resource (primary) data for five critical battery \nraw materials\nMaterial (unit)\nAnnual values (kt\u2009yr\u22121)\nCumulative values (kt)\nPotential European \nself-sufficiency by \n2030 (%)\nCurrent European \nproduction\nDemand by \n2025\nDemand by \n2035\nDemand \nby 2025\nDemand by \n2035\nCurrent European \nreserves\nEuropean \nresources\nLi (kt LCE \ncontent)\nP: <1\nR: NA\n30\u2009\u00b1\u20096\n170\u2009\u00b1\u200910\n100\u2009\u00b1\u200910\n1,230\u2009\u00b1\u2009130\n60\u2013470\n6,700\u20137,000\nP: 25\u201360\nR: 27\u201385\nNi (kt Ni \ncontent)\nP: ~50\nR: ~70\n180\u2009\u00b1\u200930\n820\u2009\u00b1\u200990\n510\u2009\u00b1\u200950\n6,170\u2009\u00b1\u2009700\n~5,200\n~12,300\nP: 20\u201323\nR: 20\u201328\nCo (kt Co \ncontent)\nP: ~50\nR: <10\n40\u2009\u00b1\u20098\n120\u2009\u00b1\u200920\n130\u2009\u00b1\u200910\n1,150\u2009\u00b1\u2009150\n90\u2013340\n520\nP: 1\u201320\nR: 20\u201337\nMn (kt Mn \ncontent)\nP: <7\nR: NA\n50\u2009\u00b1\u200910\n320\u2009\u00b1\u200940\n150\u2009\u00b1\u200910\n2,140\u2009\u00b1\u2009260\n~2,000\n~2,000\nP: 45\u2013100\nR: 28\u2013100\nGr (kt refined \ngraphite)\nP: <15\nR: NA\n260\u2009\u00b1\u200950\n1,520\u2009\u00b1\u2009120\n730\u2009\u00b1\u200980\n10,360\u2009\u00b1\u20091,070\n2,000\u20137,500\n~30,900\nP: <5\nR: 21\u201326\nAll values are rounded and specified in kilotonnes. Demand values (refined and battery-grade quality) are specified as mean\u2009\u00b1\u2009standard deviation. The ranges for potential European \nself-sufficiency for each material by 2030 (without recycling) are based on literature data; reference details and original values are provided in Supplementary Tables 22 and 23. Calculated \ndemand values and data for 2030 are provided in Supplementary Tables 25 and 26 and visualized in Supplementary Figs. 2 and 3. P, primary; R, refined, battery quality; Li, lithium; LCE, lithium \ncarbonate equivalent; Ni, nickel; Co, cobalt; Mn, manganese; Gr, natural graphite; NA, not available.\n\nNature Energy | Volume 10 | April 2025 | 526\u2013534\n530\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nshortages. Finally, the current shortage of skilled workers may continue \nto outpace the build-up in Europe15,24,25. This shortage affects scientific \nand industrial R&D at all technology readiness levels, as well as direct \nand indirect jobs in the sector, from factory build-up to cell manufac-\nturing and system-level integration. Future studies could adapt our \nprobabilistic modelling to integrate these aspects.\nFocusing on the implications for decision-makers in policy and \nindustry, Europe is striving for a competitive and sustainable battery eco-\nsystem with concrete agendas and roadmaps49, including the Net-Zero \nIndustry Act18. This includes launching several initiatives and public\u2013\nprivate partnerships, such as Batteries Europe and the Batteries Europe \nPartnership Association, or funding under the framework of the Impor-\ntant Projects of Common European Interest to align industry and policy \nperspectives on cross-cutting issues such as scale-up, sustainability,  \nrecycling and digitalization50. In addition, there are special trade and \ncooperation agreements, such as the EU\u2013UK Battery Rules of Origin.\nAlthough there is likely no single policy strategy to boost battery \necosystems, we highlight the role of industrial policies in balancing \ntrade protectionism and global competitiveness. While the current  \nEU battery policy has a supply-side emphasis with many initiatives \nfocusing on R&D50, recent initiatives such as the Strategic Technologies \nfor Europe platform51 or the Temporary Crisis and Transition Frame-\nwork52 have an explicit focus on industrial development, competitive-\nness and sovereignty. Recent examples, such as the Northvolt battery \nfactory in Germany53, indicate that such industrial policies might be \neffective in spurring domestic projects (irrespective of an individual \ncompany\u2019s future development). The global race might call for poli-\ncies that create attractive, predictable home markets33 and reduce the \nrisks for industry players, such as public\u2013private risk\u2013reward\u2013sharing \ninstruments, while purely inward-looking policies might be rather \ndetrimental17,54. Regarding prioritization, we emphasize that establish-\ning fully scaled and sustainable value chains simultaneously presents \nenormous challenges and the inherent risk of impeding fast competi-\ntiveness. China, for instance, started building internationally competi-\ntive battery value chains years ago, leads current battery R&D, employs \nan extensive specialized workforce, and is now progressing towards \nmore circular, safer and more sustainable batteries50. Finally, we high-\nlight the importance of using net materialized production capacities \nas a basis for projections rather than announced capacities, which fail \nto consider scrap rates, new technology developments, non-optimal \ncapacity utilization or delays related to construction and permits.\nFor industry players, our results indicate substantial potential for \ndomestic added value, tailoring batteries to Europe\u2019s needs, as well as \nlocalizing raw material production and battery recycling. Moreover, \nour results call for further investments in local battery production to \navoid shortages in domestic production, while keeping track of inter-\nnational developments.\nConclusions\nIn this study, we used probabilistic modelling of S-shaped produc-\ntion ramp-up and technology diffusion based on the latest empirical \ndata to project future battery demand and domestic production in \nEurope. This allowed us to evaluate Europe\u2019s prospects towards battery \nself-sufficiency via feasibility spaces and probabilities in a consistent, \ntransparent and thorough manner. Based on our results, we can draw \nfour main conclusions.\nFirst, we have shown that European demand is likely to experience \nambitious growth to at least 1.0\u2009TWh\u2009yr\u22121 by 2030. In contrast, domes-\ntic production capacities are more likely to fall behind terawatt hour \nscales by 2030, with momentum increasing after 2025 and production \ncapacities required to grow at unconventional growth rates.\nSecond, the prospects for a European footprint along the battery \nvalue chain are remarkable. On the one hand, covering at least 50\u201360% \nof its demand via domestic production by 2030 is very likely and the \ncontribution from European-affiliated companies is projected to grow \nsubstantially, while potential to improve the raw material situation in \nEurope also exists. On the other hand, 90% self-sufficiency by 2030 \nseems feasible but far from certain. We thus emphasize an increased \nlikelihood of limited European competitiveness and domestic produc-\ntion shortfalls, particularly before 2030.\nThird, the short-term volatility of battery demand, mainly caused \nby uncertain BEV popularity and availability, should not unsettle invest-\nment decisions and detract from long-cycle commitments based on \nGHG mitigation, manufacturer announcements, and ZEV targets in \nEurope and automotive core markets worldwide. Note that inter-\nnational ZEV targets will also restrict European export capacity for \nnon-BEVs in the future.\nFourth, the global battery manufacturing race, complex value \nchains, raw material dependencies and time-consuming ramp-ups \nstress the urgency of immediate action and reliable policies to lower \nrisks and ensure certain predictability. Without these, Europe\u2019s need \nto import batteries is likely to persist, or even increase, thus keeping \nEurope dependent on imports in a key technology for current and \nfuture sustainable transport and energy.\nMethods\nGeneral approach\nWe used probabilistic modelling of S-shaped production ramp-up and \ntechnology diffusion based on the latest empirical data to project future \nbattery demand and domestic production in Europe, covering the EU, \nthe European Free Trade Association (EFTA) and the United Kingdom. \nWe also translated this battery demand into the corresponding amounts \nof required critical raw materials. This probabilistic approach captures \nthe nonlinear propagation and simultaneous interaction of major \nparameters with inevitable uncertainties, in contrast to deterministic \nmodels that rely on single estimate inputs. However, the goodness of \nthe projections still depends on the quality of the input data and chosen \nmodel parameterization. A Monte Carlo simulation (N\u2009=\u20091,000) then \nallows the construction of feasibility spaces and the classification of \nfindings by probability. The model was implemented in Python. Model \ncomponents and procedures are illustrated in Supplementary Fig. 1. \nWe performed all calculations on a standard Lenovo notebook with an \ni7-8565U @1.8\u2009GHz processor and 16\u2009GB RAM (random access memory).\nParameter documentation\nWe provide comprehensive data tables for each modelling parameter, \nincluding further information and data sources, in the Supplementary \nInformation (production parameters in Supplementary Tables 3\u20137, \ndemand parameters in Supplementary Tables 8\u201318 and raw material \nparameters in Supplementary Tables 19\u201323).\nData on announced battery production capacities\nWe collected data on existing and announced battery production \nfacilities from the Fraunhofer Institute for Systems and Innovation \n(ISI) database, which lists 144 European projects for the period 2020\u2013\n2030. The database includes details of development status/risk level \n(secure, likely or insecure), technology characteristics, company infor-\nmation and, most importantly, annual production capacities for the \nannounced projects, totalling 2.55\u2009TWh\u2009yr\u22121 by 2030. The dataset for \nthis study, as of January 2024, is available at Zenodo via https://doi.org/ \n10.5281/zenodo.14505410 (ref. 55), while the latest aggregated version \ncan be accessed at https://metamarketmonitoring.de/en/.\nHistorical automotive data\nWe used data published by the European Automobile Manufacturers\u2019 \nAssociation56 and JATO Dynamics57 for general automotive market  \ndata. These data comprise European sales until 2023 and segment- \nspecific shares (first half of 2022) covering the EU, EFTA and the  \nUnited Kingdom. We differentiated between the AB segment (mini \nand small vehicles), the CM segment (compact cars and multipurpose \n\nNature Energy | Volume 10 | April 2025 | 526\u2013534\n531\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nvehicles), the DEF segment (large and premium-type vehicles) and the \nSUV segment. The SUV system is split into small and medium (SM) and \nlarge (L) vehicles.\nHistorical BEV-specific data\nWe obtained BEV-specific data (see Link et al.58), comprising sales data \nand comprehensive information on installed batteries until 2022, \nfrom the Fraunhofer ISI xEV battery database. Based on these data, \nwe calculated sales-weighted average battery capacities per segment \nand BEV-specific segment splits covering the EU, EFTA and the UK. The \ngross battery capacities were derived by assuming an average share \nof usable energy of 93% (ref. 59). The segment-specific BEV energy \nconsumption was calculated on the basis of European CO2 certification \ndata, as prescribed under Regulation (EU) 2019/631 and provided by \nthe European Environment Agency60. We evaluated all vehicles with \nspecific CO2 emissions of 0\u2009gCO2e\u2009km\u22121 (by the World Harmonized \nLight Vehicles Test Procedure). These data are available up to 2022 and \ncover the EU, the United Kingdom (until 2019), Iceland (from 2018) and \nNorway (from 2019). We assumed energy consumption improvements \nof at least 5% until 2030\u20132035 due to increased energy efficiency or \nroad load reductions. The latter mainly concern lightweight materi-\nals and decreasing battery weight61\u201363 as there is increasing evidence \nthat higher specific energy levels and enhanced system integration \novercompensate increasing battery sizes.\nModelling the effective, materialized domestic production \ncapacity\nFirst, we used the announced annual production capacities (in GWh\u2009yr\u22121) \nas the base matrix, where each project corresponds to a row and each \ncolumn to a year from 2020 to 2030. We extrapolated 2030 values to \n2035. The year in which the first batteries were announced marks the \nstart-up year.\nSecond, we used matrix multiplication to add four influencing \nfactors via probabilistic parameterization: (1) capacity utilization, \nwhere we used the overall equipment effectiveness as proxy, (2) scrap \nrates, (3) project delays or postponements and (4) materialization \ncategories. This results in the probable production capacity. Spe-\ncifically, modelling the evolution of overall equipment effectiveness \n(OEE) is announcement-specific, starting with an initial OEE (oee0) in \nthe first year. We then assumed a gradual increase to the target OEE \n(oeeT) over several years (TOEE) using linear interpolation. All three \nvalues were drawn from PERT distributions (Supplementary Table 4). \nAccordingly, the OEE matrix contains the evolution of OEE for each \nproject from its start-up year until 2035. Similarly, modelling the evo-\nlution of the scrap rate is announcement-specific, starting with high \ninitial scrap rates (sr0) in the first year. We then assumed a gradual \ndecrease to the target scrap rate (srT) over several years (TSR) using \nlinear interpolation. Again, all three values were drawn from PERT \ndistributions (Supplementary Table 5). Accordingly, the scrap matrix \ncontains the evolution of scrap rate for each project from its start-up \nyear until 2035. Modelling potential delays or postponements (tDelay) \nis announcement-specific, whereas values depend on the risk level \n(secure, likely or insecure) and were drawn from PERT distributions \n(Supplementary Table 7). Finally, modelling the probability of mate-\nrialization (pMat) is announcement-specific, whereas values depend \non the risk level (secure, likely or insecure) and were drawn from PERT \ndistributions (Supplementary Table 6).\nThird, we fitted each announcement and its probable produc-\ntion capacity using an LGM. This smooths the data and captures the \nS-curve-shaped production ramp-up. Curve fitting was conducted \nusing the Python SciPy software package64. We determined the  \nmaterialized European production capacity per year by accumu\u00adlating \nthe individual fitted announcements, where we again used curve  \nfitting to determine certain indicators, such as total growth rate and \ninflection point.\nLogistic growth model\nThe empirically observed S-shape over time emerges because of expo-\nnential growth/progress in the initial phase, which slows over time and \ngradually approaches an asymptotic maximum. The accompanying \nstandard logistic growth model is defined as follows:\nC (t) =\nCMax\n1 + e\u2212k(t\u2212tinfl)\n(1)\nwhere C(t) is the output over time t, CMax is the asymptote, k is the growth \nconstant, tinfl is the inflection point and e is Euler\u2019s number.\nOther diffusion models\nOther common diffusion models, such as the Gompertz (equation (2)) \nor Bass diffusion (equation (3)) models, allowed us to capture asym-\nmetric growth and compare the results with those of the symmetric \nLGM model. Both models are defined as follows:\nC (t) = CMax \u00d7 e\u2212e\u2212k(t\u2212tinfl)\n(2)\nC (t) = CMax \u00d7\n1 \u2212e\u2212(p+q)t\n1 +\nq\np \u00d7 e\u2212(p+q)t\n(3)\nwhere q is the coefficient of imitation and p is the the coefficient  \nof innovation.\nBattery demand from BEVs\nThe European battery demand from BEVs results from the probabil-\nistic parameterization of future BEV sales shares, battery capacities \nper segment, BEV-specific segment shares and total vehicle sales. \nFuture parameters were calculated until 2035. The interim values were \ncalculated via linear interpolation using the future target and historical \nvalues as references.\nFuture battery capacities result from segment-specific target \nranges for 2030\u20132035. Targets were inspired by the European Council for  \nAutomotive R&D (EUCAR)65, which specifies 400\u2009km for average \nshort-range vehicles and 600\u2009km for long-range vehicles. Concrete \nrange values were drawn from segment-specific PERT distributions. We \nthen calculated the gross battery capacity by using segment-specific \nenergy consumption and accounting for the average share of usable \nenergy. The segment-specific BEV energy consumption was drawn \nfrom PERT distributions based on historical values60 and potential \nefficiency improvements.\nTotal vehicle sales were projected on the basis of the correlation of \ngross domestic product (GDP) and automotive sales, with sales histori-\ncally anticipating GDP growth. Specifically, we followed GDP expecta-\ntions from the International Monetary Fund (IMF) and Organisation for \nEconomic Co-operation and Development (OECD) for the rest of the \n2020s, and matched historical year-on-year sales growth rates with his-\ntorical GDP growths to determine their relation. Concrete sales growth \nvalues were then drawn from PERT distributions using this relation.\nFor segment-specific BEV sales shares, we assumed a uniform \ndiffusion of BEVs across all segments until 2035, as well as the con-\ntinuation of ongoing trends such as the SUV boom. More precisely, \nwe initiated the BEV-specific split for each segment in 2022 and gradu-\nally aligned this share with the projected total market segment split \nuntil 2035. For the latter, we used the historical evolution of segment \nshares (compound annual growth rate for 2011\u20132022) and the total \nmarket segment shares by 2023 as references. Concrete growth rate \nvalues were drawn from PERT distributions. Finally, the total share of all  \nsegments was normalized to 100%.\nFor BEV sales shares, we fitted an LGM to historical BEV sales  \ndata (2015\u20132023) and potential future sales. Specifically, we added a \npotential inflection point and BEV sales shares by 2035 using proba-\nbilistic parameterization. This inflection point affects the shape of the \n\nNature Energy | Volume 10 | April 2025 | 526\u2013534\n532\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nBEV market diffusion and is intended to approximate various strate-\ngies from BEV boosts to delays. For 2035 and beyond, we followed  \nthe announced emission regulations of most European countries, \nmeaning that (nearly) all new cars must be ZEVs (that is, locally \nemission-free vehicles with 0\u2009gCO2e\u2009km\u22121 exhaust emissions). We \ndefined 90% BEV share as the lower limit and 100% as the upper limit. \nFinal curve fitting was achieved using the Python SciPy package64, and \nthe final asymptote (CMax), growth rate (k) and actual inflection point \n(tinfl) were calculated.\nThe final European battery demand from BEVs was then calculated \nby multiplying the battery capacities per segment, segment-specific \nsales shares, BEV sales shares and total vehicle sales.\nTotal battery demand\nThe final European battery demand from BEVs was then scaled up by \nthe demand from other first-life applications to derive the total bat-\ntery demand. This includes other automotive applications, such as \nplug-in hybrids, electrified commercial vehicles (trucks and buses) and \nlight-electric vehicles. In addition, stationary systems, from home to \nindustrial-scale buffer storages, are likely to evolve as the second largest \nmarket if electrification and the integration of renewables are pushed \nglobally and locally, especially in Europe. Concrete scaling values were \ndrawn from PERT distributions.\nRaw materials\nThe quantities of required raw materials were calculated using three \naverage cathode stoichiometries (that is, medium-nickel, high-nickel \nand iron- or manganese-based materials according to Maisel et al.66) \nand the accompanying specific material demand (in g\u2009Wh\u22121). Potential \ncathode chemistry market shares for 2035 were drawn from PERT \ndistributions and normalized, with values between 2024 and 2034 \nbeing derived via linear interpolation. All modelling parameters and \ndetailed references are provided in Supplementary Figs. 2 and 3 and \nSupplementary Tables 19\u201326. Please note that the model therefore \nestimates the amount of raw material input required to produce the \nbattery cells. The model does not, however, take into the consideration \nthe fact that the actual ramp-up and construction of European mining \ncapacity will require additional time.\nPERT distribution\nThe PERT distribution is frequently used in risk analysis and corres\u00ad\nponds to a re-parameterized/transformed beta distribution. The PERT \nprobability distribution function is defined as:\nf (x, xmin, xml, xmax) = (xmax \u2212xmin)\n1\nB (\u03b1, \u03b2)\nx \u03b1\u22121(1 \u2212x)\n\u03b2\u22121 + xmin\n(4)\nwhere x is the realized value, xmin is the minimum, xml is the most likely \nvalue and xmax is the maximum. The beta function \u0392 and shape para\u00ad\nmeters \u03b1 and \u03b2 are defined as:\nB (\u03b1, \u03b2) = \u222b\n1\n0\nt \u03b1\u22121(1 \u2212t)\n\u03b2\u22121dt\n\u03b1 = 1 + \u03bb \u00d7 xml \u2212xmin\nxmax \u2212xmin\n, \u03bb = 4\n\u03b2 = 1 + \u03bb \u00d7 xmax \u2212xml\nxmax \u2212xmin\n, \u03bb = 4\n(5)\nData availability\nAll data are available in Supplementary Tables 3\u201318 and within the \nPython model. Data can be accessed at Zenodo via https://doi.org/ \n10.5281/zenodo.14505410 (ref. 55). Source data are provided with this \npaper.\nCode availability\nThe Python model code can be accessed at Zenodo via https://doi.org/ \n10.5281/zenodo.14505410 (ref. 55).\nReferences\n1.\t\nDing, Y., Cano, Z. P., Yu, A., Lu, J. & Chen, Z. Automotive Li-ion \nbatteries: current status and future perspectives. Electrochem. \nEnergy Rev. 2, 1\u201328 (2019).\n2.\t\nJaramillo, P. et al. in Climate Change 2022: Mitigation of Climate \nChange (eds Shukla, P. R. et al.) Ch. 10 (IPCC, Cambridge Univ. \nPress, 2022).\n3.\t\nTracking Clean Energy Progress 2023 (International Energy  \nAgency, 2023); https://www.iea.org/reports/tracking-clean- \nenergy-progress-2023\n4.\t\nCreutzig, F. et al. Transport: a roadblock to climate change \nmitigation? Science 350, 911\u2013912 (2015).\n5.\t\nPl\u00f6tz, P. et al. Greenhouse gas emission budgets and policies for \nzero-carbon road transport in Europe. Clim. Policy 23, 343\u2013354 \n(2023).\n6.\t\nRoad Transport. Net Zero Roadmap (International Energy Agency, \n2023); https://www.iea.org/reports/road-transport\n7.\t\nGreenhouse Gas Emissions From Transport in Europe (European \nEnvironment Agency, 2023); https://www.eea.europa.eu/en/\nanalysis/indicators/greenhouse-gas-emissions-from-transport\n8.\t\nSantos, G. Road transport and CO2 emissions: what are the \nchallenges? Transp. Policy 59, 71\u201374 (2017).\n9.\t\nShen, C., Fadhil, I., Yang, Z. & Searle, S. The Global Automaker \nRating 2022. Who is Leading the Transition to Electric Vehicles  \n(International Council on Clean Transportation, 2023);  \nhttps://theicct.org/wp-content/uploads/2023/05/The-Global- \nAutomaker-Rating-2022_final.pdf\n10.\t Global EV Outlook 2023. Catching Up With Climate Ambitions \n(International Energy Agency, 2023); https://www.iea.org/reports/\nglobal-ev-outlook-2023\n11.\t\nUeckerdt, F. et al. Potential and risks of hydrogen-based e-fuels  \nin climate change mitigation. Nat. Clim. Change 11, 384\u2013393 \n(2021).\n12.\t Degen, F., Winter, M., Bendig, D. & T\u00fcbke, J. Energy consumption \nof current and future production of lithium-ion and post \nlithium-ion battery cells. Nat. Energy 8, 1284\u20131295 (2023).\n13.\t Frith, J. T., Lacey, M. J. & Ulissi, U. A non-academic perspective \non the future of lithium-based batteries. Nat. Commun. 14, 420 \n(2023).\n14.\t B\u00fcnting, A. et al. Resilient Supply Chains in the Battery Industry \n(VDI/VDE Innovation + Technik GmbH, 2023); https://vdivde-it.\nde/sites/default/files/document/2023-04-BZF_Studie_\nLieferketten-ENG.pdf\n15.\t Fleischman, J. et al. Battery 2030: Resilient, Sustainable, and \nCircular (McKinsey & Company, 2023); https://www.mckinsey.\ncom/industries/automotive-and-assembly/our-insights/battery-\n2030-resilient-sustainable-and-circular\n16.\t Edler, J., Blind, K., Kroll, H. & Schubert, T. Technology sovereignty \nas an emerging frame for innovation policy. Defining rationales, \nends and means. Res. Policy 52, 104765 (2023).\n17.\t March, C. & Schieferdecker, I. Technological sovereignty as \nability, not autarky. Int. Stud. Rev. https://doi.org/10.1093/isr/\nviad012 (2023).\n18.\t Regulation of the European Parliament and of the Council on \nEstablishing a Framework of Measures for Strengthening Europe\u2019s \nNet-Zero Technology Poducts Manufacturing Ecosystem (Net Zero \nIndustry Act (European Commission, 2023).\n19.\t Jewell, J. & Cherp, A. The feasibility of climate action: bridging \nthe inside and the outside view through feasibility spaces. Wiley \nInterdiscip. Rev. Clim. Change https://doi.org/10.1002/wcc.838 \n(2023).\n\nNature Energy | Volume 10 | April 2025 | 526\u2013534\n533\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\n20.\t Zielonka, N., Wen, X. & Trutnevyte, E. Probabilistic projections  \nof granular energy technology diffusion at subnational level. \nPNAS Nexus 2, pgad321 (2023).\n21.\t Odenweller, A., Ueckerdt, F., Nemet, G. F., Jensterle, M. &  \nLuderer, G. Probabilistic feasibility space of scaling up green \nhydrogen supply. Nat. Energy 7, 854\u2013865 (2022).\n22.\t Rogers, E. M. Diffusion of Innovations (Free Press of Glencoe, 1962).\n23.\t Kucharavy, D. & De Guio, R. Application of S-shaped curves. \nProcedia Eng. https://doi.org/10.1016/j.proeng.2011.03.142 (2011).\n24.\t Wolf, S. & L\u00fcken, M. in Emerging Battery Technologies to Boost the \nClean Energy Transition\u2014Cost, Sustainability, and Performance \nAnalysis (eds Passerini, S. et al.) 103\u2013118 (Springer Nature, 2024).\n25.\t Neuhausen, J., Hafner, C. & Treichel, P. Gigafactories for Batteries.  \nThe Key to Successful Battery EV Adoption (Strategy&, 2022);  \nhttps://www.strategyand.pwc.com/de/en/industries/energy- \nutilities/gigafactories-and-raw-materials.html\n26.\t Arantes, A. & Ferreira, L. M. D. Underlying causes and mitigation \nmeasures of delays in construction projects. J. Financ. Manag. \nProp. Constr. 25, 165\u2013181 (2020).\n27.\t B\u00fcnting, A., Lickert, H. & Vogl, S. Build-Up of the Battery Industry  \nin Europe\u2014Status Quo and Challenges (IPCEI Batteries, 2023);  \nhttps://www.ipcei-batteries.eu/fileadmin/Images/accompanying- \nresearch/market-updates/2023-07-BZF_Kurzinfo_Marktanalyse_\nQ2-ENG.pdf\n28.\t Vorholt, F., B\u00fcnting, A. & Bechberger, M. Turbulent Battery Cell \nMarket (IPCEI Batteries, 2023); https://www.ipcei-batteries.eu/\nfileadmin/Images/accompanying-research/publications/ \n2023-02-BZF_Kurzinfo_Marktanalyse_Q4_22-ENG.pdf\n29.\t Duffner, F., Mauler, L., Wentker, M., Leker, J. & Winter, M. \nLarge-scale automotive battery cell manufacturing: analyzing \nstrategic and operational effects on manufacturing costs. Int. J. \nProd. Econ. 232, 107982 (2021).\n30.\t Surbier, L., Alpan, G. & Blanco, E. A comparative study on \nproduction ramp-up: state-of-the-art and new challenges.  \nProd. Plan. Control 25, 1264\u20131286 (2014).\n31.\t Glock, C. H. & Grosse, E. H. Decision support models for \nproduction ramp-up: a systematic literature review. Int. J. Prod. Res. \n53, 6637\u20136651 (2015).\n32.\t Nemet, G., Greene, J., M\u00fcller-Hansen, F. & Minx, J. C. Dataset on \nthe adoption of historical technologies informs the scale-up of \nemerging carbon dioxide removal measures. Commun. Earth \nEnviron. 4, 397 (2023).\n33.\t Beuse, M., Schmidt, T. S. & Wood, V. A \u201ctechnology-smart\u201d battery \npolicy strategy for Europe. Science 361, 1075\u20131077 (2018).\n34.\t Wang, H. et al. China\u2019s electric vehicle and climate ambitions \njeopardized by surging critical material prices. Nat. Commun. 14, \n1246 (2023).\n35.\t Trost, J. N. & Dunn, J. B. Assessing the feasibility of the Inflation \nReduction Act\u2019s EV critical mineral targets. Nat. Sustain. 6, \n639\u2013643 (2023).\n36.\t Global Critical Minerals Outlook 2024 (IEA, 2024); https://www.iea.\norg/reports/global-critical-minerals-outlook-2024\n37.\t Huisman, J. et al. RMIS\u2014Raw Materials in the Battery Value Chain \n(European Union, 2020); https://publications.jrc.ec.europa.eu/\nrepository/handle/JRC118410\n38.\t Regulation of the European Parliament and the Council on \nEstablishing a Framework for Ensuring a Secure and Sustainable \nSupply of Critical Raw Materials and Amending Regulations (EU) \n168/2013, (EU) 2018/858, 2018/1724 and (EU) 2019/1020 (European \nCommission, 2023).\n39.\t Harper, G. et al. Recycling lithium-ion batteries from electric \nvehicles. Nature 575, 75\u201386 (2019).\n40.\t Sharmili, N., Nagi, R. & Wang, P. A review of research in the Li-ion \nbattery production and reverse supply chains. J. Energy Storage \n68, 107622 (2023).\n41.\t Baars, J., Domenech, T., Bleischwitz, R., Melin, H. E. &  \nHeidrich, O. Circular economy strategies for electric vehicle \nbatteries reduce reliance on raw materials. Nat. Sustain. 4, 71\u201379 \n(2021).\n42.\t Greim, P., Solomon, A. A. & Breyer, C. Assessment of lithium \ncriticality in the global energy transition and addressing policy \ngaps in transportation. Nat. Commun. 11, 4570 (2020).\n43.\t Mine 2024. Preparing for Impact (PwC, 2024); https://www.pwc.com/ \ngx/en/mining/assets/pwc-global-mine-2024.pdf\n44.\t Bomke, J.-H. & Ferk, F. European Battery Recycling Market Analysis. \nA Profitable and Sustainable Business Before 2035 (Strategy&, \n2024); https://www.strategyand.pwc.com/de/en/industries/\nautomotive/recycling-european-battery.html\n45.\t Grey, C. P. & Hall, D. S. Prospects for lithium-ion batteries and \nbeyond\u2014a 2030 vision. Nat. Commun. 11, 6279 (2020).\n46.\t Noel, L., Zarazua de Rubens, G., Sovacool, B. K. & Kester, J. Fear \nand loathing of electric vehicles: the reactionary rhetoric of range \nanxiety. Energy Res. Soc. Sci. 48, 96\u2013107 (2019).\n47.\t Herberz, M., Hahnel, U. J. J. & Brosch, T. Counteracting electric \nvehicle range concern with a scalable behavioural intervention. \nNat. Energy 7, 503\u2013510 (2022).\n48.\t B\u00fcnting, A. & Vogl, S. Battery Cell Production in Europe: \nStatus Quo and Outlook (IPCEI Batteries, 2024); https://www.\nipcei-batteries.eu/fileadmin/Images/accompanying-research/\npublications/2024-05-BZF_Kurzinfo_Marktanalyse_Q2.pdf\n49.\t Bernhart, W. & Heimes, H. Battery Monitor 2023: An Assessment \nof the Current and Future Battery Value Chain (Roland Berger, \n2023); https://www.rolandberger.com/en/Insights/Publications/\nBattery-Monitor-2023-An-assessment-of-the-current-and- \nfuture-battery-value.html\n50.\t Endo, C., Kaufmann, T., Schmuch, R. & Thielmann, A. Benchmarking \nInternational Battery Policies. A Cross Analysis of International Public \nBattery Strategies Focusing on Germany, EU, USA, South Korea,  \nJapan and China (Fraunhofer ISI, 2024); https://www.isi.fraunhofer. \nde/content/dam/isi/dokumente/cct/2024/benchmarking- \ninternational-battery-policies_2024.pdf\n51.\t Ragonnaud, G. & Mileusnic, M. Strategic Technologies for Europe \nPlatform (STEP) (European Parliament, 2023); https://www.\neuroparl.europa.eu/RegData/etudes/BRIE/2023/754547/ \nEPRS_BRI(2023)754547_EN.pdf\n52.\t Temporary Crisis and Transition Framework (European \nCommission, 2024); https://competition-policy.ec.europa.eu/\nstate-aid/temporary-crisis-and-transition-framework_en\n53.\t \u20ac902 million German state aid measure to support Northvolt. \nEuropean Commission Press Release (8 January 2024);  \nhttps://ec.europa.eu/commission/presscorner/api/files/\ndocument/print/en/ip_23_6823/IP_23_6823_EN.pdf\n54.\t Juh\u00e1sz, R., Lane, N. & Rodrik, D. The new economics of \nindustrial policy. Annu. Rev. Econom. https://doi.org/10.1146/\nannurev-economics-081023-024638 (2023).\n55.\t Link, S., Schneider, L., Weymann, L., Stephan, A. & Pl\u00f6tz, P. \nPROBAT\u2014European battery supply and demand simulation. \nZenodo https://doi.org/10.5281/zenodo.14505410 (2024).\n56.\t New car registrations\u2014monthly series available until December \n2023. European Automobile Manufacturers\u2019 Association https://\nwww.acea.auto/nav/?content=press-releases (2012\u20132023).\n57.\t European sales by segments. JATO Dynamics https://www.jato.\ncom/h1-2022-europe-by-segments/ (2022).\n58.\t Link, S., Neef, C. & Wicke, T. Trends in automotive battery cell \ndesign: a statistical analysis of empirical data. Batteries 9,  \n261 (2023).\n59.\t Wassiliadis, N. et al. Quantifying the state of the art of electric \npowertrains in battery electric vehicles: range, efficiency, and \nlifetime from component to system level of the Volkswagen ID.3. \neTransportation 12, 100167 (2022).\n\nNature Energy | Volume 10 | April 2025 | 526\u2013534\n534\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\n60.\t Monitoring of CO2 Emissions from Passenger Cars Regulation (EU) \n2019/631 (European Environment Agency, 2018\u20132022); https://\nwww.eea.europa.eu/en/datahub/datahubitem-view/fa8b1229-\n3db6-495d-b18e-9c9b3267c02b\n61.\t Zhou, W., Cleaver, C. J., Dunant, C. F., Allwood, J. M. & Lin, J. \nCost, range anxiety and future electricity supply: a review of how \ntoday\u2019s technology trends may influence the future uptake of \nBEVs. Renew. Sustain. Energy Rev. 173, 113074 (2023).\n62.\t Deng, J., Bae, C., Denlinger, A. & Miller, T. Electric vehicles \nbatteries: requirements and challenges. Joule 4, 511\u2013515 (2020).\n63.\t Hettesheimer, T. et al. Lithium-Ion Battery Roadmap\u2014\nIndustrialization Perspectives Toward 2030 (Fraunhofer ISI, 2023); \nhttps://www.isi.fraunhofer.de/content/dam/isi/dokumente/\ncct/2023/Fraunhofer-ISI_LIB-Roadmap-2023.pdf\n64.\t Virtanen, P. et al. SciPy 1.0: fundamental algorithms for scientific \ncomputing in Python. Nat. Methods 17, 261\u2013272 (2020).\n65.\t Battery Requirements for Future Automotive Applications \n(European Council for Automotive R&D, 2019); https://eucar.be/\nwp-content/uploads/2019/08/20190710-EG-BEV-FCEV-Battery- \nrequirements-FINAL.pdf\n66.\t Maisel, F., Neef, C., Marscheider-Weidemann, F. & Nissen, N. F.  \nA forecast on future raw material demand and recycling potential \nof lithium-ion batteries in electric vehicles. Resour. Conserv. \nRecycl. 192, 106920 (2023).\nAcknowledgements\nWe gratefully acknowledge funding from the Ariadne 2 (FKZ \n03SFK5D0-2, S.L. and P.P.) and BEMA (FKZ 03XP0272B, L.S., L.W. \nand A.S.) projects by the German Federal Ministry of Education and \nResearch, KAMO\u2014High-Performance Center Profilregion, funded by  \nthe Fraunhofer Gesellschaft (S.L. and P.P.), and a strategic internal \nresearch project (FKZ 4009286, A.S.) funded by Fraunhofer ISI. We thank  \nT. Wicke for providing data and also for his comments, as well as  \nG. Bowman-K\u00f6hler and L. Antill-Blum (all Fraunhofer ISI) for English editing.\nAuthor contributions\nS.L. and L.S. conceived and designed the study in consultation \nwith P.P. S.L. and L.S. collected the data, implemented the model \nand created the visualizations. S.L. wrote the original paper \nwith contributions and revisions from all authors. P.P., A.S. and \nL.W. contributed to the discussion, interpretation of findings, \nrecommendations and policy implications. P.P. supervised the study.\nFundingInformation\nOpen access funding provided by Fraunhofer-Gesellschaft zur \nF\u00f6rderung der angewandten Forschung e.V.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at  \nhttps://doi.org/10.1038/s41560-025-01722-y.\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41560-025-01722-y.\nCorrespondence and requests for materials should be addressed to \nSteffen Link.\nPeer review information Nature Energy thanks the anonymous \nreviewers for their contribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this licence, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2025\n\nNature Energy\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nExtended Data Fig. 1 | Probabilistic feasibility space (N\u2009=\u20091,000) for BEV sales shares until 2035. Logistic growth model (LGM). Historical values are in black (until \n2023) and projections are in green. Shaded areas mark all model results, while the dashed lines indicate the 5% and 95% percentiles, and the solid line indicates the \nmedian. Probability density (PD) plots for 2025, 2030, and 2035.\n\nNature Energy\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nExtended Data Fig. 2 | Probabilistic feasibility space (N\u2009=\u20091,000) for production capacities differentiated by the company\u2019s country affiliation (location of \nheadquarter (HQ)). Covering Europe (left), Asia (middle), and North America (mainly the US, right), visualized as density plots. Logistic growth model (LGM). \nProbability densities for 2025 (teal) and 2030 (purple).\n\nNature Energy\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nExtended Data Fig. 3 | Sensitivity with different diffusion models. Gompertz \ndiffusion model (Upper-A), Bass diffusion model (Middle-B), and Gompertz \ndiffusion model with limited growth rate (Lower-C) - N\u2009=\u20091,000. (1): BEV sales \nshares until 2035. Historic values are in black (until 2023) and projections are in \ngreen. Shaded areas mark all model results, while the dashed lines indicate the \n5% and 95% percentiles, and the solid line indicates the median. (2): BEV battery \ndemand (blue) versus battery production capacity (red) until 2035. Shaded \nareas mark all the potential model results, while the dashed lines indicate the \n5% and 95% percentiles, and the solid line indicates the median. (3) Comparison \nfor domestic production capacity (x-axis) versus demand (y-axis) in GWh yr\u22121 \nfor 2025 (teal triangle) and 2030 (purple diamond), including trajectories. (4) \nFeasibility space for relative coverage of total European demand for 2025 (teal) \nand 2030 (purple), including the 90% self-sufficiency level (black dashed line). \nThe results for 2025 (teal) and 2030 (purple) are shown as histograms (left y-axis, \nno ticks) and cumulative density (right y-axis).\n\nNature Energy\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nExtended Data Fig. 4 | Sensitivity without utilisation constraint. Logistic \ngrowth model (Upper-A), Gompertz diffusion model (Middle-B), and Gompertz \ndiffusion model with limited growth rate (Lower-C) - N\u2009=\u20091,000. (1): BEV battery \ndemand (blue) versus battery production capacity (red) until 2035. Shaded \nareas mark all the potential model results, while the dashed lines indicate the \n5% and 95% percentiles, and the solid line indicates the median. (2) Comparison \nfor domestic production capacity (x-axis) versus demand (y-axis) in GWh yr\u22121 \nfor 2025 (teal triangle) and 2030 (purple diamond), including trajectories. (3) \nFeasibility space for relative coverage of total European demand for 2025 (teal) \nand 2030 (purple), including the 90% self-sufficiency level (black dashed line). \nThe results for 2025 (teal) and 2030 (purple) are shown as histograms (left y-axis, \nno ticks) and cumulative density (right y-axis).\n\nNature Energy\nAnalysis\nhttps://doi.org/10.1038/s41560-025-01722-y\nExtended Data Table 1 | Statistical figures for calculated growth rates and inflection points for demand and production via \nthe Logistic Growth Model (LGM) and Gompertz (GOM)\nParameter \nTarget \nModel \nMedian \nIQR \n5% percentile \n95% percentile \nGrowth Rate \n(% per year) \nDemand \nLGM \n36.2% \n10.3% \n(43.0%-32.8%) \n31.3% \n52.9% \nGOM \n22.1% \n2.6% \n(23.6%-21.1%) \n20.5% \n26.5% \nProduction \nLGM \n61.8% \n5.8% \n(64.7%-58.9%) \n54.9% \n68.0% \nGOM \n37.3% \n5.0% \n(39.8%-34.7%) \n30.8% \n43.0% \nInflection \nPoint \n(year) \n \nDemand \nLGM \n2027.3 \n2.0 \n(2028.4-2026.4) \n2025.6 \n2029.0 \nGOM \n2025.4 \n0.5 \n(2025.6-2025.1) \n2024.6 \n2025.8 \nProduction \nLGM \n2027.6 \n0.5 \n(2027.8-2027.3) \n2027.1 \n2028.2 \nGOM \n2026.4 \n0.4 \n(2026.6-2026.2) \n2027.1 \n2026.0 \nThis involves the median (first column), the interquartile range (IQR, second columns), as well as the 5% (third column) and 95% percentiles (fourth column). Growth rates as percentage \nvalues. Inflection points in years.\n\n\n    Scientific Research Findings:", "answer": "We find that European battery cell demand will likely surpass 1.0 TWh per year by 2030, whereas domestic production capacity is expected to fall short, creating a risk of supply constraints. Although Europe can be expected to meet at least 50\u201360% of its demand through domestic production by 2030, achieving the EU\u2019s 90% self\u2011sufficiency target is feasible but uncertain, as nearly half of our modelled scenarios fail to meet this target. If Europe wants more independence from battery cell imports, our findings highlight the urgency of accelerating production capacity expansion, scaling up a battery supply chain, and implementing strong industrial policies to support competitiveness and supply sovereignty. Our approach is broadly applicable to regions aiming for battery self\u2011sufficiency and should be examined with interacting factors such as policy support and supply chain resilience. However, our analysis does not account for disruptive market shifts, policy reversals, or unexpected technological breakthroughs, which could substantially alter production and demand trajectories.", "id": 0}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 10 | April 2025 | 470\u2013478\n470\nnature energy\nArticle\nhttps://doi.org/10.1038/s41560-025-01724-w\nHow central banks manage climate and \nenergy transition risks\n \nEsther Shears\u2009\n\u200a\u20091\u2009\n, Jonas Meckling\u2009\n\u200a\u20092,3\u2009\n & Jared J. Finnegan\u2009\n\u200a\u20094\nCentral banks have begun to examine and manage climate risks, including \nboth transition risks of moving from fossil fuels to clean energy and physical \nclimate risks. Here we provide a systematic assessment of how and why central \nbanks address climate risks on the basis of an original dataset of central banks \nacross the Organization for Economic Co-operation and Development and \nGroup of 20. We show that central banks vary substantially in the extent to \nwhich they re-risk fossil fuel investments and physical risks and de-risk  \nclean energy investments. Our analysis finds that central bank climate risk  \nmanagement is not associated with a country\u2019s economic exposure to \ntransition risks, but instead with its climate politics. The results suggest that \ncentral banks may not be solely independent risk managers but also actors \nthat respond to political demands. As such, central banks may reinforce \nnational decarbonization policy, while not correcting for the lack thereof.\nEnergy transition and climate change both entail risks for the global \neconomy1. As the global economy decarbonizes, fossil fuel investments \nface stranded asset risks, that is, lost profits owing to early retirement2. \nStranded asset risks threaten financial stability. Similarly, exposure \nto climate hazards contributes to financial stability risk. Meanwhile, \nclean energy investments come with higher capital investment and \ngreater uncertainty about technology and market performance, despite \ndeclining technology costs3. Policy can help mitigate these risks4\u20137.\nOver the last decade, central banks have taken on a role in exam-\nining and managing transition risks as well as physical climate risks8. \nThese risks are not only firm-level risks but can amount to systemic \nrisks. After the financial crisis of 2008\u20132009, central banks have grown \nmore occupied with financial and macroeconomic stability, and finance \nis the transmission belt of transition risks9\u201311. Climate activists have \nwelcomed the expansion of central banks\u2019 activities to facilitate decar-\nbonization, hoping that central banks could substitute for the lack of \nstrong national climate action. Monetary conservatives, instead, have \nbeen alarmed by mission creep among central banks12,13 (B. Bremer and \nJ. Chwieroth, manuscript in preparation). Since first movers, such as \nthe Bank of England, began to explore the issue, central banks across \nthe globe have started to assess and manage climate risks. New global \nfora foster learning and co-operation among central banks, such as the \nNetwork for Greening the Financial System14,15.\nYet the response from central banks has not been uniform, some \nhave adopted measures of varying type and stringency, while other \ncentral banks have not taken any actions16\u201319. This raises the question \nof what explains central bank activity in managing climate risks. We \nconsider two sets of explanations, namely, (1) central banks respond \nto underlying economic risks or (2) central banks react to political \ndemands in addressing risks20.\nHere, we provide a systematic study of central bank management \nof climate risks. We introduce a dataset of climate risk management \nmeasures adopted by central banks across 47 Organization for Eco-\nnomic Co-operation and Development (OECD) and Group of 20 (G20) \ncountries, which represents a large country sample for such research. \nImportantly, we develop a classification system to identify actions that \nre-risk brown investments and de-risk green investments. Re-risking \nrefers to embedding transition risks and physical climate risks into \nfinancial risk management practices to ensure financial stability, \nwhereas de-risking means reducing the risk of clean energy invest-\nments, that is, the technology, market and policy risks of new clean \nenergy technologies, to facilitate decarbonization. Prior research has \nnot differentiated these two key dimensions of central bank activity, \nwhich relate to phasing out fossil fuels and phasing in clean energy.\nWe assess several economic risks and political demands. We find \nlimited evidence that economic risks are associated with central bank \nReceived: 1 March 2024\nAccepted: 24 January 2025\nPublished online: 28 February 2025\n Check for updates\n1Berkeley Economy and Society Initiative, University of California, Berkeley, Berkeley, CA, USA. 2Department of Environmental Science, Policy, and \nManagement, University of California, Berkeley, Berkeley, CA, USA. 3Harvard Business School, Harvard University, Boston, MA, USA. 4Department of \nPolitical Science, University College London, London, UK. \n\u2009e-mail: esther_shears@berkeley.edu; meckling@berkeley.edu\n\nNature Energy | Volume 10 | April 2025 | 470\u2013478\n471\nArticle\nhttps://doi.org/10.1038/s41560-025-01724-w\nthat central banks are rational technocratic agencies that follow their \nmandate to manage economic risks. This includes specifically protect-\ning financial stability23, which decarbonization and climate change can \ndisrupt15. Here, we test economic explanations for both re-risking and \nde-risking and turn towards the political explanation in the next section.\nFirst, we hypothesize that a central bank is more likely to re-risk, \nthe greater the stranded assets risks are in the economy and financial \nsystem it oversees3,24. This follows from central banks\u2019 mandate to \nensure financial stability25,26. How to best measure stranded asset risks \nis an ongoing debate, and a key part of central bank engagement with \ntransition risks is to better understand the type and magnitude of these \nrisks27. We consider two risk dimensions, the extent of fossil fuel assets \nin an economy and the relative size of the financial sector. We assume \nthat the larger the oil and gas sector in an economy, the greater the \nstranded asset risks. Similarly, a large financial sector increases the \ndirect exposure of an economy to stranded asset risks, which could \nthreaten financial stability28. We measure the size of the oil and gas \nsector by calculating the oil and gas sector share of a country\u2019s gross \ndomestic product (GDP) and the size of the financial sector as domestic \ncredit provided by the financial sector as a share of GDP.\nWe use a simple linear regression model to test the correlation of \neconomic risks with countries\u2019 re-risking scores. As we have a diverse set \nof countries in our sample, we are careful to control for factors that could \nbe associated with cross-national differences in central bank activity. \nThey are central bank independence, level of democracy, whether a cen-\ntral bank mandate includes a price stability objective only, an economic \nsupport objective and/or a sustainability objective, GDP per capita, GDP \ngrowth rate, trade share of GDP, unemployment rate, inflation rate and \nEU membership. While our dataset on central bank activities includes \n47 countries, owing to missing data our regression analyses include 41 \ncountries. See Methods and Supplementary Note 6 for further details.\nbehaviour. Among these risks, stranded asset risks and clean energy \ninvestment risks\u2014main transition risks\u2014are not associated with central \nbank actions, only physical risks are, to some extent. Instead, we find \nthat central bank actions to manage risks are significantly associated \nwith domestic climate politics\u2014existing climate policy stringency and \npublic opinion on climate change. Our results suggest that the magni-\ntude of economic risks is not associated with central bank attempts to \ncontain those risks, leaving a risk mitigation gap. Furthermore, they \nindicate that central banks may not be entirely autonomous risk man-\nagers, but rather be responsive to political demands to maintain their \nlegitimacy. Overall, our findings suggest that central banks may rein-\nforce decarbonization policy instead of correcting for the lack thereof.\nMeasuring how central banks manage climate \nrisks\nCentral banks have taken a range of actions to address transition risks \nand physical climate risks, including stress testing requirements, pur-\nchasing green bonds for their own portfolios or requiring climate risk \ndisclosure of the financial institutions they oversee16,17,21,22. We argue \nthat we need to differentiate between actions that re-risk \u2018climate bads\u2019 \n(that is, carbon-intense activities) and that de-risk \u2018climate goods\u2019 (that \nis, activities that support climate change mitigation and adaptation) \nbecause they address different sets of risks and include different policy \nmeasures (Table 1). Prior research has acknowledged that prudential \n(re-risking) and promotional (de-risking) motives exist, while we show \nthat this translates into different central bank actions19.\nRe-risking policies are targeted at adding in climate or carbon risk \nmetrics to central bank supervisory procedures (Extended Data Table 1). \nThey include, for instance, requirements for the disclosure of transition \nand physical climate risks, the inclusion of these risks in stress testing \nand to shift lending away from carbon-intensive projects. Re-risking pol-\nicies can be beneficial to countries that have a large economic exposure \nto stranded asset risks or to climate impacts and thus climate damages. \nDe-risking actions are policies that facilitate low-carbon investments. \nSuch measures include, for example, lower capital requirements for \ngreen projects, requirements for a minimum allocation of lending \ntowards green projects and investments in green bonds.\nWe construct a dataset on central bank policies on climate and \ntransition risks for the OECD and G20 countries, containing policies \nenacted as recently as August 2023 (Methods). We classify these poli-\ncies by function (re-risking, de-risking or both), cost and type of instru-\nment (Table 1, Methods and Extended Data Table 2). We aggregate \nthe policy-level dataset to the country level and calculate composite \nscores of their climate-related activity using both a re-risking score and \na de-risking score (Methods and Supplementary Note 5).\nWe find substantial variation in the extent to which countries \nre-risk, de-risk or do both (Fig. 1). First, there is a group of countries \nwith high re-risking and de-risking scores (Fig. 1, blue quadrant). These \nare mostly member states of the European Central Bank (ECB; Italy, \nGermany, France, the Netherlands and Belgium), the UK and China. A \nsecond cluster of countries with relatively less activity (a score of 10 or \nlower) in both re-risking and de-risking scores, includes the USA, South \nKorea, Costa Rica, South Africa and Russia (Fig. 1, red quadrant). A third \nset of countries clearly engage in more re-risking than de-risking (Brazil, \nSwitzerland and Sweden) (Fig. 1, yellow quadrant). Last, a set of countries \nengage primarily in de-risking (Hungary, Denmark, Japan, India and Indo-\nnesia) (Fig. 1, green quadrant). We discuss observations on trends in the \npolicy instruments that central banks use in Supplementary Note 5. This \nsubstantial cross-national variation raises the question of why central \nbanks vary in the extent to which they re-risk and de-risk transition risks.\nCentral banks and economic risks\nWe focus on two basic explanations for why central banks manage cli-\nmate risks: (1) they seek to address economic risks or (2) they respond \nto political demands for climate action. The first explanation assumes \nTable 1 | Defining re-risking and de-risking\nDefinition\nRisk type\nPolicy examples\nRe-risking\nInternalizing \nrisks of fossil fuel \ninvestments and \nphysical climate \nrisks into financial \nrisk management \npractices to \nensure financial \nstability\nStranded \nasset risks \n(transition \nrisks)\nand physical \nclimate risks\nRe-risking policies tend to \nbe supervisory in nature, \nfor example,\n- Targets added for its \nown investment portfolio \nto achieve net-zero, and \nrestrict investments in \nfossil fuels (Finland)\n- Publishes report on \ncarbon footprint of its \ncorporate bond holdings \n(Sweden)\n- Banks must now \nincorporate climate-related \nrisks into their stress tests \n(Brazil)\nDe-risking\nReducing the risk \nof clean energy \ninvestments\nClean \nenergy risks \n(transition \nrisks)\nDe-risking policies tend \nto incentivize finance for \nclean energy investments, \nfor example,\n- Lower capital \nrequirements for \nenvironmentally \nsustainable corporate \nand municipal lending \n(Hungary)\n- Invests in green bond \ninvestment fund for central \nbanks to increase share of \ngreen securities in its own \nfunds portfolio (EU)\n- Publishes guidance \nfor financial system to \nincentivize more capital \nspending for green \neconomic development \n(China)\n\nNature Energy | Volume 10 | April 2025 | 470\u2013478\n472\nArticle\nhttps://doi.org/10.1038/s41560-025-01724-w\nWe find no statistically significant correlation between the size of \neither the oil and gas sector or the financial sector and re-risking scores. \nThe central banks of economies with high exposure to stranded asset \nrisks through either a large oil and gas sector and/or a large financial \nsector do not appear to engage more in re-risking than economies with \nlow stranded asset risks.\nTo illustrate cross-national variation, we plot the relative size of \nthe financial sector against re-risking scores in Fig. 2. From here on, the \nplots in grey indicate no correlation, whereas plots in red and blue indi-\ncate significant correlation for re-risking and de-risking, respectively. \nFigure 3 plots the coefficient estimates and standard errors of these \ntwo models, with the statistically significant variables highlighted by \ncolour to distinguish the two models. Supplementary Tables 1 and 2 \ncontain the full regression results. We provide an expanded discussion \nof all hypothesized factors in Supplementary Note 3.\nSecond, we hypothesize that a central bank is more likely to de-risk \nif there is a growing green economy29. Central banks often have man-\ndates to support the domestic economy, which may lead them to \nsupport high-growth sectors, thus taking on a promotional role. We \nconsider the growth of renewable energy usage from 2018 to 2021 as \nproxy for a growing green economy because renewable energy tech-\nnologies are the most mature clean energy technologies.\nWe find no statistically significant relationship between the \ngrowth of the share of renewable energy in total primary energy use \nand de-risking scores (Figs. 3 and 4). This means that a growing renew-\nable energy industry is not associated with central banks reducing the \nrisks of clean energy investments.\nThird, we expect central banks to engage in re-risking if their \neconomies are highly exposed to physical climate hazards, such as \nstorms, droughts and wildfires. A core part of central bank re-risking \n0\n5\n10\n15\n20\n25\n30\n35\n40\nDe-risking  score\n0\n5\n10\n15\n20\n25\n30\n35\n40\nRe-risking  score\nZA\nUS\nSI\nSE\nNZ\nNO\nNL\nMX\nLT\nJP\nIT\nIN\nIE\nHU\nEU\nCR\nCH\nCA\nAT\nAR\nSK\nSA\nLU\nID\nGB\nFR\nFI\nES\nDK\nDE\nCN\nBR\nBE\nPT\nEE\nGR\nLV\nRU\nCO\nCZ\nIL\nIS\nKR\nTR\nPL\nCL\nAU\nFig. 1 | Re-risking and de-risking scores by country. A graph plotting each \ncountry\u2019s calculated re-risking and de-risking scores. Scores higher than 10 \nindicate that the country engages in substantial activity in that policy group, \nwhile scores of 10 or lower indicate rather marginal efforts. The two-digit \nInternational Organization for Standardization country codes indicate the \ncountry names. There are four clusters evident from this plot: countries that \nengage substantially in both re-risking and de-risking (blue quadrant), countries \nthat mostly re-risk (yellow quadrant), countries that mostly de-risk (green \nquadrant) and countries that engage marginally in both or either group (red \nquadrant).\n10\n20\n30\n40\n50\n60\n70\n80\n90\n100\n110\n120\n130\n140\n150\n160\n170\n180\n190\n200\nDomestic credit provided to private sector (% of GDP)\n0\n5\n10\n15\n20\n25\n30\n35\n40\nRe-risking score\nTR\nSI\nSA\nPT\nPL\nNZ\nNO\nMX\nLT\nKR\nIT\nIN\nIL\nIE\nGR\nEU\nEE\nDE\nCZ\nCR\nAT\nZA\nUS\nSK\nSE\nNL\nLV\nLU\nJP\nIS\nID\nHU\nGB\nFR\nFI\nES\nDK\nCN\nCL\nCH\nCA\nBR\nBE\nAU\nAR\nRU\nCO\nFig. 2 | Stranded asset risks and re-risking. The central bank re-risking activity against a proxy for relative size of a country\u2019s financial sector, the domestic credit \nprovided to the private sector, expressed as a percentage share of the country\u2019s GDP.\n\nNature Energy | Volume 10 | April 2025 | 470\u2013478\n473\nArticle\nhttps://doi.org/10.1038/s41560-025-01724-w\nis supervisory, that is, understanding the magnitude of these risks and \nincorporating them into financial risk management practices, such \nas stress tests. It follows that if a country is highly exposed to physical \nclimate risks, it would adopt these practices. In fact, prior research has \ndemonstrated a correlation between physical risks and central bank \nmanagement of climate risks18. We measure climate hazard exposure \nby using the exposure component of the Notre Dame Global Adaptation \nInitiative (ND-Gain) country index.\nWe do not find an association between exposure to climate hazards \nand re-risking. Instead, we find that de-risking is positively and signifi-\ncantly correlated with higher exposure to physical climate risks. This is \npuzzling and requires further analysis. One would expect that central \nbanks of economies with high exposure to climate hazards engage \nprimarily in re-risking. To illustrate cross-national variation, we plot \nthe ND-Gain exposure index (our measure for physical climate risk) \nagainst de-risking scores in Supplementary Fig. 1.\nIn sum, we find that only physical climate risks are significantly \ncorrelated with central bank de-risking activities. Transition risks\u2014\nstranded asset and clean energy investment risks\u2014are associated with \nneither re-risking actions nor de-risking actions.\nCentral banks and climate politics\nA second set of explanations for why central banks tackle climate \nrisks lies in politics30,31. Central banks could be responding to politi-\ncal demands from either policymakers and/or the public32. While cen-\ntral banks tend to have high degrees of autonomy, they are ultimately \naccountable to politicians and the public they serve33,34. Central banks \nhave a record of responding to policymaker and public pressures, includ-\ning in ways that extend beyond the scope of their primary mandate \nof price stability\u2014specifically since the 2008 financial crisis35\u201337. This \nsuggests that central banks may be increasingly sensitive to political \nforces and act strategically, in their interests of self-preservation and \nmaintaining their legitimacy38. Legitimacy in the eye of the public is a key \nstrategy to ensure their autonomy vis-\u00e0-vis politicians39\u201342. By appealing \nto issues with public salience (in this case, climate change), central banks \nmay strengthen their public legitimacy, but only insofar that engage-\nment with matters of public concern does not overshadow their primary \nmandate of ensuring price stability. We test two sources of political \ndemands\u2014policymakers and the public\u2014separately, complementing \nprior research that focuses on supply-side variables, that is, features \nof central banks, such as their mandates and level of independence16,18.\nFirst, we use the stringency of a country\u2019s climate policy as proxy \nfor policymaker demands. If policymakers\u2014meaning politicians  \nand/or bureaucrats\u2014have enacted stringent national climate policy, \nthey may expect central banks to follow suit and to support their policy \ngoals35\u201337,43,44. Absent direct policymaker influence on central banks, this \nrelationship might also exist if a central bank has a mandate to support \nthe domestic economy or support the government\u2019s economic agenda, \nas in China or the USA.\nWe find that climate policy stringency (based on an OECD index) \nis a statistically significant and positive factor for re-risking. In other \nwords, the more policymakers adopt strong climate policies, the more \nlikely central banks are to engage in re-risking. We demonstrate the \nrelationship in Fig. 5. In contrast, climate policy stringency is not cor-\nrelated with de-risking. One potential reason may be that stringent \nclimate policy often performs a de-risking function by, for instance, \nproviding clean energy subsidies and tax credits. In which case, central \nbanks do not need to step in.\nSecond, we assume that the more the public is concerned with \nclimate change, the more central banks will adopt re-risking and \nde-risking policies. In countries where climate change is a salient politi-\ncal issue and a large share of the population are concerned about the \nimpacts of climate change, there have been public calls for central \nbanks to act directly to address it45. Supportive public sentiment might \nbe a necessary condition for central banks to engage specifically in \nde-risking policies because such policies have already fallen under \nscrutiny for violating market neutrality46.\nUsing a cross-national survey on public concern about climate \nchange from the Yale Programme on Climate Change Communica-\ntion (YPCCC), we find a statistically significant positive correlation \nbetween public concern and de-risking scores (Fig. 3 and Supplemen-\ntary Table 2). This means that central banks may respond to greater \npublic concern with de-risking the clean energy transition, but not \nwith re-risking stranded assets and physical climate risks. We plot \nde-risking and public concern for climate change in Fig. 6 to illustrate \ncross-national variation in public concern and de-risking actions.\nIn sum, we find that political demands from policymakers and the \npublic are significantly and positively correlated with central banks\u2019 \nre-risking and de-risking activities, respectively.\nDiscussion\nCentral banks vary substantially in the extent to which they re-risk fossil \nfuel investments and de-risk clean energy investments. The surprising \nfinding here is that this is not significantly associated with exposure to \ntransition risks, such as stranded asset and clean energy risks. Instead, \nwe find that climate policy stringency and public concern with cli-\nmate change are associated with re-risking and de-risking activities, \nrespectively.\nOur study has limitations that future research should address. \nFirst, we provide a cross-sectional analysis of central bank actions that \nemerged only recently. As time passes, researchers need to develop \ntime series to understand the evolution of central bank climate risk \nEU member\nInflation rate\nUnemployment rate\nTrade share of GDP\nGDP growth\nGDP per capita\nMandate\u2014price stability only\nMandate\u2014support economy\nMandate\u2014sustainability\nDemocracy index\nCentral bank independence\nPublic concern for climate change\nClimate policy stringency\nPhysical climate risk\nRenewable energy growth\nSize of financial sector\nFossil fuel share of GDP\n\u201310\n0\n10\n20\n30\nEstimate\nModel\nRe-risking\nDe-risking\nFig. 3 | Estimate plot of regression models. The coefficient estimates from \nthe two linear regression models supporting our main findings. The re-risking \nmodel is shown in red and the de-risking model is shown in blue. The red circles \nor blue squares indicate the coefficient estimate values, with the lines on either \nside of the observation showing the 95% confidence intervals. The highlighted \nboxed variables indicate a statistically significant variable in the model, colour \ncoded to indicate which model (red for re-risking, blue for de-risking). Next to \nthe findings discussed in the main text, we find that GDP growth rate is a negative \nstatistically significant factor for re-risking, and that being a member of the EU \nis a statistically significant factor for de-risking. We discuss these findings in \nSupplementary Notes 1\u20134.\n\nNature Energy | Volume 10 | April 2025 | 470\u2013478\n474\nArticle\nhttps://doi.org/10.1038/s41560-025-01724-w\nmanagement, thus also increasing the number of observations for \nstatistical analysis. We are also at an early stage of understanding tran-\nsition risks. Developing transition risk indices would offer additional \nconceptualizations of risk beyond an economy\u2019s exposure to affected \nsectors as assessed in this article.\nSecond, we present cross-national correlations and theorize \nunderlying mechanisms. Future research could shed light on central \nbank decision-making in tackling transition and climate risks using \nin-depth case studies or surveys on central bank decision making. \nThis will help better understand how central bank bureaucrats think \nabout both economic risks and political demands in incorporating \nenergy and climate-related risks into their activities. In addition, as \nmore policies are enacted, research on the effectiveness of different \npolicy instruments could help shed light onto which actions have the \ngreatest impact on climate risk mitigation (Supplementary Note 5).\nThird, central banks are the primary actors in managing financial \nrisk and stability but not the only ones. To better assess the magnitude \nof the risk mitigation gap, research needs to examine the broader \nregulatory ecosystem, including financial supervisors and private \nsector actors. For example, in the USA, the Securities and Exchange \nCommission and the Commodity Futures Trading Commission are \nsetting rules on climate risk disclosure, not the Federal Reserve47,48. \nAdditionally, private actors, such as credit rating agencies, account-\ning firms and insurance companies, are working to develop their own \nclimate risk management practices. New coalitions, such as the Value \nBalancing Alliance, advance standards setting in impact accounting \nand climate-related disclosures. Yet these are all voluntary forms of \nrisk management. Understanding the broader regulatory system of \nrisk will allow for an assessment of the regulatory gap for the economy \nas a whole.\n\u201320\n\u201310\n0\n10\n20\n30\n40\n50\n60\n70\n80\n90\n100\nChange of renewable energy share of total energy (2018\u20132021) (%)\n0\n5\n10\n15\n20\n25\n30\n35\n40\nDe-risking score\nUS\nSK\nSI\nSE\nRU\nPT\nNZ\nNO\nMX\nLV\nLU\nLT\nJP\nIT\nIS\nIN\nIL\nIE\nID\nGR\nGB\nFR\nFI\nEU\nES\nDK\nDE\nCZ\nCR\nCO\nCN\nCL\nCA\nBR\nBE\nAU\nAT\nAR\nPL\nNL\nHU\nEE\nKR\nCH\nZA\nTR\n2021 renew shr\n0\n20.00\n40.00\n60.00\n80.00\n100.00\nFig. 4 | Renewable energy growth and de-risking. The central bank de-risking \nactivity against the growth of a country\u2019s renewable energy sector, specifically \nthe percentage change from 2018 to 2021 of a country\u2019s renewable energy share \nof the country\u2019s total primary energy supply. The size of the observation point \nrepresents the 2021 value of the renewable energy share of total primary energy \nsupply for each country (2021 renew shr), with the smaller circles indicating a \nsmaller share and the larger circles indicating a larger share.\n4.2\n4.5\n4.8\n5.1\n5.4\n5.7\n6.0\n6.3\n6.6\n6.9\n7.2\n7.5\nClimate policy stringency (0\u201310 index)\n0\n5\n10\n15\n20\n25\n30\n35\n40\nRe-risking score\nZA\nTR\nSK\nSI\nSE\nSA\nRU\nPT\nPL\nNZ\nMX\nLV\nLU\nLT\nKR\nJP\nIT\nIL\nIE\nHU\nGR\nGB\nES\nEE\nDK\nDE\nCZ\nCO\nCN\nCL\nBE\nAU\nAT\nUS\nNO\nNL\nID\nFR\nFI\nEU\nCR\nCH\nCA\nBR\nAR\nIN\nIS\nFig. 5 | Climate policy stringency and re-risking. Re-risking plotted against the Climate Policy Stringency index, an OECD index of adopted climate policies with a \nrange from 0 to 10, with 10 being most stringent. The observations are coloured red to indicate a significant association with re-risking.\n\nNature Energy | Volume 10 | April 2025 | 470\u2013478\n475\nArticle\nhttps://doi.org/10.1038/s41560-025-01724-w\nOur findings also have an important implication for policy. Climate \npolicy advocates have held hopes that central banks could correct for \nthe lack of national decarbonization policy. Their relative independ-\nence could theoretically provide them with greater ability to adopt \npolicies where executives or legislatures are paralysed by political \nopposition. This would most probably be the case for re-risking fossil \nfuel investments, which tends to provoke opposition from fossil fuel \nand related industries. Yet we find re-risking to be positively correlated \nwith the stringency of national climate policy. This suggests that, so \nfar, central banks complement, rather than act as a substitute for, \nregulatory or legislative policy to reduce fossil fuel dependence. This \nmakes central banks an important additional actor in decarbonization \npolicy. That said, it also cautions against high hopes for central banks \nin laggard countries to take the lead in tackling climate risks. Research \nhas found a similar pattern for transnational climate governance by \nprivate actors\u2014the more stringent national climate policy, the more \nsubnational and private actors participate in transnational climate \ngovernance49. In short, the halo effect of national climate policy and \npolitics looms large and probably shapes a country\u2019s overall decar-\nbonization ambition, including efforts by independent actors, such \nas central banks.\nThe political nature of the management of climate risks raises \nconcerns about unmanaged risks in the global economy, specifically \nstranded asset risks. These exist in economies with large oil and gas \nand/or financial sectors and low re-risking scores. We see two practi-\ncal paths to begin to address the risk mitigation gap. First, increasing \ntransparency of central bank actions on climate risk management \nwould lay the ground for building political pressure on laggard central \nbanks. This article has taken a first step towards that, while a more \ninstitutionalized effort would be to develop a central bank climate \nindex. Such an index would identify leaders and laggards and show \nprogress over time. It would provide the informational basis for a \nnumber of advocacy and market actors to build pressure on central \nbanks. For instance, private credit rating agencies may use the index \nto inform country credit ratings, thus indirectly incentivizing central \nbank actions to address climate risks.\nSecond, international organizations, such as the Bank for Interna-\ntional Settlements (BIS) or the Financial Stability Board, could move \nbeyond identifying best practices and develop standards for climate \nrisk management that members need to adhere to. Both the BIS and \nthe Financial Stability Board have begun to explore climate risks. For \nexample, the BIS has examined climate risk disclosure requirements, \ndiscussing which requirements should be standard for members and \nwhich should be at national discretion. International risk disclosure \nrequirements for central banks could potentially incentivize greater \nrisk management by central banks in countries that lack strong national \ndecarbonization policies.\nMethods\nOverview\nThis study undertakes regression analyses of an original dataset. This \nsection first describes how we created the dataset, including the sample \nof countries, data collection, dataset aggregation and outcome vari-\nable calculation. Then, it details data collection for all covariates and \ndescribes the linear regression model.\nCountries included in the dataset\nIn total, 47 countries are included in our dataset. They include the \nOECD countries plus all other countries in the G20. The countries are \nArgentina, Australia, Austria, Belgium, Brazil, Canada, Chile, China, \nColombia, Costa Rica, Czechia, Denmark, Estonia, European Union, \nFinland, France, Germany, Greece, Hungary, Iceland, India, Indone-\nsia, Ireland, Israel, Italy, Japan, Republic of Korea, Latvia, Lithuania, \nLuxembourg, Mexico, the Netherlands, New Zealand, Norway, Poland, \nPortugal, Russia, Saudi Arabia, Slovakia, Slovenia, South Africa, Spain, \nSweden, Switzerland, Turkey, the UK and the USA. As of 2022, OECD \nand G20 countries together constitute more than 84% of global GDP \nand 83% of global CO2 emissions50,51.\nData collection on central bank actions\nThe starting point of the data collection on climate risk-related central \nbank actions was the Green Monetary and Financial Policies Tracker \ncreated by the E-Axes Forum in 202152. This tracker covers G20 coun-\ntries, EU countries and some countries in Latin America. This dataset \nincludes the country, the implementing institution (whether it was the \ncentral bank or other financial regulator), the year of the policy, policy \ndescription, webpage link to the source and a policy type classification. \nThe policy type classification focused on whether the policy was a \nmonetary policy, financial policy or other, and had subclassifications \nbased on these three categories, to further identify the policy group \n(for example, collateral policy, credit operations, asset purchases, \nsupervisory guidelines, stress tests, surveys and so on).\n40\n45\n50\n55\n60\n65\n70\n75\n80\n85\n90\nShare of population concerned about climate change (%)\n0\n5\n10\n15\n20\n25\n30\n35\n40\nDe-risking score\nUS\nTR\nSK\nPL\nNZ\nJP\nIT\nGR\nCR\nCN\nCH\nCA\nBR\nAU\nAT\nAR\nZA\nSE\nSA\nPT\nNO\nNL\nMX\nLT\nKR\nIN\nIL\nIE\nID\nHU\nGB\nFR\nFI\nEU\nES\nDK\nDE\nCZ\nCO\nCL\nBE\nFig. 6 | Public concern about climate change and de-risking. The relationship between de-risking activity and a public sentiment measure from the YPCCC,  \nthe percentage share of the country\u2019s population that is concerned about climate change. The observations are coloured blue to indicate significant association  \nwith de-risking.\n\nNature Energy | Volume 10 | April 2025 | 470\u2013478\n476\nArticle\nhttps://doi.org/10.1038/s41560-025-01724-w\nWhile this tracker provided useful information, we wanted to \nexpand the dataset to include more recent policies (post-2021), a more \nexpansive country sample and create a classification system of policies \nto differentiate between re-risking and de-risking goals, so we con-\nstructed an original dataset from scratch. In the summer of 2023, we \nwent through the sample countries one by one to collect our own set of \ncentral bank climate-related actions. Via web searching on the central \nbank\u2019s website and national news sources, we searched for specific \nkeywords to identify any possible climate-related policy enacted by the \ncentral bank. These keywords were climate change, climate finance, cli-\nmate risk, sustainable, sustainable finance and green asset. Our search \nusually yielded more exhaustive findings that might not fit the scope of \nthe dataset (for example, internal sustainability efforts the central bank \nis doing to make their office buildings more energy efficient), but this \nway we felt confident our dataset was not missing any key policies. We \nalso cross-checked our policy dataset with the E-Axes Forum Tracker.\nOur policy-specific dataset includes 168 observations across the \n47 countries and identifies the country; the year of the policy (91% of \nthe policy observations are between 2019 and 2023); the name of the \npolicy; the policy description; the data source (webpage link); whether \nthe policy includes a requirement of other actors (for example, a regu-\nlatory rule for banks to comply with, or a request for information) or \nonly pertains to the central bank operations (for example, information \npublished by the bank or a change in how the bank manages its funds); \nwhether the policy aims to re-risk, de-risk or both; the policy function \nclass (information, economic or structural); and the relative \u2018cost\u2019 of \nthe policy to the central bank (the relative degree to which the action \nis costly to the central bank to enact, as interpreted by what is going \nto have a material impact on the financial system and what might have \na resource cost) (see Supplementary Note 5 for an expanded discus-\nsion of this cost concept). To review some examples of these specific \npolicies and the policy classification systems, please see Table 1 and \nExtended Data Table 1.\nDataset aggregation and dependent variable creation\nTo conduct the country-level analysis of risk management activities \nby central banks, we aggregated the policy-level dataset by country \nto create aggregate re-risking and de-risking scores (Extended Data \nTable 2). We assigned each policy observation a point value on the basis \nof its relative cost for central banks and/or complying entities to weight \nhigher cost actions more heavily. Low-cost actions have a point value of \n1, medium-cost actions have a point value of 2, high-cost actions have a \npoint value of 5 and very high-cost actions have a point value of 10. The \nrationale behind this weight (by simply tallying up observations) is that \na central bank might engage in several low- or medium-cost actions, \nsuch as publishing its own climate risk management guidance or even \nsetting up an internal climate change working group, but these actions \ndo not have immediate economic effects on the country\u2019s financial \nsector. An expanded and detailed discussion of this data aggregation \nprocess and the underlying rationale for this approach is provided in \nSupplementary Note 5.\nWe then aggregated the policies\u2019 point values on the basis \nof whether the policy is for re-risking or de-risking. In the final \ncountry-level aggregated dataset, each country has a re-risking score, a \nde-risking score and a total score (re-risking\u2009+\u2009de-risking). Supplemen-\ntary Fig. 2 maps out the total score for each country in the dataset. For \nthe European countries that are under the jurisdiction of the ECB, we \nadded the EU scores to the national scores. This then captures both the \npolicies of the national central bank of the European country as well as \nwhat the ECB implemented.\nWe also calculated alternative versions of the scores to perform \nrobustness checks. Instead of a (1, 2, 5, 10) scale for the observation \nweights (based on the policy cost), the alternate versions of these scores \napply a different weight scale: (1, 2, 3, 4), (1, 3, 5, 7) and (1, 5, 10, 15) (see \nSupplementary Note 10 for these analyses).\nData collection on explanatory variables\nOn the basis of the different initial hypotheses, we collected a broad \nset of possible explanatory variables at the country level. To capture \nstranded asset risks, we used both a measure of the size of a financial \nsector and the approximate carbon exposure of the country\u2019s GDP. \nFirst, we used the World Bank DataBank\u2019s domestic credit provided \nto the private sector measure, as a share of GDP, from the most recent \nfull coverage year in 2019 as a proxy for financial sector size53. This \nmeasure is a commonly used proxy for financial depth, which captures \nthe financial sector\u2019s size relative to the domestic economy54. We also \ncalculated the economic value contributing to a country\u2019s GDP from \nthe oil and gas sector for the year 2021, using the Global Resource \nInput\u2013Output Assessment model, a multiregional input\u2013output \ndatabase that captures input and intermediate goods trade across \nthe world55. This dataset includes not just economic value produced \nand traded, but also value added by each economy into the value of \nfinal goods. This measure best captures the economic value the oil \nand gas sector contributes to an economy. For data on the growth of \nthe renewable energy sector across countries, we used yearly data \nfrom the International Energy Agency on renewable energy\u2019s per-\ncentage share of primary energy supply and calculated the growth \nrate from 2018 to 202156. We used the 2022 exposure component \nof the ND-Gain country index to capture a country\u2019s physical risk \nof climate change impacts57. To measure climate policy stringency, \nwe used the OECD\u2019s Climate Policy Stringency of adopted policies \nindex for the most recent year available, 202058. To capture public \nsentiment about climate change, we used data from the YPCCC\u2019s \nsurvey, conducted in 2022, that produced population breakdowns \nof sentiment on climate change, by percentage59. We grouped the \nalarmed and concerned audiences together to capture the share of \nthe population concerned about climate change. This survey is the \nmost comprehensive cross-country sampling of public sentiment \nabout climate change so far, but it still is missing a handful of coun-\ntries from our dataset, most notably China and Russia. We substitute \nin a data point for China from a comparable 2022 survey from the \nInternational Monetary Fund, which includes the share of the popu-\nlation that feels that climate change will affect them or their family \nnow up through the next 10\u2009years60. We discuss this process more in \nSupplementary Note 6.\nFor the set of control variables, we used the World Bank Data-\nBank\u2019s 2022 values for a country\u2019s inflation rate, unemployment rate, \nGDP growth rate, GDP per capita and trade share (as a percentage of \nGDP)61. We sourced a 2022 democracy index from The Economist Intel-\nligence Unit, and a central bank independence index from the Quality \nof Government Institute, a dataset published by Ana Carolina Garriga in \n201662. We also created a secondary database on our sample countries\u2019 \ncentral bank mandates. The data on mandates were collected from the \nwebsites of each central bank. We gathered direct text from the central \nbank, summarized the core objectives and verified these mandates \nusing secondary sources. From these mandates, we created three \nbinary indicators for the controls: first, whether the mandate is for price \nstability only; second, whether the mandate includes a clause (can be \nsecondary in objective) to support the domestic economic policies of \nthe country; and third, whether the mandate includes any text explicitly \nthat points to sustainability objectives. For any country observation for \nthe control variables where the data point was not available given the \nyear we use, we used the data value for the most recent year available, \nto ensure the completeness of the dataset and not unnecessarily drop \ncountries from the regression analysis (specific observations are noted \nin Supplementary Note 6).\nSupplementary Note 6 presents expanded discussion of our data-\nset. We collected alternate data sources for these explanatory and \ncontrol variables as well (see Supplementary Table 4 for a discussion \nof these alternate sources and why we selected these specific variables \nfor this analysis).\n\nNature Energy | Volume 10 | April 2025 | 470\u2013478\n477\nArticle\nhttps://doi.org/10.1038/s41560-025-01724-w\nRegression analysis\nWe conducted the regression analysis in this study using R. We created \na linear regression model with robust standard errors, using the main \ndependent variables and the above-described explanatory variables \nand controls. In the regression analysis, our sample includes 41 obser-\nvations. From the 47 countries in our dataset, the countries excluded \nfrom the regression analysis owing to missing observations are Estonia, \nIceland, Latvia, Luxembourg, Russia and Slovenia. The missing data \nare owing to the variable for public sentiment about climate change, \ndiscussed above and in Supplementary Note 6. The regression sample \nof countries together constitute more than 79% of global GDP and 81% \nof global CO2 emissions50,51.\nOur use of ordinary least squares is justified in Supplementary \nNote 7. To test the robustness of our main findings, we carried out a \nnumber of checks. We tested for multicollinearity in our main model, \nran our main specification using alternate versions of the dependent \nvariable and ran an alternate model with a different measure of renew-\nable energy growth (see Supplementary Notes 8\u201311 for the results of \nthese robustness checks.)\nData availability\nWe created two datasets for this analysis. The first is a policy-specific \ndataset that includes the individual re-risking and de-risking pol-\nicy measures for all central banks in our dataset. The second is a \ncountry-level dataset that aggregates the policy observations into \naggregate re-risking and de-risking scores. This dataset also includes \nthe independent variables for our regression analysis. Both datasets \nare described in Methods and in Extended Data Tables 1 and 2. We cur-\nrently do not make the datasets publicly available owing to additional \nongoing analysis by the authors, but we make them available upon \nreasonable request.\nCode availability\nWe will provide the code for the regression analysis upon request.\nReferences\n1.\t\nIPCC Climate Change 2022: Mitigation of Climate Change (eds \nShukla, P. R. et al.) (Cambridge Univ. Press, 2022).\n2.\t\nBolton, P., Despres, M., Pereira da Silva, L. A. & Svartzman, R. The \nGreen Swan: Central Banking and Financial Stability in the Age of \nClimate Change (BIS, 2020).\n3.\t\nFinancial headwinds for renewables investors: what\u2019s the way  \nforward? IEA https://www.iea.org/commentaries/financial- \nheadwinds-for-renewables-investors-what-s-the-way-forward \n(2023).\n4.\t\nGeels, F. W. Socio-technical transitions to sustainability: a review \nof criticisms and elaborations of the multi-level perspective. Curr. \nOpin. Environ. Sustain. 39, 187\u2013201 (2019).\n5.\t\nBhandary, R. R., Gallagher, K. S. & Zhang, F. Climate finance policy in \npractice: a review of the evidence. Clim. Policy 21, 529\u2013545 (2021).\n6.\t\nBattiston, S., Monasterolo, I., Riahi, K. & van Ruijven, B. J. \nAccounting for finance is key for climate mitigation pathways. \nScience 372, 918\u2013920 (2021).\n7.\t\nGabor, D. The (European) derisking state. State Mercato 43, 53\u201384 \n(2023).\n8.\t\nOman, W., Salin, M. & Svartzman, R. Three tales of central banking \nand financial supervision for the ecological transition. Wiley \nInterdiscip. Rev. Clim. Change 15, e876 (2024).\n9.\t\nSvartzman, R., Bolton, P., Despres, M., Pereira Da Silva, L. A. & \nSamama, F. Central banks, financial stability and policy coordination \nin the age of climate uncertainty: a three-layered analytical and \noperational framework. Clim. Policy 21, 563\u2013580 (2021).\n10.\t Gruenewald, S., Knijp, G., Schoenmaker, D. & Van Tilburg, R. \nEmbracing the brave new world: a response to Demekas and \nGrippa. J. Financ. Reg. 10, 127\u2013134 (2024).\n11.\t\nBabic, M. Green finance in the global energy transition: actors, \ninstruments, and politics. Energy Res. Soc. Sci. 111, 103482 (2024).\n12.\t Thiemann, M., B\u00fcttner, T. & Kessler, O. Beyond market neutrality? \nCentral banks and the problem of climate change. Finance Soc. 9, \n14\u201334 (2023).\n13.\t Durrani, A., Rosmin, M. & Volz, U. The role of central banks in \nscaling up sustainable finance\u2014what do monetary authorities in \nthe Asia\u2013Pacific region think? J. Sustain. Finance Invest. 10, 92\u2013112 \n(2020).\n14.\t Quorning, S. The \u2018climate shift\u2019 in central banks: how field \narbitrageurs paved the way for climate stress testing. Rev. Int. \nPolitical Econ. 31, 74\u201396 (2023).\n15.\t Campiglio, E. et al. Climate change challenges for central banks \nand financial regulators. Nat. Clim. Change 8, 462\u2013468 (2018).\n16.\t Dikau, S. & Volz, U. Central bank mandates, sustainability \nobjectives and the promotion of green finance. Ecol. Econ. 184, \n107022 (2021).\n17.\t D\u2019Orazio, P. & Popoyan, L. Fostering green investments \nand tackling climate-related financial risks: which role for \nmacroprudential policies? Ecol. Econ. 160, 25\u201337 (2019).\n18.\t D\u2019Orazio, P. & Popoyan, L. Do monetary policy mandates and \nfinancial stability governance structures matter for the adoption of \nclimate-related financial policies? Int. Econ. 173, 284\u2013295 (2023).\n19.\t Baer, M., Campiglio, E. & Deyris, J. It takes two to dance: \ninstitutional dynamics and climate-related financial policies. Ecol. \nEcon. 190, 107210 (2021).\n20.\t DiLeo, M., Rudebusch, G. D. & van\u2019t Klooster, J. Why the Fed and \nECB Parted Ways on Climate Change: the Politics of Divergence \nin the Global Central Banking Community, Working Paper No. 88 \n(Hutchins Center, 2023).\n21.\t Basel Committee on Banking Supervision. Principles for the \nEffective Management and Supervision of Climate-Related \nFinancial Risks (BIS, 2022).\n22.\t McConnell, A., Yanovski, B. & Lessmann, K. Central bank \ncollateral as a green monetary policy instrument. Clim. Policy 22, \n339\u2013355 (2022).\n23.\t Goodhart, L. M. Brave new world? Macro-prudential policy and \nthe new political economy of the federal reserve. Rev. Int. Political \nEcon. 22, 280\u2013310 (2015).\n24.\t Caldecott, B. et al. Stranded Assets: a Climate Risk Challenge (IDB, \n2016).\n25.\t Smets, F. Financial stability and monetary policy: how closely \ninterlinked?. Int. J. Cent. Bank. 10, 263\u2013300 (2014).\n26.\t Central Bank Governance and Financial Stability: a Report by a \nStudy Group (BIS, 2011).\n27.\t von Dulong, A., Gard-Murray, A., Hagen, A., Jaakkola, N. & Sen, \nS. Stranded assets: research gaps and implications for climate \npolicy. Rev. Environ. Econ. Policy 17, 161\u2013169 (2023).\n28.\t Battiston, S., Mandel, A., Monasterolo, I., Sch\u00fctze, F. & Visentin, G. \nA climate stress-test of the financial system. Nat. Clim. Change 7, \n283\u2013288 (2017).\n29.\t Blinder, A. S. Central Banking in Theory and Practice (MIT, 1999).\n30.\t Young, K. Policy takers or policy makers? The lobbying of global \nbanking regulators. Bus. Horiz. 56, 691\u2013701 (2013).\n31.\t Jabko, N. & Kupzok, N. Indirect responsiveness and green central \nbanking. J. Eur. Public Policy 31, 1026\u20131050 (2024).\n32.\t Qanas, J. & Sawyer, M. \u2018Independence\u2019 of central banks and the \npolitical economy of monetary policy. Rev. Political Econ. 36, \n565\u2013580 (2023).\n33.\t Bordo, M. D. A Brief History of Central Banks (Federal Reserve Bank \nof Cleveland, 2007).\n34.\t Tooze, A. Debating Central Bank Mandates, Working Paper No.  \n01-2022 (Forum for a New Economy, 2022).\n35.\t Tooze, A. The Death of the Central Bank Myth (Foreign Policy, \n2020).\n\nNature Energy | Volume 10 | April 2025 | 470\u2013478\n478\nArticle\nhttps://doi.org/10.1038/s41560-025-01724-w\n36.\t \u00d6zg\u00f6de, O. The emergence of systemic risk: the Federal Reserve, \nbailouts, and monetary government at the limits. Socio-Econ. Rev. \n20, 2041\u20132071 (2022).\n37.\t van\u2019t Klooster, J. & Fontan, C. The myth of market neutrality: a \ncomparative study of the European Central Bank\u2019s and the Swiss \nNational Bank\u2019s corporate security purchases. New Political Econ. \n25, 865\u2013879 (2020).\n38.\t Blyth, M. Structures do not come with an instruction sheet: \ninterests, ideas, and progress in political science. Persp. Pol. 1, \n695\u2013706 (2003).\n39.\t Hielscher, K. & Markwardt, G. The role of political institutions for \nthe effectiveness of central bank independence. Eur. J. Political \nEcon. 28, 286\u2013301 (2012).\n40.\t Fern\u00e1ndez-Albertos, J. The politics of central bank independence. \nAnnu. Rev. Polit. Sci. 18, 217\u2013237 (2015).\n41.\t Macchiarelli, C., Monti, M., Wiesner, C. & Diessner, S. The \nEuropean Central Bank between the Financial Crisis and Populisms \n(Palgrave Macmillan, 2020).\n42.\t Lockwood, E. The Global Politics of Central Banking: a View from \nPolitical Science (Cornell Univ., 2016).\n43.\t Adolph, C. Bankers, Bureaucrats, and Central Bank Politics: The \nMyth of Neutrality (Cambridge Univ. Press, 2013).\n44.\t Nordhaus, W. D. The political business cycle. Rev. Econ. Stud. 42, \n169\u2013190 (1975).\n45.\t Tooze, A. A Decade after the World Bailed out Finance, It\u2019s Time for \nFinance to Bail out the World (Foreign Policy, 2019).\n46.\t Papoutsi, M., Piazzesi, M. & Schneider, M. How Unconventional Is \nGreen Monetary Policy? (International Monetary Fund, 2022).\n47.\t Behnam, R. & Litterman, B. Managing Climate Risk in the US \nFinancial System (Commodity Futures Trading Commission, \n2020).\n48.\t The Enhancement and Standardization of Climate-Related \nDisclosures for Investors, Federal Register, Vol. 89 (Securities and \nExchange Commission, 2024).\n49.\t Andonova, L. B., Hale, T. N. & Roger, C. B. National policy and \ntransnational governance of climate change: substitutes or \ncomplements? Int. Stud. Q. 61, 253\u2013268 (2017).\n50.\t GDP (current US$) The World Bank Open Data https://data.\nworldbank.org/indicator/NY.GDP.MKTP.CD (2024).\n51.\t EDGAR\u2014The Emissions Database for Global Atmospheric Research \n(European Commission, 2024); https://edgar.jrc.ec.europa.eu/\nreport_2023#data_download\n52.\t Green monetary and financial policies (GMFP) tracker. E-axes \nForum https://e-axes.org/tableau_iframe/policy-visualization/ \n(2021).\n53.\t Renewable energy. OECD Data https://data.oecd.org/energy/\nrenewable-energy.htm (2022).\n54.\t Domestic Credit to Private Sector (% of GDP) (The World Bank \nGroup, 2023); https://data.worldbank.org/indicator/FS.AST.PRVT.\nGD.ZS\n55.\t Financial Depth. Global Financial Development Report (The World \nBank Group, 2016); https://www.worldbank.org/en/publication/\ngfdr/gfdr-2016/background/financial-depth\n56.\t Global Resource Input\u2013Output Assessment. GLORIA Model \n(Industrial Ecology Virtual Laboratory, 2023); https://ielab.info/\nlabs/ielab-gloria\n57.\t Notre Dame global adaptation initiative (ND-Gain). Exposure \ncomponent of country index. University of Notre Dame https://\ngain.nd.edu/our-work/country-index/methodology/ (2024).\n58.\t Climate Actions and Policies Measurement Framework \n(OECD & IPAC, 2022); https://oecd-main.shinyapps.io/\nclimate-actions-and-policies/\n59.\t Global warming\u2019s six audiences. Yale Program on Climate Change  \nCommunication https://climatecommunication.yale.edu/ \npublications/global-warmings-six-audiences-a-cross-national- \ncomparison/ (2023).\n60.\t Dabla-Norris, E. et. al. Public Perceptions of Climate Mitigation \nPolicies: Evidence from Cross-Country Surveys. Staff Discussion \nNote (International Monetary Fund, 2023).\n61.\t DataBank. The World Bank Group https://databank.worldbank.org \n(2023).\n62.\t Garriga, A. C. Central bank independence in the world: a new \ndataset. Int. Interact. https://doi.org/10.1080/03050629. \n2016.1188813 (2016).\nAcknowledgements\nWe are grateful for feedback from R.O. Keohane and members of the \nEnergy and Environment Policy Lab at the University of California, \nBerkeley. J.M. acknowledges funding from the Climate Program of \nthe Berkeley Economy and Society Initiative and the USDA National \nInstitute of Food and Agriculture, Hatch Project (accession no. \n1020688).\nAuthor contributions\nE.S. conceived the study and led the development of theory and \ncorresponding analysis, with the guidance of J.M. and J.J.F; E.S. \ncollected the data, designed the methodology, executed the \nstatistical analysis and produced tables and figures, with the guidance \nof J.M. and J.J.F.; and E.S. and J.M. wrote the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at https://doi.org/10.1038/\ns41560-025-01724-w.\nSupplementary information The online version contains \nsupplementary material available at https://doi.org/10.1038/s41560-\n025-01724-w.\nCorrespondence and requests for materials should be addressed to \nEsther Shears or Jonas Meckling.\nPeer review information Nature Energy thanks Sajid M. Chaudhry, \nAnna Geddes Geddes and Friedemann Polzin for their contribution to \nthe peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard  \nto jurisdictional claims in published maps and institutional  \naffiliations.\nSpringer Nature or its licensor (e.g. a society or other partner)  \nholds exclusive rights to this article under a publishing agreement \nwith the author(s) or other rightsholder(s); author self-archiving  \nof the accepted manuscript version of this article is solely  \ngoverned by the terms of such publishing agreement and  \napplicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2025\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-025-01724-w\nExtended Data Table 1 | Examples of central bank measures to manage climate risks\nThere are different classes of climate-related policies, some can support both re-risking and de-risking objectives, while others are only suitable for one or the other of those goals. This table \ndescribes these different classes of policies and provides concrete examples of their re-risking and/or de-risking expression.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-025-01724-w\nExtended Data Table 2 | Classification of central bank measures\nThis table describes the different types of climate-related policies and categorizes these policies by their function, stringency, and cost. Additional details about policy cost and policy type \ndescriptive statistics are provided in Supplementary Note 5.\n\n\n    Scientific Research Findings:", "answer": "We find limited evidence that economic risks related to climate and energy are associated with central bank behaviour. While physical risks are associated with central bank actions to some extent, stranded asset risks and clean energy investment risks are not. Instead, central bank actions to manage risks are significantly and positively associated with domestic climate politics, including climate policy stringency and public concern with climate change. Our results thus suggest a risk mitigation gap between the magnitude of transition risks and central bank actions, and that central banks may not be entirely autonomous risk managers but responsive to political demands, reinforcing, instead of correcting for, lagging decarbonization policy. Our analysis is exploratory. Future research needs to move beyond cross\u2011sectional to time series analysis, investigate the underlying mechanisms, and study the broader regulatory system for climate risk, including financial supervisors and private sector institutions.", "id": 1}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 10 | March 2025 | 329\u2013341\n329\nnature energy\nhttps://doi.org/10.1038/s41560-025-01704-0\nArticle\nPower price stability and the insurance value \nof renewable technologies\n \nDaniel Navia Simon\u2009\n\u200a\u20091,2\u2009\n & Laura Diaz Anadon\u2009\n\u200a\u20091,3\nTo understand if renewables stabilize or destabilize electricity prices, we \nsimulate European power markets as projected by the National Energy \nand Climate Plans for 2030 but replicating the historical variability in \nelectricity demand, the prices of fossil fuels and weather. We propose a \n\u03b2-sensitivity metric, defined as the projected increase in the average annual \nprice of electricity when the price of natural gas increases by 1 euro. We \nshow that annual power prices spikes would be more moderate because \nthe \u03b2-sensitivity would fall from 1.4 euros to 1 euro. Deployment of solar \nphotovoltaic and wind technologies exceeding 30% of the 2030 target \nwould lower it further, below 0.5 euros. Our framework shows that this \nstabilization of prices would produce social welfare gains, that is, we find \nan insurance value of renewables. Because market mechanisms do not \ninternalize this value, we argue that it should be explicitly considered in \nenergy policy decisions.\nSeveral trends are bringing to the fore the importance of understand-\ning the influence of renewable technologies on the volatility of power \nprices. In a context where investments in renewable electricity outpace \nall other forms of capacity additions, understanding how renewables \nwill shape the stability (or volatility) of prices has important repercus-\nsions in the functioning of electricity markets, affecting issues such as \nthe investment strategies of power generators, the hedging of financial \nrisks by suppliers, demand strategies of users of electricity, the design \nof capacity mechanisms or the development of exceptional measures \nto contain electricity price spikes, among others.\nAt the same time, while renewables have long been known to be \ncritical to the achievement of climate goals and a key topic in regula-\ntion, in recent times renewable investments have become a driver of \nmacroeconomic policies and, in this context, their contribution to \nachieving more stable power prices has gained increasing relevance. \nIn September of 2022, President of the European Commission Ursula \nvon der Leyen referred to renewables as \u2018our energy insurance for the \nfuture\u20191, and this consideration has played a critical consideration in \nthe design of the European Green Deal. In the United States, the largest \nfiscal support package in recent times has been structured around sup-\nport for low-carbon technologies and included as part of the Inflation \nReduction Act, which was passed in June 2022, explicitly emphasizing \nthe connection among renewables, energy prices and broader price \nstability2. The United Nations has also frequently stressed renewables\u2019 \nrole in the path to stable power prices3. Overall, these justifications sug-\ngest renewables are increasingly seen in the context of energy security, \nwhich the International Energy Agency defines as the uninterrupted \navailability of energy sources at an affordable price4.\nIn view of the shift described in the previous paragraph, our goal \nis to understand how renewables would affect the properties of power \nprices that are more relevant for macroeconomic stability: their annual \nvolatility and their sensitivity to fluctuations in the price of natural gas. \nWe simulate future European markets and find that faster deployment \nof renewables would improve (that is, reduce) both parameters, even \nwhen accounting for the additional influence of weather factors. We \nthen show how this stabilization would be expected to lead to a poten-\ntially large reduction in the social welfare costs of macroeconomic vola-\ntility, hence creating a large insurance value of renewable investments.\nModelling price volatility in the future of EU power \nmarkets\nOur methodological design has several features that are motivated by \nour focus in macroeconomic effects of future renewable investments. \nFirst, while there is a different price of electricity for each hour of the \nReceived: 19 January 2024\nAccepted: 20 December 2024\nPublished online: 28 January 2025\n Check for updates\n1Centre for Environment, Energy and Natural Resource Governance, Department of Land Economy, University of Cambridge, Cambridge, UK. 2World \nBank, Washington DC, USA. 3Belfer Center for Science and International Affairs, Harvard Kennedy School, Harvard University, Cambridge, MA, USA. \n\u2009e-mail: dn406@cam.ac.uk\n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n330\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\nmarket implications of fossil fuel price volatility, regardless of the \norigin of EU imports of energy and, critically, of natural gas. However, \nthis topic is also clearly of great importance and has been covered \nrecently in the literature.\nBaseline assumptions and the simulation of \nvariability\nWe simulate the functioning of day-ahead electricity markets for all \nEU countries, the United Kingdom and Switzerland simultaneously, \nconsidering the expected power generation capacities in each mar-\nket, and all the interconnections\u2014both built and projected\u2014among \nthem. This area represents around 11% of global power generation, \nbut the approach we develop and apply can be used to analyse other \njurisdictions.\nThe baseline projections for key inputs we use come from the Euro-\npean Resource Adequacy Assessment18 (ERAA) of 2022 by ENTSO-E, the \nEuropean Network of Transmission System Operators for Electricity. \nWe do this to be consistent with the regulatory and planning approach \nin Europe and to make our results more directly relevant for policy \ndiscussions and because the ERAA database provides a highly detailed, \nopen-source set of information that other alternatives cannot replicate. \nAll our data are forecasts for year 2030 as constructed and used by \nENTSO-E. The European regulator forecasts generation capacity config-\nurations for 2030 based on countries\u2019 National Energy and Climate Plans \n(NECPs). For demand, ERAA projects future changes in the expected \nelectricity demand in each country (both its level and hourly profile), \naccounting for technological developments such as the expected \ngrowth in the use of heat pumps and electric vehicles, as derived from \nthe projections in the NECPs (Methods). For weather, the simulations \ndraw from the sample of weather conditions observed between 1982 and \n2016, which affect the capacity factors of variable renewable technolo-\ngies (wind and solar photovoltaic (PV)), the hydrologic conditions for \nhydropower generation and hourly demand profiles. Weather variables \nare adjusted by ENTSO-E as part of the ERAA exercise to reflect changing \nclimate and location profiles. For fuel prices, the central values are those \nused by ERAA too so that they are consistent with the projections used \nfor the rest of the variables. Throughout we use a 100 euro per ton of \nCO2 price for the European Trading System (ETS) allowances, which is \nconsistent with the ERAA exercise projections for 2030.\nWe use Monte Carlo simulation to generate random shocks to \nthese central values, which replicate the historical variance of fuel \nprices and each country\u2019s electricity demand between 1990 and 2021, \nand the correlations among them (Methods), and to replicate weather \nvariability by randomly selecting weather years. In each of the 300 rep-\netitions we run, we employ the GenX model19 to simulate the outcome \nof the market coupling at a very detailed level. We replicate the dispatch \nof power of each technology for every hour of the 8,760\u2009h in a year and \nthe marginal cost of the last technology dispatched, which will approxi-\nmate the equilibrium price that would occur in the day-ahead market if \nconditions are sufficiently competitive20. Because our process is highly \ndetailed and comprehensive in its geographic coverage of Europe, it is \ncomputationally intensive. We run our simulations using the University \nof Cambridge High Performance Computing Centre resources.\nImpacts of 2030 capacity targets on price \nvolatility\nWe first compare the properties of the distribution of annual electric-\nity prices for the European capacity mix targeted by the NECPs (plus \nUnited Kingdom and Switzerland) in 2030 and the current capacity \nmix at the start of 2024.\nOur results confirm that the targets in the NECPs would lower the \nexpected price of electricity across European markets. For the Euro-\npean aggregate, average prices would be expected to be 26% lower by \n2030 than in 2024 (Table 1). A similar reduction would be observed in \nmedian prices.\nyear, from a macroeconomic perspective, hourly price volatility5 is less \nrelevant than episodes of sustained high electricity prices lasting for \nseveral months or years, such as those experienced in 2021 and 2022, and \ntheir correlation with other energy price shocks, such as spikes in oil or \nnatural gas prices. This is exemplified by the new EU electricity market \ndesign regulation, which contemplates that an electricity price crisis \nmay only be declared if prices in wholesale electricity markets reach at \nleast two and a half times the average price during the previous five years, \nand this situation is expected to continue for at least 6 months (ref. 6).\nSecond, the large scale of the planned changes in the total capacity \nand the composition of the grids of European countries by 2030 and, \nin particular, the very large increase of renewable capacity could result \nin very different pricing dynamics from past experience7,8. Whereas \nstudies based on the history of prices in Europe provide important \ninsights5,9\u201312, extrapolation to future price behaviour from them is \nthus potentially problematic. Our design allows simulation of the \nprices resulting from future capacity, with consistent projections for \nfuture electricity demand (which include the effects of electric vehicle \nand heat pump penetration) and adjusting for changing climate. The \nEuropean Energy and Climate Plans foresee substantial reductions in \ncoal and nuclear capacity across Europe, whereas the increasing strin-\ngency of the European Trading System for CO2 could further increase \nthe marginal cost of coal generation, all of which would push for a \npotentially larger role of natural gas technologies. In view of this, we \ndevelop a measure specifically aimed at capturing the sensitivity of \nannual prices to variations in the price of natural gas. What we call the \n\u03b2-sensitivity measures the expected change in the annual average price \nof electricity after a 1 euro increase in the annual average price of natural \ngas. We argue that this is a better measure than the more commonly \nused metric of the number of hours where gas sets the marginal price.\nThird, compared to other studies that have used simulation of \nfuture capacity systems13\u201315, we consider simultaneous shocks to fuel \nprices, weather and demand to model the fundamental tension in sys-\ntems with higher renewable capacity: they are less exposed to the price \nfluctuations in fossil fuels but more exposed to weather and demand \nvariability. We replicate the historical variability and covariation of all \nthese factors, as it is a relevant and transparent benchmark. However, \nrecent research has suggested several reasons why fuel price variability \nmay be higher in the future16. This could be explored in future research \nusing the methods we propose here.\nFourth, our results provide an important input for the concept \nof an insurance value of renewables. A well-known principle in the \neconomics of uncertainty is that risk-averse agents would prefer a \nstable consumption stream to another that, having the same expected \nvalue, has a higher dispersion. Extending this insight to a social welfare \nframework, the literature, starting with Lucas17, has investigated the \neconomic costs of aggregate macroeconomic fluctuations under differ-\nent views of societal preferences. Under this lens, power price stability \nis not a goal in itself and, even if it were, there are many policies that \ncould achieve it besides increasing renewable investment. But if faster \nrenewable deployment leads to a reduction in the volatility of consump-\ntion, and if this reduction leads to an improvement in social welfare, \nthen renewable investment would have a positive societal insurance \nvalue, in addition to its positive environmental benefits (for example, \nreduced greenhouse gas emissions), health benefits (for example, from \nreduced air pollution) and other effects. Similarly to the case made \nfor advancing renewables to reap environmental benefits, if private \nagents fail to consider the benefits of stabilization, there would be an \neconomic argument to consider this insurance value in the design of \npolicies to expand renewables.\nA strong motivation behind the European Union\u2019s energy plans is \nto reduce dependency on volatile partners, particularly Russia in the \ncase of energy. We do not attempt to reflect the energy independence \nand geopolitical security dimensions of renewable investment by the \nEuropean Union in this Article; we focus instead on the electricity \n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n331\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\nThe volatility of annual prices would also be lower, but the size \nof this change is less relevant: one standard deviation in electricity \nprices would be equivalent to 25 euros in 2030, very similar to the 26.5 \neuros expected in 2024. Importantly, the reduction in electricity prices \nwould not be homogeneous across the hours of the day, particularly in \ncountries adding solar PV capacity. As an example, in Germany, prices \nwould be 64% lower in 2030 than in 2024 at 12 in the morning, whereas \nthe reduction at 7\u2009p.m. would be only 16% (Supplementary Section 8).\nGoing beyond the standard deviation as a metric to capture sta-\nbility is important because renewables can soften the incidence of \nextreme prices. We present the 85th percentile (p85) as a reference for \nhigh but reasonably frequent prices and the 95th percentile (p95) as \nthe reference for extremely high but very infrequent prices (Table 2). \nThese two metrics are better suited to capture the needs of policymak-\ners and the interests of consumers. Moreover, they allow capturing \nthe asymmetrical non-normal shape of the probability distribution \nof power prices, which may become more prominent as renewables \ngain importance in the market (Supplementary Section 4 provides \nadditional data on the distribution and skewness, kurtosis measures).\nBy 2030, for the average of Europe, prices would be expected to \nbe higher than 121 euros per MWh with a probability of 15%, and higher \nthan 139 euros per MWh with a probability of 5%. This compares with \n152 and 179 euros per MWh, respectively, in 2024. In other words, the \ngrowth in renewables foreseen by European plans, jointly with other \nchanges in capacity described in the previous section, would moderate \nprice spikes (understood as years with infrequently high prices). For the \nEuropean aggregate, spikes in annual prices could be approximately \n20% lower by 2030 than in 2024. These results confirm an underlying \ntrend towards lower prices overall and a mitigation of price spikes in \nEurope.\nIt is important to emphasize that the projected prices for electric-\nity in Tables 1 and 2 are built with stochastic shocks around the assumed \ndemand and prices of fossil fuels used in the ERAA exercise of 2022. The \ndistribution obtained through this method is, by construction, centred \naround the electricity prices that correspond to this baseline scenario. \nTo check the consistency of the modelling approach with different \ninput prices, we run simulations for 2024 conditions using the prices \nobserved during 2023 and early 2024. While this is not a full backcast-\ning exercise, the model seems to do a good job at replicating observed \nelectricity prices when provided with the observed fuel prices. Other \nimplementations of this methodology may use different baselines for \nfossil fuel prices. The information from quoted future contracts on \nenergy commodities would provide relevant complementary informa-\ntion (Supplementary Section 6).\nA different measure is required to capture the relationship \nbetween electricity prices and the price of fossil fuels and, particu-\nlarly, natural gas. Studies on the influence of the price of natural gas on \nelectricity prices have mostly considered the number of hours in a year \nwhere plants burning natural gas are the marginal price setter11. This is \na partial approach that can lead to biased policy conclusions regarding \nthe influence of natural gas21, especially in systems with higher levels \nof storage capacity (in hydro reservoirs or batteries) or in systems \nwith high interconnections. The reason is that in market equilibrium \ntwo effects are relevant: (1) the alternative price that would be set by \ngas technologies will indirectly define the bids which owners of hydro \nand batteries resources will submit; and (2) the price in one market \nTable 1 | Average, median and standard deviation of annual electricity prices in European markets\nMean (\u20ac\u2009MWh\u22121)\nMedian (\u20ac\u2009MWh\u22121)\nStandard deviation (\u20ac\u2009MWh\u22121)\n2024\n2030\nChange (%)\n2024\n2030\nChange (%)\n2024\n2030\nChange (%)\nPortugal\n139\n121\n\u221213\n134\n116\n\u221213\n31\n26\n\u221214\nSpain\n138\n120\n\u221214\n133\n114\n\u221214\n30\n26\n\u221214\nFrance\n81\n69\n\u221215\n79\n66\n\u221216\n17\n17\n\u22124\nBelgium\n129\n83\n\u221236\n127\n80\n\u221237\n23\n19\n\u221217\nNetherlands\n131\n77\n\u221241\n129\n76\n\u221241\n23\n18\n\u221221\nGermany\n134\n89\n\u221234\n132\n87\n\u221234\n22\n21\n\u22124\nItaly\n142\n136\n\u22124\n136\n130\n\u22124\n31\n30\n\u22125\nSwitzerland\n138\n128\n\u22127\n134\n124\n\u22128\n25\n25\n\u22123\nAustria\n140\n131\n\u22126\n135\n128\n\u22125\n26\n24\n\u221210\nPoland\n132\n126\n\u22124\n131\n126\n\u22124\n17\n17\n1\nCzech Republic\n135\n129\n\u22124\n132\n128\n\u22123\n20\n20\n2\nDenmark\n135\n88\n\u221235\n133\n86\n\u221235\n23\n23\n0\nSweden\n115\n39\n\u221266\n126\n12\n\u221291\n48\n49\n1\nNorway\n115\n35\n\u221269\n126\n6\n\u221295\n49\n54\n11\nFinland\n125\n46\n\u221263\n129\n24\n\u221281\n35\n44\n25\nBaltic\n139\n79\n\u221243\n136\n69\n\u221249\n24\n30\n24\nAdriatic\n140\n133\n\u22125\n135\n131\n\u22123\n26\n22\n\u221215\nOther Europe\n140\n133\n\u22125\n135\n131\n\u22123\n26\n22\n\u221215\nUnited Kingdom\n133\n72\n\u221246\n130\n70\n\u221246\n26\n18\n\u221233\nIreland\n137\n76\n\u221245\n133\n74\n\u221244\n29\n19\n\u221236\nGreece\n141\n134\n\u22125\n135\n131\n\u22123\n28\n24\n\u221216\nAll countries average (demand \nweighted)\n126\n93\n\u221226\n124\n87\n\u221230\n26\n25\n\u22127\nAll countries median\n\u221215\n\u221216\n\u22125\nThe columns show mean, median and standard deviation of prices, all of them expressed in euros per MWh, for each European country for the current (2024) capacity and grid and for the \nexpected capacity and grid in 2030, according to the NECPs. The average of all countries is weighted by their respective demands. \n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n332\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\nmay be set by a natural gas plant in another market if there is enough \ninterconnection capacity between both. Hence, it is possible that the \ncost of natural gas may be indirectly setting the marginal price most \nof the time even if the share of hours when it is producing the marginal \nelectricity unit is very low. This means that electricity prices can go up \nin response to a spike in the price of natural gas by a very large amount \neven if combined cycles are not setting the marginal price very often.\nImportantly, different weather and demand conditions would \nresult in a different reaction of electricity prices to natural gas prices. \nIn years with high demand and low wind and hydraulic reserves, for \nexample, increases in the price of natural gas would lead to a more \nintense reaction in electricity prices than in opposite conditions. \nInteresting insights can be obtained by comparing the value of this \nsensitivity between two pre-defined scenarios (for example, a high \nwind and rain year versus a low wind and rain year). Our interest, \nhowever, is to provide a measure of the sensitivity of electricity prices \nto gas prices across the stochastic range of possible scenarios. For \nthis, we define the \u03b2-sensitivity metric as the projected change in \nthe annual average electricity price when the annual average price of \nnatural gas increases by 1 euro. \u03b2-sensitivity is derived from the linear \nregression of the average annual price on the price of gas and other \ncontrol variables across all Monte Carlo simulations (Methods and \nSupplementary Section 5 provide further details on the rationale and \nestimation of \u03b2-sensitivity).\nWe find that by 2030, a 1 euro increase in the price of natural gas \nwould translate into a 1 euro increase in annual electricity prices for \nthe aggregate of European countries. This would be 40% lower than \nthe situation in 2024, where the electricity price would be expected \nto increase by 1.4 euros.\nInterpreting the results for \u03b2-sensitivity requires some clarifica-\ntion. For combined cycle gas turbines (the majority of natural gas power \nplants in 2024 and the 2030 plans), we assume a thermal efficiency of \n49% (ref. 22), implying a gas plant must burn 2.05\u2009MWh of natural gas \nto obtain 1\u2009MWh of electricity. Because of this, if the annual price of \nelectricity reacted one to one with the marginal costs of natural gas \nplants, we would expect the increase in the price of electricity of 2.05 \neuros when the price of gas went up by one euro. Our \u03b2-sensitivity esti-\nmates indicate that by 2030, the price of electricity would reflect 50% \nof the increase in the short-run marginal cost of gas plants compared \nto 70% today (that is, 0.7\u2009=\u20091.4/2.05).\nTable 2 also shows that there are important differences across \ncountries in terms of the value of the NECPs in achieving power price \nstability. The moderation in price spikes is clearer in the United King-\ndom, Ireland, Netherlands, Germany and the Nordic countries, whereas \nfor Italy, Austria, Poland, the Czech Republic and other countries in \nEastern Europe, the changes would be very small. Our simulations \nconfirm that exposure to natural gas is the main factor affecting the \nprobability of price spikes in European markets in each country (Fig. 1). \nTable 2 | Measures of the stability of electricity prices and their sensitivity to natural gas prices in 2030 versus 2024\np85 (\u20ac\u2009MWh\u22121)\np95 (\u20ac\u2009MWh\u22121)\n\u03b2-sensitivity (\u20ac\u2009MWh\u22121 change \nin electricity price per 1 \u20ac\u2009MWh\u22121 \nchange in natural gas price)\n2024\n2030\nChange (%)\n2024\n2030\nChange (%)\n2024\n2030\nChange\nPortugal\n170\n148\n\u221213\n201\n173\n\u221214\n2.1\n1.7\n\u22120.4\nSpain\n168\n146\n\u221213\n199\n171\n\u221214\n2.1\n1.7\n\u22120.4\nFrance\n98\n86\n\u221212\n113\n98\n\u221214\n0.8\n0.8\n0.0\nBelgium\n153\n102\n\u221233\n173\n118\n\u221232\n1.4\n0.9\n\u22120.5\nNetherlands\n154\n96\n\u221238\n174\n111\n\u221236\n1.4\n0.9\n\u22120.5\nGermany\n157\n109\n\u221231\n175\n128\n\u221227\n1.3\n0.9\n\u22120.4\nItaly\n173\n166\n\u22124\n205\n195\n\u22125\n2.1\n2.0\n\u22120.1\nSwitzerland\n166\n154\n\u22127\n187\n176\n\u22126\n1.7\n1.6\n\u22120.1\nAustria\n168\n156\n\u22127\n191\n175\n\u22128\n1.7\n1.5\n\u22120.3\nPoland\n149\n143\n\u22124\n163\n155\n\u22125\n0.9\n0.9\n0.0\nCzech Republic\n155\n150\n\u22123\n171\n166\n\u22123\n1.2\n1.2\n0.0\nDenmark\n158\n109\n\u221231\n176\n133\n\u221224\n1.4\n0.9\n\u22120.5\nSweden\n149\n110\n\u221226\n180\n132\n\u221227\n0.9\n\u22120.1\n\u22121.0\nNorway\n149\n118\n\u221221\n181\n137\n\u221224\n0.9\n\u22120.1\n\u22121.1\nFinland\n153\n108\n\u221229\n180\n131\n\u221227\n1.0\n0.0\n\u22121.0\nBaltic\n165\n114\n\u221231\n185\n137\n\u221226\n1.5\n0.5\n\u22121.0\nAdriatic\n168\n159\n\u22125\n188\n173\n\u22128\n1.7\n1.3\n\u22120.3\nOther Europe\n168\n159\n\u22125\n188\n173\n\u22128\n1.7\n1.3\n\u22120.3\nUnited Kingdom\n162\n90\n\u221244\n184\n103\n\u221244\n1.7\n0.9\n\u22120.8\nIreland\n168\n96\n\u221243\n197\n112\n\u221243\n2.0\n1.0\n\u22120.9\nGreece\n169\n161\n\u22125\n198\n176\n\u221211\n1.9\n1.5\n\u22120.4\nAll countries average (demand \nweighted)\n152\n121\n\u221220\n174\n139\n\u221220\n1.43\n1.00\n0.4\nAll countries median\n\u221213\n\u221214\n\u22120.38\nEach section of the table represents a different metric of the instability of annual electricity prices. The left-most section presents the 85th percentile of the simulated results, using the \ncapacity configuration in 2024 and in 2030, and the final column computes the per cent change in this percentile from one to the other. The medium section presents the same information for \nthe 95th percentile. The right-most section presents the estimates of the \u03b2-sensitivity of each country, which measures the expected increase in the annual price of electricity when the price \nof natural gas increases by 1 euro. \u03b2-sensitivity is calculated for the capacity foreseen in 2024 and in 2030, and the last column presents the change (in euros) between one and another. The \naverage of all countries is weighted by their respective power demands. \n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n333\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\nThe negative change in \u03b2-sensitivity from 2024 to 2030 is strongly cor-\nrelated with the reduction in the tails of the price distribution, both \nfor high and extreme episodes; countries that reduce the dependency \nof their markets on natural gas are those that also obtain a clearer \nmitigation of price spikes. This is an important result as it confirms the \nunderlying rationale of European renewable deployment policies. By \nreducing the dependence on natural gas, they help achieve broader \nprice stability. Because of this, in what follows, we focus on understand-\ning what drives the evolution of \u03b2-sensitivity.\nImpact of additional renewables\nWe turn to analysing the impact of accelerating variable renewable \ndeployment. We construct hypothetical scenarios where the installed \ncapacity of both wind and solar PV technologies are increased up to \n60% above the NECP targets of each country in steps of 10%. We also \ninclude scenarios where deployment is 10% and 20% lower than the \nNECP target. The capacity for all other technologies is kept at the same \nlevels foreseen in the NECPs. These simulations provide a laboratory \ntest of the pricing impacts of different levels of renewable deploy-\nment. We calculate the annual electricity prices for each country for \nthe same exact combination of climate year, demand conditions and \nprices of fossil fuels, and capacity for other technologies, so that the \nonly factor that diverges is the amount of renewable capacity in that \nscenario (Table 3).\nTo better understand the economic relevance of these results, \nwe assess the impact that higher levels of renewable capacity would \nhave on the \u03b2-sensitivity parameter. As discussed, by 2024, a 1 euro \nincrease in the price of gas would be foreseen to increase the annual \nprice of electricity by approximately 1.4 euros per MWh. By 2030, if \nthe NECP targets are met, the same shock would increase electricity \nprices by 1 euro. Deploying 10% more solar PV and wind relative to the \ntarget would additionally reduce this sensitivity to 0.8 euros. Reduc-\ning the \u03b2-sensitivity of electricity prices below 0.5 euros (that is, a \nsituation where for every euro of increase in the price of natural gas, \nthe annual price of electricity would only increase by 0.5 euros) would \nrequire deploying 30% or more renewables than in the NECPs (Fig. 2). \nConversely, if installed renewable capacity by 2030 fell 20% short of the \nenvisioned levels, the results indicate that the \u03b2-sensitivity would be \nhigher than in 2024 (1.5 versus 1.4). In this case, additional renewable \ncapacity would not be enough to offset the closure of conventional \ncapacity (particularly coal but also nuclear) between 2024 and 2030, \nresulting in a larger role for natural gas and its price.\nTable 3 also shows remarkable differences across countries; look-\ning at the European average does not tell the whole story. At +30% \ndeployment of solar PV and wind, countries such as Italy or Austria \nwould not achieve a \u03b2 lower than 0.5, whereas Spain, Portugal and the \nNordics would already be below 0.25. Achieving a \u03b2 of 0.25 consistently \nacross most markets in Europe would require between 50% and 60% of \nadditional deployment. Note that the scenarios refer to simultaneous \noverachievement of national targets by all countries. In other words, \nit is necessary for every country to deliver over the target by 50\u201360%.\nThe differences in \u03b2-sensitivity across countries can be driven by \nseveral factors, including among others the availability of coal capacity \nand interconnection capacities. Given the highly nonlinear nature of \nmarket equilibrium in power markets, a full explanation is beyond the \nscope of this paper and is a relevant area for future research. However, \none key factor to achieve a low \u03b2-sensitivity stands out: a higher number \nof hours where variable renewables set the price of electricity at their \nown, almost zero, variable costs.\nAn important implication of this result, as shown in Fig. 3, is that \nprivate, strictly market-based financing of much higher amounts of \nrenewable capacity beyond the current 2030 targets could be very dif-\nficult under the current market rules in situations where \u03b2-sensitivity \nwas very low. The reason for this is that the revenue of electricity sold \nby these technologies (that is, their captured price if selling directly in \nthe day-ahead market) would be depressed as excess capacity would \ndrive down prices during sunny and/or windy hours where they operate. \nDeployment at this scale would generate large curtailment of renewable \nelectricity, which depresses revenues for all market participants and \nwould have important implications for system stability23. The satura-\ntion or cannibalization24,25 of the price implies fewer opportunities to \nrecover capital costs for renewable technologies if they are financed \nPortugal\nSpain\nFrance\nBelgium\nNetherlands\nGermany\nItaly\nSwitzerland\nAustria\nPoland\nCzech Republic\nDenmark\nSweden\nNorway\nFinland\nBaltic\nAdriatic\nOther Europe\nUnited Kingdom\nIreland\nGreece\n\u201350\n\u201345\n\u201340\n\u201335\n\u201330\n\u201325\n\u201320\n\u201315\n\u201310\n\u20135\n0\n\u20131.2\n\u20131.0\n\u20130.8\n\u20130.6\n\u20130.4\n\u20130.2\n0\n0.2\nChange in 95th percentile (%)\nChange in \u03b2-sensitivity\nFig. 1 | Change in extreme power prices between 2030 and 2024 as a function \nof the change in the sensitivity of prices to natural gas prices. The vertical \naxis presents the change in the 95th percentile of the simulated results when \ncomparing the capacity configuration in 2024 and in 2030. The horizontal axis \nshows the change in the \u03b2-sensitivity, also calculated for these two capacity year \nreferences.\n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n334\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\nexclusively through their day-ahead market electricity sales. The larger \nthe reduction in the sensitivity, the larger the resulting potential gap in \nthe revenue of renewable plants. At the same time, prices of long-term \ncontracts would be indirectly linked with short-term prices and expec-\ntations around them, the credit risk associated with buyers of Power \nPurchase Agreements (PPAs) or the policy reversal risk of contracts for \ndifferences (CFDs) would also depend on the variability of short-term \nprices, and in practical conditions, the overall supply of long-term \ncontracts would still be conditioned by the volatility of natural gas \nprices. Hence, besides the logistical difficulties in achieving deploy-\nments 30% to 60% higher than the original NECP target by 2030, the \nfinancial viability of such deployment without policy support or other \nreforms is far from evident.\nWe must stress that the results in Fig. 3 are constructed for the \nhypothetical scenarios where additional capacity is only added to \nsolar PV and wind technologies. An important research question \nis whether other options to increase renewable capacity, for exam-\nple, with different combinations of technologies, or including addi-\ntional storage or improving interconnections, could achieve low \n\u03b2-sensitivity more effectively and generate market revenue that would \nmake them investable without policy support. To obtain some insight \ninto this question, we estimated results when storage capacity is \ndeployed jointly with the new variable renewable capacity (Sup-\nplementary Section 7). These still find a strong positive correlation \nbetween \u03b2-sensitivity and captured price, which we attribute to the \nfact that the estimates maintain the projected natural gas capacity \nin 2030 unaltered, hence limiting the amount of additional storage \nthat can be added profitably. In any event, future analysis (particularly \nin the context of the design of support mechanisms for renewables) \nshould consider a broader range of deployment possibilities and \nexplicitly evaluate their influence on the \u03b2-sensitivity parameters of \nthe EU electricity markets.\nThe insurance value of renewables\nOur results confirm that additional investments in renewable capac-\nity in the European Union would be expected to lead to a large reduc-\ntion in the volatility of electricity prices and their sensitivity to fossil \nfuel prices. The question is then to which extent this stabilization \nwould have societal value in itself, that is, whether there is an insur-\nance value of renewable investments as claimed by recent policy \nprogrammes.\nFormally, the insurance value of an increase in renewables is the \ngain of welfare associated with the stabilization of consumption ena-\nbled by an increase in renewables due to their impact on power price \nstabilization, among other possible mechanisms. The underlying \nassumption is that some form of social welfare function can account \nfor the societal value of uncertain future consumption streams. Risk \naversion implies that social welfare is reduced by an amount \u03c9 because \nof the variability of aggregate consumption, compared with a situation \nwith stable consumption over time with the same expected value. This \nTable 3 | Sensitivity of annual electricity price to a 1 euro change in the price of natural gas for different renewables \ndeployment scenarios by country\nVariable renewable capacity in 2030 versus NECP target\n\u221220%\n\u221210%\nOn target\n+10%\n+20%\n+30%\n+40%\n+50%\n+60%\nPortugal\n1.98\n1.88\n1.71\n1.12\n0.81\n0.19\n0.03\n\u22120.02\n0.02\nSpain\n1.98\n1.86\n1.69\n1.10\n0.79\n0.20\n0.04\n\u22120.01\n0.02\nFrance\n1.32\n1.01\n0.78\n0.45\n0.31\n0.16\n0.09\n0.06\n0.06\nBelgium\n1.42\n1.15\n0.94\n0.76\n0.63\n0.51\n0.44\n0.39\n0.36\nNetherlands\n1.40\n1.10\n0.88\n0.72\n0.60\n0.49\n0.43\n0.39\n0.35\nGermany\n1.42\n1.14\n0.91\n0.74\n0.57\n0.44\n0.36\n0.31\n0.26\nItaly\n2.10\n2.06\n2.02\n1.96\n1.89\n1.82\n1.74\n1.67\n1.61\nSwitzerland\n1.83\n1.71\n1.59\n1.42\n1.22\n1.00\n0.84\n0.71\n0.60\nAustria\n1.72\n1.60\n1.46\n1.32\n1.08\n0.86\n0.69\n0.59\n0.48\nPoland\n1.31\n1.05\n0.87\n0.74\n0.60\n0.49\n0.41\n0.36\n0.32\nCzech Republic\n1.56\n1.34\n1.18\n1.04\n0.86\n0.70\n0.58\n0.51\n0.45\nDenmark\n1.46\n1.13\n0.88\n0.71\n0.55\n0.41\n0.34\n0.29\n0.25\nSweden\n1.20\n0.35\n\u22120.06\n\u22120.02\n0.00\n\u22120.06\n0.01\n0.01\n0.01\nNorway\n1.22\n0.30\n\u22120.15\n\u22120.06\n\u22120.03\n\u22120.10\n\u22120.01\n0.00\n0.00\nFinland\n1.17\n0.38\n0.00\n0.03\n0.04\n\u22120.01\n0.04\n0.04\n0.03\nBaltic\n1.40\n0.85\n0.51\n0.44\n0.36\n0.26\n0.24\n0.21\n0.18\nAdriatic\n1.82\n1.55\n1.34\n1.11\n0.77\n0.55\n0.42\n0.35\n0.33\nOther Europe\n1.82\n1.55\n1.34\n1.11\n0.78\n0.55\n0.42\n0.35\n0.32\nUnited Kingdom\n1.39\n1.09\n0.88\n0.70\n0.57\n0.46\n0.40\n0.34\n0.30\nIreland\n1.50\n1.22\n1.01\n0.83\n0.70\n0.60\n0.52\n0.47\n0.42\nGreece\n1.91\n1.72\n1.49\n1.32\n1.08\n0.84\n0.60\n0.42\n0.34\nAll countries average (demand \nweighted)\n1.54\n1.22\n1.00\n0.81\n0.66\n0.49\n0.42\n0.37\n0.34\nMedian\n1.46\n1.15\n0.94\n0.76\n0.63\n0.49\n0.41\n0.35\n0.32\nThe number in each cell of the table is the \u03b2-sensitivity of each country. It measures the expected increase in the annual price of electricity when the price of natural gas increases by 1 \neuro. Each column represents a different percentage of over- or underachievement of the targets for wind and solar PV capacity in the NECPs of all member states plus United Kingdom \nand Switzerland forecasts, as derived from the ERAA exercise. The last two rows are the average of the sensitivity using demand as the weighting variable and the median sensitivity across \ncountries. \n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n335\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\nformulation faces several controversies, including particularly on the \nchoice of the discount rate and its adequate calibration to reflect risk \naversion and income uncertainty26 given equity premiums and other \nempirical evidence. Recent research tends to find that contrary to \nLucas\u2019 argument, economic volatility has large welfare costs27,28, so \nthat the value of economic stabilization policies is high. Therefore, \nthe main question in our context must be whether deploying more \nrenewables can be effective in stabilizing income and consumption \nat the aggregate level.\nIn Methods, we show that the effectiveness of additional renewa-\nbles in stabilization is actually higher, among other factors, when it \nleads to a more intense reduction in the volatility and sensitivity of \nelectricity prices to fuel prices holding the variance of the later con-\nstant. This is precisely the calculation we performed in the previous \nsections (Table 3), and the results there indicate that scenarios with \nlarge increases in renewables would result in very relevant reductions \nin both parameters. Whereas we do not quantify the insurance value \nof renewables fully in this paper, the results in Table 3 would support \nthe view that there is in fact an insurance value of renewables and that \nit can be potentially large.\nWhether it justifies different types or levels of policy support \ndepends on the extent to which private investors do not consider the \noverall economic stabilization impact of their renewable projects. \nIf they do not factor the insurance value, investment in renewables \nwould be below the societal optimum even if their full environmental \nbenefits were accounted for, for example, through a carbon tax. The \nidea that there may be an economic stabilization premium of some \nforms of energy is not a new concept in energy economics: a long tra-\ndition of research has focused on calculating the security premium of \ndomestic oil versus imports29 and estimates of this premium include \na component capturing the possibility that domestic oil production \nlowers the exposure of the economy to supply disruptions and the fact \nthat the marginal buyer does not recognize this effect when choosing \nits imports. Our results indicate that a similar exercise for renewable \ninvestments is warranted and would be an important area for further \nresearch.\nConclusions and discussion\nOverall, a realistic quantification of the effects of additional renewable \nenergy deployment through its ability to shield economies from fossil \nfuel price shocks, when also considering weather and demand volatil-\nity, suggests large stabilizing effects on electricity prices. We find that \nthe capacity expansion plans as envisioned in the NECPs would lead to \na reduction of the \u03b2-sensitivity to natural gas prices from 1.4 euros to 1 \neuro, which, in turn, would lower the extremes in prices that could be \nexpected in the future. However, we find that the resulting improve-\nment falls short of what would be required if the policy goal is to be \nclose to independence from the prices of natural gas. Reducing the \n\u03b2-sensitivity to less than 0.5 euros would require deploying 30% more \nrenewables by 2030, and going below 0.25 euros would require 60% \nadditional deployment versus the currently envisioned target. These \nare very large changes. Moreover, further increasing the capacity of \nrenewable technologies, while lowering the sensitivity and improving \nthe stability of electricity prices, results in cannibalization conditions \nthat would be associated with low market revenue for renewable pro-\nducers in day-ahead markets. Hence, the financial viability of private \ninvestments based strictly on this market for reaching the capacity \nlevels required to stabilize electricity prices is doubtful.\nThis is an important insight for ongoing debates on electricity mar-\nket reforms in the European Union. Proponents of reform argue that the \ncurrent design of markets in the European Union is not fit for the task \nof accommodating very high shares of renewables. The uniform-price \nauction or pay as clear nature of price setting in day-ahead European \nelectricity markets, coupled with strong restrictions on capacity mar-\nkets as sources of revenue for power plants, has been a hotly debated \nissue30. In this design, the marginal costs of the most expensive plants \nrequired to satisfy demand set the price for all transactions in the \nmarket. Critics of the current design argue that this feature gives a \ndisproportionate role to fossil fuel prices, and particularly natural gas, \nin the determination of electricity prices and impedes consumers from \nfully capturing the benefits of additional renewable power generation31. \nThose who defend the current market design argue, however, that set-\nting prices at the costs of the marginal producers provides the right \nincentives to cheaper technologies to enter the market at the efficient \nlevel, avoiding the pitfalls of systems where capacity decisions are taken \nwith a strong degree of administrative intervention32,33.\nIn this regard, the most recent reform of electricity markets in \nEurope6 maintain the essential elements of day-ahead markets, while \nincluding support measures for long-term PPAs and a requisite to \nuse CFDs in support mechanisms for new renewable capacity34. If the \nobjective is to use high renewable electricity capacity to reduce the \n0\n0.2\n0.4\n0.6\n0.8\n1.0\n1.2\n1.4\n1.6\n1.8\n\u20130.3\n\u20130.2\n\u20130.1\n0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n\u03b2-sensitivity \nVariable renewable capacity in 2030 versus NECP target\nFig. 2 | European average \u03b2-sensitivity of annual electricity price to a 1 euro \nchange in the price of natural gas. The figure represents the relationship \nbetween \u03b2-sensitivity of electricity prices and renewable capacity by 2030. The \nvertical axis represents the European average \u03b2-sensitivity, which measures the \nexpected increase in the annual price of electricity when the price of natural \ngas increases by 1 euro. The horizontal axis represents the per cent over- or \nunderachievement of the indicative targets for wind and solar PV capacity in the \nNational Energy and Climate Plans of all Member States (plus ERAA forecasts for \nUnited Kingdom and Switzerland). For example, 20% means that the simulated \nvariable renewable capacity is 20% higher in each country than envisioned in the \nNECP targets.\n\u20130.5\n0\n0.5\n1.0\n1.5\n2.0\n2.5\n0\n20\n40\n60\n80\n100\n120\n140\n160\nCaptured price (\u20ac MWh\u22121)\n\u03b2-sensitivity\nFig. 3 | Captured price by wind onshore in each of the scenarios. Average price \nat which wind onshore generators would sell their production (in the x axis) and \nthe corresponding \u03b2-sensitivity parameter (in the y axis). Each point in the graph \ncorresponds to a country and an assumed increase in renewable deployment. The \nrenewable deployment scenarios are the same as those described in Table 3.\n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n336\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\nsensitivity to natural gas prices, this implies depressing the prices \nachieved by renewable producers. Because the characteristics con-\ntemplated in these reforms make them mostly a risk-management \ninstrument35, the underlying financial viability of very large amounts of \nrenewables would not be substantially improved by fair-valued CFDs.\nCapacity mechanisms have also been proposed as part of these \nreforms, but recent literature has argued that they tend to favour tech-\nnologies with higher fuel costs, rather than high capital cost options \nsuch as nuclear36. In these conditions, capacity mechanisms would \nfavour technologies that increase price volatility. Our results suggest \nthat this side effect is very relevant and open the possibility that capac-\nity mechanisms without further improvements in the availability of \nrisk trading instruments could offset the price stabilization gains of \nrenewable penetration.\nFinally, the electricity price stabilization effect we document pro-\nvides preliminary support to the notion that investment in renewables \ncan reduce macroeconomic volatility. This in turn suggests that the \ninsurance value of renewables, their ability to improve social welfare \nthrough their smoothing of economic shocks, could be substantial. \nThis does not mean that power price stability is a goal in itself, to be \nattained regardless of the cost. But policy planning that focuses only \non average magnitudes such as expected total system costs could be \nbiased against renewables by failing to acknowledge the insurance \nvalue of renewables (or alternatively, the risk premium associated \nwith fossil fuel technologies). However, as of today, none of the most \nrelevant energy planning procedures in the European Union explicitly \ncontemplate the power price stability contribution of renewables. It \nis not currently considered in the elaboration of the European Union\u2019s \nNational Energy and Climate Plans nor in the European Resource Ade-\nquacy Assessment. Our results provide a justification to advance in the \npotential integration of the insurance value of renewables in the social \ncost\u2013benefit analysis of renewable investments and the institutional \ndesign of EU policies, and the concepts we advance could be applicable \nto other jurisdictions.\nMethods\nExpected changes in European markets generation capacity\nWe start by comparing the capacity plans for 2030 with the current \ncapacity mix. For 2030, we take the power generation capacity targets \nby technology published in the EU member states\u2019 National Energy \nand Climate Plans (NECPs). Introduced by Regulation (EU) 2018/1999 \nin 201837,38, the NECPs are the main long-term planning instrument \nfor member states\u2019 energy systems. The NECPs set out each country\u2019s \nobjectives for power generation capacity by technology within the \nEnergy Union39. Data for the United Kingdom and Switzerland are from \nthe European Resource Adequacy Assessment18 (ERAA).\nThe NECPs anticipate dramatic changes in the installed genera-\ntion capacity of European countries (Supplementary Section 2). We \nuse 2024 as the reference year for the current capacity mix because \nthis allows us to acknowledge generation capacity that is very close to \ncoming into market or to closing. Countries\u2019 plans foresee the doubling \nof variable renewables (wind and solar) from slightly above 500\u2009GW to \nmore than 1000\u2009GW by 2030. Solar PV would experience the largest \nabsolute growth, with the addition of more than 300\u2009GW in this decade, \nwhereas wind turbine capacity would increase by more than 230\u2009GW \n(Supplementary Section 2 provides details). The plans also envision \nadditions to hydropower capacity, battery storage and demand-side \nresponse, but on a much smaller scale. As for the retirements, all fos-\nsil fuel technologies experience reductions in capacity by 2030 in the \nNECPs, with coal being the most affected\u2014capacity would be reduced \nby 41\u2009GW, so that by 2030 coal power plant capacity would be approxi-\nmately halved. The closure plans of nuclear plants in several countries \nwould also lead to a sharp reduction total deployed capacity in this \ntechnology by 2030. Whereas there are important differences in scale \nacross countries, the broad trend towards substitution of fossil fuels \nand nuclear with variable renewables is common to all countries. The \nNECPs also reflect a large expansion in the interconnection capabilities \nof European markets, which is also a driving factor of price formation. \nAll interconnector projects in NECPs and the ERAA have been modelled \n(Supplementary Section 3).\nData sources\nWe use the databases representing the NECPs that were implemented as \npart of the ERAA18 of 2022 by ENTSO-E, the European Network of Trans-\nmission System Operators for Electricity. ERAA figures also include \npower generation capacity objectives for 2030 for two non-EU coun-\ntries (the United Kingdom and Switzerland). The initial set of NECPs \ncovering the period from 2021 to 2030 were submitted by countries in \n2019. A revision of these initial plans was ongoing at the time of writing \nthis article, but it was not possible to include the new information here \nas commission recommendations were issued only in December 2023 \nand several countries presented their plans with a delay.\nThe input data for the analysis are obtained from the 2022 ENTSO-E \nERAA databases40. The ERAA scenarios are based on the submission \nof countries NECPs, updated to account for recent trends and new \ninformation available after their publication. ERAA methodological \ndocument41 provides details on the statistical techniques and assump-\ntions used by ENTSO-E to derive each database, including how climate \nyears are considered, the treatment of electric vehicles and heat pumps \nand so on. Data in the ERAA databases are for each of Europe\u2019s bidding \nzones. Supplementary Section 1 contains further detail on the data \nsources and their treatment.\nMonte Carlo simulation design\nOur simulation approach is based on creating a large number of future \nscenarios in a way that replicates the historical variability of weather \nconditions, demand fluctuations and the prices of fossil fuels. We use \n300 future scenarios (that is, repetitions) in the analysis.\nWe replicate demand and fuel prices variability as follows. The \ninitial data include 21 annual demand data series and four annual series \nfor fossil fuel prices (natural gas, coal, oil and uranium). All of them \nare for the period between 1990 and 2021. We detrend each of these \nseries (after taking logs) using a Hodrick\u2013Prescott filter42 with lambda \nset to 100, following the standard approach for annual frequency \ndata. This procedure results in 25 residual series: each of them can be \ninterpreted as the (percentual) deviation from the respective series \ntrend. By construction, these residuals have zero or close to zero aver-\nage. We compute the variance and covariance matrix of these series: \nthe components of this matrix capture the covariance between each \nof the 21 country/areas and the covariance of each of them with each \nof the four fossil fuel prices. Elements along the diagonal capture the \nvariances of each of the 25 series.\nNote that we keep CO2 prices constant in the simulations as we \nconsider them a policy variable, which can be adjusted by modifying \nthe parameters of the ETS allowances allocation, as the European Union \nauthorities have done systematically in the last decades.\nWe randomly draw 300 repetitions from a normal multivariate \ndistribution with this same variance/covariance structure and zero \naverage. This procedure results in a set of 300 future deviation from \ntrend scenarios for each country demand and each fuel price which, by \nconstruction, replicate the variance\u2013covariance structure of national \ndemands and fuel price. These shocks capture annual frequency vari-\nations, driven by macroeconomic factors.\nAn important feature of ERAA data is that hourly capacity factors, \nweekly hydro inflows and hourly demands are provided adjusted for \ndifferent weather conditions, replicating those observed for each year \nbetween 1987 and 2016 (the last year for which data are available for \nevery country in the sample). We take full benefit of this by simulat-\ning electricity markets for randomly generated weather years, hence \naccounting for the historical variability in weather and its impact on \n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n337\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\ncapacity factors of wind and solar PV, inflows for hydropower tech-\nnologies and the demand for electricity. Each repetition is assigned \na randomly selected weather year between 1987 and 2016, using a \nuniform distribution. The weather year defines both the capacity fac-\ntors of renewables and hydro and the hourly demand of each country \n(both its total level and its hourly profile).\nEach of the 300 repetitions is thus characterized by a randomly \nselected vector of annual national demand and price deviations from \ntrends and a weather year. More specifically, each repetition is char-\nacterized by the following three factors.\nFirst, a vector with the prices of gas, oil, coal and uranium, each of \nthem constructed as the product of the central reference price (which \ncomes from the ERAA scenario) and the randomly generated deviation \nfrom trend.\nSecond, a vector of hourly capacity factors for variable renewa-\nbles and inflows into hydro generators, which is determined by the \nrandomly selected weather year, for each country.\nThird, a vector of hourly demand for each country, which is deter-\nmined as the central scenario corresponding to that weather year \n(coming from ERAA), scaled up by the randomly generated deviation \nfrom trend.\nThis procedure means that capacity factors and demand (both its \nprofile and its level) are correlated as would be warranted by weather \nvariations, whereas demand levels and fossil fuel prices are also cor-\nrelated as dictated by their historical covariance.\nWe construct the scenario configuration for each of these 300 \nrepetitions as follows: (1) generation and interconnection capacities \nare set at the levels foreseen in the NECPs as updated by the ERAA \nexercise for the horizon year being considered (2024 or 2030). (2) \nEach country/area demand is the result of taking the central forecast \nin the NECPs/ERAA databases for the horizon year (2024 or 2030), \nadjusted to reflect the corresponding weather year (randomly selected \nbetween 1987 to 2016) plus the deviation from trend resulting from the \nrandom shock to demand simulated to replicate historical volatility in \nnational demand from long-term trend. (3) Fuel prices for the simulated \nscenario-replication are the result of taking the central reference prices \nof the NECPs/ERAA exercise for the horizon year (2024 or 2030) plus \nthe deviation from trend resulting from the random shock simulated \nto replicate historical volatility in fuel prices from their trend. Each of \nthese repetitions is then used to simulate the outcome of day-ahead \nmarkets.\nThe results in Table 3 simulate the same exact 300 future scenarios \nfor hypothetical changes in the capacity mix. These changes are cal-\nculated increasing (in 10% steps) the nominal capacity of solar PV and \nwind (onshore and offshore) technologies.\nMarket simulation\nWe use GenX, an open-source electricity resource capacity expansion \nmodel, to replicate market dispatch19. GenX is a highly configurable \nsystem, which allows for complex simulation of electricity markets, \nincluding the possibility of finding optimal (that is, cost minimiz-\ning) capacity expansion plans, the simulation of operating reserves, \nunit-commitment restrictions, environmental and subsidy policies, \namong others. Our code only uses GenX dispatch module as, in our \napproach, capacity is defined by the National Energy and Climate \nPlans of EU member states, and because our capacity figures are given \nfor national aggregates (that is, not at plant-level), some of the model \nfeatures are less relevant (unit commitment, for example).\nThe simulation of market dispatch is carried out using version \n0.3.0 of GenX. The only modification to the open-source version of \nthe GenX algorithm was the addition of CO2 ETS costs to the variable \ncosts of all fossil fuel technologies. The code required for this is also \navailable via Github at https://github.com/DanielNavia1/GenX-CO2Tax.\nGiven the size of the problem, each repetition usually requires \napproximately 0.5\u2009h to solve in a modern personal computer. Because \nthere are 300 repetitions (each of them with different demand, prices, \nweather) and 11 capacity configurations (one for 2024, nine for the sen-\nsitivity scenarios as depicted in Table 3, and the 2030 target scenario \nis repeated twice), full execution of the simulations would require \napproximately 9.8 weeks if performed sequentially. To shorten this \ntime, the simulations were carried out using the University of Cam-\nbridge High Performance Computing platform through the Cambridge \nService for Data-Driven Discovery (CSD3). The repetitions were per-\nformed in parallel with each of them assigned to a node with six cores, \nand a maximum of 50 nodes were used in each point in time. With this \nconfiguration, solving all the repetitions required for the analysis \nrequired approximately half a day.\nGiven a fixed configuration of capacities for a set of zones (coun-\ntry/areas in our implementation) and transfer capacities across zones \n(interconnections in our implementation), GenX\u2019s dispatch algorithm \nminimizes the total variable costs of satisfying each zone\u2019s demand at \neach of the 8,760\u2009h in the year. The variable costs of each technology \ninclude a non-fuel variable operating cost plus a fuel cost, which is \ndefined in turn by the fuel price per thermal MWh and the thermal \nefficiency of each plant type. GenX allows for the simulation of stor-\nage technologies, including hydropower plants with reservoirs and/or \npumping capabilities and batteries. Because the algorithm minimizes \nthe total annual cost, the resulting solution assumes perfect foresight \nof hourly profiles for capacity factors, inflows and demands for the full \nyear. Hence, batteries and reservoirs will be charged and discharged \nin the optimal dispatch solution provided by GenX at the times that \nminimize the total cost of electricity generation for the full year. GenX \nincludes the possibility to curtail demand if the cost of satisfying it \nexceeds a pre-set value of loss load parameter (set at 3,000 euros in \nour analysis). GenX also allows for demand-side response capabilities, \nwhich allow to anticipate or delay certain loads at a cost. The full solu-\ntion of GenX dispatch is characterized by hourly generation profiles for \neach technology in each country/area and the hourly electricity flows \nacross countries/areas (given the restrictions imposed by network \ncapabilities). The Lagrange multiplier associated with the increase of \n1 MWh in demand for a given country at a given hour is also calculated. \nAny technology in the model can set the marginal price, including vari-\nable renewables, storage, interconnectors or demand-side response.\nThe practice in European electricity markets is to use a common \ncoupling algorithm (PCR-Euphemia43) which, every day, finds the com-\nbination of generation and demand bids that maximizes total surplus, \nsetting prices using a pay as settled mechanism. Let di (y, h, p) be the \nelectricity demanded by agent i in hour h of the year y for a price p, and \nlet sj (y, h, p) be the electricity injected to the grid by agent j, again \nspecified for hour h of year y, at sale price p. Because the algorithm \nmaximizes total surplus aggregate for all agents, the solution will yield \nthe result of a competitive market where the settlement price equalizes \naggregate demand and aggregate supply at that price. In other words, \np* (y, h), is such that \u2211i di (y, h, p\u2217) = \u2211j s j (y, h, p\u2217).\nUnder the assumption of perfect competition and perfect fore-\nsight for the full year, the dispatch outcome of GenX would maximize \nthe total economic surplus of generators and buyers in the day-ahead \nmarkets, replicating the result of a competitive market coupled with \nEuphemia. The Lagrange multipliers obtained would match the pay as \nsettled prices that would be set in this market. Hence, we take prices \nobtained from GenX dispatch as a simulation of the outcome that would \nbe obtained in European day-ahead markets. We acknowledge that \nthis is just an approximation because the trading strategies of market \nparticipants may deviate from perfect competition and the predict-\nability of demand, weather and so on is not perfect. Economic theory \nwould suggest that large deviations from the perfect competition and \nperfect information setting should be capped to the extent that market \nparticipants can learn rationally and there is the possibility of free entry. \nIn any event, the perfect competition case which we analyse here must \nbe considered as a benchmark.\n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n338\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\nIt is important to emphasize that our results are based on \ncountry-level zones, where we aggregate all bidding zones within a \ncountry into one. For some of the smaller countries, we aggregate them \ninto hypothetical regional zones. Hence, our results do not account for \ntransmission restrictions within countries and across smaller coun-\ntries. ERAA data would allow analyst to carry out the same analysis \nwith the full set of bidding zones and transmission capacities, but \nthe computational requirements of doing so would grow notably \nand make it unfeasible in the context of this paper. The impact of this \nsimplification on the results shown in this paper is likely to be different \nfor each bidding zone. In zones with high renewables and low transmis-\nsion capacity, the probability that renewable power set the price of \nelectricity would be larger than indicated in our results, hence these \nzones would be expected to have lower sensitivity to fuel prices than \nwhat we present. The reverse situation would occur in zones with low \nrenewables and low transmission capacity. Whereas our results do \nnot model balancing costs, these would also be expected to be higher \nfor bidding zones with smaller grids and high renewable penetration. \nAdditionally, we cannot model local node constraints and balancing \ncosts derived from the variability of renewables44.\nEmpirical distribution of price outcomes\nFor each of the 300 future scenarios\u2013repetitions obtained, the GenX \ndispatch model is run. Because the solution of the model is intensive in \ncomputing capability, the University of Cambridge High Performance \nComputing Resources are used for this purpose. The result is a set of \n300 full simulations of the hourly market clearing outcomes in the \nfuture horizon year (2024 and 2030), where all repetitions share the \nsame capacity configuration (including the network capabilities) but \ndiffer in the national demands, the fuel prices and the weather condi-\ntions being simulated.\nWe introduce here a stochastic characterization of electricity \nprices and their annual averages. The objective is to provide a math-\nematically solid foundation to our proposed measures of the instability \nof electricity prices.\nLet y denote a vector summarizing the demand, fuel price and \nweather conditions prevailing in an electricity market in future sce-\nnario. Let p* (y, h) be the equilibrium price set at a bidding area at \nhour h for those conditions. Given KY, the capacity mix being analysed \non the horizon year Y, over the 8,760\u2009h of a year, the GenX algorithm \nyields a corresponding number of hourly equilibrium prices for each \nbidding area. From these, P(y), the annual price of year y is defined as \nthe arithmetic average of hourly prices in the year:\nP ( y) =\n1\n8, 760\n8,760\n\u2211\nh=1\np\u2217(y, h)\n(1)\nWe are interested in the stochastic properties of the conditional \ndistribution F (P ( y) |Ky = KY). Mathematically, we seek to understand \nhow different measures of instability of annual electricity prices would \nevolve for different values of KY. Research has covered the variance, \nskewness and kurtosis of hourly, daily or weekly prices5,9,10,45,46. These \ncharacterizations of hourly prices are key for several uses, including \nthe operation of batteries or the charging of electric vehicles45.\nAs explained, for each capacity configuration, we obtain 300 dif-\nferent scenarios (that is, 300 repetitions of y), which provide an empirical \ncounterpart to the stochastic distribution F (P (y) , |, Ky = KY). Our results \nare based on the analysis of the empirical distribution so obtained.\nMeasures of volatility\nThe analysis in this article focuses on three key measures: annual vola-\ntility, percentiles of annual prices and the sensitivity to gas prices. \nThroughout, we use the subindex K to indicate that the measures of \nelectricity price are to be derived from the conditional distribution \ngiven capacity, as explained in the previous paragraph.\nAnnual volatility (\u03c3) is the standard deviation of the annual aver-\nage price. This is the common definition of volatility for a stochastic \nvariable.\n\u03c3 (K) = \u221aEK(P ( y) \u2212EK (P ( y)))\n2\n(2)\nPercentiles of the distribution of the annual price (pctile(K)). \nThe standard deviation measure is a very partial description of the \nvariability of the annual price, for two reasons. First, if the concern is \nwith the probability of annual prices being very high, this parameter \ndepends both on the standard deviation and the expected value of the \nannual price. Increasing shares of renewables may (in fact, are expected \nto) lower the expected price of electricity, so that a higher \u03c3 does not \nnecessarily mean a higher probability of extreme prices. Second, the \ndistribution of annual prices are expected to diverge from the normal \ndistribution, even under the assumption that the joint distribution of \ndemand and fossil fuel prices is normal. The estimates of the distribu-\ntion obtained with the procedures in this paper reject the null hypoth-\nesis of normality in all but two countries, with clear indications that \nprice distributions become bimodal in some cases (Supplementary \nSection 4) and a tendency for positive skewness. In view of this, we use \nthe percentiles of the distribution of the annual price on Table 2. We \nuse the 85th percentile as a reference for high but reasonably frequent \nprices and the 95th percentile as the reference for extremely high but \nvery infrequent prices.\npctile85 (K ) = p \u2236FK ( p) = 0.85\n(3)\npctile95 (K ) = p \u2236FK ( p) = 0.95\n(4)\n\u03b2-sensitivity (\u03b2(K)): note that neither the volatility nor the percen-\ntile measures can gauge the influence of natural gas prices on electricity \nprices. To measure this, we propose \u03b2 the linear sensitivity of the annual \nprice of electricity to changes in the price of natural gas. \u03b2 is derived \nfrom the linear projection of the average annual price on the price of \ngas and other control variables X.\nLK (P ( y) , |, Pgas, X) = \u03b2 (K ) Pgas + \u03c4\u2032X\n(5)\nwhere LK (P ( y) , |, Pgas, X) denotes the linear projector operator. For each \ncountry and capacity configuration, we estimate the \u03b2-sensitivity with \na regression of the observed annual price of electricity to the price of \nnatural gas, and the only control variable we include is the price of coal. \nRegarding the choice of control variables, we only include the price of \ngas and coal because they make the interpretation of \u03b2-sensitivity more \nintuitive. Suppose that we added, for example, demand variables as a \nregressor. Then, the interpretation of a \u03b2-sensitivity of 1 would be that \na 1 euro increase in the price of natural gas, holding demand factors \nconstant, would be projected to lead to a 1 euro increase in the price of \nelectricity. However, natural gas prices and aggregate activity are clearly \nrelated in Europe, and hence a shock to gas prices would lead to a reduc-\ntion in energy demand, including electricity. Because we only include \nthe price of natural gas in the regression, we can think about our esti-\nmates as best linear projections: that is, they would reflect the joint \nimpact of natural gas both through the market clearing in electricity \nmarkets and their knock-on effect on aggregate demand. A \u03b2-sensitivity \nof 1 means that one euro increase in the price of natural gas would give \na best projection on the price of electricity that is one euro lower, both \nbecause of the direct effect on market clearing and its correlation with \ndemand. Note also that given the variability of annual prices due to \nother factors (weather, demand) in the simulation, the estimates of \n\u03b2-sensitivity are subject to sampling error. This explains that the esti-\nmated coefficient may be negative in some countries, where the actual \ntrue value would be 0 or close to 0. Supplementary Section 5 provides \nfurther clarification on the properties of \u03b2-sensitivity.\n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n339\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\nRelation between power price volatility and insurance value of \nrenewables\nConsider the following idealized model for electricity price formation \nwhereby changes in the price of electricity relative to their average are \na linear function of changes in fossil fuel prices (\u0394pf), the component \nof seasonally adjusted electricity demand (d), weather conditions (w) \nand other random factors (u).\n\u0394pe = \u03b2\u0394pf + \u03c8d + \u03bbw + e\n(6)\nFor simplicity, in this model \u0394pf, d, \u03bb w, e are assumed to be uncor-\nrelated, so that the variance of electricity price (changes) is just the \ncontribution of the variability of fuels and the variability of weather, in \nboth cases considering the sensitivity to each of these factors, plus the \nvariability of the other, unspecified, factors. Note that in our simula-\ntions we do not assume this: in our Monte Carlo design the correlation \nbetween fossil fuel prices and demand replicates their covariation in the \nperiod 1990 to 2021. Even holding the volatility of fossil fuel prices con-\nstant, additional investment in renewable capacity (which we denote \nby kres) gives rise to countervailing effects as shown in equation (2):\nd\u03c32\npe\ndkres\n||||\u03c3pf=constant\n= d\u03b22\ndkres\n\u03c32\npf + d\u03c82\ndkres\n\u03c32\nd + d\u03bb2\ndkres\n\u03c32\nw\n(7)\nOn the one hand, renewables can potentially reduce the variability \nof electricity prices\u2014and the likelihood of protracted spikes in electric-\nity prices, such as those observed in 2022\u2014as they lower the sensitivity \nof power prices to future shocks in the prices of fossil fuels (lower \u03b2). \nHowever, at the same time, systems with higher renewable capacity \nmay be more exposed to fluctuations in demand and weather condi-\ntions (higher \u03c8 and \u03bb). Our simulation design allows us to obtain \nd\u03c32\npe\ndkres\n|||\u03c3pf=constant, that is, the change in the variance of electricity prices \n(Table 1 in the main text), and hence provide a quantification of which \nof the countervailing effects prevails. Moreover, whereas in this frame-\nwork, we will use variance as the key statistic affecting the insurance \nvalue, in more sophisticated formulations it would be possible to use \nother statistics describing the distribution of electricity prices, includ-\ning extreme values and the \u03b2-sensitivity. Our simulation design also \nproduces estimates of these values (Tables 2 and 3).\nWe now show how the estimates of the changes in these parameters \nwould feed into the calculation of the societal value of renewables \nattributable to their smoothing of economic volatility. The starting \npoint is a representation of societal preferences over uncertain future \nconsumption paths {Ct} that admits a representation in terms of the \nexpected discounted value of the values of an instantaneous utility \nfunction u(), using a discount factor \u03b8. The societal lifetime utility of \nan uncertain consumption path is then given by E0 [\u2211\n\u221e\n0 \u03b8tu(Ct)].\nLet \u0304Ct = E0 [Ct] so that \u2211\n\u221e\n0 \u03b8tu( \u0304Ct) is the lifetime societal utility of a \nstable, certain path in which consumption at each period matches the \nexpected value of the uncertain consumption path. The annual cost of \neconomic variability \u03c9 is the amount that would need to be subtracted \nto the certain consumption path at every period so that its utility would \nmatch that of the uncertain path:\nE0 [\u2211\n\u221e\n0 \u03b8tu(Ct)] = \u2211\n\u221e\n0 \u03b8tu( \u0304Ct \u2212\u03c9)\n(8)\nIf u() represents the preferences of a risk-averse society, it must \nbe true that \u03c9 is a positive quantity. In other words, society is worse off \nby an amount equivalent to \u03c9 every year because it gets the uncertain \npath, compared to a situation where it would get with certainty the \nexpected value of consumption.\nThe insurance value of an increase in renewables is the reduction \nof this cost of uncertainty when renewables increase: \nd\u03c9\ndkres. A useful way \nto characterize the insurance value of renewables is to decompose it \nas follows:\nd\u03c9\ndkres\n= d\u03c9\nd\u03c32\nc\nd\u03c32\nc\ndkres\n(9)\nwhere \nd\u03c9\nd\u03c32\nc is the general value of stabilization (that is, the welfare gain \nassociated with a less volatile consumption path) and \nd\u03c32\nc\ndkres is the stabi-\nlization of consumption achieved through renewable investment. As \ndiscussed in the main section of the paper, the higher the value of \nstabilization and the higher the stabilization achieved by renewables, \nthe higher the insurance value of renewables. The discussion on the \nvalue of stabilization started with Lucas scepticism, but recent evidence \npoints to large effects of economic fluctuations on societal welfare. \nThis is consistent with the prevalence of stabilization policies in the \nfiscal and monetary realms.\nWe now turn to showing the relationship between \nd\u03c32\npe\ndkres\n|||\u03c3pf=constant, \nthe parameter we estimate, and \nd\u03c32\nc\ndkres and \nd\u03c9\nd\u03c32\nc. Consider a situation where \nconsumption at time t is driven by the price of fuels and the price of \nelectricity, together with other factors (unspecified but assumed unre-\nlated with energy), for example:\n\u0394c = \u03b3e\u0394pe + \u03b3f\u0394pf + \u03f5\n(10)\nThis specification recognizes that what matters for aggregate \noutcomes is the price of energy, not only of electricity, and that differ-\nent forms of energy can have different impacts. Fossil fuels are used in \nthe economy for purposes other than electricity generation. In a social \nwelfare analysis, the general equilibrium effects of increased renewable \ninvestment cannot be neglected. Hence, a change in the variance of \nconsumption associated with faster renewable deployment impacts \nmacroeconomic volatility not only through its effect on the \u03b2 of elec-\ntricity prices to fuel prices and the sensitivity to weather and demand, \nbut also through the change in the relevance of fossil fuels\u2019 and electric-\nity prices in aggregate fluctuations (\u03b3f and \u03b3e, respectively) and the \nchange in the volatility of fossil fuel prices. For simplicity, we maintain \nthe assumption that \u0394pf, d, \u03bb w, u and  are uncorrelated, which implies \nthat the variance \u03c32\nc is the weighted sum of their variances with appro-\npriate weights derived in a straightforward fashion from \u03b2, \u03c8, \u03bb, \u03b3e and \n\u03b3f. Differentiation of \u03c32\nc with respect to kres yields the following \nexpression:\nd\u03c9\ndkres =\nd\u03c9\nd\u03c32\nc [(\u03b32\ne\nd\u03c32\npe\ndkres\n|||\u03c32\npf=constant\n) +\n(\nd\u03b32\ne\ndkres \u03c32\npe +\nd\u03b32\nf\ndkres \u03c32\npf) + [(\u03b32\ne\u03b22 + \u03bb2\nf )\ndr2\npf\ndkres ]]\n(11)\nThis expression confirms that the insurance value of renewables \nincreases, the greater the reduction in the volatility of electricity prices \ndue to investment in renewables, that is, as \nd\u03c32\npe\ndkres\n|||\u03c32\npf=constant is more nega-\ntive. Coupled with the results in the literature on stabilization \npolicies27,28, our finding of comparatively large reductions in the vari-\nability of electricity prices would suggest a large insurance value. Note, \nhowever, that other general equilibrium effects of renewable invest-\nments would need to be considered too. These could occur through \nchanges in the respective influence of electricity prices and fuel prices \nin the economy (the second term in equation (11)). Presumably, as long \nas the volatility of electricity prices remained lower than that of fossil \nfuels\u2019 prices (recall this is annual frequency data), a shift towards elec-\ntrification would increase the insurance value of renewables through \nthis effect too. However, the impact of additional renewable investment \non the volatility of fuel prices is itself a key consideration (namely, the \nthird term in equation (11)). It might occur that additional renewables, \nby lowering the demand for fossil fuels and triggering unstable equi-\nlibria in their markets, could increase the volatility of fuel prices. These \neffects, if material, would tend to offset the insurance value of \nrenewables.\n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n340\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\nData availability\nAll data used in this study are available online in the sources referenced \nin Methods. The inputs to GenX are available from the corresponding \nauthor upon reasonable request. Source data are provided with this \npaper.\nCode availability\nThe GenX electricity system capacity expansion model can be found via \nGitHub at https://github.com/GenXProject/GenX. The code required \nto read the source data into GenX, sample code for the modifications \nto GenX to run the Monte Carlo simulations and generate the rep-\netitions are available via Github at https://github.com/DanielNavia1/\nReadERAAdata.\nReferences\n1.\t\nvon der Leyen, U. Statement by President von der Leyen on \nenergy. European Commission https://ec.europa.eu/commission/\npresscorner/detail/en/speech_22_5389 (2022).\n2.\t\nBistline, J., Mehrotra, N. & Wolfram, C. Economic Implications of \nthe Climate Provisions of the Inflation Reduction Act (NBER, 2023); \nhttps://doi.org/10.3386/w31267\n3.\t\nGuterres, A. Amid backsliding on climate, the renewables effort  \nnow must be tripled. Washington Post https://www.washingtonpost. \ncom/opinions/2022/04/04/new-ipcc-climate-report-on- \naverting-catastrophe/ (2022).\n4.\t\nEnergy Security in Energy Transitions (IEA, 2022); https://www.iea. \norg/reports/world-energy-outlook-2022/energy-security-in- \nenergy-transitions\n5.\t\nWoo, C. K., Horowitz, I., Moore, J. & Pacheco, A. The impact of \nwind generation on the electricity spot-market price level and \nvariance: the Texas experience. Energy Policy 39, 3939\u20133944 \n(2011).\n6.\t\nProposal for a regulation of the European Parliament and of the \ncouncil amending regulations (EU) 2019/943 and (EU) 2019/942 \nas well as directives (EU) 2018/2001 and (EU) 2019/944 to improve \nthe union\u2019s electricity market design. Council of the European \nUnion https://data.consilium.europa.eu/doc/document/ST-14339-\n2023-INIT/en/pdf (2023).\n7.\t\nMallapragada, D. S. et al. Electricity pricing challenges in future \nrenewables-dominant power systems. Energy Econ. 126, 106981 \n(2023).\n8.\t\nSepulveda, N. A., Jenkins, J. D., Edington, A., Mallapragada, D. S. & \nLester, R. K. The design space for long-duration energy storage in \ndecarbonized power systems. Nat. Energy https://doi.org/10.1038/\ns41560-021-00796-8 (2021).\n9.\t\nKlinge Jacobsen, H. & Zvingilaite, E. Reducing the market impact \nof large shares of intermittent energy in Denmark. Energy Policy \n38, 3403\u20133413 (2010).\n10.\t Tveten, T. G., Bolkesj\u00f8, T. F., Martinsen, T. & Hvarnes, H. Solar \nfeed-in tariffs and the merit order effect: a study of the German \nelectricity market. Energy Policy 61, 761\u2013770 (2013).\n11.\t\nZakeri, B. et al. The Role of Natural Gas in Electricity Prices in \nEurope (UCL Institute for Sustainable Resources, 2023).\n12.\t Kosch, M., Blech, K. & Abrell, J. Rising electricity prices in Europe: \nthe impact of fuel and carbon prices. Preprint at SSRN  \nhttps://doi.org/10.2139/ssrn.4259209 (2022).\n13.\t Mu\u00f1oz, J. I. & Bunn, D. W. Investment risk and return under \nrenewable decarbonization of a power market. Clim. Policy 13, \n87\u2013105 (2013).\n14.\t J\u00e5stad, E. O., Trotter, I. M. & Bolkesj\u00f8, T. F. Long term power prices \nand renewable energy market values in Norway\u2014a probabilistic \napproach. Energy Econ. 112, 106182 (2022).\n15.\t Winkler, J., Pudlik, M., Ragwitz, M. & Pfluger, B. The market value \nof renewable electricity\u2014which factors really matter? Appl. \nEnergy 184, 464\u2013481 (2016).\n16.\t Mercure, J. F. et al. Reframing incentives for climate policy action. \nNat. Energy https://doi.org/10.1038/s41560-021-00934-2 (2021).\n17.\t Lucas, R. E. Models of Business Cycles (Basil Blackwell, 1987).\n18.\t European Resource Adequacy Assessment\u20142022 Edition \n(ENTSO-E, 2022); https://www.entsoe.eu/outlooks/eraa/2022/\n19.\t MIT Energy Initiative and Princeton University ZERO lab. GenX: \na configurable power system capacity expansion model for \nstudying low-carbon energy futures. Github https://github.com/\nGenXProject/GenX (2022).\n20.\t Biggar, D. R. & Hesamzadeh, M. R. The Economics of Electricity \nMarkets (Wiley, 2014).\n21.\t Blume-Werry, E., Faber, T., Hirth, L., Huber, C. & Everts, M. Eyes on \nthe price: which power generation technologies set the market \nprice. Econ. Energy Environ. Policy 10, 1\u201314 (2021).\n22.\t Plant loads, demand and efficiency of major power \nproducers. DESNZ https://assets.publishing.service.gov.uk/\nmedia/64c128ebd4051a00145a93eb/DUKES_5.10.xlsx (2023).\n23.\t Newbery, D. M. High renewable electricity penetration: marginal \ncurtailment and market failure under \u2018subsidy-free\u2019 entry. Energy \nEcon. 126, 107011 (2023).\n24.\t Pe\u00f1a, J. I., Rodr\u00edguez, R. & Mayoral, S. Cannibalization, \ndepredation, and market remuneration of power plants. Energy \nPolicy 167, 113086 (2022).\n25.\t L\u00f3pez Prol, J., Steininger, K. W. & Zilberman, D. The cannibalization \neffect of wind and solar in the California wholesale electricity \nmarket. Energy Econ. 85, 104552 (2020).\n26.\t Barrage, L. & Nordhaus, W. Policies, projections, and the social \ncost of carbon: results from the DICE-2023 model. Proc. Natl \nAcad. Sci. USA 121, e2312030121 (2024).\n27.\t Jord\u00e0, \u00d2., Schularick, M. & Taylor, A. M. Disasters everywhere: the \ncosts of business cycles reconsidered. IMF Econ. Rev. 72, 116\u2013151 \n(2024).\n28.\t Barro, R. J. Rare disasters, asset prices, and welfare costs. Am. \nEcon. Rev. 99, 243\u2013264 (2009).\n29.\t Brown, S. P. A. & Huntington, H. G. Assessing the U.S. oil security \npremium. Energy Econ. 38, 118\u2013127 (2013).\n30.\t Fabra, N. The energy transition: an industrial economics \nperspective. Int. J. Ind. Organ. 79, 102734 (2021).\n31.\t Grubb, M., Drummond, P. & Maximov, S. Separating Electricity \nfrom Gas Prices through Green Power Pools: Design Options and \nEvolution (Institute for New Economic Thinking, 2022).\n32.\t Hogan, W. W. Electricity market design and zero-marginal cost \ngeneration. Curr. Sustainable Renewable Energy Rep. 9, 15\u201326 \n(2022).\n33.\t Final Assessment of the EU Wholesale Electricity Market Design \n(ACER, 2022); https://www.acer.europa.eu/Publications/Final_\nAssessment_EU_Wholesale_Electricity_Market_Design.pdf\n34.\t Meeus, L. Electricity market reform: what is (not) in the European \nCommission proposal. EUI https://fsr.eui.eu/publications/?hand\nle=1814/75580 (2023).\n35.\t Beiter, P., Guillet, J., Jansen, M., Wilson, E. & Kitzing, L. The \nenduring role of contracts for difference in risk management and \nmarket creation for renewables. Nat. Energy https://doi.org/ \n10.1038/s41560-023-01401-w (2023).\n36.\t Mays, J., Morton, D. P. & O\u2019Neill, R. P. Asymmetric risk and fuel \nneutrality in electricity capacity markets. Nat. Energy 4, 948\u2013956 \n(2019).\n37.\t Williges, K., Van der Gaast, W., de Bruyn-Szendrei, K., Tuerk, A. & \nBachner, G. The potential for successful climate policy in national \nenergy and climate plans: highlighting key gaps and ways \nforward. Sustain. Earth 5, 1 (2022).\n38.\t Regulation on the governance of the energy union and  \nclimate action (EU)2018/1999. European Commission  \nhttps://eur-lex.europa.eu/legal-content/EN/TXT/?uri=uriserv: \nOJ.L_.2018.328.01.0001.01.ENG&toc=OJ:L:2018:328:TOC (2018).\n\nNature Energy | Volume 10 | March 2025 | 329\u2013341\n341\nArticle\nhttps://doi.org/10.1038/s41560-025-01704-0\n39.\t National energy and climate plans. European Commission  \nhttps://commission.europa.eu/energy-climate-change- \nenvironment/implementation-eu-countries/energy-and-climate- \ngovernance-and-reporting/national-energy-and-climate-plans_en \n(2023).\n40.\t European Resource Adequacy Assessment \u2013 2022 Edition \u2013 \nInput Data (ENTSO-E, 2022); https://www.entsoe.eu/outlooks/\neraa/2022/eraa-downloads/\n41.\t European Resource Adequacy Assessment \u2013 2022 Edition \u2013 Annex \n2: Methodology (ENTSO-E, 2022); https://eepublicdownloads.\nazureedge.net/clean-documents/sdc-documents/ERAA/2022/\ndata-for-publication/ERAA2022_Annex_2_Methodology.pdf\n42.\t Hodrick, R. J. & Prescott, E. C. Postwar US business cycles: an \nempirical investigation. J. Money Credit Banking 29, 1\u201316  \n(1997).\n43.\t Single Day-ahead Coupling (SDAC) (ENTSO-E, 2023); https://www.\nentsoe.eu/network_codes/cacm/implementation/sdac/\n44.\t Heptonstall, P. J. & Gross, R. J. K. A systematic review of the costs \nand impacts of integrating variable renewables into power grids. \nNat. Energy 6, 72\u201383 (2021).\n45.\t Gianfreda, A. & Bunn, D. A stochastic latent moment model  \nfor electricity price formation. Oper. Res. 66, 1189\u20131203  \n(2018).\n46.\t Gianfreda, A., Scandolo, G. & Bunn, D. W. Higher moments in the \nfundamental specification of electricity forward prices. Quant. \nFinance 22, 2063\u20132078 (2022).\nAcknowledgements\nL.D.A. acknowledges support of a Senior Fellowship award from the \nJM Keynes Fellowship Fund at the University of Cambridge.\nAuthor contributions\nD.N.S. and L.D.A. conceived this study. D.N.S. designed the \nmethodology, collected the data and led the execution of the analysis, \nwith L.D.A. supporting. D.N.S. wrote the paper with L.D.A. editing.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s41560-025-01704-0.\nCorrespondence and requests for materials should be addressed to \nDaniel Navia Simon.\nPeer review information Nature Energy thanks Angelica Gianfreda, \nWilliam Hogan and the other, anonymous reviewer(s) for their \ncontribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this licence, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2025\n\n\n    Scientific Research Findings:", "answer": "We define and estimate the \u03b2\u2011sensitivity metric, the increase in the annual average price of electricity when the price of natural gas increases by 1 euro per MWh. Achieving the 2030 capacities foreseen in system operators\u2019 resource adequacy assessments in all EU countries, the UK and Switzerland would reduce the average European \u03b2\u2011sensitivity from 1.4 to 1.0 in 2030. Deploying solar PV and wind 30% above the targets would further reduce it to 0.5, while volatility and price spikes would also be less intense. There is substantial heterogeneity across countries. The mechanisms that stabilize prices also result in lower revenues for renewable plants, limiting the potential market\u2011based penetration of these technologies. Private agents are unlikely to account for the societal value of lower macroeconomic volatility. Energy planning, market design, and support policies in Europe should consider this \u2018insurance value\u2019 of renewables explicitly, in addition to their benefits for climate mitigation.", "id": 2}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 10 | March 2025 | 380\u2013394\n380\nnature energy\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nDemand-side strategies enable rapid \nand deep cuts in buildings and transport \nemissions to 2050\n \nRik van Heerden\u2009\n\u200a\u20091\u2009\n, Oreane Y. Edelenbosch\u2009\n\u200a\u20091,2\u2009\n, Vassilis Daioglou\u2009\n\u200a\u20091,2, \nThomas Le Gallic\u2009\n\u200a\u20093,4, Luiz Bernardo Baptista\u2009\n\u200a\u20095, Alice Di Bella\u2009\n\u200a\u20096,7,8, \nFrancesco Pietro Colelli\u2009\n\u200a\u20096,7,9, Johannes Emmerling\u2009\n\u200a\u20096,7, Panagiotis Fragkos\u2009\n\u200a\u200910, \nRobin Hasse\u2009\n\u200a\u200911,12, Johanna Hoppe\u2009\n\u200a\u200911,12, Paul Kishimoto13, Florian Leblanc\u2009\n\u200a\u200914, \nJulien Lef\u00e8vre\u2009\n\u200a\u200915, Gunnar Luderer\u2009\n\u200a\u200911,12, Giacomo Marangoni6,7,16, \nAlessio Mastrucci\u2009\n\u200a\u200913, Hazel Pettifor17, Robert Pietzcker\u2009\n\u200a\u200911, Pedro Rochedo\u2009\n\u200a\u200918, \nBas van Ruijven\u2009\n\u200a\u200913, Roberto Schaeffer\u2009\n\u200a\u20095, Charlie Wilson\u2009\n\u200a\u200913,17, Sonia Yeh\u2009\n\u200a\u200919, \nEleftheria Zisarou\u2009\n\u200a\u200910 & Detlef van Vuuren\u2009\n\u200a\u20091,2\nDecarbonization of energy-using sectors is essential for tackling climate \nchange. We use an ensemble of global integrated assessment models to \nassess CO2 emissions reduction potentials in buildings and transport, \naccounting for system interactions. We focus on three intervention \nstrategies with distinct emphases: reducing or changing activity, improving \ntechnological efficiency and electrifying energy end use. We find that \nthese strategies can reduce emissions by 51\u201385% in buildings and 37\u201391% \nin transport by 2050 relative to a current policies scenario (ranges indicate \nmodel variability). Electrification has the largest potential for direct \nemissions reductions in both sectors. Interactions between the policies and \nmeasures that comprise the three strategies have a modest overall effect on \nmitigation potentials. However, combining different strategies is strongly \nbeneficial from an energy system perspective as lower electricity demand \nreduces the need for costly supply-side investments and infrastructure.\nDemand-side mitigation forms a critical part of strategies to meet the \nParis climate goals1\u20135, involving both consumer technology choices \nrelated to energy efficiency and energy sources, as well as lifestyle \nchanges. Lower energy demand reduces emissions and also allows \nfor greater flexibility in technology choices within supply sectors by \nlowering the overall energy production and associated investment \nrequirements2. Two critical demand-side sectors include buildings \n(encompassing residential and service-sector buildings) and trans-\nport (encompassing aviation, navigation and land transport) that \neach represented 29% of global final energy consumption in 20196 \nand, respectively, 19% and 7% of direct energy-related greenhouse \ngas emissions7. In the Sixth Assessment (AR6) Working Group (WG) \nIII report the Intergovernmental Panel on Climate Change (IPCC) for \nthe first time included a chapter on demand-side reductions (Ch. 5). \nIt estimates that demand-side options in buildings and land transport \ncould potentially lead to greenhouse gas emissions reductions of 66% \nand 67% by 2050, respectively2. Complementing this, a similar assess-\nment by Creutzig et al. also reports high potentials of 78% and 62% for \nthe same sectors3.\nThese figures represent median estimates derived from a review \nof \u2018bottom-up\u2019 assessments of individual demand-side mitiga-\ntion options found in existing literature. However, the reliance on \nbottom-up approaches faces limitations due to potentially inconsist-\nent assumptions regarding the effects of individual measures, varying \nReceived: 22 December 2023\nAccepted: 19 December 2024\nPublished online: 5 February 2025\n Check for updates\nA full list of affiliations appears at the end of the paper. \n\u2009e-mail: rik.vanheerden@pbl.nl; o.y.edelenbosch@uu.nl\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n381\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nAnalysis of modelled reduction strategies\nWe identified a comprehensive list of mitigation measures from rel-\nevant literature and developed scenarios to analyse their implications \nfor mitigation pathways. To ensure that the scenarios are credible and \npolicy-relevant, while accounting for all key factors, we involved poli-\ncymakers and demand-side experts. Through a survey, they evaluated \nthe feasibility of the narratives and measures and their potential for \nemissions reduction (Methods). The various emissions abatement \nmeasures are grouped into behavioural changes impacting activity, \ntechnical efficiency improvements and electrification of energy end \nuses. To analyse the impact of these groups and their interactions, we \ntranslated them into the following scenarios, each representing distinct \nintervention strategies (Table 1).\nThe activity-focused strategy (ACT) involves redesigning \nservice-provisioning systems to either reduce or shift consumption \nof energy and transport services. In buildings, this strategy includes \nreduction of average dwelling size, working in shared buildings with \nflexible use, adjusting thermostat settings to lower (heating) or higher \n(cooling) set points. In transport, it includes promotion of active modes \n(walking, biking), public and shared mobility options. Air travel is dis-\ncouraged, whereas advancements in freight logistics and speed restric-\ntions in maritime transport enable more efficient movement of goods.\nThe technology-optimizing strategy (TEC) focuses on improve-\nments in the efficiency of existing technologies. Higher levels of energy \nefficiency are achieved in both new constructions and existing build-\nings through increased renovation rates, improved thermal insulation \nand more efficient heating, ventilation and air conditioning (HVAC) \nsystems. Efficiency standards for road vehicles, aircraft and ships are \nimplemented. Environmental certification of airplanes and ships is \nrequired for using airports and ports.\nThe electrification-focused strategy (ELE) focuses on switching \nto electricity and alternative fuels. Heat pumps and electricity-based \nheating systems are widely adopted in buildings. Fossil fuels are \nphased out and new natural gas connections are banned in the global \nnorth. Passenger vehicles, light-duty trucks and ports transition to full \nelectrification. Diesel engines are phased out of heavy-duty vehicles, \nand biofuels and electrofuels are increasingly used in aviation and \nshipping.\nFinally, a combined approach, referred to as all interventions \n(ALL), integrates all the interventions from ACT, TEC and ELE.\nThese scenarios build upon the ASI (Avoid-Shift-Improve) frame-\nwork developed by Creutzig et al.3, which categorizes demand-side \nactions into the three components avoid, shift and improve. However, \nour approach provides a direct and consistent mapping to concrete \nintervention strategies and their impacts on the energy system and \nemissions (Methods). For instance, ASI categorizes installing heat \npumps as a \u2018shift\u2019 intervention, whereas adopting electric cars as \n\u2018improve\u2019. In our framework, both interventions are part of the electrifi-\ncation strategy, as they do not change the quantity or quality of energy \nservice provided and have very similar energy system implications.\nAs a starting point for the analysis, a \u2018middle-of-the-road\u2019 \nsocio-economic pathway (SSP2) is used together with current national \npolicies implementation (NPi)37,38. This includes the most important \npolicies per country adopted by national parliaments (Supplementary \nInformation 3). The scenarios use different gross domestic product \n(GDP) and population projections for each country to capture inter-\nregional heterogeneity. Subsequently, the additional policies and \nmeasures affecting demand-side sectors are explicitly modelled in \nscenarios. Although many measures are likely to impact industrial \nemissions, this impact is outside the scope of this study. To explore the \ninteractions with more stringent climate ambitions, we also consider \nscenarios aligned with 1.5\u2009\u00b0C global warming, implemented through \na carbon tax that limits the cumulative CO2 budget to 400\u2009Gt (peak) \nand 650\u2009Gt (by 2100) in the period 2020\u20132100. An overview of these \nscenarios is provided in Table 2.\nbaselines and the challenge of capturing interactions between multiple \nstrategies. System interactions that need to be considered include \nthe complementarity or overlap between options8 (for example, the \nelectrification of the vehicle fleet limits the potential for emissions \nreductions through improving combustion engine efficiency), the \ninteractions between energy-supply decarbonization and end-use \ntransformations9\u201311 and possible \u2018rebound effects\u2019 (when efficiency \nmeasures lead to unintended increase of the demand for energy ser-\nvices that can partly counterbalance the potential of activity-oriented \nmitigation options)12,13. These interactions and system impacts can \neither improve or reduce the effectiveness of individual strategies \nand have yet to be assessed in transformation pathways that require \nadequate modelling tools for doing so.\nWhereas recent literature has advanced the understanding of \ndemand-side mitigation, most scenario studies rely on single-model \napproaches, focusing on long-term global pathways3,14\u201317 and national \ntrajectories18\u201321 and critical factors such as material efficiency22,23. This \nreliance on single models limits the robustness of findings, as they do \nnot account for the structural uncertainties inherent in these models.\nIntegrated assessment models (IAMs) are specifically designed \nto capture interactions among the energy system, economy and envi-\nronment, making them valuable tools for assessing the potential of \ndemand-side intervention strategies in the buildings and transport \nsectors. These models not only facilitate the systematic assessment \nof the impacts of such strategies through scenario analysis but also \nincorporate the system-wide impacts that arise from their implemen-\ntation. Integrated modelling frameworks allow to build consolidated \ntrajectories whereas ensuring consistency not only between sectors \nand intervention strategies but also over time.\nHistorically, global IAMs have predominantly emphasized supply- \nside measures in global mitigation scenarios24,25, which has limited their \ncapacity to address demand-side pathways effectively. This has been \ncomplicated by the complexity of consumer groups and behaviour, \ndiverse sectors, services and technologies that depend on local cir-\ncumstances, climate and socio-economic conditions, infrastructures \nand technological development. However, in recent years, IAMs have \nimproved their representation of energy-demand sectors, particularly \nin the buildings and transport sectors. These advancements are driven \nby rapid technological advancements, such as electric mobility and \nheat pumps, necessitating ongoing model updates. The improvements \nencompass various aspects, including alternative electrification path-\nways in transport26,27, more diverse building types and expanded options \nfor renovations in buildings28\u201330 and improvements in the representa-\ntion of international transport31\u201334. Moreover, gradual advancements  \nhave been made in modelling behavioural transitions within IAMs35,36.\nOur study leverages the modelling improvements of multiple \nIAMs (COFFEE, IMACLIM-R, IMAGE, MESSAGEix-Buildings, PRO-\nMETHEUS, REMIND and WITCH) to analyse the relative importance \nof key demand-side mitigation strategies, examining specifically the \nareas of buildings and transport, including personal mobility and \nfreight transport. The main features of each model are provided in Sup-\nplementary Information 1. The seven models capture the interactions \nbetween measures and system effects to various degrees. Although \na complete disentangling of the contributions of these interactions \nwithin each model is beyond the scope of this study, our work aims to \nprovide a robust assessment of demand-side measures potentials and \ninteractions across the structural uncertainty of the models included. \nWe analyse interactions between strategies and apply decomposition \nanalysis to evaluate the scenario results, identifying key drivers of emis-\nsions reductions and revealing some interaction effects. Our findings \nhighlight the key role of electrification, yet underscore that a combined \napproach further reduces emissions and alleviates pressure on the \nenergy supply side. Our results also show that there is a considerable \nspread in the results across models, indicating substantial structural \nuncertainty about the complex dynamics drive the results.\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n382\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nMitigation potential and interactions\nIn the default current policy scenario (REF), direct CO2 emissions \nfrom buildings increase by \u22121% to 36% in 2030 and \u22128% to 31% in 2050 \ncompared with 2015 levels, whereas direct emissions from transport \nincrease by 5% to 32% in 2030 and \u221210% to 49% in 2050 (Fig. 1). This \nincrease is mainly caused by increasing final energy demand after \n2015. However, not all models also indicate a correspondingly strong \nincrease in emissions; in fact, some models even project a decrease \ndue to a more pronounced shift to less carbon-intense fuels. The wide \nvariation in projections across the models is closely related to how \nefficiency and changes in service demand (for example, elasticities or \nrelationships with economic activity) are modelled. In some models, \nincreasing activities result in increasing emissions, whereas in others, \nactivity growth is partially offset by efficiency improvements. Nonethe-\nless, in both sectors, final energy demand continues its upward trend \nin the majority of models.\nThe intervention strategies mitigate the increase in energy \ndemand and reduce the growth in direct emissions in both buildings \nand transport\u2014in a similar way across models. Emissions reductions \nfrom current levels are robust across models for both ELE and ALL, \nparticularly in buildings, but with more inter-model variation for the \nactivity-focused and technology-optimizing strategies. Emissions \nreduction potentials in 2030, with respect to the reference scenario \n(REF), are 3\u201316% (ACT), 3\u201319% (TEC) and 10\u201331% (ELE) of direct build-\nings emissions and 4\u201315% (ACT), 2\u201310% (TEC) and 3\u201317% (ELE) of direct \ntransport emissions (Fig. 2). The potentials become more substantial in \n2050 and reach 6\u201323% (ACT), 11\u201333% (TEC) and 45\u201377% (ELE) for build-\nings and 17\u201328% (ACT), 2\u201367% (TEC) and 22\u201386% (ELE) for transport.\nWe also estimated the impact of all measures combined as a \nproduct of the impact of each individual strategy, shown as the \nno-interaction estimate in Fig. 1. This approach assumes that the meas-\nures interact independently. Comparison with ALL shows that interac-\ntion between measures from different strategies actually play a role, \nalthough their impact on the overall mitigation potential is limited. In \nthe buildings sector, the models project an effective potential in 2050 \nthat is lower than the no-interaction estimate. However, considering \nTable 1 | Overview of the demand-side strategies and references supporting the underlying assumptions with further details \nand motivation for the assumptions provided in Methods\nScenario\nDescription and assumptions\nSupporting references\nActivity reduction and \nshifts (ACT)\nBuildings\nPolicies limiting floorspace in new constructions, along with flexible use of buildings and shared building \nspaces (such as co-housing and working), reduce per capita floorspace in both the residential and \ncommercial sectors. Shifts in household preferences and policies that limit new construction of single-family \nhouses increase the share of multi-family houses. Stimulated by information campaigns and restrictive \npolicies, the set-point temperatures in buildings shift to 20\u2009\u00b0C for heating and 25\u2009\u00b0C for cooling.\n16,70\u201373\nTransport\nMeasures such as congestion charges and growing prevalence of remote working reduce demand for \ndriving private vehicles within cities, and more bike lanes and pedestrian zones increase the adoption of \nactive modes (bicycles, e-scooters and walking), whereas improvements in public transport infrastructure, \nlast-mile services and free/lower public transport fares increase the adoption of public transport. Also \ncar-sharing/pooling is actively promoted. In the aviation sector, passenger transport is reduced by fuel taxes \n(by abolishing tax exemptions), movement taxes such as a frequent flyer levy and development of increased \nvirtual connectivity. Also freight transport is affected by fuel taxes and movement taxes. Policies encourage \ndevelopment of local manufacturing and storage, and improved road freight logistics reduce road freight \nactivity. Worldwide speed restrictions are introduced in maritime transport (slow steaming shipping) and \nshort-haul air travel is phased out by 2030.\n14,32,73\u201379\nTechnological \nimprovements (TEC)\nBuildings\nBuilding codes and standards, energy performance certification and subsidies are implemented to reduce \nenergy intensity by improving insulation levels and enhancing overall energy efficiency per surface area. \nThis results in more efficient heating, ventilation and air conditioning. Additionally, the current retrofit rate \ndoubles to 2% per year in the global north.\n16,30,73,80,82\nTransport\nEfficiency standards result in efficiency improvements across all new passenger vehicles, trucks, airplanes \nand ships. In addition, environmental certification for using airports and ports further drives efficiency \nimprovements across the entire aviation and shipping fleets.\n83,84\nElectrification and fuel \nshift (ELE)\nBuildings\nFuel mandates accelerate electrification and heating fuel switching, with heat pumps in all new buildings \nby 2030. By 2050, we assume that 70% of space and water heating is electricity based. Building regulations \nand neighbourhood-based approaches also stimulate deployment of on-site, and building-integrated \nrenewables and renewable energy (photovoltaics and thermal solar) meet 50% of the demand for cooling \nand heating in the global north by 2050. New natural gas connections for heating are banned in the global \nnorth by 2030, and non-clean heating fuels are phased out by 2050.\n71,73,85\nTransport\nFuel/technology mandates ensure full electrification of passenger vehicles and light-duty trucks by 2040, \nusing battery or fuel cell electric vehicles. In the fleet of heavy-duty vehicles, diesel engines are phased \nout by 2040. By 2030, ports are fully electrified, reducing ships\u2019 reliance on auxiliary engines, and vessels \nmeet zero-emissions berth standards by 2040. Fuel standards/mandates, infrastructure development and \nremoving blending restrictions increase the use of alternative fuels (biofuels/electrofuels) in international \ntransport. After 2050, electric short-haul planes become available.\n73,76,86\u201391\nTable 2 | Overview of the scenarios with the implementation \nof the scenarios detailed in Methods\nAdditional policies affecting \ndemand-side sectors\nClimate ambitions\nCurrent national policies (NPi)\n1.5\u2009\u00b0C scenario\nNone (REF)\nNPi-REF\n1.5C-REF\nActivity-focused (ACT)\nNPi-ACT\n1.5C-ACT\nTechnology-optimizing (TEC)\nNPi-TEC\n1.5C-TEC\nElectrification-focused (ELE)\nNPi-ELE\n1.5C-ELE\nAll interventions (ALL)\nNPi-ALL\n1.5C-ALL\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n383\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nonly the interactions represented in the models, the accumulation \nof the various measures still remains largely effective. In the trans-\nport sector, remarkably, four models simulated an effective potential \nslightly greater than the no-interaction estimate (IMAGE, IMACLIM-R, \nPROMETHEUS, REMIND).\nExamples of counteracting interactions include the diminished \nimpact of improved insulation on emissions reductions when inef-\nficient boilers are replaced with electric heat pumps. Likewise, lower-\ning set-point temperatures reduces the potential emissions savings \nthat could otherwise be achieved through enhanced insulation or \nthe adoption of heat pumps. In transport, the additional impact of \nefficiency standards is limited as the vehicle fleet becomes predomi-\nnantly electric, given the much higher efficiency of electric motors. \nConversely, certain measures can amplify the effectiveness of others, \nwhich is evident in the transport results for some models. For instance, \nas electrification policies increase the market share of electric vehicles \n(EVs), declining costs from learning effects enhance their competitive-\nness. At the same time, car-sharing services also reduce the effective \ncost of higher-priced EVs, and together these factors can accelerate \nthe transition.\nUnder a 1.5\u2009\u00b0C climate ambition, the implementation of demand- \nside measures generally results in emissions reductions greater than \nthose in 1.5C-REF, with WITCH being an exception and some mod-\nels showing only marginal differences. The COFFEE results show a \nshort-term rebound effect in oil consumption for TEC, ELE and ALL. \nWith perfect foresight, COFFEE anticipates a decrease in long-term \nfossil fuel demand, leading to price reductions and an increase in \nshort-term fossil fuel consumption.\nInterestingly, there is no consensus among the models about the \nmost effective strategy under 1.5\u2009\u00b0C climate ambition. This could be \nattributed to variations in models\u2019 responses to carbon tax, leading \nto utilization of different mitigation options. Conventional mitiga-\ntion scenarios such as 1.5C-REF include measures that are assumed to \nbe cost effective, thereby leaving part of the demand-side mitigation \npotential unexploited. Such scenarios are often implemented through \na (globally uniform) carbon tax, whereas demand-side measures\u2014such \nas accelerating technology adoption and modal shifting\u2014can be cost \ninsensitive.\nThe integration of all intervention strategies (ALL) can reverse the \ntrend of rising emissions in buildings and transport. Without additional \nclimate policies, the buildings sector could reduce CO2 emissions, on \naverage across models, by 63% (51\u201385%) and the transport sector by \n70% (37\u201391%) in 2050. The combination of strategies achieves sectoral \nemissions reductions compatible with a 1.5\u2009\u00b0C pathway (Fig. 2).\nInterestingly, these findings are comparable to\u2014but independ-\nent from\u2014reported potentials in the IPCC\u2019s AR6 WG III report that \nuse bottom-up studies to estimate emissions reduction potentials in \n2050 relative to the IEA WEO (International Energy Agency's World \nEnergy Outlook) stated-policy scenario of 66% (40\u201391%) for buildings \nand 67% (44\u201388%) for land transport2. Ranges are not explicitly men-\ntioned in the report, but we derived them by multiplying the lowest \nand highest reported potentials for the considered mitigation options, \nconsistent with the computation of median estimates. However, in \ncontrast to this study, the IPCC figures relate to total emissions includ-\ning indirect emissions from electricity generation. Whereas IAMs can \nquantify these indirect emissions, assessing them within the context \n+60\n+40\n+20\n0\n\u201320\n\u201340\nRelative to 2015 (%)\nFinal energy\nCO2 emissions\n\u201360\n\u201380\n+40\n+20\n0\n\u201320\n\u201340\n\u201360\n\u201380\n\u2013100\n2030\nBuildings\nTransport\nREF\nACT\nTEC\nELE\nAll\nNo-interaction estimate\nScenario (NPi)\na\nb\nc\nd\nCOFFEE\nIMACLIM-R\nIMAGE\nPROMETHEUS\nREMIND\nWITCH\nMESSAGEix-Buildings\nModel\n2030\n2050\nYear\n2050\nFig. 1 | Change in global final energy use and direct CO2 emissions from fuel \ncombustion in the buildings and transport sectors. a,b, Final energy use \nin buildings (a) and transport (b). c,d, CO2 emissions from buildings (c) and \ntransport (d). Results are presented relative to 2015 levels to reduce model \ndifferences resulting from calibration against different historical datasets. \nAll scenarios have current NPi. Markers indicate individual model results and \nbars depict the model ranges. The grey hatches and markers represent the \nno-interaction estimates, which approximate the combined impact of each \nindividual strategy relative to NPi-REF by multiplying their respective effects. \nMESSAGEix-Buildings results are shown, but for cross-sectoral consistency are \nnot factored into the averages and ranges (Methods). Projections for activity \npatterns in the reference scenario (floorspace, passenger-kilometres and freight \ntonne-kilometres) are provided in Supplementary Information 5. Subsectoral \nprojections for the residential, commercial, passenger and freight sectors are \navailable in Supplementary Information 8.\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n384\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nof demand-side potentials poses challenges because this relies on \nthe degree of decarbonization in the upstream electricity generation \nsector that varies vastly across scenarios and models (Supplementary \nInformation 10). When indirect emissions are considered, our study \nshows considerably lower emissions reduction potentials, particularly \nin buildings. This difference may stem from the representation of \nrooftop solar systems, which are generally not explicitly accounted for \nwithin IAMs, along with the limited representation of the commercial \nsector. Also, nuanced disparities arise from differing scopes. Whereas \nour study covers reduced international transport, improved logistics \nand fuel shifts for aviation and shipping, the IPCC synthesis includes \nadditional measures for shipping (for example, weather routing) and \nbuildings (for example, shorter showers, smarter energy use and nature- \nbased solutions).\nThe wide range of projections shown in Figs. 1 and 2 arise from \ninherent uncertainties in model dynamics and parameterizations that \nunderscore the importance of multi-model studies. Further uncertain-\nties lie in underlying socio-economic projections. For example, higher \neconomic or population growth is likely to lead to increased energy \nuse and emissions39\u201341 (although in developed countries, this link is \nless evident42). Other socio-economic factors, such as urbanization \nand household size, affect energy demand43\u201346. Higher energy demand \ncould ultimately limit the potential of the ELE scenario, if the energy \nsupply fails to meet the additional demand for cleaner fuels. Con-\nversely, lower growth could alleviate pressure on the supply of cleaner \nfuels. Moreover, energy intensity is usually lower in richer countries47, \nand higher economic growth could thus result in lower baseline energy \nintensities. This, in turn, would reduce the effectiveness of additional \npolicy measures in the ELE and TEC scenarios. Lastly, the effectiveness \nof the ACT scenario depends on the adoption of new behaviours, and \nthis process is inherently tied to various socio-economic factors.\nReducing pressure on the electricity system\nElectrification is widely recognized as a crucial strategy for emissions \nreduction48. According to most IAMs in this study, ELE yields greater \nemissions reductions by 2050 compared to the other strategies under \ncurrent policies (Fig. 2 and Supplementary Information 6). With the \nexception of MESSAGEix-Buildings and WITCH, the models project \nover 15% more emissions reductions with ELE than with the other strate-\ngies. Only MESSAGEix-Buildings demonstrates larger emissions reduc-\ntions in TEC, primarily due to the installation of heat pumps, which are \nmost effective when combined with improved insulation. In the WITCH \nmodel, emissions reductions in land transport are more pronounced \nin ELE, consistent with the other models, but bunker emissions remain \nconsiderable (Supplementary Fig. 15), probably due to the low techno-\nlogical granularity of alternative shipping technologies49.\nElectrification also poses substantial challenges, particularly \nregarding the surge in electricity demand (Fig. 3). By 2050, global \nelectricity demand increases by 8\u201316\u2009EJ per year for buildings and \n4\u201325\u2009EJ per year for transport in NPi-ELE compared to NPi-REF. The \nlarger potential increase in transport reflects its currently low elec-\ntricity share. Although growth rates from 2015 to 2050 are broadly \n\u201320\nBuildings\na\nb\nc\nd\nCurrent policies\nTransport\nBuildings\nStrategy\nREF\nACT\nTEC\nALL\nELE\n2030\n2050\nYear\n2050\n2030\nLimit to 1.5 \u00b0C\nTransport\n\u201340\n\u201360\n\u201380\n\u2013100\n\u201366% (IPCC Ch. 9: Buildings)\nCO2 emissions\nrelative to NPi reference (%)\n0\n\u201320\n\u201340\n\u201360\n\u201380\n\u2013100\n0\nFig. 2 | Deviation in global direct CO2 emissions for buildings and transport \nfor different intervention strategies relative to the NPi reference scenario. \na,b, Emissions in scenarios with current NPi from buildings (a) and transport \n(b). c,d, Emissions in the 1.5\u2009\u00b0C climate ambition scenarios from buildings (c) \nand transport (d). Boxes represent the interquartile range, with the centre line \nindicating the median and whiskers extending to the minimum and maximum \nvalues. The individual data points are shown as dots, with the boxplots based \non data from five models for buildings and six models for transport. A dashed \nline provides a comparison with estimated mitigation potentials from Ch. 9 \n(Buildings) of the IPCC\u2019s AR6 WG III report. A comparable analysis for transport is \nnot available.\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n385\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nconsistent with those observed in previous decades (Supplementary \nInformation 9), suggesting infrastructural challenges can be addressed, \nmore detailed models such as power dispatch models should be used \nto further explore this. In addition to requiring resilient energy infra-\nstructures, enhanced storage capacities and increased flexibility of the \npower system, increasing electricity demand may shift emissions from \nthe demand side to the supply side. This is mostly relevant for a weak \nclimate policy regime (NPi), whereas under 1.5\u2009\u00b0C policies, electricity \nsupply is projected to be already largely decarbonized by 2040. Moreo-\nver, under 1.5\u2009\u00b0C climate ambition, some models show lower electricity \ndemand growth due to increased availability of alternative fuels such \nas modern biomass and/or electrofuels.\nIn current climate policy scenarios (NPi), emissions reductions in \nthe ELE and ALL scenarios by 2050 are less pronounced when account-\ning for indirect emissions based on average emissions intensities \n(Fig. 4). Particularly for buildings, indirect emissions remain high in \nthe near term due to limited supply-side decarbonization and the \nsector\u2019s heavy reliance on electricity. In contrast, the transport sector \nhas a considerably larger share of fossil fuel use, which offers greater \npotential for emissions reductions through electrification, as there is \nmore opportunity to replace fossil fuels with cleaner alternatives. This \nis also reflected in the greater final energy reductions for transport than \nfor buildings, as shown in Fig. 1.\nIn light of increasing electricity demand, it is critical to expand the \ncapacity of the energy supply sector while simultaneously decarbon-\nizing it, as previous research has already suggested50\u201352. Our scenarios \nshow that integrating electrification with other demand-side strategies \n(ALL) can facilitate necessary transitions, potentially decreasing global \nelectricity demand by 10 to 39\u2009EJ per year, even under a 1.5\u2009\u00b0C climate \ntarget. Depending on the model and climate policy, this reduction \nrepresents 8 to 33% of the electricity demand for transport and build-\nings. Furthermore, the combined approach leads to over 15% more \ndirect emissions reduction compared to ELE alone in the ALL scenario.  \nThis demonstrates that a comprehensive strategy combining electri-\nfication with energy efficiency and activity-focused measures sub-\nstantially reduces the need for supply-side investments in low-carbon \ngeneration technologies, large-scale electricity storage and electricity \ninfrastructure and grids.\nDecomposing emissions reductions\nTo understand the key factors that drive emissions reductions, we \napplied decomposition analysis to passenger transport and residential \nbuildings emissions in 2050. We compared REF and ALL across five \nmodels (IMACLIM-R, IMAGE, REMIND, MESSAGEix-Buildings, WITCH) \n(Fig. 5). Efficiency gains stand out as important contributors to emis-\nsions reductions for all models. Improved efficiencies are partly a result \nof policies that promote higher energy efficiencies (TEC), such as effi-\nciency standards and building codes, but electrification (ELE) also plays \na key role and results in similar or even higher improvements: e-mobility \nis much more efficient than internal combustion engines in vehicles, \nand so too are heat pumps in comparison to boilers (Supplementary \nInformation 7 provides a decomposition of the individual strategies). \nIn addition, shared services further support efficiency gains (ACT).\nIt is important to note that the contributing elements in the \ndecomposition are interdependent. For instance, a decrease in car \ntravel will reduce more emissions in scenarios and models in which \ncars have lower fuel efficiency.\nThe contribution of electrification, defined as the ratio of electric-\nity to final energy demand, varies across models. Whereas some models \nshow decreasing emissions from electrification (WITCH, REMIND and \nIMACLIM-R in passenger transport), others show an increase (IMAGE, \nMESSAGEix-Buildings and IMACLIM-R in buildings). This discrepancy \nis due to different decarbonization rates of the power supply under \ncurrent policies and highlights that additional mitigation potential \ncould be achieved by expanding on renewable energy (Fig. 3). Likewise, \nmodels also vary in the contribution of carbon intensity reductions.\n100\n80\n60\n40\n20\n0\n40\n35\n30\n25\n20\n15\n10\n5\n0\n2030\nCurrent policies\nLimit to 1.5 \u00b0C\n2030\n2050\nYear\n2050\n2015\n2015\nElectricity demand (EJ yr\u22121)\nBuildings\nTransport\na\nb\nc\nd\nREF\nACT\nTEC\nELE\nAll\nStrategy\nCOFFEE\nIMACLIM-R\nIMAGE\nPROMETHEUS\nREMIND\nWITCH\nMESSAGEix-Buildings\nModel\nFig. 3 | Global electricity demand from buildings and transport under two \nclimate ambitions. a,b, Buildings electricity demand for scenarios with current \nnational policies implemented (a) and for the 1.5\u2009\u00b0C climate ambition scenarios \n(b). c,d, Transport electricity demand for scenarios with current national policies \nimplemented (c) and in the 1.5\u2009\u00b0C climate ambition scenarios (d). Markers \nindicate individual model results and bars depict the model ranges. A dashed line \nshows the 2015 levels. Non-electric energy demand is shown and discussed in \nSupplementary Information 11.\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n386\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nChanges in service demands, such as reduced travel and mode \nshifts lead to emissions reductions across all models. Previous research \nindicated that IAMs tend to underestimate the potential for mode \nshifting when not explicitly prescribed53 due to the complexities of \nsimulating context-dependent mode choices54. Our analysis shows that \nwhile mode shifts can be effective and useful to lower energy demand, \ntheir contribution to emissions reductions might be limited. Scenario \nimplications for the buildings sector are further explored in ref. 55.\nRegional effects\nDifferences observed in the scenario results are largely attributed to \nregional variations in the reference scenario, as depicted in Fig. 6 for \nnine regions. For the reference scenario, all models project an increas-\ning energy demand per capita for all regions except the OECD (Organi-\nsation for Economic Co-operation and Development: European Union, \nUnited States and Other OECD) for both sectors (buildings and trans-\nport) and the Africa and Middle East, other Asian countries and former \nSoviet Union countries for buildings. Figure 6 also shows that part of \nthe model spread at the global level can be explained by the regional \ndisparities, especially in the buildings sector or in transports for India. \nHowever, a broad consensus can be found across models for the most \nadvanced economies of the OECD for transport.\nRegional disparities are driven by opposing dynamics in service \ndemand and energy efficiency. As lower- and middle-income coun-\ntries get wealthier, they increasingly demand more energy services \nfor appliances, heating and cooling to support their improved living \nstandards. At the same time, the use of traditional fuels, which are \nstill commonly used for cooking in certain areas (predominantly in \nAfrica), decreases substantially until 2050. Energy demand reduc-\ntions also occur in OECD countries, where use of natural gas for space \nheating is steadily decreasing due to improved insulation levels under \ncurrent policies. In transport, per capita energy demand reduces in \nsome regions, despite rising demand for mobility, due to increased \nuse of electric vehicles with much higher efficiency56 than internal \ncombustion engine vehicles. In contrast to the global results, some \nmodels signal that emissions from buildings could be higher in ACT \nand TEC (Supplementary Fig. 22), probably due to rebound effects. \nThis is particularly noticeable in India, other Asian and Africa and \nMiddle East. An increase is also observed in transport emissions by \nCOFFEE and WITCH for the United States (ACT) and European Union \n(TEC), respectively. Carefully adapted rebound mitigation policies, \nfor example, through carbon pricing or providing consumption \ninformation, could in such cases further increase the effectiveness \nof demand-side measures57,58.\nThe combination of all interventions allows for a relative decrease \nin energy demand in 2050 for all regions and within the transport \nsector an absolute decrease compared to 2015 levels in all models, \nwith exceptions for India (all models), Other Asia (IMACLIM-R) and \nChina (IMAGE). The strongest disparities of demand-side strategies \nfor developing countries compared to more advanced economies \nreflect the strong correlation between energy demand and economic \ndevelopment59, which is likely to be a challenge for an early adoption of \ndemand-side policies in the upcoming years. On the other hand, some \ninterventions, such as reducing the floor space per capita, increasing \nthe energy efficiency in buildings and using low-carbon transport \nmodes, strongly depend on the way emerging cities will be shaped. \nThrough intelligent planning and design, such cities can lead the way \nin advancing sustainable transformations60.\nDiscussion and conclusions\nWe used seven IAMs to analyse the potential of demand-side mitigation \nstrategies. The reduction potentials of direct CO2 emissions identified \nin our study align with trends found in the synthesis of bottom-up \nstudies in the demand chapter of the IPCC\u2019s AR6 WG III report and fur-\nther highlight uncertainty and system-wide effects. Unlike the IPCC\u2019s \nestimations, our analysis builds upon a consistent set of harmonized \nscenario assumptions and includes interactions between measures \nand sectors, such as the reduced impact of lowering set-point tempera-\ntures in conjunction with enhanced insulation. Our analysis reveals \nthat interactions among different measures do play a role, yet their \neffect on overall mitigation potential is relatively small. Additionally, \nwe explicitly assessed the importance of decarbonizing electricity \ngeneration to disentangle it from demand-side mitigation potentials.\nBy using an ensemble of models, and presenting a decomposi-\ntion of the effects across models, we explored inherent uncertainty \nand diversity of results and the underlying dynamics. The wide \nvariation in projected mitigation potentials across models (51\u201385% \nfor buildings and 37\u201391% for transport) reveals the high level of \n2030\nYear\nYear\nYear\n2030\n2030\n2050\n2050\n2050\n\u201380\n\u201360\n\u201340\n\u201320\n0\n+20\n+40\n+60\na\nBuildings\nb\nTransport\nCO2 emissions\nrelative to 2015 (%)\n0\n20\n40\n60\n80\n100\n120\n140\nMt CO2 EJ\u22121\nc\nPower sector\nModel\nCOFFEE\nIMACLIM-R\nIMAGE\nPROMETHEUS\nREMIND\nWITCH\nMESSAGEix-Buildings\nScenario (NPi)\nREF\nACT\nTEC\nELE\nALL\nFig. 4 | Global CO2 emissions from buildings, transport and the power \nsector relative to 2015 levels under current NPi. a,b, Changes in global CO2 \nemissions in the buildings (a) and transport (b) sectors including emissions \nfrom electricity, heat and hydrogen generation. The darker shaded ranges \nrepresent the total emissions, combining both direct and indirect sources, with \nthe latter calculated using model-average emissions intensities. In contrast, \nthe lighter shaded ranges represent only direct emissions, matching the results \nshown in the lower panels of Fig. 1. Markers indicate individual model results \nand bars depict the model ranges. c, Global CO2 emissions from the power \nsector. The same figure for scenarios with 1.5\u2009\u00b0C climate ambition is available in \nSupplementary Information 10.\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n387\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nuncertainty that arises from the complexity of modelling future \nenergy-demand development.\nThe impact of the selected set of demand-side interventions is \nhigh: on average, across the models, 65% reduction of direct CO2 emis-\nsions in buildings and nearly 70% in transport by 2050, compared \nto a current policies scenario. Activity-focused measures result in \nreductions by 2050 of 6\u201323% for buildings and 17\u201328% for transport, \ntechnology-optimizing measures to 11\u201333% and 2\u201367%, respectively, \nand electrification-focused measures to 45\u201377% and 22\u201386%. However, \nthe success of these strategies hinges critically on the emergence \nof social innovations and implementation of policies to overcome \ncrucial barriers.\nBy disaggregating demand-side strategies, our analysis provides \nclearer insights into the distinct contributions of different policies. \nWhereas the importance of electrification and fuel shifts in reducing \ndemand-side CO2 emissions is well recognized, our results demonstrate \nthat across most models, an electrification-focused strategy yields the \ngreatest reductions in direct CO2 emissions from buildings and trans-\nport by 2050, even lowering emissions below 2015 levels. However, this \napproach also more than doubles global electricity demand by 2050 \ncompared to 2015. Focusing on increased electrification alone only \nreduces emissions if this is supplemented by a sustained effort to decar-\nbonize electricity supply. Shifts in activity patterns can also contribute \nto emissions reduction but have a greater potential in high-income \nregions, as meeting basic energy services is a higher priority in lower- \nand middle-income regions.\nIntegrated approaches combining different strategies not only \nlead to the greatest reduction in emissions but also help alleviate \nstresses on the upstream energy supply sector that may arise from \nindividual demand-side strategies, such as an increase in electricity \ndemand, storage and grids due to electrification. Furthermore, decom-\nposition analysis shows that efficiency improvements, and to a lesser \nextent activity shifts, can contribute to further emissions reductions.\nFollowing the advancements in IAMs, there are several opportu-\nnities for improving the assessment of demand-side scenarios. First, \nlifestyle changes should be better integrated into scenarios61\u201366. Reduc-\ning or shifting energy services demand requires widespread changes \nin social norms to induce more sustainable lifestyles, as discussed and \nmodelled in Supplementary Information 13. These can only partially \nbe achieved by policies and strongly depend on available organiza-\ntions and infrastructures67,68. Second, further developments should \nenhance the linkage between demand sectors and industrial and \nREF: 2050\nFloorspace\nEficiency\nElectrification\nCarbon intensity\nALL: 2050\n0\n2\n4\n6\nCO2 emissions (GtCO2)\na\nIMACLIM-R\nREF: 2050\nFloorspace\nEficiency\nElectrification\nCarbon intensity\nALL: 2050\nDirect\nElectricity generation\nb\nIMAGE\nREF: 2050\nFloorspace\nEficiency\nElectrification\nCarbon intensity\nALL: 2050\nc\nMESSAGEix-Buildings\nREF: 2050\nFloorspace\nEficiency\nElectrification\nCarbon intensity\nALL: 2050\nResidential buildings\nd\nWITCH\nCO2 emissions (GtCO2)\nREF: 2050\nActivity\nModal shift\nEficiency\nElectrification\nCarbon intensity\nALL: 2050\n0\n2\n4\n6\ne\nIMACLIM-R\nFactors\nFactors\nFactors\nFactors\nFactors\nFactors\nFactors\nFactors\nREF: 2050\nActivity\nModal shift\nEficiency\nElectrification\nCarbon intensity\nALL: 2050\nAviation\nBus\nLDV\nRail\nf\nIMAGE\nREF: 2050\nActivity\nModal shift\nEficiency\nElectrification\nCarbon intensity\nALL: 2050\ng\nREMIND\nREF: 2050\nActivity\nModal shift\nEficiency\nElectrification\nCarbon intensity\nALL: 2050\nPassenger transport\nh\nWITCH\nFig. 5 | Decomposition of global CO2 emissions for the current policy \nscenarios in 2050 from the residential sector and passenger transport. \na\u2013d, Decomposition for residential buildings in IMACLIM-R (a), IMAGE (b), \nMESSAGEix-Buildings (c) and WITCH (d). e\u2013h, Decomposition for passenger \ntransport in IMACLIM-R (e), IMAGE (f), REMIND (g) and WITCH (h). Figures \ninclude emissions from electricity generation (hatched). Changes in emissions \nare attributed to the factors activity/floorspace, modal shifts (transport only), \nefficiency, electrification and carbon intensity. Carbon intensity is the average \namount of CO2 emissions per unit energy of fuel combusted or, for emissions \nfrom electricity generation, the average emissions intensity of electricity \ngeneration. For passenger transport, four modal categories are considered: \naviation, bus, light-duty-vehicle (LDV) and rail. Floorspace use serves as a proxy \nfor activity changes in buildings because most models lack the capability to \ngenerate detailed output for service demand. This implies that all activity \nchanges not directly related to floorspace use, such as thermostat adjustments, \nare classified as efficiency improvements. Methods provide definitions of all \nfactors. Note that REMIND and MESSAGEix-Buildings are exclusively included \nfor passenger transport and buildings, respectively, due to a lack of the required \nlevel of granularity.\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n388\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nmaterial demands22. Strengthening this connection would enable a \nmore complete evaluation of emissions across the supply chain. Third, \nIAMs do not adequately consider local infrastructural challenges, \nparticularly those related to the power grid. Additional assessments \nusing more detailed models, such as power dispatch models, can help \nto identify risks, such as increased likelihood of blackouts69, that could \nimpede consumer adoption.\nThe success of policies depends on their broader support in \nsociety. Measures aimed at behavioural and lifestyle changes, but \nalso interventions such as taxing air travel and restricting low-cost \ncarbon-intensive technologies that directly influence affordability \nmight face resistance. Our study, which assumes that policies are fully \neffective, does not account for the potential resistance and partial \nimplementation that could affect the outcomes. Therefore, future \nresearch should also delve into the impact of these measures on mac-\nroeconomics, ensuring that policies are not only effective but also \nsustainable in their broader socio-economic context.\nMethods\nIdentification of key measures\nIn the initial phase of this study, an extensive list of demand-side meas-\nures was identified, based on relevant literature and Ch. 5 of IPCC\u2019s \nAR6 WG III report2 (references in Table 1). To improve the credibility \nand policy relevance of the scenarios, and to ensure that no factors \nwere overlooked, we collected input from experts in relevant areas \nrelated to climate change mitigation. This involved conducting an \nonline stakeholder survey in 2021. Experts were asked by means of a \nquestionnaire to evaluate the feasibility and effectiveness of different \nways to reduce emissions in the domains of buildings, mobility and \ninternational transport (Supplementary Table 7). Details about the \nstakeholder survey are provided in Supplementary Information 2, \nand the responses are summarized in Supplementary Tables 7 and 8. \nThis was used as input for designing the three intervention strategy \nscenarios. The full process of designing, simulating and analysing the \nscenarios is summarized in Supplementary Fig. 1.\nDefining policy scenarios and model assumptions\nOn the basis of the literature evaluations and feedback from the stake-\nholder survey, we compiled a comprehensive list of measures that \nwere considered to have substantial impact on reducing demand-side \nemissions, while also being regarded feasible. We classified these \nmeasures into three different intervention strategies, based on a \nslightly adapted version of the ASI (Avoid\u2013Shift\u2013Improve) frame-\nwork3. Categorizing policy measures in the ASI framework can be \nambiguous in some cases. For example, stimulating adoption of heat \npumps both improves the building technical systems and induces a \nshift from higher-carbon fuels to electrification for space heating. \nBecause we aimed to explore which type of policy interventions are \nmost effective, we needed a cleaner separation between the miti-\ngation options and we use a more distinct separation between the \ndemand-side measures. Table 1 summarizes the key assumptions for \nthe three intervention strategies.\nThis section presents a detailed description of the modelling \nprotocol used to assess the scenarios. We elaborate on the specific \nmeasures for each sector that the scenarios consider, as summarized \nearlier in Table 1. We explain the quantification of the measures and \nAfrica and\nMiddle\nEast\nEuropean\nUnion\nOther\nOECD\nUnited\nStates\nFormer\nSoviet\nUnion\nChina\nRegion\nIndia\nLatin and\nSouth\nAmerica\nOther Asia\nModel\nCOFFEE\nIMACLIM-R\nIMAGE\nREMIND\nWITCH\nMESSAGEix-Buildings\nScenario (NPi)\nREF\nACT\nTEC\nELE\nALL\n\u201350\n0\n+50\n+100\n+150\n+200\n+250\n\u201350\n\u201325\n0\n+25\n+50\n+75\n+100\n+125\nFinal energy per capita in 2050\nrelative to 2015 (%)\nBuildings\nTransport\na\nb\nFig. 6 | Change in final energy per capita in 2050 relative to 2015 under current \nNPi for nine regions. a, Buildings final energy use. b, Transport final energy \nuse. Results are shown for the United States, the European Union, other OECD \ncountries, countries from the Reforming Economies of the Former Soviet Union, \nChina, India, Latin and South American countries, other Asian countries and \ncountries of Africa and the Middle East. Markers indicate individual model results \nand bars depict the model ranges.\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n389\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\noutline the literature and scenarios that form the basis of the underly-\ning assumptions.\nActivity reduction and activity shifts. Policies that limit floorspace \nin new building constructions, along with flexible use of buildings \nand shared spaces (such as co-housing and co-working), reduce per \ncapita floorspace in both the residential and commercial sectors. By \n2050, we assume a regional cap of 40\u2009m2 per capita for residential \nfloorspace, based on the assumptions in the LED (low energy demand) \nand SSP1 scenarios from Fishman et al.70 and Mastrucci et al.16. For \ncommercial floorspace we assume a regional cap of 25\u2009m2 per capita \nby 205071.\nChanges in household preferences, along with policies that limit \nnew construction of single-family houses will lead to a higher share \nof multi-family housing. The proportion of the population living in \nmulti-family houses will increase by 10% across the entire housing stock \nin 2050, compared to the reference scenario. This is broadly consistent \nwith two other studies72,73.\nStimulated by policies limiting set-point temperatures and infor-\nmation campaigns, the set-point temperatures in buildings shift to \n20\u2009\u00b0C (heating) and 25\u2009\u00b0C (cooling) by 205071,73.\nDemand for private vehicles driving within cities decreases, driven \nby measures such as congestion charges and growing prevalence of \nremote working. The combination of these measures leads to a 20% \nreduction in passenger kilometres (pkm) for private cars by 2050 \nwith respect to the reference scenario. This is loosely based on the \nHigh Shift Scenario74, which assumes a 50% reduction in urban vehicle \ntravel compared to their baseline in 2050 in all regions. The High Shift \nScenario focuses on urban transport and \u2018considers what could be if the \npolicies and investments currently in place in the nations with the most \nefficient urban transport, were replicated throughout the world\u201974. \nBecause by 2050, still a third of the world population is expected to \nbe rural (according to UN projections), and distances travelled are \ngenerally longer in rural areas, we assume a lower overall reduction in \npassenger kilometres than the High Shift Scenario.\nAdditionally, ride and car-sharing associated with on-demand AV \n(autonomous vehicles), incentivized carpooling by private operators, \nPark + Ride and more HOV (high-occupancy vehicle) lanes will increase \nthe number of passengers per vehicle. We assume that the occupancy \nrate will gradually increase until 2050 by around 40% with respect to \nthe reference scenario. This is in line with results from Akimoto et al.75, \nestimating the impact of ride and car-sharing associated with fully \nautonomous cars.\nMore bike lanes and pedestrian zones increase the adoption of \nactive modes (bicycles, e-scooters and walking). Improvements in \npublic transport infrastructure, last-mile services and free/lower public \ntransport fares increase the adoption of public transport. We assume \nthat the modal shares, as a percentage of total passenger kilometres \n(pkm), for the active mode and public transport converge by 2050 to \nthe levels given in Supplementary Table 9, depending on the region. \nThe shares are based on projections from the High Shift Scenario74, \nwhere gradual shifts to public transit and active modes are assumed.\nImprovements in road freight logistics reduce road freight trans-\nport, measured in tonne kilometres (tkm), by 13.5% in 2050 with respect \nto the reference scenario, following the Modern Trucking Scenario76. \nThe Modern Trucking Scenario is an ambitious road freight scenario \nlaying out a modernization strategy aiming at increased energy security \nand prevention of climate change with \u2018rapid adoption of the techno-\nlogical and system-wide measures\u2019.\nIn the aviation sector, passenger transport is reduced through \nthe introduction of fuel taxes (by abolishing tax exemptions), the \nimplementation of movement taxes such as a frequent flyer levy and \ndevelopment of increased virtual connectivity. This results in higher \nfuel prices and reduced RPK (revenue passenger kilometres). Following \nthe Green Push Scenario14, the scenario strives for approximately 30% \n(international) and 40% (domestic) reductions of RPK on the global \nlevel by 2050 with respect to the reference scenario.\nSimilarly, freight transport is also affected by fuel taxes and move-\nment taxes. In combination with policies that encourage development \nof local manufacturing and storage, these measures result in a global \nreduction in revenue tonne kilometres (RTK) of 10% with respect to the \nreference scenario. This assumption is based on the findings of three \nstudies. One study, based on the Beyond 2\u2009\u00b0C Scenario, which is in line \nwith countries\u2019 more ambitious aspirations, projects 8% lower freight \ndemand by 2060 with respect to their baseline scenario, primarily \ndriven by diminishing trade requirements in fossil fuels77. A second \nstudy by M\u00fcller\u2013Casseres et al. shows that in a well-below 2\u2009\u00b0C SSP2 \nscenario a substantial portion of fossil energy trade can be avoided \n(20% by 2050 and 25% by 2100)32. A third study by Walsh et al. projects \na decrease of over 15% in imported and domestic trade for the United \nKingdom, based on their analysis of low carbon futures for shipping \nfrom a UK perspective78.\nWe assume that speed restrictions in maritime transport (slow \nsteaming shipping) lead to 15% reduction in energy consumption on \nfleet basis with respect to the reference scenario, based on compre-\nhensive scenarios from CE Delft assuming a 20\u201325% speed reduction79. \nThe assumed speed reduction is also in line with scenario assumptions \nfrom Walsh et al.78.\nFurther, we assume phase out of short-haul air travel by 2030 by \nclosing the price gap between rail and aviation and policies limiting \nshort-haul air travel73.\nTechnology-optimizing strategy. As a result of building codes and \nstandards, energy performance certification, subsidies and incentives, \nthe useful energy intensity per area increases and (average) U-values \ndecrease. The nearly zero-energy buildings (nZEB) level for insulation \nin new construction, representing the average for the building enve-\nlope, will be 0.3\u2009W\u2009m\u22122\u2009K\u22121 on average by 203080. Best practice examples \nof current nZEB values already reach such efficiencies today81. Energy \nsavings for renovation are at least 40% by 2030.\nAlso the ratio between final and useful energy improves by \nimprovements in HVAC (heating, ventilation, and air conditioning)73. \nConversion efficiency coefficients for air conditioning and heat pumps \nreach 6.0 by 210016,30.\nAn increased renovation rate, stimulated by subsidies and incen-\ntives, leads to a doubling of the current retrofit rate to 2% per year in \nthe global north82.\nWe assume that efficiency standards for vehicles and trucks lead \nto vehicle efficiency improvements (autonomous) of 1.5% per year \nuntil 2050. For trucks we assume even higher improvements of 2% per \nyear until 2050.\nEfficiency standards result in annual autonomous efficiency \nimprovements of 1.3% per year for new aircrafts and 1.5% per year for \nnew ships until 205083,84. The fleet efficiency improves autonomously \nas a result of environmental certification for using airports and ports. \nWe assume annual fleet efficiency improvements for new and existing \naircrafts and vessels of 0.7% per year for aviation and 1.1% per year for \nshipping until 205083,84.\nElectrification-focused strategy. Fuel mandates accelerate electrifi-\ncation in buildings and switching to cleaner heating fuels73. By 2030, all \nnew buildings adopt heat pumps, and by 2050, we assume that 70% of \nspace and water heating is electricity based71,85. Non-clean heating fuels \nare phased out by 2050, and new natural gas connections for heating \nwill be banned in the global north by 2030.\nBuilding regulations, along with neighbourhood-based \napproaches, will promote the deployment of on-site and building- \nintegrated renewable energy systems73. By 2050, renewable energy \nsources (photovoltaics and thermal solar) meet 50% of the heating and \ncooling demands in the global north.\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n390\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nFuel/technology mandates ensure full electrification of passen-\nger vehicles and light-duty trucks by 2040. We include both battery \nelectric vehicles and fuel cell electric vehicles in the electrification \ntargets. Because there are much larger hurdles for full electrification of \nheavy-duty vehicles76,86, we only assume a phase-out of diesel engines \nin the fleet of heavy-duty vehicles by 2040.\nWe assume that electric short-haul planes become available after \n2050 (broadly consistent two other studies73,87). Further, we assume \nfull electrification of ports (and a reduction of auxiliary engines \nneeded in ships) by 2030. In alignment with this, vessels are adapted \nto zero-emission berth standards by 2040. This timeline for port elec-\ntrification is loosely based on a scenario study by Gillingham et al.88 and \nthe Global EV Outlook89. Assuming that ships spend approximately \n15% of the time at berth and that 15% of their total fuel consumption \nis related to the auxiliary engine, we assume that 2.3% of the total fuel \nconsumption can be saved by cold ironing (connecting ships to the \nonshore power supply). For reference, Bouman et al. report potential \nCO2 reductions of 3\u201310% (ref. 90).\nFuel standards/mandates, infrastructure development and remov-\ning blending restrictions increase the use of alternative fuels (biofuels/\nelectrofuels)73. Following the Sustainable Development Scenario91 the \nshare of hydrogen in final energy demand grows to 40% in the aviation \nsector and to 50% in the shipping sector by 2070. The share of biofuels \nincreases to 15% for both the aviation and shipping sectors.\nModel implementation\nThe scenario implications of the scenarios listed in Table 2 are ana-\nlysed by seven global IAMs (COFFEE, IMACLIM-R, IMAGE, MESSAGEix- \nBuilding, PROMETHEUS, REMIND and WITCH), integrating the \nhigh-resolution mobility, transport and buildings modules. A descrip-\ntion of the model representation of the transport sector and the build-\nings sector is shown in Supplementary Tables 1\u20133 of the Supplementary \nMaterials.\nTwo models encountered limitations in producing output for \nboth sectors. COFFEE generated exclusively transport-related output \nvariables, whereas MESSAGEix-Buildings was utilized to implement the \nscenarios for the buildings sector only. To avoid imbalance across sec-\ntoral and total results, the results of MESSAGEix-Buildings are left out of \ncalculated averages and ranges. MESSAGEix-Buildings results are visu-\nally presented in figures, wherever possible. Moreover, due to limited \nrepresentation of other sectors, the particular MESSAGEix-Buildings \nversion used here was only able to achieve emissions consistent with \n2\u2009\u00b0C scenarios.\nThe implementation of measures largely aligns with existing lit-\nerature and scenarios previously developed to analyse demand-side \nmitigation, such as the High Shift Scenario74, the LED scenario82 and \nthe lifestyle scenario explored by Van Vuuren et al.92. An overview of \nthe implementation of the scenario measures by model is shown in \nSupplementary Table 4 (activity-focused strategy), Supplementary \nTable 5 (technology-optimizing strategy) and Supplementary Table 6 \n(electrification-focused strategy). We have made efforts to harmo-\nnize the model implementation of demand-side measures as much \nas possible across the models; however, some difference remains \ninevitable. One key distinction in the implementation is that some \nmodels use exogenous projections to model changes in energy service \n(for example, floorspace), whereas other models employ endogenous \nrepresentations of energy service and change levers, such as preference \nfactors, to attain the desired scenario changes. Next to this, differences \nin implementations also arise from the fact that some models lack the \ncapability to explicitly represent specific measures.\nThe scenarios with only current policies implemented consider \nthe current climate- energy- and land-use policies and account for \nonly those that are secured in legislative decisions, executive orders or \nequivalent93,94. No additional measures or plans, for example, regarding \nNationally Determined Contributions are considered.\nIn addition to this baseline, we consider scenarios in which global \nwarming is limited to 1.5\u2009\u00b0C with a low overshoot based on estimated \ncarbon budgets from the IPCC\u2019s AR6 WG I report95. This is implemented \nin all models through the application of a globally uniform carbon \ntax, determined by an optimization process for each model and sce-\nnario. This tax ensures that cumulative emissions from 2020 onward \nnever exceed 650\u2009Gt CO2 (peak budget) and that cumulative emissions \nbetween 2020 and 2100 are limited to 400\u2009Gt CO2 (end-of-century \nbudget). Non-CO2 gases, such as N2O, CH4 and F-gases, are priced \nequivalently to CO2 using GWP100 (global warming potential over \n100 years). Models that cannot achieve the end-of-century budgets \naim for the lowest budgets feasible. Also, carbon dioxide removal \n(CDR) deployment and net-negative emissions are limited; models aim \nfor no more than 250\u2009Gt CO2 net-negative emissions undercutting the \nend-of-century budget and aim to not have cumulated CDR (including \nAFOLU CDR) until 2100 exceed 500\u2009Gt CO2.\nNo-interaction estimate\nThe no-interaction estimate for combined emissions reductions is \nconstructed from the NPi-ACT, NPi-TEC, NPi-ELE and NPi-REF scenarios. \nThis estimate is derived by computing the ratio of CO2 emissions in each \nscenario for the year 2050 to the emissions in the reference scenario \nfor the same year:\nCO2 emissionsscenario (2050)\nCO2 emissionsNPi-REF (2050)\n(1)\nThe product of these ratios approximates the total relative reduc-\ntion in CO2 emissions. This relative reduction is then multiplied by the \nabsolute emission levels of NPi-REF in 2050 to calculate the correspond-\ning absolute emissions when all measures are combined under the \nassumption of no interaction among the strategies.\nDecomposition method\nIn this study we used the Shapley/Sun decomposition method based \non the Laspeyres index to quantify the driving components between \nthe scenarios. This method has the advantage that it is based on the \nfamiliar concept of percentage change, making it easier to interpret. \nMoreover, it can be applied to obtain a perfect additive decomposition, \nthat is, the total difference is allocated to different components and no \nunexplained residual term appears96,97.\nFor the transport sector, the activity levels are expressed in pas-\nsenger kilometres (pkm), which yields the following equation for direct \nemissions:\nCO2 emissionsdirect\n= pkm\n\u23df\nactivity\n\u00d7\n\u2211\nn\u2208modes\n\u239b\n\u239c\u239c\u239c\n\u239d\npkmn\npkm\n\u23df\u23b5\u23df\u23b5\u23df\nmodal shift\n\u00d7\nEn\npkmn\n\u23df\nefficiency\n\u00d7\nEn \u2212Eelectricity,n\nEn\n\u23df\u23b5\u23b5\u23b5\u23b5\u23df\u23b5\u23b5\u23b5\u23b5\u23df\nelectrification\n\u00d7\nCO2 emissionsdirect,n\nEn \u2212Eelectricity,n\n\u23df\u23b5\u23b5\u23b5\u23b5\u23b5\u23b5\u23df\u23b5\u23b5\u23b5\u23b5\u23b5\u23b5\u23df\ncarbon intensity\n\u239e\n\u239f\u239f\u239f\n\u23a0\n(2)\nwhere pkm denotes the number of passenger kilometres travelled in \na year, E the final energy demand in transport and Eelectricity the electric-\nity demand in transport. Electrification is treated as a distinct factor, \nfollowing an approach similar to Edelenbosch et al.98, measuring the \nshare of electricity in final energy. For simplification, alongside direct \nemissions, we include only emissions from electricity generation in the \ndecomposition. This yields:\nCO2 emissionselectricity\n= pkm\n\u23df\nactivity\n\u00d7\n\u2211\nn\u2208modes\n\u239b\n\u239c\u239c\u239c\n\u239d\npkmn\npkm\n\u23df\u23b5\u23df\u23b5\u23df\nmodal shift\n\u00d7\nEn\npkmn\n\u23df\u23b5\u23df\u23b5\u23df\nefficiency\n\u00d7\nEelectricity,n\nEn\n\u23df\u23b5\u23b5\u23df\u23b5\u23b5\u23df\nelectrification\n\u00d7\nCO2 emissionselectricity,n\nEelectricity,n\n\u23df\u23b5\u23b5\u23b5\u23b5\u23b5\u23b5\u23b5\u23df\u23b5\u23b5\u23b5\u23b5\u23b5\u23b5\u23b5\u23df\ncarbon intensity\n\u239e\n\u239f\u239f\u239f\n\u23a0\n(3)\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n391\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nThe factors electrification and carbon intensity are defined dif-\nferently from the former equation, but they share a similar meaning. \nBecause the level of detail in the buildings sector varies strongly across \nthe IAMs, a detailed decomposition in different end uses is not possible \nand we use the amount of floorspace as a proxy for activity instead. For \nthe residential sector we use the following equation:\nCO2 emissionsdirect\n= floorspace\n\u23df\u23b5\u23b5\u23df\u23b5\u23b5\u23df\nactivity\n\u00d7\nE\nfloorspace\n\u23df\u23b5\u23b5\u23df\u23b5\u23b5\u23df\nefficiency\n\u00d7\nE \u2212Eelectricity\nE\n\u23df\u23b5\u23b5\u23b5\u23df\u23b5\u23b5\u23b5\u23df\nelectrification\n\u00d7 CO2 emissionsdirect\nE \u2212Eelectricity\n\u23df\u23b5\u23b5\u23b5\u23b5\u23b5\u23df\u23b5\u23b5\u23b5\u23b5\u23b5\u23df\ncarbon intensity\n(4)\nwhere floorspace denotes the average floorspace used per capita, E the \nfinal energy demand in the residential sector and Eelectricity the electricity \ndemand in the residential sector. Similarly, this yields for emissions \nfrom electricity generation for:\nCO2 emissionselectricity\n= floorspace\n\u23df\u23b5\u23b5\u23df\u23b5\u23b5\u23df\nactivity\n\u00d7\nE\nfloorspace\n\u23df\u23b5\u23b5\u23df\u23b5\u23b5\u23df\nefficiency\n\u00d7\nEelectricity\nE\n\u23df\u23b5\u23b5\u23df\u23b5\u23b5\u23df\nelectrification\n\u00d7\nCO2 emissionselectricity\nEelectricity\n\u23df\u23b5\u23b5\u23b5\u23b5\u23b5\u23b5\u23df\u23b5\u23b5\u23b5\u23b5\u23b5\u23b5\u23df\ncarbon intensity\n(5)\nThe components are not independent from each other. For exam-\nple, subsidies for electric vehicles can lead to modal shifting, more elec-\ntrification, higher efficiency and potentially higher activity. Caution is \nneeded when interpreting results from two separate decompositions. \nSimilar changes in specific components (for example, activity levels \ndrop by 10%) may not necessarily result in identical magnitudes for \nthose components. This is because interactions with other components \n(such as efficiency) can lead to varying magnitudes for the attributed \nchange within different decomposition contexts.\nData availability\nScenario output data are available directly from the IIASA NAVIGATE \ndatabase at https://data.ece.iiasa.ac.at/navigate.\nCode availability\nExtensive documentation of integrated assessment models is available \nat https://www.iamcdocumentation.eu/. The source code for some \nmodels is available directly from their developers. Custom code for \nplotting (including post-processing steps implemented for decom-\nposition analysis) can be provided upon request.\nReferences\n1.\t\nMundaca, L., \u00dcrge-Vorsatz, D. & Wilson, C. Demand-side \napproaches for limiting global warming to 1.5\u2009\u00b0C. Energy Eff. 12, \n343\u2013362 (2019).\n2.\t\nIPCC in Climate Change 2022: Mitigation of Climate Change  \n(eds Shukla, P. R. et al.) Ch. 5 (Cambridge Univ. Press, 2023).\n3.\t\nCreutzig, F. et al. Demand-side solutions to climate change \nmitigation consistent with high levels of well-being. Nat. Clim. \nChange 12, 36\u201346 (2022).\n4.\t\nLuderer, G. et al. Residual fossil CO2 emissions in 1.5\u20132\u2009\u00b0C \npathways. Nat. Clim. Change 8, 626\u2013633 (2018).\n5.\t\nSugiyama, M. et al. High with low: harnessing the power of \ndemand-side solutions for high wellbeing with low energy and \nmaterial demand. Joule 8, 1\u20136 (2024).\n6.\t\nExtended World Energy Balances (IEA, 2023); https://doi.org/ \n10.1787/8624f431-en\n7.\t\nMinx, J. C. et al. A comprehensive and synthetic dataset for \nglobal, regional, and national greenhouse gas emissions by \nsector 1970\u20132018 with an extension to 2019. Earth Syst. Sci. Data \n13, 5213\u20135252 (2021).\n8.\t\nWiese, C., Larsen, A. & Pade, L.-L. Interaction effects of energy \nefficiency policies: a review. Energy Eff. 11, 2137\u20132156 (2018).\n9.\t\nSafarzy\u0144ska, K. & van den Bergh, J. C. J. M. A higher rebound effect \nunder bounded rationality: interactions between car mobility and \nelectricity generation. Energy Econ. 74, 179\u2013196 (2018).\n10.\t del R\u00edo, P. Analysing the interactions between renewable energy \npromotion and energy efficiency support schemes: the impact \nof different instruments and design elements. Energy Policy 38, \n4978\u20134989 (2010).\n11.\t\nLuderer, G. et al. Impact of declining renewable energy costs on \nelectrification in low-emission scenarios. Nat. Energy 7, 32\u201342 \n(2022).\n12.\t Reimers, H. et al. Indirect rebound effects on the consumer \nlevel: a state-of-the-art literature review. Cleaner Responsible \nConsumption 3, 100032 (2021).\n13.\t Sorrell, S., Gatersleben, B. & Druckman, A. The limits of energy \nsufficiency: a review of the evidence for rebound effects and \nnegative spillovers from behavioural change. Energy Res. Social \nSci. 64, 101439 (2020).\n14.\t Kikstra, J. S. et al. Climate mitigation scenarios with persistent \nCOVID-19-related energy demand changes. Nat. Energy 6, \n1114\u20131123 (2021).\n15.\t Camarasa, C. et al. A global comparison of building \ndecarbonization scenarios by 2050 towards 1.5\u20132\u2009\u00b0C targets.  \nNat. Commun. 13, 3077 (2022).\n16.\t Mastrucci, A., van Ruijven, B., Byers, E., Poblete-Cazenave, M. & \nPachauri, S. Global scenarios of residential heating and cooling \nenergy demand and CO2 emissions. Climatic Change 168, 14 \n(2021).\n17.\t Santamouris, M. & Vasilakopoulou, K. Present and future energy \nconsumption of buildings: challenges and opportunities towards \ndecarbonisation. e-Prime Adv. Electr. Eng. Electron. Energy 1, \n100002 (2021).\n18.\t Zhang, R. & Hanaoka, T. Cross-cutting scenarios and strategies \nfor designing decarbonization pathways in the transport sector \ntoward carbon neutrality. Nat. Commun. 13, 3629 (2022).\n19.\t Yan, S., de Bruin, K., Dennehy, E. & Curtis, J. Climate policies for \nfreight transport: energy and emission projections through 2050. \nTransp. Policy 107, 11\u201323 (2021).\n20.\t \u00c5kerman, J., Kamb, A., Larsson, J. & N\u00e4ss\u00e9n, J. Low-carbon \nscenarios for long-distance travel 2060. Transp. Res. Part D \nTransp. Environ. 99, 103010 (2021).\n21.\t Guo, S., Yan, D., Hu, S. & Zhang, Y. Modelling building energy \nconsumption in China under different future scenarios. Energy \n214, 119063 (2021).\n22.\t Pauliuk, S. et al. Global scenarios of resource and emission \nsavings from material efficiency in residential buildings and cars. \nNat. Commun. 12, 5097 (2021).\n23.\t Soonsawad, N., Martinez, R. M. & Schandl, H. Material demand, \nand environmental and climate implications of Australia\u2019s \nbuilding stock: current status and outlook to 2060. Resour. \nConserv. Recycl. 180, 106143 (2022).\n24.\t IPCC in Climate Change 2022: Mitigation of Climate Change  \n(eds Shukla, P. R. et al.) Ch. 3 (Cambridge Univ. Press, 2023).\n25.\t Pye, S. et al. Modelling net-zero emissions energy systems \nrequires a change in approach. Climate Policy 21, 222\u2013231 (2021).\n26.\t Rottoli, M., Dirnaichner, A., Pietzcker, R., Schreyer, F. & Luderer, G. \nAlternative electrification pathways for light-duty vehicles in the \nEuropean transport sector. Transp. Res. Part D Transp. Environ. 99, \n103005 (2021).\n27.\t Edelenbosch, O. Y., Hof, A. F., Nykvist, B., Girod, B. & van Vuuren, D. P.  \nTransport electrification: the effect of recent battery cost \nreduction on future emission scenarios. Climatic Change 151, \n95\u2013108 (2018).\n28.\t Fotiou, T., Capros, P. & Fragkos, P. Policy modelling for ambitious \nenergy efficiency investment in the EU residential buildings. \nEnergies 15, 2233 (2022).\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n392\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\n29.\t Daioglou, V., Mikropoulos, E., Gernaat, D. & van Vuuren, D. P. \nEfficiency improvement and technology choice for energy  \nand emission reductions of the residential sector. Energy 243, \n122994 (2022).\n30.\t Levesque, A. et al. How much energy will buildings consume in \n2100? A global perspective within a scenario framework. Energy \n148, 514\u2013527 (2018).\n31.\t M\u00fcller-Casseres, E. et al. Are there synergies in the \ndecarbonization of aviation and shipping? An integrated \nperspective for the case of Brazil. iScience 25, 105248 (2022).\n32.\t M\u00fcller-Casseres, E., Edelenbosch, O. Y., Szklo, A., Schaeffer, R. & \nvan Vuuren, D. P. Global futures of trade impacting the challenge \nto decarbonize the international shipping sector. Energy 237, \n121547 (2021).\n33.\t Carvalho, F. et al. Prospects for carbon-neutral maritime fuels \nproduction in Brazil. J. Cleaner Prod. 326, 129385 (2021).\n34.\t Fragkos, P. Decarbonizing the international shipping and aviation \nsectors. Energies 15, 9650 (2022).\n35.\t McCollum, D. L. et al. Improving the behavioral realism of global \nintegrated assessment models: an application to consumers\u2019 \nvehicle choices. Transp. Res. Part D Transp. Environ. 55, 322\u2013342 \n(2017).\n36.\t Pettifor, H., Wilson, C., McCollum, D. & Edelenbosch, O. Y. Modelling \nsocial influence and cultural variation in global low-carbon vehicle \ntransitions. Glob. Environ. Change 47, 76\u201387 (2017).\n37.\t Fricko, O. et al. The marker quantification of the Shared \nSocioeconomic Pathway 2: a middle-of-the-road scenario for the \n21st century. Glob. Environ. Change 42, 251\u2013267 (2017).\n38.\t Ioannis, D. Climate policy modelling protocol. Zenodo  \nhttps://doi.org/10.5281/zenodo.11657870 (2024)\n39.\t Saidi, K. & Hammami, S. The impact of CO2 emissions and \neconomic growth on energy consumption in 58 countries.  \nEnergy Reports 1, 62\u201370 (2015).\n40.\t Topolewski, \u0141. Relationship between energy consumption and \neconomic growth in European countries: evidence from dynamic \npanel data analysis. Energies 14, 3565 (2021).\n41.\t Tang, E., Peng, C. & Xu, Y. Changes of energy consumption with \neconomic development when an economy becomes more \nproductive. J. Cleaner Prod. 196, 788\u2013795 (2018).\n42.\t Waheed, R., Sarwar, S. & Wei, C. The survey of economic growth, \nenergy consumption and carbon emission. Energy Reports 5, \n1103\u20131115 (2019).\n43.\t Jones, R. V. & Lomas, K. J. Determinants of high electrical \nenergy demand in UK homes: socio-economic and dwelling \ncharacteristics. Energy Build. 101, 24\u201334 (2015).\n44.\t Bedir, M., Hasselaar, E. & Itard, L. Determinants of electricity \nconsumption in Dutch dwellings. Energy Build. 58, 194\u2013207 (2013).\n45.\t Pablo-Romero, Md. P., S\u00e1nchez-Braza, A. & Anna, G. Relationship \nbetween economic growth and residential energy use in \ntransition economies. Clim. Dev. 11, 338\u2013354 (2019).\n46.\t Komal, R. & Abbas, F. Linking financial development, economic \ngrowth and energy consumption in Pakistan. Renewable Sustain. \nEnergy Rev. 44, 211\u2013220 (2015).\n47.\t Stern, D. I. in Routledge Handbook Of Energy Economics  \n(eds Soyta\u015f, U. & Sari, R.) Ch. 2 (Taylor and Francis, 2019);  \nhttps://doi.org/10.4324/9781315459653\n48.\t Padovani, F., Sommerfeldt, N., Longobardi, F. & Pearce, J. M. \nDecarbonizing rural residential buildings in cold climates: a \ntechno-economic analysis of heating electrification. Energy Build. \n250, 111284 (2021).\n49.\t M\u00fcller-Casseres, E. et al. International shipping in a world below \n2\u2009\u00b0C. Nat. Clim. Change 14, 600\u2013607 (2024).\n50.\t Zhang, R. & Fujimori, S. The role of transport electrification in \nglobal climate change mitigation scenarios. Environ. Res. Lett. 15, \n034019 (2020).\n51.\t O\u2019Keefe, K. T. V. et al. The role of electrification and the power \nsector in U.S. carbon neutrality. Energy Clim. Change 5, 100155 \n(2024).\n52.\t Hill, G., Heidrich, O., Creutzig, F. & Blythe, P. The role of electric \nvehicles in near-term mitigation pathways and achieving the UK\u2019s \ncarbon budget. Appl. Energy 251, 113111 (2019).\n53.\t Edelenbosch, O. Y. et al. Decomposing passenger transport \nfutures: comparing results of global integrated assessment \nmodels. Transp. Res. Part D Transp. Environ. 55, 281\u2013293 (2017).\n54.\t Javaid, A., Creutzig, F. & Bamberg, S. Determinants of low-carbon \ntransport mode adoption: systematic review of reviews. Environ. \nRes. Lett. 15, 103002 (2020).\n55.\t Kriegler, E. et al. How to achieve a rapid, fair, and efficient \ntransformation to net zero emissions - Policy findings from \nthe NAVIGATE project. https://doi.org/10.48485/pik.2023.024 \n(Potsdam Institute for Climate Impact Research, 2023).\n56.\t Albatayneh, A., Assaf, M. N., Alterman, D. & Jaradat, M. \nComparison of the overall energy efficiency for internal \ncombustion engine vehicles and electric vehicles. Environ. Clim. \nTechnol. 24, 669\u2013680 (2020).\n57.\t Font Vivanco, D., Kemp, R. & van der Voet, E. How to deal with the \nrebound effect? A policy-oriented approach. Energy Policy 94, \n114\u2013125 (2016).\n58.\t Galvin, R. & Sunikka-Blank, M. Ten questions concerning \nsustainable domestic thermal retrofit policy research. Build. \nEnviron. 118, 377\u2013388 (2017).\n59.\t Semieniuk, G., Taylor, L., Rezai, A. & Foley, D. K. Plausible energy \ndemand patterns in a growing global economy with climate \npolicy. Nat. Clim. Change 11, 313\u2013318 (2021).\n60.\t Harbers, A. Does urban living reduce energy use? In Carbon \nFootprint and the Industrial Life Cycle: From Urban Planning to \nRecycling (eds \u00c1lvarez Fern\u00e1ndez, R., Zubelzu, S. & Mart\u00ednez, \nR.) 501\u2013518 (Springer, 2017); https://doi.org/10.1007/978-3-319-\n54984-2_22\n61.\t Saujot, M., Le Gallic, T. & Waisman, H. Lifestyle changes in \nmitigation pathways: policy and scientific insights. Environ. Res. \nLett. 16, 015005 (2021).\n62.\t van den Berg, N. J. et al. Improved modelling of lifestyle changes \nin integrated assessment models: cross-disciplinary insights \nfrom methodologies and theories. Energy Strat. Rev. 26, 100420 \n(2019).\n63.\t Pettifor, H., Agnew, M. & Wilson, C. A framework for measuring \nand modelling low-carbon lifestyles. Glob. Environ. Change 82, \n102739 (2023).\n64.\t Pettifor, H. et al. Endogenous simulation of low-carbon lifestyle \nchange in global climate mitigation pathways. Environ. Res. Lett. \n19, 014016 (2024).\n65.\t Hanmer, C., Wilson, C., Edelenbosch, O. Y. & van Vuuren, D. P. \nTranslating global integrated assessment model output into \nlifestyle change pathways at the country and household level. \nEnergies 15, 1650 (2022).\n66.\t van den Berg, N. J., Hof, A. F., Timmer, V., Akenji, L. & van Vuuren, \nD. P. (Path)ways to sustainable living: The impact of the SLIM \nscenarios on long-term emissions. Glob. Environ. Change 84, \n102774 (2024).\n67.\t Hirth, S. et al. Barriers and enablers of 1.5\u00b0 lifestyles: Shallow \nand deep structural factors shaping the potential for sustainable \nconsumption. Front. Sustainability https://doi.org/10.3389/\nfrsus.2023.1014662 (2023).\n68.\t Klitkou, A. et al. The interconnected dynamics of social practices \nand their implications for transformative change: a review. \nSustainable Prod. Consumption 31, 603\u2013614 (2022).\n69.\t Do, V. et al. Spatiotemporal distribution of power outages with \nclimate events and social vulnerability in the USA. Nat. Commun. \n14, 2470 (2023).\n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n393\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\n70.\t Fishman, T. et al. A comprehensive set of global scenarios of \nhousing, mobility, and material efficiency for material cycles and \nenergy systems modeling. J. Ind. Ecol. 25, 305\u2013320 (2021).\n71.\t Levesque, A., Pietzcker, R. C. & Luderer, G. Halving energy \ndemand from buildings: the impact of low consumption \npractices. Technol. Forecast. Social Change 146, 253\u2013266 (2019).\n72.\t Berrill, P. & Hertwich, E. G. Material flows and GHG emissions from \nhousing stock evolution in US counties, 2020\u201360. Build. Cities \nhttps://doi.org/10.5334/bc.126 (2021).\n73.\t Edelenbosch, O. Y. et al. Reducing sectoral hard-to-abate \nemissions to limit reliance on carbon dioxide removal. Nature \nClimate Change https://doi.org/10.1038/s41558-024-02025-y \n(2024).\n74.\t Replogle, M. A. & Fulton, L. M. A Global High Shift Scenario: \nImpacts and Potential for More Public Transport, Walking, and \nCycling with Lower Car Use (ITDP & UC Davis, 2014).\n75.\t Akimoto, K., Sano, F. & Oda, J. Impacts of ride and car-sharing \nassociated with fully autonomous cars on global energy \nconsumptions and carbon dioxide emissions. Technolog. \nForecast. Social Change 174, 121311 (2022).\n76.\t Mulholland, E., Teter, J., Cazzola, P., McDonald, Z. & \u00d3 Gallach\u00f3ir, \nB. P. The long haul towards decarbonising road freight\u2014a global \nassessment to 2050. Appl. Energy 216, 678\u2013693 (2018).\n77.\t Energy Technology Perspectives 2017 (IEA, 2017).\n78.\t Walsh, C., Mander, S. & Larkin, A. Charting a low carbon future for \nshipping: a UK perspective. Marine Policy 82, 32\u201340 (2017).\n79.\t Faber, J., Nelissen, D., Hon, G., Wang, H. & Tsimplis, M. Regulated \nSlow Steaming in Maritime Transport: An Assessment of Options, \nCosts and Benefits (CE Delft, 2012).\n80.\t Edelenbosch, O. Y., Rovelli, D., Levesque, A., Marangoni, G. & \nTavoni, M. Long term, cross-country effects of buildings insulation \npolicies. Technolog. Forecast. Social Change 170, 120887 (2021).\n81.\t D\u2019Agostino, D., Tzeiranaki, S. T., Zangheri, P. & Bertoldi, P. \nAssessing nearly zero energy buildings (NZEBs) development in \nEurope. Energy Strategy Rev. 36, 100680 (2021).\n82.\t Grubler, A. et al. A low energy demand scenario for meeting the \n1.5\u2009\u00b0C target and sustainable development goals without negative \nemission technologies. Nat. Energy 3, 515\u2013527 (2018).\n83.\t Kharina, A. & Rutherford, D. Fuel Efficiency Trends for New \nCommercial Jet Aircraft: 1960 to 2014 (International Council on \nClean Transportation, 2015).\n84.\t Wang, H. & Lutsey, N. Long-term Potential for Increased Shipping \nEfficiency Through the Adoption of Industry-leading Practices \n(International Council on Clean Transportation, 2013).\n85.\t Deason, J. & Borgeson, M. Electrification of buildings: potential, \nchallenges, and outlook. Curr. Sustainable/Renewable Energy \nRep. 6, 131\u2013139 (2019).\n86.\t Gray, N., McDonagh, S., O\u2019Shea, R., Smyth, B. & Murphy, J. D. \nDecarbonising ships, planes and trucks: an analysis of suitable \nlow-carbon fuels for the maritime, aviation and haulage sectors. \nAdv. Appl. Energy 1, 100008 (2021).\n87.\t Barzkar, A. & Ghassemi, M. Electric power systems in more and all \nelectric aircraft: a review. IEEE Access 8, 169314\u2013169332 (2020).\n88.\t Gillingham, K. T. & Huang, P. Long-run environmental and \neconomic impacts of electrifying waterborne shipping in the \nUnited States. Environ. Sci. Technol. 54, 9824\u20139833 (2020).\n89.\t Energy Technology Perspectives 2020 (IEA, 2020).\n90.\t Bouman, E. A., Lindstad, E., Rialland, A. I. & Str\u00f8mman, A. H. \nState-of-the-art technologies, measures, and potential for \nreducing GHG emissions from shipping\u2014a review. Transp. Res. \nPart D Transp. Environ. 52, 408\u2013421 (2017).\n91.\t World Energy Model Documentation (IEA, 2021).\n92.\t van Vuuren, D. P. et al. Alternative pathways to the 1.5\u2009\u00b0C target \nreduce the need for negative emission technologies. Nat. Clim. \nChange 8, 391\u2013397 (2018).\n93.\t Averchenkova, A., Fankhauser, S. & Nachmany, M. (eds) Trends in \nClimate Change Legislation (Elgar, 2017); https://doi.org/10.4337/ \n9781786435781\n94.\t Emissions Gap Report 2019 (UNEP, 2019).\n95.\t IPCC in Climate Change 2021: The Physical Science Basis (eds \nMasson-Delmotte et al.) Ch. 5 (Cambridge Univ. Press, 2023).\n96.\t Ang, B. W. Decomposition analysis for policymaking in energy: \nwhich is the preferred method? Energy Policy 32, 1131\u20131139 \n(2004).\n97.\t Ang, B. W. The LMDI approach to decomposition analysis: a \npractical guide. Energy Policy 33, 867\u2013871 (2005).\n98.\t Edelenbosch, O. Y., van Vuuren, D. P., Blok, K., Calvin, K. & \nFujimori, S. Mitigating energy demand sector emissions: the \nintegrated modelling perspective. Appl. Energy 261, 114347 \n(2020).\nAcknowledgements\nThis project received funding from the European Union\u2019s  \nHorizon Europe research and innovation programmes, grant \nagreement number 821124 (NAVIGATE; to R.v.H., O.Y.E., V.D.,  \nT.L.G., L.B.B., A.D.B., F.P.C., J.E., P.F., R.H., J.H., P.K., F.L., J.L.,  \nG.L., G.M., A.M., H.P., R.P., P.R., B.v.R., R.S., C.W., S.Y., E.Z., D.v.V.) \nand number 101081604 (PRISMA; to R.v.H., V.D., T.L.G., A.D.B., \nJ.E., P.F., F.L., J.L., G.L., A.M., H.P., R.P., B.v.R., C.W., D.v.V.). Parts \nof this work, based on preliminary results, have been previously \npublished in NAVIGATE project reports and presentations55. The PIK \nteam acknowledges funding from the German Federal Ministry of \nEducation and Research under grant agreement number 03SFK5A-2 \n(Ariadne; to J.H., R.H., G.L., R.P.). C.W. acknowledges funding  \nfrom the European Research Council under grant agreement \n101003083 (iDODDLE).\nAuthor contributions\nR.v.H., O.Y.E., G.L., A.M., B.v.R., R.S., S.Y. and D.v.V. conceived and \ndesigned the experiments. R.v.H., V.D., T.L.G., L.B.B., A.D.B., F.P.C., J.E., \nP.F., R.H., J.H., P.K., F.L., J.L., G.M., A.M., R.P., P.R. and E.Z. performed \nthe experiments. R.v.H., O.Y.E., V.D., T.L.G., L.B.B., A.D.B., F.P.C., J.E., P.F., \nR.H., J.H., P.K., F.L., J.L., G.M., A.M., R.P., P.R. and E.Z., analysed the data. \nR.v.H., H.P. and C.W. contributed materials/analysis tools. All authors \nwrote the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41560-025-01703-1.\nCorrespondence and requests for materials should be addressed to \nRik van Heerden or Oreane Y. Edelenbosch.\nPeer review information Nature Energy thanks Rui Jing, \u00c9rika Mata and \nthe other, anonymous reviewer(s) for their contribution to the peer \nreview of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \n\nNature Energy | Volume 10 | March 2025 | 380\u2013394\n394\nArticle\nhttps://doi.org/10.1038/s41560-025-01703-1\nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this licence, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2025\n1PBL Netherlands Environmental Assessment Agency, The Hague, The Netherlands. 2Copernicus Institute for Sustainable Development, Utrecht \nUniversity, Utrecht, The Netherlands. 3Ecole des Ponts Paris Tech, Centre International de Recherche sur l\u2019Environnement et le D\u00e9veloppement (CIRED), \nNogent-sur-Marne, France. 4SMASH \u2013 CIRED, Nogent-sur-Marne, France. 5Centre for Energy and Environmental Economics (Cenergia), Energy Planning \nProgram (PPE), Coppe, Universidade Federal do Rio de Janeiro, Rio de Janeiro, Brazil. 6CMCC Foundation \u2013 Euro-Mediterranean Center on Climate \nChange, Lecce, Italy. 7RFF-CMCC European Institute on Economics and the Environment, Milan, Italy. 8Department of Electronics, Information and \nBioengineering, Politecnico di Milano, Milan, Italy. 9Department of Economics, Ca\u2019 Foscari University, Venice, Italy. 10E3M-Modelling, Athens, Greece. \n11Potsdam Institute for Climate Impact Research (PIK), Potsdam, Germany. 12Global Energy Systems Analysis, Technische Universit\u00e4t Berlin, Berlin, \nGermany. 13Energy, Climate, and Environment Program, International Institute for Applied Systems Analysis (IIASA), Laxenburg, Austria. 14CNRS, CIRED, \nNogent-sur-Marne, France. 15AgroParisTech, CIRED, Nogent-sur-Marne, France. 16Faculty of Technology, Policy and Management, Delft University of \nTechnology, Delft, The Netherlands. 17University of Oxford, Environmental Change Institute, Oxford Centre for the Environment, Oxford, UK. 18Research \nand Innovation Center on CO2 and Hydrogen (RICH Center) and Management Science and Engineering Department, Khalifa University, Abu Dhabi, \nUnited Arab Emirates. 19Chalmers University of Technology, Gothenburg, Sweden. \n\u2009e-mail: rik.vanheerden@pbl.nl; o.y.edelenbosch@uu.nl\n\n\n    Scientific Research Findings:", "answer": "Our analysis shows that a comprehensive set of measures involving end users can reduce sectoral CO2\u00a0emissions by 51\u201385% in buildings and 37\u201391% in transport by 2050, compared to a scenario based on current policies. Electrifying energy end\u2011use and switching to alternative fuels delivers the largest emission reductions, though also leads to increases in electricity demand. Reducing or changing energy\u2011using activities including travel distances and mode choices can reduce energy demand and ease pressure on the electricity supply. So too can using more efficient technologies. The combination of all measures leads to significant further emission reductions, despite some offsetting interactions such as the reduced emission\u2011reduction potentials of heat pumps when homes are better insulated. It is important to note, however, that the successful implementation of these policies depends on multiple conditions, such as broader societal support, which was not explicitly considered in this study.", "id": 3}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 10 | February 2025 | 226\u2013242\n226\nnature energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nDiversity of biomass usage pathways to \nachieve emissions targets in the European \nenergy system\n \nM. Millinger\u2009\n\u200a\u20091,2\u2009\n, F. Hedenus\u2009\n\u200a\u20091, E. Zeyen\u2009\n\u200a\u20093, F. Neumann\u2009\n\u200a\u20093, L. Reichenberg1 & \nG. Berndes\u2009\n\u200a\u20091\nBiomass is a versatile renewable energy source with applications across the \nenergy system, but it is a limited resource and its usage needs prioritization. \nWe use a sector-coupled European energy system model to explore \nnear-optimal solutions for achieving emissions targets. We find that \nprovision of biogenic carbon has higher value than bioenergy provision. \nEnergy system costs increase by 20% if biomass is excluded at a net-negative \n(\u2212110%) emissions target and by 14% at a net-zero target. Dispatchable \nbioelectricity covering ~1% of total electricity generation strengthens \nsupply reliability. Otherwise, it is not crucial in which sector biomass is \nused, if combined with carbon capture to enable negative emissions and \nfeedstock for e-fuel production. A shortage of renewable electricity or \nhydrogen supply primarily increases the value of using biomass for fuel \nproduction. Results are sensitive to upstream emissions of biomass, carbon \nsequestration capacity and costs of direct air capture.\nBiomass is a diverse and versatile renewable energy source that can be \nused for various purposes1\u20133. In the electricity system, it can comple-\nment the variable renewable energy (VRE) sources of solar and wind \npower4\u20137 and provide dispatchable (firm) generation to meet demand \neven in periods of supply shortage in a VRE-based energy system8,9. \nIf used for combined heat and power (CHP), it can provide flexible \nenergy, which may be especially important during so called cold\u2013dark \ndoldrums, when space heat demand is high and electricity supply from \nwind and solar is low10,11. Biomass can also supply hydrocarbons to sec-\ntors that are challenging to electrify and where renewable alternatives \nare scarce, such as aviation and marine transport12\u201314, or plastics and \nhigh-value chemicals15\u201317. Also, it can be used to provide process heat for \nindustry18,19. All of these options can to some extent be combined with \ncarbon capture (BECC) to provide carbon for further usage (BECCU), \nor negative emissions through geological sequestration (BECCS)4,6,18,19. \nIn contrast to direct air capture (DAC), which requires a substantial \nelectricity and heat input to extract CO2 from the atmosphere20, BECC \ncaptures more concentrated CO2 in exhaust and waste streams and \nprovides net energy output along with the carbon capture.\nThe European Union and United Kingdom have adopted targets \nof net-zero greenhouse gas (GHG) emissions for all sectors to comply \nwith the Paris Agreement targets21,22. To achieve such targets, residual \nemissions, such as methane emissions in agriculture, need to be offset \nby carbon dioxide removal (CDR) from the atmosphere, where BECCS \nand DACCS emerge as key options for technical CDR23\u201325.\nBiomass is a limited resource and its use for energy can be associ-\nated with a range of positive and negative environmental, social and \neconomic effects that are context specific and depend on land type \nand climatic region, prior land use and how bioenergy feedstock and \nmanagement regimes are shaped26\u201333. Due to concerns about possible \nenvironmental impacts, insufficient emissions reductions and compe-\ntition with the food sector, EU policy has capped biofuels from food \nand feed crops and increasingly emphasizes lignocellulosic biomass, \nespecially residues and waste34,35, and prioritizes the biomass usage to \nReceived: 22 June 2023\nAccepted: 10 December 2024\nPublished online: 23 January 2025\n Check for updates\n1Department of Space Earth and Environment, Chalmers University of Technology, G\u00f6teborg, Sweden. 2Built Environment: System Transition: Energy \nSystems Analysis, RISE Research Institutes of Sweden, G\u00f6teborg, Sweden. 3Department of Digital Transformation in Energy Systems, Technische \nUniversit\u00e4t Berlin, Berlin, Germany. \n\u2009e-mail: markus.millinger@ri.se\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n227\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nsubstantially61, calling for a thorough assessment of near-optimal \nsolution spaces for biomass usage in the energy system. Also, whereas \nnet-negative emissions trajectories have been assessed widely with \nIAMs, few net-negative analyses have been performed with ESMs11,78.\nThis study addresses this gap by using a sector-coupled ESM, \nwith a more comprehensive coverage of bioenergy technologies and \nBECC than in similar modelling studies (Extended Data Table 1). The \noverarching goal of the study is to analyse effects on the system cost of \nbroad ranges of biomass supply and biomass technology deployment \nin a European energy system adhering to stringent emissions targets. \nThis is done through a detailed exploration of the near-optimal solution \nspace for biomass usage options, within the sector-coupled European \nenergy system optimization model PyPSA-Eur-Sec. The effects of dif-\nferent deployment levels for wind power, solar PV and electrolysers on \nbiomass usage, and vice versa, are also assessed. More specifically, we \nanalyse the solution space around the least-cost optimum, for system \ncost increases of 1%, 5%, 10%, through to 25%.\nScenarios with net-negative (\u2212110%) and net-zero CO2 emissions \ncompared with 1990 levels are assessed. No explicit target year is mod-\nelled, as focus is on how these targets can be met cost effectively, rather \nthan on pathways leading there. European policies prioritize emissions \nreductions over compensation of emissions through CDR22,44,79\u201382. Also, \nalthough theoretical geological storage capacities are vast, invest-\nable potentials and CO2 injection rates as indicated by historical fos-\nsil extraction may be limited48. Reflecting these aspects, in the main \nscenarios, we assess an energy system where only very little compen-\nsation of concurrent fossil fuel usage through negative emissions is \npermissible by setting carbon sequestration capacities near the limit \nfor what is necessary to achieve the respective emissions targets, while \nalso offsetting process emissions considered to be unavoidable, for \ninstance, from cement production (Extended Data Fig. 2; 600\u2009Mt\u2009CO2 \nper year for net-negative (\u2212110%) and 140\u2009Mt\u2009CO2 per year for net-zero, \nthus allowing the same slack for fossil fuel usage in both cases). The \neffect of higher assumed carbon sequestration capacities is analysed \nand discussed in a sensitivity analysis.\nReflecting current EU policy direction, bioenergy feedstock is \nassumed to consist of residues, with a domestic supply potential cor-\nresponding to the medium level in the JRC ENSPRESO database83. \nImport of non-digestible biomass represents a complementary, but \nmore expensive, feedstock source11 (Extended Data Fig. 3). The effect \nof variations to these assumptions is assessed in sensitivity analyses.\nWe find that provision of biogenic carbon for negative emissions \nand utilization has a higher value than bioenergy provision. Energy \nsystem costs increase by 20% if biomass is excluded at a net-negative \n(\u2212110%) emissions target. Dispatchable bioelectricity covering ~1% of \ntotal electricity generation strengthens supply reliability. Otherwise, \nit matters less whether biomass is used for combined heat and power, \nliquid fuel production or industrial process heat, as long as the car-\nbon content is utilized to a high extent, as facilitated through carbon \ncapture to provide renewable carbon for negative emissions or for \nproduction of fuels for further use in the energy system.\nThe cost of varying biomass use in the energy system\nIn the cost-optimal solution for the net-negative scenario, wind (54%), \nsolar photovoltaics (PV) (40%) and hydropower (5%) supply 99% of the \nwhole electricity demand at 9,250\u2009TWh (Extended Data Fig. 4), which is \nalmost three times the electricity demand in 202184,85. Biomass is mainly \nused to complement the supply of fuels and chemicals for industry, avia-\ntion and shipping, but a small share is also used to supply dispatchable \nelectricity (Extended Data Fig. 4). Some 637\u2009TWh biogas and 2,896\u2009TWh \nsolid biomass (2,172\u2009TWh imported) are used, corresponding to 29% \nof the annual primary energy consumption at 13\u2009PWh. Solid biomass \nusage amounts to about two times the 1,290\u2009TWh used in 2021, when \noverall bioenergy usage was at 1,937\u2009TWh (refs. 84,85). Around 87% of \nbiomass usage is combined with carbon capture, with the exception \nenergy applications where other alternatives are currently difficult to \nfind or considered to be too costly.\nAll of the possible biomass usage options face competition from \nelectricity-derived energy carriers and fossil fuels (Extended Data \nFig. 1). A full systems analysis of biomass allocation to different energy \nuses therefore requires broad coverage of options and sectors. Such \nanalyses have been carried out with global integrated assessment \nmodels (IAMs), with a large variety in results, but with biomass in the \nlonger term generally ending up being used for electricity and/or liquid \nfuels production36, coupled with carbon capture and storage (CCS)37,38. \nThe potential value of negative emissions from BECCS has been found \nto be very high, enabling the achievement of more ambitious climate \ntargets37,39\u201341 or delayed phase-out of fossil fuels if temperature over-\nshoot is permitted40,41. The latter raises concerns over risks in relying \non future technology deployment to compensate for earlier emissions \nand over intergenerational equity42\u201345.\nHowever, IAMs lack the spatio-temporal detail needed to capture \nthe variability of, for instance, VRE and electrolysis, and the interplay \nof these technologies with biomass options, such as dispatchable bio-\nelectricity. In addition, the costs of VRE have often been overestimated \nin IAM-based analyses46,47 and carbon sequestration capacities may be \nmore limited than what has been assumed48,49; both of these factors risk \nexaggerating the role of CCS for meeting climate targets48,50. Moreo-\nver, IAMs have until now not included carbon capture and utilization \n(CCU) or electrofuels11,47, leaving biofuels as the only non-fossil-fuel \noption, and until recently also not included DAC as an alternative CDR \noption47,51,52. Most or all of these limitations apply also to previous IAM \nanalyses focusing specifically on biomass and/or BECCS36,37,39\u201341,53\u201359, \nleading to potential biases in the cost effectiveness of biomass usage, \nBECCS and different biomass utilization pathways.\nEnergy system models (ESMs) on the other hand commonly rep-\nresent VRE explicitly, with a high spatial and temporal resolution, and \nsome ESMs have recently been enhanced to encompass all energy \nsectors simultaneously60,61. This enables a sector-coupled analysis of \nbiomass usage for energy across all sectors and of interactions with \ncompeting fuel options that can be produced from VRE sources, such \nas hydrogen and electrofuels. However these models usually include \na restricted selection of biomass applications and, in contrast to IAMs, \nonly a few studies based on sector-coupled ESMs have focused explic-\nitly on biomass, bioenergy and/or BECC11,62, and a thorough assessment \nof biomass usage including BECCUS across usage options is still lacking. \nFurther, combining bioenergy processes with conventional carbon \ncapture results in higher costs for the additional capture and heat \ninfrastructure and energy penalties to provide the substantial process \nheat needed to regenerate solvents. Such details have, to date, not been \nincluded in analyses with sector-coupled ESMs but are also lacking in \nmany IAMs (Extended Data Table 1).\nIAM and ESM studies commonly focus on the single cost-optimal \nsolution, complemented with some sensitivity analyses. However, \nsocial planning projects are subject to a plurality of economic and \nsocio-political objectives63,64, and uncertainties and objectively irrecon-\ncilable trade-offs at different levels regarding future energy systems65 \nand biomass use66\u201368 are so-called wicked facets of their planning69. \nThe sector-coupled energy system involves diverse stakeholders with \nconflicting non-economic objectives and risk perceptions, and past \nenergy transitions have been found not to follow cost-optimal paths \nin hindsight70. There is therefore a value in exploring the diversity of \nnear-optimal solutions for the energy system in general and for biomass \nusage in particular, to provide insights for policy about the flexibility \nof solutions71. Recent analyses have shown that the technology mix \nvariety of near-optimal solutions, when allowing a small system cost \nincrease, can be distinctly different from the single least-cost solution, \nin heat supply72, the power system73\u201376, in integrated assessments77 \nand in a sector-coupled European energy system61,64. The available \namount of biomass has been found to affect the manoeuvring space \n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n228\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nbeing dispatchable biomethane applications. The total annual system \ncost amounts to \u20ac822\u2009billion.\nIf overall biomass usage is restricted to current usage levels, the \nsystem cost ends up ~5% higher than without restrictions. If all biomass \n(except mandatory incineration of municipal solid waste) is excluded, \nit leads to a 20% higher system cost (Fig. 1a), or an additional cost of \n\u20ac169\u2009billion annually, roughly corresponding to European defence \nexpenses86. This is twice as much as the cost of excluding solar power \nand similar to the cost of excluding wind power, despite both of them \ncost optimally providing more primary energy (Fig. 2d,e). Excluding \nany of these primary energy sources thus leads to much higher costs. \nWind and solar power are more readily interchangeable whereas the \nsubstitution of biomass is much more expensive because biomass pro-\nvides non-fossil carbon in addition to energy, for which the substitute, \nDAC, coupled with a necessary expansion of additional energy provi-\nsion, ends up much more expensive. Wind and solar power cannot be \nexcluded simultaneously even at a 25% system cost increase (Fig. 2f).\nThe assumed biomass availability determines both relative and \nabsolute costs of excluding biomass. If biomass imports are not \nincluded as an option, excluding biomass altogether results in a 9% \nhigher system cost. If the high domestic residue potential estimate \nfrom JRC ENSPRESO is used instead of the medium potential, and \nbiomass imports are included as an option, the system cost increases \n29% when excluding biomass altogether. Excluding biomass imports \nresults in substantially higher solid biomass and CO2 prices (Table 1), \nindicating a high system pressure for increasing biomass supply.\nSubstantial flexible dispatchable methane-based power capacities \nemerge in the net-negative scenario (521\u2009GWel open-cycle gas turbines \nand gas CHPs), on par with the inflexible electricity demand peak of \n526\u2009GW (base-load household, commercial and industrial electricity \ndemand, excluding heat). This flexible power capacity is seldom used, \nwith capacity factors of 28% (Fig. 3), which renders the addition of \ncostly carbon capture to these power plants prohibitively expensive. \nFor carbon capture to be cost effective for a particular technology, \nhigh utilization rates are needed due to the high investment cost of \nthe additional infrastructure.\nAlthough only 225\u2009TWh (bio)methane is used to flexibly supple-\nment variable renewable electricity supply (covering 1% of total genera-\ntion), this option is the most costly to replace and remains longest when \nbiomass usage is minimized (Fig. 2a). Different to studies limited to the \npower system only, which indicate a substantially larger firm generation \nenergy requirement8, or IAM studies, which often obtain substantial \nbiomass use for electricity production36, this study finds lower levels \nof bioelectricity use because the sector-coupled model entails large \nflexible demand capacities such as electrolysers, heat storage, batteries \nand electric vehicles, which handle most of the variability in the power \nsystem and thus support high VRE shares (Fig. 3).\nSystem flexibility from sector coupling, energy storage and trans-\nmission reduces the dependency on biomass. With lower assumed \nflexibility, the system cost of excluding biomass increases from 20% \nto 23% (Table 1). Similar amounts of biomass are used for flexible bio-\nelectricity, but least-cost biomass usage shifts from fuel production \nto heat generation.\nExcluding biomass in the net-zero scenario increases system costs \nby 14%, substantially less than the 20% increase in the net-negative \nscenario. The cost-optimal biomass use (Fig. 4) is 36% lower than in the \nnet-negative scenario and within the range of the European Commis-\nsion net-zero scenarios (2,200\u20132,900\u2009TWh) (ref. 87). Biomass usage is \nstill cost optimally coupled with carbon capture, and solution spaces \nfor individual options are rather similar to the net-negative scenario.\nBiomass carbon is more valuable than bioenergy\nFor the net-negative scenario, 87% of biomass use is cost optimally \ncombined with carbon capture, providing 0.84\u2009Gt biogenic CO2 annu-\nally, corresponding to ~21% of total regional GHG emissions in 2021, at \n4\u2009Gt\u2009CO2-equivalent (ref. 88). The captured amount falls within projected \nfeasible CCS growth already for 2040, of 1\u20134.3\u2009Gt per year globally49, but \nwould require a ramp-up of BECC from currently near-zero commercial \ncapacity to covering almost all biomass conversion.\nRenewable carbon provision is the key system service of bio-\nmass, more so than the energy provided. The only other alternative \nfor non-fossil carbon provision is DAC, which is substantially less \nSpan at a 1%\nhigher system cost\nHigh domestic residue\npotential in JRC ENSPRESO\nTotal bioenergy 2021\nAssumed domestic residue potential\n(medium from JRC ENSPRESO)\nSolid biomass usage 2021\nCurrent solid biomass\nusage obtainable at a\n7% higher system cost\nBiomass exclusion\nat 20% higher\nsystem cost\nMaximum biomass\ndemand (net)\nMaximum \nbiomass demand \n(excluding DAC)\nMaximum biomass demand\nfor liquid fuels\nTWh biomass\nMt CO2 captured\n6,000\n\u2013110% CO2, 600 Mt CO2 storage\n(TWh biomass)\n5,000\n4,000\n3,000\n2,000\n1,000\n0\n6,000\n5,000\n4,000\n3,000\n2,000\n1,000\n0\n2,000\n1,500\n1,000\n500\n0\n0\n0\n1\n3\n5\n0\n1\n3\n5\n5\n0 1\n3\n5\n0\n1\n3\n5\n10\n10\n15\n10\n20\n15\na    All biomass excluding MSW\nb    CHP\nc    Process\n                       heat\ne    BECC\nd    Biofuel\n\u03b5 (%)\n\u03b5 (%)\n\u03b5 (%)\n\u03b5 (%)\n\u03b5 (%)\nFig. 1 | Solution spaces for different biomass usages in the net-negative \nemissions scenario with an allowed carbon sequestration potential close to \nwhat is necessary to achieve the target. The respective options are minimized \nand maximized to map out the space of feasible solutions (shaded area) for a \ngiven allowed system cost increase \u03b5 (percent deviation from the total system \ncost). a,e, The solution space when varying the amount of biomass (except \nmunicipal solid waste) (a) and the solution space when varying bioenergy with \ncarbon capture deployment (e). In a, solid biomass usage in 202184,85 is shown and \nis similar to the assumed medium domestic residue potential, in contrast to the \nhigh potential, both from JRC ENSPRESO83. Total bioenergy in 202184,85 includes \nbiomass residues and agricultural crops. b\u2013d, Horizontal lines show biomass \ndemands if the full demand for district heat (b), industrial process heat (c) and \nliquid fuels (d) would be fulfilled by solid biomass. The heat demand for CHP can \nbe expanded to supply thermal storage and the industrial process heat demand \ncan be used to supply DAC heat demand. MSW, incineration of municipal solid \nwaste, which is set to be compulsory.\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n229\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\ncost competitive than the available alternative energy provision \noptions. The value of biogenic carbon is estimated to be up to over \nthree times higher than the value of the primary energy provision at \nlow biomass usage levels in a renewable energy system, in both the \nnet-negative (Fig. 5a) and net-zero scenarios (Fig. 5c). When a higher \ncarbon sequestration allowance permits a larger amount of fossil \nfuels to be offset by the sequestration of biogenic CO2, the value of \nthe biogenic carbon is up to two times higher (Fig. 5b). With varying \namounts of biomass in the system, CO2 and solid biomass prices are \nstrongly and similarly affected, whereas the hydrogen price is sub-\nstantially less affected.\nBECC can be excluded at a 13% system cost increase (Fig. 1e), with \nmainly biofuel production decreasing whereas biogas and biomass \nusage for process heat and flexible bioelectricity remain cost effec-\ntive also without BECC (Fig. 2b). If BECC is removed, biomass can be \nexcluded within a 6% cost increase (Table 1), substantially less than with \nBECC. Capturing biogenic CO2 emissions enhances carbon utilization, \nenabling scarce renewable carbon to be used multiple times and to \nprovide negative emissions. As BECC is decreased, biomass usage also \ndecreases, and DAC increases to provide both the necessary negative \nemissions and carbon for the production of electrofuels, resulting in \nhigher total carbon capture deployment (Fig. 2b).\nIn the least-cost case, the shadow price (marginal price of an addi-\ntional MWh) of solid biomass amounts to \u20ac54\u2009MWh\u22121 (as determined \nby the import biomass price) and \u20ac135\u2009MWh\u22121 if biomass is excluded \n(Fig. 5a and Table 1). This is substantially higher than the cost of domes-\ntic residue supply (Extended Data Fig. 3 and Extended Data Table 2) or \n2020 wood chip prices at \u20ac20\u201325\u2009MWh\u22121 (ref. 89). The resulting CO2 \nprice (marginal cost of CO2 emissions) amounts to ~\u20ac260\u2009t\u22121\u2009CO2 in the \ncost-optimal case but increases to \u20ac591\u2009t\u22121\u2009CO2 if biomass is excluded \n(Fig. 5a and Table 1), indicating a high value of biomass resources to \nachieve emissions targets.\nBiomass allocation is not crucial if carbon is \ncaptured\nSolid biomass is cost optimally used for biofuel production and process \nheat (Extended Data Fig. 4), but a large range of near-optimal solutions \nfor different usage options exist (Fig. 1b\u2013d). Thus, even though it is \ncostly to exclude overall biomass usage, it is not so important in which \nsectors biomass is used.\nDistrict heat is cost optimally covered by a mix of excess heat \nfrom biofuel production and electrolysers, waste incineration, electric \nboilers and some (bio)methane-fuelled CHP (Extended Data Fig. 4). \nWhereas absent in the cost-optimal solution, solid biomass CHP can \ncover up to 50% of district heat demand within a 1% system cost increase \n(corresponding to 16% of district heat cost and thus not increasing the \nsectoral costs substantially; Fig. 1b). These CHP plants are invariably \nequipped with carbon capture, which increases capital cost substan-\ntially (Extended Data Table 1), and they are therefore run with high \ncapacity factors (>90%), coupled with heat storage, highlighting the \npriority of carbon capture over additional variation management sup-\nporting wind and solar feed-in.\nSolid biomass competes with hydrogen and (bio)methane for \nprocess heat supply in the medium-temperature range and mainly with \nelectricity for process steam production. It can cover a span of 0\u2013100% \nin these sub-sectors within the range of a 0.5% system cost increase (7% \nof costs in these sectors; Fig. 1c). Thus, a very diverse set of alternative \noptions exists within a small cost span.\nBiofuels cover a span of 20\u201361% of liquid fuel demand for aviation, \nshipping and chemicals already within the range of a 1% system cost \n6,000\n8,000\n4,000\n2,000\n0\n6,000\n8,000\n4,000\n2,000\n0\n0\n500\n1,000\nTWh biomass\nTWh\nMt CO2\n0\n0\n0\n0\n0\n0\n1\n3 5\n5\nBiofuel + electrobiofuel\nWind\nSolar\nElectrolysis\nNuclear\nBECC\nElectrofuel\nBioSNG\nBiomass CHP\nBiogas\nBiomass process heat\n5\n5\n5\n5\n10\n10\n10\n10\n10\n10\n20\n20\n20\n20\n25\n15\n15\n15\n15\n15\n\u03b5 (%)\n\u03b5 (%)\n\u03b5 (%)\n\u03b5 (%)\n\u03b5 (%)\n\u03b5 (%)\nDAC\na    All biomass excluding\nMSW\nb    BECC\nc    Electrolysis\nd    Wind\ne    Solar\nf    VRE    \nFig. 2 | Effect on the rest of the energy system when decreasing different \noptions from cost-optimal levels. Upper panels show biomass usage (TWh \nbiomass), the middle panels show other energy technologies (TWh) and the \nlower panels show carbon capture (Mt CO2 for BECC (bioenergy with carbon \ncapture) and DAC (direct air carbon capture)). a, For example, total biomass \n(excluding mandatory waste incineration) is the objective function, which is \nminimized. In the upper panel the effect on biomass usage can be observed \nas total biomass use decreases. In the middle panel the effect on other energy \ntechnologies such as wind power can be observed as total biomass use decreases, \nand in the lower panel the effect on carbon capture is shown. b\u2013e, BECC is \nminimized as shown in the lower panel (b), whereas for electrolysis (c), wind \npower (d) and solar PV (e), the minimized technology is shown as a solid line in \nthe middle panels. f, The results when minimizing variable renewables (solar PV \nand wind power). BioSNG, biogenic synthetic natural gas.\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n230\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nincrease (Fig. 1d). A wide near-optimal range for fuel supply appears: \ndirect biomass usage for liquid fuel production can be excluded at a 3% \nsystem cost increase (8% of liquid fuel supply cost), while covering the \nfull liquid fuel demand with only biofuels can be done at an 8% system \ncost increase (resulting in more than a doubling of biomass imports).\nWhen biomass use is decreased from the cost-optimal amount, \nthe primary change occurs in the production of liquids used as fuels in \naviation and shipping and as feedstock for chemicals, where bioliquids \nare replaced by electrofuels. Electricity generation from wind and solar \npower increases more than the usage of bioenergy decreases due to \nthe increasing electricity demand for electrolysis and DAC to supply \nhydrogen and carbon for electrofuel production (Fig. 2a).\nRole of biomass if VRE or electrolyser capacity is \nlimited\nThe cost-optimal results depend on a large expansion of solar and \nonshore wind power and even more so if biomass is excluded (Fig. 2a \nand Extended Data Table 3). Achieving these VRE capacities requires an \nunprecedented capacity growth at the European scale, which might be \nhindered, for instance, by industry scale-up inertia and local opposition \nwhere wind and solar projects are planned5,90\u201394. When VRE is restricted \nto a level below the least-cost case, a decrease in total electricity gen-\neration and hydrogen electrolysis to supply electrofuel production \nis observed, whereas biomass use increases to supply more biofuel \nproduction and process heat (Fig. 2f).\nResults also depend on a large capacity expansion of hydrogen \nelectrolysis, which similarly requires an unprecedented scale-up \n(Extended Data Table 3). The cost-optimal electrolysis capacity is \nmore than two times projected feasible capacity growth by 2050 for the \nEuropean Union and covering a substantial share of projected global \ncapacities95 and again even more so if biomass is excluded (Extended \nData Table 3). Decreasing electrolysis from cost-optimal levels leads to \na corresponding reduction in electricity consumption and a decrease \nin electrofuel production, which is again balanced by an expansion of \nbiomass use to supply biofuel production. Electrolysis can be excluded \nat an ~20% system cost increase, at which point biomass use is almost \ndoubled compared to cost-optimal levels (Fig. 2c). A similar magnitude \nof biomass usage emerges when VRE is minimized within the same \nsystem cost increase (Fig. 2f).\nThus VRE or electrolyser expansion inertia primarily affects liquid \nfuel supply and leads to a much higher demand for biomass if emission \ntargets are to be achieved. Vice versa, a shortage of biomass increases \nthe value of electrofuels.\nSensitivity to upstream emissions and \nsequestration capacity\nDomestic biomass resources are limited to residues, which have a \nlower risk of indirect emissions compared to dedicated energy crops, \nbut residue extraction can cause soil carbon losses and impact soil \nhealth96\u2013100, which in turn can impact yield levels, potentially leading to \nindirect emissions if production is expanded elsewhere to compensate \nfor declining harvest levels. In addition, biomass imported into Europe \nmay be associated with GHG emissions along the supply chain.\nEuropean policies can restrict imports of high GHG biomass and \nrequire domestic residue extraction to follow best-management practices \nto minimize soil impacts. Supply chain emissions can be expected to \ndecrease if other regions also adopt stringent emissions targets, and resid-\nual emissions can be offset through CDR. However, whereas innovation \nand changes in land-management practices can lead to dramatic emis-\nsions reductions, implementation may be a multi-decadal process23,101\u2013104 \nand biomass supply may still be associated with emissions due to weak \ncompliance and leakage effects105,106. Furthermore, CDR implementation \ncould offset emissions associated with other activities if it is not needed \nto offset residual emissions from the biomass supply chain.\nAssuming biomass imports to entail upstream emissions can have \na strong influence on biomass usage in the energy system, depend-\ning on the assumed allowed carbon sequestration potential (Fig 6). \nA higher allowed annual carbon sequestration potential beyond the \nrestrictive 600\u2009Mt CO2 per year limit results in similar but somewhat \nwider solution spaces for all biomass usage options (Fig. 4), as there is \nroom for using more fossil fuels if emissions are captured and stored \nor counterbalanced by negative emissions.\nIf biomass imports are assumed to be carbon neutral (that is, \nwithout upstream emissions), least-cost biomass usage amounts are \nstable across a wide range of carbon sequestration allowances (Fig. 6a). \nHowever, the cost to exclude biomass decreases substantially with \nincreasing allowed carbon sequestration potential (Fig. 6h), as it opens \nup for using fossil fuels (Fig. 6i) combined with CCS, which decreases \nthe cost effectiveness of CCU (Fig. 6f,g).\nHowever, already when assuming upstream emissions for biomass \nimports of 10.3\u2009g\u2009CO2\u2009MJ\u22121, results depend heavily on the allowed carbon \nsequestration potential. At a low carbon sequestration potential, only \nmarginal amounts of additional upstream emissions from biomass \nusage can be accommodated (Fig. 6c), and the system cost of excluding \nbiomass altogether decreases substantially from 20% to 11% (Fig. 6e). \nIn this case, especially, biofuel production decreases (Fig. 6b), and DAC \nis preferred to supply carbon for electrofuel production (Fig. 6d). At \nallowed carbon sequestration potentials of 800\u20131,600\u2009Mt\u2009CO2, there \nis room for more upstream emissions as they can be compensated for \nby using more biomass (Fig. 6a) and BECC (Fig. 6d). Higher carbon \nsequestration potentials allow for increasing use of fossil CCS (Fig. 6f,i) \ncombined with negative emissions, which are then provided by DAC \nrather than BECC if biomass entails upstream emissions (Fig. 6e).\nSensitivity to biomass and DAC cost and carbon \ncapture rate\nDAC can decrease system reliance on biomass, but BECC is more \ncompetitive than DAC for delivering renewable carbon and negative \nTable 1 | Sensitivity of near-optimal system cost and other \nmetrics\nSystem \ncost\nBiomass \nusage\nCO2 \nprice\nSolid \nbiomass \nshadow \nprice\nHigher \ncost when \nexcluding \nbiomass\nRelative \nincrease\nBiomass \namount\nBillion \u20ac\nTWh\n\u20ac t\u22121\u2009CO2\n\u20ac\u2009MWh\u22121\nBillion \u20ac\n%\nHigh + \nimport\n768\n3,561\n262\n54\n223\n29\nHigh \n\u2212 import\n788\n3,070\n368\n88\n203\n26\nMedium \n+ import\n822\n3,533\n260\n54\n169\n20\nMedium \n\u2212 import\n906\n1,372\n400\n96\n85\n9\nNone\n991\n\u2013\n591\n135\n\u2013\n\u2013\nMedium \n+ import \nand low \nflex\n911\n3,846\n292\n54\n206\n23\nMedium \n+ import \n\u2212 BECC\n930\n3,034\n367\n54\n60\n6\nMedium \n+ import \n\u2223 net zero\n756\n2,176\n260\n53\n107\n14\nAssumed biomass amounts and other variations for achieving an \u2212110% net-negative \nemissions (and in the last case for net zero). High- and medium-biomass scenarios use the \ncorresponding potentials from ref. 146 shown in Fig. 1, with (+) and without (\u2212) biomass \nimports. The details of the lower flexibility scenario are elaborated in Methods.\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n231\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\n2,500\n2,000\n1,000\n500\n250\n250\n500\n1,000\nElectricity consumption \u2190 (GW) \u2192 electricity generation\n2,000\n2,500\n07\n14\n21\n28\nFull year\nJanuary\nOCGT\nGas CHP\nWaste incineration\nOn-wind\nOf-wind\nPHS discharge\nSolar\nHydro\nV2G\nBattery discharge\nBase electricity load\nCommercial electricity load\nPHS charge\nProcess steam electricity\nHeat pump\nElectric heating\nBattery charge\nBEV charger\nElectrolysis\n0\nFig. 3 | Electricity generation and consumption. Electricity generation and \nconsumption for achieving net-negative (\u2212110%) emissions over the full year and for \nJanuary. 521 GWel dispatchable firm generation capacity emerges (open-cycle gas \nturbines (OCGT) and gas CHPs), on par with the inflexible electricity demand peak \nof 526\u2009GW (base-load household, commercial and industrial electricity demand, \nexcluding heat), but capacity factors are low, at 2% and 8%, respectively. Electricity \ngeneration and consumption are shown above and below the zero line, respectively. \nBEV, battery electric vehicle; PHS, pumped hydro storage; V2G, vehicle-to-grid.\n6,000\n5,000\n4,000\n3,000\n2,000\n1,000\n\u2013100% CO2, 140 Mt CO2 storage\n(TWh biomass)\n\u2013110% CO2, 1,500 Mt CO2 storage\n(TWh biomass)\n0\n6,000\n5,000\n4,000\n3,000\n2,000\n1,000\n0\n2,000\n1,000\n1,500\n500\n0\n2,000\n1,000\n1,500\n500\n0\n0 1\n3\n5\n10\n20\n15\n\u03b5 (%)\n\u03b5 (%)\n\u03b5 (%)\n\u03b5 (%)\n\u03b5 (%)\n0 1\n3\n5\n10\n20\n15\n\u03b5 (%)\n0 1\n3\n5\n0 1\n3\n5\n0 1\n3\n5\n10\n\u03b5 (%)\n\u03b5 (%)\n\u03b5 (%)\n0 1\n3\n5\n0 1\n3\n5\n0 1\n3\n5\n10\n0\n5\n10\n15\n\u03b5 (%)\n0\n5\n10\n15\nMt CO2 captured\nMt CO2 captured\nMaximum biomass\ndemand (net)\nMaximum biomass\ndemand (net)\nMaximum biomass\ndemand\n(excluding DAC)\nMaximum biomass\ndemand\n(excluding DAC)\nMaximum biomass demand\nfor liquid fuels\nMaximum biomass demand\nfor liquid fuels\nb    CHP\na    All biomass excluding MSW\nf    All biomass excluding MSW\ng    CHP    \nc    Process\n       heat\nh    Process\n                      heat\nd    Biofuel\ni    Biofuel\ne    BECC\nj    BECC\nFig. 4 | Solution spaces for different biomass usages. a\u2013e, In a net-zero scenario \nwith carbon sequestration capacity set to 140\u2009Mt CO2 per year, close to what \nis necessary to achieve the target (a\u2013e), and a net-negative (\u2212110%) emissions \nscenarios with an allowed carbon sequestration potential of 1,500\u2009Mt\u2009CO2 per year \n(900\u2009Mt\u2009CO2 per year more than in the base case) (f\u2013i). The respective options are \nminimized and maximized to map out the space of feasible solutions (shaded area) \nfor a given allowed system cost increase \u03b5 (percent deviation from the least-cost \nobjective value). a,f, Solution spaces when varying the amount of biomass (except \nmunicipal solid waste) in the net-zero scenario (a) and in the net-negative scenario \nwith higher carbon sequestration potential (f). b,g, Solution spaces when varying \nsolid biomass use for combined heat and power in the net-zero (b) and net-negative \n(g) scenarios. c,h, Solution spaces when varying solid biomass use for industrial \nprocess heat in the net-zero (c) and net-negative (h) scenarios. d,i, Solution spaces \nwhen varying solid biomass use for liquid fuel and chemical production in the net-\nzero (d) and net-negative (i) scenarios. e,j, Solution spaces when varying bioenergy \nwith carbon capture deployment in the net-zero (e) and net-negative (j) scenarios. \nb\u2013d,g\u2013i, Horizontal lines show biomass demands if the full demand for district heat \n(b,g), industrial process heat (c,h) and liquid fuels (d,i) would be fulfilled by solid \nbiomass. The heat demand for CHP can be expanded to supply thermal storage, \nand the industrial process heat demand can be used to supply DAC heat demand.\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n232\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nemissions across a large span of biomass import prices (Fig. 7) and cap-\nture rates (Extended Data Fig. 5). Only if DAC capital costs are assumed \nto be below about \u20ac200\u2009t\u22121\u2009CO2 per year does DAC become competitive \nat an import price of \u20ac72\u2009MWh\u22121 (3\u20133.5 times 2020 wood chip prices) \nor at carbon capture rates of 75% and below. However, there is a strong \ncost incentive for achieving high capture rates to reduce expensive \nprimary carbon input (Extended Data Fig. 5a,b) and as long as costs for \nbiomass residue collection and transport and so on are covered, bio-\nmass can remain profitable at substantially lower biomass prices. DAC \nthen serves rather as a backstop technology to prevent high scarcity \nprices for biomass and renewable carbon as raw material and enables \nthe achievement of emissions targets if biomass is too scarce107.\nThe system cost of excluding biomass varies between 13\u201325% \n(\u20ac104\u2013205\u2009billion) at DAC capital costs corresponding to \u20ac171\u2013\n800\u2009t\u22121\u2009CO2 per year and a baseline biomass import price of \u20ac54\u2009MWh\u22121. \nThus, BECC is substantially more cost competitive even with optimistic \nDAC investment costs. This is likely to inhibit the scale-up of DAC and \ntherefore its cost progression through technological learning, which is \nsubject to large uncertainties even in the case of gigaton-scale deploy-\nment108 (Extended Data Table 4). To achieve high capacity factors \nand thereby keep costs down, DAC benefits greatly from a stable and \nsubstantial supply of electricity and heat108, and the energy source can \nonly cause small or zero emissions to enable net-CO2 removal109. The \nabsence of these conditions in the short to medium term on a European \nscale prevents a large scale-up of DAC (and thereby cost progression). \nIn contrast, BECC can be scaled up already in the short term, provided \nthat sufficient biomass is available.\nDiscussion and conclusions\nExcluding biomass use in a fossil-free energy system adhering to a nega-\ntive (\u2212110%) emissions target results in a 20% higher system cost and a \nsubstantially larger and more challenging expansion of VRE, electrolys-\ners, electrofuels and DAC compared to if biomass is available. The cost \nincrease is similar to the cost of excluding wind power, and the main \nreason is the high value of renewable carbon rather than of the energy \nprovision of biomass. It matters less whether biomass is used for com-\nbined heat and power, liquid fuel production or industrial process heat, \nas long as the carbon content is utilized to a high extent, as facilitated \nthrough carbon capture to provide renewable carbon for negative emis-\nsions or for production of fuels for further use in the energy system. \nThere is large potential for carbon capture also in Fischer\u2013Tropsch \nbiofuel production, where ~70% of the biogenic carbon ends up in the \nwaste stream unless additional hydrogen is added to utilize more of the \ncarbon directly (electrobiofuels). Similar results emerge for a net-zero \nemissions target, but biomass can then be excluded at a 14% system cost \nincrease due to a lower negative emissions requirement. DAC costs are \nhighly uncertain108 and substantially affect the system cost of excluding \nbiomass, which explains some of the diversity of results in the literature \n(Extended Data Table 4), but BECC was found to remain more cost effec-\ntive even at low DAC costs across a range of assumptions.\na    Net-negative (\u2013110%) 600 Mt CO2\nb    Net-negative (\u2013110%) 1,500 Mt CO2\nc    Net-zero 140 Mt CO2\n600\nPrice ( \u20ac MWh\u22121 or t\u22121 CO2)\nIndex value normalized\nto case without biomass\n450\n300\n150\n60\n0\n0\n1,000\n2,000\n3,000\n0\n1,000\nBiomass used (excluding MSW) (TWh)\nC/H value ratio\nSolid biomass price (\u20ac MWh\u22121)\nCO2 price (\u20ac t\u22121 CO2)\nH2 price (\u20ac MWh\u22121)\n2,000\n3,000\n+117%\n+88%\n+120%\n4\n3\n2\n1\n0\n+7%\n+14%\n+11%\n+139%\n+109%\n+141%\nSolid biomass C/H value ratio\n1.0\n0.9\n0.8\n0.7\n0.6\n0.5\n0.4\n0\n1,000\n2,000\n3,000\nFig. 5 | Approximation of the carbon and energy value ratio of solid biomass \nbetween the least-cost biomass usage and when biomass, except for MSW, \nis excluded, as calculated through the resulting shadow prices for CO2 and \nH2. H2 is assumed as the reference for biomass energy, with prices consumption \nweighted over all regions and time steps. Upper panels show resulting shadow \nprices (left y axis) and the carbon/hydrogen value ratio of solid biomass (right \ny axis). Lower panels show the prices normalized to the value when biomass \nis excluded. a,b, Net-negative (\u2212110%) scenarios with carbon sequestration \ncapacity limited to 600 (a) and 1,500\u2009Mt\u2009CO2 (b). c, A net-zero scenario with \ncarbon sequestration capacity limited to 140\u2009Mt\u2009CO2. Solid biomass is assumed to \ngenerate 0.3667\u2009t\u2009CO2\u2009MWh\u22121 at combustion.\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n233\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nA shortage of either carbon-neutral biomass, VRE or hydrogen \nprimarily affects renewable liquid fuel production, which entails the \nmost conversion losses in the energy system and thereby is the marginal \nemissions abatement option11. Ramping up sufficient resources for \nliquid fuel production is therefore particularly challenging. Limiting \nfuel demand through, for instance, electrification and modal shifts \nwould make energy supply to achieve emissions targets easier, whereas \ndeveloping a portfolio of different fuel supply options appears as a \nsensible strategy to hedge against the considerable resource and tech-\nnology uncertainties. Consideration of higher CO2 sequestration levels \nwould allow for more fossil fuels combined with carbon capture, which \nwould increase the manoeuvring space and decrease the importance \nof biomass and of any renewable technology but would also rely on a \nstronger ramp-up of (BE)CCS (Fig. 6d).\nIn previous studies, biomass use has varied considerably but has \noften focused on bioelectricity and liquid fuels36, whereas power sys-\ntem studies on firm generation have indicated a value of biomass for \nhandling variability8. With a high spatio-temporal resolution and broad \ncross-sectoral coverage, our study shows that even though substantial \ndispatchable generation capacities are deployed, they are only rarely \nused, and thus only a small amount of biomass is allocated for providing \nflexible bioelectricity. However, although substantial amounts of bio-\nelectricity and CHP appear as low-priority options for biomass usage in \nthis study (similarly to Williams et al.78), many near-optimal solutions \nfor biomass usage were found to exist, with similar results for both \nrestrictive and higher carbon sequestration capacities. Thus, small \ndeviations in settings may lead to large differences in biomass usage, \nand care must be taken when deriving general conclusions.\nTotal biomass usage\n2,400\n2,200\n2,000\n1,800\n1,600\n1,400\n1,200\n1,000\n800\n600\n2,400\n2,200\n2,000\n1,800\n1,600\n1,400\n1,200\n1,000\n800\n600\n2,400\n1,000\n2,000\n1,500\n1,000\n500\n0\n20\n6\n4\n2\n0\n4\n2\n0\n1\n3\n4\n2\n1\n3\n4\n2\n0\n0\n0\n5.1\n10.3\n15.4\n20.6\n0\n5.1\n10.3\n15.4\n20.6\n0\n5.1\n10.3\n15.4\n20.6\n0\n5.1\n10.3\n15.4\n20.6\n0\n5.1\n10.3\n15.4\n20.6\n0\n5.1\n10.3\n15.4\n20.6\n0\n5.1\n10.3\n15.4\n20.6\n0\nImport biomass upstream emissions (g CO2 MJ\u22121)\nCO2 sequestration potential \n(Mt CO2 per year)\n5.1\n10.3\n15.4\n20.6\n0\n5.1\n10.3\n15.4\n20.6\n1\n3\n15\n10\n5\n0\n800\n600\n400\n200\n0\n1,000\n800\n600\n400\n200\n0\n1,000\n800\n600\n400\n200\n0\n2,200\n2,000\n1,800\n1,600\n1,400\n1,200\n1,000\n800\n600\nBiomass usage for biofuels\nSolid biomass import usage\nBECC\nDAC\nFossil and cement CC\nTotal CC and DAC\nSystem cost increase of excluding biomass\nFossil fuel usage\nPWh\nPWh\nPWh\n3.4\n3.4\n3.4\n3.4\n3.4\n3.4\n3.4\n3.5\n3.5\n3.5\n3.5\n3.5\n3.5\n3.5\n3.6\n3.7\n3.6\n3.6\n3.2\n2.4\n2.5\n2.1\n3\n3.6\n3.6\n3.8\n3.8\n2.8\n2.6\n2.6\n2.6\n2.6\n2.6\n2.8\n2.8\n2.8\n2.8\n2.8\n2.8\n2.8\n2.8\n2.8\n2.9\n2.5\n2.7\n2.7\n2.7\n2.7\n2.7\n2.7\n2.7\n2.7\n2.5\n2.5\n2.5\n2.5\n2.5\n2.5\n2.5\n2.2\n2.2\n2.2\n2.2\n2.2\n2.2\n2.3\n2.3\n2.3\n2.3\n2.3\n2.3\n2.3\n2.3\n2.1\n2.1\n2.1\n2.1\n2.0\n2.4\n2.2\n2.4\n2.4\n2.4\n2.4\n2.4\n2.4\n2.6\n2.6\n2.6\n2.6\n2.4\n3.8\n3.9\n3.9\n3.9\n3.9\n1.9\n1.9\n3.8\n617\n1,476\n1,490\n1,490\n1,490\n1,490\n1,490\n1,476\n1,476\n1,476\n1,476\n1,400\n1,400\n1,400\n1,400\n1,406\n1,600\n1,600\n1,600\n1,800\n2,000\n2,200\n2,200\n2,200\n2,206\n2,232\n2,232\n6.6\n7.2\n8.8\n6.3\n6.3\n7.3\n7.7\n8.6\n8.5\n9.6\n9.6\n7.6\n6\n5.4\n5.2\n4.6\n6.7\n4.6\n4.6\n4.1\n4.1\n4.1\n7.4\n6.4\n5.3\n4.2\n5.1\n5.2\n6.4\n6.3\n7.5\n7.6\n7.6\n7.4\n7.4\n4.1\n4.1\n5.2\n5.2\n11\n11\n11\n10\n10\n10\n13\n13\n13\n13\n13\n16\n16\n15\n15\n14\n15\n12\n12\n12\n12\n2,000\n2,000\n1,800\n1,800\n1,219\n1,230\n1,242\n1,239\n1,160\n1,182\n1,140\n1,094\n1,182\n1,188\n1,256\n1,129\n1,078\n1,173\n1,134\n1,145\n1,113\n1,080\n1,070\n444\n416\n403\n501\n669\n684\n451\n387\n316\n237\n237\n263\n259\n233\n233\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n1\n74\n367\n208\n150\n383\n583\n621\n640\n551\n551\n551\n551\n551\n149\n496\n496\n496\n496\n496\n145\n145\n145\n145\n145\n146\n151\n161\n181\n183\n182\n170\n164\n157\n202\n211\n231\n236\n238\n446\n393\n366\n526\n723\n654\n663\n442\n414\n339\n293\n392\n477\n636\n815\n742\n655\n0\n0\n0\n1.7\n925\n994\n994\n994\n994\n994\n1,007\n1,034\n1,029\n1,025\n1,061\n1,053\n1,039\n1,010\n994\n952\n959\n996\n948\n995\n926\n992\n910\n925\n925\n925\n18\n18\n18\n16\n17\n3.9\n3.9\n3.9\n3.9\n3.9\n3.7\n3.4\n3.1\n3.9\n2.9\n2.7\n1.8\n0.9\n3.9\n2.4\n2.2\n1.6\n1.3\n2.0\n1.1\n2.7\n1.8\n1.0\n0.1\n0.0\n0.7\n0.5\n0.3\n0.1\n0.0\n0.0\n0.0\n19\n20\n20\n925\n924\n950\n819\n954\n981\n930\n899\n843\n1.6\n1.6\n1.0\n894\n927\n927\n935\n935\n900\n0.7\n0.8\n0.6\n0.9\n0.9\n0.3\n894\n871\n1,036\n1,043\n1,047\n1,064\n1,064\n1,036\n869\n0.5\n1.7\n1.4\n1.5\n0.1\n0.1\n0.0\n0.0\n0.0\n0.0\n0.0\n0.2\n0.6\n0.2\n0.6\n0.8\n1.4\n1.4\n1.1\n1.0\n0.8\n0.6\n0.6\n0.7\n0.7\n0.6\n0.5\n1.3\n1\n0.72\n0.72\n0.72\n0.84\n0.86\n0.72\n1.6\n1.5\n2.3\n3.5\n3.5\n3.4\n%\nMt CO2\nMt CO2\nMt CO2\nMt CO2\nPWh\nc\nb\na\nf\ne\nd\ni\nh\ng\nFig. 6 | Heat maps of cost-optimal configurations. Heat maps show cost-optimal \nconfigurations for achieving a net-negative (\u2212110%) emissions target when \nvarying upstream emissions of biomass imports (x axis) and allowed annual \ncarbon sequestration potentials (y axis). Upstream emissions equivalent to 10% \nof the biomass CO2 emissions at full combustion corresponds to 10.3\u2009g\u2009CO2\u2009MJ\u22121. \nFor reference, estimated nitrous oxide emissions from fertilization correspond \nto 0.4\u201314\u2009g\u2009CO2\u2009equivalent\u2009MJ\u22121 for poplar or willow depending on yields and soil \nconditions, and, for example, substantially more for rape seed147. Fossil fuels in \nthis analysis are not assumed to entail upstream emissions, which presents an \noptimistic case for their performance. a,c, Total (a) and imported (c) biomass \nusage. b, Biomass usage for biofuels. d\u2013g, Least-cost deployment of BECC (d), \nDAC (e), fossil and cement carbon capture (CC) (f) and total carbon capture and \nDAC (g). h, The system cost increase of excluding biomass (except municipal \nsolid waste) compared to the least-cost solution for all combinations of carbon \nsequestration capacity and upstream emissions of biomass imports. i, Fossil  \nfuel usage.\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n234\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nGiven the many near-optimal solutions for biomass usage, local \ndifferences in terms of availability of biomass and non-fossil electricity \nand transmission bottlenecks and carbon infrastructure, the possibility \nto utilize excess heat in district heating and regional system adequacy \nconsiderations lead to local variations in cost-competitiveness of dif-\nferent options (as evidenced already by the diversity of solutions on the \ncountry-level resolution, as seen in Extended Data Fig. 6). This probably \nresults in a diversified portfolio of locally optimal biomass usages and \nspatially resolved trade-offs between using captured carbon for fuel \nproduction or for sequestration.\nNevertheless, as also found in previous studies with lower opera-\ntional resolution37,39,41, combining biomass usage with carbon capture \nwas found to be a robust strategy. Whereas these earlier studies only \ninclude CCS, this study also includes CCU and finds both very valu-\nable to enable high carbon efficiencies in a renewable energy system. \nAllowing more fossil fuel usage compensated through CDR did not \nsubstantially affect near-optimal biomass usage (in contrast to Grant \net al.48), but it was found to reduce the competitiveness of using cap-\ntured carbon to produce electrofuels (CCU), in favour of sequestering \nit (CCS). In fact, failing to apply carbon capture resulted in a consider-\nably reduced value of using biomass in the energy system. Owing to its \nhigh investment cost, carbon capture was found to be cost effective in \nprocesses running with high utilization rates and not in applications \nmanaging the integration of variable renewable electricity.\nA net-negative target for the European energy system is probably \nneeded to reach territorial net-zero emissions, considering that residual \nemissions in other sectors need to be compensated; this might exert a \nsubstantial demand pull for biomass, especially if VRE and electrolyser \ndeployment falls behind expectations. The resulting level of biomass \nusage may even exceed the lower end of estimated global biomass resi-\ndue potentials, which spans a wide range of 3\u201321\u2009PWh per year (ref. 110).\nEstimated production costs for primary non-residue biomass (for \nexample, Millinger et al.111) fall within competitive cost ranges in this \nstudy, especially if biomass residues are limited. Thus, high biomass \ndemand and prices could provide an incentive for the forest and agri-\nculture sectors to produce more primary non-residue biomass for the \nenergy system. The land carbon consequences in such a scenario are \nuncertain; studies find that biomass demand can induce changes in \nland use affecting land carbon stocks positively or negatively, depend-\ning on climate and soil conditions, historic land use, character of bio-\nmass production system being established, governance and other \ngeographically varying factors104,112\u2013116.\nThe technical BECCS potential associated with domestic biomass \nresidues in Europe (excluding forest residues) has been estimated at \n200\u2009Mt\u2009CO2 (ref. 117), which would not suffice to achieve net-negative \nemissions targets. For the new EU Renewable Energy Directive III, ener-\ngetic usage of primary forest residues was proposed to be excluded \nas an option for meeting renewable targets118, which alone have been \nestimated to amount to up to 1.6\u2009PWh per year in Europe83, or up to \n600\u2009Mt biogenic CO2 that could potentially be captured (Extended Data \nTable 2). Excluding comparably easy-to-monitor domestic resources \nmight lead to a substantially higher cost of the energy system and to \na higher demand for imported biomass and dedicated crops, with \nharder-to-foresee environmental consequences. As has been shown \nhere, biomass usage, combined with carbon capture, is cost effective as \nlong as net upstream emissions are relatively small or if negative emis-\nsions counteract limited upstream emissions. Exclusions of biomass \nsources such as primary forest residues thus need to be weighed against \nthe targets in the energy sector and the potential to achieve negative \nemissions and gauged towards achieved capacity expansion speeds \nfor VRE and electrolysis, which require an unprecedented ramp-up to \nachieve the results presented here already if biomass is not restricted.\n2,000\n3,000\n4,000\n5,000\n7,000\nDAC CAPEX (\u20ac kg\u22121 CO2 h\u22121)\n36\n54\n72\n90\n36\n54\n72\n90\n36\n54\n72\n90\n36\n54\n72\n90\n36\n54\n72\n90\n36\n54\n72\n90\nImport biomass price (\u20ac MWh\u22121)\n5.2\n4.2\n21\n13\n13\n14\n15\n19\n10\n14\n11\n7.9\n7.6\n7.8\n8.7\n8.4\n9.1\n16\n18\n20\n25\n23\n25\n4.2\n4.2\n4.2\n4.2\n4.2\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n2\n176\n369\n752\n751\n702\n610\n192\n0\n0\n0\n3.6\n1.4\n2.2\n0.0\n0.0\n0.0\n0.3\n1.5\n27\n30\n35\n1.1\n1.7\n2.0\n2.2\n2.2\n2.2\n2.2\n2.2\n2.2\n2.2\n2.2\n2.2\n2.5\n1.4\n1.4\n1.6\n827\n827\n827\n824\n887\n967\n967\n967\n967\n967\n967\n963\n999\n1,056\n1,029\n1,156\n1,185\n1,184\n1,185\n963\n963\n963\n976\n887\n887\n887\n887\n887\n825\n825\n825\n839\n827\n827\n827\n3.6\n3.6\n3.6\n3.6\n750\n295\n550\n683\n295\n345\n409\n699\n3.6\n750\n750\n750\n750\n750\n3.6\n3\n3.6\n3.6\n2.8\n3.4\n3.6\n5.2\n5.2\n5.2\n5.2\n5.2\n2,000\n3,000\n4,000\n5,000\n7,000\n1,000\n750\n500\n250\n0\n4\n2\n0\n4\n20\n30\n10\n0\n2\n0\n1,000\n1,000\n750\n500\n500\n250\n0\n0\n1,500\n1,500\nTotal biomass usage\nSolid biomass import usage\nc\nb\na\nSystem cost increase of \nexcluding biomass\nPWh\nPWh\n%\nMt CO2\nMt CO2\nMt CO2\nBECC\nDAC\ne\nf\nd\nTotal CC and DAC\nFig. 7 | Heat maps of cost-optimal configurations. Heat maps of cost-optimal \nconfigurations for achieving a net-negative (\u2212110%) emissions target when \nvarying the price of biomass imports (x axis) and DAC capital expenditure  \n(y axis). For comparison between different used metrics in literature: DAC capital \ncosts (CAPEX) of \u20ac1,500\u20137,000\u2009kg\u22121\u2009CO2\u2009h\u22121 correspond to \u20ac171\u2013800\u2009t\u22121\u2009CO2 \nper year at a full utilization rate or a capital cost of \u20ac16\u201375\u2009t\u22121\u2009CO2 (not including \noperational expenditure) with a 7% discount rate and a 20-year lifetime.  \nThe assumed capital cost of the carbon capture unit for BECC is here assumed \nat a constant \u20ac2,400\u2009kg\u22121\u2009CO2\u2009h\u22121, but it is likely that cost progressions for DAC \nwould spill over also to BECC. a, Total biomass usage. b, Biomass imports. c, The \nsystem cost increase of excluding biomass compared to the least-cost solution \nfor all combinations of carbon sequestration capacity and upstream emissions of \nbiomass imports. d\u2013f, BECC (d), DAC (e) and total CC and DAC (f).\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n235\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nMethods\nPyPSA-Eur-Sec model\nPyPSA-Eur-Sec60,119 is an open-source, sector-coupled full European \nenergy system optimization (linear programming) model including \nthe power sector, transport (including also international shipping and \naviation), space and water heating, industry and industrial feedstocks. \nThe model minimizes total system costs by co-optimizing capacity \nexpansion and operation of all energy generation and conversion and \nof storage and transmission of electricity, hydrogen and gas. The model \nis based on the Python software toolbox PyPSA (Python for Power \nSystems Analysis)120. A comprehensive description of the model can be \nfound in Neumann et al.121. A version with an extended set of biomass \nresource technology portfolio is used11, with added details for car-\nbon capture energy penalties and additional competition introduced \nfor industry heat supply and added domestic biomass pellet boilers, \nhydrogen CHPs, waste incineration and electrobiofuels (biofuels with \nextra hydrogen added to the process to utilize more of the biomass \ncarbon directly).\nA 37-node spatial resolution and a 5-h temporal resolution over a \nfull year in overnight greenfield scenarios was used, based on a trade-off \nbetween the difference in results compared to with a 1-h resolution \n(Supplementary Fig. 1) and computing time (Supplementary Fig. 2). A \ncountry-level spatial resolution was chosen for computational reasons. \nA lossy transport model for electricity transmission was used, which is \nsuitable at this resolution122, and transmission is constrained to expand \nto at most double the total line volume in 2022.\nFinal energy demands for the different sectors are calculated \nbased on the JRC IDEES database123 with additions for non-EU countries \n(refs. 10,60 provide further elaboration) and need to be met (that is, \ndemand is perfectly inelastic). However, energy carrier production \nincluding electricity, hydrogen, methane and liquid fuels is determined \nendogenously. Fossil fuels (coal, natural gas and oil) and uranium are \nincluded, as are solid biomass imports as outlined below. Technol-\nogy costs and efficiencies are elaborated on in the Supplementary \nInformation, with technology values for 2040 (given in \u20ac2015) used \nfrom the PyPSA energy system technology data set v0.6.0 (ref. 124). \nThe discount rate is uniform across countries and set to 7%, except \nfor rooftop solar PV and decentral space/water heating technologies, \nfor which it is set to 4%.\nMathematical formulation. The objective is to minimize the total \nannual energy system costs of the energy system that comprises both \ninvestment costs and operational expenditures of generation, stor-\nage, transmission and conversion infrastructure. To express both as \nannual costs, we use the annuity factor (1\u2009\u2212\u2009(1\u2009+\u2009\u03c4)\u2212n)\u2009/\u2009\u03c4 that converts \nthe upfront overnight investment of an asset to annual payments con-\nsidering its lifetime n and cost of capital \u03c4. Thus, the objective includes \non one hand the annualized capital costs c* for investments at bus i \nin generator capacity Gi,r \u2208 R+ of technology r, storage energy capac-\nity Ei,s \u2208 R+ of technology s, electricity transmission line capacities  \nP\u2113 \u2208 R+ and energy conversion and transport capacities Fk \u2208 R+ (links) \nand the variable operating costs o* for generator dispatch gi,r,t \u2208 R+ and \nlink dispatch fk,t \u2208 R+ on the other:\nmin\nG,E,P,F,g [\u2211\ni,r\nci,r \u00d7 Gi,r + \u2211\ni,s\nci,s \u00d7 Ei,s + \u2211\n\u2113\nc\u2113\u00d7 P\u2113+ \u2211\nk\nck \u00d7 Fk\n+ \u2211\nt\nwt \u00d7 (\u2211\ni,r\noi,r \u00d7 gi,r,t + \u2211\nk\nok \u00d7 fk,t)]\n(1)\nThereby, the representative time snapshots t are weighted by \nthe time span wt such that their total duration adds up to one year; \n\u2211t\u2208Twt\u2009=\u2009365\u2009\u00d7\u200924\u2009h\u2009=\u20098,760\u2009h. A bus i represents both a regional scope \nand an energy carrier. In addition to the cost-minimizing objective \nfunction, as exhaustively described in ref. 125, we further impose a set \nof linear constraints that define limits on the capacities of generation, \nstorage, conversion and transmission infrastructure from geographi-\ncal and technical potentials and the availability of variable renewable \nenergy sources for each location and point in time. Further, the limit \nfor CO2 emissions or transmission expansion is defined, along with \nstorage consistency equations, and a multi-period linearized optimal \npower flow formulation. Overall, this results in a large linear problem.\nThe modelled system represents a long-term equilibrium where \nthe zero-profit rule applies and the revenue that each generator \nreceives from the market exactly covers their costs126,127. By way of annu-\nalization of capital costs (assuming that the modelled year represents \nan average revenue year for each asset over their economical lifetime) \nand weighting of asset operation to the interannual temporal resolu-\ntion, there is therefore full cost recovery of all assets built. Prices form \nendogenously in the model based on renewable supply conditions, \nstorage and demand flexibility. Regional electricity price time series \nare retrieved from the dual value of the energy balance equations for \neach region and hour. CO2 emissions are considered in the model, \nwhereas other GHG emissions are not, and a CO2 price is calculated as \nthe shadow price of the least-cost objective function for achieving the \nset emissions target.\nNear-optimal analysis\nThe Modelling to Generate Alternatives (MGA) method70,72\u201374,128 was \nimplemented for the sector-coupled model. With this method, first \na cost-optimal result for achieving emissions targets is calculated, \nwhich gives a minimum system cost C. For notational brevity, let cTx \ndenote the linear objective function equation (1) and Ax\u2009\u2264\u2009b the set of \nadditional linear constraints in a space of continuous variables, such \nthat the minimized system cost can be represented by\nC = min\nx\n{cTx|Ax \u2264b} .\n(2)\nIn the next step, this cost is increased by \u03b5 \u2208 {0.01, 0.02, .\u2009.\u2009.} and \nset as a constraint, while minimizing or maximizing a set of variables, \nsuch as energy carriers or technologies, for example, total biomass \nusage, total biofuel production or total wind power generation, with\nxs = min\nxs\n{1Txs|Ax \u2264b, cTx \u2264(i + \u03b5) \u00d7 C}\n(3)\nxs = max\nxs\n{1Txs|Ax \u2264b, cTx \u2264(i + \u03b5) \u00d7 C}\n(4)\nBy exploring the extremes, the Pareto frontiers for a given param-\neter\u2013cost combination are mapped out. The system cost of excluding \na particular technology or resource was validated in runs where the \noption in question was excluded. To obtain shadow prices related to \nsystem cost for Fig. 5, biomass use was set as an additional constraint \nin cost-minimizing runs.\nThe model runs were performed on the Chalmers Centre for Com-\nputational Science and Engineering (C3SE) computing cluster.\nBiomass and bioenergy\nA variety of biomass categories and conversion technologies are intro-\nduced in the model. Different biomass residue types are clustered into \nthe categories solid biomass and digestible biomass (Extended Data \nFig. 4). Solid biomass can be used for a variety of applications in heat, \npower and fuel production and can optionally be combined with carbon \ncapture (Extended Data Fig. 1). Digestible biomass can be used for biogas \nproduction via anaerobic digestion, which is upgraded to pure biometh-\nane with the option to capture the waste CO2 stream. Methane can also be \nproduced via gasification of solid biomass (BioSNG), or supplied by fossil \nmethane. These routes result in the same end product, methane (CH4).\nMedium domestic (country-level) biomass residue potentials for \n2050 from the JRC ENSPRESO database were used83 nodally explicitly, \nwith a weighted average of country-level biomass costs including \n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n236\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nharvesting, collection and transport from the reference biomass sce-\nnario129. Additionally, more expensive solid biomass imports can be \nused11 (Extended Data Fig. 3 and Extended Data Table 2). Pre-treatment \nof biomass, where needed, is considered implicitly through moderate \nconversion efficiencies and spare waste heat that could be utilized. \nSmall-scale heating includes a pelletization cost of \u20ac9\u2009MWh\u22121 biomass. \nReflecting current EU policy direction, bioenergy feedstock is assumed \nto consist of residues, for which the likelihood of direct and indirect \nland-use change and land carbon changes is smaller than for dedicated \nfeedstock production. The use of residues and waste as bioenergy \nfeedstock is assumed not to influence the land carbon stock; that is, \nthe global net flow of CO2 between the atmosphere and the biosphere, \nwhich is driven by photosynthesis, respiration, decay and combustion \nof organic matter, is assumed to not be affected. In a renewable energy \nsystem, processing, conversion and transport do not cause fossil car-\nbon emissions, and residual GHG emissions associated with land use \ncan be considered to be offset through CDR if exporting countries \nadhere to net-zero or net-negative targets.\nFor biomass imports, as described in ref. 11, we assume that 175\u2009EJ \nper year of biomass can be supplied globally at a price of US$15\u2009GJ\u22121, \nusing the average of an IAM comparison54. Using regional data on \nbiomass use per capita and population estimates130, we assume that up \nto 20\u2009EJ biomass (subtracting domestic potentials) may be imported \nto Europe at a price of \u20ac15\u2009GJ\u22121 (\u20ac54\u2009MWh\u22121). For each additional EJ to \nbe imported, the price is assumed to increase by \u20ac0.25\u2009GJ\u22121, based \non the slope of the low-cost scenarios (Extended Data Fig. 3). The \nassumed import prices are substantially higher than 2020 wood chip \nprices at \u20ac20\u201325\u2009MWh\u22121 (ref. 89) (Extended Data Fig. 3). This reflects an \nincreased demand for biomass in scenarios complying with stringent \nGHG emission targets. We test the effect of this assumption on results \nin a sensitivity analysis (Fig. 7). Direct biofuel imports (and hydrogen \nderivatives) from outside Europe are not considered.\nCarbon balances of biomass\nSolid biomass carbon dioxide uptake from atmosphere, with a carbon \nshare %Csb\u2009=\u200950%, specific energy esb\u2009=\u200918\u2009GJ\u2009t\u22121 and molality \nmCO2/mC\u2009=\u200944/12 (equation (5)):\n\u03b5 sb\nat = \u2212%Csb \u00d7 3.6\nesb\n\u00d7\nmCO2\nmC\n(5)\nLiquid fuel carbon dioxide emissions (t\u2009CO2\u2009MWh\u22121) at full combus-\ntion for diesel and methane based on -CH2- simplification and specific \nenergy eCH2\u2009=\u200944\u2009GJ\u2009t\u22121 LHV for diesel and eCH4\u2009=\u200950\u2009GJ\u2009t\u22121 LHV for methane \n(equation (6)):\n\u03b5fu = 3.6\neCHx\n\u00d7\nmCO2\nmCHx\n(6)\nThe carbon efficiency \u03b7c of the conversion is estimated by  \nequation (7).\n\u03b7c = \u03b7 \u00d7 \u03b5fu\n\u03b5sb\n(7)\nThe rest is assumed to end up as CO2, of which a part \u03b5s is separated \nand possibly captured with an efficiency \u03b7\u03b5, with the remainder \u03b5v being \nvented as CO2 to the atmosphere in the exhaust gas.\nThe biogas produced from digestible biomass is assumed to contain \n60 vol% CH4 (specific energy e\u2009=\u200950\u2009GJ\u2009t\u22121, density \u03c1\u2009=\u20090.657\u2009kg\u2009mn\n\u22123) and \n40 vol% CO2 (\u03c1\u2009=\u20091.98\u2009kg\u2009mn\n\u22123), which calculates to 0.0868\u2009t\u2009CO2\u2009MWh\n\u22121\nCH4. \nThe feedstock input potentials and costs for biogas are given for \nMWhMWhCH4 and thus MWhin\u2009=\u2009MWhout for the carbon balance calcula-\ntions. Thereby the C content in the slush can be omitted, thus avoiding \nsystem boundary issues with the agricultural sector.\nCarbon capture\nProcess emissions are assumed to be captured post-combustion \nthrough solvents, which is the standard method with highest tech-\nnological readiness level in 2021131. For carbon capture in biomass \napplications, part of the biomass input is used to meet the additional \nheat demand for regenerating the solvents used for CO2 capture. Heat is \nassumed to be met by a steam boiler of the type suitable for the process \n(gas for biogas, otherwise solid biomass), with capital and operational \ncosts added accordingly. Capture rates of 95% are assumed for these \nprocesses except for biogas, where 90% is assumed. For biofuel produc-\ntion, acid gas removal (including CO2, amounting to 71% of the carbon \nin the biomass feedstock; Supplementary Table 1) from the syngas \nis assumed to be performed with the Rectisol132 (methanol-based) \nprocess, with a 90% capture rate133\u2013135 and electricity demand to cover \nfor this is assumed to be included in the base process. The effect of \ncapture rates (which can be both higher and lower than assumed here) \nis assessed in a sensitivity analysis (Extended Data Fig. 5). As a result of \nenergy penalties of the BECC processes, the efficiency of the conversion \nto the main product is decreased, and the capital cost is increased to \ncover for the additional heat demand and carbon capture infrastruc-\nture, as summarized in Extended Data Table 1. In contrast, heat demand \nfor DAC can be met by several competing process steam options as \ndescribed below.\nA scaling factor \u03b1 for the additional biomass needed to supply  \nthe steam heat demand of carbon capture is calculated as a function \n(equation (8)) of the amount of CO2 in the output stream \u03b5s (Supple-\nmentary Table 2), the required heat input for carbon capture eth,cc \n(here assumed as 0.66\u2009MWh\u2009t\u22121\u2009CO2 at 100\u2009\u00b0C (ref. 136)) and the boiler \nefficiency \u03b7th (Supplementary Table 2).\n\u03b1 =\n1\n1 + \u03b5s \u00d7 eth,cc/\u03b7th\n(8)\nThe steam efficiency \u03b7steam (equation (9)) and main product effi-\nciency \u03b7new (equation (10)) are derived as a function of the scaling \nfactor \u03b1.\n\u03b7steam = (1 \u2212\u03b1) \u00d7 \u03b7th\n(9)\n\u03b7new = \u03b1 \u00d7 \u03b7old\n(10)\nThe new investment cost CI,new is derived as a function (equation (11))  \nof the investment cost of the base plant configuration CI,old scaled \nby the updated efficiency \u03b7old\u2009/\u2009\u03b7new, with added investment costs for \nthe steam boiler CI,th scaled by the steam produced \u03b7steam\u2009/\u2009\u03b7new and \nadditional investment costs for the carbon capture unit CI,cc (here \nassumed as \u20ac2,400\u2009kg\u22121\u2009CO2\u2009h\u22121 (ref. 136)) scaled by the process CO2 \noutput stream \u03b5s.\nCI,new = CI,old \u00d7 \u03b7old/\u03b7new + CI,th \u00d7 \u03b7steam/\u03b7new + CI,cc \u00d7 \u03b5s\n(11)\nFor CHP units, the heat demand is scaled as the main product, \nwith added district heat output from the carbon capture process (here \nassumed as 0.79\u2009MWhth\u2009t\u22121\u2009CO2 at district heat temperature136).\nThese calculations result in the costs and efficiencies for the pro-\ncesses with carbon capture shown in Extended Data Table 1.\nSector-specific assumptions\nSteel production is assumed to be fully performed with hydrogen as a \nreduction agent (direct reduced iron, DRI) and electric arc furnaces, \nand the share of scrap steel increases from 40% in 2023 to 70% in the \ntarget year. Cement production entails unavoidable emissions from \ncalcination, which can be captured.\nDistrict heating is assumed to cover 30% of urban demand, whereas \nspace heating demand is assumed to decrease by 29% due to efficiency \n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n237\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\ngains through new buildings and building renovation. District heat can \nbe fulfilled by numerous options, including excess heat from electro-\nlysers and fuel production (of which 50% is assumed possible to use \nconsidering that processes are not always near district heat grids) and \nmandatory incineration of municipal solid waste.\nIndustrial heat is divided into three segments: low (process steam, \n<200\u2009\u00b0C), medium (~200\u2013500\u2009\u00b0C) and high temperature (>500\u2009\u00b0C). In \nthe low- and medium-temperature segments, biomass is an option, \nwhereas methane and hydrogen are an option in all three. Direct electri-\nfication is an option in the low-temperature (process steam) segment, \nwhereas heat pumps for process steam are not considered. Thus, solid \nbiomass competes for producing industrial process steam with electric, \nhydrogen and methane boilers and for producing medium-temperature \nprocess heat with hydrogen and methane.\nBase electricity demand for households and industry is the same \nas in 2011 (except for a subtraction of electricity used for heating, the \nsupply of which is endogenously determined), with a temporal varia-\ntion as depicted in Fig. 3.\nIt is assumed that a strong increase in the use of electric vehicles \nreduces liquid fuel demand in land transport to zero, hence reducing \nthe need for biomass and/or electricity for meeting renewable fuel \ntargets (land transport demand overall is assumed to increase by 20%). \nA liquid fuel demand is however retained in aviation (total fuel demand \nincreases by 70% compared to 2011), shipping (+50% compared to 2011, \nwith half of the fuel demand supplied by hydrogen) and in the chemical \nindustry (same demand as 2011), which can be supplied through solid \nbiomass-based liquid fuels (biofuels), electrofuels, electrobiofuels and \nfossil fuels. Transport and chemical demand is assumed as for the year \n2060 in Millinger et al.11, and recycling of plastics is not considered.\nFor a sensitivity analysis with less sector coupling, the following \noptions were turned off completely: battery electric vehicle (BEV) \ncharging demand-side management; vehicle to grid; thermal energy \nstorage; waste heat usage from biofuels, electrolysis, DAC and BioSNG; \nH2 networks; H2 underground storage in salt caverns. Further, no expan-\nsion of the electricity transmission or district heat grids from 2022 \nlevels was allowed.\nCost estimations\nCosts of energy provision to different applications are estimated as \nfollows and used for comparing cost increases to sectoral costs. The \ncost of, for instance, fuels is estimated by allocating the cost of feed-\nstocks used by the share of total feedstocks used for fuel production \nand adding investments and operational costs. For processes with \nseveral outputs, the Carnot method is used for allocating costs to \ndifferent products34.\nAssuming 60\u2009\u00b0C for heat output and 20\u2009\u00b0C for the sink, a Carnot \nfactor for heat is derived by equation (12)34:\n\u03b7c = 1 \u2212TL\nTH\n= 1 \u2212273 + 20\n273 + 60 \u22480.12\n(12)\nThe allocation factor ath is derived by considering conversion \nefficiencies for all products (equation (13)): heat \u03b7th, electricity \u03b7el, fuels \n\u03b7fu and multiplying them with their Carnot factors. Carnot factors for \nelectricity, hydrogen and fuels are set to one.\nath =\n\u03b7c\u03b7th\n\u03b7el + \u03b7fu + \u03b7c\u03b7th\n(13)\nThe share of H2 used for electrofuel production is calculated by \ndividing the electrofuel H2 demand \u03b4 H2\ni, j,t for all electrofuel technologies \nj \u2208 Fe, by the total H2 production \u03c0 H2\ni,t . The cost of H2 (including electricity, \nelectrolyser capital costs and H2 pipeline costs C j\u2208H2) is assigned to \nelectrofuels by the share of H2 used for electrofuel production. The \ntotal cost is calculated as in equation (14).\nC Fe\ntot = \u2211\ni\nC Fe\ni\n+\n\u2211i, j\u2208Fe,t \u03b4 H2\ni, j,t\n\u2211i,t \u03c0 H2\ni,t\n(\u2211\nj,i\nC el\nj,i\n\u2211i,t \u03b4 el\ni,k,tp el\ni,t\n\u2211i,t \u03c0 el\ni, j,tp el\ni,t\n+ \u2211\nj\nC j\u2208H2)\n(14)\nThe cost of solid biomass bs used to produce biofuels is assigned \nto the biofuels by the amount of solid biomass used for biofuels \u03b4 bs\nj\u2208Fb,i,t, \nwith j \u2208 Fb divided by the total amount of solid biomass used \u03c0 bs\ni,t. This \nis added to the capital cost of biomass to liquid C Fb\ni  (equation (15)). The \ncost is allocated to fuels and heat by equation (13).\nCbiofuel = (1 \u2212ath) (\u2211\ni\nC Fb\ni\n+\n\u2211i, j\u2208Fb,t \u03b4 bs\ni, j,t\n\u2211i,t \u03c0 bs\ni,t\n\u2211\ni,t\nC bs\ni,t )\n(15)\nCosts for industry heat are straightforward, as there is only one \nproduct, and they are calculated as above.\nTechnology growth rates\nHistorical technology growth rates are used to ex-post assess feasibility \nof future growth expectations and model results137, but not to restrict \nmodel results. Technology growth typically follows an S curve and \nis often estimated by a Gompertz curve. The maximum growth rate \nGgmp at the inflection point of the S curve serves as an indicator for \ncomparison to historically observed growth rates and is calculated \nby equation (16)90.\nGgmp = Lk\ne\n(16)\nwhere L is the asymptote (set to the obtained cost-optimal result \nfor individual technologies), k is the growth constant and e is Euler\u2019s \nnumber.\n\u0394t is the time (in years) it takes to grow from 10 to 90% of the \nasymptote and can be estimated by equation (17)90.\n\u0394tgmp =\nln (\nln(0.1)\nln(0.9) )\nk\n(17)\nSetting the growth constant to k\u2009=\u20090.09 gives \u0394tgmp\u2009=\u200934 years \n(equation (17)), that is, if starting at 10% in 2023, 90% of the asymp-\ntote is achieved in year 2057. This setting is used for estimation here.\nGgmp is normalized to the electricity demand at the inflection \npoint, located at 37% of the asymptote L. Demand in the base year \n\u03b40\u2009=\u20093,448\u2009TWh (ref. 84) and demand \u03b4T in the target year is set to the \nresulting electricity generation of the respective scenario.\nLimitations\nA limitation compared to IAM studies is the lack of an explicit repre-\nsentation of agriculture and forestry, and the lack of a global trade \nmodel (including economy dynamics and equilibrium modelling). An \nexpansion of the system boundaries to encompass the land-use system \nwould more accurately capture emissions flows related to biomass and \ncapture the competition between BECCUS and land-based CDR meas-\nures such as Afforestation/reforestation23. Afforestation/reforestation \nhas a substantial CDR potential24, but there are uncertainties regarding \npermanency and additionality compared to geological sequestration23. \nCombined with other CDR measures such as enhanced weathering, \nbiochar and direct ocean capture24, the necessity of achieving nega-\ntive emissions in the energy sector and thus the role of biomass may \nbe reduced.\nA chemical demand is included, but demand may deviate from the \nassumed levels, and other uses such as construction and biochar may \nalso compete for biomass residues (which stem from forestry, from \nwhich the main product is mainly used for construction), which is not \nconsidered in this work.\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n238\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nA spatially explicit representation of geological carbon seques-\ntration locations and of infrastructure for CO2 transport was not con-\nsidered and should be pursued in further work. Similarly, whereas a \nconstant cost for biomass transport to where it is used is contained in \nthe biomass residue cost, the specific transport distance required is \nnot assessed. Whereas the overall results are not notably affected by \nassumptions on biomass residue and carbon handling costs, the prac-\ntical spatial implementation probably is. Further work should assess \nspatial aspects of carbon and biomass management138,139 and connected \nlogistics challenges140. For computational reasons, a spatially more \nexplicit assessment could not be performed with the diverse biomass \ntechnology portfolio and competitions across all sectors we consider \nin this study, and for the same reason, a weighted average of domestic \nbiomass residue costs was used.\nLeakage of methane and hydrogen is not considered and presents \na risk when relying on gaseous fuels141. However, digestible biomass \ncombustion may avoid methane emissions, which would otherwise \noccur, which is valuable in its own right but not considered within the \nsystem boundaries given here.\nIn runs with higher carbon sequestration, more fossil fuels are \nused, which affects emissions of logistics of, for instance, biomass, \nwhich is not considered as it would render the optimization prob-\nlem non-convex and substantially more computationally intensive. \nUpstream emissions of fossil fuels are also not considered, which would \naffect results at higher carbon sequestration capacities.\nLand transport was exogenously assumed to be fully electrified, \nwhereas electric or hydrogen-fuelled aviation was not considered as a \nconservative assumption based on expected long lead times delaying \na substantial market penetration. Assumptions on fuel demand levels \nin these sectors influence results by affecting primary energy demand \nand costs more than any other part of the energy system11.\nFor the weather data we use historical data from 2013, which is \nregarded as a characteristic year for wind and solar resources121. Interan-\nnual weather variability has an impact on variation management and \nfirm generation but is not assessed here. Whereas firm dispatchable \ncapacities can cover for almost the whole inflexible demand already \nin the results given here, both more and less biomass would be used \nfor that purpose in more extreme years, and the assessed case may be \nseen to represent an average biomass usage case. The model was run \nwith perfect foresight of weather conditions and demands, whereas \nin reality, there is substantial uncertainty in capacity planning and \nadequacy requirements. This is especially important for combined \ncapacity and dispatch optimization as performed here and affects \nresults on firm generation requirements. However, substantial firm \ncapacity is deployed in the results but very seldom run. Given the low \nbiomass amounts, even a doubling would not affect results substan-\ntially. Also, whereas flexibility is assumed in electrolysers, BEVs and so \non, there is a part of the electricity demand that is assumed not to be \nflexible, where some flexibility (demand elasticity) could be assumed \nand which would reduce the need for firm generation. Future work \nshould assess these aspects and demand variations (flexibility and \nabsolute amounts) further.\nA greenfield assessment was performed, which does not take exist-\ning infrastructure into account, aside from existing power transmission \nlines and hydropower installations. Emphasis in this work is to assess \nthe diversity of system compositions of an energy system adhering to \nstringent emissions targets, not on the transition leading there. The \ntiming of when net-zero or net-negative targets in the energy system are \nachieved is uncertain, and the results here are not tied to a specific year. \nIf targets are achieved by 2050 or later, most current capacity would \nhave reached the end of their lifespan. In the event that targets would \nbe achieved as early as 2040, some energy infrastructure existing in \n2024 is likely to remain, but these capacities amount to a small fraction \ncompared to the capacity expansion seen in the results. Nuclear capac-\nity that began operation after 1990, or is currently under construction, \namounts to 29\u2009GW (ref. 142), able to produce 2% of the total electricity \ngeneration in the least-cost net-negative scenario. For dispatchable gas \ngenerators, 242\u2009GWel capacity was in place in the region in 2022143, or \n46% of the gas power capacity obtained in the least-cost net-negative \nscenario. Thus, accounting for any remaining gas power capacity in \nthe target year would not affect results. Therefore, we deem this to be \na mild limitation to the modelling.\nWe acknowledge the limitation of this study in focusing solely on \nthe solution space concerning the composition of a system adhering \nto very stringent emissions targets, rather than analysing the transition \ntowards achieving these targets over time. Conducting a detailed analy-\nsis of the energy transition over the years would require a reduction in \nspatial and temporal resolution to ensure computational feasibility, \nleading to the loss of detail on other valuable aspects of the analysis. \nNevertheless, we recognize the importance of further research to \nanalyse the transition dynamics in achieving emissions targets over \ntime, considering both short-term variability and long-term capacity \nplanning.\nData availability\nThe technology data can be accessed via GitHub at github.com/mill-\ningermarkus/technology-data/tree/biopower and are archived via \nZenodo at https://doi.org/10.5281/zenodo.8099703 (ref. 144). Result-\ning files and code to generate figures are archived via Zenodo at https://\ndoi.org/10.5281/zenodo.14169801 (ref. 145).\nCode availability\nThe model code can be accessed via GitHub at github.com/millinger-\nmarkus/pypsa-eur-sec/tree/mga and is archived via Zenodo at https://\ndoi.org/10.5281/zenodo.8099690 (ref. 119).\nReferences\n1.\t\nHansen, K., Mathiesen, B. V. & Skov, I. R. Full energy system \ntransition towards 100% renewable energy in Germany in 2050. \nRenewable Sustain. Energy Rev. 102, 1\u201313 (2019).\n2.\t\nPatrizio, P., Fajardy, M., Bui, M. & Dowell, N. M. CO2 mitigation \nor removal: the optimal uses of biomass in energy system \ndecarbonization. iScience 24, 102765 (2021).\n3.\t\nLauer, M. et al. The crucial role of bioenergy in a climate-neutral \nenergy system in Germany. Chem. Eng. Technol. 46, 501\u2013510 \n(2023).\n4.\t\nJohansson, V., Lehtveer, M. & G\u00f6ransson, L. Biomass in the \nelectricity system: a complement to variable renewables or a \nsource of negative emissions? Energy 168, 532\u2013541 (2019).\n5.\t\nRuhnau, O. & Qvist, S. Storage requirements in a 100% renewable \nelectricity system: extreme events and inter-annual variability. \nEnviron. Res. Lett. https://doi.org/10.1088/1748-9326/ac4dc8 \n(2022).\n6.\t\nLehtveer, M. & Fridahl, M. Managing variable renewables with \nbiomass in the European electricity system: emission targets and \ninvestment preferences. Energy 213, 118786 (2020).\n7.\t\nSchipfer, F. et al. Status of and expectations for flexible bioenergy \nto support resource efficiency and to accelerate the energy \ntransition. Renewable Sustain. Energy Rev. https://doi.org/ \n10.1016/j.rser.2022.112094 (2022).\n8.\t\nSepulveda, N. A., Jenkins, J. D., de Sisternes, F. J. & Lester, R. K.  \nThe role of firm low-carbon electricity resources in deep \ndecarbonization of power generation. Joule 2, 2403\u20132420 (2018).\n9.\t\nSepulveda, N. A., Jenkins, J. D., Edington, A., Mallapragada, D. S. & \nLester, R. K. The design space for long-duration energy storage in \ndecarbonized power systems. Nat. Energy https://doi.org/10.1038/\ns41560-021-00796-8 (2021).\n10.\t Zeyen, E., Hagenmeyer, V. & Brown, T. Mitigating heat demand \npeaks in buildings in a highly renewable European energy system. \nEnergy 231, 120784 (2021).\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n239\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\n11.\t\nMillinger, M. et al. Are biofuel mandates cost-effective? An \nanalysis of transport fuels and biomass usage to achieve \nemissions targets in the European energy system. Appl. Energy \n326, 120016 (2022).\n12.\t Korberg, A. D., Mathiesen, B. V., Clausen, L. R. & Skov, I. R. The \nrole of biomass gasification in low-carbon energy and transport \nsystems. Smart Energy 1, 100006 (2021).\n13.\t Millinger, M., Meisel, K. & Thr\u00e4n, D. Greenhouse gas abatement \noptimal deployment of biofuels from crops in Germany. \nTransp. Res. Part D Transp. Environ. https://doi.org/10.1016/j.\ntrd.2019.02.005 (2019).\n14.\t Codina Giron\u00e8s, V., Moret, S., Peduzzi, E., Nasato, M. & Mar\u00e9chal, F.  \nOptimal use of biomass in large-scale energy systems: insights for \nenergy policy. Energy 137, 789\u2013797 (2017).\n15.\t Stegmann, P., Daioglou, V., Londo, M., van, D. P. & Junginger, V. M.  \nPlastic futures and their CO2 emissions. Nature 612, 272\u2013276 \n(2022).\n16.\t Meys, R. et al. Achieving net-zero greenhouse gas emission \nplastics by a circular carbon economy. Science 374, 71\u201376 (2021).\n17.\t Zheng, J. & Suh, S. Strategies to reduce the global carbon \nfootprint of plastics. Nat. Clim. Change 9, 374\u2013378 (2019).\n18.\t Tanzer, S. E., Blok, K. & Ram\u00edrez, A. Can bioenergy with carbon \ncapture and storage result in carbon negative steel? Int. J. \nGreenhouse Gas Control 100, 103104 (2020).\n19.\t Mandova, H. et al. Achieving carbon-neutral iron and steelmaking \nin Europe through the deployment of bioenergy with carbon \ncapture and storage. J. Cleaner Prod. 218, 118\u2013129 (2019).\n20.\t Sabatino, F. et al. A comparative energy and costs assessment \nand optimization for direct air capture technologies. Joule 5, \n2047\u20132076 (2021).\n21.\t Communication from the Commission to the European Parliament, \nthe European Council, the Council, the European Economic \nand Social Committee and the Committee of the Regions\u2014The \nEuropean Green Deal (European Commission, 2019).\n22.\t Net Zero Strategy: Build Back Greener. Presented to Parliament \npursuant to Section 14 of the Climate Change Act 2008 \n(UK Government, 2021); https://www.gov.uk/government/\npublications/net-zero-strategy\n23.\t Babiker, M. et al. in Climate Change 2022: Mitigation of Climate \nChange (eds Shukla, P.R. et al.) 295\u2013408 (Cambridge Univ. Press, \n2022).\n24.\t Fuhrman, J. et al. Diverse carbon dioxide removal approaches \ncould reduce impacts on the energy\u2013water\u2013land system.  \nNat. Clim. Change 13, 341\u2013350 (2023).\n25.\t Rogelj, J. et al. Energy system transformations for limiting \nend-of-century warming to below 1.5\u2009\u00b0C. Nat. Clim. Change 5, \n519\u2013527 (2015).\n26.\t Creutzig, F. et al. Bioenergy and climate change mitigation: an \nassessment. GCB Bioenergy 7, 916\u2013944 (2015).\n27.\t Creutzig, F. Economic and ecological views on climate change \nmitigation with bioenergy and negative emissions. GCB Bioenergy \n8, 4\u201310 (2016).\n28.\t Gough, C. et al. Challenges to the use of BECCS as a keystone \ntechnology in pursuit of 1.5\u2009\u00b0C. Glob. Sustain. 1, e5 (2018).\n29.\t Smith, P. et al. Biophysical and economic limits to negative CO2 \nemissions. Nat. Clim. Change 6, 42\u201350 (2016).\n30.\t Heck, V., Gerten, D., Lucht, W. & Popp, A. Biomass-based negative \nemissions difficult to reconcile with planetary boundaries.  \nNat. Clim. Change 8, 151\u2013155 (2018).\n31.\t Calvin, K. et al. Bioenergy for climate change mitigation: \nscale and sustainability. GCB Bioenergy https://doi.org/10.1111/\ngcbb.12863 (2021).\n32.\t Fajardy, M. & Mac Dowell, N. Can BECCS deliver sustainable and \nresource efficient negative emissions? Energy Environ. Sci. 10, \n1389\u20131426 (2017).\n33.\t Donnison, C. et al. Bioenergy with carbon capture and storage \n(BECCS): finding the win-wins for energy, negative emissions and \necosystem services-size matters. GCB Bioenergy 12, 586\u2013604 \n(2020).\n34.\t European Parliament. Directive (EU) 2018/2001 of the European \nParliament and of the Council on the promotion of the use of energy \nfrom renewable sources. Off. J. Eur. Union 2018, 82\u2013209 (2018).\n35.\t The European Parliament and the Council of the European Union. \nDirective (EU) 2023/2413 of the European Parliament and of the \nCouncil of 18 October 2023 amending Directive (EU) 2018/2001, \nRegulation (EU) 2018/1999 and Directive 98/70/EC as regards \nthe promotion of energy from renewable sources, and repealing \nCouncil. Off. J. Eur. Union 2413, 1\u201377 (2023).\n36.\t Daioglou, V. et al. Bioenergy technologies in long-run climate \nchange mitigation: results from the EMF-33 study. Climatic \nChange 163, 1603\u20131620 (2020).\n37.\t Muratori, M. et al. EMF-33 insights on bioenergy with carbon \ncapture and storage (BECCS). Climatic Change 163, 1621\u20131637 \n(2020).\n38.\t Riahi, K. et al. in Climate Change 2022: Mitigation of Climate \nChange (eds Shukla, P.R. et al.) 295\u2013408 (Cambridge Univ. Press, \n2022).\n39.\t Azar, C. et al. The feasibility of low CO2 concentration targets and \nthe role of bio-energy with carbon capture and storage (BECCS). \nClimatic Change 100, 195\u2013202 (2010).\n40.\t Azar, C., Johansson, D. J. & Mattsson, N., Meeting global \ntemperature targets\u2014the role of bioenergy with carbon capture \nand storage. Environ. Res. Lett. https://doi.org/10.1088/1748-9326/ \n8/3/034004 (2013).\n41.\t Klein, D. et al. The value of bioenergy in low stabilization \nscenarios: an assessment using REMIND-MAgPIE. Climatic \nChange 123, 705\u2013718 (2014).\n42.\t Strefler, J. et al. Between Scylla and Charybdis: Delayed \nmitigation narrows the passage between large-scale CDR and \nhigh costs. Environ. Res. Lett. https://doi.org/10.1088/1748-9326/\naab2ba (2018).\n43.\t Van Vuuren, D. P. et al. Alternative pathways to the 1.5\u2009\u00b0C target \nreduce the need for negative emission technologies. Nat. Clim. \nChange 8, 391\u2013397 (2018).\n44.\t Rogelj, J. et al. A new scenario logic for the Paris Agreement \nlong-term temperature goal. Nature 573, 357\u2013363 (2019).\n45.\t Obersteiner, M. et al. How to spend a dwindling greenhouse gas \nbudget. Nat. Climate Change 8, 7\u201310 (2018).\n46.\t Xiao, M., Junne, T., Haas, J. & Klein, M. Plummeting costs of \nrenewables\u2014are energy scenarios lagging? Energy Strategy Rev. \n35, 100636 (2021).\n47.\t Luderer, G. et al. Impact of declining renewable energy costs on \nelectrification in low-emission scenarios. Nat. Energy 7, 32\u201342 \n(2022).\n48.\t Grant, N., Gambhir, A., Mittal, S., Greig, C. & K\u00f6berle, A. C. \nEnhancing the realism of decarbonisation scenarios with \npracticable regional constraints on CO2 storage capacity. \nInt. J. Greenhouse Gas Control https://doi.org/10.1016/j.\nijggc.2022.103766 (2022).\n49.\t Kazlou, T., Cherp, A. & Jewell, J. Feasible deployment of carbon \ncapture and storage and the requirements of climate targets. \nNat. Clim. Change https://doi.org/10.1038/s41558-024-02104-0 \n(2024).\n50.\t Grant, N., Hawkes, A., Napp, T. & Gambhir, A. Cost reductions in \nrenewables can substantially erode the value of carbon capture \nand storage in mitigation pathways. One Earth 4, 1588\u20131601 \n(2021).\n51.\t Realmonte, G. et al. An inter-model assessment of the role of \ndirect air capture in deep mitigation pathways. Nat. Commun. 10, \n3277 (2019).\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n240\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\n52.\t Chen, C. & Tavoni, M. Direct air capture of CO2 and climate \nstabilization: a model-based assessment. Climatic Change 118, \n59\u201372 (2013).\n53.\t Muratori, M. et al. Carbon capture and storage across fuels \nand sectors in energy system transformation pathways. Int. J. \nGreenhouse Gas Control 57, 34\u201341 (2017).\n54.\t Bauer, N. et al. Global energy sector emission reductions and \nbioenergy use: overview of the bioenergy demand phase of the \nEMF-33 model comparison. Climatic Change 163, 1553\u20131568 \n(2020).\n55.\t Leblanc, F. et al. The contribution of bioenergy to the \ndecarbonization of transport: a multi-model assessment. Climatic \nChange 170, 21 (2022).\n56.\t Humpen\u00f6der, F. et al. Investigating afforestation and bioenergy \nCCS as climate change mitigation strategies. Environ. Res. Lett. \nhttps://doi.org/10.1088/1748-9326/9/6/064029 (2014).\n57.\t Ahlgren, E. O., B\u00f6rjesson Hagberg, M. & Grahn, M. T. Transport \nbiofuels in global energy\u2013economy modelling \u2013 a review of \ncomprehensive energy systems assessment approaches. GCB \nBioenergy 9, 1168\u20131180 (2017).\n58.\t Azar, C., Lindgren, K. & Andersson, B. A. Global energy scenarios \nmeeting stringent CO2 constraints\u2014cost-effective fuel  \nchoices in the transportation sector. Energy Policy 31, 961\u2013976 \n(2003).\n59.\t Grahn, M., Azar, C., Lindgren, K., Berndes, G. & Gielen, D. Biomass \nfor heat or as transportation fuel? A comparison between two \nmodel-based studies. Biomass Bioenergy 31, 747\u2013758 (2007).\n60.\t Victoria, M., Zeyen, E. & Brown, T. Speed of technological \ntransformations required in Europe to achieve different climate \ngoals. Joule 6, 1066\u20131086 (2022).\n61.\t Pickering, B., Lombardi, F. & Pfenninger, S. Diversity of options to \neliminate fossil fuels and reach carbon neutrality across the entire \nEuropean energy system. Joule 6, 1253\u20131276 (2022).\n62.\t Wu, F., Muller, A. & Pfenninger, S. Strategic uses for ancillary \nbioenergy in a carbon-neutral and fossil-free 2050 European \nenergy system. Environ. Res. Lett. 18, 014019 (2023).\n63.\t Brill, E. D. The use of optimization models in public-sector \nplanning. Manage. Sci. 25, 413\u2013422 (1979).\n64.\t Lombardi, F., Pickering, B. & Pfenninger, S. What is redundant and \nwhat is not? Computational trade-offs in modelling to generate \nalternatives for energy infrastructure deployment. Appl. Energy \n339, 121002 (2023).\n65.\t Biehl, J. et al. Wicked facets of the German energy transition\u2014\nexamples from the electricity, heating, transport, and industry \nsectors. Int. J. Sustain. Energy 42, 1128\u20131181 (2023).\n66.\t Palmer, J. Risk governance in an age of wicked problems: lessons \nfrom the European approach to indirect land-use change. J. Risk \nRes. 15, 495\u2013513 (2012).\n67.\t Fast, S. & McCormick, K. Biofuels: from a win\u2013win solution to a \nwicked problem? Biofuels 3, 737\u2013748 (2012).\n68.\t Gamborg, C., Anker, H. T. & Sand\u00f8e, P. Ethical and legal \nchallenges in bioenergy governance: coping with value \ndisagreement and regulatory complexity. Energy Policy 69, \n326\u2013333 (2014).\n69.\t Rittel, H. W. & Webber, M. M. Dilemmas in a general theory of \nplanning. Found. Plann. Enterpr. Crit. Essays Plann. Theory 1, \n67\u2013169 (1969).\n70.\t Trutnevyte, E. Does cost optimization approximate the real-world \nenergy transition? Energy 106, 182\u2013193 (2016).\n71.\t DeCarolis, J. et al. Formalizing best practice for energy  \nsystem optimization modelling. Appl. Energy 194, 184\u2013198  \n(2017).\n72.\t Trutnevyte, E. EXPANSE methodology for evaluating the \neconomic potential of renewable energy from an energy mix \nperspective. Appl. Energy 111, 593\u2013601 (2013).\n73.\t DeCarolis, J. F., Babaee, S., Li, B. & Kanungo, S. Modelling to \ngenerate alternatives with an energy system optimization model. \nEnviron. Modell. Softw. 79, 300\u2013310 (2016).\n74.\t Neumann, F. & Brown, T. The near-optimal feasible space of a \nrenewable power system model. Electric Power Syst. Res. 190, \n106690 (2021).\n75.\t Pedersen, T. T., Victoria, M., Rasmussen, M. G. & Andresen, G. B. \nModeling all alternative solutions for highly renewable energy \nsystems. Energy 234, 121294 (2021).\n76.\t Grochowicz, A., van Greevenbroek, K., Benth, F. E. & Zeyringer, M. \nIntersecting near-optimal spaces: European power systems with \nmore resilience to weather variability. Energy Econ.118, 106496 \n(2022).\n77.\t Price, J. & Keppo, I. Modelling to generate alternatives: a \ntechnique to explore uncertainty in energy\u2013environment\u2013\neconomy models. Appl. Energy 195, 356\u2013369 (2017).\n78.\t Williams, J. H. et al. Carbon-neutral pathways for the United \nStates. AGU Adv. https://doi.org/10.1029/2020AV000284 (2021).\n79.\t European Union Regulation of the European Parliament and of \nthe Council Establishing the Framework for Achieving Climate \nNeutrality and Amending Regulations (EC) No 401/2009 and (EU) \n2018/1999 (European Climate Law, 2021).\n80.\t McLaren, D. Quantifying the potential scale of mitigation \ndeterrence from greenhouse gas removal techniques. Climatic \nChange 162, 2411\u20132428 (2020).\n81.\t Geden, O., Peters, G. P. & Scott, V. Targeting carbon dioxide \nremoval in the European Union. Clim. Policy 19, 487\u2013494 (2019).\n82.\t McLaren, D. P., Tyfield, D. P., Willis, R., Szerszynski, B. &  \nMarkusson, N. O. Beyond \u2018net-zero\u2019: a case for separate targets for \nemissions reduction and negative emissions. Front. Clim. 1, 1\u20135 \n(2019).\n83.\t ENSPRESO Integrated Data (European Commission, Joint \nResearch Centre, 2021); http://data.europa.eu/89h/88d1c405-\n0448-4c9e-b565-3c30c9b167f7\n84.\t Statistics\u2014Eurostat Data Browser (Eurostat, 2023); https://\nec.europa.eu/eurostat/databrowser/\n85.\t UK Energy In Brief (United Kingdom Department for Business, \nEnergy & Industrial Strategy, 2022); https://www.gov.uk/\ngovernment/statistics/uk-energy-in-brief-2022\n86.\t Government Expenditure on Defence\u2014Statistics Explained  \n(Eurostat, 2023); https://ec.europa.eu/eurostat/statistics- \nexplained/index.php?title=Government_expenditure_on_defence\n87.\t In-depth Analysis in Support of the Commission Communication \nCOM(2018) 773 - A Clean Planet for All A European Long-term \nStrategic Vision for a Prosperous, Modern, Competitive and \nClimate Neutral Economy (European Commission, 2018).\n88.\t Climate Watch, Historical GHG emissions (World Resources \nInstitute, 2022); https://www.climatewatchdata.org\n89.\t Biomass Markets\u2014Weekly Biomass Market News and Analysis 23-1 \n(Argus, 2023).\n90.\t Cherp, A., Vinichenko, V., Tosun, J., Gordon, J. A. & Jewell, J. \nNational growth dynamics of wind and solar power compared \nto the growth required for global climate targets. Nat. Energy 6, \n742\u2013754 (2021).\n91.\t Weinand, J. M. et al. Historic drivers of onshore wind power siting \nand inevitable future trade-offs. Environ. Res. Lett. https://doi.org/ \n10.1088/1748-9326/ac7603 (2022).\n92.\t McKenna, R. et al. Scenicness assessment of onshore  \nwind sites with geotagged photographs and impacts on  \napproval and cost-efficiency. Nat. Energy 6, 663\u2013672  \n(2021).\n93.\t McKenna, R. et al. High-resolution large-scale onshore \nwind energy assessments: a review of potential definitions, \nmethodologies and future research needs. Renewable Energy \n182, 659\u2013684 (2022).\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n241\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\n94.\t Delafield, G. et al. Conceptual framework for balancing society \nand nature in net-zero energy transitions. Environ. Sci. Policy 125, \n189\u2013201 (2021).\n95.\t Odenweller, A., Ueckerdt, F., Nemet, G. F., Jensterle, M. &  \nLuderer, G. Probabilistic feasibility space of scaling up green \nhydrogen supply. Nat. Energy https://doi.org/10.1038/s41560-022-\n01097-4 (2022).\n96.\t K\u00e4tterer, T. & Bolinder, M. A. in Understanding and Fostering Soil \nCarbon Sequestration (ed. Rumpel, C.) 453\u2013488 (Burleigh Dodds \nScience Publishing, 2022).\n97.\t Scarlat, N., Fahl, F., Lugato, E., Monforti-Ferrario, F. &  \nDallemand, J. F. Integrated and spatially explicit assessment of \nsustainable crop residues potential in Europe. Biomass Bioenergy \n122, 257\u2013269 (2019).\n98.\t Mayer, M. et al. Influence of forest management activities on soil \norganic carbon stocks: a knowledge synthesis. For. Ecol. Manage. \n466, 118127 (2020).\n99.\t M\u00e4kip\u00e4\u00e4, R. et al. How does management affect soil C \nsequestration and greenhouse gas fluxes in boreal and temperate \nforests?\u2014a review. For. Ecol. Manage. https://doi.org/10.1016/ \nj.foreco.2022.120637 (2023).\n100.\tClarke, N. et al. Effects of intensive biomass harvesting on forest \nsoils in the Nordic countries and the UK: a meta-analysis. For. Ecol. \nManage. https://doi.org/10.1016/j.foreco.2020.118877 (2021).\n101.\t Northrup, D. L., Basso, B., Wang, M. Q., Morgan, C. L. & Benfey, P. N.  \nNovel technologies for emission reduction complement \nconservation agriculture to achieve negative emissions from \nrow-crop production. Proc. Natl Acad. Sci. USA 118, e2022666118 \n(2021).\n102.\tAmundson, R. Negative emissions in agriculture are improbable \nin the near future. Proc. Natl Acad. Sci. USA 119, e2118142119  \n(2022).\n103.\tNorthrup, D. L., Basso, B., Wang, M. Q., Morgan, C. L. &  \nBenfey, P. N. Reply to Amundson: time to go to work. Proc. Natl \nAcad. Sci. USA 119, e2122842119 (2022).\n104.\tNabuurs, G.-J. et al. in Climate Change 2022: Mitigation of Climate \nChange (eds Shukla, P.R. et al.) 747\u2013860 (Cambridge Univ. Press, \n2022).\n105.\tMeyfroidt, P. et al. Focus on leakage and spillovers: informing \nland-use governance in a tele-coupled world. Environ. Res. Lett. \nhttps://doi.org/10.1088/1748-9326/ab7397 (2020).\n106.\tYu, B., Zhao, Q. & Wei, Y. M. Review of carbon leakage under \nregionally differentiated climate policies. Sci. Total Environ.782, \n146765 (2021).\n107.\t Williams, E. The Economics of Direct Air Carbon Capture and \nStorage (Global CCS Institute, 2022).\n108.\tYoung, J. et al. The cost of direct air capture and storage can be \nreduced via strategic deployment but is unlikely to fall below \nstated cost targets. One Earth 6, 899\u2013917 (2023).\n109.\tTerlouw, T., Treyer, K., Bauer, C. & Mazzotti, M. Life cycle \nassessment of direct air carbon capture and storage with \nlow-carbon energy sources. Environ. Sci. Technol. 55, 11397\u201311411 \n(2021).\n110.\t Hanssen, S. V. et al. Biomass residues as twenty-first century \nbioenergy feedstock-a comparison of eight integrated \nassessment models. Climatic Change 163, 1569\u20131586 (2020).\n111.\t Millinger, M. & Thr\u00e4n, D. Biomass price developments inhibit \nbiofuel investments and research in Germany: the crucial future \nrole of high yields. J. Cleaner Prod. 172, 1654\u20131663 (2018).\n112.\t Daigneault, A. et al. How the future of the global forest sink \ndepends on timber demand, forest management, and carbon \npolicies. Glob. Environ. Change 76, 102582 (2022).\n113.\t Favero, A., Daigneault, A. & Sohngen, B. Forests: carbon \nsequestration, biomass energy, or both? Sci. Adv. 6, eaay6792 \n(2020).\n114.\t Cintas, O. et al. The climate effect of increased forest bioenergy \nuse in Sweden: evaluation at different spatial and temporal \nscales. Wiley Interdiscip. Rev. Energy Environ. 5, 351\u2013369 (2016).\n115.\t Merfort, L. et al. State of global land regulation inadequate to \ncontrol biofuel land-use-change emissions. Nat. Clim. Change 13, \n610\u2013612 (2023).\n116.\t Englund, O. et al. Large-scale deployment of grass in crop \nrotations as a multifunctional climate mitigation strategy.  \nGCB Bioenergy 15, 166\u2013184 (2023).\n117.\t Rosa, L., Sanchez, D. L. & Mazzotti, M. Assessment of carbon \ndioxide removal potential: via BECCS in a carbon-neutral Europe. \nEnergy Environ. Sci. 14, 3086\u20133097 (2021).\n118.\t European Parliament. Amendments adopted by the European \nParliament on 14 September 2022 on the proposal for a directive \nof the European Parliament and of the Council amending \nDirective (EU) 2018/2001 of the European Parliament and of the \nCouncil, Regulation (EU) 2018/1999 of. Off. J. Eur. Union 125, \n398\u2013461 (2023).\n119.\t Brown, T. et al. millingermarkus/pypsa-eur-sec: PyPSA-Eur- \nSec-MGA-Bio. Zenodo https://doi.org/10.5281/zenodo.8099691 \n(2023).\n120.\tBrown, T., H\u00f6rsch, J. & Schlachtberger, D. PyPSA: Python for \npower system analysis. J. Open Res. Softw. https://doi.org/ \n10.5334/jors.188 (2018).\n121.\t Neumann, F., Zeyen, E., Victoria, M. & Brown, T. The potential role \nof a hydrogen network in Europe. Joule 7, 1793\u20131817 (2023).\n122.\tNeumann, F., Hagenmeyer, V. & Brown, T. Assessments of linear \npower flow and transmission loss approximations in coordinated \ncapacity expansion problems. Appl. Energy 314, 118859 (2022).\n123.\tMantzos, L. et al. JRC-IDEES: Integrated Database of the European \nEnergy Sector: Methodological Note (Publications Office of the \nEuropean Union, 2017).\n124.\tVictoria, M., Zhu, K., Zeyen, E. & Brown, T. Millingermarkus/\ntechnology-data. GitHub https://github.com/millingermarkus/\ntechnology-data/tree/biopower (2023).\n125.\tNeumann, F. & Brown, T. Broad ranges of investment \nconfigurations for renewable power systems, robust to cost \nuncertainty and near-optimality. iScience 26, 106702 (2023).\n126.\tBoiteux, M. Peak-load pricing. J. Bus. 33, 157\u2013179 (1960).\n127.\t Brown, T. & Reichenberg, L. Decreasing market value of variable \nrenewables can be avoided by policy action. Energy Econ. 100, \n105354 (2021).\n128.\tDeCarolis, J. F. Using modeling to generate alternatives (MGA) to \nexpand our thinking on energy futures. Energy Econ. 33, 145\u2013152 \n(2011).\n129.\tRuiz, P. et al. ENSPRESO\u2014an open, EU-28 wide, transparent and \ncoherent database of wind, solar and biomass energy potentials. \nEnergy Strategy Rev. 26, 100379 (2019).\n130.\tKC, S. & Lutz, W. The human core of the shared socioeconomic \npathways: population scenarios by age, sex and level of \neducation for all countries to 2100. Glob. Environ. Change 42, \n181\u2013192 (2017).\n131.\t Kearns, D., Liu, H. & Consoli, C. Technology Readiness and \nCosts of CCS (Global CCS Institute, 2021); https://www.\nglobalccsinstitute.com/wp-content/uploads/2021/03/\nTechnology-Readiness-and-Costs-for-CCS-2021-1.pdf\n132.\tRECTISOL Wash (Linde Engineering, 2023); https://www.\nlinde-engineering.com/en/process-plants/hydrogen_and_\nsynthesis_gas_plants/gas_processing/rectisol_wash/index.html\n133.\tLiu, X., Yang, S., Hu, Z. & Qian, Y. Simulation and assessment of \nan integrated acid gas removal process with higher CO2 capture \nrate. Comput. Chem. Eng. 83, 48\u201357 (2015).\n134.\tYang, S., Qian, Y. & Yang, S. Development of a full CO2 capture \nprocess based on the Rectisol wash technology. Ind. Eng. Chem. \nRes. 55, 6186\u20136193 (2016).\n\nNature Energy | Volume 10 | February 2025 | 226\u2013242\n242\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\n135.\tHannula, I. Co-production of synthetic fuels and district heat from \nbiomass residues, carbon dioxide and electricity: performance \nand cost analysis. Biomass Bioenergy 74, 26\u201346 (2015).\n136.\tTechnology Data (Danish Energy Agency, 2021); https://ens.dk/en/\nour-services/projections-and-models/technology-data/\n137.\t Jewell, J. & Cherp, A. The feasibility of climate action: bridging the \ninside and the outside view through feasibility spaces.  \nWIREs Clim.Change 14, e838 (2023).\n138.\tFreer, M., Gough, C., Welfle, A. & Lea-Langton, A. Carbon optimal \nbioenergy with carbon capture and storage supply chain \nmodelling: how far is too far? Sustain. Energy Technol. Assess. 47, \n101406 (2021).\n139.\tFreer, M., Gough, C., Welfle, A. & Lea-Langton, A. Putting \nbioenergy with carbon capture and storage in a spatial context: \nwhat should go where? Front. Clim. 4, 826982 (2022).\n140.\tLiu, B. & Rajagopal, D. Life-cycle energy and climate benefits of \nenergy recovery from wastes and biomass residues in the United \nStates. Nat. Energy 4, 700\u2013708 (2019).\n141.\t Shirizadeh, B. et al. The impact of methane leakage on the role of \nnatural gas in the European energy transition. Nat. Commun. 14, \n5756 (2023).\n142.\tGlobal Nuclear Power Tracker (Global Energy Monitor, \n2024); https://globalenergymonitor.org/projects/\nglobal-nuclear-power-tracker/\n143.\tElectricity Data Explorer (Ember, 2024); https://ember-climate.org/\ndata/data-tools/data-explorer/\n144.\tZeyen, E. et al. millingermarkus/technology-data: technology \ndata v0.5.0+bio. Zenodo https://doi.org/10.5281/zenodo.10886177 \n(2024).\n145.\tMillinger, M. Results data and plot scripts. Zenodo https://doi.org/ \n10.5281/zenodo.14169801 (2024).\n146.\tAvitabile, V. et al. Biomass Production, Supply, Uses and Flows in \nthe European Union. First Results from an Integrated Assessment \n(Publications Office of the European Union, 2023).\n147.\t Millinger, M., Meisel, K., Budzinski, M. & Thr\u00e4n, D. Relative \ngreenhouse gas abatement cost competitiveness of biofuels in \nGermany. Energies 11, 615 (2018).\n148.\tMalhotra, A. & Schmidt, T. S. Accelerating low-carbon innovation. \nJoule 4, 2259\u20132267 (2020).\n149.\tThe Chemical Engineering Plant Cost Index (CEPCI, 2023);  \nhttps://www.chemengonline.com/pci-home\n150.\tBogdanov, D. et al. Low-cost renewable electricity as the key \ndriver of the global energy transition towards sustainability. \nEnergy 227, 120467 (2021).\n151.\t Fasihi, M., Efimova, O. & Breyer, C. Techno-economic assessment \nof CO2 direct air capture plants. J. Cleaner Prod. 224, 957\u2013980 \n(2019).\n152.\tFuhrman, J. et al. The role of direct air capture and negative \nemissions technologies in the shared socioeconomic pathways \ntowards +1.5\u2009\u00b0C and +2\u2009\u00b0C futures. Environ. Res. Lett. https://doi.org/ \n10.1088/1748-9326/ac2db0 (2021).\nAcknowledgements\nWe acknowledge funding from the Swedish Energy Agency, project \nnumbers 2021-00067 (M.M., F.H., L.R.), 2023-00888 (RESILIENT, M.M.) \nand 2020-004542 (M.M., F.H., L.R.). This research was partially funded \nby CETPartnership, the Clean Energy Transition Partnership under \nthe 2022 joint call for research proposals, co-funded by the European \nCommission (grant agreement number 101069750). This research was \npartially funded by the European Union\u2019s Horizon Europe  \nresearch and innovation programme under the UPTAKE project  \n(g.a. no. 101081521). The views and opinions expressed are, however, \nthose of the author(s) only and do not necessarily reflect those of \nthe European Union or CINEA. The computations and data handling \nwere enabled by resources provided by the National Academic \nInfrastructure for Supercomputing in Sweden (NAISS) and the \nSwedish National Infrastructure for Computing (SNIC) at Chalmers \nCentre for Computational Science and Engineering (C3SE), partially \nfunded by the Swedish Research Council through grant agreement \nnumbers 2022-06725 and 2018-05973.\nAuthor contributions\nM.M. conceived and designed the research together with F.H.; M.M. \nextended the model with feedback from F.N. and E.Z.; M.M. performed \nthe modelling and analysis and created the visualizations; M.M. wrote \nthe paper with feedback from F.H., F.N., E.Z., G.B. and L.R.\nFunding\nOpen access funding provided by Chalmers University of Technology.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at  \nhttps://doi.org/10.1038/s41560-024-01693-6.\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41560-024-01693-6.\nCorrespondence and requests for materials should be addressed to \nM. Millinger.\nPeer review information Nature Energy thanks Nico Bauer, Caspar \nDonnison and Francesco Lombardi for their contribution to the peer \nreview of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard  \nto jurisdictional claims in published maps and institutional  \naffiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this licence, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2025\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nExtended Data Fig. 1 | Simplified depiction of main biomass usage options \nand competing pathways based on electricity-derived energy carriers and \nfossil fuels. Energy flows are shown, except for the dashed lines, which show \nmass flows of captured carbon (which is optional for each process). The captured \ncarbon can be utilized for hydrocarbon production (CCU), or sequestered (CCS). \nAbbreviations: AD = anaerobic digestion, CCU=carbon capture and utilization, \nCCS=carbon capture and storage, DAC=direct air capture, EV=electric vehicle, \nSMR=steam methane reforming, SNG=substitute natural gas, V2G=vehicle to grid.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nExtended Data Fig. 2 | Carbon sequestration capacity requirement for \nmeeting net-nero and net-negative (-110%) emissions targets. 12 MtCO2 carbon \nsequestration slack is added on top of the minimum amount necessary to achieve \ntargets while sequestering process emissions and negative emissions, resulting \nin 140 MtCO2/a for the net-zero scenario and 600 MtCO2/a for the net-negative \n(-110%) emissions scenario.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nExtended Data Fig. 3 | Biomass cost-supply curve. Medium domestic residues from JRC ENSPRESO83 and biomass imports as described in Methods and in11 are \nassumed. Solid biomass usage in the assessed region in 202184,85 and 2020 wood chip prices89 are also shown.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nExtended Data Fig. 4 | Sankey diagram of energy flows in the cost-optimal result for the net-negative (-110%) emissions scenario. The width corresponds to the \nenergy flow. Abbreviations: CHP=combined heat and power.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nExtended Data Fig. 5 | Heat maps of cost-optimal configurations for achieving \na net-negative (-110%) emissions target when varying carbon capture \nefficiency (x-axis) and direct air capture capital expenditure (y-axis). The \ncost of BECC is held constant but would likely in reality also experience cost \nreductions if DAC capital expenditure (CAPEX) is reduced through technological \nlearning. However, cost reductions are expected to be lower in comparison \nbecause BECC is considered to be less modular than DAC108,148. DAC CAPEX of \n1500-7000 \u20ac/kgCO2/h correspond to 171-800 \u20ac/tCO2/a at a full utilization rate,  \nor a capital cost of 16-75 \u20ac/tCO2 with a 7% discount rate and a 20 year lifetime. \nPanel (a) shows total biomass usage while (b) shows biomass imports. Panel  \n(c) shows the system cost increase of excluding biomass. Panels (d-f) show BECC \n(bioenergy with carbon capture), DAC (direct air capture) and total CC (carbon \ncapture) and DAC, respectively.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nExtended Data Fig. 6 | Nodal system cost distribution for achieving a net-negative (-110%) emissions target in the least-cost case. Above with medium domestic \nbiomass potentials and including biomass imports, and below without biomass (except municipal solid waste).\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nExtended Data Table 1 | Cost and efficiency assumptions of biomass technologies, e-fuels and direct air capture, compared \nto assumptions in key studies\nCapital expenditure (CAPEX) costs have been converted to \u20ac2015 for Luderer and Klein, who use 2005, through the Chemical Engineering Plant Cost Index (CEPCI)149 and subsequent \nconversion to \u20ac. \u03b7el denotes the conversion efficiency specifically from biomass to electricity, while \u03b7fu = fuel efficiency, \u03b7th = thermal efficiency and \u03b7cc = carbon capture efficiency. The data \nfrom Bogdanov et.al.150 is for 2040. FT = Fischer-Tropsch, CC = carbon capture, SNG = substitute natural gas, CHP = combined heat and power.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nExtended Data Table 2 | Biomass residue potentials in terms of energy and CO2\nMedium and high potentials are taken from the respective JRC ENSPRESO scenarios83. Forest residues, industry wood residues and landscape care biomass are included in the solid biomass \npotential, while manure & slurry, straw and sewage sludge are assumed to be digestible and municipal solid waste needs to be incinerated separately. For solid biomass and municipal  \nsolid waste, the CO2 is calculated at full combustion, while for digestible biomass the CO2 waste stream from the anaerobic digestion process is added to the CO2 at full combustion of the \nmethane produced.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nExtended Data Table 3 | Technology growth to achieve cost-optimal capacities in net-negative (-110%) emission scenarios\nCompared to historical precedents (solar and wind power, maximum growth rate G88) and growth projections for 2050 (electrolysers, capacity in GW95). For solar and wind, Gompertz curves \nfor technological growth have been estimated as described in the Methods section. The G-values are the highest growth at the inflection point of the S-curve, normalized to total electricity \ndemand at that point.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01693-6\nExtended Data Table 4 | Direct-Air-Capture (DAC) projections and assumptions\nFor the cost projections, a 1 GtCO2/a total scale of global DAC deployment is used for comparability between the two studies. Learning rates of 10-18% and uncertainties for the cost of a \nFirst-of-a-kind (FOAK) plant, as well as uncertain deployment magnitudes lead to a large future CAPEX cost range. Substantial uncertainties in electricity and heat input (for which some \nassume waste heat only) and the cost of energy provision lead to further differences between both cost projections108,151 and modelling studies47,51,62,78,150,152. The DAC technology assumptions \nin this study are based on values for 2040 from DEA [136]. \u2021LCOC (Levelised Cost of Carbon) indicatively estimated as follows: levelised cost calculated by annuitising CAPEX with 7% discount \nrate and a 20 year lifetime, and using base scenario model output time-averaged shadow prices for electricity (65 \u20ac/MWh), process steam (66 \u20ac/MWh) and district heat (25 \u20ac/MWh, for which \n0.1 MWh/tCO2 is output from the process and credited).\n\n\n    Scientific Research Findings:", "answer": "Excluding biomass increases energy system costs by ~20% under stringent emissions targets, similar to excluding wind power or electrolytic hydrogen, and the main value in the European energy system is the provision of carbon rather than energy. How biomass is used is less critical if carbon is captured to provide feedstock for fuels and chemicals and enable negative emissions. However, advanced biofuels and chemicals gain importance if the deployment of carbon capture, variable renewables or electrolytic hydrogen is slow. Biomass remains cost\u2011effective even when associated with some upstream emissions, and the value of biomass and carbon capture and utilization increases significantly as fossil fuels are phased out. Policymakers need to balance the risk\u2011mitigating benefits of limiting the use of biomass for energy with its role in providing renewable carbon to meet emission targets. This must be weighed against uncertainties about the possible scale\u2011up pace of direct air capture, variable renewables, electrolytic hydrogen and carbon capture and storage.", "id": 4}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 10 | January 2025 | 110\u2013123\n110\nnature energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nThe green hydrogen ambition and \nimplementation gap\n \nAdrian Odenweller\u2009\n\u200a\u20091,2\u2009\n & Falko Ueckerdt\u2009\n\u200a\u20091\nGreen hydrogen is critical for decarbonizing hard-to-electrify sectors, \nbut it faces high costs and investment risks. Here we define and quantify \nthe green hydrogen ambition and implementation gap, showing that \nmeeting hydrogen expectations will remain challenging despite surging \nannouncements of projects and subsidies. Tracking 190 projects over \n3\u2009years, we identify a wide 2023 implementation gap with only 7% of global \ncapacity announcements finished on schedule. In contrast, the 2030 \nambition gap towards 1.5\u2009\u00b0C scenarios has been gradually closing as the \nannounced project pipeline has nearly tripled to 422\u2009GW within 3\u2009years. \nHowever, we estimate that, without carbon pricing, realizing all these \nprojects would require global subsidies of US$1.3 trillion (US$0.8\u20132.6 \ntrillion range), far exceeding announced subsidies. Given past and future \nimplementation gaps, policymakers must prepare for prolonged green \nhydrogen scarcity. Policy support needs to secure hydrogen investments, \nbut should focus on applications where hydrogen is indispensable.\nThere is a widespread consensus among scientists1\u20135, industry6 and \nincreasingly also policymakers7 that green hydrogen, produced from \nrenewable electricity via electrolysis, is critical for reducing emissions \nin end-use applications that defy straightforward electrification. Addi-\ntionally, hydrogen is a promising candidate for long-duration energy \nstorage of renewables8,9 and the precursor to all electrofuels10, which are \nhighly versatile yet costly11. Consequently, policy measures to stimulate \nthe ramp-up of the hydrogen market are gaining momentum as more \nthan 40 governments have already adopted hydrogen strategies1,7. \nProminent examples are the supply-side subsidies implemented \nthrough the the US Inflation Reduction Act12 and the EU Hydrogen \nBank13. Such policy support is urgently required: to meet the median \nambition in 1.5\u2009\u00b0C scenarios, namely, 350\u2009GW by 2030, green hydrogen \nproduction needs to grow 380-fold, more than doubling each year. \nHowever, implementation is not going according to plan.\nFollowing a surge of enthusiasm14,15, the green hydrogen market \nand associated expectations have recently entered a phase of consoli-\ndation16 as high costs17,18, limited demand19 and lagging implementa-\ntion of support policies1 are hampering deployment. Shortfalls in the \nannounced deployment of electrolysers, the key component for green \nhydrogen production, are representative of the systemic challenges \nof scaling up supply, demand and infrastructure at the same time. In \n2022, instead of the 2.8\u2009GW electrolysis capacity initially announced, \neventually only 0.62\u2009GW was realized on time (Fig. 1a). Similarly, in \n2023, of the 7.1\u2009GW initially announced, only an estimated 0.92\u2009GW was \nrealized and operational. In stark contrast to these recent setbacks, \nannounced future growth rates of green hydrogen have increased \nsubstantially over the past 3\u2009years, indicating a backlog of projects as \nwell as further increasing ambition (Fig. 1b). This raises questions such \nas whether recent failure rates and the looming \u2018valley of death\u201920 can \nbe overcome to meet updated project announcements, whether the \nexpected role of hydrogen in ambitious climate change mitigation \nscenarios has changed and what plausible implementation pathways \nexist given currently announced hydrogen support policies.\nIn this paper, we structure and analyse the past and future chal-\nlenges of the nascent green hydrogen industry by introducing and \nquantifying the green hydrogen ambition and implementation gap. \nThis builds on the well-established concepts of emissions gaps21 and \nrecent extensions towards a carbon dioxide removal gap22. Looking \nback, we define the past implementation gap as the difference between \nannounced and eventually realized capacity in 2022 and 2023 (Fig. 1a). \nLooking ahead to 2030, we define the ambition gap as the difference \nReceived: 1 June 2024\nAccepted: 12 November 2024\nPublished online: 14 January 2025\n Check for updates\n1Potsdam Institute for Climate Impact Research, Member of the Leibniz Association, Potsdam, Germany. 2Global Energy Systems Analysis, Technical \nUniversity of Berlin, Berlin, Germany. \n\u2009e-mail: adrian.odenweller@pik-potsdam.de\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n111\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nadded capacity was eventually installed and operational, leading to an \noverall success rate of 7% (Fig. 3a). Furthermore, comparing announce-\nments by 2021 with the final outcome reveals that virtually no project \nannounced in 2021 was realized on time in 2023, with 86% experienc-\ning delays and 14% disappearing altogether (Fig. 3b). Similarly, of the \nprojects announced in 2022, only 3% were realized on time, with 76% \ndelayed and 21% disappearing (Fig. 3c). Projects in the feasibility study \nor concept stage almost always had a success rate of zero, implying that \nprojects announced without a final investment decision (FID) in 2021 \nor 2022 were never realized on time in 2023 (Fig. 3b,c). Across all years \nof announcement, even projects that had secured FIDs, or that were \nalready under construction, were mostly delayed or had disappeared \n(Fig. 3b\u2013d). The success rate varies by region, with projects in North \nAmerica equivalent to the global average, European projects below \naverage, Asian projects above average and a success rate of zero for Aus-\ntralian projects (Supplementary Figs. 7\u201310). On the global level, these \nhigh failure rates are not compensated by an influx of newly announced \nprojects or projects that were delayed from previous years (grey bars \nin Fig. 3a), such that a dramatic green hydrogen implementation gap \nof almost 4\u2009GW remained in 2023.\nThe low success rates of green hydrogen projects are not unique \nto the year 2023. In 2022, the overall success rate was 6%, with simi-\nlar patterns of delay and disappearance of projects over time (Sup-\nplementary Fig. 5). The high failure rates in 2022 and 2023 may be \nattributed to supply chain disruptions caused by COVID-19, surging \nelectricity prices during the European energy crisis and rising global \ninterest rates. However, in Europe, the energy crisis was also seen as an \nopportunity to accelerate green hydrogen deployment, although this \nhas yet to materialize (Supplementary Fig. 9). Considering the project \nannouncements for 2024, it remains questionable whether the more \nthan 12\u2009GW currently announced will be realized on time (Supplemen-\ntary Fig. 6). Although nearly 5\u2009GW (40%) has already achieved an FID or \nis under construction, this was also the case for project announcements \nbetween 1.5\u2009\u00b0C scenario requirements and announced projects and \nfind that it has been gradually closing in the past 3\u2009years for most sce-\nnarios (Fig. 1b). However, this has been accompanied by a widening \nfuture implementation gap, which we define as the difference between \nannounced projects and projects that are backed by policies in 2030 \n(Fig. 1b). Analysing the competition between green hydrogen (and \nhydrogen-based electrofuels) and incumbent fossil competitors across \n14 end-use sectors, we estimate that realizing all green hydrogen pro-\njects would require subsidies, or alternative policies such as end-use \nquotas, for at least another decade, even with ambitious carbon pricing \nand potentially indefinitely without. This paper is structured around \nthese three gaps and concludes with a discussion of policy implications \nto safeguard climate targets against uncertain green hydrogen supply.\nThe wide green hydrogen implementation gap in \n2022 and 2023\nGreen hydrogen project announcements reveal two opposing trends \nover the past 3\u2009years. First, there has been a notable short-term setback, \nwith capacities diminishing as projects approach their announced \nlaunch year (Fig. 2a). This trend of downward-adjusted expectations \npersists in both 2022 and 2023, indicating a dramatic green hydro-\ngen implementation gap in recent years. Second, however, this trend \nreverses from 2024 onwards, with project announcements increasing \nsteadily over the past 3\u2009years (Fig. 2b). This steep mid-term growth of \nannouncements is mostly driven by Europe, which accounts for the \nlargest share of announced capacity by 2030, followed by Australia and \nCentral and South America (Fig. 2d). These opposing trends raise the \nquestion as to whether future promises can overcome past setbacks. We \naddress this question in the next section, following the quantification \nof the 2022 and 2023 green hydrogen implementation gaps.\nTracking 190 individual green hydrogen projects announced glob-\nally for 2023 over the past 3\u2009years (Methods), we observe a substantial \nimplementation gap as only 0.3\u2009GW of the initially announced 4.3\u2009GW \n2022\n2023\n0\n5\n10\nYear\nCapacity (GW)\nProject announcements\nAnnouncements by 2021\nAnnouncements by 2022\nAnnouncements by 2023\nRealized projects\nPast green hydrogen implementation gap\na\n1.5 \u00b0C\nscenario\nrequirements\n0\n200\n400\n600\n800\n2020\n2021\n2022\n2023\n2024\n2024\n2025\n2026\n2027\n2028\n2029\n2030\nYear\nCapacity (GW)\nProject announcements\nAnnouncements by 2021\nAnnouncements by 2022\nAnnouncements by 2023\nSupported by implemented demand policies\nand announced subsidies\n2030 green hydrogen ambition and implementation gaps\nb\n 2030\nambition\ngap\n(2)\n 2030\nimplementation\ngap\n(3)\n Past\nimplementation\ngap\n(1)\nFig. 1 | The green hydrogen ambition and implementation gaps in the past \nand the future. a, Past green hydrogen implementation gaps in 2022 and 2023, \ndefined as the difference between project announcements and realized projects \n(denoted as (1), also see Fig. 3). Realized projects in 2023 show the outcome \nof project announcements by 2023, based on our own research (Methods). \nb, Green hydrogen ambition and implementation gaps in 2030. We define \nthe 2030 ambition gap as the difference between 1.5\u2009\u00b0C scenarios and project \nannouncements (denoted as (2), also see Fig. 4). The depicted data range shows \nthe IEA Net Zero Emissions by 2050 scenarios, while the full analysis includes \nfurther scenarios (Fig. 4a, Extended Data Fig. 1 and Methods). We define the \n2030 green hydrogen implementation gap as the difference between project \nannouncements and our estimate of projects that are either supported by \nimplemented demand-side policies or by currently announced subsidies \n(denoted as (3), see Fig. 5 and Supplementary Fig. 17). The black line indicates \nour central estimate and the light grey corridor indicates the uncertainty range \nspanned by the sensitivity analysis. Green hydrogen project announcements \nare displayed in terms of electrical input capacity of electrolysers. Project \nannouncements are based on three snapshots of the IEA Hydrogen Projects \nDatabase, which we have validated comprehensively (see Methods, \nSupplementary Table 1 and Supplementary Figs. 1\u20134). The dashed curve between \na and b connects the same data point in 2024 and illustrates the different y-axis \nscale between project announcements until 2024 (a) and until 2030 (b).\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n112\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nmade in 2022 for 2023, of which only 8% were completed on schedule \n(Fig. 3c). It will take some more years to determine whether the recent \nimplementation gaps were exceptions caused by unusual global events, \nor the unfortunate norm.\nSubstantial implementation gaps may be common for emerging \nenergy technologies in the early stages of technology diffusion, as large \nprojects almost always exceed their budget and run behind schedule23. \nHowever, while research has identified similarly high failure rates for \ncomplex and customized technologies24 such as carbon capture and \nstorage25, this does not apply to highly modular technologies such \nas solar photovoltaics (PV) and wind power23,26. For green hydrogen, \nrecent evidence suggests that while the mass-producible electrolyser \nstack is highly modular, other components of the electrolyser system \nand the overall green hydrogen production plant are more complex and \nrequire customization17, making them more prone to budget and time \noverrun23. As long as the underlying uncertainties remain unresolved, \npolicymakers should avoid relying solely on project announcements \nto assess progress on green hydrogen.\nApart from the unsettled question of electrolyser modularity, \nthree tangible factors contribute to the low success rate of green hydro-\ngen projects. First, cost estimates for electrolysers have recently surged \ndue to increasing equipment and financial costs1, and because only \nthe electrolyser stack may be set for rapid cost reductions24. Second, \nanalysts have observed a lack of offtake agreements19, which could arise \nfrom a limited willingness to pay for costly green hydrogen. Further-\nmore, required hydrogen end-use investments, such as transforming \nsteel production from a blast furnace to a direct reduction route, are \noften difficult to reverse and therefore pose the risk of becoming locked \ninto an expensive and potentially scarce energy carrier. Third, bridg-\ning the substantial cost gap and reducing investment risks requires \nhydrogen-specific support policies and regulation, even in countries \nwith ambitious carbon pricing27. However, lagging implementation of \nsupport policies1 and regulatory uncertainty regarding green hydrogen \nproduction standards in the European Union (EU) and the United States, \nalthough crucial to ensure climate benefits28,29, have hampered growth.\nWhat implications does the sobering track record of past project \nannouncements have for the future of green hydrogen in ambitious \nclimate change mitigation scenarios? To explore these ramifica-\ntions, we next focus on the mid-term horizon towards 2030. First, we \nprovide an overview of electrolysis requirements in 1.5\u2009\u00b0C scenarios, \nintroducing the 2030 green hydrogen ambition gap. Second, we \nanalyse the economic viability of surging project announcements \nAnnouncements\nby 2021\nAnnouncements\nby 2022\nOutcome\nin 2023\n0\n5\n10\nCapacity (GW)\nProjects until 2024 by status\na\nAnnouncements\nby 2021\nAnnouncements\nby 2022\nAnnouncements\nby 2023\n0\n100\n200\n300\n400\nCapacity (GW)\nProjects from 2024 by status\nb\n0\n5\n10\nYear\nCapacity (GW)\nProjects until 2024 by region\nc\n0\n100\n200\n300\n400\n2020\n2021\n2022\n2023\n2024\n2024\n2025\n2026\n2027\n2028\n2029\n2030\nYear\nYear\n2020\n2021\n2022\n2023\n2024\n2024\n2025\n2026\n2027\n2028\n2029\n2030\nYear\nCapacity (GW)\nProjects from 2024 by region\nd\nStatus\nConcept\nFeasibility study\nFID/construction\nOperational\nRegion\nAsia\nAustralia\nC and S America\nEurope\nMENA\nNorth America\nOther\nFig. 2 | Green hydrogen project announcements by 2021, 2022 and 2023. \na,b, Project announcements by status from 2020\u20132024 (a) and 2024\u20132030 (b). \nc,d, Project announcements by region from 2020\u20132024 (c) and 2024\u20132030 \n(d). For each year there are three bars. The left bar shows announcements by \n2021, the middle bar shows announcements by 2022 and the right bar shows \nannouncements by 2023, each of which corresponds to different project \ndatabase snapshots (Methods). Two main trends are visible. First, in 2022 and \n2023, project announcements decrease strongly as the year of project launch \napproaches (a,c), leading to a wide green hydrogen implementation gap (see \nFig. 3 and Supplementary Fig. 5). Second, after 2024, this pattern reverses as the \nproject pipeline has surged over the past 3\u2009years (b,d), thereby gradually closing \nthe green hydrogen ambition gap to 1.5\u2009\u00b0C scenarios (see Fig. 4). However, the \nvast majority of projects have not secured an FID yet (b), which gives rise to the \n2030 green hydrogen implementation gap due to a mismatch of required and \nannounced policies (see Fig. 5). In contrast to Figs. 1a and 3, this figure does not \nshow the outcome of project announcements for 2023. C and S America, Central \nand South America; MENA, Middle East and North Africa. Region mapping is \navailable in ref. 67.\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n113\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nand estimate the subsidy volumes that would be required  \nto realize all projects, leading to the 2030 green hydrogen imple-\nmentation gap.\nThe closing 2030 green hydrogen ambition gap\nComparing green hydrogen project announcements with 1.5\u2009\u00b0C sce-\nnarios, we find that the green hydrogen ambition gap for 2030 has \nbeen gradually closing over the past 3\u2009years (Fig. 4). Due to a stead-\nily growing project pipeline, the gap has already closed for most \nscenarios, including the median of both the integrated assessment \nmodel (IAM) scenarios (169\u2009GW) and the institutional and corporate  \nscenarios (350\u2009GW).\nGreen hydrogen requirements vary substantially across different \n1.5\u2009\u00b0C scenarios, consistent with previous research30 (Fig. 4a). For 2030, \nthis lack of consensus leads to an enormous range of 3\u20131,072\u2009GW for \nthe IAM scenarios and 30\u20131,016\u2009GW for the institutional and corpo-\nrate scenarios (excluding an outlier of 1,700\u2009GW), with correspond-\ning interquartile ranges of 38\u2013375 and 203\u2013655\u2009GW, respectively.  \nAnnouncements\nby 2021\nAnnouncements\nby 2022\nAnnouncements\nby 2023\nOutcome\nin 2024\nAnnouncements\nby 2021 VERSUS outcome in 2024\n0\n25\n50\n75\n100\nShare of capacity additions in 2023 (%)\nTotal\n4.3 GW\nFID/\nconstruction\nFeasibility\nstudy\nConcept\nBy status\n3.1 GW\nb\n0.7\nGW\n0.5\nGW\nAnnouncements\nby 2022 VERSUS outcome in 2024\n0\n25\n50\n75\n100\nShare of capacity additions in 2023 (%)\nTotal\n2.6 GW\nFID/\nconstruction\nFeasibility\nstudy\nConcept\nBy status\n1.2 GW\n1.2 GW\nc\n0.2\nGW\nAnnouncements\nby 2023 VERSUS outcome in 2024\n0\n25\n50\n75\n100\nShare of capacity additions in 2023 (%)\nTotal\n0.6 GW\nFID/\nconstruction\nFeasibility\nstudy\nBy status\n0.5 GW\nd\nOutcome\nDisappeared\nDelayed\nOn time\n0.1\nGW\n Green hydrogen\n(1)\nimplementation\ngap in 2023\n4.28 GW\nannounced by 2021\n2.6 GW\nannounced by 2022\n1.01 GW\nannounced by 2023\n0.3 GW\noperational as of 2024\nFID/\nconstruction\nFeasibility\nstudy\nConcept\nNew/\ndelayed\nNew/\ndelayed\nDelayed\nto 2024+\nDisappeared\nNew\nOn time\n0\n1\n2\n3\n4\n5\nCapacity additions announced for 2023 (GW)\nTracking global green hydrogen projects announced for 2023\na\nFig. 3 | The 2023 green hydrogen implementation gap. a, Sankey diagram \nshowing the development of green hydrogen projects announced for 2023 in \nterms of added electrolysis capacity (n\u2009=\u2009190). The bars show different snapshots \nof the underlying project database, where, for example, \u2018Announcements by \n2021\u2019 refers to the database published in 2021 and therefore contains project \nannouncements made by 2021 (Methods). In 2021, 4.3\u2009GW of new capacity was \nannounced to be installed in 2023. This was revised downward to 2.6\u2009GW in 2022, \nand again to 1\u2009GW in 2023. Finally, in 2024, it became clear that only 0.3\u2009GW of \nnew capacity had been installed and was operational in 2023. This results in \na green hydrogen implementation gap of almost 4\u2009GW in 2023. In contrast to \nFig. 2, this figure additionally shows the outcome of project announcements \nfor 2023 as \u2018Outcome in 2024\u2019, based on our own research (Methods). The \noutcome in 2024 refers only to projects that were included in the 2023 database. \nAdditional projects that were missing in the 2023 database could change the \nsuccess rate. b\u2013d, Percentage rates of success, delay and disappearance of \nuncertain green hydrogen projects announced to launch in 2023, comparing \nannouncements by 2021 with the outcome in 2024 (b), announcements by 2022 \nwith the outcome in 2024 (c) and announcements by 2023 with the outcome in \n2024 (d). In b\u2013d, the left panel shows the total share and the right panel shows the \ndisaggregation by status. As indicated by the horizontal whiskers at the bottom, \nthe widths of the bars in the right panels correspond to the share of the total \ncapacity (also compare with a). Within each colour band, individual projects \nare shown as segments, ordered by size. The \u2018disappeared\u2019 outcome category \ncontains projects that appeared in one database, but were absent in subsequent \ndatabases. This includes cancelled or discontinued projects, but may also be due \nto other reasons (Supplementary Note 1).\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n114\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nThis heterogeneity results from two key uncertainties. First, the pace \nat which the nascent green hydrogen value chain can be scaled up is \nhighly uncertain31, particularly as project announcements have been \na poor indicator of growth. However, to reach 1.5\u2009\u00b0C scenarios by 2030, \ngreen hydrogen would need to experience unprecedented growth \nrates (Extended Data Fig. 1a,c). Second, although evidence shows that \nhydrogen and electrofuels are promising for decarbonizing maritime \nshipping32, aviation33 and steel34, substantial uncertainty remains con-\ncerning the competition with alternative mitigation options such as \ndirect electrification, biofuels or carbon capture and storage35\u201337. This \nstructural uncertainty also persists in the long run, explaining the high \nheterogeneity until 2050 (Extended Data Fig. 1b,d).\nDespite the high heterogeneity, a notable trend emerges in a subset \nof the 1.5\u2009\u00b0C scenarios: the International Energy Agency (IEA) Net Zero \nEmissions by 2050 Scenario (NZE), which has been updated annually \nover the past 3\u2009years38\u201340, indicates a steady downward revision of \nrequired electrolysis for 2030 (Fig. 4b). This adjustment reflects recent \nsetbacks for green hydrogen and the rapid progress of competing miti-\ngation options, particularly the deep electrification of road transport \nas well as industrial and residential heat40. Meanwhile, the 2030 green \nhydrogen project pipeline has nearly tripled from 161\u2009GW to 422\u2009GW, \nsurpassing the requirements for 1.5\u2009\u00b0C in 48 of the 60 IAM scenarios, and \n9 of the 15 institutional and corporate scenarios. As a result, the green \nhydrogen ambition gap in 2030 has already closed for 60\u201380% of the \nscenarios and can be expected to close soon for the IEA NZE scenario.\nAlthough the convergence of project announcements and 1.5\u2009\u00b0C \nscenarios is encouraging, the past green hydrogen implementation \ngaps in 2022 and 2023 cast doubt on the reliability of ever-increasing \nproject announcements. Of the 422\u2009GW announced by 2030, 97% are \nstill in the concept or feasibility study phase, which have exhibited criti-\ncally insufficient success rates in the past (see the previous section). \nAchieving the level of ambition required in 1.5\u2009\u00b0C scenarios hinges on \novercoming these high failure rates. Yet, how much policy support \nwould be required to realize all project announcements?\nEstimating the 2030 green hydrogen \nimplementation gap\nThe flipside of the closing of the green hydrogen ambition gap is the \nwidening future green hydrogen implementation gap in 2030, which \nwe define as the difference between project announcements and pro-\njects that are supported by policies. In this context, we estimated the \npolicy support required to realize all 422\u2009GW of green hydrogen project \nannouncements by 2030. Modelling pay-as-bid market premium auc-\ntions, we estimated the required subsidies across 14 end-use sectors rep-\nresented in the projects database (Extended Data Fig. 2). We modelled \nthe competition between four green products (green hydrogen, plus \nthree hydrogen-based synthetic electrofuels, e-methanol, e-kerosene \nand e-methane) and five incumbent fossil competitors (natural gas, grey \nhydrogen, grey methanol, kerosene and diesel). For each end use, we \ncalculated the gradually declining cost gap between the green product \nand its fossil competitor, considering higher efficiencies of hydrogen if \napplicable (Extended Data Table 1) and accounting for end-use-specific \ntransport and storage costs (Supplementary Table 2). We explored \nthe impact of more progressive and more conservative parameter \nvalues, which cover wide ranges for green products (Extended Data \nTable 2) and fossil competitors (Extended Data Table 3). For the latter, \n0\n300\n600\n900\nIAM\nscenarios\nInstitutional and\ncorporate scenarios\nCapacity in 2030 (GW)\nGreen hydrogen in 1.5\u00b0C scenarios in 2030\na\nn = 528\nn = 788\nn = 1,168\nNZE 2021\nNZE 2022\nNZE 2023\nIncreasing\ngreen hydrogen\nproject pipeline\n Green hydrogen\n(2)\nambition gap in 2030\nDecreasing role of\ngreen hydrogen in\nIEA Net Zero Emissions\nscenarios\n0\n300\n600\n900\n2021\n2022\n2023\nYear of announcement/publication\nCapacity in 2030 (GW)\nGreen hydrogen ambition gap in 2030\nb\nProject status\nConcept\nFeasibility study\nFID/construction\nOperational\nFig. 4 | The closing green hydrogen ambition gap in 2030. a, Electrolysis \ncapacity requirements for 2030 in 1.5\u2009\u00b0C scenarios from IAMs (n\u2009=\u200960) and \nfrom institutional and corporate 1.5\u2009\u00b0C scenarios (n\u2009=\u200915), excluding one outlier \nscenario with a capacity of 1,700\u2009GW in 2030 (see Extended Data Fig. 1). Each \ndot represents one scenario. Red dots indicate the IEA NZE scenarios (b). The \nwhiskers indicate the range of capacities, 3\u20131,072\u2009GW for the IAM scenarios and \n30\u20131,016\u2009GW for the institutional and corporate scenarios, underlining the high \nuncertainty around mid-term green hydrogen deployment. The boxes indicate \nthe upper and lower quartiles, spanning the interquartile range of 38\u2013375\u2009GW \nfor the IAM scenarios and 203\u2013655\u2009GW for the institutional and corporate \nscenarios. The horizontal line inside each box indicates the median at 169 and \n350\u2009GW, respectively. For the IAM scenarios, it remains uncertain whether \nmodels explicitly represent different hydrogen applications and whether the \nresults have been vetted. When estimating the required subsidies for a 1.5\u2009\u00b0C \nscenario, we therefore only used the institutional and corporate scenarios \n(Methods and Supplementary Figs. 11 and 13). Extended Data Fig. 1 shows data \nfor all of the scenarios over time. b, Electrolysis capacity requirements in the \nIEA NZE scenarios and the project pipeline for 2030. Only the NZE scenarios \nprovide annually updated electrolysis capacity in 2030 over the past 3\u2009years. \nThe x axis shows the year of announcement of the projects database and the year \nof publication of the NZE scenarios. Individual projects are shown as segments \nwithin the coloured bars. For the NZE scenarios, the green hydrogen ambition \ngap in 2030 has gradually closed as (1) the project pipeline for 2030 has almost \ntripled in the past 3\u2009years and (2) the NZE scenarios in the past 3\u2009years show a \ndecreasing role of green hydrogen by 2030. For 80% of the IAM scenarios and \n60% of the institutional and corporate scenarios, the 2030 ambition gap has \nalready closed. However, more than 97% of the announced project capacity in \n2030 is not yet backed by an FID.\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n115\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nwe also assessed the impact of a high carbon price in line with EU cli-\nmate targets. To recover their costs, green hydrogen and electrofuel \nprojects must sell at their respective levelized costs throughout the \npayback period (see illustrative explanation in Extended Data Fig. 3). \nAssuming that offtakers are broadly not willing to pay a premium for \ngreen products, the cost gap determines the specific per-megawatt \nhour subsidy required. To estimate the total required subsidies, for \neach end use, we tracked all project announcements throughout their \npayback period and combined this vintage tracking with the cost gap \nbetween the levelized cost of the projects and the corresponding fossil \nfuel cost. Our model includes the impact of end-use-specific imple-\nmented demand-side policies, which reduce subsidy requirements  \nWithout carbon price\nWith ambitious carbon price\n0\n50\n100\n150\nYear\nSpecific cost (US$ MWhLHV\n\u20131)\nSpecific cost (US$ MWhLHV\n\u20131)\nSpecific cost (US$ MWhLHV\n\u20131)\nSpecific cost (US$ MWhLHV\n\u20131)\nGreen hydrogen versus natural gas\na\n0\n50\n100\n150\n2024\n2030\n2035\n2040\n2045\n2024\n2030\n2035\n2040\n2045\nYear\nYear\n2024\n2030\n2035\n2040\n2045\n2024\n2030\n2035\n2040\n2045\nYear\nYear\n2024\n2030\n2035\n2040\n2045\n2024\n2030\n2035\n2040\n2045\nYear\nGreen hydrogen versus natural gas\nb\nEnd uses (214 GW)\nOther industry\nPower\nGrid injection\nCHP\nDomestic heat\nIron and steel\nNA\n0\n50\n100\n150\nGreen hydrogen versus grey hydrogen\nc\n0\n50\n100\n150\nGreen hydrogen versus grey hydrogen\nd\nEnd uses (171 GW)\nAmmonia\nRefining\nBiofuels\nTotal cost\nGreen hydrogen\nNatural gas/grey hydrogen\nGreen hydrogen\nElectricity\nTransport and storage\nFixed O&M\nInvestment: stack\nInvestment: balance of plant\nNatural gas/grey hydrogen\nCO2 price\nTransport and storage\nPrice\n0\n25\n50\n75\n100\n0\n500\n1,000\n1,500\nAnnual subsidies (US$ billion yr\u20131)\nCumulative subsidies (US$ billion)\nRequired subsidies for all projects by 2030\ne\n0\n25\n50\n75\n100\n0\n500\n1,000\n1,500\nAnnual subsidies (US$ billion yr\u20131)\nCumulative subsidies (US$ billion)\nRequired subsidies for all projects by 2030\nf\nAnnual subsidies (left axis)\nProjects 2029\u20132030\nProjects 2027\u20132028\nProjects 2024\u20132026\nCumulative subsidies (right axis)\nProjects until 2030\nProjects until 2028\nProjects until 2026\n/\n/\n/\n  Green hydrogen\nimplementation gap\nin 2030\nUS$0.5 trillion\n(US$0.1\u20132.0\ntrillion)\n0\n500\n1,000\n1,500\nRequired\nwithout\ncarbon price\nRequired\nwith ambitious\ncarbon price\nAnnounced\n(BNEF)\nCumulative subsidies (US$ billion)\nRequired cumulative subsidies\nfor all projects by 2030\ng\nUS$1.3 trillion\n(US$0.8\u20132.6 trillion)\n(3)\nFig. 5 | The green hydrogen implementation gap in 2030. a\u2013d, Cost gap between \ngreen hydrogen and natural gas (a,b) and between green hydrogen and grey \nhydrogen (c,d) without carbon pricing (a,c) and with an ambitious carbon price \npathway (b,d) that is in line with reaching EU climate targets41 (US$149\u2009tCO2\n\u22121 \nin 2030, US$246\u2009tCO2\n\u22121 in 2040 and US$407\u2009tCO2\n\u22121 in 2050, see Extended Data \nTable 3). These two markets cover over 90% of the project announcements by \n2030 (Extended Data Fig. 2). The represented end uses are shown next to each \nrow. Extended Data Fig. 5 displays the full set of competition across all end uses, \ncovering four other markets and different hydrogen-based electrofuels. The \nred double-headed arrows and the light-red shading indicate the cost gap that \nneeds to be bridged by subsidies. The stacked bars indicate the decomposition \nof the LCOH and the total cost of the fossil competitor for selected years (2024, \n2030, 2035, 2040 and 2045). For easier visualization, the LCOH bar is shown on \nthe left and the fossil competitor bar on the right. Our 2030 LCOH estimates are \nin line with recent studies (see Extended Data Fig. 4). LHV, lower hydrogen value. \nO&M, operations and maintenance. CHP, combined heat and power. NA, not \navailable (end use unknown). e,f, Subsidies required to bridge the cost gap across \nall end uses to realize all project announcements until 2030 on time, without \ncarbon pricing (e) and with carbon pricing (f). The bars show the required annual \nsubsidies (left axis) and the lines show the required cumulative subsidies (right \naxis). g, Cumulative subsidies required to realize all project announcements by \n2030 compared with globally announced hydrogen subsidies as of September \n2023 from BloombergNEF (BNEF)43. Our estimate takes currently implemented \ndemand-side policies into account (see Methods and Supplementary Fig. 15). \nWithout carbon pricing, US$1.3 trillion of subsidies are required to realize all \nprojects announced until 2030 (the values in parentheses show the ranges \nof more progressive and conservative parameters, see Extended Data Fig. 6). \nNote that e and f show only the subsidies required for green hydrogen project \nannouncements until 2030. Staying on a 1.5\u2009\u00b0C scenario requires substantial \nfurther subsidies after 2030 (Supplementary Fig. 16 and Table 1).\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n116\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nby increasing the willingness to pay (Supplementary Fig. 15 and Meth-\nods) but also incur macroeconomic costs (Supplementary Note 5).\nAcross all end uses, the competitiveness analysis reveals a sub-\nstantial and prolonged cost gap between all green products and their \nrespective fossil competitors. This is exemplified by the competition \nbetween green hydrogen and natural gas, which is relevant for end uses \nsuch as industry, power and grid injection (Fig. 5a,b), as well as between \ngreen hydrogen and grey hydrogen, covering the end uses ammonia, \nrefining and some biofuel routes (Fig. 5c,d). Together, these account \nfor over 90% of the announced electrolyser capacity by 2030 (Extended \nData Fig. 2). In contrast, project announcements for electrofuels remain \nlimited, which may be due to a larger cost gap to the fossil competi-\ntors in the respective end uses (Extended Data Fig. 5g\u2013l). Without \ncarbon pricing, the cost gap between green hydrogen and natural gas \nof US$150\u2009MWh\u22121 in 2024 implies that green hydrogen is initially more \nthan seven times as expensive as natural gas (Fig. 5a), while the cost gap \nbetween green hydrogen and grey hydrogen is only slightly lower at \nUS$121\u2009MWh\u22121 in 2024 (Fig. 5c). As green hydrogen costs decrease, the \ncost gap gradually reduces, but typically prevails also into the long term. \nThis pattern holds across all end uses. Without carbon pricing, in our \ncentral estimate, no green product becomes competitive with its fossil \ncompetitor until 2050. This is robust across a wide range of progressive \nand conservative parameter values (Extended Data Fig. 5, left column).\nIn contrast, under an ambitious carbon price pathway in line with \nEU climate targets41 (US$149\u2009tCO2\n\u22121 in 2030, US$246\u2009tCO2\n\u22121 in 2040 and \nUS$407\u2009tCO2\n\u22121 in 2050, see Extended Data Table 3), green products \ngradually achieve cost parity with their fossil competitors. While the \nexact timing of cost parity remains highly uncertain, a relative sequence \nof hydrogen end-use competitiveness can be derived (Fig. 5b,d and \nExtended Data Fig. 5, right column). In our central estimate, green \nhydrogen first becomes competitive with grey hydrogen in 2034 (for \nexample, for ammonia and refining), followed by green hydrogen \nbecoming competitive with diesel in 2037 (for mobility), e-methanol \nbecoming competitive with grey methanol in 2043 (for example, for \nchemicals), and green hydrogen becoming competitive with natural gas \nin 2044 (for example, for industry and power). In our central estimate, \ne-kerosene and e-methane narrowly miss reaching cost parity with their \nfossil competitors by 2050 (Extended Data Fig. 5h,l). Thus, even with \nambitious carbon pricing, the cost gap persists for at least one decade, \ndepending on the end use and the scenario. Sustained support policies \ncomplementing carbon pricing are therefore essential to foster green \nhydrogen growth and reduce investment risks.\nThe main drivers of green hydrogen costs are electricity prices and \nelectrolyser investment costs (Fig. 5a\u2013d). For electrofuels produced \nfrom green hydrogen and renewable carbon, these two factors domi-\nnate the overall costs (Extended Data Fig. 5g\u2013l). Although electrolyser \ninvestment costs have recently surged1,17, this trend is expected to \nreverse soon due to learning by doing and economies of scale. Note \nagain that to estimate the volume of required subsidies, we considered \na scenario where all project announcements until 2030 are realized \non time, while after 2030, cost reductions are driven by the median \nelectrolysis capacity in 1.5\u2009\u00b0C scenarios (Methods and Supplementary \nFig. 11). This leads to rapidly falling electrolyser costs (Supplementary \nFig. 12). We used a payback period of 15\u2009years to calculate the levelized \ncosts (Methods), as well as to estimate the required subsidies (Extended \nData Fig. 3); this period represents the typical length of implemented \npolicy support such as auctions42 and is therefore more relevant for \ninvestment decisions than the technical lifetime. Our 2030 levelized \ncosts of green hydrogen (LCOHs) are consistent with recent studies \n(Extended Data Fig. 4).\nThe annual subsidies required to realize all project announce-\nments across all end uses by 2030 are bell-shaped, with the height and \ntiming of the peak varying by scenario (Fig. 5e,f, left axis). Without \ncarbon pricing, the required annual subsidies rise sharply to a plateau \nof around US$90 billion per year throughout the 2030s (Fig. 5e). With \ncarbon pricing, the required annual subsidies peak at US$44 billion \nper year in 2030 (Fig. 5f). The resulting cumulative subsidies for all \n422\u2009GW by 2030 follow an S curve (Fig. 5e,f, right axis). In our central \nestimate, the required cumulative subsidies are US$1.3 trillion without \ncarbon pricing and US$0.5 trillion with carbon pricing, subject to \nconsiderable uncertainty (Table 1 and Extended Data Fig. 6). However, \nthese figures only pertain to the 2030 project pipeline. Aligning green \nhydrogen with 1.5\u2009\u00b0C scenarios after 2030 would require substantially \nhigher subsidies, rising to US$9.3 trillion (US$4.2\u201317.7 trillion range) \nwithout carbon pricing by 2050 (Table 1 and Supplementary Fig. 16).\nDue to a substantial discrepancy between required and announced \nsubsidies, a wide 2030 green hydrogen implementation gap arises \n(Fig. 5g and Table 1). The cumulative subsidies required to realize all \nproject announcements by 2030 exceed currently announced subsi-\ndies, estimated at US$308 billion as of September 202343, by over 300% \nwithout carbon pricing and by over 60% without. There are counteract-\ning uncertainties regarding this estimate, as announced subsidies are \nlikely to increase in the future, but challenges may arise during their \nimplementation (Methods). Even if all currently announced global \nsubsidies were immediately available, without carbon pricing this \nwould only support 61\u2009GW (32\u2013106\u2009GW range) by 2030 (Supplemen-\ntary Fig. 17). Depending on the scenario, implemented demand-side \npolicies could support a similar share of project announcements, \nTable 1 | Estimating the 2030 green hydrogen implementation gap for different scenarios\nWithout carbon price\nWith ambitious carbon price\nCentral\nProgressive\nConservative\nCentral\nProgressive\nConservative\nGreen hydrogen project \nannouncements by 2030\nRequired total cumulative \nsubsidies (US$)\n1.3 trillion\n0.8 trillion\n2.6 trillion\n0.5 trillion\n0.1 trillion\n2.0 trillion\nAnnounced subsidies (US$)\n0.3 trillion\nImplementation gap (US$)\n1.0 trillion\n0.5 trillion\n2.3 trillion\n0.2 trillion\n0\n1.7 trillion\nGreen hydrogen scale-up \nuntil 2050 (median 1.5\u2009\u00b0C \nscenario)\nRequired cumulative subsidies  \nby 2050 (US$)\n9.3 trillion\n4.2 trillion\n17.7 trillion\n2.4 trillion\n0.1 trillion\n12.4 trillion\nAverage specific subsidies in  \n2050 (US$\u2009MWh\u22121)\n98\n35\n214\n8\n0\n119\nFor project announcements by 2030, the table shows required total cumulative subsidies (which are required until 2045, see Fig. 5), announced subsidies and the resulting implementation \ngap in terms of the missing subsidies that would be required to realize all project announcements from 2024 to 2030 on time. Without carbon pricing, there is a substantial 2030 \nimplementation gap. Even with carbon prices in line with reaching EU climate targets41 (US$149\u2009tCO2\n\u22121 in 2030, US$246\u2009tCO2\n\u22121 in 2040 and US$407\u2009tCO2\n\u22121 in 2050, see Extended Data Table 3), \nthe implementation gap only closes for the progressive scenario. Beyond 2030, we modelled the green hydrogen scale-up until 2050 by using the median of all institutional and corporate \n1.5\u2009\u00b0C scenarios (Extended Data Fig. 1a,b) and the end-use shares from the IEA NZE Scenario (Supplementary Fig. 13). For this scenario, the table shows the required cumulative subsidies by \n2050 (Supplementary Fig. 16) and the required average specific subsidies in 2050 (Extended Data Fig. 7, differentiated by green hydrogen and electrofuels). Without carbon pricing, green \nhydrogen and electrofuels require subsidies until 2050 across all end uses, leading to enormous required cumulative subsidies by 2050, as well as substantial average specific subsidies even \nin 2050, which may be required indefinitely for some applications. With ambitious carbon pricing, the required cumulative subsidies by 2050 strongly depend on the scenario.\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n117\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nunderlining the crucial role of demand-side regulation for fostering \ngreen hydrogen growth.\nOur results indicate that permanently subsidizing green hydrogen \nand electrofuels to compete with cheap fossil fuels would likely end up \nbeing prohibitively expensive in the long term, highlighting the key role \nof carbon pricing in closing the cost gap. Without carbon pricing, green \nhydrogen growth in line with the 1.5\u2009\u00b0C scenario median requires annual \nsubsidies that far exceed the historical support of solar PV and wind \n(Extended Data Fig. 7a). In particular, without carbon pricing, green \nhydrogen and electrofuels likely require subsidies until at least 2050 \n(Extended Data Fig. 7c and Table 1). In contrast, under an ambitious \ncarbon price pathway, the required green hydrogen and electrofuel \nsubsidies could remain in the same range historically observed for solar \nPV and wind, with per-megawatt hour subsidies steadily decreasing \nuntil 2050 (Extended Data Fig. 7b,d).\nDiscussion and conclusion\nThe past and future of green hydrogen is characterized by three gaps, \nreflecting the challenges of scaling-up a novel and as yet uncompeti-\ntive energy carrier that requires dedicated policy support. First, the \n2023 implementation gap shows that only 7% of initially announced \ngreen hydrogen capacity was eventually realized. Second, the 2030 \nambition gap has gradually closed over the past 3\u2009years as the project \npipeline increasingly exceeds the requirements in 1.5\u2009\u00b0C scenarios. \nThird however, this has led to a wide 2030 implementation gap as \nenormous subsidies would be required to realize all of the projects \nby 2030, and even more to put green hydrogen on track for 1.5\u2009\u00b0C in \nthe long term.\nThe high past failure rates indicate a limited reliability of project \nannouncements published by industry, which may announce green \nhydrogen projects for strategic reasons, such as raising attention or \nattracting subsidies. Although sobering, this can provide valuable \ninsights for realistic scale-up analyses of green hydrogen31 and other \nlow-carbon energy technologies in feasibility studies44\u201346, some of \nwhich45 have recently faced criticism for lacking statistical rigour47. \nOur results are particularly useful for analyses that use uncertain \nproject announcements as input data25,48. System planners, policy-\nmakers and society should interpret the increasingly steep growth \nsuggested by recent project announcements with caution, focusing \non scale-up challenges, such as lacking competitiveness and the need \nfor policy support.\nTo close the green hydrogen implementation gap, policymakers \nneed to bridge the cost gap to fossil fuels and de-risk hydrogen invest-\nments. This requires a balanced policy mix and a robust strategy to \nnavigate the following three key uncertainties and risks.\nFirst, the huge past and future implementation gaps indicate that \ngreen hydrogen will likely fall short of 1.5\u2009\u00b0C scenarios. Even if policy \nsupport is strengthened, it remains uncertain whether this would be \nsufficient to drive the necessary hydrogen investments. Realizing cur-\nrent project announcements would require unprecedented growth \nrates (Extended Data Fig. 1a,c), exceeding even the fastest-growing \nenergy technology in history, namely, solar PV. Given that green hydro-\ngen technologies are more complex, less standardizable and require \nnew infrastructure, all of which slow down technology diffusion24, \nrealizing such unprecedented growth is unlikely.\nSecond, current hydrogen policy instruments often seek to spur \nhydrogen investments by bridging the cost gap to fossil fuels through \nsupply-side subsidies such as fixed-premium auctions. However, as we \nhave shown, this approach requires not only excessive subsidy volumes \nbut also strong perseverance as policy support could be required for \nseveral decades, or even indefinitely without carbon pricing or strong \ndemand-side regulation. Subsidies for near-term green hydrogen pro-\nduction are often framed within a narrative of kickstarting a \u2018hydrogen \neconomy\u2019 through a short policy push, after which green hydrogen \nbecomes cost-competitive and scales up on its own. However, this \ncritically depends on optimistic assumptions about technology cost \nreductions, which stands in contrast to recent cost increases of electro-\nlysers1. Without ambitious cost reductions, the \u2018kickstarting\u2019 narrative \nis misleading and raises false hopes.\nThird, the primary role of hydrogen in climate change mitigation \nis to replace fossil fuels in hard-to-electrify sectors. However, strong \npolitical support for hydrogen is often accompanied by overconfi-\ndence in its potential15, resulting in conflicting visions about its future \nrole. Many global climate change mitigation scenarios show a modest \nlong-term share of hydrogen of 5\u201315% in final energy2,40,49, focusing on \nkey end uses where hydrogen is highly valuable due to a lack of alterna-\ntives5. In stark contrast, incumbent actors in gas, heat, industry and \ntransport tend to endorse a wide use of hydrogen across sectors50, even \nin end uses such as residential heat, where electrification is cheaper, \nmore efficient and readily available2,40,49,51. Uncertainties remain around \nthe role of hydrogen in complementing the electrification of heavy \ntransport and industrial heat11,35,40.\nDisregarding these uncertainties and risks, and instead focusing \non supply-side subsidies with the expectation of abundant low-cost \ngreen hydrogen in the future, risks crowding out readily available and \nmore economical options, thereby delaying climate change mitiga-\ntion. To minimize these risks while safeguarding the scale-up of green \nhydrogen, we draw two key policy conclusions.\nFirst, supply-side subsidies, which reduce the investment risk \nof electrolysis projects, should be complemented by demand-side \npolicies that guide hydrogen to its most valuable use cases by increas-\ning their willingness to pay. The benefit of demand-side measures is \nillustrated by the European Hydrogen Bank\u2019s recent inaugural auction, \nwhich resulted in surprisingly low successful bids of \u20ac0.37\u20130.48\u2009kg\u22121 \n(ref. 52) compared with a similar auction in the UK, which received only \nhigh bids equivalent to \u20ac9.40\u2009kg\u22121 (ref. 53). Aside from regional hetero-\ngeneity, this stark difference may be attributed to the EU\u2019s demand-side \nquotas, such as the mandatory 42% green hydrogen share of all hydro-\ngen used in industry by 2030 under the Renewable Energy Directive III \n(ref. 54), and mandates for hydrogen-based electrofuels under ReFu-\nelEU Aviation55 and FuelEU Maritime56 regulations. Although they incur \nmacroeconomic costs (Supplementary Note 5), demand-side policies \ncan reduce the pressure on supply-side subsidies, helping to close the \nimplementation gap.\nSecond, policymakers should plan the transition from subsidies \nto market mechanisms. In the short run, achieving rapid near-term \nhydrogen growth is crucial to keep 1.5\u2009\u00b0C scenarios within reach. This \nrequires strong policies, such as subsidies to directly bridge the cost \ngap, minimize investment risks and initialize a hydrogen market. How-\never, as hydrogen technologies and markets mature, policy support \nshould shift to market-based mechanisms to (1) reduce policy costs, (2) \nreveal the full hydrogen costs to markets and consumers, and (3) create \na level playing field with other mitigation options. The most important \ntechnology-neutral strategy is ambitious carbon pricing. However, as \ncarbon prices are currently too low and too uncertain in the future, \ncomplementary instruments are required to de-risk the remaining \nuncertainties. These include technology-neutral auctions of carbon \ncontracts for difference57, which hedge investors against unpredict-\nable prices by covering the difference between emissions abatement \ncosts and carbon prices, as well as tradable, technology-neutral quotas \nfor, for example, low-carbon materials, fostering green lead markets.\nIn summary, a comprehensive policy strategy for green hydro-\ngen should include targeted demand-side measures and a gradual \ntransition from subsidies to market mechanisms. In the short term, \nthis would de-risk early investment at manageable costs, guiding \nhydrogen to its most valuable use cases. In the long term, this would \ntransfer investment risks and competition between hydrogen and \nother mitigation options to the market, thereby establishing a cred-\nible commitment for climate change mitigation while spurring green \nhydrogen growth.\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n118\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nMethods\nOverview\nOur approach was split into three parts. First, we tracked green \nhydrogen project announcements to quantify the green hydrogen \nimplementation gap in 2022 and 2023. Second, we compared project \nannouncements with 1.5\u2009\u00b0C scenarios to show the 2030 green hydrogen \nambition gap. Third, we modelled the pay-as-bid market premium and \nestimated required subsidies using a competitiveness analysis of four \ngreen products and five fossil competitors across 14 end uses, which \nled to the 2030 green hydrogen implementation gap.\nGreen hydrogen projects database\nWe used data of electrolysis project announcements from the IEA \nHydrogen Production Projects and Infrastructure Database58 (previ-\nously called the IEA Hydrogen Projects Database), incorporating three \ndatabase snapshots from 2021, 2022 and 2023. We only included pro-\nject announcements for electrolysers that included a year of project \nlaunch, had a meaningful status (not \u2018Other\u2019 or \u2018Other/Unknown\u2019) and \nreported a capacity value. We did not filter for the type of electricity as \nthis was often unknown. These criteria led to 612 projects in the 2021 \nsnapshot, 877 projects in the 2022 snapshot and 1,265 projects in the \n2023 snapshot. In the 2023 snapshot, only a single status category was \nreported for projects that were either under construction or had an FID \n(\u2018FID/Construction\u2019). To ensure consistent status categories across all \nsnapshots, we merged the \u2018FID\u2019 and \u2018Under construction\u2019 categories \nin the 2021 and 2022 snapshots. Projects with a \u2018DEMO\u2019 status were \nallocated as \u2018Operational\u2019, \u2018FID/Construction\u2019 or \u2018Decommissioned\u2019, \ndepending on whether they were still running, announced for the future \nor had been decommissioned, respectively. We note that the \u2018Concept\u2019 \ncategory is very broadly defined with an unspecified credibility bar \nfor inclusion, while the \u2018Feasibility study\u2019 category may also contain \nprojects for which a feasibility study is planned, but has not yet started. \nConfidential projects were distributed to all regions in proportion to \nthe share of capacity from non-confidential projects, but could not be \ntracked across database snapshots.\nData quality validation\nWe conducted a comprehensive, structured and fully documented \ndata quality validation of the green hydrogen project announcements, \nmanually validating 524 project entries across all three database ver-\nsions. For projects announced for 2022 or 2023, we covered at least \n90% of the announced capacity, while for projects announced for \n2024\u20132030, we covered at least 75% of the announced capacity in \nall three database versions (Supplementary Table 1). In addition, we \nmanually verified the fate of all projects announced to launch in 2023 \nin the database published in October 2023 (Fig. 3). Note that we did not \nattempt to identify missing projects, implying that the success rate \nmay change if projects that were realized in 2023 were missing from \nthe most recent database version included in this analysis, published \nin October 2023. During the data validation, we adjusted the size of a \nproject if it was not operating at its nameplate capacity, which was the \ncase for the world\u2019s largest green hydrogen project, Sinopec Kuqa in \nChina. The data quality validation procedure is described in detail in \nSupplementary Note 1.\nTracking green hydrogen projects\nEach project has a unique reference number that stays the same \nacross all database snapshots, as confirmed by the IEA in personal \ncorrespondence. This enabled us to track the development of pro-\nject announcements over time (see Fig. 3 for projects announced for \n2023, Supplementary Fig. 5 for projects announced for 2022 and Sup-\nplementary Fig. 6 for projects announced for 2024). Supplementary \nFigs. 7\u201310 also show the 2023 project tracking for those regions that \nhave at least ten trackable project entries. We accounted for chang-\ning capacity of projects between two database snapshots by adding \ndummy projects, which are, however, not explicitly shown in the Sankey \ndiagrams for simplicity. The reported rates of disappearance, delay and \nsuccess (Fig. 3b\u2013d and Supplementary Fig. 5b, c) only refer to projects \nannounced in 2021, 2022 and 2023, respectively.\nGreen hydrogen in 1.5\u2009\u00b0C scenarios\nAs an indicator of green hydrogen requirements in stringent climate \nmitigation scenarios, we collected electrolysis capacity values from \na wide range of 1.5\u2009\u00b0C scenarios, including (1) IAM scenarios and (2) \ninstitutional and corporate scenarios (Extended Data Fig. 1). For the \nIAM scenarios, we used the IPCC AR6 Scenarios Database59 (category \nC1) as well as the Network for Greening the Financial System (NGFS) \ndataset60 (Version 4.2, the \u2018Net Zero 2050\u2019 and \u2018Low demand\u2019 scenarios). \nWe excluded IAM scenarios that always report zero electrolysis capac-\nity (or zero electrolytic hydrogen production) or, in any period from \n2025, report a value that is lower than the operational electrolysis \ncapacity in 2023. We also omitted scenarios from the NGFS project \nthat included climate damages as this is only reported by one model. \nFor the institutional and corporate scenarios, due to limited report-\ning of numerical data in text or tables, in some cases we resorted to \nextracting data from graphics using WebPlotDigitizer, which has been \nshown to be reliable61. All datasets are available via GitHub (see the \nData availability statement). If electrolysis capacity was not directly \nreported, we converted production quantities into electrolysis capac-\nity, assuming 3,750\u2009full load hours, 69% efficiency and the lower heating \nvalue of hydrogen, 33.33\u2009kWh\u2009kg\u22121. For IAM scenarios, we transformed \nthe reported hydrogen output capacity to the corresponding input \ncapacity of the electrolyser using the efficiency of 69%. Due to these \napproximations, reported electrolysis requirements in 1.5\u2009\u00b0C scenarios \nare inherently uncertain.\nModelling pay-as-bid market premiums\nTo quantify the future green hydrogen implementation gap, we devel-\noped a model of the required pay-as-bid market premiums for green \nhydrogen projects (Extended Data Fig. 3). First, we mapped each of the \n14 end-use categories from the green hydrogen projects database to \nthe competition between a green product and a fossil competitor, cov-\nering four green products (green hydrogen, e-methanol, e-kerosene \nand e-methane) and five fossil competitors (grey hydrogen, natural \ngas, grey methanol, diesel and kerosene), as shown in Extended Data \nTable 1. For projects without a designated end use, we assumed that \ngreen hydrogen competes with natural gas. Second, we calculated the \nlevelized cost of all green products (Extended Data Table 2) and the \nprices of all fossil competitors with and without an ambitious carbon \nprice pathway that is in line with EU climate targets41 (Extended Data \nTable 3). Details on these costs and prices are explained in the fol-\nlowing sections. Third, we incorporated demand-side policies such \nas end-use quotas, which increase the willingness to pay for green \nproducts and thereby reduce required policy costs (Supplementary \nFig. 15). Finally, for each end use, we estimated the required subsi-\ndies based on (1) vintage tracking of project announcements and (2) \nthe cost gap between the green product and the fossil competitor \n(Extended Data Fig. 3).\nWe included global estimates of implemented demand-side poli-\ncies in 2030 across four end uses, provided by the IEA1, which we con-\nverted into the corresponding electrolysis capacities using the lower \nheating value, as well as the full load hours and efficiencies of the \nrespective scenario. We proportionally distributed these estimates \nof electrolysis capacity that are supported by demand-side regulation \nin 2030 according to the project announcements from 2024\u20132030 \n(Supplementary Fig. 15).\nIf the capacity supported by demand-side policies exceeded the \nannounced capacity, which is the case for refining and synthetic fuels, \nwe omitted the difference, assuming that demand-side policies are \nend-use specific.\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n119\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nTo estimate the required annual subsidies, we combined these \ncomponents. As shown in Fig. 5a\u2013d and Extended Data Fig. 5, for each \nend use, the instantaneous cost gap (\u0394pt) between the levelized cost \nof the green product in year t (LCOXt) and the price of the fossil com-\npetitor (pfossil\nt\n) is given as:\n\u0394pt = LCOXt \u2212pfossil\nt\n(1)\nHowever, this cannot be used directly to estimate subsidies. As \nillustrated in Extended Data Fig. 3, a green hydrogen or electrofuel \nproject completed in year t\u2032 must sell the green product at LCOXt\u2032 for \nthe entire duration of the payback period \u03c4 to recover its costs. The \nrequired annual subsidies accumulate over time due to projects built \nin previous years. For example, in 2026, projects that were built in 2024 \nface a cost gap of LCOX2024 \u2212pfossil\n2026, projects that were built in 2025 face \na cost gap of LCOX2025 \u2212pfossil\n2026 and projects that were built in 2026 face \na cost gap of LCOX2026 \u2212pfossil\n2026. These cost gaps have to be bridged for \nthe electrolysis capacity built in the corresponding year t\u2032, denoted as \n\u0394Ct\u2032 (accounting for capacity supported by demand-side policies). For \neach end use, with electrolyser full load hours FLHH2, electrolyser effi-\nciency \u03b7H2 and payback period \u03c4, the required annual subsidy (Sannual\nt\n) \nin year t is given as:\nSannual\nt\n=\nt\n\u2211\nt\u2032=max{2024,t\u2212\u03c4+1}\n\u0394Ct\u2032 \u00d7 FLHH2,t\u2032 \u00d7 \u03b7H2,t\u2032 \u00d7 max {0, LCOXt\u2032 \u2212pfossil\nt\n}\n(2)\nNote that for subsidies in year t, only the price of the fossil competi-\ntor (pfossil\nt\n) refers to the same year t, whereas all other parameters refer \nto the year t\u2032 in which the project was built. Thus, the realization of \ngreen hydrogen projects built in the year t\u2032 requires subsidy payments \nfor the full payback period [t\u2032, t\u2032 + \u03c4) as long as LCOXt\u2032 > pfossil\nt\n. For end \nuses where the green product and the fossil competitor are not used \nthermally, we included the relative efficiency improvement of using \nthe green product over the fossil competitor, \u03b7green\nLHV /\u03b7fossil\nLHV , adjusting the \nLCOX accordingly (Extended Data Table 1). Note that for green hydro-\ngen, we denote LCOX as LCOH. Correspondingly, the required cumula-\ntive subsidies until year t (Scumulative\nt\n) are given by:\nScumulative\nt\n=\nt\n\u2211\nt\u2032=2024\nSannual\nt\u2032\n(3)\nWe show in Fig. 5e\u2013g and Extended Data Fig. 6 the required annual \nand cumulative subsidies as the sum over all end uses.\nTo analyse what would be required for a 1.5\u2009\u00b0C scenario, after 2030 \nwe used the median of the institutional and corporate 1.5\u2009\u00b0C scenarios \nfor \u0394Ct\u2032 (Extended Data Fig. 1b and Supplementary Fig. 11). To determine \nthe sectoral allocation of the overall capacity to the 14 end uses after \n2030, we used the green hydrogen end-use shares of the IEA NZE Sce-\nnario40 (Supplementary Fig. 13). The results for this 1.5\u2009\u00b0C scenario until \n2050 are presented in Supplementary Fig. 16.\nLevelized costs of green products\nFor all green products, we first calculated LCOH for each year from 2024 \nusing the annuity method and broadly following the system boundaries \noutlined in ref. 62 (for the parameters, see Extended Data Table 2), but \nadding end-use-specific transport and storage costs (Supplementary \nTable 2). Omitting time indices, the LCOH was calculated as:\nLCOH =\n1\n\u03b7H2\n{ [a (r, \u03c4) + FOMH2]\nIBOP\nFLHH2\n+ [a (r, \u03c4stack)\n+FOMH2]\nIstack\nFLHH2\n+ pelec} + VOMH2\n(4)\nwhere \u03b7H2 denotes the electrolyser efficiency, a(r, \u03c4) =\nr\n1\u2212(1+r)\n\u2212\u03c4  is the \nannuity factor, r is the cost of capital, \u03c4 is the payback period in years \n(which can be shorter than the technical lifetime), \u03c4stack is the lifetime \nof the electrolyser stack in years, FOMH2 is the fixed operation and \nmaintenance costs as a percentage of the specific investment costs, \nIBOP is the specific investment cost of the electrolyser\u2019s balance of plant \n(BOP) and other engineering work, Istack is the specific investment cost \nof the electrolyser stack, FLHH2 is the electrolysis full load hours, pelec is \nthe price of electricity and VOMH2 is the variable operation and main-\ntenance costs, which are transport and storage costs (Supplementary \nTable 2). Both IBOP and Istack relate to the electrical input capacity of the \nelectrolyser (US$\u2009kWel\n\u22121).\nThe electricity price paid by electrolysers is highly dependent on \nthe specific supply case and the regulatory definition of green hydro-\ngen with respect to spatio-temporal matching and additionality28,29. \nFlexible operation and a direct connection to a renewable energy \nsource reduces the price as electrolysers can tap into hours when \nelectricity is cheap and abundant. Grid-connected electrolysers need \nto pay grid fees on top of electricity prices, but can run at higher full \nload hours. Furthermore, stationary batteries can extend the electro-\nlyser\u2019s full load hours by providing a buffer for renewable electricity, \nbut require additional investments. While hourly energy system models \ncan represent these effects in detail28, we accounted for them in an \naggregated manner by using the same broad range of electricity prices \nas in ref. 27. This ensures high traceability of results, while still captur-\ning the effects of system heterogeneity. Further discussion is provided \nin Supplementary Note 2, while Supplementary Note 3 discusses how \nenergy system models could learn from our results.\nWe separated the total specific investments costs of the electro-\nlyser (I) into Istack and IBOP because (1) the stack needs to be replaced ear-\nlier than the rest of the electrolyser, such that we included two annuities \nin equation (4)62, and (2) the stack is much more modular and therefore \nmore susceptible to cost improvements17, which we included through \ndifferent learning rates. Technological learning reduces specific invest-\nment costs of both IBOP and Istack in year t (It) according to\nIt = I2023(\nCt\nC2023\n)\nlog2(1\u2212LR)\n(5)\nwhere I2023 denotes the investment costs in 2023, Ct denotes the global \ncumulative electrolysis capacity in year t, C2023\u2009=\u20090.92\u2009GW installed \ncapacity in 2023 and LR denotes the learning rate. Technological learn-\ning is driven by cumulative project announcements until 2030 and \nsubsequently by the median 1.5\u2009\u00b0C scenario (Supplementary Fig. 11). \nThus, electrolyser costs fall quickly (Supplementary Fig. 12).\nFor electrofuels derived from green hydrogen (e-kerosene, \ne-methanol and e-methane), the corresponding LCOX are\nLCOX = [a (r, \u03c4) + FOMX]\nIX\nFLHX\n+\npH2\n\u03b7X\n+ pCO2\u03b5X + VOMX\n(6)\nwhere FOMX represents fixed operation and maintenance costs, IX is \nthe specific investment cost of the electrofuel synthesis plant (in terms \nof electrofuel output), FLHX is the full load hours of the synthesis plant, \npH2 = LCOH \u2212VOMH2 is the price of hydrogen (that is, the LCOH without \ntransport and storage costs), \u03b7X is the synthesis energy efficiency, pCO2 \nis the price of renewable CO2 (not the carbon price of emissions), \u03b5X is \nthe CO2 intensity of the electrofuel and VOMX is the end-use-specific \ntransport and storage costs (Supplementary Table 2).\nThe price of renewable CO2, which can either come from bio-\ngenic sources or from direct air capture, is an uncertain but important \ncost component for the production of carbon-neutral electrofuels \n(Extended Data Fig. 5g\u2013l). While biogenic carbon can initially be as \ncheap as US$30\u2009tCO2\n\u22121, it likely faces availability limits such that it could \nquickly become more expensive as demand increases (see, for example, \nFig. 6.3 in ref. 63). In contrast, direct air capture is more scalable, but \ncurrently faces very high costs in the order of US$500\u20131,000\u2009tCO2\n\u22121, \n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n120\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nwhich could reduce to approximately US$300\u2009tCO2\n\u22121 once the scale of \n1\u2009GtCO2\u2009yr\u22121 is reached in the long term64, although this is again subject \nto substantial uncertainty. In our central estimate, we set the average \ncost of renewable carbon to US$200\u2009tCO2\n\u22121, which reflects the differ-\nent CO2 sources reported in electrofuel projects, while the progres-\nsive and conservative sensitivity scenarios covered a wide range of \nUS$30\u2013300\u2009tCO2\n\u22121.\nPrices of fossil competitors\nWe collected harmonized data on prices for all fossil competitors rep-\nresented in our pay-as-bid market premium model for 2024, 2030 and \n2050 (for parameters, see Extended Data Table 3), using linear interpola-\ntion in between. For natural gas, our cost estimate was the average of \nthe EU trading point Title Transfer Facility in the Netherlands and the US \ntrading point Henry Hub, using spot market prices in 2024 and future \nprices in 2030. For 2050, we used the gas price from the IEA NZE 1.5\u2009\u00b0C \nscenario40. For grey hydrogen and grey methanol, which are produced \nfrom natural gas, we first collected current prices for 2024. To ensure \ninternal consistency with natural gas prices, we then calculated the cor-\nresponding specific fixed costs in 2024, which reflect the per-megawatt \nhour capital costs associated with the synthesis plant. Assuming that \nthese stay constant, for 2030 and 2050 we inferred the price of grey \nhydrogen and grey methanol by adding the corresponding variable \ncosts, that is, the natural gas price divided by the efficiency. We pro-\nceeded similarly for kerosene and diesel, using crude oil spot and future \nprices as the reference for 2024 and 2030, respectively, while for 2050 \nwe again used the oil price from the IEA NZE 1.5\u2009\u00b0C scenario. This calibra-\ntion ensured that prices for fossil products are internally consistent.\nLast, we differentiated between scenarios without and with ambi-\ntious carbon pricing. For the latter, we used a carbon price pathway \nthat is in line with EU climate targets in the sectors covered by the EU \nEmissions Trading System, such as industry and energy supply41. The \nCO2 cost per megawatt hour of the fossil competitor is the product of \nthe emissions intensity, including upstream methane emissions for \nnatural gas, grey hydrogen and grey methanol27, and the carbon price \nper tonne of CO2. We denote the total cost as pfossil, which includes CO2 \ncosts if applicable. In addition, for natural gas, we considered grid fees \nof US$5\u2009MWh\u22121 based on ref. 65 (Supplementary Table 2).\nLimitations\nAs the quality of the data of the IEA Hydrogen Production and Infra-\nstructure Projects Database58 may be limited, we conducted a com-\nprehensive data validation (see the section \u2018Data quality validation\u2019, \nSupplementary Note 1, Supplementary Table 1 and Supplementary \nFigs. 1\u20134). Nevertheless, some errors may remain, particularly for \nsmaller projects that were not checked. In general, there are counter-\nacting uncertainties related to project announcements. On the one \nhand, the database may underestimate projects, as we verified only \nexisting entries and did not conduct research to identify potentially \nmissing projects. On the other hand, the database may include projects \nthat are no longer active, as it is often unclear if and when a project has \nbeen scrapped.\nThe quality of the data of the electrolysis requirements in 1.5\u2009\u00b0C \nscenarios is limited due to heterogeneous sources and limited numeri-\ncal reporting of the scenario data accompanying the reports. In sev-\neral cases, we had to infer electrolysis capacity from green hydrogen \nproduction values, also for IAM scenarios. Thus, Fig. 4 and Extended \nData Fig. 1 show only estimates of electrolysis capacity using publicly \navailable data and should not be interpreted as numerically exact.\nModelling the pay-as-bid market premium to estimate subsidies \nrequired several simplifications. First, although we distinguished \nbetween 14 end-use applications, four green products and five fossil \ncompetitors, we did not account for regional differences in hydrogen \nproduction costs. Our estimates can be interpreted as cross-regional \naverages. Note that our sensitivity ranges are large enough to contain \nthe regional cost heterogeneity found in GIS-based analyses66. Sec-\nond, we neglected additional end-use transformation costs, which \nare typically small or even zero, for example, for drop-in electrofu-\nels. Some applications can simply replace grey with green hydrogen \nwith no additional costs (for example, ammonia production), while \nadditional investment costs in other applications are low compared \nwith fossil applications (for example, direct reduced iron-based steel \nplants or hydrogen boilers). Third, we calculated levelized costs using \nconstant electricity prices, assuming that green hydrogen projects \nrequire new dedicated renewable energy plants or long-term con-\ntracted power-purchase agreements that deliver electricity at stable \nprices. Similarly, for electrofuels, this implies dedicated electrolysers \nor long-term contracts that deliver green hydrogen at constant prices. \nFourth, we did not consider the option that projects could pay back a \npart of the received subsidies once they are profitable relative to their \nfossil competitor in the future because this would require a contract \nfor differences that allows for this option. Fifth, we did not include \nfactors other than costs that influence the project realization as this \nwas outside the scope of this analysis. Sixth, we did not incorporate \nthe competition of green hydrogen with blue hydrogen and other \nmitigation options, which we discuss in Supplementary Note 4. Last, we \nassumed that demand-side policies directly translate into electrolysis \ncapacity without the need for additional subsidies.\nThe quality of the data of global announced hydrogen subsidies \nfrom BloombergNEF (BNEF) may be limited and will likely soon be \noutdated. The estimate for US subsidies is particularly uncertain as the \nproduction tax credits of the Inflation Reduction Act12 are uncapped \nsuch that BNEF bases their US subsidy estimates on hydrogen pro-\nject announcements. Furthermore, the tracked subsidies cover not \nonly green hydrogen but also other sources of low-carbon hydrogen, \nwhich we optimistically compared to subsidy requirements only for \ngreen hydrogen project announcements. The global subsidy volume \nof US$308 billion for low-carbon hydrogen as of September 2023 \ntherefore serves only as a snapshot. Although this figure will be out-\ndated soon, it still offers a valuable reference point. However, it should \nbe interpreted with caution as the implementation of these subsidies \nwill critically depend on future government commitments to foster \nthe hydrogen market ramp-up.\nData availability\nAll of the data are publicly available via GitHub at https://github.com/ \naodenweller/green-hydrogen-gap and via Zenodo at https://doi. \norg/10.5281/zenodo.14041796 (ref. 67). All of the data files include a \ncolumn that indicates the original source.\nCode availability\nThe R model code used to perform the analyses and produce all figures \nis available via GitHub at https://github.com/aodenweller/green-hydro \ngen-gap and via Zenodo at https://doi.org/10.5281/zenodo.14041796 \n(ref. 67).\nReferences\n1.\t\nGlobal Hydrogen Review 2023 (IEA, 2023); https://www.iea.org/ \nreports/global-hydrogen-review-2023\n2.\t\nWorld Energy Transitions Outlook 2023: 1.5\u2009\u00b0C Pathway (IRENA, \n2023); https://www.irena.org/Publications/2023/Jun/World- \nEnergy-Transitions-Outlook-2023\n3.\t\nClarke, L. et al. Climate Change 2022: Mitigation of Climate \nChange (eds Shukla, P. A. et al.) 613\u2013746 (IPCC, Cambridge Univ. \nPress, 2022); https://doi.org/10.1017/9781009157926.008\n4.\t\nDavis, S. J. et al. Net-zero emissions energy systems. Science \nhttps://doi.org/10.1126/science.aas9793 (2018).\n5.\t\nWolfram, P., Kyle, P., Fuhrman, J., O\u2019Rourke, P. & McJeon, H. \nThe hydrogen economy can reduce costs of climate change \nmitigation by up to 22%. One Earth 7, 885\u2013895 (2024).\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n121\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\n6.\t\nHydrogen Insights 2023 (Hydrogen Council and McKinsey & \nCompany, 2023); https://hydrogencouncil.com/en/hydrogen- \ninsights-2023-december-update/\n7.\t\nComparative Analysis of International Hydrogen Strategies, \nCountry Analysis 2023 (DECHEMA and acatech, 2024); https:// \nwww.wasserstoff-kompass.de/fileadmin/user_upload/img/news- \nund-media/dokumente/2023_e_H2_Laenderanalyse.pdf\n8.\t\nHunter, C. A. et al. Techno-economic analysis of long-duration \nenergy storage and flexible power generation technologies to \nsupport high-variable renewable energy grids. Joule 5, 2077\u20132101 \n(2021).\n9.\t\nDowling, J. A. et al. Role of long-duration energy storage in \nvariable renewable electricity systems. Joule 4, 1907\u20131928  \n(2020).\n10.\t Dieterich, V., Buttler, A., Hanel, A., Spliethoff, H. & Fendt, S. \nPower-to-liquid via synthesis of methanol, DME or Fischer\u2013\nTropsch-fuels: a review. Energy Environ. Sci. 13, 3207\u20133252 \n(2020).\n11.\t\nUeckerdt, F. et al. Potential and risks of hydrogen-based e-fuels in \nclimate change mitigation. Nat. Clim. Change 11, 384\u2013393 (2021).\n12.\t 117th United States Congress Inflation Reduction Act (US \nGovernment, 2022).\n13.\t Communication from the Commission to the European Parliament, \nthe European Council, the Council, the European Economic and \nSocial Committee of the Regions on the European Hydrogen Bank \n(European Commission, 2023).\n14.\t van Renssen, S. The hydrogen solution? Nat. Clim. Change 10, \n799\u2013801 (2020).\n15.\t Overhyping hydrogen as a fuel risks endangering net-zero goals. \nNature 611, 426 (2022).\n16.\t Reducing Low-Carbon Hydrogen Investment and Operating Costs. \n(Capgemini and EIT InnoEnergy, 2024); https://www.capgemini. \ncom/insights/research-library/reducing-low-carbon-hydrogen- \ninvestment-and-operating-costs/\n17.\t Achieving Affordable Green Hydrogen Production Plants (Ramboll, \n2023); https://www.ramboll.com/net-zero-explorers/what-will-it- \ntake-to-reduce-capex-in-green-hydrogen-production\n18.\t Shafiee, R. T. & Schrag, D. P. Carbon abatement costs of green \nhydrogen across end-use sectors. Joule https://doi.org/10.1016/j. \njoule.2024.09.003 (2024).\n19.\t Hydrogen offtake is tiny but growing. BloombergNEF https:// \nabout.bnef.com/blog/hydrogen-offtake-is-tiny-but-growing/ \n(2023).\n20.\t Nemet, G. F., Zipperer, V. & Kraus, M. The valley of death, \nthe technology pork barrel, and public support for large \ndemonstration projects. Energy Policy 119, 154\u2013167 (2018).\n21.\t Lessons from a Decade of Emissions Gap Assessments (United \nNations Environment Programme, 2019); https://www. \nunenvironment.org/resources/emissions-gap-report-10-year- \nsummary\n22.\t Lamb, W. F. et al. The carbon dioxide removal gap. Nat. Clim. \nChange https://doi.org/10.1038/s41558-024-01984-6 (2024).\n23.\t Flyvbjerg, B. & Gardner, D. How Big Things Get Done: The \nSurprising Factors That Determine the Fate of Every Project \nfrom Home Renovations to Space Exploration, and Everything in \nBetween (Penguin Random House, 2023).\n24.\t Malhotra, A. & Schmidt, T. S. Accelerating low-carbon innovation. \nJoule 4, 2259\u20132267 (2020).\n25.\t Kazlou, T., Cherp, A. & Jewell, J. Feasible deployment of carbon \ncapture and storage and the requirements of climate targets.  \nNat. Clim. Change 14, 1047\u20131055 (2024).\n26.\t Wilson, C. et al. Granular technologies to accelerate \ndecarbonization. Science 368, 36\u201339 (2020).\n27.\t Ueckerdt, F. et al. On the cost competitiveness of blue and green \nhydrogen. Joule 8, 104\u2013128 (2024).\n28.\t Zeyen, E., Riepin, I. & Brown, T. Temporal regulation of renewable \nsupply for electrolytic hydrogen. Environ. Res. Lett. 19, 024034 \n(2024).\n29.\t Ricks, W., Xu, Q. & Jenkins, J. D. Minimizing emissions from \ngrid-based hydrogen production in the United States. Environ. \nRes. Lett. 18, 014025 (2023).\n30.\t Quarton, C. J. et al. The curious case of the conflicting roles of \nhydrogen in global energy scenarios. Sustain. Energy Fuels 4, \n80\u201395 (2019).\n31.\t Odenweller, A., Ueckerdt, F., Nemet, G. F., Jensterle, M. &  \nLuderer, G. Probabilistic feasibility space of scaling up green \nhydrogen supply. Nat. Energy 7, 854\u2013865 (2022).\n32.\t M\u00fcller-Casseres, E. et al. International shipping in a world below \n2\u2009\u00b0C. Nat. Clim. Change https://doi.org/10.1038/s41558-024-01997- \n1 (2024).\n33.\t Dray, L. et al. Cost and emissions pathways towards net-zero \nclimate impacts in aviation. Nat. Clim. Change 12, 956\u2013962 (2022).\n34.\t Rissman, J. et al. Technologies and policies to decarbonize global \nindustry: review and assessment of mitigation drivers through \n2070. Appl. Energy 266, 114848 (2020).\n35.\t Schreyer, F. et al. Distinct roles of direct and indirect \nelectrification in pathways to a renewables-dominated European \nenergy system. One Earth 7, 226\u2013241 (2024).\n36.\t Bachorz, C., Verpoort, P., Ueckerdt, F. & Luderer, G. Exploring \ntechno-economic landscapes of abatement options for \nhard-to-electrify sectors. Preprint at Research Square  \nhttps://doi.org/10.21203/rs.3.rs-4241841/v1 (2024).\n37.\t Mignone, B. K. et al. Drivers and implications of alternative  \nroutes to fuels decarbonization in net-zero energy systems.  \nNat. Commun. 15, 3938 (2024).\n38.\t Net Zero by 2050 (IEA, 2021); https://www.iea.org/reports/net- \nzero-by-2050\n39.\t World Energy Outlook 2022 (IEA, 2022); https://www.iea.org/ \nreports/world-energy-outlook-2022\n40.\t Net Zero Roadmap: A Global Pathway to Keep the 1.5\u2009\u00b0C Goal in \nReach (IEA, 2023); https://www.iea.org/reports/net-zero- \nroadmap-a-global-pathway-to-keep-the-15-0c-goal-in-reach\n41.\t Sitarz, J., Pahle, M., Osorio, S., Luderer, G. & Pietzcker, R. EU \ncarbon prices signal high policy credibility and farsighted actors. \nNat. Energy https://doi.org/10.1038/s41560-024-01505-x (2024).\n42.\t Clean Hydrogen Monitor 2023 (Hydrogen Europe, 2023);  \nhttps://hydrogeneurope.eu/wp-content/uploads/2023/10/Clean_ \nHydrogen_Monitor_11-2023_DIGITAL.pdf\n43.\t BNEF hydrogen outlook. BloombergNEF https://www.pv- \nmagazine.com/wp-content/uploads/2022/01/Session-9_ \nPresentation-02_Adithya-Bhashyam-BNEF.pdf (2023).\n44.\t Jewell, J. & Cherp, A. The feasibility of climate action: bridging  \nthe inside and the outside view through feasibility spaces.  \nWiley Interdiscip. Rev. Clim. Change 14, e838 (2023).\n45.\t Cherp, A., Vinichenko, V., Tosun, J., Gordon, J. A. & Jewell, J. \nNational growth dynamics of wind and solar power compared \nto the growth required for global climate targets. Nat. Energy 6, \n742\u2013754 (2021).\n46.\t Zielonka, N. & Trutnevyte, E. Probabilities of reaching required \ndiffusion of granular energy technologies in European countries. \nPreprint at Research Square https://doi.org/10.21203/ \nrs.3.rs-4039857/v2 (2024).\n47.\t Baumg\u00e4rtner, C. L., Way, R., Ives, M. C. & Farmer, J. D. The need \nfor better statistical testing in data-driven energy technology \nmodeling. Joule 8, 2453\u20132466 (2024).\n48.\t Edwards, M. R. et al. Modeling direct air carbon capture and \nstorage in a 1.5\u2009\u00b0C climate future using historical analogs.  \nProc. Natl Acad. Sci. USA 121, e2215679121 (2024).\n49.\t New Energy Outlook 2024 (BloombergNEF, 2024); https://about. \nbnef.com/new-energy-outlook/\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n122\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\n50.\t Ohlendorf, N., L\u00f6hr, M. & Markard, J. Actors in multi-sector \ntransitions\u2014discourse analysis on hydrogen in Germany. Environ. \nInnov. Soc. Transit. 47, 100692 (2023).\n51.\t Rosenow, J. Is heating homes with hydrogen all but a pipe dream? \nAn evidence review. Joule https://doi.org/10.1016/j. \njoule.2022.08.015 (2022).\n52.\t European Hydrogen Bank auction provides \u20ac720 million. \nEuropean Commission (30 April 2024); https://ec.europa.eu/ \ncommission/presscorner/detail/en/IP_24_2333\n53.\t Hydrogen Production Business Model/Net Zero Hydrogen Fund: \nHAR1 successful projects. UK Department for Energy  \nSecurity and Net Zero (2023); https://www.gov.uk/government/\npublications/hydrogen-production-business-model-net- \nzero-hydrogen-fund-shortlisted-projects/hydrogen- \nproduction-business-model-net-zero-hydrogen-fund-ha\nr1-successful-projects\n54.\t Directive (EU) 2023/2413 of the European Parliament and of the \nCouncil of 18 October 2023 Amending Directive (EU) 2018/2001, \nRegulation (EU) 2018/1999 and Directive 98/70/EC as Regards \nthe Promotion of Energy from Renewable Sources, and Repealing \nCouncil Directive (EU) 2015/652 (European Parliament and Council \nof the European Union, 2023).\n55.\t Regulation (EU) 2023/2405 of the European Parliament and of \nthe Council of 18 October 2023 on Ensuring a Level Playing Field \nfor Sustainable Air Transport (ReFuelEU Aviation) (European \nParliament and Council of the European Union, 2023).\n56.\t Regulation (EU) 2023/1805 of the European Parliament and of \nthe Council of 13 September 2023 on the Use of Renewable and \nLow-Carbon Fuels in Maritime Transport, and Amending Directive \n2009/16/EC (Text with EEA Relevance) OJ L Vol. 234 (European \nParliament and Council of the European Union, 2023).\n57.\t Richstein, J. C. & Neuhoff, K. Carbon contracts-for-difference: \nhow to de-risk innovative investments for a low-carbon industry? \niScience 25, 104700 (2022).\n58.\t Hydrogen Production and Infrastructure Projects Database  \n(IEA, 2023).\n59.\t Byers, E. et al. AR6 Scenarios Database (IPCC, 2022); https://doi. \norg/10.5281/zenodo.7197970\n60.\t Richters, O. et al. NGFS Climate Scenarios Data Set Version 4.2 \n(NGFS, 2024); https://doi.org/10.5281/zenodo.10807824\n61.\t Drevon, D., Fursa, S. R. & Malcolm, A. L. Intercoder reliability  \nand validity of WebPlotDigitizer in extracting graphed data.  \nBehav. Modif. 41, 323\u2013339 (2017).\n62.\t Levelised Cost of Hydrogen. Making the Application of the LCOH \nConcept More Consistent and More Useful (Agora Industry and \nUmlaut, 2023); https://www.agora-energiewende.org/ \npublications/levelised-cost-of-hydrogen\n63.\t The Role of E-Fuels in Decarbonising Transport (IEA, 2023);  \nhttps://www.iea.org/reports/\nthe-role-of-e-fuels-in-decarbonising-transport\n64.\t Sievert, K., Schmidt, T. S. & Steffen, B. Considering technology \ncharacteristics to project future costs of direct air capture. Joule \nhttps://doi.org/10.1016/j.joule.2024.02.005 (2024).\n65.\t George, J. F., M\u00fcller, V. P., Winkler, J. & Ragwitz, M. Is blue \nhydrogen a bridging technology? The limits of a CO2 price  \nand the role of state-induced price components for green \nhydrogen production in Germany. Energy Policy 167,  \n113072 (2022).\n66.\t Pfennig, M. et al. Global GIS-based potential analysis and cost \nassessment of Power-to-X fuels in 2050. Appl. Energy 347, \n121289 (2023).\n67.\t Odenweller, A. Green-hydrogen-gap: v1.1.1. Zenodo  \nhttps://doi.org/10.5281/zenodo.14041796 (2024).\n68.\t Renewables 2023 (IEA, 2024); https://www.iea.org/reports/ \nrenewables-2023\n69.\t Fuel Cell Electric Tractor-Trailers: Technology Overview and Fuel \nEconomy (ICCT, 2022); https://theicct.org/publication/fuel-cell- \ntractor-trailer-tech-fuel-jul22/\n70.\t Jacobasch, E. et al. Economic evaluation of low-carbon \nsteelmaking via coupling of electrolysis and direct reduction.  \nJ. Clean. Prod. 328, 129502 (2021).\n71.\t Water Electrolyzer Stack Degradation (EPRI, 2022); https://www. \nepri.com/research/products/000000003002025148\n72.\t Green Hydrogen Cost Reduction: Scaling up Electrolysers  \nto Meet the 1.5\u2009\u00b0C Climate Goal (IRENA, 2020); https://www.irena.\norg/publications/2020/Dec/Green-hydrogen-cost-reduction\n73.\t The Future Cost of Electricity-Based Synthetic Fuels  \n(Agora Energiewende, Agora Verkehrswende and Frontier \nEconomics, 2018).\n74.\t Grahn, M. et al. Review of electrofuel feasibility\u2014cost and \nenvironmental impact. Prog. Energy 4, 032010 (2022).\n75.\t Stolz, B., Held, M., Georges, G. & Boulouchos, K. \nTechno-economic analysis of renewable fuels for ships  \ncarrying bulk cargo in Europe. Nat. Energy 7, 203\u2013212  \n(2022).\n76.\t E-Kerosene for Commercial Aviation, From Green Hydrogen and \nCO2 from Direct Air Capture\u2014Volumes, Cost, Area Demand and \nRenewable Energy Competition in the United States and Europe \nfrom 2030 to 2050 (dena, 2022); https://www.dena.de/ \nnewsroom/publikationsdetailansicht/pub/studie-e-kerosene-for- \ncommercial-aviation/\n77.\t E-Fuels: A Techno-Economic Assessment of European  \nDomestic Production and Imports towards 2050\u2014Update \n(Concawe, 2024); https://www.concawe.eu/wp-content/ \nuploads/Rpt_24-4-1.pdf\n78.\t World Energy Outlook 2023 (IEA, 2023).\n79.\t Jing, L. et al. Understanding variability in petroleum jet \nfuel life cycle greenhouse gas emissions to inform aviation \ndecarbonization. Nat. Commun. 13, 7853 (2022).\nAcknowledgements\nWe gratefully acknowledge funding from the Kopernikus-Projekt \nAriadne through the German Federal Ministry of Education and \nResearch (grant nos. 03SFK5A and 03SFK5A0-2, A.O. and F.U.) and the \nHyValue project (grant no. 333151, F.U.). We thank the IEA, in particular \nJ. M. Bermudez, for providing the Hydrogen Projects Database and for \nanswering related questions. We thank R. Pietzcker for brainstorming \nfigures, P. Verpoort for cost parameters and J. Sitarz for carbon price \ndata. Following the IEA\u2019s Creative Commons license, we note that this \nis a work derived by us from IEA material and we are solely liable and \nresponsible for this derived work. The derived work is not endorsed by \nthe IEA in any manner.\nAuthor contributions\nA.O. suggested the initial research question, which F.U. extended.  \nA.O. and F.U. collected cost data, while A.O. collected all other data. \nA.O. performed the analyses and created the figures. A.O. and  \nF.U. interpreted the results. A.O. wrote the paper with contributions \nfrom F.U.\nFunding\nOpen access funding provided by Potsdam-Institut f\u00fcr \nKlimafolgenforschung (PIK) e.V.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at  \nhttps://doi.org/10.1038/s41560-024-01684-7.\n\nNature Energy | Volume 10 | January 2025 | 110\u2013123\n123\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s41560-024-01684-7.\nCorrespondence and requests for materials should be addressed to \nAdrian Odenweller.\nPeer review information Nature Energy thanks Paul Wolfram, \nSebastian Zwickl-Bernhard and the other, anonymous, reviewer(s) for \ntheir contribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this licence, visit http://creativecom \nmons.org/licenses/by/4.0/.\n\u00a9 The Author(s) 2025\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nExtended Data Fig. 1 | Green hydrogen electrolysis capacity in 1.5\u2009\u00b0C scenarios. \na-b, Institutional and corporate 1.5\u2009\u00b0C scenarios. c-d, Integrated assessment \nmodel (IAM) scenarios. IAM scenarios from the IPCC AR6 Database59 only \ninclude scenarios in category C1 (1.5\u2009\u00b0C). IMP stands for illustrative mitigation \npathway. IMP-LD focusses on demand, IMP-Ren on renewable energy, and \nIMP-SP on sustainable development. NGFS stands for Network for Greening \nthe Financial System \u2013 a project that provides regularly updated IAM scenarios \nfrom different models60. We exclude scenarios, which either always report \nzero green hydrogen, or which in any year from 2025 report a lower capacity \nthan has already been realised in 2023. If scenarios do not report electrolysis \ncapacity, we convert production quantities into corresponding electrolysis \ncapacity, which implies uncertainties (see Methods). The scenarios show a very \nwide range, particularly in 2030, underlining the high uncertainty surrounding \nthe green hydrogen market ramp-up. Furthermore, panels a and c show that \nunprecedented growth rates would be required to achieve the 1.5\u2009\u00b0C scenarios in \n2030 (apart from a few IAM scenarios that report limited use of green hydrogen \nin 2030). Figure 4a depicts the distribution of these two scenario groups in 2030. \nFor the IAM scenarios, we are uncertain to what extent the different models \nexplicitly represent the different hydrogen applications and whether hydrogen \nresults have been vetted in detail. Therefore, in our analysis of required subsidies \nfor 1.5\u2009\u00b0C, we use the median of the institutional and corporate scenarios (see \nSupplementary Figure 11 and Supplementary Figure 13).\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nExtended Data Fig. 2 | Designated end uses of project announcements.  \na, Projects until 2024 by end use. b, Projects from 2024 until 2030 by end use. \nThe IEA Hydrogen Production and Infrastructure Projects Database 202358 \ndistinguishes 14 end use categories (plus one category for projects that have no \ndesignated end use). Projects can be assigned to multiple end uses, although \na single end use is most common (see Supplementary Figure 14). If a project \ncontains more than one end use, we distribute the announced capacity evenly \namong them. Natural gas and grey hydrogen emerge as the most important \ncompetitors of green hydrogen and electrofuels, covering over 90% of the \nproject pipeline in 2030. For each end use, we model the competition between \nthe corresponding green product and the fossil competitor (see Extended \nData Table 1). Notably, in the end uses refining and synthetic fuels, project \nannouncements are insufficient for the quantities required under already \nimplemented demand-side regulation (Supplementary Figure 15). Furthermore, \nin 2030 there is a substantial mismatch between the designated end uses in the \nproject announcements and the end use shares according to the IEA NZE scenario \n(Supplementary Figure 13a).\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nExtended Data Fig. 3 | Illustration of our model of the pay-as-bid market \npremium. This figure shows our the exemplary competition between green \nhydrogen and natural gas with carbon pricing in our central estimate. For \nillustration, we show how to calculate required subsidies for projects built in \n2024, and for projects built in 2035. In both cases, projects have to sell at their \nrespective LCOH for the entire payback period (dashed green horizontal lines). \nThe specific cost gap decreases due to rising natural gas prices, which increase \ndue to carbon pricing. This cost gap defines the required subsidies for projects \nbuilt in the corresponding year. For projects built in 2024, subsidies are required \nthroughout their payback period (red shaded area and red arrows). For projects \nbuilt in 2035, subsidies are required until 2047, which is when the LCOH of \nprojects built in 2035 intersects with the price of natural gas (purple shaded \narea and purple arrows). Notably, this is longer than the instantaneous cost gap \nbetween green hydrogen costs and natural gas prices would suggest as natural \ngas becomes more expensive than green hydrogen already in 2043. However, \nthis implies that only projects built after 2043 do not require any subsidies. In \norder to calculate annual subsidies for a given year, we therefore need to track \nall projects built in previous years, multiplied by the corresponding cost gap, \nwhich is the gap between the constant LCOH of those years and the current \nprice of natural gas. In Fig. 5e, f and Extended Data Figure 6 we show the sum of \nthese annual subsidies across all end use sectors. Cumulative subsidies are then \ncalculated by adding up annual subsidies over all previous years (see Methods).\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nExtended Data Fig. 4 | Levelised cost of green hydrogen in 2030 compared \nto recent studies. The filled circle indicates the central estimate (if available) \nand the vertical line indicates the range (if available). Colours indicate the \norganisation. Our 2030 LCOH estimates are shown in black on the left, and are \nin line with most recent studies. The lower end of the range corresponds to the \nprogressive scenario in the end-uses ammonia, refining and biofuels, for which \nwe do not include transport and storage costs (see Supplementary Table S2). The \nupper end of the range corresponds to the conservative scenario in the end-uses \nindustry, power, grid injection, CHP, domestic heat, iron & steel, and mobility, for \nwhich we include transport and storage costs of 20 $/MWh, based on literature \nvalues. Studies that report very low LCOH values, for example BNEF, likely do \nnot include transport and storage costs, which are however critical for a full \nassessment of hydrogen competitiveness in our model. Note that Capgemini did \nnot calculate LCOHs, but rather conducted a global survey among more than 100 \ncompanies in the hydrogen industry. The data for this figure, including sources \nand the full name of the organisations, is available on GitHub.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nExtended Data Fig. 5 | Cost gap between green products and fossil competitors \nby end uses. Next to the brace, we list all end uses represented by the competition \nin that row, showing project announcements in GW by 2030 within these end \nuses (see Extended Data Figure 2). The green and brown line show our central \nestimate, while the shaded green and brown areas represent the uncertainty \nrange spanned by the conservative and progressive scenarios. a-f, Cost gap \nbetween green hydrogen and fossil competitors, without carbon prices (left \ncolumn) and with carbon prices (right column) that are in line with reaching EU \nclimate targets41 (149 $/tCO2 in 2030, 246 $/tCO2 in 2040, 407 $/tCO2 in 2050, \nsee Extended Data Table 3). Panels a-f belong to the legend at the top. Panels a-d \nrepresent more than 90% of the total announced capacity by 2030 and are also \nshown in Fig. 5a\u2013d. In the case of mobility, the LCOH has been adjusted for the \nslightly higher end-use efficiency of hydrogen trucks in comparison to diesel \ntrucks (e-f, also see Extended Data Table 1), whereas this is for simplicity not \nvisualised here for iron & steel, which is grouped with other end uses in a-b.  \ng-l, Cost gap between different hydrogen-based electrofuels and fossil \ncompetitors, without carbon prices (left column) and with carbon prices (right \ncolumn). Panels g-l belong to the legend at the bottom. Without carbon pricing, \nno green product becomes cost competitive with its fossil competitor. With \ncarbon pricing, a distinct sequence emerges that is however characterised \nby high uncertainty. In the central estimate, green hydrogen first becomes \ncompetitive with grey hydrogen in 2034 (d), then with diesel in 2037 (f), after \nwhich e-methanol becomes competitive with grey methanol in 2043 (j), followed \nby green hydrogen becoming competitive with natural gas in 2044 (b). In the \ncentral estimate, e-kerosene and e-methane narrowly miss reaching cost parity \nwith their fossil competitor by 2050 (h,l).\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\n\u2462\n\u2462\n\u2462\nExtended Data Fig. 6 | The 2030 green hydrogen implementation gap by \nscenarios. The bars show annual subsidies required to realise all project \nannouncements by 2030, while the line shows the corresponding cumulative \nsubsidies. The left column shows results without carbon pricing, and the central \ncolumn shows results with carbon pricing in line with reaching EU climate \ntargets41 (149 $/tCO2 in 2030, 246 $/tCO2 in 2040, 407 $/tCO2 in 2050, see \nExtended Data Table 3). The right column shows how the cumulative subsidies \ncompare to currently announced global subsidies according to BNEF43.  \na-c, Central estimate (also shown in Fig. 5e\u2013g). d-f, Progressive scenario,  \ng-i, Conservative scenario. As summarised in Table 1, the size of the 2030 green \nhydrogen implementation gap strongly depends on the scenario. However, \nwithout carbon pricing there always remains a substantial gap of $1.0 trillion \nin the central estimate (c), with uncertainties ranging from $0.5 trillion in the \nprogressive scenario (f) to $2.3 trillion in the conservative scenario (i).  \nEven with ambitious global carbon pricing, the gap only closes in the progressive \nscenario (f).\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nExtended Data Fig. 7 | Required subsidies for green hydrogen and electrofuels \ncompared with historical and projected support for solar PV and wind. The \ngreen and purple line show our central estimate, while the corresponding shaded \nareas depict the range spanned by the progressive and conservative scenario. \nUntil 2030, the build-out of green hydrogen and electrofuels follows the project \nannouncements, while after 2030 it follows the median of the institutional and \ncorporate 1.5\u2009\u00b0C scenarios (Supplementary Figure 11). Similarly, until 2030, end \nuses are obtained from the project announcements (Extended Data Figure 1), \nwhile after 2030 end use shares follow the shares of the IEA NZE 1.5\u2009\u00b0C scenario40 \n(Supplementary Figure 13b). Data for historical and projected support for solar \nPV and wind is obtained from the IEA Renewables 2023 report, estimated via the \ncost difference between solar PV/wind and fossil fuel power plants68. The values \ncorrespond to the LCOE approach, not the value-adjusted LCOE approach.  \na-b, Total annual support without carbon pricing (a) and with ambitious carbon \npricing (b) in line with reaching EU climate targets41 (149 $/tCO2 in 2030,  \n246 $/tCO2 in 2040, 407 $/tCO2 in 2050, see Extended Data Table 3). Without \ncarbon pricing, required annual support in a 1.5\u2009\u00b0C scenario quickly exceeds \nhistorically observed support for solar PV and wind power (a). With carbon \npricing, total annual support could be limited to the same order of magnitude, \nalthough large uncertainties prevail (b). Note that the historical and projected \nsupport for wind power turns negative in 2022 as it is estimated from the \ndifference between the generation costs of wind and from fossil fuels68.  \nIn 2022, the energy crisis led to an unprecedented surge of natural gas prices, \nparticularly in Europe, leading to a negative estimate of policy support.  \nc-d, Relative support (per MWh) without carbon pricing (c) and with carbon \npricing (d). When calculating relative support for green hydrogen and \nelectrofuels, we exclude production that is backed by demand-side policies \nand therefore does not require subsidies. Without carbon pricing, green \nhydrogen and electrofuels require support until at least 2050, and potentially \nindefinitely (c). Due to additional conversion losses, electrofuels require higher \nspecific support. Overall, relative support is in the same order of magnitude as \nhistorically observed for solar PV and wind. With carbon pricing, the specific \nsubsidy requirements of green hydrogen and electrofuels steadily decrease \ntowards 2050, reaching zero for green hydrogen in our central estimate (d). \nFigure adapted from ref. 68 under a Creative Commons license CC BY 4.0.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nExtended Data Table 1 | Mapping of project end uses to the competition of green products and fossil competitors\nSorted by capacity of project announcements until 2030. We provide short explanations for some end use sectors: For iron & steel, production sites have to operate on the Direct Reduced \nIron (DRI) route, where hydrogen reduces iron ore to produce DRI, which is then used as a feedstock in electric arc furnaces (EAF) or occasionally in blast furnaces to produce steel. In DRI \nplants, natural gas could easily replace hydrogen as the reduction agent, making natural gas the main fossil competitor, not coal, which is used in the blast-furnace (BF) route. For other \nindustry, we assume that hydrogen replaces natural gas in providing process heat. For mobility, we assume that hydrogen is used for heavy duty vehicles such that fuel cell electric trucks \ncompete with conventional diesel trucks. For synthetic fuels, we assume they will be used primarily in aviation, making e-kerosene the green product and kerosene the fossil competitor. For \nend uses where the green product is not used thermally and thus cannot be compared with the fossil competitor using the lower heating value (LHV), we include the relative efficiency \nbetween the green product (\u03b7green\nLHV ) and the fossil competitor (\u03b7fossil\nLHV ). This applies to iron & steel and mobility. Note that we neglect additional transformation costs in end uses in cases where \nusing hydrogen requires converting end use applications (other industry, mobility, power, CHP, and domestic heat, also see Methods)69,70.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nExtended Data Table 2 | Parameters to calculate the levelised costs of green hydrogen and hydrogen-based electrofuels\nThe upper part of the table displays all parameters required to calculate the LCOH (see Methods). The central estimate column shows the values for the figures presented in the main text \n(central estimate), while the \u201cscenario range\u201d column shows the values used in the sensitivity studies (progressive and conservative scenario). Please refer to the Excel file on the GitHub \nrepository for further details regarding the sensitivity scenarios. The payback period and weighted average cost of capital are used to calculate the annuity and remain the same for both green \nhydrogen and electrofuels. The lower part of the table displays all parameters required to calculate the levelised costs (LCOX) of the green hydrogen derivatives, e-methanol, e-kerosene and \ne-methane (see Methods). Shared parameters for all derivatives are the green hydrogen price (obtained from the LCOH calculation), the renewable CO2 supply price, the full-load hours of the \nsynthesis, and FOM costs. All values in MWh refer to the respective lower heating value71\u201377.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01684-7\nExtended Data Table 3 | Parameters to calculate the total cost of the fossil competitors\nAll values in MWh refer to the respective lower heating value. The natural gas price and the oil price in 2050 are obtained from the IEA NZE 1.5\u2009\u00b0C scenario40. For grey hydrogen and grey \nmethanol, which are produced from natural gas, we ensure internal consistency with the natural gas price by calibrating the specific fixed costs in 2024, which we then use for 2030 and 2050 \n(see Methods). For kerosene and diesel in 2030 and 2050, we proceed similarly, using crude oil as the reference point. All values, sources and further comments are available in the Excel file \nprovided on the GitHub repository78,79.\n\n\n    Scientific Research Findings:", "answer": "We identify and quantify three gaps of global green hydrogen deployment. First, looking back, we find that in 2023 only 7% of the initially announced added green hydrogen capacity was eventually operational\u2014the past implementation gap. Second, looking ahead to 2030, we find that green hydrogen projects announced by industry increasingly exceed the requirements in 1.5\u00a0\u00b0C scenarios\u2014the closing ambition gap. Third, enormous subsidies of US$1.3\u00a0trillion would be required to realize all announced projects by 2030, far exceeding currently announced policy support, which we term the 2030 implementation gap. Policymakers should therefore interpret the increasingly steep growth indicated by recent project announcements with caution. To safeguard climate targets, policymakers must prepare for prolonged green hydrogen scarcity, low competitiveness, and high policy costs. Relying on abundant and cheap green hydrogen for the future risks crowding out cheaper alternatives such as end\u2011use electrification, and may endanger climate targets if hydrogen continues to fall short of expectations.", "id": 5}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 9 | October 2024 | 1230\u20131240\n1230\nnature energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01603-w\nPublic and local policymaker preferences for \nlarge-scale energy project characteristics\nHolly Caggiano\u2009\n\u200a\u20091\u2009\n, Sara M. Constantino2, Chris Greig\u2009\n\u200a\u20093 & Elke U. Weber\u2009\n\u200a\u20093,4\nRapidly building utility-scale energy infrastructure requires not only \npublic support but also political will across levels of government. Here \nwe use a conjoint experiment to assess preferences for large-scale energy \nprojects among residents and local elected officials in Pennsylvania\u2014a key \ntransition state with high solar potential where siting authority rests at \nthe local level. We find that residents prefer solar to other energy projects, \nand job creation and cooperative community ownership are associated \nwith increased support. Public and elected official support decreases \nwhen projects are owned by foreign companies. We find limited partisan \ndifferences in preferences, suggesting a path towards bipartisan support for \nsuch projects. Elected officials misperceive their constituents\u2019 preferences, \nunderestimating support for renewable energy and the importance of job \ncreation. As local officials are key decision-makers regarding infrastructure \ndevelopment, their preferences and perceptions of constituents\u2019 \npreferences may dictate which energy projects are approved and what \ncommunity benefits they deliver.\nWhile there are multiple pathways for the USA to reach net-zero \nemissions by 2050, there is broad agreement that all rely on rapid, \nlarge-scale deployment of renewable energy technology to meet \nincreased demand for electricity generation while quickly reducing \nemissions1,2. This transition will require radical technological, indus-\ntrial, economic and social shifts: most net-zero pathways call for the \nUS energy and industrial system to phase out coal use by around 2030 \nand substantially reduce oil and natural gas use by 2050. In addition \nto large-scale wind and solar energy development, transitioning to \nrenewable energy at scale requires major expansion and extension \nof transmission grids. The viability of building out the extraordinary \nlevels of infrastructure development required for deep decarboni-\nzation depends critically on social and community acceptance and \nunprecedented political will.\nDespite the fact that emissions reduction targets and renewable \nenergy production goals are often set at the national or state level, the \nplanning, siting and deployment of large-scale energy infrastructure \nprojects often happens locally, at the county or municipal level. Siting \nauthority varies widely state by state in the USA, but in the vast majority \nof states, onshore wind and solar siting authority rests either solely at \nthe local level, or authority varies on the basis of project size, where at \nleast some siting happens locally rather than at the state level3,4. Thus, \nlocal elected officials have a key role to play in the siting of large-scale \nenergy projects, including developing zoning ordinances, voting on \nproject approvals and negotiating community benefits. While some \nresearch has explored the policy preferences of local elected offi-\ncials relating to energy5, there is a notable gap in work that examines \ntheir preferences and perceptions of their constituents\u2019 preferences \nfor large-scale energy projects. This study thus seeks to understand \ncross-partisan public and local elected official preferences for char-\nacteristics of large-scale local energy projects, including distance \nfrom residential areas, employment opportunities, local benefits, \nownership structures, site type and fuel type. We focus on Pennsylvania \n(PA)\u2014a critical energy transition state in the USA2 and historical home \nof extractive energy industries\u2014where in March 2021 Governor Tom \nWolf announced the largest state government solar initiative in the \nUSA so far6. PA grants local governments authority for both onshore \nwind and solar siting3,4. Further, a recent analysis found that, across \nReceived: 16 February 2024\nAccepted: 11 July 2024\nPublished online: 1 August 2024\n Check for updates\n1School of Community and Regional Planning, University of British Columbia, Vancouver, British Columbia, Canada. 2Department of Psychology and \nSchool of Public Policy, Northeastern University, Boston, MA, USA. 3Andlinger Center for Energy and the Environment, Princeton University, Princeton, NJ, \nUSA. 4Department of Psychology, School of Public and International Affairs, Princeton University, Princeton, NJ, USA. \n\u2009e-mail: holly.caggiano@ubc.ca\n\nNature Energy | Volume 9 | October 2024 | 1230\u20131240\n1231\nArticle\nhttps://doi.org/10.1038/s41560-024-01603-w\nopposition are more nuanced, depending less on physical proximity \nto infrastructure and more on other characteristics of the project, like \nconcern that projects will decrease economic and aesthetic value of the \nland16,19. In contrast, recent survey research finds that closer proximity \nto residential areas or respondents\u2019 homes is associated with reduced \nsupport for large-scale projects, including wind turbines in Germany20 \nand renewable energy projects generally in the USA10. Further, concerns \nabout visual impacts are associated with reduced support for large-scale \nsolar in California21, and minimum distance requirements from residen-\ntial areas increase support for CCS projects in the USA22.\nIn addition to proximity to residential areas, the siting of projects \non specific types of land, such as wildlife habitats and farmland, has \nimplications for public support. Various studies point to local concerns \nabout the ecological impact of large-scale energy projects. Ecological \nimpact was the strongest predictor of opposition for wind energy \nprojects in Switzerland, Estonia and Ukraine23, and respondents in \nIreland favour wind projects located away from protected habitat \nareas24. This concern aligns with recurring narratives in the USA about \nwind turbines as harmful to threatened bird species\u2014recently, a wind \nenergy company in the USA pleaded guilty to violating the Migratory \nBird Treaty Act after their turbines killed at least 150 bald eagles25. This \nconcern also holds for solar, with respondents in the USA preferring \nthat solar farms have large buffers around wildlife habitat21. In addition \nto wildlife habitat, farmland often takes a centre role in discussions \nabout renewable energy siting. Recent studies have explored the ten-\nsions between farmers and developers, seeking to better understand \nthe intersection of farmland preservation and renewable energy sit-\ning26,27. Other studies have examined the potential for development of \nrenewables on brownfield sites or decommissioned mining sites, citing \nthe co-benefit of fewer land constraints and greater public support28,29.\nEconomic risks and opportunities associated with renewable and \nlow-carbon energy sources, specifically about employment, are par-\nticularly salient in political arguments about energy transition dynam-\nics. Studies have found that previous federal legislation in the USA, \nincluding the 2009 American Recovery and Reinvestment Act, have \nproven successful in creating energy sector jobs30. Tackling climate \nchange while creating jobs was key to US President Biden\u2019s agenda, \naiming to dispel narratives that a renewable energy transition will result \nin widespread job loss, particularly in states that are economically tied \nto the production of coal and gas. However, while net employment is \nexpected to increase as renewable energy production ramps up, this \nis true at a regional or national scale. At the local level, coal communi-\nties will probably see job losses at the individual and household level, \nresulting in loss of retail and commercial employment and decreases \nof local tax revenue bases14. Furthermore, workers justifiably fear lower \nwages and precarious job security in renewables work compared with \ntraditional energy industry jobs that are often unionized31.\nThere is potential for other local benefits, in addition to job crea-\ntion, to increase local support, including direct financial benefits, \nreduced local air pollution and environmental justice considerations. \nIn Germany, both individual and collective financial benefits, including \ndiscounted electricity rates, payments to communities and payments \nto municipalities, were found to improve acceptance of local wind \nprojects32. In the case of both wind turbines and electricity transmis-\nsion pylons, perceptions of health risks were also found to be crucial \nfor public support: hypothetical projects that cited a high frequency \nof health complaints were strongly rejected and projects that cited no \noccurrence of health complaints were accepted, and health-related \nattributes were more important than location and compensation \npayments33. Beyond just predicting support or opposition to projects, \nunderstanding preferences for local benefits also has implications \nfor equitable distribution of energy transition costs and benefits; it is \ncritical that marginalized groups already subject to historical environ-\nmental inequities do not bear the additional brunt of potential negative \ntransition impacts, including rising energy prices14.\nmore than 2,500 municipalities in the Commonwealth, 87% of zoning \ncodes provide no guidance on siting utility-scale solar7.\nWe find that direct benefits to communities, including the creation \nof permanent, union-wage jobs and cooperative community owner-\nship, increase support for energy projects. Pennsylvanians prefer \nsolar projects over wind, nuclear and natural gas power plants with \ncarbon capture and storage (CCS). Local elected officials, however, \nmisperceive the preferences of their constituents, underestimating \nsupport for renewable energy and the importance of job loss and \ncreation. The public and local elected officials have similar opinions of \nforeign-owned products, which is associated with the greatest reduc-\ntions in support. Importantly, we find limited partisan differences \nin preferences for large-scale renewable energy project characteris-\ntics, suggesting a promising path towards building bipartisan sup-\nport for such projects. Given the role of local elected officials as key \ndecision-makers regarding energy infrastructure development, their \npreferences and how they perceive their constituents\u2019 preferences may \nbe important predictors of which projects come to fruition and what \nbenefits they provide to local communities, offering opportunities to \nrealize just energy transitions.\nResults\nUnderstanding preferences for large-scale energy projects\nPolls consistently show that Americans overwhelmingly support the \ndevelopment of renewables, including wind and solar8, and are more \nlikely to support wind or solar than nuclear energy9. People who already \nlive near wind turbines tend to prefer additional wind to other types of \ncentralized power plant10. Sharpton et al.10 also found positive percep-\ntions of natural gas as an energy supply, as respondents associated it \nwith positive economic impact and local employment. In this study, in \naddition to measuring support for large-scale onshore wind and solar \nprojects, we also examine support for nuclear energy and natural gas \nto enable comparisons across technologies.\nBeyond support or opposition based on type of technology or \nenergy infrastructure, there are a wide variety of project characteristics \nthat may influence public preferences. In public discourse, large-scale \nenergy projects are routinely framed in ways that emphasize specific \nrisks or benefits to local communities, and support or opposition to \nprojects is often tied to narratives about these risks and benefits11,12. At \nthe individual level, narratives help individuals make sense of uncer-\ntainty, influence cognition and support or challenge existing power \nrelations and policy outcomes13. Narratives, in this sense, also help \nto draw attention to and evaluate justice and equity implications sur-\nrounding project development14. Thus, we can consider a package of \ncharacteristics as one narrative, and bundling different combinations \nof characteristics may help us understand which narratives influence \nsupport or opposition across which groups. In this work, we consider \ncharacteristics across six dimensions: energy source, distance from \nresidential areas, type of development site, job opportunities, local \nbenefits and project ownership. We selected these dimensions on the \nbasis of an extensive review of existing literature, as well as conversa-\ntions with experts in the field, and informed by narratives prevalent in \nrecent news stories about energy project development.\nA rich body of literature examines preferences for siting energy \ninfrastructure, with a focus on renewable energy. Early research focused \nlargely on visual aesthetics and distance from residential areas in \nresponse to \u2018not in my back yard\u2019 (NIMBY) concerns from local resi-\ndents15\u201317. NIMBY looked to be a promising explanation for the \u2018social \ngap\u2019 in renewable energy siting decisions\u2014why, when there is consistent \nbroad public support for renewable energy, are specific projects heavily \nopposed18? However, across geographical, cultural and project-specific \ncontexts, there is mixed evidence for the importance of proximity to res-\nidential areas as a predictor of opposition to large-scale energy projects. \nEarly research suggests that, while proximity can have strong effects on \npublic support for proposed projects, the \u2018selfish\u2019 components of NIMBY \n\nNature Energy | Volume 9 | October 2024 | 1230\u20131240\n1232\nArticle\nhttps://doi.org/10.1038/s41560-024-01603-w\nProject ownership is another facet of energy infrastructure devel-\nopment that may impact public support. More than 80% of US energy \ninfrastructure is owned by the private sector, including national and \ninternational companies34. Studies have found preferences for local \nownership, though this is not the norm in practice23,35. Other studies \nfind opposition to foreign ownership36, which is supported in the USA by \n\u2018America First\u2019 rhetoric and the prioritization of national industry over \nforeign competitors. While it may currently seem unlikely for large-scale \nenergy projects to be community owned, recent work has explored pos-\nsibilities for scaling up or mainstreaming community-owned energy37,38.\nWhile the trends discussed above are supported by various sam-\nples and studies, there is also heterogeneity by individual characteris-\ntics, including partisanship. While members of both parties in the USA \nlargely support renewable energy, Republican support has dropped \nin recent years39. Reasons for renewable support also differ by party: \nwhile Republicans express support when they anticipate economic \nopportunities associated with renewable energy development, Demo-\ncrats tend to focus on renewables as a climate change solution40. At the \nnational level, Republicans typically have a more favourable opinion \nof natural gas than Democrats41 and are also more likely to support \nnuclear energy9. We draw attention to partisanship, here, as climate \nand energy are often viewed as highly polarized issues and political \nidentities have been found to exert powerful influences preferences42. \nFinding opportunities for bipartisan cooperation is one avenue towards \nadvancing climate action, and these may be more likely to occur at the \nstate and local levels than in national agenda-setting43.\nIn addition to its predilection for local, decentralized decision- \nmaking, PA is an ideal study site as the state has a rich and diverse energy \nhistory. Princeton\u2019s Net-Zero America modelling identified PA as a \ntop five state for solar capacity in the decade leading up to 2050, with \none model calling for 95\u2009GW of new installed capacity2. In addition to \nbeing a key state for renewables development in the coming decades, \nfor almost two centuries PA has been central to US energy resources. \nThe state has historically been a site of extractive energy practices, \nincluding coal mining and hydraulic fracturing (fracking) of shale \ngas, which has been linked to adverse public health impacts44. Nuclear \nenergy is another component of PA\u2019s energy production portfolio, \nand indeed PA is home to Three Mile Island, the site of a partial reactor \nmeltdown and worst nuclear incident in US history. While the accident \nhas not been linked to adverse physical health impacts, scholars have \ndocumented long-term mental health consequences for locals45. In the \nface of this history, the documented health co-benefits of renewable \nenergy deployment in PA may be particularly salient46.\nPublic preferences for large-scale energy projects\nIn this study, we ask: (1) what attributes of large-scale energy projects \nimpact public and political support for their development among PA \nresidents? And (2) how do local elected officials perceive public sup-\nport for projects? Figure 1 shows the average marginal component \neffects (AMCEs) from experiment 1, with 95% confidence intervals \n(see Supplementary Table 1 for model estimates). When reporting \nAMCEs, one level per attribute is displayed as a baseline category and \nDistance\nJobs\nBenefits\nOwnership\nSite\nType\n\u22120.2\n\u22120.1\n0\n0.1\n0.2\n50 miles (not likely to be visable from your home, school or workplace)\n10 miles (potentially visable during drives or commutes in and out of the area)\n5 miles (potentially visable from your home, school or workplace)\n2 miles (probably visable from your home, school or workplace)\nNo changes to number of local energy jobs\n50 jobs lost due to retirement of an old plant\nCreation of 50 permanent jobs with union-scale wages\nCreation of 50 temporary hourly-wage jobs during construction\nReduced energy costs for all local residents\nIncreased access to afordable clean energy in marginalized communities\nIncreased overall economic activity in the community\nReduced local air pollution\nOwned by an American private company\nOwned by a foreign private company\nOwned by state government\nOwned cooperatively by community\nLow visual impact on scenic resources\nAgricultural land\nDecommissioned mining or industrial site\nLow impact on wildlife habitats\nNatural gas power plant with CCS\nLarge-scale onshore wind turbines\nLarge-scale solar energy farm\nNuclear power plant\nEstimated AMCE\nFig. 1 | AMCEs of attribute levels on public support for project development. \nThis plot reports average effects of each project element on support for the \nproject in an online sample of PA residents (n\u2009=\u2009894). Point estimates are AMCEs \nwith 95% confidence intervals (mean values\u2009\u00b1\u2009s.e.m.) for each level. Each AMCE \nestimates how inclusion of the listed project feature affects support for energy \nprojects. Each element is compared against a baseline category, represented by \npoints without lines.\n\nNature Energy | Volume 9 | October 2024 | 1230\u20131240\n1233\nArticle\nhttps://doi.org/10.1038/s41560-024-01603-w\nis set at zero. AMCEs should be interpreted as the average change in \nprobability associated with a one-unit change in the predictor variable. \nWhen reporting in text, we convert changes in probability to changes \nin percentage points (p.p.) for ease of interpretation.\nConsidering distance from residential areas, PA residents are \nless likely to support projects close to home, compared with projects \nlocated far from residential areas (50\u2009miles). Support decreases by \n7\u2009p.p. when projects are located 2\u2009miles away from residential areas, \ncompared with when they are 50\u2009miles away. Considering job oppor-\ntunities, changes to local employment opportunities are also salient \nto respondents. Using no changes to jobs as a reference category, job \nloss reduces overall support for projects by 12\u2009p.p. and the creation \nof permanent jobs increases support by 12\u2009p.p. Creating temporary \njobs increases support by 4\u2009p.p. compared with no job changes. With \nregard to local project benefits, respondents find personal monetary \nbenefits (for example, reduced costs) preferable to community or \nhealth benefits. They prefer the reference category, reduced energy \ncosts for all local residents, to increased overall economic activity in \nthe community, which is associated with a reduction of 6\u2009p.p. in support \nand reduced local air pollution, which is associated with a reduction \nof 5\u2009p.p. in support.\nVariation in ownership explains the greatest variation in support \ncompared to other attributes. Respondents prefer cooperative com-\nmunity ownership, associated with a 5\u2009p.p. increase in support rela-\ntive to American private company ownership. There is no statistically \nsignificant difference in support between the reference category and \nstate government ownership. The most pronounced result concerns \nownership by a foreign private company, which decreases support \nby 17\u2009p.p. PA residents appear to be indifferent about the type of land \non which energy projects are sited. We find no statistically significant \ndifferences in support detected between the reference category, low \nvisual impact on scenic resources and other attributes including agri-\ncultural land, decommissioned mining or industrial sites, and areas \nthat would have low impact on wildlife habitats. PA residents appear \nto support solar projects over other types of energy project, with an \nassociated 7\u2009p.p. increase in support relative to the reference category \nof natural gas power plants with CCS. Nuclear projects are associated \nwith a 6\u2009p.p. reduction in support compared with natural gas with CCS.\nFigure 2 displays the results of a subgroup analysis by political \nparty, comparing preferences of Democrats, Republicans and Inde-\npendents. Here, we report marginal means for each party on each \nattribute.\nWe find statistically significant differences in preferences as a func-\ntion of political party affiliation between Republicans and Democrats, \nRepublicans and Independents, and Independents and Democrats. \nWhile partisan differences are present, they are generally minimal and \nnot statistically significant, with their magnitudes being nearly identi-\ncal across groups. Considering distance from residential areas and site \ntypes, there is no statistically significant difference in marginal means \nacross respondents from any political party. Considering jobs, there \nis not a statistically significant difference in marginal means between \nRepublicans and Democrats, suggesting similar preferences. There are \nsmall differences between Independents and each other party that indi-\ncate Independents do not prioritize jobs as heavily (6\u2009p.p. difference \nbetween Independents and Democrats for \u2018no changes to number of \njobs\u2019, 7\u2009p.p. difference between Independents and Republicans for \u201850 \njobs lost due to retirement of an old plant\u2019). In both cases, Independents \nlean further towards the middle.\nConsidering the ownership and fuel type attributes, there are sta-\ntistically significant differences between Republicans\u2019 and Democrats\u2019 \n50 miles (not likely to be visable from your home, school or workplace)\n10 miles (potentially visable during drives or commutes in and out of the area)\n5 miles (potentially visable from your home, school or workplace)\n2 miles (probably visable from your home, school or workplace)\nNo changes to number of local energy jobs\n50 jobs lost due to retirement of an old plant\nCreation of 50 permanent jobs with union-scale wages\nCreation of 50 temporary hourly-wage jobs during construction\nReduced energy costs for all local residents\nIncreased access to afordable clean energy in marginalized communities\nIncreased overall economic activity in the community\nReduced local air pollution\nOwned by an American private company\nOwned by a foreign private company\nOwned by state government\nOwned cooperatively by community\nLow visual impact on scenic resources\nAgricultural land\nDecommissioned mining or industrial site\nLow impact on wildlife habitats\nNatural gas power plant with CCS\nLarge-scale onshore wind turbines\nLarge-scale solar energy farm\nNuclear power plant\nMarginal mean\na   Democrat leaning\nb   Republican leaning\nc   Republican \u2212 Democrat\nMarginal mean\nDistance\nJobs\nBenefits\nOwnership\nSite\nType\n0.3\n0.4\n0.5\n0.6\n0.7\n0.3\n0.4\n0.5\n0.6\n0.7\n\u22120.2\n\u22120.1\n0\n0.1\n0.2\nEstimated diference\nFig. 2 | Marginal means by respondent party (experiment 1). a\u2013c, This plot \nreports differences in subgroup preferences between Democrats (blue circle, \na) and Republicans (red square, b), along with the difference in marginal means \nbetween subgroups (black triangle, c) in an online sample of PA residents \n(n\u2009=\u2009894). Point estimates are marginal means with 95% confidence intervals \n(mean values\u2009\u00b1\u2009s.e.m.) for each level. Marginal means are centred around 50%, \nindicating no difference between levels.\n\nNature Energy | Volume 9 | October 2024 | 1230\u20131240\n1234\nArticle\nhttps://doi.org/10.1038/s41560-024-01603-w\nmarginal means in some levels. There is a 12\u2009p.p. difference between the \nparties\u2019 marginal means in the \u2018owned by an American private company \nlevel\u2019, indicating that Republicans more strongly prefer projects that \nare owned by American private companies, and an 8.4\u2009p.p. difference \nindicating that Democrats more strongly prefer projects owned by state \ngovernments. Regarding project type, Democrats have a stronger pref-\nerence for onshore wind turbines (10.6\u2009p.p. difference) and Democrats \nare less favourable towards nuclear energy than Republicans (8.6\u2009p.p. \ndifference). Similarly to Republicans, Independents favour American \nprivate companies more than Democrats (11\u2009p.p. difference) and also \nhave a less favourable opinion of onshore wind (12\u2009p.p. difference). We \ntested for differences across other subgroups including gender (Sup-\nplementary Fig. 1), education (Supplementary Fig. 2), income (Supple-\nmentary Fig. 3) and policy framing conditions (Supplementary Fig. 4).\nElected official preferences for large-scale energy projects\nIn experiment 2, we conducted a reduced version of the conjoint experi-\nment with a sample of local elected officials. For this study, we ask \nlocal elected officials not only about their own support of renewable \nprojects but also about their perceptions of constituent support for \nthese projects. As this experiment has a smaller sample size and may \nbe underpowered, we draw attention to substantive differences in \nresponses rather than solely assessing statistical significance.\nDistance from residential areas, job opportunities and site type \ndo not have a statistically significant causal effect on personal support \nfor energy projects in the overall sample of PA local elected officials \n(Fig. 3; see Supplementary Table 2 for model estimates). However, \nrespondents appear to favour projects farther away from residential \nareas, projects that create permanent jobs and projects that reduce \nresidential energy costs rather than increasing economic activity. \nLike the PA resident sample in experiment 1, ownership models are \nthe strongest predictors of local elected officials\u2019 support compared \nwith other attributes. Compared with a baseline of ownership by an \nAmerican private company, ownership by a foreign private company \nreduces support by 29\u2009p.p. Unlike the public sample in experiment 1, \nlocal elected officials in PA do not prefer solar energy to natural gas with \nCCS, where we see no statistically significant difference between levels. \nIn fact, local elected officials prefer this baseline to other options\u2014nota-\nbly, onshore wind turbines are associated with a 26\u2009p.p. reduction in \nsupport. Natural gas without CSS is associated with a 19\u2009p.p. decrease \nin support compared with natural gas with CCS. Nuclear power plants \nare associated with an 18\u2009p.p. reduction in support compared with \nnatural gas with CCS.\nSplitting the sample of public officials into Republicans and Demo-\ncrats (considering there were few Independents in this sample) and \nlooking at the marginal means, we find limited differences. In Fig. 4 \n(see Supplementary Table 3 for model estimates), the rightmost panel \ndisplays differences in marginal means between parties, with positive \nvalues indicating that Republicans prefer a level more than Democrats \nand negative values indicating that Democrats prefer a level more \nDistance\nJobs\nBenefits\nOwnership\nSite\nType\n\u22120.4\n\u22120.2\n0\n0.2\n50 miles\n2 miles\nNo net increase of permanent jobs\nNet decrease of 50 permanent jobs\nNet increase of 50 permanent jobs\nReduced residential energy costs\nIncreased economic activity\nOwned by an American private company\nOwned by a foreign private company\nOwned by state government\nOwned cooperatively by a community\nMinimal impact on scenic views\nMinimal impact on agriculture\nMinimal impact on wildlife habitats\nNatural gas power plant with CCS\nLarge-scale onshore wind turbines\nLarge-scale solar energy farm\nNatural gas power plant\nNuclear power plant\nEstimated AMCE\nElected oficials\nFig. 3 | AMCEs of attributes on local elected official support for project \ndevelopment. The figure reports average effects of each project element \non support for the project in an online sample of local elected officials in PA \n(n\u2009=\u2009206). Point estimates are AMCEs with 95% confidence intervals (mean \nvalues\u2009\u00b1\u2009s.e.m.) for each level. Each AMCE estimates how inclusion of the listed \nproject feature affects support for energy projects. Each element is compared \nagainst a baseline category, represented by points without lines.\n\nNature Energy | Volume 9 | October 2024 | 1230\u20131240\n1235\nArticle\nhttps://doi.org/10.1038/s41560-024-01603-w\nthan Republicans. For public officials, there are very few statistically \nsignificant differences in marginal means between parties. This may \nbe due to the very small sample size and lack of power for these com-\nparisons. The only attribute level that shows statistically significant \ndifferences across parties is support for projects that have minimal \nimpact on wildlife habitats, where a 17\u2009p.p. difference indicates that \nDemocrats are more concerned about protecting wildlife habitats \nwhen siting energy projects. Not statistically significant but substan-\ntively interesting, Republicans appear to support projects owned by \nAmerican private companies the most while Democrats appear to \nprefer projects owned cooperatively by communities. Additionally, \nRepublicans appear to prefer natural gas with CCS to other energy \nprojects, while Democrats prefer solar.\nIn addition to asking local elected officials which project they \npreferred, we also asked them to pick which project they think their \nconstituents would prefer (see Supplementary Table 4 for model esti-\nmates). Figure 5a displays AMCEs for experiment 1, compared with \nwhat elected officials predicted constituents would prefer (Fig. 5b). \nThese are presented as AMCEs as they are not subgroups but samples \nfrom two different populations in different surveys (with different, \nthough largely corresponding, attribute levels). Thus, we cannot make \ndirect comparisons across samples but can observe general patterns. \nIt appears that elected officials largely project their own preferences \nwhen judging those of their constituents. Considering constituent sup-\nport for jobs, local elected official predictions do not show statistically \nsignificant differences between levels or a baseline of no changes to \nnumber of jobs, while there are relatively large casual impacts across \nlevels of this attribute in PA residents. The most notable difference \nis officials\u2019 perceptions of energy type, where officials believe that \ntheir constituents do not prefer other types of energy projects to the \nbaseline category of natural gas with CCS. In the PA resident sample, \nrespondents preferred solar energy to the same baseline, suggesting \nthat elected officials underestimate support for solar renewable energy \namong their constituents.\nDiscussion\nThe results of our conjoint experiments, and samples of matched \npublic and elected officials in a key transition state, have implications \nfor both planning and policy development around energy projects \nand how dimensions of proposed projects are both decided on and \ncommunicated to local communities. Considering the results of two \nexperiments, we raise four key implications for discussion: (1) evidence \nof shared priorities across political parties, (2) the importance of and \nopportunity for labour\u2013energy coalitions, (3) the role of alternative \nownership models in just transitions and (4) the importance of local \nelected officials to energy transition research.\nConsidering increasing polarization in the USA and that federal \nclimate change policy is framed as a highly partisan issue, our results \nare promising\u2014we largely see similarity, rather than heterogeneity, \nin preferences for energy infrastructure projects across Democrats \nand Republicans in our PA resident sample. We only saw statistically \nsignificant differences between the parties\u2019 preferences across two \nattributes\u2014ownership and energy type. Considering classic partisan \nideological differences, it is not surprising that Republicans favour \nproject ownership by American private companies more than Demo-\ncrats and Democrats prefer government-owned projects more than \nRepublicans. Across energy types, Democrats have a stronger prefer-\nence for onshore wind turbines and are also less likely to prefer nuclear \n50 miles\n2 miles\nNo net increase of permanent jobs\nNet decrease of 50 permanent jobs\nNet increase of 50 permanent jobs\nReduced residential energy costs\nIncreased economic activity\nOwned by an American private company\nOwned by a foreign private company\nOwned by state government\nOwned cooperatively by a community\nMinimal impact on scenic views\nMinimal impact on agriculture\nMinimal impact on wildlife habitats\nNatural gas power plant with CCS\nLarge-scale onshore wind turbines\nLarge-scale solar energy farm\nNatural gas power plant\nNuclear power plant\nMarginal mean\nMarginal mean\nDistance\nJobs\nBenefits\nOwnership\nSite\nType\n0\n0.25\n0.50\n0.75\n0\n0.25\n0.50\n0.75\n\u22120.25\n0\n0.25\nEstimated diference\na\nDemocrat\nb\nRepublican\nc\nRepublican \u2212 Democrat\nFig. 4 | Marginal means by respondent party (experiment 2). a\u2013c, This plot \nreports differences in subgroup preferences between Democrats (blue circle, \na) and Republicans (red square, b), along with the difference in marginal means \nbetween subgroups (black triangle, c) in an online sample of local elected \nofficials in PA (n\u2009=\u2009206). Point estimates are marginal means with 95% confidence \nintervals (mean values\u2009\u00b1\u2009s.e.m.) for each level. Marginal means are centred \naround 50%, indicating no difference between levels. In c, positive differences \nindicate that Republicans prefer the level.\n\nNature Energy | Volume 9 | October 2024 | 1230\u20131240\n1236\nArticle\nhttps://doi.org/10.1038/s41560-024-01603-w\nenergy projects. Considering the wide range of project characteristics \nour experiment covered, it is encouraging that parties broadly share \nvalues around other dimensions such as employment opportunities \nand project benefits. Agreement on the importance of jobs suggests \nthat just transitions may be a bipartisan value. While this experiment \ncannot provide insight into why energy preferences seem to tran-\nscend partisanship in this sample, the results are in line with polling \nresearch that suggests renewable energy is broadly popular across \nparties nationally39. Some research has found that the reasons for \nsupport across parties differ; for example, Democrats might prioritize \nclimate change mitigation while Republicans prioritize lower costs \nand energy independence40. Further, benefits like increased employ-\nment and community development may act to bridge divides. Prefer-\nences for new development may also differ from support for fossil fuel \nphase-out\u2014one study in western Colorado found partisanship to be a \nstrong predictor of support for coal phase-out47.\nThe importance of net job creation to respondents across parties \nechoes narratives that argue for a just energy transition that supports \nlabour in the face of fossil fuel job losses. A link between \u2018green jobs\u2019 \nand \u2018good jobs\u2019 has been central to climate policy under the Biden \nAdministration. It is also likely that this narrative is particularly salient \nin PA, which has seen recent declines in nuclear, coal and natural gas \nfuel jobs48. Echoing other calls, our results suggest an opportunity \nfor labour\u2013energy coalitions and worker-led transitions49. Partner-\nships between renewable energy developers and labour organizations \nmay be one pathway towards developing projects with broad public \nsupport, as well as other investments aimed at supporting job crea-\ntion and minimizing structural unemployment, such as job training \nprogrammes.\nIn addition to employment, project ownership appears to exert \ninfluence on public support for energy projects. Projects owned coop-\neratively by communities were preferred to other forms of ownership, \nspecifically the baseline of projects owned by American private com-\npanies. This result is supported by a literature on community energy \nand energy sovereignty that examines a wide range of community \nownership and management structures50,51. Currently, most large-scale \nenergy projects require private capital for financing, but there have \nbeen various recent calls among climate activists for public ownership \nof renewable energy infrastructure52. Our findings suggest that these \ncalls may not just be from a vocal activist minority but could broadly \nincrease support for project development. It is also interesting to note \nthe lack of support for foreign-owned projects. This may have been \nparticularly salient for Americans in 2021, given increases in populist, \nanti-globalization sentiment, particularly among the far right53.\nEmployment opportunities and ownership structures stand in con-\ntrast to attributes that did not have strong effects on support among \nresident respondents, such as local project benefits, site types and \ndistance from residential areas. It is possible that responses to projects \nsited close to residential areas may be different in practice; other stud-\nies point out that people may not understand distances accurately in \nthe abstract in which they are represented in survey questions21. Consid-\nering local project benefits, respondents narrowly preferred receiving \nlow-cost energy to other benefits, including increased overall economic \nactivity in the community and reduced local air pollution. Residential \nenergy costs in PA have continued to rise from 2021 through 2022, so \nit is reasonable to believe that costs are top of mind for ratepayers in \nthe state54. This finding is interesting because state governments and \nproject developers often use narratives about local community benefits \nwhen promoting projects and seeking to gain community acceptance. \nFor example, Cottontail Solar, a large project developed by Light-\nsourceBP currently in progress in PA as a part of their state-wide solar \ninitiative, touts the local economic benefits of the project, including \n$40 million in tax revenue over the project lifespan. The website also \npoints to the number of temporary local jobs created and metric tons \nof carbon dioxide reduced each year55. Our findings suggest that other \npossible aspects of this and other projects, including permanent jobs \ncreated, electricity rate reductions and local ownership models, may \nshift support more substantially.\nWhile respondents in the PA resident sample broadly preferred \nsolar to other types of energy projects, we find a different pattern in our \nsample of local elected officials, who appear to favour natural gas with \nCSS to other types of projects. Considering that many US states defer \nsiting authority to local jurisdictions, the preferences and perceptions \nof local elected officials are an incredibly important and understudied \npiece of the transition puzzle. A recent survey of utility-scale wind and \nsolar developers found that they consider local zoning ordinances and \nDistance\nJobs\nBenefits Ownership\nSite\nType\n50 miles (not likely to be visable from your home, school or workplace)\n10 miles (potentially visable during drives or commutes in and out of the area)\n5 miles (potentially visable from your home, school or workplace)\n2 miles (probably visable from your home, school or workplace)\nNo changes to number of local energy jobs\n50 jobs lost due to retirement of an old plant\nCreation of 50 permanent jobs with union-scale wages\nCreation of 50 temporary hourly-wage jobs during construction\nReduced energy costs for all local residents\nIncreased access to afordable clean energy in marginalized communities\nIncreased overall economic activity in the community\nReduced local air pollution\nOwned by an American private company\nOwned by a foreign private company\nOwned by state government\nOwned cooperatively by community\nLow visual impact on scenic resources\nAgricultural land\nDecommissioned mining or industrial site\nLow impact on wildlife habitats\nNatural gas power plant with CCS\nLarge-scale onshore wind turbines\nLarge-scale solar energy farm\nNuclear power plant\nEstimated AMCE\na    Actual constituent\nresponse\nb    Predicted constituent\nresponse\n\u22120.2\n\u22120.1\n0\n0.1\n\u22120.3 \u22120.2 \u22120.1\n0\n0.1\n0.2\n50 miles\n2 miles\nNo net increase of permanent jobs\nNet decrease of 50 permanent jobs\nNet increase of 50 permanent jobs\nReduced residential energy costs\nIncreased economic activity\nOwned by an American private company\nOwned by a foreign private company\nOwned by state government\nOwned cooperatively by a community\nMinimal impact on scenic views\nMinimal impact on agriculture\nMinimal impact on wildlife habitats\nNatural gas power plant with CCS\nLarge-scale onshore wind turbines\nLarge-scale solar energy farm\nNatural gas power plant\nNuclear power plant\nEstimated AMCE\nFig. 5 | AMCEs of attributes on predicted constituent preferences compared \nwith actual resident preferences. a,b, Predicted constituent preferences \n(online sample of local elected officials in PA, n\u2009=\u2009206) (a) compared with \nactual resident preferences from experiment 1 (online sample of PA residents, \nn\u2009=\u2009894) (b). Point estimates are AMCEs with 95% confidence intervals (mean \nvalues\u2009\u00b1\u2009s.e.m.) for each level. Each AMCE estimates how inclusion of the listed \nproject feature affects support for energy projects. Each element is compared \nagainst a baseline category, represented by points without lines. Attribute  \nlevels are indicated on the outside panels, noting that levels differed slightly \nbetween experiments.\n\nNature Energy | Volume 9 | October 2024 | 1230\u20131240\n1237\nArticle\nhttps://doi.org/10.1038/s41560-024-01603-w\ncommunity opposition to be two of the biggest reasons for project \ndelays and cancellations56. Thus, it matters that local elected offi-\ncials have accurate perceptions of public preferences when they hold \ndecision-making power about these projects. Local elected officials \nalso report perceptions of their constituents\u2019 preferences in line with \ntheir own, underestimating support for solar projects. This suggests a \ndisconnect between local elected officials and their constituents who \nappear more tolerant and interested in renewable energy projects like \nsolar, and echoes other studies that find that local elected officials \nmisperceive public preferences regarding climate policy57.\nThese implications point to several avenues for future research. \nFuture experimental survey research might examine additional renew-\nable energy sources and legacy fossil fuel sources, including coal-fired \npower plants. We also suggest similar studies be carried out in nation-\nally representative samples, both in the USA and internationally, to \nexamine geographic heterogeneity in preferences. Further, the impor-\ntance of employment and project ownership suggest that respondent \nchoices were influenced by attributes that correspond with community \nbuy-in. The dimensions we measured do not represent every concern \nthat could be raised about a project, and these experiments do not \nconsider issues of process. For example, additional research could \nmeasure preferences for type and duration of community engagement \nfor project development. Some existing research closely examines the \nsiting process and argues for the role of perceived distributional and \nprocedural justice in building community support23. These dimen-\nsions might be better untangled in qualitative case study research that \npays close attention to processes and power dynamics across multiple \nprojects and communities58.\nWhile our study points to important implications for energy \ntransitions research, we also acknowledge limitations. First, the \nhypothetical nature of the experiment could differ from actual siting \ndecisions in practice. Some combinations of attribute levels are more \nplausible than others. For example, we are not aware of any existing \ncommunity-owned large-scale energy projects in PA. That said, we \nbelieve all combinations of attributes represent possibilities that grant \ninsight into preferences. Next, there are limitations to our sample of \nlocal elected officials in experiment 2. Local elected officials are a \ndifficult group to survey, and our sample of 206 officials is probably \nunderpowered to detect statistically significant differences across \nsubgroups; this may suggest a lack of differences where differences \nexist. Despite this, our results do provide some insight into prefer-\nences for energy projects for which local elected officials in PA often \nare key decision-makers. In this experiment, we target PA for its high \npotential for energy infrastructure development, precedent for local \ndecision-making, and historical economic ties to energy industries, \nbut results could reasonably be explored in other contexts, both within \nand beyond the USA. Given the mix of rural and urban areas and relative \ndemographic diversity of the state, we expect that many of the pat-\nterns we found could be transferable to other states. Environmental \nand economic concerns are probably salient across the USA; however, \nspecific economic context (for example, coal communities), cultural \nattitudes towards development, and political atmosphere and amount \nof existing energy infrastructure may drive different preferences that \ncan be explored in future research.\nIn this study, we assess the marginal impact of various character-\nistics of energy projects on support for project development among \nboth the public and local elected officials in a key transition state. We \nfind that the public is more likely to support solar projects, projects \nthat create permanent jobs and community-owned projects. Foreign \nownership reduces support for projects, as does job loss. On a positive \nnote, we find substantial overlap in project characteristics preferences \nbetween Republican and Democrat respondents. We also find, how-\never, that local elected officials misperceive the preferences of their \nconstituents, underestimating support for renewable energy and the \nimportance of job creation. Given the role of local elected officials as \nkey decision-makers regarding energy infrastructure development, \ntheir preferences and perceptions of their constituents\u2019 preferences \nmay play an important role when considering which energy projects \nare approved and what benefits they deliver to local communities.\nMethods\nExperimental design\nTo examine preferences for low-carbon energy infrastructure projects \nin PA, we conducted two conjoint experiments to test the effects of vari-\nous characteristics of energy projects on support for project develop-\nment. These attributes are displayed in Supplementary Table 5 and were \nselected on the basis of the above literature of previous energy siting \nstudies along with considerations about current discourse around \nenergy projects in the USA and PA. We examine the effects of these \nproject features on project approval in our samples of residents and \npolicymakers and broken down by party affiliation.\nConjoint experiments allow researchers to compare trade-offs \nacross multiple dimensions by randomly generating bundles of attrib-\nutes and asking respondents to pick between two such choice bundles. \nThis design more closely captures decision-making in realistic contexts \nthan asking respondents to independently evaluate options. Use of \nconjoint experiments to measure public preferences has increased \nin recent years across the social sciences59\u201361. Conjoint experiments \nhave high external validity and lower social desirability bias compared \nwith other methods of eliciting preferences62\u201364. In experiment 1, we \nexamine 24 project dimensions in a demographically representative \nsample of PA residents. In experiment 2, we examine a reduced set of \n19 dimensions in a smaller sample of local elected officials. We use a \nlimited number of attributes and increase the number of choice pairs \nthat respondents evaluate, to ensure that we are powered to detect \n0.07 and 0.05 effect sizes despite the smaller sample size. We selected \nattributes to drop on the basis of the results of experiment 1 (for exam-\nple, eliminating attribute levels that did not have statistically significant \nvariation, like in project benefits, and adding an additional attribute \nlevel in response to feedback from researchers experienced with elite \nsamples). We asked respondents to choose between two energy infra-\nstructure projects in their area that varied on five dimensions. The full \nlist of attributes is displayed in Supplementary Table 5. In experiment 2, \nwe additionally asked local elected officials to report their perceptions \nof constituent support for each proposed pair of projects.\nSampling\nFor experiment 1, we collected a sample of 894 Pennsylvanians that \nwas demographically representative on age, gender, ethnicity and \nhousehold income (based on 2019 American Community Survey 1-year \nestimates) and balanced on rural (25%), urban (25%) and suburban \nresidence (50%) (see demographics in Supplementary Table 6). Our \npublic sample was recruited by the survey firm Lucid. To recruit par-\nticipants, Lucid partners with multiple double opt-in panels that invite \nparticipants through emails, push notifications and in-app pop-ups. \nRespondents are incentivized through partner suppliers in the form \nof cash, gift cards or loyalty reward points. All respondents from our \npublic sample completed the survey between 7 February and 24 Febru-\nary 2022. Fifty-seven respondents who failed an attention check were \nremoved from the sample and subsequent analyses (Supplementary \nFig. 5). Mean response time was 16.9\u2009min. Each respondent was asked \nto choose between five distinct pairs of randomly generated projects, \ntesting support across a total of 8,940 projects. For both experiments, \nthe attribute levels were fully randomized within and across project \npairs, ensuring the non-parametric identification of causal effects \nof the attributes22,65. For this sample, probability weights were cre-\nated with a post-stratification raking procedure using Census vari-\nables (from 2019 American Community Survey 1-year estimates). This \nprocedure follows the methodology outlined in DeBell and Krosnick \n(2009) for the American National Elections Study66. Robustness tests \n\nNature Energy | Volume 9 | October 2024 | 1230\u20131240\n1238\nArticle\nhttps://doi.org/10.1038/s41560-024-01603-w\nincluding comparison across attention check failure and profile order \nare included in Supplementary Information.\nFor experiment 2, a sample of 219 local elected officials was \nrecruited by the survey firm CivicPulse. Thirteen respondents did not \ncomplete the conjoint experiment portion of the survey, bringing the \nanalysis sample to 206. The survey was fielded from 20 September \n2022 to 9 November 2022. The sample of respondents consisted of \nelected policymakers in PA from local governments (that is, township, \nmunicipality and county governments) with a population over 1,000 \nresidents (see demographics in Supplementary Table 7). Elected poli-\ncymakers include top elected officials and governing board members. \nEach respondent was asked to choose between four distinct pairs of \nrandomly generated projects, testing support across a total of 824 \nprojects. For this sample, probability weights were created with a \npost-stratification raking procedure using the Census and presidential \nvote share variables. Weights are calculated using local governments in \nPA with populations over 1,000 residents. This procedure follows the \nmethodology outlined in DeBell and Krosnick (2009) for the American \nNational Elections Study. Both surveys were approved by Princeton \nUniversity\u2019s institutional review board (IRB #14367: \u2018Origins, Func-\ntions, and Mechanisms of Coalitions for Change: Linking Individual \nCognition and Collective Action\u2019, principal investigator Elke Weber), \nand all participants provided informed consent. We used Qualtrics XM \n(version 2022) to collect data for both surveys.\nStatistical analysis\nStefanelli and Lukac (2020)67 provide a framework for conducting \npower analyses for conjoint experiments using simulation techniques. \nUsing their online power-analysis tool (https://mblukac.shinyapps.io/\nconjoints-power-shiny/), we find that experiment 1 is well powered. \nWith 894 participants, 5 tasks and 4 variable levels (attribute with the \nlargest number of options), the experiment has 94% power to detect \nan effect of 0.05 and 81% power to detect an effect of 0.04. Experiment \n2 is not as well powered given the smaller available sample size of local \nelected officials; with 206 participants, 4 tasks and 4 variable levels, the \nexperiment has 42% power to detect an effect of 0.05 and 78% power \nto detect an effect of 0.07 (Supplementary Fig. 6).\nWe calculate AMCEs for each option across attributes63, which meas-\nure the average causal effect of each attribute level on support for the pro-\nject relative to a baseline. Thus, we can estimate how much each attribute \nlevel increases or decreases support for a project, holding all other ele-\nments constant. The dependent variable is a binary indicator for whether \na given energy project was preferred (Y\u2009=\u20091) or not preferred (Y\u2009=\u20090). We use \nordinary least squares regression to calculate the AMCE for each attribute \nlevel, clustering standard errors at the respondent level to account for \nwithin-respondent correlations in responses. We use equation (1):\nYp = \u03b1 + \u03b2Dp + \u03b3Jp + \u03b7Bp + \u03b6Op + \u03b8Lp + \u03bbFp + \u03f5p\n(1)\nwhere Y represents the binary variable for whether the project p was \nselected. \u03b1 is the intercept. D, J, B, O, L and F are indicators for the \ndistance from residential areas; job impacts; local benefits; ownership \nmodel; site of land; and energy fuel type for p, respectively, followed \nby an error term, \u03b5. \u03b2, \u03b3, \u03b7, \u03c2, \u03b8, and \u03bb are the corresponding coefficients \nfor D, J, B, O, L, and F.\nWhile AMCEs have a clear causal interpretation, they can be mis-\nleading for subgroup analyses when one wishes to descriptively inter-\npret differences in preferences between two or more groups. Following \nthe advice of Leeper et al. (2020)68, we report marginal means instead \nof differences in AMCEs as indicators of favourability towards a given \nfeature when conducting subgroup analyses. Because levels of each \nattribute were randomly assigned, pairwise differences between two \nmarginal means for a given attribute have a direct causal interpretation. \nMarginal means have the advantage of directly presenting differences \nbetween subgroup preferences without relying on reference categories \nfor comparison, while AMCEs are interpreted relatively. Examples of \nconjoint choices seen by respondents appear in Supplementary Figs. 7 \nand 8. All data were processed and analysed using R, version 4.3.2.\nReporting summary\nFurther information on research design is available in the Nature \nPortfolio Reporting Summary linked to this article.\nData availability\nData supporting this study are openly available from OSF at https://\ndoi.org/10.17605/OSF.IO/WKGT9.\nCode availability\nCode supporting this study (including replication code for analy-\nses, figures and tables) is openly available from OSF at https://doi.\norg/10.17605/OSF.IO/WKGT9.\nReferences\n1.\t\nJenkins, J. D., Mayfield, E. N., Larson, E. D., Pacala, S. W. & Greig, C.  \nMission net-zero America: the nation-building path to a prosperous, \nnet-zero emissions economy. Joule 5, 2755\u20132761 (2021).\n2.\t\nLarson, E. et al. Net-zero America: potential pathways, \ninfrastructure, and impacts. Princeton University Andlinger Center \nfor Energy and the Environment https://netzeroamerica.princeton.\nedu/img/Princeton_NZA_Interim_Report_15_Dec_2020_FINAL.pdf \n(2021).\n3.\t\nEssa, E., Curtiss, K. & Dodinval, C. Solar siting authority across the \nUnited States. University of Michigan Center for Local, State, and \nUrban Policy https://closupstage.fordschool.umich.edu/ \nresearch/working-papers/solar-siting-authority-across- \nunited-states (2021).\n4.\t\nKahn, J. & Shields, L. State approaches to wind facility siting. \nNational Conference of State Legislatures https://www.ncsl.org/\nenergy/state-approaches-to-wind-facility-siting (2020).\n5.\t\nTumlison, C., Button, E. D., Song, G. & Kester, J. What explains \nlocal policy elites\u2019 preferences toward renewable energy/energy \nefficiency policy? Energy Policy 117, 377\u2013386 (2018).\n6.\t\nPennsylvania announces largest government solar energy \ncommitment in the U.S. Pennsylvania Office of Rural Health  \nhttps://www.porh.psu.edu/pennsylvania-announces-largest- \ngovernment-solar-energy-commitment-in-the-u-s/ (2021).\n7.\t\nBadissy, M. R. Comments for joint hearing of the Agriculture \nand Rural Affairs & Local Government Committees on \u2018Utility \nScale Solar Development & Local Government Ordinances'. \nPennsylvania State University (2021).\n8.\t\nTyson, A., Funk, C. & Kennedy, B. Americans largely favor U.S. \ntaking steps to become carbon neutral by 2050. Pew Research \nCenter Science & Society https://www.pewresearch.org/science/ \n2022/03/01/americans-largely-favor-u-s-taking-steps-to- \nbecome-carbon-neutral-by-2050/ (2022).\n9.\t\nLeppert, R. Americans continue to express mixed views about \nnuclear power. Pew Research Center https://www.pewresearch.\norg/fact-tank/2022/03/23/americans-continue-to-express-mixed- \nviews-about-nuclear-power/ (2022).\n10.\t Sharpton, T., Lawrence, T. & Hall, M. Drivers and barriers to public \nacceptance of future energy sources and grid expansion in the \nUnited States. Renew. Sustain. Energy Rev. 126, 109826 (2020).\n11.\t\nBj\u00e4rstig, T., Mancheva, I., Zachrisson, A., Neumann, W. & \nSvensson, J. Is large-scale wind power a problem, solution, or \nvictim? A frame analysis of the debate in Swedish media. Energy \nRes. Soc. Sci. 83, 102337 (2022).\n12.\t Bollman, M. Frames, fantasies, and culture: applying and \ncomparing different methodologies for identifying energy \nimaginaries in American policy discourse. Energy Res. Soc. Sci. \n84, 102380 (2022).\n\nNature Energy | Volume 9 | October 2024 | 1230\u20131240\n1239\nArticle\nhttps://doi.org/10.1038/s41560-024-01603-w\n13.\t Constantino, S. M. & Weber, E. U. Decision-making under the deep \nuncertainty of climate change: the psychological and political \nagency of narratives. Curr. Opin. Psychol. 42, 151\u2013159 (2021).\n14.\t Carley, S. & Konisky, D. M. The justice and equity implications of \nthe clean energy transition. Nat. Energy 5, 569\u2013577 (2020).\n15.\t Devine-Wright, P. Explaining \u2018NIMBY\u2019 objections to a power \nline: the role of personal, place attachment and project-related \nfactors. Environ. Behav. 45, 761\u2013781 (2013).\n16.\t van der Horst, D. NIMBY or not? Exploring the relevance of \nlocation and the politics of voiced opinions in renewable energy \nsiting controversies. Energy Policy 35, 2705\u20132714 (2007).\n17.\t Wolsink, M. Wind power and the NIMBY-myth: institutional \ncapacity and the limited significance of public support. Renew. \nEnergy 21, 49\u201364 (2000).\n18.\t Bell, D., Gray, T. & Haggett, C. The \u2018Social Gap\u2019 in wind farm siting \ndecisions: explanations and policy responses. Environ. Polit. 14, \n460\u2013477 (2005).\n19.\t Devine-Wright, P. Beyond NIMBYism: towards an integrated \nframework for understanding public perceptions of wind energy. \nWind Energy 8, 125\u2013139 (2005).\n20.\t Langer, K., Decker, T. & Menrad, K. Public participation in wind \nenergy projects located in Germany: which form of participation \nis the key to acceptance? Renew. Energy 112, 63\u201373 (2017).\n21.\t Carlisle, J. E., Solan, D., Kane, S. L. & Joe, J. Utility-scale solar and \npublic attitudes toward siting: a critical examination of proximity. \nLand Use Policy 58, 491\u2013501 (2016).\n22.\t Pianta, S., Rinscheid, A. & Weber, E. U. Carbon capture and \nstorage in the United States: perceptions, preferences, and \nlessons for policy. Energy Policy 151, 112149 (2021).\n23.\t Vuichard, P., Broughel, A., W\u00fcstenhagen, R., Tabi, A. & Knauf, J. \nKeep it local and bird-friendly: exploring the social acceptance of \nwind energy in Switzerland, Estonia, and Ukraine. Energy Res. Soc. \nSci. 88, 102508 (2022).\n24.\t Hallan, C. & Gonz\u00e1lez, A. Adaptive responses to landscape \nchanges from onshore wind energy development in the Republic \nof Ireland. Land Use Policy 97, 104751 (2020).\n25.\t The Associated Press. A Wind Energy Company Has Pleaded Guilty \nafter Killing at Least 150 Eagles (NPR, 2022).\n26.\t Moore, S., Graff, H., Ouellet, C., Leslie, S. & Olweean, D. Can \nwe have clean energy and grow our crops too? Solar siting on \nagricultural land in the United States. Energy Res. Soc. Sci. 91, \n102731 (2022).\n27.\t Pascaris, A. S., Schelly, C., Burnham, L. & Pearce, J. M. \nIntegrating solar energy with agriculture: Industry perspectives \non the market, community, and socio-political dimensions of \nagrivoltaics. Energy Res. Soc. Sci. 75, 102023 (2021).\n28.\t Adelaja, S., Shaw, J., Beyea, W. & Charles McKeown, J. D. \nRenewable energy potential on brownfield sites: a case study of \nMichigan. Energy Policy 38, 7021\u20137030 (2010).\n29.\t Spiess, T. & De Sousa, C. Barriers to renewable energy \ndevelopment on brownfields. J. Environ. Policy Plan. 18, 507\u2013534 \n(2016).\n30.\t Lim, T., Guzman, T. S. & Bowen, W. M. Rhetoric and reality: \njobs and the energy provisions of the american recovery and \nreinvestment act. Energy Policy 137, 111182 (2020).\n31.\t Jolley, G. J., Khalaf, C., Michaud, G. & Sandler, A. M. The economic, \nfiscal, and workforce impacts of coal-fired power plant closures in \nAppalachian Ohio. Reg. Sci. Policy Pract. 11, 403\u2013422 (2019).\n32.\t Knauf, J. Can\u2019t buy me acceptance? Financial benefits for wind \nenergy projects in Germany. Energy Policy 165, 112924  \n(2022).\n33.\t Zaunbrecher, B. S., Linzenich, A. & Ziefle, M. A mast is a mast is \na mast\u2026? Comparison of preferences for location-scenarios of \nelectricity pylons and wind power plants using conjoint analysis. \nEnergy Policy 105, 429\u2013439 (2017).\n34.\t National Infrastructure Protection Plan (NIPP) Energy \nSector-Specific Plan (NIPP). US Department of Homeland Security \n& US Department of Energy https://www.cisa.gov/sites/ \ndefault/files/publications/nipp-ssp-energy-2015-508.pdf (2015).\n35.\t Goedkoop, F. & Devine-Wright, P. Partnership or placation? The \nrole of trust and justice in the shared ownership of renewable \nenergy projects. Energy Res. Soc. Sci. 17, 135\u2013146 (2016).\n36.\t Venus, T. E. et al. The public\u2019s perception of run-of-the-river \nhydropower across Europe. Energy Policy 140, 111422 (2020).\n37.\t Roby, H. & Dibb, S. Future pathways to mainstreaming community \nenergy. Energy Policy 135, 111020 (2019).\n38.\t Warlenius, R. H. & Nettelbladt, S. Scaling up community wind \nenergy: the relevance of autonomy and community. Energy \nSustain. Soc. 13, 33 (2023).\n39.\t Kennedy, B. & Spencer, A. Most Americans support expanding \nsolar and wind energy, but Republican support has dropped. Pew \nResearch Center https://www.pewresearch.org/fact-tank/2021/ \n06/08/most-americans-support-expanding-solar-and-wind- \nenergy-but-republican-support-has-dropped/ (2021).\n40.\t Gustafson, A. et al. Republicans and Democrats differ in why they \nsupport renewable energy. Energy Policy 141, 111448 (2020).\n41.\t Hazboun, S. O. & Boudet, H. S. Natural gas\u2014friend or foe of the \nenvironment? Evaluating the framing contest over natural gas \nthrough a public opinion survey in the Pacific Northwest. Environ. \nSociol. 7, 368\u2013381 (2021).\n42.\t Mayer, A. National energy transition, local partisanship? Elite \ncues, community identity, and support for clean power in the \nUnited States. Energy Res. Soc. Sci. 50, 143\u2013150 (2019).\n43.\t Marshall, R. & Burgess, M. G. Advancing bipartisan \ndecarbonization policies: lessons from state-level successes and \nfailures. Clim. Change 171, 17 (2022).\n44.\t McDermott-Levy, R., Kaktins, N. & Sattler, B. Fracking, the \nenvironment, and health. Am. J. Nurs. 113, 45\u201351 (2013).\n45.\t Bromet, E. J., Parkinson, D. K. & Dunn, L. O. Long-term mental \nhealth consequences of the accident at three mile island. Int. J. \nMent. Health 19, 48\u201360 (1990).\n46.\t Dimanchev, E. G. et al. Health co-benefits of sub-national \nrenewable energy policy in the US. Environ. Res. Lett. 14, 085012 \n(2019).\n47.\t Mayer, A. More than just jobs: understanding what drives support \nfor a declining coal industry. Extr. Ind. Soc. 9, 101038 (2022).\n48.\t BW research. 2021 Pennsylvania Energy Employment \nReport. https://www.dep.pa.gov/Business/Energy/\nOfficeofPollutionPrevention/EnergyEfficiency_Environment_and_\nEconomicsInitiative/Pages/Workforce-Development.aspx (2021).\n49.\t Mijin Cha, J., Stevis, D., Vachon, T. E., Price, V. & Brescia-Weiler, M. A  \nGreen New Deal for all: the centrality of a worker and \ncommunity-led just transition in the US. Polit. Geogr. 95, 102594 \n(2022).\n50.\t Creamer, E. et al. Community energy: entanglements of \ncommunity, state, and private sector. Geogr. Compass 12, e12378 \n(2018).\n51.\t Schelly, C. et al. Energy policy for energy sovereignty: can policy \ntools enhance energy sovereignty? Sol. Energy 205, 109\u2013112 \n(2020).\n52.\t Dawson, A. People\u2019s Power: Reclaiming the Energy Commons  \n(OR Books, 2020).\n53.\t Skonieczny, A. Emotions and political narratives: populism, trump \nand trade. Polit. Gov. 6, 62\u201372 (2018).\n54.\t Electric power monthly. US Energy Information Administration \nhttps://www.eia.gov/electricity/monthly/epm_table_grapher.php \n(2022).\n55.\t Pennsylvania Cottontail Solar Farm Project | Lightsource bp. \nLightsource BP USA https://www.lightsourcebp.com/us/projects/\ncottontail-solar-farm-project/ (2020).\n\nNature Energy | Volume 9 | October 2024 | 1230\u20131240\n1240\nArticle\nhttps://doi.org/10.1038/s41560-024-01603-w\n56.\t Nilson, R., Hoen, B. & Rand, J. Survey of utility-scale wind and \nsolar developers report. Energie Technologies Area, Berkeley Lab \nhttps://emp.lbl.gov/publications/survey-utility-scale-wind-and- \nsolar (2024).\n57.\t Mildenberger, M. & Tingley, D. Beliefs about climate beliefs: the \nimportance of second-order opinions for climate politics. Br. J. \nPolit. Sci. 49, 1279\u20131307 (2019).\n58.\t Caggiano, H. & Weber, E. U. Advances in qualitative methods in \nenvironmental research. Annu. Rev. Environ. Resour. 48, 793\u2013811 \n(2023).\n59.\t Bergquist, P., Mildenberger, M. & Stokes, L. C. Combining climate, \neconomic, and social policy builds public support for climate \naction in the US. Environ. Res. Lett. 15, 054019 (2020).\n60.\t Bernauer, T. & Gampfer, R. How robust is public support for \nunilateral climate policy? Environ. Sci. Policy 54, 316\u2013330 (2015).\n61.\t Gampfer, R., Bernauer, T. & Kachi, A. Obtaining public support \nfor North\u2013South climate funding: evidence from conjoint \nexperiments in donor countries. Glob. Environ. Change 29, \n118\u2013126 (2014).\n62.\t Bansak, K., Hainmueller, J. & Hangartner, D. How economic, \nhumanitarian, and religious concerns shape European attitudes \ntoward asylum seekers. Science 354, 217\u2013222 (2016).\n63.\t Hainmueller, J., Hopkins, D. J. & Yamamoto, T. Causal inference in \nconjoint analysis: understanding multidimensional choices via \nstated preference experiments. Polit. Anal. 22, 1\u201330 (2014).\n64.\t Horiuchi, Y., Markovich, Z. D. & Yamamoto, T. Does conjoint \nanalysis mitigate social desirability bias? SSRN https://papers.\nssrn.com/abstract=3219323 (2020).\n65.\t Bechtel, M. M. & Scheve, K. F. Mass support for global climate \nagreements depends on institutional design. Proc. Natl Acad. Sci. \nUSA 110, 13763\u201313768 (2013).\n66.\t DeBell, M. & Krosnick, J. A. Computing weights for American \nnational election study survey data. ANES Technical Report Series \nAmerican National Election Studies (2009).\n67.\t Stefanelli, A. & Lukac, M. Subjects, trials, and levels: statistical \npower in conjoint experiments. Preprint at SocArXiv https://doi.\norg/10.31235/osf.io/spkcy (2020).\n68.\t Leeper, T. J., Hobolt, S. B. & Tilley, J. Measuring subgroup \npreferences in conjoint experiments. Polit. Anal. 28, 207\u2013221 (2020).\nAcknowledgements\nThis research was presented at the Association of Collegiate Schools \nof Planning Conference in October 2023, and we thank discussant \nM. Coffman for her generous feedback on the manuscript. It was \nalso presented at the 2024 American Political Science Association \nVirtual Research Meeting, and we thank participants of the \u2018Greener \nFutures in Tense Times\u2019 panel for their comments. We are grateful \nfor the assistance of various research assistants and lab managers \nin Princeton\u2019s Behavioral Science for Policy Lab and UBC\u2019s School \nof Community and Regional Planning in making this work possible, \nespecially K. Nichols and M. Lore. We appreciate the time and \nexpertise of the multiple stakeholders that reviewed early drafts of the \nconjoint experiment. Finally, we acknowledge internal funding from \nthe Andlinger Center for Energy and the Environment at Princeton \nUniversity (H.C.).\nAuthor contributions\nH.C., S.M.C., E.U.W. and C.G. conceived of the presented idea and \nresearch questions. H.C., S.M.C. and E.U.W. designed the survey \nexperiment. H.C. carried out data collection and analysis with \nsupport from S.M.C. H.C., S.M.C., E.U.W. and C.G. contributed to the \ninterpretation of results. H.C. took the lead in writing the manuscript. \nS.M.C., E.U.W. and C.G. provided critical feedback and helped shape \nthe research, analysis and manuscript.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41560-024-01603-w.\nCorrespondence and requests for materials should be addressed to \nHolly Caggiano.\nPeer review information Nature Energy thanks Salil D Benegal, Shawn \nHazboun and Adam Mayer for their contribution to the peer review of \nthis work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2024\n\n1\nnature portfolio  |  reporting summary\nMarch 2021\nCorresponding author(s):\nHolly Caggiano\nLast updated by author(s): 6/17/2024\nReporting Summary\nNature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nAll survey data were collected using Qualtrics XM (version 2022)  \nData analysis\nAll data were processed and analyzed using R, version 4.3.2. Code is available on OSF at DOI 10.17605/OSF.IO/WKGT9, https://osf.io/wkgt9/. \nPower analyses were conducted with online power-analysis tool 584 (https://mblukac.shinyapps.io/conjoints-power-shiny/)\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and \nreviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A description of any restrictions on data availability \n- For clinical datasets or third party data, please ensure that the statement adheres to our policy \n \nData supporting this study are openly available from OSF at DOI 10.17605/OSF.IO/WKGT9, https://osf.io/wkgt9/. \n\n2\nnature portfolio  |  reporting summary\nMarch 2021\nHuman research participants\nPolicy information about studies involving human research participants and Sex and Gender in Research. \nReporting on sex and gender\nParticipants in study 1 were asked to self-identify their gender. Data on sex was not collected in this study. Data on gender \nwas not collected in study 2 for local elected officials.\nPopulation characteristics\nSee above\nRecruitment\nParticipants for study 1 were recruited by the survey firm Lucid. Participants for study 2 were recruited by CivicPulse. \nEthics oversight\nBoth surveys were approved by Princeton University\u2019s Institutional Review Board (IRB#14367: \u201cOrigins, Functions, and \nMechanisms of Coalitions for Change: Linking Individual Cognition and Collective Action, PI Elke Weber).\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThis is a quantitative study that conducted a conjoint experiment in two online samples. \nResearch sample\nThe sample for study 1 (n=894) is a representative sample of Pennsylvania, USA residents based on 2019 ACS 1-year estimates, \nbalanced on rural (25%), urban (25%), and suburban (50%) residence. The sample was recruited by the firm Lucid using quota \nsampling on age, gender, ethnicity, and household income to be representative of PA residents (reported in supplemental \ninformation). The sample for study 2 is a sample of local elected officials in Pennsylvania collected by CivicPulse (n=206).  Sample \nrationale: PA is a key transition state. \nSampling strategy\nQuota sampling was used to collect data for Study 1 by the survey firm Lucid. CivicPulse recruited participants for Study 2 through \nrandom sampling from a contact list of officials in the US. For analyses, weights for both samples were calculated with post-\nstratification raking procedures. Sample size for each experiment was determiend using Stafanelli & Lukac's (2020) online power-\nanalysis tool (detailed in methods and supplement). \nData collection\nThe data was collected through anonymous online surveys with blinded treatment conditions. No researcher was present during the \nsurveys. \nTiming\nStudy 1 was fielded from February 7 to February 24, 2022. Study 2 was fielded from September 20 to November 9, 2022. \nData exclusions\nFor the PA resident sample, 57 participants who failed an attention check were excluded from the study. Robustness checks in the \nsupplementary materials include participants that failed attention checks and do not find statistically significant differences between \ngroups.  \nNon-participation\nFor study 1, Lucid only delivered data for completed surveys and did not maintain data on non-participation. For study 2, the \nCivicPulse sample contained 219 participants, but 13 did not complete the conjoint experiment portion of the survey. Since this was \nan online study that was fielded with a participant recruitment firm, we were not able to collect reasons for participants deciding not \nto complete the study. \nRandomization\nIn both experiments, each respondent was asked to choose between five distinct pairs of randomly generated projects. The attribute \nlevels were fully randomized within and across project pairs, ensuring the non-parametric identification of causal effects of the \nattributes.\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \n\n3\nnature portfolio  |  reporting summary\nMarch 2021\nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology and archaeology\nAnimals and other organisms\nClinical data\nDual use research of concern\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\n\n\n    Scientific Research Findings:", "answer": "In a recent survey study in Pennsylvania, we show that community owned projects that create permanent, union\u2011wage jobs increase support by larger margins than other project characteristics, like distance from residential areas. Regardless of their political party affiliation, respondents in the general public had similar preferences for project characteristics, which suggests that direct benefits may help build bipartisan support. Respondents prefer solar projects to wind, nuclear, and natural gas power plants with carbon capture and storage. Local elected officials, however, misperceive the preferences of their constituents, underestimating support for solar energy as well as the importance of job loss and creation. These results shed light on preferences in Pennsylvania, but they emphasize the general importance of considering variation in project features, including conferred community benefits, to build public support.", "id": 6}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 9 | August 2024 | 955\u2013963\n955\nnature energy\nhttps://doi.org/10.1038/s41560-024-01546-2\nArticle\nEvaluating community solar as a measure to \npromote equitable clean energy access\nEric O\u2019Shaughnessy\u2009\n\u200a\u20091\u2009\n, Galen Barbose\u2009\n\u200a\u20091, Sudha Kannan\u2009\n\u200a\u20092 & \nJenny Sumner2\nRooftop and community solar are alternative product classes for residential \nsolar in the United States. Community solar, where multiple households buy \nsolar from shared systems, could make solar more accessible by reducing \ninitial costs and removing adoption barriers for renters and multifamily \nbuilding occupants. Here we test whether community solar has expanded \nsolar access in the United States. On the basis of a sample of 11 states, we \nfind that community solar adopters are about 6.1 times more likely to live \nin multifamily buildings than rooftop solar adopters, 4.4 times more likely \nto rent and earn 23% less annual income. We do not find that community \nsolar expands access in terms of race. These differences are driven, roughly \nevenly, by inherent differences between the two solar products and by \npolicies to promote low-income community solar adoption. The results \nsuggest that alternative solar products can expand solar access and that \npolicy could augment such benefits.\nNearly four million residential electricity customers had adopted \nrooftop solar photovoltaics in the United States by the end of 2022  \n(ref. 1). Rooftop solar adopters tend to be more affluent than the  \ngeneral population, are less likely to rent and are less likely to self iden-\ntify as a racial minority2\u20135. Rooftop solar adoption inequity reflects \nunequal access to solar across different demographic groups6. Vari-\nous barriers restrict solar access for low- and moderate-income (LMI) \nhouseholds (for example, high up-front costs to purchase solar systems \noutright), for renters (for example, split incentives) and for multifamily \nbuilding occupants (for example, shared ownership of rooftop spaces)7. \nWhereas adoption inequity is common among emerging technologies8, \nrooftop solar adoption inequity could pose unique challenges to clean \nenergy transitions and grid decarbonization9,10. A growing number of \npolicies seek to expand solar access11, that is, increase adoption among \ndemographic groups historically underserved by rooftop solar.\nPrevious research suggests that alternative solar products can \nexpand access to solar. Specifically, the development of solar leasing \nmodels with minimal up-front costs has driven a more equitable expan-\nsion of rooftop solar12,13 and the recent emergence of solar loans may \nsimilarly address up-front cost barriers. However, rooftop solar remains \nlargely inaccessible to renters and families living in multifamily hous-\ning14,15. Another alternative class of solar products in the United States is \ncommunity solar, wherein multiple customers buy output from a single \nsolar system16. As with leasing, community solar typically entails no or \nminimal up-front costs. Unlike rooftop solar, community solar poses \nno specific barriers to adoption for renters or multifamily building \noccupants. As a result, community solar is often theorized to promote \nmore equitable solar access15,17\u201322 and is increasingly integrated into US \nsolar adoption equity policies11. The federal Inflation Reduction Act \nincludes tax credits for projects serving LMI communities or custom-\ners, and at least 17 states have incentives or regulations that promote \nLMI community solar23\u201325.\nThe impact of community solar on clean energy access has not \nbeen empirically evaluated. In this study, we fill this research gap by \nanalysing the demographic profiles of rooftop and community solar \nadopters to determine whether existing community solar projects \nhave promoted more equitable adoption. We quantitatively assess \nhousehold-level demographic data of community solar and rooftop \nsolar adopters, as does other research documenting the demographic \nprofiles of rooftop solar adopters2,4,5,26. We explore how the two cus-\ntomer groups vary in terms of median income levels, housing tenure \n(whether adopters own or rent their homes), housing type (single or \nmultifamily) and race. Our key contribution is a quantitatively rigorous \nanalysis of the prevailing hypothesis that community solar promotes \nReceived: 2 November 2023\nAccepted: 2 May 2024\nPublished online: 3 June 2024\n Check for updates\n1Lawrence Berkeley National Laboratory, Berkeley, CA, USA. 2National Renewable Energy Laboratory, Golden, CO, USA. \n\u2009e-mail: eoshaughnessy@lbl.gov\n\nNature Energy | Volume 9 | August 2024 | 955\u2013963\n956\nArticle\nhttps://doi.org/10.1038/s41560-024-01546-2\nthat at least some community solar adopters would not otherwise have \nadopted rooftop solar (Inference for solar access in Methods). We esti-\nmate all statistics based on comparisons within states to ensure sample \nindependence. We focus our analyses on 11 states where we have at least \n100 records for both adopter types (Table 2). The 11 states include some \nof the largest community and rooftop solar markets in the United States \nand comprise more than half of community solar capacity installed to \ndate33. Still, some caution is merited in extrapolating results from this \ngeographic subsample (\u2018Limitations\u2019 in Methods).\nWe explore differences in demographic characteristics using \none-sided tests for the following hypotheses: community solar adop-\nters earn less, on average, than rooftop solar adopters; community \nsolar adopters are more likely to rent than rooftop solar adopters; \ncommunity solar adopters are more likely to live in multifamily build-\nings than rooftop solar adopters; and community solar adopters are \nmore likely to identify as people of colour or Hispanic than are rooftop \nsolar adopters.\nThroughout this paper, solid points in figures indicate statistically \nsignificant results according to one-sided statistical tests, whereas \nempty points indicate insignificant results.\nThe data firmly support hypotheses 1\u20133 defined above (Fig. 1). \nWeighting the differences by state sample sizes, the data suggest \nthat community solar adopters are about 6.1 times more likely to live \nin multifamily buildings than rooftop solar adopters, 4.4 times more \nlikely to rent and earn about 23% less in annual income. At the same \ntime, the data suggest that community solar adopters are not demo-\ngraphically representative of the general population. In most states, \ncommunity solar adopters earn more than average and are less likely \nto rent and live in multifamily buildings than the general population. \nThat is, community solar expands access relative to rooftop solar \nbut is still inequitable relative to the general population. Differences \nin race are more ambiguous. The comparative statistics are mostly \nstatistically insignificant according to the one-sided test that com-\nmunity solar adopters are more likely to identify as people of colour or \nHispanic. Across all the states in the sample, rooftop solar adopters are \nabout twice as likely as community solar adopters to identify as Asian/\nAsian American or Black and about three times as likely to identify as \nHispanic (Fig. 2).\nmore equitable solar access. Whereas that hypothesis is often taken \nfor granted, the effects of community solar on clean energy access \nthus far lacks empirical evidence27,28, and there are several reasons to \nquestion whether community solar necessarily expands access. LMI \nparticipation can increase community solar costs15, largely because \nLMI customers can be more costly to acquire29,30. Profit-maximizing \ncommunity solar providers may thus prioritize marketing to relatively \naffluent customers, consistent with evidence from rooftop solar mar-\nkets31. Further, bill management and customer turnover represent \nsubstantial costs to community solar providers32. Community solar \nproviders thus face economic incentives to minimize the number of \ncustomers by maximizing energy sold per customer. Community solar \nproviders may thus avoid lighter energy users such as LMI households \nand multifamily building occupants. Providers may also perceive that \nLMI customers and renters pose higher turnover risks.\nNonetheless, we find that multifamily building occupants, renters, \nand\u2014to a lesser extent\u2014LMI households are significantly more likely to \nadopt community solar than rooftop solar. These results suggest that \ncommunity solar effectively expands access to demographic groups \nunderserved by rooftop solar. We do not, however, find evidence that \ncommunity solar expands access in terms of race. Further, we show that \nthe access benefits of community solar stem from two drivers: inherent \ndifferences in the community solar model that reduce barriers to adop-\ntion and community solar policies designed to promote LMI adoption.\nRooftop and community solar adopter \ndemographics\nWe draw on multiple sources to build rooftop and community solar \ndatasets with household-level estimates for income, housing type, \nhousing tenure and race (Table 1 and Methods; summary statistics in \nSupplementary Table 1). The community solar data represent custom-\ners of community solar projects as defined by the US Department of \nEnergy: solar projects where financial benefits flow to multiple custom-\ners within a defined geographic area. Income and race estimates are \navailable for all records, whereas housing type and tenure are incom-\nplete. The sample sizes associated with each variable in each analysis \nare described in figure captions.\nWe estimate comparative statistics of adopter demographics \nusing Wilcoxon tests (comparison of median incomes) or Pearson \nChi-squared tests (comparisons of categorical variables) (Methods). \nThe comparative statistics describe differences in adoption choices \nacross different demographic groups. Significant differences provide \nevidence that community solar expands solar access under the premise \nTable 1 | Data sources\nSource\nDescription\nN\nRooftop solar\nSolar demographics5\nDatabase of rooftop solar adopters \nwith household-level demographic \nvariables\n102,974\nCommunity solar\nState programme data\nData were obtained directly from \nstate community solar programmes in \nIllinois, Maine, New York and Oregon. \nThe data represent customers that \nwere actively enroled as of 2023.\n41,323\nSharing the Sun \ndeveloper data\nData obtained from community solar \ndevelopers as part of the National \nRenewable Energy Laboratory\u2019s \n(NREL) Sharing the Sun project. NREL \npublishes project-level data, but \ncustomer-level data are considered \nproprietary and are not publicly \navailable. Data reflect customers \nactively enroled as of 2023.\n37,391\nTable 2 | Solar sample sizes and data sources by state\nState\nCommunity solar \nsample size\nRooftop \nsolar \nsample size\nCommunity solar data \nsource(s)\nColorado\n398\n21,472\nNREL\nIllinois\n21,180\n10,774\nState of Illinois \n(Adjustable Block \nprogramme, Solar For \nAll programme), NREL\nMaine\n19,907\n1,322\nCentral Maine Power \nCompany, Versant \nPower, NREL\nMaryland\n7,779\n5,582\nNREL\nMassachusetts\n1,814\n10,611\nNREL\nMinnesota\n465\n4,319\nNREL\nNew Jersey\n519\n16,077\nNREL\nNew York\n23,375\n15,565\nNew York State  \nEnergy Research  \nand Development \nAuthority, NREL\nOregon\n2,033\n10,185\nOregon Energy Trust, \nNREL\nRhode Island\n876\n4,881\nNREL\nWashington, DC\n368\n2,186\nNREL\n\nNature Energy | Volume 9 | August 2024 | 955\u2013963\n957\nArticle\nhttps://doi.org/10.1038/s41560-024-01546-2\nWe use a conditional probability model to compare how the differ-\nent demographic factors explain household adoption decisions (Meth-\nods). The models describe the relative power of each demographic \nfactor in predicting whether a household is a community or rooftop \nsolar adopter, conditioned on correlation with the other factors (for \nexample, multifamily building occupants are more likely to rent than \nsingle-family occupants; Supplementary Table 4). The model sug-\ngests that the strongest predictors of adoption choices are race and \nhousing tenure (Fig. 3). Further, we use Akaike Information Criterion \n(AIC) scores to assess the prediction accuracy of model variations \nincluding different combinations of the demographic factors (Supple-\nmentary Table 6). The AIC scores likewise suggest that race and hous-\ning tenure are the most predictive variables. As with the comparative \nstatistics, the conditional impacts of race are directionally opposite \nto our hypothesis. In the remaining discussion, we generally focus on \nthe demographic differences that confirmed a priori hypotheses and \nreturn to a discussion of race in Conclusions.\nOverall, the results show that multifamily building occupants, \nrenters, and\u2014to a lesser extent\u2014LMI households are significantly more \nlikely to adopt community solar than rooftop solar. These results \na\n0\n90\n180\nCO\nDC\nIL\nMA\nMD\nME\nMN\nNJ\nNY\nOR\nRI\nb\n0\n20\n40\n60\nCO\nDC\nIL\nMA\nMD\nME\nMN\nNJ\nNY\nOR\nRI\nc\n0\n20\n40\n60\n80\nDC\nIL\nMA\nMD\nME\nMN\nNJ\nNY\nOR\nRI\nd\n0\n35\n70\nCO\nDC\nIL\nMA\nMD\nME\nMN\nNJ\nNY\nOR\nCO\nRI\nCommunity solar\nRooftop solar\nStatewide\nState\nState\nState\nState\nPeople of colour or Hispanic (%)\nRenters (%)\nMedian income (\u00d7 US$1,000)\nIn multifamily buildings (%)\nFig. 1 | Comparisons of demographic characteristics of community and \nrooftop solar adopters. a, Median income levels (N\u2009=\u2009181,688). b, Percentage \nof renters (N\u2009=\u2009147,881). c, Percentage of multifamily building occupants \n(N\u2009=\u2009181,672). d, Percentage of people of colour or Hispanic (N\u2009=\u2009181,688). Solid \ndiamonds indicate statistically significant (P\u2009<\u20090.05) results based on one-sided \nWilcoxon tests (income) or Pearson Chi-squared tests (all other variables). \nStatewide estimates for race are omitted for reasons explained in Methods. \nNumerical results in Supplementary Table 2. CO, Colorado; DC, Washington, DC; \nIL, Illinois; MA, Massachusetts; MD, Maryland; ME, Maine; MN, Minnesota; NJ, \nNew Jersey; NY, New York; OR, Oregon; RI, Rhode Island.\na\n0\n7\n14\nCO\nDC\nIL\nMA\nMD\nME\nMN\nNJ\nNY\nOR\nRI\nAsian/Asian\nAmerican (%)\nBlack (%)\nHispanic (%)\nb\n0\n30\n60\nCO\nDC\nIL\nMA\nMD\nME\nMN\nNJ\nNY\nOR\nRI\nc\n0\n15\n30\nCO\nDC\nIL\nMA\nMD\nME\nMN\nNJ\nNY\nOR\nRI\nCommunity solar\nState\nState\nState\nRooftop solar\nFig. 2 | Comparisons of race of rooftop and community solar adopters. a, Percentage of Asian or Asian American people. b, Percentage of Black people.  \nc, Percentage of Hispanic people. Solid diamonds indicate statistically significant (P\u2009<\u20090.05) results based on one-sided Pearson Chi-squared tests. N\u2009=\u2009181,688. \nNumerical results in Supplementary Table 3.\n\nNature Energy | Volume 9 | August 2024 | 955\u2013963\n958\nArticle\nhttps://doi.org/10.1038/s41560-024-01546-2\nsuggest that community solar expands solar access under the reason-\nable assumption that at least some of those adopters would not or could \nnot have adopted rooftop solar. That assumption probably holds for \nmost multifamily building occupants and renters, groups that face \nsubstantial barriers to rooftop solar adoption7. Many LMI households \nface those same barriers, given that LMI households are more likely \nto rent or live in multifamily buildings than are non-LMI households \n(Supplementary Table 4). In contrast, the emergence of rooftop solar \nleasing partly addressed adoption cost barriers for LMI homeowners. \nThese differences in adoption barriers may partly explain the observed \ndifferences in the demographic factors. To further explore this hypoth-\nesis, we separately identified rooftop solar system owners and lessees \nin the five states where leasing is allowed, and the data allowed us to \nidentify lessees. In three of the five states, rooftop solar lessee incomes \nmore closely resemble the typical incomes of community solar adop-\nters than rooftop solar system owners (Fig. 4). However, in those same \nthree states, rooftop solar lessees are not substantially more likely to \nrent or live in multifamily housing. The data suggest that rooftop solar \nleasing addresses barriers to adoption for LMI homeowners but does \nnot effectively address housing barriers. The renter and multifamily \nhousing market is thus the clearest market niche for community solar \nto address. Further, the data suggest that rooftop solar lessees are \nconsistently more likely to identify as people of colour or Hispanic. \nThat result suggests that rooftop solar leasing addresses adoption \nbarriers for people of colour and Hispanic households in ways that are \nnot replicated by community solar.\nInherent and policy impacts\nThe preceding analyses demonstrate that community solar adopters \ndiffer demographically from rooftop solar adopters. These differ-\nences could reflect some combination of differences between the \ntwo solar products\u2014what we refer to as \u2018inherent impacts\u2019\u2014and dif-\nferences between rooftop and community solar policies\u2014what we \nrefer to as \u2018policy impacts\u2019. LMI solar policies typically take the form of \ncustomer-level incentives (for example, reduced enrolment costs for \nLMI adopters), project-level incentives (for example, financial incen-\ntives provided to project managers for enroling LMI customers) or \ncarve outs that reserve shares of capacity or output for LMI customers25. \nSupplementary Table 8 provides a summary of LMI community solar \nprogrammes for the 11 states in our study sample. Isolating inherent \nfrom policy impacts would be useful for understanding how effectively \ncommunity solar promotes access without additional policy support \nand how impactful community solar policies have been in expanding \nsolar access.\nWe analyse inherent and policy impacts by comparing the demo-\ngraphic profiles of adopters that participated in LMI community \nsolar programmes (programme participants) and those who did not \n(non-participants). We can only precisely distinguish participants \nfrom non-participants in three states: Illinois, New York and Oregon \n(Methods). As a robustness check, we obtain similar results based on \nan analysis of inferred LMI programme participation in three other \nstates (Methods and Supplementary Fig. 1). Demographic differences \nbetween non-participant community solar and rooftop solar adopters \nprovide evidence of inherent impacts, given that LMI policies did not \ndirectly affect non-participants. Demographic differences between \nparticipant and non-participant community solar adopters provide \nevidence of policy impacts. The accuracy of that evidence depends \non reasonable assumptions around the share of participants who \nwould have otherwise adopted community solar, also known as free \nriding (Methods).\nParticipants earn significantly less and are more likely to rent and \nlive in multifamily housing than non-participants (Fig. 5). Participants \nare also significantly more likely to identify as people of colour or \nHispanic than non-participants. These results imply that community \nsolar policies have effectively expanded access in all four demo-\ngraphic dimensions. The results suggest that various approaches \ncan effectively expand access, as evidenced by the distinct LMI pro-\ngramme structures of the three states in the analysis: Illinois subsi-\ndizes LMI customer participation, whereas New York incentivizes LMI \ncustomer enrolment at the project level and Oregon administers an \nLMI carve out (Supplementary Table 8). The analysis likewise pro-\nvides evidence of the inherent impacts of community solar on adop-\nter demographics in most cases. Non-participant community solar \nadopters earn significantly less than rooftop solar adopters in Illinois \nand New York. However, non-participant community solar adopters \nearn slightly more in Oregon, suggesting that income differences in \nOregon are fully explained by policy. Differences in housing type and \ntenure remain significant for non-participants in Illinois and Oregon \n(consistent with inherent impacts) but are rendered insignificant in \nNew York. However, a robustness check to account for exceptional \ncircumstances in New York (Methods) suggests that demographic \ndifferences in that state are more similar to other states when account-\ning for geographic differences in rooftop and community solar siting \n(Supplementary Table 10).\nTo generate rough estimates of the contributions of inherent and \npolicy impacts, we calculate the effects of removing participants on \noverall demographic differences. For instance, in Illinois, the median \nincome difference between all community and rooftop solar adop-\nters is US$13,000, whereas the median income difference between \nnon-participants and rooftop adopters is US$8,000, such that policy \naccounts for about US$5,000 or 38% of the difference (assuming mini-\nmal free riding). Under that approach, if free riding is trivial, the results \nsuggest that policy explains around 67% of income differences between \ncommunity and rooftop solar adopters, 43% of the differences in hous-\ning tenure and 23% of the differences in housing type on average across \nthe three states. The fact that policy appears to contribute more to \nincome differences is not surprising given that the policies evaluated \nhere target income levels. Thus, broadly speaking, the data suggest that \npolicy impacts are the primary driver of income differences, whereas \ninherent impacts are the primary driver of differences in housing \ntype, and both impacts contribute roughly evenly to differences in \nhousing tenure.\n\u00df = 1.49\nSE = (0.03)\n0.75\n(0.02)\n0.64\n(0.02)\n\u22121.67\n(0.02)\n\u221240\n\u221220\n0\n20\n40\nRenter\nMultifamily\nLMI\nPeople of colour\nor Hispanic\nDemographic factor\nAdjusted logit coeficient\n(~percentage points)\nFig. 3 | Conditional associations between demographic factors and solar \nadoption choices. Results based on coefficients (\u00df) and standard errors (SE) \nfrom model defined in equation (1) in Methods. The dependent variable is a \nbinary variable taking on a value of 1 for community solar adopters and 0 for \nrooftop solar adopters. LMI for the purposes of this figure refers to households \nearning less than the state\u2019s median income. For simplicity, we convert the \ncoefficients to percentage point terms using the approximation of multiplying \nthe coefficients (\u00df) by 100 then dividing by four. The blue columns represent \nthe adjusted regression coefficients; the grey bars represent 95% confidence \nintervals based on regression standard errors. Full numerical results are provided \nin Supplementary Table 5. N\u2009=\u2009147,874. SE, standard error.\n\nNature Energy | Volume 9 | August 2024 | 955\u2013963\n959\nArticle\nhttps://doi.org/10.1038/s41560-024-01546-2\nConclusions\nOur results suggest that certain demographic groups are significantly \nmore likely to adopt community solar than rooftop solar. Specifically, \nwe find that community solar adopters in 11 states are about 6.1 times \nmore likely to live in multifamily buildings, 4.4 times more likely to rent, \nand earn 23% less annual income than rooftop solar adopters, on aver-\nage. These results suggest that community solar has expanded solar \naccess by removing adoption barriers for renters and multifamily build-\ning occupants, and to a lesser extent for LMI households. We cautiously \ninfer from these results that other alternative clean energy adoption \nmodels that remove such adoption barriers could similarly expand \nclean energy access. For instance, community wind programmes\u2014\nthough less common\u2014could provide similar clean energy access ben-\nefits as community solar.\nThough community solar expands access relative to rooftop solar, \ncommunity solar adopters tend to earn more than the broader popu-\nlation and are less likely to rent and live in multifamily housing. This \noutcome is not surprising given the economic incentives that commu-\nnity solar providers face (see Introduction). It is likely that community \nsolar will become more equitable over time, both because of the broad \ntendency of emerging technologies to diffuse to underserved markets \nover time8 and because of increasingly ambitious community solar \npolicies to expand access34.\nWe do not find evidence that community solar has expanded \nsolar access in terms of race. Indeed, we find that people of colour and \nHispanic households have been less likely to adopt community solar \nthan rooftop solar. The reason for these racial differences is unclear. \nOne hypothesis is that some households may be suspicious of market-\ning for relatively unknown products such as community solar29. Such \nsuspicions may be particularly prevalent among people of colour and \nHispanic households that are more likely to be victimized by fraudulent \nmarketing than White households35. At the same time, that hypothesis \nwould need to be squared with the apparent acceptance of rooftop \nsolar leasing by people of colour and Hispanic adopters (see Fig. 4). \nFuture research could explore how differences in marketing, customer \nperceptions, or other factors could explain racial differences across \nthe two solar products.\nOur data suggest that inherent differences across the two solar \nproducts largely explain historical differences in housing type and \ntenure between community and rooftop solar adopters. Evidence of \ninherent impacts on housing type and tenure is generally encouraging \nfrom a policy perspective. Inherent impacts suggest that policymak-\ners could expand solar access by creating a basic infrastructure for \ncommunity solar, such as virtual net metering, even without specific \nmeasures to promote equity, such as LMI incentives or carve outs. \nNonetheless, we find evidence that LMI community solar programmes \nincrease LMI adoption by subsidizing enrolment for LMI customers, \nincentivizing project managers to acquire LMI customers, or reserving \nshares of project capacity for LMI customers. Our results also suggest \nthat LMI community solar programmes expand access to people of \ncolour and Hispanic households. Overall, the results suggest that \ntargeted LMI community solar policy can augment the access benefits \nof community solar.\nWe conclude with suggestions for further research. Here, we have \nidentified the impacts of community solar on adopter demographics \na\n0\n95\n190\nDC\nIL\nMA\nNJ\nNY\nMedian income\n(\u00d7 US$1,000)\nRenters (%)\nMultifamily (%)\nPeople of colour or\nHispanic (%)\nb\n0\n19\n38\nDC\nIL\nMA\nNJ\nNY\nc\n0\n40\n80\nDC\nIL\nMA\nNJ\nNY\nd\n0\n40\n80\nDC\nIL\nMA\nNJ\nNY\nCommunity solar\nState\nState\nState\nState\nRooftop solar owners\nRooftop solar lessees\nFig. 4 | Comparisons of demographic characteristics across three solar \nproducts. a, Median income levels (N\u2009=\u200947,256 community, 23,133 rooftop \nowners, 18,174 rooftop lessees). b, Percentage of renters (N\u2009=\u200936,391 community, \n19,058 rooftop owners, 15,558 rooftop lessees). c, Percentage of multifamily \nbuilding occupants (N\u2009=\u200947,242 community, 23,133 rooftop owners, 18,174 \nrooftop lessees). d, Percentage of people of colour or Hispanic (N\u2009=\u200947,256 \ncommunity, 23,133 rooftop owners, 18,174 rooftop lessees). Solid diamonds \nindicate statistically significant (P\u2009<\u20090.05) differences between community \nsolar adopters and rooftop solar lessees based on one-sided Wilcoxon tests \n(income) or Pearson Chi-squared tests (all other variables). Numerical results in \nSupplementary Table 7.\na\n0\n90\n180\nIL\nNY\nOR\nMedian income (\u00d7 US$1,000)\nIn multifamily buildings (%)\nRenters (%)\nPeople of colour or\nHispanic (%)\nb\n0\n15\n30\nc\n0\n25\n50\nIL\nNY\nOR\nState\nd\n0\n30\n60\nIL\nNY\nOR\nState\nState\nState\nCommunity solar:\nLMI programme participants\nCommunity solar:\nnon-participants\nRooftop solar\nIL\nNY\nOR\nFig. 5 | Demographic characteristics in community solar and rooftop solar \nsubsamples. a, Median income levels (N\u2009=\u20092,999 participants, 19,191 non-\nparticipants, 36,110 rooftop). b, Percentage of renters (N\u2009=\u20091,817 participants, \n13,317 non-participants, 29,816 rooftop). c, Percentage of multifamily building \noccupants (N\u2009=\u20092,991 participants, 19,185 non-participants, 36,110 rooftop). \nd, Percentage of people of colour or Hispanic (N\u2009=\u20092,999 participants, 19,191 \nnon-participants, 36,110 rooftop). Solid points indicate statistically significant \n(P\u2009<\u20090.05) differences based on one-sided Wilcoxon tests (income) or Pearson Chi-\nsquared tests (all other variables). Numerical results in Supplementary Table 9.\n\nNature Energy | Volume 9 | August 2024 | 955\u2013963\n960\nArticle\nhttps://doi.org/10.1038/s41560-024-01546-2\nstemming from policies and broad differences between the prod-\nucts. Future research could explore which specific aspects of different \nsolar products most effectively promote solar access. For instance, \ndo community solar features such as cancellation terms and transfer-\nability (whether customers can keep community solar subscriptions \nwhen changing addresses) affect adopter demographics? Similarly, \nfuture research could analyse how different community solar LMI \npolicies affect adopter demographics. Our results suggest that vari-\nous approaches can be effective\u2014such as customer-level incentives, \nproject-level incentives, and LMI carve outs\u2014but future research could \nexplore whether certain approaches are more cost-effective than \nothers. Further, future research could explore how LMI community \nsolar programmes could optimize the benefits that accrue to LMI \ncustomers, such as through guaranteed bill savings and customer \nprotections. Finally, our results raise many questions about access \nto solar across race. Future research could explore why rooftop and \ncommunity solar appear to be reaching distinct racial communities \nand why policy appears to be particularly critical for expanding solar \naccess to racial minorities.\nMethods\nData\nOur data sources are defined in Table 1. We used home addresses to \nmatch adopter records at the address level to household-level vari-\nables for income, housing type and housing tenure purchased from \nExperian. Household-level incomes are estimated from individual- and \nhousehold-level variables and a proprietary model developed by Expe-\nrian, an empirically driven algorithm built to effectively predict sample \nstatistics (for example, means, medians) of solar adopter incomes. The \nincome estimates have been empirically validated in previous research \non solar adopters13. The housing type variable is based on US Postal \nService data. The housing tenure variable is based primarily on tax \nassessment and deed data. We predicted household-level racial char-\nacteristics using the \u2018wru\u2019 package in R36. The wru package estimates \ncontinuous probabilities for household race/ethnicity in five catego-\nries (Asian or Asian American, Black, Hispanic, White, other) based on \nthe surname of the household\u2019s Census tract and the surname of the \nhead of household. In cases where the surname was unavailable (12% \nof records), wru predicts race based only on the Census tract. Remov-\ning these tract-only predictions from the data does not substantially \naffect the results (Supplementary Fig. 2). The wru algorithm has been \nempirically validated to predict household race with around or above \n80% accuracy5,37. We converted the continuous probabilities to a binary \npeople of colour or Hispanic variable score based on whether some race \nother than White was assigned the greatest probability. We compare \nsolar adopter demographics to the general population using state-level \ndemographic statistics from the US Census American Community \nSurvey. However, we omit the statewide comparison for race because \nthe continuous race probabilities estimated by wru cannot be mean-\ningfully compared to the self-identified races reported in Census data.\nThe accuracy of sample statistics for the income and race estimates \nis subject to modelling error. Nonetheless, potential modelling error \nis not a material concern to the validity of the comparative statistics. \nModelling bias may yield random erroneous results in certain states, \nbut the directional consistency of comparative statistics across states \nsuggests that those results are robust to any potential modelling bias.\nUnconditional demographic differences\nWe test observed (unconditional) differences for the hypotheses \ndescribed in the main text. We use Wilcoxon rank-sum tests to test \nhypotheses for household incomes and Pearson \u03c72 tests for the cat-\negorical variables (housing tenure, housing type, race). We compare \nmedians because household income levels are not normally distrib-\nuted. We make two adjustments to ensure independence between the \ncomparison groups. First, we estimate comparative statistics within \nstates to ensure that the community and rooftop solar data are pulled \nfrom the same geographic subsamples. Second, we restrict the rooftop \nsolar adopter data to systems installed in 2022 to account for the fact \nthat our community solar data reflect samples of customers enroled in \ncommunity solar in 2022 or 2023. That temporal misalignment matters \nbecause rooftop solar adoption has become more demographically \nequitable over time5. Both restrictions are reflected in the sample sizes \nreported in Table 1.\nConditional probability model\nTo isolate the relative effects of household demographics on adoption \nchoices we use the following logit model:\np (CS) = \u03b1 + D\u03b2 + S + \u03b5\n(1)\nWhere p(CS) is the probability that a household is a community solar \nadopter (as opposed to a rooftop adopter), D is a vector of dummy \nvariables for the four demographic dimensions and S is a state random \neffect. We convert the income variable into a dummy value by bifurcat-\ning the records into households that earn more or less than the state \nmedian income. The coefficient of interest is \u03b2, which represents the \nstatistical association between household demographics and the \nhousehold\u2019s adoption choice. Note that the model is not designed for \ncausal inference. Household adoption choices are probably driven by \nnumerous idiosyncratic factors that could correlate with the demo-\ngraphic factors. The purpose of this model is to compare the rela-\ntive weights of the \u03b2 coefficients to understand which demographic \ndimensions are most strongly associated with household adoption \nchoices. We use state random effects to account for the possibility that \ncommunity solar has distinct impacts on adopter demographics in \ndifferent states with distinct policy contexts. In addition to comparing \nthe coefficients, we also implement variations of the model in equa-\ntion (1) with different combinations of the demographic factors. We \nthen compare Akaike Information Criterion (AIC) values across those \nmodels (Supplementary Table 6). The AIC is a metric that simultane-\nously measures prediction accuracy while penalizing models with more \nvariables. The AIC comparisons provide another way of comparing \nthe relative contributions of demographic differences to household \nadoption choices.\nInference about solar access\nThe comparative statistics and conditional model describe differences \nin household adoption choices. Those differences provide evidence \nthat community solar expands access (that is, promotes adoption \namong demographic groups that are underserved by rooftop solar) \nunder the premise that at least some community solar adopters would \nnot or could not have adopted rooftop solar. That is, the premise is that \nat least some community solar adopters face constraints that make \ncommunity solar the only viable option for solar adoption. Under \nthat premise, systematic differences in adopter choices partly reflect \nsystematic differences in adoption constraints, not just heterogene-\nous preferences, and community solar expands access by mitigating \nthose constraints. That premise is supported by evidence of barri-\ners to rooftop solar adoption for specific demographic groups such \nas multifamily building occupants, renters, LMI households4,7 and  \nracial minorities2.\nAnalysis of LMI programme participants and non-participants\nState programmes in Illinois, New York and Oregon provided identifiers \nfor LMI programme participants. Participants may have received finan-\ncial incentives to participate or were otherwise prioritized for adoption \nto comply with state LMI carve outs. We analyse evidence of inherent and \npolicy impacts by distinguishing participants from non-participants \nin each state: Illinois (N\u2009=\u2009918 participants; 11,143 non-participants), \nNew York (1,363 participants; 6,733 non-participants) and Oregon  \n\nNature Energy | Volume 9 | August 2024 | 955\u2013963\n961\nArticle\nhttps://doi.org/10.1038/s41560-024-01546-2\n(718 participants; 1,315 non-participants). For each state, we likewise \ncreate subsets of rooftop solar adopters who did not receive any LMI \nincentives for rooftop solar adoption. The accuracy of that analysis \ndepends on how many participants would have adopted community \nsolar without LMI programme benefits, a concept known as free riding. \nAt one extreme, inherent and policy impacts can be precisely identified \nif free riding is non-existent. However, if free riding is common, then \nthe analysis would tend to understate inherent impacts and overstate \npolicy impacts. There are at least three reasons to assume that free \nriding is not common in LMI community solar programmes. First, \nevidence that LMI programme participants are relatively difficult \nand costly to acquire29,30 suggests that participants would not other-\nwise have adopted. Second, LMI community solar programmes use \nsimilar eligibility criteria as LMI rooftop solar incentives, and available \nevidence suggests that free riding in LMI rooftop solar programmes \nis infrequent38. Third, the data suggest that participants vary signifi-\ncantly from non-participants in terms of race, indicating that LMI \nprogramme benefits are reaching a distinct population of adopters. \nLimiting the data (N\u2009=\u20097,492) to adopters earning less than 80% of their \nstates\u2019 median income (a common threshold for identifying LMI house-\nholds), participants are about 1.8 times more likely to identify as people \nof colour or Hispanic than non-participants (t\u2009=\u200914.0, two-sided). These \ndifferences suggest that LMI programmes are reaching a distinct popu-\nlation of LMI households than those adopting without programme \nbenefits. Still, some degree of free riding probably exists, meaning \nthat policy impacts are imprecisely identified. Further, at least some \ncommunity solar adopters that were eligible for LMI programmes may \nhave adopted without participating in LMI programmes. Such adopters \nwould have been, by definition, free riders if they had participated in \nthe LMI programmes, such that their classification as non-participants \nwould not affect the results.\nRobustness check of New York data sample\nAbout 75% of rooftop solar adopters in New York in the data reside \nin the relatively densely populated \u2018downstate\u2019 region of the greater \nNew York City area, whereas about 95% of community solar capac-\nity has been deployed in the less densely populated \u2018upstate\u2019 region \noutside the New York City area39. That geographic skew partly drives \nthe exceptional results for demographics differences in New York. For \ninstance, as shown in Fig. 1, the geographic skew partly explains why \nNew York rooftop solar adopters are about as likely to live in multifamily \nbuildings as community solar adopters. That geographic skew is a valid \ncomponent of the comparison: the fact that community solar projects \nare predominantly deployed in less densely populated areas affects \nthe degree to which community solar can expand solar access to mul-\ntifamily building occupants and renters. However, the valid impacts of \nthat geographic skew are exacerbated by the fact that community solar \ndata for downstate projects (about 5% of capacity) were not available \nfor this study, creating a small source of data bias. To test the robust-\nness of the New York results to this geographic skew and data bias, we \nimplement the analyses behind Figs. 1 and 5 while limiting the New York \ndata to rooftop and community solar adopters living in upstate New \nYork (specifically, all counties north of Rockland and Westchester). \nThe results of those analyses are provided in Supplementary Table 10.\nImputed carve outs\nIn addition to the three states depicted in the analysis in Fig. 5, two \nstates in our data (Colorado and Maryland) have specified percentage \npoint minimum carve outs for LMI customers. We impute an effective \nLMI carve out for Massachusetts using data compiled in the Low Income \nFinancing and Transactions (LIFT) Solar Toolkit40, downloaded on 14 \nApril 2023. The LIFT Solar Toolkit identifies project capacity reserved \nfor LMI customers. We divide the reserved LMI capacity by the total, \ncumulative community solar capacity deployed in the state based on \ndata from Connelly30. As a point of reference, the same method yielded \nan imputed carve out of 5.9% in Colorado, close to the state\u2019s mandated \ncarve out of 5%. For all three states, we isolate the bottom end of the \ncommunity solar adopter income distribution in proportion to the \ncarve out. For instance, Colorado requires that LMI households account \nfor at least 5% of community solar customers, such that in that state we \nisolate the 5% of community solar adopters with the lowest incomes \nas \u2018below\u2019 the carve out. In effect, this represents the most optimistic \nassumption for the efficacy of the LMI carve out and would thus tend \nto minimize any residual inherent impact. The results of this analysis \nare provided in Supplementary Fig. 1, where points \u2018below carve out\u2019 \nrepresent adopters below the implied carve out on the income distri-\nbution and \u2018above carve out\u2019 points represent adopters above those \nimplied carve outs.\nLimitations\nTwo limitations noted in the main text are worth expanding upon. First, \nwe analyse a geographically restricted sample of 11 states. Those 11 \nstates comprise some of the country\u2019s largest community solar mar-\nkets, accounting for around 57% of cumulatively deployed community \nsolar capacity by the end of 2022 (ref. 33). The 11 states are geographi-\ncally diverse, with at least one state in each region of the United States \nexcept the Southeast. The 11 states also reflect a diversity of commu-\nnity solar policy contexts25. The policy and geographic diversity of our \nsample and the consistency of results across these states suggest that \nour results can be generally extrapolated to community solar markets \nin other states. The external validity of the results is also supported \nby the similarities in community solar definitions and policies across \nstates34,41. Nonetheless, some caution is required in extrapolating \nour analysis to states with distinct demographic, market and policy \ncontexts. We also recognize that our analysis of policy impacts is based \non a further restricted sample of three states, though again these three \nstates reflect a diversity of LMI programme contexts (Supplementary \nTable 8). Whereas our results suggest that LMI community solar poli-\ncies have been impactful, our retrospective analysis cannot necessarily \nbe extrapolated to future LMI community solar policies, which are \nincreasingly ambitious in scope and scale34. Second, the community \nsolar data are cross-sections of households that were actively enroled \nat the time the data were generated. With those cross-sectional data, \nwe lack insights into trends in community solar adoption over time. \nThis limitation shaped our analysis of inherent and policy impacts. \nWe have used the practical method of comparing cross-sections of \ncommunity solar adopter demographics between policy programme \nparticipants and non-participants. As noted, our method only pre-\ncisely identifies policy impacts under the strict assumption of no \nfree riding in community solar LMI programmes. Whereas available \nevidence suggests that free riding is infrequent, free riding is prob-\nably non-zero and thus our policy impacts are imprecisely identified. \nAn ideal approach\u2014a suggestion for future research\u2014would be to \nmore precisely identify policy impacts through econometric analy-\nsis of changes in community solar adoption trends before and after  \npolicy implementation.\nReporting summary\nFurther information on research design is available in the Nature \nPortfolio Reporting Summary linked to this article.\nData availability\nThis work was performed using proprietary, household-level data that \ncannot be shared. However, elements of the rooftop solar data are \npublicly available from the Lawrence Berkeley National Laboratory at \nhttps://emp.lbl.gov/projects/solar-demographics-trends-and-analysis.  \nAggregated data are available in via Github at https://github.com/ \neoshaugh2/community_solar_access. The LIFT Solar Toolkit used in the \nimputed carve outs analysis (Methods) is publicly available at https:// \nlift.groundswell.org/solar-projects/.\n\nNature Energy | Volume 9 | August 2024 | 955\u2013963\n962\nArticle\nhttps://doi.org/10.1038/s41560-024-01546-2\nCode availability\nCode used in this analysis is available via Github at https://github.com/ \neoshaugh2/community_solar_access.\nReferences\n1.\t\nDavis, M. et al. US Solar Market Insight: 2022 Year in Review  \n(Wood Mackenzie, 2023).\n2.\t\nSunter, D., Castellanos, S. & Kammen, D. Disparities in rooftop \nphotovoltaics deployment in the United States by race and \nethnicity. Nat. Sustain. 2, 71\u201376 (2019).\n3.\t\nLukanov, B. & Krieger, E. Distributed solar and environmental \njustice: exploring the demographic and socio-economic  \ntrends of residential PV adoption in California. Energy Policy 134, \n110935 (2019).\n4.\t\nDarghouth, N., O\u2019Shaughnessy, E., Forrester, S. & Barbose, G. \nCharacterizing local rooftop solar adoption inequity in the US. \nEnviron. Res. Lett. 17, 034028 (2022).\n5.\t\nForrester, S., Barbose, G., O\u2019Shaughnessy, E., Darghouth, N. &  \nCrespo Monta\u00f1\u00e9s, C. Residential Solar-Adopter Income and \nDemographic Trends: 2023 Update (Lawrence Berkeley National \nLaboratory, 2023).\n6.\t\nCarley, S. & Konisky, D. M. The justice and equity implications of \nthe clean energy transition. Nat. Energy 5, 569\u2013577 (2020).\n7.\t\nHeeter, J., Sekar, A., Fekete, E., Shah, M. & Cook, J. J. Affordable \nand Accessible Solar for All: Barriers, Solutions, and On-Site \nAdoption Potential (NREL, 2021).\n8.\t\nAttanasio, O. P. & Pistaferri, L. Consumption inequality. J. Econ. \nPerspect. 30, 3\u201328 (2016).\n9.\t\nWelton, S. & Eisen, J. Clean energy justice: charting an emerging \nagenda. Harvard Environ. Law Rev. 43, 307\u2013371 (2019).\n10.\t Bidwell, D. & Sovacool, B. Uneasy tensions in energy justice and \nsystems transformation. Nat. Energy 8, 317\u2013320 (2023).\n11.\t\nZhou, S., Gao, X., Wellstead, A. M. & Min Kim, D. Operationalizing \nsocial equity in public policy design: a comparative analysis  \nof solar equity policies in the United States. Policy Stud. J. 51, \n741\u2013772 (2023).\n12.\t Drury, E. et al. The transformation of southern California\u2019s \nresidential photovoltaics market through third-party ownership. \nEnergy Policy 42, 681\u2013690 (2012).\n13.\t O\u2019Shaughnessy, E., Barbose, G., Wiser, R., Forrester, S. & \nDarghouth, N. The impact of policies and business models on \nincome equity in rooftop solar adoption. Nat. Energy 6, 84\u201391 \n(2021).\n14.\t Malhotra, R. Increasing access to solar for low-income \nhouseholds in multifamily affordable housing. In Proceedings \nof the American Solar Energy Society National Conference (eds \nGhosh, A. K. & Rixham, C.) 87\u201393. (Springer, 2022).\n15.\t Goyette, K. L. Community solar policy and the low- and \nmoderate-income customer. Nat. Res. Environ. 36, 13\u201316  \n(2021).\n16.\t Funkhouser, E., Blackburn, G., Magee, C. & Rai, V. Business model \ninnovations for deploying distributed generation: the emerging \nlandscape of community solar in the U.S. Energy Res. Social Sci. \n10, 90\u2013101 (2015).\n17.\t Hausman, N. How Community Solar can Benefit Low- and \nModerate-Income Customers (World Resources Institute, 2022).\n18.\t Heeter, J., Bird, L., O\u2019Shaughnessy, E. & Koebrich, S. Design \nand Implementation of Community Solar Programs for Low- and \nModerate-Income Customers (NREL, 2018).\n19.\t Abbott, S., Tyson, M., Popkin, M. & Farthing, A. Community \nSolar+: How the Next Generation of Community Solar can Unlock \nNew Value Streams and Help Communities Pursue Holistic \nDecarbonization (Rocky Mountain Institute, 2022).\n20.\t Michaud, G. Perspectives on community solar policy adoption \nacross the United States. Renew. Energy Focus 33, 1\u201315 (2020).\n21.\t Brown, M., Soni, A., Lapsa, M. V., Southworth, K. & Cox, M. High \nenergy burden and low-income energy affordability: conclusions \nfrom a literature review. Progr. Energy 2, 042003 (2020).\n22.\t Haynes, B. Community Solar: Expanding Access and Safeguarding \nLow-Income Families (National Consumer Law Center, 2024).\n23.\t Shared Renewables Policy Catalog (Interstate Renewable Energy \nCouncil, 2020).\n24.\t Equitable Access to Community Solar: Program Design and \nSubscription Considerations (NREL, 2021).\n25.\t Xu, K., Sumner, J., Dalecki, E. & Burton, R. Expanding Solar Access: \nState Community Solar Landscape (NREL, 2023).\n26.\t Borenstein, S. & Davis, L. The distributional effects of U.S. clean \nenergy tax credits. Tax Policy Econ. 30, 191\u2013234 (2016).\n27.\t Chwastyk, D., Leader, J., Cramer, J. & Rolph, M. Community Solar \nProgram Design Models (Smart Electric Power Alliance, 2018).\n28.\t Gallucci, M. Energy equity: bringing solar power to low-income \ncommunities. YaleEnvironment360 https://e360.yale.edu/ \nfeatures/energy-equity-bringing-solar-power-to-low-income- \ncommunities (2019).\n29.\t Lydersen, K. Bringing community solar to low-income customers \nis harder than it looks. Canary Media https://www.canarymedia. \ncom/articles/solar/bringing-community-solar-to-low-income- \ncustomers-is-harder-than-it-looks (2023).\n30.\t Connelly, C. U.S. Community Solar Market Outlook: 2023  \n(Wood Mackenzie, 2023).\n31.\t O\u2019Shaughnessy, E., Barbose, G., Wiser, R. & Forrester, S. \nIncome-targeted marketing as a supply-side barrier to \nlow-income solar adoption. iScience 24, 103137 (2021).\n32.\t Ramasamy, V. et al. U.S. Solar Photovoltaic System and Energy \nStorage Cost Benchmarks, with Minimum Sustainable Price \nAnalysis: Q1 2023 (NREL, 2023).\n33.\t Chan, G. et al. Sharing the Sun Community Solar Project data \n(December 2022). NREL Data Catalog. NREL https://data.nrel. \ngov/submissions/220 (2023).\n34.\t Connelly, C. Demystifying LMI Incentives for Community Solar \nProjects (Wood Mackenzie, 2023).\n35.\t DeLiema, M. & Witt, P. Profiling consumers who reported mass \nmarketing scams: demographic characterizations and emotional \nsentiments associated with victimization. Secur. J. https://doi. \norg/10.1057/s41284-023-00401-5 (2023).\n36.\t Khanna, K. et al. wru: who are you? Bayesian prediction of \nracial category using surname, first name, middle name, and \ngeolocation. R package version 1.0.1. CRAN https://CRAN.R- \nproject.org/package=wru (2022).\n37.\t Imai, K., Olivella, S. & Rosenman, E. T. R. Addressing census data \nproblems in race imputation via fully Bayesian improved surname \ngeocoding and name supplements. Sci. Adv. 8, eadc9824 (2022).\n38.\t O\u2019Shaughnessy, E. Rooftop solar incentives remain effective  \nfor low- and moderate-income adoption. Energy Policy 163,  \n112881 (2022).\n39.\t Statewide distributed solar projects: beginning 2000. NYSERDA \nhttps://data.ny.gov/Energy-Environment/Statewide-Distributed- \nSolar-Projects-Beginning-200/wgsj-jt5f/about_data (2024).\n40.\t LIFT Solar: solar projects. Groundswell https://lift.groundswell. \norg/ (2023).\n41.\t Xu, K., Sumner, J., Burton, R. & Dalecki, E. State policies and \nprograms for community solar. NREL Data Catalog. NREL  \nhttps://data.nrel.gov/submissions/215 (2023).\nAcknowledgements\nThis material is based upon work supported by the US Department \nof Energy\u2019s (DOE) Office of Energy Efficiency and Renewable Energy \n(EERE) under the Solar Energy Technologies Office award number \n38444 and contract number DE-AC02-05CH11231 (E.O. and G.B.). \nThis work was authored in part by the National Renewable Energy \n\nNature Energy | Volume 9 | August 2024 | 955\u2013963\n963\nArticle\nhttps://doi.org/10.1038/s41560-024-01546-2\nLaboratory, operated by Alliance for Sustainable Energy, LLC, for DOE \nunder contract number DE-AC36-08GO28308 (S.K. and J.S.). Funding \nwas provided by the DOE Office of Energy Efficiency and Renewable \nEnergy Solar Energy Technologies Office. The views expressed in \nthe article do not necessarily represent the views of the DOE or the \nUS Government. The US Government retains and the publisher, by \naccepting the article for publication, acknowledges that the US \nGovernment retains a non-exclusive, paid-up, irrevocable, worldwide \nlicense to publish or reproduce the published form of this work, or \nallow others to do so, for US Government purposes.\nAuthor contributions\nE.O. wrote the paper and led the analysis; G.B. conceived of the study, \nsupported the analysis and co-led data acquisition; and S.K. and J.S. \nled data acquisition and supported the analysis.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s41560-024-01546-2.\nCorrespondence and requests for materials should be addressed to \nEric O\u2019Shaughnessy.\nPeer review information Nature Energy thanks Michael Garvey, \nChristine Lasco Crago and the other, anonymous, reviewer(s) for their \ncontribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner)  \nholds exclusive rights to this article under a publishing  \nagreement with the author(s) or other rightsholder(s); author \nself-archiving of the accepted manuscript version of this article is \nsolely governed by the terms of such publishing agreement and \napplicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2024, corrected publication 2024 \n\n\n    Scientific Research Findings:", "answer": "Using household\u2011level data from 11\u00a0US states, we find that community solar adopters in 2023 were about 6.1\u00a0times more likely to live in multifamily buildings than rooftop solar adopters, were about 4.4\u00a0times more likely to rent, and earned about 23% less annual income. These results suggest that community solar has expanded solar access to households that would or could not otherwise have adopted solar. We also find that policies have been effective in further expanding solar access, such as programmes that reserve community solar shares for low\u2011income customers. We estimate that income\u2011targeted community solar policies explain around 70% of differences in income levels between community and rooftop solar adopters in our sample, around 40% of differences in renter rates, and around 20% of differences in housing type. The impacts of community solar on solar access may vary from the results of this study in other jurisdictions with distinct policy and socioeconomic contexts.", "id": 7}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 9 | July 2024 | 883\u2013893\n883\nnature energy\nhttps://doi.org/10.1038/s41560-024-01530-w\nAnalysis\nThe effects of corporate investment and \npublic grants on climate and energy startup \noutcomes\nKathleen M. Kennedy\u2009\n\u200a\u20091\u2009\n, Morgan R. Edwards2,3, Claudia Doblinger\u2009\n\u200a\u20094, \nZachary H. Thomas3, Maria A. Borrero1, Ellen D. Williams5,6, \nNathan E. Hultman\u2009\n\u200a\u20091 & Kavita Surana\u2009\n\u200a\u20091,7,8\u2009\nClimate and energy (climate-tech) startups can accelerate the \ncommercialization of innovative technologies but face low investment and \nhigh failure rates. Here we analyse the effects of recent growth in corporate \ninvestments, combined with public grants and other private investments, \non startup outcomes. We apply the Cox Proportional Hazards model to \na dataset of 2,910 US climate-tech startups founded 2005\u20132020. We find \nthat corporate and other private investments are significantly associated \nwith both exits (initial public offerings, mergers/acquisitions) and failures \n(bankruptcy, going out of business). While public grants are not significantly \nassociated with these outcomes, they fill important funding gaps in high-risk \nsectors. Publicly funded startups also exit at a higher rate with the addition \nof corporate investment (155% increase) compared with other private \ninvestment (78% increase). These findings highlight the roles of different \ninvestors in scaling startup technologies to meet climate goals and are robust \nacross sectors, timelines and types of public funding (national, subnational).\nStartups aiming to rapidly commercialize new climate and energy \ntechnologies, that is, climate tech, are a vital part of climate change \nmitigation, with over one-third of emissions reductions needed to reach \nnet-zero relying on technologies currently in pre-commercial stages1. \nYet despite their importance, climate-tech startups face substantial \nbarriers. They often involve hardware that requires long development \ntimelines and extensive capital\u2014and thus carry high investment risks \nwith large potential returns to society (but not necessarily to inves-\ntors)2,3. Consequently, these startups have historically suffered from \nunderinvestment and often failed to overcome the \u2018valley of death.\u2019 To \novercome these challenges and deploy at scale, climate-tech startups \nneed direct involvement from the public and private sectors. Research \nand development can be de-risked through early-stage funding (for \nexample, public grants from regional or federal agencies and equity \ninvestment from venture capital, corporations or impact investors)4. \nDemonstration phases, previously underfunded by private equity, can \nadvance with increased public co-investment enabled by government \nagencies such as the US Department of Energy Office of Clean Energy \nDemonstrations. Early deployment and growth require public policies \n(for example, regulations and market incentives for demand pull) and \nprivate sector uptake (for example, adoption by corporate \u2018first mov-\ners\u2019)5,6. With the rapidly growing diversity of public and private actors \nand public\u2013private partnerships7,8, corporations stand out in terms of \nthe number and scale of recent investments in climate-tech startups9.\nReceived: 2 May 2023\nAccepted: 15 April 2024\nPublished online: 15 May 2024\n Check for updates\n1Center for Global Sustainability, School of Public Policy, University of Maryland College Park, College Park, MD, USA. 2La Follette School of Public \nAffairs, University of Wisconsin-Madison, Madison, WI, USA. 3Nelson Institute Center for Sustainability and the Global Environment, University of \nWisconsin-Madison, Madison, WI, USA. 4Campus Straubing for Biotechnology and Sustainability and School of Management, Technical University of \nMunich, Straubing, Germany. 5Earth System Science Interdisciplinary Center, University of Maryland College Park, College Park, MD, USA. 6Department \nof Physics, University of Maryland College Park, College Park, MD, USA. 7Institute for Data, Energy and Sustainability, Vienna University of Economics and \nBusiness, Vienna, Austria. 8Complexity Science Hub, Vienna, Austria. \n\u2009e-mail: kmkenne5@umd.edu; kavita.surana@wu.ac.at\n\nNature Energy | Volume 9 | July 2024 | 883\u2013893\n884\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01530-w\nDistinct investor roles for startup sectors  \nand stages\nStartups receive financing progressively over time as they mature. \nIn the early stages, startups acquire grants that do not require them \nto give up shares or repay the investment. While their dollar value is \ntypically smaller, grants can have an outsized impact, particularly for \nhigh-risk technologies or underfunded communities and regions. As \nstartups grow, they acquire funding through equity investment rounds \n(seed, series A, series B and growth equity) where they give up shares \nin exchange for investment. Eventually, more mature startups access \ncommercial debt where they do not give up shares but must repay the \ninvestment2. Across equity investment rounds, the growth rounds \ntypically represent the largest dollar amounts as startups prove them-\nselves to investors and seek funding to expand operations and widely \ndeploy their technologies. We focus on grants and equity investment \nrather than loans because they reflect the funding required to bring \npre-commercial technologies to market.\nCorporate investors are differentiated from public grants and \nother private investors in a few key ways. First, they consistently \nincreased in number of investments and amount invested since 2005, \neven as public grants and other investors saw uneven spates of growth \nand decline, followed by continuous growth since 2016 (Fig. 1a,b).  \nSecond, corporate investment deals peak later in the startup life  \ncycle and skew towards high-value growth equity and late-stage  \nrounds rather than cheaper seed rounds (42.0% of corporate invest-\nment deals are for growth equity, compared with 28.6% for other \nprivate investment; Fig. 1c). Finally, corporations tend to invest in  \npartnership with other forms of investment (86.9% of startups funded \nby corporations also received a public grant or other private invest-\nment), whereas both public and other private investors fund a greater \nnumber of startups alone (44.5% of startups funded by other private \ninvestment and 53.1% of startups that received public grants also \nreceived funding from a different investor type; Fig. 1d). Combined, \nthese features suggest that corporations could help startups grow \nthrough difficult periods such as the infamous \u2018valley of death\u2019 by \nproviding funding in the later stages of the ten-year venture capital \ncycle, when VC funding is less available and may disproportionately \nprovide critical funds for startups that have previously been successful \nin receiving grants or other private investment (Fig. 1 and Supplemen-\ntary Fig. 1)2.\nClimate-tech startups are active in multiple sectors, and the pat-\nterns of how different groups of investors engage with these sectors \ndemonstrate their varying priorities (Figs. 2 and 3). Public grants are \nmost dominant in sectors with very few startups, with the six sec-\ntors with the highest percentage of public funding all including fewer  \nthan 60 startups (compared with an average of 145). In each of these \nsectors, public grants account for at least 24% of investments (com-\npared with 11% of investments overall). This dominance is potentially \ndue to less profitable markets or technologies that are more difficult to \ncommercialize (for example, nuclear) and suggests that public grants \nfill a key role in the climate-tech ecosystem by funding sectors that are \nless popular with traditional equity and corporate investors11,23. Other \nprivate investors make up the bulk of investments in most sectors, with \na particular dominance in certain sectors such as recycling and waste, \nwhere both corporate investment and public grants are low. This phe-\nnomenon is prominent in more recent years, as public grants receded \nin certain sectors during Cleantech 2.0 (2012\u20132020), and corporate \nand other private investors provided a larger share of investments \n(Supplementary Fig. 2).\nIn contrast to public grants, corporate investment does not show a \nclear trend based on sector size but instead depends on other strategic \nchoices that may connect to core business areas or enable expan-\nsion into new areas9. Figure 3 shows investments by different catego-\nries of corporate investors in each startup sector, revealing heavy \ninvestment in related sectors. For instance, 42% of investments from \nUnderstanding the role of different investors and their inter-\nactions is essential to effectively incentivize innovation. However, \nanalyses on climate-tech startups tend to focus on public agencies10,11 \nor financially motivated venture capital (VC)2,5 in isolation, without \nconsidering the combined effects of different investors. For example, \npublic funding agencies can de-risk technologies when securing tra-\nditional investment is difficult12 or support technologies relevant for \nstrategic national goals13. Traditional VC investors aim for high financial \nreturns from startups exiting through an initial public offering (IPO) \nor acquisition. Climate-impact investors have emerged as a subset of \nprofit-focused investors, with a focus on supporting climate tech while \nstill pursuing financial returns. Corporations, in contrast, are strategic \ninvestors that may seek to increase profits but also to achieve goals \nrelated to long-term business plans and competitive standing9,14\u201319. They \ncan help startups build commercial capacity and credibility20, increase \ninnovation output21 and provide learning benefits22, but they can also \nleave startups at risk for predatory acquisitions that undercut compe-\ntition and misappropriate intellectual property19,22. The motivations \nof these different investors collectively determine how investments \nflow to technology areas and risk levels12. These investments, in turn, \ncan have distinct effects on startup outcomes. However, policymakers \ncurrently have limited empirical evidence on how different investors \nshape startup outcomes, which could enable more effective strategies \nto address the threat of climate change.\nThe higher involvement of corporate investors in climate tech is \nrelatively recent compared with traditional VC investors. The first wave \nof investment, known as Cleantech 1.0 (2005\u20132011; Supplementary \nNote 1), resulted in many failures that scholars attribute in large part \nto the poor fit with traditional VC models that expect exits within ten \nyears2,23,24. In response to these failures, traditional venture capital \ninitially shifted to less risky software-based climate-tech investments2. \nHowever, after 2015, new investment models emerged, and corpora-\ntions in particular became increasingly involved as investors9. By 2020, \nan estimated 34% of climate-tech startups received investment from \ncorporations (representing 24% of investment dollars), compared \nwith 57% from venture capital representing 31% of investment dollars9. \nThis indicates that corporations make fewer investments, but these \ninvestments are larger and thus potentially more influential for shaping \ninnovation9. Whereas other investors such as traditional VC firms have \nbeen increasingly analysed in literature2,5, scholars and policymakers \nhave limited quantitative evidence on the effects of growing corporate \nfunding in climate tech and the interactions and synergies between \ncorporate investment and public and other private funding sources.\nIn this paper, we examine whether corporate investment, cou-\npled with public grants and other private investment, improves \nclimate-tech startup outcomes. We focus on key outcomes linked \nto the deployment of technologies (that is, exits and failures). An \nexit occurs when a startup has an IPO or merges with/is acquired \nby another entity, indicating that their product has developed suf-\nficiently to draw interest from acquirers or on the stock market. A \nfailure occurs when a startup goes bankrupt or out of business and \ntherefore is probably unable to commercialize their product. Our \nempirical approach builds on a comprehensive dataset of climate-tech \nstartups from the Cleantech Group i3 database25. The original dataset \ncontains 38,089 startups and 17,737 investors worldwide. After a com-\nprehensive cleaning and verification process (Methods), we generate \na dataset of 2,910 United States-based startups founded between 2005 \nand 2020 (with outcomes considered through 2021) that received at \nleast one grant or equity investment. These startups were supported \nby 3,979 unique investors participating in 15,108 investment deals. \nWe expand this dataset with additional data on public funding and \npatents obtained from publicly available government records. We \nuse this dataset and the Cox Proportional Hazards (CPH) model to \nperform time-to-event analysis to quantify the associations between \ndifferent funding sources and startup outcomes.\n\nNature Energy | Volume 9 | July 2024 | 883\u2013893\n885\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01530-w\ntransportation corporations go to transportation startups, followed \nby advanced materials (14%) and energy storage (14%), which are both \nnecessary inputs for electric vehicles. Similarly, agriculture corpora-\ntions invest most heavily in agriculture startups (59%) whereas clean \nenergy investors favour solar (25%) and energy storage (24%) startups. \nThis alignment of investment with sector expertise can be beneficial \nfor startups, as deep knowledge of the sector allows corporations to \nprovide specialized assistance to new companies26. This advantage may \nbe particularly strong for startups that are trying to scale, especially \ngiven that corporations invest proportionally more in growth equity \ncompared with other private investors (Supplementary Table 2). Thus, \na key role of corporations in the climate-tech ecosystem lies in identi-\nfying promising startups within their area of expertise and providing \nessential funding to grow and scale.\n1,750\na\n80\nPublic grant\nOther growth equity\nOther series B\nOther series A\nOther seed\nOther grant\nCorporate growth equity\nCorporate series B\nCorporate series A\nCorporate seed\nCorporate grant\n60\n40\n20\n0\n2,500\n2,000\n1,500\n1,000\n500\n0\n1,500\n1,250\n1,000\n750\n500\nNumber of investments\nInvestments amount (billion US$)\nNumber of investments\n250\n0\n2005\n0\n5\n10\nStartup age\n15\n2010\n2015\nYear\nYear\n2020\n2005\nCorporate\ninvestment\nOther\ninvestment\nPublic grant\n368\n203\n189\n25\n133\n669\n1,323\n2010\n2015\n2020\nb\nc\nd\nFig. 1 | Interplay between public, corporate and other private investors in \nclimate-tech startups. a, Number of investments over time. b, Dollar value of \ninvestments over time. c, Number of investments by startup age. d, Number \nof startups receiving investment from different combinations of investors. \nOther investment includes all forms of non-corporate private investment \n(Supplementary Table 1). Data for a\u2013c are provided in Supplementary Tables 2\u20135. \nSupplementary Fig. 1 provides alternate presentation of these data.\n100\n400\nCorporate investors\nOther investors\nPublic grants\nNumber of startups\n350\n300\n250\n200\n150\n100\n50\n0\n80\n60\n40\nInvestment deals (%)\nNumber of startups\n20\nClimate-tech sector\n0\nGeothermal\nNuclear\nHydro and marine power\nBiomass generation\nFuel cells and hydrogen\nWind\nUncategorized\nAir\nConventional fuels\nSmart grid\nRecycling and waste\nWater and wastewater\nBiofuels and biochemicals\nEnergy storage\nAgriculture and food\nSolar\nOther clean tech\nAdvanced materials\nTransportation\nEnergy eficiency\nFig. 2 | Investment breakdown by climate-tech sector and investor type. \nSectors are ordered by the number of startups in that sector (lowest to \nhighest). Public grants typically account for a larger portion of investment \ndeals for smaller sectors, whereas corporate investment varies across sector \nsizes. Other investors include all sources of non-corporate private investment \n(Supplementary Table 1). Data include startups founded through the full period \nof analysis (2005\u20132020).\n\nNature Energy | Volume 9 | July 2024 | 883\u2013893\n886\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01530-w\nDifferent forms of investment have synergies and \ntrade-offs\nUsing time-to-event regression analysis, we investigate the interplay \nbetween publicly funded startups and different types of private inves-\ntors, distinguishing between corporations and others, and discuss \nhow these actors jointly affect outcomes. Time-to-event models are \ncommonly used with entrepreneurial data because many of the quanti-\nties of interest for startups are highly time dependent27. Specifically, \nwe use the Cox Proportional Hazards model to analyse exit and failure \noutcomes for each of the populations summarized in Table 1. In the CPH \nmodel, time-dependent covariates such as investment are accounted \nfor by separating variables and outcome events into intervals, thus \nproviding a way to regress an outcome at a given point in time based \nsolely on past covariate values, without knowledge of future values28. \nClustering connects the identity of an individual startup between mul-\ntiple time intervals. We stratify by the year the startup was founded \nand the sector of the startup, thereby generating a separate baseline \nhazard function for each instance of these variables that accounts for \neffects caused by these variables29.\nAcross our full dataset of climate-tech startups (Model 1), corpo-\nrate investment and other private investment are both highly signifi-\ncant and positively correlated with exits and failures, whereas public \nTable 1 | Summary of model populations used in time-to-event analysis\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nDescription\nFull dataset\nStartups that received \na public grant\nStartups that did NOT \nreceive a public grant\nStartups that received a \ncorporate investment\nStartups that did NOT receive \na corporate investment\nNumber of startups\n2,910\n783\n2,127\n1,016\n1,894\nNumber of exits\n440 (15%)\n98 (13%)* \np\u2009=\u20090.020\n342 (16%)* \np\u2009=\u20090.020\n208 (20%)*** \np\u2009<\u20090.001\n232 (12%)*** \np\u2009<\u20090.001\n\u2003 Number of IPOs\n76 (2.6%)\n24 (3.0%) \np\u2009=\u20090.424\n52 (2.4%) \np\u2009=\u20090.424\n52 (5.1%)*** \np\u2009<\u20090.001\n24 (1.3%)*** \np\u2009<\u20090.001\n\u2003 Number of MAs\n372 (13%)\n78 (10%)** \np\u2009=\u20090.007\n2,934 (14%)** \np\u2009=\u20090.007\n162 (16%)*** \np\u2009<\u20090.001\n210 (11%)*** \np\u2009<\u20090.001\nNumber of failures\n591 (20%)\n154 (20%) \np\u2009=\u20090.639\n437 (21%) \np\u2009=\u20090.639\n243 (24%)*** \np\u2009<\u20090.001\n348 (18%)*** \np\u2009<\u20090.001\nNumber of private startups after \nten years\n700 (24%)\n282 (36%)*** \np\u2009<\u20090.001\n418 (20%)*** \np\u2009<\u20090.001\n187 (18%)*** \np\u2009<\u20090.001\n513 (27%)*** \np\u2009<\u20090.001\nAverage age of startups (years)\n9.16\n10.2*** \np\u2009<\u20090.001\n8.78*** \np\u2009<\u20090.001\n9.32 \np\u2009=\u20090.120\n9.08 \np\u2009=\u20090.120\nAverage number of patents\n5.63\n6.01 \np\u2009=\u20090.489\n5.48 \np\u2009=\u20090.489\n9.97*** \np\u2009<\u20090.001\n3.29*** \np\u2009<\u20090.001\nNote that subset Models 2\u20135 may overlap in coverage of startups. Percentages in parentheses indicate the percentage of startups within a model that experienced the indicated outcome. \nSignificance levels are shown for two-sided proportion tests for differences between binary variables in models (that is, Model 2 vs 3 and Model 4 vs 5), where asterisks denote *p\u2009<\u20090.05, \n**p\u2009<\u20090.01, ***p\u2009<\u20090.001. Significance levels are shown for two-sided t tests for difference between continuous variables in models. Full descriptive statistics are provided in Supplementary \nTables 6 and 7. MA = merger/acquisition.\nAdvanced materials\nAgriculture and food\nAir\nBiofuels and biochemicals\nBiomass generation\nConventional fuels\nEnergy eficiency\nEnergy storage\nFuel cells and hydrogen\nGeothermal\nHydro and marine power\nNuclear\nOther clean tech\nRecycling and waste\nSmart grid\nSolar\nTransportation\nUncategorized\nWater and wastewater\nWind\nStartup sector\nTransportation\nManufacturing, hardware,\nsemiconductors and chemicals\nFossil fuels and utilities\nFood, beverages\nand agriculture\nDigital and\nfinancial services\nClean energy\nCorporate sector\nPercentage of investments\n0\n10\n20\n30\n40\n50\n60\nFig. 3 | Corporate climate-tech investment by corporate and startup sectors. \nCorporate sectors are shown on the y axis and startup sectors are shown on the \nx axis. The size and colour of each data point corresponds to the percentage \nof investments from each corporate sector to each startup sector. Corporate \nsectors were categorized by the authors; startup sectors were determined by \nCleantech Group.\n\nNature Energy | Volume 9 | July 2024 | 883\u2013893\n887\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01530-w\ngrants are not (Table 2). Public grants also do not show any consist-\nently significant effect when grantors were separated into national, \nsubnational or international sources (Supplementary Tables 8 and 9). \nModels 2 and 3 consist of startups that did or did not receive a public \ngrant, respectively, which allows us to analyse how the presence of a \npublic grant changes the effect of other investors and key variables. \nSimilarly, Models 4 and 5 consist of startups that did or did not receive \na corporate investment, respectively. Table 2 summarizes the regres-\nsion results for each model, with corporate investment, other private \ninvestment, public grants, patent count and location in a climate-tech \n\u2018hotspot\u2019 (that is, a location with a density of two climate-tech startups \nper 100,000 population or greater) as predictor variables (Methods).\nOur results indicate that corporate investment and other private \ninvestment are highly significant for startup outcomes, in terms of \nboth success and failure, whereas public grants are not. Startups with at \nleast one corporate investment exited at a rate 110% higher than those \nwithout corporate investment (hazard ratio = 2.10). Similarly, startups \nwith at least one investment from other private investors exited at a \nrate 150% higher than those without such investments (hazard ratio = \n2.50). Each additional patent held by a startup was associated with a 2% \nincrease in rate of exits, whereas being headquartered in a climate-tech \n\u2018hotspot\u2019 location was associated with a 23% increase in exit rate. For \nthe same population of startups, corporate investment was associated \nwith a 62% increase in rate of failure, and other private investment \nwas associated with a 265% increase in rate of failure. While startups \nare high-risk investments where most will fail, and thus correlation \nwith failure is not necessarily surprising, the lower hazard of failure \nfor corporate investment compared with other private investment \nsources may indicate a role for corporations in providing more patient \ncapital. Public grants, patents and startup location were not found to \nbe significantly associated with failure.\nHowever, climate-tech sectors are not homogeneous, and startups \nthat receive a public grant are more likely to be in high-risk sectors \n(Fig. 2). Startups that received public grants remained privately held \nat the highest rate after ten years (Table 1), which could reflect longer \ntimescales associated with high-risk technologies. Our results suggest \nparticularly favourable outcomes associated with corporate invest-\nment for this group compared with other private investment for both \nexits (Model 2, 155% higher rate of exit with corporate investment \ncompared with 78% with other private investment) and failure (77% \nhigher rate of failure with corporate investment compared with 134% \nwith other private investment). For startups that do not receive public \ngrants, corporate investment continues to be significant (Model 3, 99% \nhigher rate of exit and 60% higher rate of failure), but other private \ninvestment is associated more strongly with both positive and nega-\ntive outcomes (139% higher rate of exit and 388% higher rate of failure). \nOther factors such as location (28% higher rate of exit for startups in \nhotspots) and patents (2% higher rate of exit associated with each \nadditional patent) also play a significant role.\nStartups that received corporate investment (Model 4) had the \nhighest rates of both IPO (5.1%) and merger/acquisition (16%) exits \nacross all models (Table 1), yet none of the modelled variables were \nfound to be significantly associated with these outcomes (Table 2). \nThis could be due to only 25 startups in the dataset receiving corporate \ninvestment without also receiving other private investment (Fig. 1d), \nlimiting the strength of this finding (likelihood ratio test in Table 2). \nTable 2 | Hazard ratios from CPH regression for startup exits, defined as an IPO or merger/acquisition, and failures, defined \nas bankruptcy or going out of business\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nDescription\nFull dataset\nStartups that received \na public grant\nStartups that did NOT \nreceive a public grant\nStartups that received a \ncorporate investment\nStartups that did NOT receive a \ncorporate investment\nStartup exit\nCorporate investment\n2.10*** (1.74, 2.53) \np\u2009=\u2009<0.001\n2.55*** (1.57, 4.12) \np\u2009=\u2009<0.001\n1.99*** (1.62, 2.46) \np\u2009=\u2009<0.001\n\u2013\n\u2013\nOther private investment\n2.50*** (1.91, 3.26) \np\u2009=\u2009<0.001\n1.78* (1.07, 2.96) \np\u2009=\u20090.028\n2.39*** (1.71, 3.34) \np\u2009=\u2009<0.001\n1.40 (0.967, 2.02)\np\u2009=\u20090.075\n5.11*** (3.42, 7.63)\np\u2009=\u2009<0.001\nPublic grant\n1.01 (0.808, 1.26) \np\u2009=\u20090.927\n\u2013\n\u2013\n1.29 (0.949, 1.76)\np\u2009=\u20090.104\n1.23 (0.881, 1.71)\np\u2009=\u20090.228\nPatent count\n1.02** (1.00, 1.03) \np\u2009=\u20090.009\n1.06 (0.979, 1.14) \np\u2009=\u20090.156\n1.02* (1.00, 1.03) \np\u2009=\u20090.020\n1.03 (0.993, 1.07)\np\u2009=\u20090.117\n1.02 (0.999, 1.04)\np\u2009=\u20090.064\nLocation\n1.23* (1.03, 1.47) \np\u2009=\u20090.024\n1.04 (0.699, 1.56) \np\u2009=\u20090.830\n1.28* (1.04, 1.56) \np\u2009=\u20090.019\n1.20 (0.913, 1.58)\np\u2009=\u20090.191\n1.37* (1.07, 1.76)\np\u2009=\u20090.014\nLikelihood ratio test\n119*** \np\u2009=\u2009<0.001\n28.2*** \np\u2009=\u2009<0.001\n65.9*** \np\u2009=\u2009<0.001\n9.04\np\u2009=\u20090.06\n79.1***\np\u2009=\u2009<0.001\nStartup failure\nCorporate investment\n1.62*** (1.40, 1.88) \np\u2009=\u2009<0.001\n1.77*** (1.29, 2.41) \np\u2009=\u2009<0.001\n1.60*** (1.34, 1.91) \np\u2009=\u2009<0.001\n\u2013\n\u2013\nOther private investment\n3.65*** (2.84, 4.69) \np\u2009=\u2009<0.001\n2.34*** (1.63, 3.73) \np\u2009=\u2009<0.001\n4.88*** (3.23, 7.37) \np\u2009=\u2009<0.001\n4.05*** (2.57, 6.36) \np\u2009=\u2009<0.001\n3.60*** (2.63, 4.92)\np\u2009=\u2009<0.001\nPublic grant\n1.07 (0.903, 1.28) \np\u2009=\u20090.461\n\u2013\n\u2013\n0.946 (0.736, 1.22) \np\u2009=\u20090.665\n1.09 (0.838, 1.43)\np\u2009=\u20090.510\nPatent count\n0.996 (0.976, 1.02) \np\u2009=\u20090.741\n1.01 (0.954, 1.08) \np\u2009=\u20090.640\n0.997 (0.976, 1.02) \np\u2009=\u20090.799\n0.979 (0.951, 1.01) \np\u2009=\u20090.163\n1.01* (1.00, 1.03)\np\u2009=\u20090.14\nLocation\n1.08 (0.935, 1.25) \np\u2009=\u20090.307\n1.07 (0.822, 1.39) \np\u2009=\u20090.623\n1.11 (0.933, 1.32) \np\u2009=\u20090.241\n1.10 (0.876, 1.38)\np\u2009=\u20090.415\n0.994 (0.822, 1.20)\np\u2009=\u20090.949\nLikelihood ratio test\n156.6*** \np\u2009=\u2009<0.001\n37.91*** \np\u2009=\u2009<0.001\n90.78*** \np\u2009=\u2009<0.001\n44.96***\np\u2009=\u2009<0.001\n70.27***\np\u2009=\u2009<0.001\n95% confidence intervals are shown in parentheses. Asterisks denote *p\u2009<\u20090.05, **p\u2009<\u20090.01, ***p\u2009<\u20090.001. Startup sector and founding year are accounted for as separate strata. Data included in \neach of the models are described in Table 1. Sensitivity analysis demonstrates that results are robust across multiple representations of investment variables (Supplementary Tables 10, 11, 18 \nand 19).\n\nNature Energy | Volume 9 | July 2024 | 883\u2013893\n888\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01530-w\nFailure for startups with corporate investment was only significantly \nassociated with other private investment (305% higher rate of failure \nthan startups without other private investment). Startups that did not \nreceive corporate investment (Model 5) saw the lowest rate of IPOs \n(1.3%) across all models. Exit outcomes were significantly associated \nwith other private investment (411% higher rate of exit) and location \nin a climate-tech hotspot (37% higher rate of exit). Failure for start-\nups without corporate investment was significantly associated with \nother private investment (260% higher rate of failure), as seen in other \nmodels. However, this model was the only one that found a significant \nassociation between patents and failure (1% higher rate of failure with \neach additional patent).\nEffects are robust across sectors and time periods\nDifferences in the characteristics of specific technologies in various \nclimate-tech sectors could lead to different interplays between funding \nsources at the sector level2,5. However, our results are largely consistent \nacross individual climate-tech sectors. Figure 4 shows the relationships \nbetween public grants, corporate investments and exit and failure out-\ncomes across the six largest sectors by number of startups. We generate \nsurvival curves, which represent the proportion of the population that \nhas not experienced an outcome by time t (in units of years since the \nstartup was founded). Corporate investment is consistently associated \nwith higher rates of exits (Fig. 4b), and this association is robust across \nfounding years (Supplementary Fig. 3 and Supplementary Table 12) \nEnergy eficiency\nTransportation\nAdvanced materials\nOther clean tech\nSolar\nAgriculture and food\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n0.25\n0.50\n0.75\n1.00\nTime since founding (yr)\nSurvival probability\nNo public grant\nPublic grant\na\nRelationship between public grants and startup exits\nEnergy eficiency\nTransportation\nAdvanced materials\nOther clean tech\nSolar\nAgriculture and food\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n0.25\n0.50\n0.75\n1.00\nTime since founding (yr)\nSurvival probability\nNo corporate investment\nCorporate investment\nb  Relationship between corporate investment and startup exits\nEnergy eficiency\nTransportation\nAdvanced materials\nOther clean tech\nSolar\nAgriculture and food\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n0.25\n0.50\n0.75\n1.00\nTime since founding (yr)\nSurvival probability\nNo public grant\nPublic grant\nc\n Relationship between public grants and startup failure\nEnergy eficiency\nTransportation\nAdvanced materials\nOther clean tech\nSolar\nAgriculture and food\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n4\n8\n12\n16\n0\n0.25\n0.50\n0.75\n1.00\nTime since founding (yr)\nSurvival probability\nNo corporate investment\nCorporate investment\nd\nRelationship between corporate investment and startup failure\nFig. 4 | Survival curves for exit and failure outcomes for the six largest sectors \nwith and without public grants or corporate investment. Survival curves use \nthe Kaplan\u2013Meier estimator to show the proportion of the population that has \nnot experienced the outcome of interest by a given time t, and shading represents \na 95% confidence interval. All curves begin with a survival probability of 1, which \nindicates no startups have experienced an exit or failure. a\u2013d, As startups age, \nsurvival probability decreases based on how many startups exit (a,b) or fail \n(c,d); for instance, a survival probability of 0.8 would indicate 20% of startups \nhave experienced an exit (a,b) or failure (c,d). The regressions represent the \neffect solely due to public grants and corporate investment, respectively; other \ncovariates are not included.\n\nNature Energy | Volume 9 | July 2024 | 883\u2013893\n889\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01530-w\nand other sectors (Supplementary Fig. 4). Public grants have mixed \nresults across sectors, with grants significantly related to lower rates of \nexit and failure in the energy efficiency sector, but with no significant \ntrends in other large sectors. There are also sectoral differences in the \noverall rate of exit and failure, such as particularly high rates of failure in \nthe solar sector, and relatively high rates of exits in the transportation \nsector. These differences validate our choice to assign each sector their \nown baseline hazard function to account for variation in sector-level \ncharacteristics.\nWe examine the sensitivity of our results with respect to time \nand patenting activity. Climate-tech investment has matured from \nCleantech 1.0 to Cleantech 2.0, potentially leading to different invest-\nment effects across time periods (Supplementary Note 1)2,30. To assess \npotential differences, we replicate the analysis in Table 2 for both time \nperiods. Whereas there is little difference in trends for exits, corpora-\ntions are significantly associated with failure only in the Cleantech \n1.0 period (Supplementary Tables 13 and 14). Further investigation is \nneeded to determine the causes of this change (Discussion); however, \nit may indicate a lower hazard of startup failure from future corporate \ninvestments. All other trends remain the same, indicating the analysis \npresented here is robust to temporal differences. We also examine if \nour results vary for a subset of high-patenting startups, which may \nindicate \u2018unicorn\u2019 startups that potentially skew our results (Supple-\nmentary Note 2). We find that no investment source is a significant \npredictor of exits or failures for high-patenting startups, but location \nin a climate-tech hotspot is highly significant and correlates with an \nincrease in likelihood of exit (Supplementary Table 15). Corporate \ninvestment shows a unique relationship for these high-patenting \nstartups with a negative but insignificant correlation with failure, \ncontrasted with other private investment, which still has a positive \nand significant correlation with failure as in the overall dataset (Sup-\nplementary Table 16).\nDiscussion and conclusions\nWith recent US legislation such as the Inflation Reduction Act (IRA) \nmaking historic investments in climate change mitigation31, including \nin climate-tech innovation and deployment32, this analysis provides \nan enhanced understanding of how corporate investment and public \ngrants fill private investment gaps and play complementary roles that \nsupport climate-tech startups.\nWe find that public grants are not associated with a significant like-\nlihood of either exits or failures (Models 1, 4 and 5). However, startups \nfunded by public grants represent technologies of interest to national \npolicy, and as they grow and bring in different investors, corporate \ninvestment correlates with a higher likelihood of exits and a lower \nlikelihood of failure compared with other forms of private investment \n(Model 2). This correlation is unlikely to arise purely from selection bias \nbecause both corporate and other investment are also associated with \na higher likelihood of startup failure overall. The lack of significance \nof public grants in this analysis similarly does not necessarily indicate \na lack of effect on startups. Possible explanations include preferen-\ntial selection of higher-risk startups, or differing impacts in different  \nsectors or startups choosing to pursue additional grants rather than \nan exit outcome (Supplementary Note 3)11,33.\nCorporations are significantly more likely to fund climate-tech \nstartups that achieve a successful exit (Model 1, Fig. 4) across sectors, \nreflecting a breadth of core business areas and priorities (including \nthe need to decarbonize). This correlation is also seen in startups \nthat received a public grant, a subpopulation of particular interest \nto policymakers (Model 2). Corporations consistently show a lower \ncorrelation with failure than other investors (Models 1\u20133), indicating \nthat corporations either selectively fund startups that are less likely to \nfail or can partially mitigate failure risks through their unique invest-\nment strategies. Additionally, corporate investment is not significantly \nassociated with failure in Cleantech 2.0\u2014a clear deviation from other \nresults (Models 1\u20133, Supplementary Table 14). This difference could \nbe because corporations learned from the losses in Cleantech 1.0 and \nadjusted their investment strategies or because outside pressures have \ninfluenced a change in their behaviour. There are also some sectoral \ndifferences in correlation with failure (for example, strong significance \nin energy efficiency and weak significance in advanced materials; \nFig. 4d), which could be due in part to the proportion of hardware and \nsoftware startups found in each sector, which have different likelihoods \nof exit and failure2,34.\nCorporate investors can be particularly important to the success of \ndifficult-to-scale, research-intensive technologies by providing access \nto testing or supply chains that other investors cannot. This potentially \nmeans that corporations can support startup success when investees \nare located near their own operations21,35, which may not be as tied to \ntraditional venture capital hotspots such as Silicon Valley. Our analysis \nshows evidence of this in the insignificance of location for startups with \ncorporate investment (Model 4). The absence of corporate investors \ncould be detrimental for research-intensive startups, as startups that \ndid not receive corporate investment showed a significant relationship \nbetween failure and patenting activity, a metric for innovation and a \npotential indicator of a hardware-focused company (Model 5). Con-\nversely, corporations may be particularly beneficial for high-patenting \nstartups, which showed no correlation between corporate investment \nand failure (Supplementary Table 16). Ultimately, corporations may \nmitigate risks for technologies by providing more patient capital than \nother investors and thereby offsetting the timelines for fast returns \nexpected by traditional venture capital23,24.\nOur results suggest three key recommendations for innovation- \nfocused policy, which may be particularly important in light of the \nsurge in public funding established in the IRA to leverage private capital \nand the shift in attitudes towards globalization36. First, whereas public \ngrants are not significant drivers of outcomes on their own, they prob-\nably act as catalysts for startups that work on difficult sectors before \nthey advance and acquire capital through corporate or other private \ninvestors (Fig. 2). Publicly funded startups exit at a higher rate with the \naddition of corporate investment (155% increase) compared with other \nprivate investment (78% increase) (Model 2). This suggests that increas-\ning public grant funding at both the national (for example, Advanced \nResearch Projects Agency-Energy (ARPA-E)) and regional (for example, \nNew York State Energy Research and Development Authority) level is \ncritical to de-risk technologies and create opportunities for beneficial \npublic\u2013private partnership. Second, corporate investors\u2019 association \nwith positive outcomes for grant recipients and high-patenting start-\nups (compared with other private investors) highlights the importance \nof incentivizing these investors to mobilize capital to invest in startups. \nPolicymakers can build on successes from networks enabled through \nkey grant agencies (for example, the ARPA-E annual summit) to create \nopportunities for high-risk startups. In addition, public\u2013private part-\nnerships (for example, the First Movers Coalition focused on aviation, \nsteel and other key technologies) could create similar opportunities \nin other regions or technologies. Third, the association between both \ncorporate and other private investment and increased rates of failure \nmeans that policymakers should act to minimize harmful practices \nsuch as intellectual property misappropriation (for example, misap-\npropriation of the startups\u2019 knowledge by the larger corporate partner, \nalso known as \u2018corporate sharks\u2019)22,37.\nThis analysis presents several avenues for future research that \ncould further inform evidence-based climate innovation policy. Access \nto more comprehensive data, particularly from startups that are typi-\ncally reticent to share details, would provide a deeper understanding of \nsectoral variation and technology maturity (for example, technology \nreadiness levels) for different types of investors and their impact. Such \ndata can potentially create avenues to improve models that inform \nclimate policy. It could also be used to explore different types of public \ninvestors besides grant agencies that aim to generate financial returns \n\nNature Energy | Volume 9 | July 2024 | 883\u2013893\n890\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01530-w\n(for example, green banks or public agencies that invest equity in \nstartups) or private investors (for example, financially motivated \nclimate-impact investors) to provide a more complete picture of the \npublic and private funding landscape. With more data over time, future \nwork could potentially establish causal relationships between variables \nand assess the potential impact of systematic factors such as startup \nmaturity on investment outcomes. Mixed-method analyses that com-\nbine data on individual grant programmes such as ARPA-E with insights \non specific types of investor (for example, corporate investors with \nvarious core businesses, climate-impact investors, other private inves-\ntors) could further inform public policies on how to most effectively \nsupport grant recipients through all phases of startup development. \nFinally, this pre-IRA analysis can be used for comparison with future \npost-IRA analysis to identify changes in investment patterns or other \neffects resulting from the IRA.\nMethods\nDescription of raw dataset\nData from the Cleantech Group i3 database on startups, investors, \ninitial public offerings (IPOs), mergers and acquisitions, and corporate \nrelationships was accessed on 19 January 2022. These proprietary data \nare available to subscribers at https://i3connect.com. The Cleantech \nGroup data are used in other major reports such as the International \nEnergy Agency\u2019s (IEA) Tracking Clean Energy Innovation Report and the \nSilicon Valley Bank\u2019s The Future of Climate Tech Report8,38,39. In addi-\ntion, it has been previously used in multiple peer-reviewed research \npapers3,9,10. Given the extensive use of the Cleantech Group\u2019s i3 database \nin prominent reports and peer-reviewed papers and the fact that infor-\nmation in the database is regularly updated by the database provider\u2019s \nresearch teams or by the firms themselves (with the last updated year \nreported), to our knowledge, it represents the most thorough dataset \navailable for climate tech.\nThe raw data had 38,089 startups and 17,737 investors worldwide. \nEach startup has information on sector, location, founding year, cur-\nrent status (for example, public, private or bankrupt), a description \nof the startup activity and keyword terms associated with the startup.\nData cleaning\u2014startups\nThe raw dataset was filtered to obtain a dataset that only includes \nstartups of interest to this analysis using the following steps and build-\ning on the approach in Surana et al.9. First, startups that did not have \na clear connection to climate were identified and excluded from the \ndataset. This was done using keyword terms in the description, such \nas startups focused on ride sharing for which the climate benefit is \nunclear. Supplementary Table 17 provides all filtering terms. Competi-\ntors of these excluded ventures, as identified by i3, were also excluded. \nIn addition, startups that i3 categorized in the \u2018other clean tech\u2019 sec-\ntor were excluded unless they were described by a relevant keyword \nindicating a clear climate mitigation link (Supplementary Table 17). \nManual verification revealed that many of these ventures were tech-\nnology firms with minimal or unclear climate relevance (for example, \nAirbnb). Second, startups were excluded due to lack of information if \nno sector or industry group was specified, the description or tags did \nnot include any relevant keywords and no founding year was provided. \nThird, startups were excluded from the temporal or geographical \nscope of this analysis. Startups located outside of the United States \nwere excluded. Startups founded before 2005, the year the first major \nglobal climate agreement (the Kyoto Protocol) entered into force, were \nexcluded. When the founding year was not included in the i3 data, it was \nidentified by the authors using previously cleaned data provided by \nKurowski and Doblinger and publicly available information from data-\nbases such as Pitchbook, Crunchbase and OpenCorporates. Fourth, \nentities in the dataset were also excluded if they were determined to \nbe a large group or corporation (for example, EON Group), a public \nagency (for example, the US Department of Energy) or a university or \nresearch institute rather than a startup. Finally, startups were excluded \nfrom the analysis if they had not received a single investment or grant \nduring the 2005\u20132020 period.\nIn the resulting filtered dataset, companies were identified that \nlacked data for their year of failure (defined as going out of business or \nfiling for bankruptcy) and location of company headquarters. Updates \nto this missing information were first taken from previously cleaned \ndata by Kurowski and Doblinger. The remaining missing information \nwas obtained from other databases such as Pitchbook, Crunchbase, \nOpenCorporates and publicly searchable state-level business records.\nFinally, the status of startups with no recorded activity for  \n2\u20135 years before 2021 were manually verified by the authors. Statuses \nwere updated based on public records, newspaper articles, archival \nwebsite information and comparison with the startup\u2019s status in other \ndatabases such as Crunchbase or Pitchbook.\nThe resulting dataset has 2,910 startups that were used for this \nanalysis.\nData cleaning\u2014investments and investors\nFor each startup, the Cleantech Group data provide information on \nthe investment round, the amount and the investors that participated. \nThe investments in the Cleantech Group data were filtered to only \ninclude those most relevant to startups, which were taken to be those \nclassified as grants, seed, series A, series B and growth equity. This \napproach is consistent with methodologies used by PwC, the IEA and \nother reports9,39\u201344. The i3 investor categories were then aggregated \ninto the categories of Public, Corporate and Other Investors as shown \nin Supplementary Table 1. Here we note that other types of financing \nsuch as loans, loan guarantees or project finance were not included \nas these may involve far more technically and commercially mature \ntechnologies where comprehensive data might not be available. This \nis consistent with the approach used in multiple analyses that focus \non startups9\u201311.\nCorporate investors were further manually categorized into \nsix sectors: \u2018clean energy\u2019, \u2018fossil fuels and utilities\u2019, \u2018transportation\u2019, \n\u2018manufacturing, hardware and semiconductors\u2019, \u2018food, beverages and \nagriculture\u2019 or \u2018digital and financial services\u2019, building on the approach \nin Surana et al.9. These categories were defined using the highest level \nof aggregation of the International Standard Industrial Classification \nof All Economic Activities (ISIC) Rev.4 and were chosen considering \nthe contribution of each sector to climate change and the information \navailable in the i3 and SBTi databases. Additionally, some categories \nwere aggregated or divided based on the classification used by the \nSBTi, such as \u2018food, beverages and agriculture\u2019 and \u2018transportation\u2019.\nISIC Rev.4 categories \u2018mining, oil and gas\u2019, \u2018electricity, gas, steam \nand air conditioning supply\u2019 and \u2018water supply; sewerage, waste man-\nagement and remediation activities\u2019 were mixed and transformed \ninto two categories \u2018fossil fuels and utilities\u2019 and \u2018clean energy\u2019. This \nclassification aggregates fossil fuel companies and utilities that are \nnot supplying renewable electricity in one group and establishes a \nunique category for companies that produce renewable energy or \nprovide renewable equipment such as solar panels or wind turbines. \nThe category \u2018manufacturing, hardware and semiconductors\u2019 includes \nall activities considered in the ISIC Rev.4 category \u2018manufacturing\u2019, \nexcept the production of motor vehicles that are contained in \u2018trans-\nportation\u2019. Finally, the category \u2018digital and financial services\u2019 includes \nall companies in the database that provide digital services or goods, \nfinancial services such as banks, investment and insurance firms and \nother services as corporate consultancy or real estate agencies.\nThe public grants information in the Cleantech Group data was \nsupplemented by downloading grant data from https://www.usaspend-\ning.gov/ for ARPA-E and Small Business Innovation Research (SBIR) \n(accessed 21 January 2022 and 19 January 2022, respectively). These \ngrants were matched with startups in the i3 database to verify and \nupdate existing grant information or to add missing grant data.\n\nNature Energy | Volume 9 | July 2024 | 883\u2013893\n891\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01530-w\nData cleaning\u2014patent activity\nPatent data were obtained from the US Patent and Trademark Office \n(USPTO) using the PatentsView application programming interface and \nthe R package patentsview45. Patent data were accessed on 4 May 2022. \nWe matched patent assignee names with startup names after multiple \nand extensive cleaning steps (for example, removing stems such as \n\u2018Inc.\u2019 and \u2018Ltd.\u2019), consolidating any duplicates and manual verification. \nWe used these data to identify the number of patents applied in each \nyear by each startup.\nRegression analysis\u2014exits and failure\nThe regression method used in this work for exit and failure outcomes \nis the Cox Proportional Hazards (CPH) model, which has the form:\nh (t, x1, \u2026 , xn) = ho (t) e\u2211\nn\n1 \u03b21x1\n(1)\nThe term ho (t) is the baseline hazard function, t is time and \u03b21 \u2026 \u03b2n  \nare the coefficients associated with the covariates x1 \u2026 xn. When a \ncovariate is time varying, such as with investments, this is represented \nas x(t)1 \u2026 x(t)n. This allows for recurrent event analysis of instances \nwhere the same startup received multiple investments within  \nthe studied period. The hazard ratio is defined as the exponentiated \ncoefficient e\u03b2. The model was implemented using the survival package \nand visualized using the survminer package in R46,47.\nThe applicability of the CPH model was verified by testing for \nproportionality of hazards, influential observations and nonlinearity. \nSchoenfeld residuals were used to verify the proportionality of hazards \nboth for individual variables and with a global test for the entire model, \nwhere an insignificant result means the proportional hazards assump-\ntion is valid. To check for influential observations, dfbeta values were \nvisualized and determined to be small compared with the regression \ncoefficients. Nonlinearity is not considered to be an issue for the binary \nand categorical variables used in the regressions.\nThe selection and representation of variables for the analysis \nwas based on prior literature on survival analysis and CPH models27,48, \nrelevant variables for startups in general11 and specifically in climate \nand energy3,10 and the availability of data in the Cleantech Group i3 \ndataset. Investment data represent a startup\u2019s funding strategy across \nthe range of possible investors, and data on funding rounds are used in \nstartup literature both as a predictor of future outcomes and as metric \nfor growth3,11,48. Here investment was categorized as corporate invest-\nment, other private investment or public grant (Main text provides \nreasons). Investment data were represented as time-varying binary \ncovariates in the model as in refs. 27,48. Patent count is an indicator of \ninnovation and research intensity for startups3,11. It was represented as a \ntime-dependent continuous variable indicating the number of patents \na given startup received each year48. Location is important for startups \nas it often determines access to key networks that can provide human \nand financial capital49 and sometimes highly specialized access to tech-\nnical knowledge and support through public, academic or corporate \npartnerships3,21,50. Location was represented as a constant binary vari-\nable indicating whether the startup was founded in a \u2018hotspot\u2019 state for \nclimate tech. The location in a \u2018hotspot\u2019 was operationalized as having \na density of greater than two climate-tech startups per 100,000 popu-\nlation (Supplementary Fig. 5), based on similar approaches that seek \nto account for variation in funding availability and other key factors \nbased on geographic location3,27. California, Colorado, Massachusetts \nand Washington, DC, were identified as climate-tech \u2018hotspots\u2019 under \nthis definition, which is consistent with findings in other literature8.\nAdditionally, several different ways of representing these variables \nin the model were tested to verify robustness. Each of the following \nwas tried separately and in combination when the structure of the \nmodel allowed it (clustering and frailty terms are incompatible, so \nthese could not be tried together). Investments by each investor type \nwere represented as continuous count variables rather than binary \nvariables and in US dollar amounts (Supplementary Tables 10 and 11). \nPublic grant funding was separated to identify international, national \nand subnational funding sources (Supplementary Tables 8 and 9). The \nlog of the number of patents was used instead of the count of patents. \nStartup sector was represented as a factor variable or accounted for \nthrough clustering rather than stratification. Frailty terms were used to \naccount for sector or startup specific effects. Each of these variations \nproduced similar results with coefficients of the same sign. Finally, \ninteraction terms were introduced for combinations of funding types \n(Supplementary Tables 18 and 19).\nSensitivity analysis\u2014time period and patenting activity\nWe further tested the sensitivity of our results by performing sepa-\nrate analyses of Cleantech 1.0 (startups founded 2005\u20132011) and 2.0 \n(startups founded 2012\u20132020) investment periods and by examining \nthe difference between startups that exhibit high-patenting activity vs \nlow-patenting activity. High patenting was defined as being in the top \ndecile of startups in terms of number of patents (greater than 12 patents). \nThe results of these analyses are provided in Supplementary Tables 13\u201316.\nThese robustness and sensitivity analyses attempted to account \nfor as many factors as possible. However, other potentially relevant \nvariables were not included in this analysis due to lack of data availabil-\nity, which is a common problem due to a widespread focus on secrecy \nin competitive startup sectors, or difficulty accessing granular data \nfor the large numbers of entities included in this analysis. If data were \navailable, future research could include considerations such as relation-\nships with founder networks, public\u2013private partnerships3, proximity \nto local branches of a corporate investor21, degree of digitalization34, \nstartup size and generalizability to other sectoral and country contexts.\nData availability\nThe main startup and investment dataset is proprietary and available \nfor purchase from the Cleantech Group at https://www.cleantech.com/. \nAdditional data on patents were obtained from the publicly available \nPatentsView API, and data on public grants were obtained from the \npublicly available datasets at https://www.usaspending.gov/. With \nthe available code and access to the i3 database and spreadsheets, our \nanalysis could be fully replicated.\nCode availability\nAll data-processing steps are outlined here, and the associated \ncode is publicly available at https://github.com/Climate-tech-Team/\nstartup-outcomes. Code is also provided to generate all figures in the \nmain text and Supplementary Information and to obtain patent data \nthrough the PatentsView API. With the available code and access to the \ni3 database and spreadsheets, our analysis could be fully replicated.\nReferences\n1.\t\nNet Zero Roadmap: A Global Pathway to Keep the 1.5\u2009\u00b0C Goal in \nReach; (IEA, 2023); https://iea.blob.core.windows.net/assets/ \nd954f15d-36c5-41b9-a693-9b74daef59cc/NetZeroRoadmap_ \nAGlobalPathwaytoKeepthe1.5CGoalinReach-2023Update.pdf\n2.\t\nGaddy, B. E., Sivaram, V., Jones, T. B. & Wayman, L. Venture capital \nand cleantech: the wrong model for energy innovation. Energy \nPolicy 102, 385\u2013395 (2017).\n3.\t\nDoblinger, C., Surana, K. & Anadon, L. D. Governments as \npartners: the role of alliances in U.S. cleantech startup innovation. \nRes. Policy 48, 1458\u20131475 (2019).\n4.\t\nShinkle, G. A. & Suchard, J.-A. Innovation in newly public firms: \nthe influence of government grants, venture capital, and private \nequity. Aust. J. Manage. 44, 248\u2013281 (2019).\n5.\t\nvan den Heuvel, M. & Popp, D. The Role of Venture Capital  \nand Governments in Clean Energy: Lessons from the First \nCleantech Bubble (National Bureau of Economic Research, 2022); \nhttps://doi.org/10.3386/w29919\n\nNature Energy | Volume 9 | July 2024 | 883\u2013893\n892\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01530-w\n6.\t\nMigendt, M., Polzin, F., Schock, F., T\u00e4ube, F. A. & von Flotow, P.  \nBeyond venture capital: an exploratory study of the finance- \ninnovation-policy nexus in cleantech. Ind. Corp. Change 26, \n973\u2013996 (2017).\n7.\t\nChan, G., Goldstein, A. P., Bin-Nun, A., Diaz Anadon, L. & \nNarayanamurti, V. Six principles for energy innovation. Nature \n552, 25\u201327 (2017).\n8.\t\nSurana, K. et al. Regional Clean Energy Innovation (Energy Futures \nInitiative & Univ. of Maryland Global Sustainability Initiative, \n2020); https://doi.org/10.13016/1BFC-DG1G\n9.\t\nSurana, K. et al. The role of corporate investment in start-ups for  \nclimate-tech innovation. Joule https://doi.org/10.1016/j.joule. \n2023.02.017 (2023).\n10.\t Goldstein, A., Doblinger, C., Baker, E. & Anad\u00f3n, L. D. Patenting \nand business outcomes for cleantech startups funded by the \nAdvanced Research Projects Agency-Energy. Nat. Energy 5, \n803\u2013810 (2020).\n11.\t\nHowell, S. T. Financing innovation: evidence from R&D grants.  \nAm. Econ. Rev. 107, 1136\u20131164 (2017).\n12.\t Mazzucato, M. & Semieniuk, G. Financing renewable energy: who \nis financing what and why it matters. Technol. Forecasting Social \nChange 127, 8\u201322 (2018).\n13.\t Foray, D., Mowery, D. C. & Nelson, R. R. Public R&D and social \nchallenges: what lessons from mission R&D programs? Res. Policy \n41, 1697\u20131702 (2012).\n14.\t Rossi, M., Festa, G., Solima, L. & Popa, S. Financing knowledge- \nintensive enterprises: evidence from CVCs in the US. J. Technol. \nTransf. 42, 338\u2013353 (2017).\n15.\t Chesbrough, H. W. Making sense of corporate venture capital. \nHarv. Bus. Rev. 80, 90\u201399 (2002).\n16.\t Benson, D. & Ziedonis, R. H. Corporate venture capital as a \nwindow on new technologies: implications for the performance \nof corporate investors when acquiring startups. Organ. Sci. 20, \n329\u2013351 (2008).\n17.\t Hegeman, P. D. & S\u00f8rheim, R. Why do they do it? Corporate \nventure capital investments in cleantech startups. J. Clean. Prod. \n294, 126315 (2021).\n18.\t Dushnitsky, G. & Lenox, M. J. When does corporate venture capital \ninvestment create firm value? J. Bus. Venturing 21, 753\u2013772 (2006).\n19.\t Cunningham, C., Ederer, F. & Ma, S. Killer acquisitions.  \nJ. Polit. Econ. 129, 649\u2013702 (2021).\n20.\t Maula, M., Autio, E. & Murray, G. Corporate venture capitalists \nand independent venture capitalists: what do they know, who  \ndo they know and should entrepreneurs care? Venture Cap. 7, \n3\u201321 (2005).\n21.\t Alvarez-Garrido, E. & Dushnitsky, G. Are entrepreneurial venture\u2019s \ninnovation rates sensitive to investor complementary assets? \nComparing biotech ventures backed by corporate and independent \nVCs: comparing biotech ventures backed by corporate and \nindependent VCs. Strategic Manage. J. 37, 819\u2013834 (2016).\n22.\t Maula, M. V. J., Autio, E. & Murray, G. C. Corporate venture capital \nand the balance of risks and rewards for portfolio companies. \nSpec. Issue Corp. Entrep. 24, 274\u2013286 (2009).\n23.\t Yao, T. & O\u2019Neill, H. Venture capital exit pressure and venture exit: \na board perspective. Strategic Manage. J. https://doi.org/10.1002/ \nsmj.3432 (2022).\n24.\t Masulis, R. W. & Nahata, R. Venture capital conflicts of interest: \nevidence from acquisitions of venture-backed firms. J. Financ. \nQuant. Anal. 46, 395\u2013430 (2011).\n25.\t Cleantech Group I3 Database (i3, 2024); https://www.cleantech.com/\n26.\t Park, H. D. & Steensma, H. K. When does corporate venture capital \nadd value for new ventures? Strategic Manage. J. 33, 1\u201322 (2012).\n27.\t Momtaz, P. P. Econometric models of duration data in \nentrepreneurship with an application to start-ups\u2019 time-to-funding \nby venture capitalists (VCs). J. Appl. Stat. 48, 2673\u20132694 (2021).\n28.\t Therneau, T., Crowson, C. & Atkinson, E. Using Time Dependent \nCovariates and Time Dependent Coefficients in the Cox Model \n(Comprehensive R Archive Network, 2022); https://cran.r-project. \norg/web/packages/survival/vignettes/timedep.pdf\n29.\t Cook, R. J & Lawless, J. F. The Statistical Analysis of Recurrent \nEvents; Statistics for Biology and Health (Springer, 2007).\n30.\t Cornelli, G., Frost, J., Gambacorta, L. & Merrouche, O. Climate \nTech 2.0: Social Efficiency Versus Private Returns BIS Working \nPaper number 1072 (Bank for International Settlements, 2023); \nhttps://www.bis.org/publ/work1072.pdf\n31.\t Inflation Reduction Act of 2022 (US House, 2022); https://www. \ncongress.gov/bill/117th-congress/house-bill/5376\n32.\t Infrastructure Investment and Jobs Act (US House, 2021);  \nhttps://www.congress.gov/bill/117th-congress/house-bill/ \n3684/text\n33.\t Rozansky, R. Becoming America\u2019s Seed Fund: Why NSF\u2019s \nSBIR Program Should Be a Model for the Rest of Government \n(Information Technology and Innovation Foundation, 2019); \nhttps://itif.org/publications/2019/09/26/becoming-americas- \nseed-fund-why-nsfs-sbir-program-should-be-model-rest/\n34.\t Kurowski, S. From atoms to bits: resource mobilization of \nnon-digital, hybrid, and digital cleantech startups. Heliyon 10, \ne23697 (2024).\n35.\t Mazza, P. & Shuwaikh, F. Industry-relatedness, geographic \nproximity and strategic decisions of corporate and independent \nventure capital-backed companies. J. Small Bus. Manage.  \nhttps://doi.org/10.1080/00472778.2022.2108432 (2022).\n36.\t Sj\u00f6holm, F. Navigating the New Normal: The European Union\u2019s \nChanging Stance on Globalization in the Era of Trade Conflicts \nIFN Working Paper number 1466 (Research Institute of  \nIndustrial Economics, 2023); https://www.ifn.se/media/\np34dyzls/wp1466.pdf\n37.\t Katila, R., Rosenberger, J. D. & Eisenhardt, K. M. Swimming with \nsharks: technology ventures, defense mechanisms and corporate \nrelationships. Adm. Sci. Q. 53, 295\u2013332 (2008).\n38.\t Le Marois, J.-B. & Bennett, S. Tracking Clean Energy Innovation \n(IEA, 2020); https://www.iea.org/reports/tracking-clean-energy- \ninnovation\n39.\t Belcher, K., Maitra, M., Pizzuto, E., Donovan, D. & Oftedal, E. The \nFuture of Climate Tech (Silicon Valley Bank, 2021); https://www. \nsvb.com/globalassets/trendsandinsights/reports/future-of- \nclimate/svb-future-of-climate-tech-report.pdf\n40.\t State of Climate Tech 2021 (PwC, 2021).\n41.\t Clean Energy Innovation (IEA, 2020); https://www.iea.org/reports/ \nclean-energy-innovation\n42.\t Global Climate Tech Venture Capital\u2013Full Year 2021 (HolonIQ, \n2022); https://www.holoniq.com/notes/global-climatetech- \nvc-report-full-year-2021/\n43.\t Climate Tech Investment Trends\u2013Five Years on Since the Paris \nAgreement (Dealroom.co, 2021); https://dealroom.co/blog/ \nclimate-tech-investment-trends\n44.\t Venture Capital, PE Invest $53.7 Billion in Climate Tech \n(BloombergNEF, 2022); https://www.bloomberg.com/professional/\nblog/venture-capital-pe-invest-53-7-billion-in-climate-tech/\n45.\t Baker, C. Patentsview: An R Client to the \u2018PatentsView\u2019 API  \nversion 0.3.0 (2021); https://cran.r-project.org/web/packages/ \npatentsview/index.html\n46.\t Kassambara, A., Kosinski, M. & Biecek, P. Survminer: Drawing \nSurvival Curves Using \u2018Ggplot2\u2019 version 0.4.9 (2021); https://cran. \nr-project.org/web/packages/survminer/index.html\n47.\t Therneau, T. M. & Grambsch, P. M. Modeling Survival Data: \nExtending the Cox Model (Springer, 2001).\n48.\t Keogh, D. & K.N. Johnson, D. Survival of the funded: econometric \nanalysis of startup longevity and success. J. Entrepreneurship. \nManage. Innovation 17, 29\u201349 (2021).\n\nNature Energy | Volume 9 | July 2024 | 883\u2013893\n893\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01530-w\n49.\t Saxenian, A. Regional networks and the resurgence of Silicon \nValley. Calif. Manage. Rev. 33, 89\u2013112 (1990).\n50.\t Guzman, J. & Stern, S. Where is Silicon Valley? Science 347, \n606\u2013609 (2015).\nAcknowledgements\nFunding for this research was provided by the Energy and Environment \nProgram at the Alfred P. Sloan Foundation under grant number \nG-2021-14177 (K.M.K., K.S., M.R.E., M.A.B., Z.H.T., N.E.H., E.D.W.). K.S. \nacknowledges support from the BMK (Austrian Federal Ministry \nfor Climate Action, Environment, Energy, Mobility, Innovation and \nTechnology) under the BMK endowed professorship for data-driven \nknowledge generation: climate action. We thank R. Fedorchak and R. \nLucas for their assistance with data cleaning.\nAuthor contributions\nConceptualization: K.S., M.R.E., K.M.K. Data curation: K.M.K., K.S., \nZ.H.T., M.A.B. Formal analysis: K.M.K. Funding acquisition: K.S., M.R.E., \nN.E.H. Methodology: K.M.K., K.S., C.D. Project administration: K.S., \nM.R.E., K.M.K., N.E.H. Software: K.M.K., K.S., Z.H.T. Supervision: K.S., \nK.M.K. Visualization: K.M.K. Writing\u2013original draft: K.M.K. Writing\u2013\nreview and editing: K.M.K., K.S., C.D., M.R.E., E.D.W., N.E.H.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s41560-024-01530-w.\nCorrespondence and requests for materials should be addressed to \nKathleen M. Kennedy or Kavita Surana.\nPeer review information Nature Energy thanks Hans L\u00f6\u00f6f and the \nother, anonymous, reviewer(s) for their contribution to the peer review \nof this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2024, corrected publication 2024\n\n\n    Scientific Research Findings:", "answer": "Corporate investments in climate\u2011tech have consistently increased since 2005, even when other sources of investment declined. The startups they invest in are more likely to achieve a successful exit. Public grants are not significantly associated with exits on their own but are likely to act as catalysts for high\u2011risk startups. Publicly funded startups exit at a higher rate with the addition of corporate investment (155% increase) compared with other private investment (78% increase). Additionally, corporate investment has not been significantly associated with failure in more recent years (2012\u20132020), suggesting corporations may have learned from earlier losses and could play an increasingly important role in future climate\u2011tech innovation. Although these results may indicate corporate investors are better at selecting startups destined for success, the correlations we observe are unlikely to arise purely from selection bias because investments are correlated with failure as well as success.", "id": 8}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 9 | June 2024 | 691\u2013702\n691\nnature energy\nhttps://doi.org/10.1038/s41560-024-01505-x\nArticle\nEU carbon prices signal high policy \ncredibility and farsighted actors\nJoanna Sitarz\u2009\n\u200a\u20091,2\u2009\n, Michael Pahle\u2009\n\u200a\u20091, Sebastian Osorio\u2009\n\u200a\u20091, Gunnar Luderer1,2 & \nRobert Pietzcker\u2009\n\u200a\u20091\u2009\nCarbon prices in the EU emissions trading system are a key instrument \ndriving Europe\u2019s decarbonization. Between 2017 and 2021, they surged \ntenfold, exceeding \u20ac80 tCO2\n\u22121 and reshaping investment decisions across \nthe electricity and industry sectors. What has driven this increase is an \nopen question. While it coincided with two significant reforms tightening \nthe cap (\u2018MSR reform\u2019 and \u2018Fit for 55\u2019), we argue that a reduced supply of \nallowances alone cannot fully explain the price rise. A further crucial aspect \nis that actors must have become more farsighted as the reform signalled \npolicymakers\u2019 credible long-term commitment to climate targets. This \nis consistent with model results that show historic prices can be better \nexplained with myopic actors, whereas explaining prices after the reforms \nrequires actors to be farsighted. To underline the role of credibility, we test \nwhat would happen if a crisis undermines policy credibility such that actors \nbecome myopic again, demonstrating that carbon prices could plummet \nand endanger the energy transition.\nThe EU emissions trading system (EU ETS) is a central pillar of the Euro-\npean Union\u2019s decarbonization strategy. It covers the electricity sector, \nlarge-scale industrial installations, aviation and maritime transport and \nhence controls above 40% of the European Union\u2019s total greenhouse gas \nemissions1. Over a period with two major reforms of the ETS and notably \na substantial tightening of the cap, the carbon market underwent a \nremarkable transition: carbon prices increased tenfold within four years, \nwith a first rise in 2018 from below \u20ac10 tCO2\n\u22121 to a plateau at \u20ac20\u201330 \ntCO2\n\u22121 in 2019\u20132020 and then a second, even sharper, rise during which \nprices repeatedly reached almost \u20ac100 tCO2\n\u22121 in 2021 and 20222. The \nquestion of why prices have risen so steeply is still unanswered, though, \nand a subject of debate among the scientific and policy community.\nThe literature so far identifies various factors as playing a potential \nrole in carbon price developments in general: (1) regulatory changes \n(such as the introduction of the Market Stability Reserve (MSR) or \nchanges in the linear reduction factor)3\u20135, (2) actors\u2019 behaviour (fore-\nsight horizon, hedging or participation in trade)6,7 and (3) speculation \nand external financial investors8\u201310. However, most work focuses on one \nof those aspects, provides only qualitative assessments and covers only \nthe period before the recent reforms and price increases.\nWith a view on understanding what has driven prices in the recent \nperiod, the following puzzle arises. It is economically straightforward \nthat a tightening of the long-term cap should increase current and \nexpected prices. However, past research suggests that market partici-\npants in the ETS are myopic7,11. Whereas myopia can always have an impact \non energy sector investments, it is especially relevant when the power \nsector is covered by an intertemporal emissions trading system with a \ncap that strongly tightens over time, so that future certificate scarcities \ncan influence current investments. If most market actors were myopic, a \nlong-term tightening of the cap should thus only have modest effects on \ncurrent prices, much lower than the observed increase after the reforms.\nIn light of that, we hypothesize that the reform could have had \nanother important effect on actors: making them more farsighted. \nThe reason is that through the reform EU policymakers substantially \nfirmed up the credibility of their commitment to the ETS overall. They \ndid this both explicitly, by emphasizing that the \u2018ETS is front and centre \nto all our efforts\u201912, and implicitly, by investing a lot of political capital \nin the political negotiation. More broadly, a recent empirical study \nalso shows that the European Union has currently the world\u2019s highest \nclimate policy credibility13.\nReceived: 31 March 2023\nAccepted: 13 March 2024\nPublished online: 30 May 2024\n Check for updates\n1Potsdam Institute for Climate Impact Research, Member of the Leibniz Association, Potsdam, Germany. 2Global Energy Systems Analysis, Technische \nUniversit\u00e4t Berlin, Berlin, Germany. \n\u2009e-mail: joanna.sitarz@pik-potsdam.de; robert.pietzcker@pik-potsdam.de\n\nNature Energy | Volume 9 | June 2024 | 691\u2013702\n692\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nDifferent publications furthermore suggest that the limited fore-\nsight of compliance actors contributed to low carbon prices4,6,7,18. To \nunderstand the role of foresight, one needs to consider that the EU ETS \nallows for almost unlimited forward bankability: any certificate not \nused today can be used in the future. Hence, expected future prices \nmay have a strong influence on today\u2019s prices. In contrast, in a market \nwithout bankability, a surplus of certificates over emissions would \nmean the certificate price in that year is zero, as the unused certificates \nbecome worthless at the end of the year.\nNow, many firms might not consider the long-term future (inher-\nently, or due to regulatory uncertainty and lack of policy credibility) but \nrely on short-term planning horizons of for example, five to ten years19. \nIf allowances scarcity occurs outside their planning horizon, they will \nnot anticipate it and hence don\u2019t have incentives to bank certificates \nfor the future nor decrease emissions in the short term. Consequently, \nthe carbon price will stay lower and decarbonization will be slower than \nif actors were farsighted (Fig. 2).\nThus, for many years the EU ETS failed in establishing a carbon \nprice that would drive deep decarbonization. In period (1), actors pre-\nsumably acted myopically, a behaviour leading to low carbon prices. \nHowever, just a few years later, EU ETS prices are stronger than ever2. \nWhat happened since 2017? Which mechanisms drove the rise in carbon \nprices observed in periods (2) and (3)? A plausible explanation would \nbe that prices simply increased because reforms tightened the cap20. \nHere we present a more comprehensive explanation: the reforms had \nthe side effect that market actors also became less myopic, which drove \nprices up. Therefore, we first give an overview of the most relevant \nreforms from the past years and then present our modelling results.\nThe past years were marked by numerous reforms and rapid EU \nclimate policy developments21\u201327. While it is challenging to pinpoint \none specific regulation with the highest impact on carbon prices, we \ncan, generally, speak about an intensive period in climate policy since \n2015 with two crucial ETS reform periods: the \u2018MSR reform\u2019 and the \u2018Fit \nfor 55\u2019 package, as summarized in Table 1.\nOur modelling findings are divided into two segments. We first \npresent results supporting our hypothesis that actors have extended \ntheir foresight, which strongly impacted historical carbon prices. \nHereafter, we turn to the role of external financial investors, who have \nbeen gaining attention throughout literature and media8\u201310,28\u201330, to \ndelimit their possible impact on the carbon price surge.\nFigure 3 shows our modelling results on the impact of reforms and \nactors\u2019 foresight on carbon prices (see also Extended Data Figs. 1\u20136). \nFirst of all, one can see between period (1) and period (2), when the \nMSR reform was negotiated and implemented, actors presumably \nstarted to look further into the future. When turning to period (1) \nbefore 2018, one notices that observed ETS prices are closer to the \nmodelled prices for myopic actors than to the modelled prices for \nfarsighted actors. It seems therefore plausible to assume that market \nactors behaved at least partially myopically, which is in line with ear-\nlier assessments7. For periods (2) and (3), one observes the opposite: \nboth, the 2019\u20132020 observed ETS prices of \u20ac20\u201330 tCO2\n\u22121 and the \n2021\u20132022 ones of \u20ac70\u201390 tCO2\n\u22121, are consistent with the modelled \nprices for farsighted actors (that is, perfect foresight trajectories for \nold \u2018MSR reform\u2019 targets, and new \u2018Fit for 55\u2019 targets, respectively). We \nalso calculate the Mean Average Percentage Error (MAPE) between \nthe modelled and historical prices (Extended Data Tables 1\u20134), which \nconfirms the visual conclusions drawn from Fig. 3.\nHence, regarding the first rise at the beginning of period (2), a \nhypothesis following our results is that prices increased due to a grad-\nual switch from actors\u2019 short- to long-term foresight, which might have \nbeen triggered, among other things, by the MSR reform tightening \nthe cap and strengthening the MSR. Whereas our results indicate that \nthe direct effect of the reform\u2014the tighter emissions budget\u2014cannot \nexplain the substantial increase in prices under the assumption of \ncontinued myopia, the reform might have had a strong indirect impact: \nSuch instilled credible commitment is essential to shape firms\u2019 \nexpectations about the durability of long-term policies such as the \nETS14, and indeed studies suggest that low policy credibility can be \nassociated with decreased green investments15, and that policy cred-\nibility can enhance actors\u2019 farsightedness16. The main reason is that \nlow credibility creates high regulatory uncertainty regarding a future \nsoftening of the cap or interventions to dampen high carbon prices\u2014\na major reason for myopia. Correspondingly, increasing credibility \nimplies that actors become more farsighted.\nResearch is still outstanding on whether myopia remains a prevalent \ninfluence within the current EU ETS. Equally, there has been no investiga-\ntion into whether any shifts in the foresight horizon have occurred and \ntheir potential impact on the recent surge in carbon prices.\nWe fill this gap by providing a model-based analysis of the EU ETS \nwith a specific emphasis on the influence of actors\u2019 foresight horizon. \nThe contribution of our work is threefold. We first analyse the past: \nbringing together the impact of political reforms, the foresight of \ncompliance actors and the role of external investors, we show which \nmix of those mechanisms could explain the observed strong rise in \ncarbon prices. We discuss the present: by computing marginal carbon \nprices necessary to drive the decarbonization of the electricity and \nindustry sectors in line with the new EU\u2019s 2030 goals as set in the \u2018Fit \nfor 55\u2019 package, we assess that current ETS prices correspond to the \noptimal market-efficient carbon price trajectory. We turn to the future: \nhaving understood the mechanisms that could plausibly have led to the \nobserved increase in carbon prices, we explore in how far this positive \ndevelopment is vulnerable and potentially could be reversed. We close \nwith policy recommendations on how to secure the energy transition \nin light of our results.\nFrom past to present\nWhen analysing past carbon prices (Fig. 1), one can broadly break down \nthe timeline into three periods with distinct price regimes: (1) the period \nof 2008\u20132017, in which prices first dropped and then stabilized at a \nlow level below \u20ac10 tCO2\n\u22121, (2) the period of 2018\u20132020, the first rise \nup to a plateau of \u20ac20\u201330 tCO2\n\u22121 and (3) the period since late 2020, the \nsecond rise, in which prices increased strongly and are now stabilizing \naround \u20ac70\u201390 tCO2\n\u22121. What might have been the main mechanisms \ndriving these three regimes, and, in particular, what role could actors\u2019 \nforesight have played?\nRegarding the first period (1), the common understanding is that \nprices dropped because of a high surplus of allowances that accumu-\nlated since 2008. The financial crisis reduced emissions more than \nanticipated, leaving compliance actors with a high number of unused \nallowances, hence limiting incentives to decarbonize17.\n(1)\n(2)\n(3)\n0\n25\n50\n75\n100\n2008\n2010\n2012\n2014\n2016\n2018\n2020\n2022\n2024\nYear\nCarbon price (\u20ac tCO2\n\u22121)\nFig. 1 | Evolution of carbon prices on the EU ETS from 2008 to 2023. Prices \ncorrespond to historical EU ETS allowances (EUA) prices on the EEX spot \nmarket2,68. The year tick marks the beginning of a year. We categorize the price \nevolution into three periods: (1) initial decline and stabilization, (2) first rise and \n(3) second rise.\n\nNature Energy | Volume 9 | June 2024 | 691\u2013702\n693\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nthe negotiations and ultimate implementation of the reform over 2017 \nand 2018 (Table 1) emphasized the will of EU policymakers to \u2018repair\u2019 \nthe ETS (showing \u2018the doctor has not given up on the patient\u201931), which \nstrongly increased its long-term credibility, inspiring market actors \nto show longer foresight. These findings align with previous assess-\nments, which, on the one hand, demonstrate that the MSR can lead to \nincreased carbon prices32,33, whereas, on the other hand, argue that the \neffect of the MSR reform on the emissions budget alone is unlikely to \nfully explain the surge in carbon prices6,34.\nSecondly, Fig. 3 shows that the \u2018Fit for 55\u2019 package sharply increases \nthe stringency of the EU ETS. Optimal carbon prices (that is, obtained \nunder the assumption of perfect foresight) to reach the new targets \nare substantially higher than those that were necessary for achieving \nprevious goals. In fact, modelled prices for the \u2018Fit for 55\u2019 targets for \n2020\u20132023 are in the order of \u20ac70\u201390 tCO2\n\u22121, corresponding well to \nobserved 2021\u20132023 prices on the EU ETS, thus supporting the hypothe-\nsis that actors have transitioned towards a more farsighted perspective.\nFigure 4 discusses the final point of this section: could an influx \nof long-term investors explain the strong rise of carbon prices if other \nactors had remained myopic? Here we assume external investors tem-\nporarily block a part of the allowances on the market, which then cannot \nbe used by compliance actors to cover their emissions during the period \n(Methods). This influences the price trajectory: when external investors \nbuy, prices go up; when they sell, prices can go down.\nIn reality, it is estimated that external investors currently hold \nonly around 5\u201310% of allowances futures9, consistent with the scenario \nin which 5% of auctioned allowances are bought by external financial \ninvestors. This scenario shows only a small price increase of less than \n\u20ac10 tCO2\n\u22121 in 2025 compared with the pure myopic scenario (Fig. 4). \nThus, following our results, a major contribution of external investors \nto the price rise seems unlikely. What is on the other hand possible, is \nthat they acted as a catalyser, speeding up the process of compliance \nactors switching to longer foresight and anticipating the consequences \nof the \u2018Fit for 55\u2019 package.\nTo summarize, we provide a possible explanation of the past: we \nshow that the two price rises (first to \u20ac20\u201330 tCO2\n\u22121 and more recently \nto \u20ac70\u201390 tCO2\n\u22121) are consistent with a first regulatory reform that had \nlimited impact on the cumulative certificate budget but contributed to \na switch of actors\u2019 behaviour from myopic to farsighted and a second \nreform that substantially tightened the emissions cap. Whereas exter-\nnal investors may have accelerated the transition, it seems improbable \nthat prices are artificially high solely due to their activity.\nFurthermore, our results provide insights about the present state \nof the EU ETS. Our modelling indicates that observed 2022 and 2023 \nprices of around \u20ac80 tCO2\n\u22121 put the ETS sectors on track to achieving \ntheir reduction targets set by the Climate Law, a result in line with \nearlier findings35.\nOur findings suggest that actors became farsighted, which is \nconsistent with the initially formulated hypothesis that the ETS reform \nheightened policy credibility. Overall, there are thus reasons for careful \noptimism: trust in the EU ETS revived, policy credibility seems high, \nactors are therefore farsighted and current prices are in line with EU\u2019s \na\nPerfect foresight\nTime\nMyopic foresight\nBanked allowances\nAllowances cap\nPlanned emissions (perfect foresight)\nPlanned emissions (myopic foresight)\nConsidered time horizon/budget in respective foresight mode\nHorizon 1\nConsidered\nbudget\nCumulative emissions (tCO2)\nCumulative emissions (tCO2)\nCarbon price (\u20ac t\u22121 CO2)\nCarbon price (\u20ac t\u22121 CO2)\nConsidered\nbudget\nYears into the future\nYears into the future\n5\n10\n15\n20\n25\n30\n35\n40\n(tCO2)\n(tCO2)\nb\nc\n5\n10\n15\n20\n25\n30\n35\n40\nTime\nHorizon 1\nHorizon 2\nTime\nPerfect foresight (40 years)\nMyopic foresight (ten years, revision after five years)\nFinal (realized) values\nProjected values in each horizon\nFinal (realized) values accounting for revision\nTime\nFig. 2 | Stylized emissions and carbon price trajectory with short-term, \nmyopic, and long-term, perfect, foresight. Simple cap and trade system \nwithout the Market Stability Reserve, for illustrative purposes. a, Example of a \nplanning horizon at the beginning of the transformation. With a myopic foresight \nof ten years, there is no (or very weak) incentive to reduce planned emissions. \nWith a perfect foresight, future scarcity is anticipated and planned emissions \nget reduced already in the near-term. b, Cumulative emissions over the whole \ntransformation period. Myopic foresight leads to delayed decarbonization. c, \nCarbon prices over the whole transformation period. Myopic foresight leads to \nvery low carbon prices in the near term. Short lines correspond to the specific \nhorizons: every five years a new foresight horizon of ten years starts.\n\nNature Energy | Volume 9 | June 2024 | 691\u2013702\n694\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\ngoals. However, the question arises: are these changes principally \nreversible? In particular, could credibility plummet again, implying \nthat actors return to myopic behaviour? If yes\u2014why? And what would \nit mean for carbon prices and the energy transition?\nA look into the future\nThe previous section has shown that the recent rise in ETS prices does \nnot result from an acute scarcity of allowances\u2014as their surplus is \nstill vast\u2014but can rather be explained by market actors having turned \nmore farsighted. A plausible interpretation is that this was due to the \nlong-term cap becoming considerably more credible. This may suggest \nthat from this point on, ETS prices would only increase. However, what \nwould happen if policy credibility gets shaken again due to a crisis or \npolitical backlash? How much would prices plummet and what would it \nimply for the energy transition? This is what we analyse in the following \nwith a scenario considering a shock\u2014for illustrative purposes\u2014in 2025.\nBefore turning to numerical results, to underpin our motivation \nfor analysing a shock scenario, we develop a conceptual model on how \npolicy credibility, actors\u2019 foresight and carbon prices influence each \nother (Fig. 5). The left side schematically represents the events from \n2017 to 2021. It starts from a state with low policy credibility due to the \nhuge certificate surplus, myopic market actors and ensuing low carbon \nprices. Then, the MSR reform and the higher ambition in the \u2018Fit for \n55\u2019 package substantially strengthened the policy credibility and set \nin motion a reinforcing cycle: actors extend their foresight horizon, \nwhich in turn increases carbon prices, which (1) may increase the policy \ncredibility and (2) attracts non-compliance actors to the market with \nat least partially more long-term investment strategies.\nThe right side shows a path, how the current situation could \nunravel again: a price shock or a crisis and the ensuing political reac-\ntions could potentially reduce policy credibility and trigger a relapse \ninto myopic behaviour and hence lower prices. The recent energy crisis \nserves as an illustrative example: the tenfold increase36 of European \ngas prices in 2022 put pressure on the EU ETS from several directions.\nFirst of all, rising energy prices created strong liquidity prob-\nlems for many firms37. Under liquidity problems, firms might sell \nassets not required in the short-term\u2014such as banked CO2 certificates, \nwhich could decrease prices and scare away external investors. Sec-\nondly, the rising energy prices directly created pressure to weaken \nclimate policies. As an example, Poland repeatedly proposed to freeze \nTable 1 | An intensive period in climate policy between 2015 and 2022. Developments and reforms relevant for the EU ETS\nDate\nEvent\nImpact on climate policy/the EU ETS\nDec. 2015\nAdoption of the Paris Agreement21\nGlobal climate policy: \u2018goal to limit global warming to well below 2, preferably to 1.5 degrees \nCelsius, compared to pre-industrial levels\u2019.\nOct. 2015\nDecision on the establishment of an MSR22\nEU ETS design: new mechanism with pre-defined rules addressing the high surplus of \nallowances. Depending on the total number of allowances in circulation (TNAC), allowances get \nplaced in the reserve or released from the reserve. However, as all certificates are to be released \nin the long term, this reform implies NO tightening of the intertemporal emission cap and thus \nhad little impact on market prices.\nOct. 2016\nRatification of the Paris Agreement23\nEU climate policy: all parties (including the EU) having adopted the Paris Agreement are \nrequired to submit an NDC till 2020, outlining their post-2020 climate actions.\nFeb. 2017\n\u2018MSR reform\u2019\nProposal of ETS/MSR reform for trading period \nIV (2021\u20132030)\u2014strengthening the MSR and \ntightening ETS targets24\nEU ETS design: Parliament and Council formulate their ETS/MSR reform proposals. Council \nproposes automatic cancellation of certificates in the MSR above a threshold.\nNov. 2017\n\u2018MSR reform\u2019\nFinal agreement on ETS/MSR reform for trading \nperiod IV (2021\u20132030)25,26\nEU ETS design: after six trilogues, Commission, Parliament and Council reach an agreement on \nthe ETS/MSR reform. Tightening of the cap: linear reduction factor of allowances (in percentage \npoints of 2005 cap) increases from 1.74 to 2.2. Strengthening of the MSR: higher intake and \ncertificate cancellations from 2024 on.\nMarch 2018\n\u2018MSR reform\u2019\nETS/MSR reform for trading period IV (2021\u2013\n2030) officially adopted27\nEU ETS design: directive with the ETS/MSR reform officially published.\nDec. 2019\nPresentation of the European Green Deal62\nEU climate policy: EU Commission presents EU\u2019s new climate action strategy including the goal \nof climate neutrality in 2050 and an emissions reduction of 50\u201355% until 2030, compared with \n1990 levels.\nMarch 2020\nProposal for a European Climate Law63\nEU climate policy: EU Commission presents legislative proposal of a law setting the objective \nfor the EU to become climate neutral by 2050.\nSept. 2020\nProposal to set an EU-wide 55% emissions \nreduction target for 203064\nEU climate policy: EU Commission amends the Climate Law proposal by introducing the \nupdated 2030 climate target of a net reduction of at least 55% of EU\u2019s greenhouse gas \nemissions compared with 1990 levels.\nDec. 2020\nThe Council agrees on the 55% reduction target \nfor 203065\nEU climate policy: The Council of the EU reaches an agreement on an approach on the climate \nlaw proposal, including an agreement to the updated 2030 climate target of a net reduction of \nat least 55% of EU\u2019s greenhouse gas emissions compared with 1990 levels.\nApril 2021\nFinal agreement on Climate Law66\nEU climate policy: Parliament, Council and Commission agree in trilogue negatiations on the \n\u221255% 2030 reduction target, enabling the formal adoption of the Climate Law in June 2021.\nJuly 2021\n\u2018Fit for 55\u2019 package54,67\nEU ETS design: EU Commission publishes package of legislative proposals to meet the \n2030 emissions reduction target of 55%. For the EU ETS it includes: steeper annual emission \nreductions, strengthening of the MSR, gradual removal of free allowances for the aviation \nsector and the inclusion of the maritime sector into the current EU ETS.\nDec. 2022\nAgreement on EU ETS \u2018Fit for 55\u2019 proposal39\nEU ETS design: Parliament, Council and Commission reach final agreement during trilogue \nnegotiations on the EU ETS \u2018Fit for 55\u2019 proposal. Ambitions are kept high: all main elements from \nthe initial proposal remain; the emissions cap gets slightly more tightened compared to the \ninitial proposal.\nHighlighted in bold are events directly affecting the EU ETS design.\n\nNature Energy | Volume 9 | June 2024 | 691\u2013702\n695\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\ncarbon prices at \u20ac30 tCO2\n\u22121 (ref. 29) or even temporarily suspend the \nEU ETS38. If the EU were to give in to such proposals, it would decrease \nits long-term policy credibility, and hence, following our hypothesis, \ncompliance actors\u2019 foresight. Given the trilogue results in late 202239, \nit appears the European Union managed to overcome this critical \nsituation without weakening the ETS and undermining its long-term \npolicy credibility.\nNevertheless, the future remains uncertain, with the energy crisis \nserving as just one recent illustration of potential risks. Political crisis \ncan happen anytime and history has repeatedly shown that all policy \nreforms face the threat of being undone or weakened over time40. \nThis emphasizes the importance of exploring the risks of undermined \npolicy credibility and actors returning to myopia. More specifically, to \nsafeguard against such a turn of events, it is important to quantify what \nwould be lost in terms of price degradation, and how this would slow \ndown the energy transition.\nFigure 6 shows the price trajectory of such a \u2018reversal to myopia\u2019 \nscenario. It presumes actors were myopic in the past, became farsighted \naround 2020 and turn fully myopic again in 2025. Prices could then start \nfalling, reaching a level below \u20ac30 tCO2\n\u22121 in 2025. There is currently \nno mechanism ensuring prices stay high in the next years. Assuming \nsuch a relapse into myopia really happens and prices fall in the near \nfuture below \u20ac30 tCO2\n\u22121, what would it mean for the energy transition? \nThe general impacts of myopic foresight in the energy sector have \nbeen studied in previous literature41\u201346. Nerini et al.47 show using the \ncross-sectoral capacity expansion model UK Times that myopia might \nresult in delayed climate action and higher total transformation costs, \ncompared with the pathway set by a perfect foresight model. On the \none hand, emissions abatement gets delayed. On the other hand, the \nsolutions chosen are focused on the near-term, creating lock-ins and \nnot the most efficient ones from a long-term perspective.\nTo illustrate the delayed action, we focus on the electricity sec-\ntor. The major problem we identify under the relapse to myopia is \nthat, as seen in Fig. 7, delayed investments into wind capacity in turn \ndelay the phase-out of coal power generation. As illustrated in Fig. 7a, \nour modelling shows that myopia could massively slow down wind \ncapacity expansion in the next ten years, with yearly investments \nreduced by a factor of three, compared with the cost-optimal (that \nis, perfect foresight) trajectory. The missing wind power in combi-\nnation with low carbon prices would strongly delay the phase-out \nof coal (Fig. 7b).\nThese risks are examples of what can happen in the electricity sec-\ntor. Any delay poses the risk of climate targets becoming out of reach \nor being reachable only at very high costs, as feasible roll-out rates can \nbe limited, for example, due to the availability of skilled workers or \nproduction capacities48. Likewise, the required steeper carbon price \nin the long-term might increase the likelihood of a political backlash \nthat dismantles the policy49. Hence, it is crucial to be aware that carbon \nprices could principally fall again in the near future, with strong conse-\nquences for the energy transition. Exploring potentials of a price floor \nin the ETS, proposed in the past to address the problem of myopia18 and \ndesigning complementary policy instruments to shore up the energy \ntransition thus remains critical\u2014despite currently high carbon prices.\nConclusions\nThis work proposes a quantitative explanation behind the observed rise \nin carbon prices on the EU ETS since 2017. We use a numerical model \n(see Methods) to simulate different foresight horizons of compliance \nactors and to depict the role of external investors. We show that the \ncombination of stricter EU ETS policies and changed behaviour of com-\npliance actors from myopic to farsighted can explain the rapid increase \nin carbon prices over the past years, underpinning with a quantitative \nanalysis earlier scholarly work emphasizing the role of myopia7,19. Our \nresults indicate that external investors probably only played a minor \nrole by, for example, accelerating the price rise.\nWe discuss the hypothesis that policy credibility impacts actors\u2019 \nforesight and hence carbon prices. Consequently, a glimpse into the \n2013\n0\n25\nPerfect foresight\nMyopic foresight\n50\n75\n100\n(1)\n(2)\nMSR reform\n(3)\nFit for 55\n2014\n2015\n2016\n2017\n2018\n2019\n2020\n2021\n2023\n2022\nYear\nModel results\nEU ETS\ntargets\nPre-reforms\nForesight\nPerfect\nMyopic\nReal developments\nActual EU ETS\nprices\nMSR reform\nFit for 55\nCarbon price (\u20ac tCO2\n\u22121)\nFig. 3 | Impact of reforms and actors\u2019 foresight on carbon prices. Historical \ncarbon prices on the EU ETS and modelled carbon price trajectories with the \nassumption of either perfect or myopic foresight over the three periods (1)\u2013(3) \nas defined in Fig. 1. For each period, we show the carbon price trajectories \nrequired to reach the target that was valid during that period. Thus, jumps in \nsame-coloured trajectories between one period and the next show the effect that \nthe change in the ETS targets and MSR parameters has under unchanged actor \nforesight. Myopic foresight corresponds to a rolling foresight horizon of ten \nyears (Methods provide underlying model assumptions). Historical prices are \nhistorical allowances (EUA) prices on the EEX spot market2,68. The mean average \npercentage error between modelled and historical prices is available in Extended \nData Table 1. The interaction between the foresight horizon and the MSR is shown \nin Extended Data Fig. 1. Prices are nominal until 2023 and real EUR2023 from 2023 \non (Methods \u2018Carbon prices\u2019).\n2015\n2020\n2025\n2030\n2035\n2040\nYear\n0\n150\n50\n200\n250\n100\nCarbon price (\u20ac tCO2\n\u22121)\nPerfect foresight\nMyopic foresight\nInvestors buy\n5%\n20%\nFig. 4 | Impact of external financial investors on carbon prices. Carbon price \ntrajectories assuming perfect foresight, myopic foresight and myopic foresight \nwith external investors buying 5% or 20% of yearly auctioned allowances and \nreselling them once carbon prices reach the theoretical value from the perfect \nforesight path. If external investors had started buying 5% or 20% of auctioned \nallowances from 2018 onwards, they would own, respectively, around 10% or 40% \nof today\u2019s (as of 2022) total number of allowances in circulation. All trajectories \ncorrespond to newest targets from the \u2018Fit for 55\u2019 proposal with an active MSR. \nExact assumptions on the number and timing of allowances bought and sold by \nexternal investors is available in Extended Data Fig. 2.\n\nNature Energy | Volume 9 | June 2024 | 691\u2013702\n696\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nfuture shows that carbon prices could fall again if actors become \nmyopic again (for example, due to a price shock and reduced pol-\nicy credibility). A fall back into myopia and low prices can threaten \nshort-term decarbonization efforts. To prevent such a development, \nadditional policy instruments seem advisable to stabilize expectations \nof agents. As an example, a price floor would limit the drop of carbon \nprices in the short term when long-term policy credibility is temporarily \nreduced and could even keep prices higher without being binding50.\nOverall, we find that observed 2022 and 2023 prices of around \u20ac80 \ntCO2\n\u22121 are in line with EU Climate Law reduction targets and should not \nbe artificially lowered. Compliance actors seem to trust the political \ncommitment and act farsighted\u2014a good sign for the reachability of EU\u2019s \n2030 climate targets of the ETS sectors, if the current energy crisis and \nrelated policy reactions do not undermine this mindset.\nMethods\nThe model LIMES-EU\nAll quantitative results in this work are obtained using the model \nLIMES-EU (Long-term Investment Model for the Electricity Sector), \nversion 2.38. LIMES-EU is a linear optimization modelling framework \nthat simultaneously determines cost-minimizing investment and \ndispatch decisions for generation, storage and transmission technolo-\ngies in the European electricity sector. Although its clear focus is the \nelectricity sector, the energy-intensive industry and district heating \nare also represented through marginal abatement cost curves. Com-\npared with simple emissions trading models with static exogenous cost \nabatement curves, using an energy system model such as LIMES-EU \nallows to assess not only market developments (for example, prices or \nallowances in circulation) but also the investment dynamics and path \ndependencies within the electricity sector.\nLIMES-EU allows to fully simulate the EU ETS including the Market \nStability Reserve (MSR)51. Hence, one can analyse figures such as the \nnumber of allowances in circulation, the intake by the MSR and result-\ning carbon prices. By varying the cap and MSR parameters, one can \nreproduce the state of the EU ETS between different political reforms.\nA comprehensive description of the LIMES-EU model, including \nparameters, equations and assumptions, is provided in the documenta-\ntion available from the model\u2019s website52.\nPast\nLow carbon prices\nLow carbon prices\n2013\u20132017\nPolicy\ncredibility\nFrom past to present\nActors\nextend foresight\nPotentially\nreinforcing cycle\nEnergy prices\nincrease\nActors experience\nliquidity problems\nActors become\nmore myopic\nPotentially\nreinforcing cycle\nExample: energy crisis\nExample: MSR reform\nMSR demonstrates\nmore credible policy\ncommitment\nHigher climate\nambition\nPolitical or\neconomical crisis\nHigh carbon prices\nHigher policy\ncredibility\nRisks for the future\nPresent\n2021\u20132023\nPotential future?\nLower policy\ncredibility\nFig. 5 | Role of actors\u2019 foresight and policy credibility in carbon price \nformation. A distortion from a high or low prices level can enchain a potentially \nreinforcing loop leading to a fall, or rise, in prices, respectively. Policy credibility \nplays a major role in how actors react to a distortion. The theoretical hypothesis \nis complemented by two examples: the introduction of the MSR as an example \nof increased commitment to climate targets and the current energy crisis, as an \nexample of a potential shock. Following studies highlighting the importance \nof climate policy credibility for an acting private sector13,69, we assume actors\u2019 \nforesight depends on policy credibility.\n2015\n2020\n2025\n2030\n2035\n2040\nYear\n0\n150\n50\n200\n250\n100\nCarbon price (\u20ac tCO2\n\u22121)\nPerfect foresight\nReversal to myopia\nHigh policy\ncredibility\nLow policy\ncredibility\nFig. 6 | Risk of falling EU ETS prices due to undermined policy credibility. \nCarbon price trajectories assuming perfect foresight and reversal to myopia (that \nis, actors are first fully myopic, become farsighted around 2020 and then fall \nback fully into myopia until 2025). Both trajectories correspond to newest targets \nfrom the \u2018Fit for 55\u2019 proposal with an active MSR. The carbon price trajectories \nare complemented by the hypothesis that actors\u2019 foresight strongly depends on \npolicy credibility.\n\nNature Energy | Volume 9 | June 2024 | 691\u2013702\n697\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nAll changes to LIMES version 2.38 made for the purposes of this \nstudy are described below.\nA myopic version of LIMES-EU\nRolling horizon as operationalization of myopia. Originally, LIMES-EU \nwas formulated as a perfect foresight model running in five-year steps \nfrom 2010 until 2070. For the purpose of this study, to simulate the \neffect of myopic behaviour of decisionmakers, we extend the model \nwith the option to use rolling time horizons instead of full intertem-\nporal foresight. Mathematically this means that instead of solving \none optimization problem over the whole time period from 2010 until \n2070, we solve multiple (consecutive) optimization problems, covering \nshorter time periods.\nIn our choice to implement a rolling horizon, we follow several \nother publications from our field: the rolling horizon approach (that \nis, short foresight with overlapping time steps) has already been used \nextensively as a way to represent myopia in the context of energy sys-\ntems modelling41,43,44,47. Although principally other approaches would be \npossible (for example, by varying the discount rate), we are not aware of \nany publication in our field representing myopia in a different manner.\nForesight length. All myopic foresight results in this work assume \nten-year horizons with an overlap of five years between the horizons. \nPractically it means, actors have foresight of ten years but can revise \ntheir decisions every five years. As LIMES-EU runs in five-year time \nsteps, one optimization horizon comprises always two time steps (for \nexample, (2020, 2025), covering years 2018\u20132027).\nThe literature provides different estimations on planning horizons \nof manufacturing companies, ranging between three and 12 years6. \nBocklet and Hintermayer6 and Quemin and Trotignon7 show that a \nhorizon of around ten years can best replicate EU ETS developments \n(these analyses were conducted around the time of the MSR reform). \nHence, we also chose a foresight horizon of ten years. As our model runs \nin five-year time steps, ten years is also the shortest foresight horizon \nwe can meaningfully implement (that is, which allows for an overlap) \nin LIMES-EU. Varying the length of the foresight horizon impacts the \nresults but not the general trends: the shorter the foresight, the lower \nthe near-term carbon prices and higher the delays in decarbonization47.\nWhen running in myopic foresight, the model solves consecutively \nseveral individual optimization problems. Still, some variable values \ncomputed in one optimization horizon need to be transferred into the \nnext optimization horizon. It concerns all previous capacity additions \nand decommissioning (needed to correctly compute current capaci-\nties) and emissions and banked certificates (needed for the ETS/MSR \nsimulation). For instance, for the optimization horizon (2020, 2025) \ncapacities will be fixed for 2020 and all time steps before 2020. We \nassume that dispatch decisions can still get revised every time step \n(five years), so for example, for the optimization horizon (2020, 2025), \nemissions and banked certificates values get fixed only for all time steps \nbefore 2020 but not 2020 itself.\nWhat do actors neglect and what do they still consider. In our study, \nwe use rolling horizons as a tool to represent actors\u2019 myopia due to low \ntrust in the long-term stability of the EU ETS. Hence, our main aim is to \ndepict actors that are myopic with regards to the ETS. Our modelling \napproach implies that actors don\u2019t consider any information outside \nof their ten years foresight horizon (that is, the future ETS cap and the \nfuture demand for certificates).\nNonetheless, as ETS actors are mostly large power system or manu-\nfacturing companies and salvage values (\u2018book values\u2019) are traditionally \npart of companies balance sheets, we still assume that they consider the \nfuture value of capacities also beyond the foresight horizon. Therefore, \na salvage value for the capacity stock remaining at the end the optimiza-\ntion horizon is subtracted from the cost function. In the myopic version, \nthe salvage value is considered in each time horizon. This means that \nwhen we run a diagnostic scenario where we turn off the ETS and keep \ntechnology prices constant over time, the results of the myopic mode \nexactly reproduce the results of the perfect foresight mode.\nMSR simulation. The MSR, which is originally implemented iteratively \nas a loop around the main optimization problem51, runs in the myopic \nmodel version around each time horizon.\nSpecific modelling aspects\nCarbon prices. Reported carbon prices (in \u20ac tCO2\n\u22121) represent the mar-\nginal abatement costs in a given year, which are equal to the dual value \n(shadow price) associated with the banking constraint in LIMES-EU. \nTransaction costs are neglected. Reported historic carbon prices are \nnominal, so given in \u20ac of the year in which they occurred. LIMES runs \nin real \u20ac2010, but all reported prices from LIMES until 2023 in this paper \n2020\n200\nWind onshore capacity (GW)\n300\n400\n500\n600\nPerfect foresight\nMyopic foresight\nPerfect foresight\nMyopic foresight\n200\nCoal in electricity mix (TWh yr\u22121)\n0\n400\n600\na\nb\n2025\n2030\n2035\n2040\nYear\n2020\n2025\n2030\n2035\n2040\nYear\nFig. 7 | Delays in decarbonization due to myopia. a, Expansion of wind onshore \ncapacity in the European Union under perfect and myopic foresight. b, Yearly \nelectricity generation in the European Union from black and brown coal under \nperfect and myopic foresight. Trajectories in both a and b correspond to newest \ntargets from the \u2018Fit for 55\u2019 proposal with an active MSR. Note that in both a and b, \nthe 2020 year is fixed to match real 2020 values. Additional data (solar capacity \nexpansion and total electricity mix) on perfect and myopic foresight scenarios \ncan be found in Extended Data Fig. 3.\n\nNature Energy | Volume 9 | June 2024 | 691\u2013702\n698\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nwere converted to nominal prices until 2023, adjusted for inflation using \ninflation rates provided by the Organization for Economic Co-operation \nand Development53. Computed prices after 2023 are in real \u20ac2023.\nExternal investors. To depict external investors in our model, we \nassume that the impact on carbon prices of buying/holding/selling \nEUA futures can be approximated by the assumption, external investors \nbuy/hold/sell physical allowances. As we are interested in long-term \nprice developments, we focus on external investors holding long open \nposition on EUA futures.\nTo model the impact of external investors, we implement a one-step \niteration approach. Hence, we implicitly assume that both compliance \nactors and external investors can\u2019t react the other group\u2019s action.\n(1) In a first instance, a LIMES-EU run with full myopic foresight \nwithout external investors is conducted.\n(2) The resulting carbon price trajectory pprice,CO2(ty) serves as input \nto the optimization problem from the external investors\u2019 \nperspective:\nmax\nvbought,vsold\n\u2211\nty\u2208T\n(vsold(ty)pprice,CO2(ty)\n\u2212vbought(ty)pprice,CO2(ty)) \u00d7 e\u2212i(ty\u2212ty0)\n(1)\ns.t.vbought (ty) \u2264\u03b1pauction (ty)\n(2)\nty\n\u2211\n0\nvsold(ty) \u2264\nty\u22121\n\u2211\n0\nvbought(ty)\n(3)\nvsold (ty) \u2264\u03b3 \u2211\nty\u2208T\nvsold(ty)\n(4)\nEquation (1) is the profit function: external investors want to maxi-\nmize their profit by buying allowances and selling them at a later time \nstep ty. Herein, ty \u2208[2018, \u2026 , 2040] are yearly time steps. T is the set \ncontaining all yearly steps part of the optimization. Further, vbought(ty) \nand vsold(ty) stand for the number of allowances bought and sold in  \ntime step ty. The profit gets discounted by discount rate i. We assume \ni\u2009=\u20095%, same as in the core model assumptions of LIMES-EU. Finally, \npprice,CO2 (ty) corresponds to the carbon price from a myopic run, which \ngrows at a higher rate than the discount rate of 5%.\nEquation (2) sets a limit on the number of allowances external \ninvestors can maximally buy. Herein, \u03b1 is the share of auctioned allow-\nances pauction(ty). We assume pauction to be the final number of allowances \nauctioned, after subtraction of allowances transferred into the MSR. \nIn our work, \u03b1 is varied between 5 and 20%. Equation (3) ensures the \nnumber of allowances sold is below the number of allowances external \ninvestors bought prior to time step ty.\nFinally, equation (4) limits the number of allowances that can be \nsold in a given time step ty, to prevent all of them being sold in a single \nyear. Results assume an \u03b3 of 0.2, meaning allowances need to be sold \nover minimum five years.\n(3) Having solved the optimization problem from the perspective \nof external investors, one can now conduct a new LIMES-EU run with \nfull myopic foresight and additional input on the number of allowances \nblocked by external investors.\npinvestors (t) = vbought (t) \u2212vsold(t)\n(5)\nvtnac (t) \u2212vtnac (t \u22121) = pcap (t) \u2212pinvestors (t) \u2212vemi (t)\n(6)\nHere pinvestors is the absolute number of allowances bought or sold \nby external investors. These influence the level of allowances, as shown \nin equation (7). Here vtnac(t) is the total number of allowances in circula-\ntion (TNAC) at the end of time step t, pcap(t) the total number of allow-\nances auctioned and freely allocated and vemi(t) the total emissions in \ntime step t. Here t \u2208[2010, 2015, \u2026 , 2040] are five-year time steps.\nTo capture the unpredictability of external investors on the price \nformation, we assume compliance actors can\u2019t see the realization of \npinvestors (t) before time step t. Hence, even though they have a foresight \nof ten years regarding all other model inputs, they only have a foresight \nof one LIMES-EU time step (five years) when it comes to pinvestors (t).\nIt is important to note that the way our approach is implemented, \nexternal investors behave as farsighted actors and have incentives to \nenter the market, only if compliance actors are myopic (carbon prices \ninitially lower than under the perfect foresight scenario). Hence, all \nresults showing the impact of external investors presume myopic \nforesight from compliance actors.\nAs we conduct only one iteration, we implicitly assume that exter-\nnal investors plan all their future behaviour only once and base it on \nmyopic carbon prices. In the real world, there is a constant feedback \nbetween prices and investors\u2019 buying/selling strategy. Hence, our \nmethodology does not aim to provide realistic predictions regarding \npossible behaviour of external investors. It is, however, suitable to \nshow the order of magnitude of the increase in carbon prices, assuming \nexternal investors block a certain number of certificates.\nFuture. In the \u2018Reversal to myopia\u2019 scenario from Fig. 5b, similar to \nthe full myopic version, several consecutive optimization problems \nwith ten years foresight horizons are solved, with the exception that \nthe horizon [2020, 2025] gets replaced by [2020,\u2026, 2070] to simulate \nperfect foresight in time step 2020. Afterwards, from time step 2025 \non, actors have again only myopic foresight.\nMACC curves representing industry and heating sectors. As \ndescribed in the LIMES-EU Documentation52, the industry and heat-\ning sectors are not modelled explicitly in LIMES-EU, but the cost of \nemissions abatement is approximated by marginal abatement cost \ncurves (MACCs). Originally, as they have been designed for runs start-\ning in 2020, both MACCs assumed a minimum cost of \u20ac8 tCO2\n\u22121, being a \nwell-suited assumption for benchmark modelling, in which modelled \ncarbon prices always exceed \u20ac8 tCO2\n\u22121 for relevant ETS scenarios. As in \nthis work certain counterfactual scenarios yield prices below \u20ac8 tCO2\n\u22121\n, \nwe extrapolate the MACC curves to also cover the price regime of \u20ac0\u20138 \ntCO2\n\u22121 by analysing the change in industry and heating emissions upon \nimplementation of the ETS. We thus estimate two additional emissions \nsteps of 45\u2009Mt\u2009CO2 in industry and 15\u2009Mt\u2009CO2 in heating that would be \nemitted additionally compared to historic industry/heating emissions \nwhen ETS prices remain below \u20ac5 tCO2\n\u22121 and again when they remain \nbelow \u20ac3 tCO2\n\u22121.\nScenarios\nModelling assumptions regarding calibration, policy targets and \ntechnology costs. The key assumptions behind our study\u2019s main \nscenario types are summarized in Extended Data Table 3. In Fig. 3, we \nalign scenarios with historical conditions as closely as possible, adjust-\ning variables such as ETS modelling start year and technology cost \nassumptions. Due to the five-year time steps in our model, complete \nhistorical replication and path dependency coverage may be limited \n(for example, \u2018Fit for 55\u2019 scenario starts in 2018).\nFor Figs. 4, 6 and 7, we exclusively use the \u2018Fit for 55\u2019 scenario, \nrepresenting the current EU ETS state. This simplification serves the \npurpose of preventing information overload, aligning with the figures\u2019 \nprimary objective. In Fig. 6, we extrapolate our results to 2015.\n\u2018Fit for 55\u2019 Commission\u2019s proposal vs final agreement. Extended \nData Table 4 summarizes the relevant parameters used in this study \ndefining the emissions cap and MSR functionality for the ETS state \nbetween different reforms.\n\nNature Energy | Volume 9 | June 2024 | 691\u2013702\n699\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nAll results in this study related to ETS targets from the \u2018Fit for 55\u2019 \npackage assume parameters from the Commission\u2019s proposal pub-\nlished in July 202154. As this study takes into account real ETS prices \nuntil December 2022, it is plausible to assume that until then market \nactors were basing their decisions on the Commission\u2019s proposal, not \nbeing aware yet of the upcoming changes in the final negotiations.\nFor completeness reasons, we provide a comparison of modelled \ncarbon prices according to the emissions cap from the Commission\u2019s \nproposal (used in this study) and according to the emissions cap from \nthe final ETS \u2018Fit for 55\u2019 agreement39,55 in Extended Data Fig. 4. The \nemissions cap corresponding to the final agreement can be found in \nExtended Data Table 2.\nModel validation\nGeneral modelling choices, for example, the clustering approach and \nthe representative days choice, are described in the LIMES-EU model \ndocumentation. Here we present additional validation points for the \nscenarios presented in this study. First, we show that our model can \napproximate historical developments in 2015 and 2020. Then, we \nprovide references demonstrating that our future estimates for the \nEU ETS align with other literature.\nReproducing historical developments in time step 2015. The \ncapacity spin up of LIMES-EU is fixed so that it matches the 2015 his-\ntorical mix of installed generation capacities in EU ETS countries. \nExtended Data Fig. 5 illustrates that based on this standing capacity, \nthe model-calculated dispatch then reasonable matches the historic \npower generation dispatch in EU ETS countries. The total modelled \nemissions from electricity generation in the year 2015 for EU ETS coun-\ntries covered by LIMES-EU amount to 981\u2009Mt\u2009CO2, closely aligning with \nthe historical emissions of 967\u2009Mt\u2009CO2 reported by Mantsos et al.56 \nBecause emissions from industry, heating and aviation are also cali-\nbrated to match their historical 2015 levels (as described in LIMES-EU \ndocumentation52), this calibration ensures that our model generates \nmeaningful values for total emissions in the 2015 time step. Also, the \nmodel-endogenous investments in 2015 lead to standing capacities in \n2020 that match historic wind and solar capacities in 2020. To this aim, \nwe additionally assume subsidies for electricity generated from solar \nor wind sources (\u20ac0.04\u2009kWh\u22121 for solar and \u20ac0.015\u2009kWh\u22121 for wind) to \nrepresent the various renewable subsidies that were in place in most \nEU member states. Our model, however, underestimates the capacity \nadditions of offshore wind until 2020, which took place mostly in the \nUnited Kingdom.\nReproducing historical developments in time step 2020. To validate \nthe 2020 model results, we first fix capacity spin up so that our model \nmatches the installed generation capacities for both 2015 and 2020 in \nEU ETS countries. In Extended Data Fig. 6, we show that this calibration \nenables our model to approximate EU-wide dispatch and total emis-\nsions from the electricity sector in 2020. It\u2019s important to note that our \nmodel operates in five-year steps, with time step 2020 representing the \nactual years 2018\u20132022. However, due to the exceptional circumstances \nof the COVID-19 pandemic, the year 2020 deviates from the typical \ntrends of 2018\u20132022. Hence, to validate time step 2020, we provide \nreal values for the years 2019 and 2020.\nWith respect to electricity dispatch, our model estimates lower \ngeneration from biomass compared with the International Energy \nAgency (IEA) historical data. This discrepancy may be attributed to \nseveral factors, including our reliance on the European Network of \nTransmission System Operators for Electricity (ENTSO-E) dataset for \ntotal capacities, whereas using the IEA dataset for generation values \n(as ENTSO-E lacks a Statistical Factsheet covering generation for the \nyears 2019 and 2020). Differences in values from different sources can \noften be substantial. Regarding biomass, variations may be due to, for \nexample, the way biomass co-firing in coal power plants is accounted \nfor. Nevertheless, despite minor deviations in our 2020 electricity \ndispatch from historical data, our model still provides a meaningful \nestimate of emissions. This aspect is critical for validating EU ETS mod-\nels, as it directly impacts CO2 prices, the total number of allowances in \ncirculation and the functioning of the MSR.\nEstimating future developments. While validating future projections \nis inherently impossible, we observe that LIMES-EU generally aligns \nwith findings in the literature and does not produce results that are \nfar outliers compared with other models. Osorio et al. discuss that \nLIMES-EU\u2019s estimates of MSR cancellations are consistent with other \nstudies51. Furthermore, a recent model comparison study led by Henke \net al. revealed that LIMES-EU\u2019s projections for various EU electricity \nsector variables from 2020 to 2050, such as final energy demand and \nthe share of renewable energy sources in electricity generation, are in \nline with the range provided by ten other energy systems and integrated \nassessment models57. In another model comparison study assessing \nEUA prices until 2030, LIMES-EU\u2019s estimate of \u20ac140 tCO2\n\u22121 falls within \nthe range of \u20ac80 to \u20ac160 tCO2\n\u22121 produced by six different models58.\nMethodological contribution\nWhereas the primary focus of this work lies in providing insights for \nthe ongoing debates surrounding the EU ETS, we also make a notable \nmethodological contribution. There have been other studies using \nEU ETS models that explicitly simulate the electricity sector33,59,60, \nand there have been energy systems analyses using myopic energy \nsystem models41,43,44,47. Also Nerini et al.47 pioneered the idea to compare \nmyopic and perfect foresight modes of a capacity expansion model to \nformulate more robust policies. Our study extends their approach and \nemploys both types of foresight to evaluate ex post a concrete policy \nreform to test whether the change in the observable variable\u2014in our \ncase, the EU ETS price\u2014can better be reproduced in the myopic or \nperfect foresight mode.\nData availability\nData for core model assumptions (investment costs, fuel costs and \nso on) are provided in the LIMES-EU documentation (Methods). The \ndataset containing all results displayed in this paper is publicly avail-\nable via Zenodo at https://doi.org/10.5281/zenodo.10363561 (ref. 61).\nCode availability\nThe LIMES-EU model code is available upon request from the authors. \nMoreover, a process has been started to make the model available \nunder an open-source license. When this process will be completed, \nthe code will be available for download from the PIK webpage at http://\npik-potsdam.de/limes.\nReferences\n1.\t\nEU Emissions Trading System (EU ETS) (European Commission, \n2021); https://ec.europa.eu/clima/eu-action/eu-emissions-trading- \nsystem-eu-ets_en\n2.\t\nICAP Allowance Price Explorer (ICAP, 2022); https://\nicapcarbonaction.com/en/ets-prices\n3.\t\nGerlagh, R., Heijmans, R. J. R. K. & Rosendahl, K. E. Shifting \nconcerns for the EU ETS: are carbon prices becoming too high? \nEnviron. Res. Lett. 17, 054018 (2022).\n4.\t\nWillner, M. Consulting the chrystal ball: firms\u2019 foresight and a \ncap-and-trade scheme with endogenous supply adjustments. \nPreprint at SSRN https://doi.org/10.2139/ssrn.3231927  \n(2018).\n5.\t\nPerino, G. et al. EU ETS Stability Mechanism Needs New Design. \n(CEC Dauphine Policy Brief, 2021).\n6.\t\nBocklet, J. & Hintermayer, M. How Does the EU ETS Reform Impact \nAllowance Prices? The Role of Myopia, Hedging Requirements and \nthe Hotelling Rule EWI Working Paper Series 2020/01 32 (2020).\n\nNature Energy | Volume 9 | June 2024 | 691\u2013702\n700\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\n7.\t\nQuemin, S. & Trotignon, R. Emissions trading with rolling horizons. \nJ. Econ. Dyn. Control 125, 104099 (2021).\n8.\t\nQuemin, S. & Pahle, M. Financials threaten to undermine the \nfunctioning of emissions markets. Nat. Clim. Change 13, 22\u201331 \n(2023).\n9.\t\nEmission Allowances and Associated Derivatives (European \nSecurities and Markets Authority (ESMA, 2022); https://www.\nesma.europa.eu/sites/default/files/library/esma70-445-38_final_ \nreport_on_emission_allowances_and_associated_derivatives.pdf\n10.\t Jeszke, R. & Lizak, S. Reflections on the mechanisms to protect \nagainst formation of price bubble in the EU ETS market. Environ. \nProt. Nat. Resour. 32, 8\u201317 (2021).\n11.\t\nKnopf, B. et al. Shifting paradigms in carbon pricing. \nIntereconomics 53, 135\u2013140 (2018).\n12.\t Statement by Executive Vice-President Timmermans on delivering  \nthe European Green Deal (European Commission, 2021); https://ec. \neuropa.eu/commission/presscorner/detail/en/SPEECH_21_3722\n13.\t Victor, D. G., Lumkowsky, M. & Dannenberg, A. Determining the \ncredibility of commitments in international climate policy. Nat. \nClim. Change 12, 793\u2013800 (2022).\n14.\t Dolphin, G., Pahle, M., Burtraw, D. & Kosch, M. A net-zero target \ncompels a backward induction approach to climate policy. Nat. \nClim. Change https://doi.org/10.1038/s41558-023-01798-y (2023).\n15.\t Kang, S. B. & L\u00e9tourneau, P. Investors\u2019 reaction to the government \ncredibility problem: a real option analysis of emission permit \npolicy risk. Energy Econ. 54, 96\u2013107 (2016).\n16.\t Ascher, W. Long-term strategy for sustainable development: \nstrategies to promote far-sighted action. Sustain Sci. 1, 15\u201322 \n(2006).\n17.\t\nDevelopment of EU ETS (2005\u20132020) (European Commission, 2023);  \nhttps://ec.europa.eu/clima/eu-action/eu-emissions-trading- \nsystem-eu-ets/development-eu-ets-2005-2020_en\n18.\t Flachsland, C. et al. How to avoid history repeating itself: the case \nfor an EU Emissions Trading System (EU ETS) price floor revisited. \nClim. Policy 20, 133\u2013142 (2019).\n19.\t Edenhofer, O. et al. in 12 Years of European Emissions Trading \nin Germany. Stocktaking and Perspectives for Effective Climate \nProtection 207\u2013232 (Metropolis, 2019).\n20.\t Rubin, J. D. A model of intertemporal emission trading, banking, \nand borrowing. J. Environ. Econ. Manage. 31, 269\u2013286  \n(1996).\n21.\t Paris Agreement United Nations Treaty Collection, Chapter XXVII \n7. d (UNFCCC, 2015).\n22.\t European Commission. Decision (EU) 2015/1814 concerning the \nestablishment and operation of a market stability reserve for the \nUnion greenhouse gas emission trading scheme and amending \nDirective 2003/87/EC. Off. J. Eur. Union L264, 1\u20135 (2015).\n23.\t Council of the European Union. Council Decision (EU) 2016/1841 \nof 5 October 2016 on the conclusion, on behalf of the European \nUnion, of the Paris Agreement adopted under the United Nations \nFramework Convention on Climate Change. Off. J. Eur. Union \nL282, 1\u20133 (2016).\n24.\t Proposal for a Directive of the European Parliament and of the \nCouncil Amending Directive 2003/87/EC to Enhance Cost-effective \nEmission Reductions and Low-carbon Investments (Council of the \nEuropean Union, 2017); https://data.consilium.europa.eu/doc/\ndocument/ST-6841-2017-INIT/en/pdf\n25.\t EU Emissions Trading System: Landmark Agreement Between \nParliament and Council (European Commission, 2017); https://ec. \neuropa.eu/commission/presscorner/detail/en/STATEMENT_17_4501\n26.\t Reform of the EU Emissions Trading System\u2014Council Endorses \nDeal with European Parliament (Council of the European Union, \n2017); https://www.consilium.europa.eu/en/press/press-releases/ \n2017/11/22/reform-of-the-eu-emissions-trading-system-council- \nendorses-deal-with-european-parliament/\n27.\t European Parliament. Directive (EU) 2018/410 of the European \nParliament and of the Council of 14 March 2018 amending \nDirective 2003/87/EC to enhance cost-effective emission \nreductions and low-carbon investments, and Decision (EU) \n2015/1814. Off. J. Eur. Union L76, 3\u201327 (2018).\n28.\t Abnett, K. & Strzelecki, M. Poland calls on EU to remove \n\u2018speculators\u2019 from its carbon market. Reuters (15 Feb. 2022).\n29.\t Morawiecki, M. Polish PM: the green transition cannot come at the \ncost of European security. Financial Times (4 Jul. 2022).\n30.\t Ampudia, M., Bua, G., Kapp, D. & Salakhova, D. The role of \nspeculation during the recent increase in EU emissions allowance \nprices (European Central Bank, 2022); https://www.ecb.europa.eu/ \npub/economic-bulletin/focus/2022/html/ecb.ebbox202203_ \n06~ca1e9ea13e.en.html\n31.\t Burtraw, D. Low allowance prices in the EU emissions trading \nsystem: new research on an evolving program (Resources for the \nFuture, 2015); https://www.resources.org/common-resources/\nlow-allowance-prices-in-the-eu-emissions-trading-system-new- \nresearch-on-an-evolving-program/\n32.\t Carl\u00e9n, B., Dahlqvist, A., Mandell, S. & Marklund, P. EU ETS \nemissions under the cancellation mechanism\u2014effects of national \nmeasures. Energy Policy 129, 816\u2013825 (2019).\n33.\t Anke, C.-P. & M\u00f6st, D. The expansion of RES and the EU ETS\u2014\nvaluable addition or conflicting instruments? Energy Policy 150, \n112125 (2021).\n34.\t Friedrich, M., Fries, S., Pahle, M. & Edenhofer, O. Understanding the \nexplosive trend in EU ETS prices\u2013fundamentals or speculation?  \nPreprint at https://doi.org/10.48550/arXiv.1906.10572 (2020).\n35.\t Pietzcker, R. C., Osorio, S. & Rodrigues, R. Tightening EU ETS \ntargets in line with the European Green Deal: impacts on the \ndecarbonization of the EU power sector. Appl. Energy 293, 116914 \n(2021).\n36.\t EU Natural Gas TTF (Trading Economics, 2023); https://\ntradingeconomics.com/commodity/eu-natural-gas\n37.\t Payne, J. & Zhdannikov, D. Analysis: Europe energy crisis may \ndeepen with looming liquidity crunch. Reuters (21 Sep. 2022).\n38.\t EU\u2019s von der Leyen rebuffs Polish call to suspend carbon market. \nReuters (30 Aug. 2022).\n39.\t \u2018Fit for 55\u2019: Council and Parliament Reach Provisional Deal on EU \nEmissions Trading System and the Social Climate Fund (Council \nof the European Union, 2022); https://www.consilium.europa.eu/\nen/press/press-releases/2022/12/18/fit-for-55-council-and-parl\niament-reach-provisional-deal-on-eu-emissions-trading-system- \nand-the-social-climate-fund/\n40.\t Patashnik, E. M. Reforms at Risk: What Happens After Major Policy \nChanges Are Enacted (Princeton Univ. Press, 2008).\n41.\t Babiker, M., Gurgel, A., Paltsev, S. & Reilly, J. Forward-looking \nversus recursive-dynamic modeling in climate policy analysis: a \ncomparison. Econ. Model. 26, 1341\u20131354 (2009).\n42.\t Keppo, I. & Strubegger, M. Short term decisions for long term \nproblems\u2014the effect of foresight on model based energy systems \nanalysis. Energy 35, 2033\u20132042 (2010).\n43.\t Poncelet, K., Delarue, E., Six, D. & D\u2019Haeseleer, W. Myopic \noptimization models for simulation of investment decisions in the \nelectric power sector. International Conference on the European \nEnergy Market 1\u20139 (IEEE, 2016).\n44.\t Heuberger, C. F., Staffell, I., Shah, N. & Mac Dowell, N. Impact of \nmyopic decision-making and disruptive events in power systems \nplanning. Nat. Energy 3, 634\u2013640 (2018).\n45.\t Hedenus, F., Azar, C. & Lindgren, K. Induced technological \nchange in limited foresight optimization model. Energy J. 27, \n109\u2013122 (2006).\n46.\t Gerbaulet, C., von Hirschhausen, C., Kemfert, C., Lorenz, C. &  \nOei, P. Y. European electricity sector decarbonization under \ndifferent levels of foresight. Renew. Energy 141, 973\u2013987 (2019).\n\nNature Energy | Volume 9 | June 2024 | 691\u2013702\n701\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\n47.\t Fuso Nerini, F., Keppo, I. & Strachan, N. Myopic decision making \nin energy system decarbonisation pathways. A UK case study. \nEnergy Strategy Rev. 17, 19\u201326 (2017).\n48.\t Luderer, G. et al. Economic mitigation challenges: how further \ndelay closes the door for achieving climate targets. Environ. Res. \nLett. 8, 34033 (2013).\n49.\t Pahle, M. et al. Safeguarding the energy transition against  \npolitical backlash to carbon markets. Nat. Energy 7, 290\u2013296 \n(2022).\n50.\t Salant, S., Shobe, W. & Uler, N. The effects of seemingly \nnonbinding price floors: an experimental analysis. Eur. Econ. Rev. \n159, 104583 (2023).\n51.\t Osorio, S., Tietjen, O., Pahle, M., Pietzcker, R. C. & Edenhofer, O. \nReviewing the market stability reserve in light of more ambitious \nEU ETS emission targets. Energy Policy 158, 112530 (2021).\n52.\t Osorio, S., Pietzcker, R., Levesque, A., & Sitarz, J. Documentation \nof LIMES-EU\u2014A Long-Term Electricity System Model for Europe \n(Potsdam Institute for Climate Impact Research, 2023); https://\nwww.pik-potsdam.de/en/institute/departments/transformation- \npathways/models/limes/limes-documentation-v2-38-march-2023\n53.\t Inflation (CPI) (OECD, 2023); http://data.oecd.org/price/inflation- \ncpi.htm\n54.\t Proposal for a Directive of the European Parliament and of the \nCouncil Amending Directive 2003/87/EC Establishing a System \nfor Greenhouse Gas Emission Allowance Trading within the Union, \nDecision (EU) 2015/1814 Concerning the Establishment and \nOperation of a Market Stability Research for the Union Greenhouse \nGas Emission Trading Scheme and Regulation (European \nCommission, 2021); https://eur-lex.europa.eu/legal-content/EN/ \nTXT/?uri=CELEX%3A52021PC0551\n55.\t Proposal for a Directive of the European Parliament and of the \nCouncil Amending Directive 2003/87/C Establishing a System for \nGreenhouse Gas Emission Allowance Trading within the Union, \nDecision (EU) 2015/1814 Concerning the Establishment and \nOperation of a Market Stability Reserve for the Union Greenhouse \nGas Emission Trading Scheine and Regulation (EU) 2015/757 (2021/ \n0211 (COD)) (Council of the European Union, 2023); https://data.\nconsilium.europa.eu/doc/document/ST-6210-2023-INIT/en/pdf\n56.\t Mantzos, L. et al. The JRC Integrated Database of the European \nEnergy System (JRC Publications Repository, 2018); https://\npublications.jrc.ec.europa.eu/repository/handle/JRC112474\n57.\t Henke, H. et al. The mutual benefits of comparing energy system \nmodels and integrated assessment models. Open Res Eur. 3, 69 \n(2023).\n58.\t Pahle, M., Sitarz, J., Osorio, S. & G\u00f6rlach, B. The EU-ETS price \nthrough 2030 and beyond: a closer look at drivers, models and \nassumptions. (Kopernikus-Projekt Ariadne, 2022); https://www.\necologic.eu/sites/default/files/publication/2023/30003-Ariadne- \nDocumentation_ETSWorkshopBruessel_December2022.pdf\n59.\t Bonenti, F., Oggioni, G., Allevi, E. & Marangoni, G. Evaluating the \nEU ETS impacts on profits, investments and prices of the Italian \nelectricity market. Energy Policy 59, 242\u2013256 (2013).\n60.\t Bruninx, K., Ovaere, M. & Delarue, E. The long-term impact of \nthe market stability reserve on the EU emission trading system. \nEnergy Econ. 89, 104746 (2020).\n61.\t Sitarz, J. LIMES-EU EU ETS modeling results under myopic \nand perfect foresight. Zenodo https://doi.org/10.5281/\nzenodo.10363562 (2023).\n62.\t The European Green Deal (European Commission, 2019); https://\neur-lex.europa.eu/resource.html?uri= \ncellar:b828d165-1c22-11ea-8c1f-01aa75ed71a1.0002.02/\nDOC_1&format=PDF\n63.\t Proposal for a Regulation of the European Parliament and of \nthe Council Establishing the Framework for Achieving Climate \nNeutrality and Amending Regulation (EU) 2018/1999 (European \nClimate Law) (European Commission, 2020); https://eur-lex.\neuropa.eu/legal-content/EN/TXT/PDF/?uri=CELEX:52020PC0080\n&from=EN\n64.\t Amended proposal for a regulation of the European Parliament and \nof the Council on establishing the framework for achieving climate \nneutrality and amending Regulation (EU) 2018/1999 (European \nClimate Law) (European Commission, 2020); https://eur-lex.\neuropa.eu/legal-content/EN/TXT/?uri= \nCELEX%3A52020PC0563\n65.\t Council agrees on full general approach on European climate law \nproposal (Council of the European Union, 2020); https://www.\nconsilium.europa.eu/en/press/press-releases/2020/12/17/council- \nagrees-on-full-general-approach-on-european-climate- \nlaw-proposal/\n66.\t European Commission. Regulation (EU) 2021/1119 of the \nEuropean Parliament and of the Council of 30 June 2021 \nestablishing the framework for achieving climate neutrality \nand amending Regulations (EC) No 401/2009 and (EU) \n2018/1999 (\u2018European Climate Law\u2019). Off. J. Eur. Union https://doi.\norg/10.5040/9781782258674 (2021).\n67.\t Proposal for a Decision of the European Parliament and of the \nCouncil Amending Decision (EU) 2015/1814 as Regards the Amount \nof Allowances to be Placed in the Market Stability Reserve for \nthe Union Greenhouse Gas Emission Trading Scheme until 2030 \n(European Commission, 2021); https://eur-lex.europa.eu/ \nlegal-content/EN/TXT/?uri=CELEX%3A52021PC0571\n68.\t EEX EUA Spot (EEX, 2023); https://www.eex.com/de/marktdaten/\numweltprodukte/spotmarkt (2023).\n69.\t Dolphin, G., Pahle, M., Burtraw, D. & Kosch, M. A net-zero target \ncompels a backward induction approach to climate policy. Nat. \nClim. Change 13, 1033\u20131041 (2023).\n70.\t ENTSO-E. Statistical Factsheet 2015 (ENTSO-E, 2015); https://\neepublicdownloads.entsoe.eu/clean-documents/Publications/\nStatistics/Factsheet/entsoe_sfs2015_web.pdf\n71.\t ENTSO-E. Installed Capacity per Production Type \nhttps://transparency.entsoe.eu/generation/r2/\ninstalledGenerationCapacityAggregation/show (2020).\n72.\t International Energy Agency. World Energy Balances Database \nhttps://www.iea.org/data-and-statistics/data-product/world- \nenergy-balances (2022).\n73.\t Hamner, B. et al. R: mean absolute percent error. Evaluation \nMetrics for Machine Learning (R, 2018); https://search.r-project.\norg/CRAN/refmans/Metrics/html/mape.html\n74.\t Nahmmacher, P., Schmid, E. & Knopf, B. Documentation of \nLIMES-EU: a long-term electricity system model for Europe \n(Potsdam Institute for Climate Impact Research, 2014); https://\nwww.pik-potsdam.de/en/institute/departments/transformation- \npathways/models/limes/DocumentationLIMESEU_2014.pdf\nAcknowledgements\nWe gratefully acknowledge funding from the Kopernikus-Ariadne \nproject (FKZ 03SFK5A) by the German Federal Ministry of Education \nand Research, funding from the European Union\u2019s Horizon Europe \nresearch and innovation programme under grant agreement number \n101056891 (CAPABLE) and the European Union\u2019s Horizon 2020 \nresearch and innovation programme under grant agreement numbers \n101022622 (ECEMF) and 101003815 (CAMPAIGNers).\nAuthor contributions\nR.P. and G.L. suggested the research question. J.S. and R.P. jointly \nconceived and designed the study in consultation with S.O. and \nG.L. J.S. extended the model, conducted scenario runs and created \nvisualizations. J.S., R.P., M.P., G.L. and S.O. interpreted the results. \n J.S. wrote the manuscript with contributions from R.P., M.P., S.O.  \nand G.L.\n\nNature Energy | Volume 9 | June 2024 | 691\u2013702\n702\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nFunding\nOpen access funding provided by Potsdam-Institut f\u00fcr \nKlimafolgenforschung (PIK) e.V.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at https://doi.org/10.1038/\ns41560-024-01505-x.\nCorrespondence and requests for materials should be addressed to \nJoanna Sitarz or Robert Pietzcker.\nPeer review information Nature Energy thanks Arega Getaneh Abate \nand the other, anonymous, reviewer(s) for their contribution to the \npeer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard  \nto jurisdictional claims in published maps and institutional  \naffiliations.\nOpen Access This article is licensed under a Creative  \nCommons Attribution 4.0 International License, which permits  \nuse, sharing, adaptation, distribution and reproduction in any \nmedium or format, as long as you give appropriate credit to the \noriginal author(s) and the source, provide a link to the Creative \nCommons licence, and indicate if changes were made. The  \nimages or other third party material in this article are included in the \narticle\u2019s Creative Commons licence, unless indicated otherwise in a \ncredit line to the material. If material is not included in the article\u2019s \nCreative Commons licence and your intended use is not permitted \nby statutory regulation or exceeds the permitted use, you will need \nto obtain permission directly from the copyright holder. To  \nview a copy of this licence, visit http://creativecommons.org/\nlicenses/by/4.0/.\n\u00a9 The Author(s) 2024\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nExtended Data Fig. 1 | Interactions between foresight horizon and MSR. a, \nTotal number of allowances in circulation (TNAC), theoretical cap (allowances to \nbe freely allocated and auctioned before accounting for MSR intake or outtake), \ntotal emissions, allowances taken in by the MSR for both perfect and myopic \nforesight. b, Cumulative emissions over transformation period for perfect \nand myopic foresight. The difference between cumulative theoretical cap and \ncumulative emissions corresponds to total number of allowances cancelled by \nthe MSR. All results in this figure complement Fig. 3 and correspond to runs with \nnewest targets from the \u2018Fit for 55\u2019 proposal with an active MSR.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nExtended Data Fig. 2 | Optimization problem from external financial \ninvestor\u2019s perspective. Investors optimize their profit till 2040 by buying up \nto 5% or 20% of yearly\u2019s auctioned allowances and reselling them later. Obtained \nbuying and selling strategy corresponds to assumption on external investors \nfrom Fig. 4. Before entered to the LIMES-EU model, all values are transformed to \n5-year time steps. Cap trajectory corresponds to allowances auctioned, assuming \nan MSR intake deducted from cap in years 2019\u20132023, as seen in Extended Data \nFig. 1 (myopic foresight).\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nExtended Data Fig. 3 | Delays in decarbonization due to myopia. All results in this figure complement Fig. 7 and correspond to runs with targets from the \u2018Fit for \n55\u2019 proposal with an active MSR. a, Expansion of solar capacity in the EU under perfect and myopic foresight. b, Total electricity mix. P: perfect foresight, M: myopic \nforesight.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nExtended Data Fig. 4 | \u2018Fit for 55\u2019 Commission proposal vs. Final agreement. Carbon prices corresponding to targets from the Commission\u2019s proposal (used \nthroughout the whole study) and carbon prices corresponding to targets from the final agreement after trilogues. All scenarios include the MSR.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nExtended Data Fig. 5 | Comparison of 2015 model results with historical \ndata. Scenario: myopic foresight, EU ETS pre-MSR reform. a, Emissions from \nelectricity generation in year 2015. Real emissions from the Joint Research \nCenter (JRC) Dataset IDEES56. b, Electricity dispatch in 2015. Real dispatch from \nENTSO-E Power Statistics70. c, Planned capacities for year 2020. In myopic \nmode, the model takes this investment decision in time step 2015, hence \nthe 2020 generation capacities serve to validate decisions in time step 2015. \nReal capacities from ENTSO-E Transparency Platform71. All results for EU ETS \ncountries (EU28 and Norway).\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nExtended Data Fig. 6 | Validation of historical time step 2020. Scenario: \nperfect foresight, EU ETS post-MSR reform. a, Emissions from electricity \ngeneration in year 2020 and 2019. As we are not aware of a data source providing \n2019 and 2020 emissions for the electricity sector, we estimate real electricity \nsector emissions by taking EEA ETS emissions data for \u2018combustion of fuels\u2019 and \nassuming that electricity generation accounted for 79% of this (in 2015, emissions \nfrom electricity generation reported by the JRC constituted 79% of emissions \nfrom \u2018combustion of fuels\u2019 reported by the EEA). b, Electricity dispatch in 2020. \nReal dispatch from IEA dataset72,73. All results for EU ETS countries (EU28 and \nNorway).\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nExtended Data Table 1 | Mean absolute percentage error (MAPE) of modeled CO2 prices assuming myopic or perfect \nforesight\nThis table supplements Fig.3 by providing values of the MAPE error between modeled CO2 prices and real historical EUA prices. Highlighted in bold are the runs with the lower MAPE, hence \nlower error compared to real historical EUA prices. MAPE is calculated as the average absolute percent difference between two numeric vectors73.\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nExtended Data Table 2 | Differences in EU ETS caps and MSR parameters between \u2018Fit for 55\u2019 initial proposal and final \nagreement\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nExtended Data Table 3 | Modeling assumptions across different scenarios\n\nNature Energy\nArticle\nhttps://doi.org/10.1038/s41560-024-01505-x\nExtended Data Table 4 | Parameters used in this study representing the state of the EU ETS between different reforms\n\n\n    Scientific Research Findings:", "answer": "We find that the price change following the EU ETS/MSR reform in 2018 is not mainly due to the change of the emission cap, as widely assumed. The full price increase can only be explained if the reform also made actors more farsighted. Before the reform, observed prices aligned reasonably well with simulation results for short\u2011sighted actors. After the reform, observed prices increased strongly. This strong increase diverges from what would have been only a small price increase from the change of cap if we assume actors were short\u2011sighted. However, the strong price increases can be reproduced by assuming actors are farsighted. We infer that the sizable political capital invested into the reform of the EU ETS strengthened policy credibility and thereby induced actors to account for future scarcities. The tighter EU ETS cap proposed in 2021 in the Fit\u2011for\u201155 package further increased the effect of foresight: assuming farsighted actors, carbon prices around \u20ac80\u00a0per\u00a0tCO2 emerge from our model for 2022\u20132023, consistent with actual levels observed over this period. If actors may revert to short\u2011sightedness again due to lack of policy credibility, we expect that prices will fall by around two thirds.", "id": 9}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 9 | August 2024 | 1032\u20131039\n1032\nnature energy\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01531-9\nRapidly declining costs of truck batteries \nand fuel cells enable large-scale road freight \nelectrification\nSteffen Link\u2009\n\u200a\u20091,2\u2009\n, Annegret Stephan\u2009\n\u200a\u20091, Daniel Speth\u2009\n\u200a\u20091 & Patrick Pl\u00f6tz\u2009\n\u200a\u20091\nLow-carbon road freight transport is pivotal in mitigating global warming. \nNonetheless, electrifying heavy-duty vehicles poses a tremendous \nchallenge due to high technical requirements and cost competitiveness. \nData on future truck costs are scarce and uncertain, complicating \nassessments of the future role of zero-emission truck (ZET) technologies. \nHere we derive most likely cost developments for price setting ZET \ncomponents by meta forecasting from more than 200 original sources. \nWe find that costs are primed to decline much faster than expected, with \nsignificant differences between scientific and near-market estimates. \nSpecifically, battery system costs could drop by 64% to 75% and fall below \n\u20ac150\u2009kWh\u22121 by no later than 2035, whereas fuel cell system costs may exhibit \neven higher cost reductions but are unlikely to reach \u20ac100\u2009kW\u22121 before the \nearly 2040s. This fast cost decline supports an optimistic view on the ZET \nmarket diffusion and has substantial implications for future energy and \ntransport systems.\nThe fast electrification of heavy road freight transport is pivotal in \nlimiting global warming in line with the Paris Climate Agreement1\u20133. \nThis follows since heavy-duty vehicles (HDVs) contribute a noteworthy \nproportion of annual greenhouse gas emissions despite a low share \nin the vehicle stock4. Whereas the European Union has agreed on \nambitious tail-pipe emissions reduction targets for newly sold HDVs \nof \u221243% by 2030, \u221265% by 2035 and \u221290% by 2040 (compared with \n2019/2020)5, California has effectively imposed the phase out of con-\nventional combustion trucks by 20366, with other US states expected \nto follow. Similarly, China is anticipated to tighten its tail-pipe emis-\nsions reduction targets soon to comply with its near-zero emissions \ntarget by 20607,8. These ambitions require the fast deployment of \nzero-emission trucks (ZETs), where demand from and, hence, afford-\nability for operators are key. However, high acquisition costs are cur-\nrently hampering fast ZET market diffusion9\u201313. This culminates in \nan active and cross-national debate between industry, politics and \nacademia about different measures and technological pathways of \nhow to decarbonize HDVs1,14\u201319, particularly about the respective roles \nof battery-electric trucks (BETs) and fuel cell trucks (FCETs) in future \nZET fleets.\nAlthough many studies have explored cost-reduction potentials \nusing qualitative or quantitative methods20,21, such as literature-based \nprojections, expert elicitation, detailed cost breakdowns or learning \nand experience curves, results are limited to the respective application \ncategory and system configuration22,23. For example, studies empha-\nsizing private passenger car electrification have shown that costs for \nkey components such as batteries are expected to fall substantially \nand quickly24\u201326, with increasing evidence that battery-electric vehi-\ncles will constitute the primary technology1,27. However, electrifying \nheavy commercial trucks, such as US Class 7/8 or European N2/N3, \nstill poses a tremendous challenge, particularly due to altered require-\nments limiting the transferability of passenger car findings. This can \nbe ascribed to inherently versatile operating characteristics from \nurban delivery to international long haul, distinct utilization sched-\nules, versatile value-adding duties from cargo transport to ancillary \npower- or energy-intensive services, greater lifetime mileage even \nReceived: 9 October 2023\nAccepted: 17 April 2024\nPublished online: 14 May 2024\n Check for updates\n1Fraunhofer Institute for Systems and Innovation Research ISI, Karlsruhe, Germany. 2Karlsruhe Institute of Technology (KIT), Institute of Electrical \nEngineering (ETI), Karlsruhe, Germany. \n\u2009e-mail: steffen.link@isi.fraunhofer.de\n\nNature Energy | Volume 9 | August 2024 | 1032\u20131039\n1033\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01531-9\nsettings, data originality, forecast method and source category. Addi-\ntionally, we used three robust approaches to strengthen our results by \nreducing potential outliers\u2019 contribution (Methods).\nRapidly declining battery system and fuel cell \nsystem costs\nFigure 1 illustrates that battery system costs, broken down by source \ncategories, may decline by 64% to 75% until 2050. We observe rapid and \nconsistent cost reductions per annum (p.a.), with similar patterns for \nall source categories. We find cost reductions (CRs) of around 5% p.a. \n(scientific and others) to 6.5% p.a. (near market) until 2030 and 3.3\u20134.5% \np.a. over an extended 2020\u20132050 period. Notably, near-market esti-\nmates (blue) are more optimistic, less heterogeneous and more stable \ncompared with the other categories. This consolidates into expected \ncost estimates, where near-market estimates project a decrease from \naround \u20ac2020275\u2009kWh\u22121 in 2020 to \u20ac2020140\u2009kWh\u22121 by 2030 and around \n\u20ac202070\u2009kWh\u22121 by 2050. In contrast, scientific estimates (green) indicate \na drop from roughly \u20ac2020310\u2009kWh\u22121 in 2020 to \u20ac2020180\u2009kWh\u22121 by 2030 \nand around \u20ac2020100\u2009kWh\u22121 by 2050. Other estimates (purple) also pro-\nject more conservative progress, cutting \u20ac2020200\u2009kWh\u22121 by 2030 and \napproximating \u20ac2020115\u2009kWh\u22121 by 2050. The cross-category projection \n(black) closely aligns with scientific projections.\nFigure 2 illustrates that FC system costs, broken down by source \ncategories, may decline by 65% to 85% until 2050. Notably, our obser-\nvations unveil significant heterogeneity among these categories. \nNear-market estimates (blue) initiate at approximately \u20ac2020540\u2009kW\u22121 in \n2020, undercut the \u20ac2020100\u2009kW\u22121 threshold by 2045 and attain around \n\u20ac202085\u2009kW\u22121 by 2050. This equals CRs of around 9% p.a. until 2030 and \naround 6% p.a. over 2020\u20132050. Conversely, scientific estimates (green) \ninitiate at approximately \u20ac2020\u2009kW\u22121 in 2020 and fall below \u20ac2020100\u2009kW\u22121 \nin the late 2030s, ultimately reaching around \u20ac202080\u2009kW\u22121 by 2050 and \nCRs of around 3.5% p.a. over 2020\u20132050. Other estimates (purple) are \ncentred between near-market and scientific estimates without reaching \nsub-\u20ac2020100\u2009kW\u22121 levels.\nCost overview for five major ZET components\nTable 1 shows the derived heavy ZET component costs (mean \u00b1 two \nstandard errors). Unlike for battery and FC costs, we find less substan-\ntial CR potentials for the adjacent ZET components using the same \nmethod, albeit with smaller samples and less detail (Supplementary \nFigs. 14\u201317). Precisely, costs for electric motors probably fall from \nbeyond 1\u2009million\u2009km, high reliability and longevity and an even more \npronounced cost sensitivity13,16,18,28. Hence, accurate and comprehen-\nsive data on current and projected ZET acquisition costs are essential to \nassess the future roles of these technologies. However, data are scarce \nand heterogeneous, whereas a holistic overview assessing multiple \ncomponents for heavy ZETs within a consistent scope and compara-\ntive method is missing. Thus, we address the following research ques-\ntion: what are the most likely future cost developments of central ZET \ncomponents until 2050?\nThis paper analyses projected costs for five crucial BET and/or FCET \ncomponents based on an extensive literature record: whereas we find lim-\nited cost reductions for the three adjacent components (that is, electric \ntraction motors, power electronics and high-voltage system (PE&HV) \ncomponents and hydrogen tanks), costs for battery and fuel cell (FC) sys-\ntems are primed to decline much faster than expected and due in course. \nDespite inevitable uncertainty, a rapid ZET market diffusion associated \nwith ambitious learning rates (LRs) at the required breakthrough costs \nseems within reach soon. Yet, prospects for BETs as primary technology \nseem more favourable at higher confidence, with faster availability and \nachieving cost effectiveness as of today, thus supporting an optimistic \nview on the fast decarbonization of road freight transport.\nMeta forecast using regression analysis\nMeta forecasting involves synthesizing existing projections from litera-\nture encompassing historical trends, future expectations and different \nmethods to yield novel ones. Pooling individual forecasts through this \nwell-established approach29,30 enhances accuracy, often outperforming \nindividual forecasts31.\nWe identified relevant literature via distinct search strings using \nthe Scopus and Google Scholar databases. Following Nykvist and Nils-\nson25, we considered three different source categories: near market \n(that is, market outlooks from renowned analysts and consultancies and \nindustry announcements), scientific (that is, peer-reviewed papers) and \nothers (that is, non-peer-reviewed academic publications and reports). \nAll cost values were harmonized to represent system-level original \nequipment manufacturer (OEM) purchase prices for the respective ZET \ncomponent, including profit mark-ups (+35%) and specified in 2020 \neuros (\u20ac2020). Component-specific cost developments were calculated \nby regression of log harmonized cost data and log time data using either \nweighted least squares (WLS) or ordinary least squares (OLS). Herein we \ncontrolled for several auxiliary variables such as release dates, scenario \n0\n200\n400\nBattery system costs (\u20ac2020 kWh\u20131)\n600\n800\n1,000\n1,200\n1,400\n1,600\n1,800\nHarmonized data points\nAll, N = 1,104, R2 = 0.49\nNear-market, N = 310, R2 = 0.69\nScientific, N = 339, R2 = 0.51\nOthers, N = 455, R2 = 0.4\n2010\n2015\n2020\n2025\n2030\n2035\n2040\n2045\n2050\nTime in years\n2020\n2025\n2030\nYear\n2035\n2040\n2045\n2050\n0\n100\n200\nBattery system costs\n (\u20ac2020 kWh\n\u22121)\n300\nFig. 1 | Projections for heavy ZET battery system costs. System-level costs per \nkWh of total gross battery capacity. These include, among others, battery and \nthermal management systems, cell modules, housing, connectors, wiring and \nassembly. Black circle markers represent the original harmonized data. Solid \nlines represent the regression results (log\u2013log, WLS, mean values), with the \nshaded areas (that is, error bars) in the zoom-level plot (period 2020\u20132050) as \n95% confidence intervals (mean \u00b1 2 s.e.m.). The source category is colour coded: \nnear market in blue (analysts, consultancies, industry announcements), scientific \nin green (peer-reviewed papers) and others in purple (non-peer-reviewed \nacademic publications). Total results (that is, cross-category) in black. The total \nsample covers N\u2009=\u20091,104 data points from 200 unique studies. Cross-category R\u00b2 \nis 0.49. The figure legend states the number of data points (N) and the R\u00b2 values \nper source category. Additional information is available in Supplementary  \nTables 1 and 4.\n\nNature Energy | Volume 9 | August 2024 | 1032\u20131039\n1034\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01531-9\naround \u20ac202042\u2009kW\u22121 by 2020 to \u20ac202030\u2009kW\u22121 by 2050 (\u22121.2% p.a.). For \nhydrogen tanks, we find CRs of around 2.6\u20132.9% p.a. over 2020\u20132050. \nThis translates into decreasing system costs from around \u20ac202017\u2009kWh\u22121 \n(liquid-LH2) and \u20ac202024\u2009kWh\u22121 (compressed-CH2) by 2020 to around \n\u20ac20207\u2009kWh\u22121 (LH2) and \u20ac202011\u2009kWh\u22121 (CH2) by 2050. Last, we derive stable \nsystem costs of around \u20ac202050\u2009kW\u22121 for PE&HV components.\nDifferent cost expectations for batteries and  \nfuel cells\nOur meta forecasts project a rapid cost decline for both batteries and \nFCs, while we disclose a contrasting dynamic that highlights the complex \ninterplay between scientific projections and tangible market realities \nfor emerging technologies at different stages of commercial maturity.\nOur analysis of battery cost predictions unveils that near-market \nestimates are remarkably stable over different release dates. These \nprojections are prone to only minor downward adjustments, as indi-\ncated by the difference between OLS and WLS results (Supplementary \nFigs. 5 and 6 and Supplementary Table 2), and are consistently more \noptimistic than those from scientific literature (p\u2009<\u20090.05, two-tailed \nt-test; Supplementary Table 4). Conversely, scientific cost estimates \npublished in 2010\u20132023 faced substantial downward adjustments. \nHence, battery costs have experienced a more rapid decline than ini-\ntially expected, at least in the scientific community. This echoes Nykvist \nand Nilsson25, who found similar divergences for industry vs market \nleaders vs peer-reviewed literature estimates and supports conclu-\nsions from Frith et al.32, who emphasize substantial gaps between \nacademic and industry perspectives. One explanation might be that \nnear-market sources may have more practical in-depth knowledge \nabout technologies, manufacturing or cost-saving measures and better \naccess to industry insights such as market trends, partnerships, supply \nchain dynamics or confidential pricing data. In contrast, parts of the \nscientific literature may be classified as theoretical estimates or may \nbe affected by citation patterns or time-delaying review processes, \nleading to the self-confirmation of outdated values and assumptions.\nHowever, FC system costs exhibit an inverse trend. Scientific \nestimates show higher stability and are consistently more optimistic \n(p\u2009<\u20090.05, two-tailed t-test; Supplementary Table 5) than near-market \nestimates.\nHarmonized data points\nAll, N = 424, R2 = 0.33\nNear-market, N = 96, R2 = 0.42\nScientific, N = 64, R2 = 0.26\nOthers, N = 264, R2 = 0.37\n0\n200\n400\nFuel cell system costs (\u20ac2020 kW\u20131)\n600\n800\n1,000\n1,200\n1,400\n1,600\n1,800\n2010\n2015\n2020\n2025\n2030\n2035\n2040\n2045\n2050\nTime in years\n0\n100\n200\nFuel cell system costs\n (\u20ac2020 kWh\n\u22121)\n300\n400\n2020\n2025\n2030\n2035\nYear\n2040\n2045\n2050\nFig. 2 | Projections for heavy ZET fuel cell system costs. System-level costs \nper kW of rated power. These include, among others, control systems, thermal \nmanagement, housing, hydrogen supply except for storage, air intake, including \ncompressor and humidifier and assembly. Black circle markers represent the \noriginal harmonized data. Solid lines represent the regression results (log\u2013log, \nWLS, mean values), with the shaded areas (that is, error bars) in the zoom-level \nplot (period 2020\u20132050) as 95% confidence intervals (mean \u00b1 2 s.e.m.). The \nsource category is colour coded: near market (blue: analysts, consultancies, \nindustry announcements), scientific (green: peer-reviewed papers) and others \n(purple: non-peer-reviewed academic publications). Total results (that is, \ncross-category) in black. The total sample size covers N\u2009=\u2009424 data points from 83 \nunique studies. Cross-category R\u00b2 is 0.33. The figure legend states the number of \ndata points (N) and the R\u00b2 values per source category. Additional information is \navailable in Supplementary Tables 1 and 4.\nTable 1 | Specific system-level component costs in \u20ac2020 for five major ZET components: batteries, fuel cells, electric \ntraction motors, PE&HV components, hydrogen storage tanks\nComponent\nSource category\n2020\n2030\n2040\n2050\nNumber of observations\nBattery in \u20ac2020\u2009kWh\u22121\nNear market\n275\u2009\u00b1\u20099.5\n141\u2009\u00b1\u20096.4\n94\u2009\u00b1\u20096.1\n70\u2009\u00b1\u20095.6\n310\nScientific\n310\u2009\u00b1\u200915\n178\u2009\u00b1\u20098.9\n127\u2009\u00b1\u20098.9\n100\u2009\u00b1\u20098.7\n339\nOthers\n316\u2009\u00b1\u200916\n193\u2009\u00b1\u20099.0\n143\u2009\u00b1\u20098.9\n116\u2009\u00b1\u20098.9\n455\nAll\n300\u2009\u00b1\u20098.1\n174\u2009\u00b1\u20095.0\n126\u2009\u00b1\u20095.0\n99\u2009\u00b1\u20094.9\n1,104\nFuel cell in \u20ac2020\u2009kW\u22121\nNear market\n538\u2009\u00b1\u200975\n216\u2009\u00b1\u200926\n125\u2009\u00b1\u200924\n84\u2009\u00b1\u200921\n96\nScientific\n219\u2009\u00b1\u200939\n132\u2009\u00b1\u200920\n97\u2009\u00b1\u200920\n78\u2009\u00b1\u200920\n64\nOthers\n415\u2009\u00b1\u200939\n219\u2009\u00b1\u200916\n149\u2009\u00b1\u200915\n113\u2009\u00b1\u200914\n264\nAll\n392\u2009\u00b1\u200930\n204\u2009\u00b1\u200912\n138\u2009\u00b1\u200912\n104\u2009\u00b1\u200911\n424\nElectric motor in \u20ac2020\u2009kW\u22121\nAll\n42\u2009\u00b1\u20095.3\n35\u2009\u00b1\u20093.3\n32\u2009\u00b1\u20094.3\n29\u2009\u00b1\u20095.2\n147\nPE&HV in \u20ac2020\u2009kW\u22121\nAll\n48\u2009\u00b1\u20096.5\n48\u2009\u00b1\u20095.2\n49\u2009\u00b1\u20097.8\n49\u2009\u00b1\u200910\n101\nCH2 tank in \u20ac2020\u2009kWh\u22121\nAll\n24\u2009\u00b1\u20092.1\n16\u2009\u00b1\u20091.0\n13\u2009\u00b1\u20091.1\n11\u2009\u00b1\u20091.2\n213\nLH2 tank in \u20ac2020\u2009kWh\u22121\nAll\n17\u2009\u00b1\u20091.4\n11\u2009\u00b1\u20090.6\n8.5\u2009\u00b1\u20090.7\n7.1\u2009\u00b1\u20090.8\n213\nResults are given as rounded mean value\u2009\u00b1\u20092 standard errors (95% confidence intervals) based on log\u2013log regression (WLS) for 2020, 2030, 2040 and 2050. All costs relate to either rated/\ncontinous power (kW) or gross capacity (in kWh). Values greater than ten as integers. Additional information is available in Supplementary Table 1 and Supplementary Figs. 14\u201317.\n\nNature Energy | Volume 9 | August 2024 | 1032\u20131039\n1035\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01531-9\nOne explanation for these different dynamics might be the techno-\nlogical maturity of both technologies and the related uncertainty about \nfuture pathways. The lower spread and consistency across all source \ncategories for battery system costs may indicate higher confidence \nand technological readiness, as BETs are increasingly projected to \ninherit a pivotal role in future road freight transport1\u2014at least on short \nto medium distances\u2014and have been announced in the portfolios of \nall major truck OEMs27,33. In contrast, the future role of FCETs and their \ndevelopment from laboratory to market is still more uncertain, result-\ning in larger data spread and potentially conservative projections from \nnear-market sources.\nRequired volumes, learning rates and \nbreakthrough costs\nWhereas cumulative production volumes may be the best-performing \npredictor of theoretical technology cost compared with others22,34, \nfindings are hard to generalize into policy decisions and anticipated \ntimescales. In parallel, the underlying data and thus observed cost \nestimates implicitly assume a certain relevance of battery technology \nand fuel cells for future road freight transport or even other sectors. \nThis relevance may be expressed as growing cumulative investments, \neconomies of scale, supply chain improvements, spill-over effects and \nmaterial or production improvements that would not occur without \nthis relevance. Hence, we compare our cost estimates against the poten-\ntial S-shaped market diffusion of ZETs in North America, Europe and \nChina, totalling around 2.5\u20133 million trucks produced p.a. (Methods \nand Supplementary Figs. 12 and 13).\nTo achieve battery system costs of approximately \u20ac2020150\u2009kWh\u22121 \nas indicated by our regression to be achievable between 2028 (near \nmarket) and 2032 (scientific), cumulative production volumes must \nrange from 1,300\u2009GWh (near market) to 5,200\u2009GWh (scientific). This \nyields short-term compound annual growth rates (CAGRs) of 39\u201349%. \nThose volumes may be feasible in the early 2030s if BETs take large mar-\nket shares fast, given their head start in the early 2020s. Correspond-\ning LRs would be around 16% (scientific) to 19% (near market). Falling \nbelow \u20ac2020100\u2009kWh\u22121, as indicated by our regression to be achievable \nbetween 2039 (near market) and 2049 (scientific), would require up to \n11,000\u2009GWh (near market) or even 68,000\u2009GWh (scientific), with the \nformer being probably feasible within the late 2030s given that BETs \ncomprise substantial market shares and long-term CAGRs of 25\u201329%.\nConsidering breakthrough levels, Teichert et al.35 find system costs \nof \u20ac2020120\u2013200\u2009kWh\u22121 to reach cost parity with current diesel trucks if \nfast charging becomes available. This confirms Nykvist and Olsson19, \nwho state around \u20ac2020200\u2009kWh\u22121 as the upper threshold. Basma et al.36 \nspecify system costs of around \u20ac2020100\u2009kWh\u22121 as viable tipping point for \nEuropean BETs to become cost effective, even without policy support \nthrough purchase incentives, adjusted road toll schemes or CO2 pric-\ning. For the United States, Phadke et al.37 quantify system costs below \nUS$135\u2009kWh\u22121, confirming Sripad and Viswanathan38, who quantified \nthe Tesla Semi case to be economically viable well below US$150\u2009kWh\u22121.\nSimilarly, to attain fuel cell system costs of approximately \u20ac2020150\u2009 \nkW\u22121, as our regression suggests between 2027 (scientific) and 2035 \n(near market), cumulative production volumes span from 135,000 \nunits (scientific) to 1.4 million units (near market). The latter appears \nattainable, considering the broader availability of FCET models antici-\npated to emerge in the late 2020s1, with short-term CAGRs of 35\u201346%. \nCorresponding LRs would be around 14% (scientific) to 26% (near mar-\nket). Falling below \u20ac2020100\u2009kW\u22121, as indicated by our regression to be \nachievable between 2040 (scientific) and 2045 (near market), would \nthen demand cumulative volumes from 2.3 million (scientific) to 6.8 \nmillion units (near market) and long-term CAGRs of 26\u201329%.\nOur ambitious LRs align with other technologies, with Nykvist \nand Nilsson25 suggesting conceivable LRs for batteries of 12\u201314%, and \nSchmidt et al.22 finding LRs of 9\u201318% for electrical energy storage \ntechnologies. Similarly, historical growth rates for wind and solar \ncapacity were at least 15% and often 39\u201350% (ref. 39), whereas gen-\neral technology adoption growth rates were often below 13\u201314% but \nselectively exceeded 30\u201340% (ref. 40). For BEVs, sales growth ranged \naround 25\u201355% (Supplementary Table 7), with Norway achieving nearly \n90% BEV sales within 13 years (CAGR2010\u20132023\u2009=\u200941.4%), compared with \nEurope\u2019s 25% share (CAGR2010\u20132023\u2009=\u200928.1%).\nDiscussion\nOver-extrapolation and theoretical assessments may result in deceptive \nconclusions of future performance and relevance, which is critical for \npublic and private funding and, hence, policy decisions32. Therefore, \nwe contrast our findings to target or floor costs\u2014typically defined \nby raw material and production costs\u2014to avoid excessively low-cost \nestimates22,34,41.\nRecalibrating our battery system costs to the cell level (Sup-\nplementary Table 6), we arrive at around \u20ac202090\u2009kWh\u22121 by 2030 or \n\u20ac202070\u2009kWh\u22121 by 2050. Several facts underpin the feasibility of the \nderived costs despite potential non-negligible disruptions caused \nby raw material shortages, supply chain disruptions, higher inflation \nlevels, increased energy costs or raw material shortages42\u201344, whereof \nthe latter\u2019s effect would depend on the specific chemistry. Calculated \nsystem- and cell-level costs do not fall below original scientific or \nnear-market estimates and announced industry values, as could have \nhappened by the regression. For example, the European Battery Stra-\ntegic Research and Innovation Agenda (SRIA)45 targets system costs \nof around \u20ac202075\u2009kWh\u22121 by 2025 and even below by 2030 and beyond. \nPlus, Tesla confirmed cell-level cost targets for its 4,680 cylindrical \ncells of around US$70\u2009kWh\u22121 at their third quarter Earning Call in 2022\u2014\neven before US incentive programmes such as the Inflation Reduction \nAct. The projected ready-to-drive prime costs of the first generation \n500-mile Tesla Semi truck, equipped with these cells and suspected \n800\u2013900\u2009kWh, stand at roughly US$200,000 in 202346, indicating \nassociated battery system costs below US$150\u2009kWh\u22121. However, we \nalso address potential strategies for gaining initial market shares by \ninternally subsidizing battery packs.\nFor fuel cell system costs, the US Department of Energy ultimately \ntargets HDV FC system costs of around US$201680\u2009kW\u22121 by 2030 and \nUS$201660\u2009kW\u22121 by 205047, with both targets adopted by numerous \nstudies. Similarly, the European hydrogen SRIA48 targets system costs \nof below \u20ac100\u2009kW\u22121 by 2030. This indicates a notable level of ambition \nor required federal support but might also mean that our projections \n(specified as OEM purchase price) may be quite optimistic and close \nto anticipated floor and target costs.\nOur cost-centred meta forecast focuses on most likely develop-\nments and consistent comparisons, potentially omitting scenario \ndependencies and technical particularities/trends.\nUnderlying scenarios and assumptions, such as different mitiga-\ntion pathways, can impact production volumes and, thus, component \ncost developments. However, insufficient standardized scenario clas-\nsifications and missing congruent information impede robust and \nconsistent cross-study assessments. Further regressions indicate that \noptimistic cost scenarios imply additional cost benefits for batteries \nand FCs (Supplementary Tables 4 and 5).\nThe current effects of technical designs and material choices on \ncosts and performance are well documented24,49, with pre-series and \ncommercial components disclosing real-world pricing or facilitat-\ning cost versus performance trade-off analyses35 and product tear \ndowns50,51. However, exact technical designs, materials and properties \nof next-generation components until 2050 are theoretical, remain \nhighly uncertain and are only selectively available. This impedes com-\nprehensive techno-economic assessments for single technologies \nacross 2020\u20132050.\nWhereas cost reductions and increased technical performance \nare anticipated for batteries52,53 and FCs47,48, our stable results for \nPE&HV and minor savings for electric motors indicate that technical \n\nNature Energy | Volume 9 | August 2024 | 1032\u20131039\n1036\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01531-9\nadvancements involving new components and trends may offset cost \nreductions. This may involve advancements54,55 to increase efficiency \nand thus driving range per charge, such as transitioning to higher \nvoltage, new motor topologies and materials or bidirectional system \narchitectures to facilitate Vehicle-to-Grid (V2G) applications.\nConcerning data limitations and potential bias, we state that our \ndata sample is very recent, with most sources being published between \n2018 and 2022 (Supplementary Fig. 1). Whereas we consider our find-\nings regarding battery system costs as highly robust, we acknowledge \nuncertainty for FC system costs.\nThe battery dataset, comprising N\u2009=\u2009200 sources, displays a bal-\nanced representation across all categories of sources, encompassing \nvarious methods and scenarios employed therein. Consequently, \nwe have not identified either substantial bias originating from the \ndata sample or issues from the data harmonization (Supplementary \nFigs. 2\u20134 and 11 and Supplementary Table 4).\nThe FC data sample consists N\u2009=\u200983 sources and exhibits an imbal-\nance towards other sources, particularly limited upon near-market \nestimates, limiting the significance of cross-category outcomes. Our \nfindings indicate that methods and scenario considerations exert more \npronounced effects on the results. Furthermore, the impact of data \nharmonization is more accentuated, with harmonized values being \nlower than the original ones, indicating cost developments are too low \n(15\u201325%) (Supplementary Figs. 2\u20134 and 11 and Supplementary Table 5).\nThree approaches using two techniques, namely robust norms and \noutlier removal, strengthened the validity of our original regression \nresults. Precisely, HuberT regression, RANSAC regression and WLS \nregression with only the central 50% or 80% of observations within \nfive-year time windows yield future costs close to or within our origi-\nnal prediction errors (Supplementary Figs. 7\u201310 and Supplementary \nTable 3).\nFor both battery and fuel cell data, and similar to Schmidt et al.22, we \nfind that more recent sources embed faster and larger cost-reduction \npotentials, expressed by the difference between OLS and WLS results \n(Supplementary Figs. 5 and 6 and Supplementary Table 2).\nTotal Cost of Ownership (TCO) benefits against diesel trucks (DTs) \ntypically constitute the key ZET criterion for fleet operators9\u201313, with \nother factors being also relevant. Using a recent TCO framework56 along \nwith our cost projections, we find that BETs may realize cost benefits \nversus DTs as of today. In contrast, FCETs may struggle to reach TCO \nparity throughout the 2030s because green hydrogen prices remain \nprobably too high. Herein the share of acquisition costs substantially \nrises for ZETs compared with current DTs, whereas energy storage size, \nenergy prices and mileage are the most sensitive parameters (Methods \nand Supplementary Figs. 18 and 19).\nAlongside economic considerations, technical capabilities such \nas feasible range, realizable payload, reliability, ageing behaviour \nand recharging/refuelling times are further influencing factors on \ntruck purchase decisions. Several studies indicate that current and \nannounced ZETs are already close to become technically competitive \nwith DTs19,35,49,57. Finally, infrastructure availability and user acceptance \nwill be decisive9\u201313.\nConclusion\nThis article presents a systematic overview of cost estimates for five \nmajor ZET components with meta forecasting and regression analysis. \nWe draw four conclusions from this analysis.\nFirst, we show that ZET component costs are likely to decline \nsubstantially and in due course. Precisely, future battery system costs \nare more robust, likely to fall below \u20ac2020150\u2009kWh\u22121 by no later than \n2035, and to approach or even cross \u20ac2020100\u2009kWh\u22121 upon 2050, with \nthe former corresponding to typical expected breakthrough levels. \nFC system costs are likely to reach around \u20ac2020150\u2009kW\u22121 in the late \n2030s and to approach \u20ac2020100\u2009kW\u22121 at best in the late 2040s, with \nlower values close to target and floor cost values warranting careful \nconsideration. We emphasize that calculated LRs, cost reductions and \ngrowth rates are challenging to reach, but similar ranges have been \nwitnessed for (energy) technologies in the past22,39,40, with similar to \nhigher scales for FC values.\nSecond, we find that cost predictions differ systematically between \ndifferent source categories, potentially depending on technological \nmaturity. Whereas near-market sources turn to be the most stable \nsource for batteries, their predictions are more optimistic than those \nfrom scientific sources. However, the opposite is true for fuel cells, \nwith the scientific sources being more optimistic and often close to \nfloor or target costs.\nThird, these substantial and fast cost-reduction potentials sup-\nport an optimistic outlook for both technologies. This indicates rapid \nZET market deployments that will substantially impact transporta-\ntion and energy sector players, such as value chain reconfigurations, \nestablishing national and international hydrogen ecosystems and \nelectricity infrastructure expansions from transformers to distribution \nand transmission grids.\nFinally, we highlight the competition among ZET technologies, \nraising questions about market leadership and whether we need dif-\nferent technologies. All anticipated cost reductions rely on successful \ntransitions to low-carbon road freight transport. This entails building \nlarge-scale production facilities supported by policy measures in key \nmarkets like North America, Europe and China, particularly in the \nearly market phases. These measures may include purchase subsidies, \ninfrastructure development, ZET mandates and carbon pricing. This \npolicy support may phase out later when the technology has matured \nand costs have decreased. Our TCO indicates that BETs may constitute \nthe most cost-effective pathway in reaching TCO parity with less policy \nsupport needed, in contrast to FCETs, which might require more policy \nsupport throughout the 2030s.\nMethods\nLiterature identification\nWe identified the relevant academic literature via distinct search \nstrings using the Scopus and Google Scholar databases and comple-\nmented the identified studies with commercial market outlooks from \nrenowned analysts and consultancies and industry announcements. \nTable 2 presents the search strategy, which has been used on article \ntitles, abstracts and keywords published between 2010 and May 2023. \nAvailable literature has then been analysed and filtered based on the \nabstracts. The final sample for batteries covered NB\u2009=\u2009200 sources and \nDPB\u2009=\u20091,104 data points. The final sample for fuel cells covered NFC\u2009=\u200983 \nsources and DPFC\u2009=\u2009424 data points.\nLiterature parameters\nWe collected available auxiliary data from each source to either harmo-\nnize the data or integrate those parameters as control variables in the \nregression analysis to compare and validate the results: (1) value type, \ndifferentiating between cost or price, with the latter typically includ-\ning additional overheads, mark-ups, indirect costs or supplier profits. \nHowever, we denote that both terms are also used synonymously.  \n(2) Application type, differentiating between automotive- or \nHDV-certified values, as altered requirements and scales also lead to \ndifferent configurations, designs and costs. (3) Currency and (4) refer-\nence year information were collected to ensure accurate contextualiza-\ntion and temporal accuracy. (5) Scenario information was collected, \ndifferentiating between low (mass-market), high (niche market) and \nmedium. (6) Forecast method, differentiating between literature-based \nprojections, expert elicitation, detailed bottom-up cost modelling \nor learning and experience curves. (7) Data originality, differenti-\nating between original or adopted values. (8) Integration level, dif-\nferentiating between cell- or system-level values for batteries and \nstack- or system-level values for fuel cells. This raw data are available  \nfor download, yet proprietary is anonymized.\n\nNature Energy | Volume 9 | August 2024 | 1032\u20131039\n1037\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01531-9\nLiterature classification\nIn addition to the parameters stated above, all sources have been clas-\nsified into three categories to further differentiate results (similar to \nNykvist and Nilsson25). Accordingly, we differentiated between scien-\ntific (literature) covering peer-reviewed papers and PhD theses. Others \ninvolved non-peer-reviewed academic publications such as conference \narticles, reports, working or white papers and book contributions. \nNear market involved market outlooks from renowned analysts or \nconsultancies and industry announcements.\nData harmonization\nData harmonization, based on the upper parameters, ensured that \nall data points were specified as system-level OEM purchase prices in \n\u20ac2020 for the respective component. Initially, this involved (3) currency \nconversions based on the respective historical exchange rate (subject \nto the reference year or release date) and (4) inflation adjustment to \n2020 levels (annual mean over all member states of the European Union \n(EU-27) issued by the European Commission and downloadable via \nEurostat). Potential inhomogeneities caused by (1) cost or price type, \n(2) application type or (8) integration level are harmonized based on \nstudies that explicitly state multiple values differentiated according \nto (1), (2), (8). Hence, all cost values were topped with a 35% surcharge \n(median value from battery data as no temporal trend could be iden-\ntified) to derive OEM purchase prices. Scaling automotive-certified \ncomponents to HDV-certified ones, data showed a temporal trend \nfor batteries but not for fuel cells. Hence, we used a decreasing scal-\ning factor for batteries from around 80% in 2020 to 50% in 2030 and \n20% in 2050, indicating higher integration, potential enhancements \nthrough on-purpose designs and usable synergies. For fuel cells, all \nautomotive-type values were topped with a 100% surcharge (median \nvalue). For cell-to-system or stack-to-system scaling, data showed a \nclear temporal trend for both batteries and fuel cells. Therefore, we \nused a decreasing scaling for batteries from around 40% in 2020 to \n30% in 2030 and 20% in 2050. In contrast, we used an increasing scal-\ning factor for fuel cells from 60% in 2020 to 90% in 2030 and 125% in \n2050, meaning that the cost share of the actual stacks on total system \ncosts was expected to decrease. Supplementary Figs. 2\u20134 provide more \ndetails. These harmonized data are available for download.\nRegression analysis and control variables\nResults were calculated by regression using Python statsmodels. Spe-\ncifically, we used the log harmonized cost data and log time data to \napproximate the typical learning curve shape and controlled for several \nauxiliary variables stated above. Weighting (WLS) was performed by \nsource age using the following exponential function for smoothing, \nas proposed by refs. 58,59:\nwi = 0.8\n2023\u2212yi with i \u2208{study 1, \u2026 study n}\nand yi as the respective study year\n(1)\nRobust approaches\nWe used two robust techniques, namely robust norms and outlier \nremoval, in three approaches to exclude outliers and noise, thus \nincreasing the accuracy and robustness of our original regression \nresults. First, we performed HuberT regression analyses that are less \nsensitive to outliers by minimizing a combined loss function of squared \nerrors for small residuals and absolute errors for larger residuals but \nstill using the full data sample. Second, we filtered the data by labelling \nthe central 50% (that is, values within the lower and upper quartile) and \n80% (that is, values within the 10% and 90% quantile) of observations \nas inliers and others as outliers. We then performed a WLS regression \nwith inliers only. Third, we performed a RANSAC (Random Sample Con-\nsensus) regression that iteratively selects random data subsets to fit a \nregression model, identifies inliers based on a predetermined thresh-\nold (that is, median absolute deviation), refits the model using these \ninliers and selects the best model (that is, regression coefficients) based \non the minimum absolute error. Supplementary Figs. 7\u201310 provide \nfurther details and Supplementary Table 3 provides model comparison.\nTCO framework\nWe adopted the total cost of ownership (TCO) framework (that is, \ncalculation and parameter assumptions, excluding component costs) \nfrom Noll et al.56 to calculate TCO per kilometre (\u20ac\u2009km\u22121) over the whole \nvehicle service life. This includes capital expenditures such as truck pur-\nchase (Supplementary Note 1) and resale and operating expenditures, \nsuch as energy costs, road tolls, maintenance and service. The capital \nrecovery factor discounts future payments using a specific discount \nrate. We excluded any subsidy or purchase price premiums for ZETs, \naveraged all parameters at a European level and tested our results \nagainst various energy prices and annual vehicle mileages. Supplemen-\ntary Figs. 18 and 19 and Supplementary Table 8 provide more details.\nCumulative volumes and learning rates\nWe used S-shaped diffusion curves (sigmoid functions) to obtain poten-\ntial BET/FCET shares per year, using the following function:\nyt =\ny0 \u00d7 S\ny0 + (S \u2212y0) \u00d7 e\u2212kSt\n(2)\nwhere yt (%) is the annual BET/FCET share in a particular year, S (%) is \nthe total annual market capacity (=\u2009100%), y0 (%) is the initial share for \nthe starting year and k is the growth rate. This share is then multiplied \nby the combined North American, European and greater China HDV \nproduction volume of around 2.5\u2009million\u20133\u2009million units per year to \nderive annual volumes. Following the announced or expected ZET sales \nshares for those regions, we assume most new trucks will have zero \nemissions by 2040 to 2050 (y2040\u2009>\u200990% and y2050\u2009=\u2009100%) in Europe and \nNorth America, whereas greater China will reach this threshold later \n(y2060\u2009=\u2009100%). This leads to annual growth rates between around 25% \n(China) and 40% (Europe and North America). We assume the same \nisolated market diffusion for both technologies\u2014BET and FCET\u2014to \nguarantee comparability, and we ignore any other ZET technology. The \naverage BET battery capacity was assumed to increase from 300\u2009kWh \nin 2018 to 500\u2009kWh in 2025 and 600\u2009kWh from 2030 onwards. The \ncumulative volume (in GWh) is calculated by multiplying the battery \ncapacity and annual volumes. Initial sales/production data on BETs \nand FCETs were then matched to respective reference years \n( y2018, \u2026 , y2022), which allowed calculation of learning rates by regres-\nsion of log cost data and log cumulative volumes. Supplementary \nFigs. 12 and 13 provide more details.\nData availability\nAll fuel cell and battery data presented in this study (that is, raw and \nharmonized datasets) are publicly available and can be found in the \nattached Supplementary Information. However, for raw data files, pro-\nprietary data from commercial sources, such as purchased market stud-\nies, cannot be published and corresponding records are anonymized. \nData for other components are available upon request. Source data \nare provided with this paper.\nTable 2 | Search strategy applied in the databases Scopus \nand Google Scholar\nBatteries\nFuel cells\nKeywords\n\u2018batter*\u2019 AND (\u2018cost\u2019 OR \n\u2018price\u2019) AND (\u2018truck*\u2019 OR \n\u2018heavy-duty\u2019)\n\u2018fuel cell*\u2019 AND (\u2018cost\u2019 OR \n\u2018price\u2019) AND (\u2018truck*\u2019 OR \n\u2018heavy-duty\u2019)\nField of search\nArticle title, abstract, keywords\nPeriod\n2010\u2013May 2023\n\nNature Energy | Volume 9 | August 2024 | 1032\u20131039\n1038\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01531-9\nCode availability\nThe Python model (OLS and WLS regression, robust regression and \nvisualization for Figs. 1 and 2) is available in the attached Supplemen-\ntary Information. Reproduced figures or regression results based \non the raw data files may not involve any proprietary data leading to \nchanges compared with the original results and plots in this paper.\nReferences\n1.\t\nJaramillo, P. et al. in IPCC, 2022 - Climate Change 2022: Mitigation \nof Climate Change (eds Shukla, P. R. et al.) Ch. 10 (Cambridge \nUniv. Press, 2023).\n2.\t\nPl\u00f6tz, P. et al. Greenhouse gas emission budgets and policies for \nzero-carbon road transport in Europe. Clim. Policy 23, 343\u2013354 \n(2023).\n3.\t\nCreutzig, F. et al. Energy and environment. Transport: a roadblock \nto climate change mitigation? Science 350, 911\u2013912 (2015).\n4.\t\nTracking Clean Energy Progress 2023 https://www.iea.org/reports/\ntracking-clean-energy-progress-2023 (IEA, 2023).\n5.\t\nHeavy-duty Vehicles: Council and Parliament Reach a Deal  \nto Lower CO2 Emissions from Trucks, Buses and Trailers  \nhttps://www.consilium.europa.eu/en/press/press-releases/ \n2024/01/18/heavy-duty-vehicles-council-and-parliament-reach-a- \ndeal-to-lower-co2-emissions-from-trucks-buses-and-trailers \n(European Council, 2024).\n6.\t\nCalifornia Approves World\u2019s First Regulation to Phase Out Dirty \nCombustion Trucks and Protect Public Health https://www.gov.\nca.gov/2023/04/28/california-approves-worlds-first-regulation- \nto-phase-out-dirty-combustion-trucks-and-protect-public-health/ \n(Office of Gov. Gavin Newsom, 2023).\n7.\t\nZhang, Y. & Hui, H. China\u2019s efforts to decarbonize road \ntransport: decent, but not sufficient. ICCT https://theicct.org/\nchina-ev-efforts-mar22 (2023).\n8.\t\nKhanna, N., Lu, H., Fridley, D. & Zhou, N. Near and long-term \nperspectives on strategies to decarbonize China\u2019s heavy-duty \ntrucks through 2050. Sci. Rep. 11, 20414 (2021).\n9.\t\nAnderhofstadt, B. & Spinler, S. Factors affecting the purchasing \ndecision and operation of alternative fuel-powered heavy-duty \ntrucks in Germany\u2014a Delphi study. Transp. Res. Part D Transp. \nEnviron. 73, 87\u2013107 (2019).\n10.\t Bae, Y., Mitra, S. K., Rindt, C. R. & Ritchie, S. G. Factors influencing \nalternative fuel adoption decisions in heavy-duty vehicle  \nfleets. Transp. Res. Part D Transp. Environ. 102, 103150  \n(2022).\n11.\t\nRagon, P.-L. & Rodr\u00edguez, F. Road Freight Decarbonization in \nEurope. Readiness of the European Fleets for Zero-emission \nTrucking https://theicct.org/wp-content/uploads/2022/09/\nroad-freight-decarbonization-europe-sep22.pdf (ICCT, 2022).\n12.\t Spiller, B., Lohawala, N. & DeAngeli, E. Medium- and Heavy-Duty \nVehicle Electrification: Challenges, Policy Solutions, and \nOpen Research Questions https://media.rff.org/documents/\nReport_23-03_v3.pdf (RFF, 2023).\n13.\t Muratori, M., Borlaug, B., Ledna, C., Jadun, P. & Kailas,  \nA. Road to zero: research and industry perspectives on \nzero-emission commercial vehicles. iScience 26, 106751  \n(2023).\n14.\t Pl\u00f6tz, P. Hydrogen technology is unlikely to play a major role in \nsustainable road transport. Nat. Electron. 5, 8\u201310 (2022).\n15.\t Why the battery electric drive represents the future for trucks. Traton \nhttps://traton.com/en/newsroom/current-topics/why-the-battery- \nelectric-drive-represents-the-future-for-trucks.html (2022).\n16.\t Staffell, I. et al. The role of hydrogen and fuel cells in the global \nenergy system. Energy Environ. Sci. 12, 463\u2013491 (2019).\n17.\t Meyer, T. Decarbonizing road freight transportation\u2014a \nbibliometric and network analysis. Transp. Res. Part D Transp. \nEnviron. 89, 102619 (2020).\n18.\t Borlaug, B. et al. Heavy-duty truck electrification and the impacts \nof depot charging on electricity distribution systems. Nat. Energy \n6, 673\u2013682 (2021).\n19.\t Nykvist, B. & Olsson, O. The feasibility of heavy battery electric \ntrucks. Joule 5, 901\u2013913 (2021).\n20.\t Niazi, A., Dai, J. S., Balabani, S. & Seneviratne, L. Product cost \nestimation: technique classification and methodology review.  \nJ. Manuf. Sci. Eng. 128, 563\u2013575 (2006).\n21.\t Farmer, J. D. & Lafond, F. How predictable is technological \nprogress? Res. Policy 45, 647\u2013665 (2016).\n22.\t Schmidt, O., Hawkes, A., Gambhir, A. & Staffell, I. The future cost \nof electrical energy storage based on experience rates. Nat. \nEnergy https://doi.org/10.1038/nenergy.2017.110 (2017).\n23.\t Luo, X., Wang, J., Dooner, M. & Clarke, J. Overview of current \ndevelopment in electrical energy storage technologies and the \napplication potential in power system operation. Appl. Energy \n137, 511\u2013536 (2015).\n24.\t Schmuch, R., Wagner, R., H\u00f6rpel, G., Placke, T. & Winter, M. \nPerformance and cost of materials for lithium-based rechargeable \nautomotive batteries. Nat. Energy 3, 267\u2013278 (2018).\n25.\t Nykvist, B. & Nilsson, M. Rapidly falling costs of battery packs for \nelectric vehicles. Nat. Clim. Change 5, 329\u2013332 (2015).\n26.\t Mauler, L., Duffner, F., Zeier, W. G. & Leker, J. Battery cost \nforecasting: a review of methods and results with an outlook to \n2050. Energy Environ. Sci. 14, 4712\u20134739 (2021).\n27.\t IEA. Global EV Outlook 2023. Catching Up With Climate Ambitions \n(IEA, 2023).\n28.\t Becker, H. et al. Preparing the World for Zero-emission Trucks. The \nMainstays of Commercial Road Transport Will Soon Benefit from \nCost-effective, Zero-emission Horsepower (McKinsey & Co., 2022).\n29.\t Atiya, A. F. Why does forecast combination work so well? Int. J. \nForecasting 36, 197\u2013200 (2020).\n30.\t Clemen, R. T. Combining forecasts: a review and annotated \nbibliography. Int. J. Forecasting 5, 559\u2013583 (1989).\n31.\t Timmermann, A. Chapter 4 forecast combinations. Handb. Econ. \nForecasting 1, 135\u2013196 (2006).\n32.\t Frith, J. T., Lacey, M. J. & Ulissi, U. A non-academic perspective \non the future of lithium-based batteries. Nat. Commun. 14, 420 \n(2023).\n33.\t Drive to zero\u2019s zero-emission technology inventory (ZETI). Tool \nversion 8.0. CALSTART https://globaldrivetozero.org/tools/\nzero-emission-technology-inventory/ (2022).\n34.\t Junginger, M. Technological Learning in the Transition to a \nLow-Carbon Energy System. Conceptual Issues, Empirical \nFindings, and Use in Energy Modeling (Elsevier Science & \nTechnology, 2020).\n35.\t Teichert, O., Link, S., Schneider, J., Wolff, S. & Lienkamp, M. \nTechno-economic cell selection for battery-electric long-haul \ntrucks. eTransportation 16, 100225 (2023).\n36.\t Basma, H., Saboori, A. & Rodr\u00edguez, F. Total Cost of Ownership for \nTractor-trailers in Europe: Battery-electric Versus Diesel  \nhttps://theicct.org/wp-content/uploads/2021/11/tco-bets- \neurope-1-nov21.pdf (ICCT, 2021).\n37.\t Phadke, A., Khandekar, A., Abhyankar, N., Wooley, D. & Rajagopal, \nD. Why Regional and Long-Haul Trucks are Primed for Electrification \nNow (LBNL, 2021).\n38.\t Sripad, S. & Viswanathan, V. Quantifying the economic case for \nelectric semi-trucks. ACS Energy Lett. 4, 149\u2013155 (2019).\n39.\t Odenweller, A., Ueckerdt, F., Nemet, G. F., Jensterle, M. & Luderer, \nG. Probabilistic feasibility space of scaling up green hydrogen \nsupply. Nat Energy 7, 854\u2013865 (2022).\n40.\t Nemet, G., Greene, J., M\u00fcller-Hansen, F. & Minx, J. C. Dataset on \nthe adoption of historical technologies informs the scale-up of \nemerging carbon dioxide removal measures. Commun. Earth \nEnviron. 4, 397 (2023).\n\nNature Energy | Volume 9 | August 2024 | 1032\u20131039\n1039\nAnalysis\nhttps://doi.org/10.1038/s41560-024-01531-9\n41.\t Hsieh, I.-Y. L., Pan, M. S., Chiang, Y.-M. & Green, W. H. Learning \nonly buys you so much: practical limits on battery price \nreduction. Appl. Energy 239, 218\u2013224 (2019).\n42.\t Mauler, L., Lou, X., Duffner, F. & Leker, J. Technological innovation \nvs. tightening raw material markets: falling battery costs put at \nrisk. Energy Adv. 1, 136\u2013145 (2022).\n43.\t BloombergNEF\u2019s annual battery price survey 2022. Lithium-ion \nbattery pack prices rise for first time to an average of $151/kWh. \nBloombergNEF https://about.bnef.com/blog/lithium-ion-battery- \npack-prices-rise-for-first-time-to-an-average-of-151-kwh  \n(2022).\n44.\t The Lithium-Ion (EV) Battery Market and Supply Chain. Market \nDrivers and Emerging Supply Chain Risks https://content.\nrolandberger.com/hubfs/07_presse/Roland%20Berger_The%20\nLithium-Ion%20Battery%20Market%20and%20Supply%20\nChain_2022_final.pdf (Roland Berger, 2022).\n45.\t Strategic Research and Innovation Agenda https://bepassociation.\neu/wp-content/uploads/2021/09/BATT4EU_reportA4_SRIA_V15_\nSeptember.pdf (BATT4EU, 2021).\n46.\t Kane, M. Elon Musk confirms Tesla Semi will enter the market \nlater this year. InsideEVs https://insideevs.com/news/603515/\ntesla-semi-deliveries-later-this-year/ (2022).\n47.\t Marcinkoski, J. et al. DOE Advanced Truck Technologies. \nSubsection of the Electrified Powertrain Roadmap\u2014Technical \nTargets for Hydrogen-Fueled Long-Haul Tractor-Trailer \nTrucks https://www.hydrogen.energy.gov/docs/\nhydrogenprogramlibraries/pdfs/19006_hydrogen_class8_long_\nhaul_truck_targets.pdf?Status=Master (DOE, 2019).\n48.\t Strategic Research and Innovation Agenda 2021\u20132027 Annex  \nto GB decision number CleanHydrogen-GB-2022-02  \nhttps://www.clean-hydrogen.europa.eu/about-us/\nkey-documents/strategic-research-and-innovation-agenda_en \n(Clean Hydrogen Partnership, 2022).\n49.\t Cullen, D. A. et al. New roads and challenges for fuel cells in \nheavy-duty transportation. Nat. Energy 6, 462\u2013474 (2021).\n50.\t G\u00fcnter, F. J. & Wassiliadis, N. State of the art of lithium-ion \npouch cells in automotive applications: cell teardown and \ncharacterization. J. Electrochem. Soc. 169, 30515 (2022).\n51.\t Ank, M. et al. Lithium-ion cells in automotive applications:  \nTesla 4680 cylindrical cell teardown and characterization.  \nJ. Electrochem. Soc. 170, 120536 (2023).\n52.\t Gao, Y., Pan, Z., Sun, J., Liu, Z. & Wang, J. High-energy batteries: \nbeyond lithium-ion and their long road to commercialisation. \nNano-Micro Lett. 14, 94 (2022).\n53.\t Battery Requirements for Future Automotive Applications  \nhttps://eucar.be/wp-content/uploads/2019/08/20190710-EG-BEV-\nFCEV-Battery-requirements-FINAL.pdf (EUCAR, 2019).\n54.\t Cai, W., Wu, X., Zhou, M., Liang, Y. & Wang, Y. Review and \ndevelopment of electric motor systems and electric powertrains \nfor new energy vehicles. Automot. Innov. 4, 3\u201322 (2021).\n55.\t Husain, I. et al. Electric Drive Technology Trends, Challenges, and \nOpportunities for Future Electric Vehicles (ORNL, 2021).\n56.\t Noll, B., Del Val, S., Schmidt, T. S. & Steffen, B. Analyzing \nthe competitiveness of low-carbon drive-technologies in \nroad-freight: a total cost of ownership analysis in Europe. Appl. \nEnergy 306, 118079 (2022).\n57.\t Mauler, L., Dahrendorf, L., Duffner, F., Winter, M. & Leker, J. \nCost-effective technology choice in a decarbonized and \ndiversified long-haul truck transportation sector: a US case study. \nJ. Energy Storage 46, 103891 (2022).\n58.\t Hyndman, R. J. & Athanasopoulos, G. in Forecasting: Principles \nand Practice 2nd edn Ch. 7 (OTexts, 2018).\n59.\t Guthrie, W. F. e-Handbook of Statistical Methods Ch. 6.4.3  \nhttp://www.itl.nist.gov/div898/handbook/ (NIST, 2020).\nAcknowledgements \nWe gratefully acknowledge funding from the EU STORM project \n(grant number 101006700: P.P., D.S., S.L.) by the European Union\u2019s \nHorizon 2020 research and innovation programme, the HoLa project \n(FKZ 03EMF0404A: P.P., D.S., S.L.) by the German Federal Ministry for \nDigital and Transport and the BEMA project (FKZ 03XP0272B: S.L., \nA.S.) by the German Federal Ministry for Education and Research. \nWe thank M. Wietschel and H. D\u00f6scher for their comments and \ndiscussions and S. Lange (all Fraunhofer ISI) for English editing.\nAuthor contributions\nS.L. conceived and designed the study in consultation with P.P. and \nA.S. S.L. collected the data, implemented the model and created the \nvisualizations. S.L. wrote the original manuscript with contributions \nfrom all co-authors. In particular, P.P. and A.S. contributed to the \ndiscussion, interpretation of findings, recommendations and policy \nimplications. D.S. edited the paper and contributed to the data \ncollection. P.P. supervised the study.\nFunding\nOpen access funding provided by Fraunhofer-Gesellschaft zur \nF\u00f6rderung der angewandten Forschung e.V.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains \nsupplementary material available at  \nhttps://doi.org/10.1038/s41560-024-01531-9.\nCorrespondence and requests for materials should be addressed to \nSteffen Link.\nPeer review information Nature Energy thanks Heikki Liimatainen, \nMehdi Jahangir Samet and the other, anonymous, reviewer(s) for their \ncontribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this licence, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2024, corrected publication 2024\n\n\n    Scientific Research Findings:", "answer": "We find that costs for zero-emission truck components will decline faster than expected, projecting battery costs to cut \u20ac200\u00a0kWh\u20131 soon and approach \u20ac100\u00a0kWh\u20131 towards\u00a02050. In contrast, fuel cell costs are likely to reach \u20ac150\u00a0kW\u20131 in the late\u00a02030s, albeit with greater uncertainty due to lower commercial maturity and doubts about realizing floor costs. The analysis stresses the challenge of achieving ambitious learning rates and cost reductions while highlighting the historical precedent for such advancements. Battery\u2011electric trucks are very likely the most cost\u2011effective pathway for achieving cost parity with diesel trucks. They may also require less financial incentives than fuel cell trucks, which are likely to struggle from hydrogen scarcity. Thus, prospects for battery\u2011electric trucks as the leading future truck technology seem more favourable at higher confidence, backed by faster availability and higher profitability. Additionally, other studies indicate that their technical competitiveness with diesel trucks is close for most cases  of logistics.", "id": 10}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 9 | January 2024 | 92\u2013105\n92\nnature energy\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nGeographies of regulatory disparity \nunderlying Australia\u2019s energy transition\nLee V. White\u2009\n\u200a\u20091\u2009\n, Bradley Riley\u2009\n\u200a\u20092, Sally Wilson\u2009\n\u200a\u20091,3, Francis Markham\u2009\n\u200a\u20092, \nLily O\u2019Neill\u2009\n\u200a\u20094, Michael Klerck\u2009\n\u200a\u20092,5 & Vanessa Napaltjari Davis\u2009\n\u200a\u20092,5\nDisparities in electricity retail regulatory protections will see some \nconsumers approaching energy transition from an uneven footing. Here \nwe examine the spatial organization of regulatory inequities in Australia \nby mapping electricity legal protections for settlements nationwide. \nMultiple logistic regression (n\u2009=\u20092,996) identifies the geographic and \nsocio-demographic characteristics of settlements likely to be underserved \nby regulations to: protect life-support customers, guarantee service levels, \nclarify connection requirements for rooftop solar, require disconnection \nreporting and set clear and independent complaints processes. Assessing \nwhether communities receive fewer than four of five protections, we \nfind that Indigenous communities are 15% more likely to be underserved \nacross multiple metrics and remote communities are 18% more likely to be \nunderserved. These groups overlap. Those communities whose lands are \nrich in resources necessary for energy transition are simultaneously at risk \nof non-recognition of their own energy needs under current regulation, \nrequiring policy remedies for a just transition.\nInternationally there is a movement to achieve a just transition to renew-\nable energy sources, where just transition encompasses broad elements \nof energy justice beyond employment outcomes1,2. Calls for a just tran-\nsition necessarily recognize that new energy systems will be built on \nand potentially reproduce the winners and losers of existing energy \nsystems3. Within current energy systems, groups at the spatial periph-\nery are at high risk of having their energy needs under-recognized and \nprocedurally neglected4,5. Many communities hosting new renewable \nenergy developments, particularly Indigenous communities, face pro-\ncedural injustices in the form of limited access to decision-making \nprocedures for developments on their lands6\u20138. There is a need to bet-\nter understand the spatial and socio-demographic characteristics of \ncommunities facing non-recognition in protections afforded by present \nday electricity retail regulations, wherein non-recognition refers to the \nneeds of certain groups being neglected or ignored4.\nAustralia, home to one of the oldest continuing cultures in the \nworld, is expected to play a key role in energy transition globally9, yet \nthe geographies of disparity in the present day regulations governing \nconsumer electricity retail are largely invisible. Australia is assumed \nto have achieved the goal of universal access to energy for all, with \nan electricity rate of 100% (ref. 10), but this presumed ubiquity belies \npersistent disparity in who experiences energy insecurity and where \nthey reside11\u201319. Aboriginal and Torres Strait Islander (prepay) custom-\ners in Australia\u2019s remote Northern Territory (NT) are more likely to \nexperience \u2018self-disconnection\u2019 during temperature extremes, which \nclimate change only makes more frequent19. As seen during the COVID-\n19 pandemic and recent cost-of-living crises, regulatory difference \nshapes access to financial support for essential home energy services \nsuch as refrigeration and space cooling20\u201322.\nElectricity use in modern societies is critical for many aspects \nof well-being23\u201326, and differences in regulatory protections can have \nsubstantial social impacts including by reinforcing marginalization of \nhistorically oppressed or colonized communities27. Literature map-\nping spatial differences in existing energy protections has begun to \nReceived: 17 July 2023\nAccepted: 22 November 2023\nPublished online: 15 January 2024\n Check for updates\n1School of Regulation and Global Governance (RegNet), Australian National University, Canberra, Australian Capital Territory, Australia. 2Centre for \nAboriginal Economic Policy Research (CAEPR), Australian National University, Canberra, Australian Capital Territory, Australia. 3Crawford School of Public \nPolicy, Australian National University, Canberra, Australian Capital Territory, Australia. 4Melbourne Climate Futures, University of Melbourne, Melbourne, \nVictoria, Australia. 5Tangentyere Council Aboriginal Corporation, Alice Springs, Northern Territory, Australia. \n\u2009e-mail: lee.white@anu.edu.au\n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n93\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nprotections, guaranteed service levels and disconnection reporting \nthat are ubiquitous for residential customers within urban and regional \nareas are often absent in remote settlements. Remote settlements and \nsettlements with majority Indigenous population are respectively 18% \nand 15% more likely to lack comprehensive regulatory and legal protec-\ntions compared with non-remote and non-Indigenous settlements. \nThese findings show that some communities face energy transition \nfrom an uneven footing in Australia and that action is needed to support \na just transition to avoid reproducing or exacerbating non-recognition \nin future energy systems.\nAustralian electricity retail regulation context\nAustralia is a large country (7.7\u2009million\u2009km2 compared with the 8\u2009mil-\nlion\u2009km2 of the contiguous United States), with grid infrastructure \nconcentrated on the east coast and not connected across the whole \ndemonstrate the extent of these disparities, including mapping dif-\nferences in disconnection protections in both the United States and \nEuropean Union28\u201330. Yet, nationwide mapping of regulations across \nmore granular geographies remains underexplored. Interview-based \nwork has identified that peripheral locations in Wales face challenges \naccessing energy services31 and that electricity governance arrange-\nments in Rio de Janeiro are negotiated and permitted to vary based in \npart on perceived commercial risk within different parts of the city32. \nExisting geospatial studies regularly focus on single metrics28, with an \nemerging focus on intersectionality related to: material precarities33, \nenergy and transport insecurity34, internet and energy insecurity35 and \nremedial policy during times of crisis36\u201338.\nOur study identifies settlements with fewer extant legal protec-\ntions for electricity services, mapping those at risk of further exclu-\nsion from the benefits of energy transition. We find that life support \nAustralian\nCapital\nTerritory\nVictoria\nTasmania\nWestern\nAustralia\nNorthern\nTerritory\nSouth\nAustralia\nQueensland\nNew\nSouth Wales\nTransmission lines (11 kV and up)\nPart of the NEM; consumer protections contained in NECF\nIsolated networks; consumer protections contained in NECF \u00a0\nPart of the NEM; consumer protections contained in NECF with unique state derogations for prepay customers\nIsolated networks; consumer protections contained in NECF with unique state derogations for prepay customers\nPart of the NEM; consumer protections contained in state energy laws, including the Energy Retail Code of Practice\nIsolated networks; consumer protections contained in operating licences and local codes\nSome interconnected and isolated networks; consumer protections contained in local codes\nSome interconnected and isolated networks; limited consumer protections contained in local codes\nIsolated networks; no uniform or codified consumer protections for electricity customers\n150\n140\n130\n Longitude (\u00b0 E)\nLongitude (\u00b0 E)\n120\n150\n140\n130\n120\n40\n20\nLatitude (\u00b0 S)\n30\n10\nFig. 1 | Overview of electricity retail regulatory environment and \ntransmission infrastructure by location in Australia as of July 2022. Detail \npertaining to variation in legal protections described in Figs. 2\u20135. Regulatory \nenvironment details are extracted from our data collection. Electricity \ntransmission lines shown on map are from Geoscience Australia under a Creative \nCommons license CC BY 4.0.\n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n94\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\ncountry (Fig. 1). Smaller remote settlements rely on standalone dis-\ntributed electricity networks. In Australia, states and territories (not \nthe Commonwealth) have legal jurisdiction over electricity. All states \nand mainland territories (with the exception of Western Australia (WA) \nand the NT) have opted to enact coordinating legislation, forming what \nis known as the National Electricity Market (NEM). The NEM began \noperating in 1998 through the National Electricity Law (NEL), which \ngoverns market operations in the NEM. The National Energy Customer \nFramework (NECF) including the National Energy Retail Law (NERL) and \nNational Energy Retail Rules is likewise uniform legislation in relation to \nthe retail and distribution of electricity and gas to customers connected \nto the NEM. It provides largely similar protections39 to consumers within \nthose interconnected regions (excepting Victoria and regulatory excep-\ntions; Fig. 1 and Supplementary Note 1). Evolution of electricity retail \nregulation in WA and the NT, alongside regulatory exceptions and the \nexistence of small and isolated networks within NEM states (for exam-\nple, in South Australia (SA); Supplementary Note 2), has given rise to \ndifferent electricity retail regulations across the country (Fig. 1). Not \nall settlements are covered by legislative protections for electricity, \nmost notably in the NT and WA (Fig. 1 and Supplementary Note 3).\nThe NERL recognizes the principles that the supply of energy is \nan essential service for residential customers and that disconnection \nof premises of a hardship customer due to inability to pay energy bills \nshould be a last-resort option, but Australia does not ban disconnec-\ntions for non-payment except for (most) life-support customers and \n(some) moratoria during the COVID-19 pandemic38. Some international \nregulatory environments recognize the essential nature of electricity \nmore strongly (Supplementary Note 4). Permissible payment types in \nAustralia vary by jurisdiction: prepayment metering, where customers \npay for electricity before using it and are disconnected if the meter \nruns out of \u2018credit\u2019, is allowed in some parts of the country, yet is pro-\nhibited in Australia\u2019s most populous states (Supplementary Table 1).  \nInternationally, prepay consumer protections commonly differ from \nthose available to post-pay customers40. Prepay is not a traditional \ncustomer\u2013utility relationship (advance payment contractually resem-\nbling exchange of goods rather than an essential service). It is lightly \nregulated in most countries and has generated controversies associ-\nated with customer well-being41.\nAustralian settlements underserved by electricity \nregulation\nWe reviewed each of the 284 documents recording legal protections \npertaining to 3,047 settlements across Australia as of 1 July 2022, includ-\ning those small settlements with fewer than 200 people. Of these 3,047 \nsettlements, the 51 settlements missing data on relative socio-economic \nadvantage are included in mapping but not the subsequent statistical \nanalyses. Our review indicates that an estimated 5 million Australians \n(approximately 20% of the population) are living in settlements where \nnot all customers are guaranteed protections across the five dimen-\nsions of life support, rooftop solar connection, disconnection report-\ning, guaranteed service levels and clear and independent complaints \nprocesses (Fig. 2 and Supplementary Table 2). Figure 2 summarizes the \nfindings of legal protections reviewed across these five indicators and \nillustrates the compounding disparities; a settlement was considered \nto lack protections if not all customers (both prepay and post-pay) were \nguaranteed that protection.\nWe use multiple logistic regression to examine whether remote \ncommunities and Indigenous communities are statistically more \nlikely to be underserved by electricity regulations. Five dependent \nvariables associated with legislative protections are examined: (1) \nlife-support protections, (2) guaranteed service levels, (3) clear solar \nconnection processes, (4) disconnection reporting requirements \nand (5) complaints process clarity and independence. To give con-\ntext to our regulatory review, we spoke to community and regulatory \norganizations whose remit includes electricity access (12 organizations,  \n32 individuals) who recommended creation of a sixth indicator, \u2018under-\nserved on multiple metrics\u2019, indicating that a settlement received fewer \nthan four of five protections (that is, is not a blue cross in Fig. 2; compris-\ning a population of approximately 290,000 residents). Analyses control \nfor the settlement population and the Index of Relative Socio-economic \nAdvantage and Disadvantage (IRSAD).\nRemote settlements and Indigenous settlements are more likely \nto be underserved on multiple metrics (model 6). Remote settlements \nare 18% more likely (vs urban and regional) to be underserved on mul-\ntiple metrics, and Indigenous settlements are 15% more likely (vs not \nmajority Indigenous) to be underserved on multiple metrics (margins \ncontrast, p\u2009=\u20090.000 for both). Remote settlements and Indigenous set-\ntlements are less likely to have solar connection clarity and less likely \nto have clear and independent complaints processes (models 3 and 5). \nRemote settlements are 38% less likely (vs urban and regional) to have \nsolar connection clarity and 14% less likely to have clear complaints \nprocesses (margins contrast, p\u2009=\u20090.000 for both). Indigenous settle-\nments are 48% less likely to have solar connection clarity and 10% less \nlikely to have complaints process clarity, compared with settlements \nthat are not majority Indigenous (margins contrast, p\u2009=\u20090.000 for both).\nFor three of our dependent variables, we find that being in a \nnon-remote settlement perfectly predicts success (models 1, 2 and 4). \nThat is, all settlements that are urban or regional have legally enforce-\nable protections for all customers regarding life support, guaranteed \nservice levels and disconnection reporting. For these indicators, we \nexamine variation only within remote settlements (n\u2009=\u2009610). Those \nremote settlements where over 80% of the population is Indigenous \nare less likely to have life-support protections, guaranteed service \nlevels and disconnection reporting requirements for all customers \n(models 1, 2 and 4). Compared with remote settlements that are not \nIndigenous, Indigenous settlements are 61% less likely to have life sup-\nport protections, 46% less likely to have guaranteed service levels and \n63% less likely to have disconnection reporting requirements (margins \ncontrast, p\u2009=\u20090.000 for all).\nHigher IRSAD scores (indicating a relative lack of disadvantage and \ngreater advantage in general) are correlated with higher likelihood of \nhaving life-support protections, guaranteed service levels, disconnec-\ntion reporting requirements and clear complaints processes. Higher \nIRSAD scores are likewise correlated with lower likelihood of being \nunderserved overall. Higher population is not associated with any \ndifferences in legal protections.\nThe models in Table 1 interpret settlements as having life-support \nprotections in cases where life-support registration and prepay use \nare mutually exclusive (Fig. 3a and associated text provide additional \ndetail). However, in practice, prepayment customers in remote areas \nmay still face practical challenges associated with registering for life \nsupport and associated payment plan changes. We include an addi-\ntional analysis in Supplementary Table 3 that treats life support and \nprepayment incompatibility as being consistent with an absence of \nprotection. As with our main analyses in Table 1, Indigenous and remote \nsettlements are less likely to have life-support protections and more \nlikely to lack protections across multiple dimensions.\nSettlements with fewer protections for electricity \naccess\nLife-support customers are by definition those who face increased \nrisks of morbidity and mortality when disconnected from electricity. \nDefinitions of life support vary by jurisdiction but uniformly describe \nlife-support requirements in terms of reliance on particular equipment \n(Supplementary Table 4). For 161 remote settlements where there are \nno consumer-focused regulatory frameworks, life-support protec-\ntions are unavailable for both payment types, with potentially severe \nimplications for residents (Fig. 3a). In 412 settlements prepayment is \nincompatible with life support. Three scenarios were identified where \nlife-support customers could use prepayment (Supplementary Table 5  \n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n95\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nand notes), but only one of these scenarios (for 8 SA settlements) pro-\nvides legal protections from disconnection. Mutual incompatibility of \nprepayment and life support may not provide consumer protection: \nduring consultations, one community organization described that \nconsumers could face practical difficulties in switching payment type. \nDocument review indicates that responsibility to register for life sup-\nport and navigate various procedures is allocated to the individual level, \nwith a potentially prohibitive paperwork burden to access protections.\nAs of July 2022, only the state of Victoria provides protections \nfrom disconnection for those experiencing family violence (Fig. 3b \nand Supplementary Table 6). \u2018Family violence\u2019 is expansively defined \nin many Australian jurisdictions to include all aspects of behaviour that \nseek to threaten, coerce, abuse or control a family member such that \nthat person feels fear for theirs, or another family member\u2019s, safety or \nwell-being. As an essential service, electricity access can be exploited as \na form of control42. Aboriginal and Torres Strait Islander women often \nface severe and discriminatory systemic barriers to addressing family \nviolence43. There is a pervasive fear of child removal, linked with the \ncolonial policy history of enforced Indigenous child removals (Stolen \nGeneration) and with disproportionately high rates of Aboriginal \nchildren currently in out-of-home care43. Protections that took effect \nin Victoria from 1 January 2020 require electricity retailers to have a \nfamily violence policy and outline minimum standards of assistance \nfor these vulnerable customers, including the acknowledgement of \nfamily violence as a potential cause of payment difficulty.\nIn Australia, very few settlements receive any protections from \ndisconnection upon non-payment during very hot or very cold tem-\nperatures (Fig. 3c and Supplementary Table 7). Disconnection from \nelectricity during very hot or very cold temperatures can have impacts \non mortality rates; when services are disrupted, exposure to tem-\nperature extremes can amplify risks associated with underlying health \nissues with profound adverse outcomes19,28,44\u201346. Some member states \nwithin the European Union and United States provide exemptions from \ndisconnection during extreme temperatures, with threshold cut-offs \nvarying by jurisdiction28,47,48. Although the NECF in Australia creates a \nframework to secure protections from disconnection during extreme \nweather, only SA has implemented the state-level regulations necessary \nto activate these protections, and as of July 2022 this only protected \non-grid customers who post-pay.\nMany remote settlements do not have guaranteed service levels \n(Fig. 3d and Supplementary Table 8). guaranteed service levels seek \nto compensate eligible customers for unplanned supply interrup-\ntions. Community organizations reported slow utility service response \ntimes following damage to electricity infrastructure in remote loca-\ntions, which could compound the coercive potential of electricity \nsupply disruption. The types of interruption covered by guaranteed \n0 of 5 regulations\n1 of 5 regulations\n2 of 5 regulations\n3 of 5 regulations\n4 of 5 regulations\n5 of 5 regulations \n150\n140\n130\n120\n40\n20\nLatitude (\u00b0 S)\n30\n10\n150\n140\n130\nLongitude (\u00b0 E)\nLongitude (\u00b0 E)\n120\nFig. 2 | Absence of legal protections across multiple dimensions. Considering \nwhether a settlement is underserved across multiple dimensions, compiling \n(1) life-support protections, (2) guaranteed service level, (3) solar connection \nprocess stated in contract, (4) disconnection reporting requirements and  \n(5) complaints process clarity and independence (n\u2009=\u20093,047 settlements).  \nRemote areas shown in grey.\n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n96\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nservice levels commonly do not include supply interruptions caused \nby third-party interference.\nRetailer definitions of hardship universally exclude prepayment \ncustomers (Supplementary Table 9), as hardship is defined in relation \nto either a consumers\u2019 inability to pay bills (which prepay customers do \nnot receive) or to a specified level of accrued debt (which for prepay is \nlimited to small amounts of friendly credit). This exclusion by defini-\ntion of prepay customers from hardship reporting and supports puts \nthese customers at a disadvantage relative to other payment types and \nincreases risk of non-recognition.\nSettlements with constraints to installing \nrooftop solar\nInternationally households wanting to install solar often face barriers \nto navigating the process for grid connection49\u201351. Stakeholder engage-\nments and prior Australian research52 identify these connection pro-\ncesses as an impediment facing remote and Indigenous communities \nattempting to take part in energy transition. In Australia, distribu-\ntors own the poles and wires and are responsible for the connection \nprocess. Although some states and territories have overarching legal \nrequirements for distributors to connect residential solar, these docu-\nments leave distributors with a high-to-moderate degree of discretion \nin permitting individual residential connections to networks (Fig. 4a \nand Supplementary Table 10). In full discretion cases, there is no rel-\nevant regulation that specifies conditions under which distributed \nsolar should be connected; in moderate discretion cases, a distributor \nis required to have a \u2018model standing offer\u2019 or equivalent that estab-\nlishes the conditions under which solar would be connected. Many \nsmaller remote settlements are subject to network constraints and \nnewly offered capacity allocations are oversubscribed in short order. \nSome progress is being made in this regard, for example, in WA where \nthe distributed network service provider Horizon Power has commit-\nted to a policy of no solar refusals by 202553.\nBecause distributors always retain some discretion, to understand \ncustomer ability to enforce a right to connect solar, we review standard \ncontracts that customers would navigate in the process of connecting \nresidential solar to a distributor network (such as model standing \noffers; Fig. 4b and Supplementary Table 11). The conditions under \nwhich prepay customer applications to connect would be approved \nare only clear in the contracts for two settlements. This accords with \nrecent research showing prepay customers are either precluded or face \ngreater barriers when seeking to install rooftop solar52. When distribu-\ntor standard contracts do not set out the conditions under which con-\nsumers could reasonably expect their residential solar connection to be \napproved, this presents a barrier to solar installation49\u201351. Settlements \nwithin the NEM and major networks of the NT have clear contractual \nprocesses for post-paying customers, but prepay customers may face \nchallenges installing residential solar due to lack of clarity. Outside the \nNEM, both prepay and post-pay consumers have limited recourse to \npursue residential solar grid connection in the event of a distributor \nrefusal\u2014due to the lack of a legal basis to connect, lack of standard con-\ntracts with clear parameters for connection or a combination of both.\nSettlements that face weaker reporting \nrequirements\nSubstantial geographic variation is evident in disconnection reporting \nrequirements (Fig. 5a and Supplementary Table 12). Although discon-\nnection reporting in itself does not offer a protection, the lack of report-\ning precludes efforts to secure improved protections for groups facing \nhigh disconnection rates, raising the risk of non- or mis-recognition. \nThe lack of consistent disconnection reporting for prepayment custom-\ners in Queensland (Supplementary Note 5), NT, WA and SA obscures \nthe true level of energy insecurity in these regions26,54.\nComplaints processes are an essential procedure for correcting \nunique errors (via the utility) and systemic inequities (via independent \nprocesses)7,55. Clarity and independence of process ensure consumers \nmay seek redress in the case of failure to provide protections that are \notherwise legally required. Numerous remote communities in the NT, \nWA and SA lack these complaints process protections, while they are \nprovided to Queensland card-operated communities (Fig. 5b, Sup-\nplementary Table 13 and Supplementary Note 5). Those remote com-\nmunities that do lack clear and independent complaints processes are \nthe same communities that lack other regulatory protections across \nthe categories examined. These communities, particularly Indigenous \ncommunities, face challenges in seeking remedy through complaints \nprocesses, such as lack of materials in their own languages and limited \naccess to internet18,56. Procedural injustices occur when certain groups \nhave systematically lesser access to the procedures of institutional \ngovernance and decision-making processes that are relevant to their \nneeds, resulting in marginalization and discriminatory outcomes55,57,58.\nDiscussion\nIn investigating the socio-spatial diversity of electricity retail regulation \nacross 2,996 Australian settlements nationwide, our findings reflect \na confounding albeit commonplace reality: remote communities in \nAustralia are less likely to have comprehensive regulatory protections \nfor access to electricity and the services it provides. In a disconcerting \nmeasure of indifference, remote settlements are 18% more likely to be \nunderserved across multiple metrics. Analyses further highlight the \npossibility that Indigenous peoples, whose lands are among the most \nTable 1 | Multiple logistic regression examining likelihood of remote and Indigenous settlements having protections across \nfive indicators and likelihood of being underserved by protections in multiple dimensions\n(1)\n(2)\n(3)\n(4)\n(5)\n(6)\nHas life-support \nprotections\nHas guaranteed \nservice level\nHas solar \nconnection \nclarity\nHas disconnection \nreporting requirements\nHas complaints \nprocess clarity and \nindependence\nIs underserved on \nmultiple metrics\nRemote\n(all remote)\n(all remote)\n\u22121.82*** (0.11) \n[0.000]\n(all remote)\n\u22124.11*** (0.47) [0.000]\n4.53*** (0.47) \n[0.000]\nOver 80% Indigenous\n\u22123.25*** (0.36) \n[0.000]\n\u22122.45*** (0.33) \n[0.000]\n\u22122.31*** (0.30) \n[0.000]\n\u22123.23*** (0.36) [0.000]\n\u22122.31*** (0.32) [0.000]\n2.98*** (0.34) \n[0.000]\nPopulation (1,000s)\n0.01 (0.09) \n[0.883]\n\u22120.09 (0.13) \n[0.459]\n\u22120.00 (0.00) \n[0.815]\n0.02 (0.08) [0.803]\n\u22120.00 (0.00) [0.664]\n0.00 (0.00) [0.649]\nIRSAD\n0.42** (0.15) \n[0.005]\n0.43*** (0.13) \n[0.001]\n\u22120.00 (0.07) \n[0.949]\n0.48** (0.15) [0.002]\n0.41** (0.13) [0.001]\n\u22120.39** (0.14) \n[0.004]\nPseudo R2\n0.48\n0.36\n0.19\n0.48\n0.59\n0.66\nn\n610\n610\n2,996\n610\n2,996\n2,996\nStandard errors in parentheses. +p\u2009<\u20090.10, *p\u2009<\u20090.05, **p\u2009<\u20090.01, ***p\u2009<\u20090.001. Precise two-tailed p values in square brackets. Adjustments were not made for multiple comparisons.\n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n97\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nimportant contributors to the transition to renewable energy3,59, are \nlikely to be underserved by regulations that would secure their own \nenergy needs. Our analyses find that settlements with over 80% Indig-\nenous share of population are 15% more likely to be underserved across \nmultiple metrics compared with their non-Indigenous neighbours. \nRegulatory review indicates that an estimated 5 million Australians \n(approximately one in five) are living in settlements where not all cus-\ntomers are guaranteed protections for life support, disconnection \nreporting, solar connection clarity, guaranteed service levels and \nindependent complaints processes.\nThese findings contribute to the international debate on just tran-\nsition and energy justice, and the concern that transitioning energy sys-\ntems will perpetuate existing winners and losers3. The concept of just \ntransition brings together aligned notions of environmental justice, \nclimate justice and energy justice to reflect an overarching societal goal \nthat leaves no-one behind in the process of systemic change required \nto respond to the climate crisis1,2. Though electricity regulations have \nlong been viewed as technical, in practice they are a social policy that \ncan have far-reaching impacts27. We find that many remote communi-\nties and Indigenous communities are entering energy transition from \nNeither prepay nor post-pay customers are protected\nNo protection for prepay customers, post-pay is protected\nPrepay prohibited, post-pay protected\nBoth prepay and post-pay customers are protected\nPost-pay is protected, whereas prepay and life support are mutually incompatible\na\nb\nc\nd\n150\n150\n140\n140\n130\n130\nLongitude (\u00b0 E)\nLongitude (\u00b0 E)\n120\n120\n110\n160\n10\n20\nLatitude (\u00b0 S)\nLatitude (\u00b0 S)\n30\n40\n150\n150\n140\n140\n130\n130\n120\n120\n110\n160\n10\n20\n30\n40\n150\n150\n140\n140\n130\n130\n120\n120\n110\n160\n10\n20\n30\n40\n150\n150\n140\n140\n130\nLongitude (\u00b0 E)\nLongitude (\u00b0 E)\n130\n120\n120\n110\n160\n10\n20\n30\n40\nFig. 3 | Mapping legal protections related to disconnection and service levels. a\u2013c, Mapping legal protections that establish additional protections from \ndisconnection for life-support customers (a), those experiencing family violence (b) and during extremely hot or cold temperatures (c). d, Mapping where guaranteed \nservice levels apply (n\u2009=\u20093,047 settlements).\n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n98\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nan uneven footing, lacking commensurate ability to install their own \nrooftop solar, lacking ubiquitous disconnection reporting that makes \nenergy insecurity visible, lacking procedures to redress poor service \nlevels and in some cases even lacking clear protections for life-support \ncustomers. This may lead to entrenched injustices during energy transi-\ntion60. These communities currently underserved by electricity regula-\ntions are experiencing multiple dimensions of energy injustice and as \nsuch are not being engaged in a just transition1,2.\nIndigenous lands hold many of the minerals critical for energy \ntransition59 and host rich renewable energy resources61, yet many of \nthese communities are currently underserved by electricity systems, \nface challenges installing solar52 and face frequent disconnections19. In \nAustralia, lands critical to the nation\u2019s aspirations for becoming a green \nenergy superpower9 are among the worst served by today\u2019s electricity \nretail regulations. This echoes the growing international literature on \nnon-recognition of consumer needs in energy peripheries4,5,31. Distrib-\nuted energy resources have the potential to democratize control of \nenergy systems and restore decision-making power to communities27, \nyet our mapping shows that remote communities are underserved both \nby protections for centralized energy systems and by processes that \nwould support household installation of distributed renewable energy.\nOur findings reflect that communities at the periphery geographi-\ncally and politically often face unique challenges of energy vulnerabil-\nity4,5,31. Peripheral places often hold less economic, social and political \npower5 and may be more likely to have their needs go unrecognized or \neven ignored4. Remote communities in Australia are so designated due \nto distances that people have to travel to receive basic services62, and \nalthough remote places are in no way uniformly disadvantaged, many \nof Australia\u2019s most socio-economically disadvantaged communities are \nin remote locations63. The division of jurisdictional accountabilities \nbetween the Commonwealth and the states and territories for the \nfunding of essential services in Aboriginal and Torres Strait Islander \ncommunities represents a piecemeal approach19 by settler policy-\nmakers that has too often resulted in services that do not adequately \nreflect the needs of communities themselves. There is an urgent need \nfor regulatory frameworks to be developed that better support the \nrights of Australia\u2019s First Peoples to participate in decision-making \nabout present and future energy systems.\nMany underserved communities are on lands that will experience \nsubstantial temperature increases as a result of climate change64,65, \nrequiring an ever greater reliance on electricity for cooling to maintain \nthermal safety and comfort66\u201368. The lack of regulatory protections \nin remote communities intersects with a wide array of energy justice \napplications, including energy poverty (welfare), climate change (fair-\nness and responsibility), energy resources (prosperity) and energy and \ndue process (procedural justice)55. Current Australian protections from \ndisconnection during extreme weather lag those protections granted \nin many parts of the United States and the European Union28,47, with \nmost Australian jurisdictions lacking codified protections from discon-\nnection during extreme heat or cold weather events. There is a need to \nimprove protections for all Australians, but in doing so there is a need \nto ensure that protections do not reproduce existing spatial patterns of \nunderserving remote and Indigenous communities; these communities \nare likely to experience increased extremes in a changing climate64,65.\nHaving mapped the national scale of regulatory difference for elec-\ntricity retail protections, the next analytical step will be to determine \nthe impacts of these disparities on outcomes for human communities, \nsuch as health and well-being. The absence of disconnection reporting \nfor prepay customers has wide-reaching consequences. Australia\u2019s \nClosing the Gap agreement is intended to improve life outcomes for \nAboriginal and Torres Strait Islander people69. Yet, the Commonwealth \nagency charged with monitoring progress on this area has been unable \nto report against essential services (electricity) progress due to lack \nof data70. Reporting of self-disconnections should be mandatory in \n150\n150\n140\n140\n130\nLongitude (\u00b0 E)\nLongitude (\u00b0 E)\n130\n Longitude (\u00b0 E)\n Longitude (\u00b0 E)\n120\n120\n110\n160\n10\n20\nLatitude (\u00b0 S)\n30\n40\nNo for prepay and post-pay\nNo for prepay, yes for post-pay\nYes for post-pay, prepay prohibited\nYes for both prepay and post-pay\nHigh distributor discretion\nModerate distributor discretion\na\nb\n150\n150\n140\n140\n130\n130\n120\n120\n110\n160\n10\n20\n30\n40\nFig. 4 | Mapping legal protections related to rooftop solar connection. a,b, Mapping ability to connect residential solar established by act or regulation (level \nof distributor discretion) (a) and whether the contract with the distributor lays out clear conditions under which the consumer could reasonably expect to have a \nconnection request for solar approved (b) (n\u2009=\u20093,047 settlements).\n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n99\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nall jurisdictions. Further, we recommend that Australian regulatory \nbodies require the number of registered life-support customers by \npayment type in each jurisdiction to be publicly reported; reporting \ncould highlight areas of under-registration, such as in areas where \nprepay and life-support protections are mutually exclusive. Australia\u2019s \nregulatory agencies charged with governing electricity retail protec-\ntions could take this role. Future work could consider the Australian \ngovernance structures that gave rise to the disparities in regulatory \nprotections, including use of prepayment, that we see clearly visual-\nized in our mapping.\nIn mapping the spatial arrangement of regulatory disparity, it is \nessential to acknowledge the legacy of cartography as an instrument \nof governmentality in delineating those territories stolen from Aus-\ntralia\u2019s First Peoples, whose lands they never surrendered. Even the \nmost well-intentioned map-making risks obscuring or misrepresent-\ning procedural, distributive and recognition injustices for specific \ncommunities and/or individuals. Moreover, regulation undergoes \nfrequent iteration, and there were numerous changes proposed, in \ndraft form or in the process of introduction during this review. We \nseek to dispel here any perception of a deficit narrative of Australian \nrurality and acknowledge the efforts of the many individuals, com-\nmunities, advocates, utilities and policymakers engaged in finding \nlocal solutions and alternatives to the challenges identified here. \nMethodologically, we recognize that identifying underserved com-\nmunities with a simple count of regulations provided does not capture \nthe potential for regulations to have differing extents and magni-\ntudes of impact. We opted for a simple count as the most transparent \nindicator of locations facing disadvantage in multiple areas, while \nmindful that this may not fully capture regulation in each practical \napplication. Finally, we note that the situation in Australia is unique \nto this country, its history, demography and geography. Nonetheless, \nwe hope that by identifying patterns of underserved locations and \ndemographics shown in this work, we create the impetus for future \nwork interrogating local situations globally, so as to identify dispari-\nties in current electricity regulations that may reproduce inequalities \nin transitioning systems.\nMethods\nEthics and inclusion statement\nOur research methodology is informed by the principles underpin-\nning ethical Australian Indigenous research outlined in the Australian \nInstitute of Aboriginal and Torres Strait Islander Studies Code of Ethics \nfor Aboriginal and Torres Strait Islander Research71, and our research \nteam is committed to the principles of Indigenous self-determination, \nIndigenous leadership, impact and value, sustainability and account-\nability. V.N.D. is senior Aboriginal researcher at Tangentyere Research \nHub in Mparntwe (Alice Springs) and a visiting Indigenous fellow at the \nAustralian National University (ANU) Centre for Aboriginal Economic \nPolicy Research. M.K. is senior policy manager at Tangentyere Research \nHub and a visiting fellow at the ANU\u2019s Centre for Aboriginal Economic \nPolicy Research.\nThis research both springs from and builds upon efforts by our col-\nlaborators at the Tangentyere Research Hub, starting in 2019. Research \napproach and methods were determined in collaboration with these \nlocal partners. Roles and responsibilities were agreed among col-\nlaborators early in the research, having developed out of our previous \ncollaborations on a related topic19 and included plans to centre the \nperspectives of Indigenous researchers. Capacity building included \nexchanges between researchers in Central Australia and the NT, such \nas time spent by B.R. in Alice Springs and time spent by V.N.D and M.K. \nin Canberra. Local capacity constraints may have circumscribed how \ncomprehensive the analysis could be; however, in no other ways would \nthis research have been prohibited.\nThis research occurs within Australia, and it was conducted with \nethics approval from the ANU\u2019s Research Ethics Committee approval \n2022/443. Despite the desktop, statistical, nature of the research, it was \n150\n150\n140\n140\n130\n Longitude (\u00b0 E)\n130\nLongitude (\u00b0 E)\nLongitude (\u00b0 E)\nLongitude (\u00b0 E)\n120\n120\n110\n160\n10\n20\nLatitude (\u00b0 S)\n30\n40\n150\n150\n140\n140\n130\n130\n120\n120\n110\n160\n10\n20\n30\n40\nNot required for either prepay or post-pay customers\nRequired for post-pay customers but not prepay\nRequired for prepay customers but not post-pay\nRequired for post-pay customers, and prepay is prohibited\nRequired for both post-pay and prepay customers \na\nb\nFig. 5 | Mapping legal requirements for reporting and complaints processes. a,b, Mapping legal requirements to report disconnections (a) and have clear and \nindependent complaints processes (b) (n\u2009=\u20093,047 settlements).\n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n100\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nconducted in accordance with the National Health and Medical Research \nCouncil\u2019s National Statement on Ethical Conduct in Human Research.\nGeography identification\nThe Australian Bureau of Statistics (ABS) Urban Centres and Localities \n(UCL) dataset identifies all settlements in Australia with populations \nof 200 or more people. To capture all possible settlements, including \nsmall remote locations, we developed a custom settlement classifica-\ntion based on the ABS smallest geographical units, Mesh Blocks (MBs), \nwhich cover the whole country and generally contain 30 to 60 dwell-\nings. This process identified a total of 3,089 settlements across all of \nAustralia, compared with the 1,809 settlements captured in the 2021 \nUCL dataset. We identified settlements using several steps:\n\t (1)\t Estimated residential populations (ERPs) at 30 June 2021 are \nimputed for every MB for the total population and the Indigenous \npopulation by progressively downscaling state/territory ERPs us-\ning Census counts tabulated at the SA4, SA3, SA2, SA1 and MB level.\n\t (2)\tAll MBs within a UCL are allocated to that UCL.\n\t (3)\tAll remaining MBs classified by the ABS as being primarily used \nfor residential purposes are grouped into clusters based on \nspatial contiguity.\n\t (4)\tClusters of MBs that are contiguous with a UCL are allocated to \nthat UCL.\n\t (5)\tUnallocated MB clusters are classified based on OpenStreetMap \ndata. Specifically, a place name is allocated to a cluster of MBs \nif the MBs intersect an OpenStreetMap node with a \u2018place\u2019 tag \ncontaining any of the values \u2018city\u2019, \u2018town\u2019, \u2018village\u2019, \u2018hamlet\u2019 or \n\u2018isolated_dwelling\u2019 and with a \u2018name\u2019 tag.\n\t (6)\tUnallocated MB clusters with a total ERP of 20 or less are \nexcluded.\n\t (7)\tUnallocated MB clusters that are within 10\u2009km of a UCL are al-\nlocated to that UCL.\n\t (8)\tUnallocated MB clusters that are within 10\u2009km of a named Open-\nStreetMap place node (as above) are allocated to that place.\n\t (9)\tOutliers (for example, prisons, a fracking field) were manually \nremoved.\n\t(10)\tManual checking against satellite imagery and gazetteers was \nundertaken, especially of those MBs allocated to a settlement \non the basis of distance to the closest named place. Numerous \nplace nodes were added manually in OpenStreetMap based on a \nvisual inspection of interim results.\nThis method treats Indigenous communities and small non- \nIndigenous settlements identically. We caution that this method does \nnot capture all populated small settlements, such as remote Indigenous \nhomelands that are seasonally populated living areas on Traditional \nLands. Some smaller communities are not identified under our geo-\ngraphic methods as discrete settlements, instead being merged with \nlarger nearby settlements. This means that some small settlements \ndo not appear by name, despite having distinct regulatory regimes \nidentifiable in other documentation. For example, Acacia Larrakia and \nKybrook Farm in the NT are not discretely identified, despite having a \ndistinct set of regulation. We note these agglomerated settlements as \nbeing a limitation of our method. Given these limitations, our analyses \nprobably under-represent the extent to which smaller communities are \nunderserved by current regulation.\nRemoteness and socio-demographic variables\nFor each identified settlement, we calculated the latitude and longitude \nof its centroid. The ABS remoteness classifications were then added \nvia spatial join of Remoteness Area 2021 ABS shapefile to geographic \ncentre point of each settlement. These classifications are Major Cities, \nInner Regional, Outer Regional, Remote and Very Remote. We identi-\nfied that in rare cases, the shape of a settlement caused the centre \npoint to fall into a neighbouring remoteness category. Therefore, we \ncross-checked all settlements within 1\u2009km of the nearest remoteness \nboundary against their remoteness category in the ABS\u2019 ArcGIS Online \nMap Viewer with the 2021 remoteness boundaries as a layer.\nWe calculated socio-demographic variables for each settle-\nment based on SA1-level Socio-Economic Indexes for Areas (SEIFA) \nindicators from ABS 2021: IRSAD, Index of Relative Socio-economic \nDisadvantage, Index of Economic Resources and Index of Educa-\ntion and Occupation. Each MB was assigned the SEIFA score of its \nenclosing SA1. Settlement-wide SEIFA scores were then calculated \nas a population-weighted average of its constituent MBs. Statistical \nanalyses use IRSAD as a control variable. This ABS indicator summa-\nrizes information about the economic and social conditions of people \nand households within an area, including both relative advantage and \ndisadvantage indicators. This is used in place of individual indicators \nsuch as income, employment and housing tenure due to the likelihood \nthat those variables are highly collinear.\nRemoteness is coded dichotomously such as that Remote or Very \nRemote settlements are coded 1 whereas Major Cities, Inner Regional \nand Outer Regional are coded 0. Indigenous share of the population is \ncoded as a dichotomous variable where 1 denotes a settlement where \nover 80% of the population is Indigenous, and 0 denotes otherwise. This \nthreshold was chosen to capture discrete Indigenous communities, that \nis, those predominantly Indigenous communities that are established on \nAboriginal land and have historically had housing or infrastructure that \nis managed on a community basis. Pre-coarsening, each settlement\u2019s \ntotal ERP, Indigenous ERP and percentage of a settlement that identify \nas Indigenous were calculated by summing the ERPs associated with \neach MB that comprise the settlement (step 1 in geography identifica-\ntion). These settlements formed the basis of our subsequent regulatory \ncoding and analysis and were manually matched by settlement name.\nNational review of retail legal protections\nData were collected during October 2021 to February 2023 and \nincluded review of 1,159 regulatory documents (284 of which are \nlegal documents used for protection coding). Our review focused on \nconsumer-facing electricity retail regulation (such as the NERL) as \nopposed to energy regulation more broadly (such as the NEL). Where \ncategories of interest for electricity services fell within distributor remit \n(that is, solar connections), we reviewed the appropriate documents \nassociated with electricity regulation (such as the National Electricity \nRules made under the NEL). We mapped 12 categories, four of which \nwere combined into a single indicator (\u2018minimum complaints protec-\ntions\u2019). Data collection was completed during October 2021 to February \n2023 and included review of 284 legal documents to identify protec-\ntions in each settlement. Regulatory environment at the settlement \nlevel was cross-checked with review of over 800 further documents \nto ensure no exceptions were overlooked. Settlements were coded \nbased on their legal protections up to and including 1 July 2022. Regula-\ntion undergoes frequent iteration, and there were numerous pending \nchanges proposed in draft form or in the process of introduction during \nour review (notably in remote WA, though this is unlikely to change \ncommunity status in the short term; Supplementary Note 6).\nLegal protections were mapped for 3,047 of the geographically \nidentified settlements. We excluded 42 settlements from mapping \ndespite meeting geography criteria (Supplementary Table 14). Our \nreview began by identifying those legal documents applicable to each \nsettlement nationally following a two-step process, with some itera-\ntion. We first identified the acts, regulations, rules, determinations \nand orders governing electricity supply to residential customers in \nAustralia, through top-down document identification of statutory \nframeworks governing the electricity industry in each state and  \nterritory\u2014including legislation implementing the NERL and National \nEnergy Retail Rules within Australia\u2019s NEM. We then reviewed core acts \nand associated regulations to identify which conditions or areas were \nexempted, excluded or not explicitly included in relevant coverages. \n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n101\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nSecond, we identified any licences, exemptions, codes and guidelines \nthat form part of the regulatory frameworks; in many cases these docu-\nments identify settlements by name that subsequently facilitated cod-\ning at settlement level. If the regulatory context for a settlement could \nnot be identified through these steps, document searches continued \nwith expansion to other publicly available documents, including gov-\nernment and retailer/distributor websites and annual reports; however, \nthese documents were used only to provide context on regulatory \nenvironment because they do not have legally enforceable status to \nprovide consumer protections.\nStatutory and contractual interpretation\nThe process of coding protections received in each settlement involved \ninterpretation of legislation and contracts, where legal principles \nnecessarily apply, thus we followed norms of legal interpretation in our \ndocument analysis and coding. In all cases, this involved the applica-\ntion of basic principles of statutory and contractual interpretation.\nLegislative documents, such as the NERL, provide clear statu-\ntory protections and are legally enforceable, hence these documents \nformed the basis of our coding. Our coding questions (as specified in \nrespective Supplementary Tables) generally followed the format of \u2018Are \nprotections required by act, regulation, code or licence condition?\u2019, \nwith all coded responses referring to specific provisions of legislative \ninstruments (Data Availability statement to download the full dataset). \nFor legislation, analysis proceeded by examination of \u2018the words, their \ncontext and the purpose of the legislation\u2019 without reliance on extrinsic \nmaterials72. In some cases, a relevant protection arose directly from \none legislative provision; in other cases, it involved the combination \nof primary legislative provisions and associated subordinate legisla-\ntion (for example, extreme weather protections present in the NERL \nbut which are only enlivened by local subordinate legislation in SA).\nLegislative instruments regarding consumer right to connect \nresidential solar give discretion to distributors (Supplementary Table \n10); hence we consulted distributor contracts that customers must \nnavigate when installing solar to determine consumer rights (Sup-\nplementary Table 11). For contracts, analysis concerned the meaning \nof words in written contracts construed \u2018according to the strict, plain, \ncommon meaning of the words themselves\u201973. In some cases, external \ndocuments were referred to directly in the contract and were therefore \n\u2018read in\u2019 and naturally included in analysis, for example, utility hardship \npolicies (Supplementary Table 4). When reviewing standard retailer \ncontracts, we identified the exclusive retailer or retailer of last resort \n(the local area retailer obligated to provide households with an elec-\ntricity contract where another retailer fails) and coded based on their \nregulatory context to represent a putative case. We did not consider \n\u2018market retail contracts\u2019 (under which retailers can offer special plans \nand bundles) when coding legal protections.\nAlongside legal documents (such as legislation and customer \ncontracts), our regulatory review included quasi-regulatory docu-\nments (such as utility policies and utility web pages). Quasi-regulatory \ndocuments were incorporated only where necessary to give context \nto legal documents, such as to identify arrangements in some off-grid \nsettlements that lack legislative transparency, before proceeding to \nexclusion of that settlement (Supplementary Table 14) or where they \nwere referred to in legal documents, such as life-support procedures \nor hardship policies. Quasi-regulatory documents were primarily used \nas a cross-check to ensure that no relevant legal documents had been \noverlooked in searching. Where further disambiguation was required, \ngoverning bodies were contacted directly for clarity (via email or tel-\nephone). A settlement was never coded as receiving protection only on \nthe basis of a quasi-regulatory document, but in some cases these docu-\nments corroborated a lack of definitive protections conferred by legal \ndocuments. In cases where a quasi-regulatory document suggested a \nprotection that could not be clearly confirmed in a legal document, \nthat settlement was coded as lacking protections.\nEngagement with community and regulatory organizations\nIndicators were developed iteratively during three rounds of engage-\nment and consultations with 32 intermediaries from energy, hous-\ning, health and social service organizations operating at national and \nsub-national levels representing a diversity of constituents and loca-\ntions. We engaged with 12 organizations one to three times over the \ncourse of the project in semi-structured 1-h long discussions. Stake-\nholders included: the Northern Territory Council of Social Service, \nthe South Australian Council of Social Service, the Western Australian \nCouncil of Social Service, Original Power, the First Nations Clean Energy \nNetwork, Tangentyere Council Research Hub, Indigenous Consumer \nAssistance Network, Weipa Community Care, Energy Consumers Aus-\ntralia (ECA), Australian Energy Regulator and one other who requested \nanonymity. Before engagement commenced, these organizations all \nreceived a project information sheet and were read a consent form \nscript; options were offered for anonymity, attribution at organiza-\ntional level and attribution at individual level.\nOur initial list of indicators included life-support protections, \nhardship policies, protections from disconnection, redress of elec-\ntricity service or access issues and access to solar; this was based on \nenergy justice concerns identified in prior literature. Review of regu-\nlatory documents prompted re-evaluation of some of these (hard-\nship policies were not sufficiently precise to map), and refinement of \nothers (the only protections related to disconnection that could be \nidentified were related to life support, extreme weather and reporting \nof disconnections; redress of electricity service or access issues was \nrefined to complaints process and independence based on language \nused in regulatory documents). Stakeholder consultation reinforced \nthe necessity of including access to solar, and regulatory review trian-\ngulated the need to focus on distributor contracts to establish this. \nIn conversations with stakeholders, we also identified the need to \ninclude guaranteed service levels and family violence policies in review. \nExcepting these two additions, stakeholders agreed that our initial \nlist and refinements covered the key areas of interest. Stakeholders \nalso repeatedly reinforced the importance of an indicator to visualize \nwhether multiple protections were absent.\nDue to this primary intent of stakeholder engagement to ensure \ncompleteness of indicator selection rather than to comprise a form \nof data collection or a formal mode of analysis, we took detailed min-\nutes of each stakeholder meeting but did not record or transcribe \nour discussions with stakeholders. Where key insights emerged from \nthese engagements, any references included in the manuscript were \ndouble-checked with individuals for accuracy.\nData preparation for mapping and statistical analysis\nThe summarized legislative situation for each settlement was recorded \nby a team member with legal expertise. These summaries are reported \nin Supplementary Tables 1, 5\u20138 and 10\u201313. Detailed categories were \nthen simplified based on agreement between at least two our team \nmembers, with summaries recorded in these Supplementary Tables.\nData underwent two steps of simplification. First, we created the \ncategories used in individual maps (Figs. 3\u20135), that is, neither post-pay \nnor prepay customers are protected (0); no protection for prepay cus-\ntomers, post-pay protected (1); prepayment is prohibited, post-pay pro-\ntected (2) and both post-pay and prepay customers are protected (3).  \nSome maps required additional categories, that is, prepayment and \nlife support are mutually incompatible (4) and procedure required for \nprepay but not post-pay customers (1.5).\nThis was then further aggregated for statistical analyses in Table 1  \nmodels 1\u20135, simplified to a dichotomous yes (1) or no (0), using the \nprinciple of minimum protection available to all customers in the settle-\nment. We aggregated the map codes described in Supplementary Tables \n5, 8 and 11\u201313. Settlements are coded 1 where all customers, that is, both \npost-pay and prepay, receive protection (map codes 2 and 3; life-support \ncode 4 is coded 1 for Table 1 and 0 for Supplementary Table 3).  \n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n102\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nWe coded settlements 0 where not all customers receive protection \n(map codes 0, 1 or 1.5). This approach extended to locations where \nprepay is in theory permitted but not currently in use. In such cases if \nthe protection was not clearly available to (hypothetical future) prepay \ncustomers but was available to post-pay customers, we coded 0.\nWe examined whether the use of prepayment meters is expressly \nprohibited (prepayment prohibited) by legislation, code or licence \nconditions. Supplementary Table 1 describes simplified coding. In \ncases where prepay or equivalent meters were permitted but not cur-\nrently operational, we coded 0 (not prohibited).\nWe examined whether protections from disconnection for life- \nsupport customers are required by legislation, code or licence condi-\ntion in the event of non-payment. Supplementary Table 5 details the \nsimplified codes and map codes assigned to each category. Life-support \ncoding includes a \u201891\u2019 simplified code for prepay customers where life \nsupport and prepay are mutually exclusive. For mapping, this 91 code \nbecame either a 4 (if prepay and life support were mutually exclusive \nbut post-pay is protected) or a 0 (if prepay and life support were mutu-\nally exclusive and post-pay is not protected). In rare cases, contracts \nafforded a legal protection even in the absence of legislative require-\nments. This is the case, for example, for remote communities served \nby Jacana in the NT, where Jacana offers contractual protections for \nlife-support customers equivalent to those for communities con-\nnected to major grid networks. Our mapping of life-support protec-\ntions focuses on those in legislation, code or licence condition, but \nwe note that in certain cases customer protections may also arise at a \n(more changeable) contractual level.\nWe assessed whether there is a legal requirement for the retailer \nto have a family violence policy pursuant to act, regulation or code \nand investigated this separately for prepay and post-pay customers. \nSupplementary Table 6 details the simplified coding and the map code \nassigned to each category.\nWe examined whether the retailer is legally required to provide \nprotections from disconnection for non-payment during an extreme \nweather event pursuant to legislation, code or licence condition. Sup-\nplementary Table 7 details the simplified codes and map codes assigned \nto each category.\nWe examine whether act, code or licence condition establishes \na guaranteed service level scheme, which the distributor is legally \nrequired to adhere to, covering unplanned interruptions in the cus-\ntomer\u2019s electricity supply. Supplementary Table 8 details the simplified \ncodes and map codes assigned to each category.\nLegislation may provide a right to connect solar, but this is always \nat some degree of distributor discretion for technical and economic \nconsiderations. There is no distinction in these regulatory documents \nbetween prepay and post-pay customers. We examine the degree to \nwhich relevant legislation specifies a standard set of conditions under \nwhich the distributor is required to connect residential solar (Supple-\nmentary Table 10; this variable represents an exception to the coding \nscheme). We interpret the absence of a legislative right to connect or \nbroad qualifying parameters (for example, which apply in isolated \nnetworks in Queensland) as \u2018high distributor discretion\u2019. In contrast, \n\u2018moderate discretion\u2019 arises where there is a requirement for a model \nstanding offer (for example, National Electricity Rules, Chapter 5A \nfor NEM interconnected locations) or there are other clear qualifying \nparameters (for example, off-grid Tasmania).\nBecause legislation always allows distributor discretion in enact-\ning customer right to connect solar, we review the connection terms \nin the model standing offer contracts (or equivalent) that a customer \nwould navigate when attempting to establish a solar connection with \nthe distributor (that is, the clarity in connecting solar). These standard \ncontracts represent a point at which the distributor can legally refuse a \nconnection if the residential solar system does not meet their require-\nments (such as by falling outside conditions specified in the model \nstanding offer). Specifically, we consider whether these contracts \narticulate those conditions under which a solar connection (through \nthe associated contract) could be applied for with a reasonable expec-\ntation of success, such as system size, and inverter requirements and \nexport limits. Review assessed whether the contract (such as a model \nstanding offer) that a customer would refer to when connecting solar to \na distribution network had clear eligibility criteria laid out under which \nthe consumer could reasonably expect the distributor to approve a \nconnection request for solar. For example, Essential Energy\u2019s licensed \ndistribution area in New South Wales has a model standing offer for \nbasic (post-pay) connection services, and the standing offer contains \nthe terms and conditions for solar connections\u2014this provides clarity \nand was coded as \u20181\u2019. Supplementary Table 11 describes the simplified \ncoding and map coding for prepay and post-pay customers.\nWe examined whether act, regulation or code legally requires \nthe retailer to report total numbers of customer disconnections \nfor non-payment (disconnection reporting), coding separately for \npost-pay and prepay customers. Supplementary Table 12 reports sim-\nplified coding and map coding.\nFour indicators were combined to understand minimum com-\nplaints protections. We examined the complaints resolution process \nand ombudsman process for both distributors and retailers to deter-\nmine whether there is a requirement by act, regulation, code or licence \ncondition for the retailer and/or distributor to (1) have and publish \ncustomer complaints/dispute resolution procedures and (2) be subject \nto an independent investigation and resolution process in relation to \ncustomer complaints/disputes. These protections were synonymous \nin most, but not all, cases. Given the similarity and close relation of \nthese four indicators, we created a \u2018minimum complaints process\u2019 \nindicator for each settlement that was assigned the lowest value given \nto any of these four component indicators. We consider both pre-\npay and post-pay protections, and Supplementary Table 13 describes \nthe simplified coding and map coding for categories of settlement. \nWhere procedures were not clearly required or where procedures \nwere required but publishing or making the procedures available to \ncustomers was not, settlements were assigned a simplified code of 0. \nIn cases where dispute resolution required retailers\u2019 participation only \nif requested in writing by the regulator and cases with thresholds for \ncustomer inclusion, we assigned a simplified code of 0 due to the high \nbarrier customers may face.\nMost settlements (92%) have four or five of the legal protections \nfor life support, guaranteed service levels, solar connection, discon-\nnection reporting and complaints process clarity and independence. \nFigure 2 shows distribution of these differences. We create an indicator \nfor underserved (in multiple areas) that is coded 1 if settlements have \nzero to three of these protections and 0 if settlements have four to five \nof these protections.\nStatistical analysis\nStata MP 17.0 is used for all statistical analysis. Statistical analysis is \nlimited to those variables for which we can identify differences between \npost-pay and prepay customers and those variables where visible vari-\nation is found during geographic mapping. We thus exclude family vio-\nlence, extreme weather and legislative degree of distributor discretion \nin solar connection from statistical analysis. We examine life-support \nprotections, guaranteed service levels, solar connection clarity, discon-\nnection reporting requirements, clarity and independence of com-\nplaints process and whether a settlement is underserved on multiple \nmetrics. The dataset is largely the same as that used for mapping, with \na further 51 settlements excluded due to lack of key socio-demographic \ndata (IRSAD), for a final sample of n\u2009=\u20092,996 settlements.\nMultiple logistic regression is used to assess the extent to which a \nsettlement being remote or Indigenous (where over 80% of the popula-\ntion identified as Aboriginal and Torres Strait Islanders) is associated \nwith greater likelihood of a settlement lacking each of the five tested pro-\ntections and for the aggregate indicator. Stata estimates equation (1),  \n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n103\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nwhere p is the expected probability that the outcome is present (that \nis, of having a regulatory protection for either life-support protec-\ntions, guaranteed service level, solar connection clarity, disconnec-\ntion reporting requirements, clarity and independence of complaints \nprocess or the aggregated indicator for settlements underserved on \nmultiple metrics); X1 through X4 are distinct independent variables \n(1, remote (dichotomous); 2, majority Indigenous (dichotomous);  \n3, population (continuous) and 4, IRSAD (continuous)); \u03b21 through \u03b24 \nare the regression coefficients associated with corresponding variables \nand \u03b20 is the intercept.\np =\nexp(\u03b20 + \u03b21X1 + \u03b22X2 + \u03b23X3 + \u03b24X4)\n1 + exp(\u03b20 + \u03b21X1 + \u03b22X2 + \u03b23X3 + \u03b24X4)\n(1)\nTesting the variance inflation factor (VIF) for multicollinearity in \nregressions indicates VIF of 0 to 2 for each independent variable within \nour dataset. By most rules of thumb, this is a low VIF\u2014generally, VIF only \nmerits further investigation for variables over 10. Multicollinearity of \nour independent variables is thus unlikely to impact interpretation of \nour model results.\nThe logit command in Stata is used, providing logistic regression \ncoefficients (not odds ratios). The margins (contrast) post-estimation \ncommand is then used to calculate the likelihood of groups lacking \nprotections in comparison to counterparts. Interpretation of these \nlikelihoods is that unconditional on other variables, settlements that \nare remote (Indigenous) are x% more likely to have protections, com-\npared with settlements that are not remote (not Indigenous).\nReporting summary\nFurther information on research design is available in the Nature Port-\nfolio Reporting Summary linked to this article.\nData availability\nThe data underlying this project are available via Figshare at  \nhttps://doi.org/10.6084/m9.figshare.24550585. We make available both \n(1) the customized geographies and their associated socio-demographic \ndata and (2) the full underlying regulatory coding, preserving the origi-\nnal richness before we applied our simplification criteria. The Stata do \nfile is included in the Figshare upload for ease of adapting the data for \nfurther use. Source data are provided with this paper.\nReferences\n1.\t\nHeffron, R. J. & McCauley, D. What is the \u2018just transition\u2019? \nGeoforum 88, 74\u201377 (2018).\n2.\t\nJohansson, V. Just transition as an evolving concept in \ninternational climate law. J. Environ. Law 35, 229\u2013249  \n(2023).\n3.\t\nCarley, S. & Konisky, D. M. The justice and equity implications of \nthe clean energy transition. Nat. Energy 5, 569\u2013577 (2020).\n4.\t\nBouzarovski, S. & Simcock, N. Spatializing energy justice. Energy \nPolicy 107, 640\u2013648 (2017).\n5.\t\nO\u2019Sullivan, K., Golubchikov, O. & Mehmood, A. Uneven energy \ntransitions: understanding continued energy peripheralization in \nrural communities. Energy Policy 138, 111288 (2020).\n6.\t\nZ\u00e1rate-Toledo, E., Pati\u00f1o, R. & Fraga, J. Justice, social exclusion \nand indigenous opposition: a case study of wind energy \ndevelopment on the isthmus of Tehuantepec, Mexico. Energy Res. \nSocial Sci. 54, 1\u201311 (2019).\n7.\t\nWalker, G. Environmental Justice: Concepts, Evidence and Politics \n(Taylor & Francis Group, 2012).\n8.\t\nThorburn, K., O\u2019Neill, L., Hunt, J. & Riley, B. Renewable Energy \nProjects on the Indigenous Estate: Identifying Risks and \nOpportunities of Utility-Scale and Dispersed Models (Centre for \nAboriginal EconomicPolicy Research, Australian National Univ., \n2019); https://doi.org/10.25911/5dbaaa5c47c6a\n9.\t\nBurke, P. J. et al. Contributing to regional decarbonization: \nAustralia\u2019s potential to supply zero-carbon commodities to the \nAsia-Pacific. Energy 248, 123563 (2022).\n10.\t Tracking SDG 7: The Energy Progress Report (IEA, IRENA, UNSD, \nWorld Bank & WHO, 2022); https://trackingsdg7.esmap.org/data/\nfiles/download-documents/sdg7-report2022-full_report.pdf\n11.\t\nDavis, V. N., Lee, W. & Riley, B. Temperature extremes \nexacerbate energy insecurity\u2014Australia needs to better \nsupport remote Indigenous communities to prepare \nnow. Springer Nature Sustainability Community http://\nsustainabilitycommunity.springernature.com/posts/ \ntemperature-extremes-exacerbate-energy-insecurity-australia - \nneeds-to-better-support-remote-indigenous-communities-to- \nprepare-now-3d514d55-d403-4931-9200-44ccef116e4c  \n(2021).\n12.\t Vera-Toscano, E. & Brown, H. Empirical evidence on the incidence \nand persistence of energy poverty in Australia. Aust. Econ. Rev. \n55, 515\u2013529 (2022).\n13.\t Liu, E. & Judd, B. Regional variations in the experiences of energy \npoverty across Australia. Energy Build. 196, 293\u2013298 (2019).\n14.\t Hammerle, M. & Burke, P. J. Solar PV and energy poverty in \nAustralia\u2019s residential sector. Aust. J. Agric. Resour. Econ. 66, \n822\u2013841 (2022).\n15.\t Grealy, L. Enforced commensuration and the bureaucratic \ninvention of household energy insecurity. Aust. Geogr. 54, 155\u2013172 \n(2023).\n16.\t Awaworyi Churchill, S. & Smyth, R. Energy poverty and health: \npanel data evidence from Australia. Energy Econ. 97, 105219 \n(2021).\n17.\t Stuck in the Heat: Lived Experiences of Public Housing Tenants \nin the Kimberley (KCLS, 2022); https://static1.squarespace.com/\nstatic/56aae0e04d088e4dfa68396f/t/6385f2f85f679917d0fb7b\n2e/1669722882386/Stuck+in+the+Heat+2022.pdf\n18.\t Dwyer, A. & Vernes, T. Power Usage in the Bidyadanga Community \nand its Relationship to Community Health and Well-being (Univ. Notre  \nDame Australia, 2016); https://www.notredame.edu.au/__data/\nassets/pdf_file/0011/2333/Report-Power-Usage-Project.pdf\n19.\t Longden, T. et al. Energy insecurity during temperature extremes \nin remote Australia. Nat. Energy 7, 43\u201354 (2022).\n20.\t Cost of Living Report: General Update Tracking Changes in \nthe Cost of Living, Particularly for Northern Territorians on Low \nIncomes and Northern Territorians Experiencing Vulnerability and \nDisadvantage (NTCOSS, 2022).\n21.\t Helping People in Need During a Cost-of-Living Crisis: Findings  \nfrom the Australian Community Sector Survey\u2014ACOSS (ACOSS, \n2022); https://www.acoss.org.au/helping-people-in-need- \nduring-a- cost-of-living-crisis-findings-from-the-australian- \ncommunity-sector-survey/\n22.\t Cassidy, C. \u2018I\u2019m alive, and that\u2019s it\u2019: rising cost of living puts \npressure on people already struggling. Guardian (27 April 2022).\n23.\t Walker, G. & Day, R. Fuel poverty as injustice: integrating \ndistribution, recognition and procedure in the struggle for \naffordable warmth. Energy Policy 49, 69\u201375 (2012).\n24.\t Day, R., Walker, G. & Simcock, N. Conceptualising energy use and \nenergy poverty using a capabilities framework. Energy Policy 93, \n255\u2013264 (2016).\n25.\t Walker, G. & Day, R. Necessary energy uses and a minimum \nstandard of living in the United Kingdom: energy justice or \nescalating expectations? Energy Res. Social Sci. 18, 129\u2013138 \n(2016).\n26.\t Simcock, N., Frankowski, J. & Bouzarovski, S. Rendered invisible: \ninstitutional misrecognition and the reproduction of energy \npoverty. Geoforum 124, 1\u20139 (2021).\n27.\t Chan, G. & Klass, A. B. Regulating for energy justice. N. Y. Univ. \nLaw Rev. 97, 1426\u20131506 (2022).\n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n104\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\n28.\t Flaherty, M., Carley, S. & Konisky, D. M. Electric utility \ndisconnection policy and vulnerable populations. Electr. J. 33, \n106859 (2020).\n29.\t Verclas, K. & Hsieh, E. From utility disconnection to universal \naccess. Electr. J. 31, 1\u20138 (2018).\n30.\t Graff, M., Carley, S., Konisky, D. M. & Memmott, T. Opportunities \nto advance research on energy insecurity. Nat. Energy 8, 550\u2013553 \n(2023).\n31.\t Golubchikov, O. & O\u2019Sullivan, K. Energy periphery: uneven \ndevelopment and the precarious geographies of low-carbon \ntransition. Energy Build 211, 109818 (2020).\n32.\t Pilo\u2019, F. Negotiating networked infrastructural inequalities: \ngovernance, electricity access, and space in Rio de Janeiro. \nEnviron. Plan. C 39, 265\u2013281 (2021).\n33.\t Whittle, H. J. et al. Precarity and health: theorizing the intersection \nof multiple material-need insecurities, stigma, and illness  \namong women in the United States. Social Sci. Med. 245,  \n112683 (2020).\n34.\t Simcock, N. et al. Identifying double energy vulnerability: a \nsystematic and narrative review of groups at-risk of energy and \ntransport poverty in the global north. Energy Res. Social Sci. 82, \n102351 (2021).\n35.\t Chen, C., Greig, J., Nelson, H. & Li, F. When disadvantage collides: \nthe concentrated effects of energy insecurity and internet \nburdens in the United States. Energy Res. Social Sci. 91, 102713 \n(2022).\n36.\t Memmott, T., Carley, S., Graff, M. & Konisky, D. M. Utility \ndisconnection protections and the incidence of energy insecurity \nin the United States. iScience 26, 106244 (2023).\n37.\t Bednar, D. J. & Reames, T. G. Fleeting energy protections: state \nand utility level policy responses to energy poverty in the United \nStates during COVID-19. Energy Res. Social Sci. 99, 103045 \n(2023).\n38.\t Riley, B. et al. Disconnected during disruption: energy insecurity \nof Indigenous Australian prepay customers during the COVID-19 \npandemic. Energy Res. Social Sci. 99, 103049 (2023).\n39.\t Legislation (AEMC, 2023); https://www.aemc.gov.au/regulation/\nlegislation\n40.\t Telles Esteves, G. R., Cyrino Oliveira, F. L., Antunes, C. H. &  \nSouza, R. C. An overview of electricity prepayment experiences \nand the Brazilian new regulatory framework. Renew. Sustain. \nEnergy Rev. 54, 704\u2013722 (2016).\n41.\t Prepayment Rules and Protections\u2014A Call for Evidence\u2014Response \nfrom Citizens Advice to Ofgem (Citizens Advice, 2023); https://\nwww.citizensadvice.org.uk/Global/CitizensAdvice/Energy/\nCitizens%20Advice%20response%20to%20Prepayment%20\nrules%20and%20Protections%20-%20call%20for%20\nevidence%20March%202023.pdf\n42.\t Protecting Customers Affected by Family Violence (AEMC, 2022); \nhttps://www.aemc.gov.au/sites/default/files/2022-09/RRC%20\n0042%20Information%20Sheet_Protecting%20customers%20\naffected%20by%20family%20violence_clean%20copy.pdf\n43.\t Langton, M. et al. Improving Family Violence Legal and Support \nServices for Aboriginal and Torres Strait Islander Women \n(Australia\u2019s National Research Organisation for Women\u2019s Safety \nLimited, 2020); https://anrowsdev.wpenginepowered.com/\nwp-content/uploads/2019/02/AT.19.03_Langton_RR-FVsupport- \nWomen.pdf\n44.\t Barreca, A., Park, R. J. & Stainier, P. High temperatures and \nelectricity disconnections for low-income homes in California. \nNat. Energy 7, 1052\u20131064 (2022).\n45.\t Longden, T. The impact of temperature on mortality across \ndifferent climate zones. Climatic Change 157, 221\u2013242 (2019).\n46.\t Friedman, C. Unsafe temperatures, going without necessities, and \nunpayable bills: energy insecurity of people with disabilities in the \nUnited States during the COVID-19 pandemic. Energy Res. Social \nSci. 92, 102806 (2022).\n47.\t Dobbins, A., Fuso Nerini, F., Deane, P. & Pye, S. Strengthening the \nEU response to energy poverty. Nat. Energy 4, 2\u20135 (2019).\n48.\t Kyprianou, I. et al. Energy poverty policies and measures in 5 EU \ncountries: a comparative study. Energy Build. 196, 46\u201360 (2019).\n49.\t Shrimali, G. & Jenner, S. The impact of state policy on deployment \nand cost of solar photovoltaic technology in the U.S.: a \nsector-specific empirical analysis. Renew. Energy 60, 679\u2013690 \n(2013).\n50.\t Stevens, K. A., Iman, S. & Davis, K. O. The cost of utility discretion \non residential solar requirements. Renew. Sustain. Energy Rev. \n159, 112231 (2022).\n51.\t Carley, S. Distributed generation: an empirical analysis of primary \nmotivators. Energy Policy 37, 1648\u20131659 (2009).\n52.\t Riley, B. et al. Connected: rooftop solar, prepay and reducing \nenergy insecurity in remote Australia. Aust. Geogr. 54, 325\u2013346 \n(2023).\n53.\t Annual Report 2021/22 Results (Horizon Power, 2022); https://\nwww.horizonpower.com.au/about-us/our-performance/\n54.\t Fraser, N. From Redistribution to Recognition?: Dilemmas of justice \nin a \u2018postsocialist\u2019 age. New Soc. Theory Reader https://doi.org/ \n10.4324/9781003060963-30 (2020).\n55.\t Sovacool, B. K. & Dworkin, M. H. Energy justice: conceptual \ninsights and practical applications. Appl. Energy 142, 435\u2013444 \n(2015).\n56.\t Park, S. Digital inequalities in rural Australia: a double jeopardy \nof remoteness and social exclusion. J. Rural Stud. 54, 399\u2013407 \n(2017).\n57.\t McCauley, D. et al. Energy justice in the transition to low carbon \nenergy systems: exploring key themes in interdisciplinary \nresearch. Appl. Energy 233\u2013234, 916\u2013921 (2019).\n58.\t Yenneti, K. & Day, R. Procedural (in)justice in the implementation \nof solar energy: the case of Charanaka solar park, Gujarat, India. \nEnergy Policy 86, 664\u2013673 (2015).\n59.\t Owen, J. R. et al. Energy transition minerals and their intersection \nwith land-connected peoples. Nat. Sustain. 6, 203\u2013211 (2023).\n60.\t Tornel, C. Decolonizing energy justice from the ground up: \npolitical ecology, ontology, and energy landscapes. Prog. Hum. \nGeogr. 47, 43\u201365 (2023).\n61.\t O\u2019Neill, L. et al. Renewable energy development on the \nIndigenous estate: free, prior and informed consent and best \npractice in agreement-making in Australia. Energy Res. Social Sci. \n81, 102252 (2021).\n62.\t Accessibility/Remoteness Index of Australia (ARIA+) (The Australian \nCentre for Housing Research & University of Adelaide, 2021); \nhttps://able.adelaide.edu.au/housing-research/data-gateway/aria\n63.\t Socio-Economic Indexes for Areas (SEIFA), Australia, 2021 \n(Australian Bureau of Statistics, 2023); https://www.\nabs.gov.au/statistics/people/people-and-communities/\nsocio-economic-indexes-areas-seifa-australia/latest-release\n64.\t Mora, C. et al. Global risk of deadly heat. Nat. Clim. Change 7, \n501\u2013506 (2017).\n65.\t Lansbury Hall, N. & Crosby, L. Climate change impacts on health \nin remote indigenous communities in Australia. Int. J. Environ. \nHealth Res. 32, 487\u2013502 (2022).\n66.\t Weeramanthri, T. S., Quilty, S. & Campbell, S. L. Climate, extreme \nheat and human health: risks and lessons for Australia. Med. J. \nAust. 215, 393\u2013395 (2021).\n67.\t Zander, K. K., Shalley, F., Taylor, A., Tan, G. & Dyrting, S. \u2018Run \nair-conditioning all day\u2019: adaptation pathways to increasing heat \nin the Northern Territory of Australia. Sustainable Cities Soc. 74, \n103194 (2021).\n68.\t Quilty, S. et al. The relative value of sociocultural and \ninfrastructural adaptations to heat in a very hot climate in \n\nNature Energy | Volume 9 | January 2024 | 92\u2013105\n105\nArticle\nhttps://doi.org/10.1038/s41560-023-01422-5\nnorthern Australia: a case time series of heat-associated mortality. \nLancet Planet. Health 7, e684\u2013e693 (2023).\n69.\t Objective and Outcomes: Closing the Gap (Department of the \nPrime Minister and Cabinet, 2020); https://www.closingthegap.\ngov.au/national-agreement/national-agreement-closing-the-gap/ \n3-objective-and-outcomes\n70.\t Socioeconomic Outcome Area 9: Aboriginal and Torres Strait \nIslander People Secure Appropriate, Affordable Housing that is \nAligned with their Priorities and Need (Closing the Gap Information \nRepository\u2014Productivity Commission, 2023); https://www.\npc.gov.au/closing-the-gap-data/dashboard/socioeconomic/\noutcome-area9\n71.\t AIATSIS Code of Ethics for Aboriginal and Torres Strait Islander \nResearch (AIATSIS, 2020); https://aiatsis.gov.au/sites/default/\nfiles/2020-10/aiatsis-code-ethics.pdf\n72.\t Middleton, J. Statutory interpretation: mostly common sense? \nMelbourne Univ. Law Rev. 40, 626\u2013656 (2017).\n73.\t Edelman, J. in The World of Maritime and Commercial Law: Essays \nin Honour of Francis Rose (eds Mitchell, C. & Watterson, S.) \n243\u2013258 (Hart Publishing, 2020).\nAcknowledgements\nL.V.W., B.R., L.O., F.M. and S.W. received funding from Energy \nConsumers Australia (ECA) under grant ARFEB22001 as part of its \ngrants process for consumer advocacy projects and research projects \nfor the benefit of consumers of electricity and natural gas. The views \nexpressed in this document do not necessarily reflect the views of \nECA. The research was also supported by the Australian National \nUniversity\u2019s Zero-Carbon Energy for the Asia-Pacific Grand Challenge \nand by Melbourne Climate Futures at the University of Melbourne. \nWe thank, for their continued collaboration and support, the staff at \nthe Tangentyere Research Hub (Mparntwe/Alice Springs), Julalikari \nAboriginal Corporation (Warumungu/Tennant Creek), Original Power \nand the First Nations Clean Energy Network. We also thank those \norganizations that engaged throughout the project, namely the \nNorthern Territory Council of Social Service (NTCOSS), the South \nAustralian Council of Social Service (SACOSS), the Western Australian \nCouncil of Social Service (WACOSS), Original Power, the First Nations \nClean Energy Network (FNCEN), Tangentyere Council Research Hub, \nIndigenous Consumer Assistance Network (ICAN), Weipa Community \nCare, ECA and the Australian Energy Regulator (AER).\nAuthor contributions\nL.V.W., B.R., S.W., L.O., M.K. and V.N.D. contributed to conceptualization \nof the research. We especially note the contributions of M.K. and \nV.N.D. in shaping our understanding of the key issues faced by \nIndigenous communities in the NT. F.M. developed the customized \ngeographies essential to settlement identification and worked with \nS.W. to validate and refine this new methodology. S.W. led data \ncuration, that is, reviewed regulatory documents and established the \ndataset. L.V.W. and S.W. reviewed the regulation dataset to establish \nand implement mapping and analysis protocols. L.V.W., S.W. and \nB.R. conducted three rounds of engagement and consultations with \n32 intermediaries from energy, housing, health and social service \norganizations. L.V.W., S.W., B.R., F.M. and L.O. wrote the initial draft of \nthe paper, and L.V.W., B.R., S.W., F.M., L.O., M.K. and V.N.D. contributed \nto subsequent review and revisions. L.V.W., S.W., B.R., F.M. and L.O. \nacquired funding for the work. L.V.W. conducted statistical analysis \nand geographic information systems (GIS) mapping, and her role \nincluded supervision. L.V.W. and S.W. managed project administration. \nF.M. contributed to GIS mapping.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41560-023-01422-5.\nCorrespondence and requests for materials should be addressed  \nto Lee V. White.\nPeer review information Nature Energy thanks Ciaran O\u2019Faircheallaigh, \nNicola Willand and the other, anonymous, reviewer(s) for their \ncontribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons license, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons license, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons license and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this license, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2024\n\n\n    Scientific Research Findings:", "answer": "We mapped five categories of regulatory protection for household electricity consumers in Australia:\u00a01) life support protections against disconnection;\u00a02) guaranteed minimum service levels;\u00a03) mandated disconnection reporting;\u00a04) complaints process clarity and independence; and\u00a05) clear contractual guidelines for rooftop solar connection. Remote communities are\u00a018% more likely to receive fewer than four of these five protections compared to urban or regional communities. Indigenous communities are\u00a015% more likely to be underserved compared to communities that are not majority Indigenous. These groups overlap. Approximately\u00a01 in\u00a05 Australians live in settlements where not all consumers have all five of the protections examined, while all urban and regional settlements are legally required to protect life support customers, guarantee service levels, and report disconnections. Only\u00a02 of the\u00a0631 settlements where prepayment can operate have clearly outlined conditions for prepay customers to connect rooftop solar.", "id": 11}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 8 | August 2023 | 850\u2013858\n850\nnature energy\nArticle\nhttps://doi.org/10.1038/s41560-023-01298-5\nIncreasing the reach of low-income energy \nprogrammes through behaviourally \ninformed peer referral\nKimberly S. Wolske\u2009\n\u200a\u20091\u2009\n, Annika Todd-Blick\u2009\n\u200a\u20092 & Emma Tome\u2009\n\u200a\u20093\nSubsidized energy assistance programmes are a popular policy tool for \npromoting energy justice, but, like other social benefits programmes, are \noften undersubscribed. To improve uptake, some programmes have turned \nto social influence strategies, such as asking programme participants \nto refer their peers. Here, through a field experiment with California\u2019s \nlow-income solar programme (N\u2009=\u20097,676), we show that referral behaviour \ndepends on how existing participants are approached. Adding behavioural \nscience strategies to a referral reward increases peer referral rates, referral \nquality and ultimately solar adoption. Compared with only reminding \nexisting adopters of a potential US$200 reward for referrals that result in \nadoption, adding an appeal to reciprocity through a non-contingent US$1 \ngift\u2014and further combining this gift with a simplified referral process\u2014\nleads to 2.6\u20135.2 times as many solar contracts. These results highlight the \npotential of behaviourally informed peer referral programmes to accelerate \nequitable access to clean energy.\nTo promote an equitable energy transition1, many governments have \nenacted policies to provide energy technologies at reduced or zero \ncost to low-income households. In Western countries, these policies \noften take the form of subsidies for home retrofits2\u20136, heating-system \nand efficiency upgrades3,4,7,8, and solar panels3,7,9\u2014with some policies, \nsuch as the US Inflation Reduction Act, covering 100% of costs for house-\nholds below certain income thresholds. However, even when goods \nand services are \u2018free\u2019, attracting qualified households can be difficult. \nStudies show that low-income individuals often fail to take advantage \nof the social benefits programmes for which they qualify10, includ-\ning cash11 and food assistance programmes12, and subsidized health-\ncare13,14. Indeed, research on low-income energy programmes finds \nthat participation rates can be quite low1,15, owing to non-economic \nbarriers such as lack of information, high transaction costs1,15, the need \nfor home repair, and distrust of programme providers1,16,17. Eligible \nhouseholds may also question the need for unfamiliar technologies \n(for example, heat pumps) when their homes already provide the same \nservices (that is, heat)18. Further complicating matters, administrators \nof social benefits programmes often have incomplete information \nabout which individuals qualify for services10,19, leading budgets to be \nspent inefficiently on outreach.\nOne strategy to address these challenges is peer referral. We define \npeer referral narrowly to mean when existing programme participants \nprovide names to the programme provider of others who might enrol; \nwe exclude informal word-of-mouth. Compared with other outreach \nstrategies, peer referral can more efficiently find programme partici-\npants20,21: referrers know eligibility requirements and can more readily \nidentify qualified peers. Peer referral also leverages social influence, \na widely recognized driver of energy technology adoption22,23 that \nmay be especially influential among low-to-moderate income (LMI) \nhouseholds16,24. Since individuals often know who referred them, a \npeer\u2019s nomination may signal endorsement for the programme, reduce \ndistrust of the provider and ultimately increase participation. And in \ncontrast to other peer-based strategies (for example, ambassador \nprogrammes25), peer referral programmes may be easier to scale: there \nis no need to identify and train individuals willing to become ongoing \nReceived: 20 January 2023\nAccepted: 7 June 2023\nPublished online: 6 July 2023\n Check for updates\n1Harris School of Public Policy, University of Chicago, Chicago, IL, USA. 2Ernest Orlando Lawrence Berkeley National Laboratory, Berkeley, CA, USA. \n3National Renewable Energy Laboratory, Golden, CO, USA. \n\u2009e-mail: wolske@uchicago.edu\n\nNature Energy | Volume 8 | August 2023 | 850\u2013858\n851\nArticle\nhttps://doi.org/10.1038/s41560-023-01298-5\nto overcome these challenges by streamlining the process for par-\nticipation and making pertinent details more salient and easier to \ncomprehend. Examples include simplifying the application process \nfor college financial aid55, presenting pre-selected retirement savings \nplans to reduce choice overload37, and reformatting notices about tax \nbills56 and student absenteeism57 to focus attention on the most critical \ninformation. Making peer referral easier may be particularly important \nfor subsidized energy programmes since existing participants may \nbe time-constrained and the benefits of referral accrue primarily to \nanother person.\nOur findings show that adding behavioural interventions to a \nreferral programme multiplies the effect of providing financial rewards \nalone. Incorporating an appeal to reciprocity and further simplifying \nthe referral process generated 2.0 and 7.5 times as many referrals. The \ncontrol condition, on average, generated one referral from every 106 \nexisting clients, while the reciprocity treatment led to one referral \nfrom every 52 clients, and reciprocity + simplification produced one \nreferral for every 14 clients. By virtue of generating so many referrals, \nthe reciprocity + simplification treatment led to 5.2 times as many solar \ncontracts at a lower cost per contract than the baseline rewards condi-\ntion. The reciprocity condition generated 2.6 as many contracts as the \ncontrol, while also increasing the average quality of nominations. Taken \ntogether, our results demonstrate that administrators of low-income \nenergy programmes have low-cost, easy-to-implement tools available \nto boost programme outcomes by increasing engagement of existing \nprogramme participants.\nReferral programme design\nWe test reciprocity and simplification in a large, pre-registered ran-\ndomized controlled field experiment. All California clients of the \nnon-profit (N\u2009=\u20097,680) were randomized to receive one of three mailers \nhighlighting the US$200 referral rewards programme (final N\u2009=\u20097,676; \nMethods). The control condition mimicked the non-profit\u2019s standard \napproach: a postcard describing the US$200 contingent reward, with a \ntoll-free number and website to provide referrals (Fig. 1). The reciprocity \ncondition was a letter with the same information and a US$1 bill framed \nas a gift. The reciprocity + simplification condition added a referral \nslip with space for three nominations and a stamped return envelope \n(though recipients could still call or go online). The slip, thus, simpli-\nfied the referral process by reducing the steps involved and making \nsalient what information referrers would need to provide. Eleven days \nafter the initial mailing, all conditions received a reminder postcard, \nwith treatment postcards reminding recipients of the gift they had \nreceived. Given the total population size, we were underpowered to \ninclude a fourth simplification-only condition (that is, referral slip and \nenvelope without US$1).\nWe evaluate each condition on three outcomes: referral behaviour \n(response rate and number of names provided), the quality of refer-\nrals (do individuals nominate people who meet qualification criteria) \nand how many referred individuals sign solar contracts. Moreover, to \ndetermine whether treatments merely shifted the timing of intended \nreferrals sooner, without yielding better outcomes in the aggregate58, \nwe compare results at two timepoints post-intervention: 17\u2009weeks \n(when the referral programme was next marketed) and 9\u2009months. \nAlthough the non-profit sent additional reminders about the refer-\nral programme to subsets of clientele during the 9\u2009month follow-up, \nour randomization procedure ensured that the impact of additional \ncampaigns would be evenly distributed across treatments (Methods).\nReferral behaviour\nFigure 2a shows the response rates by condition. During the first \n17\u2009weeks, 74% of clients in the reciprocity + simplification condition \nreferred by slip rather than by phone or webform, affirming the slip\u2019s \nconvenience. The odds of referring were five times as high in the reci-\nprocity + simplification condition compared with the control (response \nprogramme advocates. Asking existing clients for nominations instead \nof word-of-mouth referrals also removes the onus of convincing peers \nto contact the programme provider.\nDespite its potential benefits, methods to increase peer referral \nfor social benefit programmes have received little attention. The few \nstudies so far test the efficacy of referral incentives for subsidized \nhealth programmes, with mixed results and sometimes limited sam-\nple sizes19,26. Most studies aimed at improving programme uptake \ninstead focus on directly recruiting new participants\u2014overlooking \nthe potential benefits of tapping existing programme beneficiaries \nto reach their peers.\nIn this Article, we test strategies to improve take up of low-income \nsolar programmes by increasing referrals from LMI solar adopters. We \ndo so in partnership with GRID Alternatives, the non-profit organiza-\ntion that administers the California Solar Initiative\u2019s Single-family \nAffordable Solar Homes (SASH) Program. SASH provides rooftop \nphotovoltaic (PV) systems at no cost to qualified LMI homeowners \n(Methods). Though there are debates about how government funding \nshould be spent on solar deployment27,28, programmes like SASH exist \nto address inequities in access to this technology. Most US PV instal-\nlations have been concentrated among moderate-to-high-income \nhouseholds29\u201333, despite analyses showing that LMI housing stock offers \nsubstantial suitable roof space30,34.\nRecognizing that referred households were more likely to go solar \nthan cold contacts35, GRID Alternatives created a referral rewards \nprogramme: existing LMI solar adopters are offered a conditional \nUS$200 incentive for each referred household that installs solar. Such \na programme is in line with the for-profit solar industry, though the \nlatter typically offers higher rewards (US$500\u20132,000).\nRather than offer more expensive rewards\u2014which are often mon-\netarily infeasible for programme administrators\u2014we test whether \nincorporating behavioural insights into the design of referral rewards \nprogrammes can increase peer referral. Behavioural science interven-\ntions\u2014or \u2018nudges\u2019\u2014use insights from psychology and behavioural \neconomics to increase engagement in desirable behaviours. Interven-\ntions often involve simplifying application processes and communica-\ntions11,36,37, altering the motivation to act (without incentives)38,39, or \nemploying strategies that make following through on intentions more \nlikely40,41. Such interventions often produce small to medium changes in \nbehaviour42,43 at low incremental costs, making them potentially more \ncost-effective than traditional financial inducements44. Two types of \nbehavioural intervention may be especially effective at increasing peer \nreferral: appealing to reciprocity and simplification.\nReciprocity is the idea that people feel obligated to act in kind to \nothers\u2019 generosity45\u201349. If someone does us a favour or gives us a gift, \nwe feel the need to return the kindness. Because this norm, which is \ndeeply embedded in social life, feels uncomfortable to violate, it can \nbe leveraged in interventions to increase compliance with a request49. \nIn our case, by including a small unconditional gift with the request \nfor referrals, we sought to invoke a sense of indebtedness that par-\nticipants could repay by referring. While this strategy is commonly \nused to increase charitable giving50\u201352 or pro-social behaviour47, peer \nreferral for energy assistance programmes presents a unique context: \nthe intervention target has already received a large gift (in our study, \na PV system with an estimated installed value of US$16,400\u201327,500). \nReciprocity appeals may consequently remind recipients of what they \nhave already received, thereby increasing the psychological cost of \nnon-compliance\u2014and, thus, referrals.\nThe second strategy we test is simplification, which involves mak-\ning the target behaviour easier11,53. Aspects of a programme that are \ncomplex or make participation cumbersome can create friction that \nimpedes action. Such factors have deterred participation in a variety \nof public benefits programmes11,15 and can be especially detrimental \nto low-income individuals11 who may have scarce mental bandwidth54 \nto decipher programme details. Simplification interventions seek \n\nNature Energy | Volume 8 | August 2023 | 850\u2013858\n852\nArticle\nhttps://doi.org/10.1038/s41560-023-01298-5\nrates of 4.22% versus 0.86%, odds ratio (OR) 95% confidence interval \n(CI) 3.16\u20137.88, P\u2009<\u20090.0001) and three times as high compared with reci-\nprocity alone (4.22% versus 1.45%, OR 95% CI 2.04\u20134.33, P\u2009<\u20090.0001; \nSupplementary Table 2). Clients in the reciprocity condition were  \n1.7 times as likely to refer than those in the control, though the com-\nparison was less precise (OR 95% CI 0.99\u20132.84, P\u2009=\u20090.054). Nine months \nafter the campaign, the differences in response rates across conditions \nnarrowed but were significantly different at P\u2009<\u20090.05 (Fig. 2a and Sup-\nplementary Table 3), indicating that the treatments had lasting effects \nand did not simply shift intended referrals sooner.\nThe reciprocity + simplification condition yielded 181 referred \nnames in the first 17\u2009weeks, 7.5 times more than the control (24 \nnames, incidence rate ratio (IRR) 95% CI 4.66\u201312.19, P\u2009<\u20090.0001) and \n3.7 times more than reciprocity alone (49 names, IRR 95% CI 2.41\u20135.65, \nP\u2009<\u20090.0001), while reciprocity yielded 2 times more than the control \n(IRR 95% CI 1.16\u20133.59, P\u2009=\u20090.013; Supplementary Table 8). Among indi-\nviduals who responded to the campaign, clients in the reciprocity + \nsimplification condition nominated more individuals (1.68 names per \nreferring client) than either the reciprocity or control conditions (1.32 \nnames in reciprocity, IRR 1.27, 95% CI 1.02\u20131.56, P\u2009=\u20090.029; 1.09 names in \ncontrol, IRR 1.54, 95% CI 1.31\u20131.80, P\u2009<\u20090.0001; Supplementary Table 8).\nOne explanation for the number of names generated in the reci-\nprocity + simplification condition is that the slip provided space for \nthree referrals, perhaps unintentionally signalling that nominating \nthree people was the default expectation59,60. By contrast, the webform \nallowed one referral at a time (though it could easily be completed \nagain). Similarly, individuals who phoned in referrals may have done \nso as soon as one name came to mind.\nAt 9\u2009months, the total number of nominations remained signifi-\ncantly different across conditions (Fig. 2b and Supplementary Table \n9), providing additional evidence that the treatments did not merely \nshift nominations sooner. Between 17\u2009weeks and 9\u2009months, reciproc-\nity + simplification continued to outperform the control, yielding 2.2 \nas many additional names (55 versus 25 names, IRR 95% CI 1.30\u20133.71, \nP\u2009=\u20090.003; Supplementary Table 10), only 11 of which came by slip. The \nreciprocity-only condition yielded 37 additional names but could not be \nstatistically distinguished from either group (versus control, P\u2009=\u20090.157; \nversus reciprocity + simplification, P\u2009=\u20090.113).\nQuality of nominations\nOne potential drawback of the slip\u2019s format is that it could lead to \nlower-quality referrals if it encouraged clients to list multiple names \nRefer by:\nShare \nGRID with \nyour \nfriends and \nfamily and \nearn $200\nShare \nGRID with \nyour \nfriends and \nfamily and \nearn $200\nShare \nGRID with \nyour \nfriends and \nfamily and \nearn $200\nReminder of conditional \nUS$200 reward\nReminder of conditional \nUS$200 reward\n+ \nUS$1 non-contingent gift\n\u201cWe have enclosed a small $1 gift to thank \nyou for being part of the GRID community.\u201d\nReminder of conditional \nUS$200 reward\n+ \nUS$1 non-contingent gift\n\u201cWe have enclosed a small $1 gift to thank \nyou for being part of the GRID community.\u201d\n+\nReferral slip and stamped return envelope\nControl\nN = 2,558\nReciprocity\nN = 2,558\nReciprocity + simplification\nN = 2,560\nReferrals:\nYour Name:\nRefer by:\nRefer by:\nPhone\nWeb\nPhone\nWeb\nPhone\nWeb\nMail\nFig. 1 | Illustration of experimental conditions. Clients of the non-profit \nwere randomized to receive one of three mailers about the referral reward \nprogramme. All mailers reminded clients that they could receive a US$200 \nreward for each referral that resulted in a solar installation. The control group \nreceived the standard postcard reminder, which included a toll-free phone \nnumber and website for providing referrals. The reciprocity condition provided \nthe same information in a letter along with a US$1 that was described as a gift. \nThe reciprocity + simplification condition was the same as the reciprocity \ntreatment but added a referral slip and stamped return envelope. Eleven days \nafter the initial mailing, all clients received a postcard reminder about the referral \nprogramme. For reciprocity and reciprocity + simplification, the postcard \nreminded clients of the gift they had received: \u2018We hope you liked the small gift \nwe sent in the mail\u2019.\n\nNature Energy | Volume 8 | August 2023 | 850\u2013858\n853\nArticle\nhttps://doi.org/10.1038/s41560-023-01298-5\nwithout considering who would qualify. To evaluate this, we compared \nthe proportion of referrals per client in each condition who lived in geo-\ngraphic areas eligible for solar subsidies (Methods and Supplementary \nTables 8 and 9). As shown in Fig. 2b, reciprocity + simplification had a \nstatistically equivalent proportion of referrals living in eligible areas \nas the control (69% versus 71% at 17\u2009weeks, P\u2009=\u20090.793; 73% versus 67% \nat 9\u2009months, P\u2009=\u20090.691). In contrast, the reciprocity condition yielded \na significantly higher proportion of eligible nominations at both time-\npoints, representing a 23\u201327% improvement over the control at 17\u2009weeks \nand 9\u2009months, respectively, and a 28% to 22% improvement over reci-\nprocity + simplification. We note, however, that because reciproc-\nity + simplification generated far more nominations in total, it led to  \n7.3 times more qualified nominations than the control at 17\u2009weeks  \n(124 versus 17; IRR 95% CI 4.17\u201312.74, P\u2009<\u20090.0001) and three times as many \nthan reciprocity (124 versus 42; IRR 95% CI 1.89\u20134.61, P\u2009<\u20090.0001). Simi-\nlarly, reciprocity alone yielded 2.5 times as many qualified nominations \nas the control (95% CI 1.32\u20134.61, P\u2009=\u20090.005). The differences between \nconditions largely persisted at 9\u2009months (Supplementary Table 9).\nNew solar contracts\nWe also examined the number of signed solar contracts resulting from \neach condition. For referrals generated within 17\u2009weeks (Fig. 2d), reci-\nprocity + simplification led to 26 contracts: 5.2 times the control (IRR \n95% CI 1.95\u201313.8, P\u2009<\u20090.001) and 2.0 times reciprocity alone (IRR 95% \nCI 0.94\u20134.27, P\u2009=\u20090.074; Supplementary Table 8). The reciprocity con-\ndition resulted in 2.6 times more contracts than the control, though \nthe comparison was less precise (IRR 95% CI 0.89\u20137.60, P\u2009=\u20090.081). At \n9\u2009months, the differences between conditions were all different at \nP\u2009<\u20090.05 (Supplementary Table 9).\nUsing back-of-the-envelope calculations (Supplementary  \nTable 6), we estimate that the reciprocity + simplification condi-\ntion, though more expensive in terms of material costs and staff  \ntime needed to screen referrals, was less expensive per resulting  \ncontract than the control, saving 17.8% at 17\u2009weeks (US$522 ver-\nsus $635). Reciprocity alone cost 2.5% more per contract than the  \ncontrol (US$651 versus US$635), but generated more than twice as \nmany contracts.\n0.86%\n1.45%\n4.22%\nP < 0.0001\nP < 0.0001\nP = 0.054\n1.72%\n2.54%\n5.74%\nP < 0.0001\nP = 0.044\nP < 0.0001\nReciprocity\nReciprocity +\nsimplification\n17 weeks\nReciprocity\nReciprocity +\nsimplification\n9 months\nControl\nControl\n1\n2\n3\n4\n5\n6\n7\nResponse rate (%)\n(71%)\n(86%)t\n(69%)t\n24\n49\n181\nP = 0.013\nP < 0.0001\nP < 0.0001\n(67%)\n(90%)c,t\n(73%)t\n49\n86\n236\nP < 0.0001\nP = 0.008\nP < 0.0001\nReciprocity\nReciprocity +\nsimplification\n17 weeks\nReciprocity\nReciprocity +\nsimplification\n9 months\nControl\nControl\n0\n100\n200\n300\nNumber of nominations\n(% in qualifying area)\n0\n50\n100\n150\n200\n250\n1 October 2018\n1 January 2019\n1 April 2019\n1 July 2019\nReferral date\nControl\nReciprocity\nReciprocity + \nsimplification\nSubset of \nR+S that \nused slip\nReciprocity\nReciprocity +\nsimplification\n17 weeks\nReciprocity\nReciprocity +\nsimplification\n9 months\nControl\nControl\nP = 0.081\nP = 0.074\nP < 0.001\nP = 0.024\nP = 0.034\nP < 0.0001\n5\n13\n26\n12\n26\n49\nNumber of solar contracts \nCumulative number of nominations\na\nb\nc\nd\n60\n40\n20\n0\nControl\nReciprocity\nReciprocity + simplification\nFig. 2 | Effects of mailer campaigns on response rate, number and quality of \nreferrals made, and the number of resulting solar contracts. Comparison of \nexperimental conditions (control N\u2009=\u20092,558; reciprocity N\u2009=\u20092,558; reciprocity  \n+ simplification N\u2009=\u20092,560) at 17\u2009weeks and 9\u2009months after mailers were sent. \na, The response rate per condition, which is the percentage of people who \nresponded to the mailer by providing at least one nomination. b, The overall \nnumber of nominations per condition and, within that, the percentage of  \nthose nominations that lived in qualified areas for subsidized solar (in darker \nshading). c, The cumulative number of nominations in each treatment over  \ntime, and, for reciprocity + simplification, which referrals came in by slip.  \nd, The number of signed solar contracts generated from referrals during each \nperiod. Statistical specifications and test statistics were: logistic regression \nwith Firth\u2019s penalization to correct for rare events (\u2018Firth\u2019s logistic\u2019) and pairwise \ncomparisons with z test statistics (a); Poisson count models to test (i) differences \nbetween the number of nominations, and (ii) differences between the proportion \nof those nominations that lived in a qualified area, with z tests for pairwise \ncomparisons (b); Poisson count models to test the difference in resulting solar \ncontracts, and pairwise comparisons using z statistics (d). For Poisson models, \nstandard errors were estimated using the robust Huber/White/sandwich method. \nAll tests were two-sided, and no adjustment for multiple comparisons was made. \nFor b, brackets show whether the count of nominations differed by treatment; for \nthe percentage of nominations living in qualified areas, superscript c indicates \nreciprocity is significantly different from the control at P\u2009<\u20090.05 and superscript \nt indicates reciprocity is significantly different from reciprocity + simplification \nat P\u2009<\u20090.05 during the specified time periods. Supplementary Tables 2 and \n3 report test statistics, CIs and effect sizes for comparisons shown in a, and \nSupplementary Tables 8 and 9 report the same for b and d.\n\nNature Energy | Volume 8 | August 2023 | 850\u2013858\n854\nArticle\nhttps://doi.org/10.1038/s41560-023-01298-5\nPercentage of first-time referrers\nIn a non-pre-registered analysis, we examined heterogeneity with \nrespect to whether clients had previously referred (N\u2009=\u2009796) or not \n(N\u2009=\u20096,875). This type of heterogeneity may be of interest to programme \nmanagers looking to engage more clients in peer referral or to trim costs \nby selectively targeting treatments. In all conditions, previous referrers \nwere more likely to refer than clients who had never referred (Fig. 3 and \nSupplementary Table 7). Among previous referrers, only reciprocity \n+ simplification led to a significantly greater percentage of responses \nthan the control condition (OR 2.76, 95% CI 1.14\u20136.71, P\u2009=\u20090.025). Among \nindividuals who had never referred before, the odds of referring were \nnearly sixfold higher in the reciprocity + simplification treatment \ncompared with the control (OR 5.92, 95% CI 3.41\u201310.28, P\u2009<\u20090.0001), \nand 1.9 times higher in reciprocity than the control, though the latter \ncomparison was less precise (OR 1.87, 95% CI 0.99\u20133.51, P\u2009=\u20090.052).\nDiscussion and conclusions\nWe demonstrate a cost-effective strategy to increase take up of energy \nassistance programmes through behaviourally informed peer referral \nprogrammes. Using low-income rooftop solar as our study context, we \nshow that adding low-touch interventions to a referral rewards pro-\ngramme increases its effectiveness several fold within a short time span, \nultimately leading to more LMI solar installations. Relative to simply \nreminding clients of referral rewards, adding an appeal to reciprocity \nand combining that appeal with simplification generated, respectively, \n1.7 and 5 times the response rate (P\u2009=\u20090.054 and P\u2009<\u20090.0001); 2.0 and \n7.5 times as many referrals (P\u2009=\u20090.013 and P\u2009<\u20090.0001); and 2.6 and 5.2 \ntimes as many solar contracts (P\u2009=\u20090.081 and P\u2009<\u20090.001) during the first \n17\u2009weeks following the campaign. Furthermore, reciprocity + simpli-\nfication generated nearly six times as many referrals from individuals \nwho had never referred before compared to the non-profit\u2019s standard \npractice. The reciprocity + simplification treatment achieved the above \nresults at less cost per solar contract than the control, while reciprocity \nalone cost only marginally more.\nThe effects of reciprocity + simplification over time shed light on \nwhat impedes peer referral. The large, immediate effect of the referral \nslip suggests that having to phone in referrals or visit a website cre-\nates friction. The effect of reciprocity + simplification was also long \nlasting, with referrals continuing to come in at a higher rate than the \ncontrol, even months after receiving the dollar and slip. As fewer slips \nwere returned over time, this finding may indicate that the slips, by \nlisting what information was needed to nominate someone, reduced \nuncertainty about what referring entails. This could explain why the \nreciprocity + simplification treatment was most effective at getting \nfirst-time referrers.\nOur results further reveal potential trade-offs in terms of the quan-\ntity and quality of referrals generated. Compared with only sending \nreminders of the referral rewards programme, adding an appeal to \nreciprocity increased the number of referrals but also the propor-\ntion who were likely to qualify for solar subsidies. The salience of the \ndollar along with language thanking recipients for being part of the \nnon-profit\u2019s solar community may have encouraged recipients to think \nmore deeply about who among their peers might live in qualified areas. \nFurther adding a referral slip and stamped return envelope counter-\nacted this effect, as the proportion of quality referrals was lower but \nequivalent to the control. The ease of referring by slip, however, more \nthan compensated for this dip in quality by generating considerably \nmore nominations and ultimately more installations.\nAdditional research is needed to understand the difference in \nreferral quality between treatments. By providing three spaces to list \nnominations, the slips may have focused clients on quantity over qual-\nity. The return stamped envelope may have also encouraged clients to \nsubmit referrals quickly, without fully considering who among their \npeers lived in qualifying areas. Future work could examine whether \n2.59%\n3.49%\n6.72%\n0.66%\n1.22%\n3.76%\nP = 0.052 P < 0.0001\nP < 0.0001 \nP < 0.0001 \nP < 0.0001 \nP = 0.020\nP = 0.105\nP = 0.025\n5.56%\n5.81%\n11.19%\n1.27%\n2.17%\n4.94%\nP = 0.032\nP = 0.018\n17 weeks\n9 months\n15\n10\n5\n0\n5\n10\nResponse rate among those\nwho had referred before\n(dark shading)\nResponse rate among those \nwho had never referred before\n(light shading)\nControl\nReciprocity\nReciprocity +\nsimplification\nP = 0.837\nP = 0.483\nFig. 3 | Response rates, by condition, among clients who had never \nreferred before and among those who had referred at least once before the \nexperiment. This figure shows the efficacy of each experimental condition on \nresponse rate among existing solar adopters who had never referred before \n(\u2018never referred\u2019, in lighter shading; N\u2009=\u20096,875) and among those who had referred \nat least once before the experiment (\u2018referred before\u2019, in darker shading; N\u2009=\u2009796). \nFive referrers in the reciprocity + simplification treatment were excluded from \nthis analysis as they returned their referrals slips anonymously and could not be \ncategorized. The response rate is calculated within each group (that is, among \nindividuals who had never referred before, what percentage responded to the \nmailer campaign). Logistic regressions with Firth\u2019s penalization to correct for \nbias in estimating rare events were used to test the differences between response \nrates with z test statistics (two-tailed tests). No adjustments were made for \nmultiple comparisons. For test statistics and ORs, see Supplementary Table 7.\n\nNature Energy | Volume 8 | August 2023 | 850\u2013858\n855\nArticle\nhttps://doi.org/10.1038/s41560-023-01298-5\nmaking eligibility criteria salient could offset these effects. This may \nbe particularly beneficial when programmes, like this one, have com-\nplex geographical restrictions that are probably unfamiliar to most \npeople (for example, disadvantaged community designations and \nutility service areas).\nThe quality versus quantity trade-offs we observed have practical \nimplications for programme managers depending on their goals and \nresource constraints. If the goal is to attract as many referrals as pos-\nsible or to attract referrals quickly (for example, because funding will \nexpire), reciprocity + simplification offers the most cost-effective way \nto do so. The drawback is that it requires a higher initial outlay of funds. \nIf the goal is to provide the highest impact within a constrained budget, \nreciprocity alone leads to the greatest proportion of high-quality refer-\nrals at lower upfront costs. Another consideration is target audience. \nIf programme managers seek to attract new referrers, reciprocity + \nsimplification could be used to selectively target these clients. Since \nthe slip seems to improve knowledge of what referring entails, this may \nbe a wise investment to increase the efficacy of less expensive referral \ncampaigns in the future.\nAnecdotal evidence suggests that the US$1 gift was effective at \ngenerating referrals, in part because it reminded recipients of the \nmuch larger gift they had already received: a solar PV system. A hand-\nful of clients returned the dollar with their completed referral slips, a \ngesture that prior work suggests is guilt motivated61. If having received \na prior gift augments the effect of a reciprocity appeal, then research is \nneeded to understand to what extent the value of that prior gift matters. \nAlthough other efficiency improvements may have a smaller economic \nvalue than a PV system, the quality-of-life improvements they bring \nmay have similar or greater psychological value. For these reasons, \nwe expect our results to generalize to energy subsidy programmes \nseen globally that provide low-income homeowners subsidies for \nhigh-value goods like home retrofits2\u20136, heating equipment3\u20135,7\u20139 and \nefficiency upgrades3,5,7,9. Like the solar programme we studied, such \nprogrammes often involve high interaction between the programme \nprovider and participating households, which might strengthen the \nfeeling of obligation to comply with the reciprocity appeal.\nOur findings invite future research about whether the interven-\ntions can be used repeatedly. Both strategies may work best as one-off \nor infrequent interventions, particularly to attract first-time referrers. \nProviding referrals slips too often may diminish the salience of the \nrequest or degrade the quality of nominations. Likewise, repeated  \nuse of the US$1 gift may provoke resentment if clients start to perceive \nthe gesture as manipulative62. Both hypotheses could be tested in future \nreferral programmes. Research is also needed to understand the effect \nof the slip, absent the gift, and how much of the observed effects are \ndue to the referral reward, which was common to all conditions.\nWe see these questions as opportunities to continue exploring \nmore efficacious ways of expanding the reach of low-income energy \nprogrammes. As more policies like the US Inflation Reduction Act, UK \nHelp to Heat scheme and others come online to address inequities \nin access to clean energy technologies, administrators will face the \nchallenge of efficiently implementing these programmes to maximize \ntheir benefits. Growing evidence points to the value of engaging social \nnetworks16,63 to persuade hard-to-reach populations of a programme\u2019s \nbenefits. Our findings demonstrate that how existing programme \nparticipants are engaged in these efforts matters. Relying on financial \nrewards alone leaves many peer referrals\u2014and subsequently LMI solar \ninstallations\u2014unrealized. Complementing rewards with programme \nsimplification and an appeal to reciprocity seems to motivate existing \nLMI adopters to pay it forward\u2014resulting in five times as many solar \ninstallations. The results further show that reciprocity and simplifica-\ntion have trade-offs in terms of the timing, quality and relative cost of \nthe referrals generated\u2014suggesting that managers of energy assistance \nprogrammes have flexibility for improving peer referral depending on \ntheir objectives and constraints.\nMethods\nDesign and sample\nThe experiment was implemented in September 2018. Participants \nincluded the entire population of California LMI homeowners who \nhad adopted solar with the non-profit GRID Alternatives, had inter-\nconnected their solar system between 2004 and 1 June 2018, and for \nwhom GRID had a valid mailing address, N\u2009=\u20097,680. The vast majority \nof these customers qualified for California\u2019s Single-family Affordable \nSolar Homes (SASH) programme, of which GRID Alternatives is the \nprogramme administrator. SASH provides incentives at US$3.00\u2009W\u22121 \nto homeowners with household incomes below 80% of the area median \nincome (AMI), who live in a home defined as affordable housing by \nCalifornia Public Utilities Code 2852, and who receive electrical service \nfrom Pacific Gas & Electric, Southern California Edison or San Diego Gas \n& Electric. Applicant roofs are also screened for solar suitability, and \nsolar installations must meet a minimum performance requirement. \nFor our study population, 32% of the population had a household \nincome below 30% AMI, 27.4% were between 30% and 50% AMI, 38.5% \nwere between 50% and 80% AMI, and 2.1% were missing income data. \nWhere SASH incentives are insufficient to cover the entire cost of a \nPV system and installation, GRID sources supplementary funding to \nprovide the PV system at no cost.\nIntervention\nWe randomly assigned the 7,680 clients to three groups: control \n(N\u2009=\u20092,559), reciprocity (N\u2009=\u20092,560) and reciprocity + simplification \n(N\u2009=\u20092,561). The control condition mimicked GRID\u2019s standard proce-\ndure of sending clients occasional postcards to remind them that they \ncould earn US$200 for successfully referring someone else to GRID. As \nis standard, the postcard included a toll-free number as well as a web \naddress where clients could fill out referral information. The reciprocity \ncondition received the same information but in letter form along with \na US$1 bill. The dollar was framed as a gift: \u2018We have enclosed a small \n$1 gift to thank you for being part of the GRID community. Thanks in \nadvance for your referrals!\u2019 The reciprocity + simplification group \nreceived the US$1 gift along with a referral slip and first class-stamped \nreturn envelope. Eleven days later, each group received a reminder \npostcard; for the treatment conditions, the postcard reminded recipi-\nents of the gift they had received: \u2018We hope you liked the small gift we \nsent in the mail. Thank you for your referrals!\u2019 Materials were provided \nin English or Spanish depending on the known language preference of \nthe household. The study was approved for exemption by the Institu-\ntional Review Board at Lawrence Berkeley National Laboratory (Human \nSubjects Committee number 381H25JE23). As the experimental treat-\nments were in line with standard operating practices of the non-profit \npartner, informed consent was not required. Our methods and analysis \nplan were pre-registered with the American Economic Association\u2019s \nRCT Registry (https://www.socialscienceregistry.org/trials/4114).\nRandomization procedure\nWe used a stratified complete random assignment method that ensures \na balanced panel: that is, each condition had an equal number of house-\nholds, and there were no statistically significant differences between \nthe groups on pre-treatment variables that might influence referral \nbehaviour and quality. We used the R package randomizr along with \na \u2018max\u2013min\u2019 technique. Rather than stratifying the randomization \non the basis of all pre-treatment variables, this technique selects a \nrandom seed to perform an unstratified randomization on the study \npopulation, calculates the P value for each pairwise comparison of the \nthree conditions for each pre-treatment stratification variable, and \nthen determines the minimum of these P values (that is, the one with \nthe lowest P value, showing the biggest statistical difference). This is \nthen repeated 100,000 times (where each of the 100,000 randomiza-\ntions results in one minimum P value). Then, the seed that generated \nthe randomization with the highest minimum P value (that is, the \n\nNature Energy | Volume 8 | August 2023 | 850\u2013858\n856\nArticle\nhttps://doi.org/10.1038/s41560-023-01298-5\nmaximum of the minimum P values) is chosen as the final seed for \ncreating random groups64.\nPre-treatment stratification variables included the client\u2019s geo-\ngraphic region; race/ethnicity; age (whether clients were above 62 years \nold); primary spoken language (English and/or Spanish, or some other \nlanguage); whether they had given previous referrals; the date their \nPV system was interconnected; local population density; qualifying \nincome; and the CalEnviroScreen percentile for the tract where they \nlived. This latter score is a cumulative impacts indicator of the extent \nto which a community is socio-economically and environmentally \ndisadvantaged65. Tracts with scores in the top 25% are termed \u2018disadvan-\ntaged\u2019 and qualify for a new solar incentive through the Disadvantaged \nCommunities Single-family Affordable Solar Homes (DAC-SASH) pro-\ngramme (a more geographically restricted version of the original SASH \nprogramme). A higher CalEnviroScreen score served as a proxy for how \nlikely nearby friends and neighbours were to qualify for DAC-SASH. \nWe also stratified on whether clients had an email address to account \nfor their comfort using online technology in the referral process. As \nshown in Supplementary Table 1, there were no statistically signifi-\ncant differences between the groups on any pairwise comparison of \npre-treatment stratification variables.\nThis randomization method enabled us to continue comparing \ntreatment conditions after 17\u2009weeks, when the non-profit resumed \nstandard marketing of its referral programme with subsets of its cli-\nentele. Geographic areas were targeted on the basis of whether they \nqualified for the DAC-SASH programme. Because our experiment \nrandomized at the level of households, and stratified at a geographic \nregion level, the regions had a roughly equal mix of households exposed \nto either the control, reciprocity, or reciprocity + simplification treat-\nments. As a result, any additional marketing material sent by the \nnon-profit would have been equally distributed to the three condi-\ntions. Moreover, treated households were geographically dispersed, \nminimizing risk of between-treatment contamination (for further \ndiscussion, see Supplementary Note 3).\nData and exclusion criteria\nGRID Alternatives tracks every point of contact with existing and poten-\ntial clients (including referrals) in a Salesforce database. We analysed \na de-identified version of data extracted from this database. Typically, \nonce referred to the non-profit, individuals are contacted to determine \nwhether their household meets the qualification criteria for subsi-\ndized solar (for example, they own their home, meet household income \nrequirements, have a good roof and live in a qualifying area). Unfortu-\nnately for our study, some individuals were never screened; the dra-\nmatic increase in referred names from the reciprocity + simplification \ncondition outpaced the non-profit\u2019s ability to process them as well as \navailable funding for subsidized PV. Of the nominations that were not \nscreened (94 during the first 17\u2009weeks; 122 by 9\u2009months), most came \nfrom the reciprocity + simplification condition (86.2% of non-screened \nat 17\u2009weeks, 77.9% at 9\u2009months; Supplementary Table 5). As this issue was \nonly discovered at the start of the coronavirus pandemic, the non-profit \nwas unable to recontact these individuals. This forced us to deviate from \nour analysis plan and proxy for quality by examining whether nominated \nindividuals lived in areas that qualify for low-income solar subsidies \nthrough SASH and DAC-SASH. See Supplementary Note 1.\nAfter consultation with GRID Alternatives, we excluded \nsuper-referrers from analysis. These were individuals known to the \nnon-profit to be well connected in the community with a history of \nmaking many referrals at once. We defined super-referrers as those who \nhad previously given 20 or more referrals (regardless of whether they \nreferred in the present study). This led to excluding three clients, one \nin each condition. Supplementary Note 2 provides further detail and \nshows that the results are robust even if the definition of super-referrer \nis lowered to 8 or more previous referrals (the next highest number of \npast referrals given by any person in our experiment).\nThere were also two returned slips (with six nominations total) \nfrom the reciprocity + simplification treatment that, through clerical \nerror, were not entered into Salesforce when they were received. As \nwe are unable to pinpoint when these slips came in, we have excluded \nthem from our analyses to avoid distorting conclusions we might draw \nabout the temporal effects of the treatment. With these slips included, \nthe total response rate for the reciprocity + simplification treatment, \nduring the 9\u2009months for which we have data, was 5.82%, resulting in \n242 nominated individuals; the total number of signed solar contracts \nremained unchanged.\nFinally, we excluded one outlier from the reciprocity condi-\ntion. This individual provided one referral during the first 17\u2009weeks, \nwhich resulted in a solar contract, and an additional 19 nominations \n4\u2009months later over the course of 3\u2009days. Of these 19 nominations,  \n18 resulted in solar contracts. We considered it improbable that such \na large within-subject effect would be due to the reciprocity treat-\nment, 7\u2009months after exposure. For comparison, the median number \nof nominations over the 9\u2009month period (conditional on referring)  \nwas one.\nAnalysis\nData were analysed with Stata 17 according to our pre-registration plan \nusing basic ordinary least squares (OLS) specifications (for analyses of \nthe number of referrals) and logit specifications (for the response rate \nanalysis of whether a client made any referrals), with two-tailed tests. \nFull regression results are in Supplementary Tables 2 and 3. The only \npredictors in the models are treatment conditions, except for the het-\nerogeneity analysis which also accounts for whether a client had given \nreferrals before the experiment (Supplementary Table 7). The dataset \ncontained one observation per treated client; each client had variables \nindicating the number of nominations, and the regressions therefore \nrepresent standard errors clustered by client. For OLS specifications, \nstandard errors were estimated using the robust Huber/White/sand-\nwich method. For the logit specification, we used Firth\u2019s penalization \nto correct for bias in estimating rare events.\nAs a robustness check, we re-specified our analyses using Poisson \ncount models with robust standard errors (Supplementary Tables \n8\u201310). There were no substantive changes in results, with P values \nrarely changing before the third decimal. As we find IRR a more intui-\ntive and practically meaningful measure of effect size, we report the \nPoisson results in the paper instead of the OLS results, but the reader \ncan find both in Supplementary Information. For more details, see \nSupplementary Note 1.\nReporting summary\nFurther information on research design is available in the Nature Port-\nfolio Reporting Summary linked to this article.\nData availability\nExported data from the Salesforce database are not publicly available \ndue to ethical and privacy concerns. An anonymized, public version \nof the dataset with referral behaviour and resulting solar contracts, \nwithout any personally identifiable information, is available on the \nOpen Science Framework (https://osf.io/x4sqp/). The time-stamped \ndata needed to generate Fig. 2c are proprietary to the non-profit part-\nner. Zip-code-level data used in Supplementary Note 3 are not publicly \navailable to protect the privacy of participants.\nCode availability\nReplication code to generate results from the anonymized, publicly \navailable data is at https://osf.io/x4sqp/.\nReferences\n1.\t\nCarley, S. & Konisky, D. M. The justice and equity implications of \nthe clean energy transition. Nat. Energy 5, 569\u2013577 (2020).\n\nNature Energy | Volume 8 | August 2023 | 850\u2013858\n857\nArticle\nhttps://doi.org/10.1038/s41560-023-01298-5\n2.\t\nHelp to Heat Schemes. UK Government https://www.gov.uk/ \ngovernment/collections/find-energy-grants-for-you-home- \nhelp-to-heat (2023).\n3.\t\nInflation Reduction Act of 2022, Pub.L. No. 117\u2013169. (Sec. 50122). \nhttps://www.congress.gov/bill/117th-congress/house-bill/5376 \n(2022).\n4.\t\nWarmer Kiwi Homes programme. EECA https://www.\neeca.govt.nz/co-funding/insulation-and-heater-grants/\nwarmer-kiwi-homes-programme/ (accessed 30 March 2023).\n5.\t\nFully funded energy upgrades scheme. Sustainable \nEnergy Authority of Ireland https://www.seai.ie/grants/\nhome-energy-grants/fully-funded-upgrades-for-eligible-homes/ \n(accessed 27 March 2023).\n6.\t\nEnergy Conservation and Production Act, 42 U.S.C. \u00a76863 - \nWeatherization Program. https://www.govinfo.gov/content/pkg/\nUSCODE-2021-title42/html/USCODE-2021-title42-chap81- \nsubchapIII-partA-sec6863.htm (2021).\n7.\t\nHome Energy Support Program. ACT Government https://\nwww.climatechoices.act.gov.au/policy-programs/\nhome-energy-support-rebates-for-homeowners (2022).\n8.\t\nNew Green Savings Programme. State Environmental \nFund of the Czech Republic https://www.sfzp.cz/en/\nadministered-programmes/new-green-savings-programme/ \n(2023).\n9.\t\nRebate swap for solar. NSW Government https://www.energy.\nnsw.gov.au/households/rebates-grants-and-schemes/\nrebate-swap-solar (2023).\n10.\t Currie, J. The Take Up of Social Benefits (Russell Sage Foundation, \n2006).\n11.\t\nBhargava, S. & Manoli, D. Psychological frictions and the \nincomplete take-up of social benefits: evidence from an IRS field \nexperiment. Am. Econ. Rev. 105, 3489\u20133529 (2015).\n12.\t Finkelstein, A. & Notowidigdo, M. J. Take-up and targeting: \nexperimental evidence from SNAP. Q. J. Econ. 134, 1505\u20131556 \n(2019).\n13.\t Nonzee, N. J. et al. Delays in cancer care among low-income \nminorities despite access. J. Women\u2019s Health 24, 506\u2013514  \n(2015).\n14.\t Remler, D. K. & Glied, S. A. What other programs can teach us: \nincreasing participation in health insurance programs. Am. J. \nPublic Health 93, 67\u201374 (2003).\n15.\t Fowlie, M., Greenstone, M. & Wolfram, C. Are the non-monetary \ncosts of energy efficiency investments large? Understanding low \ntake-up of a free energy efficiency program. Am. Econ. Rev. 105, \n201\u2013204 (2015).\n16.\t Reames, T. G. A community-based approach to low-income \nresidential energy efficiency participation barriers. Local Environ. \n21, 1449\u20131466 (2016).\n17.\t Fox, N. Increasing solar entitlement and decreasing energy \nvulnerability in a low-income community by adopting the \nProsuming Project. Nat. Energy 8, 74\u201383 (2023).\n18.\t Rogers, E. M. Diffusion of Innovations 5th edn (Simon and \nSchuster, 2003).\n19.\t Southwell, B. G., Slater, J. S., Nelson, C. L. & Rothman, A. J. \nDoes it pay to pay people to share information? Using financial \nincentives to promote peer referral for mammography among the \nunderinsured. Am. J. Health Promot 26, 348\u2013351 (2012).\n20.\t Gwadz, M. et al. Public health benefit of peer-referral strategies \nfor detecting undiagnosed HIV infection among high-risk \nheterosexuals in New York City. J. Acquir. Immune Defic. Syndr. 74, \n499\u2013507 (2017).\n21.\t Golden, M. R. et al. Peer referral for HIV case-finding among men \nwho have sex with men. AIDS 20, 1961\u20131968 (2006).\n22.\t Wolske, K. S., Gillingham, K. T. & Schultz, P. W. Peer influence on \nhousehold energy behaviours. Nat. Energy 5, 202\u2013212 (2020).\n23.\t Bollinger, B. & Gillingham, K. Peer effects in the diffusion of solar \nphotovoltaic panels. Mark. Sci. 31, 900\u2013912 (2012).\n24.\t Wolske, K. S. More alike than different: profiles of high-income \nand low-income rooftop solar adopters in the United States. \nEnergy Res. Soc. Sci. 63, 101399 (2020).\n25.\t Kraft-Todd, G. T., Bollinger, B., Gillingham, K., Lamp, S. & Rand, \nD. G. Credibility-enhancing displays promote the provision of \nnon-normative public goods. Nature 563, 245\u2013248 (2018).\n26.\t Zanolini, A. et al. Feasibility and effectiveness of a peer referral \nincentive intervention to promote male circumcision uptake in \nZambia. J. Acquir. Immune Defic. Syndr. 72, S262\u2013S268 (2016).\n27.\t Borenstein, S. It\u2019s time for rooftop solar to compete with other \nrenewables. Nat. Energy 7, 298\u2013298 (2022).\n28.\t O\u2019Shaughnessy, E. Curbing rooftop solar is a poor way to promote \nequity. Nat. Energy 7, 299\u2013299 (2022).\n29.\t Barbose, G., Forrester, S., Darghouth, N. & Hoen, B. Income trends \namong U.S. residential rooftop solar adopters. US Department of \nEnergy https://doi.org/10.2172/1603637 (2020).\n30.\t Reames, T. G. Distributional disparities in residential rooftop \nsolar potential and penetration in four cities in the United States. \nEnergy Res. Soc. Sci. 69, 101612 (2020).\n31.\t Lukanov, B. R. & Krieger, E. M. Distributed solar and environmental \njustice: exploring the demographic and socio-economic trends \nof residential PV adoption in California. Energy Policy 134, 110935 \n(2019).\n32.\t Darghouth, N. R., O\u2019Shaughnessy, E., Forrester, S. & Barbose, G. \nCharacterizing local rooftop solar adoption inequity in the US. \nEnviron. Res. Lett. 17, 034028 (2022).\n33.\t Barbose, G., Forrester, S., O\u2019Shaughnessy, E. & Darghouth, N. \nResidential solar-adopter income and demographic trends: \n2021 update. Lawrence Berkeley National Laboratory https://\nescholarship.org/uc/item/473933mw (2021).\n34.\t Sigrin, B. O. & Mooney, M. E. Rooftop solar technical potential  \nfor low-to-moderate income households in the United States.  \nUS Department of Energy https://doi.org/10.2172/1434891 (2018).\n35.\t Sigrin, B., Sekar, A. & Tome, E. The solar influencer next door: \npredicting low income solar referrals and leads. Energy Res. Soc. \nSci. 86, 102417 (2022).\n36.\t Bergman, P., Lasky-Fink, J. & Rogers, T. Simplification and defaults \naffect adoption and impact of technology, but decision makers \ndo not realize it. Organ. Behav. Hum. Decis. Process. 158, 66\u201379 \n(2019).\n37.\t Beshears, J., Choi, J. J., Laibson, D. & Madrian, B. C. Simplification \nand saving. J. Econ. Behav. Organ. 95, 130\u2013145 (2013).\n38.\t Brandon, A., List, J. A., Metcalfe, R. D., Price, M. K. & Rundhammer, \nF. Testing for crowd out in social nudges: evidence from a natural \nfield experiment in the market for electricity. Proc. Natl Acad. Sci. \nUSA 116, 5293\u20135298 (2019).\n39.\t Milkman, K. L. et al. A megastudy of text-based nudges \nencouraging patients to get vaccinated at an upcoming doctor\u2019s \nappointment. Proc. Natl Acad. Sci. USA 118, e2101165118 (2021).\n40.\t Rogers, T., Milkman, K. L., John, L. K. & Norton, M. I. Beyond \ngood intentions: prompting people to make plans improves \nfollow-through on important tasks. Behav. Sci. Policy Assoc. 1, \n33\u201341 (2016).\n41.\t Rogers, T. & Milkman, K. L. Reminders through association. \nPsychol. Sci. 27, 973\u2013986 (2016).\n42.\t Mertens, S., Herberz, M., Hahnel, U. J. J. & Brosch, T. The \neffectiveness of nudging: a meta-analysis of choice architecture \ninterventions across behavioral domains. Proc. Natl Acad. Sci. \nUSA 119, e2107346118 (2022).\n43.\t DellaVigna, S. & Linos, E. RCTs to scale: comprehensive evidence \nfrom two nudge units. Econometrica 90, 81\u2013116 (2022).\n44.\t Benartzi, S. et al. Should governments invest more in nudging? \nPsychol. Sci. 28, 1041\u20131055 (2017).\n\nNature Energy | Volume 8 | August 2023 | 850\u2013858\n858\nArticle\nhttps://doi.org/10.1038/s41560-023-01298-5\n45.\t Trivers, R. L. The evolution of reciprocal altruism. Q. Rev. Biol. 46, \n35\u201357 (1971).\n46.\t Axelrod, R. & Hamilton, W. D. The evolution of cooperation. \nScience 211, 1390\u20131396 (1981).\n47.\t Goldstein, N. J., Griskevicius, V. & Cialdini, R. B. Reciprocity by \nproxy: a novel influence strategy for stimulating cooperation. \nAdm. Sci. Q. 56, 441\u2013473 (2011).\n48.\t Gouldner, A. W. The norm of reciprocity: a preliminary statement. \nAm. Sociol. Rev. 25, 161\u2013178 (1960).\n49.\t Cialdini, R. B. & Trost, M. R. in The Handbook of Social Psychology \n4th edn, Vols. 1\u20132 (eds Gilbert, D. T. et al.) 151\u2013192 (McGraw-Hill, \n1998).\n50.\t Yin, B. (Miranda), Li, Y. J. & Singh, S. Coins are cold and cards are \ncaring: the effect of pregiving incentives on charity perceptions, \nrelationship norms, and donation behavior. J. Mark. 84, 57\u201373 \n(2020).\n51.\t Alpizar, F., Carlsson, F. & Johansson-Stenman, O. Anonymity, \nreciprocity, and conformity: evidence from voluntary \ncontributions to a national park in Costa Rica. J. Public Econ. 92, \n1047\u20131060 (2008).\n52.\t Falk, A. Gift exchange in the field. Econometrica 75, 1501\u20131511 \n(2007).\n53.\t White, K., Habib, R. & Dahl, D. W. A review and framework for \nthinking about the drivers of prosocial consumer behavior.  \nJ. Assoc. Consum. Res. 5, 2\u201318 (2019).\n54.\t Mani, A., Mullainathan, S., Shafir, E. & Zhao, J. Poverty impedes \ncognitive function. Science 341, 976\u2013980 (2013).\n55.\t Bettinger, E. P., Long, B. T., Oreopoulos, P. & Sanbonmatsu, L. \nThe role of application assistance and information in college \ndecisions: results from the H&R Block FAFSA experiment.  \nQ. J. Econ. 127, 1205\u20131242 (2012).\n56.\t John, P. & Blume, T. How best to nudge taxpayers? The impact of \nmessage simplification and descriptive social norms on payment \nrates in a central London local authority. J. Behav. Public Admin. \nhttps://doi.org/10.30636/jbpa.11.10 (2018).\n57.\t Lasky-Fink, J., Robinson, C. D., Chang, H. N.-L. & Rogers, T. \nUsing behavioral insights to improve school administrative \ncommunications: the case of truancy notifications. Educ. Res. 50, \n442\u2013450 (2021).\n58.\t Gneezy, U. & List, J. A. Putting behavioral economics to work: \ntesting for gift exchange in labor markets using field experiments. \nEconometrica 74, 1365\u20131384 (2006).\n59.\t Madrian, B. C. & Shea, D. F. The power of suggestion: inertia \nin 401(k) participation and savings behavior. Q. J. Econ. 116, \n1149\u20131187 (2001).\n60.\t Madrian, B. C. Applying insights from behavioral economics to \npolicy design. Annu. Rev. Econ. 6, 663\u2013688 (2014).\n61.\t Zlatev, J. J. & Rogers, T. Returnable reciprocity: returnable gifts \nare more effective than unreturnable gifts at promoting virtuous \nbehaviors. Organ. Behav. Hum. Decis. Process. 161, 74\u201384 (2020).\n62.\t O\u2019Keefe, D. J. in The Persuasion Handbook: Developments in \nTheory and Practice (eds Dillard, J. P. & Pfau, M.) 329\u2013344 (Sage \nPublications, 2002).\n63.\t Brown, M. A., Soni, A., Lapsa, M. V., Southworth, K. & Cox, M. High \nenergy burden and low-income energy affordability: conclusions \nfrom a literature review. Prog. Energy 2, 042003 (2020).\n64.\t Bruhn, M. & McKenzie, D. In pursuit of balance: randomization \nin practice in development field experiments. Am. Econ. J. Appl. \nEcon. 1, 200\u2013232 (2009).\n65.\t CalEnviroScreen 3.0. California Office of Environmental Health \nHazard Assessment https://oehha.ca.gov/calenviroscreen/report/\ncalenviroscreen-30 (2017).\nAcknowledgements\nWe thank GRID Alternatives (particularly J. Coleman, Z. Franklin,  \nA. Kim and L. Nobel) and B. Sigrin at the National Renewable Energy \nLaboratory for their collaboration. We also thank K. Aves and  \nR. Walatka for their assistance in creating figures. This work was \nsupported by the National Renewable Energy Laboratory, operated \nby Alliance for Sustainable Energy, LLC, for the US Department of \nEnergy (DOE) under contract no. DE-AC36-08GO28308 (K.S.W. and \nE.T.). Funding was provided by the US Department of Energy Office \nof Energy Efficiency and Renewable Energy Solar Energy Technology \nOffice. This manuscript was also authored by an employee of \nLawrence Berkeley National Laboratory under contract no. DE-AC36-\n08GO28308 with the US Department of Energy (A.T.-B.). The views \nexpressed in the article do not necessarily represent the views of \nthe DOE or the US Government. The US Government retains and the \npublisher, by accepting the article for publication, acknowledges \nthat the US Government retains a non-exclusive, paid-up, irrevocable, \nworldwide licence to publish or reproduce the published form of this \nwork, or allow others to do so, for US Government purposes.\nAuthor contributions\nK.S.W. and A.T.-B. conceived and implemented the study, with E.T. \nhandling randomization of postal mail recipients. A.T.-B. and E.T. \ncleaned the data, which A.T.-B. analysed. K.S.W. drafted the manuscript \nwith revisions from A.T.-B. E.T. contributed to the methods section.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains \nsupplementary material available at  \nhttps://doi.org/10.1038/s41560-023-01298-5.\nCorrespondence and requests for materials should be addressed to \nKimberly S. Wolske.\nPeer review information Nature Energy thanks Rohan Best,  \nMahelet G. Fikru and Stefan Lamp for their contribution to the peer \nreview of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2023\n\n\n    Scientific Research Findings:", "answer": "We identify cost\u2011effective, scalable strategies to improve the efficacy of a peer referral programme for fully\u2011subsidized low\u2011income solar in California. The baseline programme (\u2018control\u2019) offered existing programme participants a US$200 reward for every referral that resulted in a solar installation. Adding a token gift upfront with a reminder about the programme (to evoke a sense of \u2018reciprocity\u2019) led to 1.7\u00a0times the response rate, 2\u00a0times the number of referrals and\u00a02.6\u00a0times as many solar contracts. A third \u2018reciprocity and simplification\u2019 group in which the gift was combined with a mailable referral slip (instead of just phone or web referrals) led to nearly\u00a05\u00a0times the response rate,\u00a07.5\u00a0times as many referrals, and\u00a05.2\u00a0times as many solar contracts \u2014 making it more cost\u2011effective than the baseline programme. It was also more effective at eliciting referrals from participants who had not previously referred. The results highlight strategies that could be adapted to other energy assistance programmes for electrification measures, heating and insulation upgrades, and electric vehicles.", "id": 12}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 8 | June 2023 | 597\u2013609\n597\nnature energy\nhttps://doi.org/10.1038/s41560-023-01259-y\nArticle\nEquitable low-carbon transition pathways \nfor California\u2019s oil extraction\nRanjit Deshmukh\u2009\n\u200a\u20091,2,3,12\u2009\n, Paige Weber\u2009\n\u200a\u20092,4,12\u2009\n, Olivier Deschenes\u2009\n\u200a\u20092,5,6, \nDanae Hernandez-Cortes\u2009\n\u200a\u20092,7,8, Tia Kordell\u2009\n\u200a\u20091,2,9, Ruiwen Lee1,2,9, \nChristopher Malloy\u2009\n\u200a\u20092,5, Tracey Mangin\u2009\n\u200a\u20091,2,9, Measrainsey Meng\u2009\n\u200a\u20092,3,9, \nSandy Sum\u2009\n\u200a\u20091,2, Vincent Thivierge\u2009\n\u200a\u20091,2, Anagha Uppal2,10, David W. Lea11  \n& Kyle C. Meng\u2009\n\u200a\u20091,2,5,6,12\u2009\nOil supply-side policies\u2014setbacks, excise taxes and carbon taxes\u2014are \nincreasingly considered for decarbonizing the transportation sector. \nUnderstanding not only how such policies reduce oil extraction and \ngreenhouse gas (GHG) emissions but also which communities receive the \nresulting health benefits and labour-market impacts is crucial for designing \neffective and equitable decarbonization pathways. Here we combine \nan empirical field-level oil-production model, an air pollution model \nand an employment model to characterize spatially explicit 2020\u20132045 \ndecarbonization scenarios from various policies applied to California, a \nmajor oil producer with ambitious decarbonization goals. We find setbacks \ngenerate the largest avoided mortality benefits from reduced air pollution \nand the largest lost worker compensation, followed by excise and carbon \ntaxes. Setbacks also yield the highest share of health benefits and the lowest \nshare of lost worker compensation borne by disadvantaged communities. \nHowever, currently proposed setbacks may fail to meet California\u2019s GHG \ntargets, requiring either longer setbacks or additional supply-side policies.\nAcross many industrialized economies, climate policies are increas-\ningly focused on the transportation sector, which lags behind the \nlevel and pace of decarbonization observed in other sectors. Indeed, \nbetween 2010 and 2019, while non-transportation greenhouse gas \n(GHG) emissions have fallen by 6% across Organisation for Eco-\nnomic Co-operation and Development countries, GHG emissions \nfrom transportation have risen by 6% (ref. 1). Today, the transporta-\ntion sector is responsible for the largest share of GHG emissions  \nin the United States and the European Union at 28% and 24%, \nrespectively, and an even larger share in California (40%), the region \nof focus in this study1,2.\nTo date, transportation climate-policy debates have primarily \nfocused on demand-side policies to reduce fossil fuel consumption, \nsuch as fuel taxes, vehicle fuel-economy standards, low-carbon fuel \nstandards and electric vehicle subsidies3\u20139. In recent years, attention \nhas turned towards supply-side policies that directly reduce fossil fuel \nproduction. These policies can take different forms. Some directly ban \nextraction from specific oil fields, such as oil-well setbacks targeted at \nReceived: 9 August 2022\nAccepted: 5 April 2023\nPublished online: 18 May 2023\n Check for updates\n1Bren School of Environmental Science and Management, University of California Santa Barbara, Santa Barbara, CA, USA. 2Environmental Markets Lab \n(emLab), University of California Santa Barbara, Santa Barbara, CA, USA. 3Environmental Studies Department, University of California Santa Barbara, Santa \nBarbara, CA, USA. 4Department of Economics, University of North Carolina, Chapel Hill, NC, USA. 5Department of Economics, University of California Santa \nBarbara, Santa Barbara, CA, USA. 6National Bureau of Economic Research, Cambridge, MA, USA. 7School for the Future of Innovation in Society, Arizona \nState University, Tempe, AZ, USA. 8School of Sustainability, Arizona State University, Tempe, AZ, USA. 9Marine Science Institute, University of California \nSanta Barbara, Santa Barbara, CA, USA. 10Department of Geography, University of California Santa Barbara, Santa Barbara, CA, USA. 11Department of Earth \nScience, University of California Santa Barbara, Santa Barbara, CA, USA. 12These authors contributed equally: Ranjit Deshmukh, Paige Weber, Kyle Meng. \n\u2009e-mail: rdeshmukh@ucsb.edu; paigeweber@unc.edu; kmeng@bren.ucsb.edu\n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n598\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\nhealthcare facilities and playgrounds; (2) an excise tax on each barrel \nof crude oil extracted and (3) a carbon tax on GHG emissions from oil \nextraction. We find that a setback policy provides greater statewide \nhealth benefits but also larger lost worker compensation compared \nwith a carbon or excise tax that achieves the same 2045 GHG emissions \ntarget. In general, setback policies also have better equity outcomes \nas disadvantaged communities accrue a larger share of health benefits \nand a smaller share of loss in worker compensation. By contrast, a \ncarbon tax imposes the smallest statewide worker compensation loss \namong the three policies. Finally, currently proposed setback distances \napplied to only new wells will be unable to meet California\u2019s decarboni-\nzation goals. To do so requires setbacks with a distance greater than  \n1 mile, applied to both new and existing wells and/or combined with a \ncarbon or excise tax.\nCrude oil production and GHG emissions \npathways\nWe develop spatially and temporally explicit pathways that reduce \nCalifornia\u2019s oil extraction in response to various supply-side interven-\ntions\u2014well setbacks, excise tax and carbon tax\u2014between 2020 and \n2045. Our approach has two components and is summarized in Fig. 1. \nFor all oil fields in California (Fig. 1a), we first construct an empirically \nestimated model of crude oil-well entry (Fig. 1b), production and exit \nat the oil-field level to project how various supply-side policies and \nmacroeconomic conditions affect oil production across California oil \nfields out to 2045 (Methods and Supplementary Notes 8\u201311, 16 and 17).  \nIn our second step, we insert field-level predictions of oil production \nfrom our empirical model into: (1) an air pollution model, InMAP (Inter-\nvention Model for Air Pollution)22, to characterize how air pollution \nemissions from oil fields disperse across the state (Fig. 1c,d and Sup-\nplementary Note 13) and (2) an employment input\u2013output model, \nIMPLAN (Impact Analysis for Planning)23,24, which uses fixed multipli-\ners to quantify local employment changes in the oil-extraction sector \n(\u2018direct\u2019), in sectors that provide inputs to oil extraction (\u2018indirect\u2019) \nand in sectors where these workers spend income (\u2018induced\u2019) (Fig. 1e \nand Supplementary Note 14). Together, these components provide an \nempirically based analysis of how supply-side policies could alter not \njust oil production across oil fields but also the spatial distribution of \nhealth impacts from air pollution and employment across California.\nFor well setbacks, we consider three setback distances\u20141,000\u2009feet, \n2,500\u2009feet and 1\u2009mile\u2014which encompass distances currently considered \nin policy proposals25\u201328. To ensure policy comparability, we set excise \ntaxes as a percentage of oil price fixed across all years and carbon taxes \nwhich increase at an annual rate of 7% to levels that result in the same \n2045 statewide GHG emissions as our three setback-distance policies \n(Supplementary Note 17). We further consider a fourth excise- and \ncarbon-tax level that achieves a 90% GHG emissions reduction by 2045 \ncompared with 2019 levels, inline with California\u2019s target for in-state \nfinished-fuel demand2.\nEach combination of policy intervention\u2014setbacks, excise tax \nand carbon tax\u2014and the 2045 annual GHG emissions target result in \na unique spatial and temporal pattern of oil production, benefits and \ncosts. We model these patterns across California for the 2020\u20132045 \nperiod, focusing on avoided mortality due to reduced PM2.5 emissions \nand avoided global climate damages from reduced GHG emissions \non the benefits side and lost earnings from the oil-extraction sector \non the cost side. We analyse these policy scenarios using a common \nbenchmark projection of global oil prices out to 2045 (US Energy \nInformation Administration\u2019s (EIA) reference oil-price projection29). \nSensitivity analysis results using higher and lower projected oil prices \nare shown in the Supplementary Information.\nCalifornia\u2019s oil production peaked in 1985 and has been declining \nsince then30. Our projection of statewide oil production to 2045 under \na business-as-usual (BAU) scenario continues this trend (Fig. 2). In this \nno-supply-side policy BAU scenario, oil production in 2045 decreases \nfields located near where people live and work. Other policies reduce \nextraction by targeting oil fields according to their extraction costs, \neither on a per barrel basis as with an excise (or severance) tax or on \na per GHG-emissions basis as with a carbon tax. Thus, for the same \noverall GHG emissions target, different supply-side policies can gener-\nate distinct aggregate and distributional consequences by reducing \nproduction from different oil fields.\nTwo primary considerations arise when evaluating supply-side \npolicies. The first is the relative effectiveness of each policy type in \nreducing oil production and associated GHG emissions, which to date, \nhas received limited empirical analysis10\u201312. The second pertains to the \nancillary benefits and costs of each policy and how they are distributed \nacross different communities. In particular, oil extraction tends to \nbe highly spatially concentrated in certain areas, employing a local \nworkforce and generating air pollution impacting nearby residents. \nDepending on how oil extraction is spatially located in relation to \nworkers and households, different supply-side policies can have dif-\nferent aggregate and distributional consequences in terms of health \nbenefits and labour-market impacts. For example, for the same overall \nGHG emissions target, a policy that phases out more labour-intensive \noil fields may have higher lost worker compensation than other poli-\ncies. Likewise, a policy that bans oil fields near where disadvantaged \nhouseholds reside may generate larger overall health benefits and \nhealth equity gains. Quantifying such potential consequences is critical \nfor informing the design of supply-side policies. More broadly, there \nis a need to understand if and how effectiveness in GHG emissions \nreductions and distributional consequences trade off across different \noil supply-side policies.\nPrevious decarbonization studies employ either Integrated \nAssessment Models, which are combined energy, economy and climate \nmodels13,14, or macro energy-system models15\u201317 that model regional \nenergy systems. These models typically simulate or optimize energy \ninfrastructure investments and retirements to meet certain GHG \nemissions-reduction targets by assuming that fossil fuel extraction \nwill be phased out and replaced by cleaner alternatives. Such models \ntypically do not explicitly consider how specific supply-side policies \n(other than a carbon tax) can yield different decarbonization outcomes \nfor fossil fuel extraction. Furthermore, most energy or economic mod-\nels lack the fine spatial resolution needed to examine the distributional \noutcomes of alternative policies over time. For example, existing stud-\nies on the distributional and equity consequences of phasing fossil fuel \nproduction including oil extraction use only the petroleum basin or \ncounty level and not the oil-field and census-tract-level representation \nfor fuel production and air pollution exposure, respectively15,18, which \nis critical to accurately estimate energy production, health effects and \nequity outcomes of decarbonization pathways.\nThis paper examines the effectiveness and distributional con-\nsequences of potential supply-side policies intended to phase out \noil extraction across California. As the world\u2019s fifth-largest economy \nand the United States\u2019 seventh-largest oil-producing state, California \nprovides a unique setting to study supply-side policies. The state is cur-\nrently implementing some of the world\u2019s most ambitious climate poli-\ncies with a statewide carbon-neutrality goal by 2045. This includes an \nactive debate over various supply-side policies to dramatically reduce \noil extraction, with an explicit interest in examining resulting labour \nand health equity consequences and their distribution across the \nstate19\u201321.We improve upon previous studies by developing an empiri-\ncally estimated model of crude oil-well entry (drilling), production and \nexit (retirement) at the oil-field level along with an air pollution model \nto quantify health effects at the census-tract level and an employment \ninput\u2013output model to determine employment impacts at the county \nlevel. We examine three supply-side policy interventions that have \nbeen widely debated in California and elsewhere: (1) well setbacks \nthat require new oil wells to be located beyond a specified minimum \ndistance from sensitive sites such as occupied dwellings, schools, \n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n599\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\nby 57% compared with 2019 levels. Associated GHG emissions decline \nby 53%, which is well short of California\u2019s decarbonization targets.\nSupply-side policies lower statewide crude oil production but \nwith different temporal and spatial patterns (Fig. 2a and Supplemen-\ntary Fig. 17). Setbacks applied to new wells, excise taxes applied per \nunit of production and carbon taxes applied per tonne of GHG emis-\nsions lead to continuous declines that outpace that of the BAU trajec-\ntory, albeit with different pathways. In general, a setback and an excise \ntax result in lower oil production in each year when compared with a \ncarbon tax that is calibrated to achieve the same 2045 GHG emissions \ntarget. This is because a carbon tax on extraction emissions targets  \noil fields with higher GHG emissions intensities, whereas a setback \ntargets oil fields in more populated areas and an excise tax targets \nproduction declines among more costly oil fields. Supplementary \nFig. 1 shows that the relationship between production costs and emis-\nsions intensities is not systematic. As a result, the fields that reduce \nproduction under a carbon tax will be unique from the fields that \nreduce production under an excise tax that achieves an equivalent \nreduction in carbon emissions.\nThere is close correspondence between statewide oil production \nand emissions pathways (Fig. 2b). As with oil production, setbacks, \nexcise taxes and carbon taxes induce a continuous decline. By con-\nstruction, because excise- and carbon-tax levels were calibrated to \nresult in the same 2045 GHG emissions as the corresponding setback \ndistances, the GHG emissions trajectories of setbacks, excise taxes and \ncarbon taxes are more closely aligned than oil-production trajectories. \nCumulative 2020\u20132045 GHG emissions reductions from carbon taxes \nare consistently lower than setbacks and excise taxes for each 2045 \nGHG emissions target, irrespective of the oil-price projections (Fig. 2c \nand Supplementary Figs. 24 and 25). However, excise taxes, depend-\ning on the tax level required to meet the GHG emissions target under \ndifferent oil prices, could have slightly lower or higher cumulative \nGHG emissions compared to setbacks. When considering alternative \noil-price projections, annual GHG emissions reduction in 2045 for a \n1\u2009mile setback is substantially lower (33%) under EIA\u2019s high oil-price \nprojection (Supplementary Fig. 24), while it nearly reaches the 90% \nreduction target under EIA\u2019s low oil-price projection (89% reduction) \n(Supplementary Fig. 25).\nHealth, labour and avoided climate change \nimpacts\nReduced crude oil production from supply-side policies have associ-\nated health benefits, labour-market impacts and benefits from avoided \nclimate change damages. We estimate statewide health benefits from \ncumulative avoided mortality resulting from lower air pollution levels, \ncosts from lost total labour compensation and benefits from avoided \nclimate change damages due to abated GHGs, priced at the social cost \nof carbon31, both total (Fig. 3a\u2013c) and per unit of cumulative avoided \nGHG emissions over 2020\u20132045 for each scenario (Fig. 3d\u2013f). The \ncosts and benefits are relative to the BAU scenario and estimated in \nnet-present-value terms, valued in 2019 US dollars (Supplementary \nNotes 13\u201315).\n33\n34\n35\n36\n37\n122\n121\n120\n119\n118\nPM2.5 concentration of\nall oil-field emissions\n0\n1\n2\n3\n4\n1978\n1990\n2000\nYear\n2010\n2020\nNumber of new oil wells (thousand)\nLatitude (\u00b0 N)\nLatitude (\u00b0 N)\nLatitude (\u00b0 N)\nLatitude (\u00b0 N)\nObserved\nEstimated\nObserved and modelled\noil wells\nOil production (million barrels)\n0\n1\n2\n3\nPM2.5 concentration from\nVentura cluster\n0.00025\n0.00125\nPM2.5 (\u00b5g m\n\u20133)\n34\n33\n35\n36\n37\n122\n121\n120\nLongitude (\u00b0 W)\nLongitude (\u00b0 W)\nLongitude (\u00b0 W)\nLongitude (\u00b0 W)\n119\n118\nOil fields and disadvantaged\ncommunities\nDAC\n0\n5\n10\n15\nOil production (million barrels)\n34\n33\n35\n36\n37\n122\n121\n120\n119\n118\n34\n35\nc\na\nb\nd\ne\n119.5\n119.0\n118.5\nVentura\nFresno\nKern\nLos Angeles\nMonterey\nVentura\nPopulation\u2212weighted\nPM2.5 (\u00b5g m\n\u22123)\n0\n500\n1,000\n1,500\n0\n500\n1,000\nMillion US$\nFresno\nLos Angeles\nMonterey\nVentura\nKern\nWorker compensation from\nall oil production\nFig. 1 | Summary of data and methods. a, Oil production in 2019 by field. Grey \nshaded areas indicate census tracts with disadvantaged communities (DAC), \nas defined by CalEnviroScreen. b, Observed and estimated historical oil-well \nentry across California (Supplementary Note 9). c, Particulate matter (PM2.5) \nconcentration by census tract for a 1\u2009tonne pulse of PM2.5 emissions from the \nVentura cluster. Points indicate location of 2019 oil production from oil fields \nwithin the cluster. d, PM2.5 concentration by census tract associated with all \n2019 oil production. e, Worker compensation by county associated with all \n2019 oil production.\n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n600\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\nWe note that health benefits denominated in monetized avoided \nmortality from air-quality improvements and lost worker compensa-\ntion from oil extraction reported here do not provide a full account of \nstatewide benefits and costs under each supply-side policy. Reductions \nin ambient air pollution can bring a wide range of health benefits, \nincluding reduced morbidity, asthma attacks and other respiratory \ndiseases and lower hospital and medication expenses. For example, \nreduced activity in the oil and gas extraction sectors may reduce \nground-level ozone concentrations, which may lead to additional \nhealth benefits that are not accounted for in our study32. To the extent \nthat other ambient air pollutants such as ozone travel similarly to PM2.5, \nthe disadvantaged communities vs non-disadvantaged communities \ncontrasted in the estimated health benefits should be a reasonable \napproximation of the full health benefits comparison despite focusing \nonly on primary and secondary PM2.5.\nWe focus on monetized avoided mortality alone to measure the \nbenefits of air-quality improvements because the previous literature \nhas shown that monetized avoided mortality is by far the largest ben-\nefit33. Premature mortality is also the health end point for which there \nis the most scientific consensus supporting the causal link between \nair pollution (in particular PM2.5) and the end point33. There are also \npotential benefits associated with non-health impacts through changes \nin agricultural and labour productivity34,35. Likewise, we are unable to \naccount for the possible re-employment of oil-extraction workers that \nmay find employment in other sectors. Unfortunately, little is known \non re-employment rates and wages for former oil-extraction workers to \ninform such calculations. Thus, our estimates represent lower bounds \nof potential health benefits and upper bounds of potential employment \nand worker compensation losses. Lastly, considerable uncertainty \nexists in the value of the social cost of carbon, a key ingredient in how \navoided climate damages are calculated31. For these reasons, we present \nour health, labour and avoided climate damage values separately in \nFig. 3, without attempting to conduct a full cost\u2013benefit analysis. We \ninstead focus on the relative rankings of each benefit and cost across \nthe three supply-side policies examined.\nAmong policies, setbacks consistently achieve the greatest health \nbenefits, both in total and per unit of cumulative avoided GHG emis-\nsions (Fig. 3a,d). This result validates the intent behind setbacks, a \nBAU\nBAU\nBAU\n2020\n60\n70\nGHG emissions-reduction target (%, 2045 vs 2019)\n80\n90\nBarrels (million)\nMt CO2e\n0\n0\n40\n80\n120\na\nc\nb\n160\nMt CO2e\n150\n175\n200\n250\n225\n275\n5\n10\n15\n2025\n2030\n2035\n2040\n2045\n2020\n2045 GHG emissions-reduction target\nPolicy\n65% (= 2,500 foot setback)\n72% (= 1 mile setback)\n90%\nCarbon tax\nExcise tax\nSetback (new wells)\nOil production\nCumulative GHG emissions\nYear\nYear\nGHG emissions\n2025\n2030\n2035\n2040\n2045\nFig. 2 | California crude oil production and associated GHG emissions \npathways. Annual California oil production and GHG emissions under BAU \nand three supply-side policies\u2014setbacks applied to new wells, excise tax on oil \nproduction and carbon tax on emissions from oil extraction. Excise and carbon \ntaxes are calibrated to meet 62% (=\u20091,000\u2009foot setback), 65% (=\u20092,500\u2009foot \nsetback), 72% (=\u20091\u2009mile setback) and 90% GHG emissions reduction by 2045 \nrelative to 2020. a, Crude oil production. b, GHG emissions from crude oil \nproduction. c, Cumulative 2020\u20132045 GHG emissions. Data for 62% GHG \nemissions-reduction scenario (=\u20091,000\u2009foot setback) not shown in a and b for \nvisual clarity. Setback distances are limited to 1\u2009mile or below, and thus, a setback \nthat meets a 90% 2045 GHG emissions target is not modelled. Total number of oil \nfields in the model is 263. CO2e = carbon dioxide equivalent.\n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n601\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\npolicy designed specifically for improving health outcomes by elimi-\nnating oil extraction from fields that are situated near residences, \nschools and other locations where people live and work. However, per \nunit of cumulative avoided GHG emissions, longer-distance setbacks \nyield smaller health benefits (Fig. 3d) because the marginal pollution \nfrom avoided wells affects a smaller number of people.\nFor statewide worker compensation losses, the pattern flips \nacross supply-side policies. For a given 2045 GHG emissions target, \nsetbacks consistently generate slightly higher worker compensation \nlosses across the state than excise taxes, which exceed that for carbon \ntaxes (Fig. 3b). This is because setbacks experience a drop in produc-\ntion larger than excise and carbon taxes designed to meet the same \n2045 GHG emissions target, and they affect wells in counties that \nhave a higher employment intensity (jobs per barrel of oil produced). \nExcise taxes lead to greater worker compensation loss because they \nare less cost effective at targeting GHG emissions reductions com-\npared with carbon taxes, requiring a larger drop in oil production \nand associated employment losses to meet the same GHG emissions \ntarget. The ranking across policies is preserved when considering \nworker compensation losses per unit of cumulative avoided GHG \nemissions (Fig. 3e).\nFor avoided climate change damages, setbacks deliver slightly \ngreater cumulative benefits for each 2045 GHG emissions target com-\npared with excise and carbon taxes (Fig. 3c). These differences are \neven smaller across policies on a per unit of cumulative avoided GHG \nemissions basis (Fig. 3f).\nThe relative ranking for the health impacts from the three \nsupply-side policies remains the same under the EIA\u2019s high and low \noil-price projections, although the average magnitude of these benefits \nand costs are correspondingly higher or lower than the reference EIA \noil-price projection (Supplementary Figs. 26 and 27). Cumulative lost \nworker compensation and avoided climate damages remain the lowest \nfor carbon taxes across high and low oil-price projections (Supplemen-\ntary Figs. 26 and 27).\nDrivers of health and labour outcomes across \npolicies\nThe ranking of health benefits and labour costs shown in Fig. 3 across \nsupply-side policies occurs because each policy targets different \naspects of crude oil production and thus the sequence and timing of \nwell entry, production and retirements across oil fields. To explore \nthis further, we sort oil fields according to the characteristic directly \ntargeted by each policy. Specifically, these characteristics, shown on \nthe x axis across the columns of Fig. 4, include an oil-field cluster\u2019s: (1) \narea share near sensitive sites, (2) per barrel cost of extraction per barrel \nand (3) GHG emissions intensity per barrel. These characteristics are \ndirectly affected by a setback, an excise tax and a carbon tax. Under each \npolicy, oil fields on the left of the x axis retire first, moving rightward \nas stringency tightens. For example, for a particular setback distance \n(2,500\u2009feet in Fig. 4a,d), fields with a greater share of their area near \nsensitive sites will experience greater reduction in oil production than \nfields with areas less affected by the same setback. The latter fields that \nGHG emissions-reduction target (%, 2045 vs 2019)\n70\na\n Health: avoided mortality\nb\nLabour: lost worker compensation\nc\n Climate: avoided damage\nd\ne\nf\n0\n0.5\n1.0\nNPV (billions of US$)\nNPV (millions of US$)\nper avoided GHG Mt CO2e\n1.5\n2.0\n\u201315\n\u201310\n\u20135\n0\n5\n4\n3\n2\n1\n0\n\u2013150\n\u2013100\n\u201350\n0\n0\n10\n20\n40\n30\n50\n0\n10\n20\n40\n30\n50\n80\n90\n70\n80\n90\n70\n80\n90\n70\n80\n90\n70\n80\n90\n70\n80\n90\nPolicy\nSetback (new wells)\nExcise tax\nCarbon tax\nFig. 3 | Health, labour and climate impacts from California\u2019s oil-production \npathways under different policies relative to BAU. a\u2013c, Total health benefits \nfrom avoided mortality (a), total lost worker compensation (b) and avoided \nclimate damages valued at the social cost of carbon over 2020\u20132045 (c) under \nthree supply-side policies\u2014setbacks applied to new wells, excise tax on oil \nproduction and carbon tax on emissions from oil extraction\u2014relative to BAU  \nto meet four 2045 GHG emissions targets. d\u2013f, Panels replicate a (d), b (e) and  \nc (f) but normalized by cumulative avoided GHG emissions over 2020\u20132045.  \nNo setback distance equivalent to 90% 2045 GHG emissions target is applied. \nTotal number of oil fields in the model is 263. Net present values (NPV) are in  \n2019 US dollars, estimated using a discount rate of 3%.\n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n602\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\nare farther from sensitive sites will be increasingly affected as setback \ndistances increase. Likewise, under a low excise tax, the oil fields that \ninitially phase out production are those with higher extraction costs. \nAs the excise tax increases, oil fields with lower extraction costs incre-\nmentally phase out production. A similar pattern holds for carbon taxes \nand their effect on oil fields with varying GHG intensities.\nTo understand how policies differ in terms of statewide health ben-\nefits, the y axis in the top panels of Fig. 3 shows the number of affected \nindividuals per unit of pollution for each oil field. Because of the down-\nward relationship shown in Fig. 4a, shorter distance setbacks initially \naffect oil fields that are upwind of more population-dense locations. \nAs setback distances increase, the marginal oil field that is phased out \nis upwind of fewer people, explaining why the health benefit per unit of \ncumulative avoided GHG emissions falls with more stringent setbacks \n(Fig. 3d). By contrast, the relationships between population affected \nby pollution and costs of extraction and GHG intensity of oil fields \nare both upward sloping (Fig. 4b,c). This is reflected in the increas-\ning health benefits, in both total and per unit of cumulative avoided \nGHG emissions, with increasing stringency of excise and carbon taxes  \n(Fig. 4a,d). In other words, as excise and carbon taxes increase, the \nmarginal oil field that exits production is upwind of more people.\nTo understand patterns in labour-market impacts, we explore cor-\nrelations between employment intensity in the oil-extraction sector at \nthe county level in total job losses per million barrels of oil produced \nand the three oil-field characteristics (Fig. 4d\u2013f). The employment \nimpacts reported in this study are driven by IMPLAN multipliers that \naccount for direct, indirect and induced jobs. As shown in Fig. 4, oil \nfields that are more impacted by setbacks have a greater employment \nintensity (jobs per million barrels), reflecting larger multipliers and \ncounty population. For example, oil fields in Los Angeles County are \naffected more by shorter setbacks because a larger population in the \ncounty lives close to oil fields, but they also create more direct, indirect \nand induced jobs based on IMPLAN\u2019s data. The downward relationship \nin Fig. 4d explains why employment loss per GHG emissions reduction \nis the highest at shorter setback distances (Fig. 3d). Shorter setbacks \ninduce more labour-intensive oil fields to exit production first, followed \nby less labour-intensive fields as setback distances increase. Again, \nby contrast, Fig. 4e,f is upward sloping, indicating that with excise \nand carbon taxes, less labour-intensive oil fields go out of production \nfirst. This is consistent with statewide labour costs, in both the total \nand per unit of cumulative avoided GHG emissions basis, increas-\ning (more negative) in Fig. 4b,e as excise- and carbon-tax stringency \nincreases. Higher excise and carbon taxes incrementally induce more \nlabour-intensive fields to go out of production.\nCounty-level outcomes are similarly driven by county and oil-field \ncharacteristics. Comparing California\u2019s three highest oil-producing \ncounties in 2019, production in Los Angeles County has lower average \ncosts per barrel and lower average GHG emissions intensity compared \nwith Kern or Monterey counties (Supplementary Figs. 19 and 20) but \ngreater health impacts (mortality) and employment intensity per barrel \nof oil production (Supplementary Figs. 21\u201323). Under a setback policy, \noil production in denser Los Angeles County is affected more than \nKern and Monterey counties (Supplementary Fig. 18), which results \nin greater health benefits but also higher labour impacts compared \nwith the excise- and carbon-tax policies. Because the average cost \nof oil production and GHG emissions intensities in oil fields in Kern \nand Monterey counties are greater than Los Angeles County, both \nthe excise- and carbon-tax policies result in lower health benefits and \nlabour impacts compared to the setback policy.\nEquity impacts of supply-side policies\nTo understand the equity impacts of supply-side policies, we exam-\nine how the statewide health and labour consequences of each decar-\nbonization pathway are distributed spatially across the state. We use \n200\n400\n600\n800\n1,000\n1,200\nPopulation afected by\npollution\n0\n0.2\n0.4\n0.6\n0.8\n1\na\n\u2013500\n0\n500\n1,000\n20\n40\n60\n80\nb\n\u2013500\n0\n500\n1,000\n20\n40\n60\n80\nc\n\u20130.0005\n0\n0.0005\n0.0010\nEmployment intensity\n(jobs per million barrels)\n0\n0.2\n0.4\n0.6\n0.8\n1\nShare of area afected by setback\nd\n\u20130.0005\n0\n0.0005\n0.0010\n18\n22\n26\n30\n34\n38\nCost of extraction (US$ per barrel)\ne\n\u20130.0002\n0\n0.0002\n0.0004\n0.0006\n0.0008\n20\n40\n60\nGHG intensity (kg CO2e per barrel)\nf\nFig. 4 | Correlations between health and labour impacts with oil-field \ncharacteristics. a\u2013c, Correlation between statewide population affected by \na 1\u2009tonne pulse of PM2.5 from an oil-field cluster on the y axis and that cluster\u2019s \nshare of area affected by setback (at 2,500\u2009feet) in blue (a); cost of extraction \n(in US dollars per barrel) in red (b); and GHG intensity (in kg CO2e per barrel) \nin orange (c) on x axes. d\u2013f, Replicates a\u2013c but with employment intensity (in \njobs per million barrels of oil produced) on the y axis at the county level. Total \nnumber of oil fields in the model is 263. All oil-field characteristics shown here are \nestimates from 2020. Data are presented as mean values \u00b1\u20091.96\u2009\u00d7\u2009standard errors \nof measurement (SEM).\n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n603\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\nCalifornia\u2019s legal definition of whether a census tract is a \u2018disadvantaged\u2019 \ncommunity (DAC) using CalEnviroScreen, a scoring system based on \nmultiple-pollution exposure and socioeconomic indicators developed \nby the California Environmental Protection Agency36. For each policy \nscenario, we estimate the share of the total statewide health benefits \nand employment losses in oil extraction borne by communities living \nin disadvantaged community census tracts (Fig. 5a,b).\nThe DAC\u2019s share of health benefits is consistently larger under a \nsetback than under excise and carbon taxes for a given 2045 GHG emis-\nsions target. This share is largest at lower setback distances or equiva-\nlently less stringent 2045 GHG emissions targets and decreases as the \nsetback distance increases. For excise and carbon taxes, the DAC\u2019s share \nof benefits is relatively unaffected by the stringency of the 2045 GHG \nemissions target. The lost worker compensation is largest for setbacks \nat the statewide level. However, the share of total lost worker compen-\nsation from workers in DACs is consistently lower under setbacks than \nunder excise and carbon taxes. Thus, for any given 2045 GHG emissions \ntarget, a greater share of health benefits and a lower share of worker \ncompensation impacts are experienced by DACs under a setback than \nunder excise and carbon taxes. This result holds even under the EIA\u2019s \nhigh and low oil-price projections (Supplementary Figs. 28 and 29).\nSetbacks applied to all versus only new wells\nAlthough most existing and proposed setback policies apply to only \nnew wells, applying setbacks additionally to existing wells could be an \nimportant policy instrument to further mitigate GHG emissions and \nimprove health outcomes of neighbouring communities that have \nhistorically borne the burden of local pollution from oil extraction. To \nunderstand the health, labour and equity consequences of setbacks \non all wells, we also model a setback policy that affects both new and \nexisting wells applied in 2020.\nIn comparison to setbacks on only new wells, applying setbacks to \nall wells predictably results in greater oil-production declines and emis-\nsions reductions. As discussed earlier, setbacks applied to only new \nwells result in a continuous decline in oil production and GHG emissions \n(Fig. 6). In contrast, setbacks applied to all wells induce an immediate \ndrop in statewide oil production and associated GHG emissions in 2020 \nas existing wells within the setback distance fall out of production. \nThis drop is then followed by a gradual decline thereafter that tracks \nthe BAU trajectory. Oil production and GHG emissions reductions \nincrease as setbacks get longer. Although a 1-mile setback, the largest \nconsidered in this study, applied to all wells achieves a substantially \ngreater GHG emissions reduction (81%) by 2045 compared with the \nsame setback on new wells (72%), it still falls short of meeting the 90% \nreduction target (Fig. 6b). However, the cumulative GHG emissions \nreduction over 2020\u20132045 for the 1-mile setback applied to all wells is \non par with those of excise and carbon taxes that result in a 90% annual \nGHG emissions reduction in 2045 (Fig. 2c).\nSetbacks applied to all wells result in fewer premature deaths but \nalso greater total lost worker compensation compared with setbacks \non only new wells (Fig. 6). Setbacks on all wells have better equity out-\ncomes by accruing a greater share of avoided mortality benefits and \na lower share of lost worker compensation to disadvantaged commu-\nnities. Thus, setbacks applied to all wells in general would yield more \npronounced health and labour-market consequences than setbacks \napplied to just new wells.\nDiscussion and conclusions\nBy quantifying the trade-offs across different supply-side policies, we \nfind that for California, an oil-well-setback policy applied to new wells \nprovides greater health benefits compared to a carbon- or excise-tax \npolicy designed to achieve the same 2045 GHG emissions-reduction \ntarget. A setback policy also produces equity gains as DACs accrue \ngreater health benefits and lower employment costs than other com-\nmunities under a setback compared to excise and carbon taxes.\nYet a setback policy imposes the largest statewide loss of worker \ncompensation among the three policies for the reference oil-price \nprojection. Moreover, on its own, a setback policy applied to new wells \nachieves only a 72% GHG emissions reduction in 2045 compared with \n2019 for a 1-mile setback, a distance larger than the maximum 3,200\u2009feet \ncurrently proposed in California28. GHG emissions reductions would \nbe even lower under higher global crude oil prices. While a setback \npolicy is generally advocated by stakeholders based on public health \nconcerns, it will need to either impose greater distances, be applied \nto both new and existing wells or be combined with an appropriate \nexcise or carbon tax to meet California\u2019s decarbonization goals (Sup-\nplementary Figs. 30\u201335).\nWhereas carbon taxes and excise taxes are both able to achieve \nmore aggressive annual GHG emissions reductions, that is, 90% GHG \nemissions reduction by 2045 compared with 2019, the tax values \nrequired to achieve 90% decarbonization are higher than those con-\nsidered in current policies. The carbon tax required to drive a 90% \nGHG emissions reduction by 2045 starts at US$250\u2009t\u22121\u2009CO2e in 2020 \nand increases to US$1,330\u2009t\u22121\u2009CO2e in 2045. This trajectory is nearly \nfour times higher than the allowance price ceiling under California\u2019s \ncap-and-trade system that starts at US$65\u2009t\u22121\u2009CO2e in 2021 and rises to \nUS$330\u2009t\u22121\u2009CO2e by 2045, assuming an annual real growth rate of 5% \nand an inflation rate of 2% (ref. 37). Similarly, none of the excise taxes \ncurrently in effect across 27 US states exceed 10% of the oil price38, \nwhich is far lower than the 67% tax we find is required to achieve a \n90% GHG emissions-reduction target by 2045 under EIA\u2019s reference \noil-price projection.\nFinally, our results indicate that combining a setback with a car-\nbon tax could achieve the state\u2019s GHG emissions target while yielding \ngreater statewide health benefits, lower statewide worker compensa-\ntion losses and larger equity gains compared to having just a carbon \ntax or excise tax alone. However, if the setbacks are applied to only new \nwells, the carbon-tax trajectory would still need to be three times higher \nthan currently permitted under California\u2019s cap-and-trade system \n(Supplementary Fig. 16). For the two trajectories to be similar, setbacks \nwould need to be applied to both existing and new wells.\nAlthough we examined only the impacts of PM2.5 on health out-\ncomes, oil extraction also emits other toxic pollutants, including ben-\nzene, ethylbenzene and n-hexane, which are known to cause cancer and \nother serious health effects39. Setbacks will not only reduce exposure \nto PM2.5 pollution but will also decrease exposure to these other toxic \npollutants and thus could lead to larger health benefits as oil extraction \nis phased out. To realize the health and climate benefits of setbacks \nGHG emissions-reduction target (%, 2045 vs 2019)\nDAC share\n0.35\na\nb\n0.32\n0.30\n0.28\n0.25\n60\n70\n80\n90\n60\n70\n80\n90\n0.40\n0.38\n0.35\n0.33\n0.30\nCarbon tax\nExcise tax\nSetback (new wells)\nPolicy\nHealth: avoided mortality\nLabour: lost worker compensation\nFig. 5 | DACs\u2019 share of health and labour impacts. a,b, Share of avoided \nmortality benefits borne by individuals (a) and share of foregone oil-extraction \nearnings borne by workers in DACs (b) under setbacks, excise tax and carbon tax \nfor different 2045 GHG-reduction targets.\n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n604\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\nestimated in this study, setbacks will need to be applied to both exist-\ning and new wells, unlike most existing and proposed regulations that \napply setbacks to only new wells.\nTwo other supply-side policies that we do not examine in this study \ninclude limiting producer subsidies14,40 and restricting development of \noil fields, either by compensating resource owners for not exploiting \ntheir fuel resources, buying and retiring resource rights or limiting new \nleases on government lands10,41. The former is similar to imposing an \nexcise tax on production, whereas the latter requires rules to prioritize \nfields for constraining development, similar to a setback policy that is \nconsidered in this study.\nThe effectiveness and equity trade-offs across various oil \nsupply-side policies must be ultimately considered in tandem with oil \ndemand-side policies, without which global GHG emissions reductions \nmay be limited when oil markets are global. For example, demand-side \npolicies from any jurisdiction alone may yield limited GHG emissions \nreductions if other jurisdictions increase oil demand in response to \nlower global oil prices11,42,43. Similarly, restricting only oil supply in a \nPolicy:\n160\nBarrels (million)\nNPV (billions of US$)\nDAC share\nMt CO2 e\n120\n80\n40\n0\n2020\n60\n65\n70\n75\n80\n65\n70\n75\n80\n65\n70\n75\n80\n65\n70\n75\n80\n65\n70\n75\n80\n2025\n2030\n2035\n2040\n2045\nBAU\nBAU\n0\n0.30\n0.31\n0.32\n0.33\n0.34\n0.35\n1\n2\n3\n275\n250\n225\n200\n175\n150\n\u201315\n0.30\n0.31\n0.32\n0.33\n0.34\n0.35\n\u201310\n\u20135\n0\nSetback (new wells)\nSetback (all wells)\nSetback distance:\n2,500 feet\n5,280 feet\n1,000 feet\n2,500 feet\n5,280 feet\n+\nCumulative GHG emissions\nOil production\na\nb\nc\nd\ne\nf\nLabour: lost worker compensation\nGHG emissions-reduction target (%, 2045 vs 2019)\nYear\nGHG emissions-reduction target (%, 2045 vs 2019)\nGHG emissions-reduction target (%, 2045 vs 2019)\nHealth: avoided mortality\nDAC share labour: lost worker compensation\nDAC share health: avoided mortality\nFig. 6 | Comparison between setback policies applied to new and all wells. \na\u2013f, Three setback distances\u20141,000-foot setback, 2,500-foot setback and \n1-mile setback\u2014applied to new and all (new and existing) wells. Oil-production \npathways (a), cumulative GHG emissions over 2020\u20132045 (b), total health \nbenefits from avoided mortality (c), total lost worker compensation (d), share of \navoided mortality benefits borne by individuals in DACs (e) and share of foregone \noil-extraction earnings borne by workers in DACs (f) under the three setbacks. \nTotal number of oil fields in the model is 263. Net present values are in 2019  \nUS dollars, estimated using a discount rate of 3%.\n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n605\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\nsingle jurisdiction without efforts to limit oil demand in that jurisdic-\ntion will result in an increase in oil exports from elsewhere, with some \namount of local GHG emissions reduction replaced by increased GHG \nemissions elsewhere. By coordinating oil supply- and demand-side \npolicies, it is possible for a jurisdiction\u2019s oil supply and demand curves \nto jointly shift in a manner that leaves the global oil price unchanged \nand avoid GHG leakage to other jurisdictions.\nAdditionally, demand and supply policies that simply reduce GHG \nemissions from transportation fuels may have limited GHG emissions \nreductions if there is not an economy-wide climate policy, such as \na carbon price, that ensures any energy source that replaces oil for \ntransportation, such as electricity, is not more carbon intensive. For \nexample, a transition from oil to electricity in transportation may have \nlimited climate benefits if the electricity is produced primarily by coal. \nFuture research should assess the resulting effectiveness and equity \nconsequences of having multiple complementary climate policies.\nSuch future analyses can take advantage of the methodological \napproach developed in this paper. Across many settings and sectors, \nstakeholders are asking decarbonization policies to take into account \nnot just their GHG emissions consequences but also how the local \ncosts and benefits of these policies are distributed spatially and across \ndifferent demographic groups. This paper provides a step forward \nin that direction by combining an empirical-based, spatially explicit \nenergy production model with state-of-the-art air pollution transport \nmodelling to quantify health benefits at a fine spatial scale and an \nemployment model to quantify local labour-market consequences. \nOur framework can be applied to other decarbonization policies at \nvarious scales such as studying the distributional consequences of \ndecarbonizing other forms of fossil fuel extraction, electricity produc-\ntion or manufacturing activity. More broadly, in many settings that \nalready exhibit socioeconomic inequities, there is an increasing need \nto understand whether decarbonization policies themselves would \nexacerbate or narrow such inequities. This study and its methodology \nprovides a path forward for such analyses.\nMethods\nModelling framework\nTo estimate the health and labour consequences of supply-side policies, \nwe build an empirically validated model of oil production to estimate \nfield-level oil production and GHG emissions pathways under varying \npolicy scenarios. These estimates drive our projections of pollution \ndispersion, mortality effects and local employment, which are used \nto quantify health and labour impacts under different policy and GHG \nemissions-target scenarios. We further examine the equity impacts of \nthese scenarios focusing on how health and labour impacts are distrib-\nuted between disadvantaged and other communities. Throughout, we \nuse nominal prices in both the estimation and projection parts of the \nanalysis. When presenting health and labour impacts, we calculate net \npresent discounted values in 2019 US dollars after applying a discount \nrate of 3% and an inflation rate of 2%.\nSupply-side policies and oil-price forecasts\nWe model the impacts of three policies\u2014setbacks, an excise tax and a \ncarbon tax\u2014on California\u2019s oil sector. A setback policy prohibits oil (and \ngas) extraction within a specified distance from sensitive sites including \noccupied dwellings, schools, healthcare facilities and playgrounds. We \nmodel two setback scenarios: (1) setbacks that apply to new wells only \n(main results) and (2) setbacks that apply to new and existing wells or \nall wells. We model setbacks on new wells by proportionally reducing \nfield-level future new well entry based on the relative field area covered \nby a given setback buffer. For existing wells, setbacks are implemented \nin our model by removing those within the setback distance from future \nproduction. We consider setback distances of 1,000\u2009feet, 2,500\u2009feet \nand 1\u2009mile. We assume only vertical drilling in the setback analysis. \nHorizontal and directional drilling from pads outside of the setback \ndistance could access additional sub-surface oil resources within the \nsetback distance, reducing our estimates of the health and equity ben-\nefits of setbacks, especially for shorter setback distances44. However, \nthe costs and extent of adoption of horizontal drilling are uncertain for \nCalifornia and thus are not included in this study. The excise-tax policy \nimposes a tax on each barrel of crude oil extracted. In our projection \nperiod, we apply a constant tax rate to the oil price each year. This is \nconsistent with historical proposals for excise taxes on California oil \nextraction45. The carbon-tax policy imposes a tax on the GHG emissions \nfrom the oil-extraction site. We consider only direct GHG emissions, \nexcluding methane emissions due to a lack of reliable oil-field-specific \ndata. All carbon-tax trajectories increase at an annual rate of 7%, the \nsum of a 5% real growth rate and 2% inflation rate per year (ref. 46). We \ndetermine the excise-tax rates applied to the oil price and carbon taxes \nthat result in the following 2045 statewide GHG emissions targets using \nan optimization function: (1) 2045 statewide GHG emissions associated \nwith the three setback distances (Supplementary Table 4) and (2) a 90% \nreduction in statewide GHG emissions compared with 2019. The excise \nand carbon taxes are shown in Supplementary Figs. 15 and 16 and are \ninputs to the oil-extraction model and affect future well entry and exit. \nSupplementary Note 17 provides more details.\nFor 2020\u20132045 macroeconomic conditions, we assume three \nBrent spot crude oil nominal price trajectories (reference, low and \nhigh) obtained from the EIA\u2019s Annual Energy Outlook 2021 forecast \n(Supplementary Fig. 13) (ref. 29). For scenarios that do not include \na carbon tax, we apply a baseline nominal carbon price equal to Cali-\nfornia\u2019s cap-and-trade allowance price floor (Supplementary Fig. 14). \nSupplementary Note 16 provides more details.\nOil-production model\nThe model of oil production has three components: (1) well entry,  \n(2) annual production after entry and (3) well exit.\nWe model new well entry by estimating a Poisson model of well \nentry using data on historical production from existing wells and fields, \ncosts and crude oil nominal prices. Specifically, we estimate annual new \nwell entry in an oil field as a function of oil prices, field-level capital and \noperational expenditures (Supplementary Figs. 2\u20134) and field-level \ndepletion. Details are provided in Supplementary Note 9. This model is \nestimated using well-entry data between 1977 and 2019 from California\u2019s \nDepartment of Conservation\u2019s WellSTAR database47. Supplementary \nNotes 1 and 3\u20135 provide more information on the input data. Capital \nand operational expenditure data are from the subscription-based data \nprovider Rystad Energy (Supplementary Note 2). Model estimates are \nprovided in Supplementary Table 1.\nAfter estimating the well-entry model, we predict annual well \nentries for the 2020\u20132045 projection period using forecasted nominal \nprices and prescribed policy conditions. Field-level operational costs \nare modified each year based on the relevant carbon and excise tax. The \nsetback policy constrains projected new well entry in a given field by \nreducing the number of predicted new wells by the percentage of field \narea covered by a setback. Figure 1 and Supplementary Fig. 5 compare \nthe predicted and observed entry at the state level and for each top \nfield category, respectively.\nTo predict annual oil production after well entry, we estimate \noil-production decline curves at the field and vintage level for both \nexisting (that is, pre-2020 entry) and new wells (that is, wells that enter \nduring 2020\u20132045). Production from oil wells often follow a declin-\ning profile of production until the wells exit48,49. For existing wells, we \nestimate the decline-curve parameters using historical oil-production \ndata (Supplementary Note 10) and apply them to the decline-curve \nequations to estimate future annual production at the field-vintage \nlevel. To predict future production from new wells, we extrapolate \nhistorical parameters using a linear regression model to obtain val-\nues for the 2020\u20132045 forecast period. In each forecast year for each \nfield, we use the corresponding extrapolated decline parameters and \n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n606\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\ndecline-curve equations to determine field-vintage-level production \nfrom the year the wells enter through the end of the projection period. \nWe repeat this process for all forecast years. Modelled production \ndecline curves and actual production for two fields are shown in Sup-\nplementary Figs. 6 and 7.\nBecause most wells that idle for a long time stop producing alto-\ngether50, we use historical data on wells that idled continuously for ten \nyears as a proxy for wells that stop producing and exit. We model well \nexits as a function of the nominal oil price, nominal field-level opera-\ntional costs and field-level depletion. We estimate the parameters of the \nmodel using historical data from 1977 to 2019 and apply the parameters \nto predict future well exit in the period 2020\u20132045, again modifying \nfield-level operational costs each year based on the relevant carbon and \nexcise taxes. Supplementary Note 11 provides details. Model estimates \nare provided in Supplementary Table 1. Supplementary Figs. 8 and 9 \ncompare the predicted and observed exit at the state level and for each \ntop field category, respectively.\nTo account for well exits and setbacks, we adjust the predicted \nproduction from both existing and new vintages. We assume that each \nwell in a given field-vintage produces the same amount of oil. Each \nyear the exit model predicts the number of wells that exit from each \nfield. We then remove these wells in order of vintage, starting with the \noldest. For vintages that experience well exit, future production is \ncorrespondingly decreased to account for the reduction in number of \nwells in production. Similarly, for existing vintages we adjust predicted \nproduction to account for wells prohibited from future production \ndue to setbacks by reducing production volumes proportionally by \nthe number of wells removed by the setback. Supplementary Note 8 \nprovides more details about the oil-production model.\nGHG emissions\nWe estimate GHG emissions associated with oil extraction using \nfield-specific GHG emissions factors. We first estimate historical GHG \nemissions factors using the Oil Production Greenhouse Gas Emis-\nsion Estimator (OPGEE) model v2.0 from the California Air Resources \nBoard51,52 (Supplementary Fig. 10 provides 2015 data). The OPGEE \nmodel is an engineering-based life-cycle assessment tool for the meas-\nurement of GHG emissions from the production, processing and trans-\nport of crude oil. Using the OPGEE model and oil-extraction data from \nthe California Department of Conservation, we model field-level GHG \nemissions for the years 2000, 2005, 2010, 2012, 2014, 2016 and 2018. \nWe consider only upstream emissions from exploration, drilling, crude \nproduction, surface processing, maintenance operations, waste treat-\nment/disposal and other small sources (as modelled by OPGEE). To \nobtain emissions factors for oil fields that were not modelled by OPGEE, \nwe apply the median emissions factors for the fields that were mod-\nelled, separated by the use of steam injection (Supplementary Note 12  \nprovides more information). To estimate the field-level GHG emis-\nsions for the projection period (2020\u20132045), we average the historical \nemissions factors for each year, again separated by fields based on the \nuse of steam injection. We then linearly regress the average emissions \nfactors and extrapolate over the projection period. Last, we apply the \npercent change in emissions factor between each forecast year to the \nfield-level historical emissions factors from 2018 onwards to determine \nfield-level emissions factors for each forecast year. Supplementary \nNote 12 provides more details.\nHealth impacts\nWe first estimate PM2.5 emissions from oil production for each oil-field \ncluster (set of oil fields clustered by geographical proximity; Sup-\nplementary Fig. 11) using average emissions factors obtained from \na nationwide US sample53 (Supplementary Table 2). Using average \nPM2.5 emissions factors is a limitation of the study due to the lack \nof field-specific PM2.5 emissions factors. In practice, actual emis-\nsions factors are probably highly heterogeneous across oil fields. \nEmissions-factor heterogeneity can arise from differences across PM2.5 \nemissions sources\u2014which include on-site fossil fuel combustion from \nprocessing plants, generators, pumps, compressors and drilling rigs, \nflaring, gas venting, dust from heavy vehicles and secondary formation \nfrom ambient conditions\u2014and across well vintages and operators53,54. \nWhether such heterogeneity is consequential for air-quality disparities \nshould be a subject of future research as field-level emissions data \nbecome available.\nNext, we model pollution dispersal using the Intervention Model \nfor Air Pollution (InMAP) to obtain PM2.5 concentration from oil pro-\nduction at the census-tract level for each projection year55. InMAP is a \nreduced-complexity dispersal model based on the Weather Research \nand Forecasting model coupled with Chemistry (WRF-Chem) that mod-\nels secondary PM2.5 concentrations developed by ref. 22. We followed the \nmethods of ref. 55 and ran InMAP individually for each cluster and pol-\nlutant combination to obtain a source receptor matrix for all the extrac-\ntion clusters. We then quantify the avoided mortality associated with \nchanges in ambient PM2.5 exposure at the census-tract level compared \nwith the BAU scenario56,57 using a mortality concentration-response \nfunction adapted from ref. 58. This function estimates avoided mortal-\nity using population projections (Supplementary Fig. 12), a baseline \nmortality rate from 2015, the percentage change in mortality associ-\nated with a 1\u2009\u03bcg\u2009m\u22123 increase in PM2.5 exposure (0.0058 from ref. 59)  \nand our estimated changes in ambient concentrations of PM2.5. \nLast, we estimate the monetized values of avoided mortality using a \nUS$9.4\u2009million (in 2019 dollars) value obtained from ref. 60. All mortal-\nity benefits are then summed over the 2020\u20132045 projection period \nand presented in net present value terms. Supplementary Notes 6  \nand 13 provide more details.\nLabour impacts\nWe quantify changes in employment and worker compensation using \nan economic input\u2013output model from IMPLAN61,62. IMPLAN uses over \n90 sources of employment data to construct measures of county-level \nemployment and compensation based on sector-specific revenue \ninputs. Supplementary Table 3 summarizes the input specifications for \nthe labour analysis. Oil production and oil prices from the projected \npathways serve as the inputs to IMPLAN, which then computes result-\ning employment in full-time equivalent job years and total employee \ncompensation supported by the oil and gas industry for each county \nwith active oil and gas operations in the state. IMPLAN uses fixed multi-\npliers to quantify local employment changes in the oil-extraction sector \n(\u2018direct\u2019), in sectors that provide inputs to oil extraction (\u2018indirect\u2019) and \nin sectors where these workers spend income (\u2018induced\u2019). Similar to \nother input\u2013output models, IMPLAN is based on a static framework \nwhere the underlying multipliers are fixed and do not change with \nthe economic environment, which is a limitation of this model. This \nimplies, for example, that inflation, changes in labour productivity and \ngeographical or temporal shocks to labour markets, all of which could \nbe the result of some of the supply-side policies we consider, cannot \nbe incorporated in the labour-market impact analysis. Supplementary \nNotes 7 and 14 provide more details.\nEquity impacts\nTo quantify distributional impacts, we use California\u2019s legal definition \nof a \u2018disadvantaged\u2019 community (DAC) using CalEnviroScreen, a scor-\ning system based on multiple-pollution exposure and socioeconomic \nindicators developed by the California Environmental Protection \nAgency36. The following indicators are considered for the DAC defi-\nnition: ozone concentration, PM2.5 concentration, diesel emissions, \npesticide use, toxic releases, traffic, drinking water quality, cleanup \nsites, groundwater threats, hazardous waste facilities, impaired water \nbodies, solid waste sites, asthma rate, cardiovascular disease rate, low \nbirth weight percent, educational attainment, housing burden, linguis-\ntic isolation, poverty percent and percent unemployed. A census tract \n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n607\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\nis considered disadvantaged if it has a CalEnviroScreen score above the \ntop 25th percentile (ref. 63). We calculate the DACs ratio of health and \nlabour impacts (that is, the share of impacts experienced by DACs) by \ncalculating the ratio of the impact experienced by DAC census tracts \nto the total statewide impact. Supplementary Note 18 provides more \ndetails. Supplementary Notes 19, 36 and 37 show the advantages of \nfiner spatial resolution analysis (census-tract level) and the errors that \nmay be introduced by a coarser analysis conducted at the county level, \nespecially in the ranking of equity outcomes.\nData availability\nData on assets and asset-level costs from Rystad Energy and employ-\nment and worker compensation data from IMPLAN are proprietary. \nAll other datasets are publicly available and were collected online \nfrom California Department of Conservation, US Energy Information \nAdministration, International Energy Agency, California Air Resources \nBoard, California Office of Environmental Health Hazard Assessment, \nCalifornia Department of Finance, the Environmental Benefits Mapping \nand Analysis Program - Community Edition (BenMAP-CE), National His-\ntorical Geographic Information System, Congressional Budget Office, \nInMAP and the US Census Bureau. All publicly available datasets are \navailable on Zenodo at https://doi.org/10.5281/zenodo.7742802 with \nthe exception of InMAP and BenMAP-CE data, which can be downloaded \ndirectly from the software. The Zenodo repository includes raw input \ndata files that are not proprietary, intermediate data files to run the \nmodels and final results files to create the figures. A detailed readme \nfile includes descriptions of all data used in the study. Source data are \nprovided with this paper.\nCode availability\nAll code used to conduct the study is available at https://github.com/ \nemlab-ucsb/ca-transport-supply-decarb.\nReferences\n1.\t\nOECD.Stat (OECD, 2022); stats.oecd.org\n2.\t\nCalifornia GHG Emissions Inventory Data (CARB, 2022); https:// \nww2.arb.ca.gov/ghg-inventory-data\n3.\t\nHensher, D. A. Climate change, enhanced greenhouse gas \nemissions and passenger transport\u2014what can we do to make a \ndifference? Transp. Res. Part D Transp. Environ. 13, 95\u2013111 (2008).\n4.\t\nSperling, D. & Eggert, A. California\u2019s climate and energy policy for \ntransportation. Energy Strategy Rev. 5, 88\u201394 (2014).\n5.\t\nCreutzig, F. et al. Transport: a roadblock to climate change \nmitigation? Science 350, 911\u2013912 (2015).\n6.\t\nMorrow, W. R., Marano, J., Hasanbeigi, A., Masanet, E. & Sathaye, J. \nEfficiency improvement and CO2 emission reduction potentials in \nthe United States petroleum refining industry. Energy 93, 95\u2013105 \n(2015).\n7.\t\nLepitzki, J. & Axsen, J. The role of a low carbon fuel standard in \nachieving long-term GHG reduction targets. Energy Policy 119, \n423\u2013440 (2018).\n8.\t\nJenn, A., Azevedo, I. L. & Michalek, J. J. Alternative-fuel-vehicle \npolicy interactions increase U.S. greenhouse gas emissions. \nTransp. Res. Part A Policy Pract. 124, 396\u2013407 (2019).\n9.\t\nAndersson, O. & B\u00f6rjesson, P. The greenhouse gas emissions of \nan electrified vehicle combined with renewable fuels: life cycle \nassessment and policy implications. Appl. Energy 289, 116621 \n(2021).\n10.\t Lazarus, M. & van Asselt, H. Fossil fuel supply and climate policy: \nexploring the road less taken. Climatic Change 150, 1\u201313 (2018).\n11.\t\nF\u00e6hn, T., Hagem, C., Lindholt, L., M\u00e6land, S. & Rosendahl, K. \nE. Climate policies in a fossil fuel producing country\u2019s: demand \nversus supply side policies. Energy J. 38, 77\u2013102 (2017).\n12.\t Kunce, M. Effectiveness of severance tax incentives in the U.S.  \noil industry. Int. Tax Public Finance 10, 565\u2013587 (2003).\n13.\t McCollum, D. L. et al. Energy investment needs for fulfilling the \nParis Agreement and achieving the Sustainable Development \nGoals. Nat. Energy 3, 589\u2013599 (2018).\n14.\t Schwanitz, V. J., Piontek, F., Bertram, C. & Luderer, G. Long-term \nclimate policy implications of phasing out fossil fuel subsidies. \nEnergy Policy 67, 882\u2013894 (2014).\n15.\t Larson, E. et al. Net-Zero America: Potential Pathways, \nInfrastructure, and Impacts (Princeton Univ., 2020).\n16.\t Williams, J. H. et al. Carbon-neutral pathways for the United \nStates. AGU Advances 2, e2020AV000284 (2021).\n17.\t Brown, A. L. et al. Driving California\u2019s Transportation Emissions to \nZero (Institute of Transportation Studies, Univ. of California, 2021); \nhttps://doi.org/10.7922/G2MC8X9X\n18.\t Mayfield, E., Jenkins, J., Larson, E. & Greig, C. Labor pathways to \nachieve net-zero emissions in the United States by mid-century. \nEnergy Policy 177, 113516 (2023).\n19.\t California\u2019s Draft 2022 Scoping Plan (CARB, 2022).\n20.\t Governor Newsom Takes Action to Phase Out Oil Extraction in \nCalifornia (Office of Governor Gavin Newsom, 2021); https://www. \ngov.ca.gov/2021/04/23/governor-newsom-takes-action-to-phase- \nout-oil-extraction-in-california/\n21.\t Deschenes, O. et al. Enhancing equity while eliminating emissions \nin California\u2019s supply of transportation fuels. Zenodo https://doi. \norg/10.5281/zenodo.4707966 (Environmental Markets Lab, Univ.  \nof California, 2021).\n22.\t Tessum, C. W., Hill, J. D. & Marshall, J. D. InMAP: a model for air \npollution interventions. PLoS ONE 12, e0176131 (2017).\n23.\t Squibb, J. & Thorvaldson, J. IMPLAN\u2019s Gravity Model and Trade \nFlow RPCs (IMPLAN Group, 2020).\n24.\t IMPLAN Data Sources (IMPLAN Group, 2020); https://implanhelp. \nzendesk.com/hc/en-us/articles/115009674448-IMPLAN-Data- \nSources\n25.\t Lewis, C., Greiner, L. H. & Brown, D. R. Setback distances for \nunconventional oil and gas development: Delphi study results. \nPLoS ONE 13, e0202462 (2018).\n26.\t Ferrar, K. People and Production: Reducing Risk in California \nExtraction Fractracker (Fractracker, 2020); https://www.\nfractracker.org/2020/12/people-and-production/\n27.\t Colorado Oil & Gas Conservation Commission Unanimously \nAdopts SB 19-181 New Mission Change Rules, Alternative Location \nAnalysis and Cumulative Impacts (COGCC, 2020).\n28.\t Draft Rule for Protection of Communities and Workers from Health \nand Safety Impacts from Oil and Gas Production Operations \nPre-rulemaking Release for Public Review and Consultation \n(Department of Conservation, 2022); https://www.conservation. \nca.gov/calgem/Documents/public-health/PHRM%20Draft%20 \nRule.pdf\n29.\t Annual Energy Outlook 2021\u2014Table: Table 12. Petroleum and Other \nLiquids Prices (EIA, 2021); https://www.eia.gov/outlooks/aeo/data/ \nbrowser/#/?id=12-AEO2021&region=0-0&cases=highprice~lowpric \ne~aeo2020ref&start=2019&end=2045&f=A&sourcekey=0\n30.\t WellSTAR Oil and Gas Well Monthly Production (Department of \nConservation, 2020); https://wellstar-public.conservation.ca.gov/ \nGeneral/Home/PublicLanding\n31.\t Technical Update of the Social Cost of Carbon for Regulatory \nImpact Analysis\u2014Under Executive Order 12866 (US Government \nInteragency Working Group on Social Cost of Greenhouse Gases, \n2016); https://www.epa.gov/sites/production/files/2016-12/ \ndocuments/sc_co2_tsd_august_2016.pdf\n32.\t Regulatory Impact Analysis for the Proposed Standards of \nPerformance for New, Reconstructed, and Modified Sources and \nEmissions Guidelines for Existing Sources: Oil and Natural Gas \nSector Climate Review (US EPA, 2021).\n33.\t The Benefits and Costs of the Clean Air Act from 1990 to 2020  \n(US EPA, 2011).\n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n608\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\n34.\t Neidell, M. Air Pollution and Worker Productivity (IZA World of \nLabor, 2017).\n35.\t Lobell, D. B., Di Tommaso, S. & Burney, J. A. Globally ubiquitous \nnegative effects of nitrogen dioxide on crop growth. Sci. Adv. 8, \neabm9909 (2022).\n36.\t Rodriquez, M. & Zeise, L. CalEnviroScreen 3.0 Report (California \nEPA & Office of Environmental Health Hazard Assessment, 2017).\n37.\t Final Regulation Order: California Cap on Greenhouse Gas \nEmissions and Market-Based Compliance Mechanisms (CARB, \n2018); https://ww3.arb.ca.gov/regact/2018/capandtrade18/ \nct18fro.pdf?_ga=2.258682314.1729598153.1606172336- \n1333792675.1605911480\n38.\t Oil Severance Tax States (Virginia Department of Taxation, 2020); \nhttps://www.tax.virginia.gov/sites/default/files/inline-files/Oil%20 \nSeverance%20Tax%20States%20Matrix.pdf\n39.\t Basic Information about Oil and Natural Gas Air Pollution Standards \n(US EPA, 2022); https://www.epa.gov/controlling-air-pollution- \noil-and-natural-gas-industry/basic-information-about-oil-and- \nnatural-gas#:~:text=In%20addition%20to%20helping%20 \nform,and%20other%20serious%20health%20effects\n40.\t Monasterolo, I. & Raberto, M. The impact of phasing out fossil \nfuel subsidies on the low-carbon transition. Energy Policy 124, \n355\u2013370 (2019).\n41.\t Erickson, P., Lazarus, M. & Piggot, G. Limiting fossil fuel \nproduction as the next big step in climate policy. Nat. Clim. \nChange 8, 1037\u20132043 (2018).\n42.\t Felder, S. & Rutherford, T. F. Unilateral CO2 reductions and carbon \nleakage: the consequences of international trade in oil and basic \nmaterials. J. Environ. Econ. Manage. 25, 162\u2013176 (1993).\n43.\t Sinn, H.-W. Public policies against global warming: a supply side \napproach. Int. Tax Public Finance 15, 360\u2013394 (2008).\n44.\t Ericson, S. J., Kaffine, D. T. & Maniloff, P. Costs of increasing oil and \ngas setbacks are initially modest but rise sharply. Energy Policy \n146, 111749 (2020).\n45.\t Elkind, E. N. & Lamm, T. Legal Grounds: Law and Policy Options \nto Facilitate a Phase-Out of Fossil Fuel Production in California \n(Berkeley Center for Law, Energy and the Environment, 2020); \nhttps://www.law.berkeley.edu/wp-content/uploads/2020/04/ \nLegal-Grounds.pdf\n46.\t 2020 Annual Auction Reserve Price Notice (CARB, 2019).\n47.\t Geologic Energy Management Division. GIS Mapping. California \nDepartment of Conservation https://www.conservation.ca.gov/ \ncalgem/maps (2020).\n48.\t EIA Production Decline Curve Analysis (EIA, 2020); https://www. \neia.gov/analysis/drilling/curve_analysis/\n49.\t Anderson, S. T., Kellogg, R. & Salant, S. W. Hotelling under \npressure. J. Polit. Econ. 126, 984\u20131026 (2018).\n50.\t Muehlenbachs, L. A dynamic model of cleanup: estimating sunk \ncosts in oil and gas production. Int. Econ. Rev. 56, 155\u2013185 (2015).\n51.\t El-Houjeiri, H. M., Masnadi, M. S., Vafi, K., Duffy, J. & Brandt, A. R. \nOil Production Greenhouse Gas Emissions Estimator OPGEE v2.0 \n(2017); https://eao.stanford.edu/research-project/opgee-oil- \nproduction-greenhouse-gas-emissions-estimator\n52.\t Duffy, J. Staff Report: Calculating Carbon Intensity Values of Crude \nOil Supplied to California Refineries (CARB, 2015); https://ww3.arb. \nca.gov/fuels/lcfs/peerreview/050515staffreport_opgee.pdf\n53.\t Jaramillo, P. & Muller, N. Z. Air pollution emissions and damages \nfrom energy production in the U.S.: 2002\u20132011. Energy Policy 90, \n202\u2013211 (2016).\n54.\t Gonzalez, D. J. et al. Upstream oil and gas production and ambient \nair pollution in California. Sci. Total Environ. 806, 150298 (2022).\n55.\t Goodkind, A. L., Tessum, C. W., Coggins, J. S., Hill, J. D. & Marshall, \nJ. D. Fine-scale damage estimates of particulate matter air \npollution reveal opportunities for location-specific mitigation of \nemissions. Proc. Natl Acad. Sci. USA 116, 8775\u20138780 (2019).\n56.\t Sacks, J. D. et al. The environmental Benefits Mapping and \nAnalysis Program\u2014Community Edition (BenMAP-CE): a tool \nto estimate the health and economic benefits of reducing air \npollution. Environ. Modell. Software 104, 118\u2013129 (2018).\n57.\t Estimating the Health Benefits Associated with Reductions in PM \nand NOx Emissions (CARB, 2019); https://ww2.arb.ca.gov/ \nsites/default/files/2019-08/Estimating%20the%20Health%20 \nBenefits%20Associated%20with%20Reductions%20in%20 \nPM%20and%20NOX%20Emissions%20-%20Detailed%20 \nDescription_0.pdf\n58.\t Shapiro, J. S. & Walker, R. Is Air Pollution Regulation Too Stringent? \nWorking Paper 28199 (National Bureau of Economic Research, \n2020); https://www.nber.org/papers/w28199\n59.\t Krewski, D. et al. Extended follow-up and spatial analysis of the \nAmerican Cancer Society study linking particulate air pollution \nand mortality. Res. Rep. 140, 5\u2013114 (2009).\n60.\t Mortality Risk Valuation (US EPA, 2014); https://www.epa.gov/ \nenvironmental-economics/mortality-risk-valuation\n61.\t Clouse, C. IMPLAN to FTE & Income Conversions (IMPLAN Group,  \n2019); http://implanhelp.zendesk.com/hc/en-us/articles/ \n115002782053\n62.\t Clouse, C. Understanding Labor Income (LI), Employee \nCompensation (EC), and Proprietor Income (PI) (IMPLAN Group, \n2020); https://implanhelp.zendesk.com/hc/en-us/articles/36002\n4509374-Understanding-Labor-Income-LI-Employee-Compensati\non-EC-and-Proprietor-Income-PI-\n63.\t California Communities Environmental Health Screening Tool \n(CalEnviroScreen 3.0) (California OEHHA, 2018); https://oehha.ca. \ngov/calenviroscreen/report/calenviroscreen-30\nAcknowledgements\nWe thank the state of California for supporting this work through the \nGreenhouse Gas Reduction Fund. The state of California assumes \nno liability for the contents or use of this study. The study does not \nreflect the official views or policies of the state of California. We also \nthank the California Environmental Protection Agency, California \nState Transportation Agency, California Air Resources Board, \nCalifornia Energy Commission, California Natural Resources Agency, \nCalifornia Workforce Development Board, California Department of \nConservation, California Governor\u2019s Office of Business and Economic \nDevelopment, California Office of Environmental Health Hazard \nAssessment and Office of Planning and Research and the California \nGovernor\u2019s Office of Planning and Research for providing support, \ndata and valuable feedback. We thank M. Clemence and E. O\u2019Reilly for \nmanaging and supporting the project.\nAuthor contributions\nR.D., P.W., K.C.M., O.D. and D.W.L. conceptualized the study and \nacquired the funding. R.D., P.W., K.C.M., O.D., D.H.-C., R.L., C.M., T.M., \nM.M. and V.T. developed the methodology and software. R.D., P.W., \nK.C.M., O.D., D.H.-C., R.L., C.M., T.M., M.M., S.S., V.T. and A.U. conducted \nthe formal analysis. D.H.-C., R.L., C.M., T.M., M.M., S.S., V.T. and A.U. \ncurated the data. R.D., P.W., K.C.M., O.D., D.W.L., D.H.-C., T.K., R.L., C.M., \nT.M., M.M. and V.T. wrote and edited the paper. K.C.M., O.D., D.W.L., P.W. \nand R.D. supervised the project.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s41560-023-01259-y.\nCorrespondence and requests for materials should be addressed to \nRanjit Deshmukh, Paige Weber or Kyle C. Meng.\n\nNature Energy | Volume 8 | June 2023 | 597\u2013609\n609\nArticle\nhttps://doi.org/10.1038/s41560-023-01259-y\nPeer review information Nature Energy thanks Peter Maniloff,  \nErin Mayfield and the other, anonymous, reviewer(s) for their \ncontribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) \nholds exclusive rights to this article under a publishing  \nagreement with the author(s) or other rightsholder(s); author \nself-archiving of the accepted manuscript version of this article  \nis solely governed by the terms of such publishing agreement  \nand applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2023\n\n\n    Scientific Research Findings:", "answer": "For a statewide\u00a02045\u00a0GHG target, we find that setbacks applied to new oil wells generate the largest health benefits in terms of avoided mortality from reduced particulate matter air pollution, but also the largest lost worker compensation. This is followed by excise taxes and carbon taxes. Setbacks also have the most favourable equity outcomes by yielding the highest share of health benefits and the lowest share of lost worker compensation borne by disadvantaged communities. However, even a\u00a01\u2011mile setback \u2014 the largest considered in this study and much larger than the\u00a03,200\u00a0ft\u00a0(975\u00a0m) currently proposed in California \u2014 may fail to meet California\u2019s ambitious\u00a090%\u00a0GHG reduction target by\u00a02045. Meeting this target will require combining setbacks applied to both existing and new oil wells with other supply\u2011side policies such as excise taxes and carbon taxes.", "id": 13}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Energy | Volume 7 | October 2022 | 989\u2013997  \n989\nnature energy\nhttps://doi.org/10.1038/s41560-022-01122-6\nAnalysis\nThe effect of European fuel-tax cuts on the oil \nincome of Russia\nJohan Gars\u2009\n\u200a\u20091, Daniel Spiro\u2009\n\u200a\u20092 and Henrik Wachtmeister\u2009\n\u200a\u20093\u2009\nFollowing Russia\u2019s invasion of Ukraine, there has been a surge in transport \nfuel prices. Consequently, many European Union (EU) countries are cutting \ntaxes on petrol and diesel to shield consumers. Using standard theory \nand empirical estimates, here we assess how such tax cuts influence the \noil income in Russia. We find that an EU-wide tax cut of \u20ac0.20\u2009l\u22121 increases \nRussia\u2019s oil profits by around \u20ac8\u2009million per day in the short and long term. \nThis is equivalent to \u20ac3,100\u2009million per year, 0.2% of Russia\u2019s gross domestic \nproduct or 5% of its military spending. We show that a cash transfer to EU \ncitizens\u2014with a fiscal burden equivalent to the tax cut\u2014reduces these side \neffects to a fraction.\nFollowing Russia\u2018s military attack on Ukraine, the European Union \n(EU) and United States have imposed a large number of sanctions \non Russia1,2. The attack has also led to a negative supply shock of oil, \npartly because Russia\u2019s ability to export has been hampered by the \nlack of will to insure Russian ships3, but also due to industry prepa-\nration for the upcoming EU oil import ban4. Together with surging \npost-pandemic demand, this has led to very high prices of transport \nfuels5,6. In response, a large number of European countries are either \ndiscussing or have already implemented a reduction in fuel taxes to \nhelp consumers cope with high prices. These include Austria, Belgium, \nFrance, Germany, Italy, the Netherlands, Romania and Sweden (see \nthe Methods section \u2018Fuel price and taxes\u2019 and, for example, refs. 7\u20139  \nfor details). Such tax reductions have problematic consequences \nsince they increase demand, thus making current supply even more \nscarce. Some of the tax reduction will be attenuated by an increase in \nthe underlying oil price, leading to increased profits for oil produc-\ners. Here, we assess the magnitude of this effect using basic theory \nand empirical estimates from the oil sector. We ask: \u2018How much does \nthe oil income in Russia increase following fuel-tax reductions in the \nEU?\u2019 Knowing the answer to this question is highly relevant to policy \nas Russia\u2019s oil profits may undermine the geopolitical interests of \nthe EU, reduce the effectiveness of the EU\u2019s sanctions and ultimately \nimprove the ability of Russia to wage war.\nHere, we show that an EU-wide fuel-tax cut equivalent to \u20ac0.20\u2009l\u22121 \nwould increase Russia\u2019s oil profits by \u20ac36\u2009million per day in the first \nmonth, \u20ac8.4\u2009million per day during the rest of the first year and \u20ac8.2\u2009mil-\nlion per day beyond the first year. The additional profits are equivalent \nto 0.2% of Russia\u2019s gross domestic product (GDP) and 5% of its defence \nspending. The fiscal cost to the EU would be \u20ac170\u2009million per day dur-\ning the first year. An alternative policy with an equivalent fiscal cost \nis studied as well: providing EU citizens with cash transfers. Such a \npolicy yields a fraction of the tax cut\u2019s profits to Russia and is ulti-\nmately more flexible for citizens as they can use the cash on anything  \nthey please.\nTheoretical approach\nHere, we describe how we derive the effects of a decreased fuel tax in \nthe EU on Russian oil profits. Our analysis uses a standard model of \nsupply and demand for oil. Our approach is similar to those of Erickson \nand Lazarus10 and Faehn et al.11. To analyse the effects of the EU\u2019s tax, \nwe distinguish between oil demand for road transport fuels in the EU \nand remaining global oil demand. Similarly, to focus on the effects on \nRussian revenues, we distinguish between oil supply from Russia and \nsupply from the rest of the world. We also consider the alternative \npolicy of income transfers to households.\nThe first step is to derive how the global oil market responds to \nchanges in the EU\u2019s road transport fuel tax. More detailed derivations \nare provided in the Methods. The global per-unit crude oil price is \ndenoted p. To make the crude oil usable as fuel for end users, it must \nbe refined, transported and so on. For the EU, we assume a per-unit cost \nc for this (in the Methods, we consider analytically the case where these \ncosts are instead proportional to the oil price, and assess this case \nquantitatively in Supplementary Note 3). Additionally, road fuel users \nin the EU pay a fuel excise tax \u03c4 per unit of fuel, as well as the value-added \ntax (VAT) rate vEU. The market equilibrium is found by equating the \ndemand for crude oil for road transport in the EU, \nReceived: 28 April 2022\nAccepted: 15 August 2022\nPublished online: 15 September 2022\n Check for updates\n1Beijer Institute, The Royal Swedish Academy of Sciences, Stockholm, Sweden. 2Department of Economics, Uppsala University, Uppsala, Sweden. \n3Department of Earth Sciences, Uppsala University, Uppsala, Sweden. \n\u2009e-mail: henrik.wachtmeister@geo.uu.se\n\nNature Energy | Volume 7 | October 2022 | 989\u2013997 \n990\nAnalysis\nhttps://doi.org/10.1038/s41560-022-01122-6\nwith respect to \u03c4 and making a linear approximation gives that Russian \noil profits change by\n\u0394\u03c4\u03c0RU \u2248(1 +\np\u2212e\np \u03b5S,RU) SRU (p) \u0394\u03c4p.\n(6)\nAs an explicit aim of the tax cut is to shield consumers from increas-\ning fuel costs, an alternative way to do so is to give general income \ntransfers to people that correspond to the reduction in tax revenues \nthat would result from a decreased fuel tax. From a welfare perspective, \nthis is preferable since people can then choose how to use the money. \nTo the extent that the tax cut is supposed to appease particular groups, \nit is also possible to direct the income transfers to these groups. Such \noptions include giving money to all car owners (not based on their \ndriving), giving money to particular regions where the population is \nmore reliant on cars (for example, rural areas) or reimbursing the tax \ncollected in a region or municipality to that same region or munici-\npality. Another rationale to uphold fuel taxes is to internalize climate \ndamages from fossil fuel use (that is, Pigouvian taxes). From the per-\nspective of this Analysis, an additional effect of an income transfer is \nthat a share smaller than one will be spent on fuel; hence, the increase \nin Russian oil profits will be smaller. How much smaller will be assessed \nquantitatively.\nTo analyse this alternative policy option, we assume that road fuel \ndemand in the EU now depends on disposable income I in addition to \nthe fuel price. Differentiating the market equilibrium condition with \nrespect to I and treating the equilibrium price as a function of income, \nwe get that the oil price change due to an income change is\ndp\ndI = p\nI\nx\u03b5I,EU\ny\u03b5S,RU + (1 \u2212y) \u03b5S,ROW \u2212x\np\np+c+\u03c4 \u0303\u03b5D,EU \u2212(1 \u2212x) \u03b5D,ROW\n,\n(7)\nwhere \u03b5I,EU is the income elasticity of road fuel demand in the EU. The \neffects of a change in the disposable income \u0394I on the oil price, p, and \nthe fuel price in the EU, f, are\n\u0394Ip \u2248\ndp\ndI \u0394I and \u0394If \u2248(1 + vEU) \u0394Ip.\n(8)\nThe effect on Russian oil profits due to the income transfer is\n\u0394I\u03c0RU \u2248(1 +\np\u2212e\np \u03b5S,RU) SRU\u0394Ip.\n(9)\nDEU((1 + vEU)(p + c + \u03c4)) (in many EU countries, VAT also applies to the \nexcise tax; hence, (1\u2009+\u2009vEU) multiplies \u03c4), and the remaining global \ndemand for crude oil, DROW(p), to the supply from Russia, SRU(p), and \nfrom the rest of the world, SROW(p):\nDEU ((1 + vEU) (p + c + \u03c4)) + DROW (p) = SRU (p) + SROW (p)\n(1)\nSince our focus here is on tax cuts on transport fuel, DEU should be \nunderstood as the demand for crude oil to be used as oil-based road \ntransport fuel in the EU (that is, mainly petrol and diesel). With some \nabuse of technical differences, we will often refer to it only as fuel. DROW \nshould be understood as the global demand for all other oil products \nexcept road fuel in the EU (that is, it also includes non-road oil products \nin the EU). We assume that the EU\u2019s crude demand depends on the fuel \nprice, including costs and taxes, while we assume that rest of the world \ndemand depends on the oil price.\nThe effect of a change in the tax on the equilibrium price can then \nbe found by treating the price as a function of the tax, differentiating \nthe equilibrium condition fully with respect to the tax, solving for the \nderivative of the price with respect to the tax and rewriting. Let x denote \nthe share of global demand for oil that comes from road fuel demand \nin the EU and let y denote Russia\u2019s share of the global oil supply. The \nmarket response to a change in the tax depends on the supply and \ndemand elasticities in the submarkets. Let \u0303\u03b5D,EU, \u03b5D,ROW, \u03b5S,RU and \u03b5S,ROW \ndenote the demand elasticities in the EU and the rest of the world and \nthe supply elasticities of Russia and the rest of the world, respectively. \nThe demand elasticity for the EU measures the response in demanded \nquantities of oil to changes in the end user prices, including additional \ncosts and taxes, while the other elasticities measure responses of quan-\ntities of oil to changes in the oil price. Using this notation, the effect of \na tax change on the oil price is given by the derivative\ndp\nd\u03c4 =\nx\np\np+c+\u03c4 \u0303\u03b5D,EU\ny\u03b5S,RU+(1\u2212y)\u03b5S,ROW\u2212x\np\np+c+\u03c4 \u0303\u03b5D,EU\u2212(1\u2212x)\u03b5D,ROW .\n(2)\nNote the ratio multiplying the EU demand elasticity. This ratio \ncorrects for the fact that demand depends on the price including the \nsupply costs and taxes and hence that a change in the global oil price p \nwill have a smaller percentage effect on consumer prices (see Methods \nfor further discussion of this).\nThe effect of a tax change \u0394\u03c4 on the oil price can be approximated \nlinearly as\n\u0394\u03c4p \u2248\ndp\nd\u03c4 \u0394\u03c4.\n(3)\nLet the fuel price in the EU be denoted f \u2261(1 + vEU) (p + c + \u03c4). The \nchange in the fuel price is\n\u0394\u03c4f \u2248(1 + vEU) (\u0394\u03c4p + \u0394\u03c4) .\n(4)\nThe EU\u2019s fuel tax revenues are TEU \u2261( vEU(p + c)+ (1 + vEU)\u03c4 ) DEU \nand the fiscal cost of the tax change (that is, the lost tax revenue) can \nbe found by differentiating with respect to \u03c4\u2014taking into account that \nthe oil price depends on the tax\u2014and making a linear approximation. \nThe fiscal cost is then\n\u0394\u03c4TEU \u2248[1 + (vEU +\n\u03c4+vEU(p+c+\u03c4)\np+c+\u03c4\n\u03b5D,EU) (1 +\ndp\nd\u03c4 )] DEU\u0394\u03c4.\n(5)\nFinally, we translate an oil price change into a change in Russian \noil profits. The oil profits of Russia are \u03c0RU = (p \u2212e)SRU(p), where e \nrepresents the oil extraction costs, which are assumed to be constant \nper unit. This assumption is realistic under the production changes \nconsidered here. Again, treating p as a function of \u03c4, differentiating \nBarrels of oil\nPrice of oil\nDEU\nDROW\nDEU+ROW\nS\nFig. 1 | Illustration of supply and demand in the long term. Oil demand for \nroad transport in Europe (DEU) increases by a tax cut (the DEU demand line shifts \nto the right from the solid to the dashed line position). Consequently, total world \noil demand (DEU+ROW) increases by an equal amount (the DEU+ROW line shifts to the \nright from the solid to the dashed line position). World oil supply (S) is somewhat \nelastic and a new, higher, equilibrium price is attained at a higher quantity level \n(dashed lines).\n\nNature Energy | Volume 7 | October 2022 | 989\u2013997  \n991\nAnalysis\nhttps://doi.org/10.1038/s41560-022-01122-6\nData and estimates vary over different time \nhorizons\nWe quantitatively assess the effects described theoretically in three \ndifferent cases: the very short term, the short term and the long term. \nHere, we sketch the qualitative effects in the different time horizons. \nThe size of the effects depends on the numerical values, which are pre-\nsented at the end of this section. We start by describing the long-term \neffects. Note that we use long term to describe the effects beyond 1\u2009year, \nto be thought of as in 1\u20133\u2009years but not more. Other studies deriving \nprice elasticity estimates often use long term to describe effects on \nlonger time horizons. In the Methods, we discuss our selection and \nrange of elasticity estimates in more detail and how they relate to the \ntime horizon.\nIn the long term, supply is somewhat elastic and demand is rather \nelastic too. This is because producers have time to adjust their produc-\ntion and make some capacity investments. Likewise, consumers can \nacquire new habits or find solutions based on a new fuel price, and \nthose in the process of buying a new vehicle will take the fuel price into \naccount (see, for example, ref. 12). This is the case illustrated in Fig. 1. \nThe grey lines show the demand from the EU and the rest of the world, \nrespectively. The black lines represent global demand and supply. A \ntax reduction in the EU shifts the EU\u2019s demand outwards (dashed grey \nline). This in turn increases global demand by an equivalent amount \n(dashed black line). There are two effects of this: an increase in the oil \nprice (a shift along the vertical axis) and an increase in the quantity of \noil produced (a shift along the horizontal axis).\nIn the short term, to be thought of as 1\u201312\u2009months, the supply \nelasticity in terms of quantity is lower than in the long term. This can \nbe seen in Fig. 2a, where supply is illustrated as fixed. Demand elastic-\nity is lower than in the long term since most of the consumer choice \nregards how much to drive rather than what vehicle to buy or how \nto change long-term habits. Since, as in the figure, supply is fixed, a \nreduced tax results in increased oil price but no increase in production \nor consumption. In practice, some of the increased demand from the \nEU is attenuated by decreased consumption in the rest of the world.\nIn the very short term, to be thought of as up to 1\u2009month, there \nare limitations on how much oil, which was originally meant for other \nmarkets, can be redirected quickly to the EU. The reason for this is that \nbilateral contracts of supply can be viewed as partly fixed; oil tankers \non their way to one country cannot in the very short term easily be \nsent elsewhere. To capture this, we model this as though the EU is an \nisolated oil market. The supply is fixed, both in terms of quantity and \nwho supplies the oil. Implicitly, this also means that the oil price in the \nEU may differ from the oil price elsewhere. This case is illustrated in \nFig. 2b. Importantly, here Russia is relatively a much larger supplier \nthan on the global market. Demand is also less elastic than in the short \nterm. We view our modelling here as a thought experiment. Reality in \nthe very short term probably lies somewhere in between our modelling \nhere and the short-term scenario described above. One simplification \nand limitation of our model is that we do not consider oil invento-\nries. We discuss the potential effects of an EU import embargo in the  \nnext section.\nThe parameters, quantities and shares in equations (1)\u2013(9) are \nbased on previous research and current data. They are reported in Table \n1. Where relevant, we distinguish between very-short-term, short-term \nand long-term elasticities and shares. Motivations for the values are \nprovided in the table, with further information available in Methods.\nIn a sensitivity analysis in Supplementary Note 2, we perturb the \nkey parameters to show how this affects our results.\nEffects of fuel-tax cuts on Russian oil income\nHere, we analyse the effects of a fuel-tax cut in the very short term, the \nshort term and the long term. The tax cut considered is equivalent to \n\u20ac0.20\u2009l\u22121 including VAT (that is, it amounts to \u20ac0.2/(1\u2009+\u2009VAT)). This is \nbased on a weighted average of currently announced tax cuts in EU \ncountries (equivalent to roughly 10% of the price) and on the possibil-\nity that all countries cut the taxes to the EU\u2019s minimum level (see the \nMethods section \u2018Fuel price and taxes\u2019 for details).\nThe results in the very short term are presented in the top row of \nTable 2. Of the 20\u2009cents of tax reduction, 7\u2009cents are passed through to \noil suppliers. Russia, being an important supplier, attains a large share \nof the fiscal cost of the policy, making an additional \u20ac36\u2009million per \nday. Apart from financing Russia, the policy is also quite ineffective \nin lowering consumer prices in the very short term; consumers only \nexperience 12\u2009cents of reduction per litre despite the tax reduction \nbeing 20\u2009cents. The results here take into account Russia\u2019s current \nreduced supply to the EU (see the Methods section \u2018Size of markets \nand Russian export declines\u2019).\nIn the short-term case, to be thought of as the remaining part of \nthe first year, the consumer price in the EU is reduced by almost the \nfull tax reduction and the now global oil price is increased by much \nless than in the very short term. Nevertheless Russia\u2014a large supplier \nBarrels of oil\nBarrels of oil\nPrice of oil\nPrice of oil\nb\na\nDEU\nDROW\nDEU\nSEU\nDEU+ROW\nS\nFig. 2 | Illustration of supply and demand in the short term and in the very \nshort term. a, In the short term, oil demand for road transport in the EU (DEU) \nincreases by a tax cut (the DEU demand curve shifts to the right from the solid \nto the dashed line position). Consequently, total world oil demand (DEU+ROW) \nincreases by an equal amount (the DEU+ROW demand curve shifts to the right from \nthe solid to the dashed position). World oil supply (S) is fixed (inelastic) and a \nnew, higher, equilibrium price is attained (dashed line) at the same quantity level. \nb, In the very short term, oil demand for road transport in Europe (DEU) increases \nby a tax cut as in the short term, but due to transport rigidities, the EU is modelled \nas an isolated market. Hence, the total demand equals EU demand (DEU). Supply \nto the EU (SEU) is fixed (inelastic) and a new, higher, equilibrium price is attained \n(dashed line) at the same quantity level.\n\nNature Energy | Volume 7 | October 2022 | 989\u2013997 \n992\nAnalysis\nhttps://doi.org/10.1038/s41560-022-01122-6\nalso globally\u2014is still receiving sizable additional profits (\u20ac8.4\u2009million \nper day or \u20ac3.1\u2009billion in year equivalents).\nIn the long-term case, to be thought of as beyond 1\u2009year and up to \n3\u2009years, supply becomes somewhat elastic and so does demand. The \nprice effects are again smaller and the fiscal cost to the EU is smaller \nthan in the very short and short term. This is because, instead of an oil \nprice increase, there is an increase in the supply. Russia\u2019s additional oil \nprofits are still sizeable at \u20ac8.2\u2009million per day or \u20ac3.0\u2009billion per year.\nIn Supplementary Note 2, we carry out a sensitivity analysis of \nthese results with respect to other parameter values. The results are \ngenerally not very sensitive in the very short term. The long-term and \nshort-term results are more sensitive. The sensitivity analysis sug-\ngests that Russia\u2019s short-term profit gains can be one-third compared \nwith those using our preferred parameter values (reported here in \nthe main text). However, they may also be around 70% higher. We also \ninvestigate whether the discount on Russian oil (through the Urals \nprice) is likely to change our results (Supplementary Note 4) and let the \ncosts c be proportional to the oil price (Supplementary Note 3). With \nregard to the Urals price, while the time window is too short to infer \nthe long-term effects, our analysis suggests that an increased global oil \nprice (for example, due to a tax cut, as considered here) will lead to an \nequivalent increase in the Urals price. Hence, our results are probably \nnot affected by this discount. We refer the reader to Supplementary \nNote 4 for more details.\nA final issue of robustness is whether the results change due to \nan import embargo. On 3 June 2022, the EU announced an extensive \nbut incomplete import embargo on Russian oil4 to be imposed with \na delay of several months. Had the embargo been imposed before \nthe tax cuts, the particular effects relying on the rigidity of transport \nwould disappear. It would make the mechanisms and results in the very \nshort term identical to those in the short term, which are independent \nof whether an oil import embargo exists or not. However, since the \nimport embargo will take force with a long delay and since most of the \ntax cuts have already been implemented, we view the very short-term \nresults presented here as an assessment of what has possibly already \nTable 1 | Parameters for equations (1)\u2013(9) describing Russia\u2019s oil income\nParameter\nVariable\nVery short term\nShort term\nLong term\nReference\nDemand elasticity of road fuel in the EU with \nrespect to price\n\u0303\u03b5D,EU\n\u22120.25\n\u22120.25\n\u22120.9\nSee Methods\nDemand elasticity of oil globally (excluding \nEU road fuel) with respect to price\n\u03b5D,ROW\n\u22120.125\n\u22120.125\n\u22120.45\nSee Methods\nSupply elasticity of Russian oil with respect \nto price\n\u03b5S,RU\n0\n0\n0.13\nSee Methods\nSupply elasticity of global oil (excluding \nRussia) with respect to price\n\u03b5S,ROW\n0\n0\n0.13\nSee Methods\nDemand elasticity of road fuel in EU with \nrespect to income\n\u03b5I,EU\n1\n1\n1\nRef. 15\nBase fuel excise tax in the EU (\u20ac\u2009l\u22121)\n\u03c4\n0.55\n0.55\n0.55\nWeighted average of all countries\u2019 \ntaxes; see Methods\nBase oil price per litre of fuel (\u20ac\u2009l\u22121)\np\n0.58\n0.58\n0.58\nOil price\u2009=\u2009$110 per barrel; foreign \nexchange rate\u2009=\u2009\u20ac0.9 per $; 170\u2009l fuel \nper barrel\nOther costs, such as refining (\u20ac\u2009l\u22121)\nc\n0.45\n0.45\n0.45\nFollows from p\u2009+\u2009\u03c4\u2009+\u2009c; see Methods\nVAT rate\nvEU\n0.2\n0.2\n0.2\nBase consumer price excluding VAT (\u20ac\u2009l\u22121)\np\u2009+\u2009\u03c4\u2009+\u2009c\n1.58\n1.58\n1.58\nSum of oil price, fuel tax and other \ncosts; see Methods\nOil extraction cost in Russia (\u20ac per barrel)\ne\n17\n17l\n17\nRystad UCube database14\nEU road fuel demand share of global oil \nconsumption (%)\nx\n48\n5.7\n5.7\nRefs. 31,32\nRussian oil exports per day (million litres of \nfuel; million barrels of oil)\nSRU\n540; 3.2\n1,200; 7\n1,400; 8\nIEA38 and case assumptions; see \nMethods\nGlobal oil supply, excluding Russian exports, \nper day (million litres of fuel; million barrels \nof oil)\nSROW\n13,000; 7.6\n16,000; 92\n16,000; 92\nFollows from SRU,S\nRoad fuel demand in the EU per day (million \nlitres of fuel; million barrels of oil)\nDEU\n870; 5.1\n960; 5.6\n970; 5.7\nFollows from x, D\nGlobal oil demand, excluding EU road fuel \ndemand, per day (million litres of fuel; \nmillion barrels of oil)\nDROW\n950; 5.6\n16,000; 93.4\n16,000; 94.3\nFollows from x, D\nRussia\u2019s exports as a share of the global \nsupply (%)\ny\n29\n7\n8\nFollows from SRU,S\nTable 2 | Effect of an EU fuel-tax cut of \u20ac0.20\u2009l\u22121\nOil price \nchange \n(cents per \nlitre)\nEU fuel \nprice \nchange \n(cents per \nlitre)\nFiscal cost \nto the EU \n(million \nEuros per \nday)\nProfit gain \nfor Russia \n(million \nEuros per \nday)\nProfit gain \nfor Russia \n(million \nEuros per \nyear)\nVery short \nterm\n6.7\n\u221212\n150\n36\n\u2013\nShort term\n0.7\n\u221219\n170\n8.4\n3,100\nLong term\n0.55\n\u221219\n115\n8.2\n3,000\n\nNature Energy | Volume 7 | October 2022 | 989\u2013997  \n993\nAnalysis\nhttps://doi.org/10.1038/s41560-022-01122-6\nmaterialized. It should be noted that the results for the short and long \nterm are driven by an increase in the global oil price due to higher EU \ndemand. Furthermore, in such time frames, transport routes can be \nadjusted. Hence, the results in the short and long term are independ-\nent of whether or not the EU imposes an import embargo on Russia. As \nlong as there are sufficiently many or large buyers of Russian oil, the \nmarket will adjust the trade flows.\nAre the additional profits for Russia large? We now discuss this \nfrom a few different perspectives. First note that in the very short term, \na large share of the EU\u2019s fiscal cost (24%) is sent to Russia. In the short \nterm and long term, much less is sent, but still around 5\u20137% of what is \nmeant to help European consumers is instead going to Russia.\nThe additional Russian profits are sizeable compared with Rus-\nsia\u2019s pre-invasion GDP, which was about \u20ac3.7\u2009billion per day. The EU\u2019s \ntax cut increases Russia\u2019s GDP by ~1% in the very short term versus \n0.2% in the short term and the long term. We can also compare them \nwith Russia\u2019s military spending, which was about \u20ac160\u2009million per day \npre-invasion (based on ref. 13, the average yearly military spending in \n2015\u20132020 was US$65 billion; a $ to \u20ac exchange rate of 0.9 makes this \n\u20ac160\u2009million per day). The daily profit increase then corresponds to \n23, 5 and 5% of military spending in the very short term, short term and \nlong term, respectively.\nAlmost all of these revenues stay in the Russian economy as 93% of \nRussian production is owned by Russian companies14 (both state owned \nand privately owned). Rosneft and Gazprom\u2014the two main majority \nstate-owned companies\u2014produce ~46% of Russia\u2019s oil. The average \ngovernment take for oil (that is, combined state taxes, tariffs and so \non) in Russia is 50% of total revenues14. Hence, almost all oil revenues \nstay in the Russian economy while 50% goes directly to the Russian \nstate as taxes and fees and an additional 23% goes to state-controlled \ncompanies.\nFinally, it should be noted that the effects are linear in the size of the \ntax cut. This implies that, should the tax cut be twice as large, the Russian \nadditional profits will be twice as large too, and vice versa in case the tax \ncut is half of what we study. This follows from our linear approximation \n(see the section \u2018Theoretical approach\u2019). This is a reasonable assump-\ntion as long as the effects on aggregate global demand and supply are \nsmall, which indeed is the case in both the short- and long-term cases. \nIn the very short term (where the EU is a submarket), changes in the tax \ncut may not scale linearly. This is because, if the tax cut is sufficiently \nlarge, the EU demand change is sizeable. Hence, a doubling of the tax cut \ncannot be assumed to imply doubling of Russian profits; it can be more \nand it can be less than doubling. One issue that plays a role is whether the \nelasticities are constant over the demand and supply curves. To date, \nthis is a largely unresolved question in the literature.\nEffects of an alternative direct cash transfer policy\nAs seen, a fuel-tax cut in the EU provides Russia with a large additional \nincome. Furthermore, part of the help meant for EU fuel consumers in \nthe form of a tax cut is instead passed through to increased oil prices, \nespecially in the very short term. The question is then whether it is pos-\nsible to help EU consumers in a way that does not benefit Russia. Here, \nwe look at one such alternative, namely in the form of a cash transfer \nto consumers with an equivalent fiscal burden to the tax cut. That is, \nwe tie the transfer to \u20ac150\u2009million, \u20ac170\u2009million and \u20ac115\u2009million in the \nvery short, short and long term, respectively.\nThe results are presented in Table 3. The increased income leads \nto an increased demand for fuel. However, since the cash can be spent \non anything, most of it is used for other things. Hence, the fuel price \nincreases only marginally (1.1\u2009cents in the very short term and much \nless in the long term). One perspective on this, which highlights the \nmain benefit of cash transfers compared with fuel-tax cuts, is that \nconsumers, through these policies, receive the equivalent of 20\u2009cents \nper litre of fuel consumed on average. To receive the benefits of a tax \ncut, consumers have to buy fuel. If they receive it in the form of cash \ninstead, they can choose to spend it all on fuel, to spend none of it on \nfuel or somewhere in between; cash is a more flexible currency than a \ntax cut. Even a person who spends the whole transfer on fuel will gain \nfrom a cash transfer since the fuel price increases only by 1.1\u2009cents. \nTherefore, this allows for varying preferences in the population (it is \nalso possible to direct the cash to particular groups that are hit harder \nby the price increase; see \u2018Theoretical approach\u2019). Another benefit of a \ncash transfer is that it avoids decreasing the fuel tax (that is, a Pigouvian \ntax), thus abating climate concerns.\nPerhaps most importantly for the subject matter here, Russia\u2019s \nprofit gains are substantially lower with the cash transfer (in the short \nterm ~15% of the profits received from a tax cut and in the long term \nmuch less). It can be noted that Russia\u2019s profits under a cash transfer are \nalso substantially lower when compared with the lowest profits of a tax \ncut considered in the sensitivity analysis (see Supplementary Note 2).\nThe conclusion that an income transfer yields much lower Russian \nprofits is not likely to change even if we consider other expenditures \nof households in the EU. The reason for this is that an income elastic-\nity of 1, as we use here15, implies that fuel\u2019s share of income is constant \nwhen income increases. Put differently, since private road-transport \nfuels correspond to 7% of consumer spending in the EU (figure for \n2008 from ref. 16), only 7% of the cash transfer would go to road fuel. \nTransport accounts for 13.2% of households\u2019 expenditures17, not all of \nwhich goes to oil products. Elasticity for other transport is not very \ndifferent from that for road fuels (for aviation income, elasticity is \naround 1 (ref. 18)). Hence, the amount of cash transfer that may go to oil \nproducts is effectively capped by oil\u2019s cost share in the EU. Elasticities \nfor other categories of consumer spending vary. For food, elasticity is \naround 0.5 in the EU19 and it accounts for 14.8% of expenditures, while \nhousing-related expenditures correspond to 25.7% (a small share of \nwhich is gas for heating). Various smaller categories correspond to \nthe remaining 48% (ref. 20).\nConclusion\nWe have analysed how much a fuel-tax cut in the EU will increase Rus-\nsian income from oil. The effects are substantial. A tax cut of \u20ac0.20 \nincreases Russia\u2019s daily oil profits by \u20ac8.4\u2009million during the first year \nand \u20ac8.2\u2009million if it remains longer (in the very short term, the daily \nprofit increase can be substantially higher).\nWe show that a fiscally equivalent cash transfer can achieve similar \nalleviation to consumers to a tax cut but with a fraction of the increased \nprofits for Russia.\nMethods\nDerivation of model\nHere, we provide derivations of the model used in the analysis. We \ndistinguish between the demand for crude oil for vehicle fuel in the EU, \nDEU, and the remaining oil demand DROW. Note that the rest of the world \ndemand includes oil demand in the EU for uses other than vehicle fuel. \nSince we want to study the effects of taxes on vehicle fuel in the EU, we \nexplicitly consider these and assume that the demand for oil for fuel in \nthe EU depends on the price, including refinement, transportation and \nTable 3 | Effect of a fiscal equivalent cash transfer\nOil price \nchange \n(cents per \nlitre)\nEU fuel \nprice \nchange \n(cents per \nlitre)\nFiscal cost \nto the EU \n(million \nEuros per \nday)\nProfit gain \nfor Russia \n(million \nEuros per \nday)\nProfit gain \nfor Russia \n(million \nEuros per \nyear)\nVery short \nterm\n0.91\n1.1\n150\n4.9\n\u2013\nShort term\n0.11\n0.13\n170\n1.3\n470\nLong term\n0.016\n0.019\n115\n0.24\n88\n\nNature Energy | Volume 7 | October 2022 | 989\u2013997 \n994\nAnalysis\nhttps://doi.org/10.1038/s41560-022-01122-6\ntaxes. Let p denote the crude oil price, v the VAT rate and \u03c4 the per-unit \ntax on vehicle fuel in the EU. The costs for refinement and transpor-\ntation could have one per-unit component, c, and one component \nproportional to the price, z. We thus have the fuel price\nf = (1 + v) ((1 + z) p + c + \u03c4)\n(10)\nand crude demand in the EU depends on this price and on income I, \nDEU(f, I). The rest of the world demand depends directly on the crude \noil price DROW(p). In the baseline model, we focus on per-unit costs, \ncorresponding to z\u2009=\u20090. In the Supplementary Information, we instead \nconsider completely proportional costs, corresponding to c\u2009=\u20090.\nThe supply is a function of the crude oil price and we distinguish \nbetween oil supply from Russia, SRU(p), and supply from the rest of the \nworld, SROW(p).\nThe world market price of crude oil is determined by the equilib-\nrium in the global oil market:\nDEU (f, I) + DROW (p) = SRU (p) + SROW (p) .\n(11)\nModel of fuel-tax cut\nTreating the equilibrium price p as a function of \u03c4, and differentiating \nboth sides fully with respect to \u03c4, taking equation (10) into account, \ngives\n(1 + v)\n\u2202DEU(f,I)\n\u2202f\n((1 + z)\ndp\nd\u03c4 + 1) + D\u2032\nROW (p)\ndp\nd\u03c4 = (S\u2032\nRU (p) + S\u2032\nROW (p))\ndp\nd\u03c4\n(12)\nLet D and S denote total demand and supply and let x and y denote \nshares of demand and supply as given by\nDEU = xD, DROW = (1 \u2212x) D, SRU = yS and SROW = (1 \u2212y) S.\n(13)\nMultiplying the left- and right-hand sides of equation (12) by p/D \nand p/S, respectively (with S\u2009=\u2009D in equilibrium), yields\ndp\nd\u03c4 =\n(1 + v)\np\nD\n\u2202DEU(f,I)\n\u2202f\np\nS S\u2032\nRU (p) +\np\nS S\u2032\nROW (p) \u2212(1 + v) (1 + z)\np\nD\n\u2202DEU(f,I)\n\u2202f\n\u2212\np\nD D\u2032\nROW (p)\n.\nUsing equation (13) and multiplying the terms containing the \nderivative of DEU by f/f, we get\ndp\nd\u03c4 =\n(1+v)p\nf\nxf\nDEU\n\u2202DEU(f,I)\n\u2202f\nyp\nSRU S\u2032\nRU (p) +\n(1\u2212y)p\nSROW S\u2032\nROW (p) \u2212\n(1+v)(1+z)p\nf\nxf\nDEU\n\u2202DEU(f,I)\n\u2202f\n\u2212\n(1\u2212x)p\nDROW D\u2032\nROW (p)\n.\nUsing equation (10), we arrive at\ndp\nd\u03c4 =\nx\np\n(1+z)p+c+\u03c4 \u0303\u03b5D,EU\ny\u03b5S,RU + (1 \u2212y) \u03b5S,ROW \u2212x\n(1+z)p\n(1+z)p+c+\u03c4 \u0303\u03b5D,EU \u2212(1 \u2212x) \u03b5D,ROW\n,\n(14)\nwhere we have the price elasticities of supply\n\u03b5S,RU \u2261\np\nSRU S\u2032\nRU (p) and \u03b5S,ROW \u2261\np\nSROW S\u2032\nROW (p)\n(15)\nand demand\n\u0303\u03b5D,EU \u2261\nf\nDEU\n\u2202DEU(f,I)\n\u2202f\nand \u03b5D,ROW \u2261\np\nDROW D\u2032\nROW (p) .\n(16)\nWe can differentiate the fuel price from equation (10) with respect \nto \u03c4 to get\ndf\nd\u03c4 = (1 + v) ((1 + z) dp\nd\u03c4 + 1) .\nThe changes in the oil price p and the EU fuel price f induced by a \nchange \u0394\u03c4 in the tax can be linearly approximated as\n\u0394\u03c4p \u2248\ndp\nd\u03c4 \u0394\u03c4 and \u0394\u03c4f \u2248(1 + v) ((1 + z) \u0394\u03c4p + \u0394\u03c4) .\n(17)\nThe EU tax revenues associated with fuel contain both the direct \nVAT revenue and the excise duty with VAT applied to it:\nTEU = (v ((1 + z)p + c) + (1 + v) \u03c4) DEU (f, I) .\nDifferentiating this fully with respect to \u03c4 gives\ndTEU\nd\u03c4\n= [\u03c4 + v ((1 + z)p + c + \u03c4)]\n\u2202DEU(f,I)\n\u2202f\n[(1 + v) ((1 + z)\ndp\nd\u03c4 + 1)]\n+ [v ((1 + z)\ndp\nd\u03c4 + 1) + 1] DEU\n= (1 + v)\n\u03c4+v((1+z)p+c+\u03c4)\nf\nf\nDEU\n\u2202DEU(f,I)\n\u2202f\n[((1 + z)\ndp\nd\u03c4 + 1)] DEU\n+ [v ((1 + z)\ndp\nd\u03c4 + 1) + 1] DEU\n= [1 + (\n\u03c4+v((1+z)p+c+\u03c4)\n(1+z)p+c+\u03c4\n\u0303\u03b5D,EU + v) ((1 + z)\ndp\nd\u03c4 + 1)] DEU.\nMultiplying by \u0394\u03c4 and using \u0394\u03c4\u2009f from equation (17) delivers a linear \napproximation of the change in tax revenues from the tax change \u0394\u03c4:\n\u0394\u03c4TEU \u2248[\u0394\u03c4 + ( \u03c4 + v ((1 + z)p + c + \u03c4)\n(1 + z)p + c + \u03c4\n\u0303\u03b5D,EU + v) ((1 + z)\u0394\u03c4p + \u0394\u03c4)] DEU.\nThe Russian oil profits are\n\u03c0RU = (p \u2212e) SRU (p) ,\nwhere e represents Russian extraction costs that are assumed to be \nconstant.\nTreating p as a function of \u03c4 and differentiating fully with respect \nto \u03c4 gives\nd\u03c0RU\nd\u03c4\n= dp\nd\u03c4 SRU + (p \u2212e) S\u2032\nRU (p) dp\nd\u03c4 = (1 + p \u2212e\np\n\u03b5S,RU) SRU\ndp\nd\u03c4\nand, using \u0394\u03c4p from equation (17), a linear approximation of the change \nin Russian oil profits is\n\u0394\u03c4\u03c0RU \u2248(1 + p \u2212e\np\n\u03b5S,RU) SRU\u0394\u03c4p.\nModel of income transfer\nHere, we consider the effects of transferring income to households \ninstead of lowering the fuel tax. Treating the equilibrium oil price as \na function of I, differentiating equation (11) fully with respect to I and \nusing equation (10) gives\n\u2202DEU(f,I)\n\u2202I\n+\n\u2202DEU(f,I)\n\u2202f\n(1 + v) (1 + z)\ndp\ndI + D\u2032\nROW (p)\ndp\ndI = (S\u2032\nRU (p) + S\u2032\nROW (p))\ndp\ndI .\n(18)\nMultiplying the left- and right-hand sides of equation (18) by p/D \nand p/S, respectively (with S\u2009=\u2009D in equilibrium), we get\ndp\ndI =\np\nD\n\u2202DEU(f,I)\n\u2202I\np\nS S\u2032\nRU (p) +\np\nS S\u2032\nROW (p) \u2212(1 + v) (1 + z)\np\nD\n\u2202DEU(f,I)\n\u2202f\n\u2212\np\nD D\u2032\nROW (p)\n.\n\nNature Energy | Volume 7 | October 2022 | 989\u2013997  \n995\nAnalysis\nhttps://doi.org/10.1038/s41560-022-01122-6\nUsing equation (13) yields\ndp\ndI =\np\nI\nxI\nDEU\n\u2202DEU(f,I)\n\u2202I\nyp\nSRU S\u2032\nRU (p) +\n(1\u2212y)p\nSROW S\u2032\nROW (p) \u2212\n(1+v)(1+z)p\nf\nxf\nDEU\n\u2202DEU(f,I)\n\u2202f\n\u2212\n(1\u2212x)p\nDROW D\u2032\nROW (p)\n.\nUsing equations (10), (15) and (16), we arrive at\ndp\ndI = p\nI\nx\u03b5I,EU\ny\u03b5S,RU + (1 \u2212y) \u03b5S,ROW \u2212x\n(1+z)p\n(1+z)p+c+\u03c4 \u0303\u03b5D,EU \u2212(1 \u2212x) \u03b5D,ROW\n,\nwhere the income elasticity of demand in the EU is\n\u03b5I,EU \u2261\nI\nDEU\n\u2202DEU (f, I)\n\u2202I\n.\nThe responses of the oil price p and the EU fuel price f to an income \nchange \u0394I can be linearly approximated as\n\u0394Ip \u2248\ndp\ndI \u0394I and \u0394If \u2248(1 + v) (1 + z) \u0394Ip.\n(19)\nTreating p as a function of I, differentiating oil profits \n\u03c0RU = (p \u2212e) SRU (p) fully with respect to I and using the Russian supply \nelasticity in equation (15) gives\nd\u03c0RU\ndI\n= dp\ndI SRU + (p \u2212e) S\u2032\nRU (p) dp\ndI = (1 + p \u2212e\np\n\u03b5S,RU) SRU\ndp\ndI\nand, using equation (19), a linear approximation of the change in Rus-\nsian oil profits is\n\u0394I\u03c0RU \u2248(1 + p \u2212e\np\n\u03b5S,RU) SRU\u0394Ip.\nParameter values of elasticity of demand\nThe price elasticity of demand for road transport fuels (gasoline and \ndiesel) and crude oil is low in the short term but increases with time. In \nthe short term, many fuel consumers can only drive less to reduce con-\nsumption, while in the longer term many can shift to more efficient vehi-\ncles or change their commuting distance or mode of transportation.\nSeveral studies compile existing estimates of the demand elas-\nticity of gasoline, diesel and crude oil (see, for example, refs. 21\u201324). \nThese estimates are derived using different methods over different \ntime periods and locations. Short-term gasoline elasticity estimates \nrange from \u22120.04 to \u22120.5 (refs. 23,24) with several review studies deriv-\ning averages around \u22120.25 (ref. 24). Aklilu24 also provides additional \noriginal estimates for EU countries using recent data, finding an EU \naverage short-term gasoline elasticity of \u22120.255. We use \u22120.25 in our \ncalculations, in line with these recent EU estimates as well as the wider \nliterature samples.\nLong-term gasoline demand elasticity estimates range from \u22120.2 \nto \u22121 (ref. 23). Aklilu24 compiles review studies with even higher ranges \nbut with averages around \u22120.7. Aklilu\u2019s own empirical study finds a \nlong-term EU average of \u22120.88. We use \u22120.9 in our calculations based \non these results.\nCrude oil demand elasticity is usually found to be lower than that \nof gasoline. This is to be expected since crude oil is only a part of the \ngasoline price. If we assume that crude oil represents half the cost of \nretail gasoline, a 10% increase in the price of crude oil would translate \nto a 5% increase in the price of gasoline, and the demand elasticities for \noil would be about half those for gasoline21. Caldara et al.22 compile 31 \nstudies for short-term world oil demand in the range of \u22120.04 to \u22120.9 \nwith a mean of \u22120.22 and a median of \u22120.13. We follow Hamilton21 and \nuse half of the gasoline estimates for wider oil demand (that is, \u22120.125 \nfor our short-term oil demand elasticity and \u22120.45 for our long-term \noil demand elasticity), which is also in line with the estimates of  \nCaldara et al.22.\nIn the sensitivity analysis in Supplementary Note 2, we also explore \nlower and higher values in the range found in the literature.\nParameter values of elasticity of supply\nThe price elasticity of global oil supply is low (close to zero) in the short \nterm and grows only slowly in the longer term. New conventional oil \nfields take several years to bring into production and additional supply \nin the short term (within 1\u2009year) can only come from either inventory, \npolitically withheld supply (including, for example, Saudi Arabian spare \ncapacity), shale oil production and infill drilling in conventional fields.\nCompared with demand elasticity studies, supply elasticity studies \nare rare. Caldara et al.22 compile six studies applying different methods \nand data and find short-term (within 1\u2009year) supply elasticities in the \nrange of 0\u20130.27. These estimates are based on historical data and do \nnot necessarily reflect current or future oil supply. As a complement, \nwe also rely on modelled forward-looking estimates of supply elasticity \nderived by Wachtmeister25 using a bottom-up modelling framework \nand Rystad UCube field-by-field data, as well as our own judgement of \nthe current oil market outlook.\nFor our very-short-term and short-term scenarios, we use a supply \nelasticity of 0 for both Russia and the rest of the world. This corre-\nsponds to a hypothetical scenario with no inventory draw, no addi-\ntional production by the Organization of the Petroleum Exporting \nCountries and a timeframe below 6\u2009months where the shale response is \nstill low. In the sensitivity analysis (Supplementary Note 2), we present \na short-term case using 0.1, which can be seen as reflecting a 12-month \nshale response and/or a stronger response from the Organization of \nthe Petroleum Exporting Countries.\nFor our long-term scenario, we use 0.13, which is in line with a \nmodelled 3-year horizon estimate of global supply by Wachtmeister25. A \nvalue of 0.13 is also used as a central estimate by Erickson and Lazarus10, \neven though their studied time horizon is longer than ours. In the sen-\nsitivity analysis (see Supplementary Note 2), we explore 0.2 as a higher \nestimate, reflecting a stronger supply response.\nNote again that our long-term scenario reflects a supply response \nin 1\u20133\u2009years. Other studies reporting long-term supply elasticity esti-\nmates might use long term to describe longer time horizons. For exam-\nple, Gately26, Brook et al.27 and Erickson and Lazarus10 use long term to \ndescribe responses up to 15\u2009years ahead.\nOther costs and refinery margins\nWe translate oil production (crude oil, condensate and natural gas liq-\nuids) in barrels per day to a corresponding volume of refined products. \nWe make the simplifying assumption that one barrel of oil yields 170\u2009l \nof products and fuels that can be sold to consumers. In the base case, \nthe variable production cost of these fuels is assumed to correspond \ndirectly to the crude oil price (that is, the variable fuel production cost \nis the global oil price per barrel (Brent; in US$ per barrel) measured in \nEuros per litre of fuel product (p)). The retail fuel price (consumer price) \nis then the variable fuel production cost (oil price, p) plus the other, fixed, \nproduction costs, c (refining, transport, margins and so on), plus the fuel \ntax, \u03c4, then VAT is applied to all of these. In the sensitivity analysis (see \nSupplementary Note 3), we explore other variable production costs (z) \nand discuss which case is more likely. Our base case c of \u20ac0.45\u2009l\u22121 is derived \nbackwards from a consumer price of \u20ac1.9\u2009l\u22121. c thus includes the current \nrefinery margins (the value difference between crude oil and refined \nproducts), which vary in time and are currently at historically high levels.\nSize of markets and Russian export declines\nWe assume that Russia has already lost 1\u2009million barrels per day of \noil exports based on recent export data28 and analysis29. In January, \nbefore the war, the global supply and demand of oil was estimated \nto be 100\u2009million barrels per day (ref. 30). Consequently, we assume \n\nNature Energy | Volume 7 | October 2022 | 989\u2013997 \n996\nAnalysis\nhttps://doi.org/10.1038/s41560-022-01122-6\nthe current oil market to be 99\u2009million barrels per day. Road trans-\nport fuel\u2019s share of total EU oil consumption is 47.5% (ref. 31). The \nEU\u2019s share of global oil consumption is 12% (ref. 32), which yields our \nx\u2009=\u20090.475\u2009\u00d7\u20090.12\u2009=\u20095.7%. The EU imports, in normal times, ~35% of its oil \nfrom Russia33. For the very short term, we assume that the reduction in \nRussia\u2019s export (1\u2009million barrels per day) fully accrues to the EU. This \nimplies y\u2009=\u200929% in the very short term.\nFuel price and taxes\nFor our analysis, we need to construct an EU-level fuel price and fuel \ntax. However, fuel prices and taxes differ between the EU countries. \nWe construct EU-level values using weighted averages based on each \ncountry\u2019s share of total EU fuel consumption using data from Eurostat34. \nThe data underlying the quantification in this section are summarized \nin Supplementary Table 1.\nFor the country-level fuel price, we use data from the European \nCommission (from the fuel-prices.eu database35). We use the average \nprices for the month of May 2022 (see columns 3 and 4 of Supplemen-\ntary Table 1) and obtain weighted averages of \u20ac1.89\u2009l\u22121 for petrol and \n\u20ac1.87 for diesel. Hence, we choose \u20ac1.9 as transport fuel price.\nIn assessing the EU\u2019s tax reduction, we want to emphasize that \nannounced and implemented tax reductions come in several forms \n(for example, excise duty reduction and directed VAT reductions), new \nones are being suggested and their time spans are varied. Hence, there \nis scarcely any way to assess the final aggregate outcome until pos-\nsibly several years from the time of writing. Instead, to obtain a rough \nestimate of what the final outcome may be, we look at two scenarios.\nThe first scenario is based on the possibility that all EU countries \nreduce their excise tax to the EU minimum level (\u20ac0.359\u2009l\u22121 for unleaded \npetrol and \u20ac0.33\u2009l\u22121 for diesel). In this scenario, we use the countries\u2019 \ntax levels from July 2021 (ref. 36) and reduce them to the minimum level \n(we use the numbers for unleaded petrol and gas oil for propellant use \n(diesel); in some instances, different tax levels apply to different sub-\ncategories, in which case we take an average). The data are summarized \nin columns 5\u20138 of Supplementary Table 1. We then obtain a weighted \naverage tax reduction of \u20ac0.24 for petrol and \u20ac0.15 for diesel (these \ndo not include the indirect effects on VAT that apply to excise duties).\nThe second scenario (columns 9 and 10 in Supplementary Table 1) \ntakes the post-invasion announcements to date (mid-June 2022). For \nthis, we take the compilation of Transport & Environment37, cross-check \ntheir entries with news articles and add directed VAT reductions for \nfuels (for Estonia and Romania). For most countries, we verify the \nnumbers from Transport & Environment37 (see the footnote of Sup-\nplementary Table 1). We interpret (but ultimately cannot verify) the \nnumbers to exclude the indirect effects on VAT payments in those cases \n(such as Germany) where VAT applies to the excise duty (the number \nfor Sweden in Supplementary Table 1 includes this indirect VAT effect). \nThe most consequential decision is probably our choice to set Poland\u2019s \npre-invasion VAT reduction to zero. In calculating VAT reductions in \nEuros, we use the average price in January and February 2022 (this \nis conservative since prices were lower than post-invasion). We then \nobtain a weighted average tax reduction of \u20ac0.165\u2009l\u22121 for petrol and \n\u20ac0.132\u2009l\u22121 for diesel. We also calculate the percentage reduction in fuel \ntaxes. We obtain percentage reductions for petrol and diesel of 9.7 and \n8.2%, respectively, compared with the price in January and February, \n8.8 and 7.4%, respectively, compared with the 3-month pre-reduction \naverage price and 5 and 4.3%, respectively, compared with the price in \nMay. These numbers are not shown in Supplementary Table 1.\nBased on these scenarios, each with its own caveats, we use a tax \nreduction of \u20ac0.20 (including indirect VAT effects) for the analysis.\nData availability\nAll of the data generated or analysed during this study are included \nin this published article (and its Supplementary Information files) or \nare publicly available.\nCode availability\nModel implementation code written in MATLAB is available as Sup-\nplementary Code.\nReferences\n1.\t\nSanctions Adopted Following Russia\u2019s Military Aggression Against \nUkraine (European Commission. 2022); https://ec.europa.eu/ \ninfo/business-economy-euro/banking-and-finance/international- \nrelations/restrictive-measures-sanctions/sanctions-adopted- \nfollowing-russias-military-aggression-against-ukraine_en\n2.\t\nUkraine-/Russia-Related Sanctions (U.S. Department of the \nTreasury, 2022); https://home.treasury.gov/policy-issues/ \nfinancial-sanctions/sanctions-programs-and-country- \ninformation/ukraine-russia-related-sanctions\n3.\t\nSaul, J. All at sea: Russian-linked oil tanker seeks a port. Reuters \nhttps://www.reuters.com/business/energy/all-sea-russian-linked- \noil-tanker-seeks-port-2022-03-09/ (2022).\n4.\t\nRussia\u2019s War on Ukraine: EU Adopts Sixth Package of Sanctions \nAgainst Russia (European Commission, 2022); https://ec.europa. \neu/commission/presscorner/detail/en/IP_22_2802\n5.\t\nAndersson, J. & Tippmann, C. The impact of rising gasoline prices \non Swedish households\u2014is this time different? Free Network \nhttps://freepolicybriefs.org/2022/05/02/rising-gasoline-prices/ \n(2022).\n6.\t\nMartin, J. Gas prices hit new record sparking fears over bill rises. \nBBC News https://www.bbc.com/news/business-60613855 \n(2022).\n7.\t\nJones, G. Analysis: climate goals take second place as EU states \ncut petrol prices. Reuters https://www.reuters.com/business/ \nenergy/climate-goals-take-second-place-eu-states-cut-petrol- \nprices-2022-03-22/ (2022).\n8.\t\nChambers, M. German finance minister plans gasoline discount. \nReuters https://www.reuters.com/business/energy/german- \nfinance-minister-plans-gasoline-discount-bild-2022-03-13/ \n(2022).\n9.\t\nDe Beaupuy, F. France plans $2.2 billion fuel rebate in bid to help \nmotorists. Bloomberg UK https://www.bloomberg.com/news/\narticles/2022-03-12/france-plans-2-2-billion-fuel-rebate-in-bid-t\no-help-motorists (2022).\n10.\t Erickson, P. & Lazarus, M. Impact of the Keystone XL pipeline \non global oil markets and greenhouse gas emissions. Nat. Clim. \nChange 4, 778\u2013781 (2014)\n11.\t\nFaehn, T., Hagem, C., Lindholt, L., M\u00e6land, S. & Einar  \nRosendahl, K.Climate policies in a fossil fuel producing  \ncountry\u2014demand versus supply side policies. Energy J. 38, \n77\u2013102 (2017).\n12.\t Severen, C. & van Benthem, A. A. Formative experiences and the \nprice of gasoline. J. Appl. Econ. 14, 256\u2013284 (2022).\n13.\t Military expenditure (current USD)\u2014Russian Federation. The \nWorld Bank https://data.worldbank.org/indicator/MS.MIL.XPND.\nCD?locations=RU (2022).\n14.\t UCube (Rystad Energy, 2022); https://www.rystadenergy.com/\nenergy-themes/oil\u2013gas/upstream/u-cube/\n15.\t Dahl, C. A. Measuring global gasoline and diesel price and \nincome elasticities. Energy Policy 41, 2\u201313 (2012).\n16.\t Expenditure on Personal Mobility (European Environment Agency, \n2011); https://www.eea.europa.eu/data-and-maps/indicators/\nexpenditure-on-personal-mobility-2/assessment\n17.\t Transport costs EU households over \u20ac1.1 trillion. Eurostat  \nhttps://ec.europa.eu/eurostat/web/products-eurostat-news/-/\nddn-20200108-1 (2020).\n18.\t Hanson, D., Toru Delibasi, T., Gatti, M. & Cohen, S.How do \nchanges in economic activity affect air passenger traffic? The \nuse of state-dependent income elasticities to improve aviation \nforecasts, J. Air Transp. Manag. 98, 102147 (2022).\n\nNature Energy | Volume 7 | October 2022 | 989\u2013997  \n997\nAnalysis\nhttps://doi.org/10.1038/s41560-022-01122-6\n19.\t Femenia, F.A meta-analysis of the price and income elasticities of \nfood demand. Ger. J. Agric. Econ. 68, 77\u201398 (2019).\n20.\t Household consumption by purpose. Eurostat https://ec.europa.\neu/eurostat/statistics-explained/index.php?title=Household_\nconsumption_by_purpose (2021).\n21.\t Hamilton, J. D.Causes and consequences of the oil shock of 2007\u2013\n08, Brookings Pap. Econ. Act. 40, 215\u2013283 (2009).\n22.\t Caldara, D., Cavallo, M. & Iacoviello, M. Oil price elasticities and \noil price fluctuations. J. Monet. Econ. 103, 1\u201320 (2019).\n23.\t H\u00f6ssinger, R., Link, C., Sonntag, A. & Stark, J. Estimating the  \nprice elasticity of fuel demand with stated preferences derived \nfrom a situational approach. Transp. Res. A Policy Pract. 103, \n154\u2013171 (2017).\n24.\t Aklilu, A. Z. Gasoline and diesel demand in the EU: implications \nfor the 2030 emission goal. Renew. Sustain. Energy Rev. 118, \n109530 (2020).\n25.\t Wachtmeister, H. World Oil Supply in the 21st Century: a Bottom-up \nPerspective. PhD thesis, Uppsala Univ. (2020).\n26.\t Gately, D.OPEC\u2019s incentives for faster output growth. Energy J. 25, \n75\u201396 (2004).\n27.\t Brook, A. M., Price, R., Sutherland, D., Westerlund, N. & Andr\u00e9, C. \nOil Price Developments: Drivers, Economic Consequences and \nPolicy Responses (OECD, 2004); https://www.oecd-ilibrary.org/\ncontent/paper/303505385758\n28.\t Wech, D. Russian crude exports remain high. Vortexa https://www.\nvortexa.com/insights/russian-crude-exports-remain-high/ (2022).\n29.\t Oil Market Report\u2014May 2022 (IEA, 2022); https://www.iea.org/\nreports/oil-market-report-may-2022\n30.\t Oil Market Report\u2014February 2022 (IEA, 2022); https://www.iea. \norg/reports/oil-market-report-february-2022\n31.\t Oil and petroleum products\u2014a statistical overview. Eurostat \nhttps://ec.europa.eu/eurostat/statistics-explained/index.php? \ntitle=Oil_and_petroleum_products_-_a_statistical_overview (2022).\n32.\t Statistical review of world energy 2021 BP http://www.bp.com/en/ \nglobal/corporate/energy-economics/statistical-review-of-world- \nenergy.html (2021).\n33.\t Statistical review of world energy 2020 BP http://www.bp.com/ \nen/global/corporate/about-bp/energy-economics/statistical- \nreview-of-world-energy-2013.html (2020).\n34.\t Final energy consumption in transport by type of fuel. Eurostat \nhttps://ec.europa.eu/eurostat/databrowser/view/ten00126/\ndefault/table?lang=en (2022).\n35.\t Fuel Prices Archive by Country (European Commission, 2022); \nhttps://www.fuel-prices.eu/archive/30-05-2022/\n36.\t Excise Duty Tables: Part II Energy Products and Electricity \n(European Commission, 2021); https://ec.europa.eu/taxation_\ncustoms/system/files/2021-09/excise_duties-part_ii_energy_\nproducts_en.pdf\n37.\t Taxpayers face \u20ac9bn bill for fuel tax cuts skewed towards the \nrich. Transport & Environment https://www.transportenvironment.\norg/discover/taxpayers-face-e9bn-bill-for-fuel-tax-cuts-sk\newed-towards-the-rich-study-finds/ (2022).\n38.\t Oil market and Russian supply\u2014Russian supplies to global  \nenergy markets. IEA https://www.iea.org/reports/russian-supplies- \nto-global-energy-markets/oil-market-and-russian-supply-2 (2022).\nAcknowledgements\nWe thank S. O\u2019Brien for excellent research assistance; N. Rossbach, \nP. Olsson and J. Norberg at the Swedish Defence Research Agency \nfor valuable input on Russian military spending; M. H\u00f6\u00f6k for valuable \nsupport; and P. Bansal for important comments. This work has been \nsupported by Formas under grant number 2020-00371 (to J.G.  \nand D.S.).\nAuthor contributions\nAll of the authors contributed equally to the project. J.G., D.S. and H.W. \ntogether designed the study, developed the methodology, interpreted \nthe results and wrote and edited the manuscript. H.W. led the data and \ninput parameter work, J.G. led the model implementation. D.S. led the \ncoordination of the project.\nFunding\nOpen access funding provided by Uppsala University.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41560-022-01122-6.\nCorrespondence and requests for materials should be addressed to \nHenrik Wachtmeister.\nPeer review information Nature Energy thanks Michael Ross, Michael \nPlante and the other, anonymous, reviewer(s) for their contribution to \nthe peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard  \nto jurisdictional claims in published maps and institutional  \naffiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons license, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons license, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons license and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this license, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2022\n\n\n    Scientific Research Findings:", "answer": "We first estimate that EU countries on average already have, or intend to, cut fuel taxes by around\u00a020\u00a0eurocents per litre. We then find that this will lead to an increased oil price with significantly increased Russian oil profits. In the first year and up to three years later, Russia\u2019s oil profits can be expected to increase by more than\u00a08\u00a0million euros per day. This is equivalent to around\u00a03,000\u00a0million euros in a year,\u00a00.2%\u00a0of Russia\u2019s pre\u2011invasion GDP and\u00a05%\u00a0of its estimated pre\u2011invasion military spending. This is independent of whether the EU implements an import ban on Russian oil. Our analysis thus indicates that lowering fuel taxes undermines the EU\u2019s efforts to restrict Russia\u2019s economic abilities. As households are being hit significantly by the increased fuel price, we study an alternative policy measure: direct income transfers to households with the same fiscal burden (\u20ac170\u00a0million per day) as the tax cut. Our analysis finds that Russian oil revenues then increase by only a sixth compared to the tax cut. Such income transfers are more flexible as households can use them for road fuel as well as for anything else. The transfers can also be targeted at those hit particularly hard by fuel\u2011price increases.", "id": 14}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-022-00994-y\n1Institute for Sustainable Economic Development, University of Natural Resources and Life Sciences, Vienna, Austria. 2Institute of Statistics, University of \nNatural Resources and Life Sciences, Vienna, Austria. \u2709e-mail: katharina.gruber@boku.ac.at\nW\neather extremes such as storms can strongly affect the \nreliability of power systems1. The increasing use of vari-\nable renewable energies additionally exposes power sys-\ntems to hazards caused by weather extremes2,3. However, recently \na gas-power-dominated system was deeply impaired by a weather \nextreme: a cold spell over Texas between 10 February and 20 \nFebruary 2021 with temperatures far below 0\u2009\u00b0C caused a failure of \nlarge parts of the Texan power system. The combination of extraor-\ndinarily high winter electricity demand and more importantly the \nfailure of substantial power generation capacities, both due to low \ntemperatures, resulted in up to 4.5 million Texans being cut off from \ntheir electricity supply4.\nA first retrospective analysis of the event by Busby et al.5 dis-\ncusses the magnitude of the event and its causes, indicating that the \ntotal economic loss amounted to US$130 billion and that the out-\nage of gas power plants was mainly responsible for the high defi-\ncits in power generation capacity. Wu et al.6 provide a power grid \nsimulation to conduct a very detailed analysis of the 2021 event, \nand Doss-Gollin et al.7 have shown that lower temperatures than \nthose in February 2021 have been observed in the past 71\u2009years, and \nheating demand predicted from temperature data would also have \nbeen higher in the past, although the 2021 freeze event was compa-\nrably long. These previous studies indicate a striking gap between \nthe occurrence probability of such an event, its large-scale economic \nand social cost and the lack of winterization efforts. However, none \nof these studies assessed whether the economic incentives for power \ncompanies to invest in winterization have been sufficient, when the \n2021 event is put into a long-term climatic context. As winterization \nwas not strongly enforced by regulation in Texas, power generators \nhad to rely on the incentives provided by the energy-only market to \narrive at investment decisions. These incentives consist mainly of \nregulated price spikes at the spot market when generation capacity \nis scarce8.\nHere we assess how revenues from winterization compare to \nits cost for power companies. Technically, we combine estimates \nof temperature-dependent load with a model of power plant out-\nages, taking into account 71\u2009years (1950\u20132021) of past climate from \nreanalysis data. Subsequently, we discuss the 2021 event in detail, \nanalyse the long-term frequency of such events and determine \nrevenues from and cost of winterization. Furthermore, we discuss \npotential reasons for under-investment.\nThe 2021 event in a long-term climatic context\nStarting on 10 February 2021, temperatures in Texas decreased, \ncausing load to increase from around 40\u2009GW to over 70\u2009GW by \n14\u201315 February. On 15 February, the freeze reached a critical level \nand, consequently, substantial shares of generation capacities failed. \nAvailable capacities dropped below demand, leading to a sustained \ndeficit in power generation capacity (Fig. 1). Consequently, roll-\ning blackouts had to be implemented to stabilize the grid, and \nscarcity prices at the power market increased to the upper limit of \nUS$9,000\u2009MWh\u20131. The deficit event continued until 19 February, \nwhen rising temperatures allowed the system to recover.\nIn the following, we compare events in Februrary 2021 to the \nperiod 2004\u20132020, as the Electric Reliability Council of Texas \n(ERCOT) provides hourly data on system operation in this period. \nFurthermore, the values of loss of load, capacity failures and \ndemand prediction in this section rely on our simulation and may, \ntherefore, differ from ERCOT reports to some extent. Because we \nfocus on estimating the long-term frequency of such events, we did \nnot aim to reproduce the February 2021 event in detail. We find that \nthe highest predicted demand in the February 2021 event was well \nabove the highest load observed in winter in that period. However, \nour estimate is in the range of observed extreme summer loads \n(Supplementary Note 1 and Supplementary Fig. 9).\nBesides leading to high electricity demand, the low temperatures \nalso caused substantial outages of generation capacities. As a result, \nloss of load occurred in 106\u2009hours. Based on the predicted demand \nand the observed load, we estimate that in total, 1.45\u2009TWh of load \nwere affected by blackouts. Busby et al.5 estimate the social cost of \nthe power outages at US$130 billion. Therefore, the deficit cost of \naround US$87,000\u2009MWh\u20131 is one magnitude higher than the value \nof loss of load used by the Texan market regulator ERCOT, that is, \nUS$9,000\u2009MWh\u20131, in 2021.\nOutages of gas generation capacities increased rapidly when \nthe average temperature weighted by installed capacities at gas \nProfitability and investment risk of Texan power \nsystem winterization\nKatharina Gruber\u200a \u200a1\u2009\u2709, Tobias Gauster2, Gregor Laaha\u200a \u200a2, Peter Regner\u200a \u200a1 and Johannes Schmidt\u200a \u200a1\nA lack of winterization of power system infrastructure resulted in substantial rolling blackouts in Texas in 2021, but debate \nabout the cost of winterization continues. Here we assess if incentives for winterization on the energy-only market are suffi-\ncient. We combine power demand estimates with estimates of power plant outages to derive power deficits and scarcity prices. \nExpected profits from winterization of a large share of existing capacity are positive. However, investment risk is high due to \nthe low frequency of freeze events, potentially explaining the under-investment, as do the high discount rates and uncertainty \nabout power generation failure under cold temperatures. As the social cost of power deficits is one to two orders of magnitude \nhigher than the winterization cost, regulatory enforcement of winterization is welfare enhancing. Current legislation can be \nimproved by emphasizing the winterization of gas power plants and infrastructure.\nNature Energy | VOL 7 | May 2022 | 409\u2013416 | www.nature.com/natureenergy\n409\n\nArticles\nNATurE EnErgy\npower plant locations dropped below \u20138.8\u2009\u00b0C, which is a record low \ncompared to the period 2004\u20132020 (Supplementary Note 1 and \nSupplementary Fig. 10). The outages were related to the freezing \nof power plants and of gas supply infrastructure, including produc-\ntion equipment at gas fields. Power plants outages increased rapidly \nwhen average temperature weighted by gas production at gas field \nlocations dropped below \u201310.9\u2009\u00b0C, a record low compared to the \nperiod 2004\u20132020 (Supplementary Fig. 10). Therefore, gas supply \ninfrastructure played an important role in causing the outages, as \nconfirmed by ERCOT\u2019s classification of around 8\u2009GW of gas power \noutages being related to limited fuel supply9. Coal generation capac-\nity came offline at average temperatures weighted by coal plant loca-\ntions of below \u201310.2\u2009\u00b0C. This temperature is at the very lower end \nof observed temperatures in the period 2004\u20132021. For both coal \nand gas, recovery time was substantial. Even when temperatures \nincreased to over 0\u2009\u00b0C, 11.3\u2009GW of thermal power plants\u2014that is, \n18% of total available thermal capacity\u2014stayed offline for another \n16\u2009hours10.\nAt the time of the failure, temperatures at wind parks in Southern \nTexas were at the very lower end of the temperature range observed \nin the period 2004\u20132020. However, the average wind park tempera-\nture in Northern Texas when wind power plants started to fail was \njust below 0\u2009\u00b0C and well within the range of previously observed \nlow temperatures. On 13 February, when gas outages summed up \nto only 5\u2009GW, ERCOT already reported 13\u2009GW of wind capacity \noutages (Fig. 1). This represented a loss of 3.3\u2009GW of wind power \nproduction on average at the prevailing wind conditions. However, \nlater on, temperatures reached record lows at wind power plant sites \nin Northern Texas, too.\nOur simulations of loss-of-load events using climate data from \n71\u2009years shows that the 2021 event was a record one. In total, we \nestimate that eight other severe power deficit events would have \noccurred in the current system if it had existed from 1950 to 2021, \nassuming the climate conditions of 1950\u20132021 (Fig. 2). The second \nlargest power deficit event at 1.26\u2009TWh is predicted when using  \nclimate data from 1983, assuming installed generation capacities \nas in February 2021. Furthermore, we observe 17 minor events. \nHowever, as the sum of the deficits of all 17 minor events is less than \n1% of the sum of the deficits of the nine largest events, we exclude \nthem from further analysis.\nIn our model simulations, the loss-of-load event has a duration \nof 106\u2009hours and causes an aggregated deficit of 1.49\u2009TWh, at a peak \ncapacity deficit of 31.3\u2009GW. There are several events with similar \npeak capacity deficits identified in the 1950\u20132021 period, and also \nevents with a comparably long duration, but none with a compa-\nrably high amount of loss of load. In the largest events before 2021 \n(1962 and 1983), 250\u2009GWh less lost load results from our simula-\ntion (Fig. 2). The year 1989 was the last time a similar freeze event \noccurred.\nThe 2021 record-high loss of load was not caused by the freeze \nmagnitude alone but by a combination of a long, relatively cold freeze \nevent and an inopportune timing of the freeze peak. According to \nFig. 1, the system failure occurred early and was prolonged by a  \nlong freeze period afterward. This is in contrast to other years when \n80\na\nb\nc\n70\n60\n50\n(GW)\n40\n40\n30\n20\n10\n(GW)\n0\n13 Feb\n14 Feb\n15 Feb\n16 Feb\n17 Feb\nOutage capacity (ERCOT)\nTemperature\nPower deficit\nLoad (ERCOT)\nAverage temperature\nweighted by\nPopulation\nGas power plants\nGasfields\nCoal power plants\nCoal parks north\nCoal parks south\nPredicted demand\nPredicted available\ncapacity\nPredicted deficit\nGas\nCoal\nWind\nOther\nPeriod of deficit event (ERCOT)\nPeriod of predicted deficit event\n18 Feb\n19 Feb\n20 Feb\n21 Feb\n15\n10\n5\n0\n(\u00b0C)\n\u20135\n\u201310\n\u201315\n\u201320\nFig. 1 | The February 2021 event in retrospect. a, Observed load, predicted demand and predicted available capacity during the February (Feb) 2021 event. \nb, Temperatures during the February 2021 event. Average temperature weighted by population: weights are determined according to population density \nin a square kilometre grid. Average temperature weighted by wind, gas and coal power plants: weights are determined according to installed capacity at \nplant locations. Average temperature weighted by gas fields: weights are determined according to annual gas production at gas field locations. c, Capacity \noutages during the February 2021 event.\nNature Energy | VOL 7 | May 2022 | 409\u2013416 | www.nature.com/natureenergy\n410\n\nArticles\nNATurE EnErgy\ntemperatures recovered more quickly after temperature minima had \nbeen reached (for example, in 1951 and 1963). This finding is sup-\nported by the extreme value statistics of the freeze spells shown in \nFig. 3. The 2021 event was the second longest freeze event in seven \ndecades. It has a return period of 37\u2009years. Other events, however, \nwere colder (1951, 1989) or had higher frost sums (1951, 1983).\nIn terms of load, the highest predicted winter load in 2021 was \nslightly lower than the highest predicted winter load in the complete \n71\u2009year time series (Supplementary Fig. 9), as there were lower tem-\nperatures in earlier years during the 1989 event.\nFreeze events may have decreased due to global warming. \nHowever, it has been shown11 that extremely cold events in the \nnorthern hemisphere have increased over the past 40\u2009years. Our \nanalysis for Texas does not indicate any significant trend in defi-\ncit events (Fig. 4) or freeze events (Supplementary Table 2). Still, \naverage temperatures in Texas significantly increased due to cli-\nmate change since 1951 (Supplementary Note 4 and Supplementary \nFig. 13). This result is confirmed by others; however, the increase \nin mean temperature is not genuinely transferable to extreme tem-\nperatures12. A stratified analysis of annual freeze events (minimum \nannual temperature) below temperature thresholds from 0 to \u201310\u2009\u00b0C \nreveals that there is indeed no significant observed change of severe \nfreeze events below \u20132\u2009\u00b0C (Supplementary Note 5). Only very mild \nfreeze events showed a significant attenuation (2.6\u2009\u00b0C over the past \nseven decades), but such events are irrelevant to freeze-related fail-\nures of the power system comparable to the 2021 event.\nComparing revenues to cost of winterization\nIn the following, we assess how much revenue could have been \nearned by power generators by winterizing their capacities under \nCapacity de\ufb01cit event\nTemperature event\n(1a)\n20\nEvent 1 (1951)\nDuration: 129 h\n\u2013123 \u00b0C\n29.4 GW\n\u20138.6 \u00b0C\n22.2 GW\n\u20139.3 \u00b0C\n23.9 GW\n\u201310.2 \u00b0C\n29.1 GW\n\u20139.2 \u00b0C\n27.2 GW\n\u201314.2 \u00b0C\n29.1 GW\n\u20139.6 \u00b0C\n27.5 GW\n\u201311.3 \u00b0C\n28.4 GW\n\u201312 \u00b0C\n31.3 GW\nDuration: 85 h\nDuration: 77 h\nDuration: 51 h\nDuration: 72 h\nDuration: 200 h\nDuration: 72 h\nDuration: 82 h\nDuration: 132 h\nDuration: 48 h\nDuration: 73 h\nDuration: 106 h\nDuration: 36 h\nDuration: 46 h\nDuration: 120 h\nDuration: 65 h\nDuration: 68 h\nDuration: 33 h\nEvent 2 (1962)\nEvent 3 (1963)\nEvent 4 (1979)\nEvent 5 (1982)\nEvent 6 (1983)\nEvent 7 (1985)\nEvent 8 (1989)\nEvent 9 (2021)\n10\n0\n\u201310\n\u201320\n25\n0\n\u201325\n\u201350\n27\nJan\n29\nJan\n31\nJan\n02\nFeb\n04\nFeb\n06\nFeb\n06\nJan\n08\nJan\n10\nJan\n12\nJan\n14\nJan\n16\nJan\n21\nJan\n23\nJan\n25\nJan\n27\nJan\n29\nJan\n31\nJan\n28\nDec\n30\nDec\n01\nJan\n03\nJan\n05\nJan\n07\nJan\n09\nJan\n07\nJan\n09\nJan\n11\nJan\n13\nJan\n15\nJan\n17\nJan\n20\nDec\n22\nDec\n24\nDec\n26\nDec\n28\nDec\n30\nDec\n28\nJan\n30\nJan\n01\nFeb\n03\nFeb\n05\nFeb\n07\nFeb\n18\nDec\n20\nDec\n22\nDec\n24\nDec\n26\nDec\n28\nDec\n11\nFeb\n13\nFeb\n15\nFeb\n17\nFeb\n19\nFeb\n21\nFeb\n(1b)\n(2a)\n(2b)\n(3a)\n(3b)\n(4a)\n(4b)\n(5a)\n(5b)\n(6a)\n(6b)\n(7a)\n(7b)\n(8a)\n(8b)\n(9a)\n(9b)\nTemperature (\u00b0C)\nCapacity\ndeficit (GW)\n20\n10\n0\n\u201310\n\u201320\n25\n0\n\u201325\n\u201350\nTemperature (\u00b0C)\nCapacity\ndeficit (GW)\n20\n10\n0\n\u201310\n\u201320\n25\n0\n\u201325\n\u201350\nTemperature (\u00b0C)\nCapacity\ndeficit (GW)\nFig. 2 | Analysis of extreme temperatures and simulated loss-of-load events. Average temperature weighted by population and predicted capacity deficit \nof severe freeze events within the 1950\u20132021 period. Labels in the graph refer to temperature minima and deficit maxima, which are circled: temperature \n(1a) and capacity deficit (1b) of the 1951 event; temperature (2a) and capacity deficit (2b) of the 1961 event; temperature (3a) and capacity deficit (3b) of \nthe 1963 event; temperature (4a) and capacity deficit (4b) of the 1979 event; temperature (5a) and capacity deficit (5b) of the 1982 event; temperature \n(6a) and capacity deficit (6b) of the 1983 event; temperature (7a) and capacity deficit (7b) of the 1985 event; temperature (8a) and capacity deficit (8b) \nof the 1989 event; and temperature (9a) and capacity deficit (9b) of the 2021 event. Jan, January; Dec, December.\nNature Energy | VOL 7 | May 2022 | 409\u2013416 | www.nature.com/natureenergy\n411\n\nArticles\nNATurE EnErgy\ncurrent regulation, assuming perfect competition and using the \n71 past meteorological years. For that purpose, we use a chain of \nstatistical and simulation models to derive loss-of-load events and \nrevenues from winterization (Supplementary Fig. 1). The revenues \nresult from the scarcity price mechanism implemented by ERCOT: \nit increases prices automatically if spare capacity falls below a cer-\ntain threshold8. The upper price limit is set to US$9,000\u2009MWh\u20131 \nif spare capacity is below 2\u2009GW (Supplementary Fig. 8). If power \noperators winterize, they will be able to generate during times of \nhigh scarcity, generating additional revenue.\nRevenue from winterization is high but shows strong variability. \nFor the first winterized gigawatt of gas power capacity, the expected \nrevenue over a 30\u2009year period is US$1.06 billion per gigawatt, but \ndrops to US$0.52 billion per gigawatt at 14\u2009GW of winterization \n(Fig. 5). Revenue for winterization of a coal power plant is slightly \nlower per gigawatt, and revenue for winterization of a wind power \nplant is substantially lower. For all technologies, the spread of rev-\nenues is high: the revenue at the 68% confidence interval is reduced \nor increased by half of the expected revenue. In 1.2% of all cases, \nthere is no deficit event in a 30\u2009year period, implying zero revenue \nfrom winterization.\nReturn period in years\nFrost duration (h)\nTemperature\nReturn period in years\nCapacity deficit duration (h)\nLoss of load\nFrost sum (\u00b0C h)\nTotal loss of load (GWh)\nProbability\nMinimum temperature (\u00b0C)\nProbability\nMax capacity deficit (GW)\n200\na\nc\ne\nb\nd\nf\n150\n100\n50\n0\n100\n50\n0\n1,500\n1,000\n500\n30\n20\n10\n0\n0\n900\n600\n300\n0\n0\n\u20135\n\u201310\n\u201315\n0.01 0.1\n1.5 2\n5\n10\n20\n50\n100\n1983\n2021\n1973\n1951\n1962\n1962\n1983\n1951\n1962\n1983\n1989\n1962\n2021\n1983\n2021\n1962\n1951\n1989\n1983\n1989\n2021\n1951\n2021\n1951\n1989\n1989\n2021\n1951\n1983\n1989\n1.5 2\n5\n10\n20\n50 100\n0.5\n0.8\n0.9 0.95\n0.99\n0.01 0.1\n0.5\n0.8\n0.9 0.95\n0.99\nFig. 3 | Extreme value statistics of temperature and power deficits. a,c,e, Average temperature weighted by population of freeze events in Texas from \nseven decades of climate data. b,d,f, Predicted capacity deficit of freeze events in Texas from seven decades of climate data. Shown are the empirical \n(circles) and theoretical (lines) quantile functions of the three deficit characteristics: duration (a,b), severity (c,d) and intensity (d,e). The return period \n(inferred from a generalized extreme value distribution) and magnitude of the 2021 event are annotated with the help of the blue dashed lines. Very small \nevents lasting less than six hours are removed.\n0\n500\n1,000\n1,500\n2,000\n1960\n1980\n2000\n2020\nYear\nLoss of load (GWh)\nFig. 4 | Trend in loss-of-load events. Loss-of-load events in the period \n1950\u20132021 with a non-significant trend line (two-sided t-test with no \nadjustments results in P value\u2009=\u20090.31). Data are presented as mean values \nwith 95% confidence interval.\nNature Energy | VOL 7 | May 2022 | 409\u2013416 | www.nature.com/natureenergy\n412\n\nArticles\nNATurE EnErgy\nSubstantial winterization measures can be implemented under \nour estimates of expected revenue, assuming that variable operating \ncost is low for power generators. This is true for nuclear, wind, pho-\ntovoltaic and coal power plants. However, gas power plants may face \nhigh spot market prices for gas during cold spells, as observed dur-\ning the 2021 event. We assume here that gas plant operators there-\nfore have physically or financially hedged against high gas prices. We \nestimate that the winterization of gas wells\u2014or building 250\u2009GWh of \npipe gas storage at gas power plant locations\u2014in combination with \nwinterization of gas power plants will cost about US$450 million per \ngigawatt (Supplementary Note 6). This cost is below the revenue up \nto the 15th gigawatt of winterized capacity. Winterization of coal \nand wind power plants is substantially cheaper, as no or very lim-\nited fuel supply infrastructure must be winterized. Winterization \ncost assumed at 10% of initial plant investments of coal power plants \nis far below revenues up to the full winterization of all failed coal \ncapacity. In fact, one could assume that winterization costs 30% of \ninitial plant investments, and even under this assumption winteriza-\ntion cost would be lower than the revenue for the completely winter-\nized capacity. For wind turbines, our estimates of revenue are half \nthose of coal, but are still substantially higher than the cost of win-\nterization, which is reported to be 5% of investment cost13.\nThe assumed discount rate has a strong impact on results. When \nincreasing the rate from 5% to 10%, the revenue for winterizing the \nfirst gigawatt of gas power plants is reduced from US$1.06 billion to \nUS$0.65 billion. While winterization of coal and wind power still \nfully pays off under these assumptions, the profitable winterization \nof gas infrastructure and gas power is reduced from 15\u2009GW to 7\u2009GW.\nBefore 2021, these estimates might have been lower, as  \nsince 1989 no climatic event of a similar magnitude to that of  \n2021 had been observed. When dropping 2021 from our set of \nevents, our estimates of expected revenue fall by 17.6% on aver-\nage, causing only 13\u2009GW of gas power to be profitably winterized. \nStill, when taking into account expected revenue, a substantial \namount of capacity should have been winterized. Further assess-\nment of the uncertainties, such as onset and recovery temperatures \n(Supplementary Fig. 11), in our modelling approach can be found \nin Supplementary Note 2.\nPotential reasons for the lack of winterization\nWe have shown that the Texas loss-of-load event in February 2021 \nwas among the top three extreme events when simulating the power \ngeneration system under climate conditions of the last 71\u2009years. \nNevertheless, winterization of power generation infrastructure was \n0.8\na\nb\nc\nd\n0.6\n0.4\n(billion US$ GW\u20131)\n0.2\n0\n2.5\n2.0\n1.5\n(billion US$ GW\u20131)\n1.0\n0\n0.5\n2.5\n2.0\n1.5\n1.0\n0\n5\n2\n4\n2\n4\n6\n8\n10\n12\n10\n15\nWinterized capacity (GW)\nWinterization cost\nConfidence interval = 95%\nConfidence interval = 68%\nExpected revenue over a 30 year period\nWinterized capacity (GW)\nWinterized capacity (GW)\nGas\nCoal\nWind north\nWind south\nWinterized capacity (GW)\n1.0\n1.5\n2.0\n2.5\n3.0\n0.5\n0.8\n0.6\n0.4\n0.2\n0\nFig. 5 | Revenues from winterization. a\u2013d, Comparison of revenues (billion US dollars per gigawatt of winterized capacity in a 30\u2009year period) to marginal \ncost for winterizing existing wind power in the north (a) and south (b), as well as gas (c) and coal (d) power generation capacity.\nNature Energy | VOL 7 | May 2022 | 409\u2013416 | www.nature.com/natureenergy\n413\n\nArticles\nNATurE EnErgy\nalready profitable before the event took place, taking into account \nthe past climate, potential climate-change impacts and current \nregulatory conditions. So why did power generators not winterize? \nWe identify several possible reasons for this gap between potential \nrevenue and observed winterization efforts.\nFirst, while expected revenue is above winterization cost, its vari-\nance is high, in particular for gas power plants: for this technology, \nthe winterization of the first gigawatt in 16% of all cases and the \nwinterization of the tenth gigawatt in 36.2% of all cases results in \nnegative profits. Additionally, owners of gas power plants have to \nsecure their gas supply and may be exposed to high gas prices dur-\ning cold spells. In contrast, the risk of not investing in winterization \nis low14. Even if a catastrophic failure occurs, the associated social \ncost is not born by the power generators. Potential costs include \nminor fines and damages to power plant equipment. A risk-averse \ninvestor may therefore decide against winterization. For owners of \ncoal and gas power plants, another fact may reduce expected rev-\nenue: we assumed a 30\u2009year lifetime for all power plants, but gas and \ncoal generation infrastructure is partly old, and winterization may \nnot pay off, if the plants are retired soon. For owners of wind power \nplants, however, we see the highest incentives and the lowest risk: \nthe fleet is comparably young, winterization cost is low and reve-\nnues are substantially higher. Even with relatively high risk aversion, \nwinterization seems to be a rational choice for wind power plant \ninvestors, especially if new wind parks are built.\nSecond, even though the historical temperature time series \nclearly indicates that the 2021 event could have been expected, the \nassociated outages of power plants may have been underestimated \nby power generators. This is supported by King et al.15, who show \nthat for a high number of plants, rated temperatures of failing power \nplants were in some instances substantially lower than prevailing \ntemperatures during outages in February 2021. As no major outage \nsuch as the one in February 2021 had been observed before, plant \nowners may have assumed that current power plant standards were \nreliable enough.\nThird, high discount rates also imply that winterization becomes \nsubstantially less attractive. At a 10% discount rate, our estimates of \nrevenues drop by about 38.8%. Therefore, alternative investments \nwith a higher return on investment and potentially lower risk may \nbe preferred if limited capital is available, in particular for the more \ncostly winterization of gas power generation.\nFourth, our calculation holds only under perfect competition. \nOwners of large generation assets have less incentive to additionally \nwinterize (Supplementary Fig. 12), as this would partly reduce scar-\ncity prices for the already winterized part of their fleet. However, \nthere is little concentration in the market in Texas, and we there-\nfore see strategic behaviour as a potentially less important reason \n(Supplementary Note 3).\nConclusions\nThe total cost of extreme freeze events to society is at least one order \nof magnitude higher than the winterization cost. Winterization is \ntherefore highly welfare enhancing. We have shown that current \nregulation provided a sufficient incentive for large-scale winteriza-\ntion for risk neutral investors, but risk aversion, lack of knowledge \nabout potential outages or higher-yielding alternative investments \nmay have impeded generators from making more effort to winter-\nize. A more stringent regulation of winterization therefore seems to \nbe necessary. In June 2021, Texas Senate Bills 2 and 3 became effec-\ntive. They require the winterization of power plant infrastructure. \nHowever, gas power plants do not have to fully winterize; instead, a \ncommittee will define which installations have to16. As the failure of \nthe gas power infrastructure by far had the largest impact on deficits \nin our simulations, we emphasize here that this impact should be \nstrongly considered when defining rules for enforcing winterization \nof power plants and associated infrastructure.\nWinterization of power generation units, however, may also \ncome with downsides during periods of warm temperatures, as \nmeasures used to increase performance during cold weather, such \nas integrating gas turbines into insulated buildings, may make the \ncooling of power plants more complex. Apart from winterization, \nother mitigation measures, in particular strong demand response \nprograms and an expansion of transmission capacities to neigh-\nbouring states, may therefore become important6. These mitigation \nmeasures may be substantially less costly, and they will be beneficial \nnot only during cold spells. Both options can make the system more \nresilient against other variations in power generation, in particu-\nlar taking into account the ongoing transition to a larger share of \nrenewable energies in the power generation mix. During extreme \nfreeze events, demand response may have to focus on industrial \napplications and less on households as electric heating in particular \ncannot be fully postponed during freeze events.\nOf course, our results have to be considered in light of a con-\ntinuously evolving power system. We assumed 30\u2009years of lifetime \nfor all installed capacities; however, some capacities, in particular \ncoal power plants, may soon be retired, and therefore, their winter-\nization may not be profitable. As the winterization of new capac-\nity is cheaper and easier to implement than the winterization of \nexisting capacity, winterization standards for installing new power \nplants and associated infrastructure should have a high priority. The \nongoing transformation of the Texan power system can therefore \nbe considered an opportunity to ensure robustness during future  \nfreeze events.\nMethods\nEstimation of temperature-induced electricity deficits. To estimate the amount \nof deficits in the power system, we simulated the difference between the expected, \ntemperature-dependent electricity demand and the available generation capacity. \nWe used a regression model to simulate the demand from observations of past \nload in the years 2012\u20132020. Available generation capacity was obtained from the \nexpected available capacity according to reports, reduced by outages related to \ntemperatures.\nAccording to ERCOT17, 67.5\u2009GW of thermal capacity was available during the \n2021 winter, but 4\u2009GW of this may have been under maintenance. We therefore \nassumed a value of 63.5\u2009GW of available thermal capacity before the freeze event.\nDemand prediction. We predicted demand from winter temperatures using \na regression model (equation (4) in the Supplementary Methods). To avoid \nthe overestimation of loads at low temperatures with this model, a threshold \ntemperature at \u20138\u2009\u00b0C was introduced, under which the temperature dependency of \nthe load was kept constant (Supplementary Fig. 2).\nThe model performed well for different temperature ranges in terms of \naverage bias (Supplementary Fig. 3), although at low temperatures we slightly \nunderestimated the load. We therefore also ran the whole model chain with an \nalternative specification of the demand model where the temperature impact \ndid not flatten off at \u20138\u2009\u00b0C. The results are reported in our sensitivity analysis, \nindicating that the estimated deficits do not strongly change when using a \ndifferent specification for the demand model (<10% difference). Testing the \nmodel out-of-sample for January 2021 delivered a high fit with a coefficient \nof determination (R2) of 0.92 and a root mean square error of 1.31\u2009GW. A \ncross-validation for other years (Supplementary Table 1) indicated a good fit.\nTemperature-dependent generator outages. We estimated large-scale \ninfrastructure failure aggregated by plant category. We derived the temperature, \nat which the largest increase in outages occurred for each power plant category, \nfrom the 2021 outage data (Supplementary Figs. 4\u20137). Furthermore, we estimated \na constant outage level in terms of tripping capacity. Finally, we also defined a \nconstant recovery rate, which describes how outages decrease after the recovery \ntemperature is reached. This approach will omit smaller outages, but it accounts \nbetter for the inter-dependency of failures. Furthermore, the uncertainty in the \ndata did not allow us to derive outage curves for individual plants. A detailed \noutline of how we derived the outage parameters from the 2021 data in Texas is \ngiven in the Supplementary Methods. The resulting outage models were applied to \nthe 71\u2009years of climate data to obtain the capacity availability during this period.\nEstimating revenues from winterization. We determined the revenues from the \nscarcity price mechanism implemented by ERCOT8. This mechanism comes into \neffect whenever spare capacity falls below 8\u2009GW. In this case, power market prices \nare regulated and set to fixed (high) values depending on the spare capacity. The \nNature Energy | VOL 7 | May 2022 | 409\u2013416 | www.nature.com/natureenergy\n414\n\nArticles\nNATurE EnErgy\nlower the spare capacity, the higher the price, reaching US$9,000\u2009MWh\u20131 when \nspare capacity falls under 2\u2009GW (Supplementary Fig. 8). The difference between \navailable capacity and demand was used to determine the current scarcity price \nand thus the revenue from winterization for each additional gigawatt of installed \ncapacity for the 71 available weather years. The total revenue for a 30\u2009year \ninvestment period with 5% discount rate was calculated. We did this for 10,000 \nscenarios, drawing randomly 30\u2009years from the available 71\u2009years to simulate \ndifferent realizations of climate.\nWe emphasize here that this calculation only holds under a perfectly \ncompetitive market. For generators with large capacities that are already partly \nwinterized, additional winterization may yield negative revenues, as market prices \nfor existing winterized assets are reduced by the additional winterization. We \ndiscuss this in more detail in Supplementary Note 3.\nFrequency analysis of freeze and power-deficit events. The frequency analysis of \nextreme events follows well-established methods of hydrological drought analysis18, \nwhere deficit events are defined as periods when the variable of interest is below a \ncertain threshold. Here, we use two different threshold concepts. First, we analyse \ntemperature and define a constant threshold of 0\u2009\u00b0C to define deficit events, in \nanalogy to drought events in drought statistics. Second, deficit events in the power \nsystem resulting from low temperatures are defined as periods when the capacity \ndeficit is >0\u2009GW (equation (1) in the Supplementary Methods). In each case, the \nresult is a derived deficit time series, which is further investigated using Yevjevich\u2019s \ntheory of runs19. During a freeze period, minor thaw episodes or other disturbances \nmay split an event into several smaller events. As a remedy, pooling procedures \nhave been recommended20. In this study, an inter-event time criterion of one day \nis used to define the deficit event series. In the case where multiple events occur in \na year, the event with the absolutely largest accumulated deficit is used for further \nanalysis. The derived series are characterized by three deficit characteristics: \nduration (measured in hours), intensity (minimum temperature and maximum \npower deficit for temperature and load deficit time series, respectively) and severity \n(aggregated frost sum and power deficit over the event for temperature and load \ndeficit time series, respectively), each of which constitutes an annual extreme value \nseries. These are further analysed using extreme value statistics to determine the \nreturn period of each freeze and capacity-deficit characteristic according to natural \nhazard management standards. Minor capacity-deficit events (with a duration \nof <6\u2009hours) are excluded as these are not extreme events and would distort \nthe extreme value modelling. Our analysis was conducted using the R-software \npackage lfstat (ref. 21), which provides a collection of state-of-the-art methods that \nare fully described in the World Meteorological Organization\u2019s manual on low flow \nestimation and prediction22.\nData. The temperature at 2\u2009m above ground is taken from the ERA5 reanalysis23. \nWe calculate the average temperature over Texas weighted by population density24 \nto derive a temperature index for modelling electricity demand. For estimating \noutages in the power system due to low temperatures, we derive the average \ntemperature weighted by the capacities of wind25, coal and natural gas power \nplants26, which were the power generation technologies most affected by failure \nduring the extreme temperature event of February 2021. For wind power plants, \nwe split Texas into northern and southern regions (along the latitude of 30\u00b0), since \ntemperatures at wind parks in the northern and southern regions of Texas differ \nsubstantially. Since the failure of the power system is also related to infrastructure \nat gas fields supplying these power plants27, we determine an average gas-field \ntemperature index, weighted by the distribution of natural gas production by \ncounty28 to complement our analysis. Load data used for demand prediction were \nretrieved from ERCOT29 for the period January 2004\u2013February 2021. Since the \nfocus of this study is on freeze events in winter, only winter load data (December\u2013\nFebruary) is used. Outage data is provided in the period since 10 February 2021 by \nERCOT10 and is aggregated by power generation technology for the analysis.\nData availability\nAggregated climate data from ERA5 as well as results from the analysis, such as \nestimated load and threshold time series resulting from available capacity reduced \nby estimated outages and marginal winterization cost under different scenarios, \nare provided openly to the community on Zenodo: https://doi.org/10.5281/\nzenodo.5902745. Data from public institutions, in particular ERCOT, the Energy \nInformation Administration and the Texas Railroad Commission, are not available \nunder an open license. However, within the description of the repository, links to \ndata sources and the whole code, including download scripts, are provided so that \nour analysis, and in particular all figures, can be fully reproduced. Source data are \nprovided with this paper.\nCode availability\nCode is published in a Github repository. The repository can be found at https://\ngithub.com/inwe-boku/texas-power-outages.\nReceived: 16 April 2021; Accepted: 9 February 2022;  \nPublished online: 4 April 2022\nReferences\n\t1.\t Bennett, J. A. et al. Extending energy system modelling to include extreme \nweather risks and application to hurricane events in Puerto Rico. Nat. Energy \n6, 240\u2013249 (2021).\n\t2.\t Thornton, H. E., Scaife, A. A., Hoskins, B. J. & Brayshaw, D. J. The \nrelationship between wind power, electricity demand and winter weather \npatterns in Great Britain. Environ. Res. Lett. 12, 064017 (2017).\n\t3.\t H\u00f6ltinger, S. et al. The impact of climatic extreme events on the feasibility  \nof fully renewable power systems: a case study for Sweden. Energy 178, \n695\u2013713 (2019).\n\t4.\t Pallone, F. MEMORANDUM Hearing on \u201cPower Struggle: Examining the \n2021 Texas Grid Failure\u201d https://energycommerce.house.gov/sites/democrats.\nenergycommerce.house.gov/files/documents/Briefing%20Memo_OI%20\nHearing_2021.03.24.pdf (2021).\n\t5.\t Busby, J. W. et al. Cascading risks: understanding the 2021 winter blackout in \nTexas. Energy Res. Soc. Sci. 77, 102106 (2021).\n\t6.\t Wu, D. et al. An open-source extendable model and corrective measure \nassessment of the 2021 Texas power outage. Adv. Appl. Energy 4, 100056 \n(2021).\n\t7.\t Doss-Gollin, J., Farnham, D., Lall, U. & Modi, V. How unprecedented was the \nFebruary 2021 Texas cold snap? Environ. Res. Lett. 16, 064056 (2021).\n\t8.\t Bajo-Buenestado, R. Operating reserve demand curve, scarcity pricing and \nintermittent generation: lessons from the Texas ERCOT experience. Energy \nPolicy 149, 112057 (2021).\n\t9.\t D\u2019Andrea, A. C. Preliminary Report on Causes of Generator Outages and \nDerates for Operating Days February 14\u201319, 2021 Extreme Cold Weather \nEvent. Access to the document is geo-restricted to the US (ERCOT, 2021); \nhttps://www.ercot.com/files/docs/2021/04/06/51878_ERCOT_Letter_re_\nPreliminary_Report_on_Outage_Causes.pdf\n\t10.\tERCOT. Generation Resource and Energy Storage Resource Outages and \nDerates, February 14-19, 2021. Access to the document is geo-restricted to the \nUS (ERCOT, 2021); https://www.ercot.com/files/docs/2021/03/04/ERCOT_\nLetter_Re_Feb_2021_Generator_Outages.pdf\n\t11.\tCohen, J., Agel, L., Barlow, M., Garfinkel, C. I. & White, I. Linking Arctic \nvariability and change with extreme winter weather in the United States. \nScience 373, 1116\u20131121 (2021).\n\t12.\tSheridan, S. C. & Lee, C. C. Temporal trends in absolute and relative extreme \ntemperature events across North America. J. Geophys. Res. Atmospheres 123, \n11889\u201311898 (2018).\n\t13.\tStarn, J. & Chia, K. Sweden shows Texas how to keep turbines going in icy \nweather. Bloomberg https://www.bloomberg.com/news/articles/2021- \n02-16/sweden-shows-texas-how-to-keep-turbines-spinning-in-icy-weather \n(2021).\n\t14.\tMays, J. et al. Private risk and social resilience in liberalized electricity \nmarkets. Joule https://doi.org/10.1016/j.joule.2022.01.004 (2021).\n\t15.\tKing, C. W. et al. The Timeline and Events of the February 2021 Texas Electric \nGrid Blackouts (Univ. of Texas at Austin Energy Institute, 2021).\n\t16.\tZou, I. Texas power generation companies will have to better prepare for \nextreme weather under bills Gov. Greg Abbott signed into law. The Texas \nTribune https://www.texastribune.org/2021/06/08/greg-abbott-texas-power- \ngrid-ercot/ (2021).\n\t17.\tERCOT. Seasonal Assessment of Resource Adequacy for the ERCOT Region \n(SARA) Winter 2020/2021. Access to the document is geo-restricted to the US \n(ERCOT, 2020); https://www.ercot.com/files/docs/2020/11/05/SARA- \nFinalWinter2020-2021.pdf\n\t18.\tLaaha, G. et al. The European 2015 drought from a hydrological perspective. \nHydrol. Earth Syst. Sci. 21, 3001\u20133024 (2017).\n\t19.\tYevjevich, V. M. An Objective Approach to Definitions and Investigations of \nContinental Hydrologic Droughts. PhD thesis, Colorado State Univ. (1967).\n\t20.\tTallaksen, L. & van Lanen, H. (eds) Hydrological Drought. Processes and \nEstimation Methods for Streamflow and Groundwater (Elsevier, 2004).\n\t21.\tKoffler, D., Gauster, T. & Laaha, G. lfstat: calculation of low flow statistics  \nfor daily stream flow data https://CRAN.R-project.org/package=lfstat  \n(2016).\n\t22.\tGustard, A. & Demuth, S. Manual on Low-Flow Estimation and Prediction \n(World Meteorological Organization, 2008).\n\t23.\tERA5 Monthly Averaged Data on Single Levels from 1979 to Present \n(Copernicus Climate Change Service, 2019).\n\t24.\tGridded Population of the World (GPW) v.4.11 (Center For International \nEarth Science Information Network at Columbia Univ., 2018).\n\t25.\tHoen, B. et al. United States Wind Turbine Database (US Geological Survey, \n2018); https://doi.org/10.5066/F7TX3DN0\n\t26.\tMaps (US Energy Information Administration, 2020); https://www.eia.gov/\nmaps/layer_info-m.php\n\t27.\tDouglas, E. Gov. Greg Abbott wants power companies to \u201cwinterize.\u201d Texas\u2019 \ntrack record won\u2019t make that easy. The Texas Tribune https://www.\ntexastribune.org/2021/02/20/texas-power-grid-winterize/ (2021).\n\t28.\tTexas Oil and Gas Production by County (RRC, 2020); https://www.rrc.state.\ntx.us/media/qcpp3bau/2020-12-monthly-production-county-gas.pdf\nNature Energy | VOL 7 | May 2022 | 409\u2013416 | www.nature.com/natureenergy\n415\n\nArticles\nNATurE EnErgy\n\t29.\tHourly Load Data Archives \u2013 ERCOT. Access to the document is geo-restricted \nto the US (ERCOT, 2021); http://www.ercot.com/gridinfo/load/load_hist\nAcknowledgements\nWe gratefully acknowledge support from the European Research Council (\u2018reFUEL\u2019 \nERC2017-STG 758149; J.S.). We are grateful to E. Virg\u00fcez, who provided spatial locations \nof power plant outages and with whom we exchanged early results, and to S. Wehrle with \nwhom we extensively discussed the paper.\nAuthor contributions\nConceptualization: K.G., T.G., P.R., G.L. and J.S.; software: K.G., T.G., P.R. and J.S.; \ninvestigation: K.G., T.G., P.R., G.L. and J.S.; writing the original draught: K.G., G.L., P.R. \nand J.S.; reviewing and editing the paper: K.G., T.G., G.L., P.R. and J.S.; visualization: \nK.G., T.G. and P.R.; supervision: J.S.; and funding acquisition: J.S. K.G. and J.S. equally \ncontributed as first authors to the manuscript.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary material \navailable at https://doi.org/10.1038/s41560-022-00994-y.\nCorrespondence and requests for materials should be addressed to Katharina Gruber.\nPeer review information Nature Energy thanks Peter Cramton, Le Xie and the other, \nanonymous, reviewer(s) for their contribution to the peer review of this work.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2022\nNature Energy | VOL 7 | May 2022 | 409\u2013416 | www.nature.com/natureenergy\n416\n\n\n    Scientific Research Findings:", "answer": "We find that since\u00a01950, eight events similar to the one observed in\u00a02021 would have occurred if the past climate had met the current power system and power markets. We further confirm previous findings that the frequency of cold events did not significantly decrease in the past seven decades, although the population\u2011weighted mean temperature in Texas has increased by\u00a00.017\u00a0\u00b0C per year on average. Under reasonable assumptions on discount rates and winterization cost, the frequency of cold events in Texas is sufficiently high to generate positive expected profits from investments into winterization under the current scarcity price mechanism. Importantly, this result is qualitatively robust to changes in assumptions on power plant outages and to changes in how climate warming\u2011induced mean temperature changes will affect extreme cold events. Consequently, the under\u2011investment can be explained by the high variance of profits from winterization, which may become negative due to the low frequency of cold events in some of our simulations.", "id": 15}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-021-00942-2\n1Crawford School of Public Policy, Australian National University, Canberra, Australian Capital Territory, Australia. 2Zero-Carbon Energy for the Asia-Pacific \nGrand Challenge, Australian National University, Canberra, Australian Capital Territory, Australia. 3Research School of Population Health, Australian \nNational University, Canberra, Australian Capital Territory, Australia. 4Centre for Aboriginal Economic Policy Research (CAEPR), Australian National \nUniversity, Canberra, Australian Capital Territory, Australia. 5School of Regulation and Global Governance (RegNet), Australian National University, \nCanberra, Australian Capital Territory, Australia. 6Tangentyere Council Aboriginal Corporation, Alice Springs, Northern Territory, Australia. 7Julalikari \nCouncil Aboriginal Corporation, Tennant Creek, Northern Territory, Australia. \u2709e-mail: michael.klerck@tangentyere.org.au\nI\nndigenous communities in remote Australia face temperature \nextremes that can increase their use of electricity and amplify their \nrisk of being disconnected. Energy is a necessary resource for work, \neducation, participation in social life and for maintaining healthy liv-\ning practices at home1\u20136. Energy insecurity remains a pressing issue \nglobally, including in countries with an abundance of wealth and \nresources4,7\u201310. It can be defined as \u2018an inability to meet basic household \nenergy needs\u20195 and is broadly synonymous with the concept of energy \npoverty11\u201313. Insufficient access to energy has been linked to poor \nhealth (both mental and physical) as energy is required to maintain \nessential services, including food security, lighting, essential medical \nequipment and thermal comfort/safety during extreme weather4,5,14\u201322.\nThere is a need to better understand the extent of energy insecu-\nrity experienced by Australia\u2019s remote Indigenous communities, in \nparticular the role that temperature plays in shaping energy insecu-\nrity. The vulnerabilities associated with energy insecurity vary spa-\ntially on the basis of underlying characteristics, which can be highly \nregionalized and locally specific23. Socio-economic, demographic \nand behavioural factors, as well as occupancy and structural char-\nacteristics (including the size, type and quality of housing stock and \nappliances), are all key drivers of energy consumption; while the \nprevailing temperature can affect the security of electricity supply \ndue to the cost of heating or cooling24,25.\nTemperature extremes are likely to act as a risk multiplier, wors-\nening energy insecurity for those at greatest risk as \u2018vulnerable \nhouseholds typically live in poorer quality housing, and have least \nresource or opportunity to invest in improvements to its efficiency \nand heating technology\u20196. The importance of access to energy has \nprompted governments worldwide to implement policies maintain-\ning this access, many with special attention to reducing the health \neffects of heat and cold7,26,27.\nThe climate of the Northern Territory (NT) ranges from equato-\nrial and tropical regions in the north to hot dry grassland regions \nin Central Australia (Fig. 1a). Remote Indigenous communities in \nthe NT are mostly off-grid and unregulated by the guidelines of the \nAustralian Energy Regulator28. In a situation unusual in Australia, \nremote living residents prepay for access to electricity and regularly \nexperience disconnection on non-payment. Distant from Australia\u2019s \nurban centres and major electricity grids, these communities have \nlong relied on diesel and gas-fired generators. In recent years, there \nhas been incremental integration of renewable energy into these \nisolated, high-cost electricity networks.\nAustralia\u2019s remote Indigenous communities face some of the \nhighest temperatures nationally and are vulnerable to the effects of \na warming climate (Fig. 1b\u2013d)29. Exposure to extreme temperatures \nhas been associated with a range of adverse health outcomes and \ndeath22,30\u201333. In the three hottest climate zones in Australia, between \n4.5 and 9.1% of all deaths were associated with heat-related mortal-\nity, which is an estimate that is much higher than in other Australian \nregions (2% nationally)31. The challenge of maintaining thermal \ncomfort and safety during temperature extremes is a pressing issue \nfor author N.F.J.: \u201cWe can\u2019t do anything about climate change except \nturn the power up, but it costs a lot too, don\u2019t forget that. Electricity, \nyou\u2019re using more power when you turn that air conditioner up!.\u201d \nTemperatures over 35\u2009\u00b0C, and even over 40\u2009\u00b0C, are increasingly com-\nmon in the NT as the climate changes (Fig. 1d). There is a need \nto better understand how extreme temperatures already shape the \nexperience of energy insecurity in remote Indigenous communities.\nBecause of the health implications of energy disconnection \nand the subsequent loss of essential services, there are questions \naround how strongly disconnection events relate to temperature \nand whether disconnections occur more frequently during extreme \nEnergy insecurity during temperature extremes in \nremote Australia\nThomas Longden\u200a \u200a1,2, Simon Quilty\u200a \u200a3, Brad Riley\u200a \u200a2,4, Lee V. White\u200a \u200a2,5, Michael Klerck\u200a \u200a4,6\u2009\u2709, \nVanessa Napaltjari Davis\u200a \u200a6 and Norman Frank Jupurrurla\u200a \u200a7\nIndigenous communities in remote Australia face dangerous temperature extremes. These extremes are associated with \nincreased risk of mortality and ill health. For many households, temperature extremes increase both their reliance on those \nservices that energy provides, and the risk of those services being disconnected. Poor quality housing, low incomes, poor health \nand energy insecurity associated with prepayment all exacerbate the risk of temperature-related harm. Here we use daily smart \nmeter data for 3,300 households and regression analysis to assess the relationship between temperature, electricity use and \ndisconnection in 28 remote communities. We find that nearly all households (91%) experienced a disconnection from electric-\nity during the 2018\u20132019 financial year. Almost three quarters of households (74%) were disconnected more than ten times. \nHouseholds with high electricity use located in the central climate zones had a one in three chance of a same-day disconnection \non very hot or very cold days. A broad suite of interrelated policy responses is required to reduce the frequency, duration and \nnegative effects of disconnection from electricity for remote-living Indigenous residents.\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n43\n\nArticles\nNATure Energy\ntemperatures. In this paper, we assess the relationship between tem-\nperature, electricity use and disconnection using daily smart meter \ndata for 28 remote Indigenous communities in the NT, all of which \nare using prepayment metering. We then present how many discon-\nnection events occur during example temperatures to indicate the \nextent and severity of energy insecurity attributable to temperature \nextremes. Quantification of energy insecurity and how it is related \nto temperature thresholds could support future policy responses.\nEnergy injustice and a history of policy exceptionalism\nSimilar to Indigenous peoples worldwide, communities in the \nremote NT have long been at greater risk across the three dimen-\nsions of energy injustice. Energy justice is concerned with the dis-\ntribution of costs, benefits and risks in energy transitions and is a \nprinciple that arises from theories of distributional justice2,6,9,34\u201337. \nDistributional injustices in energy systems can be produced by \nsystemic inequalities that arise from ongoing procedural injustices \n(which is a concept rooted in the failure to accord some groups of \npeople equal rights and respect)38 and recognition injustices (when \ncertain groups face a lack of cultural respect and are excluded from \ndecision-making)39. Procedural injustices extend to differences in \nlegal rights; distributional injustices are reflected in differences in \nthe quality of housing that affects energy use; and injustices in rec-\nognition include a lack of acknowledgement of the unique needs \nof specific communities, including those associated with energy \naccess, use and practices16. Some groups, such as remote Australian \nIndigenous communities, consistently have less access to energy \nefficient housing, less ability to shape the electricity systems that \nthey are connected to and less ability to pay for higher electricity \ncosts that may result from certain systems6,9. These aspects tend to \nbe intertwined.\nFor author V.N.D., who works on issues related to energy, hous-\ning and social justice in Central Australia, maintaining access to \nelectricity during temperature extremes represents a complex suite \nof interrelated challenges: \u201cOlder houses had solar hot water and \npot belly stoves for the winter. We could collect wood and the sun \nheated the water. The new houses built by the Government since the \nIntervention (in 2007) have electric hot water heaters and no pot \nbelly stoves. When the old houses were upgraded, pot belly stoves \nwere removed. Our houses don\u2019t have heating anymore. Most resi-\ndents don\u2019t have much money, so residents buy cheap fan heaters \nand air-cons. The problem with these is that they are expensive to \nrun. Our houses have become expensive to heat and expensive to \ncool and we run out of money for electricity. When the power goes \noff it is bad for our health, the food gets spoiled, we can\u2019t wash our \nclothes and we can\u2019t wash our kids.\u201d\nResidents of remote communities live in extremely challenging \nsocio-economic circumstances and in housing that can be extremely \na\nb\nc\nd\nNT\nNT\nNT\nNT\nClimate classification of Australia\nProjection: Lambert conformal with\nstandard parallels 10\u00b0 S, 40\u00b0 S.\nBased on a modified Koeppen\nclassification system and\non a standard 30-year\nclimatology (1961\u20131990)\nCommonweath of Australia, 2005\nLowest minimum temperature\n1 July 2018 to 30 June 2019\nAustralian Bureau of Meteorology\nHighest maximum temperature\n1 July 2018 to 30 June 2019\nAustralian Bureau of Meteorology\nMaximum temperature anomalies\n(1961\u20131990 climate)\n1 July 2018 to 30 June 2019\nAustralian Bureau of Meteorology\n3.0 \u00b0C\n2.5 \u00b0C\n2.0 \u00b0C\n1.5 \u00b0C\n1.0 \u00b0C\n0.5 \u00b0C\n0 \u00b0C\n\u20130.5 \u00b0C\n\u20131.0 \u00b0C\n\u20131.5 \u00b0C\n\u20132.0 \u00b0C\n\u20132.5 \u00b0C\n\u20133.0 \u00b0C\n45 \u00b0C\n42 \u00b0C\n39 \u00b0C\n36 \u00b0C\n33 \u00b0C\n30 \u00b0C\n27 \u00b0C\n24 \u00b0C\n21 \u00b0C\n18 \u00b0C\n15 \u00b0C\n12 \u00b0C\n9 \u00b0C\n6 \u00b0C\n3 \u00b0C\n0 \u00b0C\n\u20133 \u00b0C\n\u20136 \u00b0C\n45 \u00b0C\n42 \u00b0C\n39 \u00b0C\n36 \u00b0C\n33 \u00b0C\n30 \u00b0C\n27 \u00b0C\n24 \u00b0C\n21 \u00b0C\n18 \u00b0C\n15 \u00b0C\n12 \u00b0C\n9 \u00b0C\n6 \u00b0C\n3 \u00b0C\n0 \u00b0C\n\u20133 \u00b0C\n\u20136 \u00b0C\nClimate classes\nEquatorial\nTropical\nSubtropical\nDesert\nGrassland\nTemperate\nRainforest (monsoonal)\nRainforest (monsoonal)\nRainforest (persistently wet)\nSavannah\nSavannah\nNo dry season\nDistinctly dry summer\nDistinctly dry winter\nModerately dry winter\nHot (persistently dry)\nHot (persistently dry)\nWarm (persistently dry)\nWarm (persistently dry)\nWarm (summer drought)\nNo dry season (hot summer)\nModerately dry winter\n(hot summer)\nDistinctly dry (and hot)\nsummer\nNo dry season\n(warm summer)\nModerately dry winter\n(warm summer)\nDistinctly dry (and warm)\nsummer\nNo dry season\n(mild summer\nDistinctly dry (and mild)\nsummer\nNo dry season\n(cool summer)\nHot (summer drought)\nHot (summer drought)\nHot (winter drought)\nHot (winter drought)\n\u00a9\nFig. 1 | NT compared with other Australian regions. a, Climate zones. b, Highest maximum temperatures. c, Lowest minimum temperatures. d, The \n12-monthly mean maximum temperature anomaly for Australia compared with 1961 to 1990. Panels reproduced with permission from the Australian \nGovernment Bureau of Meteorology under Creative Commons license CC-BY 3.0 AU: a, ref. 73; b, ref. 72; c, ref. 74; d, ref. 75.\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n44\n\nArticles\nNATure Energy\ncrowded (see Supplementary Table 1 for example statistics from the \n2016 census). Housing quality is often poor across these remote \ncommunities. Approximately half of Indigenous households \nin the NT fall below the poverty line. The National Indigenous \nReform Agreement (Closing the Gap), the core of the Australian \nGovernment\u2019s agenda to address social and health inequities  \nfacing Indigenous Australians, identifies \u2018healthy homes\u2019 as a key \npriority for healthy living practices40,41. Nine priorities are identified, \nwith six pertaining to electricity systems, which are power connec-\ntion, electrical safety, heating for showering, facilities to wash chil-\ndren, laundry facilities and facilities to store food and prepare food \n(including refrigeration)41.\nMoreover, Aboriginal and Torres Strait Islander peoples have \nexperienced frequent changes in the policy environment regulat-\ning their lives and lands, via a non-Indigenous regime of law. Many \ncommunities in the NT have been subject to frequent changes in \nregulatory practices for electricity and regressive changes in pro-\ncedural and recognition justice aspects, such as cycles of gain then \nsubsequent loss of representation in governing bodies37. Some of \nthis complex regulatory and legislative history is summarized in \na non-exhaustive timeline of key developments in policy affecting \nIndigenous peoples in the Territory between 1967 and 2021 (Fig. 2). \nThis includes the unilateral introduction of a \u2018user-pays\u2019 model for \nenergy provision in 1992. Author and Warumungu elder N.F.J. has \nlived experience of these deep structural imbalances that impact, \ntoo-often detrimentally, on the lives and livelihoods of Indigenous \ncommunities in the NT: \u201cI reckon the Government doesn\u2019t want \nto listen to Wumpurrarni (Indigenous) people because I reckon \nthey\u2019ve had enough and they\u2019re just ignoring us now, they think we \nget everything for free but we struggle for that. Policy is like a bible, \nfor Government, it tells them how they run things, how they can do \nthings. If they don\u2019t have a policy, they don\u2019t know how things run. \nAnd if they have a policy they can jam you and that\u2019s what happens \nto us, they jam us all the time.\u201d\nLimited protections and prepaying for power in the  \nremote NT\nPrepayment for electricity is uncommon in urban Australia. It is \nheavily regulated in most jurisdictions on the basis of concerns for \nwellbeing42\u201345. It remains disproportionately common in small and \nwidely dispersed remote communities across the NT, Queensland, \nWestern Australia and South Australia. In the NT prepayment is \nnot limited to urban town camps and remote communities. It is also \nused in urban and regional settings, including Darwin, Palmerston, \nNhulunbuy, Katherine, Tennant Creek and Alice Springs. Many of \nthese communities have prepaid electricity services as their only \noption for service provision46. There is considerable variation in \nthe operation of services and available protections for prepayment \nconsumers subnationally. As an example, in other parts of Australia \nwhere consumers are protected by the Australian Energy Regulator \nguidelines, people cannot be disconnected from electricity when life \nsupport medical equipment is being used43. This protection is not \ncomprehensively applied in remote NT communities47,48.\nIn previous international studies, rates of disconnection among \nprepayment households ranged from 10% to 53% for the UK, \nGermany and New Zealand49\u201352. Prepayment disconnection num-\nbers for Australia are not systematically collected or reported by \nregulators or providers and estimates are scarce. Previous analysis \nin the NT found prepayment disconnection rates between 59% \nand 91% (ref. 53). In comparison, the rate of \u2018raised disconnec-\ntions\u2019 for postpayment households in other Australian regions \nthat are most at risk of disconnection ranged from 3% to 30% with \nlarge variation in disconnection rates associated with local and \nregional socio-economic characteristics and whether smart meters \nwere commonly used54. The St Vincent de Paul Society and Alviss \nConsulting report defines a \u2018raised disconnection\u2019 as the case when \na \u2018retailer raises a service order with the relevant network busi-\nness\u2019. These raised disconnections may be rejected, cancelled or \ncompleted. They may be rejected on the basis of an invalid address \nor when disconnection is prohibited for medical reasons. And the \ndisconnection request can be cancelled by the retailer when the \npayment issue has been resolved. This can be a full payment or the \nestablishment of a payment plan54. These rates of disconnections are \nthe higher end of estimates as they are those that were found for the \ntop 30 postcode regions from four Australian areas (that is, New \nSouth Wales, Victoria, South East Queensland and South Australia). \nAcross the eastern (and most populous) parts of Australia, the rate \nof completed disconnections for postpayment households was 1% \n(ref. 55).\nWhile data are scarce, issues with prepayment and disconnection \nin other regions of Australia have been noted by key organizations, \nincluding the Essential Services Commission of South Australia44, \nEnergy and Water Ombudsman New South Wales45, the Queensland \nCouncil of Social Service56 and Bushlight46. The last two raised these \nconcerns specifically in relation to remote Indigenous communities.\nDisconnection as an indicator of energy insecurity\nWhile energy insecurity describes more than disconnection rates \nalone, having an electricity connection is the first part of being \nable to access electricity to meet household energy needs. Here we \nassess disconnection rates as a proxy for energy insecurity, while \nnoting that other factors also contribute to the disconnection rates \nobserved. This may include housing design, construction and insu-\nlation; sociodemographic factors such as income and health; and \nentrenched regulatory structures.\nWhen the households in our dataset run out of credit, they face \nimmediate disconnection between 10:00 and 14:00. Outside these \nhours, credit is extended and the accumulated debt is automatically \ndeducted from the subsequent purchase of energy credit. Energy \nservices are not recommenced until the accrued debt is paid57.\nFor the houses in our study with complete data for the 2018\u2013\n2019 financial year (July 2018 to June 2019), 91% of households \nexperienced a disconnection. Most disconnections were same-day \ndisconnections (92% of the total) where the meter was reconnected \nwhen credit was restored on the same day. A total of 71% of house-\nholds experienced a same-day disconnection more than ten times in \na 12-month period. These rates of disconnection are much higher \nthan Australian and international examples (mentioned above). On \naverage, a same-day disconnection lasted for almost 3\u2009h. Multi-day \ndisconnections were less common (0.9% of days), but two thirds \nof households experienced this type of disconnection, which lasted \novernight or longer. Of all households, 7% experienced a multi-day \ndisconnection more than ten times in 1\u2009year. These disconnections \ncould last for many days (the all-region average was almost 4\u2009days). \nTable 1 shows how these disconnections differed by climate zones.\nElectricity use during temperature extremes\nTo confirm our expectation that temperature affects electricity use, \nwe used linear regression with panel-corrected standard errors to \nprovide estimates for seven daily temperature ranges. These tem-\nperature ranges are specified in the regressions as the daily aver-\nage temperature for the same day as electricity use and the daily \naverage temperature for the 2\u2009days before. We estimate these regres-\nsions for all climate regions pooled together, and for each climate \nregion separately. The regressions were also estimated for all houses \nand then re-estimated for different levels of the average daily use of \nelectricity. Our data did not include any information on household \ncharacteristics, so we have used the average level of daily electricity \nuse to group the households. Determinants of the average daily load \ninclude occupancy and the use of appliances. For more information, \nrefer to the methods section for more details on the grouping of \nhouseholds with high/low electricity use.\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n45\n\nArticles\nNATure Energy\nReferendum \nNationwide referendum in which \nAustralians voted to amend the \nconstitution to count Aboriginal \nand Torres Strait Islander peoples \nin the national census, granting full \nrights as citizens and compelling \nthe Commonwealth Government to \ntake action in Aboriginal affairs.\nNorthern Territory (Self-\nGovernment) Act 1978\nNorthern Territory Government \nformed with the granting of self- \ngovernment.\nNorthern Territory Electricity \nCommission established.\nThe Northern Territory \u2018Intervention\u2019\nA complex range of measures enacted \nthrough the Northern Territory National \nEmergency Response Act 2007. The \nintervention effected changes to welfare \npayments, education, employment and \nhealth services in remote NT communities. \nCompulsorily acquired five-year leases over \n70 communities, implemented universal \nwelfare quarantining and removed the \nlegislative protection of the Racial\nDiscrimination Act (Cwlth) 1975 in relation\nto these changes.  \nCommonwealth Aboriginal\nLand Rights (Northern\nTerritory) Act 1976   \nIn 1976, the Commonwealth \npassed through parliament \nland rights legislation over \nthe Northern Territory,\nbeginning the process of \nrecognizing the continued \ncollective ownership of \n\u2018Country\u2019 by Aboriginal\nTraditional Owners. \nIndigenous Essential Services (IES)  \nIn 2003, Power and Water \nCorporation established IES Pty Ltd.\nA not-for-profit entity, IES is \nresponsible for the delivery of \nessential services to remote \ncommunities.\n1976\nPower and Water Authority\n(PAWA)\nNorthern Territory Electricity \nCommission and Water \nAuthority amalgamate to form \nPower and Water Authority \n(PAWA). By 1987, Northern\nTerritory Government indicates\nits intention to begin charging for\nelectricity on remote communities.\nAboriginal and Torres \nStrait Islander Commission\n(ATSIC) (1990\u20132004)\nThe first national \nrepresentative body to give \nAboriginal and Torres \nStrait Islander Australians \ndecision-making capacities.\nATSIC was established in \n1990 replacing DAA. ATSIC \nwas abolished in 2004.\nManymak energy efficiency programme \n\u2018Dharray Manymakkung Pawaw Ga Gapuw\u2019\nmeans \u2018looking after power and water\u2019 in \nthe East Arnhem Land language of \nDjampbarrpuyngu. Energy efficiency \nprogramme ran in six communities between \n2013 and 2015 assisting households in energy \nefficiency practices.\nMabo versus Queensland (No. 2) 1992\nThe landmark Australian high court ruling of Mabo versus \nQueensland (No. 2) overturned the fiction of \u2018terra nullius\u2019 \n(a land \u2018belonging to no-one\u2019) and recognized the fact that \nAboriginal and Torres Strait Islander peoples had lived in \nAustralia for thousands of years, enjoying rights to lands \naccording to their own laws and customs. It ruled that \nAboriginal and Torres Strait Islander property rights, \ntermed \u2018native title\u2019, could survive colonial annexation.\nNative Title Act (Cwlth) 1993\nNative Title was introduced as a result of the \nMabo decision (1992). Native Title is the \nrecognition that Aboriginal and Torres Strait \nIslander peoples have rights and interests to \nland and waters according to their traditional \nlaw and customs as set out in Australian law. \nNative Title is governed by the Native Title Act\n(Cwlth) 1993.  \nUnited Nations Declaration of the Rights of Indigenous \nPeoples (2007) (UNDRIP)\nOn 3 April 2009, Australia reversed its initial (2007)\nopposition, to express support for the United Nations \nDeclaration on the Rights of Indigenous Peoples.  \nStudy period\nJanuary 2018\u2013\nJuly 2019.\nUluru Statement from the Heart \nOn 26 May 2017, delegates \nfrom the First Nations National  \nConstitutional Convention \ncalled for the establishment of a\n\u2018First Nations Voice\u2019 to be \nenshrined in the Constitution. \nSeeks a Makarrata Commission \nto supervise agreement making \nbetween governments and First \nNations and \u2018truth-telling\u2019 of \nAustralia\u2019s shared history.\nGovernment issued a statement\nrejecting the Uluru Statement\nfrom the Heart on the 26\nOctober 2017.\nRacial Discrimination Act 1975 (Cwlth)\n(RDA)   \nMaking racial discrimination in certain \ncontexts illegal in Australia, overriding \nstate and territory legislation to the extent \nof any inconsistency. \n1967\n1975\n1990\n1992\n2007\n1978\n2009\n2017\n1986\nThe introduction of \u2018user-pays\u2019 model \nin NT remote communities\nDomestic residential users in remote \nAboriginal communities begin paying\nfor power 1 January 1992.\nIncremental introduction of token and \ncard prepayment metering.   \n1992\nIndigenous Community \nEngineering Guidelines \n(ICEG)\nUnregulated by the\nUtilities Commission, the \nutility reports on minimum \nservice levels as outlined in \nthe ICEG.   \nPower and Water \nCorporation (PWC)\n1 July 2002\nbecame the first \ngovernment-owned \ncorporation in the \nTerritory, replacing \nPAWA as the \nresponsible\nelectricity provider for\nremote communities.\nSolar Energy Transformation Program \n(SETuP) is one of the largest isolated off-grid\nsolar programmes in remote communities, \nintegrating 10 MW of solar photovoltaics in 25 \ncommunities.  \nRollout of \u2018smart\u2019 prepayment metering to \nreplace token prepayment.\n2002\n2003\n2014\n2015\n2018\n2020\nClosing the Gap in partnership\n27 July 2020 the Australian\nGovernment committed to the National \nAgreement on Closing the Gap in \nIndigenous health inequality (in \npartnership) through: shared decision- \nmaking, building the community \ncontrolled sector, transforming \ngovernment organizations and shared \naccess to data and information at a \nregional level. \nOffice of Aboriginal Affairs\n(OAA) 1967\u20131972\nDepartment of Aboriginal \naffairs (DAA) 1972\u20131990\nOffice of Indigenous Policy \nCoordination (OIPC) 2004\u20132011\nFamilies, and Housing, Community\nServices and Indigenous Affairs\n(FaHCSIA) 2006\u20132014\nAboriginal and Torres\nStrait Islander\nCommission \n(ATSIC)1990\u20132004\nDepartment \nof Prime\nMinister and\nCabinet\n(DPM&C) \n2014\u2013present\nNational \nIndigenous\nAustralians\nAgency\n(NIAA) 2019\u2013\npresent\nFig. 2 | Timeline of the complex history of Indigenous policy in Australia. Key developments in Indigenous policy, including remote community energy \npolicy, between 1967 and 2021. Boxes at the bottom are the federal departments or agencies responsible for Indigenous policy.\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n46\n\nArticles\nNATure Energy\nThe regression estimates confirm that electricity use differs \nnotably by temperature, climate zone and month (Fig. 3). Estimates \nfor temperature-related increases in electricity use are shown in \nFig. 3, with the number of days that these temperatures occurred \nand the estimates for the monthly change in electricity use (with-\nout the daily temperature effect). These estimates for temperature \nand monthly effects should be interpreted in relation to the refer-\nence temperature range (daily average temperatures between 20\u00b0C  \nand 25\u2009\u00b0C).\nGiven that the climate across the northern half of the NT is \ncharacterized by tropical heat (and mild cool nights only during \na short winter season), daily electricity use was on average higher \nin the hottest periods of the year (November to March). This sea-\nsonal increase in electricity consumption was most pronounced in \nhigh-use households, which also experienced a reciprocal reduc-\ntion in monthly electricity consumption in the cooler months. As \nexpected in the NT, which generally has a prevailing hot climate, \nhousehold energy expenditure is greatest during hotter weather due \nto the need for cooling.\nBeyond seasonal effects, the need for heating and cooling can \ninfluence the daily use of electricity. The hot persistently dry \ngrassland climate zone is the region that predominantly deter-\nmines the all-regions result (Fig. 3b). It is the combination of the \nCentral Australian climate zones shown in Fig. 1a. For this climate  \nzone, which unlike the other regions experiences cold nights, the \nhouseholds with the highest electricity use (top tenth percentile of \naverage daily load) increased their electricity use by 30\u2009kWh (on \naverage) on the coldest of nights (between 0\u00b0C and 10\u2009\u00b0C). The \naverage increase was 17\u2009kWh across all houses in this climate zone. \nExtremely hot days with average temperatures between 30\u00b0C and \n40\u2009\u00b0C corresponded to a 16\u201319\u2009kWh increase (on average) for the \nhouseholds with the highest electricity use. When considering all \nhouses, the average increase was 6\u20138\u2009kWh.\nDisconnection during temperature extremes\nTo assess the question of whether temperature influences the prob-\nability of disconnection, we used random-effects probit regressions \nto estimate the probability of same-day and multi-day disconnec-\ntions. Figure 4 shows the estimates for temperature-related increases \nin the probability of a same-day disconnection, which includes daily \nestimates and the estimates for the monthly change in disconnec-\ntions (without the daily temperature effect). The estimates are inter-\npreted in relation to a reference temperature range (daily average \ntemperatures between 20\u00b0C and 25\u2009\u00b0C). These estimates are also \nre-estimated by the level of electricity use and climate zones.\nThe probability of a same-day disconnection occurring on any \ngiven day (except during weekends and public holidays when dis-\nconnection is prohibited) is high (0.04\u20130.06) and increases on the \nfirst day that credit can expire, predominantly the next business day \n(approximately 0.19). This is captured in our results, with a large \nTable 1 | Disconnection rates by climate zone between July 2018 and June 2019 (n\u2009=\u20091,045,725)\nEquatorial climate \nzone (most northern)\nCoastal tropical \nclimate zone\nSavannah tropical \nclimate zone\nHot persistently dry \ngrassland climate \nzone (most southern)\nAll regions\nSame-day disconnections\nPercentage of households experiencing a \nsame-day disconnection at least once\n85\n94\n90\n90\n89\nPercentage of households experiencing a \nsame-day disconnection at least ten times\n60\n77\n75\n75\n71\nPercentage of days in the sample with a \ndisconnection\n7\n11\n11\n10\n10\nAverage length of disconnection (hours)\n2.92\n2.89\n2.69\n2.19\n2.67\nMulti-day disconnections\nPercentage of households experiencing \nmulti-day disconnection at least once\n47\n68\n71\n83\n66\nPercentage of households experiencing \nmulti-day disconnection at least ten times\n2\n4\n9\n16\n7\nPercentage of days in the sample with a \ndisconnection\n<1\n1\n1\n2\n1\nAverage length of disconnection (hours)\n102.23\n87.82\n73.29\n125.49\n98.48\nAny type of disconnection\nPercentage of households experiencing any  \ntype of disconnection at least once\n87\n95\n91\n92\n91\nPercentage of households experiencing any  \ntype of disconnection at least ten times\n63\n80\n78\n78\n74\nPercentage of days in the sample with a \ndisconnection\n7\n11\n12\n12\n10\nAverage length of disconnection (hours)\n8.60\n8.46\n8.39\n17.66\n10.62\nDaily electricity use and expenditure\nDaily electricity use (kWh) (s.d.)\n21.25 (16.46)\n23.29 (15.24)\n25.73 (16.86)\n26.95 (22.05)\n24.13 (17.75)\nDaily expenditure (AUD$) (s.d.)\n6.09 (4.71)\n6.67 (4.37)\n7.37 (4.83)\n7.72 (6.31)\n6.91 (5.08)\nCount of observations for balanced panel for \n2018\u20132019 financial year\n306,600\n229,585\n296,380\n213,160\n1,045,725\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n47\n\nArticles\nNATure Energy\na\nb\nSelected regression estimates by level of avergae daily load (ADL) (with 95% CI in grey)\n30\n20\n10\n0\n0\n\u201320\n\u201310\n0\n10\n400\n800\n1,200\n0\u201310\nDaily electricity use (kWh)\nDaily electricity use (kWh)\n\u201320\n\u201310\n0\n10\nDaily electricity use (kWh)\n0\n25\n50\n75\n100\n0\n10\n20\n30\nDaily electricity use (kWh)\nNumber of observations\n(thousands)\nNumber of observations\n(thousands)\n10\u201315\nAll houses\nHigh ADL (90\u2013100%)\nLow ADL (0\u201310%)\nSelected regression estimates by level of avergae daily load (ADL) (with 95% CI in grey)\nAll houses\nHigh ADL (90\u2013100%)\nLow ADL (0\u201310%)\n15\u201320\n20\u201325\n25\u201330\n30\u201335\n35\u201340\nHigh\nAll\nLow\nHigh\nAll\nLow\nHigh\nAll\nLow\nHigh\nAll\nLow\nDaily average temperature (\u00b0C)\n0\u201310\n10\u201315\n15\u201320\n20\u201325\n25\u201330\n30\u201335\n35\u201340\nDaily average temperature (\u00b0C)\n0\u201310\n10\u201315\n15\u201320\n20\u201325\n25\u201330\n30\u201335\n35\u201340\nDaily average temperature (\u00b0C)\n0\u201310\n10\u201315\n15\u201320\n20\u201325\n25\u201330\n30\u201335\n35\u201340\nDaily average temperature (\u00b0C)\nJune\nJuly\nAugust\nSeptember\nOctober\nNovember\nDecember\nJanuary\nFebruary\nMarch\nApril\nMay\nMonth\nJune\nJuly\nAugust\nSeptember\nOctober\nNovember\nDecember\nJanuary\nFebruary\nMarch\nApril\nMay\nMonth\nFig. 3 | Daily electricity use by temperature and month. a, All regions. b, Hot persistently dry grassland climate zone. These are the coefficient estimates and  \n95% confidence intervals from multiple regressions for a sample with 1,674,786 daily observations across 3,300 houses. Regressions were grouped by percen\u00ad\ntile of electricity use (that is, average daily load) and climate zones. Estimates for all houses, low-electricity-use households and high-electricity-use households \nare shown here (Supplementary Tables 2\u20136 have all of the electricity use estimates). Temperature bins are specified using daily average temperatures (in \u00b0C). \nTemperature-based estimates are for a three-day period (that is, temperature on the day of electricity use and the two days before). Temperature estimates are \ninterpreted using 20\u201325\u2009\u00b0C as the reference temperature range. Monthly estimates are interpreted using April as the reference month.\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n48\n\nArticles\nNATure Energy\nincrease in disconnections occurring on Monday and the day after a \npublic holiday (Fig. 4). There is a significant relationship with tem-\nperature that is most notable for the households with the highest \nelectricity use in the two southern climate zones.\nFor the full sample, there was a one in seventeen chance (prob-\nability of 0.06) of a same-day disconnection occurring on moder-\nate days with average temperatures between 20\u00b0C and 25\u2009\u00b0C. This \nincreased to a one in eleven chance (probability of 0.9) on hot days \nwith average temperatures between 35\u00b0C and 40\u2009\u00b0C. A series of cold \nnights had a significant effect with an almost one in six chance \n(probability of 0.18) of a same-day disconnection occurring on cold \ndays with average temperatures between 0\u00b0C and 10\u2009\u00b0C.\nThe households with the highest electricity use had a much \ngreater probability of same-day disconnection. For this group, there \nwas almost a one in seven chance (a probability of 0.15) of experi-\nencing a same-day disconnection on moderate days with average \ntemperatures between 20\u00b0C and 25\u2009\u00b0C. This increased to one in \nthree (probability of 0.35\u20130.39) for the coldest temperatures (0\u00b0C \nto 15\u2009\u00b0C) and one in four (probability of 0.24 to 0.27) for the hottest \ntemperatures (30\u00b0C to 40\u2009\u00b0C).\nClimate zones also influenced the probability of a same-day dis-\nconnection occurring (Fig. 4b and Supplementary Information). \nFor households with the highest electricity use in the southern-most \nclimate zone (that is, hot persistently dry grasslands shown in Fig. \n4b), a one in seven chance (probability of 0.14) of same-day discon-\nnection for temperatures between 20\u00b0C and 25\u2009\u00b0C, increased to one \nin three (probability of 0.31) for the coldest temperatures (0\u00b0C to \n10\u2009\u00b0C) and one in four (probability of 0.23) for the hottest tempera-\ntures (30\u00b0C to 40\u2009\u00b0C). For the households with the highest electric-\nity use in the savannah tropical climate zone, there was a one in \nfour chance (probability of 0.23) of disconnection for temperatures \nbetween 20\u00b0C and 25\u2009\u00b0C, which increased to one in three (probabil-\nity of 0.37 to 0.39) for the hottest temperatures (3\u00b0C to 40\u2009\u00b0C).\nOnly a weak relationship between temperature and multi-day \ndisconnections was found. The estimation results are provided in \nSupplementary Table 12. While rarer (approximately one-tenth as \ncommon), multi-day disconnection events lasted for an average of \n4\u2009days (Table 2).\nNumber of disconnections during temperature extremes\nTemperature-related disconnections are driven by an increased \nneed for electricity to maintain thermal comfort and safety during \nextreme temperatures. We now focus on the proportion of discon-\nnections that occurred during hot and cold temperatures for two \nreasons. First, these are critical events where expenditure on energy \nhas increased due to cooling/heating, leading to a disconnection that \ncompromises the other functioning of the home, including refrig-\neration, lighting and life support medical equipment (for example, \noxygen concentrators, sleep apnoea machines, home renal dialysis \nequipment). Second, there is a concern about the impact on health. \nProtections internationally include several examples of restricting \ndisconnection for vulnerable customers, including on the basis of \nhealth risks and outdoor temperatures7,26,48. These protections can \ninclude disconnection prohibitions based on the time of year (for \nexample, no disconnections during winter months in cold climates), \non reaching specific temperature thresholds, and on declarations of \nextreme weather events (for example, no disconnections during a \ndeclared heat wave event)26. Using example temperature thresholds \nto determine the number of temperature-related disconnections, \nwe find that over 49,000 incidences of disconnection (29% of dis-\nconnections) occurred during hot and cold temperature extremes \n(Table 2). We examine both 35\u00b0C and 40\u2009\u00b0C as the threshold for \nextreme heat.\nDiscussion\nWe begin to address the need to better understand how temperature \naffects energy insecurity in Australia\u2019s remote communities by exam-\nining (1) whether temperature affects electricity use, (2) whether \ntemperature influences the probability of disconnection and (3) the \nproportion of temperature-related disconnections (that is, discon-\nnections that occur during extreme temperatures). Temperature is \nconfirmed to effect electricity use. Correspondingly, disconnections \nare more likely during extreme temperatures. We find that in the \n28 remote Indigenous communities that are the focus of this study, \ndisconnections increase from an already high baseline of one in \nseventeen during mild temperatures (20\u201325\u2009\u00b0C), to a one in eleven \nchance of disconnection during hot days (34\u201340\u2009\u00b0C) and a one in \nsix chance during cold days (0\u201310\u2009\u00b0C). Disconnection occurs more \nfrequently for households with the highest electricity use in the cen-\ntral climate zones, which had a one in three chance of a same-day \ndisconnection on very hot or very cold days. This indicates that \nhouseholds are having trouble cooling/heating their homes, which \nin turn compromises access to other essential services including \nrefrigeration, lighting and essential medical devices. While the level \nof energy service that is viewed as \u2018essential\u2019 can vary over time and \nwith changing social norms58, a complete loss of access to energy \nservices constitutes a level of energy insecurity that can harm well-\nbeing2. In the financial year July 2018 to June 2019, disconnection \nwas experienced by 91% of households in the remote NT communi-\nties that we have data for.\nThere are multiple levels of energy injustice in remote \nIndigenous Australia, and the effects of climate change will exac-\nerbate pre-existing energy insecurity and subsequent effects on \nhealth and wellbeing. In considering how to address these issues, it \nneeds to be recognized that there is a disproportionate prevalence \nof prepayment metering in remote Indigenous communities com-\npared to the rest of Australia46. There are questions about whether \nprepayment is a good option for remote communities that already \nface compounding distributional injustices. While some studies \nfind that prepayment may be preferred to the accrual of unsustain-\nable debts50,51, this is only a weak endorsement of prepayment when \ncompared to worse options. The framing of household electricity \npayment for communities needs to be extended beyond individual \nfiscal responsibility to incorporate a broader economic lens that \naccounts for the effects of frequent disconnection from the services \nthat energy provides on Indigenous wellbeing.\nIn considering such distributional injustices, procedural injus-\ntices first need to be addressed by supporting participatory com-\nmunity engagement in energy policy development (for example, \nincrease local access to data/information and community consulta-\ntions). Indigenous Australians have distinct societal values and per-\nspectives of wellbeing, and for progress across all spheres of inequity \nFig. 4 | Probability of a same-day disconnection by temperature, day and month. a, All regions. b, Hot persistently dry grassland climate zone. \nCoefficient estimates and 95% confidence intervals from multiple regressions for a sample with 1,674,786 daily observations across 3,300 houses are \npresented above. Regressions were grouped by percentile of electricity use (that is, average daily load) and by climate zone. Estimates for all houses, \nlow-electricity-use households and high-electricity-use households are shown here (Supplementary Tables 7\u201311 have all same-day disconnection \nestimates). Temperature bins are specified using daily average temperatures (in \u00b0C). Temperature-based estimates are for a three-day period (that is, \ntemperature on the day of disconnection and the two days before). Temperature estimates are interpreted using 20\u201325\u2009\u00b0C as the reference temperature \nrange. Day of the week estimates are interpreted using Wednesday as the reference day. Monthly estimates are interpreted using April as the reference \nmonth.\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n49\n\nArticles\nNATure Energy\nand injustice faced by Indigenous communities there needs to be a \nrecognition of these world views in the context of the realities that \nthese communities face59.\nThe Partnership Agreement on Closing the Gap calls for the \ngreater sharing of, and access to, data and information to support \nshared decision-making (Fig. 2)60. With respect to disconnections, \na\nb\n0\u201310\n10\u201315\n15\u201320\n20\u201325\n25\u201330\n30\u201335\n35\u201340\n0.6\n0.6\n0.6\n0.6\n0.4\n0.4\n0.4\n0.4\n0.4\n0.4\n0.2\n0.2\n0.2\n0.1\n0.3\n0.2\n0.2\n0.2\n0.3\n0.1\n0\n0\n0\n0\n0\n0\nProbability\nProbability\nProbability\nProbability\nProbability\nProbability\nDaily average temperature (\u00b0C)\nDay of the week (including public holidays)\nSun\nSat\nFri\nThu\nWed\nTue\nMon\nDay after public holiday\nPublic holiday\nJune\nJuly\nAugust\nSeptember\nOctober\nNovember\nMonth\nDecember\nJanuary\nFebruary\nMarch\nApril\nMay\nSelected regression estimates by level of average daily load (ADL) (with 95% CI in grey)\nAll houses\nHigh ADL (90\u2013100%)\nLow ADL (0\u201310%)\n0\u201310\n10\u201315\n15\u201320\n20\u201325\n25\u201330\n30\u201335\n35\u201340\nDaily average temperature (\u00b0C)\nDay of the week (including public holidays)\nSun\nSat\nFri\nThu\nWed\nTue\nMon\nDay after public holiday\nPublic holiday\nJune\nJuly\nAugust\nSeptember\nOctober\nNovember\nMonth\nDecember\nJanuary\nFebruary\nMarch\nApril\nMay\nSelected regression estimates by level of average daily load (ADL) (with 95% CI in grey)\nAll houses\nLow ADL (0\u201310%)\nHigh\nAll\nLow\nHigh\nAll\nLow\nHigh\nAll\nLow\nHigh\nAll\nLow\nHigh ADL (90\u2013100%)\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n50\n\nArticles\nNATure Energy\na key aim will be reducing the frequency, duration and effects of \ndisconnection. This might include: improving the accessibility and \naffordability of energy through changes to tariffs or direct access to \nthe benefits of renewable energy such as residential rooftop solar \non community housing; improving energy efficiency of infrastruc-\nture, buildings and appliances; and improving energy provision for \nparticular critical needs, for example, disconnection prohibitions or \ntariff reductions during temperature extremes, protections for criti-\ncal care customers and the use of protected circuits for refrigeration, \nlighting and essential medical equipment.\nIn addition to better policy, the language around disconnection \nevents needs to recognize and reflect community experience. The \nterm \u2018self-disconnection\u2019, while in common use, is a misrepresenta-\ntion as it incorrectly implies that households were making a vol-\nuntary choice to disconnect themselves21,50,61. The term \u2018involuntary \nself-disconnection\u2019 emphasizes \u2018that the household has not chosen \nto cease their electricity supply\u2019 62.\nLimitations\nOn considering these results there are limitations that need to \nbe considered. First, we were unable to identify the reasons for \nmulti-day disconnections. Once a disconnection occurs, whether it \nbecomes a multi-day disconnection will depend on a number of fac-\ntors including whether residents can \u2018top up\u2019 credit, pay off \u2018friendly \ncredit\u2019 debt, or make other arrangements. Inter- and intraregional \nmobility in remote and rural Australia is common (refs. 63\u201365 and \nE. Ings, personal communication), which could influence the onset \nand length of multi-day disconnections. Residents may choose to \ntemporarily relocate to the residence of kin for socio-cultural rea-\nsons or because of a disconnection, and disconnection-associated \nrelocations may themselves exacerbate overcrowding and increase \nelectricity demand on households that are not yet disconnected. \nFuture studies should focus on the experiences of energy poverty in \nthese communities and further investigate these issues.\nSecond, we did not have information on the socio-economic and \ndemographic composition of households. To address this, we used \nstatistical approaches to control for differences across households \nand estimated the effect of temperature on the basis of the usual \nlevel of electricity use and disconnections (during moderate tem-\nperatures between 20\u00b0C and 25\u2009\u00b0C). As a result, when discussing \nthe estimates, we provide the likelihood of disconnections during \nmoderate temperatures. Electricity consumption data were used to \nestimate relationships for different groups of households. There will \nbe additional determinants for the differences in electricity use and \ndisconnection, which should be investigated further.\nThird, this study uses data from 3,300 smart prepayment meters \nand finds that over 170,000 disconnections occurred across 28 com-\nmunities over a period spanning 18\u2009months. Note that our dataset \nis unbalanced due to the timing of the roll out of smart meters in \nthe NT. Thus, we underestimate the total number of disconnections \nin the NT as there are many more households currently using pre-\npayment metering than just those represented in this study53. The \nvulnerability of prepayment customers is often overlooked by gov-\nernment reporting. Further research on prepayment and discon-\nnection in other jurisdictions is needed, as is greater understanding \nof the direct health effects of these disconnections.\nConclusions\nAustralia could do much better at providing protections from dis-\nconnection. Policymakers are beginning to consider the importance \nof electricity to wellbeing in approaches that seek to limit the fre-\nquency and duration of disconnection events, particularly in relation \nto temperature extremes and wellbeing. In the USA, for example, \ndespite not having a formalized definition of energy poverty or fed-\neral level protections, many states have utility regulation policies that \nprotect customers from disconnection of service in certain cases, \nincluding extreme temperatures4,26. Some state consumer protec-\ntions target vulnerable groups, such as in the state of Texas where \nprepayment-meter enrolment is prohibted for those diagnosed with \nsevere medical conditions that require electricity services to main-\ntain temperatures or run devices26,66. Many European Union states \nhave also introduced protection from disconnection7, many with \nparticular focus on extreme temperatures and vulnerable groups.\nIn Australia, the Essential Service Commission (for Victoria) \nobserves that \u201ccustomers who are disconnected from electricity or \ngas can face significant risks to their welfare\u2026 disconnection for \nTable 2 | Electricity use and disconnections by climate zones\nEquatorial climate zone \n(most northern)\nCoastal tropical \nclimate zone\nSavannah tropical \nclimate zone\nHot persistently dry  \ngrassland climate zone  \n(most southern)\nAll regions\nPercentage of disconnections above maximum temperature of 35\u2009\u00b0C\nPercentage of same-day disconnections\n7\n12\n34\n42\n24\nPercentage of multi-day disconnections\n6\n11\n36\n40\n28\nPercentage of all disconnections\n7\n12\n34\n42\n25\nPercentage of disconnections above maximum temperature of 40\u2009\u00b0C\nPercentage of same-day disconnections\n0\n0\n4\n18\n5\nPercentage of multi-day disconnections\n0\n0\n3\n18\n7\nPercentage of all disconnections\n0\n0\n4\n18\n5\nPercentage of disconnections below minimum temperature of 0\u2009\u00b0C\nPercentage of same-day disconnections\n0\n0\n5\n12\n4\nPercentage of multi-day disconnections\n0\n0\n6\n11\n6\nPercentage of all disconnections\n0\n0\n5\n12\n4\nNumber of disconnections\nNumber of same-day disconnections\n32,133\n39,212\n54,309\n30,281\n155,935\nNumber of multi-day disconnections\n1,971\n2,787\n4,841\n4,692\n14,291\nNumber of all disconnections\n34,104\n41,999\n59,150\n34,973\n170,226\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n51\n\nArticles\nNATure Energy\nnon-payment reasons should only ever be a last resort\u201d67. Australia\u2019s \nNational Energy Retail Rules require that the retailer not arrange for \nthe de-energization of premises having life support equipment or \nduring an extreme weather event68, but this is not comprehensively \napplied in remote NT communities47,48.\nEnergy insecurity in remote Indigenous Australia remains a press-\ning issue. Access to energy has been identified as a key part of the \u2018criti-\ncal healthy living priorities\u2019 for remote living Indigenous Australians3. \nEnsuring access to essential services is an important prerequisite to \nimproving health and wellbeing outcomes69. The focus should not be \nsolely on electricity provision itself, but on maximizing the benefit \nthat households receive from their electricity use and this includes a \nrange of essential services, including heating and cooling.\nMethods\nEthics statement. This research was conducted with ethics approval from the \nCentral Australian Health Research Ethics Committee Centre for Remote Health \n(CA-20-3809).\nAutoethnographic quotes. While this is predominantly a quantitative study, our \nresearch methodology is based on principles of social justice where Aboriginal \npeople participate in setting the research agenda70 and we supplement our analysis \nwith insights through autoethnography71. These quotes are included by authors, \nN.F.J. who is a community elder and V.N.D. who is a senior Aboriginal researcher. \nThis paper has more than a single authorial voice, coming as it does from multiple \nperspectives. These perspectives are from researchers who work at an academic \ninstitution (T.L., B.R. and L.V.W.), a regional hospital (S.Q.) and community-based \norganizations (S.Q., M.K. and V.N.D.).\nDataset. The data used in this paper were sourced from the NT Government \nowned utility Power and Water Corporation, the Australian Bureau of Meteorology \n(BOM) and the Australian Bureau of Statistics. Daily electricity use data for 3,300 \nhouseholds with a smart prepayment meter were matched to temperature data \nfrom the closest weather station. For cases where there were no temperature data \nfor that day, the next closest weather station was used (6.1% of all observations). If \nthat still resulted in a missing value, then the average for that climate zone was used \n(0.3% of observations). Data on disconnections were provided along with the time \nand date that the electricity service was discontinued and subsequently restored. \nThese cases of disconnection were aggregated into daily data and separated into \ntwo variables on the basis of whether an electricity service was restored to the \nhousehold on the same day or not. Selected summary statistics for these data by \nclimate zones are shown in the paper in Tables 1 and 2.\nThe climate zones we used were sourced from the Australian BOM31. We \nmade some modifications to the zones mapped by BOM, which was to reclassify \nall of the mainland communities that were within 20\u2009km of the coast as a \u2018coastal \ntropical climate zone\u2019 and combine the Central Australian climate zones into one \nregion that we called the \u2018hot persistently dry grassland climate zone\u2019. The second \nmodification was due to sample size and a similarity in temperatures. The other \nclimate zones are those prescribed by BOM, shown in Fig. 1c.\nExtreme temperatures in remote communities. Those communities that we focus \non are typically exposed to extremely hot days and cold nights. Supplementary \nTable 13 shows the breakdown of key temperature statistics by climate zone for \nthese communities. Figure 1 presents maps produced by the Australian BOM \nshowing how temperature differs across both Australia and the NT72. Differences \nin temperature indicators vary across climate zones (shown in Fig. 1a\u2013c). Central \nAustralia experiences prolonged hot daytime temperatures in summer and cold \n(below zero) nights in winter. For example, the hottest day (46\u2009\u00b0C) and coldest \nnight (\u22124\u2009\u00b0C) in our dataset both occurred in the southern-most \u2018hot persistently \ndry grassland climate zone\u2019. Northern regions of the NT experience the southern \nextent of the tropical monsoon, which brings seasonal cyclonic activity and \nafternoon storms. Average temperatures decrease north to south. The highest \nmaximum temperature (lowest minimum temperature) increases (decreases) as \nyou move from the north to south (Fig. 1b,c and Supplementary Table 13). The \nregressions were run using daily average temperatures.\nGrouping by average daily electricity use. Energy insecurity and disconnection \nrates are likely to be determined by a range of factors, including the usual level \nof electricity use and the inability to pay. While we do not have household data \non occupancy or income, we are able to categorize the meters into groups using \naverage daily electricity use. The average daily load will be a function of the types \nof appliance, intensity of use and the number of residents. The percentile groupings \nused are shown in Supplementary Table 14 along with the aggregate expenditure \non electricity. Note that the data are not a balanced panel due to the incremental \nroll out of smart meters across communities and we excluded those meters with \nless than 100 observations.\nRegression analysis. To estimate the relationship between electricity use and \ntemperature we used linear regression with panel-corrected standard errors that \naccounted for heteroscedasticity and autocorrelation. The 28 communities were \nthe level for the panel correction. The software that was used to estimate these \nrelationships was Stata MP 16.1. The \u2018xtpcse\u2019 command was used to estimate the \nresults shown in Fig. 3 and Supplementary Tables 2\u20136. We also tested for normality \nin linear panel-data models using the \u2018xtsktest\u2019 command, which did not reject \nthe null hypothesis of normality (Chi-square statistic of 3.28\u20133.42 (P value of \n0.18\u20130.19)). The \u2018xtserial\u2019 command was used to perform the Wooldridge test for \nautocorrelation in panel data, which rejected the null hypothesis of no first-order \nautocorrelation (F-statistic of 6,694.17 (P of 0.00)). We used random-effects probit \nregressions to estimate the relationship between the probability of same-day/\nmulti-day disconnection and temperature. These estimates were estimated \nusing the \u2018xtprobit\u2019 command in Stata and we clustered the standard errors \nby community to control for regional differences. The results of the same-day \ndisconnection estimations are shown in Fig. 4 and Supplementary Tables 7\u201311. \nWhen discussing these estimates in the paper we compare them to the likelihood \nof disconnection during moderate temperatures. The probabilies of disconnection \nevents were converted into an odds ratio (for example, 1:2) and then reported as \nthe chance of a disconnection (for example, one in three). Multi-day disconnection \nestimates are included in Supplementary Table 12, but are not discussed in the \npaper as there was no clear relationship between multi-day disconnections and \ndaily average temperature. Note that the regression estimates were graphed using \nthe \u2018ggplot\u2019 command in R.\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nThe data used in this paper are not freely available and were sourced from the data \ncustodians. The electricity data were sourced from Power and Water Corporation \n(https://www.powerwater.com.au/) and the Australian BOM (http://www.bom.gov.\nau/). We note that access to data will be a key part of local communities helping \nto develop appropriate policy responses to the challenges outlined in this paper. \nThe Partnership Agreement on Closing the Gap calls for the greater sharing of, \nand access to, data and information at a regional level noting that \u201cdisaggregated \ndata and information is most useful to Aboriginal and Torres Strait Islander \norganizations and communities to obtain a comprehensive picture of what is \nhappening in their communities and to support decision-making\u201d60.\nCode availability\nThe code used to estimate the regressions (in Stata MP 16.1) and create the \ngraphics (in R v.4.1.1) is available on request. The \u2018xtpcse\u2019 and \u2018xtprobit\u2019 commands \nin Stata MP 16.1 were used for the regressions. The regression estimates were \ngraphed using the \u2018ggplot\u2019 command in R. Statistical tests for normality and \nautocorrelation were performed in Stata MP 16.1 using the \u2018xtsktest\u2019 and \u2018xtserial\u2019.\nReceived: 16 April 2021; Accepted: 20 October 2021;  \nPublished online: 16 December 2021\nReferences\n\t1.\t Bouzarovski, S. & Petrova, S. A global perspective on domestic energy \ndeprivation: overcoming the energy poverty-fuel poverty binary. Energy Res. \nSoc. Sci. 10, 31\u201340 (2015).\n\t2.\t Day, R., Walker, G. & Simcock, N. Conceptualising energy use and energy \npoverty using a capabilities framework. Energy Policy 93, 255\u2013264 (2016).\n\t3.\t Standen, J. C. et al. Prioritising housing maintenance to improve health in \nindigenous communities in NSW over 20 years. Int. J. Environ. Res. Public \nHealth 17, 5946 (2020).\n\t4.\t Bednar, D. J. & Reames, T. G. Recognition of and response to energy poverty \nin the United States. Nat. Energy 5, 432\u2013439 (2020).\n\t5.\t Hern\u00e1ndez, D. Understanding \u2018energy insecurity\u2019 and why it matters to \nhealth. Soc. Sci. Med. 167, 1\u201310 (2016).\n\t6.\t Walker, G. & Day, R. Fuel poverty as injustice: Integrating distribution, \nrecognition and procedure in the struggle for affordable warmth. Energy \nPolicy 49, 69\u201375 (2012).\n\t7.\t Dobbins, A., Fuso Nerini, F., Deane, P. & Pye, S. Strengthening the EU \nresponse to energy poverty. Nat. Energy 4, 2\u20135 (2019).\n\t8.\t Carley, S. & Konisky, D. M. The justice and equity implications of the clean \nenergy transition. Nat. Energy 5, 569\u2013577 (2020).\n\t9.\t Sovacool, B. K. & Dworkin, M. H. Energy justice: conceptual insights and \npractical applications. Appl. Energy 142, 435\u2013444 (2015).\n\t10.\tMemmott, T., Carley, S., Graff, M. & Konisky, D. M. Sociodemographic \ndisparities in energy insecurity among low-income households before and \nduring the COVID-19 pandemic. Nat. Energy https://doi.org/10.1038/\ns41560-020-00763-9 (2021).\n\t11.\tAnderson, W., White, V. & Finney, A. Coping with low incomes and cold \nhomes. Energy Policy 49, 40\u201352 (2012).\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n52\n\nArticles\nNATure Energy\n\t12.\tBouzarovski, S., Petrova, S. & Sarlamanov, R. Energy poverty policies in the \nEU: a critical perspective. Energy Policy 49, 76\u201382 (2012).\n\t13.\tSnell, C., Bevan, M. & Thomson, H. Justice, fuel poverty and disabled people \nin England. Energy Res. Soc. Sci. 10, 123\u2013132 (2015).\n\t14.\tFrank, D. A. et al. Heat or eat: the low income home energy assistance \nprogram and nutritional and health risks among children less than 3 years of \nage. Pediatrics 118, e1293\u2013e1302 (2006).\n\t15.\tCook, J. T. et al. A brief indicator of household energy security: associations \nwith food security, child health, and child development in US infants and \ntoddlers. Pediatrics 122, e867-75 (2008).\n\t16.\tNord, M. & Kantor, L. S. Seasonal variation in food insecurity is associated \nwith heating and cooling costs among low-income elderly Americans. J. Nutr. \n136, 2939\u20132944 (2006).\n\t17.\tReames, T. G. Targeting energy justice: exploring spatial, racial/ethnic and \nsocioeconomic disparities in urban residential heating energy efficiency. \nEnergy Policy 97, 549\u2013558 (2016).\n\t18.\tAlola, A. A. et al. Cooling and heating degree days in the US: the role of \nmacroeconomic variables and its impact on environmental sustainability. Sci. \nTotal Environ. 695, 133832 (2019).\n\t19.\tShin, M. & Do, S. L. Prediction of cooling energy use in buildings using an \nenthalpy-based cooling degree days method in a hot and humid climate. \nEnergy Build. 110, 57\u201370 (2016).\n\t20.\tDe Rosa, M., Bianco, V., Scarpa, F. & Tagliafico, L. A. Heating and cooling \nbuilding energy demand evaluation; a simplified model and a modified \ndegree days approach. Appl. Energy 128, 217\u2013229 (2014).\n\t21.\tO\u2019Sullivan, K. C., Howden-Chapman, P. L. & Fougere, G. Making the \nconnection: the relationship between fuel poverty, electricity disconnection, \nand prepayment metering. Energy Policy 39, 733\u2013741 (2011).\n\t22.\tHealy, J. D. Excess winter mortality in Europe: a cross country analysis \nidentifying key risk factors. J. Epidemiol. Commun. Health 57, 784\u2013789 (2003).\n\t23.\tRobinson, C., Lindley, S. & Bouzarovski, S. The spatially varying components of \nvulnerability to energy poverty. Ann. Am. Assoc. Geogr. 109, 1188\u20131207 (2019).\n\t24.\tO\u2019Sullivan, K. C., Stanley, J., Fougere, G. & Howden-Chapman, P. Heating \npractices and self-disconnection among electricity prepayment meter \nconsumers in New Zealand: a follow-up survey. Util. Policy 41, 139\u2013147 (2016).\n\t25.\tJessel, S., Sawyer, S. & Hern\u00e1ndez, D. Energy, poverty, and health in climate \nchange: a comprehensive review of an emerging literature. Front. Public \nHealth https://doi.org/10.3389/fpubh.2019.00357 (2019).\n\t26.\tFlaherty, M., Carley, S. & Konisky, D. M. Electric utility disconnection policy \nand vulnerable populations. Electr. J. 33, 106859 (2020).\n\t27.\tDobbins, A., Pye, S., De Miglio, R. & Brajkovic, J. Measures to Protect \nVulnerable Consumers in the Energy Sector: An Assessment of Disconnection \nSafeguards, Social Tariffs and Financial Transfers Vol. 8 (Insight-E, 2016).\n\t28.\tRegulation\u2014unregulated networks. Power and Water Corporation https://\nwww.powerwater.com.au/about/regulation (2021).\n\t29.\tClimate Change in the Northern Territory \u2013 State of the Science and Climate \nChange Impacts (NESP ESCC Hub, 2020).\n\t30.\tKovats, R. S. & Hajat, S. Heat stress and public health: a critical review. Annu. \nRev. Public Health 29, 41\u201355 (2008).\n\t31.\tLongden, T. The impact of temperature on mortality across different climate \nzones. Climatic Change 157, 221\u2013242 (2019).\n\t32.\tPatz, J. A., Campbell-Lendrum, D., Holloway, T. & Foley, J. A. Impact of \nregional climate change on human health. Nature 438, 310\u2013317 (2005).\n\t33.\tMora, C. et al. Global risk of deadly heat. Nat. Clim. Change 7, 501\u2013506 (2017).\n\t34.\tNussbaum, M. C. & Sen, A. The Quality of Life (Oxford Univ. Press, 2004).\n\t35.\tSen, A. Inequality Reexamined (Oxford Univ. Press, 1992).\n\t36.\tNussbaum, M. C. Creating Capabilities: The Human Development Approach \n(Harvard Univ. Press, 2011).\n\t37.\tZahurul, S. et al. Future strategic plan analysis for integrating distributed \nrenewable generation to smart grid through wireless sensor network: \nMalaysia prospect. Renew. Sustain. Energy Rev. 53, 978\u2013992 (2016).\n\t38.\tFraser, N. Justice Interruptus: Critical Reflections on the \u2018Postsocialist\u2019 \nCondition (Routledge, 1997).\n\t39.\tSchlosberg, D. Defining Environmental Justice: Theories, Movements and \nNature (Oxford Univ. Press, 2007).\n\t40.\tCouncil of Australian Governments National Indigenous Reform Agreement \n(Closing the Gap) (CAG, 2009).\n\t41.\tDepartment of Families Community Services and Indigenous Affairs National \nIndigenous Housing Guide: Improving the Living Environment for Safety, \nHealth and Sustainability 3rd edn (DSS, Australian Government, 2012).\n\t42.\tNational Energy Retail Law (South Australia) Act 2011 (Government of South \nAustralia, 2011); https://www.legislation.sa.gov.au/LZ/C/A/NATIONAL%20\nENERGY%20RETAIL%20LAW%20(SOUTH%20AUSTRALIA)%20ACT%20\n2011/CURRENT/2011.6.AUTH.PDF\n\t43.\tAustralian Energy Market Commission National Energy Retail Rules Version \n27 (AEMC, 2021).\n\t44.\tPrepayment Meter System Code: Draft Final Decision (Essential Services \nCommission of South Australia, 2005); https://www.escosa.sa.gov.au/ArticleD\nocuments/805/050308-PrePaymentMeter_CombDraftFinalDec_Code.pdf.\naspx?Embed=Y\n\t45.\tPrepayment Meters: Discussion Paper (Energy and Water Ombudsman NSW, \n2014).\n\t46.\tMcKenzie, M. Pre-Payment Meters and Energy Efficiency in Indigenous \nHouseholds (Centre for Appropriate Technology, 2013); https://cfat.org.au/\npre-payment-meters-and-energy-efficiency-in-indigenous-households\n\t47.\tElectricity Retail Supply Code Review Issues Paper (Utilities Commission of \nthe Northern Territory, 2021); https://utilicom.nt.gov.au/__data/assets/pdf_\nfile/0019/1011655/Issues-Paper-Electricity-Retail-Supply-Code-Review.pdf\n\t48.\tThomson, M. Review of the Retail Supply Code\u2014further information. \nUtilities Commission of the Northern Territory https://utilicom.nt.gov.au/__\ndata/assets/pdf_file/0010/747568/\nPower-and-Water-Corporation-Submission-additional.pdf (2019).\n\t49.\tVyas, D. Topping-up or Dropping-Out: Self-Disconnection Among Prepayment \nMeter Users (Citizens Advice Bureau, 2014); https://www.citizensadvice.org.\nuk/global/migrated_documents/corporate/topping-up-or-dropping-out.pdf\n\t50.\tO\u2019Sullivan, K. C., Howden-Chapman, P. L., Fougere, G. M., Hales, S. & \nStanley, J. Empowered? Examining self-disconnection in a postal survey of \nelectricity prepayment meter consumers in New Zealand. Energy Policy 52, \n277\u2013287 (2013).\n\t51.\tWagner, O. & Wiegand, J. Prepayment metering: household experiences in \nGermany. Renew. Sustain. Energy Rev. 98, 407\u2013414 (2018).\n\t52.\tBrutscher, P.-B. Self-Disconnection Among Pre-Payment Customers\u2014A \nBehavioural Analysis (Univ. Cambridge, 2012); https://aspace.repository.cam.\nac.uk/handle/1810/257126\n\t53.\tSubmission 165.2 to the Inquiry into Homelessness in Australia (Tangentyere \nCouncil, 2021); https://www.aph.gov.au/Parliamentary_Business/Committees/\nHouse/Social_Policy_and_Legal_Affairs/HomelessnessinAustralia/Submissions\n\t54.\tHouseholds in the Dark II: Mapping Electricity Disconnections in South \nAustralia, Victoria, New South Wales and South East Queensland (St Vincent \nde Paul Society & Alviss Consulting, 2019); https://www.vinnies.org.au/icms_\ndocs/310289_Households_in_the_Dark_II_2019.pdf\n\t55.\tRetail Performance Data Snapshot 2018\u221219 (AER, 2019); https://www.aer.gov.\nau/system/files/AER%20Payment%20difficulties%20and%20hardship%20\ndata%20by%20jurisdiction%202018-19.pdf\n\t56.\tEmpowering Remote Communities (Queensland Council of Social Service, 2014).\n\t57.\tPrepayment Power Meters (Power and Water Corporation, 2021); https://www.\npowerwater.com.au/customers/power/power-meters/prepayment-power-meter\ns#Remote\n\t58.\tWalker, G., Simcock, N. & Day, R. Necessary energy uses and a minimum \nstandard of living in the United Kingdom: energy justice or escalating \nexpectations? Energy Res. Soc. Sci. 18, 129\u2013138 (2016).\n\t59.\tKingsley, J., Townsend, M., Henderson-Wilson, C. & Bolam, B. Developing an \nexploratory framework linking australian aboriginal peoples\u2019 connection to \ncountry and concepts of wellbeing. Int. J. Environ. Res. Public Health 10, \n678\u2013698 (2013).\n\t60.\tPriority Reforms. Closing the Gap (Australian Government, 2021); https://\nwww.closingthegap.gov.au/priority-reforms\n\t61.\tRocha, M., Baddeley, M., Pollitt, M. & Weeks, M. Addressing \nself-disconnection among prepayment energy consumers: a behavioural \napproach. Energy Econ. 81, 273\u2013286 (2019).\n\t62.\tNTCOSS Cost of Living Report\u2014Issue 26 (NTCOSS, 2019).\n\t63.\tMemmott, P. et al. NATSISS Crowding Data: What Does it Assume and How \nCan We Challenge the Orthodoxy? (Australian National Univ., 2019); https://\npress-files.anu.edu.au/downloads/press/p206931/html/ch12.\nhtml?referer=&page=19\n\t64.\tMemmott, P., Long, S. & Thomson, L. Mobility of Aboriginal People in Rural \nand Remote Australia Research and Policy Bulletin (AHURI, 2006); https://\nwww.ahuri.edu.au/research/research-and-policy-bulletins/69\n\t65.\tMemmott, P., Long, S., Bell, M., Taylor, J. & Brown, D. Between Places: \nIndigenous Mobility in Remote and Rural Australia (The Australian Housing \nand Urban Research Institute, 2004).\n\t66.\tPublic Utility Commission (PUC) Rules Ch. 25 (Public Utility Commission of \nTexas, 2021); https://www.puc.texas.gov/agency/rulesnlaws/subrules/electric/\nch25complete.pdf?v=20200511\n\t67.\tVictorian Energy Market Report 2019\u201320 (Essential Services Commission, \n2020).\n\t68.\tNational Energy Retail Rules\u2014Current (AEMC, 2021).\n\t69.\tHunt, J., Riley, B., O\u2019Neill, L. & Maynard, G. Transition to renewable energy \nand indigenous people in Northern Australia: enhancing or inhibiting \ncapabilities? J. Hum. Dev. Capab. https://doi.org/10.1080/19452829.2021.1901\n670 (2021).\n\t70.\tCoghlan, D. & Brydon-Miller, M. The SAGE Encyclopedia of Action Research \n(Sage, 2014).\n\t71.\tAdams, T. E., Ellis, C. & Jones, S. H. in The International Encyclopedia of \nCommunication Research Methods (ed. Matthes, J.) https://doi.\norg/10.1002/9781118901731.iecrm0011 (2017).\n\t72.\tClimate maps: temperature archive\u2014twelve-monthly highest maximum \ntemperature for Australia. Australian Government Bureau of Meteorology \nhttp://www.bom.gov.au/jsp/awap/temp/archive.jsp?colour=colour&map=maxe\nxtrm%2Fhi&year=2019&month=6&period=12month&area=nat (2019).\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n53\n\nArticles\nNATure Energy\n\t73.\tAustralian climate zones. Australian Government Bureau of Meteorology \nhttp://www.bom.gov.au/jsp/ncc/climate_averages/climate-classifications/index.\njsp?maptype=kpn#maps (2005).\n\t74.\tClimate maps: temperature archive\u2014twelve-monthly lowest minimum \ntemperature for Australia. Australian Government Bureau of  \nMeteorology http://www.bom.gov.au/jsp/awap/temp/archive.jsp?colour=colour\n&map=minextrm%2Flow&year=2019&month=6&period=12month&area=n\nat (2019).\n\t75.\tClimate maps: temperature archive\u2014maximum temperature anomaly. \nAustralian Government Bureau of Meteorology http://www.bom.gov.au/jsp/\nawap/temp/archive.jsp?colour=colour&map=maxanom&year=2019&month=\n6&period=12month&area=nat (2021).\nAcknowledgements\nWe acknowledge and thank the Power and Water Corporation for providing  \nthe electricity data. We acknowledge the Australian BOM and thank them for  \nproviding temperature data and maps. A range of people provided useful advice that \nhelped to shape this paper and they include E. Ings and J. Hulcombe. We thank the  \nBoard and staff of Tangentyere Council Aboriginal Corporation in Mparntwe  \n(Alice Springs) and the Board and staff of Julalikari Council Aboriginal Corporation \nin Tennant Creek. Our research methodology was informed by the principles \nunderpinning ethical Australian Indigenous research outlined in the AIATSIS \nCode of Ethics for Aboriginal and Torres Strait Islander Research (AIATSIS 2020): \nIndigenous self-determination, Indigenous leadership, impact and value, sustainability \nand accountability. We acknowledge and thank colleagues at the Australian National \nUniversity. T.L., B.R. and L.V.W. thank their colleagues from the ANU Grand Challenge \nZero-Carbon Energy for the Asia-Pacific and the Institute for Climate, Energy and \nDisaster Solutions.\nAuthor contributions\nAll authors contributed to the conceptualization of the research. We especially note the \ncontributions of N.F.J. and V.N.D. in shaping our understanding of the key issues faced \nby Indigenous communities in the NT. S.Q. acquired the key data, and T.L performed \nthe analysis. T.L., S.Q., B.R., L.V.W. and M.K. wrote the initial draft of the manuscript, \nand all authors contributed to the review and revision. N.F.J. and V.N.D. were engaged \nin discussions on key issues with the other members of the authorship team and their \nselected quotes are provided to highlight those themes and the issues that they raised as \nbeing the most important (refer to autoethnographic data section for more information).\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data are available for this paper at https://doi.org/10.1038/\ns41560-021-00942-2.\nSupplementary information The online version contains supplementary material \navailable at https://doi.org/10.1038/s41560-021-00942-2.\nCorrespondence and requests for materials should be addressed to Michael Klerck.\nPeer review information Nature Energy thanks Kimberley O\u2019Sullivan, Sangeetha \nChandrashekeran and Stefan Bouzarovski for their contribution to the peer review of \nthis work\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2021\nNature Energy | VOL 7 | January 2022 | 43\u201354 | www.nature.com/natureenergy\n54\n\n\n    Scientific Research Findings:", "answer": "Among\u00a028 remote communities in the Northern Territory, we found that\u00a091% of households experienced a disconnection event at least once during the\u00a02018/19 financial year;\u00a074% of households were disconnected over\u00a010\u00a0times, and\u00a029% of all disconnections occurred during extreme temperatures. In mild temperatures (20\u201325\u00a0\u00b0C), households had a\u00a01\u00a0in\u00a017 chance of disconnection on a given day. This increased to a\u00a01\u00a0in\u00a011 chance during hot days (34\u201340\u00a0\u00b0C) and a\u00a01\u00a0in\u00a06 chance during cold days (0\u201310\u00a0\u00b0C). Households with high electricity use in the central Australian climate zones had a\u00a01\u00a0in\u00a03 chance of a same\u2011day disconnection during temperature extremes. Energy insecurity is worsened when energy use is heightened owing to heating or cooling needs. Our analysis does not explore all of the complexities underlying energy insecurity in these communities, but we expect that these findings will inform discussions of energy insecurity in regions with extreme temperatures.", "id": 16}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Analysis\nhttps://doi.org/10.1038/s41560-021-00911-9\n1Energy, Climate, and Environment Program, International Institute of Applied Systems Analysis, Laxenburg, Austria. 2Population and Just Societies \nProgram, International Institute of Applied Systems Analysis, Laxenburg, Austria. 3Climate Analytics, Berlin, Germany. \u2709e-mail: pachauri@iiasa.ac.at\nE\nnergy for cooking is a most fundamental need. Yet today, over \nthree billion people still cook by burning wood on open fires \nand in smoky stoves. The enormous social, public health and \nenvironmental benefits of transitioning to cleaner cooking underpin \nthe inclusion of a universal access target for this under the United \nNations Sustainable Development Goal 7 (SDG\u20097)1\u20133. Even before \nthe COVID-19 pandemic, data showed that efforts to provide clean \nfuels and stoves have been lagging far behind those aimed at extend-\ning electricity access4. A recent report claims that this sluggish prog-\nress in providing clean cooking access is costing the world more \nthan US$2 trillion each year as a result of health impacts, produc-\ntivity losses and environmental degradation5. Mounting evidence \nof the impacts of air pollution exposure on COVID-19 mortality \nmakes universal access to clean cooking services more urgent6,7. \nDespite this, emerging evidence suggests that the economic fallout \nof the pandemic might pose a further setback to efforts to reach this \ngoal, as many are forced to climb down the energy ladder8\u201310.\nPrevious literature analysing scenarios for achieving univer-\nsal access to modern energy services have focused predominantly \non electricity supply to assess the cost-effectiveness of alternative \noptions to provide connections11\u201314. There is a paucity of studies \nanalysing clean cooking scenarios, particularly at a global scale15,16. \nExisting studies that focus on cooking access scenarios are limited \nin their representation of multiple cooking fuel use (fuel stacking), \npopulation heterogeneity and affordability constraints, which are \ncritical to understanding whether people will regularly use new \nfuels or stoves after they acquire them17\u201319. The limited existing evi-\ndence, preceding the pandemic, suggests that the world is far off the \nmark of the SDG\u20097 goal, with nations in sub-Saharan Africa pro-\njected to not achieve this target even in 20505,20.\nIn this study we explored clean cooking access until 2050 under \nreference scenarios of socioeconomic and demographic change, \nambitious climate mitigation policy scenarios and a slow economic \nrecovery from the COVID-19 pandemic scenario (see Methods \nfor scenario details). We applied existing microdata-based cook-\ning choice and demand models that explicitly consider fuel stack-\ning and represent affordability constraints for urban and rural  \npopulations, capturing heterogeneity in household preferences \nacross the entire income distribution21,22. We find that a slow recov-\nery from the pandemic and fuel price changes because of ambitious \nclimate mitigation policy could substantially retard progress in \nachieving universal clean cooking access if additional policies related \nto energy access and poverty alleviation are not simultaneously pur-\nsued. Those most at risk of not being able to afford to transition to \nclean cooking are low-income households in sub-Saharan Africa \n(AFR), developing Asia, and Latin America and the Caribbean \n(LAM). A faster transition to clean cooking fuels can attenuate \nfuture growth in cooking energy demand, especially in regions that \ncurrently depend largely on biomass and other solid fuels.\nPopulations without access to clean cooking\nIn what follows, we define clean cooking as cooking with modern \nfuels such as liquid petroleum gas (LPG), electricity and piped gas, \nwhich when used in modern stoves result in little to no household \npollution. Newer options, for example, bioethanol or solar electric, \nmight become viable in the future, but were not included in our \nanalysis as we detected no use of these in the empirical datasets that \nwe employed. All other fuels, including solid biomass-based fire-\nwood or charcoal and coal, are categorized as polluting, because \nprevailing stove technologies that use these result in pollution levels \nthat exceed World Health Organization indoor air quality guide-\nlines for household fuel combustion23. We categorize a household \nas cooking poor if it depends on polluting cooking fuels, that is, on \nfuels other than those we define as clean, for half or more of its cook-\ning energy consumption. In our analysis we compared an especially \nconstructed COVID-19 recovery scenario (COVID) with three ref-\nerence scenarios, namely sustainability (SSP1), middle-of-the-road \n(SSP2) and regional rivalry (SSP3), from the Shared Socioeconomic \nPathways (SSP) framework24. We further compared our recov-\nery and reference scenarios with a set of scenarios that impose an \nambitious climate mitigation policy that limits warming to 2\u2009\u00b0C by \nthe end of the century (CP2C) using a previous formulation that \nassumes a regionally differentiated carbon price trajectory that rises \ngradually over time25 (see Methods).\nAccess to clean cooking services in energy and \nemission scenarios after COVID-19\nShonali Pachauri\u200a \u200a1\u2009\u2709, Miguel Poblete-Cazenave\u200a \u200a1, Arda Aktas2 and Matthew J. Gidden\u200a \u200a1,3\nSlow progress in expanding clean cooking access is hindering progress on health, gender, equity, climate and air quality goals \nglobally. Despite a rising population share with clean cooking access, the number of cooking poor remains stagnant. In this \nstudy we explored clean cooking access until 2050 under three reference scenarios, a COVID-19 recovery scenario and ambi-\ntious climate mitigation policy scenarios. Our analysis shows that universal access may not be achieved even in 2050. A pro-\ntracted recession after the pandemic could leave an additional 470 million people unable to afford clean cooking services in \n2030 relative to a reference scenario, with populations in sub-Saharan Africa and Asia the worst affected. Ambitious climate \nmitigation needs to be twinned with robust energy access policies to prevent an additional 200 million people being unable to \ntransition to clean cooking in 2030. Our findings underline the need for immediate acceleration in efforts to make clean cooking \naccessible and affordable to all.\nNature Energy | VOL 6 | November 2021 | 1067\u20131076 | www.nature.com/natureenergy\n1067\n\nAnalysis\nNAture Energy\nOur findings show that the share of the population with access to \nclean cooking rises under all scenarios until 2050, but no scenario \nmeets the rate of improvement required to achieve the SDG\u20097 2030 \ntarget (Fig. 1 and Supplementary Fig. 1). Even under our most opti-\nmistic reference growth scenario SSP1, we found that close to 38% \nof the global population could continue to remain cooking poor in \n2030. Slower growth and urbanization under the SSP2 and SSP3 \nreference scenarios could leave an additional 1.2\u20133.9% of the popu-\nlation unable to afford clean cooking in 2030. We found universal \naccess may not be achieved even in 2050.\nUnder our COVID recovery scenario, an additional 470 mil-\nlion people may remain cooking poor in 2030 as compared with \nunder SSP3, our most pessimistic reference growth scenario. The \nCOVID scenario has a persistent impact, as, even though average \nincome levels are assumed to revert to the reference SSP3 trend \nin 2040, income inequality remains higher until the middle of the \ncentury, leaving more families dependent on biomass even in 2050. \nAmbitious climate mitigation policy in the absence of additional \ntargeted support policies could also make transitioning to clean \ncooking more difficult for about 200 million people, specifically \nOther Paci\ufb01c Asia\nMiddle East and North Africa\nCentrally Planned Asia and China\nSub-Saharan Africa\nSouth Asia\nLatin America and the Caribbean\n2010\n2020\n2030\n2040\n2050\n2010\n2020\n2030\n2040\n2050\n2010\n2020\n2030\n2040\n2050\n0\n25\n50\n75\n100\n0\n25\n50\n75\n100\nYear\nYear\nYear\nCooking poor population (%)\nCooking poor population (%)\nSSP1\nSSP2\nSSP3\nCOVID\nb\na \n0\n50\nCooking poor population (%)\n100\nFig. 1 | Cooking poor populations. a, Percentage of cooking poor population in 2030 by MESSAGEix regions under the SSP2 reference scenario. b, \nPercentage of cooking poor population until 2050 by scenario in selected MESSAGEix regions. The bars depict shares in reference scenarios and the \ncrosses above the bars depict shares under the climate mitigation policy scenarios.\nNature Energy | VOL 6 | November 2021 | 1067\u20131076 | www.nature.com/natureenergy\n1068\n\nAnalysis\nNAture Energy\nthose in Centrally Planned Asia (CPA) and South Asia (SAS), where \nfossil fuel demand for cooking is currently high and most house-\nholds are unable to afford electric cooking.\nIn AFR, we observed only very moderate improvements in access \nto clean cooking services over time as population growth in this \nregion outpaces the rate of transition to better stoves and fuels in all \nscenarios. In contrast, in CPA, SAS and Other Pacific Asia (PAS), clean \ncooking access becomes increasingly affordable, especially under an \noptimistic SSP1 reference scenario. Overall, differences on account of \nincome growth, distribution and urbanization in our reference sce-\nnarios impact clean cooking access more than shifts in fuel prices as \na result of ambitious climate mitigation policy. However, in regions \nof developing and emerging Asia, climate mitigation policy could \nincrease the cost of clean cooking services. Implementing additional \nsupport policies to make clean cooking affordable will be essential to \nachieve both climate goals and SDG\u20097 simultaneously in these regions.\nSSP1\nSSP2\nSSP3\nCOVID\nREF\nCP2C\n2010\n2020\n2030\n2040\n2050 2010\n2020\n2030\n2040\n2050 2010\n2020\n2030\n2040\n2050 2010\n2020\n2030\n2040\n2050\n0\n5\n10\n15\n0\n5\n10\n15\nYear\nYear\nYear\nYear\nTotal fuel consumption (EJ)\nTotal fuel consumption (EJ)\nFirewood\nCharcoal/Coal\nKerosene\nLPG/Natural gas\nElectricity\na\nSSP3-CP2C\nSSP2-CP2C\nSSP1-CP2C\nSSP3\nSSP2\nSSP1\nCOVID\nReference year\n0\n0.5\n1.0\n1.5\n2.0\nPer-capita fuel consumption (GJ)\nb\nSSP3-CP2C\nSSP2-CP2C\nSSP1-CP2C\nSSP3\nSSP2\nSSP1\nCOVID\nReference year\n0\n1\n2\n3\nPer-capita fuel consumption (GJ)\nc \nFig. 2 | Total and average cooking energy demand. a, Total cooking energy demand until 2050 under different scenarios. b, Average cooking energy \ndemand per capita in the reference year 2010, and in 2030 for biomass-dependent regions (AFR, LAM, PAS and SAS). c, Average cooking energy demand \nper capita in the reference year 2010, and in 2030 in other regions (CPA, MEA and the rest of the world).\nNature Energy | VOL 6 | November 2021 | 1067\u20131076 | www.nature.com/natureenergy\n1069\n\nAnalysis\nNAture Energy\nFinal cooking energy demand\nIn many developing regions, cooking is still the most energy-intensive \nactivity in homes. Figure 2 shows future total cooking energy \ndemand by scenario. We see a decline in biomass use over time in \nall reference scenarios with a faster phase out under SSP1 compared \nwith the other reference scenarios (Fig. 2a). Under the COVID sce-\nnario, we observe a much larger share of solid fuels, and higher total \ncooking energy demand because of the inefficiency of these fuels. \nThis large COVID effect could result in an increase in biomass use \nuntil 2030, with a light rebound in 2040, when income levels are \nassumed to go back to the reference SSP3 trend. Climate mitigation \npolicies could attenuate the transition away from polluting stoves by \nmaking oil and gas-based fuels more expensive. The higher prices \ncould also result in lower average per capita final cooking energy \nFirewood\nCharcoal/Coal\nKerosene\nLPG/Natural gas\nElectricity\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nFuel distribution (%)\nFuel distribution (%)\nSub\u2212Saharan Africa\na\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nCentrally Planned Asia and China\nb\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nLatin America and the Caribbean\nc\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nFuel distribution (%)\nMiddle East and North Africa\nd\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nFuel distribution (%)\nOther Pacific Asia\ne\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nSouth Asia\nf\nFig. 3 | Total cooking energy mix in 2030 for rural populations. a\u2013f, Distribution of cooking fuels by income and scenario in 2030 for rural households in \ndifferent regions, overlaid with the population income distribution: AFR (a), CPA (b), LAM (c), MEA (d), PAS (e) and SAS (f).\nNature Energy | VOL 6 | November 2021 | 1067\u20131076 | www.nature.com/natureenergy\n1070\n\nAnalysis\nNAture Energy\ndemand under the climate mitigation policy scenarios as compared \nwith the reference scenarios and a much smaller proportion of gas \nin the cooking energy mix.\nWe find that the rate of change in cooking energy demand across \nregions varies. In developed regions, there is little change in the ref-\nerence scenarios as there is little transition in fuels over time, and \npopulation and urbanization remain quite stable. In emerging and \ndeveloping regions, for example, CPA, income growth and urban-\nization result in shifts from less efficient to more efficient fuels, but \npopulation shifts could mean relatively little change in total cooking \nenergy demand. On the other hand, in AFR, total cooking energy \ndemand could increase over time on account of rapid population \ngrowth and a slow transition away from polluting stoves. We also \nfind demand may rise initially in the region, as populations move \nFirewood\nCharcoal/Coal\nKerosene\nLPG/Natural gas\nElectricity\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nSub\u2212Saharan Africa\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nCentrally Planned Asia and China\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nFuel distribution (%)\nLatin America and the Caribbean\na\nb\nc\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nMiddle East and North Africa\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n2.5\n5.0\n7.5\n10.0\n12.5\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nOther Pacific Asia\nCOVID\nSSP1\nSSP2\nSSP3\nREF\nCP2C\n0\n0.25\n0.50\n0.75\n1.00\n0\n0.25\n0.50\n0.75\n1.00\nlog[Expenditure (US$2010 person\u20131 yr\u20131)]\nSouth Asia\nd\ne\nf\nFig. 4 | Total cooking energy mix in 2030 for urban populations. a\u2013f, Distribution of cooking fuels by income and scenario in 2030 for urban households in \ndifferent regions, overlaid with the population income distribution: AFR (a), CPA (b), LAM (c), MEA (d), PAS (e) and SAS (f).\nNature Energy | VOL 6 | November 2021 | 1067\u20131076 | www.nature.com/natureenergy\n1071\n\nAnalysis\nNAture Energy\nout of extreme poverty. Cooking energy demand can rise if house-\nholds increase cooking frequency, shift diets to eating different \nfoods or simply cook more food.\nDifferences in the average per capita cooking energy demand \nin the year 2030 by scenario are depicted in Fig. 2b,c. We distin-\nguished regions that are more dependent on biomass (that is, AFR, \nLAM, PAS and SAS) from other regions of the world that are not.  \nThe differences in average cooking energy demand between the \ntwo are stark. Furthermore, the differences between scenarios are \nmore pronounced for the non-biomass-dependent regions (Fig. \n2c). Under the climate mitigation policy scenarios, we find cook-\ning energy demand could be lower than in the reference scenarios, \nas the consumption of gas may decrease substantially, more so in \nregions that are non-biomass-dependent to begin with, and there \n0\n500\n1,000\n1,500\n2,000\n0\n500\n1,000\n1,500\n2,000\nPopulation spending less that US$5 per capita per day (millions)\nPopulation using solid fuels (millions)\nAFR\nCPA\nLAM\nMEA\nPAS\nSAS\nSSP1 \nSSP1-CP2C\nSSP2 \nSSP2-CP2C\nSSP3 \nSSP3-CP2C\nCOVID \nCOVID-CP2C\na\nOther Pacific Asia\nMiddle East and North Africa\nCentrally Planned Asia and China\nSub\u2212Saharan Africa\nSouth Asia\nLatin America and the Caribbean\nSSP1\nSSP2\nSSP3\nCOVID\nSSP1-CP2C\nSSP2-CP2C\nSSP3-CP2C\nCOVID-CP2C\nCOVID-CP2C\nCOVID-CP2C\nSSP1\nSSP2\nSSP3\nCOVID\nSSP1-CP2C\nSSP2-CP2C\nSSP3-CP2C\nSSP1\nSSP2\nSSP3\nCOVID\nSSP1-CP2C\nSSP2-CP2C\nSSP3-CP2C\n\u2212750\n\u2212500\n\u2212250\n0\n250\n500\n\u2212750\n\u2212500\n\u2212250\n0\n250\n500\nMillions of people\nMillions of people\nRural, less than US$5 person\u20131 day\u20131\nRural, more than US$5 person\u20131 day\u20131\nUrban, less than US$5 person\u20131 day\u20131\nUrban, more than US$5 person\u20131 day\u20131\nb\nFig. 5 | Relationship between income poor and cooking poor populations by region and scenario. a, Cooking poor populations plotted against income \npoor in 2030. The diagonal line represents equal numbers of cooking poor and income poor populations. b, Relative changes in cooking poor populations \nbetween 2010 and 2030 by scenario and region.\nNature Energy | VOL 6 | November 2021 | 1067\u20131076 | www.nature.com/natureenergy\n1072\n\nAnalysis\nNAture Energy\ncould be an increase in electricity consumption instead. In the \nCOVID scenario, we find cooking demand could be lower than in \nthe reference scenarios in the biomass-dependent regions, because \nof lower income levels (Fig. 2b). In the non-biomass-dependent \nregions too, biomass demand could be slightly higher in the COVID \nscenario. More biomass use is also likely in the SSP3 reference sce-\nnario as compared with the SSP1 reference scenario.\nCooking energy transitions\nIn line with the literature, we found that rising incomes and urban-\nization drive a transition to cleaner fuels and stoves17,26,27. However, \nwithin and across regions, the nature and pace of this transition \nvary vastly depending on several factors beyond income. In Figs. 3 \nand 4 we present the total cooking energy mix in 2030 for the full \nincome distribution by region and scenario, separately for rural and \nurban populations, respectively. Income is presented in the logarith-\nmic scale to better visualize the transitions at lower levels of income.\nClear differences in the pace and nature of the transition in cook-\ning fuels across regions are evident. In rural CPA in the COVID sce-\nnario (top left panel of Fig. 3b), we see most households with very \nlow incomes per capita (approximately less than US$5 per capita per \nday). We see a transition from high firewood dependence at lower \nincome levels to almost equal shares of firewood, gas and electric-\nity at the highest income levels. Households with middle-income \nlevels could still rely mostly on biomass fuels. By contrast, among \nurban households (top left panel of Fig. 4b), income per capita \nlevels are higher (up to approximately US$17 per capita per day). \nThese households are likely to depend mostly on gas and electricity. \nThe pattern varies greatly across other regions, even at comparable \nincome levels. Nevertheless, overall, we see a strong income effect \non the choice of cooking fuels, with households with higher levels of \nincome transitioning to either gas or electricity in all regions, except \nin the Middle East and North Africa (MEA), a region rich in fos-\nsil fuels and poor in biomass. In MEA, higher-income households \ncould continue to use cheap kerosene, in line with what we observed \nin the empirical data (see Methods section).\nFor lower-income rural households, especially in AFR, SAS and, \nto a lesser extent, LAM, we find a large share of total cooking energy \ndemand could still be biomass-based even in 2050, particularly in \nthe COVID scenario. In the climate mitigation policy scenarios \ntoo, an increase in fossil fuel prices could increase the dependence \non biomass, with the price effect dominating the income effect in \ndetermining the choice of fuels. Considering regional heterogene-\nity, in AFR and SAS, regions that are most acutely dependent on \nbiomass today, we find price sensitivity could be higher under the \nclimate mitigation policy scenarios.\nIn addition to income, our analysis shows a clear urban\u2013rural \ndivide in fuel choice, with even richer households in rural areas \nlikely to continue relying on solid fuels because of their easy access \nand the poor accessibility to cleaner alternatives. Indeed, as we can \nsee from the example of CPA, there are stark differences in the \nchoice of cooking fuels between urban and rural households even \nat the same income levels. We observed this in other regions of the \nworld as well (Supplementary Fig. 3).\nFinally, we find energy prices also affect cooking fuel transitions. \nFor instance, in SAS, we find that in the climate mitigation policy sce-\nnarios, rising fossil fuel prices could push LPG out of reach of many. \nHowever, future transitions in this region remain the most uncer-\ntain, as there are currently strong policies to expand LPG access to \neven rural households in India. The effect of these policies are only \nbecoming evident now and are not reflected in the data that we used \nto estimate the parameters for this region (see Methods section).\nPopulations most at risk of being cooking poor\nOur findings on the transition in cooking fuels discussed above \nsuggest that access to clean cooking is clearly a poverty issue. This \nis further illustrated in Fig. 5. In Fig. 5a we show the relationship \nbetween cooking poor and income poor populations (defined \nas those earning less than US$5 per capita per day). We see that \nin AFR, and to a lesser extent in LAM and PAS, income poverty \nstrongly correlates with cooking poverty. In other regions, this is \nless so, as the number of poor individuals is much lower than those \ndependent on solid fuels. This can be explained by the existence of \nnatural resources that result in lower prices of liquid fuels in some \nregions (for example, MEA), or sustained public policies aimed at \npoverty alleviation and increasing access to clean cooking in others \n(for example, in CPA and SAS).\nIn AFR, we find that future income growth may not compensate \nfor the effect of population growth, so that the number of cooking \npoor could increase in all scenarios. What is also evident from Fig. \n5b is that AFR could remain largely rural in 2030, and between 94% \n(in SSP1) and 98% (in COVID) of rural households could earn less \nthan US$5 per capita per day and remain cooking poor in 2030. In \nSAS too, 78% (in SSP1) to 90% (in COVID) of the rural population \ncould earn less than US$5 per capita per day and remain cooking \npoor in 2030. In other regions of the Global South, we found that \nabout 15% more people could remain income poor in the COVID \nscenario compared with in the SSP3 scenario.\nIn other regions of Asia, specifically in CPA, but also in PAS, \nthe number of cooking poor is likely to decline between 2010 and \n2030. In these regions, not only do we find that many move out of \npoverty, but also rising urbanization makes clean fuels more acces-\nsible and affordable. For example, in CPA, in the COVID scenario, \nwe observe that a little less than half of the rural population could \nearn less than US$5 per capita per day and depend on solid fuels, \nbut the percentage of urban households in this category is likely to \nbe negligible.\nWe find outcomes differ by scenario in MEA, LAM and SAS. \nIn these regions, in the SSP1 and SSP2 scenarios we observe there \ncould be a reduction in cooking poor, but in the SSP3 and COVID \nscenarios we find that the extent of both poverty and cooking pov-\nerty could rise. In MEA and SAS this is largely explained by dif-\nferences in urbanization between the SSP1 and SSP2 scenarios as \ncompared with the SSP3 and COVID scenarios. In LAM, however, \nwe find even urban households could remain cooking poor, as \nincome inequality is higher in this region, even in countries that \nhave a gross domestic product (GDP) per capita comparable to \nthat in most developed regions. Therefore, we find many people \ncould remain income poor even in 2030. Especially in the SSP3 \nand COVID scenarios, we find that 60% of rural and about 50% of \nurban populations could earn less than US$5 per capita per day and \ndepend on solid fuels for cooking.\nIn general, our findings show that clean cooking access in the \nCOVID scenario is likely to be lower than under the climate miti-\ngation policy scenarios. In SAS, however, the urban poor could be \nparticularly affected by rising gas prices under the climate mitiga-\ntion policy scenarios, with the dependency on solid fuels for this \ngroup almost doubling. However, even in this region, solid fuel \ndependency is much higher at 7% in the COVID scenario in 2030 \nas compared with under the climate mitigation policy scenarios (2% \nin SSP3-CP2C).\nDiscussion and conclusions\nOur analysis provides new insights into how access to clean cook-\ning services may change under alternative reference scenarios, a \nslow pandemic recovery scenario and climate mitigation policy \nscenarios. Our findings show that the SDG\u20097 target of universal \nclean cooking access by 2030 could be out of reach under all the \nscenarios we explored. In regions that currently have the highest \naccess gaps, specifically AFR, SAS, PAS and LAM, universal access \nmay not be achieved even in 2050. A protracted recession follow-\ning the pandemic could further retard progress towards achieving \nNature Energy | VOL 6 | November 2021 | 1067\u20131076 | www.nature.com/natureenergy\n1073\n\nAnalysis\nNAture Energy\nextensive subsidies. But sustaining these subsidies can become a fis-\ncal burden. Better targeting of subsidies, through efforts like India\u2019s \nGive It Up campaign, and better market segmentation could help37. \nIn addition, directing green and climate funds and revenues from \ncarbon pricing to the clean cooking sector can be another way to \nincrease financing to this sector38.\nDespite the challenges in extending clean cooking access, recent \nadvances in technologies39 and new payment and financing mod-\nels40 can help clean cooking services reach even low-income house-\nholds. In urban centres, providing piped gas to dense settlements \nand introducing smaller LPG cylinder sizes, pay-for-service financ-\ning models, smart metering for gas with electronic payment options \nas well as more reliable and affordable electricity can be instrumen-\ntal in encouraging a more rapid transition to clean cooking. For \nrural regions, awareness-raising and behaviour change campaigns \nare also important to ensure that those who gain access use new \nstoves regularly41. Our results suggest a need for much greater prior-\nitization and coordinated policies to provide access to clean cooking \nglobally, with efforts targeted at the most disadvantaged, specifically \nthe poorest regions and populations. This will require considerable \nupscaling of investment, capacity and commitment, but can result \nin big benefits for planetary and population health and wellbeing.\nMethods\nData sources. We used microdata from a large set of nationally representative \nhousehold surveys of different countries to estimate energy and cooking \ntechnology choices for regions of the world. We aggregated nations in line with \nthe 11 regions of the MESSAGEix-GLOBIOM model42. The datasets used to \nrepresent each of these regions are presented in Supplementary Table 1, with a \nfocus on regions of the world where access to clean cooking is lacking. Regions \nwhere access is not an issue (that is, North America (NAM), Western Europe \n(WEU) and Pacific OECD (PAO)) were included in our analysis and modelled \nindependently, but are presented clubbed together in a single other or rest of the \nworld region in some results.\nNone of these surveys report energy consumption for cooking purposes \nseparately, but instead include data on total consumption or expenditure by fuel. \nThis was not an issue for the modelling approach we followed for the Global North \ncountries (as described further in the following subsection), but was relevant \nfor regions of the Global South. Therefore, we used two different approaches to \nseparate energy for cooking from energy for other end uses. Both approaches \nrequired us to create profiles of consumers who use the fuel for cooking and those \nwho do not. Depending on the particulars of the specific national dataset, we then \neither used a simple regression approach to back out representative non-cooking \nconsumption of households depending on a set of characteristics, or we calculated \nsimple averages differentiated by income categories for urban and rural households \nseparately. Importantly, the simulation-based estimation approach we followed \nlargely minimizes the potential biases of such estimation, as we first obtained a \nguesstimate of cooking consumption using unbiased simulators derived from the \nfirst- and second-order moments of the empirical distributions, and next matched \nthe non-biased first moments from the simulated data with the empirical data \nwhile minimizing errors in the second moments.\nModels and estimation. We used two different models for different regions of the \nworld. For the Global South, where energy stacking is more prevalent, we used \na previously developed model of household cooking fuel choices that allows for \nmultiple fuel use21. For the Global North, where households generally use only \none type of fuel, we used a different more general model of household energy and \nappliance choices22.\nBoth models follow a simulation-based structural econometric approach, \nwherein a large set of simulated households are constructed to mimic the \ncharacteristics found in the empirical data. Households choose between cooking \nalternatives to maximize utility given budget constraints and the prices they face \nin the market. All the simulated households have different characteristics (for \nexample, income, household size and location, among others), and therefore the \nmodels capture heterogeneity with respect to these characteristics. The models, \nalthough different, work by backing out household preferences for different \ncooking fuels and technologies from the choices observed in the empirical data \nsources using different simulation-based estimation techniques. Therefore, the \nmodels only include options that appear in the different sources of household data \n(Supplementary Table 1), namely, firewood, charcoal and coal (usually lumped \ntogether), kerosene, LPG and natural gas (also usually lumped together), and \nelectricity. Unfortunately, this implies that preferences for fuels that are lumped \ntogether cannot be distinguished from each other, and therefore are modelled \nas a single option. Also, other potential viable alternatives (for example, biogas \nthis goal. Our findings resonate with recent analysis that suggests \nthat the aftermath of the pandemic could push half a billion people \nback into extreme poverty28. Climate mitigation policies may also \nhamper a transition to fossil-based cleaner burning cooking fuels \nby making these more expensive and putting them out of reach of \nmany if not twinned with appropriate additional policies specifi-\ncally targeting the cooking poor in AFR, parts of developing Asia \nand LAM.\nThe location and supply of fuels and their affordability are critical \ndrivers of the trends that we observed. In MEA, where biomass sup-\nply is limited, we found dependence on cheap kerosene could remain \nhigh. In AFR and LAM, where many households are remotely \nlocated, have easy access to biomass resources and limited access to \nmodern cooking services, we found biomass dependency is likely \nto persist. For these populations, the inconveniences and health \nimpacts of cooking with polluting stoves need to be factored into the \nhousehold choice decision. This requires committed and sustained \npolicies to provide easy access to modern fuels and stoves at afford-\nable prices, and information and behaviour change messaging. If \nanything, studies in regions that are enforcing such policies now (for \nexample, India) show that these behaviours can be persistent, and \nproviding households access to new stoves and cleaner burning fuels \nalone may not be enough to promote their regular use29,30.\nOur analysis has some limitations that point to important ave-\nnues for future research. Our model can capture a wide variety of \nheterogeneity in circumstances and population characteristics, but \nour results are only as good as the input data used. As our methods \nare based on empirical data, we could only include cooking options \nthat exist in the datasets that we employed. Although newer alter-\nnatives that might become viable in the future were not explicitly \nmodelled, our analysis can inform policy of the price points and \nincome levels at which populations in different regions will be able \nto afford new fuels and technologies. Because cooking behaviours \nappear to be quite persistent, future research should consider using \npanel or pseudo-panel datasets that can better capture longitudinal \nshifts. Our analysis could also be further expanded to better assess \nhow societal changes, such as better education and women\u2019s empow-\nerment and labour force participation, relate to cooking energy \nchoices. We did not include institutional capacity and governance \nconstraints that might limit the expansion of supply of certain cook-\ning options or the effectiveness of policies in specific contexts. In \naddition, our analysis did not assume any climate feedback on bio-\nmass availability. Both unsustainable harvesting and future climate \nchange could reduce the availability of abundant biomass resources \nin certain regions. Alternatively, more sustainable land and biomass \nmanagement could make supplies more abundant and encourage \nnew biomass-based clean cooking options31,32.\nInsights from our analysis emphasize the increased urgency to \naddress this issue, which has been left on the back burner for far too \nlong. A recent report has highlighted the chronic underinvestment \nin the sector, particularly in regions where this is needed most33. \nRecent estimates suggest that achieving universal access to clean \ncooking services by 2030 will require US$4.5\u20139.8 billion annu-\nally5,33. We estimate an average expenditure gap of US$10.5 billion \nannually between 2020 and 2030, comparing the reference SSP3 and \nCOVID scenarios. This includes the cost of stoves and expenditure \non fuels, and is in line with previous estimates. These estimates are \nan order of magnitude lower than the US$2 trillion of estimated \nlosses incurred each year from a lack of access5. Recent commit-\nments pledged to COVID-19 recovery funds are in the range \nUS$9\u201315 trillion34,35. Directing even a fraction of 1% of these funds \nto eradicating cooking poverty could help bridge the financing gap \nto meet the SDG\u20097 target.\nCountries like Brazil, India and Indonesia have expanded access \nto clean cooking services considerably in recent years36. This has \nbeen achieved through strong government commitment and  \nNature Energy | VOL 6 | November 2021 | 1067\u20131076 | www.nature.com/natureenergy\n1074\n\nAnalysis\nNAture Energy\nwhere \u2018cook\u2019 refers to the particular cooking appliance used by the household and \ni takes the value 1 if households use an electric stove and 2 if they use another fuel. \nThe unknown preference parameters (\u03b1 and the \u03d5s and \u03bbs) are estimated following \na similar procedure to the one described above, albeit using a different estimation \ntechnique, namely indirect inference44, which is a generalized version of the \nmethod of simulated moments. The empirical fit of the model can be found in the \nSupplementary Data file.\nScenario design and future simulations. We considered three different future \nreference scenarios following the narratives of the Shared Socioeconomic \nPathways, specifically sustainability (SSP1), middle-of-the-road (SSP2) and \nregional rivalry (SSP3). Although the population45 and urbanization46 projections \nby country for the SSP scenarios that we used are as described in the references \ncited here, a special algorithm was applied to obtain separate income distributions \nfor urban and rural households. Specifically, a machine learning algorithm47 \nwas applied to a large longitudinal dataset including average household income \nand Gini coefficient estimates for urban and rural households of all regions of \nthe world to obtain associations between these indicators and other SSP drivers. \nThe algorithm was trained on data from 1988 to 2010, and data from 2011 to \n2015 were used for validation. These associations were posteriorly applied to the \ncorresponding scenarios25,45,46,48 to obtain the future income distributions for urban \nand rural households separately in all countries and regions.\nThe COVID scenario deserves special attention. For its elaboration, the \noriginal GDP trend for the reference SSP3 scenario from 2010 to 2019 was retained \n(as this is closer to the actual GDP trends in the past few years). For the years \n2020 and 2021, we then used the most recent national GDP growth estimates \nfrom the World Bank49. We assumed convergence in GDP growth trends in the \nyear 2040, considering a 20-year protracted recovery period from the pandemic \nworldwide. The revised GDP estimates were used to adjust the level of average \nhousehold income in each region and period. The standard deviation, which is \nused to estimate inequality, was also adjusted in line with its association with GDP, \nurbanization and year. The Supplementary Data file contains values of all scenario \ndrivers for the period of analysis.\nWe also considered ambitious climate mitigation policy scenarios that limit \nglobal warming to below 2\u2009\u00b0C by the end of the century. The climate mitigation \npolicy simulated in this analysis considers regionally differentiated carbon \nprice trajectories as estimated by McCollum et al.25. Details of all key scenario \ndrivers and assumptions are presented in the Supplementary Data file. The \nmicrodata-based structural estimation models described above are soft-linked \nto the MESSAGEix-GLOBIOM integrated assessment model42, which allows \nmacroeconomic feedbacks via energy prices to be captured. These reflect changes \nin the energy supply mix and land use to meet estimated demands at least cost.\nData availability\nLinks to the micro datasets that were used in the analysis are included in the \nSupplementary Information, when available. Given that some of these datasets are \nnot publicly available, the data used for the estimation module is only available \nfrom the corresponding author upon reasonable request. The simulated datasets \ngenerated during the current study are also available from the corresponding \nauthor upon reasonable request. All estimated moments, scenario assumptions and \nthe datasets underlying the plots are available in the Supplementary Data file.\nCode availability\nThe codes used during the current study are available from the corresponding \nauthor upon reasonable request.\nReceived: 14 December 2020; Accepted: 31 August 2021;  \nPublished online: 28 October 2021\nReferences\n\t1.\t Rosenthal, J., Quinn, A., Grieshop, A. P., Pillarisetti, A. & Glass, R. I. Clean \ncooking and the SDGs: integrated analytical approaches to guide energy \ninterventions for health and environment goals. Energy Sustain. Dev. 42, \n152\u2013159 (2018).\n\t2.\t Maji, P. & Kandlikar, M. Quantifying the air quality, climate and equity \nimplications of India\u2019s household energy transition. Energy Sustain. Dev. 55, \n37\u201347 (2020).\n\t3.\t Watts, N. et al. The 2020 report of The Lancet Countdown on health and \nclimate change: responding to converging crises. Lancet 397, 129\u2013170 (2021).\n\t4.\t IEA, IRENA, UNSD, WB & WHO Tracking SDG 7: The Energy Progress \nReport 2020 (World Bank, 2020); https://openknowledge.worldbank.org/\nhandle/10986/33822\n\t5.\t Energy Sector Management Assistance Program The State of Access to Modern \nEnergy Cooking Services (World Bank, 2020); http://documents.worldbank.\norg/curated/en/937141600195758792/The-State-of-Access-to-Modern-Energy- \nCooking-Services\n\t6.\t Fattorini, D. & Regoli, F. Role of the chronic air pollution levels in the \nCovid-19 outbreak risk in Italy. Environ. Pollut. 264, 114732 (2020).\nand ethanol) cannot be included in the model. However, given that preferences \nare estimated independently for different regions of the world, the methods we \nemployed allow for capturing local behaviours and preferences, which add more \nrelevant layers of heterogeneity to our results.\nModel of cooking fuel choices. For the Global South, we applied a simulation-based \nstructural econometrics model of household cooking fuel choice (details of the \nmodel are available elsewhere21). In brief, the approach assumes households \nmaximize their utility U by choosing between the consumption of cooking fuels \n(allowing for fuel stacking, if households prefer to do so) and other items:\nmax\nC,F U (C, F) =\n\uf8ee\n\uf8f0C\u03b1\n\uf8eb\n\uf8ed\nNf\n\u2211\nf=1\nefFf\n\uf8f6\n\uf8f8\n1\u2212\u03b1\uf8f9\n\uf8fb\n\u03b3\n[\u03c7 (F1...FNf\n)]1\u2212\u03b3\nsubject to:\npcC +\nNf\n\u2211\nf=1\n(pfFf + Af\n)\n= I\nC, Ff \u22650\nwhere C is the consumption of items other than cooking fuels, pc is the price of \nother consumption goods, pf is the price of cooking fuel f, Ff is the consumption of \ncooking fuel f, Af is an annualized representation of the cost of a cooking stove of \nfuel f, I is the household expenditure and \u03c7 (F1...FNf\n) is a function that represents \npreference shifts for reasons other than the consumption of the fuel, such as time \nsavings or health impacts.\nThe unobserved preference parameters of the model (\u03b1, \u03b3, K, \u03b41f and \u03b42f) are \nbacked-out from empirical data using the method of simulated moments43, a \nsimulation-based estimation procedure that seeks to identify numerical parameter \nvalues that match a set of selected moments from the empirical data with the same \nset of moments computed from the simulated dataset of individuals with similar \ncharacteristics to their empirical counterparts in the data, whose fuel choices are \nthen determined using the model described above.\nThe parameters were estimated independently for each region. To be \nrepresentative, several sources of data were used for regions with very \nheterogeneous realities (for example, urbanization rates, natural environments, \nand supply of and accessibility to different cooking fuels and technologies). The fit \nof the model can be seen graphically in Supplementary Fig. 2. Naturally, regions \nthat present non-smooth distributions of cooking fuels over income were harder \nto match (that is, CPA, LAM and PAS). Overall, the observed trends over income \nwere appropriately captured by the model. The full set of matched moments are \npresented in the Supplementary Data file.\nModel of energy consumption and appliance choices. Cooking choices and demand \nin the Global North were estimated using a different simulation-based structural \neconometrics model that reflects the multiple choices of households in terms of \nfuels and appliances for different end uses22. In this model, households, based on \ntheir characteristics, make a discrete choice between available cooking technologies \nin each region, which is finally reflected in their total energy consumption for the \nrespective fuel. Specifically, if households choose an electric stove, their electricity \nconsumption is calculated as:\nx1 = \u03d50 +\nm\n\u2211\nj=1\n\u03b4j\u03d5j +\n\uf8ee\n\uf8f0\u03bb1p1 + \u03bb2p2 + \u03bb3 \u03c7 + \u03bb4\n\uf8eb\n\uf8edy \u2212\u03c1\nm\n\u2211\nj=1\nKj\u03b4j\n\uf8f6\n\uf8f8\n\uf8f9\n\uf8fb,\nwhereas if they choose a gas or biomass stove, their total alternative fuel \nconsumption is calculated as:\nx2 = \u03bb2\n\u03bb4 (\u03b1 \u22121) + \u03b1\n\u03bb4\n(\n\u03d50 + \u03bb1\n\u03bb4 + \u03bb3 \u03c7\n)\n1\np2 + \u03b1\u03bb1\n\u03bb4\np1\np2\n+ \u03b1\np2\n[\ny \u2212\u03c1\nm\n\u2211\nj=1\nKj\u03b4j +\nm\n\u2211\nj=1\n\u03d5j\n\u03bb4 \u03b4j\n]\n,\nwhere consumption xi is modelled as the sum of the base consumption of electricity \n\u03d50 and consumption due to ownership of appliances \u03b4j\u03d5j, where \u03b4j represents the \nownership of appliance j\u2009\u2208\u2009m and \u03d5j is the average electricity consumption of the \nappliance. In addition, each household\u2019s consumption is affected by the price of \nelectricity p1, the price of alternative fuels p2, a vector of household characteristics  \n\u03c7 and the household\u2019s disposable income (that is, total income y minus the \nannualized investment cost of electric appliances \u03c1\nm\n\u2211\nj=1\nKj\u03b4j).\nFinally, the actual cooking energy consumption is calculated as:\nEcook =\n\u03b4cook\u03d5cook\n\u03d50 + \u2211m\nj=1 \u03b4j\u03d5j\nxi,\nNature Energy | VOL 6 | November 2021 | 1067\u20131076 | www.nature.com/natureenergy\n1075\n\nAnalysis\nNAture Energy\n\t33.\tEnergizing Finance: Understanding the Landscape (Sustainable Energy for All, \n2020); https://www.seforall.org/system/files/2020-11/EF-2020-UL-SEforALL_0.\npdf\n\t34.\tAndrijevic, M., Schleussner, C.-F., Gidden, M. J., McCollum, D. L. & Rogelj, J. \nCOVID-19 recovery funds dwarf clean energy investment needs. Science 370, \n298\u2013300 (2020).\n\t35.\tSovacool, B. K., Furszyfer Del Rio, D. & Griffiths, S. Contextualizing the \nCovid-19 pandemic for a carbon-constrained world: insights for sustainability \ntransitions, energy justice, and research methodology. Energy Res. Soc. Sci. 68, \n101701 (2020).\n\t36.\tGoldemberg, J., Martinez-Gomez, J., Sagar, A. & Smith, K. R. Household air \npollution, health, and climate change: cleaning the air. Environ. Res. Lett. 13, \n30201 (2018).\n\t37.\tTripathi, A. & Sagar, A. Ujjwala, V2.0: What Should Be Done Next? \n(Collaborative Clean Air Policy Centre, 2019); https://ccapc.org.in/s/\nUjjwala-V20-Jun-19b.pdf\n\t38.\tCameron, C. et al. Policy trade-offs between climate mitigation and clean \ncook-stove access in South Asia. Nat. Energy 1, 15010 (2016).\n\t39.\tBatchelor, S. et al. Solar e-cooking: a proposition for solar home system \nintegrated clean cooking. Energies 11, 2933 (2018).\n\t40.\tShupler, M. et al. Pay-as-you-go liquefied petroleum gas supports sustainable \nclean cooking in Kenyan informal urban settlement during COVID-19 \nlockdown. Appl. Energy 292, 116769 (2021).\n\t41.\tBarnes, B. et al. Behavior change communication: a key ingredient for \nadvancing clean cooking. J. Health Commun. 20, 3\u20135 (2015).\n\t42.\tKrey, V. et al. {MESSAGEix-GLOBIOM} Documentation - 2020 release. \n(IIASA, 2020); https://doi.org/10.22022/IACC/03-2021.17115\n\t43.\tMcFadden, D. A method of simulated moments for estimation of discrete \nresponse models without numerical integration. Econometrica 57, 995\u20131026 \n(1989).\n\t44.\tGourieroux, C., Monfort, A. & Renault, E. Indirect inference. J. Appl. \nEconom. 8, S85\u2013S118 (1993).\n\t45.\tK.C., S. & Lutz, W. The human core of the shared socioeconomic pathways: \npopulation scenarios by age, sex and level of education for all countries to \n2100. Glob. Environ. Change 42, 181\u2013192 (2017).\n\t46.\tJiang, L. & O\u2019Neill, B. C. Global urbanization projections for the Shared \nSocioeconomic Pathways. Glob. Environ. Change 42, 193\u2013199 (2017).\n\t47.\tFriedman, J. H. Stochastic gradient boosting. Comput. Stat. Data Anal. 38, \n367\u2013378 (2002).\n\t48.\tRao, N. D., Sauer, P., Gidden, M. & Riahi, K. Income inequality projections \nfor the Shared Socioeconomic Pathways (SSPs). Futures 105, 27\u201339 (2019).\n\t49.\tGlobal outlook: pandemic, recession: the global economy in crisis. In Global \nEconomic Prospects 1\u201366 (World Bank, 2020); https://doi.org/10.1596/978-1- \n4648-1553-9_ch1\nAcknowledgements\nM.P.-C. received funding from the European Union\u2019s Horizon 2020 research and \ninnovation programme under grant agreement no. 821124 (NAVIGATE). This work \nwas partially funded by the contributions of the National Member Organizations of the \nInternational Institute of Applied Systems Analysis.\nAuthor contributions\nS.P. and M.P.-C. conceived the initial framework. S.P. and M.P.-C. designed the research. \nM.P.-C., A.A. and M.J.G. prepared the data. M.P.-C. and A.A. performed the modelling, \nwrote the codes and carried out the analysis. S.P. and M.P.-C. led the writing of the paper \nwith all other authors contributing to the writing, revisions and editing.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary material \navailable at https://doi.org/10.1038/s41560-021-00911-9.\nCorrespondence and requests for materials should be addressed to Shonali Pachauri.\nPeer review information Nature Energy thanks Joshua Rosenthal, Francis Johnson and \nthe other, anonymous, reviewer(s) for their contribution to the peer review of this work.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2021\n\t7.\t Pozzer, A. et al. Regional and global contributions of air pollution to risk of \ndeath from COVID-19. Cardiovasc. Res. 116, 2247\u20132253 (2020).\n\t8.\t The Covid-19 Crisis is Reversing Progress on Energy Access in Africa (IEA, \n2020); https://www.iea.org/articles/the-covid-19-crisis-is-reversing-progress- \non-energy-access-in-africa\n\t9.\t Boza-Kiss, B., Pachauri, S. & Zimm, C. Deprivations and inequities in cities \nviewed through a pandemic lens. Front. Sustain. Cities 3, 15 (2021).\n\t10.\tShupler, M. et al. COVID-19 impacts on household energy & food security in \na Kenyan informal settlement: the need for integrated approaches to the \nSDGs. Renew. Sustain. Energy Rev. 144, 111018 (2021).\n\t11.\tDagnachew, A. G. et al. The role of decentralized systems in providing \nuniversal electricity access in sub-Saharan Africa \u2013 a model-based approach. \nEnergy 139, 184\u2013195 (2017).\n\t12.\tDagnachew, A. G., Lucas, P. L., Hof, A. F. & van Vuuren, D. P. Trade-offs and \nsynergies between universal electricity access and climate change mitigation \nin sub-Saharan Africa. Energy Policy 114, 355\u2013366 (2018).\n\t13.\tPanos, E., Densing, M. & Volkart, K. Access to electricity in the World \nEnergy Council\u2019s global energy scenarios: an outlook for developing regions \nuntil 2030. Energy Strategy Rev. 9, 28\u201349 (2016).\n\t14.\tMoksnes, N., Korkovelos, A., Mentis, D. & Howells, M. Electrification \npathways for Kenya\u2013linking spatial electrification analysis and medium to \nlong term energy planning. Environ. Res. Lett. 12, 95008 (2017).\n\t15.\tDagnachew, A. G., Hof, A. F., Lucas, P. L. & van Vuuren, D. P. Scenario \nanalysis for promoting clean cooking in sub-Saharan Africa: costs and \nbenefits. Energy 192, 116641 (2020).\n\t16.\tPachauri, S., Rao, N. D. & Cameron, C. Outlook for modern cooking energy \naccess in Central America. PLoS ONE13, e0197974 (2018).\n\t17.\tShankar, A. V. et al. Everybody stacks: lessons from household energy case \nstudies to inform design principles for clean energy transitions. Energy Policy \n141, 111468 (2020).\n\t18.\tGould, C. F., Hou, X., Richmond, J., Sharma, A. & Urpelainen, J. Jointly \nmodeling the adoption and use of clean cooking fuels in rural India. Environ. \nRes. Commun. 2, 85004 (2020).\n\t19.\tJeuland, M., Soo, J.-S. T. & Shindell, D. The need for policies to reduce the \ncosts of cleaner cooking in low income settings: implications from systematic \nanalysis of costs and benefits. Energy Policy 121, 275\u2013285 (2018).\n\t20.\tDagnachew, A. G. et al. Integrating energy access, efficiency and renewable \nenergy policies in sub-Saharan Africa: a model-based analysis. Environ. Res. \nLett. 15, 125010 (2020).\n\t21.\tPoblete-Cazenave, M. & Pachauri, S. A structural model of cooking fuel \nchoices in developing countries. Energy Econ. 75, 449\u2013463 (2018).\n\t22.\tPoblete-Cazenave, M. & Pachauri, S. A model of energy poverty and access: \nestimating household electricity demand and appliance ownership. Energy \nEcon. 98, 105266 (2021).\n\t23.\tWHO Guidelines for Indoor Air Quality: Household Fuel Combustion (World \nHealth Organization, 2014); https://www.who.int/airpollution/guidelines/\nhousehold-fuel-combustion/IAQ_HHFC_guidelines.pdf\n\t24.\tO\u2019Neill, B. C. et al. A new scenario framework for climate change research: \nthe concept of shared socioeconomic pathways. Clim. Change 122, 387\u2013400 \n(2014).\n\t25.\tMcCollum, D. L. et al. Energy investment needs for fulfilling the Paris \nAgreement and achieving the Sustainable Development Goals. Nat. Energy 3, \n589\u2013599 (2018).\n\t26.\tFoell, W., Pachauri, S., Spreng, D. & Zerriffi, H. Household cooking fuels and \ntechnologies in developing economies. Energy Policy 39, 7487\u20137496 (2011).\n\t27.\tCoelho, S. T., Sanches-Pereira, A., Tudeschini, L. G. & Goldemberg, J. The \nenergy transition history of fuelwood replacement for liquefied petroleum gas \nin Brazilian households from 1920 to 2016. Energy Policy 123, 41\u201352 (2018).\n\t28.\tSumner, A., Hoy, C. & Ortiz-Juarez, E. Estimates of the Impact of COVID-19 \non Global Poverty WIDER Working Paper 2020/43 (UNU-WIDER, 2020); \nhttps://doi.org/10.35188/UNU-WIDER/2020/800-9\n\t29.\tKar, A., Pachauri, S., Bailis, R. & Zerriffi, H. Using sales data to assess \ncooking gas adoption and the impact of India\u2019s Ujjwala programme in rural \nKarnataka. Nat. Energy 4, 806\u2013814 (2019).\n\t30.\tMani, S., Jain, A., Tripathi, S. & Gould, C. F. Sustained LPG use requires \nprogress on broader development outcomes. Nat. Energy 5, 430\u2013431 (2020).\n\t31.\tVan de Ven, D.-J. et al. Integrated policy assessment and optimisation over \nmultiple sustainable development goals in Eastern Africa. Environ. Res. Lett. \n14, 94001 (2019).\n\t32.\tHosier, R., Kappen, J., Hyseni, B., Tao, N. & Usui, K. Scalable Business Models \nfor Alternative Biomass Cooking Fuels and Their Potential in Sub-Saharan \nAfrica (World Bank, 2017); https://openknowledge.worldbank.org/\nhandle/10986/28595\nNature Energy | VOL 6 | November 2021 | 1067\u20131076 | www.nature.com/natureenergy\n1076\n\n\n    Scientific Research Findings:", "answer": "We explore clean cooking access until\u00a02050 under alternative future scenarios of socioeconomic and demographic change, COVID\u201119 recovery and ambitious climate mitigation. We find that the population share with access to clean cooking improves in all scenarios relative to today, but the target of universal access by\u00a02030 is not reached even in our most optimistic growth and low inequality scenario. About\u00a0470\u00a0million more people could be pushed into cooking\u2011fuel poverty by\u00a02030, exacerbating global inequities, in a slow pandemic recovery scenario that accounts for\u00a02020 and\u00a02021 GDP estimates and assumes a\u00a020\u2011year recovery period, relative to a pessimistic growth scenario that assumes no pandemic shock. We find that populations in sub\u2011Saharan Africa, developing Asia and Latin America are the worst affected. Cooking poverty strongly correlates with income poverty, particularly in sub\u2011Saharan Africa. Ambitious climate mitigation, without additional policies and financial support, could also make clean cooking unaffordable for about\u00a0200\u00a0million people by\u00a02030. A transition to clean cooking can reduce future demand for cooking energy, specifically in regions that currently rely heavily on biomass.", "id": 17}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-021-00781-1\n1Sustainability Research Institute, School of Earth and Environment, University of Leeds, Leeds, UK. 2Institute for Transport Studies, University of Leeds, \nLeeds, UK. 3Grantham Institute, Imperial College London, London, UK. 4Energy Futures Lab, Faculty of Engineering, Imperial College London, London, UK. \n5Chemical Engineering Department, Imperial College London, London, UK. 6, Waikato, New Zealand. \u2709e-mail: s.hall@leeds.ac.uk\nD\nomestic consumers\u2019 choice of energy companies in com-\npetitive markets affects energy transitions1,2. Energy utilities \nhave adopted a volume-sale business model to reduce costs \nand attract consumers3\u20137. This incumbent business model is threat-\nened by low-carbon transitions8,9, because large renewables reduce \nthe profitability of legacy fossil plants10; prosumers and microgen-\neration reduce predictability and volume sales6,11; more suppliers \nare entering the market7,12; and more flexible demand competes \nwith traditional plants for flexibility services13. These pressures are \nforcing the volume-sale model to change and adapt, to survive the \nenergy transition14.\nThe changes utilities make to survive in low-carbon transitions \ncan also change the consumer energy contract15. New consumer con-\ntracts are emerging that integrate decentralized renewables9,16, move \nto energy-as-a-service as opposed to pay-per-kilowatt tariffs17,18, \nand reward consumer behaviour change19\u201321. The energy transition \nin liberalized markets is shaped by these trends22. However, little \nhas been done to understand consumer preferences for these new \nbusiness models and how these consumer preferences could affect  \nthe transition.\nMost research on consumer preferences in energy retail markets \nrelates to switching suppliers23. This work shows that consumers do \nnot respond well to price signals, and the number of households \nthat enter the market to find a better deal is lower than regulators \nwould hope or economists predicted24\u201327. Recent research has turned \nto behavioural economics28 to explain these trends. Behavioural \neconomics recognizes that consumers are not perfectly rational;  \nthey operate in dynamic markets, where options are too complex to \nfully process29.\nThis complexity leads consumers to adopt satisficing over opti-\nmizing behaviour. Satisficing behaviour aims at satisfactory levels of \nperformance given existing resources and imperfect information30. \nThere is strong evidence that consumers are satisficing in their \nchoice of energy supplier29,31,32. Consumers are likely to make deci-\nsions based on simplification strategies that reduce the complex-\nity of decision-making and fall back on heuristics, such as trust in  \ninstitutions, to distinguish between options30,33. Consumers also \nexhibit status-quo bias and loss aversion behaviour in making \nenergy choices34. However, this satisficing behaviour comes at a \nprice, as even consumers on the poorest energy contracts report \nhigh confidence that they are on the best deal for them, and report \nhigh trust in their current supplier23.\nConsumer satisficing is problematic enough when consumers \nare choosing between contracts competing only on price. This prob-\nlem may be compounded if consumers are being asked to choose \nbetween diverse contracts created by utility business model inno-\nvation. Yet, there has been little work to explore the types of util-\nity energy contracts consumers are likely to opt for when presented \nwith a range of possible offers. This is the problem we address in \nthis study. We explicitly assume that utility companies and their \nconsumers are operating under uncertainty, searching for satisfac-\ntory strategies in the market, and that these decisions will evolve \nwith each other to affect the direction and outcomes of energy \ntransitions. To explore this relationship, we define the business \nmodels that utilities are exploring to respond to the pressures of a \nlow-carbon transition and which contractual attributes they might \noffer consumers, and test how consumers respond to these new \nbusiness models, and which consumers prefer which new contracts. \nWe then use the results to explore the implications of these data on \nenergy transitions and market regulation.\nWe found that utilities are developing new ways of electrify-\ning heat and transport, servitising the energy contract, bundling \nenergy services with other infrastructure services and facilitating \npeer-to-peer (P2P) platforms. We also identified four consumer \nsegments with varying appetites for new utility contracts. The seg-\nment with the highest appetite for new models is also the smallest \n(16% of respondents), suggesting that new utility business models \nmight only have a limited niche to expand into. The other three \nsegments face barriers to participation (based on tenancy type \nor income levels) or barriers to acceptance (based on social trust \nor market engagement). Based on our segmentation, we define \nthree challenges for the energy transition. First, there is potential \nMatching consumer segments to innovative utility \nbusiness models\nStephen Hall\u200a \u200a1\u2009\u2709, Jillian Anable2, Jeffrey Hardy\u200a \u200a3, Mark Workman4, Christoph Mazur5 and \nYvonne Matthews6\nEnergy as a service, smart home opportunities and electrification of heat and transport can lead to new ways of switching sup-\nplier or choosing new energy contracts. Here, we used business model collaboration workshops to create archetypes of new \nutility business models, which were then tested with a representative sample of British energy consumers to explore their \nattractiveness to different segments of society. We show that some of these segments have a substantial appetite for new busi-\nness models. However, the segments that choose these models are more likely to be affluent, educated homeowners. Without \nintervention, innovation in utility business models risks exacerbating existing social inequalities, as lower incomes, lower home \nownership and low education result in lower preferences for, or no ability to engage with, new utility business models. We also \nfind that institutional trust beyond the energy sector is a key driver of consumer segmentation.\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n349\n\nArticles\nNaTurE EnErgy\nfor market innovation to stall due to the most receptive segments \nbeing relatively small and/or most likely to rent and the most disen-\ngaged segments being most likely to own homes and have the abil-\nity to include building fabric alteration in their energy contracting. \nSecond, there is a social trust barrier to overcome that leads to low \nconfidence in new utility business models but also applies across \nother societal institutions. Finally, the culmination of these issues \ncould lock in existing social inequalities and lock out some sectors \nof society from participating in low-carbon transitions.\nBusiness model generation\nWe adopted a collaborative business model innovation process35 to \nexplore how utility business models can evolve to meet the chal-\nlenges of a low-carbon transition. We followed Rohrbeck et al.36, \nwho used this approach to support collaborative business model \ngeneration in the German utilities sector. The workshop was under-\ntaken on 15 June 2016 with 38 industry, academic and government \nstakeholders. We identified 11 future utility business model arche-\ntypes that responded to a hierarchy of threats to the current util-\nity business model. We use the term archetype to describe a new \nutility business model. Rohrbeck et al.36 suggested that collabora-\ntive business model generation is done in three stages: (1) idea gen-\neration; (2) prioritization; and (3) validation. As this research was \ntime constrained to a one-day workshop, we used a project steering \ngroup, comprising three utility executives, two infrastructure con-\nsultants and two energy financiers drawn from the Energy Research \nPartnership, for initial idea generation and took these ideas on util-\nity business models for further development in the workshop (see \nSupplementary Methods for details of the workshop process).\nStage 1 prioritized the systemic challenges to the incumbent utility \nbusiness model (see Supplementary Fig. 6). The six highest-priority \nthreats to the incumbent volume-sale utility model were: (1) pol-\nicy uncertainty; (2) large penetration of intermittent renewables;  \n(3) demand-side management; (4) diversifying the supply mar-\nket; (5) cost of capital; and (6) increasing micro-decentralized \noptimization.\nIn stage 2, groups were asked to explore the implications of six \nproposed innovative utility business models designed to address \nthese systemic challenges. These business models were visualized in \ncomponent diagrams37. The six initial business model concepts and \ncomponent diagrams were generated by the project steering group \nand were refined and expanded by workshop participants. Figure 1  \nshows the resultant utility business models and Supplementary  \nFigs. 7\u201317 show all of the business models considered in the workshop.\nThe five business model archetypes shown in Fig. 1 were derived \nfrom the workshop in June 2016 and set the business models to be \nresearched to the middle of 2019. During this time, some of these \npropositions have been tested by utilities in the United Kingdom and \nelsewhere and none have been rendered obsolete or unlikely since \ntheir conception. Indeed, new electric vehicle tariffs are beginning \nto enter the market38, P2P trials are ongoing39, automated-switching \nmodels that show the early stages of the third-party-control arche-\ntype are emerging40, and bundled retrofit and energy service models \nare being piloted41.\nThe attributes of the business models that could be presented \nto a consumer in a switching situation were then developed by the \nresearch team and are shown in Table 1, along with a control busi-\nness model: same but smart (SBS).\nConceptual model development\nThe attributes of each business model were presented to 2,024 \nBritish residential utility bill payers in a questionnaire survey. The \ntheoretical basis of the questionnaire design was the technology \nacceptance model (TAM)42. The value of the TAM, or adapted vari-\nants of it, has been demonstrated in several technology contexts, \nincluding the uptake of IT services, e-commerce and smart grids43. \nUsing an adapted version of the TAM in the technology-driven con-\ntext of innovative utility business models is justified given that each \nbusiness model involves different engagement with technologically \nmediated, energy-smart behaviour44.\nThe original TAM model explains willingness to adopt a tech-\nnology by two factors: (1) perceived ease of use; and (2) perceived \nusefulness. Expanded versions of the TAM have found additional \nconstructs to add explanatory power and these capture individual \nexperience as well as beliefs about how the innovation under study \nwould perform in relation to multiple societal factors. We ensured \nthat our measure of perceived usefulness includes specific measures \nof expected value for adopters of energy contracts as they involve \nP2P\nESC\n3PC\nP2P customers directly buy,\nsell or swap electricity with\neach other\nAn ESC delivers energy services to\ncustomers, such as comfort and\nillumination, rather than units of\nenergy like a traditional supplier\nA third party, such as a price\ncomparison website, takes\ndecisions on consumers\u2019 behalf,\nsuch as automatically switching\nenergy suppliers\nPure low-carbon generator\nProducing low-carbon\npower and selling directly to\nlarge customers or\nwholesale market\nNew electrifier\nTraditional utility that is\nhelping consumers to switch to\nelectric heat and mobility,\nincluding installing equipment\nand automating DSR\nFig. 1 | Utility business model archetypes. Overview of the business models synthesized from the business model collaboration workshop. The pure \nlow-carbon generator was not analysed in the consumer-facing experiment as it sells only to wholesale markets. DSR, demand side response. Figure \nreproduced from ref. 77, Elsevier.\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n350\n\nArticles\nNaTurE EnErgy\nboth technology and service components45. The model draws on \ndiffusion of innovations theory by capturing notions of perceived \nease of use and complexity46 and the theories of reasoned action \nor planned behaviour47. These social-psychological theories have \ninformed behavioural economic approaches, which recognize that \nconsumers are not perfectly economically rational; they operate in \ndynamic markets, where choice is expanding and options are too \ncomplex to process fully29.\nThrough a combination of our domain-specific knowledge of \nenergy efficiency and tariff-switching behaviours within current \nenergy markets, together with the components of trust48 and per-\nsonal innovativeness49, both of which have been proven to enhance \nthe TAM, we developed a conceptual framework on which to base \nthe questionnaire design and analysis (see Methods). The key con-\nstructs consist of: experience (current knowledge, engagement and \nstatus-quo bias); salient beliefs (concern about the future, and green \nTable 1 | Consumer-facing attributes of each new business model archetype\nNew utility \nbusiness model\nConsumer-facing attributes\nSBSa\nYou are free to switch companies as and when you want to. You will have a smart meter with live information at home and on your \nphone. Your supplier can see how and when you use electricity. You can change your behaviour (not use the washing machine and so \non) when you see that electricity is cheaper.\nNew electrifier\nYou have a 2-year contract. You receive a discount for switching your home from gas to electric heat. It will cost about the same as \nnow. You might have some new features installed, such as electric radiators or a heat pump. Your supplier can pause your heating \noccasionally for up to 15\u2009min at a time or take control of when to charge your electric car to help you avoid paying the highest prices, \nalthough you can opt out of this.\nESC\nYou have a 10-year contract. Your energy bills are guaranteed to be lower than you are currently paying for the duration of the \ncontract. You receive one bill for all of your light, heat and any electric car needs. You might have some new features installed, such \nas insulation and a home energy management system. Your supplier can pause your heating and appliances (such as your fridge) \noccasionally for up to 15\u2009min at a time or take control of when to charge your electric car to help you avoid paying the highest prices, \nalthough you can opt out of this.\nP2P\nYou have no contract with any energy supplier. You use an app on your phone to choose, based on price, type or location of energy, \nwho to buy energy from. For example, you might want local green energy, even though it might not be the cheapest. You can change \nwho you receive your energy from as often as you like. If you have a solar panel on your roof, you can make money by selling the \nenergy from it through the app.\n3PC\nYou have a multi-year contract. You tell the company how you want to live your life and it makes decisions on your behalf to deliver this. \nYou receive one bill for all of your energy, broadband, TV, mobile phone, electric vehicle and water services. Your company may offer \nto install equipment, such as insulation and a home energy management system, to make your home more efficient and smarter. Your \ncompany can pause your heating and appliances (such as your fridge) occasionally for up to 15\u2009min at a time or take control of when to \ncharge your electric car to help you avoid paying the highest prices, although you can opt out of this.\naControl archetype used to reflect the contracts available to consumers today.\n0\n10\n20\n30\n40\n50\n60\n70\n80\n90\n100\nSBS\nIntention to adopt (%)\nP2P\n3PC\nLikely\nNeutral\nUnlikeley\nUtility business model name\nNew electrifier\nESC\nFig. 2 | Likelihood of adopting each archetype. Graph of the percentage of participants (n\u2009=\u20092,024) who were likely, unlikely or neutral in their intention \nto adopt each archetype. Responses to the question \u201cIf this option were available today, what is the likelihood you would sign up to it?\u201d were scored on a \nsliding scale from 0\u20131 (very unlikely to very likely), then coded as unlikely (0\u20130.4), neutral (0.5) or likely (0.6\u20131).\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n351\n\nArticles\nNaTurE EnErgy\nbeliefs); personal innovativeness (openness to new technology); \ntrust (trust in operators and perceived risk of system use); perceived \nease of use; perceived usefulness (including the value perceived \nfrom adopting the contracts); and intention (stated likelihood  \nto adopt).\nArchetype preference\nSurvey participants were presented with the information in Table 1 on \neach archetype and were asked to score it on several semantic differen-\ntial attribute scales. This exercise provided for some reflection on, and \nassimilation of, each option. Intention to adopt was measured using a \nEngaged but cautious\n(n = 706; 35%)\nAspiring opt-outs\n(n = 537; 27%)\nUnconvinced and\nunmotivated (n = 449;\n22%)\nPragmatic innovators\n(n = 330; 16%)\nI really want to save the\nplanet but I already shop\naround to get the best deals\nand I am not sure these new\ndeals will be any better.\nI don\u2019t trust any of the\nexisting companies and I am\nworried about future price\nrises. I am quite attracted to\nthe idea of opting out and\ngoing it alone if I could.\nScience and technology are\nthe answers to\nenvironmental problems\nand I am open to change so\nlong as it leads to more\nchoice and freedom.\nThere is not much you could\ndo to interest me in my\nenergy use. These new\ncompanies will peddle a\nload of green nonsense\nwhile exploiting your data.\nFig. 3 | Consumer segments. Label, size and defining statement for each consumer segment.\n0\n10\n20\n30\n40\n50\n60\n70\n80\n90\nEngaged but cautious\nAspiring opt-outs\nUnconvinced and\nunmotivated\nPragmatic innovators\nProbability of adoption (%)\nSBS\nNew electrifier\nESC\nP2P\n3PC\nConsumer segment title\nFig. 4 | Probability of adoption of each archetype in each consumer segment. Probability of adoption was measured as the proportion of times an \narchetype was chosen (out of four eligible times for each archetype for each person). Using ANOVA, the segments differ on all five archetypes at \nP\u2009>\u20090.0001, with the exception of new electrifier, with P\u2009=\u20090.21 (F\u2009=\u200995.138, d.f.\u2009=\u20093 and P\u2009>\u20090.0001 for SBS; F\u2009=\u20091.509, d.f.\u2009=\u20093 and P\u2009=\u20090.21 for new electrifier; \nF\u2009=\u200936.898, d.f.\u2009=\u20093 and P\u2009>\u20090.0001 for ESC; F\u2009=\u200922.477, d.f.\u2009=\u20093 and P\u2009>\u20090.0001 for P2P; and F\u2009=\u200937.408, d.f.\u2009=\u20093 and P\u2009>\u20090.0001 for 3PC).\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n352\n\nArticles\nNaTurE EnErgy\nTable 2 | Profile statistics for each segment on demographic characteristics and clustering variables\n(A) Engaged but \ncautiousa\n(B) Aspiring \nopt-outsa\n(C) Unconvinced \nand unmotivateda\n(D) Pragmatic \ninnovatorsa\nX2/F valueb\nd.f.\nP\u2009value\nDemographics\n\u2002Male (%)\n45.9\n42.5c\n55.9d\n52.7\n24.424\n6\nP\u2009<\u20090.001\n\u2002Owner occupier (%)\n66.1\n53.1c\n71.7d\n58.8\n43.459\n6\nP\u2009<\u20090.001\n\u2002With children (%)\n23.8\n28.7\n12.9c\n49.4d\n134.081\n3\nP\u2009<\u20090.001\n\u2002Degree education or  \nabove (%)\n38.7\n35.4\n28.3c\n43.6d\n24.399\n9\nP\u2009<\u20090.001\n\u2002In work (%)\n51.1\n60.0\n44.8c\n75.5d\n83.908\n4\nP\u2009<\u20090.001\n\u2002>\u00a360,000 per annum  \nincome (%)\n15.5\n10.2c\n11.1\n18.7d\n43.039\n12\nP\u2009<\u20090.001\n\u2002Age (mean (s.d.) years)\n49.6 (15.8)B,C,D\n44.2 (14.9)A,C,D\n55.1c (14.0)A,B,D\n36.6d (13.0)A,B,C\n113.413\n3\nP\u2009<\u20090.001\nExperience/engagement\n\u2002Never switched supplier (%)\n27.2c\n49.2d\n35.9\n31.5\n24.034\n6\nP\u2009<\u20090.001\n\u2002Satisfaction with supply \n(mean (s.d.))#\n4.1d (0.5)B,C\n3.1c (0.8)A,C,D\n3.5 (0.7)B\n4.0 (0.6)B\n262.091\n3\nP\u2009<\u20090.001\n\u2002Number of energy actions \n(mean (s.d.))3\n3.2 (0.5)C,D\n3.2 (0.5) C,D\n3.1c (0.5)A,B,D\n3.6d (0.7)A,B,D\n58.788\n3\nP\u2009<\u20090.001\n\u2002Think about electricity  \n(mean (s.d.))\n2.7 (0.8)B,C,D\n3.4 (0.9)A,C,D\n2.3c (0.8)A,B,D\n3.8d (0.9)A,B,C\n254.647\n3\nP\u2009<\u20090.001\n\u2002Willingness to think more \nabout electricity use  \n(mean (s.d.))\n2.6 (0.8)B,C,D\n3.1 (0.9)A,C,D\n2.1c (0.8)A,B,D\n3.7d (0.9)A,B,C\n286.575\n3\nP\u2009<\u20090.001\n\u2002Energy engagement  \n(mean (s.d.))\n3.6 (0.9)B,C,D\n3.3 (0.9)A,C,D\n3.1c (1.0)A,B,D\n3.8d (0.7)A,B,C\n5.794\n3\nP\u2009<\u20090.001\nSalient beliefs\n\u2002Worried about price now \n(mean (s.d.))\n1.6c (0.8)B,D\n3.1d (1.2)A,C,D\n1.7 (1.1)B,D\n2.6 (1.3)A,B,C\n275.571\n3\nP\u2009<\u20090.001\n\u2002Worried for future  \n(mean (s.d.))\n2.5c (0.9)B,D\n3.8d (0.9)A,C,D\n2.6 (1.1)B,D\n3.4 (1.1)A,B,C\n223.035\n3\nP\u2009<\u20090.001\n\u2002Environmentally responsible \n(mean (s.d.))\n3.8 (0.6)C\n3.8 (0.6)C\n3.2c (0.7)A,B,D\n3.9d (0.6)C\n101.669\n3\nP\u2009<\u20090.001\n\u2002Environmentally sceptical \n(mean (s.d.))\n2.4c (0.6)B,C,D\n2.8 (0.6)A,D\n2.7 (0.6)A,D\n3.4d (0.8)A,B,C\n191.267\n3\nP\u2009<\u20090.001\n\u2002Descriptive norm  \n(mean (s.d.))\n3.3 (0.9)D\n3.2c (0.9)D\n3.3 (0.9)D\n3.6d (0.9)A,B,C\n16.928\n3\nP\u2009<\u20090.001\nPersonal innovativeness\n\u2002Adopts latest technology \n(mean (s.d.))\n2.7 (1.0)C,D\n2.7 (1.1)C,D\n2.1c (0.9)A,B,D\n3.8d (1.0)A,B,C\n176.039\n3\nP\u2009<\u20090.001\nTrust\n\u2002Trusts own energy company \n(mean (s.d.))\n3.8 (0.7)B,C,D\n2.8c (0.9)A,C,D\n3.1 (1.0)A,B,D\n4.1d (0.8)A,B,C\n233.642\n3\nP\u2009<\u20090.001\n\u2002Trusts other energy \ncompanies (mean (s.d.))\n3.2 (0.7)B,C,D\n2.6c (0.8)A,D\n2.6c (0.8)A,D\n3.7d (0.8)A,B,C\n225.853\n3\nP\u2009<\u20090.001\n\u2002Willing to share info  \n(mean (s.d.))\n3.3 (0.8)B,C,D\n3.1 (0.8)A,C,D\n2.3c (0.9)A,B,D\n3.9d (0.7)A,B,C\n249.425\n3\nP\u2009<\u20090.001\nPerceived ease of use\n\u2002SBS (mean (s.d.))\n0.7d (0.2)B,C,D\n0.6 (0.2)A,C\n0.4c (0.2)A,B,D\n0.6 (0.2)A,C\n8.906\n3\nP\u2009<0.001\n\u2002New electrifier (mean  \n(s.d.))\n0.5 (0.2)B,C,D\n0.5 (0.2)A,C,D\n0.3c (0.2)A,B,D\n0.6d (0.2)A,B,C\n165.999\n3\nP\u2009<\u20090.001\n\u2002ESC (mean (s.d.))\n0.4 (0.2)B,C,D\n0.4 (0.2)A,C,D\n0.3c (0.2)A,B,D\n0.6d (0.2)A,B,C\n136.260\n3\nP\u2009<\u20090.001\n\u2002P2P (mean (s.d.))\n0.4 (0.2)C,D\n0.4 (0.2)C,D\n0.2c (0.2)A,B,D\n0.6d (0.2)A,B,C\n135.804\n3\nP\u2009<\u20090.001\n\u20023PC (mean (s.d.))\n0.5 (0.2)B,C,D\n0.4 (0.2)A,C,D\n0.3c (0.2)A,B,D\n0.6d (0.2)A,B,C\n115.045\n3\nP\u2009<\u20090.001\nContinued\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n353\n\nArticles\nNaTurE EnErgy\nlikelihood-to-adopt scale. Participants were asked: \u201cIf this option was \navailable today, what is the likelihood that you would sign up for it?\u201d. \nResponses were made using a sliding scale from 0 (not at all likely) to \n1 (very likely), which was then banded to create a three-point scale \n(likely, neutral or unlikely). Figure 2 shows that the control case SBS \n(see Table 1) performed best, followed by P2P, with energy service \ncompany (ESC) showing the lowest overall attractiveness.\nConsumer segments\nWe then identified energy consumer segments based on consum-\ners\u2019 likelihood of adopting the new energy contracts. Segmentation \nstudies have been used to identify different types of consumer in \nrelation to household electricity storage50, acceptance of demand \ncontrol51, energy conservation behaviours52 or acceptance of smart \ngrids. Some studies have used measured or metred consumption \ndata combined with only basic household and building character-\nistics to identify groups of energy-consuming households53. Others \nhave used a richer suite of variables, based on qualitative surveys \nand constructs from behavioural psychology, to capture the psycho-\nsocial antecedents of likely market acceptance and targets for behav-\niour change52. The bases used to cluster the sample populations in \nthese studies were multifarious, ranging among general values, life-\nstyles, general patterns of energy-consuming behaviours, attitudes \nto environmental issues and specific energy-related behaviours. \nSome focused the clustering on the antecedents of behaviour only43, \nwhereas others mixed both the antecedents and the intentions or \nbehaviours themselves52, as we have done here.\nThe cluster analysis identified four groups individually repre-\nsenting between 16 and 35% of the sample. The groups were profiled \nusing the variables used to create them as well as other factors, such \nas demographic characteristics and current energy use. Profiling \nconsisted of characterizing each segment individually using descrip-\ntive statistics as well as comparisons with the other segments using \nmeasures of variance and association. Each of these segments was \ngiven a short name and a representative narrative statement (Fig. 3).\nSegment archetype preferences\nIn a paired comparison task, respondents were forced to choose \none or the other archetype in each of the ten paired cases (order \nrandomized). Each archetype was presented using the short para-\ngraph in Table 1, with two paragraphs displayed side by side. While \nthis method means that the key attributes are less directly compa-\nrable than can be the case with some forms of conjoint analysis, it \nwas designed to promote engagement with the task by requiring \nidentification and consideration of salient characteristics of each  \nbusiness model.\nFigure 4 shows the number of times each business model was \nchosen as a proportion of the number of times it was available to be \nchosen (out of four eligible times for each archetype for each per-\nson). This unweighted probability shows that there was a statistically  \n(A) Engaged but \ncautiousa\n(B) Aspiring \nopt-outsa\n(C) Unconvinced \nand unmotivateda\n(D) Pragmatic \ninnovatorsa\nX2/F valueb\nd.f.\nP\u2009value\nPerceived usefulness\n\u2002Perceived savings (mean \n(s.d.))\n2.4c (0.9)B,C,D\n3.5d (0.8)A,C\n3.1 (1.0)A,B,D\n3.5d (0.9)A,C\n158.602\n3\nP\u2009<\u20090.001\n\u2002Less energy to save money \n(mean (s.d.))\n3.9 (0.7)B,C\n4.2 (0.7)A,C\n3.5c (0.8)A,B,D\n4.1d (0.8)C\n93.795\n3\nP\u2009<\u20090.001\n\u2002Less energy for environment \n(mean (s.d.))\n3.6 (0.9)C,D\n3.7 (0.9)C\n2.7c (1.0)A,B,D\n3.9d (0.9)A,C\n132.611\n3\nP\u2009<\u20090.001\n\u2002Smart control benefits  \n(mean (s.d.))\n3.2 (1.0)C,D\n3.0 (1.0)C,D\n2.0c (1.1)A,B,D\n4.1d (0.8)A,B,C\n280.458\n3\nP\u2009<\u20090.001\nIntention\n\u2002Weighted probabilitye: SBS \n(mean (s.d.))\n0.56d (0.27)B,C,D\n0.45 (0.26)A,C,D\n0.27c (0.24)A,B\n0.32 (0.23)A,B\n158.955\n3\nP\u2009<\u20090.001\n\u2002Weighted probabilitye: new \nelectrifier (mean (s.d.))\n0.20 (0.18)C,D\n0.22 (0.19)C,D\n0.10c (0.14)A,B,D\n0.27d (0.20)A,B,C\n160.901\n3\nP\u2009<\u20090.001\n\u2002Weighted probabilitye: ESC \n(mean (s.d.))\n0.15 (0.20)C,D\n0.12 (0.17)C,D\n0.07c (0.13)A,B,D\n0.31d (0.25)A,B,C\n179.495\n3\nP\u2009<\u20090.001\n\u2002Weighted probabilitye: P2P \n(mean (s.d.))\n0.22 (0.24)B,C\n0.32d (0.27)A,C,D\n0.14c (0.19)A,B,D\n0.25 (0.23)B,C\n129.876\n3\nP\u2009<\u20090.001\n\u2002Weighted probabilitye: 3PC \n(mean (s.d.))\n0.13 (0.17)C,D\n0.14 (0.17)C,D\n0.06c (0.11)A,B,D\n0.29d (0.24)A,B,D\n170.559\n3\nP\u2009<\u20090.001\n\u2002Spread of scores (mean \n(s.d.))\n0.55d (0.26)B,C,D\n0.49 (0.26)A,C,D\n0.30 (0.24)A,B\n0.29c (0.21)A,B\n145.771\n3\nP\u2009<\u20090.001\n\u2002Consistency of scoring \n(mean (s.d.))\n0.25c (0.64)D\n0.34 (0.80)D\n0.31 (0.72)D\n1.27d (1.45)A,B,C\n114.550\n3\nP\u2009<\u20090.001\nCluster size (n)\n706\n537\n449\n330\nCluster share (%)\n35\n27\n22\n16\naThe homogeneity-of-variance test was met in each case to apply Tukey\u2019s post-hoc test alongside the ANOVA test. Different upper-case letters indicate statistically significant differences between specific \nsegments (as labelled A\u2013D in the header) using this test (P\u2009<\u20090.05). bOne-way ANOVA, producing an F statistic or X2 value, was conducted to evaluate the relationship between cluster membership and \neach variable. cValue of the segment with the lowest score for each segmentation variable. In all cases, the lower the mean score, the lower the agreement on each statement. For an explanation of each \nvariable name and the measurement scale used, see Supplementary Table 3. dValue of the segment with the highest score for each segmentation variable. In all cases, the higher the mean score, the greater \nthe agreement or value on this construct or issue. eWeighted probability refers to the probability score (proportion of times chosen (0\u20131)) multiplied by the stated likelihood-of-adoption score (0\u20131).\nTable 2 | Profile statistics for each segment on demographic characteristics and clustering variables (Continued)\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n354\n\nArticles\nNaTurE EnErgy\nsignificant difference between at least two of the segments for all \nbut the new electrifier archetype. Table 2 provides these prob-\nability variables weighted by the stated likelihood-to-adopt score, \nto account for the fact that the paired comparison exercise forces \npeople to make a decision of some kind. Here, there are differ-\nences between at least two of the segments on all of the archetypes. \nThis shows that some segments are more discriminating across the \narchetypes than others.\nThe status-quo option (SBS) is the standout preference of three \nof the segments, although engaged but cautious people are still sub-\nstantially more likely than all other groups to choose this option. \nThis segment demonstrates the joint-highest tendency to switch \nutility providers now and achieve high levels of satisfaction and \nengagement by doing so. They therefore represent an engaged and \ninformed group of consumers, but do not perceive the current sys-\ntem to be problematic enough to need to change.\nP2P stands out as being the second most favourable option for \nthree out of four of these groups. The aspiring opt-outs are the most \nenthusiastic about this archetype (and the least in favour of the \nESC), motivated by a chance to break free from large utility compa-\nnies and achieve greater cost savings.\nAlthough the differences in preference for the new electrifier \narchetype are less strong and more dependent on which measure-\nment of likelihood to adopt that we use, this option is neverthe-\nless consistently perceived more favourably than ESC for all but \nthe pragmatic innovators, for whom it is no more or less acceptable \nthan any other. The long contract term involved in the ESC seems \nto be off-putting to both those who are less trusting of large energy \nsuppliers (unconvinced and unmotivated or aspiring opt-outs) or \nthose who are more trusting but still want to take some control and \nlook for the best deals when they can (engaged but cautious or prag-\nmatic innovators).\nThe pragmatic innovators have the highest appetite for new types \nof contract overall, although they have a lower tendency than the \nother three groups for the P2P solution. The members of this seg-\nment are already markedly more likely to have adopted new tech-\nnology, such as solar photovoltaics or electric vehicles, and express \ngreat faith in scientific and technological solutions. However, \nthey are shown to have greater levels of trust in larger companies \nthan all of the other segments, and feel less convinced that there \nare large financial savings to be made from choices about energy \nsuppliers. This might explain why the greater level of hands-on \ndecision-making involved in P2P is not regarded by pragmatic \ninnovators as so necessary to overcome trust and financial issues as \nit is by the aspiring opt-outs or even the unconvinced and unmoti-\nvated. While the unconvinced and unmotivated appear marginally \nmore receptive to this option than both the engaged but cautious \nand pragmatic innovators, when the scores are weighted by the \nstated likelihood to adopt (Table 2), their tendency to say they are \nunlikely to adopt any of the models means they are the least enthu-\nsiastic for all of the options.\nDemographic, experiential and attitudinal characteristics\nTable 2 displays characteristic variables for each segment, orga-\nnized by the constructs of the conceptual model in addition to \nsome demographic variables that were not used in the segmenta-\ntion. Figure 5 focuses on one construct that proved to be one of the \nstrongest predictors of segment membership. Out of the 23 predic-\ntor variables of the segments identified using discriminant analysis, \ntrust in one\u2019s existing supplier is ranked fifth and trust in other sup-\npliers is ranked sixth. Figure 5 shows further differences between \nthe segments on trust of additional societal institutions using a \nquestion that has been used to understand comparative switch-\ning behaviour in a number of different markets and to set this in  \ncontext of a consumer\u2019s general trust tendency54. In Table 3, we \nbring together the analysis of all of the characteristics to provide \na brief summary profile of each segment and the implications for \nenergy market transitions.\n0\n1\n2\n3\n4\n5\nOwn energy company\nOther energy companies\nMobile phone providers\nInsurance companies\nBanks\nBroadband providers\nCharities\nCar manufacturers\nTrain companies\nPolice\nEngaged but cautious\nAspiring opt-outs\nUnconvinced and unmotivated\nPragmatic innovators\nFig. 5 | Trust across societal institutions by consumer segment. For each institution, we asked the question: \u201cTo what extent do you trust or distrust \nthe following types of organization to treat you in a fair and honest way?\u201d. This was measured on a five-point scale from 1 (distrust strongly) to 5 (trust \nstrongly). Using ANOVA, the segments differ significantly on all of the ten institutions at P\u2009>\u20090.0001. Supplementary Table 5 shows the mean values, s.d., \nANOVA and Tukey\u2019s post-hoc tests for these ten comparisons.\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n355\n\nArticles\nNaTurE EnErgy\nTable 3 | Summary profiles of each segment and implications for energy market transitions\nSummary profile\nImplications for energy market transitions\nEngaged but \ncautious \n(35%)\n\u2022 These consumers are engaged in their energy and are the most likely to shop around \nand switch supplier, finding it relatively straightforward to understand tariff information \nand calculate potential savings. They have high trust in their own and other suppliers and \nhigh willingness to share information with them. However, because they feel satisfied that \nthey can get what they want out of the current market and that they are paying the least \nfor their energy, they have the highest satisfaction levels and are not enthusiastic about \nalternative business archetypes.\n\u2022 They appear more motivated to get the best price for their energy than to save money \nthrough demand reduction or by being early adopters of alternative energy sources. Their \nmotivations for being price conscious do not appear to be driven strongly by affordability \nor fuel poverty.\n\u2022 They are the most environmentally conscious and most consistently minded towards \ngreen energy, although they are cautious about paying more. Although they are \nenvironmentally motivated, they do not want to work very hard for this. For instance, they \nare the most likely to admit that they could reduce their energy demand, despite being \nhighly likely to say that reducing energy would make them feel good. However, they are \nclear about not being overly willing to spend time thinking about the energy they use, \nand this is reflected in them being unconvinced by smart meters and not standing out for \nundertaking many other energy-saving behaviours.\n\u2022 Tendency to be older, female, childless, educated, in employment and owner occupiers \nwith an average income.\n\u2022 This group stands out as the highest scorer \non SBS, with only moderate rankings of all \nof the other archetypes. Thus, although they \nare motivated to shop around and inform \nthemselves about tariffs, their satisfaction \nwith the status quo appears to lead to \nresistance towards new business models\n\u2022 Importantly, the two segments with \nabove-average owner occupation status \n(engaged but cautious and unconvinced and \nunmotivated) are the least likely to choose \nthe two archetypes that require alterations \nto building fabric (that is, ESC and 3PC).\nAspiring \nopt-outs \n(27%)\n\u2022 These consumers are the least likely to have switched supplier, despite the lowest \nsatisfaction and the highest tendency to say they have thought about switching. Those \nwho have switched have the greatest tendency to say that they found it difficult and \nbelieve tariffs to be deliberately confusing, with little financial gain from one company \nor tariff to another. They are the most likely to say that they find their bills difficult to \nunderstand. This is also reflected in them showing the lowest levels of trust in their own \nand other suppliers and the greatest resistance to suppliers controlling appliances.\n\u2022 They do have some tendency to think about their energy use and do show the highest \nengagement in energy saving, as they are very motivated to save money. They are paying \nabove average for their energy bills, yet are by far the least likely to say they are able to \nkeep warm in winter, with a high proportion citing cost as the reason. Despite being the \nmost worried about keeping warm in winter, they are by far the most likely of all of the \nsegments to say they would cut down on the amount of energy used if prices were to rise \nby 20%.\n\u2022 They are only moderately environmentally concerned and have the lowest faith in \nscience and technology, not considering themselves to be early adopters of technology. \nThey are reluctant to pay more for environmental gain.\n\u2022 Tendency to be younger, female, with children and less educated and to have average \nemployment, low income and high rent.\n\u2022 Highest scorer for P2P compared with \nother segments, although this segment as a \nwhole scores highest for SBS and lowest for \nESC and 3PC.\n\u2022 It is interesting that those who currently \nhave the greatest difficulty navigating \nthe energy market would want to take \nresponsibility for a more interactive market \nexperience. However, it makes sense in that \nthey do not trust the larger suppliers and \nthis would enable them to opt out of their \ncontrol.\n\u2022 Their low trust is likely to have led to ESC \nbeing their least favoured archetype, as this \nwould involve being tied in for a long period. \nNew electrifier is their third most preferred \noption, perhaps based on their expressed \ndesire for improvements to their home \nheating systems.\nUnconvinced \nand \nunmotivated \n(22%)\n\u2022 This group are indifferent about their energy use and seem to have paid very little \nattention to it in terms of their supply or their own behaviour.\n\u2022 Although this group have switched suppliers in the past at an average frequency for \nthis sample and they are not particularly likely to say that it is too hard to switch, they \nhave low satisfaction with their current supplier and are the least likely to know details \nsuch as the tariff they are on or the insulation level of their home, or to believe that they \nunderstand their energy.\n\u2022 They are not motivated, by money savings or climate change, to change their behaviour, \nand are the least likely to be paying attention to any deals or their own energy use. They \nare the least likely to say they are thinking about electricity use now and want to think \nmore about it in the future or that they want to change anything about their home energy \nsystem. They are the least likely to have a smart meter or to say they want one, with the \nhighest likelihood to say that this kind of information is not at all helpful and the lowest \nlikelihood (8%) of saying they want one. They have the lowest willingness to share their \nenergy data with others or to have suppliers control their appliances.\n\u2022 They consider themselves to be late adopters of technology.\n\u2022 This group are the oldest and have a tendency to be out of work or retired, male and \nchildless, with low education and slightly below-average income, although they are more \nlikely to be owner occupiers.\n\u2022 Although P2P is the second most popular \narchetype after SNS for this group, their \ntendency to say they are unlikely to actually \nadopt any of them gives them the lowest \nlikelihood of choosing any of the business \nmodels. They consistently rate all of the \narchetypes as complicated, with the most \nextreme scores on this measurement of \nall the segments. SBS is rated as the least \ncomplicated and P2P as the most.\n\u2022 This group of consumers appear to be \ndifficult to hook into any alternative models \nas they are not engaged in their energy \nuse or particularly motivated by cost or \nenvironmental benefits. Even though it \nwould seem that they do not want to put \nthe effort into anything that involves them \nhaving to think much about their supply, \nthey are also very negative towards the idea \nof their data being shared or appliances \nbeing controlled.\nContinued\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n356\n\nArticles\nNaTurE EnErgy\nDiscussion\nThese data show that there is some consumer demand for innovative \nretail energy contracts. Although it is clear that the business archetype \nthat represents only an incremental change from the status quo (SBS) \nhas proven the most popular, there are clear indications that different \nconsumers would be attracted to some other types of contractual rela-\ntionship. Even the least innovative segment in our sample (engaged \nbut cautious) rejects the SBS model 15% of the time, and this increases \nto 45% for the most innovative segment (pragmatic innovators). \nThese data show some desire for tying in new services, appliances and \neven building works into an energy bill. This goes well beyond what \nenergy retail market regulation was initially designed to achieve55.\nAs flexible tariffs, home retrofit/appliance bundles and potential \nP2P trading are all now technically possible56, and these data show \nthey each can find a compatible consumer segment, there is sub-\nstantial potential for market disruption.\nGiven that the UK retail market already experiences regulatory \nchallenges from expanding conventional competition55, this latent \ndemand for innovative business models complicates the regulatory \ntask further and invites a policy and regulatory response. From \nthese data, we extract three challenges, as outlined below.\nPragmatic innovators show a high preference for two utility \nbusiness model archetypes that are radical departures from the cur-\nrent utility model: ESC and third-party control (3PC). Importantly, \nthe two segments with above-average owner occupation status \n(engaged but cautious (66% owned outright or mortgaged) and \nunconvinced and unmotivated (72%)) are the least likely to choose \nthe two archetypes that require alterations to building fabric (that \nis, ESC and 3PC). This is a problem when these business models \nrely on changes to the building fabric that are paid for by long-term \nenergy bill savings. UK housing tenure trends show a reduction in \nowner occupation and in increase in private rented tenure, espe-\ncially for younger demographics57. Since both pragmatic innovators \nand aspiring opt-outs are younger than the sample average, and of \nbelow-average or decreasing likelihood to own their own home, this \nmeans there is a disconnect between the segments that find such \nbusiness models attractive and their ability to sign up to them when \nin rented tenure.\nFor aspiring opt-outs, who are younger, have lower income and \nare probably renting, the opportunity to benefit from their preferred \narchetype of P2P may be limited as they are the least likely to have their \nown microgeneration to trade. With an average of 36% of the whole \nstudy sample in rented tenure, a large proportion of consumers are \nunable to access many of the benefits of new utility energy contracts.\nTrust in one\u2019s own and other suppliers is a significant driver of \nsegment membership. However, it may not be trust in the utilities \nalone that drives segments to prefer different business models, but \nalso trust across societal institutions. Other recent work has pointed \nto the role of trust and legitimacy between consumers and the \nevolving retail energy market58. However, the data reported here \nshow remarkably consistent trust scores across societal institutions \nby segment, in sectors as diverse as broadband provision, insurance, \nbanks and car manufacturers.\nWhile societal legitimacy and institutional trust is important \nfor energy transitions59, this work shows that the sector may make \nmore progress on business model innovation by targeting segments \nthat already trust the system, as opposed to working to change the \nperceptions of aspiring opt-outs or unconvinced and unmotivated \nconsumers. The low trust levels displayed by these segments may \nbe relatively fixed across institutions, resistant to trust messages \nfrom the energy community, and not entirely unfounded, given \nthe Competition and Markets Authority found that UK utilities \nwere serially overcharging less engaged consumers by hundreds of \npounds per year7.\nWe may expect utilities to focus new tariffs and services towards \npragmatic innovators and aspiring opt-outs, who make up 43% of \nthe sample and are much more likely to choose a new type of utility \nbusiness model in a future switching decision. Further investigation \nof the pragmatic innovators shows they are also amenable to direct \nload control of appliances and most likely to already have (or to pur-\nchase in the next 10\u2009years) a plug-in electric car. This segment also \nreports the highest income; segments with higher income find this \ninnovation most attractive and are likely to have the highest ability \nand desire to purchase smart, flexible appliances and vehicles first. \nTherefore, to mitigate the risk that the monetary benefits of flexible \nelectricity tariffs are most likely to be captured by higher-income \nSummary profile\nImplications for energy market transitions\nPragmatic \ninnovators \n(16%)\n\u2022 This group of consumers are engaged and potentially active but discerning.\n\u2022 They have the joint-highest experience of switching suppliers so far, have very high \nawareness and understanding of their energy tariff and are the most likely to be actively \nundertaking energy-saving activities now and spending time thinking about their electricity \nuse. They are the most likely to say their homes are already insulated and to have a smart \nmeter, and they are by far the most likely to have a source of heat or electricity individual \nto the dwelling, such as solar panels. They are also the most likely to own alternatively \nfuelled cars. Thus, they are early adopters and perceive themselves as such.\n\u2022 Interestingly, they have high trust in energy suppliers and do not believe that switching \nsuppliers brings much benefit because they are too similar. This means they have the \ngreatest preparedness to share information with these suppliers and for suppliers to \ncontrol appliances.\n\u2022 They have a complex view about the environment. They say they are motivated to do \nsomething about it and believe that issues are urgent, but do not believe people\u2019s freedoms \nshould be curtailed and that the problems may be slightly exaggerated.\n\u2022 They do believe that reducing energy use can save money and be environmentally \nbeneficial, and they are the most likely to say that reducing energy consumption would \nmake them feel good. However, they are also the most likely to say they could not reduce \ntheir energy use any further.\n\u2022 This group are the youngest, with few above 55\u2009years. They are both men and women, \nwith a tendency to have children. They have the highest education, highest income and \nhighest employment and tend to be owner occupiers.\n\u2022 Fairly equal choices among the options. \nHighest of all of the groups on 3PC, which \nmay have appeal across the variety of \nlifestyle services involved in a working family \nhome with some innovation tendencies.\n\u2022 Largest number of reversals, showing some \nambivalence in choice. Thus, this groups \nappears the least wedded to the status quo \nand can see merit in a number of different \nsolutions.\nSee Supplementary Table 4 for a longer version and the Supplementary Data for background response scores.\nTable 3 | Summary profiles of each segment and implications for energy market transitions (Continued)\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n357\n\nArticles\nNaTurE EnErgy\ngroups, groups such as the aspiring opt-outs, who also desire flex-\nibility and autonomy but more specifically the ability to manage \ntheir more limited income more freely, need to be targeted directly \nand given the means to act on their preferences. This segment \nexpresses the greatest preference for the P2P archetype, yet has the \nleast understanding of current bills, as well as the lowest trust scores \nand lowest incomes. This highlights a particular risk as P2P offers \nmay only be beneficial to flexible, active consumers or those with \nmicrogeneration to trade. While the current energy market suffers \nfrom inequitable outcomes for unengaged consumers, the types \nof utility business model innovation explored here may only risk \nentrenching poor distributional outcomes.\nHere, we have shown that some consumers do want the types \nof business model being developed by utilities in response to the \npressures of existing low-carbon energy policies. While the entry \nof these new types of contractual relation into the energy market \nmay be disruptive by itself, it also poses at least three specific chal-\nlenges to the UK sector, which under similar demographics and \nmarket conditions may be present in other liberalized, decarbon-\nizing power sectors. First, there is the potential for market innova-\ntion to stall due to the most receptive segments being most likely \nto rent and the most disengaged segments being most likely to own \nhomes and have the power to opt for contracts that include build-\ning alteration. Second, there is a social trust barrier to overcome \nin the energy sector, with low confidence across societal institu-\ntions for some segments. Finally, there is a real risk that the cul-\nmination of these issues could lock out some sectors of society, \nincluding low-income, low-information and renting demograph-\nics, from participating in low-carbon transitions. As the market \ndiversifies and contracts become more complex, consumers may \nrely even more on heuristics to make decisions, introducing more \ncomplex consumer risks. The challenge for regulatory institutions \nis to recognize these risks and evolve the regulatory model of the \nretail market.\nMethods\nSurvey design. The research aim for the questionnaire was to elicit the preferences \nof domestic energy consumers across a set of business archetypes and to identify \na rich set of possible explanatory factors for these choices. The explanatory factors \nincluded in the study were based on the technology assessment model combined \nwith a literature review, and detailed multiple situational, demographic and \nattitudinal antecedents of engagement with electricity and energy usage in the \nhome, including energy efficiency and tariff-switching behaviours43,60\u201362. These \nfindings were used to inform the design of the survey, including some batteries \nof questions already tried and tested in the UK context in the examination of \ndomestic energy usage61,63\u201365. The novelty in the present study is the profiling \nof potential future consumer groupings based on a theoretically underpinned \nconceptual framework of individual motivations involved in opting for different \nways of engaging with energy supply arrangements.\nThe questionnaire contained nine sections consisting of: (1) domestic \narrangements, including housing type and occupancy; (2) supply of energy to \nthe home, including billing, tariff and expenditure, and comprehension of these \nelements; (3) heating methods and fuel, including perceived levels of insulation \nand satisfaction; (4) perceived comfort and affordability, concern for and likely \nresponse to future energy prices, experience with smart metering and demand \ncontrol; (5) reasons for choice of supplier, switching behaviour, satisfaction and \ntrust in energy suppliers and other types of service provider; (6) business archetype \nattribute rating, paired choice experiment and overall likelihood to adopt each \nmodel; (7) current car ownership and usage, willingness to adopt electric vehicles \nand shared mobility; (8) attitudes towards different forms of energy generation, \nenergy supplier regulation and environmental beliefs/concerns; and (9) individual \nand household sociodemographic characteristics.\nAddressing limitations of hypothetical choices. Eliciting beliefs about and \npreferences towards a set of products that are not currently available and also \nrepresent an area of consumption (domestic energy use) in which consumers tend \nto be minimally engaged is a serious research challenge. Three principal strategies \nwere used to mitigate this:\n\t1.\t\nSampling of householders at least partially responsible for (and there-\nfore most likely to be engaged with) energy supply in the household (see \n\u2018Sampling\u2019)\n\t2.\t\nBuilding up engagement with the issues successively through the question-\nnaire by starting with factual questions about current arrangements for en-\nergy supply against which people can mentally benchmark future archetypes\n\t3.\t\nInclusion of topic areas that have been found by other studies to result in a di-\nversity of opinion relating to trust in organizations, environmental behaviour \nand cost of energy\nAltogether, the questionnaire took a minimum of 15\u2009min to complete, with a \nmedian duration of 30\u2009min.\nData collection. The study employed a market research company (Accent) \nto programme the online survey and organize the data collection. Accent \nspecialize in online stated-preference surveys with randomization and have \naccess to the Survey Sampling International (SSI) global market research panel of \ndemographically diverse adults (>18\u2009years of age) who have voluntarily subscribed \nto undertake such research. People register by providing varying amounts of \npersonal and demographic information that is later used to select participants for \nspecific surveys. Thus, as is typical with such panels, the SSI panel does not use \nprobability-based recruitment. Nevertheless, the panel from which this study\u2019s \nparticipants were selected was large enough to enable the selection of a nationally \nrepresentative sample or sample representative of sub-groups that reflect the actual \nbreakdown of their key demographics (for example, age, gender, region, social \ngrade, ethnicity, disability and so on).\nParticipants were sent an electronic communication (email or through a \nphone app) asking them to participate, and were rewarded with a small incentive \n(approximately \u00a30.85) for completion of the survey. The survey underwent a pilot \ntest in March 2017, during which 66 participants completed this version. This led \nto some removal of attitudinal questions eliciting a disproportionate number of \nneutral responses. The final questionnaire was in the field between 30 March and \n13 April 2017.\nSampling. The aim was to obtain a representative sample of electricity bill payers \nin England, Wales and Scotland. To screen for such people, we used the question \n\u201cWhich one of the following describes your level of involvement in decisions about \nwhich company your household uses to supply gas and/or electricity?\u201d, to which \nparticipants could respond (1) \u201cIt is my responsibility entirely\u201d; (2) \u201cI have equal \nresponsibility with someone else in the household\u201d; (3) \u201cI have some involvement \nin the decision\u201d; or (4) \u201cI have no involvement at all\u201d. If a participant answered with \nthe fourth option, the survey was terminated.\nAs there are no reliable statistics against which to benchmark our achieved \ndemographic profile of bill payers to achieve representativeness among them, \nwe used a proportionate stratified sampling approach to mirror the general \npopulation66, with quotas based on age (six groups), gender and residential location \n(11 government office regions). Thus, the sample itself was made up of bill payers \nonly, but unlikely to be entirely representative of them. For example, the number of \n18- to 24-year-old respondents had to be specifically boosted as a disproportionate \nnumber of this age group were otherwise consistently screened out due to being \nless likely to fall into this group. However, given that the focus of this research \nrequires future-facing preferences, over-representing this group was justified.\nSSI sent a total of 41,579 invitations, from which 3,552 surveys were started and \n2,090 were completed (that is, a 5% response rate and a 59% completion rate). While \nthis appears low as a response rate, it was not possible to calculate one accurately \ndue to the need to consider the incidence rate (the proportion of respondents \ncontacted who qualified for the survey) and the fact that the survey closed as soon \nas the required sample size had been achieved. Once full quotas were considered, \nthe incidence rate was 57%. The final achieved sample was 2,024 respondents, \nsince some were removed (n\u2009=\u200966) who were deemed to have provided incomplete \nor invalid data based on straight liner variables detected to indicate whether \nparticipants had given the same response for all sub-questions in a relevant block.\nParticipant characteristics. As discussed, our sample comprised only domestic \nenergy bill payers, but with their proportions set through quota sampling to match \nthe characteristics of the general population. Comparisons with national data \nfor England, Wales and Scotland67\u201369 suggested that, compared with the general \npopulation, participants: were somewhat better educated; were more likely to be on \nhome duties; included higher shares of people among the higher earners (>\u00a340,000 \nper household per year); and had slightly higher car accessibility. In addition, more \nowned outright (no mortgage), but also, more rented their households. Otherwise, \nthe study population was largely representative in its demographic, socioeconomic \nand geographic characteristics. Therefore, no post-stratification weighting was \napplied. See Supplementary Table 1 for details.\nThe sample was segmented using hierarchical cluster analysis (HCA) followed \nby non-HCA. The clustering variables were those representing the constructs in \nthe conceptual model, which are a compilation of beliefs, along with experiential \nand preference factors. Note that we chose to include a measure of likelihood to \nadopt among the clustering variables. The mixing of generic and specific attitudinal \nvariables in this way has a solid conceptual justification. If based only on general \nantecedent constructs to intention (that is, not including the likelihood variable), \nthe segmentation model could fail to identify that there is more than one group \nexhibiting similar core beliefs or psychological processes but expressing different \nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n358\n\nArticles\nNaTurE EnErgy\nintentions or behaviours due to important contextual or experiential antecedents. \nThe converse may also be present, whereby groups with similar intentions may \nbe present, but with markedly different beliefs or experience. If only one or the \nother type of variable were put into the segmentation, there would be a risk that \nsuch multifaceted attitude\u2013behaviour linkages could be hidden among generic \ngroupings. It is important to realize that we have not developed a segmentation \nmodel here that we claim will stand the test of time. The purpose is to understand \nnow what the appetite for change is, for what change and by whom. We therefore \nbelieve that our model can serve as a practical tool that offers a robust building \nblock for the development of strategy in the initial market evolution of new utility \nbusiness models.\nIdentification of preference factors. The first stage in identifying variables \nto use as the bases for the cluster analysis was the development of measures of \npreference for the five business archetypes. Five measures were derived from the \nvarious questions on the survey. (1) Likelihood to adopt each archetype, using a \nsingle question asked after each archetype, was scored on a number of sematic \ndifferential attribute scales. (2) Adoption spread was the highest value of likely \nminus the lowest for each individual to generate a signal of preference certainty. If \na person scored one or more options very highly in terms of likelihood to adopt, \nand another option very low, they would have a larger spread than someone who \nscored all options similarly. (3) Probability of adoption was the number of times \nchosen during the paired comparison experiments divided by the number of times \navailable to be chosen. (4) Weighted probability of adoption was calculated by \nmultiplying the probability of adoption score by the likelihood score. (5) Preference \nstability was a test of internal consistency calculated by using an excel macro to \ndetect reversals between each combination of three options and then adding up \nthe number per person. The majority of people were perfectly internally consistent \n(that is, 76% did not reverse their ranking of options across the three possible \ncombinations with each option).\nData reduction. The attitudinal variables on the survey were subjected to data \nreduction in order to reduce the variables to a smaller set of underlying dimensions \nto be used in the subsequent segmentation. In factor analysis, variables that showed \nsimilar patterns of variation across respondents were assumed to be associated with \nthe same underlying construct. Principal axis factoring was used in SPSS Statistics \nversion 22.0 (IBM), chosen to account for some non-normal distribution in the \ndata. Rotation of the final solution was necessary to clarify the underlying structure \nand produce a set of arbitrary factors that provide the clearest conceptual picture \nof the relationships among the items70. A direct oblimin rotation was used as this \nmaximizes the variance of the loadings within factors across variables so that each \nof the original items loads on only one factor. Oblimin rotation also allows the \nfactors to correlate, which avoids unnecessary loss of information in orthogonal \nmethods and thus leads to more reproducible solutions71.\nThe three batteries of questions, each subject to the factor analysis \n(amounting to 30 statements in total, each measured on a five-point scale \n(usually, \u201cstrongly agree\u201d to \u201cstrongly disagree\u201d)), were: (1) attitudes towards \nthe environment and renewable energy; (2) general approaches to purchase \ndecision-making; and (3) process of choosing energy suppliers. Each of these was \nsubject to a series of analyses that were run iteratively, each time excluding items \nwith low communalities (h\u2009<\u20090.5). Communalities identify the items\u2019 variance \nand thus the ones that form highly consistent scales that discriminate well in the \nclustering procedure72. Each component was subject to a reliability test using \nCronbach\u2019s \u03b1, which measures internal consistency based on item correlation. \nAlpha coefficients range from 0\u20131 and may be used to describe the reliability \nof each factor. A value of 0.5 is generally regarded as an acceptable reliability \ncoefficient73 and was used here.\nThe three batteries of questions were reduced to six latent constructs:  \n(1) green urgency; (2) green scepticism; (3) information seeking when purchasing; \n(4) inspiration seeking; (5) perceived savings; and (6) engagement with energy \nusage discerning trustworthiness and the quality of service suppliers. Six questions \nhad been discarded from the analysis as inspection of the correlation matrix \nrevealed that they were not significantly related to any other items in this set or \nwere found to be responsible for a lower \u03b1 value. These factors are a very valuable \nset of internally consistent constructs to be used in further analyses to understand \nconsumer perceptions and motivations. Factor scores, as opposed to summated \nscale scores, were computed for each respondent in the dataset to be used in the \ncluster analysis.\nSupplementary Table 2 itemizes the constructs, with the interpretive label, \nfactor loadings, Cronbach\u2019s \u03b1 score and percentage of variance explained in each of \nthe three batches of questions.\nCluster analysis. A two-stage cluster analysis common in market research74 was \nperformed to identify segments of potential consumers of the business archetypes. \nSegmentation market research begins with the assumption that there is little value \nin targeting the average customer and more value in treating different people in \ndifferent ways because they are motivated by varying rationales75. HCA is used first \nin an exploratory structure-seeking phase, followed by the iterative partitioning \nmethod (k-means) to fine-tune the analysis73.\nHCA was performed on a set of 31 standardized variables comprising those \ndetailed in Supplementary Table 3 and chosen to represent the constructs of \nthe conceptual model. The variables were standardized and subjected to HCA, \napplying Ward\u2019s method. The squared Euclidean distance was used as the \nproximity measure in the clustering procedure. This gave an indication of how the \nsample population was partitioning and hinted at the optimal number of clusters to \nbe used in the second stage. Following visual inspection of the HCA dendrogram \n(Supplementary Fig. 4), a range of cluster solutions were chosen (from three to six \nsegments) to be carried to the next step.\nThe k-means clustering took the cluster centres of the HCA cluster solution as \ninput and re-clustered the sample according to the squared Euclidean distance from \nthe centres. Since HCA does not correct cluster assignments, k-means can generate \nmore homogeneous groups and hence improved solutions whereby variability \nwithin clusters is minimized while maximizing the variability between them73. \nThe number of clusters was specified at three, four, five or six. Selection criteria \nsuggested in the literature ranged from highly subjective to complex mathematical \nprocedures. The agglomeration schedule (Supplementary Fig. 5)76 showed a jump \nin which the value of the error coefficient nearly doubled when four clusters were \nreduced to three, giving a strong indication that a four-cluster solution made \nthe most sense. This was confirmed by inspecting one-way analysis of variance \n(ANOVA) tables generated for the different solutions using the 31 variables used to \ncreate them to ensure high variability between groups compared with within them \nand the presence of discrete and concentrated clusters. This procedure indicated \nthat the four-cluster solution may be marginally better than both the three- and \nfive-cluster solutions, although significant and meaningful differences were found \nacross all 30 segmentation variables for all solutions. Since the four-cluster solution \nseparated clusters with particularly distinctive views on adoption likelihood, this \nsolution was selected. Two participants were not clustered due to some missing data \non key variables. Hence, the sample size for the cluster profiling was 2,022.\nCharacterizing the segments. The cluster analysis concluded that four relatively \nstable groups could be identified ranging from 16\u201335% of the sample. The groups \nwere profiled on the variables used to create them, as well as other factors such \nas demographic characteristics. ANOVA with Tukey\u2019s post-hoc tests were used \nfor means comparisons, and two-tailed chi-squared tests were used where we \ncompared relative frequencies of categorical variables across clusters.\nWe also undertook a discriminant analysis to test whether it was possible \nto define a set of variables that could predict group membership with adequate \nreliability without including the specific willingness-to-adopt (likelihood) variables \nin the analysis. The discriminant analysis confirmed that segment membership \ncould be predicted by only using the antecedent model constructs used to form \nthe clusters, as well as some demographic variables we had found to strongly \ndiscriminate across segments. As explained in Supplementary Note 1, a satisfactory \nsolution that would predict group membership with 80% accuracy was found using \n23 of the original 31 variables.\nEthics. The field-work proposal was reviewed and approved by the University of \nLeeds AREA Faculty Research Ethics Committee (AREA 16-155). The research \nwas undertaken in line with the requirements of the international quality standard \nfor market, opinion and social research (ISO 20252:2012), to which Accent is \nregistered and audited annually. Accent is registered under the Data Protection  \nAct 1998.\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nThe relevant survey data, including all raw data, generated or analysed during \nthis study are included in the Supplementary Data file. Data generated in the \nconstruction of business model archetypes are summarized in the Supplementary \nInformation. The data that support the plots within this paper and other findings of \nthis study are available from the corresponding author upon reasonable request.\nReceived: 18 April 2019; Accepted: 18 January 2021;  \nPublished online: 1 March 2021\nReferences\n\t1.\t Schot, J., Kanger, L. & Verbong, G. The roles of users in shaping transitions to \nnew energy systems. Nat. Energy 1, 16054 (2016).\n\t2.\t Future Energy Scenarios (National Grid ESO, 2020); https://www.\nnationalgrideso.com/document/173821/download\n\t3.\t Retail Markets Monitoring Report CEER Report C17-MMR-83-02 (Council of \nEuropean Energy Regulators, 2017): https://www.ceer.eu/documents/ \n104400/-/-/56216063-66c8-0469-7aa0-9f321b196f9f\n\t4.\t Horta\u00e7su, A., Madanizadeh, S. A. & Puller, S. L. Power to choose? An analysis \nof consumer inertia in the residential electricity market. Am. Econ. J. Econ. \nPolicy 9, 192\u2013226 (2017).\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n359\n\nArticles\nNaTurE EnErgy\n\t5.\t Sen, A., Nepal, R. & Jamasb, T. Reforming Electricity Reforms? Empirical \nEvidence from Asian Economies (Oxford Institute for Energy Studies, 2016); \nh\u00adt\u00adt\u00adp\u00ads\u00ad:\u00ad/\u00ad/w\u00adw\u00adw\u00ad.\u00ado\u00adx\u00adf\u00ador\u00add\u00ade\u00adn\u00ade\u00adr\u00adg\u00ady.\u00ado\u00adr\u00adg\u00ad/\u00adw\u00adp\u00adcm\u00ads\u00ad/\u00adw\u00adp\u00ad-\u00adc\u00adon\u00adt\u00ade\u00adnt\u00ad/\u00adu\u00adpl\u00ado\u00ada\u00adds\u00ad/\u00ad2\u00ad01\u00ad6\u00ad/\u00ad02\u00ad/\u00adR\u00adef\u00ado\u00adr\u00adm\ni\u00adn\u00adg\u00ad\u00ad-E\u00adl\u00ade\u00adct\u00adr\u00adi\u00adci\u00adt\u00ady\u00ad-R\u00ade\u00adf\u00ador\u00adms\u00ad-E\u00admp\u00adir\u00adical-Evidence-from-Asian-Economies-EL-18.pdf\n\t6.\t Van Beek, M, Holst, A & Keeble, J. Low Carbon, High Stakes\u2014Do You Have \nthe Power to Transform? (Accenture Strategy, 2015).\n\t7.\t Energy Market Investigation Final Report (Competition and Markets \nAuthority, 2016).\n\t8.\t The Road Ahead: Gaining Momentum from Energy Transformation (PwC, \n2014); https://www.pwc.com/gx/en/utilities/publications/assets/pwc-the- \nroad-ahead.pdf\n\t9.\t Bryant, S. T., Straker, K. & Wrigley, C. The typologies of power: energy utility \nbusiness models in an increasingly renewable sector. J. Clean. Prod. 195, \n1032\u20131046 (2018).\n\t10.\tHawker, G., Bell, K. & Gill, S. Electricity security in the European Union\u2014\nthe conflict between national capacity mechanisms and the single market. \nEnergy Res. Soc. Sci. 24, 51\u201358 (2017).\n\t11.\tParag, Y. & Sovacool, B. K. Electricity market design for the prosumer era. \nNat. Energy 1, 16032 (2016).\n\t12.\tBritton, J., Hardy, J., Mitchell, C. & Hoggett, R. Changing Actor Dynamics and \nEmerging Value Propositions in the UK Electricity Retail Market IGov Working \nPaper (2019); http://projects.exeter.ac.uk/igov/report-analysis-of-uk-electricity- \nsystem-actors/\n\t13.\tBurger, S., Chaves-\u00c1vila, J. P., Batlle, C. & P\u00e9rez-Arriaga, I. J. A review of the \nvalue of aggregators in electricity systems. Renew. Sustain. Energy Rev. 77, \n395\u2013405 (2017).\n\t14.\tWard, A., Thomas, N. & Chazan, G. European utilities primed for \nconsolidation in shift to renewables. Financial Times (27 June 2017).\n\t15.\tWorld Energy Investment (International Energy Agency, 2017).\n\t16.\tHall, S. & Roelich, K. Business model innovation in electricity supply \nmarkets: the role of complex value in the United Kingdom. Energy Policy 92, \n286\u2013298 (2016).\n\t17.\tRoelich, K. et al. Towards resource-efficient and service-oriented  \nintegrated infrastructure operation. Technol. Forecast. Soc. Change 92,  \n40\u201352 (2015).\n\t18.\tHelms, T. Asset transformation and the challenges to servitize a utility \nbusiness model. Energy Policy 91, 98\u2013112 (2016).\n\t19.\tHelms, T., Loock, M. & Bohnsack, R. Timing-based business models for \nflexibility creation in the electric power sector. Energy Policy 92,  \n348\u2013358 (2016).\n\t20.\tWhite, L. V. & Sintov, N. D.Inaccurate consumer perceptions of monetary \nsavings in a demand-side response programme predict programme \nacceptance. Nat. Energy 3, 1101\u20131108 (2018).\n\t21.\tEid, C., Codani, P., Perez, Y., Reneses, J. & Hakvoort, R. Managing electric \nflexibility from distributed energy resources: a review of incentives for market \ndesign. Renew. Sustain. Energy Rev. 64, 237\u2013247 (2016).\n\t22.\tFrei, F., Sinsel, S. R., Hanafy, A. & Hoppmann, J. Leaders or laggards? The \nevolution of electric utilities\u2019 business portfolios during the energy transition. \nEnergy Policy 120, 655\u2013665 (2018).\n\t23.\tWilliams, B. & Waring, G. Consumer Engagement In The Energy Market  \n2018: Report on a Survey of Energy Consumers (GfK, 2018); https://www.\nofgem.gov.uk/system/files/docs/2018/10/consumer_engagement_survey_2018_\nreport_0.pdf\n\t24.\tFlores, M. & Waddams Price, C. The role of attitudes and marketing in \nconsumer behaviours in the British retail electricity market. Energy J. 39, \n153\u2013179 (2018).\n\t25.\tHorta\u00e7su, A., Madanizadeh, S. A. & Puller, S. L. Power to choose? An analysis \nof consumer inertia in the residential electricity market. Am. Econ. J. Econ. \nPolicy 9, 192\u2013226 (2017).\n\t26.\tDefeuilley, C. Retail competition in electricity markets. Energy Policy 37, \n377\u2013386 (2009).\n\t27.\tWaddams Price, C. Back to the future? Regulating residential energy markets. \nInt. J. Econ. Bus. 25, 147\u2013155 (2018).\n\t28.\tLoewenstein, G. Experimental economics from the vantage\u2010point of \nbehavioural economics. Econ. J. 109, 25\u201334 (1999).\n\t29.\tFrederiks, E. R., Stenner, K. & Hobman, E. V. Household energy use: applying \nbehavioural economics to understand consumer decision-making and \nbehaviour. Renew. Sustain. Energy Rev. 41, 1385\u20131394 (2015).\n\t30.\tSimon, H. A. A behavioral model of rational choice. Q. J. Econ. 69,  \n99\u2013118 (1955).\n\t31.\tPichert, D. & Katsikopoulos, K. V. Green defaults: information presentation \nand pro-environmental behaviour. J. Environ. Psychol. 28, 63\u201373 (2008).\n\t32.\tDeller, D. et al. Fairness in Retail Energy Markets? Evidence from the UK \n(Centre for Competition Policy, 2018).\n\t33.\tShafir, E., Simonson, I. & Tversky, A. Reason-based choice. Cognition 49, \n11\u201336 (1993).\n\t34.\tHobman, E. V., Frederiks, E. R., Stenner, K. & Meikle, S. Uptake and usage of \ncost-reflective electricity pricing: insights from psychology and behavioural \neconomics. Renew. Sustain. Energy Rev. 57, 455\u2013467 (2016).\n\t35.\tEppler, M. J., Hoffmann, F. & Bresciani, S. New business models through \ncollaborative idea generation. Int. J. Innov. Manag. 15, 1323\u20131341 (2011).\n\t36.\tRohrbeck, R., Konnertz, L. & Knab, S. Collaborative business modelling for \nsystemic and sustainability innovations. Int. J. Technol. Manag. 63,  \n4\u201323 (2013).\n\t37.\tPieroni, M. P., McAloone, T. & Pigosso, D. A. Business model innovation for \ncircular economy and sustainability: a review of approaches. J. Clean. Prod. \n215, 198\u2013216 (2019).\n\t38.\tBurnet, F. Take Charge: An Analysis of the Domestic Electric Vehicle Tariff \nMarket (Citizens Advice, 2019); https://www.citizensadvice.org.uk/Global/\nCitizensAdvice/Energy/Take%20Charge%20-%20EV%20tariff%20report.pdf\n\t39.\tHahnel, U. J., Herberz, M., Pena-Bello, A., Parra, D. & Brosch, T. Becoming \nprosumer: revealing trading preferences and decision-making strategies in \npeer-to-peer energy communities. Energy Policy 137, 111098 (2020).\n\t40.\tRetail Research into Customer Switching and Supply Disintermediation: Final \nReport: Disintermediation (ESP Consulting, 2018); https://www.ofgem.gov.uk/\nsystem/files/docs/2018/07/retail_research_-_report_on_supply_\ndisintermediation.pdf\n\t41.\tBrown, D. Business models for residential retrofit in the UK: a critical \nassessment of five key archetypes. Energy Effic. 11, 1497\u20131517 (2018).\n\t42.\tDavis, F. D. Perceived usefulness, perceived ease of use, and user acceptance \nof information technology. MIS Q. 13, 319\u2013340 (1989).\n\t43.\tCurtius, H. C., K\u00fcnzel, K. & Loock, M.Generic customer segments and \nbusiness models for smart grids. der markt 51, 63\u201374 (2012).\n\t44.\tFell, M. J., Shipworth, D., Huebner, G. M. & Elwell, C. A. Public acceptability \nof domestic demand-side response in Great Britain: the role of automation \nand direct load control. Energy Res. Soc. Sci. 9, 72\u201384 (2015).\n\t45.\tVenkatesh, V. & Davis, F. D. A theoretical extension of the technology \nacceptance model: four longitudinal field studies. Manag. Sci. 46,  \n186\u2013204 (2000).\n\t46.\tRogers, E. M. Diffusion of Innovations 4th edn (The Free Press, 1995).\n\t47.\tAjzen, I.The theory of planned behavior. Organ. Behav. Hum. Decis. Process. \n50, 179\u2013211 (1991).\n\t48.\tPavlou, P. A. Consumer acceptance of electronic commerce: integrating trust \nand risk with the technology acceptance model. Int. J. Electron. Commer. 7, \n101\u2013134 (2003).\n\t49.\tAgarwal, R. & Prasad, J. A conceptual and operational definition of personal \ninnovativeness in the domain of information technology. Inf. Syst. Res. 9, \n204\u2013215 (1998).\n\t50.\tAgnew, S. & Dargusch, P. Consumer preferences for household-level battery \nenergy storage. Renew. Sustain. Energy Rev. 75, 609\u2013617 (2017).\n\t51.\tXu, X., Chen, C. F., Zhu, X. & Hu, Q. Promoting acceptance of direct load \ncontrol programs in the United States: financial incentive versus control \noption. Energy 147, 1278\u20131287 (2018).\n\t52.\tS\u00fctterlin, B. & Brunner, T. A. & Siegrist, M.Who puts the most energy into \nenergy conservation? A segmentation of energy consumers based on \nenergy-related characteristics. Energy Policy 39, 8137\u20138152 (2011).\n\t53.\tAlbert, A. & Rajagopal, R. Smart meter driven segmentation: what your \nconsumption says about you. IEEE Trans. Power Syst. 28, 4019\u20134030 (2013).\n\t54.\tMoon, N. & McHugh, S. Energy Market Investigation. Technical Report on a \nSurvey Conducted for the Competition and Markets Authority by GfK NOP \n(GfK, 2015).\n\t55.\tLittlechild, S. Promoting competition and protecting customers? Regulation \nof the GB retail energy market 2008\u20132016. J. Regul. Econ. 55, 107\u2013139 (2019).\n\t56.\tMazur, C., Hall, S., Hardy, J. & Workman, M. Technology is not a barrier: a \nsurvey of energy system technologies required for innovative electricity \nbusiness models driving the low carbon energy revolution. Energies 12,  \n428 (2019).\n\t57.\tBarton, C. Home Ownership and Renting: Demographics Briefing Paper \nNumber CBP 7706 (House of Commons Library, 2017).\n\t58.\tStenner, K., Frederiks, E. R., Hobman, E. V. & Cook, S. Willingness to \nparticipate in direct load control: the role of consumer distrust. Appl. Energy \n189, 76\u201388 (2017).\n\t59.\tPerlaviciute, G. & Steg, L. Contextual and psychological factors shaping \nevaluations and acceptability of energy alternatives: integrated review and \nresearch agenda. Renew. Sustain. Energy Rev. 35, 361\u2013381 (2014).\n\t60.\tDECC Public Attitudes Tracker - Wave 1 Summary of Key Issues (Department \nof Energy and Climate Change, 2012).\n\t61.\tDemski, C., Butler, C., Parkhill, K. A., Spence, A. & Pidgeon, N. F. Public \nvalues for energy system change. Glob. Environ. Change 34, 59\u201369 (2015).\n\t62.\tSehic, S., Ashworth, P. & Harris, J. Understanding the Socio-Economic \nChallenges for Energy Storage Uptake (The University of Queensland, 2017).\n\t63.\tDECC Public Attitudes Tracking Survey\u2014Wave 2 Questionnaire (Department \nof Energy and Climate Change, 2012).\n\t64.\tConsumer Engagement in the Energy Market Since the Retail Market Review. \n2016 Survey Findings (Ofgem, 2016).\n\t65.\tWaring, B., Silk, A. & Waring, G. Consumer Engagement in the Energy Market \n2017 (Ofgem, 2017).\n\t66.\tHenry, G. T. Practical Sampling (Sage, 1990).\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n360\n\nArticles\nNaTurE EnErgy\n\t67.\tHouseholds Below Average Income: An Analysis of the UK Income Distribution: \n1994/95 to 2015/16 (Department for Work and Pensions, 2017).\n\t68.\tScotland\u2019s Census 2011 (National Records of Scotland, 2018); http://www.\nscotlandscensus.gov.uk/\n\t69.\tCensus 2011 (Office of National Statistics, 2017); https://www.nomisweb.co.uk\n\t70.\tDe Vellis, R. F. Scale Development: Theory and Applications (Applied Social \nResearch Methods) Vol. 26 (Sage, 1991).\n\t71.\tCostello, A. B. & Osborne, J. W.Best practices in exploratory factor analysis: \nfour recommendations for getting the most from your analysis. Pract. Assess. \nRes. Eval. 10, 1\u20139 (2005).\n\t72.\tGorsuch, R. L. Factor Analysis 2nd edn (Lawrence Erlbaum Associates, 1983).\n\t73.\tHair, J., Anderson, R., Tatham, R. & Black, W. Multivariate Data Analysis \n(Prentice Hall, 1998).\n\t74.\tSarstedt, M. & Mooi, E. A Concise Guide to Market Research (Springer, 2014).\n\t75.\tAnable, J. Complacent car addicts or aspiring environmentalists? Identifying \ntravel behaviour segments using attitude theory. Transp. Policy 12,  \n65\u201378 (2005).\n\t76.\tMilligan, G. W. & Cooper, M. C. An examination of procedures for determining \nthe number of clusters in a data set. Psychometrika 50, 159\u2013179 (1985).\n\t77.\tHall, S., Mazur, C., Hardy, J., Workman, M. & Powell, M. Prioritising  \nbusiness model innovation: what needs to change in the United Kingdom \nenergy system to grow low carbon entrepreneurship? Energy Res. Soc. Sci. 60, \n101317 (2020).\nAcknowledgements\nThis research was partly funded by the Engineering and Physical Sciences Research \nCouncil (grant EP/N029488/1), Economic and Social Research Council (grant ES/\nM500562/1) and UK Research Councils (grants EPSRC EP/L024756/1 and NERC NE/\nG007748/1) as part of the UK Energy Research Centre (UKERC).\nAuthor contributions\nS.H. led the development of the Utility 2050 research process and led the literature \nreview discussion and conclusions. J.A. led the survey design with input on business \nmodel characteristics and question generation from J.H., S.H., C.M., and M.W.; J.A. led \nthe statistical analysis and segmentation exercise. M.W., C.M., J.A. and J.H. contributed \nsubstantive analysis, redrafting and editing. Y.M. aided early analysis and drafting.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary material \navailable at https://doi.org/10.1038/s41560-021-00781-1.\nCorrespondence and requests for materials should be addressed to S.H.\nPeer review information Nature Energy thanks Rachel Bray, Laura Olkkonen and the \nother, anonymous, reviewer(s) for their contribution to the peer review of this work.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2021, corrected \npublication 2021\nNature Energy | VOL 6 | April 2021 | 349\u2013361 | www.nature.com/natureenergy\n361\n\n1\nnature research  |  reporting summary\nOctober 2018\nCorresponding author(s):\nStephen Hall\nLast updated by author(s): 06-01-2021\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nProvide a description of all commercial, open source and custom code used to collect the data in this study, specifying the version used OR \nstate that no software was used.\nData analysis\nProvide a description of all commercial, open source and custom code used to analyse the data in this study, specifying the version used \nOR state that no software was used.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. \nWe strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nRelevant survey data, including all raw data, generated or analysed during this study are included in this published article in the Supplementary Data file. Data \ngenerated in the construction of business model archetypes is summarised in the Supplementary Information.\n\n2\nnature research  |  reporting summary\nOctober 2018\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThe study aimed to create archetypes of utility business models that are more compatible with a low carbon future than the current \nutility business model. The study then used a survey method to explore how a representative sample of the domestic British energy retail \nmarket (domestic consumers) responded to these models, what preferences they had, what consumer segments emerged and what this \ncould mean for the regulation of the British energy retail market in a low carbon future.  \nResearch sample\nThe aim was for a representative sample of electricity bill payers in England, Wales and Scotland. To screen for such people, we used the \nfollowing question: \nSample screening question \n\u201cWhich one of the following describes your level of involvement in decisions about which company your household uses to supply gas \nand/or electricity? \n1. It is my responsibility entirely \n2. I have equal responsibility with someone else in the household \n3. I have some involvement in the decision \nI have no involvement at all  THANK & CLOSE \nSampling strategy\nAs there are no reliable statistics against which to benchmark our achieved demographic profile of bill payers to achieve \nrepresentativeness among them, we used a proportionate stratified sampling approach to mirror the general population with quotas \nbased on age (six groups), gender and residential location (eleven Government Office Regions). Thus the sample itself is made up only of \nbill payers but unlikely to be entirely representative of them. For example, the number of 18-24-year-old respondents had to be \nspecifically boosted as a disproportionate number of this age group were otherwise consistently screened out due to being less likely to \nfall in to this group. However, given the focus of this research requires future facing preferences, over-representing this group was \njustified.  \nSSI sent a total of 41,579 invites, of whom 3,552 started and 2090 completed the survey (i.e. a 5% response rate and a 59% completion \nrate). Whilst this appears as a low response rate, it is not possible to calculate one with on-line samples due to the need to consider the \nincidence rate (the proportion of respondents contacted who qualify for the survey) and the fact the survey closes as soon as the \nrequired sample size has been achieved. Once full quotas are considered, the incidence rate was 57%. The final achieved sample was \n2,024 respondents after some were removed (N=66) who were deemed to have provided incomplete or invalid data based on \u2018straight \nliner variables\u2019 detected to indicate whether participants had given the same response for all sub questions in a relevant block.  \n \nData collection\nThe study employed a market research company (Accent) to programme the online survey and organise the data collection. Accent \nspecialise in online stated preference surveys with randomisation and have access to the SSI\u2019s global market research panel of \ndemographically diverse adults (>18 years of age) who have voluntarily subscribed to undertake such research. People register by \nproviding varying amounts of personal and demographic information that is later used to select participants for specific surveys. Thus, as \nis typical with such panels, the SSI panel does not use probability-based recruitment. Nevertheless, the panel from which this study\u2019s \nparticipants were selected is large enough to enable the selection of a nationally representative sample or sample representative of sub-\ngroups that reflect the actual breakdown of their key demographics (e.g. age, gender, region, social grade, ethnicity, disability etc.). \nTiming\nParticipants were sent an electronic communication (email or through a phone app) to be asked to participate and were rewarded a \nsmall incentive (approximately 0.85 GBP) for completion of the survey. The survey underwent a pilot test in March 2017 during which 66 \nparticipants completed this version. This led to some removal of attitudinal questions eliciting a disproportionate number of neutral \nresponses. The final questionnaire was in the field between 30th March and 13th April 2017. \nData exclusions\nN=66 participants were excluded having been deemed to have provided incomplete or invalid data based on \u2018straight liner variables\u2019 \ndetected to indicate whether participants had given the same response for all sub questions in a relevant block. \nNon-participation\nN/A\nRandomization\nIn a paired comparison task, respondents were forced to choose one or the other archetype in each of the ten paired cases (order \nrandomized). \nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \n\n3\nnature research  |  reporting summary\nOctober 2018\nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology\nAnimals and other organisms\nHuman research participants\nClinical data\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nHuman research participants\nPolicy information about studies involving human research participants\nPopulation characteristics\nAs there are no reliable statistics against which to benchmark our achieved demographic profile of bill payers to achieve \nrepresentativeness among them, we used a proportionate stratified sampling approach to mirror the general population with \nquotas based on age (six groups), gender and residential location (eleven Government Office Regions). Thus the sample itself is \nmade up only of bill payers but unlikely to be entirely representative of them. For example, the number of 18-24-year-old \nrespondents had to be specifically boosted as a disproportionate number of this age group were otherwise consistently screened \nout due to being less likely to fall in to this group. However, given the focus of this research requires future facing preferences, \nover-representing this group was justified. \nRecruitment\nUndertaken by a panel provider (Accent)\nEthics oversight\nUniversity of Leeds\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\n\n\n    Scientific Research Findings:", "answer": "We identified four new energy business models that could emerge in UK utilities:\u00a0A \u2018new electrifier\u2019, helping consumers to switch to electric heat and mobility; an \u2018energy service company\u2019, which uses long\u2011term contracts to finance smart retrofits; a peer\u2011to\u2011peer model where consumers generate and trade energy with each other; and a \u2018third party control\u2019 model where consumers allow a third party to meet their energy and other utility needs, taking decisions on their behalf. We presented these new models to existing consumers alongside a fifth \u2018control\u2019 model, \u2018same but smart\u2019 (a standard electricity tariff with a smart meter). We found that innovative energy contracts appeal differently to four specific consumer segments, each defined by a combination of characteristics relating to demographics, income, education, trust and willingness to innovate. Despite a willingness to change, individual circumstances (such as not owning their own home) will lock some consumers out of new offers. Other consumers are satisfied with the status quo or have trust issues with the energy market and are unlikely to engage with innovative offers. Retail energy market regulation is currently insufficient to acknowledge this complexity and the risk of poor social outcomes. These relations were observed in the UK retail energy market but could apply to other nations with liberalized energy systems.", "id": 18}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-020-00717-1\n1Energy Economics and Regulation Group, Department of Technology, Management and Economics, DTU Technical University of Denmark, Kongens \nLyngby, Denmark. 2European Commission\u2014Joint Research Centre, Petten, the Netherlands. 3National Renewable Energy Laboratory, Golden, CO, USA. \n\u2709e-mail: lkit@dtu.dk\nW\nind energy has emerged as a mainstream electricity pro-\nduction technology. Although globally it accounts for \napproximately 10% of electricity production capacity, it \nnow has a leading market share in some pioneering countries1,2. As \ntechnology rapidly improves and operating facilities increase in age, \nthe modernization of wind facilities becomes increasingly relevant. \nRepowering, that is, the combined activity of dismantling or refur-\nbishing and commissioning of wind turbines, has been an active \nsegment of the industry since the 1990s, when the first commer-\ncial wind farms (commissioned in the late 1970s and early 1980s) \nbegan to require substantial reinvestments and modernization, \nbut the scale and impact of repowering on the global wind indus-\ntry is expected to grow substantially in the next decade. Moreover, \nalthough repowering activity to date has been concentrated in a \nfew markets and driven by relatively unique policy incentives, by \nthe late 2020s repowering could become a primary source of busi-\nness activity for the wind industry across Europe, North America \nand China, and a key to optimizing wind energy utilization in the \ncontext of energy transition. Today, only a few countries have wind \nenergy fleets that support substantial repowering activity2. In this \ncontext, empirically derived insights from repowering activities are \nincreasingly informative. Denmark is a prime case for analysis: wind \nenergy has emerged as the primary electricity production technol-\nogy, having reached a 47% market share in 20193, and it features the \noldest wind turbine fleet in the world1,4 (Table 1). Further factors \nthat contribute to a robust repowering market in Denmark are the \nsize of the country, with its consequent space limitations, and its \ncontinued desire to expand wind energy production5.\nRepowering has historically been viewed as a means to increase \nthe productivity of existing projects, with improved profitability \nas the primary decision driver4,6. Although the motivation behind \nmuch repowering activity is an improved business case for an exist-\ning facility, repowering offers an array of benefits when compared \nwith greenfield development, which include lower implementation \nbarriers through existing grid connections, long-term empirical \nunderstanding of the available wind resource at a given site, rela-\ntively well-known environmental impacts and accustomed neigh-\nbours. Further, subsidy schemes and short-term market incentives \nmay encourage repowering, and include incentive programmes \ndirected at repowering (Supplementary Note 1), market premiums \nfor new projects6 or favourable tax policy. Recent repowering in the \nUnited States, for example, was predominantly driven by the ability \nto requalify for full value and tenure of the production tax incentive \nprogramme5.\nFrom a societal perspective, repowering may offer environ-\nmental and social improvements through the replacement of older \n(faster-rotating and noisier) turbines with fewer and more efficient \nnewer turbines\u2014although the new (taller) turbines may intro-\nduce new negative impacts, largely related to visibility over longer \ndistances7. It is further suggested that birds and bats benefit from \nrepowering, although challenges remain in the monitoring and \nassessment of bird fatalities and collision rates before and after \nrepowering8. Repowering is also particularly relevant in offering the \npotential to increase wind energy production from a limited area \nof available land, so that even when \u201cvirtually all the good wind \nresource sites have been taken\u201d (using the words of a representative \nfrom the Danish wind industry)9, increases in wind energy produc-\ntion may still be facilitated.\nPrevious scientific studies on repowering can be broadly catego-\nrized into five groups: feasibility-type studies on single projects4,10\u201315; \noptimization of repowering decisions, for example, regarding \ntime and wind farm topology4,16\u201319; repowering statistics, such as \nincreased capacities or productivity, on a turbine-by-turbine com-\nparison20,21; research on public acceptance6 and rejection factors \nfor repowering projects7,22 and impact assessments, which address \nlife-cycle emissions23, visual impacts24 and impacts on birds and \nbats25\u201328. Few studies explore techno-economic and regulatory driv-\ners of repowering decisions5,6 and policy options16,29. All the studies \nMultifaceted drivers for onshore wind energy \nrepowering and their implications for energy \ntransition\nLena Kitzing\u200a \u200a1\u2009\u2709, Morten Kofoed Jensen1, Thomas Telsnig2 and Eric Lantz3\nWind energy is anticipated to become a backbone of the future energy system. Ageing wind turbine fleets, increasing land-use \nconstraints and rising relevance of societal factors make the deployment of land-based (onshore) wind energy ever more com-\nplicated. Consequently, repowering is expected to become a rapidly growing point of focus for the wind industry. Here we \npropose a more holistic and socially informed project-level approach to analyse repowering activity that enables a more robust \nunderstanding of the process and potentials. We demonstrate that for wind pioneer in Denmark, only 67% of the capacity \nremoved in repowering projects was related to the physical space needed for a new turbine. Other factors that drive repowering \ninclude regulation (for example, noise-related, 8\u201317%), development principles (for example, aesthetics, 7\u201320%) and political \nbargaining (4\u201313%). The recognition of repowering as a negotiated process between host communities and wind developers \nwill probably be critical to unlock the full potential of wind energy in the future.\nNature Energy | VOL 5 | December 2020 | 1012\u20131021 | www.nature.com/natureenergy\n1012\n\nArticles\nNATure Energy\nfocus their scope towards the replacement of turbines on existing \nsites or consider a turbine-by-turbine replacement. We label this \ndominant approach as an \u2018on-site approach to repowering\u2019, and \nshow that this is insufficient for a robust understanding of repower-\ning decisions, processes and outcomes.\nWe propose a more holistic and socially informed perspective \non repowering based on comprehensive project-level information. \nSpecifically, this approach considers multifaceted drivers for repow-\nering decisions and hence identifies dismantled turbines that are \nlocated at the same site as the new development project (on-site) as \nwell as existing turbines in other locations that need dismantling so \nnew turbines can be installed, that is, we considered all the activity \nthat can be conditionally associated with the repowering project, \nregardless of spatial or temporal proximity between the existing \nand the new turbines, as illustrated in Fig. 1. We then identified \ncategories for dismantling reasons in repowering projects, espe-\ncially for those turbines physically located outside the immedi-\nate project development site (off-site). Applying our approach to \nDenmark as an empirical case, we quantified repowering projects \nin their entirety. Through data collection from municipal plans30,31, \nproject publications, geospatial analysis and semi-structured inter-\nviews with developers covering 91% of all dismantled capacity in \nrepowering projects, we identified the full breadth of repowering \nprojects and determined dismantling reasons. We revealed sev-\neral notable differences between the traditional on-site and our \nmore holistic and socially informed repowering perspective. Our \nwork emphasizes that a pure technology perspective alone cannot \nexplain implementation pathways and must be supplemented with \nthe political and social dimensions. Ultimately, this work enables \nmore-informed estimations of the future global repowering mar-\nket. It also informs the challenges and opportunities presented by \nrepowering as a means to increase renewable energy production.\nRepowering projects in Denmark and their market \nsignificance\nOur synthesis database contains 102 wind energy projects that were \ndeveloped in Denmark between 2012 and 2019. Table 2 provides an \noverview of our database by greenfield projects, repowering proj-\nects and dismantled non-repowered turbines, which are turbines \ndismantled without relation to any new project. Eight projects are \nexcluded from the analysis (Methods).\nMore than a third (38%) of all new wind energy developments \nwere repowering projects. Only 10% more capacity was developed \nin greenfield projects than in repowering. In repowering projects, \nthe net capacity additions amounted to 576.8\u2009MW, whereas, at the \nsame time, the net number of turbines decreased by 1. Overall, we \nobserve 1.3\u2009GW of net capacity additions and a net reduction of 109 \nturbines, which considerably reduces the number of turbines physi-\ncally present in the landscape. This shows that newer, more-efficient \nturbines have replaced earlier, less-efficient turbines over time.\nTurbines removed in repowering projects are, on average, \n5.8 years younger than those dismantled on a stand-alone basis \n(non-repowering). Interviewees confirmed that most of the dis-\nmantled turbines in repowering projects had not reached the end \nof their operational lifetime, but were dismantled prematurely \nso that the new project could be executed. This emphasizes the \nimportance of distinguishing between end-of-life decisions and \nrepowering decisions, which involve multiple social and economic \nconsiderations.\nFurther, dismantled repowering turbines are, on average, 3.4 \ntimes as large as dismantled non-repowered turbines. Although \nrepowering comprises 66% of the total dismantled capacity, it only \ncomprises 37% of the total number of dismantled turbines. This \nsize difference between dismantling in repowering and as a separate \nactivity meant we further analysed both the capacity and the num-\nber of turbines.\nRepowering has played an increasingly important role in \nDenmark since 2012, both in terms of capacity (Table 3) and the \nnumber of turbines (Table 4). Repowering shares have steadily \nincreased for gross added capacity and the number of installed \nturbines. The relatively low starting value in 2012, of 23%, may be \ndue to the removal of a pre-existing repowering incentive scheme \nthat was in place from 2005 to 201132, which left a weak market \nfor repowering at the beginning of the observed period. In the last \nobserved year, the repowering market share jumped to an unprec-\nedented level of 86% of gross added capacity (Table 3), or 87% of \nadded wind turbines (Table 4).\nRepowering shares in the dismantling of turbines were even \nmore notable, although they varied over time. Repowering was \nthe principal reason for capacity dismantling in five of the eight \nyears investigated, and the repowering share in capacity reductions \nranged between 17% in 2015 and 92% in 2013. Variations in repow-\nering shares for the number of dismantled turbines follow a differ-\nent pattern, as turbine sizes have increased in recent years.\nMore detailed results from our dataset, including on market \nshares of project developers and turbine manufacturers, can be \nfound in Supplementary Note 3.\nCategorizing dismantled turbines into repowering reasons\nAs argued earlier, more holistic and socially informed repower-\ning analysis must include all the dismantled turbines related to a \nrepowering project regardless of distance and time. For a deeper \nunderstanding of the reasons behind dismantling turbines in \nrepowering projects, we conducted in-depth interviews with lead-\ning wind developers in Denmark, which covered 91% of all the \ncapacity dismantled in repowering projects in the investigated time \nframe. From the interviewee responses, eight mutually exclusive \ncategories emerged for dismantling existing turbines in repowering \nprojects, as summarized in Table 5.\nIt is noteworthy that none of the interviewees named profitabil-\nity or other economic reasons as drivers to dismantle turbines in \nrepowering projects. On the contrary, talking about the market for \nrepowering, one interviewee stated that greenfield projects generally \nprovide the highest profits. Repowering is seen as a costly necessity, \nnot least because the existing turbines often have to be purchased at \na high cost before they can be dismantled. This is notwithstanding \nthe fact that some older wind turbines may not have been feasible \nto operate much longer, and in general may not have fulfilled the \nowner\u2019s expectations, for example, in regard to production volumes, \nfailure rate and downtime.\nAlthough we focused our study on repowering from the  \ndevelopers\u2019 perspective, others who investigated the attitudes of \nlocal politicians towards wind energy projects and repowering, for \nTable 1 | Turbine fleet age structure in leading countries for \nonshore wind energy\nDenmark\nGermany\nSpain\nEU28\nUSA\nChina\nCumulative \ncapacity \ninstalled in \n2019 (GW)\n4.4\n53.2\n23.5\n160.7\n97.7\n206.8\nShare of cumulative capacity\n\u2002>10 years\n55%\n43%\n73%\n39%\n34%\n7%\n\u2002>15 years\n53%\n26%\n27%\n17%\n6%\n0.4%\n\u2002>20 years\n23%\n4%\n3%\n3%\n1%\n0.2%\nThe table shows the cumulative capacity of installed onshore wind energy and the age distribution \nof installed turbines in selected countries. Data from refs. 1,42,43. EU-28, European Union 28 \ncountries.\nNature Energy | VOL 5 | December 2020 | 1012\u20131021 | www.nature.com/natureenergy\n1013\n\nArticles\nNATure Energy\nexample, in the Czech Republic, found that they equally considered \neconomic value creation and displeasure of local residents and other \nlocal stakeholders22. Concrete reasons that are likely to contribute \nto dismantling requests in our context were found to involve the \nopposition of local residents, local landscape disruption and impact \non well-being (noise)22.\nTable 6 shows the results of our categorization effort. We found \nthat space is the strongest motivation for dismantling turbines in \nrepowering projects in Denmark, with 67% of the total capacity \nand 63% of the total number of turbines, respectively. Space is, of \ncourse, a well-known reason given that most repowering projects \nare developed from existing wind farm sites. This category gener-\nally corresponds to what is typically captured by analyses using an \non-site approach.\nThe second major individual reason for dismantling turbines  \nin repowering is a violation of the cumulative noise emissions, \nwhich comprises 13% of the total number of turbines and 8%  \nof the full capacity dismantled. Hence, noise regulation in Denmark \nhas a considerable impact on repowering projects. This impact \nwas also underlined by several of the interviewees. Here, it is rel-\nevant to note that noise regulation differs from country to coun-\ntry. Interestingly, turbines removed because of noise have the  \nsmallest average capacity size (435\u2009kW) and highest average age \n(22.1 years, Fig. 2b).\nIt was sometimes impossible for developers to point to a specific \nreason for dismantling a particular turbine, especially when aes-\nthetics and politics were involved. We thus analysed the remaining \ncategories jointly. The individual category aesthetics comprises 9% \nAesthetics\nPolitical\nbargaining\nNoise\nNoise\nDismantled turbine\nCommissioned turbine\n4 times the total height of \ncommissioned turbines\n28 times the total height of \ncommissioned turbines\n1.5 times the total height of \ncommissioned turbines\nCumulative noise violation \nzone  \nSpace\nSettlements\nFig. 1 | Illustrative map of a typical repowering project that includes the on-site as well as off-site dismantling of turbines. This representative example \nfor a (fictive) Danish repowering project is based on the median repowering project information from our dataset, in which six 600\u2009kW turbines \ncommissioned in 1997 are replaced by five new 3.3\u2009MW turbines in 2016. Annual energy production increased by a factor of 6.5 and operations and \nmaintenance costs declined by about 20%, which results in an increased annual net cash flow of up to \u20ac0.9 million assuming a high electricity price \nscenario. The overall levelized cost of energy is \u20ac38\u201342\u2009MWh\u20131, approximately 70\u201378% of the original project overall levelized cost of energy (see \nSupplementary Note 2 for more details). Note that wind farm layout and scales are chosen to aid visualization and not to represent a realistic case.\nTable 2 | Onshore wind energy projects developed in Denmark 2012\u20132019\nNumber \nof \nprojects\nTotal \nnumber of \nturbines\nAverage \nnumber of \nturbines \nper project\nRange of \nnumber of \nturbines \nper project\nTotal \ncapacity \n(MW)\nAverage \ncapacity \nper project \n(MW)\nAverage \ncapacity \nper turbine \n(MW)\nRange of \ncapacity per \nproject (MW)\nAverage \nturbine age at \ndismantling (yr)\nGreenfield projects\n58\n272\n4.7\n1\u221221\n804.3\n13.9\n3.0\n2\u221246.8\nRepowering projects \n(net)\n36\n\u22121\n0\n\u221214\u221216\n576.8\n16.04\n0.15\u221265.9\nRepowering projects \n(gross additions)\n220\n6.1\n1\u221222\n731.9\n20.3\n3.3\n0.2\u221277.4\nRepowering projects \n(gross reductions)\n221\n6.1\n1\u221235\n155.1\n4.3\n0.7\n0.075\u221223.0\n18.6\nNon-repowered \ndismantled turbines\n380\n78.8\n0.2\n24.4\nSee Methods for details on the derived numbers, and Supplementary Note 3 for more detailed results.\nNature Energy | VOL 5 | December 2020 | 1012\u20131021 | www.nature.com/natureenergy\n1014\n\nArticles\nNATure Energy\nof the total number of turbines and 7% of the total capacity dis-\nmantled. However, aesthetics also play a role in three mixed catego-\nries (aesthetics or noise, politics or aesthetics and politics, aesthetics \nor noise). Assuming that all the turbines in these three categories \nwere removed because of aesthetics, it would become the second \nmost pronounced reason, comprising 21% of the total number of \nturbines and 20% of the total capacity dismantled. The individual \ncategory politics comprises 3% of the total number of turbines and \n4% of the total capacity dismantled. If all the turbines in the respec-\ntive mixed categories (politics or aesthetics and politics, aesthetics \nor noise) were dismantled because of politics, the category politics \nwould contain 10% of the total number of dismantled turbines or \n13% of the removed capacity.\nMajor differences in distance, age and repowering factor\nThe distance between dismantled and newly installed turbines dif-\nfers considerably across the categories (Fig. 2a), of which the cat-\negory \u2018politics\u2019 has the highest spread. With a dismantled turbine \nlocated 4.6\u2009km (31 times the total height) away from the nearest new \nturbine, it is apparent that politics can target a broad range of exist-\ning turbines. As stated by one interviewee, politics seem spatially \nunrestricted within the municipality because voluntary agreements \nbetween developers and politicians can include the removal of any \nexisting turbines within the municipal area. Some interviewees \nperceived that some municipalities had an informal one-for-one \nrequirement, and one interviewee confirmed that this was the very \nreason why they chose to dismantle a specific older existing turbine \nlocated in the municipality. Concerning the one-for-one require-\nment, another interviewee elaborated on a current project (to be \ncommissioned in 2022), in which, as part of a voluntary agreement, \nthey would dismantle several turbines located more than 10\u2009km \naway from the site of the new project.\nIn contrast to the informal nature of politics-related agreements, \nformal regulations are in place that restrict the farthest impact  \ndistances of noise and aesthetics: the relationship between existing \nand planned turbines only needs to be evaluated within a radius \nof 28 times the total height of the planned turbines33. Here, the \nobserved maximum removal distances are 2.3\u2009km (15 times the \ntotal height) for noise and 2.8\u2009km (18.5 times the total height) for \naesthetics.\nTurbine age at dismantling varies considerably across the catego-\nries (Fig. 2b). The space category shows some of the highest and \nlowest dismantling ages, which range from 3.8 to 32.9 years with a \nmedian value of 17.4 years. The average age for the noise category is \n22.1 years, the highest of all groups, which suggests that older tur-\nbines more commonly induce a violation of noise emission limits. \nThose older turbines are mechanically louder and have less sound \ninsulation than newer turbines, as stated by one of the interviewees.\nIn seven of the eight categories, both the median and average \nvalue for the operating turbine lifetime are <20 years. This is sur-\nprising because turbines installed in Denmark between 1996 and \n2008 are guaranteed government support for 20 years34. Hence, \nmany of these turbines forego support payments because of dis-\nmantling. This reinforces the idea that repowering is seldom a tech-\nnical end-of-lifetime decision or even an end-of-support decision. \nOne interviewee stated that they continuously monitor their fleet \nof existing turbines and evaluate the turbines\u2019 economic viability to \ninclude them or not in a new repowering project; hence, the current \nage of the existing turbine is, in itself, largely irrelevant, and merely \na parameter in the broader economic evaluation.\nOur analysis reveals that turbines in repowering projects are dis-\nmantled on average 5.8 years earlier than those separately disman-\ntled (non-repowering) (Table 2). These \u2018lost\u2019 years of production \nconstitute foregone value creation. Of course, the lost production \nTable 3 | Annual wind turbine capacity additions and reductions in megawatts\n2012\n2013\n2014\n2015\n2016\n2017\n2018\n2019\nTotal\nRepowering, gross additions\n30.0\n122.4\n40.5\n92.9\n104.8\n126.9\n108.3\n106.2\n732\nRepowering, gross reductions\n\u22120.8\n\u221221.7\n\u221230.8\n\u22122.7\n\u221221.4\n\u221247.1\n\u22128.5\n\u221222.3\n\u2212155\nGreenfield\n103.2\n167.3\n57.5\n110.9\n94.3\n174.9\n79.5\n16.8\n804\nSeparately dismantled turbines\n\u22121.0\n\u22121.8\n\u22126.2\n\u221213.5\n\u221230.0\n\u221218.8\n\u22123.0\n\u22124.4\n\u221279\nAnnual net capacity instalment\n131\n266\n61\n188\n148\n236\n176\n96\n1302\nShare of repowering: commissioned \n(%)\n23\n42\n41\n46\n53\n42\n58\n86\n48\nShare of repowering: dismantled (%) 43\n92\n83\n17\n42\n71\n74\n84\n66\nThe data exclude test and small-scale (household) turbines (<25\u2009kW) (Methods).\nTable 4 | Annual number of wind turbine additions and reductions\n2012\n2013\n2014\n2015\n2016\n2017\n2018\n2019\nTotal\nRepowering, gross additions\n10\n38\n13\n31\n32\n37\n33\n26\n220\nRepowering, gross reductions\n\u22121\n\u221226\n\u221253\n\u22126\n\u221232\n\u221267\n\u22128\n\u221228\n\u2212221\nGreenfield\n34\n63\n21\n38\n37\n51\n24\n4\n272\nSeparately dismantled turbines\n\u22123\n\u22129\n\u221237\n\u221270\n\u2212145\n\u221294\n\u221211\n\u221211\n\u2212380\nAnnual net turbine instalment\n40\n66\n\u221256\n\u22127\n\u2212108\n\u221273\n38\n\u22129\n\u2212109\nShare of repowering: commissioned \n(%)\n23\n38\n38\n45\n46\n42\n58\n87\n45\nShare of repowering: dismantled (%)\n25\n74\n59\n8\n18\n42\n42\n72\n37\nThe data exclude test and small-scale (household) turbines (<25\u2009kW) (Methods).\nNature Energy | VOL 5 | December 2020 | 1012\u20131021 | www.nature.com/natureenergy\n1015\n\nArticles\nNATure Energy\nis offset by the additional value created from the larger and more  \nefficient new turbines. However, it can be argued that for the tur-\nbines that fall within the more intangible dismantling categories, \nwhich include aesthetics and politics, the value of lost production \nshould be considered and weighed against the societal willingness \nto pay for the turbine removal. Approximating the value of lost \nproduction through electricity market prices, we estimate between \n\u20ac0.27 and 0.43 million per turbine is lost as a result of premature \ndismantling\u2014the lower value reflects the average of all the cat-\negories and the higher one that of the three categories that capture \nnon-physical reasons for dismantling (that is, aesthetics, politics, \nand aesthetics and politics). This value compares with estimated \noperations and maintenance costs of \u20ac0.25 to 0.29 million per tur-\nbine over a period of 5.8 years35. The general comparability of these \nranges points toward economic as well as societal factors as drivers \nof turbine dismantling in repowering projects.\nThe presented empirical data demonstrate the difficulties in \nanalysing repowering strictly from an on-site perspective. Likewise, \nsimply broadening the scope to a larger radius around the new \ndevelopment site comes with considerable limitations, as illustrated \nin Table 7, which depicts the total number of dismantled turbines \nand capacity as a function of proximity to the nearest new turbine \nwithin each repowering project, along with the share of dismantled \nturbines captured by the respective approach.\nWe observe that increasing the radius of investigation around \nthe newly developed turbine also increases the likelihood of captur-\ning a higher share of relevant dismantled turbines. However, even \na radius equal to ten times the total height of a new turbine fails \nto capture 10% of the associated dismantled turbines. Furthermore, \nsuch a simplistic radius approach bears the risk of including unre-\nlated dismantling of turbines in the repowering statistics. Therefore, \nit is still necessary to manually establish a causal relationship  \nbetween a commissioned and dismantled turbine within each \nrepowering project.\nThe on-site approach equally falls short of estimating a realis-\ntic net repowering factor, which is the ratio between commissioned \nand dismantled values (number of turbines or capacity), and hence \ndenotes the amount commissioned for every unit dismantled. \nIncluding all the causally identified dismantled turbines in repow-\nering projects, we obtain a net capacity repowering factor of 4.72, \nin contrast to 13.10 if using a radius derived from one times the \ntotal height of the new turbines. If we only consider those turbines \nthat have been identified in the space category, the net repowering  \nfactor would equal 7.05. Hence, project-level information and \ncausal relationships are crucial to determine the full extent of  \nrepowering impacts.\nDiscussion and conclusions\nThis work proposes a more holistic and socially informed \nproject-based perspective on wind plant repowering that starts \nfrom the development of new projects and explores the need for \nand drivers of removing existing turbines to enable the realization \nof new wind energy facilities. Our results suggest that to focus on \nend-of-life considerations and assuming a simple on-site localized \nreplacement of existing turbines is insufficient to truly understand \nthe repowering impacts and opportunities. This is demonstrated \nby an application to Denmark, in which more than a third of all \nwind energy development projects between 2012 and 2019 involved \nrepowering. Only 67% of the capacity removed in repowering \nprojects was related to turbines that occupied the physical space  \nTable 5 | Reasons for dismantling existing wind turbines in repowering projects as derived from interviews\nSpace\nPhysical requirement for the installation of new turbines, such as a space requirement for new foundations, access roads, \nsetbacks and grid infrastructure. Aside from the stated reasons by the interviewees, we included dismantled turbines \nwithin the radius of 1.5 times the total height of the new wind turbine also to be in this category, as a minimum distance \nrequired to prevent the hazard from a collapse of turbine towers.\nNoise\nViolation of the cumulative noise emissions within a radius of 28 times the total height of the planned wind farms. A noise \nemission difference less than 15\u2009dB between existing wind turbines and planned wind turbines at an inhabited property \ninduces a cumulative noise calculation. If this calculation exceeds the noise limit, either noise reduction on the new \nturbines or dismantling of the existing turbines are conceivable options.\nAesthetics\nA disturbance of the scenic or aesthetic values in the landscape. The Environmental Impact Assessment (EIA) requires \na visual impact analysis that concerns disturbances of aesthetic values within a radius of 28 times the total height of \nthe planned wind turbines. Focusing on location, design and the interaction between existing turbines and planned wind \nturbines, the visual impact analysis can result in the dismantling of turbines.\nAesthetics or noise\nEither a violation of noise emissions or a disturbance of the scenic or aesthetic values in the landscape caused the \ndeveloper to dismantle existing wind turbines. Although the interviewees could not single out one reason for the \ndismantling, they explicitly negated politics as a reason for turbines in this category.\nPolitics\nDuring local development and planning dialogues, informal requests by local politicians, proactive suggestions by \ndevelopers or voluntary agreements by both parties are put forward to satisfy the agreement on the development of the \nnew project. All of these actions involve the dismantling of existing wind turbines located in the municipality based on \nsubjective intentions as a condition for the new project.\nPolitics or aesthetics\nThe interviewees could not point out precisely if the main driver was politics or aesthetics. Some mentioned that local \npoliticians tend to use the disturbance of aesthetic values in the landscape to argue for the removal of existing wind \nturbines, which made selecting between aesthetics or politics difficult.\nPolitics, aesthetics or noise\nFor some turbines, a differentiation between the dismantling reasons is unobtainable. A combination of factors may \nprompt the developer to plan the dismantling of existing wind turbines before the permit proposal. In the pre-permit \nprocess, the developer identifies problematic wind turbines knowing that they will be chosen for dismantling as a result \nof politics, noise or aesthetics, and thereby save time on wind farm design, noise calculations and the visual impact \nassessment.\nUnclassified\nUnassigned reasons for dismantling wind turbines. This category contains those wind turbines for which our interviews or \nthe local development plans do not cover the motivations behind the dismantling.\nSee Methods for a more comprehensive description of the reasons.\nNature Energy | VOL 5 | December 2020 | 1012\u20131021 | www.nature.com/natureenergy\n1016\n\nArticles\nNATure Energy\nneeded for a new turbine. We identified several other reasons for \ndismantling, which included regulation (for example, noise-related, \n8\u201317% of removed capacity) and development principles (for exam-\nple, aesthetics, 7\u201320% of removed capacity). Also, political bargain-\ning around the dismantling of unwanted older turbines was not \nuncommon (4\u201313% of removed capacity).\nThe broader perspective offered by our approach highlights how \nthe repowering process is not a matter of turbine replacement but \na negotiation between the developer and the local host community \nabout how wind energy should coexist with the community given \nthe changes that the introduction of the new technology might \ninduce. There is no doubt that repowering can increase the total \ncontribution of wind energy from a given land area. However, we \ndemonstrate that this contribution is not as high as theoretically \npossible based on physical considerations alone. Considering only \nthe on-site replacement of turbines might lead to the mispercep-\ntion that a net repowering factor of 7.05 could have been achieved \nin Denmark, that is, that the installed capacity after completing \nan average repowering project would be 7.05 times higher than \nthe original capacity. In fact, by establishing causal relationships \nbetween repowering projects and all dismantled turbines, which \ninclude the additional off-site capacity removed, we observed that \nthe actual net capacity repowering factor per project was only \n4.72. Although the repowering potential within a project may still \nbe considerable, it is weakened by the conditional dismantling of \nunwanted off-site turbines and the associated land use and societal \nconsiderations that this dismantling may serve. In addition, tur-\nbines dismantled in repowering projects achieved an average oper-\nating age of 5.8 years less than those dismantled on a stand-alone \nbasis (non-repowering) and are often dismantled before reaching \nthe end of their operational life. Accordingly, repowering may arrive \nearlier and with a stronger force than would be predicted when tak-\ning an end-of-life perspective.\nNotwithstanding the relatively lower net repowering factor and \nshorter lifetime of the dismantled repowered turbines, our analy-\nsis also indicates that repowering, in combination with techno-\nlogical innovation, can help reduce the impacts of wind turbines  \non neighbouring communities. In a sense, repowering provides \nTable 6 | Wind turbines dismantled in repowering projects by dismantling reason\nSpace\nNoise\nAesthetics\nAesthetics or \nnoise\nPolitics\nPolitics or \naesthetics\nPolitics, \naesthetics or \nnoise\nUnclassified\nNumber of turbines\n139\n28\n20\n10\n6\n6\n10\n2\nCapacity (MW)\n103.8\n12.2\n11.3\n6.7\n6.3\n5.9\n7.4\n1.5\nAverage capacity per \nturbine (MW)\n0.747\n0.435\n0.566\n0.669\n1.050\n0.983\n0.740\n0.750\nShare of number (%)\n63\n13\n9\n5\n3\n3\n5\n0.9\nShare of capacity (%)\n67\n8\n7\n4\n4\n4\n5\n1.0\nSpace\nNoise\nAesthetics\nAesthetics or noise\nPolitics\nPolitics or aesthetics\nPolitics, aesthetics or noise\nUnclassified\n0\n500\n1,000\n1,500\n2,000\n2,500\n3,000\n3,500\n4,000\n4,500\n5\n10\n15\n20\n25\n30\nDistance (m)\nYears\nSpace\nNoise\nAesthetics\nAesthetics or noise\nPolitics\nPolitics or aesthetics\nPolitics, aesthetics or noise\nUnclassified\nb\na\n75%\n25%\nMin\nMax\nMedianMean\n75%\n25%\nMin\nMax\nMedianMean\nFig. 2 | Distribution of wind turbines in repowering projects by dismantling category. a, The nearest distance between dismantled and installed wind \nturbines. b, The age of dismantled turbines.\nNature Energy | VOL 5 | December 2020 | 1012\u20131021 | www.nature.com/natureenergy\n1017\n\nArticles\nNATure Energy\ndual benefits to society\u2014an increased clean energy supply and \nfewer turbines present in the landscape. Although these benefits do \nnot come without a trade-off (such as higher visibility over longer \ndistances), they are substantial in regions where space is limited and \nan increased deployment of wind energy is desired. This situation \nmay be increasingly experienced with a growing reliance on renew-\nable power to serve clean energy needs.\nAs Denmark is a global pioneer of wind energy and repowering, \nthe experiences observed here are anticipated to become increasingly \nrelevant for other countries and regions. Of course, we acknowledge \nthat many of the non-space-related reasons for dismantling turbines \nare idiosyncratic and depend on the physical, political and social \nlandscape in a country. Nevertheless, the diversity of repowering \nand dismantling drivers captured here illuminates the complex \nand multifaceted decision process that constitutes repowering and \nillustrates considerable implications for policymakers and planners  \nwho try to understand the potential of wind power to serve their \nenergy needs.\nMethods\nEthical compliance. We have complied with all relevant ethical regulations. The \ninterviews were conducted in Denmark by the Technical University of Denmark \naccording to the guidelines. All interactions followed Chatham House rules. We \nhave obtained informed consent from all interview participants.\nApproach. We took a new wind energy development project as the starting \npoint of the analysis. From this lens, we considered all the turbines that need \nto be dismantled to enable the undertaking of the new project (as a conditional \nrequirement). We used the term dismantling, but, in principle, our approach \ncan capture both the complete dismantling of turbines (also referred to as full \nrepowering4), and the installation of new equipment (for example, drivetrain \nand rotor) on an existing tower and/or foundation (also referred to as partial \nrepowering4). We emphasize that the analysis must include the dismantling of \nexisting turbines not only at the location of the new project undertaking (on-site), \nbut also those located elsewhere (off-site). Furthermore, existing turbines may \nbe dismantled years before the new turbines are installed. Therefore, repowering \nmust reflect a combined action of dismantling existing turbines and establishing \nnew turbines, regardless of spatial or temporal proximity. We thus investigated \nthe conditional relationship between capacity reduction and capacity addition. \nThis enabled us to reveal notable differences between repowering projects \n(comprising both commissioning new turbines and dismantling existing turbines) \nand greenfield projects (only comprising the commissioning of new turbines) for \nseveral key indicators, such as project size and lifetime.\nApplying our approach on an empirical case, we created a synthesized database \nof all the approved Danish wind projects between 2012 and 2019. We constructed \nthe database by using publicly available sources and semi-structured interviews \nwith leading wind developers in Denmark. Our database has two key advantages. \nFirst, it is exhaustive on a project-level basis, whereas previous databases focused \nsolely on technical specifications or the physical planning of wind turbines. \nSecond, our database allows for a comprehensive overview of the extent of \nrepowering projects in Denmark, whereas most previous studies only provide a \ngeneral overview (in the form of \u2018bulk\u2019 sizes or turbine-per-turbine comparisons). \nOur database is uniquely suited to identify the reasons for dismantling turbines in \nrepowering projects. We created our database between May 2018 and September \n2019 in three major stages. In the first stage, we manually extracted data and \ncoded the database. In the second stage, we interviewed leading wind developers \nto collect qualitative data on the motivations behind dismantling turbines in \nrepowering projects. In the third stage, we synthesized and analysed the rich data \nfrom the interviews into our database.\nData selection. We limited our time horizon to the period between 2012 \nand 2019. First, we excluded the years before 2001 because of discontinued \nregulations that may have distorted repowering decisions from today\u2019s perspective. \nSecond, we also excluded the years from 2001 to 2011 as, during that period, \nthe Danish government provided policy incentive programmes for repowering \n(Supplementary Note 1) that could have confounded our analysis. Third, we \nexcluded projects that involved research and test turbines, as these turbines \nare subject to different regulations. Of the 102 identified wind projects, 8 were \nexperimental test projects, that is, turbines established and dismantled at test sites. \nIn total, 70 test turbines were dismantled during the investigated period. Fourth, \nwe excluded household wind turbines, that is, turbines with a generation capacity \nof 25\u2009kW or less, as there are different regulations for these types of turbines. \nDuring the investigated period, 77 household turbines were dismantled. Fifth, we \nexcluded a dismantling project that was part of an infrastructure development \nproject, in which 26 turbines were dismantled because of the construction of the \nFemern tunnel in Southern Denmark. Overall, we determined that 221 out of the \nrelevant 601 dismantled turbines were repowering related.\nNote that we included all wind energy projects that were developed, approved \nand project execution started up to the end of 2019. We included two repowering \nprojects, with a total of 106.2\u2009MW, for which the commission date was rescheduled \nTable 7 | Dismantled wind turbines in repowering projects as a function of proximity to the nearest commissioned new turbine by \ncapacity and number of turbines\nProximity to the nearest commissioned repowering turbine\nCategory \nspace\nAll \ndismantled \nturbines\n<1\u2009\u00d7\u2009total \nheight of a \nnew turbine\n<2\u2009\u00d7\u2009total \nheight of a \nnew turbine\n<3\u2009\u00d7\u2009total \nheight of a \nnew turbine\n<4\u2009\u00d7\u2009total \nheight of a new \nturbine\n<5\u2009\u00d7\u2009total \nheight of a \nnew turbine\n<10\u2009\u00d7\u2009total \nheight of a new \nturbine\nNumber of \ndismantled \nturbines identified\n75\n147\n168\n174\n179\n199\n139\n221\nShare of \ndismantled \nturbines captured \n(%)\n33.9\n66.5\n76.0\n78.3\n81.0\n90.0\n63\n100\nDismantled \ncapacity identified \n(MW)\n55.9\n110.1\n125.6\n129.9\n134.1\n144.5\n103.8\n155.1\nShare of \ndismantled \ncapacity \ncaptured(%)\n36.0\n71.0\n80.9\n83.8\n86.4\n93.2\n67\n100\nNet number of \nturbine repowering \nfactor\n2.93\n1.50\n1.31\n1.26\n1.23\n1.11\n1.58\n1\nNet capacity \nrepowering factor\n13.10\n6.65\n5.83\n5.63\n5.46\n5.06\n7.05\n4.72\nThe bottom two rows denote the net repowering factor, that is, the ratio between commissioned and dismantled turbine capacity in repowering projects\nNature Energy | VOL 5 | December 2020 | 1012\u20131021 | www.nature.com/natureenergy\n1018\n\nArticles\nNATure Energy\nfrom 2019 to primo 2020. Decommissioning of turbines for one project \ncommenced in November 2019, whereas the other was postponed into 2020.\nData sources. We constructed the database by using three primary sources. \nThe first source was the master data registry of wind turbines from the Danish \nEnergy Agency31. The master data registry is a comprehensive dataset that \nconsists of postconstruction information. The dataset contains information, such \nas identification number, technical specifications of the turbine and Universal \nTransverse Mercator (UTM) coordinates, on all commissioned and dismantled \nturbines in Denmark. The Danish Energy Agency periodically updates the dataset \nunder European Union directive 2001/77/EC36. The second source was local \ndevelopment plans for wind projects, which we retracted online from the Danish \nBusiness Authority30. Local development plans set out the framework for future \nplanned projects and describe the project; they comprise information on the \nlocation, appearance, number of commissioned and, in most cases, also dismantled \nturbines and an EIA. The third source was interviews with repowering wind \ndevelopers in Denmark.\nData collection by desktop research. We manually reviewed all the approved and \nundertaken local development plans for wind turbines from 2012 to 2019. For each \nlocal development plan, we proceeded as follows.\nUsing spatial analysis in the proximity of the local development plan, we \nidentified and connected commissioned and dismantled turbines from the master \ndata registry to the respective local development plan in our database. To identify \nthe commissioned turbines, the local development plan specifies exact locations \non maps. By translating these locations into UTM coordinates in Google Earth, \nwe were able to locate the matching commissioned turbines in the master data \nregistry of wind turbines. The local development plan often does not provide exact \nlocations for the dismantled turbines. Instead, the local development plan describes \nthe location of the existing turbines within the text. Consequently, using the \nEuclidean distance formula with the UTM coordinates, we calculated the distance \nbetween each dismantled turbine and the centre point of the local development \nplan. Next, we sorted the distances according to the shortest value. Using the sorted \nlist of dismantled wind turbines, we used Google Earth to find existing turbines \nthat best matched the location of those mentioned in the local development plan. \nThis evaluation served two purposes, to identify the turbines mentioned in the \nlocal development plan and to check if there were any unmentioned dismantled \nwind turbines in the nearby area that had a possible connection to the local \ndevelopment plan.\nAfter the spatial analysis, we categorized the project from the local \ndevelopment plan as either a greenfield, repowering or demonstration project. \nThen, we searched through the local development plan and the EIA to identify \nthe reason for dismantling each turbine in repowering projects. If the local \ndevelopment plan or the EIA mentioned a specific reason, we included that \ninformation in our database, mostly for later verification in the interviews. Finally, \nwe searched through the local development plan to obtain any contact details \nregarding the developer of the project. To account for missing details on the \ndeveloper, we also drew on web searches of property valuation settlements, news \narticles and company announcements.\nTo ensure that we identified all the approved wind farm projects from 2012 to \n2019, we consistently checked if we had assigned all commissioned wind turbines \nfrom the master data registry of wind turbines to a local development plan in our \ndatabase. In the case of any unassigned commissioned wind turbines, we used the \nassociated UTM coordinates together with an interactive map provided by the \nDanish Business Authority32 to detect if any local development plan existed in the \narea. As a result, we identified all wind projects in Denmark from 2012 to 2019.\nData collection by interviews. We conducted five in-depth semi-structured \ninterviews with the leading wind developers in Denmark, which comprised \n52.8% of all the repowering projects between 2012 and 2019 and 72.4% of all \nthe dismantled turbines. Additionally, we conducted eight informal interviews \nwith local wind developers, which comprised 22.2% of the approved repowering \nprojects in the same period, or 16.3% of the dismantled turbines. Overall, our \ninterviews cover 89% of all the dismantled turbines in repowering projects in the \ndesignated time frame, and 91% of the dismantled capacity in repowering projects. \nFor each project covered, we received responses on all the turbines in the respective \nproject. Using our interviews and the special planning reports, we identified a \nspecific reason for 94% of the turbines dismantled in repowering projects.\nWe conducted the interviews by phone between March 2019 and September \n2019. The in-depth interviews lasted 45\u201370\u2009min. We recorded and transcribed the \ninterviews. The informal interviews lasted 10\u221225\u2009min. Here, we took handwritten \nnotes. All the interviews were conducted in Danish. In both interview types, we \ninformed the interviewees that we would reveal neither their identity nor their \naffiliation, according to the Chatham House Rule37.\nUsing a protocol, we structured the in-depth interviews into two main parts. In \nthe first part, we asked the interviewees five open questions:\n\t(1)\t How would you define a repowering project?\n\t(2)\t Can you delineate the motivations behind the dismantling of wind turbines \nduring a repowering project?\n\t(3)\t What would be a required minimum space between planned and existing \nwind turbines?\n\t(4)\t Does your company have any guidelines in selecting existing wind turbines \nfor dismantling?\n\t(5)\t What happened with the dismantled wind turbines?\nIn the second part of the protocol, we asked project-specific questions to \nestablish the motivations behind dismantling each turbine in the respective \nrepowering projects. In the informal interviews, we only focused on the second \npart of the protocol.\nAfter the interviews, we analysed the transcripts and notes to identify and \nclassify the reasons for dismantling wind turbines in repowering projects. Based \non this, we found eight mutually exclusive categories for dismantling turbines in \nrepowering projects, as described below.\nSpace. Commonly, developers construct repowering projects at existing project \nsites. Therefore, it is often necessary to dismantle existing wind turbines on that \nsite because of physical constraints. Developers reported that these physical \nconstraints cover the land use needed for the installation and operation of the \nindividual new planned wind turbine, such as a space requirement for access roads, \nsetbacks, foundation and grid infrastructure.\nAside from the statements given by the interviewees and for consistency \nreasons, we categorized all the existing wind turbines dismantled within the \nproximity of 1.5 times the total height of the new wind turbine in this category. \nWe consider the value of 1.5 as a minimum distance required to prevent a collapse \nof the turbine tower, a crash of a wind turbine or a detached blade from the hub. \nNote that this is our consolidated estimate of a minimum distance. When asked \nabout a required minimum distance between planned and existing wind turbines, \nthe interviewees gave no fully conclusive answer. In an additional legal analysis (in \nwhich we reviewed primary sources of law), we found no state-level legislation or \nstatutes concerning a minimum distance between wind turbines. Additionally, we \nexamined legal statutes in all municipalities involved with repowering projects. We \nonly found guidelines that specified a distance that corresponded to 3\u22125 times the \nrotor diameter between turbines when planning new wind farms38. These guidelines \ndo not consider the relationship between existing and planned wind turbines where \nthe hub height differs, and wake effects might be negligible. Another guideline \nstates a minimum distance of one times the total height of the wind turbine to main \nroads and railroads, in case of crash down or falling ice debris38.\nNoise. Wind turbines generate noise as a result of moving mechanical parts and \nthe aerodynamic effects of air passing over the blades. According to Danish noise \nregulation, a developer must calculate the cumulative noise emissions from the \nplanned and existing wind turbines at a property if the difference in noise emission \nfrom the planned and the existing wind turbines is less than 15\u2009dB (ref. 39). The \ncumulative noise emissions include outdoor open space areas near the property \nand low-frequency noise indoors. If the cumulative noise emissions violate the \nnoise limits, the developer can either choose to reduce the noise from the planned \nwind turbines, and thereby reduce the generation output, or dismantle the existing \nwind turbines.\nAesthetics. Aesthetics are defined as the visible impact on the landscape of \nindividual wind turbines. The growing height and rotor diameter means wind \nturbines are more visible over greater distances within the local community. \nSome citizens feel affected by the presence of wind turbines with regard to a \n\u2018disfigurement\u2019 of the landscape. This feeling is subjective: some see wind turbines \nas a positive element in the transition to sustainability, others may see them as a \ndisturbance to scenic or aesthetic values.\nUsing a visual impact analysis, the EIA includes an evaluation of the \ndisturbances of aesthetic values within 28 times the total height of the planned \nwind turbines. The evaluation focuses on location, design and the interaction \nbetween the proposed new wind turbines and the existing infrastructure. Several \ndevelopers mentioned in our interviews that the difference in height and rotational \nspeed between planned and existing wind turbines are two of the main aesthetic \nfactors that cause the dismantling of existing turbines in repowering projects. \nOther aesthetic factors include the number of turbines in a location, the distance \nbetween farms and the array layout in relation to other existing wind farms. \nGuidelines exist to portray and evaluate the visual impact40; however, the final \ndecision rests with the municipality.\nAesthetics or noise. This category comprises those turbines that interviewees  \ncould not place into only one of the two categories, but for which they  \nwere certain it was one of these two reasons. Note that for a turbine to be placed in \nthis category, interviewees had to explicitly exclude politics as a possible reason for \ndismantling.\nPolitics. Politics may have an impact on the dismantling of existing turbines during \nthe local development and planning dialogues, where informal requests by local \npoliticians, proactive suggestions by developers or voluntary agreements by both \nparties come into play. In Denmark, wind project developers must obtain approval \nfrom the majority of local politicians in the municipality. Therefore, the developers \nNature Energy | VOL 5 | December 2020 | 1012\u20131021 | www.nature.com/natureenergy\n1019\n\nArticles\nNATure Energy\nengage with individual politicians during the planning process. During these \nnegotiations, developers are sometimes met with informal requests to dismantle \nexisting turbines located in the municipality. Developers see these informal \nrequests as a requirement to achieve support for the project and the dismantling as \na necessary action to receive approval for the project.\nPolitics or aesthetics. This category comprises those turbines that are dismantled \neither because of aesthetics or politics. Interviewees mentioned that local \npoliticians tend to use the disturbance of aesthetic values in the landscape to argue \nfor the removal of existing turbines. Using the visual impact illustrations from \nthe EIA, local politicians would address their view on the disturbance in aesthetic \nvalues and argue for the dismantling of existing turbines. As no firm regulation \nexists, but only guidelines for evaluating visual impact, it was often impossible \nfor interviewees to select between aesthetic or political motivations behind the \ndismantling requirements of a specific turbine.\nPolitics, aesthetics or noise. For turbines in this category, interviewees excluded \nspace as a reason for dismantling, but could not point to any other individual \nreason. Many of the interviewees noted that, during the early evaluation of the new \ndevelopment site, potentially problematic existing turbines are already identified \non a more intuitive basis or from experience. Assuming the removal of those \nexisting turbines before entering the permitting process can save time on wind \nfarm design, noise calculations and the visual impact assessment. This makes \nit difficult to (in retrospect) point towards a specific dismantling reason, as, for \nexample, the respective analyses for noise and aesthetics were not undertaken. For \nturbines in this category, interviewees could not rule out politics as a reason, either. \nHere, some noted that it was common to proactively offer dismantling turbines \nthat are known to be unwanted by politicians in the municipality to create a good \nstart for negotiation during the permitting process.\nUnclassified. All the dismantled turbines that the local development plans or the \ninterviews did not cover are put into this category.\nHaving identified and defined the categories for dismantling turbines in \nrepowering projects, we analysed the interview transcripts and notes  \nand added categories to the identified dismantled turbines in our database.  \nOf the 221 dismantled repowering turbines, only two turbines remain  \nwithout category.\nValue of lost production. We approximated the value of lost production \nspecifically for each turbine dismantled in repowering projects by multiplying \nits annual production averaged over its lifetime (obtained from the master data \nregistry of wind turbines31) with the wind-weighted electricity price41 for Denmark \nin each year from the dismantling of the wind turbine until the respective wind \nturbine would have reached a lifetime of 24.5 years (which corresponds to the \naverage lifetime of non-repowered dismantled turbines). We approximated \noperations and maintenance cost assuming \u20ac30\u201335\u2009MWh\u20131 (ref. 35), an expectedly \nhigh level because of the turbine age (which is between 18.6 and 24.4 years during \nthe lost years of production).\nReporting summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nThe dataset generated during the current study is published as Supplementary Data \n1 (spreadsheet format). Although the dataset contains all quantitative data collected \nfrom the different sources, we cannot disclose specific disaggregate information \nfrom the interviews, which have been conducted under the Chatham House rules. \nInterview transcripts and notes may be requested from the authors, but will only be \nhanded out after explicit consent from the interviewees. Source data are provided \nwith this paper.\nReceived: 13 March 2020; Accepted: 25 September 2020;  \nPublished online: 2 November 2020\nReferences\n\t1.\t Global Wind Energy Report 2018 (Global Wind Council Energy, 2019); \nhttps://gwec.net/wp-content/uploads/2019/04/\nGWEC-Global-Wind-Report-2018.pdf\n\t2.\t Wind Energy in Europe in 2018\u2014Trends and Statistics (WindEurope, 2019); \nwww.windeurope.org/wp-content/uploads/files/about-wind/statistics/\nWindEurope-Annual-Statistics-2018.pdf\n\t3.\t Rekord Lav CO2-Udledning fra Danskernes Elforbrug i 2019 (Energinet, 2020); \nwww.energinet.dk/Om-nyheder/Nyheder/2020/01/16/Rekord-lav-C\nO2udledning-fra-danskernes-elforbrug-i-2019\n\t4.\t Lantz, E., Leventhal, M. & Baring-Gould, I. Wind Power Project Repowering: \nFinancial Feasibility, Decision Drivers, and Supply Chain Effects. Prepared \nunder Task no. WE11.0630, Technical Report NREL/TP-6A20-60535 (NREL, \n2013); www.nrel.gov/docs/fy14osti/60535.pdf\n\t5.\t Silvosa, A. C., G\u00f3mez, G. I. & Del R\u00edo, P. Analyzing the techno-economic \ndeterminants for the repowering of Wind Farms. Eng. Econ. 58,  \n282\u2013303 (2013).\n\t6.\t Ziegler, L., Gonzalez, E., Rubert, T., Smolka, U. & Melero, J. J. Lifetime \nextension of onshore wind turbines: a review covering Germany,  \nSpain, Denmark, and the UK. Renew. Sustain. Energy Rev. 82,  \n1261\u20131271 (2018).\n\t7.\t Scherhaufer, P., H\u00f6ltinger, S., Salak, B., Schauppenlehner, T. & Schmidt, J. \nPatterns of acceptance and non-acceptance within energy landscapes:  \na case study on wind energy expansion in Austria. Energy Policy 109, \n863\u2013870 (2017).\n\t8.\t Smallwood, K. S. in Wind Energy and Wildlife Interactions (ed. K\u00f6ppel. J.) \n175\u2013187 (Springer 2017).\n\t9.\t Steininger, G. Wind turbine repowering is on the horizon (L. Buchsbaum and \nS. Patel, interviewers) POWER Magazine (2016); https://www.powermag.com/\nwind-turbine-repowering-horizon/\n\t10.\tVillena-Ruiz, R., Ramirez, F. J., Honrubia-Escribano, A. & G\u00f3mez-L\u00e1zaro, E. \nA techno-economic analysis of a real wind farm repowering experience: the \nMalpica case. Energy Convers. Manag. 172, 182\u2013199 (2018).\n\t11.\tFilgueira, A., Seijo, M. A., Munoz, E., Castro, L. & Piegari, L. Technical and \neconomic study of two repowered wind farms in Bustelo and San Xo\u00e1n, \n24.7\u2009MW and 15.84\u2009MW respectively. In 2009 International Conference on \nClean Electrical Power (ICCEP) (ICCEP, 2009); https://doi.org/10.1109/\nICCEP.2009.5211998\n\t12.\tPrabu, T. & Kottayil, S. K. Repowering a wind farm\u2014a techno-economic \napproach. Wind Eng. 39, 385\u2013397 (2015).\n\t13.\tNivedh, B., Devi, R. & Sreevalsan, E. Repowering of wind farms\u2014a case \nstudy. Wind Eng. 37, 137\u2013150 (2013).\n\t14.\tPaul, A. & Prabu, T. Technical and economic feasibility study on repowering \nof wind farms. Indian J. Sci. Technol. 9, 1\u20139 (2016).\n\t15.\tGoyal, M. Repowering\u2014next big thing in India. Renew. Sustain. Energy Rev. \n14, 1400\u20131409 (2010).\n\t16.\tDel R\u00edo, P., Calvo Silvosa, A. & Iglesias G\u00f3mez, G. Policies and design \nelements for the repowering of wind farms: a qualitative analysis of different \noptions. Energy Policy 39, 1897\u20131908 (2011).\n\t17.\tSantos-Alamillos, F. J., Thomaidis, N. S., Usaola-Garc\u00eda, J., Ruiz-Arias, J. A. & \nPozo-V\u00e1zquez, D. Exploring the mean-variance portfolio optimization \napproach for planning wind repowering actions in Spain. Renew. Energy 106, \n335\u2013342 (2017).\n\t18.\tM\u00f6ller, B. Changing wind-power landscapes: regional assessment of visual \nimpact on land use and population in Northern Jutland, Denmark. Appl. \nEnergy 83, 477\u2013494 (2006).\n\t19.\tHimpler, S. & Madlener, R. Optimal timing of wind farm repowering: a \ntwo-factor real options analysis. J. Energy Mark. 7, 3\u201334 (2014).\n\t20.\tWind Energy in Europe: Outlook to 2023 (WindEurope, 2019); https://\nwindeurope.org/about-wind/reports/\nwind-energy-in-europe-outlook-to-2023/#download\n\t21.\tLacal-Ar\u00e1ntegu, R., Iglesias, A. & Mar\u00eda Yusta, J. Technology effects in \nrepowering wind turbines. Wind Energy 23, 660\u2013675 (2019).\n\t22.\tFrant\u00e1l, B. Have local government and public expectations of wind energy \nproject benefits been met? Implications for repowering schemes. J. Environ. \nPolicy Plan. 17, 217\u2013236 (2015).\n\t23.\tMart\u00ednez, E., Latorre-Biel, J. I., Jim\u00e9nez, E., Sanz, F. & Blanco, J. Life cycle \nassessment of a wind farm repowering process. Renew. Sustain. Energy Rev. \n93, 260\u2013271 (2018).\n\t24.\tManchado, C. et al. Wind farm repowering guided by visual impact criteria. \nRenew. Energy 135, 197\u2013207 (2019).\n\t25.\tH\u00f6tker, H., Thomsen, K.-M. & Jeromin, H. Impacts on Biodiversity of \nExploitation of Renewable Energy Sources: the Example of Birds and Bats\u2014\nFacts, Gaps in Knowledge, Demands for Further Research, and Ornithological \nGuidelines for the Development of Renewable Energy Exploitation \n(Michael-Otto-Institut for the Nature and Biodiversity Conservation  \nUnion, 2006).\n\t26.\tSmallwood, K. S. & Karas, B. Avian and bat fatality rates at old-generation \nand repowered wind turbines in California. J. Wildl. Manage. 73,  \n1062\u20131071 (2009).\n\t27.\tMarques, A. T. et al. Understanding bird collisions at wind farms: an updated \nreview on the causes and possible mitigation strategies. Biol. Conserv. 179, \n40\u201352 (2014).\n\t28.\tFerri, V., Battisti, C. & Soccini, C. Bats in a Mediterranean mountainous \nlandscape: does wind farm repowering induce changes at assemblage and \nspecies level? Environ. Manage. 57, 1240\u20131246 (2016).\n\t29.\tColmenar-Santos, A., Camp\u00ed\u00f1ez-Romero, S., P\u00e9rez-Molina, C. & Mur-P\u00e9rez, \nF. Repowering: An actual possibility for wind energy in Spain  \nin a new scenario without feed-in-tariffs. Renew. Sust. Energy Rev. 41, \n319\u2013337 (2015).\n\t30.\tSearchlist over all Plan Proposals and Plans Adopted Pursuant to the \nPlanning Act (Danish Business Authority, 2019); kort.plandata.dk/\nsearchlist/#/\nNature Energy | VOL 5 | December 2020 | 1012\u20131021 | www.nature.com/natureenergy\n1020\n\nArticles\nNATure Energy\n\t31.\tMaster Data Register for Wind Turbines\u2014Data on Operating and \nDecommissioned Wind Turbines (Ultimo 02 2020) (Danish Energy Agency, \naccessed 26 March 2020; www.ens.dk/en/our-services/\nstatistics-data-key-figures-and-energy-maps/overview-energy-sector\n\t32.\tAfgifts og Tilskudsanalysen p\u00e5 Energiomr\u00e5det\u2014Delanalyse 6, Fremtidigt Tilskud \ntil Landvind (Danish Ministry of Taxation, 2018); www.skm.dk/media/5361/\nafgifts-og-tilskudsanalysen-delanalyse-6.pdf\n\t33.\tBekendtg\u00f8relse om Planl\u00e6gning for og Tilladelse til Opstilling af Vindm\u00f8ller \n(Danish Ministry of Industry, Business and Financial Affairs, 2019); www.\nretsinformation.dk/Forms/R0710.aspx?id=210040\n\t34.\tVindm\u00f8lleindustrien som Historisk Flagskib (Danish Energy Agency, 2011); \nwww.ens.dk/sites/ens.dk/files/Vindenergi/vindmoelleindustrien_historisk_\nflagskib.pdf\n\t35.\tWind Energy Generation (Strategic Energy Technologies Information System, \nEuropean Commission, 2013); setis.ec.europa.eu/system/files/Technology_\nInformation_Sheet_Wind_Energy_Generation.pdf\n\t36.\tDirective 2001/77/EC on the Promotion of Electricity Produced from Renewable \nEnergy Sources in the Internal Energy Market OJ L 283/33 (European \nParliament and Council, 2001).\n\t37.\tChatham House Rule (Chatham House, 2020); www.chathamhouse.org/\nchatham-house-rule\n\t38.\tVejledning om planl\u00e6gning for og tilladelse til opstilling af vindm\u00f8ller  \n(Danish Nature Agency under the Ministry of Environment and Food of \nDenmark., 2015); www.naturstyrelsen.dk/media/131731/vejledning_ \n06012015_web.pdf\n\t39.\tSt\u00f8j fra Vindm\u00f8ller\u2013Vejledning fra Milj\u00f8styrelsen nr 1 (Environmental \nProtection Agency under the Ministry of Environment and Food of \nDenmark, 2012); www2.mst.dk/Udgiv/\npublikationer/2012/05/978-87-92903-08-2.pdf\n\t40.\tVejledning om VVM i Planloven (Danish Nature Agency under the Ministry \nof Environment and Food of Denmark, 2009); www.naturstyrelsen.dk/media/\nnst/9948968/vvm_vejledning2.pdf\n\t41.\tIntegration of Wind Power (Ea Energianalyse, 2015); www.ea-energianalyse.\ndk/wp-content/uploads/2020/02/1550_Integration_vindkraft_viking_link_og_\nandre_tiltag-2.pdf\n\t42.\tUihlein, A., Telsnig, T. & Vazquez Hernandez, C. JRC Wind Energy Database \n(Joint Research Centre of the European Commission, 2019).\n\t43.\tLacal-Ar\u00e1ntegui, R. & Uihlein, A. Repowering wind turbines\u2014analysis of the \neffects of technology substitution in repowered wind farms. Appl. Energy 23, \n660\u2013675 (2020).\nAcknowledgements\nThis work constitutes a contribution to the research in the international working group \nIEA TCP Wind Task 26 (Cost of Wind Energy). We thank all the members of the group \nfor their extraordinary collaboration and essential comments on our work that formed \nthis article. The work was in part funded by the Danish Public Energy Technology \nDevelopment and Demonstration Program (EUDP), project no. 64018-0577. This Article \nconstitutes a contribution from the European Commission to IEA Task 26 research. \nThe views expressed are purely those of the authors and may not in any circumstances \nbe regarded as stating an official position of the European Commission. This work \nwas authored (in part) by the National Renewable Energy Laboratory, operated by the \nAlliance for Sustainable Energy, LLC, for the US Department of Energy under Contract \nno. DE-AC36-08GO28308. Funding was provided by the US Department of Energy \nOffice of Energy Efficiency and Renewable Energy Wind Energy Technologies Office. \nThe views expressed in the Article do not necessarily represent the views of the DOE or \nthe US Government. The US Government retains and the publisher, by accepting the \narticle for publication, acknowledges that the US Government retains a nonexclusive, \npaid-up, irrevocable, worldwide license to publish or reproduce the published form of \nthis work, or allow others to do so, for US Government purposes.\nAuthor contributions\nL.K. and M.K.J. conceived the study, developed the analysis and undertook the \ninterviews. M.K.J. led the data processing and analysis, with support from L.K. E.L. and \nT.T. contributed with international data. All the authors contributed to data analysis and \ninterpretation. All the authors wrote and edited the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-020-00717-1.\nCorrespondence and requests for materials should be addressed to L.K.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 National Renewable Energy Laboratory under exclusive licence to Springer Nature \nLimited 2020\nNature Energy | VOL 5 | December 2020 | 1012\u20131021 | www.nature.com/natureenergy\n1021\n\n1\nnature research  |  reporting summary\nApril 2020\nCorresponding author(s):\nLena Kitzing\nLast updated by author(s): Aug 20, 2020\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nWe manually collected data from different online sources and interviews (see below).\nData analysis\nn/a. Manual data processing, see below.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and \nreviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\n We collected data manually from different sources:  \n1) master data registry of wind turbines from the Danish Energy Agency (Excel sheet, online available, regularly updated, version),  \nhttps://ens.dk/en/our-services/statistics-data-key-figures-and-energy-maps/overview-energy-sector, downloadable excel sheet, Version upladed on 25th of March \n2020. \n2) local development plans from the Danish Business Authority ('Searchlist over all plan proposals and plans adopted pursuant to the Planning Act', PDF documents), \nhttp://kort.plandata.dk/searchlist/#/  \n3) semi-structured interviews with project developers (see below) \n \n\n2\nnature research  |  reporting summary\nApril 2020\nWe make the final dataset available online together with the published paper. The data set contains the full project data and all data underlying the figures and \ntables in the manuscript. \n \nWhile the data set contains all quantitative data collected from the different sources, we cannot disclose specific disaggregate information from the interviews, \nwhich have been conducted under the Chatham House rules. Interview transcripts and notes may be requested from the authors but will only be handed out after \nexplicit consent from the interviewees.\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nMixed-methods study with quantitative data collection in both manual online data search and interviews, as well as qualitative \nclassification of turbine dismantling reasons by interviewees.\nResearch sample\nOur quantitative data sample contains all wind development projects in Denmark between 2012 and 2019. \nOur interviews covered 27 out of 36 repowering projects in total (19 ( 52.8%) projects were covered by in-depth interviews;   \n8 (22.2%) projects were covered by shorter phone-conversations), corresponding to 91% of all dismantled turbine capacity in \nrepowering projects in Denmark between 2012 and 2019.\nSampling strategy\nWe excluded the years before 2001 because of discontinued support regulations that would have produced different and obsolete \nrepowering decisions. For the same reason, we also excluded the years 2001 to 2011, because political incentive programs for \nrepowering were in place. \nWe included all wind energy projects that have been developed, approved, and where project execution has started until the end of \n2019. We have included two repowering projects, with a total of 106.2 MW, for which the commission date was rescheduled from \n2019 to primo 2020. \nThe sample of developers we conducted interviews with was based on a major criterion: they needed to have developed at least one \nrepowering project in Denmark during 2012-2019. We conducted extended semi-structured interviews with all developers who are \noperating nationwide and added shorter interviews with local developers until we reached a broad coverage of projects, \nrepresenting above 90% of all dismantled capacity in repowering projects.\nData collection\nTo ensure that we had constructed a comprehensive dataset, we consistently checked if the two primary data sources aligned, i.e. \nthat we had assigned all commissioned wind turbines from the master data registry to a local development plan. We ensured \nalignment in an iterative process, in which we conducted an online spatial analysis or all still unassigned commissioned wind turbines \nused the associated UTM coordinates (which we translated from the master registry using Google Earth) together with an interactive \nmap provided by the Danish Business Authority to detect if any local development plan existed in the area. We then checked the \nrespective additional local development plans from the search list. We continued until all turbines were assigned a project. \nIn the interviews, we asked specifically for each project about the association of specific turbines with the project and their individual \ndismantling reasons. For each project covered, we received responses on all turbines in the particular project. We asked open-ended \nquestions about the reasons for dismantling each turbine and constructed the classification categories after the interviews had been \nfinalised. The interviews were conducted in Danish.\nTiming\nThe main elements of the quantitative data were collected between May 2018 and January 2019, and then updated with concurrent \nevents until December 2019. In the end of March 2020, we implemented some adjustments that were made to the master registry of \nwind turbines (version of 25 March 2020). \nThe interviews were conducted between March 2019 and September 2019.\nData exclusions\nOut of the 102 identified wind projects, we excluded 8 experimental test projects comprising dismantling of 70 test turbines, as these \nturbines are subject to different regulations. We also excluded 77 dismantled household wind turbines, i.e., turbines with a \ngeneration capacity of 25 kW or less, as there are also different regulations for these types of turbines. Further, we excluded a \ndismantling project that was part of an infrastructure development project, where 26 turbines were dismantled because of the \nconstruction of the Femern tunnel in Southern Denmark. \nNon-participation\nAll project developers we succesfully established contact with have agreed to participate in the interviews. We received responses on \nall specific turbines in all discussed projects.\nRandomization\nnot applicable\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \n\n3\nnature research  |  reporting summary\nApril 2020\nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology and archaeology\nAnimals and other organisms\nHuman research participants\nClinical data\nDual use research of concern\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nHuman research participants\nPolicy information about studies involving human research participants\nPopulation characteristics\nOur research focused on business activities for which we collected information through human interaction (interviews) with \ncompany representatives. The humans were, thus, information providers and not the subject of our investigation. We had 12 \nindividual participants in the interviews and phone conversations. All participants are business professionals, working as \nproject development experts in their respective companies. 11 participants were male, 1 participant female.\nRecruitment\nWe obtained information about the responsible project developers for each relevant company via the internet. Typically, \ncontact details were provided on project homepages. We approached potential interviewees by e-mail and by phone. In few \ncases, the approached person established contact to a more suitable interviewee in the company. \nEthics oversight\nInteraction with human participants was undertaken by DTU according to Danish regulation. All interactions followed \nChatham House rules. Informed consent was obtained from all participants. All data is anonymised and stored on Danish \ninstitutional servers according to GDPR regulation.\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\n\n\n    Scientific Research Findings:", "answer": "Through detailed project\u2011level analysis, our study reveals that wind energy repowering involves consideration of more than just end\u2011of\u2011life replacement and space constraints. Dismantling decisions are multifaceted and result in the dismantling of a larger number of turbines and at an earlier age than physically required. We show this for Denmark, where more than a third (38%) of recent wind energy projects (2012\u20132019) involved repowering. We found that repowered turbines were dismantled at an average age of\u00a018.6\u00a0years \u2014\u00a05.8\u00a0years less than those without repowering. Only two thirds (67%) of the capacity dismantled in repowering projects was related to physical space needed for new turbines. This resulted in considerably lower achieved net capacity increases per project, with only\u00a04.72\u00a0megawatts commissioned per dismantled megawatt, as compared to\u00a07.05\u00a0megawatts if only space\u2011related dismantling were considered. Dismantling reasons besides physical space needs included regulation (noise\u2011related,\u00a08\u201317% of capacity), development principles (aesthetics,\u00a07\u201320%), and political bargaining (4\u201313%). We find that repowering is also a negotiated process between developer and host community, used to reduce community impacts of wind turbines.", "id": 19}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-020-00719-z\n1RFF-CMCC European Institute on Economics and the Environment (EIEE), Centro Euro-Mediterraneo sui Cambiamenti Climatici, Milan, Italy. 2Department \nof Management, Economics and Industrial Engineering, Politecnico di Milano, Milan, Italy. 3Department of Economics, Management and Quantitative \nMethods (DEMM), University of Milan, Milan, Italy. 4Fondazione Pesenti, Milan, Italy. \u2709e-mail: giovanna.dadda@unimi.it\nS\nocial information programmes that provide information on \nthe actions or beliefs of others are widely used interventions \nto foster behavioural change in several domains1\u20134, which \ninclude residential resource conservation5\u201312. These programmes \ntypically feature two forms of feedback: descriptive and injunc-\ntive. In the case of residential energy, descriptive feedback generally \ntakes the form of information on other households\u2019 average energy \nconsumption, whereas injunctive feedback provides social approval \nfor energy savings. The combination of descriptive and injunctive \nfeedback within the standard design of social information pro-\ngrammes was inspired by the finding that descriptive information \nalone leads those who use less energy (low energy users) to increase \ntheir consumption, and that the addition of injunctive feedback in \nsupport of an energy conservation norm prevents this boomerang \neffect13. Injunctive information thus counterbalances descriptive \ninformation. However, the experimental evidence on the impact of \ndescriptive and injunctive information when they exert opposing \ninfluences on behaviour is mixed and mainly focuses on short-term \nor self-reported outcomes14,15.\nGiven the wide adoption of communication campaigns that \nrely on social information to promote behavioural change among \nboth policymakers and private firms, it is important to understand \nhow different programme features interact in real-world settings16. \nIndeed, impact evaluations of similar programmes find that they are \neffective in fostering energy savings, but that effect sizes vary widely \nacross contexts and individuals17. Prominent explanations for such \ndifferential responses rely on the heterogeneity of consumers\u2019 traits, \nsuch as beliefs18,19, misperceptions of one\u2019s compliance with the \nsocial norm20 or personal values21,22.\nHere we focus on the varying effect of specific features of these \nmessages, and particularly on how the salience, strength and consis-\ntency of the feedback they contain differ, and thus affect behaviour \ndifferently, across users. This could inform a more effective design \nand targeting of messages and provide more specific and nuanced \nguidance to prevent similar information campaigns from back-\nfiring16. First, we exploited the features of the standard design of \nhome energy reports and isolate the impact of changes in injunctive  \nfeedback. Specifically, we examined whether reinforcing the injunc-\ntive feedback has different effects on electricity use if it is accompa-\nnied by consistent descriptive feedback\u2014as is the case for those who \nuse more energy (high energy users), for whom both the injunc-\ntive and descriptive information encourage energy conservation\u2014\nor contrasting descriptive feedback\u2014as is the case for low energy \nusers, for whom conforming with the descriptive feedback entails \nconsumption increases, at odds with the injunctive feedback that \npraises energy saving. Second, we randomized descriptive or injunc-\ntive information that primes a social norm of energy conservation, \nand evaluated the effect of strengthening the injunctive feedback in \nthe presence of either the descriptive or the injunctive prime.\nWe propose a conceptual framework for understanding how dif-\nferent features of social information programmes impact energy \nconservation that can be articulated in a set of hypotheses, illus-\ntrated in Fig. 1. First, the effectiveness of a normative message is \nmaximized by the inclusion of consistent feedback of different types \n(that is, injunctive and descriptive; Fig. 1a). Second, when injunc-\ntive and descriptive feedbacks are in contrast (Fig. 1b), the strength \nof each single piece of information matters. The strength of the nor-\nmative feedback may depend on several factors highlighted in the \nliterature, from the recipient\u2019s beliefs on what relevant others think \nis socially approved of18,23 to the degree of consensus or ambigu-\nity around the norm conveyed by the information16,24. In our set-\nting, we hypothesize that the effect of the descriptive information \nincreases according to the difference between an individual\u2019s elec-\ntricity consumption and the average consumption of the reference \ngroup. The effect of injunctive information instead varies accord-\ning to the strength of social approval conveyed through visual cues \nand encouragement messages. Third, additional pieces of consistent \nfeedback of the same type produce smaller savings (Fig. 1c).\nOur results are in line with these hypotheses. First, we found \nsuggestive evidence that the standard social information message \ninduces larger savings among high electricity users who are exposed \nto consistent descriptive and injunctive feedback, compared with \nlow electricity users who are exposed to contrasting descriptive and \ninjunctive feedback. More importantly, reinforcing the injunctive \nThe interaction of descriptive and injunctive social \nnorms in promoting energy conservation\nJacopo Bonan1,2, Cristina Cattaneo\u200a \u200a1, Giovanna d\u2019Adda\u200a \u200a1,3,4\u2009\u2709 and Massimo Tavoni1,2\nBehavioural interventions that leverage social norms are widely used to foster energy conservation. For instance, home energy \nreports combine information on others\u2019 behaviour (descriptive feedback) and approval for norm compliant behaviour (injunc-\ntive feedback). In a randomized controlled trial, we investigated how descriptive and injunctive feedbacks interact to affect \nelectricity use, and evaluate the effects of additional normative feedback presented in the form of descriptive or injunctive \nenergy conservation norm primes. We found that consistent descriptive and injunctive feedback boosts the effectiveness of \nsocial information in inducing energy conservation. When descriptive and injunctive feedback are in conflict, conservation \nbehaviour is a function of the relative strength of the two types of feedback. Additional normative feedback produces smaller \ngains when it reinforces existing information of the same type. These results suggest complementarities between different \ntypes of normative messages rather than superiority of any one kind of feedback.\nNature Energy | VOL 5 | November 2020 | 900\u2013909 | www.nature.com/natureenergy\n900\n\nArticles\nNATure Energy\nfeedback has the largest effect among low electricity users exposed \nto a consistent descriptive prime. These findings are in line with \nthe notion that injunctive and descriptive feedbacks have a larger \nimpact on electricity conservation when they pull behaviour in the \nsame rather than in opposite directions. Second, reinforcing the \ninjunctive feedback led to a reduction in consumption, but only \namong customers with low electricity usage. This shows that the \nrelative strengths of the different types of feedback matters when \nthey are contrasting. Such a reinforcement has no effect on custom-\ners with high consumption. This demonstrates the limited effect of \nreinforcing one type of normative feedback within a message that \nalready contains two consistent pieces of normative information of \ndifferent types. This is further confirmed by the finding that rein-\nforcing the injunctive feedback has no effect among users exposed \nto a consistent injunctive prime. Together these findings suggest \nthat additional pieces of feedback have a larger impact when they \npull behaviour in the same direction and are of different types. \nOverall, our results support the presence of synergies between dif-\nferent types of feedback rather than the primacy of any one type of \nfeedback.\nField experiment\nOur setting consists of a randomized controlled trial implemented \nby an Italian energy company that provides almost half-a-million \nhouseholds with information on their electricity use relative to \nthat of their neighbours6,17. The social information is included in \na Home Energy Report distributed to customers via email (eHER). \nThe programme was rolled out in 2016 and involved 464,523 cus-\ntomers (n\u2009=\u2009418,178 treatment, n\u2009=\u200946,345 control). The core fea-\nture of the eHER is the neighbour comparison, which combines \ndescriptive and injunctive normative information. The descriptive \nnorm graphically compares the customer\u2019s electricity use over the \nprevious month with the average use in two reference groups: 100 \nsimilar customers who live nearby (that is, neighbours) and the \n15% most-efficient neighbours. The injunctive norm takes the form \nof thumbs-up symbols next to the descriptive norm graph: three \nthumbs up (\u2018excellent\u2019) for users who consume within the top 15% \nmost-efficient neighbours, two thumbs up (\u2018good\u2019) for those more \nefficient than the average neighbour and one thumbs up (\u2018you can \ndo better\u2019) for the others. Figures 2a,b shows the eHER for users \nreceiving three and two thumbs up, respectively.\nWe collaborated with the energy company to augment this set-up \nwith a message displayed at the bottom of the eHER delivered in \nApril\u2013May 2018. The utility randomly allocated half of the treated \nsample at that time (n\u2009=\u2009256,487) to receive either the descriptive \n(n\u2009=\u2009127,899) or the injunctive (n\u2009=\u2009128,588) message priming an \nenergy conservation norm (Fig. 2c, Supplementary Methods and \nSupplementary Fig. 1). The descriptive norm prime emphasizes that \na large majority of customers try to save energy, that is, adopt behav-\niours consistent with a social norm of saving electricity. The injunc-\ntive norm prime claims that a majority of customers hold electricity \nsaving as a personal value, which thus supports the belief that elec-\ntricity saving is approved by relevant others. The two primes use \nfellow customers of the same utility as the reference group. The \ninformation on energy saving behaviours and values featured in the \nprimes was taken from an online survey that we conducted with \nabout 3,000 utility customers (Methods).\nWe have access to data on monthly electricity consumption from \nJuly 2015 to December 2019. The daily average electricity usage, nor-\nmalized with respect to the control group consumption in the inter-\nvention period, was our main outcome variable. Pre-intervention \ndaily electricity usage in a month was calculated over the period \nJuly 2015 to June 2016. Our data also include information on the \ncontents of customers\u2019 reports and on whether customers open or \nclick on them. We provide details on the programme implementa-\ntion and data in the Methods, and descriptive statistics and balance \ntests in Supplementary Tables 1 and 2 and Supplementary Note 1. \nSamples are balanced across all available dimensions.\nImpact of the social information programme\nThe impact evaluation of the standard programme indicates a statis-\ntically significant reduction of normalized electricity usage in its first \nyear (coefficient\u2009=\u2009\u22120.353, standard error (s.e.)\u2009=\u20090.113, P\u2009=\u20090.002; \nequation (1) in Methods and Supplementary Table 3, column 1). \nThe impact of the treatment increases with baseline consumption, \nalthough this result is not robust to the measure of electricity con-\nsumption used, that is, discrete or continuous (Supplementary Table 3,  \ncolumns 2 and 5); its statistical significance varies with how the \nsample is defined (Supplementary Table 4) and with the time frame \nconsidered (Supplementary Tables 5\u20137) and it does not always hold \nafter multiple hypotheses corrections. Exploiting data on engage-\nment with the reports and on changes in feedback over time, we \nfound that the impact of social information is magnified among \nusers who actually read it and who experience upgrades in feedback \n(see Supplementary Tables 8 and 9 and Supplementary Note 2 for \nfurther details).\na\nb\nc\nHypothesis: combined effect is\nmaximized\nHypothesis: combined effect\ndepends on relative strength\nHypothesis: smaller combined\neffect\nEnergy conservation\nSign, strength and \namount of feedback\nEnergy conservation\nEnergy conservation\nSign, strength and \namount of feedback\nSign, strength and \namount of feedback\nFig. 1 | Hypothesized impact of injunctive and descriptive feedbacks in social information messages. a\u2013c, Hypothesized effects of consistent injunctive \nand descriptive (blue) feedbacks (a), contrasting injunctive (red) and descriptive feedbacks (b) and additional consistent feedbacks of the same type \n(c). The black curve represents the overall impact of the normative message. Injunctive feedback is shown in red, descriptive feedback is shown in blue \nand their combined effect is shown in purple. The horizontal axes indicate the sign, strength and amount of normative feedback, where positive feedback \nvalues imply messages that encourage electricity savings. The vertical axes represent the behavioural outcome where positive values are associated with \nbehaviour in compliance with the norm, which in our case is energy conservation.\nNature Energy | VOL 5 | November 2020 | 900\u2013909 | www.nature.com/natureenergy\n901\n\nArticles\nNATure Energy\nAlthough the magnitude of the average savings from the pro-\ngramme (\u22120.353%) is outside the range of those generated by similar \nones in the United States (minimum\u2009=\u20090.88%, maximum\u2009=\u20092.55%) \n(ref. 18), they are in line with the existing evidence from Europe19. \nVarious factors, such as lower average consumption in Europe than \nthat in the United States, the specific features of the programme we \nstudied or differences in beliefs across contexts, may be responsible \nfor these differences. The heterogeneous effects, although not robust \nand only marginally statistically significant, are qualitatively in line \nwith the existing evidence on the larger impact of social informa-\ntion on high electricity users17,20,25 and on the absence of boomerang \neffects among low users13.\nThese results provide initial, albeit weak, support for our concep-\ntual framework. For high users, normative and injunctive feedbacks \npull behaviour in the same direction, which results in a reduction \nin electricity almost twice as large as that in the average treatment \neffect. For low electricity users, conforming to the reference groups\u2019 \nbehaviour motivates a consumption increase (\u2019boomerang\u2019), but \nthe injunctive feedback included in the eHER counterbalances the \nnegative effect of the descriptive feedback. The injunctive feedback \ntherefore induces stronger behavioural reactions among high elec-\ntricity users, who are also exposed to the supporting descriptive \nfeedback, than that among low electricity users, for whom the two \ntypes of feedback are at odds. Although such an interpretation is \nonly suggestive based on the evidence presented so far, it shows how \nestablished findings are consistent with our conceptual framework.\nImpact of strengthening the injunctive feedback\nOur conceptualization can guide the analysis and interpreta-\ntion of the effect of intensifying the injunctive feedback, with the \ndescriptive feedback kept unchanged, for low and high users. We \nisolated the causal impact of the strength of the injunctive feed-\nback via a regression discontinuity (RD) estimation (Methods). We \nexploited the fact that the injunctive feedback (number of thumbs \nup) changes discretely as a customer\u2019s consumption crosses the two \nthresholds given by the average consumption of the efficient neigh-\nbours (three versus two thumbs-up cutoff), and the neighbours\u2019 \naverage consumption (two versus one thumbs-up cutoff), whereas \nthe descriptive feedback (bars of electricity use) changes continu-\nously across the thresholds. We focused on customers around the \nthresholds, whose assignment to a given injunctive feedback cat-\negory was plausibly random. Indeed, although customers that \nbelong to the three normative feedback categories differ on average \nalong various baseline characteristics (Supplementary Note 3 and \nSupplementary Table 10), individuals close to the thresholds are \nsimilar (Supplementary Tables 11 and 12, columns 1\u20133).\nWe conducted two separate RD estimations, one for each cutoff, \non the sample of treated customers who received the eHER sent in \nApril\u2013May 2018 (n\u2009=\u2009256,487) to allow a direct comparison with the \nanalysis presented below. In each estimation, we compared users \nin the two feedback categories adjacent to the cutoff and estimated \nthe marginal effect of receiving one additional thumbs up. Figure 3  \npresents the results in terms of level changes in electricity usage. \nAlthough there are no statistically significant changes in the effect \nof the eHER when crossing the threshold between the one and \ntwo thumbs up (Fig. 3a), the discrete shift in the injunctive norm \nreduces electricity use when moving from the two to three thumbs \nup (Fig. 3b). The corresponding empirical estimates are presented \nin Table 1 (columns 1 and 2).\nWe can attribute these effects to the social information contained \nin the report rather than to other content, namely electricity sav-\ning tips\u2014tips can only be accessed through a clickable link on the \nb\nInjunctive feedback  \nDescriptive feedback\nLink to tips\nDescriptive prime: Are you reducing energy consumption in your house? More than 80% of [name of utility]\ncustomers take actions to save energy.* Even little deeds can have a large impact. Discover our tips to \nconsume less and better.\nInjunctive prime: Is saving energy important to you? For more than 80% of [name of utility] customers saving \nenergy is an important value.* Even little deeds can have a large impact. Discover our tips to consume less and better.\n*Survey conducted with a representative sample of [name of utility] customers.\nLink to tips\nDescriptive or injunctive \nnorm prime (randomized)\nDescriptive or injunctive \nnorm prime (randomized)\nFIND WAYS TO CONSUME LESS AND BETTER\nFIND WAYS TO CONSUME LESS AND BETTER\na\nc\nFig. 2 | Home Energy Report. a,b, Layout and content of a Home Energy Report for a user receiving three thumbs up (a) and a user receiving two \nthumbs up (b). Both versions of the report contain the injunctive feedback, that is, the thumbs up (top), and the descriptive feedback, that is, the energy \nconsumption bars (bottom). c, Text of the randomized norm primes. Credit: Copyright 2016-2020 Oracle. All rights reserved.\nNature Energy | VOL 5 | November 2020 | 900\u2013909 | www.nature.com/natureenergy\n902\n\nArticles\nNATure Energy\nreport and we see no difference in click shares across the cutoffs \n(Supplementary Table 12). The impact of the shift in injunctive \nfeedback is persistent even after 6 and 12 months (Supplementary \nTable 13). The results are robust to different specifications of the RD \nestimate (Table 1).\nThese findings are consistent with our second hypothesis. \nStrengthening the approval for electricity savings delivered through \nthe injunctive feedback affects customers around the three versus \ntwo thumbs-up cutoff more because, for them, it reinforces the rela-\ntive strength of the injunctive information in the presence of con-\nflicting injunctive and descriptive feedback. On the contrary, and in \nagreement with our third hypothesis, for consumers around the two \nversus one thumbs-up cutoff, the reinforcement of the injunctive \nfeedback has a smaller marginal effect, as it adds strength to already \nconsistent feedback types.\nImpact of additional injunctive and descriptive feedback\nTo further test the impact of an additional piece of normative infor-\nmation, we combined the features of the standard eHER with the \nrandomized addition of descriptive and injunctive information \nthrough the primes. To exploit the interaction between the discon-\ntinuities in the eHER\u2019s injunctive feedback and the randomly deliv-\nered primes, we repeated the RD analysis across the two cutoffs (two \nversus one and three versus two thumbs up) separately for the sub-\nsamples of customers who received the two types of prime.\nThe results are reported in Fig. 4 and Table 1 (columns 3\u20136). \nAcross the cutoff between one and two thumbs up, we observe \nno statistically significant changes in consumption, regardless of \nwhether the descriptive (Fig. 4a) or the injunctive (Fig. 4b) prime \nis present. Conversely, a discrete shift in the injunctive feedback \nacross the three versus two cutoff causes electricity reduction, but \nonly when combined with the descriptive prime that nudges energy \nefficiency (Fig. 4c). The results are robust to adjustments for mul-\ntiple hypothesis testing (Table 1) and are persistent over longer time \nhorizons (Supplementary Table 13).\nThis evidence, consistent with our first hypothesis, suggests \nsynergies between different types of feedback: adding support-\nive descriptive information increases the impact of the shift in \ninjunctive feedback. The marginal contribution of additional \nfeedback of the same type instead decreases: customers exposed \nto the injunctive prime do not react to the reinforcement of the \ninjunctive feedback across the three versus two thumbs-up cutoffs  \n(Fig. 4d). Similarly, strengthening the injunctive feedback across the \ntwo versus one thumbs-up threshold makes no difference, regard-\nless of whether the descriptive (Fig. 4a) or injunctive information \n(Fig. 4b) is added. In this case, the descriptive and injunctive feed-\nback within the standard neighbour comparison already pull behav-\niour in the same direction. Further priming either type of normative \nfeedback does not generate incremental electricity conservation.\nTo determine whether the overall effect of crossing the three \nversus two thumbs-up threshold is exclusively due to the presence \nwithin the eHER of the descriptive prime, we performed the RD \nestimation on a standard eHER (February\u2013March 2018). We found \nstatistically significant effects (coefficient\u2009=\u2009\u22120.855, s.e.\u2009=\u20090.368, \nP\u2009=\u20090.02; Table 2, column 1). Therefore, the effect of reinforcing the \ninjunctive feedback for low users does not depend on the presence \nof the descriptive prime within the report. In addition, we observe \nthat the overall effect of crossing the three versus two thumbs-up \n5.5\n6\n6.5\n7\nAverage daily electricity usage (kWh)\n\u201330\n\u201320\n\u201310\n0\n10\n20\n30\nElectricity usage with respect to cutoff (kWh)\na\n3\n3.5\n4\n4.5\n5\nAverage daily electricity usage (kWh)\n\u201330\n\u201320\n\u201310\n0\n10\n20\n30\nElectricity usage with respect to cutoff (kWh)\nb\nFig. 3 | Impact of the injunctive feedback on electricity usage. a,b, Each dot represents the average daily electricity usage in the 3 months after the receipt \nof the April\u2013May 2018 eHER around the two versus one thumbs-up cutoff (n\u2009=\u2009216,328) (a) and three versus two thumbs-up cutoff (n\u2009=\u2009130,466) (b) \nwithin evenly spaced bins. The solid line represents the local linear fit, estimated separately on either side of the cutoff and the shaded area shows 95% \nconfidence intervals. The number of bins was selected through the integrated mean squared error (m.s.e). The running variable (horizontal axis) reports \nthe individual difference between each customer\u2019s monthly consumption in the period reported in the eHER and the relevant cutoff. The cutoff is then \nrepresented by the vertical line, which is set to zero. For positive values of the score, customers get an extra thumbs up with respect to those with negative \nvalues of the score. In a, customers on the right of the cutoff consume less than the average neighbour and more than the top 15% most efficient and get \ntwo thumbs up. In b, customers on the right of the cutoff consume within the 15% most efficient neighbours and get three thumbs up. Bandwidths (BWs) \nand kernel are set following the data-driven process described for formal RD estimations of impacts reported in Table 1.\nNature Energy | VOL 5 | November 2020 | 900\u2013909 | www.nature.com/natureenergy\n903\n\nArticles\nNATure Energy\nthreshold within the standard report is smaller than the same effect \nwhen the report is augmented with the descriptive prime (coeffi-\ncient\u2009=\u2009\u22122.426, s.e.\u2009=\u20090.619, P\u2009<\u20090.01; Table 1, column 3). This con-\nfirms that the combination of consistent descriptive and injunctive \ninformation boosts the effectiveness of social information in induc-\ning energy conservation.\nThe RD estimation provides a robust identification of the causal \neffects, which are, however, only local. We complemented it with an \nestimation of the heterogeneous impact of the primes by the num-\nber of thumbs up that customers receive (equation (3) in Methods). \nWe obtained similar results: although the descriptive prime does not \ninfluence consumption on average (coefficient\u2009=\u20090.088, s.e.\u2009=\u20090.149, \nP\u2009=\u20090.554; Table 3, column 1), it led to a negative and statistically sig-\nnificant decrease in consumption among customers who received \nthree thumbs up (coefficient\u2009=\u2009\u22120.959, s.e.\u2009=\u20090.284, P\u2009<\u20090.001;  \nTable 3, column 4). This negative effect was persistent over 6, 12 and \n18 months (Supplementary Note 4 and Supplementary Table 14). \nInterestingly, the effect of the descriptive prime on the entire group \nof customers who received three thumbs up is smaller than the RD \nestimates for the three versus two thumbs-up threshold combined \nwith the descriptive prime. Although these simple heterogeneity \nresults should be taken with caution, as thumbs up are correlated \nwith customers\u2019 characteristics (Supplementary Table 10), they can \nbe interpreted in light of our conceptual framework and suggest a \npotential determinant of the strength of descriptive information. \nThe sample of customers who received three thumbs up includes \nthe most efficient users, who are far from the three versus two \nthumbs-up threshold. The further customers are from the thresh-\nold, the larger the deviation between their own consumption and \nthe average electricity use, and therefore, we argue, the larger the \nstrength of the descriptive norm contained in the standard eHER. \nIn other words, we suggest that in our setting conformity motives \nbecome more influential the further away individuals are from the \ndescriptive norm.\nThis interpretation was confirmed by analysing specifically the \neffect of the descriptive prime among customers who experienced \nan upgrade in the injunctive feedback (from two to three thumbs up) \nrelative to the previous report. These customers are likely to overlap \nwith the customers included in the RD estimation, as being close \nto the three versus two thumbs-up threshold may result in down-\ngrades and upgrades between reports. The effect of the descriptive \nprime on these customers is in line with the RD estimates for the \nTable 1 | Regression discontinuity estimates of the impact of the injunctive norm and normative prime message on electricity usage\nAll\nDescriptive prime\nInjunctive prime\n1\n2\n3\n4\n5\n6\nThree versus two thumbs up\nConventional\n\u22121.217***\n\u22121.165**\n\u22122.426***\n\u22122.388***\n\u22120.0550\n0.209\n(0.461)\n(0.456)\n(0.619)\n(0.611)\n(0.686)\n(0.676)\n[0.001]\n[0.001]\n[1]\n[0.66]\nRobust bias-corrected\n\u22121.130**\n\u22121.077**\n\u22122.264***\n\u22122.349***\n\u22120.0470\n0.310\n(0.461)\n(0.456)\n(0.619)\n(0.611)\n(0.686)\n(0.676)\n[0.002]\n[0.001]\n[1]\n[0.68]\nObservations\n130,466\n130,466\n65,091\n65,091\n65,305\n65,305\nBW select method\n1\u2009m.s.e.\n2 m.s.e.\n1\u2009m.s.e.\n2\u2009m.s.e.\n1\u2009m.s.e.\n2\u2009m.s.e.\nBW above\n30.59\n35.67\n34.50\n40.27\n27.94\n39.15\nBW below\n30.59\n26.95\n34.50\n28.89\n27.94\n23.14\nEffective number of observations above\n31,226\n36,752\n17,852\n21,096\n10,698\n20,632\nEffective number of observations below\n23,216\n20,908\n12,669\n11,051\n13,900\n9,392\nTwo versus one thumbs up\nConventional\n0.254\n\u22120.564\n\u22120.426\n\u22121.046*\n0.491\n\u22120.161\n(0.531)\n(0.434)\n(0.704)\n(0.601)\n(0.691)\n(0.619)\n[1]\n[0.14]\n[1]\n[0.66]\nRobust bias-corrected\n0.488\n\u22120.409\n\u22120.118\n\u22120.883\n0.731\n\u22120.150\n(0.531)\n(0.434)\n(0.704)\n(0.601)\n(0.691)\n(0.619)\n[1]\n[0.27]\n[0.771]\n[0.68]\nObservations\n216,328\n216,328\n107,729\n107,729\n108,477\n108,477\nBW select method\n1\u2009m.s.e.\n2\u2009m.s.e.\n1\u2009m.s.e.\n2\u2009m.s.e.\n1\u2009m.s.e.\n2\u2009m.s.e.\nBW above\n30.10\n76.51\n35.93\n84.88\n33.80\n67.97\nBW below\n30.10\n33.90\n35.93\n37.67\n33.80\n30.86\nEffective number of observations above\n32,148\n40,038\n18,765\n22,115\n20,015\n18,306\nEffective number of observations below\n35,704\n69,226\n21,124\n36,967\n17,843\n31,980\nThe table shows the impact of the injunctive norm, that is, the number of thumbs up, on electricity usage, overall and by normative prime message. The outcome variable of the RD estimations is the \naverage daily energy usage (kWh) in the 3\u2009months after the receipt of the eHER augmented with the normative prime, normalized by the average control group consumption in the post period, around the \nthree versus two thumbs up and two versus one thumbs up. The estimations in columns 1 and 2 are on the whole sample, in columns 3 and 4 are on the sample of customers who received the descriptive \nnorm prime and in columns 5 and 6 are on the sample of those who received the injunctive norm prime. The first row shows the conventional RD estimate for the thumbs-up comparison, whereas the \nsecond corrects for bias48. BWs were selected to minimize the m.s.e.48,49. Odd columns use the same BW on either side of the cutoff, whereas even columns estimate separate BWs. Standard errors are \nclustered at the customer level in parentheses. FDR (false discovery rate)-adjusted q values are given in square brackets. ***P < 0.01, **P < 0.05, *P < 0.1.\nNature Energy | VOL 5 | November 2020 | 900\u2013909 | www.nature.com/natureenergy\n904\n\nArticles\nNATure Energy\nthree versus two thumbs-up threshold combined with the descrip-\ntive prime (Fig. 5) and larger than the average impact of the descrip-\ntive prime on the three thumbs-up subsample. Other factors may \ncontribute to these results, but we note that they are consistent with \nour argument that, when descriptive and normative expectations \ndiverge, the resulting behaviour is a function of the relative strength \nof the two types of feedback.\nConclusions\nOur findings have implications for the design of social information \nprogrammes that rely on the combination of different types of norms \nto maximize behavioural change. Similar programmes are used in \nseveral domains, such as tax compliance26,27, charitable giving28 or \nwater conservation29. According to our conceptual framework and \nempirical results, no single type of normative information is more \neffective in absolute terms. Policymakers should, instead, pay atten-\ntion to the type of normative feedback they include in their commu-\nnication, strive to diversify them, avoid conflicting information when \nit mitigates the desirable effects and exploit it otherwise, be aware of \nthe diminishing returns from additional pieces of social information \nand of the varying strength of conformity motives across individuals.\nOf course, our results may be specific to the context that we stud-\nied, and particularly to the formulation of injunctive and descrip-\ntive feedbacks that characterize the energy efficiency programme \nwe evaluated and the normative primes we designed. For example, \nthe wording and graphical representation of the injunctive feedback \nin the eHER of this study differ from those of widely evaluated stan-\ndard social information programmes6,17,18,21,30. Further investigations \n5.5\n6\n6.5\n7\nAverage daily\nelectricity usage (kWh)\n\u201330\n\u201320\n\u201310\n0\n10\n20\n30\n5.5\n6\n6.5\n7\n\u201330\n\u201320\n\u201310\n0\n10\n20\n30\n3\n3.5\n4\n4.5\n5\nAverage daily\nelectricity usage (kWh)\n\u201330\n\u201320\n\u201310\n0\n10\n20\n30\nElectricity usage with respect to cutoff (kWh)\n3\n3.5\n4\n4.5\n5\n\u201330\n\u201320\n\u201310\n0\n10\n20\n30\nElectricity usage with respect to cutoff (kWh)\na\nb\nc\nd\nFig. 4 | Heterogeneous impact of the normative primes at different injunctive feedback cutoffs. Each dot represents the average daily electricity  \nusage (kWh) in the 3 months after the receipt of the eHER augmented with the normative prime, within evenly spaced bins. The solid line represents  \nlocal linear fit, estimated separately on either side of the cutoff with the shaded area showing 95% confidence intervals. The number of bins is  \nselected through the integrated m.s.e. The running variable (horizontal axis) reports the individual difference between each customer\u2019s monthly \nconsumption (kWh) in the period reported in the eHER and the relevant cutoff. The running variable (horizontal axis) reports the individual difference \nbetween each customer\u2019s monthly consumption (kWh) in the period reported in the eHER and the relevant cutoff. For positive values of the score, \ncustomers get an extra thumbs up with respect to those with negative values of the score. a, Customers who received the descriptive norm prime around \nthe two versus one thumbs-up cutoff (n\u2009=\u2009107,729). b, Injunctive norm prime around the two versus one thumbs-up cutoff (n\u2009=\u2009108,477). c, Descriptive \nnorm prime around the three versus two thumbs-up cutoff (n\u2009=\u200965,091). d, Injunctive norm prime around the three versus two thumbs-up cutoff \n(n\u2009=\u200965,305).\nTable 2 | Regression discontinuity estimates of the impact of \nthe injunctive norm on electricity usage\nThree versus two \nthumbs up\nTwo versus one \nthumbs up\n1\n2\n3\n4\nConventional\n\u22120.855**\n\u22120.834**\n0.0321\n0.00172\n(0.368)\n(0.342)\n(0.400)\n(0.345)\nRobust bias-corrected\n\u22120.683*\n\u22120.701**\n0.167\n0.131\n(0.368)\n(0.342)\n(0.400)\n(0.345)\nObservations\n134,970\n134,970\n224,212\n224,212\nBW select method\n1\u2009m.s.e.\n2\u2009m.s.e.\n1\u2009m.s.e.\n2\u2009m.s.e.\nBW above\n25.07\n35.54\n26.49\n77.43\nBW below\n25.07\n21.87\n26.49\n24.06\nEffective number of \nobservations above\n25,013\n35,922\n28,920\n71,838\nEffective number of \nobservations below\n20,674\n17,915\n30,963\n28,136\nThis table shows RD estimation of average daily energy usage in the 3 months after the receipt \nof the eHER in February\u2013March 2018 (that is, the one preceding the eHER augmented with the \nnormative prime), normalized by the average control group consumption in the post period, around \nthe three versus two thumbs-up cutoff (columns 1 and 2) and two versus one thumbs-up cutoff \n(columns 3 and 4). The first row shows the conventional RD estimate, whereas the second corrects \nfor bias48. BWs were selected to minimize the m.s.e.48,49. Odd columns use the same BW on either \nside of the cutoff, whereas even columns estimate separate BWs. Standard errors are clustered at \nthe customer level in parentheses. ***P < 0.01, **P < 0.05, *P < 0.1.\nNature Energy | VOL 5 | November 2020 | 900\u2013909 | www.nature.com/natureenergy\n905\n\nArticles\nNATure Energy\nin other behavioural domains and of alternative formulations of \ndescriptive and injunctive feedback are needed to verify the gener-\nalizability of our findings.\nFinally, although we identify potential determinants of the \nstrength of the social information messages, we are far from for-\nmulating a comprehensive theoretical model. Such a model should \nincorporate other important insights from the social information \nliterature, for instance, on the role of individual descriptive and nor-\nmative second-order beliefs, from the perceived consensus around \nthe norm or from misperceptions that concern one\u2019s own compli-\nance with the norm. Similarly, within our setting, due to lack of data, \nwe can only test a few implications of our conceptual framework \nand cannot control for the influence of other important determi-\nnants of the impact of social information identified in the literature. \nThe predictions from a more comprehensive model should, instead, \nbe subject to a systematic experimental investigation.\nMethods\nEthics statement. Ethical approval for the use of the data that support the findings \nof this study was granted by the Institutional Review Board at Politecnico di \nMilano (approval number 04/2017). Consent for their administrative data to \nbe used for the research was given by the users as part of the privacy consent \nstatements that they signed with the utility.\nProgramme details. We evaluated a social information programme designed and \nimplemented by the utility and Opower (acquired by Oracle in 2016). The eHER \nthat constitutes the core feature of the programme differed from the standard \nOpower HER, evaluated in other works6,17,18,21,30, under three respects: first, it was \ndelivered by email rather than by post, hence the notation eHER; second, it did \nnot feature a section with energy saving tips\u2014tips could be consulted by interested \ncustomers within their personal area on the utility\u2019s website, accessible through \na clickable link on the eHER; third, the normative feedback was given through \nthumbs-up symbols accompanied by the expressions \u2018excellent\u2019, \u2018good\u2019 and \u2018you can \ndo better\u2019, rather than through the standard smiley faces coupled with \u2018great\u2019, \u2018good\u2019 \nand \u2018above average\u2019. The first two differences are consistent with the objective to \nfoster customers\u2019 digital engagement, which the utility primarily wanted to achieve \nthrough the programme, whereas the third is the result of focus groups conducted \nby the utility and Opower to define the design of the eHER.\nOur augmented eHER added a simple treatment to this basic set-up in the \nform of a message at the bottom of the report. We proposed a formulation of the \ndescriptive and injunctive normative messages based on previously collected survey \ndata and collaborated with Opower and the utility to finalize the wording, layout \nand graphical aspects of the messages. Opower and the utility were responsible for \nthe randomization of the normative messages and the implementation of the test.\nIn addition to the experimental test discussed in the present study, we \nmanipulated in a similar way the November\u2013December 2017 eHER. The experiment \naimed to test the impact of environmental identity on energy conservation. It \nfeatured a treatment message that primed individual environmental self-identity \nand a control message that encouraged energy conservation22. Given that both the \nNovember\u2013December 2017 and the April\u2013May 2018 primes were randomized, \nparticipation in the environmental prime test should not affect the results presented \nhere. Nevertheless, we support this claim through further tests reported below.\nSample. Our sample of analysis is represented by the entire eligible customer base \nat the time of the start of the programme (n\u2009=\u2009464,523). Eligibility criteria were \nestablished and verified by the utility and Opower, and included availability of a \ncontact email address and a set of technical requirements, such as living in single \nfamily homes, having one year of pretreatment consumption data without missing, \nnegative or abnormally high usage, and having a sufficient number of neighbours\u2014\ndefined as customers who lived within a 10\u2009km radius and were similar in terms \nof type of housing and any other characteristic available to the utility) for the \nneighbour comparison. The utility is present over the full national territory and \nthe programme was targeted to all eligible customers regardless of their area of \nresidence (Supplementary Fig. 2 shows the study sample distribution across Italian \nmunicipalities). Moreover, to foster energy conservation was not the main goal of \nthe programme. These considerations alleviate concerns of site selection bias31.\nEligible customers were randomly assigned to the treatment (n\u2009=\u2009418,178) and \ncontrol (n\u2009=\u200946,345) groups by the utility through the minmax t-statistic algorithm, \ndepending on baseline consumption and geographical location32. The small relative \nsize of the control group was determined by the utility in collaboration with \nOpower, with the goal to minimize the number of customers who did not receive \nthe programme, but avoid issues of statistical power in the evaluation of its impact. \nThe experimental design could eventually capture a minimum detectable effect with \n90% power and 5% significance of about 0.36%. As for heterogeneous effects by \npretreatment usage, the minimum detectable effects ranged from 0.8 to 1.2%. These \neffect sizes are relatively small with respect to those found in the literature19,33\u201335.\nTable 3 | Impact of the descriptive versus injunctive messages \non electricity usage\nAll\nOne \nthumbs up\nTwo \nthumbs up\nThree \nthumbs up\n1\n2\n3\n4\nPost\n\u22120.300***\n\u22120.081\n\u22120.459***\n\u22121.190***\n(0.113)\n(0.194)\n(0.128)\n(0.213)\nDescriptive \nprime\u2009\u00d7\u2009post\n0.088\n0.225\n0.296*\n\u22120.959***\n(0.149)\n(0.252)\n(0.170)\n(0.284)\n[0.142]\n[0.09\n[0.003]\nConstant\n102.302*** 133.685***\n81.618***\n50.329***\n(0.044)\n(0.076)\n(0.052)\n(0.080)\nObservations\n2,783,190\n1,358,951\n1,002,787\n421,452\nR-squared\n0.109\n0.171\n0.093\n0.044\nNumber of  \ncustomers\n256,487\n125,249\n92,407\n38,831\nThe dependent variable is the average daily electricity usage (kWh), main and heterogeneous \neffects, normalized by average control group consumption in the post period. The reference \nperiod for the analysis is October 2017 to August 2018 (3\u2009months impact). All the specifications \ninclude customer fixed effects and month by year fixed effects. Standard errors are clustered \nat the customer level in parentheses and FDR-adjusted q values in square brackets. ***P\u2009<\u20090.01, \n**P\u2009<\u20090.05, *P\u2009<\u20090.1.\n2\n0\n\u20132\n\u20134\nRegression specifications\n95% Cl\n99% Cl\n90% Cl\nPoint estimate\nAII\nNS\nNS 1T\nNS 2T\nNS 3T\nSU 2T\nSU 3T\nSD 1T\nSD 2T\nSU\nSD\nTreatment effect\nFig. 5 | Prime impact by up- and downgrades in the injunctive feedback \ncategory. The figure plots the coefficient of the descriptive norm\u2009\u00d7\u2009post \nperiod on average daily electricity usage (kWh), normalized by the average \ncontrol group consumption in the post period, estimating equation (3) for \nthe following subsamples. All, whole sample of customers included in the \nnormative prime trial (n\u2009=\u20092,783,190); NS, customers who experienced no \nswitch in the number of thumbs in the eHER augmented with the normative \nprime with respect to the previous one (n\u2009=\u20092,334,191); SU, switch up, \nthat is, an improvement in the number of thumbs in the eHER augmented \nwith the normative prime with respect to the previous one (n\u2009=\u2009239,501); \nSD, switch down, that is, a decrease in the number of thumbs in the \neHER augmented with the normative prime with respect to the previous \none (n\u2009=\u2009209,498); NS 1T, no switch and one thumbs up in the eHER \naugmented with the normative prime (n\u2009=\u20091,232,602); NS 2T, no switch \nand two thumbs up (n\u2009=\u2009773,165); NS 3T, no switch and three thumbs up \n(n\u2009=\u2009328,424); SU 2T, switch up and two thumbs up (n\u2009=\u2009146,473); SU \n3T, switch up and three thumbs up (n\u2009=\u200993,028); SD 1T, switch down and \none thumbs up (n\u2009=\u2009126,349); SD 2T, switch down and two thumbs up \n(n\u2009=\u200983,149). CI, confidence interval.\nNature Energy | VOL 5 | November 2020 | 900\u2013909 | www.nature.com/natureenergy\n906\n\nArticles\nNATure Energy\nThe programme was rolled-out in three waves: July (39% of treated \ncustomers), October (33%) and December 2016 (28%). After that, customers \nreceived reports bimonthly. Conditional on being assigned to the treated group, \n91% of customers received at least one eHER, on average 8.8 over the first \n24\u2009months of the programme. In April and May 2018, the report augmented with \nthe experimental prime was sent to 256,487 programme participants, randomly \nassigned to receive the descriptive (n\u2009=\u2009127,899) or injunctive (n\u2009=\u2009128,588) \nmessage. The sample of customers who received the augmented eHER was smaller \nthan the entire sample of programme participants at that time (n\u2009=\u2009348,131)\u2014\nas the implementation of the programme was not under the research team\u2019s \ncontrol, we cannot document the reasons for such a discrepancy. We did, \nhowever, examine individual correlates of the receipt of the augmented eHER \n(Supplementary Note 5 and Supplementary Table 15).\nAbout 21.5% of customers left the dataset between the launch of the \nprogramme and August 2018, primarily due to termination of the contract with \nthe utility. We checked that the attrition was non-differential by treatment status \nusing a two-sided t-test (Supplementary Table 1). In Supplementary Table 1, we \nalso report a similar test as to whether attrition in the 3\u2009months after the delivery of \nthe augmented eHER, equal to about 5.8%, correlated with the experimental prime \ntreatment. We explored individual correlates of attrition at different points in time \nusing regression analysis (Supplementary Table 16). These allowed us to assess the \nextent to which attrition was a threat to the internal and external validity.\nData. We had access to historical electricity consumption data from July 2015 \nto December 2019 for all the customers. These data were provided by the utility, \nafter being verified by the electricity distributor, and constituted our main \noutcome variable. Similarly to other works, we computed the average daily \nelectricity usage in a month from the total monthly consumption and normalized \nit by dividing by the average post-period control group consumption and then \nmultiplying by 100 (ref. 17). We also computed the average daily pretreatment \nelectricity usage as the mean of the average daily consumption in a month \nbetween July 2015 and June 2016.\nWe also had access to information on the treated customers\u2019 engagement with \nthe reports and on the reports\u2019 contents. These data were provided by Opower. We \nknew when a eHER was sent and when a customer opened or clicked on a report, \nalthough we did not know which one. By opening the reports, customers were \nable to view the neighbour comparison; by clicking on it, customers were directed \nto their personal page on the utility\u2019s website, where further information, such as \nenergy saving tips and bills, was available. On average, 64 and 30% of the customers \nopened and clicked on an eHER, respectively, at least once over the two years after \nthe programme launch. As for the eHER augmented with the normative prime, 55 \nand 9.7% of customers opened and clicked at least once in the two months after its \nreceipt, respectively.\nThe reports\u2019 content data include customers\u2019 relative performance within \nthe reference group and the type of feedback they received within each eHER, \nin terms of number of thumbs up. In general, considering all the reports in the \nfirst 24\u2009months, customers received 20% of reports with three thumbs up, 34% \nwith two thumbs up and 45% with one thumbs up, consistent with the definition \nof the three feedback categories. Overall, 38% of customers received the same \nfeedback throughout the period, and the remaining experienced some change. \nIn the majority of cases, customers experienced both upgrades and downgrades. \nSpecifically to the eHER augmented with the randomized prime, 15.1% of \ncustomers received three thumbs up, 36.1% received two thumbs up and 48.8% \none thumbs up. As time-invariant controls, we used dummies for the main \ngeographical areas in Italy, that is, the northeast, northwest, central, south and \nthe islands, and the population of the municipality in which the customers lived, \nobtained by matching the contract municipality data with data on municipalities\u2019 \ncharacteristics36. We missed information on the geographical location of 5,675 \ncustomers (about 1.2% of the sample), equally distributed by treatment status \n(P\u2009=\u20090.273). This reduced the sample size to 459,088 customers whenever the \nanalysis featured geographical controls.\nFinally, we used data from an online survey conducted with a representative \nsubsample of about 3,000 utility customers in April 2017 to inform the design \nof the norm primes. In particular, questions on actual pro-environmental \nbehaviour\u2014such as turning off the lights when leaving a room and hanging the \nclothes to dry instead of using dryer\u2014were used to design the descriptive prime, \nwhereas questions on personal values related to energy conservation\u2014such as \nwhether the customer feels personally responsible to try to save energy, whether \nthe customer would act according to the customer\u2019s principles if energy was saved \nand whether the customer feels morally obliged to save energy\u2014informed the \ndesign of the injunctive prime. The survey questions used to elicit personal values \nand norms were taken from established survey instruments, such as the World \nValues Survey, and from published studies in environmental psychology37,38. More \ndetails on the survey can be found in Bonan et\u00a0al.22.\nBalance. We checked that the Opower treatment and control group, and the \nsamples of customers assigned to the descriptive or injunctive prime, were \nbalanced across the observable characteristics through two-sided t-tests \n(Supplementary Table 1). In addition, we tested for the randomization balance \nacross the subgroups identified by the combination of the April\u2013May 2018 \nnormative prime treatments and the November\u2013December 2017 environmental \nprime treatments, mentioned above (Supplementary Table 2). This was done \nthrough an F-test of joint significance of subtreatment dummies regressed on the \nobservable characteristics. We confirmed that balance generally holds.\nImpact of the programme. We evaluated the impact of the programme by \nestimating the following model:\nyit \u00bc \u03b21Postit \u00fe \u03b22Programi \u00b4 Postit \u00fe ht \u00fe gi \u00fe \u03b5it\n\u00f01\u00de\nwhere yit is the customer\u2019s i normalized average daily consumption in month t. \nProgram is a treatment indicator, Post is a dummy variable which becomes 1 when \ncustomers receive the first eHER. The coefficient \u03b21 captures the effect of any time \nvariant factors affecting consumption after the start of the programme, while \u03b22 \nisolates the impact of the programme on treated customers. \u0190it is the error term. \nGiven the staggered phase-in of the programme, to allow the identification through \ndifference-in-differences, we randomly assigned control customers to the three \nprogramme start waves, with the same proportions as treated customers. This \nimplies that Post becomes 1 at the beginning of each wave for the same share of \ntreated and control customers. The regression also included month-by-year fixed \neffects, ht, and household fixed effects, gi. Standard errors were clustered at the level \nof household to allow for the presence of within-customer correlation over time in \nthe error term39. The average treatment effects can be interpreted as the percentage \nchange (Supplementary Table 3).\nWe examined the differential response to the eHER depending on pretreatment \nelectricity use by interacting it with the Program\u2009\u00d7\u2009Post dummy. We expressed \nelectricity usage with a continuous variable and with dummies for the quartiles. \nTo allow for differential post-treatment trends, we also interacted the Post dummy \nwith pretreatment usage. To adjust for multiple hypothesis testing, in the subgroup \nanalysis we computed the sharpened two-stage q values (FDR-adjusted q values)40. \nWe analysed ex post power by calculating the minimum detectable effects with a \n90% power and 5% significance41. As a robustness check, we estimated the main \nand heterogeneous effects of the programme on the subsample of customers for \nwhom information on geographical location was available (Supplementary Table 4).\nWe evaluated the main and heterogeneous impacts of social information \nover the first and the first two years of the programme. In a further analysis, we \nextended the treatment period until December 2019 and explored in greater detail \nthe persistence of the treatment effects (Supplementary Tables 5\u20137).\nAs in other works6,17, our treatment effects are intent-to-treat estimates \ncomputed on the full sample of eligible customers, regardless of whether they opted \nout of the programme or did not open the reports. We kept customers who did not \nreceive or read reports in the analysis to maintain the balance between the treatment \nand control group and avoid selection issues affecting our results. By doing this, \nwe were likely to underestimate the effect of the programme on the group of \ncustomers initially assigned to receive the eHER and who actually saw the treatment \ncommunication. In an additional analysis, we examined the role of the engagement \nwith the programme on treatment impacts by instrumenting opening the eHER and \nclicking on it with the treatment status, and reporting the local average treatment \neffect (Supplementary Table 8). We did this in cross-section. The outcome variable \nwas the normalized average electricity usage in the 24\u2009months after the launch of the \nprogramme. The specifications also included pretreatment electricity usage.\nWe further exploited our data on the content of the reports, specifically on \nthe number of thumbs up received within each eHER, to examine heterogeneous \neffects of the treatment depending on whether the customers were upgraded \nor downgraded with respect to the previous report. We did it by restricting the \nanalysis to the months when the eHERs were sent and by focusing on normalized \ndaily electricity usage in the 3\u2009months after each report. Specifications include \ncustomer and month-by-year fixed effects (Supplementary Table 9).\nRegression discontinuity estimation. We used RD to estimate the impact of \nchanges in the injunctive feedback included in the neighbour comparison. We \njustified this approach by showing that customers who received one, two or three \nthumbs up within the April\u2013May 2018 eHER were different in many respects, \nwhich may also correlate with the impact of the treatment (Supplementary Table 10).  \nThe RD approach allowed us to eliminate the influence of confounding factors on \nthe estimated effect of changing the feedback category, as it focused on customers \nfor whom the number of thumbs up is random. The price we paid for the improved \nidentification of the effects is that the impacts estimated through RD are local, \nspecific to a neighbourhood of the thresholds.\nIn our RD framework, the running variables, Xi1 and Xi2, are the customer\u2019s i \nmonthly electricity usage (kWh). The cutoffs, ci1 and ci2, are the electricity usage of \nthe 15th percentile and the overall average electricity usage among the neighbours, \nrespectively. The assignment variable Ti1 takes the value of 1 when the customer\u2019s \nusage lies above ci1 (Xi1\u2009>\u2009ci1) and 0 otherwise. Similarly, T21 takes the value of 1 \nwhen the customer\u2019s usage lies between ci1 and ci2 and 0 otherwise. We estimated \nthe equations:\nYi \u00bc \u03bb1\n0 \u00fe \u03b71Ti1 \u00fe f Xi1 \u001d ci1\n\u00f0\n\u00de \u00fe \u03b51\n1\n\u00f02A\u00de\nYi \u00bc \u03bb2\n0 \u00fe \u03b72Ti2 \u00fe f Xi2 \u001d ci2\n\u00f0\n\u00de \u00fe \u03b52\n1\n\u00f02B\u00de\nNature Energy | VOL 5 | November 2020 | 900\u2013909 | www.nature.com/natureenergy\n907\n\nArticles\nNATure Energy\nwhere Yi is the customer\u2019s i-normalized average daily electricity usage in the \n3\u2009months after the receipt of the report that contained the normative prime. \nModel (2A) is estimated from customers who received either three or two thumbs \nup (for example, with Xi1\u2009>\u2009ci1 or ci1\u2009<Xi1\u2009<ci2), whereas model (2B) includes those \nwho received either two or one thumbs up (for example, Xi1\u2009<\u2009ci1). The parameter \n\u03bb0 is a constant, while \u03b7 captures the effect of obtaining an extra thumbs up. \n\u03b51\n1\nI\n and \u03b52\n1\nI\n are the error terms. Our main specification used a non-parametric \napproach (local polynomial point estimation), which amounts to fitting two \nlinear regressions on customers respectively close to the left and to the right \nof the cutoff42,43. This approach used only observations within a BW from the \ncutoff. The selection of the BW was data driven and aims to minimize the m.s.e. \nof the local polynomial RD estimator. We computed the point estimates using \nboth one common BW on both sides of the cutoff and two distinct BWs. We \nused a linear polynomial for f(.), which represents a good trade-off between \nsimplicity, precision and stability44. Observations within the BW were weighted \nthrough a triangular kernel function, which assigned zero weight to observations \noutside the BW and positive weights within it, which decreased symmetrically \nand linearly as the distance from the cutoff increased. The triangular kernel had \ndesirable asymptotic optimality properties44. We show point estimates under \nconventional (from a parametric least-squares estimation) and bias-corrected \nrobust standard errors, which deliver a valid inference under the m.s.e.-optimal \nBW selection43 (Table 1).\nIdentification within RD design relies on the assumption that customers have \nno full control over the running variable and hence are unable to manipulate \ntheir position in the distribution of electricity usage within their reference group. \nManipulation should not be a threat to our identification for several reasons. First, \ncustomers do not know the algorithm for the selection of the reference group and \nassignment of thumbs up. Second, although people may be able to control their \nown electricity usage, they have no control on their neighbours\u2019 consumption, \nwhich determines the unpredictability of the consumption level associated with \nthe different cutoffs every round. Third, people do not gain any direct benefit from \nan extra thumbs up, apart from individual satisfaction, and hence the incentives to \ngame the system should be minimal. We tested this identification assumption by \nchecking for discontinuities in the density of the running variable at the thresholds, \nusing a non-parametric density estimator based on local polynomial techniques45. \nWe did the same with the pretreatment covariates, following a procedure suggested \nin the literature46 (Supplementary Table 11).\nWe also performed an RD estimation on the probability that customers opened \nor clicked in the April\u2013May 2018 eHER. This analysis aimed to support the claim \nthat the effects we found on consumption were due to the neighbour comparison, \nrather than to discontinuous changes in unobserved determinants of open rates or \nthe likelihood of viewing electricity saving tips at the thresholds (Supplementary \nTable 12).\nWe repeated the test of the RD identifying assumption and the RD estimation \nusing a standard eHER, delivered in February\u2013March 2018, to rule out that our \nresults are due to the presence of the normative primes (Supplementary Note 6 and \nSupplementary Tables 17 and 2) and RD effects over the following 6 and 12\u2009months \n(Supplementary Table 13).\nImpact of the normative primes. We repeated the RD estimation separately on \nthe subsamples of customers randomly allocated to the descriptive and injunctive \nprimes to obtain local marginal effects of moving across the injunctive feedback \ncategories by prime treatment. We adjusted for multiple comparisons40.\nWe also assessed the impact of receiving the eHER augmented by the \ndescriptive versus injunctive prime, by estimating the following model:\nyit \u00bc \u03b11Postit \u00fe \u03b12Descriptive normi \u00b4 Postit \u00fe ht \u00fe gi \u00fe \u03bcit\n\u00f03\u00de\nDescriptive_norm takes the value of 1 for the descriptive prime and 0 for \nthe injunctive prime. Post becomes one when customers receive the eHER that \ncontains the prime (roughly half in April and half in May 2018). The coefficient \n\u03b11 captures the change in consumption after receiving the eHER for customers \nassigned to receive the injunctive prime, while \u03b12 reveals any differential effect of \nreceiving the descriptive prime. Standard errors are clustered at the household level \n(Table 3). The regression was run on the sample of treated customers at that time \n(n\u2009=\u2009256,487) and considered the period that spanned from October 2017 to August \n2018, but we also explored the persistence of the effects until December 2019, that \nis, 18 months after the prime (Supplementary Table 14). We also estimated equation \n(3) for the subsamples of customers who received one, two or three thumbs up, \noverall and separately for customers upgraded or downgraded with respect to the \nprevious report, to examine the heterogeneous impact of the augmented eHER.\nWe tested that these results were not due to interaction effects with the \nNovember\u2013December 2017 environmental identity test by adding to equation (3) \nan interaction term between Post, Descriptive_norm and a dummy equal to 1 if \na customer was assigned to the environmental prime treatment (Supplementary \nTable 18).\nPreregistration. We analysed the impact of the standard and augmented eHER \nusing non-experimental RD techniques. In terms of individual level data, our \nanalysis relied only on customers\u2019 consumption, location and content of the eHER. \nFor these reasons, the analysis we present was not prespecified in a pre-analysis \nplan: detailed prespecification is warranted when, within the impact evaluation of \nfield experiments, subgroup analysis is expected to be important and there is the \npossibility to cherry-pick the dimensions of heterogeneity to focus on or that a \nparty to the study has a vested interest47. These conditions did not apply to our case.\nReporting summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nThe data that support the findings of this study are proprietary data of the energy \ncompany and cannot be shared publicly. To inquire about access to the proprietary \ndata, please contact M.T.\nCode availability\nThe replication code is available on Open Science Framework at https://osf.io/\nwz8gb/.\nReceived: 30 April 2020; Accepted: 28 September 2020;  \nPublished online: 2 November 2020\nReferences\n\t1.\t Frey, B. S. & Meier, S. Social comparisons and pro-social behavior: testing \n\u2018conditional coop-eration\u2019 in a field experiment. Am. Econ. Rev. 94, \n1717\u20131722 (2004).\n\t2.\t Shang, J. & Croson, R. A field experiment in charitable contribution: the \nimpact of social information on the voluntary provision of public goods. \nEcon. J. 119, 1422\u20131439 (2009).\n\t3.\t Gerber, A. S. & Rogers, T. Descriptive social norms and motivation to vote: \neverybody\u2019s voting and so should you. J. Polit. 71, 178\u2013191 (2009).\n\t4.\t Beshears, J., Choi, J. J., Laibson, D., Madrian, B. C. & Milkman, K. L. The \neffect of providing peer information on retirement savings decisions.  \nJ. Finance 70, 1161\u20131201 (2015).\n\t5.\t Allcott, H., Mullainathan, S. & Taubinsky, D. Externalizing the Internality \n(New York Univ., 2011).\n\t6.\t Allcott, H. & Rogers, T. The short-run and long-run effects of behavioral \ninterventions: experimental evidence from energy conservation. Am. Econ. \nRev. 104, 3003\u20133037 (2014).\n\t7.\t Ayres, I., Raseman, S. & Shih, A. Evidence from two large field experiments \nthat peer comparison feedback can reduce residential energy usage. J. Law \nEcon. Organ. 29, 992\u20131022 (2013).\n\t8.\t Brent, D. A., Cook, J. H. & Olsen, S. Social comparisons, household water \nuse, and participation in utility conservation programs: evidence from three \nrandomized trials. J. Assoc. Environ. Resour. Econ. 2, 597\u2013627 (2015).\n\t9.\t Ferraro, P. J. & Price, M. K. Using nonpecuniary strategies to influence \nbehavior: evidence from a large-scale field experiment. Rev. Econ. Stat. 95, \n64\u201373 (2013).\n\t10.\tFerraro, P. J., Miranda, J. J. & Price, M. K. The persistence of treatment effects \nwith norm-based policy instruments: evidence from a randomized \nenvironmental policy experiment. Am. Econ. Rev. 101, 318\u2013322 (2011).\n\t11.\tFerraro, P. J. & Miranda, J. J. Heterogeneous treatment effects and \nmechanisms in information-based environmental policies: evidence from a \nlarge-scale field experiment. Resour. Energy Econ. 35, 356\u2013379 (2013).\n\t12.\tJaime Torres, M. M. & Carlsson, F. Direct and spillover effects of a social \ninformation campaign on residential water-savings. J. Environ. Econ. Manag. \n92, 222\u2013243 (2018).\n\t13.\tSchultz, P. W., Nolan, J. M., Cialdini, R. B., Goldstein, N. J. & Griskevicius, V. \nThe constructive, destructive, and reconstructive power of social norms. \nPsychol. Sci. 18, 429\u2013434 (2007).\n\t14.\tBicchieri, C. & Xiao, E. Do the right thing: but only if others do so. J. Behav. \nDecis. Mak. 22, 191\u2013208 (2009).\n\t15.\tMeisel, M. K. & Goodie, A. S. Descriptive and injunctive social norms\u2019 \ninteractive role in gambling behavior. Psychol. Addict. Behav. 28,  \n592\u2013598 (2014).\n\t16.\tBiccheri, C. & Dimant, E. Nudging with care: the risks and benefits of social \ninformation. Public Choice https://doi.org/10.1007/s11127-019-00684-6 \n(2019).\n\t17.\tAllcott, H. Social norms and energy conservation. J. Public Econ. 95, \n1082\u20131095 (2011).\n\t18.\tJachimowicz, J. M., Hauser, O. P., O\u2019Brien, J. D., Sherman, E. & Galinsky, A. D. \nThe critical role of second-order normative beliefs in predicting energy \nconservation. Nat. Hum. Behav. 2, 757\u2013764 (2018).\n\t19.\tAndor, M. A., Gerster, A., Peters, J. & Schmidt, C. M. Social norms and energy \nconservation beyond the US. J. Environ. Econ. Manag. 103, 102351 (2010).\n\t20.\tByrne, D. P., Nauze, A. L. & Martin, L. A. Tell me something I don\u2019t already \nknow: informedness and the impact of information programs. Rev. Econ. Stat. \n100, 510\u2013527 (2018).\nNature Energy | VOL 5 | November 2020 | 900\u2013909 | www.nature.com/natureenergy\n908\n\nArticles\nNATure Energy\n\t21.\tCosta, D. L. & Kahn, M. E. Energy conservation \u2018nudges\u2019 and environmentalist \nideology: evidence from a randomized residential electricity field experiment. \nJ. Eur. Econ. Assoc. 11, 680\u2013702 (2013).\n\t22.\tBonan, J., Cattaneo, C., D\u2019Adda, G. & Tavoni, M. Can We Make Social \nInformation Programs More Effective? The Role of Identity and Values Working \nPaper 19-21 (Resources for the Future, 2019).\n\t23.\tBursztyn, L., Gonz\u00e1lez, A. L. & Yanagizawa-Drott, D. Misperceived Social \nNorms: Female Labor Force Participation in Saudi Arabia Working Paper \n24736 (National Bureau of Economic Research, 2018); https://doi.\norg/10.3386/w24736\n\t24.\tD\u2019Adda, G., Dufwenberg, M., Passarelli, F. & Tabellini, G. Partial Norms SSRN \nScholarly Paper ID 3362021 (Social Science Research Network, 2019).\n\t25.\tList, J. A., Metcalfe, R. D., Price, M. K. & Rundhammer, F. Harnessing Policy \nComplementarities to Conserve Energy: Evidence from a Natural Field \nExperiment Working Paper 23355 (National Bureau of Economic Research, \n2017); https://doi.org/10.3386/w23355.00000\n\t26.\tHallsworth, M., List, J. A., Metcalfe, R. D. & Vlaev, I. The behavioralist as tax \ncollector: using natural field experiments to enhance tax compliance. J. Public \nEcon. 148, 14\u201331 (2017).\n\t27.\tBobek, D. D., Hageman, A. M. & Kelliher, C. F. Analyzing the role of social \nnorms in tax compliance behavior. J. Bus. Ethics 115, 451\u2013468 (2013).\n\t28.\tKrupka, E. L. & Croson, R. T. A. The differential impact of social norms cues \non charitable contributions. J. Econ. Behav. Organ. 128, 149\u2013158 (2016).\n\t29.\tBahnot, S. Isolating the effect of injunctive norms on conservation behavior: \nnew evidence from a field experiment in California. Organ. Behav. Hum. \nDecis. Process https://doi.org/10.1016/j.obhdp.2018.11.002 (2018).\n\t30.\tAllcott, H. & Taubinsky, D. Evaluating behaviorally motivated policy: \nexperimental evidence from the lightbulb market. Am. Econ. Rev. 105, \n2501\u20132538 (2015).\n\t31.\tAllcott, H. Site selection bias in program evaluation. Q. J. Econ. 130, \n1117\u20131165 (2015).\n\t32.\tBruhn, M. & McKenzie, D. In pursuit of balance: randomization in practice \nin development field experiments. Am. Econ. J. Appl. Econ. 1, 200\u2013232 (2009).\n\t33.\tAndor, M. A. & Fels, K. M. Behavioral economics and energy \nconservation\u2014a systematic review of non-price interventions and their causal \neffects. Ecol. Econ. 148, 178\u2013210 (2018).\n\t34.\tFarrow, K., Grolleau, G. & Ibanez, L. Social norms and pro-environmental \nbehavior: a review of the evidence. Ecol. Econ. 140, 1\u201313 (2017).\n\t35.\tDelmas, M. A., Fischlein, M. & Asensio, O. I. Information strategies and \nenergy conservation behavior: a meta-analysis of experimental studies from \n1975 to 2012. Energy Policy 61, 729\u2013739 (2013).\n\t36.\tGuiso, L., Sapienza, P. & Zingales, L. Long-term persistence. J. Eur. Econ. \nAssoc. 14, 1401\u20131436 (2016).\n\t37.\tSteg, L., Perlaviciute, G., van\u00a0der Werff, E. & Lurvink, J. The significance of \nhedonic values for environmentally relevant attitudes, preferences, and \nactions. Environ. Behav. 46, 163\u2013192 (2014).\n\t38.\tvan\u00a0der Werff, E., Steg, L. & Keizer, K. The value of environmental \nself-identity: the relationship between biospheric values, environmental \nself-identity and environmental preferences, intentions and behaviour.  \nJ. Environ. Psychol. 34, 55\u201363 (2013).\n\t39.\tBertrand, M., Duflo, E. & Mullainathan, S. How much should we trust \ndifferences-in-differences estimates? Q. J. Econ. 119, 249\u2013275 (2004).\n\t40.\tBenjamini, Y., Krieger, A. M. & Yekutieli, D. Adaptive linear step-up \nprocedures that control the false discovery rate. Biometrika 93,  \n491\u2013507 (2006).\n\t41.\tHaushofer, J. & Shapiro, J. The short-term impact of unconditional cash \ntransfers to the poor: experimental evidence from Kenya. Q. J. Econ. 131, \n1973\u20132042 (2016).\n\t42.\tImbens, G. W. & Lemieux, T. Regression discontinuity designs: a guide to \npractice. J. Econom. 142, 615\u2013635 (2008).\n\t43.\tCalonico, S., Cattaneo, M. D. & Titiunik, R. Robust nonparametric  \nconfidence intervals for regression- discontinuity designs. Econometrica 82, \n2295\u20132326 (2014).\n\t44.\tCattaneo, M. D., Idrobo, N. & Titiunik, R. A Practical Introduction to \nRegression Discontinuity Designs: Foundations (Cambridge  \nUniversity Press, 2019).\n\t45.\tCattaneo, M. D., Jansson, M. & Ma, X. Simple local polynomial density \nestimators. J. Am. Stat. Assoc. 0, 1\u20137 (2019).\n\t46.\tCanay, I. A. & Kamat, V. Approximate permutation tests and induced order \nstatistics in the regression discontinuity design. Rev. Econ. Stud. 85, \n1577\u20131608 (2018).\n\t47.\tBanerjee, A. et\u00a0al. In Praise of Moderation: Suggestions for the Scope and Use \nof Pre-analysis Plans for RCTs in Economics Working Paper w26993 (National \nBureau of Economic Research, 2020).\n\t48.\tCalonico, S., Cattaneo, M. D. & Titiunik, R. Robust nonparametric  \nconfidence intervals for regression\u2013discontinuity designs. Econometrica 82, \n2295\u20132326 (2014).\n\t49.\tCalonico, S., Cattaneo, M. D., Farrell, M. H. & Titiunik, R. Regression \ndiscontinuity designs using covariates. Rev. Econ. Stat. 101, 442\u2013451 (2019).\nAcknowledgements\nThe research leading to these results received funding from the European Research \nCouncil under the European Union\u2019s Seventh Framework Program (FP7/2007-2013)/\nERC grant agreement no. 336155\u2014project COBHAM, \u2018The role of consumer behaviour \nand heterogeneity in the integrated assessment of energy and climate policies\u2019.\nAuthor contributions\nJ.B., C.C., G.D. and M.T. conceived and designed the experiments. J.B. analysed the data. \nJ.B. and G.D. contributed the analysis tools. J.B., C.C., G.D. and M.T. wrote the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-020-00719-z.\nCorrespondence and requests for materials should be addressed to G.d.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2020\nNature Energy | VOL 5 | November 2020 | 900\u2013909 | www.nature.com/natureenergy\n909\n\n1\nnature research  |  reporting summary\nApril 2020\nCorresponding author(s):\nGiovanna d'Adda\nLast updated by author(s): May 5, 2020\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nNo software was used for data collection.\nData analysis\nStata 15 was used for all statistical analysis and for the production of graphs.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and \nreviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nThe data that support the findings of this study were used under license for the current study. Because of the privacy rules in Europe and the non-disclosure \nagreement with company, the data are not publicly available. Data are however available from the authors upon reasonable request and with permission of the \ncompany. In addition, to ensure that the results can be replicated, the code is available upon request.\n\n2\nnature research  |  reporting summary\nApril 2020\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThe study uses quantitative data generated by a field experiment (Randomized Control Trial): the dataset is a panel including monthly \nelectricity consumption and details on the content of email communication sent to a sample of utility customers.\nResearch sample\nOur sample of analysis is represented by 464,763 customers eligible for the the social information program and with non-missing \nelectricity consumption over the entire period of analysis. We have access to this sample thanks to an existing agreement with a large \nItalian energy utility. Since eligibility requirements include availability of an email contact, we imagine that the sample is not fully \nrepresentative of the entire customer base of the utility nor of the entire population of energy users in Italy. The administrative data \nwe use is restricted to the RCT sample, and only contains information on customers' consumption and area of residence: they thus \ndo not allow us to conduct formal tests of representativeness. However, the size of the sample guarantees that our results are \ngeneralizable to important, and growing, segments of the population.\nSampling strategy\nThe sample used in the study is the entire population of customers targeted by the utility's social information program at the time \nthe program was launched. The size of the sample, eligibility criteria and randomization into treatments were all determined by the \nutility and the private consultants that designed and implemented the social information program. Where the research team had any \ninput on the sampling, i.e., in the design of the normative prime experiment, we requested that treatment and control groups be of \nthe same sizes, in order to maximise statistical power.\nData collection\nThe data used in the study are exclusively administrative data collected directly by the partner utility, and transferred to the research \nteam through secure data sharing protocols.\nTiming\nConsumption data are available from July 2015 to August 2018. Data on the content of program communication are available for the \nduration of the program: July 2016 to August 2018.\nData exclusions\nNo data were excluded from the analysis.\nNon-participation\nNon-participation in the study occurrs when users terminate their contract with the utility. Even though attrition is non-differential \nby treatment, we keep attriters in the sample in order to avoid any selection bias affecting the results.\nRandomization\nRandomization was implemented by the utility through an algorithm (minmax t-statistic), which conducts 1000 randomizations and \nselects the most balanced draw, along baseline consumption and geographic location (Bruhn and McKenzie 2009).\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology and archaeology\nAnimals and other organisms\nHuman research participants\nClinical data\nDual use research of concern\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nHuman research participants\nPolicy information about studies involving human research participants\nPopulation characteristics\nWe only have access to information on the location of utility customers, at the regional level, and on their consumption \nbefore the start of the program. The geographical distribution of users is as follows: 29 per cent from North-West Italy, 12.2 \n\n3\nnature research  |  reporting summary\nApril 2020\nper cent from the North East, 31.2 per cent from the Center, and the remaining from the South and Islands. Average baseline \nconsumption is 6.3 kWh.\nRecruitment\nInclusion in the sample is based exclusively on technical eligibility criteria set by the utility: households must have a valid \nname and email address as of June 2016, live in single-family homes, have at least one year of valid pre-experiment energy \nconsumption data, have no negative electricity meter reads, at least one meter read in the previous three months, no \nsignificant gaps nor extreme peaks in usage history, and exactly one account per customer per location. While these criteria \nmay result in the study sample not being perfectly representative of the full sample of utility customers, no self-selection in \nthe study is present. \nEthics oversight\nWe obtained ethics approval for the study from Milan Polytechnic's Ethics Committee, with approval number 4/2017.\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\n\n\n    Scientific Research Findings:", "answer": "Energy customers who received two different types of social feedback (descriptive and injunctive) encouraging them to save energy reduced their consumption more than low-energy users for whom conforming with the descriptive feedback would entail consumption increases, at odds with the injunctive feedback praising energy saving. The addition of a second piece of information of the same type (for example, adding a second descriptive messages that encouraged energy saving) had a limited impact. When feedback was inconsistent, the piece of feedback delivering the strongest message prevailed, where strength reflected the difference between the user\u2019s energy consumption and that of their neighbours (descriptive feedback) and the intensity of social approval conveyed through visual cues (injunctive feedback). These results suggest the significance of synergies between different types of feedback, rather than the superiority of any one type of feedback. The results may be specific to the precise wording and graphical representations used to provide feedback in our HER, and may not generalize to the whole customer base.", "id": 20}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Analysis\nhttps://doi.org/10.1038/s41560-020-00683-8\n1Department of Mechanical and Industrial Engineering, University of Massachusetts Amherst, Amherst, MA, USA. 2Harvard Kennedy School, Belfer \nCenter for Science and International Affairs, Harvard University, Cambridge, MA, USA. 3Campus Straubing for Biotechnology and Sustainability, Innovation \nand Technology Management, Technical University of Munich, Straubing, Germany. 4Centre for Environment, Energy and Natural Resource Governance, \nDepartment of Land Economy, University of Cambridge, Cambridge, UK. \u2709e-mail: goldstein@umass.edu\nA\nRPA-E was established in the US Department of Energy \n(DOE) in 2009, in recognition of the urgent need for accel-\nerating energy technology innovation. The agency was cre-\nated to fund breakthrough energy research with greater flexibility \nto overcome the stovepiping that afflicts traditional DOE funding \noffices1\u20133. For over two decades, experts have called for increased \ngovernment spending on energy research and development (R&D)4\u20137  \nand for stronger mission-oriented innovation support policies, \nincluding government action to create and shape markets8,9. It is \nwell known that market failures due to knowledge spillovers lead to \nunderinvestment in research by the private sector10,11. In addition, \nexperimentation is fundamental to innovation and entrepreneur-\nship, which are characterized by high uncertainty and low success \nrates12,13, and thus, public intervention can be essential for encour-\naging radical innovation when the cost of experimentation is high14. \nParticularly for cleantech, where societal benefits are generally high \ncompared to private returns because of the environmental external-\nities of air pollution and climate change, governments can be seen as \na risk-tolerant investor with the ability to de-risk or shape markets \nfor new technology9.\nHowever, concerns over the effectiveness of government subsi-\ndies for applied energy R&D in private firms have long been part \nof the political and academic debate. Very recently, for example, \na proposal by the US Office of Management and Budget to elimi-\nnate funding for ARPA-E claimed that \u201cthe private sector is better \npositioned to finance disruptive energy research and development \nand to commercialize innovative technologies\u201d15. Nevertheless, a \nvast amount of past research from various European countries has \ngenerally indicated that public R&D grants do not substitute private \nR&D spending16\u201319. Public direct investments can rather mobilize \nprivate investments in specific sectors such as renewable energy20. \nStudies have found that US government partnerships in the form of \njoint development and licensing21 and federal research subsidies for \nsmall businesses22 can increase patenting and private financing for \nUS cleantech firms. In a recent study of UK firms across all sectors, \nsmall firms substantially increased their internal R&D spending in \nresponse to direct grants and tax credits for R&D23.\nEven though the positive effect of public R&D grants to com-\npanies has been highlighted in recent studies, we still know little \nabout how specific government R&D programmes impact clean-\ntech innovation\u2014especially programmes that provide more than \njust the financial resources, such as R&D grants and tax credits, \noffered by traditional subsidy programmes. ARPA-E was designed \nin the model of the US Defense Advanced Research Projects Agency \n(DARPA), with expert programme staff who are hired on short-term \nrotation and empowered to craft solicitations in an area of technical \nneed, select proposals and actively manage projects24,25. Based on \nDARPA\u2019s long track record of accomplishments, this funding model \nwas expected to produce breakthrough innovations with the poten-\ntial to transform the energy market. In the case of ARPA-E, which \nhas become the often-discussed poster child of mission-oriented \ninnovation in the US and abroad26,27, it is important to understand \nwhether and how ARPA-E has supported innovation, for the sake of \nlearning and informing future initiatives28,29.\nWe are particularly interested in innovation outcomes for startup \ncompanies, which can be more flexible and agile than larger firms or \npublic organizations. Startups have the ability to quickly respond to \nmarket opportunities and can provide solutions to urgent problems \nas long as they have access to adequate resources30,31. Public organi-\nzations and incumbent firms also play a key role for energy innova-\ntion, given their scale and breadth of capabilities. However, previous \nresearch suggests that larger organizations respond to innova-\ntion opportunities later than entrepreneurial firms, due to inertia  \nand legacy systems; and when they do move into new products or \nPatenting and business outcomes for cleantech \nstartups funded by the Advanced Research \nProjects Agency-Energy\nAnna Goldstein\u200a \u200a1,2\u2009\u2709, Claudia Doblinger\u200a \u200a3, Erin Baker\u200a \u200a1 and Laura D\u00edaz Anad\u00f3n2,4\nInnovation to reduce the cost of clean technologies has large environmental and societal benefits. Governments can play \nan important role in helping cleantech startups innovate and overcome risks involved in technology development. Here we \nexamine the impact of the US Advanced Research Projects Agency-Energy (ARPA-E) on two outcomes for startup companies: \ninnovation (measured by patenting activity) and business success (measured by venture capital funding raised, survival, and \nacquisition or initial public offering). We compare 25 startups funded by ARPA-E in 2010 to rejected ARPA-E applicants, start-\nups funded by a related government programme and other comparable cleantech startups. We find that ARPA-E awardees have \na strong innovation advantage over all the comparison groups. However, while we find that ARPA-E awardees performed better \nthan rejected applicants in terms of post-award business success, we do not detect significant differences compared to other \ncleantech startups. These findings suggest that ARPA-E was not able to fully address the \u2018valley of death\u2019 for cleantech startups \nwithin 10\u201315\u2009yr after founding.\nNature Energy | VOL 5 | October 2020 | 803\u2013810 | www.nature.com/natureenergy\n803\n\nAnalysis\nNature Energy\nmarkets, they often harvest innovation from startups32\u201335. Exploring \nthe drivers for startup innovation and success is therefore a valuable \npursuit, especially given these firms\u2019 growing potential for trans-\nforming the energy sector36. Moreover, startups typically operate \nunder a short-term imperative for achieving success as a result of \nresource constraints. Thus, although innovation in the energy sec-\ntor can take many years (and sometimes even decades), there are \nimportant outcomes for startups in cleantech that can be observed \nin the relatively short term21.\nIn this paper, we investigate these short-term (5\u201310\u2009yr) out-\ncomes\u2014specifically innovation, as measured by patenting activity, \nand business success, as measured by acquisition or initial public \noffering (IPO), survival and venture capital (VC) raised\u2014for the \nearliest cohort of ARPA-E startups, funded in 2010. Using a data-\nset combining information from Cleantech Group\u2019s i3 platform and \nDOE financial assistance records, as well as substantive information \non startups from public information sources, we ask whether and \nto what extent the performance of the cleantech startups funded by \nARPA-E differs from that of comparable firms, controlling for key \ncharacteristics, including cleantech subsector, age, prior patenting \nand prior VC investment. We find that ARPA-E\u2019s startup awardees \nproduced significantly more patents post-2010 than similar com-\npanies did (P\u2009<\u20090.01). However, while ARPA-E awardees performed \nbetter than the group of rejected ARPA-E applicants in terms of their \nability to attract VC investment post-award and their likelihood of \nsurviving, being acquired or going public, they had no advantage \nover the average similar cleantech company in these dimensions. \nThese findings indicate that barriers remain for innovative clean-\ntech companies in gaining market traction.\nARPA-E in the context of US cleantech innovation\nBefore ARPA-E was established, there was a rise of VC funding for \ncleantech from 2006 to 2008 in the US. VC investors accustomed \nto funding the commercialization of medical and software technol-\nogy turned their attention to cleantech, expecting to earn positive \nreturns from a few high-value companies in a portfolio. Returns on \nthese cleantech investments, however, turned out to be compara-\ntively low, due to longer development cycles and the high capital \nintensity of energy innovation, as well as competition from incum-\nbent technologies producing the same commodities (for example, \nelectricity and liquid fuels)37. Subsequent VC funding for cleantech \ndropped dramatically.\nShortly after the exodus of cleantech VCs, there was a surge in \nUS funding for public energy R&D through DOE. The American \nReinvestment and Recovery Act (ARRA), which was passed to \naddress the financial crises of 2009, also authorized roughly US$60 \nbillion of stimulus spending on clean energy investments in the \n2010 fiscal year (October 1, 2009, through September 30, 2010)38. \nA portion of ARRA funding established ARPA-E to fund high-risk \nR&D on energy technologies. Other ARRA funds were distributed \nthrough applied technology offices such as the Office of Energy \nEfficiency and Renewable Energy (EERE); Supplementary Fig. 1 \nillustrates the funding surge for US DOE R&D and demonstration \nin 2010.\nWe construct a dataset containing 1,287 US cleantech compa-\nnies founded from 2005 to 2010, that is, firms that were startups \n(five years old or younger) in 201021. These firms are assigned to \none of four groups: (1) startups funded by ARPA-E in 2010 (listed \nin Supplementary Table 2); (2) startups whose applications for \nfunding were denied by ARPA-E in 2010\u2014importantly, our sam-\nple only includes rejected applicants who were \u2018encouraged\u2019 by \nARPA-E, suggesting that they were judged to be high quality rela-\ntive to the larger pool of rejected applicants; (3) startups funded by \nEERE in 2010; and (4) startups that were not identified as ARPA-E \napplicants and also did not receive EERE funding in 2010, although \nthey may have received other types of federal assistance (labelled \nas \u2018Other cleantech\u2019). Supplementary Table 1 provides descrip-\ntive statistics of each group, and Supplementary Data 1 contains a \nde-identified dataset of all 1,287 firms and their post-2010 patent-\ning and business outcomes.\nWe also collect data on companies that received awards from \nthe DOE Small Business Innovation Research (SBIR) programme. \nPrevious research on this programme found it had a positive impact \non the likelihood of VC investment and patenting for small busi-\nnesses22. However, in our data, we find that SBIR startups as a whole \nare fairly dissimilar to ARPA-E awardees and other cleantech start-\nups. Of the 82 startups funded by DOE SBIR in 2010, we are able \nto identify only 60% of them as cleantech companies, compared to \n93% of EERE startups and 100% of ARPA-E startups from the same \nyear. When we perform coarsened exact matching on the ARPA-E \n2010 awardees (see Methods for more detail), only five SBIR com-\npanies remain in the matched sample. For this reason, we exclude \nSBIR awardees from our main analysis; more information on these \ncompanies is provided in Supplementary Note 1.\nAmong the four groups of cleantech startups in our main dataset, \nwe find differences in emphasis across various cleantech subsectors \n(Fig. 1 and Supplementary Table 3). Both ARPA-E awardees and \nARPA-E rejected applicants were concentrated in energy storage; \nARPA-E funded nine storage startups in 2010\u2014over one-third of its \nawards in that year, while such companies composed only 6% of all \ncleantech startups. This emphasis arises largely from two targeted \ntechnical programmes at ARPA-E for battery technology in 2010\u2014\none for electric vehicles and one for grid-scale storage.\nOne important reason for this emphasis is that technology prior-\nities at ARPA-E are determined in a bottom-up fashion, as they aim \nto fund \u2018white space\u2019, that is, technical areas that are overlooked by \nother funding sources39. This is a complementary approach to that \nof other offices in DOE, like EERE, where funds are allocated based \non road-mapping of existing technological trajectories\u2014in other \nwords, EERE funds are more likely to be subject to path depen-\ndencies in the programmes when compared to ARPA-E. EERE\u2019s \ntop-down approach is reflected in its 2010 startup portfolio, which \nwas more similar to the overall makeup of US cleantech companies. \nSolar photovoltaics, for example, composed one-fifth of all clean-\ntech startups in 2010; EERE funded eight solar startups, whereas \nARPA-E funded only one.\nWe also find differences across groups in terms of the pre-2010 \ncompany profile: over half of ARPA-E awardees had already received \nprivate VC investment (56%), and over half had filed at least one \npatent (also 56%). Among applicants rejected by ARPA-E, the  \nproportion of companies with pre-2010 patenting and private funding  \n0\n20\n40\n60\n80\n100\nPercent of companies\nOther\ncleantech\nEERE\n2010\nARPA\u2212E\nrejected\nARPA\u2212E\n2010\nEnergy efficiency\nSmart grid\nBiofuels and biochemicals\nConventional fuels\nBiomass generation\nFuel cells and hydrogen\nGeothermal\nHydro and marine power\nNuclear\nSolar\nWind\nEnergy storage\nTransportation\nAdvanced materials\nAgriculture and food\nAir\nRecycling and waste\nWater and wastewater\nMiscellaneous cleantech\nNumber of\ncompanies\n25\n39\n70\n1,155\nFig. 1 | Subsectors represented in our dataset of US cleantech startup \nfirms in 2010. Each company was assigned to one of 19 cleantech \nsubsectors used by Cleantech Group\u2019s i3 platform. See Methods for more \ndetail on these categorizations.\nNature Energy | VOL 5 | October 2020 | 803\u2013810 | www.nature.com/natureenergy\n804\n\nAnalysis\nNature Energy\nwere lower (38% and 18%, respectively). It appears that ARPA-E \nselections were preferential towards companies with pre-award pat-\nents or financing, either because these attributes were used directly \nby ARPA-E staff as a signal of quality, or because they happen to \ncorrelate with the actual selection criteria.\nARPA-E startups patented more than similar firms\nPatents are an important part of the value chain for new ideas in \nenergy technology, and protecting intellectual property allows start-\nups to demonstrate their value to potential investors and/or compa-\nnies that might acquire them40\u201342. If the ARPA-E funding mechanism \nis effective at supporting breakthrough R&D, then ARPA-E startups \nshould show a higher rate of post-award patenting compared to \nrejected ARPA-E applicants as well as to similar cleantech startups.\nWe count the successful US patent applications that were \nfiled by each company in a given year, and we find that 80% of \nARPA-E-funded startups had filed a patent after their award in \n2010\u2014a greater proportion than any other group. However, it is not \nclear from these descriptive statistics whether ARPA-E\u2019s advantage \nin patenting is simply a result of selecting companies with more \nprior patenting activity. To test this possibility, we model post-2010 \npatenting activity using a Poisson regression with a control variable \nfor pre-2010 company patenting, as well as age, subsector and pre-\n2010 VC activity (Table 1).\nWe find that the patenting advantage for ARPA-E startups holds \nin regression analysis controlling for various factors, including \npre-award patenting behaviour. A coefficient of 0.7 corresponds \nto an incidence rate ratio of 2, meaning that startups who received \nARPA-E awards in 2010 went on to file successful patent applications \nat twice the rate of similar cleantech companies on average (Models \n2 and 3 in Table 1), even accounting for their pre-2010 patenting \nactivity. We repeat the regressions using a subsample of companies \nthat has been balanced on observable co-variates between ARPA-E \nand non-ARPA-E companies, using coarsened exact matching43. The \nmatched results confirm our finding (Models 4 and 5 in Table 1);  \nafter their award in 2010, ARPA-E companies filed patents at \nroughly double the rate that would be expected based on their age, \nsubsector and pre-2010 company profile. Confidence intervals for \nthe ARPA-E 2010 coefficient in Model 5 (not shown) indicate an \nincidence rate ratio between 1.5 and 3.2.\nBecause companies that apply for government grants differ \nfrom other companies in several ways44,45, we compare ARPA-E \nawardees in 2010 to companies that were rejected by ARPA-E in \nthe same year. These rejected applicants were publicly \u2018encouraged\u2019 \nby ARPA-E, which indicates that they were close to receiving the \naward. We take this approach to be approximating the idea behind \nregression discontinuity design methods, although the agency does \nnot assign differentiated scores within the rejected group or the \nawardees. We find that rejected ARPA-E applicants were no more \nlikely to patent than other cleantech startups. This indicates that \nthe increased post-award patenting activity of ARPA-E companies \nis not a product of self-selection of more innovative companies \ninto the applicant pool. We also test whether increased patenting \nis observed in the set of companies funded by EERE in 2010; these \ncompanies showed no advantage over other companies in terms of \npost-award patenting activity.\nPanel regressions of annual patents filed illustrate the consis-\ntently elevated patenting activity by ARPA-E-funded startups from \n2011 to 2014 (Fig. 2a; regression coefficients are in Supplementary \nTable 4). In Supplementary Table 5, we test for the likelihood of \nany patenting from the four groups of startups, and we compare \ncitation-weighted patent counts as well. The increased patenting \nactivity for ARPA-E startups compared to rejected applicants, EERE \nawardees and the wider pool of cleantech startups is robust to these \nalternative measurements. Although our main analyses exclude \nSBIR awardees, results from an expanded dataset including those \nTable 1 | ARPA-E patenting outcomes\nDependent variable\nPost-2010 \npatents\nPost-2010 patents\nPost-2010 patents\nPost-2010 patents\nPost-2010 patents\nForm\nPoisson\nPoisson\nPoisson\nPoisson\nPoisson\nModel\n(1)\n(2)\n(3)\n(4)\n(5)\nGroup:\nARPA-E 2010\n1.393***\n0.716***\n0.712***\n0.805***\n0.793***\n(0.391)\n(0.223)\n(0.208)\n(0.193)\n(0.196)\nARPA-E rejected\n0.392\n0.343\n\u22120.525\n(0.365)\n(0.312)\n(0.377)\nEERE 2010\n0.419\n0.081\n\u22120.171\n(0.370)\n(0.175)\n(0.184)\nCompany age in 2010\n0.160***\n\u22120.140**\n\u22120.141**\n\u22120.128\n\u22120.170\n(0.053)\n(0.060)\n(0.060)\n(0.108)\n(0.117)\nln(pre-2010 patents + 1)\n0.843***\n0.842***\n1.047***\n1.073***\n(0.101)\n(0.101)\n(0.090)\n(0.090)\nln(pre-2010 VC\u2009+\u2009$1)\n0.035***\n0.035***\n0.007\n0.006\n(0.012)\n(0.012)\n(0.020)\n(0.021)\nSubsector fixed effects\nY\nY\nY\nY\nY\nCoarsened exact matching\nY\nY\nObservations\n1,287\n1,287\n1,287\n238\n238\nPseudo R2\n0.128\n0.401\n0.402\n0.560\n0.563\nNotes: direct coefficients reported. Robust standard errors in parentheses. Incidence rate ratios for binary variables can be obtained by exponentiating the coefficient. *P\u2009<\u20090.10, **P\u2009<\u20090.05, ***P\u2009<\u20090.01. \nNon-matched sample is the full dataset of US cleantech startups in 2010. Pre-2010 counts are cumulative from company founding through 2009. Post-2010 patent counts are cumulative 2011 through \n2014. Base group is Other cleantech. Coarsened exact matching for ARPA-E versus non-ARPA-E companies was performed using the following criteria: subsector, company age (three cohorts), any pre-\n2010 patent and any pre-2010 VC.\nNature Energy | VOL 5 | October 2020 | 803\u2013810 | www.nature.com/natureenergy\n805\n\nAnalysis\nNature Energy\ncompanies is shown in Supplementary Table 6; SBIR awardees pro-\nduced patents at roughly half the rate of other cleantech startups, \non average.\nThe increased patenting by ARPA-E companies could be \nexplained by either a treatment effect or a selection effect. If par-\nticipation in the ARPA-E programme causes companies to produce \nmore patents, this treatment effect could be ascribed to ARPA-E\u2019s \nunique programme design. ARPA-E and EERE both provide finan-\ncial resources in the form of direct grants, but when ARPA-E was \nestablished in 2010, it was set apart by offering commercialization \nassistance and active project management by staff with technical \nexpertise; these interventions can be considered additional resources \nfor the firm. Furthermore, programme directors are empowered to \nmodify or terminate award agreements based on project perfor-\nmance46. This leverage may allow them to induce greater patenting \nsimply by encouraging awardees to patent; patents are a celebrated \noutput for ARPA-E as a whole47.\nHowever, we are not able to rule out the explanation that \nARPA-E chooses to fund companies with a higher propensity to \npatent. We address some obvious potential selection effects by \ncontrolling for pre-award company profiles and comparing to \nrejected applicants, but it is still possible that ARPA-E staff select \nmore-innovative companies based on attributes other than prior \npatenting activity or VC funds. A previous study of ARPA-E project \nselection practices revealed that the main driver of funding deci-\nsions at ARPA-E was the individual judgement of the programme \ndirector, rather than observable characteristics or even external \npeer review scores48. This use of individual discretion by ARPA-E \nstaff may allow them to judge research proposals on future pat-\nenting potential based on additional information beyond previous \npatents and private funding.\nARPA-E startups were no more successful than similar \nfirms\nARPA-E has highlighted the ability of its awardees to leverage their \nARPA-E award into follow-on funding47. One of the agency\u2019s goals \nis to fill a funding gap by de-risking technologies enough to make \nthem suitable for private investment and commercialization. If this \nmission were fully accomplished, then we would expect to see an \nincreased likelihood of success (measured in different ways) for \nARPA-E startups compared to similar companies, perhaps com-\nmensurate with the increased patenting activity measured above.\nWe measure three post-2010 business outcomes associated with \nincreased likelihood of market success for our sample of cleantech \nstartups: acquisition/IPO, survival through 2019 and amount of VC \nfunding raised through 2017 (Table 2). As with our patenting analysis \nabove, we run regressions for these outcomes, with controls for firm \nage and cleantech subsector to account for age- and sector-specific \neffects, as well as prior patenting and VC funds raised. We find no \nstatistically significant differences between ARPA-E companies and \nnon-ARPA-E companies for the three market success indicators, \nin either the whole sample or the matched subsample. A greater \nproportion of ARPA-E startups (52%) raised VC funds post-2010 \ncompared to the Other cleantech group (40%), but this difference \nis within the error of measurement when accounting for the firm\u2019s \npre-award profile. This nonsignificant difference could be mask-\ning an actual effect, due to the small sample size and large standard \nerror; confidence intervals for ARPA-E 2010 in Model 2 of Table 2 \n(not shown) yield an incidence rate ratio between 0.7 and 5.1. Thus, \nthis is a tentative result, and we hope that future research with a \nlarger sample of companies will be able to either identify an effect or \nbetter establish the lack thereof, for VC funding as well as acquisi-\ntion/IPO and survival.\nWe show two sets of alternative tests in Supplementary Table 7: \nlogit regressions for the binary outcome of any VC funding, and \nPoisson regressions for all private finance deal totals (including \ncommon stock and private equity as well as VC). The results are \nconsistent\u2014ARPA-E-funded companies were not measurably more \nsuccessful at raising private funds than similar startups from the \nOther cleantech group, despite their increased patenting activity \npost-award. Panel regressions also show no clear difference between \nARPA-E companies and others in annual VC funding 2011\u20132017 \n(Fig. 2b; regression coefficients are in Supplementary Table 4).\nWe find a positive and significant rate of VC funding for the \nEERE 2010 group compared to the Other cleantech group (P\u2009<\u20090.05) \nin the matched subsample, but a Wald test fails to reject the hypoth-\nesis that the coefficients for ARPA-E 2010 and EERE 2010 are \nequal (P\u2009=\u20090.90). On the other hand, we find significantly less busi-\nness success for firms whose application for ARPA-E funding was \ndenied in 2010. The coefficient on post-2010 VC funding for the \nARPA-E rejected group is significantly negative even when account-\ning for their lower rates of prior VC funding (P\u2009<\u20090.05); the rate of \nVC funding for this group is roughly one-tenth the funding rate of \nthe Other cleantech group (Model 2 of Table 2). A Wald test rejects \nthe hypothesis that the coefficients for ARPA-E 2010 and ARPA-E \nrejected are equal (P\u2009=\u20090.014), meaning that ARPA-E-rejected appli-\ncants performed significantly worse than did awardees.\nThe average rejected ARPA-E applicant was therefore less likely \nto succeed than other cleantech firms, while ARPA-E awardees \nperformed on par with similar cleantech firms. As with our patent-\ning results, we are not able to fully establish whether the difference \nbetween ARPA-E awardees and rejected applicants is due to treat-\nment, selection or both. However, also as with patenting, several \nfactors would seem to contradict a strong selection effect. First, the \nrejected applicants in this study are relatively high quality, having \nbeen encouraged by ARPA-E. Furthermore, the comparison already \naccounts for prior patenting and VC funding, which eliminates \nsome obvious mechanisms for selection. Finally, ARPA-E markets \n\u221210\n0\n10\n20\n30\n40\nAverage annual VC raised per firm\n(predictive margins, million US$)\n2011\n2012\n2013\n2014\n2015\n2016\n2017\nYear announced\nARPA\u2212E 2010\nARPA\u2212E rejected\nEERE 2010\nOther cleantech\n0\n1\n2\n3\n4\n5\nAverage annual patent filings\nper firm (predictive margins)\n2011\n2012\n2013\n2014\nYear filed\na\nb\nFig. 2 | Annual outcomes for 2010 cleantech startups by company type. \na,b, Plots of predictive margins for patents filed (a) and VC funds raised \n(b) are based on panel regressions controlling for subsector, company \nage and pre-2010 levels of patenting and VC funding, within a matched \nsubsample of cleantech companies; see Supplementary Table 4 for more \ndetail. Outcome data are cut off in 2014 for patents and 2017 for funding \ndue to a lag in observations. Predictive margins are the average predicted \noutcomes for companies in each group, using the observed values of all \nother variables. Error bars show 95% confidence intervals.\nNature Energy | VOL 5 | October 2020 | 803\u2013810 | www.nature.com/natureenergy\n806\n\nAnalysis\nNature Energy\nitself as a funder of high-risk ideas39, and previous research on \nARPA-E selection practices indicates that they do in fact prefer to \nfund proposals with greater uncertainty48.\nDiscussion\nStartups with an ARPA-E award were issued more patents in the \nyears following their award compared to similar cleantech com-\npanies. Although we are cautious in over-interpreting the size of \nthe effect, given the small sample of 25 ARPA-E companies, the \ntrend towards greater patenting is significantly positive (P\u2009<\u20090.01). \nWe show that the higher rate of patenting is not shared by rejected \nARPA-E applicants, indicating that the effect is not caused by a \nhigher propensity to patent among firms that self-select into the \nARPA-E applicant pool. Furthermore, the patenting advantage is \nnot shared by similar startups with R&D awards from elsewhere in \nDOE, indicating that increased patenting is not caused simply by \nthe infusion of public financial resources. However, two possible \nexplanations remain: treatment (ARPA-E awards caused cleantech \nstartups to produce more patents) and/or selection (ARPA-E pro-\nvided a subsidy to more innovative startups, who would have pro-\nduced more patents anyway).\nWithout a randomized assignment of ARPA-E awards among a \nset of eligible firms49,50, we are not able to disentangle selection and \ntreatment as possible explanations for the increased patenting post-\n2010 among ARPA-E awardees relative to other cleantech startups. \nGiven our understanding that ARPA-E awardees are selected to be \nthe best proposed ideas for solving a technical challenge, the treat-\nment is far from random. Therefore, we face an unavoidable risk of \nbias from omitted variables, like project quality.\nAnd yet, if ARPA-E programme directors are skilled at predict-\ning future patenting outcomes and choose to fund those compa-\nnies with the greatest potential for future patenting, this could be \nan important result in itself given the environmental and societal \nimportance of cleantech innovation. Predicting research outcomes \nand the success of technical projects is notoriously difficult, even for \nindustry experts or venture capitalists51. Although many ARPA-E \nstartups had private funding prior to their award, those investors \nmay not have been willing to support the particular R&D projects \nfunded by ARPA-E\u2014projects with transformative potential but \ngreater risk. In a report from the US Government Accountability \nOffice, investment firms stated that they \u201cgenerally do not fund \nprojects that rely on unproven technologies\u201d52. One public company \nsaid that ARPA-E projects would not meet the criteria for internal \ninvestment because \u201cthe rate of return on investment required by its \nmanagement was at least 20 percent per year\u201d52. ARPA-E funding \nallowed startups to pursue mission-aligned R&D projects without \nrequiring short-term profitability, and we find that these startups \nwere more inventive than others.\nDespite the demonstrated advantage of ARPA-E awardees in \nproducing inventions, we find no clear evidence that they perform \ndifferently from similar cleantech startups as a whole in terms of \nacquisition/IPO, survival or VC funding post-award within 10\u201315\u2009yr \nof founding. We face the limitations of a small sample of ARPA-E \ncompanies, which partly explains the large standard error on these \nmeasurements, and a relatively short period of time following \nARPA-E funding for cleantech innovations to pay off. Nonetheless, \nwe do measure significantly better business outcomes for ARPA-E \nstartups than for our sample of high-quality rejected applicants. \nAlthough we cannot confirm the presence of a treatment effect, we \nnote that such an effect could be consistent with the lack of business \nadvantage for the kinds of firms that apply for ARPA-E support.  \nIf ARPA-E attracted applicants whose plans for technology develop-\nment were riskier or more capital intensive, then these firms may \nhave been generally at a disadvantage in terms of business success. \nA positive treatment effect could have allowed ARPA-E awardees to \novercome negative factors afflicting ARPA-E applicants and to catch \nup with other firms, while not fully closing the gap between public \nand private funding sources for these innovative cleantech startups.\nTable 2 | ARPA-E business outcomes\nDependent variable\nPost-2010 VC\nAcquired or IPO\nSurvival\nForm\nPoisson\nPoisson\nLogit\nLogit\nLogit\nLogit\nModel\n(1)\n(2)\n(3)\n(4)\n(5)\n(6)\nGroup:\nARPA-E 2010\n0.053\n0.643\n\u22120.663\n\u22120.330\n0.250\n0.308\n(0.620)\n(0.506)\n(0.528)\n(0.632)\n(0.501)\n(0.645)\nARPA-E rejected\n\u22121.407**\n\u22122.231**\n\u22121.716***\n\u22121.420*\n\u22121.010***\n\u22121.165**\n(0.641)\n(1.068)\n(0.615)\n(0.784)\n(0.347)\n(0.589)\nEERE 2010\n\u22120.665\n0.569**\n\u22120.057\n\u22120.029\n\u22120.321\n\u22120.176\n(0.432)\n(0.229)\n(0.296)\n(0.729)\n(0.296)\n(1.274)\nCompany age in 2010\n\u22120.620*\n0.093\n0.186***\n0.217\n0.065\n\u22120.241\n(0.374)\n(0.161)\n(0.044)\n(0.187)\n(0.048)\n(0.213)\nln(pre-2010 patents + 1)\n0.536\n0.291**\n0.109\n0.315\n0.234**\n0.820**\n(0.339)\n(0.129)\n(0.092)\n(0.253)\n(0.114)\n(0.327)\nln(pre-2010 VC\u2009+\u2009$1)\n0.169***\n0.044\n0.010\n0.018\n\u22120.023**\n\u22120.057*\n(0.046)\n(0.032)\n(0.010)\n(0.028)\n(0.011)\n(0.035)\nSubsector fixed effects\nY\nY\nY\nY\nY\nY\nCoarsened exact matching\nY\nY\nY\nObservations\n1,287\n238\n1,282\n238\n1,287\n228\nPseudo R2\n0.456\n0.370\n0.047\n0.109\n0.033\n0.164\nNotes: direct coefficients reported. Robust standard errors in parentheses. Incidence rate ratios (Models 1 and 2) or odds ratios (Models 3\u20136) for binary variables can be obtained by exponentiating the \ncoefficient. *P\u2009<\u20090.10, **P\u2009<\u20090.05, ***P\u2009<\u20090.01. Non-matched sample is the full dataset of cleantech startups in 2010. Pre-2010 counts are cumulative from company founding through 2009. Post-2010 VC \nfunding totals are cumulative US dollars 2011 through 2017. Coarsened exact matching for ARPA-E versus non-ARPA-E companies was performed using the following criteria: subsector, company age  \n(three cohorts), any pre-2010 patent and any pre-2010 VC.\nNature Energy | VOL 5 | October 2020 | 803\u2013810 | www.nature.com/natureenergy\n807\n\nAnalysis\nNature Energy\nHardware-based energy technology innovation requires large \namounts of capital and long timescales to become fully commercial-\nized. Private investors have shown some renewed interest in energy \ntechnologies, but their interest appears limited to digital innova-\ntions such as demand-management software53. More than half of the  \nstartups in our dataset are manufacturing firms, according to their \nNorth American Industry Classification System (NAICS) codes. It \nis not surprising if ARPA-E alone has not fully solved the \u2018valley \nof death\u2019 problem for innovative cleantech companies, which has \nbeen shown to be especially acute in the demonstration phase54. \nComplementary innovation policies, such as increased funding \nfor demonstration and commercialization, in-kind support from \nnational laboratories and targeted procurement programmes, may \nbe needed to allow scale-up beyond the R&D phase and to ensure \nthat cleantech innovations can leverage additional private financing \nand transition to the market.\nTen years after ARPA-E first began issuing awards, our findings \nprovide important contributions to the ongoing discussions in the \nUS about the role of governments as early investors in clean energy \ntechnologies. Our results will also inform other governments around \nthe world that are interested in replicating or adapting the ARPA \nmodel as a mechanism for supporting mission-oriented innova-\ntion, as well as researchers interested in the design of public fund-\ning mechanisms for supporting R&D to advance energy-related and \nindustrial policy goals.\nMethods\nIdentification of cleantech companies. Cleantech is not a conventional market \ncategory; it cuts across many traditional industries and sectors, including \nagriculture, electricity, transportation, water and waste management. While \n\u2018clean energy technology\u2019 does not fully encompass the range of technologies \nthat can be considered cleantech, the two terms are sometimes used as \ninterchangeably37. Cleantech companies are distinguished by offering reduced \nenvironmental impact compared to the status quo, although this reduced \nenvironmental impact of a cleantech company\u2019s product or service may or may \nnot have market value. The label of \u2018clean\u2019 has value to the extent that customers \nand investors recognize the environmental harm from existing technologies. \nFor this reason, it can be difficult to identify a given company as cleantech, even \nknowing its product offerings, without knowing how the company signals its \nenvironmental value and impact.\nIn this work, cleantech firms were identified from three data sources: the \nCleantech Group i3 dataset, recent DOE awardees and applicants for ARPA-E \nfunding in fiscal year 2010. The commercial i3 dataset contains a list of cleantech \ncompanies \u201cfocused on emerging and innovative companies\u201d. The Cleantech Group \nwebsite states that they \u201ctypically track start-ups and SMEs from the seed stage to \nlate-stage private equity \u2013 and everything in between\u201d55. This database was selected \nfor its inclusion of detailed information on subsector and founding year and its \nfocus on entrepreneurial firms.\nTo identify companies recently funded by DOE, we obtained transaction data \non federal financial assistance from fiscal years 2009\u20132016 on USAspending.gov. \nFollowing the method of Goldstein and Narayanamurti3, we removed duplicate \ntransactions and combined transactions to create an award-level dataset. This \ndataset includes grants, which entail relatively unfettered financial support, as well \nas cooperative agreements, which require substantial involvement of the funding \nagency; we refer to these collectively as \u2018awards\u2019. We exclude contracts, which \nare used for government procurement. We limit our analysis to awards given \nto businesses by ARPA-E and EERE, with a start date in the range of fiscal year \n2010\u20132015. In our analysis, we pay special attention to awards that began in fiscal \nyear 2010 (October 1, 2009, through September 30, 2010), which was the year of \nthe ARRA stimulus and the first year of operation for ARPA-E.\nA final set of cleantech startups are identified as rejected ARPA-E applicants. \nUnfunded applicants are typically not publicly announced, but in 2011 and 2012, \nARPA-E posted a list of \u2018encouraged applicants\u2019 on its website. These applicants \nagreed to make a description of their proposed technology publicly available, in \norder to \u201chelp foster connections that will accelerate technology development\u201d. \nIn total, 100 unfunded projects were listed prior to September 2011. Based on \nthe dates when various ARPA-E awards were announced, we conclude that \nthese applications were to one of ARPA-E\u2019s first seven technical programmes, \nall of which were initiated in fiscal year 2010: OPEN 2009, BEEST, Electrofuels, \nIMPACCT, ADEPT, BEETIT and GRIDS. Based on internal data obtained from \nARPA-E48, we find that the \u2018encouraged applicants\u2019 sample is only 18% of the 565 \nrejected applications to the 2010 ARPA-E funding opportunities.\nOne startup in our dataset was funded by both ARPA-E and EERE in fiscal \nyear 2010, and one ARPA-E-rejected applicant was funded by EERE in that same \nyear. For the purpose of panel regressions and the bar plots in Fig. 1, companies are \nassigned to only one group; these overlapping cases are assigned to ARPA-E 2010 \nand ARPA-E rejected, respectively.\nCompany characteristics. We combined several datasets to determine the \nfounding year of each cleantech company identified by the methods above.  \nThe i3 dataset contains data on company founding year. USAspending.gov does \nnot include information on the age of a business; DOE awardees and rejected \nARPA-E applicants required manual research to determine founding year. When \nnot available on the company\u2019s website, founding year was obtained from Orbis, \nFactset, public news sources and/or state business registration documents. We limit \nour dataset to companies that were startups in 2010, that is, with a founding year in \nthe range 2005\u20132010.\nWe also used manual research to determine each company\u2019s subsector within \ncleantech. The i3 dataset includes a categorization for \u2018primary sector\u2019 with 19 \ndistinct labels; this label allows greater cohesion than the NAICS subsector, of \nwhich there are at least 50 unique categories represented within our dataset. The \n19 cleantech subsectors include clean energy categories, which compose  \nthe majority of the companies in our dataset (88%), as well as several categories \nthat have broader application beyond clean energy: Advanced Materials, \nAgriculture and Food, Air, Recycling and Waste, Water and Wastewater, and \nMiscellaneous Cleantech.\nWe checked the categorization for each company, based on the company \nwebsite and other publicly available information online. For companies that were \nnot in the i3 dataset, the cleantech subsector was determined by a review of the \ncompany\u2019s website as of 2017. If the website was not active as of 2017, we used \nthe most recent archive on The Wayback Machine (https://web.archive.org/). We \nfocused on the company\u2019s self-description and degree of emphasis on various \ncleantech applications for their technology; our determination does not account for \npossible changes in technology area over time.\nSome companies were found to operate primarily outside the 19 cleantech \nsubsector classifications. We exclude these non-cleantech companies from our \ndataset.\nOutcome variables\u2212patents. We searched the US Patent and Trademark Office \nwebsite for patents assigned to each cleantech startup, issued on or before April 20, \n2017. Design patents were excluded from the dataset. Observations were manually \nreviewed if the name of the assignee was not an exact match for the name of the \ncompany, or if the year of the patent filing was prior to the recorded founding year \nof the company. Patents are counted by the year in which the application was filed. \nObservations were collected through December 31, 2014, to account for the lag \nin observing a successful patent application. A citation-weighted count was also \ngenerated, following Trajtenberg\u2019s method of adding the annual patent count to a \nsum of all forward citations to those patents56. Citations to patents were observed \nthrough April 1, 2017.\nOutcome variables\u2212private investments. We collected data on private funding \nfrom three sources: Thomson ONE, Crunchbase and Preqin. To obtain the \ntotal funding received by a company at the end of a given calendar year, we use \nthe maximum amount recorded from any of the three sources. We limited our \nobservations to funding deals announced between January 1, 2004, and December \n31, 2017. For VC funding, we include the following deal types: Venture Capital \nEquity Investment (Thomson ONE); Seed, Series A-I, Venture\u2014Series Unknown \n(Crunchbase); and Seed, Series A-I (Preqin). For private funding variables, we \ninclude all VC funding deals, plus the following deal types as well: Common Stock, \nConvertible Preferred Stock, Preferred Shares and Warrants (Thomson ONE); \nAngel, Convertible Note, Corporate Round, Equity Crowdfunding, Funding \nRound, Post-IPO Equity, Pre-Seed and Private Equity (Crunchbase); Add-on, \nAngel, Growth Capital/Expansion, Private Investment in Public Equity (PIPE) and \nUnspecified Round (Preqin).\nBecause funding decisions may be influenced by the signalling effect \nof government grants, we aimed to observe fundraising activities for DOE \nawardees before their grants were announced publicly. We consider funding \nannounced on or before December 31, 2009, to be pre-award funding. This \nmethod introduces two countervailing sources of error. On one hand, it \noverestimates pre-award funding for companies whose awards were announced \nin 2009, by including investments that may have been influenced by that \nannouncement; for example, ARPA-E grants in the OPEN 2009 programme were \nannounced on October 26, 2009. On the other hand, our method underestimates \npre-award funding for companies whose awards were announced in 2010, by \nexcluding investments that were made before the announcement. For example, \nsome ARPA-E programmes announced awards in spring and summer of 2010 \n(April 29 for BEEST, Electrofuels and IMPACCT; July 12 for BEETIT, GRIDS \nand ADEPT.)\nFor consistency with the funding data, we limit pre-award patent observations \nto those filed on or before December 31, 2009. For both patents and private \nfunding, we ignore observations (patents filed or funding announced) in \ncalendar year 2010. We test the main results for robustness with two alternative \nspecifications in Supplementary Table 8: pre-award outcomes including 2010, and \npost-award outcomes including 2010.\nNature Energy | VOL 5 | October 2020 | 803\u2013810 | www.nature.com/natureenergy\n808\n\nAnalysis\nNature Energy\nOutcome variables\u2014exit and survival. We further collected data from public \nsources on two events for each company in the dataset: whether a company had \nbeen acquired or issued an IPO (indicating an \u2018exit\u2019 for investors) and whether \nthey remained in business as of December 2019. Data on acquisition and IPO were \ncollected from i3, Crunchbase, Thomson ONE and Orbis, and complemented with \na Web search. The outcome was operationalized as a binary variable equalling 1 for \neither an acquisition or an IPO, and 0 otherwise.\nSurvival is an important indicator of performance in the startup context, \nwhere many companies end in failure (bankruptcy, closure or liquidation)57,58. We \ncompiled data on whether or not a company was still in business as of December \n2019 in three steps. First, we searched for each company on the Web using \nkeywords for bankruptcy, closure and liquidation. We also checked whether the \nfirms still had webpages and when they were last updated, using The Wayback \nMachine and Google cached results. Finally, for companies with outdated or \nnonexistent online records, we searched their status in state-level business \nregistration databases. Acquired companies may be absorbed into their new parent \ncompany and no longer observable as an independent entity, so we operationalized \nfirm survival to also account for acquisition/IPO outcomes; survival is a binary \nvariable that takes on the value 1 if a firm was still active or had undergone \nacquisition/IPO, and 0 otherwise.\nRegressions. To estimate differences between ARPA-E-funded firms and the \nother groups of cleantech startups, we conducted a set of regression analyses. We \nestimate post-2010 patent counts and funding amounts (Yi) for firm i using the \nfollowing Poisson model:\nln Yij\n\u001f\n\u001e\n\u00bc \u03b20 \u00fe \u03b21\u00f0ARPA-E\u00dei \u00fe \u03b22\u00f0ARPA-E rejected\u00dei \u00fe \u03b23\u00f0EERE\u00dei\n\u00fe\u03b24Age \u00fe \u03b25 ln Pi \u00fe 1\n\u00f0\n\u00de \u00fe \u03b26 ln Zi \u00fe 1\n\u00f0\n\u00de \u00fe \u03c6j\nwhere \u03b21\u20133 are the coefficients of interest for three groups of cleantech startups, \ncompared to the Other cleantech group. \u03b24\u20136 are the coefficients controlling for \ncompany age, prior patenting and prior VC funding, respectively. Pi is the number \nof patents filed pre-2010, Zi is the amount of pre-2010 VC funding and \u03c6j is a fixed \neffect for the cleantech subsector of the firm (j). Subsector fixed effects are not \nreported for brevity, but outcomes differ dramatically across subsectors; descriptive \nstatistics are shown in Supplementary Table 9.\nAll regression tables report the coefficients directly. For Poisson regressions, \ncoefficients can be converted to incidence rate ratios between firms in different \ngroups\u2014for example, ARPA-E and non-ARPA-E firms\u2014by exponentiating the \nrelevant \u03b2 coefficient. We estimate the likelihood of acquisition/IPO and firm \nsurvival using a logit model, with the same set of explanatory and control variables \nas in equation (1). Here, the exponentiated coefficient for binary independent \nvariables (for example ARPA-E 2010) indicates the odds ratio of an event for \nstartups in a given group compared to the Other cleantech group.\nCoarsened exact matching is performed to reduce imbalance between \nARPA-E and non-ARPA-E cleantech startups. The cem package in Stata is used \nto determine importance weights59. The criteria for company matching are as \nfollows: 19 individual cleantech subsectors, three age cohorts (founded 2005\u20132006, \n2007\u22122008 and 2009\u20132010), whether they had any pre-2010 patents and whether \nthey had any pre-2010 VC funding. The matched dataset contains 238 companies \n(including 23 of 25 ARPA-E companies) across 17 matched strata. The multivariate \nL1 distance is reduced to 1.25\u2009\u00d7\u200910\u201316 from 0.731 pre-matching43.\nAlternative regression models are used to check for robustness. We convert \nour cross-sectional data on startups to a panel dataset, and we repeat the Poisson \nregressions above with company-years as the unit of analysis, for years starting  \nin 2011 (Supplementary Table 4). We estimate logit regressions for whether or  \nnot a firm patented or received any VC funding (Supplementary Tables 5 and 7).  \nWe conduct Tobit analysis to account for left-censored data on the latent \nvariable of propensity to patent or raise VC funds (Supplementary Table 10), \nwith a lower limit of 0 patents filed or $0 raised. The coefficient in this case is \nan estimate of linear increase in the uncensored latent variable. We also estimate \ndifference-in-differences to account for pre-2010 and post-2010 time trends \n(Supplementary Table 11), where the coefficient is an estimate of linear increase in \nthe outcome variable.\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nThe data on cleantech firm identification used in this study were made available to us \nby Cleantech Group under a restricted license and are therefore not publicly available. \nSupplementary Data 1 contains an anonymized version of our company-level \nanalytical sample with the names of the firms removed. Supplementary Data 2 \ncontains the company-year panel dataset. The full dataset is available from the \nauthors upon reasonable request and with permission of Cleantech Group.\nCode availability\nSupplementary Code 1 contains the Stata code that enables the reproduction of  \nour main analysis.\nReceived: 19 July 2019; Accepted: 30 July 2020;  \nPublished online: 14 September 2020\nReferences\n\t1.\t America COMPETES Act (110th US Congress, 2007).\n\t2.\t Anad\u00f3n, L. D., Chan, G. A., Bin-nun, A. Y. & Narayanamurti, V. The pressing \nenergy innovation challenge of the US National Laboratories. Nat. Energy \nhttps://doi.org/10.1038/nenergy.2016.117 (2016).\n\t3.\t Goldstein, A. P. & Narayanamurti, V. Simultaneous pursuit of discovery and \ninvention in the US department of energy. Res. Policy https://doi.org/10.1016/ \nj.respol.2018.05.005 (2018).\n\t4.\t Anad\u00f3n, L. D., Bunn, M. & Narayanamurti, V. Transforming U.S. Energy \nInnovation (Cambridge Univ. Press, 2014).\n\t5.\t Accelerating the Pace of Change in Energy Technologies Through an Integrated \nFederal Energy Policy (President\u2019s Council of Advisors on Science and \nTechnology (PCAST), 2010).\n\t6.\t A Business Plan for America\u2019s Energy Future (American Energy Innovation \nCouncil, 2010).\n\t7.\t Jaffe, A. B., Newell, R. G. & Stavins, R. N. A tale of two market failures: \ntechnology and environmental policy. Ecol. Econ. 54, 164\u2013174 (2005).\n\t8.\t Kattel, R. & Mazzucato, M. Mission-oriented innovation policy and dynamic \ncapabilities in the public sector. Ind. Corp. Chang. 27, 787\u2013801 (2018).\n\t9.\t Mazzucato, M. The Entrepreneurial State: Debunking Public vs. Private Sector \nMyths (Anthem Press, 2013).\n\t10.\tArrow, K. in The Rate and Direction of Inventive Activity: Economic and Social \nFactors 609\u2013626 (Princeton University Press, 1962).\n\t11.\tNelson, R. R. The simple economics of basic scientific research. J. Polit. Econ. \n67, 297\u2013306 (1959).\n\t12.\tKerr, W. R., Nanda, R. & Rhodes-Kropf, M. Entrepreneurship as \nexperimentation. J. Econ. Perspect. 28, 25\u201348 (2014).\n\t13.\tHall, R. E. & Woodward, S. E. The burden of the nondiversifiable risk of \nentrepreneurship. Am. Econ. Rev. 100, 1163\u20131194 (2010).\n\t14.\tNanda, R. & Rhodes-Kropf, M. Financing risk and innovation. Manag. Sci. \n63, 901\u2013918 (2017).\n\t15.\tAmerica First: A Budget Blueprint to Make America Great Again (US Office of \nManagement and Budget, 2017).\n\t16.\tBronzini, R. & Piselli, P. The impact of R&D subsidies on firm innovation. \nRes. Policy 45, 442\u2013457 (2016).\n\t17.\tAerts, K. & Schmidt, T. S. Two for the price of one?: additionality effects of \nR&D subsidies: a comparison between Flanders and Germany. Res. Policy 37, \n806\u2013822 (2008).\n\t18.\tGonz\u00e1lez, X. & Paz\u00f3, C. Do public subsidies stimulate private R&D spending? \nRes. Policy 37, 371\u2013389 (2008).\n\t19.\tHussinger, K. R&D and subsidies at the firm level: an application of \nparametric and semiparametric two-step selection models. J. Appl. Econom. \n23, 729\u2013747 (2008).\n\t20.\tDeleidi, M., Mazzucato, M. & Semieniuk, G. Neither crowding in nor out: \npublic direct investment mobilising private investment into renewable \nelectricity projects. Energy Policy https://doi.org/10.1016/j.enpol.2019.111195 \n(2019).\n\t21.\tDoblinger, C., Surana, K. & Anad\u00f3n, L. D. Governments as partners: the \nrole of alliances in U.S. cleantech startup innovation. Res. Policy 48, \n1458\u20131475 (2019).\n\t22.\tHowell, S. T. Financing innovation: evidence from R&D grants. Am. Econ. \nRev. 107, 1136\u20131164 (2017).\n\t23.\tPless, J. Are\u201cComplementary Policies\u201d Substitutes? Evidence from R&D Subsidies \nin the UK (Jacquelyn Pless, 2018); http://jacquelynpless.com/research/\n\t24.\tAzoulay, P., Fuchs, E. R. H., Goldstein, A. P. & Kearney, M. Funding \nbreakthrough research: promises and challenges of the \u2018ARPA Model\u2019.  \nInnov. Policy Econ. 19, 69\u201396 (2019).\n\t25.\tFuchs, E. R. H. Rethinking the role of the state in technology development: \nDARPA and the case for embedded network governance. Res. Policy 39, \n1133\u20131147 (2010).\n\t26.\tStokstad, E. UK cues up big funding increases for R&D. Science https://doi.\norg/10.1126/science.abb6928 (2020).\n\t27.\tBonvillian, W. B. DARPA and its ARPA-E and IARPA clones: a unique \ninnovation organization model. Ind. Corp. Chang. 27, 897\u2013914 (2018).\n\t28.\tChan, G. A., Goldstein, A. P., Bin-Nun, A., Anad\u00f3n, L. D. & Narayanamurti, \nV. Six principles for energy innovation. Nature 552, 25\u201327 (2017).\n\t29.\tPless, J., Hepburn, C. & Farrell, N. Bringing rigour to energy innovation \npolicy evaluation. Nat. Energy https://doi.org/10.1038/s41560-020-0557-1 \n(2020).\n\t30.\tAudretsch, D. B., Keilbach, M. C. & Lehmann, E. E. Entrepreneurship and \nEconomic Growth (Oxford Univ. Press, 2006).\n\t31.\tBettis, R. A. & Hitt, M. A. The new competitive landscape. Strateg. Manag. J. \n16, 7\u201319 (1995).\n\t32.\tHockerts, K. & W\u00fcstenhagen, R. Greening Goliaths versus emerging \nDavids\u2014theorizing about the role of incumbents and new entrants in \nsustainable entrepreneurship. J. Bus. Ventur. 25, 481\u2013492 (2010).\nNature Energy | VOL 5 | October 2020 | 803\u2013810 | www.nature.com/natureenergy\n809\n\nAnalysis\nNature Energy\n\t33.\tDushnitsky, G. & Lenox, M. J. When do incumbents learn from \nentrepreneurial ventures?: corporate venture capital and investing firm \ninnovation rates. Res. Policy 34, 615\u2013639 (2005).\n\t34.\tMarkard, J. & Truffer, B. Innovation processes in large technical systems: \nmarket liberalization as a driver for radical change? Res. Policy 35,  \n609\u2013625 (2006).\n\t35.\tDoblinger, C. & Soppe, B. Change-actors in the U.S. electric energy system: \nthe role of environmental groups in utility adoption and diffusion of wind \npower. Energy Policy 61, 274\u2013284 (2013).\n\t36.\tPopp, D., Pless, J., Hascic, I. & Johnstone, N. Innovation and Entrepreneurship \nin the Energy Sector NBER Working Paper No. 27145 (National Bureau of \nEconomic Research, 2020).\n\t37.\tGaddy, B. E., Sivaram, V., Jones, T. B. & Wayman, L. Venture capital and \ncleantech: the wrong model for energy innovation. Energy Policy 102, \n385\u2013395 (2017).\n\t38.\tA Retrospective Assessment of Clean Energy Investments in the Recovery Act \n(Council of Economic Advisors, 2016).\n\t39.\tAn Assessment of ARPA-E (National Academies of Sciences, Engineering, and \nMedicine, 2017).\n\t40.\tIslam, M., Fremeth, A. & Marcus, A. Signaling by early stage startups: US \ngovernment research grants and venture capital funding. J. Bus. Ventur. 33, \n35\u201351 (2018).\n\t41.\tConti, A., Thursby, J. & Thursby, M. Patents as signals for startup financing.  \nJ. Ind. Econ. 61, 592\u2013622 (2013).\n\t42.\tGuzman, J. & Stern, S. Where is silicon valley? Science 347, 606\u2013609 (2015).\n\t43.\tIacus, S. M., King, G. & Porro, G. Multivariate matching methods that are \nmonotonic imbalance bounding. J. Am. Stat. Assoc. 106, 345\u2013361 (2011).\n\t44.\tFeldman, M. P. & Kelley, M. R. The ex ante assessment of knowledge \nspillovers: government R&D policy, economic incentives and private firm \nbehavior. Res. Policy 35, 1509\u20131521 (2006).\n\t45.\tConti, A. Entrepreneurial finance and the effects of restrictions on \ngovernment R & D subsidies. Organ. Sci. 29, 134\u2013153 (2018).\n\t46.\tGoldstein, A. P. & Kearney, M. Know when to fold \u2019em: an empirical \ndescription of risk management in public research funding. Res. Policy 49, \n103873 (2020).\n\t47.\tARPA-E Impacts: A Sampling of Project Outcomes Vol. 3 (ARPA-E, 2018).\n\t48.\tGoldstein, A. P. & Kearney, M. Uncertainty and individual discretion in \nallocating research funds. SSRN https://doi.org/10.2139/ssrn.3012169 (2017).\n\t49.\tAzoulay, P. Research efficiency: turn the scientific method on ourselves. \nNature 484, 31\u201332 (2012).\n\t50.\tJaffe, A. B. Building programme evaluation into the design of public \nresearch-support programmes. Oxf. Rev. Econ. Policy 18, 22\u201334 (2002).\n\t51.\tScott, E., Shu, P. & Lubynsky, R. Entrepreneurial uncertainty and expert \nevaluation: an empirical analysis. Manage. Sci. 66, 1005\u20131507 (2018).\n\t52.\tAdvanced Research Projects Agency\u2013Energy Could Benefit from Information on \nApplicants\u2019 Prior Funding (Government Accountability Office, 2012).\n\t53.\tBumpus, A. & Comello, S. Emerging clean energy technology investment \ntrends. Nat. Clim. Chang. 7, 382\u2013385 (2017).\n\t54.\tNemet, G. F., Zipperer, V. & Kraus, M. The valley of death, the technology \npork barrel, and public support for large demonstration projects. Energy \nPolicy 119, 154\u2013167 (2018).\n\t55.\tOur Coverage (Cleantech Group, 2017).\n\t56.\tTrajtenberg, M. Patents, Citations and Innovations: Tracing the Links  \n(National Bureau of Economic Research, 1987).\n\t57.\tClarysse, B., Wright, M., Bruneel, J. & Mahajan, A. Creating value in \necosystems: crossing the chasm between knowledge and business ecosystems. \nRes. Policy 43, 1164\u20131176 (2014).\n\t58.\tHannan, M. T. & Freeman, J. Organizational Ecology (Harvard Univ.  \nPress, 1989).\n\t59.\tBlackwell, M., Iacus, S., King, G. & Porro, G. CEM: coarsened exact matching \nin Stata. Stata J. 9, 524\u2013546 (2009).\nAcknowledgements\nWe thank Cleantech Group for providing access to their i3 database. We also acknowledge \nexcellent research assistance by A. Amini, B. Dirscherl, P. Gigglinger, A. Hammerstingl, \nS. Kurowski, C. Lex and N. Schlosser. We acknowledge funding from the EU Framework \nProgramme for Research and Innovation H2020 under grant agreement no. 730403 \n(INNOPATHS) and from the Belfer Center\u2019s Science, Technology, and Public Policy \nProgram. We also acknowledge a grant from John and Elizabeth Armstrong.\nAuthor contributions\nA.G., C.D. and L.D.A. designed the study. A.G. and C.D. collected data. A.G. analysed \ndata and ran statistical tests. A.G. and C.D. wrote the paper. L.D.A and E.B. guided the \nstudy and edited the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-020-00683-8.\nCorrespondence and requests for materials should be addressed to A.G.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2020\nNature Energy | VOL 5 | October 2020 | 803\u2013810 | www.nature.com/natureenergy\n810\n\n1\nnature research  |  reporting summary\nOctober 2018\nCorresponding author(s):\nAnna P. Goldstein\nLast updated by author(s): Jul 29, 2020\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nWe used Python code to access public information on issued patents\nData analysis\nWe used Stata code to construct our datasets and run regressions\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. \nWe strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nWe have included a Data Availability statement that describes the elements of our dataset that are publicly available.\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\n\n2\nnature research  |  reporting summary\nOctober 2018\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThe study applies a quantitative design based on secondary data. We use Poisson regressions for our statistical analyses.\nResearch sample\nOur database contains 1,287 cleantech companies founded in the Unites States from 2005 to 2010. We include startups that were \nfunded by another office in the US Department of Energy (DOE), those that were publicly announced as rejected ARPA-E applicants, and \nthe larger universe of non-DOE-funded cleantech startups based on information from Cleantech Group. The Cleantech Group i3 platform \nis a rich and thorough collection of information on the global cleantech sector. It applies various mechanisms to collect information on \nfirms, alliances between firms, industry development, and investments and this information is regularly updated and checked either by a \nresearch team or directly by the firms. By combining multiple datasets, we were able to design a representative sample of the universe of \nUS cleantech startups that were founded between 2005 and 2010, including 25 that were funded by ARPA-E in 2010. \nSampling strategy\nWe started by classifying the firms that are funded by the US DOE as part of the ARPA-E or EERE programs into startups and older firms. \nWe considered a firm to be a startup as of 2010 if it was founded between 2005 and 2010 (inclusive). We further expanded our dataset \nby adding non-DOE funded startups (founded between 2005 and 2010) from Cleantech Group that operate in the US and belong to the \nfollowing (cleantech) sub-sectors: Energy Efficiency, Smart Grid, Biofuels & Biochemicals, Conventional Fuels, Biomass Generation, Fuel \nCells & Hydrogen, Geothermal, Hydro & Marine Power, Nuclear, Solar, Wind, Energy Storage, Transportation, Advanced Materials, \nAgriculture & Food, Air, Recycling & Waste, Water & Wastewater, or Miscellaneous. \nData collection\nThe data collection combined secondary data from different sources:  \na) Proprietary with paid access: Cleantech Group i3, Crunchbase, Thomson ONE, Preqin, Orbis \nb) Publicly available data: DOE funding on USAspending.gov, DOE award characterization on Portfolio Analysis And Management System \nwebsite, patents on USPTO website\nTiming\nObservations of company identification and outcomes began Oct. 1, 2016 and ended Dec. 31, 2019. \nData exclusions\nFirms outside of cleantech were excluded, as well as those founded before 2005 or after 2010.\nNon-participation\nThis study did not involve participant interaction.\nRandomization\nThis is an observational study with no experimental groups.\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology\nAnimals and other organisms\nHuman research participants\nClinical data\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\n\n\n    Scientific Research Findings:", "answer": "We find that startups that received funding from ARPA-E in its first year of operation filed twice the number of patents in subsequent years compared to other similar firms without ARPA-E funding. This was true even when accounting for the age of the firm, its cleantech sub-sector (for example, energy efficiency or energy storage), and its pre-2010 patenting and venture capital (VC) financing. ARPA-E is therefore either directly enhancing the ability of awardees to innovate and/or selecting highly innovative firms to receive non-dilutive cash and other forms of support. ARPA-E startups also had better business outcomes, including VC fundraising, survival and being acquired or going public, compared to firms that applied for ARPA-E funding in 2010 and were \u2018encouraged\u2019 by ARPA-E but not ultimately funded. However, ARPA-E startups did not have better business outcomes than the larger pool of US cleantech startups that did not apply for ARPA-E funds.", "id": 21}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-020-0596-7\n1Council on Energy, Environment and Water, New Delhi, India. 2Department of Environmental Health Sciences, Columbia Mailman School of Public Health, \nNew York, NY, USA. \u2709e-mail: sunil.mani@ceew.in; saurabhtripathi.in@gmail.com\nI\nndia is in the midst of the largest household energy transition \nto date, with about one hundred million households acquiring a \nliquified petroleum gas (LPG) connection since 20161. In 2011, \nthree-fifths of Indian households\u2014primarily in rural areas\u2014relied \non solid fuels to meet their daily cooking and heating needs2, result-\ning in substantial negative impacts on health3, the economy4 and the \nenvironment5. The Government of India\u2019s flagship cooking energy \nprogramme\u2014Pradhan Mantri Ujjwala Yojana (PMUY)\u2014has \nsought to alleviate the public health burden of household air pollu-\ntion by providing a subsidy and loan for the upfront cost of adopting \nan LPG connection, which typically includes a double-burner gas \nstove, a 14.2-kilogram gas cylinder, a gas regulator and a hose pipe. \nAlthough over 80 million households have benefited from PMUY6, \nsteep recurring expenses and the poor availability of LPG in rural \nareas continue to hinder the extent to which households use LPG \nfor all cooking needs7,8.\nFollowing the connection-oriented approach of PMUY, most \nevaluations of the programme have discussed enrolment rates \nusing government-provided data9\u201311. Few studies have examined \nLPG use after adoption, and most of these have focused on average \ncylinder refill rates10,12\u201314. Government records indicate that about \none-quarter of PMUY beneficiaries purchased five or more refills \neach year, suggesting the use of LPG as their primary cooking fuel, \nwhile another one-quarter did not purchase any refills in the first \nyear, suggesting sparing use10,14. Recently, researchers used refill \nrecords in Karnataka to show that the average annual LPG consump-\ntion for PMUY customers was 2.3 cylinders (32.7\u2009kg per year), com-\npared with 4.7 cylinders (66.7\u2009kg per year) for general (non-PMUY) \ncustomers15. According to a government audit, PMUY customers \naveraged 2.98 cylinder refills per year (42.3\u2009kg per year) and general \ncustomers averaged 6.73 cylinder refills per year (95.6\u2009kg per year) \nacross India between March 2018 and March 20191.\nAdoption of a clean cooking fuel like LPG is a necessary first step \nin ensuring smoke-free kitchens and reducing the drudgery of bio-\nmass collection. However, both sustained use of cleaner fuels16,17 and \ncessation of solid fuel use are required18 to achieve the stated health \noutcomes of PMUY. Many households simultaneously use multiple  \nfuels\u2014termed fuel stacking\u2014rather than moving step-by-step \nalong an energy ladder towards cleaner fuels19.\nUsing panel data collected from the rural areas of six major \nenergy access-deprived states in India in 2014\u20132015 and 2018, we \nassess the determinants of upward shifts in LPG use\u2014from use as a \nminority fuel for cooking only, to the primary fuel and then as the \nexclusive cooking fuel. By examining dynamic fuel stacking instead \nof LPG consumption, this study identifies factors that help house-\nholds to transition away from solid fuel use and towards cleaner \nfuels. We report a number of important findings. First, PMUY ben-\neficiaries have considerably lower odds of primary and exclusive use \nof LPG than general customers. Second, the village-level penetra-\ntion of LPG as a primary fuel and the age of the LPG connection are \nboth positively associated with sustained use. Third, we find that \neasy access to free-of-cost biomass is a major deterrent to the exclu-\nsive use of LPG. Finally, households depending on agriculture and \nlabour as the primary source of income have much lower odds of \nincreased use of LPG.\nReview of the literature on household cooking fuel choice\nThe body of research characterizing the determinants of household \ncooking fuel choice has been summarized elsewhere20\u201322. We outline \na few econometric approaches that have explained the determinants \nof cooking fuel use in India. Previous work\u2014while employing rig-\norous analytical approaches\u2014has been limited, primarily owing to \ntwo main gaps.\nFirst, studies seeking to understand the drivers for adoption and \nuse of clean cooking fuels, as well as enablers of the reduction and \ncessation of solid fuel use, are primarily based on cross-sectional \ndata. Such studies are inherently unable to observe household-level \ntransitions. Many studies of cooking fuel choice in India have also \nbeen geographically limited, with relatively limited sample sizes23,24. \nOther studies have employed nationally representative micro-\neconomic survey data that benefit from large sample sizes, such \nas those from the National Sample Survey (NSS)25\u201327, the India \nHuman Development Survey28\u201330 and the National Family Health \nSurvey30\u201333. All three of these surveys have multiple waves, enabling  \nThe drivers of sustained use of liquified petroleum \ngas in India\nSunil Mani\u200a \u200a1\u2009\u2709, Abhishek Jain1, Saurabh Tripathi\u200a \u200a1\u2009\u2709 and Carlos F. Gould\u200a \u200a2\nNinety-five per cent of Indian households now have access to liquified petroleum gas (LPG), with 80 million acquiring it under \nthe Pradhan Mantri Ujjwala Yojana (PMUY) since 2016. Still, having a connection is not enough to eliminate household air pol-\nlution. Studying panel data from rural households in six major states from 2014\u20132015 and 2018, we assess the determinants of \ncooking energy transition from solid fuels to LPG. We find that PMUY beneficiaries have much lower odds of using LPG as the \nprimary or exclusive fuel compared with general customers, irrespective of their economic status. Village-level penetration of \nLPG as a primary fuel and the years of LPG use positively influence its sustained use, while ease of access to freely available \nbiomass and reliance on uncertain and irregular income sources hinder LPG use. The findings highlight the need to interlace \ncooking fuel policies with rural development, to enable a complete transition towards cleaner cooking fuels.\nNature Energy | VOL 5 | June 2020 | 450\u2013457 | www.nature.com/natureenergy\n450\n\nArticles\nNaTUre EnerGY\nresearchers to leverage multiple rounds of data34\u201336; although, only \nthe India Human Development Survey has a panel structure.\nTwo recent studies using multiple NSS rounds are notable. The \nfirst study separately analyses the 50th round (1993\u20131994), the \n61st round, (2004\u20132005) and the 66th round (2009\u20132010) to assess \nthe determinants of multiple cooking fuel use, and, separately, the \nexclusive use of clean cooking fuels37. The authors find a strong \nassociation between increasing household expenditures and clean \ncooking fuel use. However, this study neither models the data in \na panel structure, nor models household-level fuel switching. The \nsecond study creates a synthetic panel of 100,000 households using \nthe 55th NSS round (1999\u20132000) and the 68th round (2011\u20132012), \nby matching households on the basis of state, urban/rural and \ncaste38. The authors model the extent to which LPG has offset fire-\nwood use to estimate the change in carbon dioxide emissions due \nto increasing LPG adoption and use. This study does not primar-\nily study the determinants of fuel switching or fuel stacking. Taken \ntogether, these studies have contributed to our understanding of the \ndifferences between LPG adopters and non-adopters, but have not \nassessed household-level cooking fuel transitions.\nSecond, many previous analyses have focused only on the house-\nhold\u2019s primary cooking fuel as the variable of interest. Current evi-\ndence suggests that the full health benefits of clean cooking fuel \nuse are only obtained when used nearly exclusively16,18. Yet, many \nsurveys have collected data only on households\u2019 primary cooking \nfuel; although this may be changing with recent efforts to incor-\nporate multiple fuel use into surveys, worldwide39. Previous studies \nhave assessed cooking fuel choice using several different outcomes: \nownership of a clean cooking fuel; clean cooking fuel as the primary \ncooking fuel; and, in some cases, ordered fuel choice25,40.\nTwo additional studies employing panel multinomial logit \nmodels using multiple rounds of survey data are noteworthy. The \nfirst uses three rounds of surveys to assess the socio-economic \nfactors that determine cooking fuel choice in urban households \nin Ethiopia41. The authors find that household expenditures, fuel \nprices and education are key determinants of fuel stacking behav-\niour. The second study uses three waves of panel data in Tanzania to \nmodel fuel choice for cooking and lighting42. This study takes sev-\neral approaches to assessing fuel stacking and incorporates impor-\ntant contextual covariates, such as fuel prices and local luminosity \ndata. However, neither study directly models households switching \nfuels, and both include kerosene as a clean cooking fuel, despite  \nevidence of substantial health risk from kerosene combustion due to \nits polluting emissions43. Still, efforts to formally model fuel stacking \nin India are limited28.\nResearch design\nOur analysis is based on two waves of household surveys carried out \nin six historically energy-poor states in India\u2014Bihar, Jharkhand, \nMadhya Pradesh, Odisha, Uttar Pradesh and West Bengal\u2014collec-\ntively representing 40% of the Indian population (Supplementary \nTable 1). The first wave covered 8,563 households and was collected \nbetween November 2014 and February 2015. The second wave cov-\nered 9,072 households\u20147,317 of which were also surveyed in wave 1 \n\u2014and was gathered between March and June 2018. The survey \ndesign is described in full in Supplementary Note 1.\nTo simultaneously study increased LPG use and displacement of \nsolid fuel use, we specify a categorical outcome variable: LPG as a \nminority cooking fuel (secondary to solid fuels), LPG as a primary \ncooking fuel (with solid fuels retained as secondary options) and \nLPG as the exclusive cooking fuel.\nWe use two approaches to model the determinants of increased \nLPG use. In the first approach, we use a generalized-ordered logis-\ntic regression to study the cross-section of LPG-owning households \nin wave 2. We consider 4,102 households that reported LPG use \nin wave 2, but that may or may not have used LPG in wave 1. We \nonly include households that were surveyed in both waves despite \nanalysing the cross-section of wave 2, owing to the inclusion of \ncertain lagged (wave 1) covariates. We identify covariates from the \nliterature, including socio-economic and demographic contextual  \nfactors and household experience with cooking fuels (discussed in \nfull in Methods).\nIn the second approach, we use a panel-ordered logistic regres-\nsion to assess the household-level determinants of increasing the \nshare of LPG in a household\u2019s cooking fuel mix. We consider 1,411 \nhouseholds that reported LPG use during both survey waves (dis-\ntribution by category in Supplementary Table 2). While the focus of \nthis study is understanding the factors for upward movement, we \ninclude movements up, down and across to account for all variation \nin household-level changes between the two survey waves.\nWe also undertake subsample analyses to assess whether the \nexplanatory factors influence households differently depending on \ntheir starting point of LPG use (minority users versus primary users \nin wave 1) (Supplementary Table 3).\nIncreases in LPG use between 2015 and 2018\nThe overall sample (n\u2009=\u200917,635) witnessed a dramatic increase in \nLPG adoption between 2015 and 2018\u2014from 22% LPG users in \nwave 1 to 58% in wave 2. PMUY has been an important driver of \nthis shift: 43% of new LPG owners in wave 2 were enroled through \nPMUY. However, a lower proportion of PMUY beneficiaries were \nexclusive LPG users compared to general customers (Fig. 1).\nIn addition to a greater number of households having LPG in \nwave 2, we also observe an upward progression of LPG-use category. \nAmong the panel subset, exclusive use was the smallest category in \nwave 1, but the largest in wave 2 (Fig. 2). Still, there was some down-\nward movement from primary to minority use, and from exclusive \nuse to primary or minority use.\nDeterminants of LPG use and fuel stacking\nCross-sectional generalized-ordered logistic regressions show that \nhouseholds acquiring their LPG connections through PMUY have \nsignificantly lower odds of being primary users compared to gen-\neral customers (odds ratio (OR), 0.64; 95% confidence interval \n(CI), 0.52\u20130.78) (Table 1). The odds drop further still for exclu-\nsive use (OR, 0.56; 95% CI, 0.45\u20130.71). The lower odds of PMUY \nhouseholds using more LPG is neither a factor of their relatively \npoor socio-economic status, nor of the infancy of their connection, \n37%\n(N = 1,519)\n32%\n(N = 1,306)\n31%\n(N = 1,277)\n0\n25\n50\n75\n100\nAll LPG users\n(N = 4,102)\nPercentage of households\n56%\n(N = 687)\n30%\n(N = 369)\n14%\n(N = 170)\n29%\n(N = 832)\n33%\n(N = 937)\n39%\n(N = 1,107)\nPMUY\nhouseholds\n(N = 1,226)\nGeneral\nhouseholds\n(N = 2,876)\nExclusive use\nPrimary use\nMinority use\nFig. 1 | Distribution of LPG-use categories among LPG households in 2018. \nPercentages refer to the fraction of households in each LPG-use category \nin wave 2 (2018). Sample sizes are included in parentheses. Black lines \nindicate standard errors. Using a two-sample test of proportions (which \nuses the t-statistic), we find that the difference between the proportions of \nexclusive users of LPG among PMUY households and general customers is \nstatistically significant.\nNature Energy | VOL 5 | June 2020 | 450\u2013457 | www.nature.com/natureenergy\n451\n\nArticles\nNaTUre EnerGY\nas both of these factors are controlled for in the model. Rather, we \nsuggest that households that made a decision to spend the upfront \ncost and effort to procure an LPG connection are probably more \nwilling and more financially prepared to use LPG for the majority \nof their cooking needs than those who have been provided with a \nfree connection.\nWe also observe village-level popularity of LPG as a primary fuel \ninfluencing LPG use. The fraction of households in a participant\u2019s \nvillage using LPG as their primary cooking fuel is significantly posi-\ntively associated with that household being a primary or exclusive \nLPG user (Table 1). Our panel model corroborates this finding. \nHouseholds in a village where the percentage of primary LPG users \nincreased by one percentage point had 1.046 (95% CI, 1.041\u20131.052) \ntimes higher odds of moving to primary or exclusive LPG use than \nhouseholds in villages where there was no change in percentage of \nprimary LPG users (Table 2). The positive effect of village-level pen-\netration of LPG as a primary cooking fuel suggests a possible peer \neffect of LPG use and the importance of community norms in influ-\nencing LPG use. Although we account for geographic heterogeneity \nat the state level, higher percentages of primary users of LPG may \nalso be related to village-level variation in the cost and availability of \nbiomass, infrastructure quality or LPG availability.\nAs hypothesized, we find that households relying on agriculture \nor on labour as their primary source of income had lower odds of \nusing LPG as their primary and exclusive fuel compared to house-\nholds getting their primary income from salaried employment \n(Table 1). Agriculture-dependent households may have lower odds \ndue to the unpredictability and irregularity of cash flow, which may \nmake the large, \u2018lumpy\u2018 LPG refill payments a challenge. Indeed, \nhouseholds where the primary income changed to agriculture or \nlabour between wave 1 and wave 2 had lower odds of increasing \ntheir LPG-use category than those whose occupation stayed the \nsame or changed away from agriculture or labour (Table 2).\nYears of owning LPG was positively associated with higher use \ncategory in both models. This effect of time may be because house-\nholds get used to the convenience of LPG-based cooking, have more \ntime to adjust their cooking habits and practices to LPG and/or have \nmore time to align their household expenditures and cash flows \nwith the recurring expense of LPG refills.\nCounterintuitively, we did not find involvement of women \nin household decisions to purchase durable goods to be a reli-\nable predictor of household use of LPG after adoption. Further \nresearch on the involvement of women in cooking energy-related \ndecision-making and its impact on fuel choice is warranted.\nDistance travelled to procure LPG was not significantly associ-\nated with higher odds of using LPG for primary or exclusive use. \nHowever, households that travelled one kilometre less to procure \nLPG in wave 2 than in wave 1 had somewhat higher odds (OR, 1.04; \n95% CI, 1.01\u20131.08) of moving up from minority use than those for \nwhom distance did not change (Supplementary Table 3).\nAs hypothesized, households that owned cattle had considerably \nlower odds of being primary and exclusive users of LPG, probably \nbecause of easy access to dung cakes (Table 1). The magnitude of \nthe association was higher for exclusive use compared to primary \nuse, implying that the push towards complete cessation of solid fuel \nuse is likely to face strong resistance from households that have easy \naccess to biomass. Furthermore, households that acquired cattle \nbetween the two waves had lower odds of shifting to primary and \nexclusive use than other households (Table 2).\nWe find that households that collected firewood daily or a few \ntimes a week (a proxy for easy availability of biomass) in wave 1 had \nlower odds of exclusive use of LPG in wave 2 than those who col-\nlected firewood less frequently (Table 1). While easier availability of \nfree-of-cost firewood impedes exclusive use of LPG, we did not find \nit hindering use of LPG as a primary fuel. Surprisingly, the effect of \nweekly expenditure on traditional biomass for cooking in wave 1 \non the odds of increased use of LPG in wave 2 was not statistically \nsignificant (Table 1). This finding is contrary to our hypothesis and \nwarrants further research.\nRespondents\u2019 perceptions that LPG affected their health less \nthan traditional cook stoves in wave 1 was not associated with LPG \nuse in wave 2 (Supplementary Table 4). However, their percep-\ntions in wave 2 were positively associated with LPG use in wave 2 \n(Supplementary Table 5). It is possible that knowledge of the health \nbenefits of LPG is a consequence of LPG use, as opposed to a driver \nof its use, suggesting that health-related awareness in isolation may \nnot be sufficient to nudge sustained use of LPG.\nLimitations and future work\nBefore concluding, we note some limitations of our research. We \nrely on primary survey data for our analysis, including self-reported \ninformation on quantitative variables, which could be affected by \ntypical limitations and biases of survey-based reporting. However, on \nthe basis of a couple of robustness checks for our dependent variable \n(Supplementary Table 6\u20139), we can still claim that our self-reported \nvariable is reasonable for this analysis. While approaches such as \nstove-use monitors and in-field fuel measurements can potentially \nlead to even more precise estimates, these measurement strategies \nare costly and can be prone to the Hawthorne effect, where partici-\npants behave differently to usual because they are being observed44.\nIn addition, we focused on fuel stacking, but only pertaining to \ncooking. Future studies may investigate the end uses for each stove \nor fuel in the household. At times, solid fuels are also used for other \nheating applications\u2014space heating, water heating or food prepara-\ntion of animals\u2014that are not easily replicated by LPG or other clean \ncooking fuels. Elsewhere, such non-cooking end uses\u2014or even \npreparation of specific dishes\u2014have been motivators for sustained \nsolid fuel use, even after long-term LPG adoption and use45,46. The \nspecific nature of household solid fuel uses can potentially inform \nlocally targeted interventions to facilitate complete transitions to \nclean household energy.\nRecommendations and conclusions\nIndia has witnessed a revolution in clean cooking energy adop-\ntion between 2016 and 2019 owing to the Pradhan Mantri Ujjwala \nYojana programme. With 95% of households now having access to \nLPG, the policy focus in India is shifting from initial adoption to \nthe sustained use of LPG to reduce air pollution exposure. Using \ncross-sectional and panel analyses, we examine and identify factors \nMinority use\nN = 570\nPrimary use\nN = 546\nExclusive use\nN = 295\nMinority use\nN = 371\nPrimary use\nN = 457\nExclusive use\nN = 583\n2015\n2018\nFig. 2 | Household-level shifts in fuel stacking from 2015 to 2018 in the \npanel subset. The nodes on the left of the alluvial diagram refer to LPG-use \ncategory status in wave 1, and those on the right of the chart refer to \ncategory status in wave 2. Sample sizes in each category in each wave are \nindicated below the labels. The links flowing from the left nodes to the right \nnodes indicate upward, lateral and downward movement of households in \neach LPG-use category from wave 1 to wave 2. We discuss specific patterns \nof downward movement in the Methods.\nNature Energy | VOL 5 | June 2020 | 450\u2013457 | www.nature.com/natureenergy\n452\n\nArticles\nNaTUre EnerGY\nbetween and within households that drive sustained use of LPG and \nresult in the complete cessation of solid fuel use.\nOur results show that PMUY beneficiaries have systematically \nlower odds of using LPG for primary and exclusive cooking needs \nthan general customers, even after controlling for socio-economic dif-\nferences. Without targeted support for their energy transition, mean-\ningful reductions in household air pollution among the majority of \n80 million PMUY households will probably remain out of reach. A \ncommon notion worth assessing in future quantitative and qualitative \nstudies is that the decision and effort to procure an LPG connection \nby general customers meant their greater preparedness (behaviourally \nand financially) to use LPG as their primary fuel for cooking.\nGiven that village-level penetration of LPG as the primary fuel \nhas a strong positive association with household-level LPG use, we \nsuggest that policies with community-level targeting could effec-\ntively increase LPG use. Thanks to the age of the LPG connection \nhaving a positive influence on its use, we may witness increasing use \nof LPG among recent adopters over time. However, targeted policies \nto further improve affordability of LPG refills for poor households \ncan accelerate the transition.\nTable 1 | Explaining LPG use in wave 2 (cross-section model)\nOdds of primary use as opposed to minority use\nOdds of exclusive use as opposed to primary use\nLPG-use category: dependent \nvariable\nOR (s.e.m.)\nP\u2009>\u2009z\n95% CI\nOR (s.e.m.)\nP\u2009>\u2009z\n95% CI\nWhether household is a PMUY \nbeneficiary\n0.641 (0.066)\n<0.001***\n0.523\n0.784\n0.564 (0.068)\n<0.001***\n0.446\n0.714\nPercentage of households in a \nvillage using LPG as their primary \ncooking fuel\n1.043 (0.002)\n<0.001***\n1.039\n1.048\n1.027 (0.002)\n<0.001***\n1.023\n1.031\nWeekly expenditure on biomass \n(wave 1) (ln\u2009+\u20091)\n1.012 (0.016)\n0.467\n0.981\n1.044\n0.982 (0.017)\n0.270\n0.950\n1.015\nWhether firewood is collected \nmultiple times a week (wave 1)\n0.895 (0.085)\n0.245\n0.743\n1.079\n0.788 (0.081)\n0.021**\n0.643\n0.965\nWhether women are involved in the \nhousehold decision-making\n1.135 (0.099)\n0.149\n0.956\n1.348\n0.873 (0.082)\n0.146\n0.727\n1.048\nNumber of years that the household \nhas had LPG (ln)\n1.260 (0.088)\n0.001***\n1.099\n1.445\n1.226 (0.089)\n0.005***\n1.064\n1.412\nOne-way distance to procure LPG \ncylinders (km)\n1.010 (0.007)\n0.155\n0.996\n1.023\n1.000 (0.008)\n0.984\n0.985\n1.015\nWhether the household owns cattle\n0.447 (0.039)\n<0.001***\n0.377\n0.530\n0.423 (0.036)\n<0.001***\n0.358\n0.499\nWhether the household owns land\n0.838 (0.084)\n0.079*\n0.687\n1.021\n0.835 (0.086)\n0.080*\n0.683\n1.022\nEconomic status index\n1.147 (0.030)\n<0.001***\n1.089\n1.208\n1.171 (0.030)\n<0.001***\n1.113\n1.232\nHousehold size (ln\u2009+\u20091)\n0.604 (0.067)\n<0.001***\n0.487\n0.750\n0.484 (0.053)\n<0.001***\n0.390\n0.600\nEducation of the household head (by category; base category is 12th standard and above)\n No education\n0.742 (0.086)\n0.010**\n0.592\n0.931\n0.742 (0.090)\n0.014**\n0.585\n0.941\n Up to 5th standard\n0.919 (0.103)\n0.449\n0.737\n1.145\n0.755 (0.086)\n0.013**\n0.602\n0.941\n Between 5th and 10th standard\n1.019 (0.127)\n0.882\n0.797\n1.302\n1.005 (0.121)\n0.959\n0.785\n1.259\nCaste of the household head (by category; base category is OBCs and general caste together)\n Scheduled caste\n0.984 (0.099)\n0.877\n0.808\n1.200\n0.948 (0.104)\n0.568\n0.757\n1.165\n Scheduled tribe\n0.936 (0.165)\n0.707\n0.662\n1.323\n0.964 (0.184)\n0.810\n0.658\n1.388\nPrimary source of income of the household (by category; base category is salaried occupation)\n Agriculture on own land or \nleased land\n0.595 (0.094)\n0.001***\n0.437\n0.810\n0.619 (0.086)\n<0.001***\n0.469\n0.806\n Casual agricultural or daily-wage \nlabour\n0.496 (0.082)\n<0.001***\n0.358\n0.686\n0.610 (0.092)\n0.001***\n0.449\n0.809\n Own business\n0.850 (0.159)\n0.385\n0.588\n1.227\n0.905 (0.144)\n0.500\n0.658\n1.226\n Others\n0.989 (0.296)\n0.971\n0.551\n1.777\n0.920 (0.246)\n0.729\n0.540\n1.538\nNumber of households (n)\n4,102\nState fixed effects\nYes\nlog-likelihood\n\u22123,603.996\nProbability\u2009>\u2009\u03c72\n<0.0010\nPseudo R2\n0.1981\nResults from the cross-section generalized-ordered logistic regression model are shown here. The response variable is the LPG-use category in wave 2. The model explains determinants of households being \nprimary users of LPG (compared to minority users), and of being exclusive users of LPG (compared to primary users) in wave 2. There are 4,102 households, all of which were surveyed in both waves, and \nwere using LPG in wave 2. Standard errors of the mean (s.e.m.) are in parentheses after the OR values. ***P\u2009<\u20090.01, **P\u2009<\u20090.05, *P\u2009<\u20090.10.\nNature Energy | VOL 5 | June 2020 | 450\u2013457 | www.nature.com/natureenergy\n453\n\nArticles\nNaTUre EnerGY\nOwnership of cattle and a high frequency of firewood collection\u2014\nboth indicators of easy access to free-of-cost biomass\u2014considerably \nlower the odds of exclusive LPG use. Creating opportunities for \nhouseholds to sell the biomass for commercial purposes\u2014such as \nalternative transport or industrial fuel\u2014could create an opportunity \ncost for biomass to facilitate greater use of modern cooking fuels.\nReduction in the distance travelled to procure LPG increases a \nminority user household\u2019s odds of moving to primary or exclusive \nuse of the fuel. While the government has taken steps to improve \nfuel accessibility in recent years, further efforts are required to \nincrease the density of distribution outlets in rural areas. To reduce \nthe distance travelled in remote areas, the government and oil mar-\nketing companies could pilot locally tailored business models, such \nas extending LPG distribution to village-level entrepreneurs and \nlocal cooperatives.\nFinally, the underlying economic transitions of India\u2019s rural \neconomy cannot be ignored. Households relying on agriculture and \nlabour as the primary source of income need livelihood support that \nenables predictable and regular cash flow to facilitate sustained use \nof clean cooking fuels. Convergence across government schemes on \nrural livelihoods and employment guarantees with clean cooking \nfuel promotion could become an important driver of the transition \naway from polluting solid fuels.\nOur findings underscore that there is no \u2018silver bullet\u2019 that will \nyield exclusive clean cooking fuel use in rural India. PMUY has \nresulted in a tremendous national transition towards LPG, and we \nneed multipronged approaches to accelerate its sustained use over \ntime. Only by going beyond cooking fuel policies, and interlacing \nthem with overall rural development priorities, can India move \nforward in enabling a complete transition towards clean cooking  \nfuels for all.\nMethods\nResponse variable. We report here on the specifications of the response, \nexplanatory and control variables. Data collection in wave 1 (refs. 47,48) and \nwave 2 (refs. 49,50) of ACCESS have been described previously, and in further \ndepth in Supplementary Note 1. The survey questionnaire used is available in \nSupplementary Note 2.\nWe define three mutually exclusive categories of LPG use to assess factors that \ndetermined progression towards exclusive use of LPG amongst rural households \nin India between 2014 and 2018. These categories are used as the response \nvariable in the panel and cross-sectional analysis: LPG as a minority cooking fuel \n(secondary to solid fuels), LPG as a primary cooking fuel (with solid fuels retained \nas secondary options) and LPG as the exclusive cooking fuel.\nThese categories are based on self-reported responses. Households are \nclassified as exclusive users if they report LPG use but do not use any other fuel \nsource for cooking. Primary users are those that state that LPG is their \u2018primary \ncooking fuel\u2019, but also report using other cooking fuels. Minority users are the \nremaining households, which use LPG but report some other fuel as their primary \ncooking fuel.\nElsewhere, continuous variables have been used to model LPG consumption \nafter adoption. While outcomes such as \u2018refills per year\u2019 or \u2018kilograms of LPG \nconsumed per year\u2019 capture overall LPG use, they are limited because they do \nnot directly account for solid fuel use. Intuitively, increases in LPG use have been \nshown to yield decreases in solid fuel use; however, the extent to which there is \nperfect displacement is unknown. Furthermore, continued solid fuel use may not \nbe a function of needing to meet household energy demands exclusively, but is \nprobably also about preferences and meeting household end uses, such as cooking \nspecific dishes or non-cooking energy demands such as space or water heating. \nTherefore, we use the three categories outlined above to capture dynamic shifts in \ncooking fuel stacking patterns. As a first-degree check, we assessed our dependent \nvariable among 1,411 panel households by observing the changes in the amount \nof firewood (kg per person per month), dung cakes (pieces per person per month) \nand LPG (kg per person per year) used for cooking within households by each LPG \ncategory shift. This fuel displacement analysis is reported in Supplementary  \nTable 6, and clearly shows that all the upward (downward) movements in LPG \ncategory among panel households also come up with reduction (increase) in \ntraditional biomass consumption.\nTo further ensure robustness of our choice of dependent variable, we assessed \nthe deviation between self-reported primary fuel and the primary fuel based on \nuseful energy analysis (considering quantity of fuel and stove efficiencies). We \nfound the self-reported variable to be reasonable for our analysis (Supplementary \nTables 7\u20139).\nWe noticed some downward movement across the categories between the two \nwaves. There were few systematic differences between households that regressed \nand those that made lateral or upward movement. Among those that moved \ndown from exclusive use, the proportion that owned cattle increased from 45% to \n58%, while cattle ownership declined for all other households in the panel subset. \nFurthermore, those that moved down from exclusive and primary use had a \nTable 2 | Explaining household-level change in LPG use between wave 1 and wave 2\nDependent variable: LPG-use category\nOR (s.e.m.)\nP\u2009>\u2009z\n95% CI\nPercentage of households in a village using LPG as their primary cooking fuel 1.046 (0.003)\n<0.001***\n1.041\n1.052\nWomen are involved in the household decision-making\n0.784 (0.086)\n0.027**\n0.633\n0.972\nNumber of years that the household has had LPG (ln\u2009+\u20091)\n1.161 (0.066)\n0.008***\n1.040\n1.297\nOne-way distance to procure LPG cylinders (km)\n1.000 (0.008)\n0.970\n0.985\n1.015\nHousehold owns cattle\n0.275 (0.031)\n<0.001***\n0.221\n0.344\nHousehold owns land\n1.070 (0.148)\n0.623\n0.816\n1.403\nEconomic status index\n1.204 (0.032)\n<0.001***\n1.143\n1.269\nHousehold size (ln\u2009+\u20091)\n0.510 (0.065)\n<0.001***\n0.398\n0.655\nPrimary source of income of the household\n Agriculture on own land or leased land\n0.662 (0.093)\n0.003***\n0.502\n0.872\n Casual agricultural or daily-wage labour\n0.643 (0.100)\n0.005***\n0.474\n0.872\n Salaried occupation\n1.192 (0.190)\n0.271\n0.872\n1.629\nNumber of observations\n2,821\nNumber of households\n1,411\nState fixed effects\nYes\nlog-likelihood\n\u22122,515.184\nWald \u03c72\n577.569\nProbability\u2009>\u2009\u03c72\n<0.0010\nThe table shows results from the panel-ordered logistic regression model. The response variable is the LPG-use category in both waves of survey. The model explains household-level determinants of the \nmovement to higher LPG-use categories from wave 1 to wave 2. It includes 1,411 households, all which used LPG in both waves. The s.e.m. are in parentheses after the OR values. ***P\u2009<\u20090.01, **P\u2009<\u20090.05, *P\u2009<\u20090.10.\nNature Energy | VOL 5 | June 2020 | 450\u2013457 | www.nature.com/natureenergy\n454\n\nArticles\nNaTUre EnerGY\nsharper drop in their asset index score than those that did not regress. Last, among \nthose that regressed, the proportion relying on agriculture for primary income \nincreased, whereas that statistic decreased for the households that remained in the \nsame category or moved up a category between the two waves.\nExplanatory variables. Here, we outline the specification of explanatory variables \nincluded in our models, as well as discuss their inclusion in previous studies. \nThe variables were chosen on the basis of the existing literature on cooking \nenergy adoption and use. Supplementary Table 10 (cross-section subset) and \nSupplementary Table 11 (panel subset) contain descriptive summaries of all \nexplanatory variables, along with the hypothesized direction of association between \nthe covariate and LPG-use category.\nWe include an indicator variable for whether a household is a PMUY \nbeneficiary. Elsewhere, PMUY beneficiaries have been shown to utilize LPG after \nadoption differently from general consumers; often purchasing fewer refills relative \nto their peers15. Some PMUY beneficiaries may be financially unable to support \nregular cooking with LPG, and without the programme may not have purchased an \nLPG connection. In addition, it is possible that PMUY beneficiaries have different \nperceptions, attitudes or abilities compared to general customers, although this \nremains unexplored. We found that PMUY households were generally from the \nsame rural villages as general customers (Supplementary Fig. 1), indicating no \ngeographical bias in sampling. In the cross-section subset, 45% of minority users, \n26% of primary users and 13% of exclusive users of LPG were PMUY households. \nIn this subset, we found 134 households that identified as PMUY beneficiaries \nin wave 2, while also having reported LPG use in wave 1. Since PMUY only \nstarted after wave 1 of the survey, we force-coded these households as general \n(non-PMUY) customers to maintain internal consistency. We note the specific \ntiming of data collection in both wave 1 and wave 2 in relation to when PMUY \nenroled the beneficiaries in each state in Supplementary Table 12.\nWe include the percentage of households in a village using LPG as their \nprimary cooking fuel as a potential measure of local popularity of LPG, which, as \na community norm, could influence household use of LPG. In addition to the peer \neffect, we also expect this covariate to capture unobserved village-level variations, \nsuch as local road infrastructure and availability of LPG. Empirical evidence \nregarding the associations between LPG penetration in a community or community \nnorms is otherwise limited, but there is some positive evidence for increased \nprobability of LPG ownership with higher LPG community penetration51. There are \nongoing trials testing the effects of peer effect and community norms52,53.\nWe include the natural log of total weekly expenditure on biomass procurement \nin wave 1. We expect that households that spent a higher amount on biomass in \nwave 1, might have higher odds of increased use of LPG in wave 2 because they \nwould find LPG to be relatively cheaper54. We find that households that are primary \nor exclusive users of LPG in wave 2 had a higher mean spending on biomass in \nwave 1 than minority users of LPG. Of the households in the cross-section subset, \n57% did not spend any money on procuring biomass in wave 1, which made the \nnatural log calculations impossible. Therefore, before converting the values into \ntheir logarithmic form they were increased by one. In the cross-section analysis, we \nuse the lagged (wave 1) value of this covariate instead of the wave 2 value because \nwe expect it to be endogenous with LPG-use category. For the same reason, we do \nnot consider this covariate for the panel analysis.\nTo capture historical availability of biomass, we also include whether firewood \nwas collected multiple times a week in wave 1 as a binary variable\u2014where \u20181\u2019 is \nassigned to households that collected firewood daily or a few times in a week, \nand \u20180\u2019 is assigned for any other frequency. High frequency may indicate ease of \nfirewood collection, and the resistance a household is likely to face in transitioning \nto the exclusive use of LPG. As above, we use the lagged (wave 1) value of this \ncovariate instead of the wave 2 value in the cross-section analysis because we \nexpect it to be endogenous with LPG-use category. Similarly, we do not consider \nthis covariate for the panel analysis.\nThe involvement of women in household decision-making may be an important \ndeterminant of LPG use. We asked respondents who made the decision to purchase \ndurable goods; responses were categorized as (1) a woman was involved in \nhousehold decision-making (woman alone or joint with male household head), or \n(2) a woman is not involved in household decision-making. In the panel subset, \nthere is a net increase in the proportion of households where women were involved \nin household decision-making between the two waves, from 20 to 28%. Women are \noften the primary cooks and, as such, may understand the benefits of transitioning \nto modern fuels, such as LPG, better. If women are involved in household \ndecision-making, then their opinions can influence cooking fuel choice. A previous \nstudy showed that households where women were involved in decision-making \nhad higher odds of owning LPG among the first wave of ACCESS55. Elsewhere, it \nhas been relatively uncommon to directly model women\u2019s decision-making power \nin studies of cooking fuel choice20,56. However, in some cases households headed by \na woman have been more likely to have a clean cooking fuel25,57, and in others less \nlikely, compared with households headed by a man58,59.\nWe include the natural log of the number of years that a household has had \nLPG to capture shifts in attitudes or in abilities for using cooking gas efficiently \ndue to increased familiarity with the fuel over time. Increased ability is often cited \nas an important aspect of behavioural change52,60. This covariate required changes \nsimilar to those in the biomass expenditure covariate. While studying data for \nall households that were surveyed in both the waves, we came across internal \ninconsistencies in this covariate. First, for households that had an LPG connection \nin both waves, the reported age in wave 1 and wave 2 should have differed by no \nmore than 3.5 years\u2014the time between the two waves of surveys. However, the \ndifference was off for most of the households. Therefore, we assumed the age \nreported in wave 1 to be true and added 3.5 to that to arrive at the estimated age \nin wave 2. Second, for households that did not use LPG in wave 1 but did so in \nwave 2, the age of connection\u2014as reported in wave 2\u2014could not be higher than \n3.5 years. However, for 22% of the households that reported a higher value, age of \nconnection was capped at 3.5.\nWe asked participants whether they considered LPG to be better than \ntraditional cook stoves from the viewpoint of impact on health, coding responses \ninto a binary variable where \u20181\u2019 is assigned to households that found LPG better, \nand \u20180\u2019 to those that thought LPG was not better. Positive perceptions of LPG and \nan understanding of the negative health impacts of traditional solid fuel cooking \npractices may indicate a household valuing the benefits of LPG adoption and use. \nElsewhere, such positive perceptions have been observed among LPG owners. We \nuse this covariate twice in the cross-section subset\u2014first, the response in wave 1 \nto assess the causal impact on LPG category in wave 2, and second, the response in \nwave 2 to confirm if perception is endogenous with use. This covariate is only used \nin the alternate cross-section specifications (Supplementary Tables 4 and 5).\nTo account for the difficulty associated with accessing LPG cylinders in rural \nareas, we include the self-reported one-way distance covered by households \nto procure LPG cylinders (in kilometres). Fuel accessibility is a well-defined \nconstraint to LPG adoption and use8. There is increasing empirical evidence for \ndirect associations between fuel accessibility and use51,61,62 and interventions testing \npotential solutions52. In the cross-section subset, we find that 43% of households \nreceived doorstep delivery of LPG cylinders, and while 22% of households \ntravelled up to 3\u2009km (one-way) to procure LPG, 24% travelled more than 5\u2009km. \nWe also find that the proportion of households receiving doorstep delivery has \nincreased unevenly across categories of LPG use between the two waves\u2014from \n12% to 29% for minority users, but from 24% to 57% for exclusive users\u2014even \nas the distance travelled to procure LPG for those who do not receive doorstep \ndelivery has declined evenly across all three categories. We acknowledge the \npotential for measurement error, since not all households may accurately report \nthe distance in kilometres. Although we cannot quantify the extent of this error in \nthe cross-section subset, we expect a degree of self-correction in the panel subset, \nwhere respondents might have committed a similar random error in both waves, \nbalancing out the error in the observations.\nTo capture easier availability of free-of-cost biomass, we include two binary \nvariables\u2014ownership of cattle and ownership of land. In the cross-section, 68% \nown land and 57% own cattle. Literature suggests that these may be significant \npredictors of continued biomass use24. We account for access to land because many \nhouseholds cooking with firewood or crop residue obtain it from their own land63.\nControl variables. We use a set of variables to control for socio-economic factors \nthat are likely to impact regular use of LPG by rural households, and include \nstate-level fixed effects to account for variation in unspecified state-level factors \nthat could affect household consumption of LPG48,64. Supplementary Table 10 \n(cross-section subset) and Supplementary Table 11 (panel subset) also contain \ndescriptive summaries of all control variables, along with the hypothesized \ndirection of association between the covariate and LPG-use category.\nWe utilize an economic status index as a measure of a household\u2019s relative \nwealth and economic status, based on the Filmer and Pritchett65 approach. Such \nindices are commonly used across studies in regions where fixed incomes are \nuncommon65\u201367. Overall measures of wealth and income, including asset indices, \nhave been predominately positively associated with ownership and use of clean \ncooking fuels57,68,69. The list of variables included in the economic status index \ncan be found in Supplementary Note 3 and summarized in both survey waves in \nSupplementary Table 13.\nThe level of education of the household head is used as a measure of general \nawareness in the household and the ability to make informed decisions on clean \ncooking energy. Higher levels of education of the household head have often been \npositively associated with adoption of clean cooking energy27,70,71. We use four \ncategories of education: (1) no education, (2) up to 5th standard, (3) between \n5th and 10th standard and (4) 12th standard and above (the base category). A \nsignificantly lower proportion of households where the household head has not \nreceived a formal education are exclusive users of LPG than those where the \nhousehold head has studied beyond the 5th standard.\nThe caste of the respondent is used as a categorical control variable to account \nfor the impact of longstanding social hierarchies on access to LPG connections \nand refills. Caste is a social stratification system that has been negatively associated \nwith clean cooking energy adoption66. The sample has households belonging to \nfour caste categories: (a) scheduled caste (SC), (b) scheduled tribe (ST), (c) other \nbackward classes (OBC) and (d) general. We assess the likelihood of sustained use \nof LPG for SCs and STs, against the combined base category of OBCs and general, \nbecause SCs and STs are historically considered to be the more systematically \ndisadvantaged communities.\nNature Energy | VOL 5 | June 2020 | 450\u2013457 | www.nature.com/natureenergy\n455\n\nArticles\nNaTUre EnerGY\nThe primary source of income of the household is used as a categorical control \nvariable for the varying impact of different cash-flow structures on a household\u2019s \nability to pay for the recurring cost of cooking gas. We use five categories:  \n(1) agriculture on own or leased land, (2) labour (agricultural or daily wage),  \n(3) salaried employment, (4) own business and (5) others, with salaried \nemployment as the base category. \u2018Other\u2019 sources of primary income mostly \nincluded cattle rearing, employment in religious institutions, driving and pension. \nA larger proportion of households with salaried employment and business as \nprimary sources of income are using LPG as an exclusive fuel, while a larger \nproportion of those reliant on agriculture and labour are using LPG as a minority \nfuel. Dependence on agricultural or daily-wage labour has been negatively \nassociated with clean cooking energy access, which is probably attributable to lower \nsocio-economic status or inconsistent cash flow21,72. In other cases, the primary \nsource of income has not been statistically significantly associated with fuel choice \nor is absent from the analyses20. In the panel model, we change this covariate by \nincluding binary variables for the three most theoretically relevant primary sources \nof income from an affordability (cash flow) perspective on LPG uptake: agriculture \non own or leased land, labour (agricultural or daily wage) and salaried employment.\nThe natural log of household size is included to account for its potential \nimpact on LPG use. In previous studies, household size has been both positively23 \nand negatively40,68 associated with clean cooking energy use. In some cases, larger \nhouseholds may seek faster or more cooking options (a positive association with \ncleaner cooking), while in others it is possible that larger households require the \nsubstantially greater capacity of traditional stoves to handle large quantities of food, \nand are less likely to see the benefit of a possibly limited cleaner cook stove, or have \nmore available labour for fuelwood collection (negative association). There are \nother studies that have found no association between household size and cooking \nfuel choice73. This covariate required changes similar to those in the biomass \nexpenditure covariate.\nTests of model assumptions, multicollinearity and robustness. The \ngeneralized-ordered logistic regression enables us to relax the parallel lines \nassumptions of ordered logistic regressions because we do not see a theoretical \njustification to impose the parallel lines restriction. However, we also use an  \nautofit script in Stata that uses a Wald test to determine which covariates  \nhave different beta coefficients by outcome category. Then, the autofit script  \nrelaxes the parallel lines assumption for these covariates only. Supplementary  \nTable 14 shares the results from the Wald tests for parallel lines applied to \neach covariate. Supplementary Table 15 shows results from the regression no \nautofit script and parallel lines assumption imposed (Supplementary Table 15). \nWhile general directions and magnitudes of the associations are similar to our \nmain findings presented in Table 1, we note that our main results with relaxed \nassumptions show meaningful variation in the coefficients by outcome for some \nvariables, such as household size, education of the household head and primary \nsource of income.\nWe also tested for the assumption of multicollinearity after finalizing variables \nby estimating the variance inflationary factor for each covariate. We found \nacceptably low levels of multicollinearity in our model (Supplementary Tables 16 \nand 17). To account for potential residual spatial autocorrelation at the village level \n(beyond what we have captured using state-level fixed effects in the cross-sectional \nmodel and other village-level covariates), we carried out an additional \ngeneralized-ordered logistic regression clustering standard errors at the village level \n(Supplementary Table 18). Coefficients are similar to those presented in Table 1.\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nThe data that support the findings of this study (both of panel and cross-sectional \nanalysis) are made available through Figshare at https://doi.org/10.6084/\nm9.figshare.9810170.\nCode availability\nThe Stata.do files that format, clean and analyse the merged and appended datasets \nare available through Figshare at https://doi.org/10.6084/m9.figshare.9810167, \nwhile the R scripts that produce the figures are available at https://doi.org/10.6084/\nm9.figshare.11838963. Information on unique identifiers between the datasets is \navailable in the Stata.do file.\nReceived: 16 August 2019; Accepted: 28 February 2020;  \nPublished online: 27 April 2020\nReferences\n\t1.\t Report of the Comptroller and Auditor General of India on Pradhan Mantri \nUjjwala Yojana (Comptroller and Auditor General of India, 2019); https://cag.\ngov.in/content/report-no14-2019-performance-audit-pradhan-mantri-ujjwala-\nyojana-ministry-petroleum-and\n\t2.\t Census of India 2011 (Office of the Registrar General & Census \nCommissioner, India & Ministry of Home Affairs, Government of India, \n2013); http://censusindia.gov.in/Tables_Published/H-Series/houselist_main.\nhtml\n\t3.\t Balakrishnan, K. et\u00a0al. The impact of air pollution on deaths, disease burden, \nand life expectancy across the states of India: the Global Burden of Disease \nStudy 2017. Lancet Planet. Health 3, e26\u2013e39 (2019).\n\t4.\t Jeuland, M., Pattanayak, S. K. & Bluffstone, R. The economics of household \nair pollution. Annu. Rev. Resour. Econ. 7, 81\u2013108 (2015).\n\t5.\t Bailis, R., Wang, Y., Drigo, R., Ghilardi, A. & Masera, O. Getting the numbers \nright: revisiting woodfuel sustainability in the developing world. Environ. Res. \nLett. 12, 115002 (2017).\n\t6.\t Bhaskar, U. NDA Ujjwala surpasses targets, provides 80.33 million LPG \nconnections. Mint (9 September 2019).\n\t7.\t Gould, C. F. & Urpelainen, J. LPG as a clean cooking fuel: adoption, use, and \nimpact in rural India. Energy Policy 122, 395\u2013408 (2018).\n\t8.\t Kumar, P., Kaushalendra Rao, R. & Reddy, N. H. Sustained uptake of LPG as \ncleaner cooking fuel in rural India: role of affordability, accessibility, and \nawareness. World Dev. Perspect. 4, 33\u201337 (2016).\n\t9.\t Dabage, A., Sreenivas, A. & Josey, A. What has Pradhan Mantra Ujjwala \nYojana achieved so far? Econ. Polit. Wkly 53, 69\u201375 (2018).\n\t10.\tKar, A., Singh, D., Pachauri, S., Bailis, R. & Zerriffi, H. in The Ujjwala Saga \n\u2013 Unending Happiness & Health 16\u201321 (Ministry of Petroleum and Natural \nGas, Government of India, 2019).\n\t11.\tMinistry of Petroleum and Natural Gas. 7.23 Crore Connections released \nunder Pradhan Mantri Ujjwala Yojana. Press Information Bureau Government \nof India  https://pib.gov.in/newsite/PrintRelease.aspx?relid=190670 (2019).\n\t12.\tAhmad, N., Sharma, S. & Singh, D. A. K. Pradhan Mantri Ujjwala Yojana \n(PMUY) step towards social inclusion in India. Int. J. Trend Res. Dev. 5, \n46\u201349 (2018).\n\t13.\tGiri, A. & Aadil, A. Pradhan Mantri Ujjwala Yojana: A Demand-side \nDiagnostic Study of LPG Refills Policy Brief (MicroSave, 2018).\n\t14.\tSmith, K. R. in Making of New India: Transformation Under Modi \nGovernment (eds Debroy, B. et\u00a0al.) 401\u2013410 (Dr.\u00a0Syama Prasad Mookerjee \nResearch Foundation and Wisdom Tree, 2018).\n\t15.\tKar, A., Pachauri, S., Bailis, R. & Zerriffi, H. Using sales data to assess \ncooking gas adoption and the impact of India\u2019s Ujjwala programme in rural \nKarnataka. Nat. Energy 4, 806\u2013814 (2019).\n\t16.\tPope, D., Bruce, N., Dherani, M., Jagoe, K. & Rehfuess, E. Real-life \neffectiveness of \u2018improved\u2019 stoves and clean fuels in reducing PM 2.5 and CO: \nsystematic review and meta-analysis. Environ. Int. 101, 7\u201318 (2017).\n\t17.\tSambandam, S. et\u00a0al. Can currently available advanced combustion biomass \ncook-stoves provide health relevant exposure reductions? Results from initial \nassessment of select commercial models in India. EcoHealth 12, 25\u201341 (2015).\n\t18.\tJohnson, M. A. & Chiang, R. A. Quantitative guidance for stove usage and \nperformance to achieve health and environmental targets. Environ. Health \nPerspect. 123, 820\u2013826 (2015).\n\t19.\tRuiz-Mercado, I. & Masera, O. Patterns of stove use in the context of \nfuel-device stacking: rationale and implications. EcoHealth 12,  \n42\u201356 (2015).\n\t20.\tMuller, C. & Yan, H. Household fuel use in developing countries: Review of \ntheory and evidence. Energy Econ. 70, 429\u2013439 (2018).\n\t21.\tLewis, J. J. & Pattanayak, S. K. Who adopts improved fuels and cookstoves? A \nsystematic review. Environ. Health Perspect. 120, 637\u2013645 (2012).\n\t22.\tPuzzolo, E., Pope, D., Stanistreet, D., Rehfuess, E. A. & Bruce, N. G. Clean \nfuels for resource-poor settings: A systematic review of barriers and enablers \nto adoption and sustained use. Environ. Res. 146, 218\u2013234 (2016).\n\t23.\tGupta, G. & K\u00f6hlin, G. Preferences for domestic fuel: Analysis with \nsocio-economic factors and rankings in Kolkata, India. Ecol. Econ. 57, \n107\u2013121 (2006).\n\t24.\tJoon, V., Chandra, A. & Bhattacharya, M. Household energy consumption \npattern and socio-cultural dimensions associated with it: A case study of \nrural Haryana, India. Biomass Bioenergy 33, 1509\u20131512 (2009).\n\t25.\tFarsi, M., Filippini, M. & Pachauri, S. Fuel choices in urban Indian \nhouseholds. Environ. Dev. Econ. 12, 757\u2013774 (2007).\n\t26.\tKemmler, A. Factors influencing household access to electricity in India. \nEnergy Sustain. Dev. 11, 13\u201320 (2007).\n\t27.\tPandey, V. L. & Chaubal, A. Comprehending household cooking energy \nchoice in rural India. Biomass Bioenergy 35, 4724\u20134731 (2011).\n\t28.\tAhmad, S. & Puppim de Oliveira, J. A. Fuel switching in slum and non-slum \nhouseholds in urban India. J. Clean. Prod. 94, 130\u2013136 (2015).\n\t29.\tKhandker, S. R., Barnes, D. F. & Samad, H. A. Energy Poverty in Rural and \nUrban India: Are the Energy Poor also Income Poor? (The World Bank, 2010); \nh\u00adt\u00adt\u00adp\u00ad:\u00ad/\u00ad/\u00addo\u00adc\u00adu\u00adm\u00ade\u00adn\u00adt\u00ads.\u00adw\u00ado\u00adrl\u00add\u00adb\u00adan\u00adk\u00ad.\u00ador\u00adg\u00ad/\u00adcu\u00adr\u00ada\u00adte\u00add\u00ad/\u00aden\u00ad/\u00ad9\u00ad20\u00ad9\u00ad0\u00ad14\u00ad6\u00ad8\u00ad25\u00ad9\u00ad7\u00ad74\u00ad5\u00ad5\u00ad7/\u00adE\u00adn\u00ader\u00adg\u00ad\ny\u00ad-p\u00ado\u00adv\u00ader\u00adt\u00ady\u00ad-i\u00adn\u00ad-\u00adru\u00adr\u00ada\u00adl-\u00adan\u00add-\u00adur\u00adba\u00adn-India-are-the-energy-poor-also-income-poor \n(2010).\n\t30.\tKishore, A. & Spears, D. Having a Son Promotes Clean Cooking Fuel Use in \nUrban India: Women\u2019s Status and Son Preference. Econ. Dev. Cult. Change 62, \n673\u2013699 (2014).\nNature Energy | VOL 5 | June 2020 | 450\u2013457 | www.nature.com/natureenergy\n456\n\nArticles\nNaTUre EnerGY\n\t31.\tEpstein, M. B. et\u00a0al. Household fuels, low birth weight, and neonatal death in \nIndia: The separate impacts of biomass, kerosene, and coal. Int. J. Hyg. \nEnviron. Health 216, 523\u2013532 (2013).\n\t32.\tMishra, V. K. & Retherford, R. D. Cooking smoke increases the risk  \nof acute respiratory infection in children. Natl. Fam. Health Surv. Bull. 8,  \n1\u20134 (1997).\n\t33.\tMishra, V. K., Retherford, R. D. & Smith, K. R. Biomass cooking fuels and \nprevalence of tuberculosis in India. Int. J. Infect. Dis. 3, 119\u2013129 (1999).\n\t34.\tCheng, C. & Urpelainen, J. Fuel stacking in India: Changes in the cooking \nand lighting mix, 1987\u20132010. Energy 76, 306\u2013317 (2014).\n\t35.\tGangopadhyay, S., Ramaswami, B. & Wadhwa, W. Reducing subsidies on \nhousehold fuels in India: how will it affect the poor? Energy Policy 33, \n2326\u20132336 (2005).\n\t36.\tViswanathan, B. & Kavi Kumar, K. S. Cooking fuel use patterns in India: \n1983\u20132000. Energy Policy 33, 1021\u20131036 (2005).\n\t37.\tSwarup, V. A. & Rao, K. R. An Econometric Approach to Analysis of Trends \nand Patterns of Household Fuel Choices in India. Indian Econ. Rev. 50, \n105\u2013129 (2015).\n\t38.\tSingh, D., Pachauri, S. & Zerriffi, H. Environmental payoffs of LPG cooking \nin India. Environ. Res. Lett. 12, 115003 (2017).\n\t39.\tHarmonized household energy survey questions. World Health Organization \nhttp://www.who.int/airpollution/household/survey-harmonization/en/ (2018).\n\t40.\tRao, M. N. & Reddy, B. S. Variations in energy use by Indian households: An \nanalysis of micro level data. Energy 32, 143\u2013153 (2007).\n\t41.\tAlem, Y., Beyene, A. D., K\u00f6hlin, G. & Mekonnen, A. Modeling household \ncooking fuel choice: A panel multinomial logit approach. Energy Econ. 59, \n129\u2013137 (2016).\n\t42.\tChoumert-Nkolo, J., Combes Motel, P. & Le Roux, L. Stacking up the ladder: \nA panel data analysis of Tanzanian household energy choices. World Dev. \n115, 222\u2013235 (2019).\n\t43.\tLam, N. L., Smith, K. R., Gauthier, A. & Bates, M. N. Kerosene: a review of \nhousehold uses and their hazards in low- and middle-income countries.  \nJ. Toxicol. Environ. Health B 15, 396\u2013432 (2012).\n\t44.\tSimons, A. M., Beltramo, T., Blalock, G. & Levine, D. I. Using unobtrusive \nsensors to measure and minimize Hawthorne effects: evidence from \ncookstoves. J. Environ. Econ. Manag. 86, 68\u201380 (2017).\n\t45.\tDickinson, K. L. et\u00a0al. Adoption of improved biomass stoves and stove/fuel \nstacking in the REACCTING intervention study in Northern Ghana.  \nEnergy Policy 130, 361\u2013374 (2019).\n\t46.\tGould, C. F. et\u00a0al. Household fuel mixes in peri-urban and rural Ecuador: \nExplaining the context of LPG, patterns of continued firewood use, and the \nchallenges of induction cooking. Energy Policy 136, 111053 (2020).\n\t47.\tAklin, M., Cheng, C., Ganesan, K., Jain, A. & Urpelainen, J. Access to Clean \nCooking Energy and Electricity: Survey of States in India (ACCESS) (Harvard \nDataverse, 2016); https://doi.org/10.7910/DVN/0NV9LF.\n\t48.\tAklin, M., Cheng, C., Urpelainen, J., Ganesan, K. & Jain, A. Factors affecting \nhousehold satisfaction with electricity supply in rural India. Nat. Energy 1, \n16170 (2016).\n\t49.\tJain, A. et\u00a0al. Access to Clean Cooking Energy and Electricity: Survey of States \n(CEEW, 2015); https://www.ceew.in/publications/access-clean-cooking \n-energy-and-electricity (2018)\n\t50.\tMani, S. et\u00a0al. Access to Clean Cooking Energy and Electricity: Survey  \nof States in India 2018 (ACCESS 2018) (Harvard Dataverse, 2019);  \nhttps://doi.org/10.7910/DVN/AHFINM\n\t51.\tDalaba, M. et\u00a0al. Liquified petroleum gas (LPG) supply and demand for \ncooking in Northern Ghana. EcoHealth 15, 716\u2013728 (2018).\n\t52.\tCarri\u00f3n, D. et\u00a0al. Enhancing LPG adoption in Ghana (ELAG): a factorial \ncluster-randomized controlled trial to Enhance LPG Adoption & Sustained \nuse. BMC Public Health 18, 689 (2018).\n\t53.\tDickinson, K. L. et\u00a0al. Prices, peers, and perceptions (P3): study protocol for \nimproved biomass cookstove project in northern Ghana. BMC Public Health \n18, 1209 (2018).\n\t54.\tSehjpal, R., Ramji, A., Soni, A. & Kumar, A. Going beyond incomes: \nDimensions of cooking energy transitions in rural India. Energy 68,  \n470\u2013477 (2014).\n\t55.\tGould, C. F. & Urpelainen, J. The gendered nature of liquefied petroleum gas \nstove adoption and use in rural India. J. Dev. Stud. 0, 1\u201321 (2019).\n\t56.\tPachauri, S. & Rao, N. D. Gender impacts and determinants of energy \npoverty: are we asking the right questions? Curr. Opin. Environ. Sustain. 5, \n205\u2013215 (2013).\n\t57.\tRahut, D. B., Behera, B. & Ali, A. Patterns and determinants of household use \nof fuels for cooking: empirical evidence from sub-Saharan Africa. Energy 117, \n93\u2013104 (2016).\n\t58.\tAbebaw, D. Household determinants of fuelwood choice in urban Ethiopia: a \ncase study of Jimma Town. J. Dev. Areas 41, 117\u2013126 (2007).\n\t59.\tOuedraogo, B. Household energy preferences for cooking in urban \nOuagadougou, Burkina Faso. Energy Policy 34, 3787\u20133795 (2006).\n\t60.\tKar, A. & Zerriffi, H. From cookstove acquisition to cooking transition: \nframing the behavioural aspects of cookstove interventions. Energy Res. Soc. \nSci. 42, 22\u201333 (2018).\n\t61.\tPattanayak, S. K. et\u00a0al. Experimental evidence on promotion of electric  \nand improved biomass cookstoves. Proc. Natl Acad. Sci. USA 116, \n13282\u201313287 (2019).\n\t62.\tPillarisetti, A. et\u00a0al. Promoting LPG usage during pregnancy: a pilot study in \nrural Maharashtra, India. Environ. Int. 127, 540\u2013549 (2019).\n\t63.\tBehera, B., Rahut, D. B., Jeetendra, A. & Ali, A. Household collection and use \nof biomass energy sources in South Asia. Energy 85, 468\u2013480 (2015).\n\t64.\tBaqui\u00e9, S. & Urpelainen, J. Access to modern fuels and satisfaction with \ncooking arrangements: Survey evidence from rural India. Energy Sustain. Dev. \n38, 34\u201347 (2017).\n\t65.\tFilmer, D. & Pritchett, L. Estimating wealth effects without expenditure \ndata\u2014Or tears: An application to educational enrollments in states of India. \nDemography 38, 115\u2013132 (2001).\n\t66.\tMenghwani, V. et\u00a0al. Determinants of cookstoves and fuel choice among rural \nhouseholds in India. EcoHealth 16, 21\u201360 (2019).\n\t67.\tVyas, S. & Kumaranayake, L. Constructing socio-economic status indices: \nhow to use principal components analysis. Health Policy Plan. 21,  \n459\u2013468 (2006).\n\t68.\tPaudel, U., Khatri, U. & Pant, K. P. Understanding the determinants of \nhousehold cooking fuel choice in Afghanistan: A multinomial logit \nestimation. Energy 156, 55\u201362 (2018).\n\t69.\tRavindra, K., Kaur-Sidhu, M., Mor, S. & John, S. Trend in household  \nenergy consumption pattern in India: A case study on the influence of \nsocio-cultural factors for the choice of clean fuel use. J. Clean. Prod. 213, \n1024\u20131034 (2019).\n\t70.\tvan\u00a0der Kroon, B., Brouwer, R. & van Beukering, P. J. H. The energy ladder: \nTheoretical myth or empirical truth? Results from a meta-analysis.  \nRenew. Sustain. Energy Rev. 20, 504\u2013513 (2013).\n\t71.\tPeng, W., Zerriffi, H. & Pan, J. Household level fuel switching in rural Hubei. \nEnergy Sustain. Dev. 14, 238\u2013244 (2010).\n\t72.\tShi, X., Heerink, N. & Qu, F. The role of off-farm employment in the rural \nenergy consumption transition \u2014 A village-level analysis in Jiangxi Province, \nChina. China Econ. Rev. 20, 350\u2013359 (2009).\n\t73.\tHeltberg, R. Fuel switching: evidence from eight developing countries.  \nEnergy Econ. 26, 869\u2013887 (2004).\nAcknowledgements\nThe Council on Energy, Environment and Water supported time spent by S.M., A.J. and \nS.T. on this research. The data collection was supported by the Shakti Sustainable Energy \nFoundation and the National University of Singapore. C.F.G. is supported by the United \nStates National Institute of Environmental Health Sciences grant no. T32 ES007322.\nAuthor contributions\nA.J. conceptualized the study and led the design of the work. A.J., S.T. and S.M. \ncontributed to the collection of data. S.M. and A.J. led the data analysis, with input from \nall team members. S.T. led the interpretation of results and the writing of the manuscript, \nwith input from all team members. C.F.G. contributed to writing the manuscript, led \nreviewing of the literature and designed the figures. All authors discussed the results and \ncommented on the manuscript. Please direct any comments or requests for data used in \nthe figures or analysis to S.M. (sunil.mani@ceew.in).\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-020-0596-7.\nCorrespondence and requests for materials should be addressed to S.M. or S.T.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2020\nNature Energy | VOL 5 | June 2020 | 450\u2013457 | www.nature.com/natureenergy\n457\n\n1\nnature research  |  reporting summary\nOctober 2018\nCorresponding author(s):\nSaurabh Tripathi\nLast updated by author(s): Feb 13, 2020\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nThe first wave of data (of 8,563 households) was collected in 2014-15 using a paper-based survey; the second wave of data (of 9,072 \nhouseholds) was collected using the app-based SurveyCTO software. Data from both waves was converted into .csv files by the survey \ncompany, Morsel Research and Development, and shared with the researchers for analysis.\nData analysis\nThe combined panel data set (consisting of both waves of data) was analysed using MS Excel and Stata 16.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. \nWe strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nThe data that support the findings of this study (both of panel and cross-sectional analysis) are made available through Figshare at https://doi.org/10.6084/\nm9.figshare.9810170. The Stata .do files that format, clean, and analyse the merged and appended data sets are available through Figshare at https://\ndoi.org/10.6084/m9.figshare.9810167 while the R scripts that produce Figures are available at https://doi.org/10.6084/m9.figshare.11838963. Information on \nunique identifiers between the data sets are available in the Stata .do file. \u2003\n\n2\nnature research  |  reporting summary\nOctober 2018\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThe study design, which is primarily quantitative, comprised of cross-sectional statistical analysis of the latest wave of survey data, and \nthe panel analysis over the two waves of data (from 2014-15 and 2018).\nResearch sample\nThe sample for this article is rural households from six states in India - Bihar, Jharkhand, Madhya Pradesh, Odisha, Uttar Pradesh, and \nWest Bengal. These states were selected because they contain 51% of India's rural population and are historically some of the most \nenergy-access deprived states in the country. In the first wave, we surveyed 714 villages in 51 districts across the six states, reaching out \nto 8,563 households. In the second wave, we doubled the sample in Odisha (which was earlier 504 households), and surveyed in 756 \nvillages across 54 districts, and reached out to 9,072 households. In each surveyed village, we surveyed 12 households at random. Over \nthe two waves we surveyed same households (to the extent possible), making it a panel dataset. We made replacements at the village \nlevel in case we were not able to survey the old household in the second wave. The sample is representative at the district level as well as \nat the state level. We primarily surveyed the head of the household in our survey.\nSampling strategy\nOwing to budgetary constraints, in the first wave, we sampled one district from each administrative division of each state, except for in \nWest Bengal, which only has three large divisions, where we sampled two districts in each division. In the second wave, we sampled two \ndistricts in each division of West Bengal as well as Odisha, thereby doubling the sample in Odisha. This was done in order to improve \nrepresentation of the sample in Odisha. Using Census 2011 data, each district was chosen with a probability proportional to its \npopulation relative to the division population. Once the district is selected, we split the total population of each district into two groups \nof villages such that one comprises primarily large villages and other comprises primarily small, but the population in each group is more \nor less equal. Although the number of households in each group is the same, the group with large villages has fewer villages. Seven \nvillages were then sampled from each group with probability proportional to population. This approach ensured that the sampling is self-\nweighted within a district, yet guaranteed both small and large villages are well represented in the sample. In each village, we surveyed \n12 households at random.\nData collection\nThe first wave of data was collected through a pen-and-paper-based survey, wherein enumerators (contracted by the survey company - \nMorsel Research and Development) interviewed the head of the household (if not available, an adult member) of each sampled \nhousehold using a questionnaire that was prepared by the authors and their collaborators and tested in the field pilots. The second wave \nof data was collected using the app-based SurveyCTO software by the enumerators. This allowed for real-time monitoring of data quality \nand enumerator bias, which were used to commission re-surveys. About 5 per cent of the sample was resurveyed in each wave. Each of \nthe enumerators were trained and evaluated by the researchers in both waves. The researchers were not present at the time of the \nsurvey and hence did not interfere with the process of data collection. The data was converted into .csv files by the survey company and \nshared with the researchers for analysis a few weeks post-data collection.\nTiming\nThe first wave of data was collected from November 2014 to February 2015. The second wave was collected from March to June 2018.\nData exclusions\nFor the analysis contained in this article, two subsets were used. In the first, i.e. cross-section subset (4,102 households), only those that \nused LPG in Wave 2, yet may or may not have used LPG in Wave 1 were considered. Further, we included only those that were surveyed \nin both waves because of the inclusion of certain lagged (Wave 1) covariates. In the second panel subset (1,411 households), only those \nthat used LPG in both waves of survey were included, in order to determine factors that determine upward mobility in LPG-use.\nNon-participation\nIn the second wave of the survey, we were able to retain about 85% of the households that participated in the first wave. Drop outs were \nprimarily owing to enumerators finding the particular household's door locked at the time of visit, or owing to them not finding the \nparticular household available at the address provided (perhaps because the family has shifted to a different location). In these cases, the \nenumerators were asked to randomly select and interview a new household from the same village. However, for the purpose of this \narticle, only households that were interviewed in both waves of survey were included.\nRandomization\nNot applicable.\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \n\n3\nnature research  |  reporting summary\nOctober 2018\nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology\nAnimals and other organisms\nHuman research participants\nClinical data\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\n\n\n    Scientific Research Findings:", "answer": "We estimate the odds of a PMUY beneficiary using LPG for all cooking needs are about 56% lower than those of a general customer, controlling for baseline socioeconomic and geographic differences. We find that households with irregular and uncertain cash flows \u2014 those dependent on agriculture or on daily wages \u2014 have lower odds of using LPG as their main cooking fuel, perhaps owing to the recurring and inflexible cost of LPG refills. Households in villages with a greater proportion of LPG primary users have higher odds of increased LPG use, suggesting a possible peer-effect or influence of other village-specific factors such as access to biomass and LPG availability. Further, cattle ownership \u2014 which facilitates access to dung cakes \u2014 and easy access to firewood are major hindrances to increased LPG use.", "id": 22}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-020-0600-2\n1School of Public Health, University of California, Berkeley, CA, USA. 2Columbia University Mailman School of Public Health, New York, NY, USA. \n3Department of Environmental Health, Harvard TH Chan School of Public Health, Boston, MA, USA. 4Department of Statistics and Data Sciences and \nDepartment of Women\u2019s Health, University of Texas, Austin, TX, USA. 5Department of Environmental and Radiological Sciences, Colorado State University, \nFort Collins, CO, USA. 6Propeller Health, San Francisco, CA, USA. 7Louisville Metro Department of Public Health and Wellness, Louisville, KY, USA. \n8Christina Lee Brown Environme Institute, University of Louisville, Louisville, KY, USA. 9Louisville Metro Office of Civic Innovation, Louisville, KY, USA. \n10Family Allergy & Asthma, Louisville, KY, USA. \u2709e-mail: jac2250@cumc.columbia.edu\nC\noal-fired power plants provide a large amount of electric-\nity worldwide. In 2015, they produced six trillion megawatt \n(MW) hours and 25% of the global supply1. Simultaneously, \ncoal-fired power plants emit air pollution, which included 63% of \nthe US economy-wide sulfur dioxide (SO2) emissions in 20142, as \nwell as nitrogen oxides (NOx), PM2.5 (particulate matter less than \n2.5\u2009\u03bcm in diameter) and PM10 (particulate matter between 2.5 and \n10\u2009\u03bcm in diameter), mercury, acid gases, polycyclic aromatic hydro-\ncarbons and volatile organic compounds3. Such pollutants are asso-\nciated with increased asthma symptoms, emergency room visits \n(ERVs), hospitalizations and mortality4\u20138.\nAmong populations that live near coal-fired power plants and \nfossil fuel refineries, some, though not all9,10, epidemiological stud-\nies found a relationship between higher SO2 levels and uncontrolled \nasthma11, respiratory symptoms12\u201315 and respiratory-related hospi-\ntalizations16. Residential proximity to such facilities alone, without \nassessed air quality, was also identified as a risk factor for asthma \nexacerbation17\u201319.\nPrior scholarship related to asthma and coal-fired power plant \nexposures often consisted of observational and cross-sectional \nstudies that considered single air pollutants (usually SO2) and \nhospitalizations, pulmonary function or symptoms alone. Studies \nof symptoms lacked objective measures and usually relied on par-\nticipant diaries. Some studies overcame a portion of these limita-\ntions; for example, Smargiassi et\u00a0al.16 used a case-crossover design to \nevaluate the relationship between SO2 concentrations and ERVs and \nhospitalizations among young children who lived near a refinery. \nTo build on these prior studies, we incorporate improved exposure \nand outcome assessment and capitalize on recent abrupt changes in \ncoal-fired power plant emissions.\nBetween 2000 and 2015 in the United States, 49.5\u2009GW of capacity \nfrom coal-fired generators retired at 146 coal-fired power plants1,20 \nand many generating units installed flue-gas desulfurization systems \n(alternatively, SO2 emission \u2018controls\u2019) to comply with regulations \nfrom the US Environmental Protection Agency (USEPA) and indi-\nvidual states, which include the Acid Rain Program, the Clean Air \nInterstate Rule and Mercury and Air Toxics Standards2. The discrete \nnature of these energy transitions and the ensuing abrupt changes in \nemissions present circumstances for a \u2018natural experiment\u2019 to study \nrelated changes in health in the time frame and population exposed \nto emissions from the coal-fired power plants21,22. The key feature is \nthat the transition-induced change in exposure occurs for reasons \nunrelated to a health investigation and produce exposure changes \nthat more closely resemble an experiment than a typical observa-\ntional study. Such a natural experiment supports the study of the \ninfluence of coal-fired power plant emissions on asthma outcomes \nmore directly, without relying on exposure\u2013response functions esti-\nmated with different populations across time and space with vary-\ning levels of pollution exposure23. Previous studies framed in this \nmanner used a steel mill closure24, a ban on coal sales25, the Olympic \nGames26,27 and the Nitrogen Oxides Budget Program28 to study the \nrelationship between air pollution and respiratory-related medica-\ntion expenditures, hospitalizations and death.\nWe used retirements and SO2 emissions control installations \nat four coal-fired power plant facilities near Louisville, Kentucky, \nbetween 2013 and 2016 to frame a natural experiment, which  \nImproved asthma outcomes observed in the \nvicinity of coal power plant retirement, retrofit \nand conversion to natural gas\nJoan A. Casey\u200a \u200a1,2\u2009\u2709, Jason G. Su1, Lucas R. F. Henneman\u200a \u200a3, Corwin Zigler4, Andreas M. Neophytou1,5, \nRalph Catalano\u200a \u200a1, Rahul Gondalia6, Yu-Ting Chen7, Leanne Kaye6, Sarah S. Moyer7, Veronica Combs8, \nGrace Simrall9, Ted Smith8, James Sublett10 and Meredith A. Barrett6\nCoal-fired power plants release substantial air pollution, which included over 60% of US sulfur dioxide emissions in 2014. Such \nair pollution may exacerbate asthma, but direct studies of the health impacts linked to power plant air pollution are rare. Here \nwe take advantage of a natural experiment in Louisville, Kentucky, where one coal-fired power plant was retired and converted \nto natural gas, and three others installed SO2 emission control systems between 2013 and 2016. Dispersion modelling indicated \nthat exposure to SO2 emissions from these power plants decreased after the energy transitions. We used several analysis strate-\ngies, which include difference-in-differences, first-difference and interrupted time-series modelling to show that the emissions \ncontrol installations and plant retirements are associated with a reduced asthma disease burden related to hospitalizations and \nemergency room visits at the ZIP-code level, and to individual-level medication use as measured by digital medication sensors.\nNature Energy | VOL 5 | May 2020 | 398\u2013408 | www.nature.com/natureenergy\n398\n\nArticles\nNaTUre Energy\ngenerated abrupt changes in air pollution exposures that varied \nin extent across the Louisville area. The circumstances of these \nenergy transitions and their resemblance to an experiment that \n(approximately) randomizes exposure changes motivates a vari-\nety of analysis approaches that each mitigate some of the common \nthreats to observational studies. Our approaches include compar-\ning asthma-related end-points in areas more and less exposed to \nchanges in air pollution and individuals before and after the dis-\ncrete transition events to evaluate impacts on the asthma-related \nend-points. Kentucky has historically ranked among the top five \nin the United States for SO2, NO2 and PM10 emissions from power \ngeneration29. We confirmed air-quality improvements after the \npower plant energy transitions using a longitudinal coal-fired \npower plant emission exposure model30. Analyses that leveraged \ndifferent elements of spatial and temporal variability in exposure \ndemonstrated that the coal-fired power plant energy transitions \nwere associated with reductions in asthma-related hospitalizations \nand ERVs at the ZIP-code level and asthma symptoms measured \nat the individual level using digital medication sensors within the \nLouisville metropolitan area.\nImpact of retirement, retrofit and conversion on emissions\nFour coal-fired power plants\u2014Cane Run, Clifty Creek, Mill \nCreek and Rockport\u2014with emissions that impacted air quality in \nJefferson County, Kentucky (which fully contains Louisville), either \nretired or installed SO2 controls between 2012 and 2016 (Fig. 1). We \nquantified monthly ZIP-code variability in SO2 emissions exposure \nfrom these plants using the HYSPLIT Average Dispersion (HyADS) \nmodel, which aggregates the results from 146,000 HYSPLIT runs \nper power plant30. During the study period, the quarterly median \nunitless HyADS exposure was 6,553 (interquartile range (IQR), \n2,283 and 9,702; maximum, 31,781). Power plant energy transitions \nresulted in declining levels over time (Fig. 2 and Supplementary \nFigs. 1 and 2). Data from the Louisville Metro Air Pollution Control \nDistrict confirmed that retirements at Cane Run and scrubbers at \nMill Creek reduced the annual SO2 emissions by 9.6 and 12.9 mil-\nlion kg, respectively (Supplementary Fig. 3). Comparing years pre- \nand post-control installations, SO2 emissions also declined at Clifty \nCreek (\u221290%) and Rockport (\u201350%). These four facilities contrib-\nuted 36, 30 and 16% of the total average ZIP-code HyADS exposure \nin Louisville in 2012, 2014 and 2016, respectively. The transitions \nat the four facilities contributed to an overall decline in coal-fired \npower plant emissions exposures in Jefferson County.\nHyADS emissions exposures peaked annually during the third \nquarter (April\u2013June, Fig. 2). In a companion analysis, we found that \nmeteorological variability contributed more than emissions reduc-\ntions to changing HyADS in 2012, 2013 and 201431. Table 1 shows \nthat, accounting for seasonality and meteorological factors, the \naverage level of HyADS exposure decreased substantially after three \nof the four energy transitions, with a 55% decline from baseline after \nthe second quarter of 2015 (Q2-2015).\nObserved changes in ZIP-code-level asthma outcomes\nThe median (IQR) quarterly asthma hospitalizations/ERV counts \nacross the 35 ZIP codes in Jefferson County between 2012 and 2016 \nwas 16 (range 9\u201331), and the counts declined county-wide over time \n(Fig. 3 and Supplementary Table 1). Between 2012 and 2016 the rates \nof uninsured and unemployed individuals also fell. Quarter 4 typi-\ncally exhibited hospitalization and ERV values higher than those of \nother quarters after adjusting for ZIP code and annual specific means.\nThree of the four energy transitions were associated with reduc-\ntions in ZIP-code-level asthma hospitalizations and ERVs (Table 2), \nwhich correspond to the three transitions associated with reduced \nHyADS (Table 1). The largest reduction in risk came after the \nQ2-2015 transitions, relative risk (RR)\u2009=\u20090.81; 95% confidence inter-\nval (CI), 0.70, 0.92. We therefore focused the next stage of analysis \non the Q2-2015 transitions at Cane Run, Mill Creek and Rockport. \nDuring the time surrounding the Q2-2015 transition (Q2-2014 to \nQ2-2016), the average quarterly ZIP code HyADS reduction was \n25,281 (standard deviation (s.d.)\u2009=\u20093,638).\nIn a difference-in-differences framework, we categorized ZIP \ncodes based on their pre-period HyADS exposure (high versus low). \nThe results indicated that the Q2-2015 energy transition reduced \nasthma hospitalizations and ERVs by an additional 2.8 visits per ZIP \ncode per quarter in areas with high pre-transition exposure relative to \nareas with a lower pre-transition exposure (Fig. 4). When we specified \npre-transition HyADS as a continuous variable, results were similar \nwhen converted into a comparable scale (\u22120.4 (95% CI: \u22120.2, \u22120.7)) \nasthma hospitalizations and ERVs per ZIP code per 1,000-unit higher \npre-period HyADS exposure; Supplementary Fig. 4).\nWith a first-difference linear regression model, we found that a \n1,000-unit ZIP-code-level reduction in HyADS exposure from the \nRockport\n99 million kg\n100\n105\n110\n2012 HyADS, 1,000s\n115\n2012 total SO2 emissions (million kg)\n10\n3\n30\n30\n60 km\nClifty Creek\n96 million kg\nCane Run\n6 million kg\nMill Creek\n28 million kg\n0\n1,000\nIndiana\nKentucky\nFig. 1 | Power plant locations, emissions and exposure. The locations of the four coal-fired power plant facilities, their total SO2 emissions (larger grey \ncircles indicate higher emissions) and their total HyADS exposures at the ZCTA level within Jefferson County in 2012. Unit and facility data downloaded \nfrom the USEPA Air Markets Program50.\nNature Energy | VOL 5 | May 2020 | 398\u2013408 | www.nature.com/natureenergy\n399\n\nArticles\nNaTUre Energy\nyear prior to the year after the Q2-2015 energy transitions resulted \nin, on average, 2.2 fewer asthma hospitalizations and ERVs (95% \nCI: \u22124.5, 0.2) per ZIP code per year and a first-difference model \nthat specified categories of \u0394HyADS showed the largest effect for \nthe highest \u0394HyADS category (Supplementary Fig. 4). Inferences \nremained stable in sensitivity analyses using baseline popula-\ntion weights instead of adjusting models for baseline population \n(Supplementary Table 2).\nObserved changes in individual-level asthma symptoms\nWe identified 207 study participants in the AIR Louisville pro-\ngramme who were under observation during the year prior to and \nthe year after the 8\u2009June\u20092016 Mill Creek power plant scrubber instal-\nlation. Most participants were female (67%), of white race/ethnicity \n(63%) and middle-aged (average age of 45 years) (Supplementary \nTable 3). Participants\u2019 median monthly HyADS exposure was \n1,915 (IQR: 1,172, 3,050) (Supplementary Fig. 5). Participants had \na daily mean of one short-acting beta agonist (SABA) inhaler use \n(s.d.\u2009=\u20091.5) and a median of no SABA uses (IQR\u2009=\u20090, 1). From visual \ninspection, daily rescue inhaler use was more prevalent and variable \nearlier in the study period, probably driven by the smaller number \nof enrolled participants (Supplementary Fig. 6) and later trended \ndownward (Supplementary Fig. 7).\nIn a within-person conditional quasi-Poisson model, we \nobserved a reduced monthly average daily SABA use associated \nwith a reduced monthly HyADS exposure (RR\u2009=\u20090.94, 95% CI: 0.89, \n0.98, for each 1,000-unit decrease in HyADS).\nAs the June\u20092016 Mill Creek scrubber installation resulted in \nrelatively uniform reductions in HyADS exposure across Jefferson \ncounty (Fig. 5b), we used an interrupted time-series framework. \nWe identified a level shift in SABA use (at the time of scrubber \ninstallation) and a possible slope change (decreasing trend in \nSABA use) (Fig. 6). The scrubber installation was associated with \na 17% reduction in monthly average daily SABA use (RR\u2009=\u20090.83, \n95% CI: 0.69, 1.00) and a 2% reduction (95% CI: \u22125%, 1%) for each \nmonth thereafter.\nIn two sensitivity analyses, we evaluated the change in odds of \nhaving any daily SABA use (monthly average of \u22651\u2009use\u2009day\u20131 ver-\nsus <1\u2009use\u2009day\u20131) and high daily SABA use (monthly average of \n\u22654\u2009uses\u2009day\u20131 versus <4\u2009uses\u2009day\u20131) at the time of scrubber instal-\nlation. We found an apparent larger immediate effect on higher \nmonthly average daily SABA use (32% reduction (odds ratio\u2009=\u20090.68, \n95% CI: 0.45, 1.02]) and a trend in reduced monthly average daily use \nof \u22651 uses (17% reduction each month after the scrubber installation \n(odds ratio\u2009=\u20090.83, 95% CI: 0.71, 0.97]) (Supplementary Table 4).\nDiscussion\nThe top four coal-fired power plants in terms of emissions that \naffect air quality in Louisville in 2012 either retired or installed SO2 \nemission controls between 2012 and 2016. The resulting air-qual-\nity improvements translated into reductions in both acute asthma \noutcomes\u2014measured by quarterly ZIP-code-level asthma-related \nQuarter 3\n2012\n0\n1 \u00d7 104\n2 \u00d7 104\na\nb\nUnit ID\n983-1\n983-2\n983-3\n983-4\n1,363-4\n1,363-5\n1,363-6\n1,364-1\n1,364-2\n1,364-3\n1,364-4\n6,166-MB1\n6,166-MB2\n983-5\n983-6\n\u20131,250\n\u20131,000\n\u2013750\n\u2013500\n\u2013250\n0\n2013\n2014\n2015\n2016\n2017\n2015\u20132014\nMean reduction = 60%\nMean reduction = 86%\nHyADS absolute change for selected units\nHyADS, unitless\nMean reduction = 90%\nMean reduction = 76%\n2015\u20132014\n2016\u20132015\n2016\u20132015\nQuarter 4\nQuarter 1\nQuarter 2\nFig. 2 | Quarterly mean ZIP code-level HyADS exposure in Jefferson County. a, From 2012 to 2017 stratified by coal-fired power plant unit. Clifty Creek, \n983; Cane Run, 1,363; Mill Creek, 1,364; Rockport, 6,166. b, Absolute change in quarterly ZIP code tabulation area-level HyADS and mean percentage \nreduction across the two stated years in the stated quarter.\nTable 1 | Coal-fired power plant events and HyADS emissions \nexposures\nCoal-fired power plant transition date\n%\u0394 HyADS (95% CI)\n\u2002Q2-2013\n11 (6, 16)\n\u2002Q4-2014\n\u221227 (\u221231, \u221223)\n\u2002Q2-2015\n\u221255 (\u221261, \u221247)\n\u2002Q2-2016\n\u221229 (\u221233, \u221224)\nAssociation between coal-fired power plant events and ZIP-code-level HyADS emissions exposures \nin the 35 Jefferson County ZIP codes with a population greater than 5. %\u0394 was calculated relative \nto countywide average HyADS level in 2012 (11,392 units). Estimates provided from an ordinary \nleast squares linear regression model adjusted for quarterly mean temperature, wind speed, \nrelative humidity and atmospheric pressure, and quarter and ZIP code. Includes Liang and Zeger \ncluster-robust standard errors. Q1, quarter 1 (Jan\u2013Mar); Q2, quarter 2 (Apr\u2013June); Q3, quarter 3 \n(July\u2013Sept); and Q4, quarter 4 (Oct\u2013Dec).\nNature Energy | VOL 5 | May 2020 | 398\u2013408 | www.nature.com/natureenergy\n400\n\nArticles\nNaTUre Energy\nhospitalizations and ERVs\u2014and daily symptoms, as measured by \nsensor-collected SABA use. In the spring of 2015, coal-fired power \nplant units retired at Cane Run and SO2 controls were installed \nat Mill Creek and Rockport. These energy transitions resulted in \napproximately three fewer hospitalizations and ERVs per ZIP code \nper quarter in the following year, which translates into nearly 400 \n(\u22122.8 per ZIP quarter\u2009\u00d7\u20094 quarters\u2009\u00d7\u200935 ZIPs) avoided hospitaliza-\ntions and ERVs in Jefferson County annually. At the individual level, \nthe Mill Creek SO2 scrubber installed in June\u20092016 was associated \nwith immediate reductions in SABA use and a marginally declining \ntrend in use through May\u20092017.\nUniquely, this study found that exposure reductions to coal-fired \npower plant emissions were consistently associated with a reduced \nasthma morbidity at two spatial and temporal resolutions: monthly \naverage daily SABA use within individuals and acute quarterly \nERVs and hospitalizations at the ZIP-code level. Acute outcomes, \nsuch as hospitalizations, represent an important, yet more infre-\nquent, severe and costly32 signal of asthma morbidity. SABA use can \nrepresent the daily burden of disease and the future risk of adverse \noutcomes33. Our analyses used digital medication sensors to objec-\ntively track the time and location of SABA use and avoid the poten-\ntial recall bias associated with the diaries used in previous studies of \nshort-term air pollution levels and asthma15,34,35.\nIn 2014, US power plants accounted for 64% of SO2, 14% of NOx \nand 7% of PM2.5 economy-wide emissions, of which coal power \ncomprised 98% of SO2 and >82% of both NOx and PM2.5 power \nplant emissions2. Exposures to these air pollutants, even at relatively \nlow levels, are associated with asthma morbidity7,8,10,36. Four main \nmechanisms\u2014oxidative stress, airway remodelling, inflammation \nor immunological response and an enhanced response to inhaled \nallergens37\u2014likely explain how air pollution contributes to asthma \nonset and exacerbation. Laboratory studies of SO2 exposure among \nasthmatics have demonstrated bronchoconstriction in humans38 \nand airway inflammatory and immune responses in rats39.\nIn a 2017 report for the US Department of Energy, Massetti et\u00a0al. \ncharacterized SO2 as the main source of air-pollution-related eco-\nnomic damage associated with coal-fired power plants due to its \nhigh volume of emissions and because SO2 is a PM2.5 precursor2. \nAmong populations that live close to coal-fired power plants, SO2 is \nassociated with wheeze and asthma prevalence13,40, asthma attacks \nand ERVs12,15,16, although it is likely that other pollutants play a \nrole. Coal electricity generation contributes to anthropogenic PM2.5 \nlevels (which account for 1\u20133% of all asthma-related ERVs in the \nAmericas41) and NO2 exposures (which may cause 19% of asthma \nincidence in high-income countries42). Among children who lived \nnear Israel\u2019s 2580\u2009MW Orot Rabin coal-fired plant, a combined NOx \nand SO2 exposure was more strongly associated with a reduced pul-\nmonary function test performance than either pollutant alone14.\nSeveral studies characterized exposure via residential proximity \nto coal-fired power plants, a metric that incorporates, albeit crudely, \n1,000\na\nb\n750\n500\n250\n0\nQ1-2012\nQ2-2012\nQ3-2012\nQ4-2012\nQ1-2013\nQ2-2013\nQ3-2013\nQ4-2013\nQ1-2014\nQ2-2014\nQ3-2014\nQ4-2014\nQ1-2015\nQ2-2015\nQ3-2015\nQ4-2015\nQ1-2016\nQ2-2016\nQ3-2016\nQ4-2016\nCount\nIndicates the quarter-year\nin which the SO2 controls were\ninstalled: 3 units in Q1-2013, 3\nin Q2-2013\u20132, 1 in Q4-2014, 3 in\nQ2-2015, 1 in Q2-2016 (n = 1)\nand Q2-2015 when 2 coal-fired\nunits were retired.\n0\n20\n40\n60\n80\nQ1-2012\nQ2-2012\nQ3-2012\nQ4-2012\nQ1-2013\nQ2-2013\nQ3-2013\nQ4-2013\nQ1-2014\nQ2-2014\nQ3-2014\nQ4-2014\nQ1-2015\nQ2-2015\nQ3-2015\nQ4-2015\nQ1-2016\nQ2-2016\nQ3-2016\nQ4-2016\nQuarterly Louisville-wide\nasthma hospitalizations and ERVs, N\nFig. 3 | Quarterly ZIP-code-level counts of asthma hospitalizations and \nERVs in Jefferson County. a, Countywide from 2012 to 2016. b, By ZCTA \nfrom 2012 to 2016. Data provided by the Louisville Metro Department of \nPublic Health and Wellness.\nTable 2 | Relative risk of ZIP-code-level asthma hospitalization \nand ERV\nCoal-fired power plant transition date\nRR (95% CI)\n\u2002Q2-2013\n1.08 (0.96, 1.19)\n\u2002Q4-2014\n0.90 (0.75, 1.06)\n\u2002Q2-2015\n0.81 (0.70, 0.92)\n\u2002Q2-2016\n0.91 (0.80, 1.03)\nThese estimates of relative risk are related to each coal-fired power plant event in the 35 Jefferson \nZIP codes with a population greater than 5. They are provided from a quasi-Poisson regression \nmodel with a ZIP-code-level annual population offset and adjusted for annual percentage non-\nHispanic black individuals, unemployed individuals, individuals living below the federal poverty \nthreshold, quarterly mean temperature, wind speed, relative humidity and atmospheric pressure, \nand year, quarter and ZIP code. Includes Liang and Zeger cluster-robust standard errors.\nNature Energy | VOL 5 | May 2020 | 398\u2013408 | www.nature.com/natureenergy\n401\n\nArticles\nNaTUre Energy\nmultiple coal-related pollutants17\u201319. The continuous and quantita-\ntive HyADS metric improves the exposure characterization around \ncoal-fired power plants and attempts to isolate the effect of collective \ncoal emissions on health. However, the temporal resolution of our \ndata and the HyADS metric make it difficult to compare our find-\nings, in magnitude, to those of prior research that reports the rela-\ntionships between coal-fired power plant exposure and increased \ndiary-reported respiratory symptoms12,13,40 and SABA use15, as well \nas ERVs and hospitalizations16,19.\nEpidemiologists cannot randomize individuals to differing lev-\nels of air pollution, which can make causal inference challenging. \nQuasi-experimental designs23 that leverage the circumstances of nat-\nural experiments provide one alternative method that, when paired \nwith the appropriate analytical approaches, can mimic pseudo-ran-\ndomization of individuals and populations to varying levels of envi-\nronmental exposures21,22. Compared to traditional observational  \nstudies, quasi-experimental techniques are better equipped to han-\ndle unobserved differences in populations that might otherwise \nconfound exposure\u2013outcome associations43. A few prior studies \nof asthma exacerbations used quasi-experimental designs24,26\u201328, \nfor example, monitoring changes in asthma-related hospitaliza-\ntions and acute care visits before, during and after an Olympic \nGames that reduced local air pollution26,27. Environmental expo-\nsures tend to follow a social gradient44 and co-occurring factors, \nsuch as high body mass index, smoking, low socio-economic sta-\ntus and area-level racial segregation are risk factors for asthma45. \nTherefore, an abrupt change in an environmental exposure \nimproves our ability to disentangle the social and environmental \ndeterminants of health.\nTo our knowledge, no study has focused on installations for the \ncontrol of SO2 emissions or on power plant retirements to assess \nthe potential asthma-related health benefits. We examined multi-\nple interventions across several years and tracked the intervention \nimpacts at several phases (that is, intervention, emissions, exposure \nand health outcome). An important analytical decision in quasi-\nexperimental studies is the choice of control group. Our difference-\nin-differences analysis compared changes between ZIP codes that \nwere initially most exposed to the power plants that generated the \ntransition and control ZIP codes that were comparatively less subject \nto exposure from these plants. In our first-difference analysis, we \nused ZIP codes that experienced less reduction in HyADS exposure \nduring the study period to control for time trends in hospitaliza-\ntions and ERVs and other factors over the study period, and thereby \nminimized the threat of confounding due to such factors. Both \ndesigns control for the variation in observed and unobserved fixed \ncharacteristics of place46. At the individual level, we implemented \na within-person analysis, which regards each person as his or her \nown control, to reduce the threat of bias associated with confound-\ning variables related to differences across people. This case time-\nseries analysis only controlled for the overall linear temporal trends. \nIndividual- and ZIP-code-level analyses consistently supported the \nnotion that the energy transitions improved asthma outcomes.\nThis study has limitations. The AIR Louisville cohort partici-\npants were enrolled through a number of channels and did not have \nstandardized clinical oversight. We lacked access to individual-level \ndata on participants (for example, healthcare utilization, socio-eco-\nnomic status, smoking or tobacco-smoke exposure), although most \nof these factors are considered time-invariant and controlled for by \nQ1-2012\n10\nAverage ZIP-code asthma count\n20\nHigh\nLow\nUnadjusted\nAdjusted\n30\na\nb\n0\n\u20132\n\u20134\n\u20136\n\u20138\n\u2206 quarterly asthma hospitalizations/ERVs\nQ2-2012\nQ3-2012\nQ4-2012\nQ1-2013\nQ2-2013\nQ3-2013\nQ4-2013\nQ1-2014\nQ2-2014\nQ3-2014\nQ4-2014\nQ1-2015\nQ2-2015\nQ3-2015\nQ4-2015\nQ1-2016\nQ2-2016\nQ3-2016\nQ4-2016\nHyADS group\nRetrofits and\nretirements\nFig. 4 | Spring 2015 coal-fired power plant events and counts of ZIP-code-level asthma hospitalization and ERVs. a, Trends in count of ZIP-code-level \nasthma hospitalization and ERVs from Q1-2012 to Q4-2016. The dashed line notes Q2-2015, where the transition took place late in the quarter. Prior \nto Q2-2015, trends in ZIP-code-level counts of asthma events appear parallel, which provides evidence that we meet the parallel trends assumption of \ndifference-in-differences analysis. b, Difference-in-differences results with 95% CI from an ordinary least squares model (equation (3)) with a binary \nspecification of pre-period HyADS (high (\u226532,500) versus low (<32,500)). The model was adjusted for annual total population, percentage non-Hispanic \nblack individuals, unemployed individuals, individuals living below the federal poverty threshold, quarterly mean temperature, wind speed, relative \nhumidity and atmospheric pressure, and included fixed effects for year, quarter and ZIP code. Liang and Zeger cluster-robust standard errors were used.\nNature Energy | VOL 5 | May 2020 | 398\u2013408 | www.nature.com/natureenergy\n402\n\nArticles\nNaTUre Energy\nthe study design. Time-varying factors, however, both at the indi-\nvidual and area level could have confounded our results. A prior \nstudy found that the digital health intervention (that is, Propeller \ninhaler system) may have helped reduce SABA use, with the larg-\nest improvements during the first few months47. If enrolment in the \nprogramme coincided with plant changes, this could have resulted \nin overestimated effects. However, our within-person interrupted \ntime-series design suggested an abrupt change in SABA use at the \ntime of a SO2-control installation and a continued decline thereaf-\nter. We assumed a linear trend in outcome, although changes may \nhave followed a non-linear pattern. Although the results in Fig. 6 \nsuggest that the largest change in SABA use corresponds to the larg-\nest drop in HyADS in Q2-2015, we lacked the individual-level data \nin the pre-period to assess this statistically. At the ZIP-code level, \ndue to data limitations, we could not consider asthma-related hospi-\ntalizations and ERVs separately, nor could we track multiple events \nwithin individuals. Finally, we were not able to assess cumulative \nimpacts; for example, the relationship between chronic coal-fired \npower-plant-related exposures and asthma prevalence or the co-\noccurrence of asthma and chronic disease. As the evidence is mixed \nMean reduction = 58%\nMean reduction = 32%\nMean reduction = 82%\nMean reduction = 82%\nMean reduction = 92%\nMean reduction = 88%\nMean reduction = 100%\nMean reduction = 84%\nMean reduction = 88%\nMill Creek unit 3 HyADS absolute change\n\u20136,000\n\u20134,000\n\u20132,000\n0\nJanuary\nFebruary\nUnit ID\nMill creek\nscrubber\ninstalled\n1,364-1\n6,000\na\nb\n4,000\n2,000\nCoal-fired power plant exposure\n0\nJune 2015\nSeptember 2015\nDecember 2015\nMarch 2016\nJune 2016\nSeptember 2016\nDecember 2016\nMarch 2017\nJune 2017\n1,364-2\n1,364-3\n1,364-4\nMarch\nApril\nMay\nJune\nJuly\nAugust\nSeptember\nOctober\nNovember\nDecember\nMean reduction = 88%\nMean reduction = 96%\nMean reduction = 58%\nFig. 5 | Monthly mean ZIP-code-level HyADS exposure from the Mill Creek power plant in Jefferson County, 2015\u20132017. a, Stratified by Mill  \nCreek power plant unit. b, Average change in monthly mean ZIP code tabulation area-level coal-fired power plant air pollution exposure (HyADS)  \nbetween 2015 and 2016.\nNature Energy | VOL 5 | May 2020 | 398\u2013408 | www.nature.com/natureenergy\n403\n\nArticles\nNaTUre Energy\nregarding the association between long-term air pollution exposure \nand asthma10,48, this presents an area for future research.\nOur study focused specifically on the impact of four coal-fired \npower plants, but took place against the backdrop of changing emis-\nsions at other power plants and reductions in vehicle emissions. We \ndid not explicitly include air pollutants that might affect asthmat-\nics, for example, SO2, PM2.5, NO2 or ozone, but HyADS does repre-\nsent an index of the mixture of pollutants emitted by power plants. \nChanges in air quality unrelated to power plants but that trended \nin time were controlled for in our models. Although the HyADS \nmodel provided a summary of coal-fired power-plant-specific \nexposure, it did so at the monthly level, which limited our ability \nto assess the relationship between short-term exposure fluctuations \nand SABA use.\nIn conclusion, this study showed reductions in coal-fired power \nplant air pollution exposure after retirements and SO2 control \ninstallations near Louisville. These reductions appear to translate \ninto substantially fewer asthma-related ERVs and hospitalizations, \nas well as fewer average daily SABA uses. Given that 20.4 million \nadults, or about 9% of the population, suffered from current asthma \nin 201649, the shift in US energy trends away from coal-fired elec-\ntricity generation may reduce asthma morbidity below otherwise \nexpected levels. Future research should evaluate this potential \nimpact so that plant controls and retirement sites can be phased to \naffect neighbourhoods and schools at the highest risk for asthma.\nMethods\nThe natural experiment. The circumstances that support this natural experiment \narose in Jefferson County, which fully contains Louisville, a city that covers  \n842\u2009km2 and had a population of 600,000 people in 2010. We hypothesized  \nthat (1) coal-fired power plant emissions and subsequent population exposures \ndropped after the coal-fired power plant energy transitions and (2) the lower \nexposures resulting from energy transitions translated into fewer healthcare \nutilization events and symptoms. Our analyses used two health outcomes (acute \nasthma-related healthcare utilization and asthma symptoms) at two spatial scales \n(ZIP-code and individual level) and employed a measure of coal-fired power plant \nemission exposure.\nExposure to power plant emissions. From the USEPA Air Markets Program50, \nwe downloaded data on power plant facilities (which typically comprise multiple \ngenerating units) nationwide with at least one unit using coal as its primary  \nfuel between 2008 and 2017. These data included power plant latitude and \nlongitude, number of units and fuel type by unit, monthly SO2 emissions, SO2 \ncontrol type and installation date, and retirement date. We obtained further \ninformation on the timing of the Jefferson County coal-fired power plant control \ninstallations, retirements and SO2 emissions from the Louisville Metro Air \nPollution Control District.\nWe estimated the exposure to power plant emissions over time using the \nrecently developed HyADS model30. HyADS uses the HYSPLIT air parcel \ntrajectory model51 to quantify the extent to which any power plant influences air \nquality across US ZIP code tabulation areas (ZCTAs), which we mapped to ZIP \ncodes using a crosswalk (https://www.udsmapper.org/zcta-crosswalk.cfm) (see \nSupplementary Note 1 and Henneman et\u00a0al.30 for details). This method produces \na unitless measure of emissions influence that is highly correlated with related \nmeasured and modelled exposure metrics for coal emissions exposure30,52. HyADS \nhas previously been applied to estimate US-wide health benefits achieved through \nreduced coal-fired power plant SO2 emissions53.\nWe ran HyADS for each of the over 1,009 US coal-fired power plant units in \noperation (located at 478 facilities) nationwide in 2012, 2014 and 2016. In 2012, \nwe ranked each unit\u2019s exposure contribution in each Louisville ZIP code and \nidentified the four facilities that had the highest influence on Louisville ZIP codes \n(Supplementary Fig. 8). These four facilities subsequently underwent retirement \nor control installations before 2017. We ran HyADS for each facility\u2019s units for \nyears 2012\u20132017 to estimate the ZIP-code-level monthly (used in individual-\nlevel analyses) and quarterly (as the average of monthly, used in ZIP-code-level \nanalyses) influence of each unit\u2019s emissions on Louisville air quality over time \n(Supplementary Data).\nEmissions exposures for the four plants of interest. The four highest-ranked \nfacilities by HyADS impact in 2012 included Cane Run, a 943\u2009MW plant located \n14\u2009km southwest of downtown Louisville; Clifty Creek, a 1,303\u2009MW plant located \nabout 75\u2009km northeast of Louisville that had the greatest impact of any coal-fired \npower plant in the country on east Jefferson County; Mill Creek, a 1,717\u2009MW plant \nlocated 25\u2009km southwest of downtown Louisville that had the greatest impact of \nany coal-fired power plant in the country on west Jefferson County and Rockport, \na 2,600\u2009MW plant located about 100\u2009km west of Louisville.\nClifty Creek installed wet limestone scrubbers for units 1\u20133 on 20 March\u20092013 \nand for units 4\u20136 on 15\u2009May\u20092013. The Mill Creek plant installed wet limestone \nscrubbers for unit 4 on 9\u2009December\u20092014, for units 1 and 2 on 27\u2009May\u20092015 and for \nunit 3 on 8\u2009June\u2009201654. Cane Run retired its three active units in May 2015, after \nwhich a 650\u2009MW natural gas combined cycle plant began operating at the site. \nRockport installed dry sodium scrubbers on its two coal units in April 2015.\nArea-level confounders of asthma symptoms. We assembled several indicators \nof community socio-economic status and demographic composition at the ZIP-\ncode level from the US American Community Survey (ACS)55. These are variables \npotentially associated with coal-fired power plant exposure that might also predict \nasthma exacerbations56 and included total population, number of non-Hispanic \nblack individuals, number of individuals without health insurance coverage, \nnumber of individuals 16\u2009years and older that were unemployed, number of \nindividuals 25 years and older without a high school diploma or equivalent and \nnumber of individuals with income in the previous 12 months below the federal \npoverty level. To create a time-varying dataset, we linked multiple 5-yr surveys \nbecause annual ZCTA-level estimates were not available. For example, we used \nthe 2008\u20132012 ACS to estimate ZCTA characteristics in 2012 and the 2009\u20132013 \nACS for 2013, which we then assigned to ZIP codes using a crosswalk. Finally, we \ndownloaded ZIP-code-level meteorological data on ambient temperature, relative \nhumidity, windspeed and atmospheric pressure from the USEPA57.\nJefferson County quarterly ZIP-code-level asthma data. The Louisville Metro \nDepartment of Public Health and Wellness provided quarterly combined counts \nof asthma-related hospitalizations and ERVs among all ages for the years 2012\u2013\n2016 for all Louisville ZIP codes, which we restricted to the 35 ZIP codes with \npopulation greater than 5 in 2012. Hospitalizations and ERVs were considered \nasthma related if the following International Classification of Diseases 9 or 10 \ndiagnosis codes were present in the first through to the third diagnosis positions: \n493.XX or J45.X. To present health data spatially, we used a crosswalk to map ZIP \ncodes to ZCTAs.\nParticipant recruitment for individual-level asthma data. Data for the \nindividual-level analyses came from the AIR Louisville pilot (2012\u20132014) and \nfull programme (2015\u20132017)47. Eligible participants (n\u2009=\u20091,021 from pilot and full \nprogramme) were enrolled from 2012 to 2016 through multiple channels, which \nincluded employer partner wellness programmes, clinics, community events and \nsocial media campaigns. Inclusion criteria consisted of a self-reported diagnosis \nof asthma, a current prescription for a compatible inhaled medication for asthma \nand a home or work address in Jefferson County. Additional details on the study \nare given elsewhere47. Participants all agreed to and signed Propeller Health\u2019s \nTerms of Service, which stated that their data may be used in an aggregated \nRR = 0.83, \n95% CI: \n0.69, 1.00 \nRR = 0.98, \n95% CI: 0.95, 1.01 \nPre\nMay 2015\n1\n2\nMonthly average daily SABA use\nAugust 2015\nNovember 2015\nFebruary 2016\nMay 2016\nAugust 2016\nNovember 2016\nFebruary 2017\nMay 2017\nPost\nMill Creek\nscrubber\ninstalled\nFig. 6 | Monthly average daily SABA use before and after the June\u20092016 \nMill Creek SO2 scrubber installation. Orange represents the period prior to \nthe scrubber installation and navy blue that after it. Points are the monthly \naverage daily predicted SABA use from the adjusted model. Relative risks \n(RR) are from a conditional Poisson case interrupted time-series model, \nwhich should be interpreted within individual. The model was adjusted for \ntemperature, humidity, windspeed, ambient pollen (grass, tree and weed), \nmould counts and long-term and seasonal trends. Lagged residuals were \nused to account for autocorrelation.\nNature Energy | VOL 5 | May 2020 | 398\u2013408 | www.nature.com/natureenergy\n404\n\nArticles\nNaTUre Energy\nfashion for public-health-oriented analyses. We also received a waiver of consent \nand exemption (PRH1-17-508) from the Copernicus Group Independent Review \nBoard. The protocol was additionally approved by the University of California \nBerkeley Committee for Protection of Human Subjects.\nWe included participants in the analyses if they had inhaler-use data  \ncollected in the year before and the year after the coal-fired power plant energy \ntransitions. We focused on the 8\u2009June\u20092016 SO2 scrubber installation for Mill \nCreek Unit 3, because this date supported the largest participation cohort for the \nindividual-level analyses (N\u2009=\u2009207, 20.3% of the entire AIR Louisville cohort) \n(Supplementary Fig. 9).\nDigital health platform. All participants received digital sensors (Propeller Health, \nMadison) to attach to their inhaled SABA (that is, \u2018rescue\u2019) medications. The sensor \nand platform comprised a US Food and Drug Administration-cleared digital health \nintervention that combined inhaler sensors, mobile apps, web-based dashboards \nand predictive analytics for patients and clinicians47,58,59. The sensor objectively \nmonitored the use of inhalers, capturing the date, time of day and number of \nactuations, and wirelessly transmitted these data back to secure servers through \na smartphone application or hub base station. Actuations that occurred within \n2\u2009min were grouped into a single-use event, and therefore each use could represent \nmultiple actuations. The sensors regularly transmitted medication use data back \nto the server (\u2018sync\u2019) through the smartphone or hub, and also transmitted a \n\u2018heartbeat\u2019 signal, which reported sensor battery life, confirmed no actuations \nhad occurred since the previous sync and recorded the participant\u2019s GPS (global \npositioning system) location from which we assigned HyADS exposure estimates. \nHeartbeats occurred approximately every 3\u2009h .\nThe majority (98%) of participants entered the cohort in 2015 or 2016 with a \nmean (standard deviation) duration of follow-up of 602 (321) days from first to last \nday under observation. We used the monthly average count of SABA use events per \nperson per day as our outcome of interest.\nIndividual meteorological data and plant-specific exposures. Over 85% of \nSABA-use events and heartbeats had a recorded location. For those SABA events \nwithout a GPS location (about 6%), we retro-filled the location information with \nthe most recent recorded location within 24\u2009h before or after the index event. If \nno location was available, we assigned the participant\u2019s home address location. \nAt the time and place of inhaler use or sensor heartbeat, we assigned hourly \nmeteorological data, which included temperature, wind speed, atmospheric \npressure and relative humidity from the National Oceanic and Atmospheric \nAdministration for the years 2008\u2013201660. These data were averaged first to daily \nand then to monthly means. A specialty clinic, Family Allergy & Asthma, provided \ndaily counts of grass, weed and tree pollen, as well as of mould spores from their \nNational Allergy Bureau-certified pollen-counting station in Louisville. We \nassigned each SABA use or heartbeat event daily pollen or mould counts based \non the event date, with events that occurred on the same day assigned the same \npollen or mould counts. For analyses, we took the monthly mean of the pollen or \nmould count. We assigned monthly HyADS exposures based on the ZIP code in \nwhich the participant spent the most time that month. We linked a measure of \ncommunity social vulnerability in 2014 from the US Centers for Disease Control \nand Prevention to each individual\u2019s census tract of residence61.\nStatistical analysis. We conducted ZIP-code and individual-level analyses to test \nthe hypothesis that coal-fired power plant energy transitions altered air quality \nand thereby asthma morbidity among populations living nearby. In Supplementary \nTable 5, we provide a roadmap for all the analyses, which includes questions, \ndatasets, analyses and locations of results. We performed analyses using R \nStatistical Software version 3.5.1 (R Core Team (2018) https://www.R-project.org/). \nIndividual-level analyses were conducted using the gnm package (https://cran.\nrproject.org/package=gnm). All the tests were two-sided.\nZIP-code level HyADS analysis. We first tested whether the energy transitions \nwere associated with reduced HyADS exposures throughout Louisville. To do so, \nwe used linear regression to fit the equation:\nHyADSpqr \u00bc \u03b20 \u00fe \u03b21Transition1pqr \u00fe \u03b22Transition2pqr \u00fe \u03b23Transition3pqr\n\u00fe\u03b24Transition4pqr \u00fe \u03bbpqr \u00fe Pp \u00fe Qq \u00fe \u03b5pqr\n\u00f01\u00de\nHyADSpqr is the level of HyADS in ZIP code p, quarter q, and year r. \nTransition1pqr\u2013Transition4pqr are indicator variables equal to 1 if the energy \ntransition has occurred and 0 otherwise. For example, Transition1pqr equals 1 after \nQ2-2013 and 0 beforehand, Transition2pqr equals 1 after Q4-2014 and 0 beforehand. \nPp and Qq are indicator variables for the ZIP code and quarter of year, respectively. \n\u03bbpqr is a set of quarterly meteorological variables that include temperature, wind \nspeed, relative humidity and atmospheric pressure. Coefficients \u03b21\u2013\u03b24 can be \ninterpreted as the change in HyADS associated with their respective energy \ntransition; negative values indicate decreased HyADS post-transition.\nZIP-code-level quasi-Poisson asthma analysis. We then used three approaches \nto estimate the relationship between energy transitions and ZIP-code-level asthma \nhospitalizations and ERV counts. The first was a generalized linear model with a \nquasi-Poisson distribution, which accommodates overdispersion, and a log link:\nlog E Asthmapqr\n\u001f\n\u001e\n\u001d\n\u001c\n\u00bc log totpoppr\n\u001b\n\u001a\n\u00fe \u03b20 \u00fe \u03b21Transition1pqr \u00fe \u03b22Transition2pqr\n\u00fe\u03b23Transition3pqr \u00fe \u03b24Transition4pqr \u00fe \u03bbpqr \u00fe \u03b8pr \u00fe Pp \u00fe Qq \u00fe Rr\n\u00f02\u00de\nwhere E(Asthmapqr) represents the expected quarterly asthma hospitalization and \nERV counts, totpoppr is the estimated ZIP-code-level population. \u03b8pr is a set of \ndemographic variables from the ACS that includes percentage non-Hispanic black \nindividuals, percentage individuals living in poverty and percentage unemployed \nindividuals. Rr is the year indicator variable. We included population as an offset.\nZIP-code-level difference-in-differences asthma analysis. A second analysis \nof the natural experiment posed by the Q2-2015 transitions used a quasi-\nexperimental difference-in-differences design to evaluate how asthma outcomes \ndifferentially changed in ZIP codes defined as either \u2018low\u2019 or \u2018high\u2019 HyADS \nexposure based on the average pre-period (Q2-2014 to Q2-2015) HyADS exposure. \nThe post-period spanned the year after Q2-2015, so the analysis covered the time \nperiod that surrounded the installation of the scrubbers on units 1 and 2 of the \nMill Creek plant and both units of the Rockport plant, and the retirement of Cane \nRun\u2019s three units. Difference-in-differences analysis is commonly used in studies of \nnatural experiments and can effectively eliminate both observed and unobserved \nconfounding variables that do not vary in time46,62. We categorized ZIP codes as \n\u2018low\u2019 when their pre-period average HyADS was <32,500 (near the median) and \n\u2018high\u2019 when their pre-period HyADS was \u226532,500, selecting the cutoff based on \nthe change\u2019s distribution (Supplementary Fig. 10). By estimating the change in \nasthma hospitalization and ERVs from the pre- to the post-period (difference 1) \nand subtracting off the difference between the exposed and control exposure ZIP \ncodes (difference 2), we estimated the effect of the Q2-2015 energy transitions in \nJefferson County. We used the parametric equation:\nAsthmapqr \u00bc\n\u03b20 \u00fe \u03b21Hpqr \u00fe \u03b22Cpqr \u00fe \u03b23Hpqr \u00b4 Cpqr \u00fe \u03bbpqr \u00fe \u03b8pr\n\u00fePp \u00fe Qq \u00fe Rr \u00fe \u03b5pqr\n\u00f03\u00de\nwhere Hpqr is an indicator variable equal to one for ZIP codes with pre-period \nHyADS\u2009\u2265\u200932,500 (that is, exposed). The group of ZIP codes with pre-period \nHyADS <32,500 served as the control group because these ZIP codes were less \nexposed in the pre-period and therefore benefitted less from the Q2-2015 energy \ntransitions, but experienced similar secular trends as the high pre-period HyADS \ngroup. Cpqr is an indicator variable equal to one when the quarter is after Q2-2015. \n\u03b23 represents the difference-in-differences estimate of interest. We also specified \na continuous model by comparing ZIP codes with differing levels of pre-period \nHyADS rather than creating a cut point. To do so, we replaced the Hpqr indicator \nvariable with the pre-period continuous HyADS level. We opted to include the \ndescribed set of potential confounding variables in equation (3) based on a\u00a0priori \nhypotheses, causal diagrams63 and the correlation structure of potential covariates \n(Supplementary Fig. 11). For example, annual unemployment and uninsured status \nhad a Spearman correlation of 0.84, so we included only unemployment in our \nmodels to avoid instability in the parameter estimation due to multicollinearity. \nDifference-in-difference analysis relies on the parallel trends assumption, that is, \nin the absence of intervention, trends in the outcome in the pre-period continue62. \nWe visually inspected trends over time by low and high pre-Q2-2015 HyADS and \nfound no evidence of a violation (Fig. 4a).\nZIP-code-level first-difference analysis. In a secondary analysis, we used a \nfirst-difference design64 to evaluate the relationship between the pre-post change \nin HyADS and pre-post change in ZIP-code-level hospitalization and ERVs \n(Supplementary Note 2). This model took the form:\n\u0394Asthmapr \u00bc \u03b20 \u00fe \u03b21\u0394HyADSpr \u00fe \u0394\u03bbpr \u00fe \u0394\u03b8pr \u00fe \u0394\u03b5pr\n\u00f04\u00de\nwhere \u0394Asthmapr represents the differences between counts of ZIP-code-level \nasthma hospitalization and ERVs in the year prior to Q2-2015 and the year after \nQ2-2015 and \u0394HyADSpr represents the ZIP-code-level difference in HyADS \nexposure to the three facilities with energy transitions from the year prior to \nQ2-2015 to the year after Q2-2015. We took differences in the meteorological and \ndemographic variables (\u0394\u03bbpr and \u0394\u03b8pr) by subtracting the average value between \nQ2-2014 and Q1-2015 from the average value between Q3-2015 and Q2-2016. \nThis estimator is unbiased when \u03b5pr is independent of \u0394HyADSpr, \u0394\u03bbpr and \u0394\u03b8pr. \nIn a separate model, we allowed for non-linearity in the association between \n\u0394HyADSpr and \u0394Asthmapr by specifying quintiles of \u0394HyADSpr. In both the \ndifference-in-differences and first-difference models, we accounted for correlation \nwithin the ZIP codes using Liang and Zeger cluster-robust standard errors65. We \nalso completed a sensitivity analysis using baseline population weights rather than \nadjusting equations (3) or (4) for estimated total population.\nIndividual-level case time-series analysis. We used a case time-series analysis that \nincluded 207 participants under observation in the year prior and year after the \nNature Energy | VOL 5 | May 2020 | 398\u2013408 | www.nature.com/natureenergy\n405\n\nArticles\nNaTUre Energy\nJune\u20092016 Mill Creek scrubber installation. The case time-series design is a within-\nperson analysis that allows for the control of individual-level confounders similar \nto other case-only methods such as case-crossover and self-controlled case-series \ndesigns66,67. Unlike the case-crossover design, the case time-series design does not \nrely on the assumptions of risk-set sampling68. It maintains the temporal structure \nof the time-series format, which allows for modelling of trends over time. It can \nalso accommodate count data and account for overdispersion and autocorrelation \nin counts within stratum (that is, the individual). The conditional Poisson model \nassumes no unmeasured confounding that is not homogeneous within the strata. \nWe adjusted for measured time-varying meteorological variables and seasonal \ntrends. The model also assumes no residual autocorrelation of counts (to this end \nwe adjust for the lagged residuals). We also relaxed the assumption of the Poisson \ndistribution to allow for overdispersion of counts and assessed the relationship \nbetween exposure to coal-fired power plant emissions and monthly average daily \nSABA use with the conditional quasi-Poisson model:\nlog E Inhaler useim\n\u00f0\n\u00de\n\u00bd\n\ue08a\u00bc \u03b1i \u00fe \u03b20 \u00fe \u03b21T \u00fe \u03b22HyADSim \u00fe \u03bbim \u00fe sm \u00fe rim\u00191\n\u00f05\u00de\nassuming a quasi-Poisson distribution to account for overdispersion. Inhaler_useim \nis the average daily count of SABA uses for individual i in month m (conditional \non the total number of SABA uses for individual i), \u03b1i is a parameter for individual \nlevel effects (see equation (6) for more detail), T is the number of days elapsed \nsince May 2015, HyADSim is the total HyADS exposure in participant i\u2019s ZIP code \nin month m, \u03bbim is a vector of meteorological variables including natural cubic \nsplines for temperature, relative humidity, wind speed, atmospheric pressure and \nmould counts, and linear terms for ambient pollen concentrations (that is, weed, \ntree and grass) and sm is a harmonic term with 2 sine\u2013cosine pairs and a period of \n12 months to account for seasonal trends. The model adjusts for autocorrelation \nby adding the residuals rim\u20131 for individual i in month m\u2009\u2013\u20091, which were estimated \nas the residuals of a model fitted as in equation (5) without the residual term and \nlagged by one month66,69. The parameters \u03b1i are not estimated by the model, but are \nrather conditioned out, by conditioning on the sum of the total number of inhaler \nuses for each individual i, P\ni\nInhaler useim\nI\n, which results in a multinomial model \nwith the likelihood:\nInhaler useimj P\nm\nInhaler useim \ue019Multinomial \u03c0m\n\u00f0\n\u00de\n\u03c0m \u00bc exp\u00f0\u03b20\u00fe\u03b21T\u00fe\u03b22HyADSim\u00fe\u03bbim\u00fesm\u00de\nP\njexp\u00f0\u03b20\u00fe\u03b21T\u00fe\u03b22HyADSij\u00fe\u03bbij\u00fesj\u00de\n\u00f06\u00de\nwhere j \u2208 the subset of m with observations for each participant i. Next, we sought \nto directly test the relationship between the June\u20092016 Mill Creek energy transition \nand SABA use. Unlike the Q2-2015 energy transitions, the June\u20092016 Mill Creek \nunit 3 scrubber installation resulted in fairly uniform reductions in exposure to \nplant emissions across Jefferson County (Fig. 4b). This precluded the assembly of \na control group based on exposure change analogous to that in the ZIP-code-level \ndifference-in-differences analysis. Instead, we implemented an interrupted time-\nseries design, which construes each person\u2019s study time during the pre-installation \nperiod as a control for his or her time during the post-installation period. To do so, \nwe modified equation (5) by replacing HyADSim with a single indicator variable, \ncntrlim, of exposure to the SO2 control installation at Mill Creek in June\u20092016, which \ntook the value of 1 beginning in May 2016 (Mill Creek unit 3 was turned off in May \nfor the June control installation) and 0 otherwise. We also added an interaction \nbetween cntrlim and T to estimate trends in changing SABA use over time. A \nnegative coefficient for the cntrlim variable indicates that the average number of \nrescue inhaler uses decreased after the scrubber installation. Likewise, a negative \u03b2 \non cntrlimT indicates a downward linear trend in rescue inhaler use over time after \nthe installation.\nTo assess the sensitivity of the individual-level analysis, we evaluated the \nassociation between the June\u20092016 Mill Creek event and two different binary \nspecifications of SABA use: any use and high use. We followed a similar model \nspecification to that in equation (5), but used a logistic regression model and \nsubstituted for the Inhaler useim variable an any-SABA-use variable (monthly \naverage of \u22651\u2009use\u2009day\u20131 versus <1\u2009use\u2009day\u20131) and a high-SABA-use variable (monthly \naverage of \u22654\u2009uses\u2009day\u20131 versus <4 uses day\u20131).\nReporting summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nThe ZIP-code-level asthma hospitalization and ERV data are available from the \nauthors following the submission of an analysis proposal and written approval \ngranted by the Louisville Metro Public Health and Wellness. The AIR Louisville \nmonthly medication use data are considered Protected Health Information under \nthe Health Insurance Portability and Accountability Act of 1996 (HIPAA) in the \nUnited States, and as such may be accessible from the authors for analysis only \nafter specific written authorization of access following HIPAA guidelines and \nInstitutional Review Board approval. We provide Jefferson County ZIP-code-\nlevel monthly HyADS estimates on GitHub at https://github.com/joanacasey/\nky_asthma_coal.\nCode availability\nAn R package is available on GitHub for running the HyADS model (https://github.\ncom/lhenneman/disperseR). We also provide analysis code on GitHub at https://\ngithub.com/joanacasey/ky_asthma_coal.\nReceived: 11 May 2019; Accepted: 6 March 2020;  \nPublished online: 13 April 2020\nReferences\n\t1.\t IEA Statistics (International Energy Agency, 2018); http://www.iea.org/\nstatistics/\n\t2.\t Massetti, E. et\u00a0al. Environmental Quality and the US Power Sector: Air Quality, \nWater Quality, Land Use and Environmental Justice (Oak Ridge National \nLaboratory, 2017); https://www.energy.gov/sites/prod/files/2017/01/f34/\nEnvironment%20Baseline%20Vol.%202\u2013Environmental%20Quality%20\nand%20the%20U.S.%20Power%20Sector\u2013Air%20Quality%2C%20Water%20\nQuality%2C%20Land%20Use%2C%20and%20Environmental%20Justice.pdf\n\t3.\t Toxic Air: The Case for Cleaning Up Coal-Fired Power Plants (American Lung \nAssociation, 2011); http://www.lung.org/assets/documents/healthy-air/\ntoxic-air-report.pdf\n\t4.\t Zheng, X.-Y et\u00a0al. Association between air pollutants and asthma emergency \nroom visits and hospital admissions in time series studies: a systematic review \nand meta-analysis. PLoS ONE 10, e0138146 (2015).\n\t5.\t Orellano, P., Quaranta, N., Reynoso, J., Balbi, B. & Vasquez, J. Effect of outdoor \nair pollution on asthma exacerbations in children and adults: systematic review \nand multilevel meta-analysis. PLoS ONE 12, e0174050 (2017).\n\t6.\t Williams, A. M., Phaneuf, D. J., Barrett, M. A. & Su, J. G. Short-term impact \nof PM2.5 on contemporaneous asthma medication use: behavior and the value \nof pollution reductions. Proc. Natl Acad. Sci. USA 116, 5246\u20135253 (2019).\n\t7.\t Integrated Science Assessment for Oxides of Nitrogen-Health Criteria (US \nEnvironmental Protection Agency, 2016); http://ofmpub.epa.gov/eims/\neimscomm.getfile?p_download_id=526855\n\t8.\t Integrated Review Plan for the Primary National Ambient Air Quality Standard \nfor Sulfur Dioxide EPA-452/R-14-007 (US Environmental Protection Agency, \n2014); https://www3.epa.gov/ttn/naaqs/standards/so2/\ndata/20141028so2reviewplan.pdf\n\t9.\t Ramadour, M. et\u00a0al. Prevalence of asthma and rhinitis in relation to \nlong-term exposure to gaseous air pollutants. Allergy 55, 1163\u20131169 (2000).\n\t10.\tGuarnieri, M. & Balmes, J. R. Outdoor air pollution and asthma. Lancet 383, \n1581\u20131592 (2014).\n\t11.\tDeger, L. et\u00a0al. Active and uncontrolled asthma among children exposed to \nair stack emissions of sulphur dioxide from petroleum refineries in Montreal, \nQuebec: a cross-sectional study. Can. Respir. J. 19, 97\u2013102 (2012).\n\t12.\tCharpin, D. et\u00a0al. Respiratory symptoms and air pollution changes in children: \nthe Gardanne Coal-Basin Study. Arch. Environ. Health 43, 22\u201327 (1988).\n\t13.\tSchenker, M. B., Speizer, F. E., Samet, J. M., Gruhl, J. & Batterman, S. Health \neffects of air pollution due to coal combustion in the Chestnut Ridge Region \nof Pennsylvania: results of cross-sectional analysis in adults. Arch. Environ. \nHealth 38, 325\u2013330 (1983).\n\t14.\tDubnov, J. et\u00a0al. Estimating the effect of air pollution from a coal-fired power \nstation on the development of children\u2019s pulmonary function. Environ. Res. \n103, 87\u201398 (2007).\n\t15.\tCohen, A. A., Bromberg, S., Buechley, R. W., Heiderscheit, L. T. & Shy, C. M. \nAsthma and air pollution from a coal-fueled power plant. Am. J. Public \nHealth 62, 1181\u20131188 (1972).\n\t16.\tSmargiassi, A. et\u00a0al. Risk of asthmatic episodes in children exposed to sulfur \ndioxide stack emissions from a refinery point source in Montreal, Canada. \nEnviron. Health Perspect. 117, 653 (2009).\n\t17.\tRodriguez-Villamizar, L. A., Rosychuk, R. J., Osornio-Vargas, A., Villeneuve, \nP. J. & Rowe, B. H. Proximity to two main sources of industrial outdoor air \npollution and emergency department visits for childhood asthma in \nEdmonton, Canada. Can. J. Public Health 108, e523\u2013e529 (2017).\n\t18.\tMiddleton, N., Kolokotroni, O., Lamnisos, D., Koutrakis, P. & Yiallouros, P. \nK. Prevalence of asthma and respiratory symptoms in 15\u201317-year-old \nGreek-Cypriots by proximity of their community of residence to power \nplants: Cyprus 2006\u201307. Public Health 128, 288\u2013296 (2014).\n\t19.\tLiu, X., Lessner, L. & Carpenter, D. O. Association between residential \nproximity to fuel-fired power plants and hospitalization rate for respiratory \ndiseases. Environ. Health Perspect. 120, 807\u2013810 (2012).\n\t20.\tForm EIA-860 (US Energy Information Administration, 2019); https://www.\neia.gov/electricity/data/eia860/\n\t21.\tHumphreys, D. K., Panter, J., Sahlqvist, S., Goodman, A. & Ogilvie, D. \nChanging the environment to improve population health: a framework for \nconsidering exposure in natural experimental studies. J. Epidemiol. \nCommunity Health 70, 941\u2013946 (2016).\n\t22.\tRich, D. Q. Accountability studies of air pollution and health effects: lessons \nlearned and recommendations for future natural experiment opportunities. \nEnviron. Int. 100, 62\u201378 (2017).\nNature Energy | VOL 5 | May 2020 | 398\u2013408 | www.nature.com/natureenergy\n406\n\nArticles\nNaTUre Energy\n\t23.\tZigler, C. M. & Dominici, F. Point: clarifying policy evidence with potential-\noutcomes thinking\u2014beyond exposure\u2013response estimation in air pollution \nepidemiology. Am. J. Epidemiol. 180, 1133\u20131140 (2014).\n\t24.\tPope, C. A. 3rd Respiratory disease associated with community air pollution \nand a steel mill, Utah Valley. Am. J. Public Health 79, 623\u2013628 (1989).\n\t25.\tClancy, L., Goodman, P., Sinclair, H. & Dockery, D. W. Effect of air-pollution \ncontrol on death rates in Dublin, Ireland: an intervention study. Lancet 360, \n1210\u20131214 (2002).\n\t26.\tFriedman, M. S., Powell, K. E., Hutwagner, L., Graham, L. M. & Teague, W. \nG. Impact of changes in transportation and commuting behaviors during the \n1996 Summer Olympic Games in Atlanta on air quality and childhood \nasthma. J. Am. Med. Assoc. 285, 897\u2013905 (2001).\n\t27.\tLi, Y., Wang, W., Kan, H., Xu, X. & Chen, B. Air quality and outpatient visits \nfor asthma in adults during the 2008 Summer Olympic Games in Beijing.  \nSci. Total. Environ. 408, 1226\u20131227 (2010).\n\t28.\tDeschenes, O., Greenstone, M. & Shapiro, J. S. Defensive investments and the \ndemand for air quality: evidence from the NOx budget program. Am. Econ. \nRev. 107, 2958\u20132989 (2017).\n\t29.\tNational Emissions Inventory (US Environmental Protection Agency, 2011); \nhttps://www.epa.gov/air-emissions-inventories/2011-national-emissions-\ninventory-nei-data\n\t30.\tHenneman, L. R. F., Choirat, C., Ivey, C. E., Cummiskey, K. & Zigler, C. M. \nCharacterizing population exposure to coal emissions sources in the United \nStates using the HyADS model. Atmos. Environ. 203, 271\u2013280 (2019).\n\t31.\tHenneman, L. R. F., Mickley, L. J. & Zigler, C. M. Air pollution accountability \nof energy transitions: the relative importance of wind fields and emissions in \nexposure changes. Environ. Res. Lett. 14, 115003 (2019).\n\t32.\tZein, J. G. et\u00a0al. Impact of age and sex on outcomes and hospital cost of acute \nasthma in the United States, 2011-2012. PLoS ONE 11, e0157301 (2016).\n\t33.\tPatel, M. et\u00a0al. Metrics of salbutamol use as predictors of future adverse \noutcomes in asthma. Clin. Exp. Allergy 43, 1144\u20131151 (2013).\n\t34.\tPrieto-Parra, L. et\u00a0al. Air pollution, PM2.5 composition, source factors, and \nrespiratory symptoms in asthmatic and nonasthmatic children in Santiago, \nChile. Environ. Int 101, 190\u2013200 (2017).\n\t35.\tSchildcrout, J. S. et\u00a0al. Ambient air pollution and asthma exacerbations in \nchildren: an eight-city analysis. Am. J. Epidemiol. 164, 505\u2013517 (2006).\n\t36.\tU.S. Environmental Protection Agency. Federal Registrar. Primary National \nAmbient Air Quality Standard for Sulfur Dioxide, Final Rule, 40 CFR Parts 50, \n53, and 58., <https://www.govinfo.gov/content/pkg/FR-2010-06-22/\npdf/2010-13947.pdf> (2010).\n\t37.\tGowers, A. M. et\u00a0al. Does outdoor air pollution induce new cases of asthma? \nBiological plausibility and evidence; a review. Respirology 17, 887\u2013898 (2012).\n\t38.\tJohns, D. O. & Linn, W. S. A review of controlled human SO2 exposure \nstudies contributing to the US EPA integrated science assessment for sulfur \noxides. Inhal. Toxicol. 23, 33\u201343 (2011).\n\t39.\tLi, R. et\u00a0al. Effect of sulfur dioxide on inflammatory and immune regulation \nin asthmatic rats. Chemosphere 112, 296\u2013304 (2014).\n\t40.\tAmster, E. D., Haim, M., Dubnov, J. & Broday, D. M. Contribution of \nnitrogen oxide and sulfur dioxide exposure from power plant emissions on \nrespiratory symptom and disease prevalence. Envrion. Pollut. 186,  \n20\u201328 (2014).\n\t41.\tAnenberg, S. C. et\u00a0al. Estimates of the global burden of ambient PM2.5, ozone, \nand NO2 on asthma incidence and emergency room visits. Environ. Health \nPerspect. 126, 107004 (2018).\n\t42.\tAchakulwisut, P., Brauer, M., Hystad, P. & Anenberg, S. C. Global, national, \nand urban burdens of paediatric asthma incidence attributable to ambient \nNO2 pollution: estimates from global datasets. Lancet Planet Health 3, \ne166\u2013e178 (2019).\n\t43.\tDominici, F., Greenstone, M. & Sunstein, C. R. Particulate matter matters. \nScience 344, 257\u2013259 (2014).\n\t44.\tCushing, L., Morello-Frosch, R., Wander, M. & Pastor, M. The haves, the \nhave-nots, and the health of everyone: the relationship between social inequality \nand environmental quality. Annu. Rev. Public Health 36, 193\u2013209 (2015).\n\t45.\tBaltrus, P. et\u00a0al. Individual and county level predictors of asthma related \nemergency department visits among children on Medicaid: a multilevel \napproach. J. Asthma 54, 53\u201361 (2017).\n\t46.\tCraig, P., Katikireddi, S. V., Leyland, A. & Popham, F. Natural experiments: \nan overview of methods, approaches, and contributions to public health \nintervention research. Annu. Rev. Public Health 38, 39\u201356 (2017).\n\t47.\tBarrett, M. et\u00a0al. AIR Louisville: addressing asthma with technology, \ncrowdsourcing, cross-sector collaboration, and policy. Health Affairs 37, \n525\u2013534 (2018).\n\t48.\tAnderson, H. R., Favarato, G. & Atkinson, R. W. Long-term exposure to \noutdoor air pollution and the prevalence of asthma: meta-analysis of \nmulti-community prevalence studies. Air Qual. Atmos. Health 6,  \n57\u201368 (2013).\n\t49.\t2016 National Health Interview Survey (NHIS) Data (US Centers for  \nDisease Control & Prevention, 2016); https://www.cdc.gov/asthma/nhis/2016/\ntable3-1.htm\n\t50.\tAir Markets Program Data (US Environmental Protection Agency, 2018); \nhttps://ampd.epa.gov/ampd/\n\t51.\tStein, A. et\u00a0al. NOAA\u2019s HYSPLIT atmospheric transport and dispersion \nmodeling system. Bull. Am. Meteorol. Soc. 96, 2059\u20132077 (2015).\n\t52.\tHenneman, L. R. F., Dedoussi, I. C., Casey, J. A., Choirat, C. & Zigler, C. M. \nComparisons of simple and complex methods for quantifying exposure to \npoint source air pollution emissions. J. Expo. Sci. Environ. Epidemiol.  \nhttps://doi.org/10.1038/s41370-020-0219-1 (2020).\n\t53.\tHenneman, L. R. F., Choirat, C. & Zigler, C. M. Accountability assessment of \nhealth improvements in the United States associated with reduced coal \nemissions between 2005 and 2012. Epidemiology 30, 477\u2013485 (2019).\n\t54.\tMill Creek Station Wins 2016 Project of the Year Award (Louisville Gas & \nElectric, 2017); https://lge-ku.com/newsroom/articles/2017/01/09/mill-creek-\nstation-wins-2016-project-year-award\n\t55.\tManson, S., Schroeder, J., Riper, D. V. & Ruggles, S. IPUMS National \nHistorical Geographic Information System: Version 13.0 (IPUMS, 2018); \nhttps://doi.org/10.18128/D050.V12.0\n\t56.\tGottlieb, D. J., Beiser, A. S. & O\u2019Connor, G. T. Poverty, race, and medication \nuse are correlates of asthma hospitalization rates: a small area analysis in \nBoston. Chest 108, 28\u201335 (1995).\n\t57.\tAir Data Pre-generated Data Files (US Environmental Protection Agency) \nhttps://aqs.epa.gov/aqsweb/airdata/download_files.html\n\t58.\tMerchant, R. K., Inamdar, R. & Quade, R. C. Effectiveness of population \nhealth management using the propeller health asthma platform: a \nrandomized clinical trial. J. Allergy Clin. Immunol. Pract. 4,  \n455\u2013463 (2016).\n\t59.\tVan Sickle, D., Magzamen, S., Truelove, S. & Morrison, T. Remote monitoring \nof inhaled bronchodilator use and weekly feedback about asthma \nmanagement: an open-group, short-term pilot study of the impact on asthma \ncontrol. PLoS ONE 8, e55335 (2013).\n\t60.\tNational Centers for Environmental Information, Weather and Climate Data \n(National Oceanic and Atmospheric Administration, accessed 13 February \n2017); https://www.ncdc.noaa.gov/orders/qclcd/\n\t61.\tCDC\u2019s Social Vulnerability Index (SVI) (US Centers for Disease Control & \nPrevention, 2016); https://svi.cdc.gov/data-and-tools-download.html\n\t62.\tAngrist, J. D. & Pischke, J.-S. Mostly Harmless Econometrics: An Empiricist\u2019s \nCompanion (Princeton Univ. Press, 2009).\n\t63.\tGlymour, M. & Greenland, S. in Modern Epidemiology 3rd edn (eds \nRothman, K. J., Greenland, S. & Lash, T. L.) 183\u2013209 (Lippincott Williams \nand Wilkins, 2008).\n\t64.\tPope, C. A. 3rd, Ezzati, M. & Dockery, D. W. Fine-particulate air  \npollution and life expectancy in the United States. N. Engl. J. Med. 360, \n376\u2013386 (2009).\n\t65.\tAbadie, A., Athey, S., Imbens, G. W. & Wooldridge, J. When Should You \nAdjust Standard Errors for Clustering? (NBER, 2017); https://www.nber.org/\npapers/w24003\n\t66.\tArmstrong, B. G., Gasparrini, A. & Tobias, A. Conditional Poisson models: a \nflexible alternative to conditional logistic case cross-over analysis. BMC Med. \nRes. Methodol. 14, 122 (2014).\n\t67.\tBernal, J. L., Cummins, S. & Gasparrini, A. Interrupted time series regression \nfor the evaluation of public health interventions: a tutorial. Int. J. Epidemiol. \n46, 348\u2013355 (2017).\n\t68.\tPetersen, I., Douglas, I. & Whitaker, H. Self controlled case series methods: \nan alternative to standard epidemiological study designs. Br. Med. J. 354, \ni4515 (2016).\n\t69.\tBrumback, B. A. et\u00a0al. Transitional regression models, with application to \nenvironmental time series. J. Am. Stat. Assoc. 95, 16\u201327 (2000).\nAcknowledgements\nWe acknowledge the network of local partners that made the AIR Louisville program \npossible, which include the Center for Healthy Air, Water and Soil, Louisville Metro, \nthe Community Foundation of Louisville and all the AIR Louisville participants. \nPartners within the Louisville Metro Government include G. Fischer, the Office of \nCivic Innovation, Louisville Metro Department of Public Health and Wellness, the \nOffice of Sustainability, the Office of Advanced Planning, the Louisville Jefferson \nCounty Information Consortium and Louisville Forward. Specifically, K. Talley, \nM. King and R. Hamilton of the Air Pollution Control District provided critical \ninformation and review of this manuscript. We also thank P. Tarini and O. W\u00f3jcik \nat the Robert Wood Johnson Foundation for helpful guidance throughout the \nproject. The main funding for the project was provided by the Robert Wood Johnson \nFoundation. Support was also provided by the Foundation for a Healthy Kentucky, \nNorton Healthcare Foundation, Owsley Brown Charitable Foundation, the American \nLung Association, the National Institute of Environmental Health Sciences  \n(J.A.C., K99/R00 ES027023; A.M.N, K99/R00 ES027511; C.Z., R01 ES026217)  \nand the USEPA (C.Z., EPA 83587201). The contents of this work are solely the \nresponsibility of the grantee and do not necessarily represent the official views  \nof the USEPA or the Robert Wood Johnson Foundation. Further, the USEPA  \ndoes not endorse the purchase of any commercial products or services mentioned  \nin the publication.\nNature Energy | VOL 5 | May 2020 | 398\u2013408 | www.nature.com/natureenergy\n407\n\nArticles\nNaTUre Energy\nAuthor contributions\nJ.A.C., T.S. and M.A.B. secured funding for the study. J.A.C., J.G.S., L.R.F.H., C.Z., \nA.M.N., R.C., S.S.M., Y.-T.C., J.S. and M.A.B. designed the study and provided exposure \nand outcome data. J.A.C. and A.M.N. carried out the statistical analyses. J.A.C., J.G.S., \nL.R.F.H., C.Z., A.M.N., R.C., R.G., Y.-T.C., L.K., S.S.M., V.C., G.S., T.S., J.S. and M.A.B. \nreviewed and critically revised the manuscript.\nCompeting interests\nM.A.B., R.G. and L.K. are salaried employees of Propeller Health and J.G.S. receives \nlimited funding from Propeller Health to conduct analyses.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-020-0600-2.\nCorrespondence and requests for materials should be addressed to J.A.C.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2020\nNature Energy | VOL 5 | May 2020 | 398\u2013408 | www.nature.com/natureenergy\n408\n\n1\nnature research  |  reporting summary\nOctober 2018\nCorresponding author(s):\nJoan A. Casey\nLast updated by author(s): Feb 12, 2020\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nAsthma data were provided by the Louisville Metro Public Health and Wellness and Propeller Health. \nInputs for the HyADS model were downloaded from US Environmental Protection Agency's Air Markets Program Database. \nAn R package is available on Github for running the HyADS model (https://github.com/lhenneman/hyspdisp). \nData analysis\nWe used R statistical software and provide code for the main analyses on GitHub at https://github.com/joanacasey/ky_asthma_coal.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. \nWe strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nThe ZIP code-level asthma hospitalization/ED visit data are available following approval by the Louisville Metro Public Health and Wellness from the authors.  \nThe AIR Louisville monthly medication use data are considered Protected Health Information (PHI) under the Health Insurance Portability and Accountability Act of \n1996 (HIPAA) in the U.S., and as such may be accessible from the authors for analysis only after specific written authorization of access following HIPAA guidelines \nand IRB approval.  \nWe provide Jefferson County ZIP code-level monthly HyADS estimates on GitHub at https://github.com/joanacasey/ky_asthma_coal.  \nOther environmental data are publicly-available for download or available from the authors upon request. \n\n2\nnature research  |  reporting summary\nOctober 2018\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThis study uses two types of data: ZIP code level asthma hospitalization/emergency department quarterly counts from 2012-2016 and \nindividual-level daily rescue inhaler use (aggregated to monthly averages) data from the AIR Louisville Cohort.   \nData are longitudinal and quantitative.\nResearch sample\nZIP code-level data: The Louisville Metro Department of Public Health and Wellness provided quarterly ZIP code-level combined counts \nof asthma-related hospitalizations and ED visits for the years 2012\u20132016. Hospitalizations and ED visits were considered asthma-related if \nthe following International Classification of Diseases 9 (ICD-9) or ICD-10 diagnosis codes were present in diagnosis positions 1\u20133: 493.XX \nor J45.X. The ZIP code-level data includes all asthma-related hospitalizations and ED visits in Jefferson county over the study period. The \nmedian (IQR) ZIP population in 2012 was 21,500 people (13727, 30352).  \n \nIndividual-level data: Our analysis included 207 participants. The average age was 44.7 years (SD = 17.6) and 67% of participants were \nwomen. We used this study sample because these individuals were under observation during the year before and after the June 2016 \nMill Creek sulfur dioxide scrubber installation. Participants were enrolled from 2012 to 2016 through a number of channels, including \nemployer partner wellness programs, clinics, community events, and social media campaigns. Inclusion criteria consisted of a self-\nreported diagnosis of asthma, a current prescription for a compatible inhaled medication for asthma, and living or working in Jefferson \nCounty, Kentucky. No incentives were provided for participation, but participants could benefit from no-cost use of the digital health \nintervention and contribute to a community program. The full AIR Louisville cohort contained 1039 individuals, mean age 34.9 years (SD \n= 17.6) and also primarily female (n = 659 [63%]). Therefore, our subsample was slightly older but with a similar sex-distribution \ncompared to the full cohort.   \nSampling strategy\nZIP code-level: we obtained all hospitalization/ED visits during the study period. \n \nIndividual-level: we used the portion of the existing AIR Louisville cohort that had data overlapping the relevant time period (n = 207). \nThe original cohort (n = 1039 participants) made efforts to achieve geographic, socioeconomic and demographic representation similar \nto that of Jefferson County. \nData collection\nZIP code-level: The Louisville Metro Department of Public Health and Wellness provided the asthma data collected from local hospitals \nbased on diagnostic codes. \n \nIndividual-level: Eligible participants were enrolled in the AIR Louisville pilot program from June 2012\u2013December 2016. Basic \ndemographic information (e.g., age, sex) was collected. All participants received digital sensors (Propeller Health, Madison, WI) to attach \nto their inhaled medications for asthma, including short-acting beta agonist (SABA, i.e., \u201crescue\u201d) medications. The sensor and platform \ncomprised a U.S. Food and Drug Administration-cleared digital health intervention that combined inhaler sensors, mobile apps, web-\nbased dashboards and predictive analytics for patients and clinicians. The sensor monitored the use of inhalers, capturing the date, time \nof day, and number of actuations, and wirelessly transmitted these data back to secure servers through a smartphone application or hub \nbase station. Actuations occurring within 2 minutes were grouped into a single use event, therefore each use could represent multiple \nactuations. The sensors regularly transmitted medication use data back to the server (\u201csync\u201d) through the smartphone or hub, and also \ntransmitted a \u201cheartbeat\u201d signal, which reported sensor battery life and confirmed no actuations had occurred since the last sync. \nHeartbeats occurred approximately every 3 hours. Over 85% of SABA use events and heartbeats had a recorded location. For those SABA \nevents missing a GPS location (about 6%), we retro-filled location information with the most recent recorded location within 24 hours \nbefore or after the index event. If no location was available, we assigned the participant\u2019s home address location. Researchers involved in \nthis data collection were not aware of the current study question/hypotheses.\nTiming\nZIP code-level: collection of hospitalization/ED visit data spanned January 1, 2012 to December 31, 2016. \n \nIndividual-level: Eligible participants were enrolled in the AIR Louisville pilot program from June 2012\u2013December 2016. For the present \nanalysis, we only included individuals (n = 207 [20% of the original cohort]) that had observations in the year prior (June 2015-June 2015) \nand year following (June 2016-June 2017) the June 2016 Mill Creek scrubber installation. \nData exclusions\nZIP code-level: no data were excluded. \n \nIndividual-level: we excluded participants from the AIR Louisville cohort who did not have observations in the year before and after the \nJune 2016 Mill Creek scrubber installation. This was determined a priori.\nNon-participation\nThe AIR Louisville cohort used open enrollment between June 2012-December 2016. For the overall cohort, the median follow-up was \n436 days (IQR: 188-628). In our subset (n = 207), median follow-up was 549 days (409-740). 81% of our subset had follow-up through the \nlast quarter of 2016. Our within-person analysis should help prevent extrapolation beyond the data (since some individuals were lost to \nfollow-up prior to Q4-2016).\n\n3\nnature research  |  reporting summary\nOctober 2018\nRandomization\nParticipants were not randomized.\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology\nAnimals and other organisms\nHuman research participants\nClinical data\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nHuman research participants\nPolicy information about studies involving human research participants\nPopulation characteristics\nSee above.\nRecruitment\nZIP code-level: no recruitment \n \nIndividual-level: Eligible participants were enrolled in the AIR Louisville pilot program from 2012\u20132016 through multiple \nchannels, including employer partner wellness programs, clinics, community events, and social media campaigns. Inclusion \ncriteria consisted of a self-reported diagnosis of asthma, a current prescription for a compatible inhaled medication for asthma, \nand a home or work address in Jefferson County, Kentucky. No incentives were provided for participation, but participants could \nbenefit from no-cost use of the digital health intervention and contribute to a community program. All participants agreed to \nand signed Propeller Health\u2019s Terms of Service, which clearly states that their data may be used in an aggregated fashion for \npublic health-oriented analyses. We also received a waiver of consent from the Copernicus Review Board.  \n \nAn attempt was made via recruitment to achieve geographic, socioeconomic and demographic representation similar to that of \nJefferson County. Participants who enrolled via social media campaigns may have been self-selecting as more interested in \ntechnology than other enrollees.\nEthics oversight\nThe protocol was approved by the Copernicus Independent Review Board and the UC Berkeley Committee for Protection of \nHuman Subjects. \nNote that full information on the approval of the study protocol must also be provided in the manuscript.\n\n\n    Scientific Research Findings:", "answer": "Our study took advantage of a natural experiment that occurred between 2013 and 2016, when four coal-fired power plants in the Louisville, Kentucky area retired, transitioned to natural gas or installed SO2emissions controls. We found that these coal-fired power plant changes translated into reduced asthma exacerbations among the local community. Among ZIP codes more exposed to coal-fired power plant emissions, there were three fewer hospitalizations or emergency department visits per ZIP code per quarter in the year following a major transition, which translates into nearly 400 avoided hospitalizations and ER visits each year across Jefferson County, where Louisville is located. At the level of individuals, a 2016 emissions control installation was associated with a 17% reduction in overall counts of rescue inhaler use in the following month, and a 32% reduction in the odds of having high rescue inhaler use (defined as 4 puffs per day on average per month).", "id": 23}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-020-0589-6\n1Marron Institute of Urban Management and Center for Urban Science & Progress, New York University, New York, NY, USA. 2Guarini Center on \nEnvironmental, Energy and Land Use Law, NYU School of Law, New York, NY, USA. 3Department of Civil and Urban Engineering and Center for Urban \nScience & Progress, New York University, New York, NY, USA. \u2709e-mail: ckontokosta@nyu.edu\nE\nnergy use in buildings accounts for a significant proportion \nof urban greenhouse gas (GHG) emissions, particularly in \nhigh-density cities1. For example, New York City\u2019s (NYC\u2019s) \nmost recent carbon inventory estimates that building energy use \nis responsible for approximately 67% of citywide emissions2. Given \nthe substantial contribution of the built environment to global GHG \nemissions, city policymakers have made increasing building energy \nefficiency a central component of long-term sustainability goals.\nAs new construction represents a small fraction of the building \nstock of a given city in any year, city energy policies are increasingly \nfocused on ways to improve the efficiency of existing buildings3. \nInformational energy regulations, which are premised on the idea \nthat an absence of data and transparency causes suboptimal invest-\nment in energy efficiency4, have become popular policy instruments \nto encourage market-based interventions for energy use reductions. \nMore than 20 cities in the United States, including Austin, Chicago \nand San Francisco, have adopted energy informational policies in \nrecent years, and the pace of adoption continues to increase5.\nNYC has been a leader in this regard, and has implemented two \ninformational regulations in its efforts to reduce building energy \nuse and cut GHG emissions. The first, set forth in Local Law 84 of \n2009 (LL84), requires property owners of large buildings to release \nannual energy consumption data used to benchmark building \nenergy performance. The second, known as Local Law 87 (LL87), \nintroduced a mandatory energy audit requirement for buildings \nlarger than 50,000\u2009ft2. Each covered property must conduct an audit, \nalso referred to as an Energy Efficiency Report, once every 10 years \nand report its findings, which include detailed energy end-use infor-\nmation and recommended energy conservation measures (ECMs). \nRoughly 10% of regulated buildings have been required to conduct \nan audit each year since 2013, and annual deadlines are randomly \nassigned based on the last digit of the property\u2019s Borough-Block-Lot \n(BBL) tax parcel identifier. LL87 also requires owners to implement \ncertain retrocommissioning measures to \u2018tune-up\u2019 existing systems \nat the time of audit, such as to ensure that light fixtures are clean and \nwater pumps are operating as designed6.\nEarly studies indicate that energy disclosure is correlated with \nmeaningful reductions in building energy use7\u20139, although recent \nevidence suggests that there are differential impacts of the policy10. \nYet few, if any, studies have examined the effect of a mandatory \naudit policy on energy use in commercial and multifamily resi-\ndential buildings. This is an important omission: if information \ngained from the required energy audits leads property owners to \ninvest in energy efficiency improvements that they would not other-\nwise implement, policymakers may be justified in expanding audit \nrequirements to a broader range of properties and other cities may \nconsider these requirements in their carbon-reduction-policy tool-\nkit. A positive effect of mandatory audits on energy efficiency may \nalso provide support for requiring more rigorous audits, including \nthe consideration of \u2018deep energy retrofits\u2019, which will be necessary \nto achieve citywide 80%, and greater, carbon emissions reduction \ntargets. However, if mandatory audits\u2014which can be quite costly \nand time-consuming for property owners, especially of smaller \nbuildings11\u2014do not meaningfully influence behaviour or invest-\nment decisions, policymakers may do well to simply adopt an \nenergy performance standard and allow the market to determine \nthe optimal way to meet the requirement. Notably, in April 2019 \nNYC passed a law that caps the amount of fossil-fuel based energy \nthat large buildings can consume without paying a fine12. This law, \nwhich effectively established performance standards for the build-\nings it covers, may obviate the need for additional mandates.\nThis study seeks to inform urban energy policy decisions by \ncomparing energy use in properties that have performed a manda-\ntory energy audit with those that have not. Specifically, we anal-\nysed energy benchmarking data collected by NYC annually under \nLL84 from 2011 to 2016 to investigate whether properties that \nconducted an audit exhibited greater average reductions in energy \nuse than similar, non-audited properties. We collected detailed \naudit report data from mandatory LL87 audits conducted in 2013 \nand 2014 through a randomly assigned allocation process, which \nresulted in an integrated sample of 3,981 buildings. The analy-\nsis examines two primary building types (office and multifamily  \nThe impact of mandatory energy audits on \nbuilding energy use\nConstantine E. Kontokosta\u200a \u200a1\u2009\u2709, Danielle Spiegel-Feld2 and Sokratis Papadopoulos3\nCities are increasingly adopting energy policies that reduce information asymmetries and knowledge gaps through data trans-\nparency, including energy disclosure and mandatory audit requirements for existing buildings. Although such audits impose \nnon-trivial costs on building owners, their energy use impacts have not been empirically evaluated. Here we examine the effect \nof a large-scale mandatory audit policy\u2014New York City\u2019s Local Law 87\u2014on building energy use, using detailed audit and energy \ndata between 2011 and 2016 for approximately 4,000 buildings. This specific policy context, in which the compliance year is \nrandomly assigned, provides a unique opportunity to explore the audit effect without the self-selection bias found in studies \nof voluntary audit policies. We find energy use reductions of approximately \u20132.5% for multifamily residential buildings and \n\u20134.9% for office buildings. The results suggest that mandatory audits, by themselves, create an insufficient incentive to invest \nin energy efficiency at the scale needed to meet citywide carbon-reduction goals.\nNature Energy | VOL 5 | April 2020 | 309\u2013316 | www.nature.com/natureenergy\n309\n\nArticles\nNATure EnergY\nhousing), while it controls for multiple time-invariant and time-\nvarying attributes to evaluate whether energy audits have dif-\nferential impacts across market segments. We also attempt to \ncontextualize the audit effect by disaggregating the potential \nimpacts of retrocommissioning activities from those attributable \nto ECMs. Our results indicate postaudit energy use reductions of \n2.5% for multifamily and 4.9% for office buildings compared to \nnon-audited properties. The magnitude of savings is found to be \nconsistent with what could be achieved through low-cost and no-\ncost energy efficiency improvements.\nEnergy retrofits and the energy efficiency gap\nNumerous studies have demonstrated that energy efficiency retro-\nfits can substantially reduce building energy use13,14. In fact, building \nscience research indicates that \u2018deep\u2019 retrofits can reduce building \nenergy use by 50% or more15,16. Given that commercial and resi-\ndential buildings are responsible for approximately 40% of the total \nenergy consumption in the United States17, to implement even a \nfraction of these measures would make a material contribution to \ngovernments\u2019 GHG reduction goals. As a case in point, one study \nreports that, by 2050, more than 11% of US energy use could be \nsaved through building retrofits alone18.\nMany of the available energy efficiency retrofit measures, whether \ntechnological or behavioural, are shown to have relatively short \npayback periods, after which they are expected to confer net cost \nsavings19. For larger commercial or multifamily buildings, several \ncommon ECMs, such as upgrading building controls and replacing \nlighting, are estimated to have payback periods of five years or less \nand, in many cases, less than one year20,21. Despite the apparent eco-\nnomic case for retrofits, adoption has been slow22,23. This disconnect \nbetween the availability of cost-effective energy efficient technolo-\ngies and their diffusion and use in the marketplace is often referred \nto as the \u2018energy efficiency gap\u20194.\nScholars have long theorized that the energy efficiency gap is \ndue, at least in part, to information deficits11,24\u201326. Three distinct \ntypes of deficits are believed to be at play. First, building own-\ners may lack information about the energy performance of their \nbuildings, which includes information about strategies to reduce \nconsumption, and thus may not be knowledgeable about energy \nefficiency investment opportunities. Second, even where building \nowners are so informed, they may withhold relevant information \nfrom their tenants. Landlords may have an incentive to withhold \nsuch information where tenants pay their own utility bills because, \nin such instances, landlords will bear the cost of energy efficiency \nupgrades, whereas their tenants receive the benefits through \nlower utility payments27. This misalignment of incentives between \nlandlords and tenants is typically referred to as the \u2018split incen-\ntive\u2019 problem25. Finally, the absence of energy use information in \nthe marketplace creates barriers to accounting for building energy \nefficiency in investment or tenant and homebuyer locational deci-\nsions. Labelling and rating systems, such as the US Environmental \nProtection Agency\u2019s ENERGY STAR programme and the US Green \nBuilding Council\u2019s Leadership in Energy and Environmental Design \nrating system, are designed to overcome this knowledge gap, similar \nto nutrition labelling or vehicle fuel efficiency ratings28,29. However, \nthese programmes, and others, are constrained by the self-selection \nof participants and non-trivial flaws in data quality and the statisti-\ncal methods used to calculate scores30,31.\nThe literature contains substantial evidence to support the exis-\ntence of persistent information deficits32,33. Even more concern-\ning is how information asymmetries might impact lower-income \nhouseholds, for whom a lack of knowledge about efficiency oppor-\ntunities or mismatched utility subsidies and payment programmes \nmay hinder measures for energy cost reduction34\u201336. Recent stud-\nies of energy efficiency in subsidized housing found these proper-\nties to be far less efficient than similar market-rate housing, which  \nsuggests that significant opportunities exist to improve the quality \nof the low-income housing stock37,38.\nEffects of energy disclosure and audit policies\nPolicies that require energy audits, such as LL87, and the disclosure \nof energy performance information, like LL84, are designed to help \novercome information gaps39. Yet the two types of policies employ \ndifferent tactics to do so. In particular, although audit requirements \naim to improve building owners\u2019 awareness of cost-effective retrofit \nopportunities, disclosure policies seek to reduce information asym-\nmetries between building owners and prospective investors or renters \nso that this information can factor into decision-making processes40.\nA growing body of evidence suggests that disclosure policies can \ngenerate efficiency improvements. For instance, a recent study by \nMeng et\u00a0al. revealed that LL84 produced a 6% reduction in energy \nuse intensity (EUI) in the first three years after the policy took effect, \nand a 14% reduction over the first four years8. Papadopoulos et\u00a0al. \nalso found energy use reductions over time in the buildings covered \nby LL84, but identified distinct clusters of buildings that actually \nincreased their consumption during the same time period10. A num-\nber of other studies also demonstrate that energy information dis-\nclosure impacts real-estate prices, which is assumed to reflect rental \npremiums, higher occupancy rates or lower operating expenses in \nmore efficient buildings41,42.\nIn contrast, relatively few scholars have examined the impact of \nenergy audits on energy use, and most of those who have done so \nfocused on voluntary audits in the industrial sector, rather than on \nmandatory audits of buildings43,44. With respect to the building sec-\ntor, a study of municipalities in northern Italy found that municipal \nauthorities were significantly more likely to make energy efficiency \nimprovements to public buildings after having conducted an energy \naudit45. Conversely, Murphy46 found that voluntary energy audits in \nprivate homes in the Netherlands had little, if any, impact.\nThe paucity of scholarship that evaluates the impact of manda-\ntory building energy audits is surprising given the increasing preva-\nlence of such policies across the United States. As of 2017, at least \neight American jurisdictions require building owners to conduct \nenergy audits in some form40. Some of these cities have released \ncompliance data and early results from their audit programmes, but \nthis information has not been formally studied. The lack of analysis \nis striking given the costs associated with energy audits. In NYC, for \nexample, the audits required by LL87 have been estimated to cost \napproximately $0.15\u2009ft\u20132 (ref. 11). Applying this cost estimate to our \ndataset of properties that were audited in 2013 or 2014 indicates \nthat an average size multifamily building (126,368\u2009ft2) and office \nbuilding (412,430\u2009ft2) would pay $17,691 and $57,740, respectively, \nfor their audits. Although these are relatively small figures when \ncompared to the value or net operating income of buildings of this \nsize, the time, cost and regulatory oversight create the possibility of \nundue burdens for some building owners.\nThe question, then, is whether mandatory energy audits lead \nto energy use reductions over time. If so, does the magnitude of \nthe effect justify the cost and time associated with conducting and \nreporting the results of the audit? We use a two-way mixed design \nanalysis of variance (ANOVA) and Bayesian regression to evalu-\nate the impact of building audits on energy consumption pre- and \npostaudit compared with a control group not subject to the audit \nrequirement. The analysis is based on energy use and audit compli-\nance data for 3,981 large multifamily and office buildings in NYC \nfrom 2011 to 2016.\nEvidence of postaudit energy use reductions\nFrom Fig. 1, we observe a clear decrease in EUI for audited (treat-\nment) multifamily buildings in the post-audit period, when com-\npared to non-audited buildings. For office buildings, there is a \nreduction in EUI for audited properties between the pre-audit and \nNature Energy | VOL 5 | April 2020 | 309\u2013316 | www.nature.com/natureenergy\n310\n\nArticles\nNATure EnergY\nintervention periods, after which EUI values between the control \nand treatment groups converge, with audited properties exhibit-\ning greater uncertainty. Given higher initial average EUI, office \nbuildings experience larger absolute decreases, on average, in \nenergy use over time when compared to multifamily residential \nbuildings. The ANOVA results, as shown in Table 1, demonstrate \nstatistically significant coefficients for the interaction term (time \nperiod and intervention) for both office and multifamily buildings \nat the 95 and 90% confidence levels, respectively. Although these \nresults suggest that audits do have an impact on energy use over \ntime, they represent a relatively coarse quantification of the audit \npolicy\u2019s impact because they do not account for other factors that \ncould influence building energy use in the postaudit period. The \nBayesian regression results discussed below account for both the \neffect of energy audits and the dynamic control variables that affect \nbuilding energy use over time.\nIn Table 2, we show that Bayesian regression model coefficient \nmeans and the 95% highest posterior density (HPD) intervals for \nNYC\u2019s office and multifamily housing building stock. We note \nthat in both cases, when controlling for factors that might affect \nchanges in energy consumption over time, the average value of the \ncoefficient related to audits (that is, audited property) is negative. \nSpecifically, audited office buildings tend to reduce their EUI by \n4.86% compared to non-audited properties, and multifamily prop-\nerties are found to have a 2.47% reduction. From the intercept of the \ntwo models, however, we notice diverging results. Although office \nbuildings, on average, reduced their energy consumption, multi-\nfamily properties, on average, increased their energy consumption \nduring the study period. Therefore, in the multifamily housing case, \naudits result in a smaller increase in energy use than would other-\nwise be expected.\nFocusing on the coefficient of interest for the effect of energy \naudits on change in energy consumption over time, we show the \ncoefficient posterior distributions for multifamily and office build-\nings, respectively, in Figs. 2 and 3. The three Markov Chains con-\nverged well and are stationary, which means that there are no \ndrifts and discrepancies in the mean and standard deviation of \nthe distributions. To answer the question of \u2018how certain are we \nthat audits have a negative impact on energy consumption over \ntime?\u2019, we calculate the density of the posterior distribution that \nTable 1 | Mixed design ANOVA results\u2014statistical significance\nMultifamily housing\nOffice\nTreatment (P value)\n0.352\n0.749\nTime (P value)\n0.706\n0\nInteraction (treatment\u2009\u00d7\u2009time)  \n(P value)\n0.085\n0.047\n86\na\nb\n115\n110\n105\n100\n95\n90\n85\n85\nEnergy use intensity (kBtu ft\u20132)\nEnergy use intensity (kBtu ft\u20132)\n84\n83\n82\nControl\nAudit\nControl\nAudit\nPre\nIntervention\nTime period\nPost\nPre\nIntervention\nTime period\nPost\nFig. 1 | EUI distribution pre-, during and postaudit period. a,b, Multifamily housing (a) and office buildings (b). The error bars correspond to the 95% \nconfidence intervals. Although there is an observable difference in control and audited office buildings during the pre-audit period, the overlapping \nconfidence intervals make this difference not statistically significant. The t-test results in Table 3 further validate the above argument. Btu, British \nthermal units.\nTable 2 | Bayesian regression results demonstrate the effect of \naudits on energy use\nMultifamily housing\nOffice\nMean \ncoefficient\n95% HPD \ninterval\nMean \ncoefficient\n95% \nHPD \ninterval\nIntercept\n4.27\n(2.32, 6.22)\n\u20137.81\n(\u201319.46, \n4.99)\nAudited property\n\u20132.47\n(\u20134.45, \n\u20130.44)\n\u20134.86\n(\u201310.49, \n0.83)\nNumber of floors\n\u20130.01\n(\u20130.12, 0.10)\n0.01\n(\u20130.28, \n0.26)\nBuilding age\n\u20130.02\n(\u20130.05, 0.01)\n0.07\n(0, 0.15)\nGross floor area\n0\n(0, 0)\n0\n(0, 0)\nElectricity as \nprimary fuel \nsource\n8.28\n(5.83, 10.84)\n\u20133.58\n(\u20139.26, \n1.86)\nMean pre-audit \nEUI\n\u20130.36\n(\u20130.39, \n\u20130.33)\n\u20130.14\n(\u20130.19, \n\u20130.09)\nProperty market \nvalue\n0\n(\u20130.01, 0.01)\n\u20130.01\n(\u20130.04, \n0.02)\nWorker density \ndifference\n\u2013\n\u2013\n0.04\n(\u20130.10, \n0.17)\nComputer density \ndifference\n\u2013\n\u2013\n0.04\n(\u20130.09, \n0.18)\nOperating hours \ndifference\n\u2013\n\u2013\n\u20130.04\n(\u20130.14, \n0.06)\nNature Energy | VOL 5 | April 2020 | 309\u2013316 | www.nature.com/natureenergy\n311\n\nArticles\nNATure EnergY\nfalls below the threshold of zero, which for multifamily housing is \nP(\u03b2audited\u2009<\u20090)\u2009=\u20090.990 and for office is P(\u03b2audited\u2009<\u20090)\u2009=\u20090.958, where \u03b2 is \nthe coefficient for the effect of energy audits on energy use.\nIn an attempt to associate the observed savings with potential \nretrofit actions, we estimate from the audit report data the average \nexpected EUI improvement possible through recommended low-\ncost ECMs (those with payback periods of less than two years) and \nretrocommissioning. The expected savings from low-cost ECMs are \nfound to be 4.56% for multifamily and 1.87% for office buildings, \nwith retrocommissioning activities associated with approximately \n2% savings in both building types. Based on the magnitude of the \naudit-impact coefficients, these figures suggest office buildings \nexhibit, on average, energy savings that are consistent with those \nexpected from recommended low-cost measures. For residen-\ntial buildings, however, as the audit coefficient is lower than that \nexpected from low-cost ECM adoption, the impact of the manda-\ntory audit is negligible in relation to identified savings opportunities.\nFinally, we link the savings associated with energy audits to a \nfinancial consideration that is often overlooked in the retrofit \ndecision: the cost burden of the audit itself. According to the US \nDepartment of Energy, the cost of a building energy audit ranges \nbetween $0.12 and 0.50\u2009ft\u20132 (ref. 47), whereas NYC market-specific \nestimates set the cost at $0.15\u2009ft\u20132 (refs 11,48). Given the average \nenergy savings attributed to audits from the Bayesian model dis-\ncussed above, combined with building fuel mix and energy cost esti-\nmates from the US Environmental Protection Agency49, we find that \nthe average annual energy cost savings due to auditing for the NYC \nbuilding stock are $0.121\u2009ft\u20132 for office and $0.038\u2009ft\u20132 for residential \nbuildings. Therefore, especially for residential properties, the rela-\ntively high payback period of the energy audit (four years or more, \non average) is an important consideration in the cost-benefit analy-\nsis of mandatory audit policies.\nDiscussion and policy implications\nOur analysis finds that buildings that conduct a mandatory audit \nreduce their energy use over time more than non-audited buildings. \nHowever, it is important to note that the magnitude of the audit \neffect is consistent with, or less than, the expected savings from low-\ncost ECMs and retrocommissioning. This effect is particularly lim-\nited for multifamily buildings, which exhibited percentage decreases \nin EUI postaudit of approximately half those of office buildings. We \nconsider two explanations for the relatively modest effect of manda-\ntory audits.\nFirst, it is possible that LL87 audits are, in fact, influencing prop-\nerty owners\u2019 plans for energy efficiency improvements, but capital \ncycles for investment in the relevant measures are too long to be \ncaptured by the temporal period of our analysis. Stated otherwise, it \nmay be that after reviewing a required LL87 audit, a property owner \ndecides to eventually invest in some relatively costly recommended \nimprovement, such as new heating, ventilation and air conditioning \nequipment, but plans to do so (or did do so) after 2016, which is the \nlast year for which we have EUI data in this analysis.\nNotably, however, we did not find a greater difference between \nenergy consumption in audited and non-audited properties when \nwe included only those buildings that conducted audits in 2013, \nrather than the original 2013 and 2014 combined treatment group. \nThis finding suggests that the investment cycle hypothesis may not \nbe a significant factor, as these buildings had an additional year to \nimplement audit recommendations. Of course, capital cycles may \nstill be too long for certain capital-intensive ECMs to be captured \nduring this time period. However, many ECMs with non-trivial \nenergy savings have relatively low first costs and/or have short pay-\nback periods.\nThe second possible explanation is that energy audits do not \nmotivate property owners to invest in ECMs that they would not \notherwise pursue. There are at least three reasons why this might \nbe the case.\nPoor audit quality. The most straightforward explanation as to why \nLL87 audits may not be encouraging meaningful energy savings is \nthat the audit process and reports are not of sufficient quality to \nreduce the uncertainty that buildings owners have regarding the cost \nand energy savings of particular ECMs50. A case study of 30 commer-\ncial and residential audits conducted in buildings across the United \nStates revealed widespread shortcomings, which included missed \nECMs and overestimated savings51. It is possible that the audits being \nproduced in compliance with LL87 are similarly lacking.\nInsufficient economic incentives for investment. Assuming man-\ndatory audits are effective in identifying substantial cost-effective \nECMs, other capital improvements, such as renovating a lobby or \nadding new amenities, may generate larger risk-adjusted returns on \ninvestment. This may be driven, in part, by the perceived uncer-\ntainty around energy retrofit investments, and thus inflate the \nrequired rate of return (or hurdle rate) for energy investments when \ncompared with more traditional building improvements. Building \nowners, then, may lack sufficient economic incentive, in the absence \nof strong energy pricing signals, public incentives or regulatory \nmandates, to implement energy improvements over other capital \nimprovements33.\nIt is important for local policymakers to understand whether this \nis, indeed, a barrier to the implementation of recommended ECMs. \nIf so, cities will need to strengthen the incentives for efficiency \nimprovements before audit mandates can have a significant impact. \nFor instance, to require property owners to assess more ambitious \nECMs in their audits, as has been proposed (for example, NYC\u2019s \nOne City Built to Last sustainability plan52), will not lead to actual \nenergy savings if property owners do not have an incentive, and the \naccess to capital, to invest in energy efficiency in the first place.\nMotivation of property owners. A final reason as to why LL87 \naudits may not generate substantial energy savings relates to the \nmandatory nature of the programme. Clearly, not all property own-\ners are equally interested in energy efficiency and those who are \nTable 3 | Mean values and t-test results\nMultifamily housing\nOffice\nVariable (mean value)\nControl\nTreatment\nMatched\nControl\nTreatment\nMatched\nBuilding age\n65\n66\n66\n75\n77\n77\nNumber of floors\n11\n11\n11\n21\n24\n22\nGross floor area (ft2)\n133,801\n146,207\n132,196\n410,318\n470,741\n371,380\nElectricity as primary fuel source\n0.11\n0.1\n0.1\n0.72\n0.8\n0.8\nProperty market value ($\u2009ft\u20132)\n64\n60.6\n61.5\n139.7a\n165.4\n144.4\nMean pre-audit EUI (kBtu\u2009ft\u20132)\n84.7\n84.5\n84.1\n96\n101.9\n90.3\naDifference significant at 95% confidence level (P value<0.05)\nNature Energy | VOL 5 | April 2020 | 309\u2013316 | www.nature.com/natureenergy\n312\n\nArticles\nNATure EnergY\nmost motivated will probably begin exploring energy saving oppor-\ntunities voluntarily, without waiting for the obligation to conduct \nan audit26,53. It can be argued that much of the reduction in energy \nuse over time observed in cities with energy disclosure laws can be \nattributed to this self-selection by owners more sensitive to envi-\nronmental concerns or, conversely, the public perception of a lack \nof interest in environmental responsibility. If this assumption holds, \nthen the information produced by mandatory audits should be most \nuseful for those owners who are not otherwise motivated to reduce \ntheir energy consumption. Requiring property owners to devote \ntime and money to conduct an audit may provide little value in \nshifting behaviour. Indeed, it may further discourage energy invest-\nments out of contempt for the mandatory nature of the process and \ncreate an incentive to simply do the minimum reporting necessary \nto comply with the law. These attitudinal factors could potentially \nexplain the discrepancy between the findings in our study, which \ninvestigates the impact of mandatory energy audits on energy con-\nsumption, and previous studies that have examined responses to \nvoluntary, subsidized audit programmes and found a robust adop-\ntion of the recommended measures43,54.\nIndependent of the exact cause, if mandatory audits do not \nencourage investment in ECMs proportionate to the time and cost \nof conducting the audit, policymakers should consider whether a \nmandatory policy is the most efficient strategy to encourage energy \nefficiency in buildings. Although mandatory audits do produce \nsubstantial amounts of data about building systems and energy \nreduction potentials, which can have a significant value for policy-\nmaking, it may be possible to collect these data through less costly \nmeans, and complete coverage of a city\u2019s building stock may not \nbe necessary to make reasonable inferences about where savings \nopportunities exist. For example, a recent study demonstrates how \nmachine learning can be used to predict ECM recommendations \nand cost-savings opportunities with a high degree of certainty based \non simple surveys of building systems55. Such \u2018automated\u2019 audits \ncould significantly reduce the time and cost associated with build-\ning assessments.\nConclusions and future work\nClimate action is becoming an increasingly pressing priority for city \nleaders. Given the substantial contribution building energy use has \non GHG emissions in urban areas, cities are focusing their efforts on \nincreasing the efficiency of the building sector through a range of \ndata-driven policy and regulatory tools. Although energy reporting \nis a relatively low-cost mandate, required audits impose non-trivial \ncosts in terms of capital and time for building owners. The question \nwe explore here is whether this requirement actually advances the \ngoal of reducing energy consumption in buildings.\nWe find that, for the time period studied between 2011 and \n2016, mandatory energy audits had a modest negative impact on \nenergy consumption in office and residential buildings in NYC, \n0.40\na\nb\n1\n0\n\u20131\n\u20132\n\u20133\n\u20134\n\u20135\n\u20136\n\u20137\n0\n250\n500\n750\n1,000 1,250\nSample\n1,500 1,750 2,000\nMC 1\nMC 2\nMC 3\nMC 1\nMC 2\nMC 3\n0.35\n0.30\n0.25\nDensity\n0.20\n0.15\n0.10\n0.05\n0.00\n\u20138\n\u20136\n\u20134\nAudit coefficient\nAudit coefficient\n\u20132\n0\n2\nFig. 2 | Audit coefficient distribution (multifamily housing). a,b, Multifamily housing audit coefficient distribution (a) and sampled values (b) for the \nthree Markov Chains (MC) used to train the regression model.\n0.14\n0.12\n0.10\n0.08\nDensity\n0.06\n0.04\n0.02\n0.00\n\u201315\n\u201310\n\u20135\nAudit coefficient\n0\n5\n0\n\u20135\n\u201310\n\u201315\n5\n0\n250\n500\n750\n1,000\nSample\n1,250\n1,500\n1,750\n2,000\nMC 1\nMC 2\nMC 3\nMC 1\nMC 2\nMC 3\nAudit coefficient\na\nb\nFig. 3 | Audit coefficient distribution (office). a,b, Office audit coefficient distribution (a) and sampled values (b) for the three Markov Chains used to \ntrain the regression model.\nNature Energy | VOL 5 | April 2020 | 309\u2013316 | www.nature.com/natureenergy\n313\n\nArticles\nNATure EnergY\nat a magnitude consistent with the savings potential of low-cost \nECMs and retrocommissioning activities. The result reinforces the \nhypothesis that audits, by themselves, provide only limited incen-\ntive to invest in energy efficiency upgrades. Ultimately, building \nowners remain constrained by factors that audit information alone \nmay not overcome, such as limited access to capital, uncertainty in \nsavings projections, opportunity costs and weak pricing signals in \nenergy markets.\nTo begin to address these economic and behavioural barriers, \ncities need to develop a comprehensive strategy to support energy \nefficiency in the building sector that starts with a foundation of data \ntransparency and rigorous analytics. Energy disclosure mandates \nare an important first step: once the data are available, buildings can \nbe evaluated on their energy performance and compared with their \npeers, which creates a \u2018grading\u2019 scheme that can help to shift indi-\nvidual and collective decision-making56. Following this, cities can \nconsider performance targets and provide financial and regulatory \nincentives to motivate building owners to improve their energy effi-\nciency and also ensure that regulations are in place to require poorly \nperforming buildings to improve when owners do not respond to \nincentives. Audit requirements, then, could be used to target deep \nretrofits, focusing on ECM opportunities that could achieve 30% \nor greater savings, and automated or virtual audits could replace \nthe existing need for traditional audit mandates. Similarly, because \naudit policies produce significant data on building systems and \noperating characteristics\u2014information that is useful for a range of \ncity agencies, but often difficult to collect\u2013mandatory requirements \ncould be replaced by incentives for voluntarily reporting audit data.\nOur analysis here is an initial attempt to empirically investigate \nthe impact of a citywide mandatory energy audit law. As we con-\ntinue our research, future work will extend this analysis as addi-\ntional years of data become available. A detailed survey of building \nowners and managers would help to understand the motivations \nfor and barriers to implementing energy efficiency upgrades and \nprovide the needed context to develop a comprehensive city energy \nstrategy. A systematic evaluation of audit quality should also be con-\nducted to ensure the accuracy of audit reports, which would help \nto provide owners with greater certainty on the validity of the pro-\nposed recommendations. Finally, machine learning and artificial \nintelligence applications should be explored more fully to under-\nstand if data-driven modelling can enhance, complement or even \nreplace more manual processes for collecting relevant building data \nand developing ECM recommendations. Together, these steps could \nbe used to develop evidenced-based and data-driven policies for \ncity climate action.\nMethods\nData description. Our energy performance data (total energy use and EUI) and \naudit data were collected by the NYC Department of Finance and the Department \nof Buildings pursuant to LL84 and LL87. Data were provided by the NYC Mayor\u2019s \nOffice of Sustainability subsequent to a data sharing request. The LL84 dataset \nincludes all covered buildings, which are defined as buildings with greater than \n50,000\u2009ft2 of gross floor area, that submitted energy use data in calendar years 2011 \nthrough 2016. EUI is defined as the total annual energy consumption divided by \nthe gross floor area of the building, and we utilized weather-normalized site EUI \nto capture the direct consumption reported through utility bills adjusted for the \ntotal number of heating degree days and cooling degree days in a given year57. In \naddition to the total energy use and EUI values, the dataset contains building-\nspecific features, which include physical (age, gross floor area and so on) and \noperational (occupancy density, weekly operating hours, conditioned spaces and so \non) characteristics.\nWe performed preprocessing of the LL84 dataset prior to our analysis. First, as \nEUI data are self-reported, we identified and removed misreported and erroneous \nentries. Specifically, we applied a logarithmic transformation to the EUI data and \nfiltered outlier values that fell outside the threshold of two standard deviations \nfrom the mean58. Second, using the unique Borough Block Lot property identifier, \nwe merged energy and building attribute data with tax lot and zoning information \nprovided by the NYC\u2019s Primary Land Use Tax Lot Output database to identify \nadditional building characteristics, such as assessed value. Finally, we integrated \nthe merged dataset with information from individual audit reports submitted to \nthe Department of Buildings as per LL87 requirements to identify properties that \nconducted an energy audit in calendar years 2013 or 2014, and to analyse building-\nspecific ECM recommendations and savings potentials.\nAfter our data processing steps, we analysed whether the audited properties \nin our sample demonstrated larger percentage reductions in site EUI between the \npre- and postaudit period than those of similar buildings that did not perform an \naudit during the study period. We defined the EUI percentage change for each \nbuilding as the difference between the mean EUI during the two years prior to \nthe audit (2011 and 2012) and the two years after the audit (2015 and 2016). We \nused the two-year average to account for anomalous variations in building energy \nconsumption that could occur in a given year (for example, Hurricane Sandy had \na non-trivial impact on energy use in buildings in the impacted areas in 2012). We \nfocused on multifamily and office buildings, as the two types account for more \nthan 90% of the total LL84 covered properties by quantity and aggregate energy \nconsumption. The merged dataset contains 3,981 properties, which include 3,563 \nmultifamily buildings and 418 office buildings.\nMethodology. To assess the impact of energy audits on building EUI over time, \nwe split the data into three time periods: pre-audit (average EUI in years 2011 and \n2012), intervention (average EUI in years 2013 and 2014) and postaudit (average \nEUI in years 2015 and 2016). As a first step to test the significance of the audit \neffect, we used a two-way mixed design ANOVA in which the dependent variable \nis EUI, the within-group variable is time (with three levels as mentioned above) \nand the between-group variable is the intervention (with two levels, which indicate \naudited and non-audited buildings).\nOne significant limitation of this approach is that it does not account for \nadditional variables, besides time and intervention, that might be associated with \nchanges in EUI, such as changes in occupancy characteristics.\nBayesian regression. To more comprehensively examine the impact of the audit \npolicy, we used a Bayesian regression model to quantify the impact of energy \naudits on energy reduction over time while controlling for additional time-varying \n(dynamic) covariates that can influence changes in energy consumption. Bayesian \nstatistics and probabilistic programming allow us not only to assess the effect \nof audits on energy use, but also to quantify the uncertainty of the estimated \nparameters59,60. Bayesian regression is a useful alternative to frequentist methods \nfor policy analysis, as it provides a more intuitive, probabilistic output rather than \nrelying on fixed P-value thresholds. We formulate our problem as follows:\ny \ue019N\u00f0\u03b2TX; \u03c32\ny\u00de\n\u00f01\u00de\nThe output y, which is the percentage difference in EUI between the pre- and \npostaudit period, is generated from a normal distribution (N) with a mean equal \nto the transpose (T) of the coefficients\u2019 (\u03b2) matrix multiplied by the independent \nvariables\u2019 (X) matrix and variance \u03c32\ny\nI\n.\nThe frequentist approach to this problem estimates the model parameters using \nthe maximum likelihood estimation method as static values: ^\u03b2 \u00bc \u00f0XTX\u00de\n\u001c1XTy\nI\n. This constitutes the fundamental difference between Bayesian and frequentist \nstatistics; Bayesian makes a statement of probability about a parameter value \ngiven a fixed credible region ( P\u00f0\u03b8j credibleregion \u00de\nI\n), whereas frequentist makes \na statement of probability about the confidence interval given a fixed parameter \nvalue ( P\u00f0 confidenceinterval j\u03b8\u00de\nI\n)61,62. Therefore, in the context of our policy \nevaluation research problem, Bayesian statistics allow us to ask and answer \nthe question \u2018how certain are we that the effect of energy audits on energy \nconsumption over time is negative?\u2019.\nThe independent variables we use in our model cover a range of building \ncharacteristics, changes in occupancy and the audit treatment effect. We control \nfor non-varying characteristics, which include building age, size and EUI levels \nprior to the audit, as well as time-varying occupancy factors, which include worker \ndensity and operating hours, that might cause changes in the EUI over time. Note \nthat occupancy variables are not available for multifamily buildings, and hence \nthey are omitted from the residential model.\nWe used the No-U-Turn Sampler developed by Hoffman and Gelman63, an \nefficient Markov Chain Monte Carlo algorithm, to draw 2,000 posterior samples \nfor the office and multifamily housing stock models. Monte Carlo is a general \napproach of drawing random samples, whereas Markov Chain revolves around \nthe concept that the next sample to be drawn is independent of the past, based \nsolely on the present sample (the Markov process). We repeated the process for \nthree chains to assess the robustness of our estimated model parameters. Another \nadvantage of Bayesian regression is the ability to include prior knowledge regarding \nthe model parameters\u2019 distribution. In this work, as NYC is among the first \nlarge cities to enact a mandatory energy audit law, we do not provide any prior \ninformation in the algorithm (an uninformative prior).\nTo measure the uncertainty in the model estimates, we calculated the HPD \nintervals, which is the range of values that cover the maximum distribution \ndensity under a predefined probability. Let f(x) be the density function of \nthe random variable X; then the 95% highest density region is the subset \nR\u00f0f 95\u00de \u00bc\nx : f \u00f0x\u00de\u2265f 95\n\u001f\n\u001e\nI\n, where P X 2 R\u00f0f 95\u00de\n\u001f\n\u22650:95\nI\n (ref. 64).\nNature Energy | VOL 5 | April 2020 | 309\u2013316 | www.nature.com/natureenergy\n314\n\nArticles\nNATure EnergY\nPrior to fitting the Bayesian model, we examined the treatment and control \ngroups to verify that the inherent variability between audited and non-audited \nbuilding characteristics is random and not attributable to some unobserved \nselection bias. As the year a building must comply with LL87 and submit its audit \ndata is determined based on the last digit of its tax lot number, the selection and \nreporting process is essentially random. However, the concern is that the treatment \n(audited) group is not a randomized sample of the population and therefore \nfundamentally different than the control (non-audited) group. To account for \nthis, Table 3 shows the mean values of the model covariates that do not change \nover time for audited (treatment), non-audited (control) and a set of matched \nnon-audited properties using propensity scores with 1:1 nearest-neighbour \nmatching65,66. We ran t-tests to examine the difference in means between the \ncontrol and matched sample with that of the treatment group. With the exception \nof the assessed value per square foot for office buildings, we found no statistically \nsignificant differences in the treatment and control groups, either with or without \nmatching. Therefore, we are confident in the random assignment of buildings to \nthe treatment and control samples.\nData availability\nAll data, except LL87 data, are available through the NYC Open Data Portal at \nhttps://opendata.cityofnewyork.us/. LL87 data are available upon reasonable \nrequest, and with permission, from the NYC Mayor\u2019s Office of Sustainability. The \ndata that support the plots within this Article and other findings of this study can \nbe obtained from the NYC Mayor\u2019s Office of Sustainability or the authors, upon \npermission from the NYC Mayor\u2019s Office of Sustainability who own the data.\nCode availability\nAny applicable code relevant to the findings is available from the authors upon \nreasonable request.\nReceived: 27 August 2019; Accepted: 26 February 2020;  \nPublished online: 30 March 2020\nReferences\n\t1.\t Glaeser, E. L. & Kahn, M. E. The greenness of cities: carbon dioxide \nemissions and urban development. J. Urban Econ. 67, 404\u2013418 (2010).\n\t2.\t Inventory of New York City Greenhouse Gas Emissions (City of New York, \n2017); https://www.dec.ny.gov/docs/administration_pdf/nycghg.pdf\n\t3.\t Burr, A., Keicher, C. & Lawrence, J. Commercial Building Energy Rating and \nDisclosure Policies (Institute for Market Transformation, 2013).\n\t4.\t Jaffe, A. B. & Stavins, R. N. The energy-efficiency gap What does it mean? \nEnergy Policy 22, 804\u2013810 (1994).\n\t5.\t Map: U.S. City, County, and State Policies for Existing Buildings: Benchmarking, \nTransparency and Beyond (Institute for Market Transformation, 2017);  \nhttps://www.imt.org/resources/map-u-s-building-benchmarking-policies\n\t6.\t LL87: energy audits & retro-commissioning NYC Mayor\u2019s Office of \nSustainability http://www.nyc.gov/html/gbee/html/plan/ll87.shtml (2020).\n\t7.\t New York City\u2019s Energy and Water Use 2014 and 2015 Report (City of New \nYork, 2017); http://www.nyc.gov/html/gbee/downloads/pdf/UGC-\nBenchmarking-Report-101617-FINAL.pdf\n\t8.\t Meng, T., Hsu, D. & Han, A. Estimating energy savings from benchmarking \npolicies in New York City. Energy 133, 415\u2013423 (2017).\n\t9.\t Palmer, K. & Walls, M. Using information to close the energy efficiency gap: \na review of benchmarking and disclosure ordinances. Energy Effic. 10, \n673\u2013691 (2017).\n\t10.\tPapadopoulos, S., Bonczak, B. & Kontokosta, C. E. Pattern recognition in \nbuilding energy performance over time using energy benchmarking data. \nAppl. Energy 221, 576\u2013586 (2018).\n\t11.\tHsu, D. How much information disclosure of building energy performance is \nnecessary? Energy Policy 64, 263\u2013272 (2014).\n\t12.\tSpiegel-Feld, D. Local Law 97: Emissions trading for buildings? NYU Law \nReview Online 94, 148\u2013168 (2019).\n\t13.\tChidiac, S., Catania, E., Morofsky, E. & Foo, S. Effectiveness of single and \nmultiple energy retrofit measures on the energy consumption of office \nbuildings. Energy 36, 5037\u20135052 (2011).\n\t14.\tArdente, F., Beccali, M., Cellura, M. & Mistretta, M. Energy and \nenvironmental benefits in public buildings as a result of retrofit actions. \nRenew. Sust. Energy Rev. 15, 460\u2013470 (2011).\n\t15.\tHestnes, A. G. & Kofoed, N. U. Effective retrofitting scenarios for energy \nefficiency and comfort: results of the design and evaluation activities within \nthe office project. Build. Environ. 37, 569\u2013574 (2002).\n\t16.\t\u00dcrge-Vorsatz, D., Harvey, L. D. D., Mirasgedis, S. & Levine, M. D. Mitigating \nCO2 emissions from energy use in the world\u2019s buildings. Build. Res. Inf. 35, \n379\u2013398 (2007).\n\t17.\tUse of Energy in the United States Explained (U.S. Energy Information \nAdministration, 2017); https://www.eia.gov/energyexplained/index.\ncfm?page=us_energy_use\n\t18.\tNadel, S. Pathway to Cutting Energy Use and Carbon Emissions in Half \n(American Council for an Energy-Efficient Economy, 2016).\n\t19.\tBradshaw, J. L., Bou-Zeid, E. & Harris, R. H. Greenhouse gas mitigation \nbenefits and cost-effectiveness of weatherization treatments for low-income, \nAmerican, urban housing stocks. Energy Buildings 128, 911\u2013920 (2016).\n\t20.\tFulton, M. et\u00a0al. United States Building Energy Efficiency Retrofits: Market \nSizing and Financing Models (Deutsche Bank Climate Change Advisors, 2012).\n\t21.\tChidiac, S., Catania, E., Morofsky, E. & Foo, S. A screening methodology for \nimplementing cost effective energy retrofit measures in Canadian office \nbuildings. Energy Build. 43, 614\u2013620 (2011).\n\t22.\tIm, J., Seo, Y., Cetin, K. S. & Singh, J. Energy efficiency in US residential \nrental housing: adoption rates and impact on rent. Appl. Energy 205, \n1021\u20131033 (2017).\n\t23.\tRyghaug, M. & S\u00f8rensen, K. H. How energy efficiency fails in the building \nindustry. Energy Policy 37, 984\u2013991 (2009).\n\t24.\tHirst, E. & Brown, M. Closing the efficiency gap: barriers to the efficient use \nof energy. Resour Conserv. Recycl. 3, 267\u2013281 (1990).\n\t25.\tGillingham, K. and Palmer, K. Bridging the Energy Efficiency Gap Discussion \npaper 13\u201302 (Resources for the Future, 2013).\n\t26.\tKontokosta, C. E. Energy disclosure, market behavior, and the building data \necosystem. Ann. NY Acad. Sci. 1295, 34\u201343 (2013).\n\t27.\tHope, A. J. & Booth, A. Attitudes and behaviours of private sector landlords \ntowards the energy efficiency of tenanted homes. Energy Policy 75, 369\u2013378 \n(2014).\n\t28.\tElbel, B., Kersh, R., Brescoll, V. L. & Dixon, L. B. Calorie labeling and food \nchoices: a first look at the effects on low-income people in New York City. \nHealth Affairs 28, w1110\u2013w1121 (2009).\n\t29.\tDavis, L. W. & Metcalf, G. E. Does better information lead to better choices? \nEvidence from energy-efficiency labels. J. Assoc. Environ. Resour. Econ. 3, \n589\u2013625 (2016).\n\t30.\tNewsham, G. R., Mancini, S. & Birt, B. J. Do LEED-certified buildings save \nenergy? Yes, but. Energy Build. 41, 897\u2013905 (2009).\n\t31.\tKontokosta, C. E. A market-specific methodology for a commercial building \nenergy performance index. J. Real Estate Finance Econ. 51, 288\u2013316 (2015).\n\t32.\tNiemeyer, S. Consumer voices: adoption of residential energy-efficient \npractices. Int. J. Consum. Stud. 34, 140\u2013145 (2010).\n\t33.\tKontokosta, C. E. Modeling the energy retrofit decision in commercial office \nbuildings. Energy Build. 131, 1\u201320 (2016).\n\t34.\tLaskari, M., Karatasou, S. & Santamouris, M. The design of an energy and \nwater advice programme for low-income households. Energy Build. 110, \n426\u2013434 (2016).\n\t35.\tKontokosta, C. E., Reina, V. J. & Bonczak, B. Energy cost burdens for \nlow-income and minority households: evidence from energy benchmarking \nand audit data in five US cities. J. Am. Plann. Assoc. 86, 89\u2013105 (2020).\n\t36.\tSantamouris, M. et\u00a0al. On the relation between the energy and social \ncharacteristics of the residential sector. Energy Build. 39, 893\u2013905 (2007).\n\t37.\tBird, S. & Hernandez, S. Policy options for the split incentive: increasing \nenergy efficiency for low-income renters. Energy Policy 48, 506\u2013514 (2012).\n\t38.\tReina, V. J. & Kontokosta, C. E. Low hanging fruit? Regulations and energy \nefficiency in subsidized multifamily housing. Energy Policy 106, 505\u2013513 \n(2017).\n\t39.\tRamos, A., Gago, A., Labandeira, X. & Linares, P. The role of information for \nenergy efficiency in the residential sector. Energy Econ. 52, S17\u2013S29 (2015).\n\t40.\tMims, N. et\u00a0al. Evaluation of US Building Energy Benchmarking and \nTransparency Programs: Attributes, Impacts, and Best Practices (Lawrence \nBerkeley National Laboratory, 2017).\n\t41.\tJensen, O. M., Hansen, A. R. & Kragh, J. Market response to the public \ndisplay of energy performance rating at property sales. Energy Policy 93, \n229\u2013235 (2016).\n\t42.\tFuerst, F. & McAllister, P. Eco-labeling in commercial office markets: do \nLEED and energy star offices obtain multiple premiums? Ecol. Econ. 70, \n1220\u20131230 (2011).\n\t43.\tAnderson, S. T. & Newell, R. G. Information programs for technology \nadoption: the case of energy-efficiency audits. Resour. Energy Econ. 26, 27\u201350 \n(2004).\n\t44.\tFleiter, T., Gruber, E., Eichhammer, W. & Worrell, E. The German energy \naudit program for firms\u2014a cost-effective way to improve energy efficiency? \nEnergy Effic. 5, 447\u2013469 (2012).\n\t45.\tAnnunziata, E., Rizzi, F. & Frey, M. Enhancing energy efficiency in public \nbuildings: the role of local energy audit programmes. Energy Policy 69, \n364\u2013373 (2014).\n\t46.\tMurphy, L. The influence of energy audits on the energy efficiency \ninvestments of private owner-occupied households in the Netherlands. Energy \nPolicy 65, 398\u2013407 (2014).\n\t47.\tBaechler, M.C. A Guide to Energy Audits PNNL-20956 (Pacific Northwest \nNational Laboratory, 2011).\n\t48.\tVarone, S. & Varsalona, P. Complying with Local Law 87/09 (Rand \nEngineering and Architecture, 2018); https://randpc.com/ask/mep/\nlocal-law-8709\nNature Energy | VOL 5 | April 2020 | 309\u2013316 | www.nature.com/natureenergy\n315\n\nArticles\nNATure EnergY\n\t49.\tEstimating Fuel Mix and Energy Cost (Environmental Protection Agency, \n2018); https://portfoliomanager.energystar.gov/pdf/reference/FuelMixand \nCost.pdf\n\t50.\tKrutwig, M. C. & Tant\u0103u, A. Obligatory versus voluntary energy audits: are \nthere differences in quality?. Proc. Int. Conf. Bus. Excellence 12, 522\u2013532 (2018).\n\t51.\tShapiro, I. 10 common problems in energy audits. ASHRAE J. 53, 26\u201331 (2011).\n\t52.\tOne City Built to Last (City of New York, 2014); https://www1.nyc.gov/assets/\nbuilttolast/downloads/OneCity.pdf\n\t53.\tNair, G., Gustavsson, L. & Mahapatra, K. Factors influencing energy efficiency \ninvestments in existing Swedish residential buildings. Energy Policy 38, \n2956\u20132963 (2010).\n\t54.\tHarris, J., Anderson, J. & Shafron, W. Investment in energy efficiency: a \nsurvey of Australian firms. Energy Policy 28, 867\u2013876 (2000).\n\t55.\tMarasco, D. E. & Kontokosta, C. E. Applications of machine learning \nmethods to identifying and predicting building retrofit opportunities. Energy \nBuild. 128, 431\u2013441 (2016).\n\t56.\tPapadopoulos, S. & Kontokosta, C. E. Grading buildings on energy \nperformance using city benchmarking data. Appl, Energy 233, 244\u2013253 (2019).\n\t57.\tEnergy Star Portfolio Manager: Climate and Weather Technical reference (US \nEnvironmental Protection Agency, 2019).\n\t58.\tKontokosta, C. E. & Tull, C. A data-driven predictive model of city-scale \nenergy use in buildings. Appl. Energy 197, 303\u2013317 (2017).\n\t59.\tEfron, B. Bayesians frequentists, and scientists. J. Am. Stat. Assoc. 100,  \n1\u20135 (2005).\n\t60.\tClark, J. S. Why environmental scientists are becoming Bayesians. Ecol. Lett. \n8, 2\u201314 (2005).\n\t61.\tMurphy, K. P. Machine Learning: a Probabilistic Perspective (MIT press, 2012).\n\t62.\tGelman, A. et\u00a0al. Bayesian Data Analysis (Chapman and Hall/CRC, 2013).\n\t63.\tHoffman, M. D. & Gelman, A. The No-U-Turn Sampler: adaptively  \nsetting path lengths in Hamiltonian Monte Carlo. J. Mach. Learn. Res. 15, \n1593\u20131623 (2014).\n\t64.\tHyndman, R. J. Computing and graphing highest density regions. Am. Stat. \n50, 120\u2013126 (1996).\n\t65.\tDehejia, R. H. & Wahba, S. Propensity score-matching methods for \nnonexperimental causal studies. Rev. Econ. Stat. 84, 151\u2013161 (2002).\n\t66.\tRosenbaum, P. R. & Rubin, D. B. The central role of the propensity score in \nobservational studies for causal effects. Biometrika 70, 41\u201355 (1983).\nAcknowledgements\nWe thank the NYC Mayor\u2019s Office of Sustainability for sharing the LL84 benchmarking \nand LL87 energy audit data used in this study. We also thank participants at the \u2018Next \nGeneration Building Efficiency Policies\u2019 workshop hosted by the NYU Law School and \nNYU Marron Institute for their feedback on preliminary versions of this work, and \nK. Hoffman for useful discussions on methodology. The research was supported by \nNational Science Foundation Grant no. 1653773 and by the Sloan Foundation. All errors \nremain our own.\nAuthor contributions\nAll the authors contributed equally to all aspects of the work presented in this article.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nCorrespondence and requests for materials should be addressed to C.E.K.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2020\nNature Energy | VOL 5 | April 2020 | 309\u2013316 | www.nature.com/natureenergy\n316\n\n\n    Scientific Research Findings:", "answer": "We find that, for the time period studied between 2011 and 2016, mandatory energy audits had a modest negative impact on energy consumption in office and residential buildings in New York City. The magnitude was consistent with the savings potential of low-cost energy conservation measures and retro-commissioning activities. Audited buildings are found to have energy use reductions post-audit of 2.5% for multifamily and 4.9% for office buildings beyond what would otherwise be expected. The results reinforce the hypothesis that audits, by themselves, provide only limited incentive to invest in energy efficiency upgrades. Ultimately, building owners remain constrained by factors that audit information alone may not overcome, such as limited access to capital, uncertainty in savings projections, opportunity costs and weak pricing signals in energy markets. However, our analysis is constrained by data limitations that prevent us from examining the adoption of specific energy conservation measures over the study period.", "id": 24}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-020-0549-1\nDepartment of Earth and Environment, Boston University, Boston, MA, USA. *e-mail: kaufmann@bu.edu\nB\netween the second quarters of 2007 and 2008, the nominal \nprice for a barrel of crude oil rises from US$65 to US$124. \nNine months and a financial crisis later, the same barrel sells \nfor about US$43. Such rapid changes suggest a speculative bubble, \nwhich the US Commodity Futures Trading Commission defines as \na rapid run-up in prices caused by excessive buying that is unrelated \nto any of the basic, underlying factors affecting the supply or demand \nfor a commodity (https://www.cftc.gov/ConsumerProtection/\nEducationCenter/CFTCGlossary/glossary_s.html). But empirical \nevidence for a speculative bubble is mixed1,2.\nNor is the 2007\u20132008 spike and collapse the first time that prices \nseem to move away from a trading range implied by market funda-\nmentals. Between the second quarters of 1973 and 1974, nominal \noil prices nearly triple, from about US$3.50 to US$10 per barrel. \nThis increase coincides with production reductions and embargoes \nby the Organization of the Petroleum Exporting Countries (OPEC) \nthat reduce net production by about 3.1 million barrels per day \n(mbd), which represents about 5.5% of total output. But histori-\ncal perspectives suggest that these reductions are not the primary \ndriver of the price increase, \u201cit was not loss of supply, but fear of \npossible loss that drove up the price\u201d (ref. 3).\nBeyond market fundamentals, oil prices may respond to geo-\npolitical events4,5 and structural changes6. The structure of the oil \nmarket changes when the \u2018Seven Sisters\u2019, a group of seven interna-\ntional oil companies that dominated the world oil market between \nthe 1940s and the early 1970s, sign 50-50 profit-sharing agreements \nwith producing countries and changes again when producing coun-\ntries nationalize properties owned by the Seven Sisters. But analy-\nses that focus on such changes largely ignore market fundamentals, \nwhich they represent by assuming that prices revert to a constant \nmean or follow a deterministic trend.\nHere, we identify the price effects of speculative bubbles and \nchanges in market governance and property rights by defining \nregimes as two or more consecutive quarters when oil prices move \naway from the level implied by market fundamentals. To identify \nregimes, we start with a simple assumption: in a world without \nregimes, a stable long-run cointegrating relation translates market \nfundamentals into an equilibrium price for oil. Observed prices \nmove towards these ever-changing equilibria as represented by an \nerror correction model (ECM). Statistically significant deviations \nfrom market fundamentals (that is, regimes) are identified by an \nindicator saturation technique. This methodology identifies nine \nregimes that we tie to a speculative bubble in 2007\u20132008 (and 2010), \nchanges in property rights, such as OPEC nations gaining control \nof oil resources from the Seven Sisters, and changes in market gov-\nernance, such as the extensive US energy legislation in the 1970s. \nConversely, not all sharp price changes constitute regimes: market \nfundamentals account for the 13% decline in real prices during the \nAsian financial crisis, which indicates that market fundamentals can \naccount for some of the price changes that are attributed to geopo-\nlitical events by previous analyses.\nA cointegrating relation for price\nBecause a cointegrating relation for price cannot be chosen a\u00a0priori, \nwe estimate 58 possible specifications (Supplementary Table 1).  \nOf these, we focus on nine (Models 1\u20139), which specify a single \nmeasure of capacity utilization by the Texas Railroad Commission \n(UtilTRC) and OPEC (UtilOPEC), and include inventories (Days) \nand refinery utilization rates (UtilRef), that cointegrate with and \nhave a statistically measurable relation with oil prices. ECMs find \nthat oil prices adjust towards the equilibrium implied by the cointe-\ngrating relation (Supplementary Table 2). After we allow for regime \nchanges, UtilOPEC no longer has a statistically measurable long-\nrun relation with real oil prices (Table 1). Similarly, the statistical \nsignificance of UtilTRC disappears when specified as a cubic term. \nFinally, the statistical significance of Days disappears in many mod-\nels that use a non-linear specification of UtilTRC or UtilOPEC.\nOf the nine models estimated, the elements of the cointegrat-\ning relation in the simplest Model 1 indicate that a 1-day increase \nin crude oil inventories reduces the equilibrium price by about \nUS$0.14, a one percentage point increase in refinery utilization \nrates reduces the equilibrium price by about US$0.92 and a one per-\ncentage point increase in capacity utilization by the Texas Railroad \nCommission (TRC) raises the equilibrium price by about US$0.21 \n(Supplementary Note 3). We focus on Model 1 because it is the \nsimplest. But this focus does not affect the discussion that follows \nbecause results generally are similar across all models (Table 1 and \nSupplementary Table 3).\nOil price regimes and their role in price diversions \nfrom market fundamentals\nRobert. K. Kaufmann\u200a \u200a\u2009* and Caitlin Connelly\nSpeculative bubbles, market governance and property rights are thought to affect oil prices, but their timing and magnitude \nare uncertain. Here, we quantify these effects using econometric techniques that identify periods between 1938:1 and 2018:3 \n(denoting year:quarter) when prices strayed from the levels implied by market fundamentals. We identify nine price regimes \nthat are associated with the Organization of the Petroleum Exporting Countries gaining control over the marginal supply of crude \noil, US energy legislation, a precautionary demand shock, the Arab Spring and speculative bubbles. These bubbles raised real oil \nprices by US$14.31 and US$4.65 per barrel in 2007:4\u20132008:3 and 2010:1\u20132011:1, respectively, which transferred US$42.8 bil-\nlion from US consumers to US oil producers and US$87.4 billion from the US economy to oil exporting nations. Conversely, some \nsharp changes, such as the price decline associated with the Asian financial crisis, can be explained by market fundamentals.\nNature Energy | VOL 5 | February 2020 | 141\u2013149 | www.nature.com/natureenergy\n141\n\nArticles\nNATUrE EnErgy\nECMs indicate that prices adjust towards the equilibrium that is \nimplied by the cointegrating relation (Table 1 and Supplementary \nTable 2) in about four quarters (ln(2)/0.17). The first differences \nof UtilRef, UtilTRC, Days and Price have a short-run relation with \nPrice in all ECMs, whereas there is no short-run relation with \nUtilOPEC in any ECM.\nIdentifying regimes\nThe saturation indicator technique identifies periods when prices \nmove away from the level implied by market fundamentals for two \nor more consecutive quarters. Regimes start in the quarter when a \nstep (SIS) changes the equilibrium price and end when the next step  \neither moves the equilibrium price back towards or further away \nfrom the equilibrium that is implied by market fundamentals. For \nmodels that specify a linear (or squared) relation between price and  \ncapacity utilization by producer organizations, the gets procedure \nwithin the R package of that name identifies 12 steps (Table 1), \nwhich define nine regimes (Fig. 1). For each regime, the devia-\ntion from the equilibrium price implied by market fundamentals  \n(rectangles in Fig. 1) is given by the accumulated value of individual \nsteps. When the sum of steps is not statistically different from zero \n(that is, the sum of step 1 and step 2), no regime is present.\nThe start and end dates of the regimes identified by the nine  \nmodels are similar, except for Regime 5, which starts either in \nthe first or third quarter of 1981, and the end of Regime 8, which \nstarts in the third or fourth quarter of 2009 or the first quarter of \n2010 (Table 1). Eight regimes are identified in models that specify  \nUtilTRC3. Finally, the regimes identified by Model 1 are robust  \nto the sample period, the variables used to measure price and/\nor crude oil inventories, and the frequency of the observations \n(Supplementary Note 4).\nInterpreting regimes\nResults for the cointegration/ECMs are consistent with Occam\u2019s \nrazor; market fundamentals account for much of the variation in oil \nprices in 139 (or 138) of the 323 quarters in the 1938:1\u20132018:3 sam-\nple period. During the other 184 quarters, the price implied by the \ncointegrating relation is altered by regimes (Table 2). Most regimes \nTable 1 | Regression results for different models of equilibrium oil price\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nModel 6\nModel 7\nModel 8\nModel 9\nDynamic ordinary least squares (DOLS) estimate of the cointegrating relation (see Equation (4) in the Methods section)\nConstant\n0.258\n0.210\n0.209\n0.242\n0.204\n0.211\n0.379*\n\u22120.086\n\u22120.089\nDays\n\u22120.190*\n\u22120.177*\n\u22120.178*\n\u22120.134\n\u22120.130\n\u22120.137\n\u22120.318***\n\u22120.078\n\u22120.081\nUtilRef\n\u22120.528***\n\u22120.471***\n\u22120.473***\n\u22120.487***\n\u22120.437***\n\u22120.443***\n\u22120.440***\n\u22120.393***\n\u22120.400***\nUtilTRC\n0.504***\n0.450***\n0.452***\nUtilTRC2\n0.497***\n0.457***\n0.469***\nUtilTRC3\n\u22120.013\n0.066\n0.072\nUtilOPEC\n0.010\n0.071\n0.022\nUtilOPEC2\n0.076\n0.054\n0.013\nUtilOPEC3\n0.084\n0.066\n0.025\nSIS1946 Q1\n\u22120.874***\n\u22120.838***\n\u22120.854***\n\u22120.862***\n\u22120.834***\n\u22120.839***\n\u22120.660***\n\u22120.576***\n\u22120.574***\nSIS1949 Q4\n0.700***\n0.632***\n0.636***\n0.624***\n0.636***\n0.646***\n0.475***\n0.483***\nSIS1953 Q1\n\u22120.468***\n\u22120.473***\nSIS1957 Q3\n\u22120.635***\nSIS1970 Q3\n\u22121.041***\n\u22120.947***\n\u22120.950***\n\u22120.948***\n\u22120.871***\n\u22120.892***\nSIS1974 Q1\n0.929***\n0.952***\n0.969***\n0.833***\n0.857***\n0.872***\n1.110***\n0.933***\n0.941***\nSIS1979 Q3\n1.304***\n1.233***\n1.233***\n1.336***\n1.271***\n1.267***\n1.312***\n1.300***\nSIS1981 Q1\n\u22121.754***\n\u22121.789***\nSIS1981 Q3\n\u22121.717***\n\u22121.725**\n\u22121.756***\n\u22121.761***\n\u22121.764***\n\u22121.778***\n\u22121.780***\nSIS2004 Q3\n1.205***\n1.242***\n1.230***\n1.244***\n1.279***\n1.259***\n1.278***\n1.335***\n1.317***\nSIS2007 Q4\n1.564***\n1.608***\n1.599***\n1.599***\n1.623***\n1.613***\n1.637***\n1.653***\n1.642***\nSIS2008 Q4\n\u22122.269***\n\u22122.398***\n\u22122.377***\n\u22122.418***\n\u22122.537***\n\u22122.358***\n\u22122.404***\n\u22122.556***\n\u22122.380***\nSIS2009 Q3\n1.134***\n1.141***\nSIS2009 Q4\n1.090***\n1.080***\n1.106***\n1.058***\n1.100***\n1.062***\nSIS2010 Q1\n1.029***\nSIS2011 Q2\n0.613***\n0.648***\n0.636***\n0.670***\n0.731***\n0.633***\n0.673***\n0.728***\n0.635***\nSIS2014 Q4\n\u22121.564***\n\u22121.596***\n\u22121.601***\n\u22121.603***\n\u22121.628***\n\u22121.627***\n\u22121.548***\n\u22121.664***\n\u22121.661***\nOrdinary least squares (OLS) estimate of ECM (see equation (6) in the Methods section)\n\u03bct\u22121\n\u22120.17**\n\u22120.19***\n\u22120.19***\n\u22120.16**\n\u22120.21***\n\u22120.18***\n\u22120.19***\n\u22120.17***\n\u22120.16***\nIndependent variables include US refinery utilization rates (UtilRef), days of forward consumption in US oil inventories (Days), capacity utilization by Texas Railroad Commission (UtilTRC) and capacity \nutilization by OPEC (UtilOPEC). DOLS regression results for the sample period 1938:1\u20132018:3 (denoting year:quarter) for the cointegrating relation given by Models 1\u20139 (Equation (4)) and the error \ncorrection term (\u03c1) that is associated with the lagged equilibrium error (\u03bct\u22121) from the ECM (equation (6)). Note that all models represent cointegrating relations for price. Periods when price changes \nfrom the equilibrium price in a statistically significant fashion (steps) are given by step indicator saturation (SIS) followed by the year and quarter when the deviation occurs. Test statistics reject the null \nhypothesis at *P\u2009<\u20090.05, ** P\u2009<\u20090.01 and ***P\u2009<\u20090.001.\nNature Energy | VOL 5 | February 2020 | 141\u2013149 | www.nature.com/natureenergy\n142\n\nArticles\nNATUrE EnErgy\noccur after OPEC gains control over the marginal supply of crude \noil in the early 1970s (Regime 2).\nOnce the cointegrating relation allows for regimes, UtilOPEC \ndoes not have a statistically measurable long- or short-run relation \nwith prices. The lack of a significant relation does not indicate that \nOPEC is unimportant. Rather, the lack of a relation may be caused \nby changes in OPEC behaviour, such as a change from acting as the \nmarginal supplier to defending a fair share of the market (and vice \nversa). Such changes alter the relation between UtilOPEC and Price; \nthe relation would be positive when OPEC acts as the marginal sup-\nplier and negative when OPEC acts to defend a fair share of the mar-\nket. Such changes in sign prevent the statistical methodology from \nquantifying a coefficient for UtilOPEC that applies over the entire \nsample period.\nTo understand price regimes in Model 1, we relate each to real-\nworld events and explain why some changes noted by previous \nanalyses do not appear in Table 2. For regimes prior to 2000, expla-\nnations borrow heavily from a chronology of the world oil market7; \nafter 2000, explanations rely heavily on articles in the industry \nnewsletter Petroleum Intelligence Weekly. We recognize that associ-\nating real-world events to regimes represents post hoc explanations. \nThis cannot be avoided because there is no a\u00a0priori definition for a \nprice regime and it is very difficult to quantify time series that could \nrepresent such regimes.\nRegime 1, 1946:1\u20131949:3, witnesses a price reduction by an aver-\nage of US$9.26 relative to the equilibrium price implied by market \nfundamentals (Supplementary Table 3). This reduction is opposite \nthe price increase that is associated with the elimination of price \ncontrols on motor gasoline after World War II. During the war, con-\ntrols create a lack of capacity, as indicated by chronic shortages in \ndrilled wells and transportation facilities8,9. When controls are lifted, \nprice increases are suppressed by non-market interventions. For \nexample, Standard Oil of Indiana and Phillips Petroleum Company \nration gasoline to dealers in June 1947 (ref. 10). To stimulate new \ncapacity, the US federal government relaxes anti-trust legislation to \n\u201330\n\u201320\n\u201310\n0\n10\n20\n30\n40\n50\n60\n1938:1\n1948:1\n1958:1\n1968:1\n1978:1\nYear:quarter\n1988:1\n1998:1\n2008:1\n2018:1\nPrice (US$1982 per barrel)\nRegime 1 US$9.26\nRegime 2 US$11.03\nRegime 3\nUS$1.19\nRegime 4\nUS$12.63\nRegime 5 US$5.95\nQuarterly change\nin price\nRegime 6\nUS$6.82\nRegime 9\nUS$17.85\nRegime 7 US$23.39\nRegime 8 US$10.90\nSimulated\nprice\nObserved price\nEquilibrium \nprice\nFig. 1 | Modelled and observed prices under different regimes. The nine regimes identified from Model 1 by the indicator saturation technique. The real \n(US$1982 per barrel) price for crude oil simulated by the cointegrating relation (blue line) and ECM (red line) and the observed price (black circles). Note that \nthe size of quarterly changes in real oil prices (black line) increase after OPEC takes control over real oil prices during Regime 2. Note that the model simulates \nobserved prices fairly well during Regime 5 without separate regimes for the 1986 price collapse, the Asian financial crisis and the first Persian Gulf War.\nTable 2 | Nine regimes identified by Model 1\nRegime\nStart\nEnd\nCause\nRegime 1\n1946:1\n1949:3\nRationing and price suppression\nRegime 2\n1970:3\n1974:1\nOPEC gains control over \ndomestic resources\nRegime 3\n1974:1\n1979:2\nUS price controls\nRegime 4\n1979:3\n1980:4\nPanic buying after the  \nIranian Revolution\nRegime 5\n1981:1\n2004:2\nOPEC fair share of the market\nRegime 6\n2004:3\n2007:3\nReallocation of quotas for  \nOPEC members\nRegime 7\n2007:4\n2008:3\nSpeculative bubble\nRegime 8\n2010:1\n2011:1\nSpeculative bubble\nRegime 9\n2011:2\n2014:3\nArab Spring\nRegimes start in the quarter when a step changes the equilibrium price for two or more  \nquarters. Regimes end when the next step either moves the equilibrium price back towards or \nfurther away from the equilibrium that is implied by market fundamentals. The cause for  \nthe regime is based on real-world events. Explanations for regimes prior to 2000 borrow  \nheavily from a chronology of the world oil market7; after 2000, explanations rely heavily on  \narticles in the industry newsletter Petroleum Intelligence Weekly. We recognize that associating  \nreal-world events to regimes represents post hoc explanations. This cannot be avoided because \nthere is no a\u00a0priori definition for a price regime and it is very difficult to quantify time series  \nthat could represent such regimes.\nNature Energy | VOL 5 | February 2020 | 141\u2013149 | www.nature.com/natureenergy\n143\n\nArticles\nNATUrE EnErgy\nenhance cooperation among US oil firms. But one form of coopera-\ntion is forbidden: \u201cIn February 1948, the attorney general explicitly \nexcluded from antitrust prosecution cooperation among oil compa-\nnies so long as this cooperation attempted to alleviate shortages by \nany mechanism other than raising price.\u201d (ref. 8).\nRegime 1 ends in 1949:3 when the second step increases prices \nby US$7.42, which moves the equilibrium price back to the level \nimplied by market fundamentals. We attribute the end of Regime 1 \nto higher production and lower consumption, which relax the need \nfor market interventions. Informal rationing ends when the 1949 \nrecession reduces consumption. Its effect on the supply/demand \nbalance is reinforced by greater US output and the 1950 opening of \nthe Trans-Arabian Pipeline.\nRegime 2, 1970:3\u20131974:1, often is associated with a series of \nOPEC-induced price rises4. Here, Regime 2 represents an US$11.03 \nprice reduction. We reconcile this seeming contradiction by arguing \nthat Regime 2 is the period when the marginal supply of crude oil \nshifts geographically from the TRC (US) to OPEC, but this change \nin market fundamentals does not affect price because production \nagreements allow international oil companies to set production \nfrom fields in OPEC nations. Regime 2 ends when OPEC nations \ngain control over marginal supply and demonstrate this control by \ncutting production and imposing an embargo on the US in 1973:4.\nBefore the start of Regime 2, UtilTRC rises rapidly and reaches \n100% in the middle of Regime 2, 1972:2 (Fig. 2). The loss of spare \ncapacity means that Texas no longer produces the marginal barrel of \noil, it comes from fields in OPEC nations. But this increase in UtilTRC \ndoes not increase price as indicated by the cointegrating relation \nbecause concession agreements allow international oil companies \nto expand production in low-cost OPEC nations. Concession agree-\nments specify the terms for international oil companies to pay host \ncountries for the oil produced, but allow international oil companies \nto control production. Because the Seven Sisters control production, \nthere is no \u2018market mechanism\u2019 to raise the price when UtilTRC rises \nto capacity and the marginal supply shifts from the TRC to OPEC.\nBut the ability of international oil companies to control produc-\ntion from fields in OPEC nations ends during Regime 2. OPEC \nnations nationalize oil holdings and/or replace concessions with \nparticipation agreements7. These changes are not fully appreciated \nby the market; their import is recognized by a select few who wield \nlittle influence. In November 1970, James Placke, a petroleum offi-\ncer in the US embassy in Tripoli, warns of potential consequences, \nbut there is little reaction7. Similarly, few heed a warning that the US \nis losing control over the marginal supply of oil11. Instead, Regime 2 \nends in 1974:1 when OPEC demonstrates control over its produc-\ntion and hence marginal supply by cutting production and Saudi \nArabia stops selling oil to the US and a few other nations.\nRegime 3, 1974:1\u20131979:2, starts when OPEC demonstrates con-\ntrol over marginal supply, which raises the price by US$9.84. But \nthis increase leaves prices slightly (US$1.190.51) below the price \nimplied by market fundamentals. Lower prices may be associated \nwith non-market responses to OPEC\u2019s control over marginal supply. \nUS energy legislation in the 1970s is the largest peacetime inter-\nference with the economy, which has many (contradictory) effects, \nincluding lower prices12,13. For example, divisions between \u2018new\u2019 and \n\u2018old\u2019 oil keep the wellhead price for much of US production (includ-\ning West Texas intermediate, WTI) below world prices.\nRegime 4, 1979:3\u20131980:4, is associated with a precautionary \ndemand shock that raises the price by US$12.63 relative to that indi-\ncated by market fundamentals. Prior to the start of Regime 4, revo-\nlutionary chaos in Iran subsides and Iranian production of crude oil \nrises from 0.73\u2009mbd in February 1979 to 4.2\u2009mbd in April 1979 (ref. 14).  \nCommensurate with this increase, Saudi Arabia cuts production by \nabout 2\u2009mbd back to its official ceiling of 8.5\u2009mbd. This rebalancing \n\u201cwas the moment at which some kind of order, well short of disaster \nmight have been reestablished\u201d (ref. 7).\nBut order is not re-established. Instead, panic buying pushes pur-\nchases well beyond current consumption such that inventories rise \nsignificantly7. Consistent with previous analyses15,16, we argue that \nRegime 4 is a precautionary demand shock. During Regime 4, US oil \ninventories rise from 15 to 23\u2009days of forward consumption (Fig. 3).  \nThe positive price effect of this inventory build is not consistent \nwith the negative relation that is implied by market fundamentals17.\nRegime 5, 1981:1 (or 1981:3)\u20132004:2, starts with a US$18.58 \nprice drop due to declines in both inventories and consumption18. \nThis drop pushes prices US$5.95 below the levels indicated by the \ncointegrating relation (Fig. 1). Regime 5 includes the 1986 change in \nOPEC strategy, when OPEC abandons its official price and instead \nseeks a fair share of the market. But we argue that this change starts \nbefore 1986, when OPEC\u2019s official price simply formalizes changes \nin refined product prices19,20. Note that Regime 5 includes several \nlarge price changes, such as the 1986 price collapse, the Asian finan-\ncial crisis and the first Persian Gulf War. But these events do not \nchange prices in a statistically significant fashion relative to that pre-\ndicted by Regime 5 because they are largely captured by the proxies \nfor market fundamentals in the cointegrating relation (Fig. 1). For \nexample, the 13% price reduction during the Asian financial crisis is \nassociated with increases in Days and UtilRef.\nRegime 6, 2004:3\u20132007:3, is defined by a US$12.77 price increase. \nOPEC reallocates national quotas in a way that enhances its ability \nto control output, which pushes the price US$6.82 above that indi-\ncated by market fundamentals.\nAt its December 2004 meeting, OPEC suspends its US$22\u201329 \nprice range and reallocates production quotas based on existing \nrates of production as opposed to the existing quotas. This change \nis described as a \u201cpragmatic way of gradually leaving the old sys-\ntem behind without an open squabble over quota redistributions\u201d \n(ref. 21). Although OPEC cannot enforce quotas, quotas influence \nobserved levels of production22,23. As such, harmonizing quotas \nwith operable capacity reduces the incentive to cheat and increases \nOPEC\u2019s ability to control marginal supply.\n20\n30\n40\n50\n60\n70\n80\n90\n100\n110\n1938:1\n1948:1\n1958:1\n1968:1\nYear:quarter\n1978:1\n1988:1\n1998:1\n2008:1\nOil supply capacity utilized (%)\nUtilOPEC\nUtilTRC\nRegime 1\nFig. 2 | OPEC gains control over marginal supply from the TRC. \nThe percentage of capacity allowed to operate by the Texas Railroad \nCommission (UtilTRC, black line) and the percentage of capacity allowed \nto operate by OPEC (UtilOPEC, red line). The two lines cross in 1973:1. As \nUtilTRC approaches 100%, the TRC loses control over marginal supply. \nAfter Regime 2, control over the marginal supply shifts to OPEC because \nUtilOPEC is well below 100%. The timing of Regime 2 (a US$11.03 \nreduction in price) is indicated by the area shaded blue.\nNature Energy | VOL 5 | February 2020 | 141\u2013149 | www.nature.com/natureenergy\n144\n\nArticles\nNATUrE EnErgy\nRegime 7, 2007:4\u20132008:3, raises the price by US$16.57 relative to \nRegime 6. We postulate that this increase is a speculative bubble, in \nwhich noise investors raise crude oil prices relative to the equilib-\nrium that is implied by the proxies for market fundamentals in the \ncointegrating relation. A speculative bubble is consistent with the \nexplanations offered by industry and academia. One quarter into \nRegime 7, Rex Tillerson, then CEO of Exxon Mobil states \u201cI cannot \nexplain why we have US$70 oil. The fundamentals behind supply \nand demand do not support US$70 oil. The fundamentals support \nsomething much less.\u201d (ref. 24). Academic researchers recognize the \npotential importance of a speculative bubble: the peer-review lit-\nerature contains seven articles that mention \u2018oil\u2019, \u2018prices\u2019 and \u2018specu-\nlation\u2019 before the end of Regime 7. As of February 26, 2019, 173 \narticles appear.\nThe role of a speculative bubble can be tested indirectly by exam-\nining the relation between Regime 7 (and Regime 8, see below) and \nthe discussions of a speculative bubble by industry experts. To eval-\nuate this relation, we compile information on the number of articles \nin the industry newsletter Petroleum Intelligence Weekly that include \nthe words \u2018oil\u2019, \u2018price\u2019 and \u2018speculation\u2019, with speculation referring to \noil prices as opposed to guessing about future events. If our hypoth-\nesis about a speculative bubble is correct, Regime 7 (and Regime 8) \nshould be marked by an increase in the number of articles about \nspeculation. Consistent with the visual impression given by Fig. 4,  \na logit model (Regime 7t \u00bc \u03b1 \u00fe \u03b2*Articlet \u00fe \u03b5t\nI\n) indicates that a \nbinary variable for Regime 7 is related to the number of articles that \nmention speculation (^\u03b2 \u00bc 1:00; t \u00bc 2:89; p<0:01\nI\n).\nWe recognize that a speculative bubble may be a catch-all for \nthe inability to describe the effect of market fundamentals (or some \nother non-market force) on prices during Regime 7. But it is diffi-\ncult to relate a speculative bubble to a concrete measure. Theoretical \nmodels indicate that a speculative bubble moves prices away from \nfundamentals by increasing the risk premium for arbitrageurs25. \nConsistent with this mechanism, Regime 7 is associated with an \nincrease in the risk premia that is caused by and causes a complex \nseries of changes in long and short positions by both non-commer-\ncial and commercial traders26. These complex sequences may be one \nreason that simple measures of trader positions fail to validate the \nimportance of speculation1,2.\nRegime 7 ends with the start of the financial crisis in 2008:4. The \nfinancial crisis reduces interest rates and convenience yields, which \nreduces returns to holding oil as a commodity relative to a financial \nasset27. This changes the correlation between returns to crude oil \nand the S&P 500 from negative to positive, which means that lower \nequity prices at the start of the financial crisis correlate with lower \noil prices. Note that the end of Regime 7 returns the equilibrium \nprice to the level implied by market fundamentals.\nRegime 8, 2010:1\u20132011:1, raises the equilibrium price by \nUS$10.90 relative to that implied by market fundamentals. Although \nthis period coincides with cuts in production by OPEC, their effects \non market fundamentals are offset by increased production and \nslowed demand such that inventories rise28. As such, reduced pro-\nduction by OPEC probably cannot account for Regime 8. Instead, \nwe argue that Regime 8 also is associated with a speculative bubble.\nBefore Regime 8 starts, the divergence between market funda-\nmentals and prices is highlighted by price forecasts that consistently \nunder-predict observed increases. The Center for Global Energy \nStudies warns that nominal prices are unlikely to rise above US$40 \nper barrel29 and market analysts forecast nominal prices to average \nUS$50 during 2009 (US$42 to US$57)30. By the first quarter of 2010, \nthe price of WTI rises beyond US$75 and remains there for the rest \nof the year.\nWhen asked about these price increases, the Secretary General \nof OPEC, Abdullah al-Badri, states, \u201cIf there is a problem with fun-\ndamentals then we would be concerned, but if the fundamentals are \nokay and there is enough oil in the market and the prices shoots to \nUS$147 as what happened in 2008, this is not our problem. This is \na speculation problem.\u201d (ref. 31). Consistent with this interpretation, \nFig. 4 shows a second rise in articles that mention \u2018speculation\u2019, but \nthe relation between a binary variable for Regime 8 and the number \nDays of forward consumption\n0\n5\n10\n15\n20\n25\n30\n1975:1\n1976:1\n1977:1\n1978:1\n1979:1\n1980:1\n1981:1\n1982:1\n1983:1\n1984:1\n1985:1\nRegime 4\nYear:quarter\nFig. 3 | Precautionary demand and Regime 4. Uncertainty about the \nability to purchase oil induces firms to increase their inventories, which is \nindicated by a steady increase in days of forward consumption (Days) in US \ninventories (black line) during Regime 4 (area shaded red). This increase \nrepresents a precautionary demand shock that raises the equilibrium price \nby US$12.63 relative to that simulated by market fundamentals. Inventories \ndecline to a more \u2018normal level\u2019 after Regime 4 because of fears about \nsupply diminish.\n0\n1\n2\n3\n4\n5\n6\n7\n8\n9\n1998:1\n1999:1\n2000:1\n2001:1\n2002:1\n2003:1\n2004:1\n2005:1\n2006:1\n2007:1\n2008:1\n2009:1\n2010:1\n2011:1\n2012:1\n2013:1\n2014:1\n2015:1\n2016:1\n2017:1\n2018:1\nNumber of articles\nYear:quarter\nFig. 4 | Discussions of a speculative bubble. The degree to which the \nindustry thinks that a speculative bubble inflates the price of oil, as \nindicated by the number of articles in Petroleum Intelligence Weekly that \ncontain the words \u2018oil\u2019, \u2018price\u2019 and \u2018speculation\u2019, with speculation referring \nto price speculation. The first set of red bars occur during Regime 7 and \nthe second set of red bars occur during Regime 8. A logit model indicates \nthat Regime 7 is related to the number of articles (P\u2009<\u20090.01), whereas the \nrelation with Regime 8 is significant at P\u2009=\u20090.06.\nNature Energy | VOL 5 | February 2020 | 141\u2013149 | www.nature.com/natureenergy\n145\n\nArticles\nNATUrE EnErgy\nof articles that mention \u2018speculation\u2019 is significant only at P\u2009=\u20090.10 \n(^\u03b2 \u00bc 0:38; t \u00bc 1:93; P<0:06\nI\n).\nRegime 9, 2011:2\u20132014:3, raises the equilibrium price by  \nUS$6.95 relative to Regime 8. We argue that Regime 9 is associated \nwith the Arab Spring, which refers to political changes in Tunisia, \nEgypt and Libya, and worries about political stability throughout \nthe region. These political changes alter market fundamentals, \nbut they are not captured by proxies in the cointegrating rela-\ntion. There is little change in UtilOPEC because political changes \nin Libya reduce both production and capacity (as reported by the \nEnergy Information Administration, EIA). Nor are these political \nchanges captured by crude oil inventories (Days). In June 2011, the \nInternational Energy Agency releases 60 million barrels from stra-\ntegic reserves, including 30 million barrels from the US Strategic \nPetroleum Reserve (SPR)32. But the SPR is not part of Days, and so \nthis release is not represented in the cointegrating relation.\nRegime 9 ends (2014:4) with the 2014 price collapse, which is \ntriggered by increasing production from tight formations in the \nUS, which rises 3.6\u2009mbd between April 2011 and November 2014. \nThis forces OPEC to choose between defending price by cutting \nproduction to accommodate this additional supply or defending its \nshare of the market by letting the increased production of tight oil \nreduce prices33\u201335. By choosing not to cut production, OPEC hopes \nthat lower prices will reduce the profitability of producing oil from \ntight formations, which would lower production. But this strategy is \nnot successful. After a brief decline, tight oil production resumes its \nincrease because technical innovations increase the productivity of \nrigs drilled into tight formations and the break-even price (~US$50 \nnominal) is lower than previously thought and is consistent with the \nprices implied by the market fundamentals that prevail after Regime \n9 (ref. 36).\nThe presence of a low-cost alternative prompts discussion that \nthe United States now controls marginal supply37. But tight oil can-\nnot be considered marginal supply because there is no spare capac-\nity38. Instead, tight oil may act in the medium term to impose some \ndiscipline on prices. As such, it may discourage OPEC actions to \ncreate the next \u2018oil price regime.\u2019 And this may be one reason that \nthe end of Regime 9 returns oil prices back to the level implied by \nmarket fundamentals, where they remain through to the end of the \nsample period 2018:3.\nConclusion\nWe associate two regimes (Regimes 7 and 8) with a speculative \nbubble, in which prices move away from the level implied by the \nbasic, underlying factors that affect the supply or demand for a com-\nmodity. A speculative bubble is different from speculative trading, \nwhich is an integral part of price discovery in spot and future mar-\nkets39\u201341. Speculative trading is critical because it allows commer-\ncial participants to hedge, which transfers risk to non-commercial \nparticipants42. This transfer increases the efficiency of commercial \noperations.\nBut if spot and futures markets do not operate properly, a specu-\nlative bubble can impose significant costs. During Regimes 7 and 8, \nspeculative bubbles raise prices by US$14.31 and US$4.65 per bar-\nrel, respectively (Table 3 and Supplementary Note 3). These price \nincreases transfer US$42.9 billion from US consumers to US oil \nproducers and transfer US$87.4 billion from the US economy to oil \nexporting nations. These totals suggest transfers that reduce total \nsocial welfare. Such welfare losses suggest that regulations, which \nprevent speculative bubbles, can avoid significant costs.\nMethods\nData. In this study we compile quarterly observations for oil prices and proxies \nfor the balance between oil supply and demand from 1938:1 through 2018:3, \nwhich is the entire period for which observations are available. This long sample \nperiod poses two challenges. First, many time series are not available over the \nentire sample period. Instead, we use observations for the data available, even \nwhen they are not the best choice. For example, we proxy market fundamentals \nusing observations for US oil inventories and utilization rates for US refineries. We \nprefer observations for global (or the Organisation for Economic Co-operation \nand Development, OECD) inventories and global rates of refinery utilization, but \nobservations are limited; observations for OECD inventories start in the 1980s.\nFor some time series, the frequency of observations changes during the \nsample period. For these series, we use a simple algorithm to interpolate annual \nTable 3 | The effect of speculative bubbles on the price of crude oil and the monies transferred between consumers and producers\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nModel 6\nModel 7\nModel 8\nModel 9\n\u2206Price (US$1982 per barrel)\n Regime 7\n14.31\n13.21\n13.24\n10.39\n13.05\n13.35\n13.04\n13.34\n13.40\n Long-run\n16.57\n17.04\n16.94\n16.94\n17.20\n17.09\n17.34\n17.51\n17.40\n Regime 8\n4.65\n5.08\n5.05\n5.16\n6.33\n5.04\n5.01\n6.62\n6.08\n Long-run\n10.90\n11.55\n11.44\n11.72\n12.01\n11.21\n11.65\n12.09\n11.25\nDomestic transfers (Billion US$1982)\n Regime 7\n29.05\n26.83\n26.89\n21.12\n26.52\n27.11\n26.49\n27.09\n27.21\n Regime 8\n13.80\n15.05\n14.96\n15.28\n19.59\n14.91\n14.82\n19.29\n23.82\n Total\n42.85\n41.88\n41.85\n36.40\n46.11\n42.01\n41.31\n46.37\n51.03\nForeign transfers (Billion US$1982)\n Regime 7\n64.77\n59.78\n59.90\n47.03\n59.07\n60.39\n59.00\n60.35\n60.63\n Regime 8\n22.66\n24.79\n24.65\n25.20\n33.10\n24.57\n24.41\n32.63\n23.82\n Total\n87.43\n84.57\n84.55\n72.22\n92.18\n84.96\n83.41\n92.98\n84.45\nTotal transfer (Billion US$1982)\n Regime 7\n93.83\n86.61\n86.79\n68.15\n85.59\n87.50\n85.49\n87.44\n87.85\n Regime 8\n36.45\n39.84\n39.61\n40.48\n52.69\n39.47\n39.23\n51.92\n47.63\n Total\n130.28\n126.45\n126.39\n108.63\n138.28\n126.97\n124.72\n139.35\n135.48\nThe first panel reports the increases in the equilibrium and predicted price during Regimes 7 and 8. The second panel reports the total funds transferred from US consumers to US producers due to the price \nincreases listed in the first panel. The third panel reports the total funds transferred from the US economy to foreign producers due to the price increases listed in the first panel. The fourth panel reports \nthe sum of funds transferred from consumers to producers as reported in the second and third panels. These totals suggest transfers that reduce total social welfare. Such welfare losses suggest that \nregulations, which prevent speculative bubbles, can avoid significant costs.\nNature Energy | VOL 5 | February 2020 | 141\u2013149 | www.nature.com/natureenergy\n146\n\nArticles\nNATUrE EnErgy\nobservations into quarterly observations. To interpolate annual observations for \nUS oil inventories before 1972 into quarterly observations, we create an adjustment \nfactor (QAdj) for each quarter q (q\u2009=\u20091, 2, 3, 4) as follows:\nQAdjq \u00bc\nPEnd\nY\u00bcStart\nXY;q\n\u001fXY\nEnd \u001e Start\n\u00f01\u00de\nin which X is the observed value for one of the indepdent variables X, in year Y for \nquarter q over the sample period for which quarterly observations are available. \nThese adjustment factors are used to convert annual values \u001fXY\nI\n, which represent the \naverage for year Y, into quarterly values (~XY;q\nI\n) as follows:\n~XY;q \u00bc QAdjq*\ue016XY\n\u00f02\u00de\nNominal oil prices are measured using quarterly observations for West Texas \nintermediate (WTI) (http://www.globalfinancialdata.com)43. We recognize that \nprices are specific to the quality of the crude oil44, that prices between 1973 and \n1981 are affected by Federal efforts to control the price of old and new oil produced \nin the US13, and that WTI becomes a weaker benchmark after 2011 such that the \nEIA abandons WTI as its reference oil price in the 2013 Annual Energy Outlook45. \nNonetheless, this is the only measure for the price of crude oil that is available for \nthe entire sample period. Nominal prices are deflated (base year 1982) by the US \ncity average for all items (CUUR0000SA0), which is obtained from the Bureau of \nLabor Statistics46 to calculate the real price (Price).\nWithout regimes, real prices for crude oil are determined by the balance \nbetween supply and demand. We proxy this balance with capacity utilization \nby producer organizations that control marginal supply, inventories of crude oil \nand rates of refinery utilization. We specify capacity utilization by two producer \norganizations: the TRC and OPEC. We expect prices to have a positive albeit non-\nlinear relation with capacity utilization47\u201349. Higher rates of capacity utilization by \nthe marginal supplier signal less spare capacity. This suggests that consumption \nis high relative to supply, supply disruptions cannot be offset by increasing \nproduction elsewhere and there is relatively little opportunity for individual \nproducers to cheat on production agreements. Under these conditions, prices \ntend to rise. Conversely, low rates of capacity utilization by the marginal supplier \nindicate higher amounts of spare capacity. This suggests that consumption is \nlow relative to supply, supply disruptions can be accommodated by increased \nproduction elsewhere and there is greater opportunity for individual producers to \ncheat on production agreements. Under these conditions, prices tend to decline.\nThe TRC\u2019s control over the marginal supply of crude oil is conferred by \nthe Connelly Hot Oil Act of 1935, which gives the TRC and other US state \ncommissions the power to open and close existing capacity with the goal of \nstabilizing price50. To stabilize price, the TRC determines the rate of output that \nwould balance supply and demand. This rate is enforced by controlling the number \nof days per month that well owners are allowed to pump oil, which is known as \nthe market demand factor. We divide this market demand factor by the number of \ndays in a month to calculate the percentage of Texas capacity that the TRC allows \nto operate (UtilTRC), which varies between 26 and 101% (Fig. 2). Monthly data \nfrom January 1938 through December 1986 are obtained from the TRC. Monthly \nobservations of 100% from March 1977 through December 1986 are extended \nthrough September 2018, which indicates that producers are free to choose the rate \nat which they operate.\nThis loss of spare capacity gives OPEC control over the marginal supply of \ncrude oil. We calculate capacity utilization by OPEC (UtilOPEC) as the quotient \nof OPEC production of crude oil18 and OPEC\u2019s operable capacity51 (Fig. 2). For the \nperiod before OPEC is founded in September 1960, we extrapolate the value from \n1960:1, 97.1%.\nWe also proxy the balance between supply and demand with inventories \nof crude oil. Inventories rise when demand falls relative to supply, which puts \ndownward pressure on price17. Conversely, prices rise when inventories fall, which \noccurs when demand rises relative to supply. This balance between demand and \nsupply is measured by days of forward consumption (Days), which is calculated \nby dividing barrels of crude oil in US inventories (excluding SPR)52,53 by the US \nconsumption of refined petroleum products.\nFinally, we proxy the balance between supply and demand with the utilization \nrates for US refiners (UtilRef), which are calculated by dividing US refinery runs \nby US refinery capacity54,55 and multiplying by 100. UtilRef is expected to have \na negative relation with price49. As utilization rates near capacity, there is less \ndemand for new supply, which reduces price.\nTo eliminate the effects of inverting matrices with elements that differ greatly \nin size due to different units of measure, each of the time series described above is \nstandardized as follows:\nxt \u00bc\nyt \u001e \u001fy\n\u00f0\n\u00de\n\ufb03\ufb03\ufb03\ufb03\ufb03\ufb03\ufb03\ufb03\ufb03\ufb03\ufb03\ufb03\ufb03\ufb03\nVar\u00f0y\u00de\np\n\u00f03\u00de\nin which yt is the value (in original units), \u001fy\nI\n is the average value over the sample \nperiod and Var(y) is the variance over the sample period. We recognize that non-\nstationary time series do not have a constant mean or variance, rather the sample \nmean and variances are used in a linear transformation, given by equation (3), to \nharmonize values across time series.\nWe recognize that misreporting data could contribute to regimes. If OPEC \nunder-reports oil production, this could affect the relation between market \nfundamentals (as reported by OPEC) and oil prices. But this effect could go beyond \ndata artefacts. If buyers believe that OPEC is producing less oil (as indicated by \nits announcement) than it actually does, buyers may act accordingly and thereby \nboost prices. Nonetheless, we do not believe that our regimes reflect poor data \nbecause observations for many variables are compiled from primary sources and \nnot official sources, such as OPEC announcements.\nEconometric methodology. The first step of the econometric methodology \nidentifies a cointegrating relation for the real price of oil from an unrestricted \nmodel, which is given by:\nPricet \u00bc f UtilTRCt; UtilTRC2\nt ; UtilTRC3\nt ; UtilOPECt; UtilOPEC2\nt ;\n\u001f\nUtilOPEC3\nt ; Dayst; UtilRef t\n\u001e\n\u00f04\u00de\nfor which the variables have been defined previously.\nWe identify a final version of equation (4) by estimating versions that specify all \npossible combinations of utilization rates by the TRC and OPEC (Supplementary \nNote 1) and evaluating each model against three criteria: do the variables \ncointegrate, do the variables have a statistically measurable relation with oil prices \nand do oil prices error correct to disequilibrium in the cointegrating relation?\nWe evaluate cointegration by estimating each possible specification using OLS \nand testing the residual for a unit root with four statistics (PT, DFGLS, QT and \nDFGLSu)56. Statistics that reject the null hypothesis indicate that the residual does \nnot contain a unit root, which indicates that the variables cointegrate and therefore \nrepresent a long-run cointegrating relation for price.\nThe long-run relation between Price and the proxies for the balance \nbetween supply and demand are estimated using DOLS57. Lags and leads for \nthe first differences of the independent variables are chosen using the Bayesian \ninformation criterion58. We test the null hypothesis that independent variable \ni has no statistically measurable relation with Price (\u03b2i\u2009=\u20090) with a t-test that is \ncalculated with a standard error, which is robust to the presence of autocorrelation \nand heteroscedasticity in the regression residual59. Rejecting the null hypothesis \nindicates that independent variable i has a statistically measurable relation  \nwith Price.\nWe test whether Price adjusts to disequilibrium in the cointegrating  \nrelation by estimating an ECM. Disequilibrium (\u03bc) is calculated by subtracting  \nthe equilibrium price implied by the cointegrating relation from observed  \nprice (^\u03bct \u00bc Pricet \u001e ^Pequilt\nI\n). Disequilibrium from the cointegrating  \nrelation estimated by the saturation indicator technique is calculated as  \n\u03bct \u00bc Price \u001e\nd\n^Pequilt \u00fe P\nt\ni\u00bc1\nStept\nI\n in which Step is a step estimated by the indicator \nsaturation technique.\nDisequilibrium is specified in an ECM as follows:\n\u0394Pricet \u00bc \u03b1 \u00fe \u03c1^\u03bct\u001d1 \u00fe\nX\ns\ni\u00bc1\n\u03b3t \u0394Pricet\u001d1 \u00fe\nX\nn\nj\u00bc1\nX\ns\ni\u00bc1\n\u03b8j;t \u0394Xj;t\u001d1 \u00fe \u03b5t\n\u00f05\u00de\nin which X is a vector of n independent variables in the cointegrating relation, S is \nchosen using the Akaike information criterion60, \u0394 is the first difference operator \n(xt\u2013xt\u22121), \u03b1, \u03c1, \u03b3 and \u03b8 are regression coefficients estimated using OLS, and \u03b5 is the \nregression residual. Price adjusts towards the equilibrium that is implied by the \ncointegrating relation if the error correction mechanism \u22121\u2009\u2264\u2009\u03c1\u2009<\u20090.\nFor models in which the proxies for market fundamentals cointegrate with \nprice, have a statistically measurable relation with price and oil prices error correct \nto disequilibrium in the cointegrating relation, we identify price regimes using an \nindicator saturation technique that is implemented in the R package gets (refs. 61,62). \nWe \u2018fix\u2019 the independent variables used in the DOLS estimate (along with the lags \nand leads of the first differences), use a P\u2009=\u20090.01 significance level and allow gets to \nchoose from a full set of impulses and/or steps. Impulses are a one-quarter change \nin the equilibrium price for crude oil relative to that simulated by the proxies for \nmarket fundamentals in the cointegrating relation, while steps are changes that \npersist for two or more consecutive quarters. Impulses and steps are evaluated \niteratively for every possible quarter. The method used to calculate the statistical \nsignificance of impulses and steps is summarized in Supplementary Note 2.\nShort-run price dynamics are quantified by an ECM which is given by:\n\u0394Pricet \u00bc \u03b1i \u00fe \u03c1\u03bci;t\u001d1 \u00fe P\ns\nj\u00bc1\n\u03a8j\u0394Pricet\u001dj \u00fe P\ns\nj\u00bc1\n\u03bej\u0394Dayst\u001dj\n\u00fe P\ns\nj\u00bc1\n\u03b3j\u0394UtilRef t\u001dj\nP\ns\nj\u00bc1\n\u03b8j\u0394UtilTRCt\u001di \u00fe P\ns\nj\u00bc1\n\u03bbj\u0394UtilOPECt\u001dj\n\u00fe P\nk\nR\u00bc1\nP\ns\nj\u00bc0\n\u03c8R;j\u0394StepR;t\u001dj \u00fe \u03c5i;t\n\u00f06\u00de\nin which Step is each of the k steps in the cointegrating relation as defined above. \nThe number of lags (s) is chosen using the Akaike information criterion60 and the \nstandard errors are robust to the presence of autocorrelation and heteroscedasticity \nin the regression residual59.\nNature Energy | VOL 5 | February 2020 | 141\u2013149 | www.nature.com/natureenergy\n147\n\nArticles\nNATUrE EnErgy\nData availability\nData on US petroleum stocks, US refinery utilization rates, prices for WTI (earlier \nobservations from the Federal Reserve Bank of St. Louis43), OPEC production \nand OPEC capacity are available from US EIA (ref. 32), data for the US city average \nfor all items are available from the US Bureau of Labor Statistics46 and the TRC \ndemand factors are available from the Texas Railroad Commission. These data and \nthe computer code can be obtained from authors upon reasonable request. Source \ndata for Figs. 1\u20134 are provided with this paper.\nCode availability\nThe code can be obtained from the authors upon reasonable request.\nReceived: 20 May 2019; Accepted: 3 January 2020;  \nPublished online: 17 February 2020\nReferences\n\t1.\t Boyd, N. E., Harris, J. H. & Li, B. An update on speculation and \nfinancialization in commodity markets. J. Commod. Mark. 10, 91\u2013104 (2018).\n\t2.\t Haase, M., Zimmerman, Y. S. & Zimmerman, H. The impact of speculation \non commodity futures markets \u2013 A review of the finding of 100 empirical \nstudies. J. Commod. Mark. 2, 3\u201315 (2016).\n\t3.\t Maugeri, L. The Age of Oil: The Mythology, History and Future of the World\u2019s \nMost Controversial Resource (Praeger, 2006).\n\t4.\t Noguera, J. Oil prices: breaks and trends. Energy Econ. 37, 60\u201367 (2013).\n\t5.\t Noguera-Santaella, J. Geopolitics and the oil price. Econ. Model. 52,  \n301\u2013309 (2016).\n\t6.\t Noguera, J. The Seven Sisters versus OPEC: solving the mystery of the \npetroleum market structure. Energy Econ. 37, 60\u201367 (2017).\n\t7.\t Yergin, D. The Prize: The Epic Quest for Oil, Money, & Power (Simon & \nSchuster, 1991).\n\t8.\t Hamilton, J. D. Oil and the macroeconomy since World War II. J. Polit. Econ. \n91, 228\u2013248 (1983).\n\t9.\t Hamilton, J. D. Historical causes of postwar oil shocks and recessions. Energy \nJ. 6, 97\u2013116 (1985).\n\t10.\tHamilton, J. D. in Routledge Handbook of Major Events in Economic History \n(eds Parker, R. E. & Whaples, R.) 239\u2013265 (Routledge Taylor and Francis \nGroup, 2013).\n\t11.\tAkins, J. E. The oil crisis: this time the wolf is here. Foreign Aff. 51,  \n463\u2013490 (1973).\n\t12.\tMurphy, R. P. Removing the 1970s crude oil price controls: lessons for \nfree-market reform. J. Priv. Enterp. 33, 63\u201378 (2018).\n\t13.\tErickson, E. W. et\u00a0al. The political economy of crude oil price controls.  \nNat. Resour. J. 18, 787\u2013801 (1978).\n\t14.\tTotal Energy: Iranian Crude Oil Production (US Energy Information \nAdministration); https://www.eia.gov/totalenergy/data/browser/index.php?tbl\n=T11.01A#/?f=M&start=197812&end=201901&charted=13\n\t15.\tKilian, L. Not all oil price shocks are alike: disentangling demand and supply \nshocks in the crude oil market. Am. Econ. Rev. 99, 1053\u20131069 (2009).\n\t16.\tAlquist, R. & Kilian, L. What do we learn from the price of crude oil futures? \nJ. Appl. Econ. 25, 539\u2013573 (2010).\n\t17.\tPindyck, R. The dynamics of commodity spot and futures markets: a primer. \nEnergy J. 22, 1\u201329 (2001).\n\t18.\tTotal Energy: Petroleum Products Supplied by Type (US Energy Information \nAdministration); https://www.eia.gov/totalenergy/data/browser/index.php?tbl\n=T03.05#/?f=M&start=197301&end=201806&charted=9-13\n\t19.\tLowinger, T. & Ram, R. Product value as a determinant of OPEC\u2019s official \ncrude oil prices: additional evidence. Rev. Econ. Stat. 66, 691\u2013695 (1984).\n\t20.\tVerleger, P. K. The determinants of official OPEC crude prices. Rev. Econ. \nStat. 64, 177\u2013182 (1982).\n\t21.\tOPEC Cairo Cuts May Solve Quota Quandry (Petroleum Intelligence Weekly, \n2004).\n\t22.\tParnes, D. Heterogeneous noncompliance with OPEC\u2019s oil production cuts. \nEnergy Econ. 78, 289\u2013300 (2019).\n\t23.\tKaufmann, R. K., Bradford, A., Belanger, L. H., Mclaughlin, J. P. & Miki, Y. \nDeterminants of OPEC production: implications for OPEC behavior. Energy \nEcon. 30, 333\u2013351 (2008).\n\t24.\tFundamentals don\u2019t justify $70 oil price \u2013 Exxon CEO. Reuters; https://\nuk.reuters.com/article/exxon-oil-price/update-2-fundamentals-dont-justify-\n70-oil-price-exxon-ceo-idUKN0720445120070907\n\t25.\tDe Long, J. B., Shleifer, A., Summers, L. H. & Waldmann, R. J. Noise trader \nrisk in financial markets. J. Polit. Econ. 98, 703\u2013738 (1990).\n\t26.\tTokic, D. Rational destabilizing speculation, positive feedback trading, and \nthe oil bubble of 2008. Energy Policy 39, 2051\u20132061 (2011).\n\t27.\tKolodziej, M., Kaufmann, R. K., Kulatilaka, N., Bicchetti, D. & Maystre, N. \nCrude oil: commodity or financial asset? Energy Econ. 46, 216\u2013223 (2014).\n\t28.\tUNITED STATES (Petroleum Intelligence Weekly, 2009).\n\t29.\tOPEC Will Have Hard Time Eeviving Prices, CGES Warns (Petroleum \nIntelligence Weekly, 2009).\n\t30.\tPundits See Oil Price Around $50 This Year (Petroleum Intelligence Weekly, \n2009).\n\t31.\tFactbox: OPEC ministerial comments at Quito meeting. Reuters https://www.\nreuters.com/article/us-opec-comments/factbox-opec-ministerial-comments-\nat-quito-meeting-idUSTRE6BA1UA20101211 (2010).\n\t32.\tInternational Energy Agency members release strategic petroleum stocks. \nToday in Energy (US Energy Information Administration, accessed 30 \nNovember 2019); https://www.eia.gov/todayinenergy/detail.php?id=1950\n\t33.\tManescu, C. B. & Nu\u00f1o, G. Quantitative effects of the shale oil revoluton. \nEnergy Policy 86, 855\u2013866 (2015).\n\t34.\tAnsari, D. OPEC, Saudi Arabia, and the shale revolution: insights from \nequilibrium modelling and oil politics. Energy Policy 111, 166\u2013178 (2017).\n\t35.\tBehar, A. & Ritz, R. A. OPEC vs US shale: analyzing the shift to a market-\nshare strategy. Energy Econ. 63, 185\u2013198 (2017).\n\t36.\tAnsari, E. & Kaufmann, R. K. The effect of oil and gas price and price \nvolatility on rig activity in tight formations and OPEC strategy. Nat. Energy 4, \n321\u2013328 (2019).\n\t37.\tSheikhs v shale - The new economics of oil. The Economist https://www.\neconomist.com/leaders/2014/12/04/sheikhs-v-shale (2014).\n\t38.\tKleinberg, R. L., Paltsev, S., Ebinger, C. K. E., Hobbs, D. A. & Boersma, T. \nTight oil market dynamics: Benchmarks, breakeven points, and inelasticities. \nEnergy Econ. 70, 70\u201383 (2018).\n\t39.\tShrestha, K. Price discovery in energy markets. Energy Econ. 45,  \n229\u2013233 (2014).\n\t40.\tSilv\u00e9rio, R. & Szklo, A. The effect of the financial sector on the evolution  \nof oil prices: analysis of the contribution of the futures market to  \nthe price discovery process in the WTI spot market. Energy Econ. 34, \n1799\u20131808 (2012).\n\t41.\tKaufmann, R. K. & Ullman, B. Oil prices, speculation, and fundamentals: \ninterpreting causal relations among spot and futures prices. Energy Econ. 31, \n550\u2013558 (2009).\n\t42.\tEderington, L. & Lee, J. H. Who trades futures and how: evidence from the \nheating oil futures market. J. Bus. 75, 353\u2013373 (2002).\n\t43.\tWTI - Economic Data Series (FRED, St. Louis Fed, accessed January 2019); \nhttps://fred.stlouisfed.org/tags/series?t=wti;http://www.imf.org/external/np/\nres/commod/index.aspx\n\t44.\tKaufmann, R. K. Price differences among crude oils: the private costs of \nsupply disruptions. Energy Econ. 56, 1\u20138 (2016).\n\t45.\tScheitrum, D. P. & Carter, C. A. WTI and Brent future pricing structure. \nEnergy Econ. 72, 462\u2013469 (2018).\n\t46.\tBureau of Labor Statistics Data (US Bureau of Labour Statistics, accessed \nJanuary 2019); https://data.bls.gov/cgi-bin/surveymost\n\t47.\tKaufmann, R. K. A model of the world oil market for Project LINK \nintegrating economics, geology and politics. Econ. Model. 12, 165\u2013178 (1995).\n\t48.\tKaufmann, R. K., Dees, S., Karadeloglou, P., S\u00e1nchez, M. & Sanchez, M. Does \nOPEC matter? An econometric analysis of oil prices. Energy J. 25,  \n67\u201390 (2004).\n\t49.\tKaufmann, R. K., Dees, S., Gasteuil, A. & Mann, M. The role of refinery \nutilization futures markets, and non-linearities. Energy Econ. 30, 2609\u20132622 \n(2008).\n\t50.\tPrindle, D. F. Petroleum Politics and the Texas Railroad Commission (Univ. of \nTexas Press, 1981).\n\t51.\t3c. OPEC Crude Oil (excluding condensates) Supply. STEO Data Browser (US \nEnergy Information Administration); https://www.eia.gov/outlooks/steo/data/\nbrowser/#/?v=7&f=M&s=0&start=201401&end=201912&linechart=CO\nPR_AG~COPC_OPEC&maptype=0&ctype=linechart&map\n\t52.\tPetroleum Trade: Exports by Country Destination (US Energy Information \nAdministration); https://www.eia.gov/totalenergy/data/monthly/pdf/sec3_13.\npdf\n\t53.\tPetroleum Facts and Figures (American Petroleum Institute, 1971).\n\t54.\tAmerican Petroleum Institute in Basic Petroleum Data Book Section VIII \n(American Petroleum Institute, 1981).\n\t55.\tU.S. Refinery Utilization and Capacity (US Energy Information \nAdministration); https://www.eia.gov/dnav/pet/pet_pnp_unc_dcu_nus_a.htm\n\t56.\tEliott, G., Rothenberg, T. J. & Stock, J. H. Efficient tests for an autoregressive \nunit root. Econometrica 64, 813\u2013836 (1996).\n\t57.\tStock, J. H. & Watson, M. W. A simple estimator of cointegrating vectors in \nhigher order integrated systems. Econometrica 61, 783\u2013820 (1993).\n\t58.\tSchwarz, G. Estimating the dimension of a model. Annu. Stat. 6,  \n461\u2013464 (1978).\n\t59.\tNewey, W. K. & West, K. D. A simple positive semi-definite heteroskedasticity \nand autocorrelation consistent covariance matrix. Econometrica 55,  \n703\u2013708 (1987).\n\t60.\tAkaike, H. in 2nd International Symposium on Information Theory (eds \nPetrov, P. N. & Csaki, F.) 267\u2013281 (Akad\u00e9miai Kiado, 1973).\n\t61.\tGeneral-to-Specific (GETS) Modelling and Indicator Saturation Methods \nv0.21 (CRAN, 2019); https://cran.r-project.org/web/packages/gets/index.html\n\t62.\tCastle, J., Doornik, J., Hendry, D. & Pretis, F. Detecting location shifts during \nmodel selection by step-indicator saturation. Econometrics 3, 240\u2013264 (2015).\nNature Energy | VOL 5 | February 2020 | 141\u2013149 | www.nature.com/natureenergy\n148\n\nArticles\nNATUrE EnErgy\nAcknowledgements\nWe thank F. Pretis for his help with the R package gets. We also thank C. A. S. Hall, J. \nNoguera, the members of Project LINK and members of the Economics Department at \nthe University of Victoria for comments on preliminary versions of this work. Any errors \nthat remain are our responsibility.\nAuthor contributions\nThis project was conceived by R.K. and modified by C.C. C.C. took the lead in compiling \nthe data used in the analysis reported in the main text; R.K. compiled the data used in \nthe analysis reported in the Supplementary Information. C.C. took the lead in estimating \nthe statistical models in R; R.K. adapted this code for the analyses in the Supplementary \nInformation. R.K. took the lead in writing the manuscript and designing the figures, with \nsignificant input from C.C.\nCompeting interests\nThe authors have no financial or non-financial interests associated with the material in \nthis manuscript.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-020-0549-1.\nCorrespondence and requests for materials should be addressed to R.K.K.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2020\nNature Energy | VOL 5 | February 2020 | 141\u2013149 | www.nature.com/natureenergy\n149\n\n\n    Scientific Research Findings:", "answer": " Proxies for supply and demand accurately account for oil prices in 138 of 313 quarters from January 1938 to March 2018. In the other 184 quarters, reduced accuracy creates nine regimes when oil prices deviate from the level implied by supply and demand. Of these nine regimes, two are associated with policy interventions to suppress price increases. The other seven are associated with non-market forces, such as strategic changes by OPEC, panic buying or speculative bubbles. This historical analysis of the oil market indicates that factors other than supply and demand have a large effect on prices for extended periods (with the shortest regime being a year and the longest being more than two decades). By summing the additional price paid for oil consumed in the US, we show that in recent regimes, speculative bubbles transferred US$42.9 billion from US consumers to US oil producers and transferred US$87.4 billion from the US economy to oil exporting nations. This suggests significant economic losses.", "id": 25}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Analysis\nhttps://doi.org/10.1038/s41560-019-0546-4\n1Tyndall Centre for Climate Change Research, Department for Mechanical, Aerospace and Civil Engineering, Faculty of Science and Engineering,  \nUniversity of Manchester, Manchester, UK. 2Department of Economics, School of Social Sciences, University of Manchester, Manchester, UK.  \n3Hunter Centre for Entrepreneurship, University of Strathclyde Business School, Glasgow, UK. 4Grantham Institute, Imperial College London, London, UK. \n*e-mail: timothy.braunholtz-speight@manchester.ac.uk; maria.sharmina@manchester.ac.uk; edward.manderson@manchester.ac.uk\nL\nocal energy projects delivered by community groups could \npotentially play a pivotal role in realizing the transition to a \nlow-carbon energy future. Community energy schemes offer \nan alternative to large-scale energy provision, with various forms \nof community energy already found across Europe, North America \nand elsewhere1\u20135. In the UK, the term \u2018community energy' is gener-\nally associated with small civil society organizations and/or social \nenterprises running projects that encourage energy saving and \nefficiency, or that generate renewable electricity. These projects are \ntypically grounded in the motivation to accelerate decarbonization \nthrough both decentralization and democratization of the energy \nsystem, and address issues such as fuel poverty and energy justice6\u201311.\nThe growth of the sector in the UK has been driven by a combi-\nnation of the decreasing cost of renewable energy technologies and \ngovernment policies1,4 (see Table 1). However, more recently, gov-\nernment support for small-scale renewables has been substantially \nscaled back1,12. Most notably, Feed-in Tariff scheme (FITs) rates \nfell more than 50% from 2015 to 2016 for many technologies (see \nSupplementary Table 2) and the scheme is now closed to new proj-\nects. In this challenging low-subsidy environment, project develop-\nment and investment have slowed significantly13.\nThese recent developments emphasize the importance of under-\nstanding how the community energy sector is financed. However, \nthere is currently very limited empirical evidence on the financing \nof community energy activities6,14. Studies in Germany and Belgium \nnote the mixture of motivations reported by citizens investing in \ncommunity energy15,16; further studies in Germany note the sub-\nstantial size of the renewable energy cooperative sector there17 and \nits success in raising finance from cooperative members18. Much of \nthe literature on the UK community energy sector focusses on the \nqualitative exploration of the definitions, motivations and challenges \nfor projects, and how they engage with questions of justice and pov-\nerty6\u20138,10,11,19\u201327. Nevertheless, some studies have examined the sec-\ntor\u2019s finance and business models9,13,28\u201333. In particular, sector-wide  \nsurveys have played an important role in establishing the size and \nstructure of the sector, gathering some data on finance; however, \nthese surveys do not offer project-level analysis on financial perfor-\nmance9,13,33. A government-convened working group on community \nenergy finance offered insight into the difficulties faced by com-\nmunity projects, but did not present an analysis of empirical data29. \nWe are aware of only one previous quantitative study of community \nenergy business models at project level, which compared the costs of \ncommunity- and commercially-owned wind and hydro energy proj-\nects in Scotland30,31. The study provides valuable detailed evidence \non the distinctiveness of community energy, finding that commu-\nnity projects face additional risks and transaction costs compared \nwith commercial projects. However, its scope does not include other \naspects of business models (such as finance or revenue), other busi-\nness model types and projects in other parts of the UK.\nIn this analysis, we fill an important gap in the community \nenergy literature by providing a UK-wide analysis of the financing \nmechanisms and financial performance of individual community \nenergy projects. We also systematically characterize community \nenergy business models using quantitative methods. We perform \nour analysis using a dataset on the financial characteristics of com-\nmunity energy projects, collected in a survey of the UK community \nenergy sector undertaken by the authors in 2017\u22122018.\nThe Financing Community Energy survey\nWe use data from the new Financing Community Energy survey \nof the UK community energy sector conducted in 2017\u22122018. The \nsurvey structure is based on the Business Model Canvas34, which \nanalyses organizations\u2019 value propositions and associated activities, \ncustomers, resources, and costs and revenues (see Supplementary \nMethods for the survey questions). For each project\u2019s energy gen-\neration and financial flows (costs and revenues), we collected data \nfor only the most recent 12-month period for which they were \navailable, to minimize the administrative burden on participants \nBusiness models and financial characteristics of \ncommunity energy in the UK\nTim Braunholtz-Speight\u200a \u200a1*, Maria Sharmina\u200a \u200a1*, Edward Manderson\u200a \u200a2*, Carly McLachlan1, \nMatthew Hannon3, Jeff Hardy\u200a \u200a4 and Sarah Mander1\nCommunity energy projects take a decentralized and participatory approach to low-carbon energy. Here we present a quantita-\ntive analysis of business models, financing mechanisms and financial performance of UK community energy projects, based on \na new survey. We find that business models depend on technology, project size and the fine-tuning of operations to local con-\ntexts. Although larger projects rely more on loans, community shares are the most common and cheapest financial instrument \nin the sector. Community energy has pioneered low-cost citizen finance for renewables, but its future is threatened by reduc-\ntions, and instability, in policy support. Over 90% of the projects in our sample make a financial surplus during our single-year \nsnapshot, but this falls to just 20% if we remove income from price guarantee mechanisms, such as the Feed-in Tariff scheme. \nRenewed support and/or business model innovations are therefore needed for the sector to realize its potential contribution \nto the low-carbon energy transition.\nNature Energy | VOL 5 | FebruarY 2020 | 169\u2013177 | www.nature.com/natureenergy\n169\n\nAnalysis\nNATURe EneRgy\nand maximize the number of projects included in the dataset. Our \nanalysis of the community energy sector is therefore cross-sectional \nrather than longitudinal (see Methods for further discussion).\nWe received substantive responses to our survey on 145 projects \nfrom 48 organizations. To obtain more data on certain technologies, \nwe supplemented the survey data with information on a further \neight projects from organizations\u2019 published financial statements \nand reports. This extra data give a total of 153 projects and 56 orga-\nnizations in our dataset. However, as published documents provide \nless extensive data than our survey, the data on the additional eight \nprojects are used only when we provide summary statistics on proj-\nect characteristics by technology.\nOf the 153 projects, the majority (139) are electricity generation \nprojects, with a total capacity of about 41\u2009MW. Two of these projects \nalso involved heat generation, and there are a further five that are \nexclusively heat generation projects. Nine projects involve no direct \nenergy generation. Other surveys of the sector found 228 commu-\nnity energy organizations in England, Wales and Northern Ireland \nin 2018, of which 204 were engaged in electricity generation, with \na total of 168\u2009MW operational capacity13. In Scotland, there was a \ntotal of 81\u2009MW of community-owned renewable electricity gen-\nerating capacity in 201728. Our survey dataset therefore captures \napproximately one-sixth of the UK community energy sector in \nterms of installed generation capacity.\nOur analysis first provides a taxonomy of the project business \nmodels employed in the community energy sector. We then anal-\nyse the financial characteristics of community energy projects, \nthe mechanisms by which they have been financed, the price they \ncharge their customers and the importance of FITs and other incen-\ntive schemes to project revenues.\nA taxonomy of community energy projects\nTo shed further light on the structure of the sector we applied cluster \nanalysis to produce the first ever quantitative data-driven taxonomy \nof community energy business models in the UK. A cluster analysis \nof the survey results helps to identify similarities and differences \nbetween community energy projects, to complement case study \nresearch. The clusters also present a \u2018menu\u2019 of business models that \ncan be used to inform the design of new projects.\nOur taxonomy has two parts, based on two runs of the cluster \nanalysis. The first run used project-only variables, omitting the \nproject location and all the variables that related to the organization \nrunning the project. This analysis produced three broad clusters \n(Table 2) shaped largely by the type of energy activity undertaken \n(generation versus demand-side activities) and by the type of tech-\nnology employed.\nThe second cluster analysis run used all the variables, includ-\ning those relating to the organization as a whole (such as turnover, \nnumber of members and legal structure) and the project location. \nHere, the three broad clusters splintered into many smaller ones \nproducing 12 clusters in total (Table 3).\nThe two runs of the cluster analysis map closely on to one another \n(Table 3). Taken together, they suggest that while technology and \nactivity are important drivers of business models, within the three \nbroad clusters, community energy organizations have fine-tuned \ntheir business models. This fine-tuning includes different means of \naccessing finance and other resources (such as varying reliance on \ncommunity shares, loans and grants), and offering a range of value \npropositions to different customers (such as funding other local \nprojects, providing educational opportunities, cutting CO2 emis-\nsions and reducing energy bills).\nOf the 12 clusters, the \u2018demand-side services\u2019 and \u2018energy as a \nsideline\u2019 projects stand out as significantly different from the others, \nand they also form the third cluster in the project-level cluster anal-\nysis. The other ten clusters are differentiated partly by technology, \nwith a clear divide between solar rooftop and other generation tech-\nnologies, also reflected in the project-level cluster analysis. Other \nvariables, such as whether the organization runs multiple projects, \nhas paid staff or is entirely volunteer-run, the type of customers it \ndeals with and how it finances its projects, are part of the fine-tun-\ning we noted above.\nWe find that the most common aspects across the current busi-\nness models in the community energy sector include a predomi-\nnance of electricity generation particularly through solar PV, not \nhaving charitable status and not being linked to a charitable \u2018parent\u2019 \nbody, employing three full-time staff at most on average (although \nthere are rare cases of employing up to ten staff), relying on at least \nsome volunteers (and up to 90 in some projects), relying on FITs for \nrevenue and community shares for finance, mainly working with \none type of customer only and typically emphasizing environmental \nvalue propositions over social and economic value propositions.\nCosts, revenues and performance of energy generation \nprojects\nOur sample includes 84 solar rooftop, 15 wind, 12 hydro, 4 solar \nground-mount and 2 biomass projects with sufficient data to calcu-\nlate financial performance. Table 4 presents summary statistics on \nthe average project characteristics by technology. (The table does \nnot include data on the two biomass heat projects due to the risk of \ncompromising data confidentiality.)\nWe find that there is substantial heterogeneity in the size, costs \nand revenues of community energy projects across the different \ntechnologies. Wind and solar ground-mount projects tend to be \nsubstantially larger than others in terms of generation capacity and \nperformance, costs and revenues. The mean solar rooftop project is \nsmaller, at 74\u2009kW capacity, but this size remains much larger than \ntypical UK domestic solar rooftop capacity35 (\u2009<4\u2009kW).\nTable 4 also presents two measures of the financial performance \nof the projects. Annual costs per unit generated are highest for wind \nprojects and lowest for solar rooftop and biomass. In contrast, the \nreturn on capital expenditure (CAPEX) is higher for the average \nwind project than the average solar or hydro project. With the caveat \nthat the sample is very small, biomass heat compares very favourably \nin terms of return on capital with other technologies (21\u2009pence per \u00a3 \nCAPEX). The differences in performance observed across technolo-\ngies may reflect various factors, including project-specific charac-\nteristics such as age and size, both of which may have a significant \nimpact on original capital expenditure figures, organization-specific \ncharacteristics, such as the expertise of personnel and learning by \ndoing, as well as the features of the technologies themselves.\nWe find that the average annual financing costs across all proj-\nects is \u00a346,500 per annum (excluding projects with zero financing \ncosts). The average total CAPEX across these projects is \u00a3865,900. \nTherefore, community energy projects on average face annual \nfinancing costs equal to about 5% of their initial total CAPEX.\nTable 1 | Principal forms of government-mandated price  \nsupport for small-scale low-carbon energy provision referred  \nto in this paper\nScheme\nDates open to \nnew projects\nEnergy scope\nRenewables obligation (RO)\n2002\u20132017\nElectricity generation\nFeed-in Tariff scheme\n2010\u20132019\nElectricity generation\nRenewable heat incentive (RHI)\n2014\u20132021\nHeat generation\nNotes: These schemes offer eligible projects a guaranteed price for the electricity or heat they \ngenerate. In some cases the schemes operate differently across the UK\u2019s devolved nations, \nand each scheme involves a complex set of scale and technology bandings, and eligibility and \nadministrative requirements. Further information on these schemes can be found on the Ofgem \nwebsite67 and in previous work by the authors1.\nNature Energy | VOL 5 | FebruarY 2020 | 169\u2013177 | www.nature.com/natureenergy\n170\n\nAnalysis\nNATURe EneRgy\nCommunity energy project financing\nData on financing mechanisms are available for 136 projects (89% \nof the total). Around three-quarters of projects (77%) use just one \nor two external financing instruments to fund their projects. Over \none-third (37%) of projects also use the organization\u2019s pre-existing \nfunds to undertake a project. Community shares36 are the most  \nTable 2 | Taxonomy of UK community energy from cluster analysis of survey data using only project-level variables\nCluster name\nNo. of projects\nNo. of organizations\nKey characteristics of clusters\nStandalone renewables\n20\n16\nWind, hydro and ground-mounted solar electricity generation projects, \nand one biomass heat network\nOn-site customer renewables\n85\n22\nAlmost entirely rooftop solar photovoltaic (PV) projects, but also one \nsolar thermal, one hydro, one wind and one biomass boiler. Electricity \nsold to customer in building (solar rooftop) or by private wire\nDemand-side activities\n14\n10\nA mixture of energy efficiency and fuel poverty advice projects, and \nrenewable energy generated for own use (rather than selling)\nTable 3 | Taxonomy of UK community energy from cluster analysis of survey data using project, organization and location variables\nProject-level cluster \n(Table 2)\nCluster name\nNo. of \nProjects\nEnergy activities\nOrganization\nFinance and resources\nRevenues and \ncustomer sectors\nStandalone renewables\nMulti-financed \nhydro and wind\n4\nHydro and wind \nelectricity generation \non small/medium\u2013\nlarge scale\nCo-ops and other \ncompanies, single \nprojects only\nMix of financing \ninstruments; often \npay for sites and other \nresources\nFITs only\nStandalone renewables \nand on-site customer \nrenewables\nLarge wind selling \nto grid\n5\nWind with mean \ncapacity of >2\u2009MW\nCo-ops and other \ncompanies with \npart-time staff\nAll use loans, some \ncommunity shares\nRO, energy sales to \nwholesalers or local \ngrid\nOn-site customer \nrenewables\nMedium-scale \ngeneration with \nmixed financing\n9\nWind, hydro, solar \nground-mount and \nbiomass heat \u2013 mean \ncapacity >1\u2009MW\nCo-ops and other \ncompanies with \nsome paid staff\nAll use loans, some \ncommunity shares\nFITs and sales to \nenergy companies\nSmall/medium \nsolar rooftop\n9\nSolar rooftop  \nPV \u2013 mix of scales\nVolunteer-run \nco-ops\nCommunity shares, \nmost resources free\nFITs and energy sales \nto mix of sectors\nMulti-site solar on \npublic sector roofs\n35\nSolar rooftop PV \nmostly <50\u2009kW \ncapacity \u2013 sometimes \nwith energy efficiency \nalso\nAll co-ops with some \npaid employment, \nrunning multiple \nprojects\nCommunity shares; \nsites free, some \nresources free or \nin-house (for example, \nlegal services)\nFITs and energy \nsales, mostly to \npublic sector\nProfessionalized \nsolar rooftop \nco-ops\n13\nSolar rooftop  \nPV \u2013 mix of scales\nCo-ops with multiple \nprojects and paid \nstaff\nCommunity shares \nmain financing \ninstrument; many \nresources paid for\nFITs and energy sales \nto mix of sectors\nSmall multi-project \ngeneration for third \nsector groups\n12\nMostly solar rooftop \nPV (but some heat)\nVolunteer co-ops \nrunning multiple \nprojects\nMix of financing \ninstruments\nFITs and energy sales \nto third sector\nSmall solar rooftop\n9\nSolar rooftop PV \nmostly <50\u2009kW \ncapacity\nMostly volunteer-\nrun co-ops\nFinance mostly \ncommunity shares; \nmix of free and paid \nresources\nFITs and energy sales \nto mix of sectors\nSmaller-scale \nmulti-project \nco-ops\n4\nMultiple solar rooftop \nPV <50\u2009kW capacity\nVolunteer-run \nco-ops\nAll using community \nshares, some loans \nand grants also\nMix of public, private \nand third sector \ncustomers\nDemand-side activities\nDemand-side \nservices\n6\nEnergy efficiency \nadvice, installation \nand fuel poverty \nreduction work\nMostly co-ops, paid \nstaff and volunteers\nGrants and loans\nCustomer fees, \nenergy services \ncontracts, some work \nfree of charge\nEnergy as a sideline\n7\nRooftop solar PV, heat \nand electricity storage\nSmall-scale third \nsector sports and \nleisure clubs\nGrants and self-\nfinancing\nFITs and savings on \nown energy bills\nAll three project-level \nclusters\nMulti-tech \ngeneration including \npartnerships\n6\nHydro, wind and solar \nrooftop PV\nCompanies with paid \nstaff\nGrant and loan finance; \nsites free, other \nresources paid or \nin-house\nHydro and wind sold \nto grid, solar to local \ncustomer\nNote: For further explanation of the terms used, see Methods section on data gathering.\nNature Energy | VOL 5 | FebruarY 2020 | 169\u2013177 | www.nature.com/natureenergy\n171\n\nAnalysis\nNATURe EneRgy\nfrequently used instrument, with 102 issues of community shares in \nour dataset. In addition, there are 73 loans, 54 grants and 9 \u2018other\u2019 \ninstruments, mostly bonds.\nFocusing on operational energy generation projects (121 of the \n136 with financing data), the size of the CAPEX is related to how \nthe finance is raised: larger projects rely more heavily on loans, and \nsmaller projects rely more on community shares. There seems to \nbe a threshold around a CAPEX of \u00a3200,000: 88% of generation \nprojects above this threshold use some loan finance, but only 17% \nof projects below this threshold reported using loans. However, \ncommunity shares still account for a significant proportion of the \ntotal capital raised for all but the largest scales of project, as Fig. 1 \nshows. Community shares account for almost all the finance raised \nby projects with a CAPEX of less than \u00a3200,000 (the majority of \nprojects), but a much smaller proportion of the total finance raised \nby projects costing over \u00a31.5\u2009million. Grants, such as the Rural and \nUrban Community Energy Funds37,38, form a relatively small part of \ntotal capital for all project scales. However, they may play a signifi-\ncant role in de-risking projects39, as they are often used to finance \nthe earliest \u2013 and riskiest \u2013 stages of project development. They are \nalso important sources of funding for projects in the demand-side \nservices cluster.\nWe use regression analysis to consider whether there is a sta-\ntistical relationship between the cost of finance and the instru-\nment type (see Methods for details). From this analysis, we find \nthat community shares charge an interest rate that is 2 percentage \npoints lower than loans on average. (Community shares typically \npay interest rather than dividends, see Supplementary Note 1.) To \nput this finding into perspective, the average size of the financing \ninstruments in the regression sample is about \u00a3306,000; therefore, \nthe first year\u2019s interest payments on this would be about \u00a36,200 \nlower if financed by community shares rather than loans (see \nMethods section for details).\nThese findings are striking because, unlike conventional equity, \ncommunity shares are neither saleable to third parties for profit, \nnor do they necessarily give the holder a claim to the proceeds of \na sale of the issuing company\u2019s assets36. Therefore, the prospect \nof capital gains, which might encourage conventional sharehold-\ners to accept lower interest payments, is not available to commu-\nnity shareholders. Further, there do not appear to be many cases \nof community shares refinancing risky early-stage loans: most \nprojects that issued community shares did not use loans at all. We \nexplore alternative explanations for the interest rate difference in \nthe Discussion section.\nWe now turn to the marketing mechanisms employed to attract \nfunds. Despite the growth of online alternative finance platforms, \nsuch as Ethex or Crowdfunder, that can reach potential investors \nacross the UK, around half the community share issues in our data-\nset were made using local marketing only, for example, through \nlocal newspapers and the organization\u2019s own website. Many others \nwere marketed nationally, but through community energy networks \nrather than general alternative finance online platforms. There is a \nclear gap between the scale of funds raised by these different mecha-\nnisms, with general large-scale marketing raising the largest sums \n(see Table 5). However, our data show that local marketing has the \nlowest mean interest rates, with rates on average 0.8% lower than \nenergy-specific UK-wide marketing (a significant difference at the \n1% significance level). It is also notable that locally marketed com-\nmunity shares raised enough to cover project CAPEX for 32 of the \n43 projects in the table (74% of these projects). This suggests that \nmany community energy projects have succeeded in raising the \ncapital they need through relatively cheap local finance.\nImportance of price guarantee schemes to project finances\nAs the FIT and RO schemes are now closed to new projects, we \nexamine the importance of revenue from these schemes to commu-\nnity energy business models. The overwhelming majority of genera-\ntion projects in our dataset accessed FITs, RHI or RO revenues (only \ntwo projects did not). Of these, we used 110 projects with sufficient \ndetail on annual costs and revenues to perform a simple calculation \nto examine their dependency on these schemes (note that existing \nprojects are not affected by cuts to FITs rates and the closure of the \nFIT and RO schemes, see Methods for details). We find that 92% of \nthese projects (101 projects) were in financial surplus (that is, total \nannual revenues exceeded total annual costs) for the year for which \ndata were provided; however, after removing the price scheme rev-\nenues, only a fifth of the projects (22 projects) were in surplus. As \nthese projects were designed to draw on FITs or similar revenue \nstreams, it is not surprising that removing those revenues would \npush many projects into deficit. Yet, it is notable that 22 projects do \nnot suffer this fate in our exercise, and so in the rest of this section \nwe examine their characteristics in more detail.\nOf the 22 projects, 5 were commissioned in the 2\u2009years prior to \nthe survey date and were financed primarily by community shares \nbut reported no financing costs. Quite often, community shares are \nTable 4 | Energy generation project characteristics by \ntechnology\nHydro\nWind\nSolar \nground\nSolar \nroof\nNumber of projects in sample\n12\n15\n4\n84\nCapacity (kW)\n163\n1,862\n3,428\n74\n(162)\n(2,741)\n(2,304)\n(168)\nTotal capital expenditure  \n(\u00a3 thousands)\n1,097\n3,255\n5,992\n87\n(979)\n(4,187)\n(6,387)\n(170)\nAnnual operating costs  \n(\u00a3 thousands)\n25\n136\n172\n3\n(32)\n(182)\n(205)\n(13)\nAnnual financing costs  \n(\u00a3 thousands)\n29\n161\n318\n4\n(37)\n(219)\n(370)\n(9)\nAnnual generation (MW\u2009h)\n472\n4,317\n3,385\n56\n(458)\n(7,512)\n(2,289)\n(111)\nAnnual revenue (\u00a3 thousands)\n91\n552\n730\n11\n(93)\n(818)\n(900)\n(25)\nAnnual surplus (\u00a3 thousands)\n37\n256\n240\n4\n(39)\n(489)\n(328)\n(7)\nCapacity factor\n0.36\n0.27\n0.11\n0.09\n(0.15)\n(0.14)\n(0.01)\n(0.03)\nAnnual cost per kW\u2009h (\u00a3)\n0.12\n0.15\n0.12\n0.11\n(0.05)\n(0.18)\n(0.10)\n(0.09)\nReturn on capital costs (\u00a3)\n0.11\n0.18\n0.10\n0.12\n(0.05)\n(0.10)\n(0.03)\n(0.05)\nNotes: Table shows mean characteristics with standard deviations (the square root of the variance) \nin parentheses. Only data from fully operational projects are included. Some projects for which \nrevenue data are missing are excluded. Data for eight projects taken directly from organizations\u2019 \npublished financial statements and reports are only used in this table. The table does not include \ndata on the two biomass heat projects due to the risk of compromising data confidentiality. \nOperating costs refer to expenditure on running the project. Financing costs include all repayments \nof borrowing and payments to shareholders. Operating costs do not include financing costs. \nAnnual surplus\u2009=\u2009(annual revenue\u2009\u2013\u2009annual operating costs\u2009\u2013\u2009annual financing costs). Capacity \nfactor\u2009=\u2009[annual generation/(365\u2009\u00d7\u200924\u2009\u00d7\u2009capacity)]. Annual cost per kW\u2009h\u2009=\u2009(annual operating \ncosts\u2009+\u2009annual financing costs)/annual generation. See Methods section for a discussion of this \nmetric and a comparison with the levelized cost of energy (LCOE) metric. Return on capital \ncosts\u2009=\u2009(annual revenue/total capital costs). Costs, revenues and generation figures are based on a \nsingle year of project data. See Methods section for further discussion of the implications.\nNature Energy | VOL 5 | FebruarY 2020 | 169\u2013177 | www.nature.com/natureenergy\n172\n\nAnalysis\nNATURe EneRgy\nissued with the stipulation that they pay no interest for the first year \nor two of generation36, which may be the case with these projects. \nSubtracting community share interest payments, at the rate given \nfor each project, from these project revenues leaves three out of five \nof them making a loss without FITs revenue.\nFour projects, including one of the five just mentioned, had addi-\ntional revenue that was not directly linked to levels of energy gen-\neration (for example, environmental grant funding or land rental \nto a commercial partner); two other projects gained revenue in the \nform of savings on the organization\u2019s own energy costs. The remain-\ning 12 projects were all solar rooftop projects that sold electricity to \nthe owners or occupiers of the building where the solar panels were \nlocated. For 10 out of these 12 projects, the customer used at least \n80% of the electricity generated, at a price between wholesale and \ntypical retail prices (see the next section).\nThis analysis suggests only a small number of existing projects, \nthose selling most of the electricity they generate to on-site custom-\ners, could generate a financial surplus without price support. Further, \nthe 22 projects highlighted above are mostly small scale (less than \n50\u2009kW generation capacity), and 18 were left with an annual surplus \nof less than \u00a33,000 without price support revenues. In the context \nof such small surpluses, year-on-year variation in weather condi-\ntions and operational costs can have a significant bearing on project \nfinancial performance. The long-term price guarantee offered by \nschemes such as the FITs plays a significant role in de-risking proj-\nects and attracting finance40,41.\nPrices paid by community renewable energy customers\nCommunity energy generation projects sell to a range of custom-\ners, including energy companies, other companies, community \nand third sector organizations, and public sector bodies. We find \nthat energy companies pay the lowest rates on average, equal to \njust 5.03\u2009pence per kW\u2009h in our sample (Table 6). This low rate is \nto be expected, as projects selling to energy companies are compet-\ning with wholesale rather than retail prices. Of community energy\u2019s \nretail customers, 6 out of 25 community or third sector customers \nin our dataset receive energy for free (\u2018zero rate\u2019 customers); the \nremaining 19 customers pay an average rate equal to 6.33\u2009pence per \nkW\u2009h. Private sector companies that are not energy companies pay \na slightly higher rate equal to 6.87\u2009pence per kW\u2009h. Public sector \norganizations pay 2.28\u2009pence per kW\u2009h or 45% more on average for \ntheir energy than energy companies (and 0.99\u2009pence per kW\u2009h more \nthan community or third sector customers). However, this rate may \nstill represent a significant saving on retail market electricity prices: \naverage non-domestic electricity prices were over 10\u2009pence per kW\u2009h \nfor most of the period (2015\u20132017) to which these data relate42.\n120\n100\n80\n60\n40\n20\n0\nPercentage of CAPEX (%)\n<200\n(90 projects)\n200\u20131,499\n(11 projects)\n>1,500\n(10 projects)\nProject CAPEX range (\u00a3 thousands)\nBonds\nLoans\nGrants\nCommunity\nshares\nFig. 1 | Percentage of capital raised by different instruments in relation \nto the scale of project capital expenditure. For each size category of \nproject CAPEX, the chart shows the proportion of total finance raised for \nall projects in that CAPEX range by different instruments (namely, loans, \ncommunity shares and grants). Where less than 100% of CAPEX is shown \nas being raised, this is due to some instruments that raised only relatively \nsmall sums being omitted from the figure. Where more than 100% of \nCAPEX is raised, these organizations retain surplus funds for reinvestment \nin future projects, in agreement with investors. The chart is based on 111 \nenergy generation projects with sufficient data on financing and CAPEX to \nperform the analysis.\nTable 5 | Community share issues analysed by marketing \nmechanism\nNo. of \nshare \nissues\n% of \ntotal\nTotal \namount \nraised (\u00a3)\nMean \namount \nraised \nper share \nissue (\u00a3)\nMean \ninterest \nrate on \nshares (%)\nGeneral online \nplatforms \u2013 UK wide\n9\n10\n6,362,856\n706,984\n4.33 (1.00)\nEnergy-specific \nmarketing \u2013 UK \nwide\n38\n42\n7,881,930\n207,419\n5.06 (0.33)\nLocal marketing\n43\n48\n4,248,498\n98,802\n4.26 (0.83)\nTotal\n90\n100\n18,493,284 205,481\n4.60 (0.79)\nNotes: We find that the difference in mean interest rate between energy-specific marketing and \nlocal marketing is statistically significant at the 1% level (t-statistic of 5.64). The difference in \nmean interest rate between energy-specific marketing and general online platforms is statistically \nsignificant at the 10% level (t-statistic of 2.14). The difference in mean interest rate between \nlocal marketing and general online platforms is insignificant. For the mean interest rate on shares, \nstandard deviations (the square root of the variance) are given in parentheses.\nTable 6 | Energy prices charged by community renewable \ngenerators by type of customer\nCustomer type\nAverage \ncustomer \nrate (pence \nper kW\u2009h)\nNo. of \ncustomers\nExcluding recipients of free \nenergy\nAverage \ncustomer rate \n(pence per \nkW\u2009h)\nNo. of \ncustomers\nEnergy company 5.03 (0.66)\n12\n5.03 (0.66)\n12\nOther private \nsector company\n6.24 (3.48)\n11\n6.87 (2.95)\n10\nCommunity or \nthird sector\n4.81 (3.39)\n25\n6.33 (2.28)\n19\nPublic sector\n7.32 (1.54)\n51\n7.32 (1.54)\n51\nNotes: Table shows mean customer rates by customer type, with standard deviations (the square \nroot of the variance) in parentheses.\nNature Energy | VOL 5 | FebruarY 2020 | 169\u2013177 | www.nature.com/natureenergy\n173\n\nAnalysis\nNATURe EneRgy\nDiscussion\nThis paper sheds light on how community energy organizations \nhave developed small-scale energy projects, often with significant \ncitizen funding, in an energy system dominated by large-scale \nactors and commercial finance that is the UK energy system32. We \nfind that, while organizations fine-tune the details of their business \nmodels to their context, the sector is dominated by renewable elec-\ntricity generation for which two basic project business models have \nbeen developed. First, larger projects have become increasingly pro-\nfessionalized and \u2018bankable\u2019, as shown by their ability to raise com-\nmercial loans alongside citizen finance. Second, many organizations \nrun rooftop solar PV projects that supply an on-site customer as \nwell as the grid, and are small enough to be mainly funded through \ncommunity share issues. Whilst international comparisons are lim-\nited by the scarcity of literature, we note that these UK community \nsolar projects appear to be similar to German renewable energy \ncooperatives in terms of financing structure and the cost of capital, \nalthough German cooperatives tend to be larger in terms of capi-\ntal raised17,18. We further find evidence that UK community energy \nprojects benefit from a local discount in fund-raising, with locally \nmarketed community share issues the cheapest category of finance \nin our dataset (other than grants).\nOur analysis shows that bringing social finance approaches into \nthe renewables sector has helped community energy to pioneer \ninnovations that can make a significant contribution to the energy \ntransition. In the field of social finance, innovations such as crowd-\nfunding and community shares have emerged as a response to the \ndifficulties that social enterprises have with accessing finance from \ntraditional lenders43,44. Meanwhile, in the energy sector, it is argued \nthat progress towards a low-carbon transition is hampered by domi-\nnant actors being \u2018locked-in\u2019 to the existing system by short-term \neconomic pressures45\u201347. However, expanding and diversifying the \nenergy investor base can increase the flow of finance into renew-\nables and other transition technologies48, because different actors \ninvest according to different criteria49,50.\nWe suggest that, through its emphasis on environmental and \nsocial value propositions, community energy has developed alter-\nnative investment criteria that have successfully lowered financing \ncosts for small-scale renewables through diversifying the investor \nbase. Previous research in the UK51, Belgium15 and Germany16,18 \nfinds that people invest in community energy projects for a mix of \nfinancial and non-financial reasons, and local investors may invest \nlarger sums15. Our analysis shows that, while it is clear from previ-\nous research that community projects can face additional costs and \nchallenges29\u201331, a community approach may also bring some finan-\ncial advantages.\nHowever, most of the business models in our data were built \nin an energy market where revenue was substantially de-risked by \nprice guarantee schemes40 such as FITs. While citizens\u2019 investment \nmotivations may be mixed, the financial security offered by such \nschemes was likely particularly important for people investing their \nown money41. What can our study say about future prospects for \ncommunity energy in contexts like the UK (and also Germany52,53), \nwhere FIT schemes are now closed?\nFirst, we note that renewable heat and self-financing demand-\nside projects are currently a rarity in the UK sector13. Yet the proj-\nects of this type in our dataset show a financial surplus. The growing \navailability of technologies (for example, light-emitting diode (LED) \nlights for energy efficiency) and the continued financial support for \nrenewable heat (RHI for heat generation), coupled with the reduc-\ntion in support for renewable electricity, may lead to growth in these \nactivities as community energy groups seek new business models54.\nSecond, even discounting those with special circumstances, we \nfind that 11% of renewable energy projects still showed an annual \nsurplus without FITs (or other price scheme) revenues. This finding \nsuggests that some new renewable electricity projects may be viable \nin a post-FITs world. Our analysis indicates the key elements of \npost-FITs renewable energy business models to be rooftop solar PV \nas the generation technology, a building with high energy demand \nas the site and a customer willing to pay. Our data show public sec-\ntor bodies pay, on average, the highest prices for community-gener-\nated electricity, but we also find projects on private sector rooftops \nthat show a surplus without FITs revenues.\nHowever, without some form of price stabilizer it is hard to see \nthe number of projects using community renewable energy business \nmodels returning to its previous rate of growth in the short term. \nOne source of price stability could be a floor price for exported \nelectricity, as suggested by community energy sector associations55. \nAnother mechanism might be Contracts for Difference auctions. \nThe UK already runs such auctions for large-scale renewables and \nhas opened them to remote island wind56. The auctions may benefit \nsome future community projects using the standalone renewables \nbusiness model, but could benefit many more if other technologies \nwere also able to participate.\nThird, policy could encourage, or even mandate, public sector \nbodies to purchase community-generated energy on long-term con-\ntracts. Given the growth of the community energy sector to date, \nthe low cost of community capital and the wider social benefits it  \noffers, these three measures would appear to be promising routes \nforward for both expanding renewable generation capacity and  \nsupporting the delivery of positive social impacts through the \nenergy transition.\nMethods\nSurvey design and data collection. This survey formed part of the Financing \nCommunity Energy research project led by Professor McLachlan, which was \nfunded as part of the UK Energy Research Centre (UKERC) research programme. \nIn the early stages of this research project, Community Energy England (CEE) \nand Community Energy Wales (CEW) launched their State of the Sector Survey \n2017 (SOTS 2017), which addressed some of the same topics. The Financing \nCommunity Energy project signed a Memorandum of Understanding with CEE \nto share survey data where possible, to maximize the benefit from the two data \ncollection exercises.\nThe survey questionnaire covered characteristics of community energy \norganizations, and of the projects they run. With regard to organizations, \nit included the legal structure, annual turnover, numbers of paid staff and \nvolunteers, and numbers of members. In relation to each project, topics included \nenergy activities (including electricity or heat generation and energy efficiency), \nownership (sole or partnership type), financing (for example, details of each \ninstrument type, value, terms), resources employed (including sites, technical, \nfinancial and legal services, general administration), costs (operating and \nfinancing), revenues (values and sources), value propositions (a range of economic, \nsocial and environmental propositions), customers (for example, types and rates \npaid) and other beneficiaries.\nThese categories were based on the Business Model Canvas approach to \nanalysing business models34, adjusted to take account of the project\u2019s particular \ninterest in financing mechanisms, and the characteristics of the community energy \nsector as the project team understood it.\nThe format of some of the questions was designed to complement the SOTS \n2017 survey to facilitate data sharing. Pre-set multiple choice formats were used \nas far as possible to facilitate data coding and quantitative analysis. Some free-text \nqualitative questions were also included, particularly in relation to organizations\u2019 \nfuture plans. The responses to these questions have been fed into other parts of the \noverall research project and are not reported in this paper.\nThe survey sample was constructed with reference to the SOTS 2017 \nrespondents list, data on community energy organizations in Scotland held by the \nsocial enterprise consultancy SCENE, and through internet searches, searching \nattendance lists at sector events, and Local Energy Scotland sending a survey link \nto their members through their newsletter.\nThe survey received research ethics approval from the University of \nManchester in October 2017. Informed consent was obtained in writing from all \nsurvey participants. The questionnaire was piloted in October/November 2017 \nwith three community energy organizations. Only minor changes were made after \nthe pilot process, and the pilot data form part of the survey dataset analysed in \nthis paper. The full survey was launched in November 2017 and closed in May \n2018. During January and February 2018 it was suspended in England, Wales and \nNorthern Ireland to avoid an overlap with the 2018 iteration of the SOTS.\nThe survey was available to complete online or by telephone interview with \nthe project team. Two methods of completing the survey were offered because the \nteam were conscious that community energy is a heavily surveyed sector. Allowing \nNature Energy | VOL 5 | FebruarY 2020 | 169\u2013177 | www.nature.com/natureenergy\n174\n\nAnalysis\nNATURe EneRgy\nresearch participants to choose the most convenient participation method ensured \nthe survey achieved sufficient responses for a meaningful quantitative data analysis, \nwhile also reducing the administrative burden on research participants. Ideally, we \nwould use only one method of data collection, because using different methods \nmay affect the quality of the data. Although we attempted to minimize this concern \nby ensuring that the online and telephone data followed a standardized framework, \nwe cannot rule out that inconsistent data collection methods have resulted in \nmeasurement error in our data. This is a limitation of the study.\nIn total the researchers contacted 280 organizations, of which 83 responded \nand 48 completed the survey, providing data on 145 projects. Not all projects are \nincluded in all the analysis presented here. Complete data were not available for \nsome projects, limiting the kinds of analysis that could be performed. Further, \nsome projects were classified as \u2018stalled or on hold\u2019, and so by their nature did \nnot have complete data. Data were collected on an additional eight projects using \npublished accounts and reports only. These data are presented in Table 4, to \nprovide greater coverage of the hydro, wind and solar ground-mount technologies, \nbut are not otherwise used in the analysis.\nCluster analysis. We used data on 119 projects for cluster analysis, as the projects \nthat had missing values related to any of the business model aspects had to be \nexcluded. Missing values were predominantly due to respondents not having access \nto some of the information required in the survey, for example, because of lost \npaperwork or limited documenting.\nUnlike in the performance and financial analysis (please see the next section), \nthe two bioenergy projects were not excluded from the cluster analysis, as their \nanonymity would be preserved when aggregated into clusters. However, we did \nre-run the cluster analysis without these two projects to test whether the results \nwould change. The exclusion of these two projects did not affect the composition of \nthe clusters in any of the runs (that is, either in the run with all variables, including \nthose related to the organizational level and location, or in the run with only \nproject-level variables). Excluding those two projects also had little effect on the \nsilhouette coefficients and the shapes of the clusters on the t-distributed stochastic \nneighbour embedding (t-SNE) plots. As the presence or otherwise of the bioenergy \nprojects did not affect our cluster analysis results, we kept those projects in the \nsample. The analysis runs with those projects excluded are available on request \nfrom the authors.\nThe cluster analysis was performed using R 3.6.157 and the packages dplyr58, \ncluster59, factoextra60, ggplot261, Rtsne62, dbscan63, fpc64 and clustMixType65. \nPartitioning around medoids (PAM), hierarchical agglomerative clustering (HAC), \ndensity-based clustering and k-prototypes clustering were the four clustering \nmethods applied to the dataset.\nWe included different combinations of organization-level variables (for \nexample, the legal structure of the organization) and project-level variables (for \nexample, the type of energy activity and type of customer) in several analysis \nruns. The first run used all 48 variables, the second run omitted the variables \nfor organization turnover and project location, and the third run omitted all the \norganization-level variables (such as turnover, number of members, number of \nvolunteers, number of staff employed, ownership structure, charitable status and \nyear of foundation) and project location, and used only the 40 variables relating to \nthe operation of individual projects.\nBefore running the analyses, we created a heatmap of the dissimilarity in the \ncommunity energy dataset using the daisy function with Gower distance that can \nhandle mixed types of variables. The heatmap (Supplementary Fig. 1) demonstrated \nthat the dataset did contain patterns compared with a heatmap of random data. We \nthen performed a sanity check on the dissimilarity matrix through outputting the \nmost and the least similar pairs of projects, with the expected results.\nWe used two key types of validation statistics to compare the results of the \nfour clustering methods: the within-cluster sum of squares (WSS) and the average \nsilhouette width (see Supplementary Table 1). The WSS was significantly better \nfor PAM (1.8901) than for the next best method using this metric, HAC (4.5846). \nThe average silhouette width for PAM (0.3058) was slightly lower than for HAC \n(0.3798), but significantly better than for the next best method using this metric, \ndensity-based clustering (0.1485). The validation results were similar for the \nanalysis run that included only project variables, that is, with organization and \nlocation variables excluded. In this run, density-based clustering showed somewhat \nbetter average silhouette width (0.2910) than PAM (0.2690). However, in this \nanalysis run, the density-based clustering method only yielded one cluster of 68 \nprojects, with the remaining 51 projects designated as outliers, which did not \nprovide meaningful insights. On the basis of these validation statistics (reported in \nSupplementary Table 1), the PAM clustering results were not a statistical artefact, \nand hence we selected PAM as our main clustering method.\nFor PAM clustering, we calculated and plotted silhouette width (Supplementary \nFig. 2) to select the optimal number of clusters. Twelve clusters corresponded \nto the highest silhouette width (0.4054) for the first analysis run in which all \nvariables were used, with thirteen and three clusters for the second and third runs, \nrespectively (with the highest silhouette widths of 0.4327 and 0.4026, respectively). \nThe results of the first two analysis runs were very similar to each other, therefore \nwe omitted the second analysis run with 13 clusters as it did not add any extra \ninsights to the results.\nWe then visualized the clustering using the t-SNE technique. The t-SNE \ntechnique decreases the number of dimensions while preserving the structure of \nthe dataset. The resulting figures for both the twelve- and three-cluster runs are \npresented in Supplementary Fig. 3. The figures illustrate which business models \nare well defined and distinct (for example, clusters 4, 11 and 12 in Supplementary \nFig. 3a) and which business models are more diffuse and might share similarities \nwith other types (for example, clusters 6, 8 and 9 in Supplementary Fig. 3a). Similar \nplots for other clustering methods (Supplementary Fig. 4a,b) give a less clear-cut \nallocation of projects into clusters.\nIn relation to the variables present within the clusters, it is important to \nexplain the value proposition variable. We constructed a list of value propositions \npotentially offered by community energy organizations to their customers based \non a review of the wider community energy literature. Survey participants were \nasked to say which of these value propositions they felt were important in their \ncustomers\u2019 decision to use their services (for example, to buy electricity). In \nthe cluster analysis, the value propositions were categorized as environmental, \neconomic or social, and projects were coded according to whether participants \nselected environmental, economic or social propositions (or a mixture) as \nimportant. Environmental value propositions included providing renewable \nelectricity, reducing CO2 emissions and tackling climate change. Economic value \npropositions included electricity generation regardless of origin, reducing energy \nbills, dealing with a known trusted organization, benefiting local economy, \nenabling customer to meet planning requirements and enhancing customer \nreputation. Finally, social value propositions included bringing a community \ntogether, generating community benefit and providing educational benefits. \nIn further research customer perspectives of the value propositions offered by \ncommunity energy organizations could be investigated.\nPerformance and financial analysis. This paper uses data collected for a single \nyear of project operation. Therefore, we provide a cross-sectional analysis that \ninvolves looking at the sector at a moment in time rather than assessing how it \nchanges over time. It is particularly important to bear this in mind for the project \nperformance characteristics presented in Table 4: because generation, revenue and \noperating costs may vary considerably from one year to the next, these data may \nnot be representative of project performance in other years. Future research may \nwish to address these issues by collecting survey data on project performance over \na number of years to construct a panel dataset.\nAlthough data were collected between November 2017 and May 2018, as noted \nabove, the data do not relate to the performance of projects during the months the \nsurvey was open. Rather, organizations reported data that relate to a 12-month \nperiod, more specifically, the most recent financial year for which data on the \nproject were available. As the data measure project performance over a 12-month \nperiod, they will reflect project performance over a sustained period of time rather \nthan during an individual month or season of the year. Furthermore, we do not \ntypically expect that community energy projects will vary that much in terms of \ntheir performance from one year to the next, especially in a systematic way (such \nthat variations over time would not average out across projects when performing \nstatistical tests, for example, when performing t-tests of means). Nonetheless, we \ncannot be sure that information during one 12-month period is representative of a \ndifferent 12-month period. This is an issue with any cross-sectional dataset.\nThe absence of data with a time dimension means we also do not aim to assess \nperformance over the lifetime of a project, for example, by measuring the internal \nrate of return. Likewise, we analyse costs in terms of cost per unit of kW\u2009h of \ngeneration rather than the LCOE. Cost per kW\u2009h is a similar metric to an LCOE in \nthat it involves dividing operating costs and a capital cost recovery component (in \nthe form of an annual financing cost) by electricity generation. However, unlike the \nLCOE, it provides an annual snapshot rather than discounting the predicted costs \nand generation over a project\u2019s entire lifetime.\nTo better understand the importance of financing characteristics for \ncommunity energy projects, we explore whether there is a statistical relationship \nbetween the interest rate (cost of finance) and the instrument type. We are \nparticularly interested in comparing community shares with loans because the \nmajority of community energy projects are financed using these instruments. \n(Grants are also a common source of finance but do not charge interest.) To do \nthis we first note that the mean interest rates for community shares and loans in \nour sample are 4.58% and 5.58%, respectively. The difference in means is 1.01 \npercentage points. Performing a t-test on the equality of the mean rates, we find \nthat the means are statistically different at the 1% significance level (t-statistic of \n3.03). Thus, community shares charge on average a statistically lower interest rate \nthan loans.\nA comparison of means may, however, be misleading because the size of the \nfinance obtained and the financing term (duration) may also influence the interest \nrate. We therefore compare the difference in interest rate between community \nshares and loans while holding these other characteristics constant. We do this \nby estimating a linear regression model. We proceed by defining three dummy \nvariables that capture the instrument type:\nCommunityShare\u2009=\u20091 if the financing instrument is community shares, and \nCommunityShare\u2009=\u20090 if it is not community shares.\nBonds\u2009=\u20091 if the financing instrument is bonds, and Bonds\u2009=\u20090 if it is not bonds.\nNature Energy | VOL 5 | FebruarY 2020 | 169\u2013177 | www.nature.com/natureenergy\n175\n\nAnalysis\nNATURe EneRgy\nLoans\u2009=\u20091 if the financing instrument is loans, and Loans\u2009=\u20090 if it is not loans.\nAlthough all three instrument types are included in our model, we need \nto include only two of these three dummy variables in the regression equation \n(further explanation can be found in ref. 66). We choose to include the \nCommunityShare and Bonds dummy variables. Therefore, Loans is the base group \n(or benchmark or omitted group) and is the group against which comparisons are \nmade. We choose loans as the base group because we are especially interested in \nlooking at the difference in interest rate between community shares and loans.\nWe then estimate the linear regression model given by equation (1):\nIRi \u00bc \u03b21 \u00fe \u03b22CommunitySharei \u00fe \u03b23Bondsi \u00fe \u03b24Sizei \u00fe \u03b25Durationi \u00fe \u03f5i\n\u00f01\u00de\nwhere the dependent variable IR is the financing interest rate of financing source i. \nIR is a continuous variable that can take non-integer values. CommunityShare and \nBonds are defined above. As explained previously, we compare how these financing \ninstruments are associated with the interest rate relative to the omitted category, \nwhich is loans. In equation (1) we also include the variables Size and Duration \nto control for the size and duration of the financing instrument, respectively. \nSize is defined as the monetary value of the financing source (in \u00a3 millions) and \nDuration is a dummy variable equal to 1 if the finance term is 240\u2009months or more, \nor indefinite/not specified, and 0 if a relatively short-term duration (less than \n240\u2009months). \u03b21 to \u03b25 are coefficients to be estimated. Finally, \u03b5 is an error term. We \nestimate regression equation (1) using ordinary least squares.\nEach observation on financing source i belongs to an organization that may \nuse one or more sources of finance for its community energy project(s). Outcomes \nfor different financing sources within organizations are likely to be correlated. \nAs we cannot assume that the error term is independently distributed within \norganizations, we cluster standard errors at the organization level.\nIn equation (1), the continuous variables (IR and Size) enter in levels. An \nalternative approach that allows for a non-linear relationship between the \ndependent and explanatory variables is to enter the continuous variables in \nlogarithms. We find our results are robust if we use a logarithmic functional form \n(the results are available on request). However, here we present the results with \nvariables in levels because in this case the coefficients have a percentage  \npoint interpretation.\nWe now present the results from the estimation of regression (1). Here \nwe report the estimated coefficients with cluster-robust standard errors in \nparentheses. For each estimated coefficient, we also report the t-statistic that \nwe calculate to test the null hypothesis that the population coefficient equals \nzero. We find, ^\u03b21 \u00bc 5:124 \u00f00:585\u00de\nI\n and t-statistic\u2009=\u20098.76, ^\u03b22 \u00bc \u001e2:016 \u00f00:706\u00de\nI\n and \nt-statistic\u2009=\u2009\u20132.85, ^\u03b23 \u00bc \u001e0:653 \u00f00:667\u00de\nI\n and t-statistic\u2009=\u2009\u20130.98, ^\u03b24 \u00bc 0:185 \u00f00:091\u00de\nI\n \nand t-statistic\u2009=\u20092.02, and ^\u03b25 \u00bc 1:452 \u00f00:777\u00de\nI\n and t-statistic\u2009=\u20091.87. Finally, the \nR-squared from the regression is 0.2457 and there are 118 observations.\nThe estimated coefficient for the dummy variable CommunityShare (\u20132.016) \nindicates that there is a difference in the interest rate between community shares \nand loans of 2.016 percentage points on average in our sample while holding \nconstant the size and duration of the finance. The t-statistic indicates that this \ndifference is statistically significant at the 1% level. To put this finding into \nperspective, the average size of an individual financial instrument (that is, a \nsingle loan or share issue) in the regression sample is about \u00a3306,000. Therefore, \nfor the average project, the annual interest payment for the first year would be \non average lower by \u00a36,168.96 (2.016% of \u00a3306,000) if financed by community \nshares rather than loans. This does not take into account compound interest and \nrepayments in later years of a project; it is simply intended to illustrate what the \ninterest rate differential between loans and community shares means, in terms of \nactual amounts a community energy project might pay in interest on the initial \nprincipal sum. In the paper, the figures given are rounded for greater readability: \nthus, we mention a \u201c2 percentage points\u201d difference in interest rates and an average \nrepayment differential of \u201cabout \u00a36,200\u201d.\nWe also find that projects financed by bonds do not have an interest rate that is \nsignificantly different from loans. In addition, we find that instruments that have a \nlonger duration and larger value have higher interest rates on average.\nWe investigate whether these results are sensitive to outliers. We do not find \nany evidence of observations with large estimated residuals that may affect the \nestimates. We also investigate the distribution of the dependent and explanatory \nvariables by inspecting the raw data and by using a leverage-versus-squared-\nresidual plot. From this analysis we identify two observations with large leverage \non the estimated coefficients due to outlying values for the explanatory variables. \nHowever, our central findings on the difference in the interest rate between \ncommunity shares and loans are robust to dropping these observations from the \nanalysis. Therefore, they do not affect our conclusions.\nThe impact of the removal of price guarantee schemes is calculated by simply \nsubtracting all price guarantee scheme revenue (FITs, RHI or RO) from total \nproject revenue, project by project, for the single year of revenue data that we \ncollected. It is important to note that, for the FITs, projects retain the tariff rate for \nwhich they initially qualified for the rest of their lifetime, including an inflation \nadjustment; unlike the RO, the FITs is not subject to annual variations in price due \nto market conditions. The RO scheme revenues are affected by year-to-year market \nvariation, but this variation is not itself affected by the scheme being closed to new \nentrants. Therefore, the data do not only reflect the performance of community \nenergy under the tariff rates available to new projects at the time of data collection.\nThis analysis allows an appreciation of the extent to which actual projects \nare reliant on price scheme revenues. There is no consideration for how projects \nmight have been designed if the schemes had not been available, which is a more \ncomplex question. Therefore, these results do not in themselves show that it would \nbe impossible to design a future project to make a financial surplus without a price \nguarantee scheme; nor, given that we have just one year\u2019s data, do they test the \n\u2018viability' of a project over its lifetime.\nTo investigate whether different types of customers pay different rates for \ncommunity-generated energy, we calculate mean rates paid by the four different \ntypes of customer (energy companies, other private sector, public sector, \ncommunity and third sector). As noted in the main body of the text (Table 5), we \nfind that the mean rates differ, with the mean rate lowest for energy companies and \nhighest for public sector customers. Performing a t-test on the equality of the mean \nrates paid by energy companies and public sector organizations, we find that the \nmeans are statistically different at the 1% significance level (t-statistic of 3.69).\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nSource data for Fig. 1 are provided with the paper. The data gathered in the \nFinancing Community Energy survey can be found at https://doi.org/10.5286/\nukerc.edc.000007, or by searching \u2018Financing Community Energy\u2019 on the UK \nEnergy Research Centre\u2019s Energy Data Centre: https://ukerc.rl.ac.uk/\nDue to the terms under which the data were collected, individual project records \ncannot be made public and are therefore not be available. However, aggregated \nrecords of small numbers of similar projects are available.\nReceived: 10 May 2019; Accepted: 20 December 2019;  \nPublished online: 10 February 2020\nReferences\n\t1.\t Braunholtz-Speight, T. et\u00a0al. The Evolution of Community Energy in the UK \n(UK Energy Research Centre, 2018).\n\t2.\t Hoicka, C. & MacArthur, J. From tip to toes: mapping community energy \nmodels in Canada and New Zealand. Energy Policy 121, 162\u2013174 (2018).\n\t3.\t Peters, M. et\u00a0al. Community solar initiatives in the United States of America: \ncomparisons with \u2013 and lessons for \u2013 the UK and other European countries. \nEnergy Policy 121, 355\u2013364 (2018).\n\t4.\t Berka A. in Handbuch Energiewende und Partizipation (eds Holstenkamp, L. \n& Radtke, J.) 1011\u20131035 (Springer VS, 2017).\n\t5.\t Chapman, A. Community Energy in the UK (Green European  \nFoundation, 2019).\n\t6.\t Busch, H. et\u00a0al. Scientific Review Paper on CE Drivers and Barriers \n(Co2mmunity Project, 2019).\n\t7.\t Forman, A. Energy justice at the end of the wire: enacting community energy \nand equity in Wales. Energy Policy 107, 649\u2013657 (2017).\n\t8.\t Bauwens, T., Gotchev, B. & Holstenkamp, L. What drives the development  \nof community energy in Europe? The case of wind power cooperatives. \nEnergy Res. Soc. Sci. 13, 136\u2013147 (2016).\n\t9.\t Seyfang, J., Park, J. J. & Smith, A. A thousand flowers blooming? An \nexamination of community energy in the UK. Energy Policy 61, 977\u2013989 (2013).\n\t10.\tSaunders, R. W., Gross, R. J. K. & Wade, J. Can premium tariffs for \nmicrogeneration and small-scale renewable heat help the fuel poor, and if so, \nhow? Case studies of innovative finance for community energy schemes in \nthe UK. Energy Policy 42, 78\u201388 (2012).\n\t11.\tWalker, G. & Devine-Wright, P. Community renewable energy: what should it \nmean? Energy Policy 36, 497\u2013500 (2008).\n\t12.\tREView: Renewable Energy View 2016 (Renewable Energy Association, 2016).\n\t13.\tState of the Sector Report 2018 (Community Energy England and \nCommunity Energy Wales, 2018).\n\t14.\tvan\u00a0der Schoor, T. & Scholtens, L. J. B. Scientific approaches of community \nenergy: a literature review (Centre for Energy Economics Research, 2019).\n\t15.\tBauwens, T. Analyzing the determinants of the size of investments by \ncommunity renewable energy members: findings and policy implications \nfrom Flanders. Energy Policy 129, 841\u2013852 (2019).\n\t16.\tHolstenkamp, L. & Kahla, F. What are community energy companies trying \nto accomplish? An empirical investigation of investment motives in the \nGerman case. Energy Policy 97, 112\u2013122 (2016).\n\t17.\tYildiz, O. Financing renewable energy infrastructures via financial citizen \nparticipation \u2013 the case of Germany. Renew. Energy 68, 677\u2013685 (2014).\n\t18.\tYildiz, O. et\u00a0al. Renewable energy cooperatives as gatekeepers or facilitators? \nRecent developments in Germany and a multidisciplinary research agenda. \nEnergy Res. Soc. Sci. 6, 59\u201373 (2015).\n\t19.\tCreamer E. et\u00a0al. Community energy: entanglements of community, state and \nprivate sector. Geogr. Compass 12, e12378 (2018).\nNature Energy | VOL 5 | FebruarY 2020 | 169\u2013177 | www.nature.com/natureenergy\n176\n\nAnalysis\nNATURe EneRgy\n\t20.\tHaf, S., Parkhill, K., McDonald, M. & Griffiths, G. Distributing power? \nCommunity energy projects\u2019 experience of planning, policy and incumbents \nin the devolved nations of Scotland and Wales. J. Environ. Plan. Manag. 62, \n921\u2013938 (2019).\n\t21.\tvan Veelen, B. Making sense of the Scottish community energy sector \u2013 an \norganising typology. Scott. Geogr. J. 133, 1\u201320 (2017).\n\t22.\tBomberg, E. & McEwen, N. Mobilising community energy. Energy Policy 51, \n435\u2013444 (2012).\n\t23.\tStrachan, P. A. et\u00a0al. Promoting community renewable energy in a corporate \nenergy world. Sustain. Dev. 23, 96\u2013109 (2015).\n\t24.\tHaggett, C. et\u00a0al. Community energy in Scotland: the social factors for \nsuccess (ClimateXChange, 2013).\n\t25.\tPark, J. J. Fostering community energy and equal opportunities between \ncommunities. Local Environ. 17, 387\u2013408 (2012).\n\t26.\tHoffman, S. & High-Pippert, A. From private lives to collective action: \nrecruitment and participation incentives for a community energy program. \nEnergy Policy 38, 7567\u20137574 (2010).\n\t27.\tWalker, G. What are the barriers and incentives for community-owned means \nof energy production and use? Energy Policy 36, 4401\u20134405 (2008).\n\t28.\tCommunity and Locally Owned Renewable Energy in Scotland at June 2018 \n(Energy Saving Trust, 2018).\n\t29.\tCommunity Energy Finance Roundtable: Final Report and Recommendations \n(Department for Energy and Climate Change, 2014).\n\t30.\tBerka, A. et\u00a0al. A comparative analysis of the costs of onshore wind energy: is \nthere a case for community-specific policy support? Energy Policy 106, \n394\u2013403 (2017).\n\t31.\tHarnmeijer, J. et\u00a0al. The comparative costs of community and commercial \nrenewable energy projects in Scotland (ClimateXChange, 2015).\n\t32.\tHall, S., Foxon, T. J. & Bolton, R. Financing the civic energy sector: how \nfinancial institutions affect ownership models in Germany and the United \nKingdom. Energy Res. Soc. Sci. 12, 5\u201315 (2015).\n\t33.\tDatabuild Research & Solutions Ltd and Energy Saving Trust. Community \nEnergy in the UK: Part 2 (Department for Energy and Climate Change, 2014) .\n\t34.\tOsterwalder, A. & Pigneur, Y. Business Model Generation: A Handbook for \nVisionaries, Game Changers, and Challengers. (John Wiley and Sons, 2010).\n\t35.\tSolar Photovoltaic Deployment (Department for Business, Energy and \nIndustrial Strategy, 2019).\n\t36.\tThe Community Shares Handbook (Community Shares Unit, 2015).\n\t37.\tUrban community energy fund. GOV.UK (Department for Business Energy \nand Industrial Strategy, 2016); https://www.gov.uk/guidance/urban-\ncommunity-energy-fund\n\t38.\tRural community energy fund. GOV.UK (Department for Business Energy \nand Industrial Strategy, 2019); https://www.gov.uk/guidance/rural-\ncommunity-energy-fund\n\t39.\tNolden, C. Governing community energy\u2014Feed-in tariffs and the \ndevelopment of community wind energy schemes in the United Kingdom \nand Germany. Energy Policy 63, 543\u2013552 (2013).\n\t40.\tMitchell, C., Bauknecht, D. & Connor, P. M. Effectiveness through risk \nreduction: a comparison of the renewable obligation in England and Wales \nand the feed-in system in Germany. Energy Policy 34, 297\u2013305 (2006).\n\t41.\tCouture, T. & Gagnon, Y. An analysis of feed-in tariff remuneration  \nmodels: implications for renewable energy investment. Energy Policy 38, \n955\u2013965 (2010).\n\t42.\tPrices of fuels purchased by non-domestic consumers in the UK (Department \nfor Business, Energy and Industrial Strategy, 2018).\n\t43.\tLehner, O. Crowdfunding social ventures: a model and research agenda. \nVenture Cap. 15, 289\u2013311 (2013).\n\t44.\tFloyd, D., Gregory, D. & Wilson, N. After the Gold Rush: Report of The \nAlternative Commission on Social Investment (The Alternative Commission \non Social Investment, 2015).\n\t45.\tHall S., Foxon, T. J. & Bolton, R. Investing in low-carbon transitions: energy \nfinance as an adaptive market. Climate Policy 17, 280\u2013298 (2017).\n\t46.\tBlyth, W., McCarthy, R. & Gross, R. J. K. Financing the UK power sector: is \nthe money available? Energy Policy 87, 607\u2013622 (2015).\n\t47.\tUnruh, G. Understanding carbon lock-in. Energy Policy 28, 817\u2013830 (2000).\n\t48.\tMcInerney, C. & Bunn, D. W. Expansion of the investor base for the energy \ntransition. Energy Policy 129, 1240\u20131244 (2019).\n\t49.\tMazzucato, M. & Semieniuk, G. Financing renewable energy: who is \nfinancing what and why it matters? Technol. Forecast. Soc. Change 127,  \n8\u201322 (2018).\n\t50.\tBolton, R. & Foxon, T. A socio-technical perspective on low carbon \ninvestment challenges \u2013 insights for UK energy policy. Environ. Innov. Soc.\nTransit. 14, 165\u2013181 (2015).\n\t51.\tInside the Market Report: June 2015 (Community Shares Unit, 2015).\n\t52.\tLundberg, L. Auctions for all? Reviewing the German wind power auctions in \n2017. Energy Policy 128, 449\u2013458 (2019).\n\t53.\tLeiren, M. D. & Reimer, I. Historical institutionalist perspective on the shift \nfrom feed-in tariffs towards auctioning in German renewable energy policy. \nEnergy Res. Soc. Sci. 43, 33\u201340 (2018).\n\t54.\tMirzania, P., Ford, A., Andrews, D., Ofori, G. & Maidment, G. The impact of \npolicy changes: the opportunities of community renewable energy projects in \nthe UK and the barriers they face. Energy Policy 129, 1282\u20131296 (2019).\n\t55.\tJoint response to the Department for Business, Energy and Industrial Strategy \nConsultation on a Smart Export Guarantee (Community Energy England, \nCommunity Energy Wales and Community Energy Scotland, 2019).\n\t56.\tContracts for Difference (CfD): Draft Budget Notice for the third allocation \nround (Department for Business, Energy and Industrial Strategy, 2018).\n\t57.\tR Core Team R: A language and environment for statistical computing  \n(R Foundation for Statistical Computing, 2019).\n\t58.\tWickham, H., Francois, R., Henry, L. & Muller, K. dplyr: A grammar of data \nmanipulation. R package version 0.7.1. (2017).\n\t59.\tMaechler, M., Rousseeuw, P., Struyf, A., Hubert, M. & Hornik, K. cluster: \nCluster analysis basics and extensions. R package version 2.0.4. (2016).\n\t60.\tKassambara, A. & Mundt, F. factoextra: Extract and visualize the results of \nmultivariate data analyses. R package version 3.1.2. (2017).\n\t61.\tWickham, H. ggplot2: Elegant Graphics for Data Analysis (Springer, 2009).\n\t62.\tKrijthe, J. H. Rtsne: T-distributed stochastic neighbor embedding using a \nBarnes-Hut implementation. R package version 0.15. (2015).\n\t63.\tHahsler, M. & Piekenbrock, M. dbscan: Density based clustering of \napplications with noise (DBSCAN) and related algorithms (2019).\n\t64.\tHennig, C. fpc: Flexible procedures for clustering. R package version  \n2.2-3 ed. (2019).\n\t65.\tSzepannek, G. clustMixType: user-friendly clustering of mixed-type data in R. \nR J. 10, 200\u2013208 (2018).\n\t66.\tWooldridge, J. Introduction to Econometrics: Europe, Middle East and Africa \nedition (Cengage Learning, 2014).\n\t67.\tEnvironmental Programmes (Ofgem, 2019); https://www.ofgem.gov.uk/\nenvironmental-programmes\nAcknowledgements\nThis research was funded as part of the UK Energy Research Centre Phase 3 research \nprogramme (grant number EP/L024756/1). C. Birch and C. Walsh, doctoral researchers \nat the Tyndall Centre for Climate Change Research, University of Manchester, also \nprovided research assistance for the survey. We are grateful for comments from seminar \nparticipants at several universities and conferences. Any remaining errors are our own. \nWe would like to thank the survey participants for their time and data, and Community \nEnergy England, in particular E. Bridge and J. Hall, for their practical support and for \nsharing their State of the Sector data with the research team (further details on this is in \nthe Methods section). We would like to thank SCENE (and in particular J. Harnmeijer \nand S. Robinson) for additional data which, although they are not analysed in this paper, \nwere very helpful to us in framing our survey work, as noted in the Methods section. We \nwould also like to thank Local Energy Scotland for assistance with marketing the survey.\nAuthor contributions\nC.M. led the research project. C.M., S.M., M.S., E.M., J.H., M.H. and T.B-S. contributed \nto the conception, framing and design of the survey research. T.B-S. conducted the \nsurvey and supervised the work of C.B. and C.W.; he also conducted the analysis of \nthe composition of project finances and the impact of price support mechanisms. M.S. \ndesigned and conducted the cluster analysis. E.M. conducted the econometric analyses \nof financing instrument interest rates and provided descriptive statistical analysis. All \nauthors jointly wrote the paper: T.B-S. led the writing; M.S., E.M., C.M., M.H., J.H. and \nS.M. contributed text and extensive comments on the structure and content of several \ndrafts of the paper.\nCompeting interests\nC.M. is Chair of the Trustees of the climate change charity Possible (formerly 10:10), \nand a director of Community Energy North. Both of these roles are unpaid. M.H. is an \nunpaid Trustee of South Seeds, Glasgow, a community environmental charity with a \nfocus on energy. J.H. is a non-executive director of Public Power Solutions Limited, a \nrenewable energy developer that has worked with community groups.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-019-0546-4.\nCorrespondence and requests for materials should be addressed to T.B.-S., M.S. or E.M.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2020\nNature Energy | VOL 5 | FebruarY 2020 | 169\u2013177 | www.nature.com/natureenergy\n177\n\n1\nnature research  |  reporting summary\nOctober 2018\nCorresponding author(s):\nTim Braunholtz-Speight\nLast updated by author(s): Oct 15, 2019\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nData was collected using SelectSurvey.NETv4.075.003, and skype and telephone interviews with data entered into Microsoft Excel 2010.\nData analysis\nR 3.6.0 was used to perform cluster analysis, using the following R packages: dplyr, cluster, factoextra, ggplot2, Rtsne, and clustMixType. \nMicrosoft Excel 2010 was used for the descriptive statistical analyses of financing mechanisms in relation to CAPEX (Figure 1), community \nshare marketing mechanisms (Table 6), and for the analysis of project revenues in relation to the Feed-in Tariff scheme.  \nStata / SE 14.2 was used for generating data presented on energy generation project characteristics (Table 5) and energy prices by \ncustomer type (Table 7), and for the regression analysis on the determinants of the cost of finance (Equation 1 in Methods).\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. \nWe strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nThe data gathered in the Financing Community Energy survey will be available via the UK Energy Research Centre\u2019s Energy Data Centre: https://ukerc.rl.ac.uk/. Due \nto the terms under which the data were collected, individual project records cannot be made public and will therefore not be available. However, aggregated \nrecords of small numbers of similar projects will be available. In addition, the (aggregated) source data used to generate Figure 1 is available as a standalone file, \nand has been uploaded to the Nature Energy system.\n\n2\nnature research  |  reporting summary\nOctober 2018\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nQuantitative survey of activities, structures and financial characteristics of community energy companies in the UK.\nResearch sample\nCommunity energy companies in the UK. Sample is not representative. Study aim was to investigate the financial performance and \nbusiness models of community energy companies in the UK. Individuals participating in the research were participating in their capacity \nas board members or members of staff of community energy companies.\nSampling strategy\nWe aimed to collect data on a minimum of 100 active projects in order to conduct meaningful econometric and cluster analyses. We had \nno maximum sample size, and attempted to collect data on as many projects as possible. The breakdown of activity types recorded in the \nCommunity Energy England State of the Sector 2017, and in data made available for Scotland by Scene Connect, was used as a reference \npoint for the distribution of the different activities, and renewable energy technologies, present in the sector. However, as no definitive \ncensus of the sector exists, the researchers did not calculate precise quotas of different types of project. The study is presented with the \ncaveat that the dataset is not necessarily representative of the sector.\nData collection\nCommunity energy companies were contacted via email and, if they indicated they were willing to participate in the survey, they were \noffered the choice of completing an online questionnaire on their own, or having the questionnaire administered by a researcher over \nthe telephone, or via skype.  \n \nThe questionnaire was constructed using Select Survey software, and hosted on the University of Manchester website. Pre-set multiple \nchoice formats were used as far as possible to facilitate data coding and quantitative analysis. Some free text qualitative questions were \nalso included, particularly in relation to organisations\u2019 future plans: responses to these questions have been fed into other parts of the \noverall research project and are not reported in the paper. \n \nResponses were downloaded into an individual excel spreadsheet for each participating company, from which a master sheet was \nconstructed. Researchers conducting telephone interviews either entered data into the online questionnaire, or directly onto a \nspreadsheet using the same data categories as the online questionnaire. \n \nData collected through online or telephone-administered questionnaires were augmented with data provided under agreement with \nCommunity Energy England from their State of the Sector survey 2017, and data collected from company documents (e.g. annual \nreports) - either provided by study participants, or accessed online. All of this was explained to participants. The rationale was to \nminimise the demands on participants\u2019 time. \n \nFollowing the completion of survey-based data collection, data was collected for 8 additional renewable energy projects exclusively from \npublicly-available online sources (e.g. company annual reports). \nTiming\nThe questionnaire was piloted in October \u2013 November 2017 with three community energy organisations. Only minor changes were made \nafter the piloting process, and the pilot data forms part of the survey dataset analysed in this paper. The full survey was launched in \nNovember 2017 and closed in May 2018. During January and February 2018 it was suspended in England, Wales and Northern Ireland to \navoid an overlap with the 2018 iteration of Community Energy England\u2019s \u2018State of the Sector\u2019 survey.\nData exclusions\nData were excluded from the analyses when there was not sufficient data on a particular project to perform the analysis, or where the \nanalysis was not relevant to a particular project. In all cases, the total number of projects analysed, and the rationale on which projects \nwere selected for analysis, is included in the text of the paper. Data were collected on 20 projects that were not fully operational - either \nin active development, or 'on hold' - but these were excluded from most analyses in the paper which focus on the characteristics and \nperformance of operational projects. (These data will be used for other parts of the overall research project.) \n \nThe details of data exclusions for specific analyses presented in the paper are: \nCluster analysis (Tables 2 and 3): 35 projects were excluded for insufficient data. \nFinancing mechanisms vs CAPEX (Figure 1): 42 projects were excluded. 27 were excluded for insufficient data; 15 were excluded as the \nanalysis focussed on energy generation projects only. \nProject characteristics (Table 4): data from projects that are not fully operational were excluded.  \nCommunity share marketing mechanisms (Table 5): 63 projects were excluded, as they did not use community shares to raise funds. \nProject reliance on FITs, RHI or RO revenues: 43 projects were excluded, either because they did not use these revenue schemes, or \nthere was insufficient data to perform the analysis. \nRegression analysis on cost of finance (equation (1)): projects with missing data were excluded.\nNon-participation\nResponse rate was 17%. 280 organisations were contacted and 48 organisations completed the survey. A further 18 declined to \nparticipate, 18 responded expressing interest in the survey but did not complete the questionnaire, and the remaining 196 did not \nrespond at all.\n\n3\nnature research  |  reporting summary\nOctober 2018\nRandomization\nNo allocation into experimental groups\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology\nAnimals and other organisms\nHuman research participants\nClinical data\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nAntibodies\nAntibodies used\nDescribe all antibodies used in the study; as applicable, provide supplier name, catalog number, clone name, and lot number.\nValidation\nDescribe the validation of each primary antibody for the species and application, noting any validation statements on the \nmanufacturer\u2019s website, relevant citations, antibody profiles in online databases, or data provided in the manuscript.\nEukaryotic cell lines\nPolicy information about cell lines\nCell line source(s)\nState the source of each cell line used.\nAuthentication\nDescribe the authentication procedures for each cell line used OR declare that none of the cell lines used were authenticated.\nMycoplasma contamination\nConfirm that all cell lines tested negative for mycoplasma contamination OR describe the results of the testing for \nmycoplasma contamination OR declare that the cell lines were not tested for mycoplasma contamination.\nCommonly misidentified lines\n(See ICLAC register)\nName any commonly misidentified cell lines used in the study and provide a rationale for their use.\nPalaeontology\nSpecimen provenance\nProvide provenance information for specimens and describe permits that were obtained for the work (including the name of the \nissuing authority, the date of issue, and any identifying information).\nSpecimen deposition\nIndicate where the specimens have been deposited to permit free access by other researchers.\nDating methods\nIf new dates are provided, describe how they were obtained (e.g. collection, storage, sample pretreatment and measurement), \nwhere they were obtained (i.e. lab name), the calibration program and the protocol for quality assurance OR state that no new \ndates are provided.\nTick this box to confirm that the raw and calibrated dates are available in the paper or in Supplementary Information.\nAnimals and other organisms\nPolicy information about studies involving animals; ARRIVE guidelines recommended for reporting animal research\nLaboratory animals\nFor laboratory animals, report species, strain, sex and age OR state that the study did not involve laboratory animals.\nWild animals\nProvide details on animals observed in or captured in the field; report species, sex and age where possible. Describe how animals \nwere caught and transported and what happened to captive animals after the study (if killed, explain why and describe method; if \nreleased, say where and when) OR state that the study did not involve wild animals.\nField-collected samples\nFor laboratory work with field-collected samples, describe all relevant parameters such as housing, maintenance, temperature, \nphotoperiod and end-of-experiment protocol OR state that the study did not involve samples collected from the field.\nEthics oversight\nIdentify the organization(s) that approved or provided guidance on the study protocol, OR state that no ethical approval or \n\n4\nnature research  |  reporting summary\nOctober 2018\nEthics oversight\nguidance was required and explain why not.\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nHuman research participants\nPolicy information about studies involving human research participants\nPopulation characteristics\nSee above.\nRecruitment\nParticipants were recruited by email invitation. Emails were obtained under agreement from Community Energy England, and \nfrom online searches. The study was also publicised via the Local Energy Scotland email newsletter, with a link and contact \ndetails of the researchers.  \n \nThese recruitment processes might favour established groups, already in contact with other sector bodies, rather than the very \nsmallest groups, or emerging groups. However, as the study focussed on groups engaged in trading activities, and most of our \nanalyses required at least one year\u2019s trading data, we believe that the possible lack of contact with some very new or very small \ngroups did not significantly affect the results. Nevertheless, we did gather data on a number of non-operational projects that \nwere in development or 'on hold', and we did receive a few initial responses from groups that were too new or too small to have \nrelevant data for us. We therefore believe that our recruitment procedures reached a wide range of potential research \nparticipants.  \nEthics oversight\nThe study received research ethics approval from the University of Manchester.\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nClinical data\nPolicy information about clinical studies\nAll manuscripts should comply with the ICMJE guidelines for publication of clinical research and a completed CONSORT checklist must be included with all submissions.\nClinical trial registration\nProvide the trial registration number from ClinicalTrials.gov or an equivalent agency.\nStudy protocol\nNote where the full trial protocol can be accessed OR if not available, explain why.\nData collection\nDescribe the settings and locales of data collection, noting the time periods of recruitment and data collection.\nOutcomes\nDescribe how you pre-defined primary and secondary outcome measures and how you assessed these measures.\nChIP-seq\nData deposition\nConfirm that both raw and final processed data have been deposited in a public database such as GEO.\nConfirm that you have deposited or provided access to graph files (e.g. BED files) for the called peaks.\nData access links \nMay remain private before publication.\nFor \"Initial submission\" or \"Revised version\" documents, provide reviewer access links.  For your \"Final submission\" document, \nprovide a link to the deposited data.\nFiles in database submission\nProvide a list of all files available in the database submission.\nGenome browser session \n(e.g. UCSC)\nProvide a link to an anonymized genome browser session for \"Initial submission\" and \"Revised version\" documents only, to \nenable peer review.  Write \"no longer applicable\" for \"Final submission\" documents.\nMethodology\nReplicates\nDescribe the experimental replicates, specifying number, type and replicate agreement.\nSequencing depth\nDescribe the sequencing depth for each experiment, providing the total number of reads, uniquely mapped reads, length of \nreads and whether they were paired- or single-end.\nAntibodies\nDescribe the antibodies used for the ChIP-seq experiments; as applicable, provide supplier name, catalog number, clone \nname, and lot number.\nPeak calling parameters\nSpecify the command line program and parameters used for read mapping and peak calling, including the ChIP, control and \nindex files used.\nData quality\nDescribe the methods used to ensure data quality in full detail, including how many peaks are at FDR 5% and above 5-fold \nenrichment.\n\n5\nnature research  |  reporting summary\nOctober 2018\nSoftware\nDescribe the software used to collect and analyze the ChIP-seq data. For custom code that has been deposited into a \ncommunity repository, provide accession details.\nFlow Cytometry\nPlots\nConfirm that:\nThe axis labels state the marker and fluorochrome used (e.g. CD4-FITC).\nThe axis scales are clearly visible. Include numbers along axes only for bottom left plot of group (a 'group' is an analysis of identical markers).\nAll plots are contour plots with outliers or pseudocolor plots.\nA numerical value for number of cells or percentage (with statistics) is provided.\nMethodology\nSample preparation\nDescribe the sample preparation, detailing the biological source of the cells and any tissue processing steps used.\nInstrument\nIdentify the instrument used for data collection, specifying make and model number.\nSoftware\nDescribe the software used to collect and analyze the flow cytometry data. For custom code that has been deposited into a \ncommunity repository, provide accession details.\nCell population abundance\nDescribe the abundance of the relevant cell populations within post-sort fractions, providing details on the purity of the samples \nand how it was determined.\nGating strategy\nDescribe the gating strategy used for all relevant experiments, specifying the preliminary FSC/SSC gates of the starting cell \npopulation, indicating where boundaries between \"positive\" and \"negative\" staining cell populations are defined.\nTick this box to confirm that a figure exemplifying the gating strategy is provided in the Supplementary Information.\nMagnetic resonance imaging\nExperimental design\nDesign type\nIndicate task or resting state; event-related or block design.\nDesign specifications\nSpecify the number of blocks, trials or experimental units per session and/or subject, and specify the length of each trial \nor block (if trials are blocked) and interval between trials.\nBehavioral performance measures\nState number and/or type of variables recorded (e.g. correct button press, response time) and what statistics were used \nto establish that the subjects were performing the task as expected (e.g. mean, range, and/or standard deviation across \nsubjects).\nAcquisition\nImaging type(s)\nSpecify: functional, structural, diffusion, perfusion.\nField strength\nSpecify in Tesla\nSequence & imaging parameters\nSpecify the pulse sequence type (gradient echo, spin echo, etc.), imaging type (EPI, spiral, etc.), field of view, matrix size, \nslice thickness, orientation and TE/TR/flip angle.\nArea of acquisition\nState whether a whole brain scan was used OR define the area of acquisition, describing how the region was determined.\nDiffusion MRI\nUsed\nNot used\nPreprocessing\nPreprocessing software\nProvide detail on software version and revision number and on specific parameters (model/functions, brain extraction, \nsegmentation, smoothing kernel size, etc.).\nNormalization\nIf data were normalized/standardized, describe the approach(es): specify linear or non-linear and define image types \nused for transformation OR indicate that data were not normalized and explain rationale for lack of normalization.\nNormalization template\nDescribe the template used for normalization/transformation, specifying subject space or group standardized space (e.g. \noriginal Talairach, MNI305, ICBM152) OR indicate that the data were not normalized.\n\n6\nnature research  |  reporting summary\nOctober 2018\nNoise and artifact removal\nDescribe your procedure(s) for artifact and structured noise removal, specifying motion parameters, tissue signals and \nphysiological signals (heart rate, respiration).\nVolume censoring\nDefine your software and/or method and criteria for volume censoring, and state the extent of such censoring.\nStatistical modeling & inference\nModel type and settings\nSpecify type (mass univariate, multivariate, RSA, predictive, etc.) and describe essential details of the model at the first \nand second levels (e.g. fixed, random or mixed effects; drift or auto-correlation).\nEffect(s) tested\nDefine precise effect in terms of the task or stimulus conditions instead of psychological concepts and indicate whether \nANOVA or factorial designs were used.\nSpecify type of analysis:\nWhole brain\nROI-based\nBoth\nStatistic type for inference\n(See Eklund et al. 2016)\nSpecify voxel-wise or cluster-wise and report all relevant parameters for cluster-wise methods.\nCorrection\nDescribe the type of correction and how it is obtained for multiple comparisons (e.g. FWE, FDR, permutation or Monte \nCarlo).\nModels & analysis\nn/a Involved in the study\nFunctional and/or effective connectivity\nGraph analysis\nMultivariate modeling or predictive analysis\nFunctional and/or effective connectivity\nReport the measures of dependence used and the model details (e.g. Pearson correlation, partial \ncorrelation, mutual information).\nGraph analysis\nReport the dependent variable and connectivity measure, specifying weighted graph or binarized graph, \nsubject- or group-level, and the global and/or node summaries used (e.g. clustering coefficient, efficiency, \netc.).\nMultivariate modeling and predictive analysis\nSpecify independent variables, features extraction and dimension reduction, model, training and evaluation \nmetrics.\n\n\n    Scientific Research Findings:", "answer": "The UK community energy sector is dominated by renewable electricity generation. Activities addressing demand-side issues, such as energy efficiency or fuel poverty, are mostly cross-subsidized from renewables revenue or grant funded, although a few groups do run financially self-sustaining demand-side projects. For renewables, two basic business models exist. First, larger projects supplying the grid, like wind or solar farms, are increasingly professionalized and \u2018bankable\u2019: they raise commercial loans alongside citizen finance. Second, rooftop solar photovoltaic projects, supplying an on-site customer as well as the grid, are small enough to be funded primarily through community share issues. In both cases, community shares represent a low-cost source of finance: we find that on average, they offer interest rates two percentage points lower than loans, making them the cheapest form of capital (other than grants). However, these two business models rely on price guarantee schemes, such as the Feed-in Tariff. Over 90% of the projects in our sample made a financial surplus in our single-year snapshot, but this falls to just 20% if we remove Feed-in Tariff income.", "id": 26}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Analysis\nhttps://doi.org/10.1038/s41560-019-0429-8\n1Institute for Resources, Environment and Sustainability,  University of British Columbia, Vancouver, British Columbia, Canada. 2International Institute \nfor Applied Systems Analysis, Laxenburg, Austria. 3Stockholm Environment Institute, United States Centre, Somerville, MA, USA. 4Forest Resources \nManagement, Faculty of Forestry,  University of British Columbia, Vancouver, British Columbia, Canada. *e-mail: abhishek.kar@alumni.ubc.ca\nC\nlean cooking energy transitions away from solid fuels present \na major public policy challenge for the Global South1,2. About \n2.9 billion people lack access to clean cooking fuels such as \ngas and electricity and burn solid fuels in open fires and other simple \ndevices3. Exposure to household air pollution from solid fuel com-\nbustion is associated with 1.6 million avoidable deaths each year4 and \nresults in a global welfare loss of roughly US$1.5 trillion5. Worldwide, \n~30% of wood fuels burnt for cooking are not sustainably harvested. \nThis contributes to forest degradation and global emissions of  \n1.0\u20131.2\u2009GtCO2e6. Women often engage in solid fuel collection, thereby  \nlimiting their economic and educational opportunities7,8.\nTo tackle these negative externalities, India recently launched the \nPradhan Mantri Ujjwala Yojana (hereafter, PMUY) to promote lique-\nfied petroleum gas (LPG)1,9. The programme provides poor women \nwith a one-time subsidy and an optional loan to cover the initial up-\nfront cost1,10 (see Supplementary Figs. 1\u20134 in Supplementary Note 1 \nfor details on PMUY, and Supplementary Table 1 in Supplementary \nNote 2 for details of enrolment costs). PMUY\u2019s deployment in low-\nincome, rural settings is unmatched in terms of scale and pace in \nthe history of modern cooking energy access. Seventy million poor \nhouseholds11 took advantage of PMUY to purchase LPG stove kits \nwithin 35 months since its launch. PMUY has garnered interna-\ntional attention12,13 and enquiries from other countries hoping to \nreplicate elements of the programme14.\nHowever, there has been no objective evaluation of the extent \nto which PMUY has actually induced a transition away from solid \nfuels15. Assessments of LPG use, beyond just acquisition, are critical, \nas the health, social, economic and environmental benefits of a tran-\nsition2 are conditional on the extent to which solid fuels are replaced \nby LPG16. Typically, a clean cooking transition starts with the pur-\nchase of an LPG starter kit and partial fuel stacking. For a successful \ntransition, this needs to be followed by sustained and growing LPG \nuse17, and eventually, complete solid fuel displacement by LPG18,19.\nAt a national level, aggregate state-level consumption data  \n(A. Jindal, personal communication) for PMUY beneficiaries (as of \nNovember 2018) show that about half of the total beneficiaries from \n30 states have been enrolled for at least 1 year. Of the PMUY benefi-\nciaries who have completed at least 1 year, 28% purchased 5 or more \ncylinders, while 24% did not return for even a single refill purchase \nin their first year (see Supplementary Note 1 for background infor-\nmation on India\u2019s LPG market and state-level patterns of use).\nNotably, LPG is a commercial fuel and sold in cylinders (14.2\u2009kg \nor 5\u2009kg). All beneficiaries buy the first (installation) cylinder with \nthe stove kit and have to purchase refill cylinders for continued use20. \nHereafter, cylinder implies 14.2\u2009kg and total LPG purchase includes \ninstallation cylinder unless mentioned otherwise. Typically, 10\u201312 \ncylinders are required annually to use LPG as an exclusive cooking \nfuel for an average-sized family20. PMUY beneficiaries do not get \nincentives for refill purchases, beyond a subsidy that is available to \nall LPG consumers.\nFor the purposes of this study, we view transitions as a two-\nstep process\u2014acquisition of an LPG kit and sustained use through \nthe regular purchase of LPG19. The core premise of PMUY1 is that \n\u2018deprived\u2019 women aspire to use LPG21 but cannot afford the up-\nfront capital cost22. In other words, the growth in the number of \nPMUY consumers should ideally be additive (that is, in addition to \nany business-as-usual (BAU) growth that would have happened in \nthe absence of PMUY). However, it is reported that \u2018many well-to-\ndo [sic] households\u2014who could otherwise afford LPG\u2019 benefitted \nfrom PMUY23. If PMUY targets consumers who could and would \nhave purchased LPG without financial incentives, \u2018subsidy target-\ning\u2019 emerges as an important line of inquiry. Hence, from a policy \nevaluation perspective, to understand the impact of PMUY, assess-\ning the growth in the number of customers after accounting for the \nBAU trend is required to assess the actual \u2018purchase\u2019 transition.\nThere is also an assumption in the PMUY promotions24,25 that \nonce the poor are unburdened with the capital outlay for stoves, they \nwill be able to buy enough LPG to \u2018lead a smoke-free, less polluted, \nconvenient and healthy life\u201926. Past studies17\u201319 suggest, however, \nthat purchases by low-income rural households do not necessarily \nUsing sales data to assess cooking gas adoption \nand the impact of India\u2019s Ujjwala programme in \nrural Karnataka\nAbhishek Kar\u200a \u200a1*, Shonali Pachauri\u200a \u200a2, Rob Bailis\u200a \u200a3 and Hisham Zerriffi\u200a \u200a4\nMore than 70 million poor women in India have received liquefied petroleum gas (LPG) stoves within the first 35 months under \na government programme, Pradhan Mantri Ujjwala Yojana (PMUY). Here, we analyse multi-year LPG sales data from a district \nin Karnataka to assess enrolment and consumption trends for both PMUY beneficiaries and general (non-PMUY) rural consum-\ners. We find rapid growth in enrolments of LPG consumers, but this is not matched by an increase in LPG sales, suggesting that \nLPG access has not induced a full transition away from the use of polluting solid fuels. The number of LPG refills among PMUY \nbeneficiaries is less than half that of rural general consumers. We also find no observable increase in LPG consumption among \ngeneral rural consumers with years of experience. These results suggest that mid-course policy revisions to encourage regular \nLPG use are needed for both PMUY and general rural consumers.\nNature Energy | VOL 4 | SEPTEMBER 2019 | 806\u2013814 | www.nature.com/natureenergy\n806\n\nAnalysis\nNaTUre EnerGY\nresult in sustained use23,27,28. The gap between consumption (refill \npurchases) and consumer growth rates are important indicators of \nthe impact of PMUY\u2019s one-time capital cost incentive on promoting \nregular usage (repeat refill purchases).\nIndia, like many developing countries, has a large urban\u2013rural \ndisparity in LPG use. Urban populations have greater access to \nLPG distributors, lower access to solid fuels, higher incomes and \nhigher opportunity costs of labour1,18,28,29. Urban women also have \nhigher income and education levels. These indicators are tradition-\nally associated with greater decision-making autonomy and lower \ngender disparity30. Hence, comparing rural PMUY beneficiary con-\nsumption levels against urban or national averages may be mislead-\ning, especially in the short term, and LPG consumption by general \nconsumers in rural areas may provide a better benchmark.\nThe literature on technology adoption, in general, suggests that \nincreasing familiarity with technology can help reduce people\u2019s \nattachment to earlier technologies due to greater comfort, expertise \nand confidence that come with experience31,32. A recent study sug-\ngests that LPG consumption in general households is strongly cor-\nrelated with experience, with a median difference of two cylinders \nbetween consumers after their first year and consumers after their \ntenth year23. This remains to be assessed for PMUY beneficiaries.\nPast literature also suggests that household demographics, fuel \nprices, socio-economics and climatic conditions influence the con-\nsumption of clean cooking fuels33\u201335. A better understanding of the \nquantitative impact of LPG price and seasonality on LPG purchases \nis, however, largely absent in the literature, as most studies are cross-\nsectional in nature36. To design better promotional campaigns and \noptimize pricing policies, understanding the relative effect of price \nand seasonality can be informative.\nLiterature hitherto published on PMUY15,20,23,28,37,38 has primarily \nfocused on national-level LPG connections, average consumption \nor in-depth surveys of PMUY beneficiaries from rural areas based \non self-reported consumption data. These prior works are either too \ngeneral (national-level and average consumption figures) or need to \nbe complemented with other data sources on consumption. Here, we \nuse an original dataset of LPG sales transactions records (>200,000) \nfor more than 25,000 rural (PMUY and general) consumers from \nKoppal district in Karnataka state to address these knowledge gaps \n(see Methods) to assess the efficacy of PMUY in promoting a cook-\ning energy transition, and the determinants of adoption, which can \ninform potential mid-course corrections, and similar future initia-\ntives. The dataset used contains details for all LPG sales transactions \nin Koppal district (see Supplementary Note 3 and Supplementary \nTable 2 for a socio-economic profile of the region) between January \n2016 and December 2018 for three distributors of a public sector \nLPG marketing company (Indian Oil) and data from earlier than \n2016 for distributors 1 and 2 (D1 and D2; see Methods). While \nthe results presented here are specific to the conditions within the \nKoppal region, they address key questions around LPG use and \npoint to implementation issues that may have broader implications. \nThey also demonstrate the value of large consumer purchase data \nanalysis to support transparent programme evaluation.\nPMUY and LPG consumer growth\nPMUY was launched in Koppal in June 2017. Within 16 months, \nPMUY beneficiaries in this region exceeded the number of general \nrural consumers. By the end of the available data window (December \n2018), there were approximately 15,000 PMUY customers and \n12,500 general customers in the database. The median monthly \ngrowth rate in PMUY customers was approximately six times \nthat of the general customers over the same time period (Fig. 1)  \nand twice that of the general customers in the pre-PMUY period.\nHowever, that rapid increase in PMUY customers is also associ-\nated with a drop in the growth of general customers. The median \nvalue for the distribution of month-to-month general consumer \ngrowth rate decreased from 2.4% to 0.9% after PMUY was rolled \nout (Fig. 1). Overall, the compounded monthly enrolment growth \nrate among general consumers dropped by half from 3.2% to 1.5% \nduring this time period. Under a BAU trend, with 3.2% monthly \ngrowth, the number of (general) consumers ought to have increased \nfrom 9,629 consumers (July 2017) to 16,423 (December 2018). \nInstead, the general consumer numbers reached only 12,364. The \ncurrent number of actual consumers (27,431) would probably not \nhave been reached until April 2020. Thus, after adjusting for the \nBAU trend, we estimate that PMUY has been able to fast-track LPG \nconsumer enrolments by 16 months, but at the same time there was \na drop in general customers from the expected number.\nWe consider two explanations for the decline in growth of gen-\neral consumers. First, some PMUY-eligible consumers would have \nbecome consumers regardless but are now enrolled as PMUY con-\nsumers and are, essentially, \u2018free-riders\u2019 on the programme. If the \nlower rate of growth of general consumers is due to a diversion of \npotential general consumers to PMUY beneficiaries, then PMUY\u2019s \nabsolute addition of 15,067 consumers should be adjusted for the \nloss in general consumers to examine PMUY\u2019s effective contribu-\ntion to uptake of LPG in rural households. This adjustment suggests \nthat the effective impact of PMUY in expanding the consumer base \nis only 73% of the absolute number of enrolled PMUY consumers. \nA second explanation emerges from interviews with distributors \nand executives from oil marketing companies (OMCs)\u2014the three \npublic sector companies responsible for household LPG distribu-\ntion and marketing. These informants suggest that people ineligible \nfor PMUY may have adopted a \u2018wait-and-watch\u2019 approach, hoping \nthat rules would change, making them eligible in the future. Indeed, \nthe eligibility has already been expanded two times in the last 18 \nmonths (see Supplementary Note 1). A combination of both factors \nmay be at play. Hence, PMUY enrolment numbers are not strictly \nadditive in nature.\nWe also find the rapid expansion in LPG users attributable to \nenrolment in PMUY has not resulted in a comparable increase in \nLPG sales (Fig. 2). There were only ~50 daily refill sales recorded \nfor the 15,000 PMUY customers in December 2018. By con-\ntrast, the 12,000 general consumers purchase ~150 refills daily \n(see Supplementary Fig. 5). This is not surprising, as the poor-\nest households were specifically targeted by the programme and \ntheir rapid inclusion into the consumer pool has reduced average \nconsumption levels. Thus, while PMUY has been successful in \npromoting LPG enrolment among the rural poor1,20, their enrol-\nment has led to an overall decline (see Supplementary Fig. 6 in \nSupplementary Note 4) in consumption among households with \nan LPG connection.\nLPG use in PMUY beneficiaries versus general consumers\nWe compare normalized and adjusted monthly refill sales for PMUY \nand general consumers (see Methods). Figure 3 shows monthly \nsales from January 2016 to December 2018 (equally split between \nthe pre-PMUY and post-PMUY period) normalized by the number \nof registered consumers.\nIn recent months, PMUY refill sales have fluctuated around \n100 cylinders per 1,000 consumers (suggesting rare use), while the \nmonthly refill sales for general consumers is around 400 cylinders \nper 1,000 consumers. This suggests that average general consumers \nuse LPG as a secondary or primary cooking fuel. Considering that \nthe national (urban\u2009+\u2009rural) monthly average is about 600 cylinders \nper 1,000 consumers, there is considerable scope to encourage more \nregular use among general consumers in rural areas.\nTo avoid capturing any effect of experience, we also compare \nLPG consumption trends of both general and PMUY customers for \njust their first year as customers (Fig. 4). The mean consumption \nrate for PMUY and general customers is 2.3 and 4.7, respectively \n(this includes refill purchases as in Fig. 3 plus the initial cylinder). \nNature Energy | VOL 4 | SEPTEMBER 2019 | 806\u2013814 | www.nature.com/natureenergy\n807\n\nAnalysis\nNaTUre EnerGY\nThe gap between PMUY and general consumers widens moving \nacross the purchase spectrum (at 5th and 95th percentiles, the gap \nincreases from 1 to 5 cylinders). Notably, 35% of PMUY consumers \npurchased no refills in their first year. Only 7% of PMUY consumers \nhave purchased 4 or more cylinders, which is the median purchase \nlevel for general consumers.\nLPG use over time and the effect of experience\nAn increase in LPG consumption with experience would conform \nto patterns of technology adoption in other sectors, but there are \nfew studies that directly measure this for cooking fuels. One recent \nIndian study does find an increase in LPG consumption with expe-\nrience, based on self-reported data23. As PMUY consumers in this \nregion have not yet completed two years since enrolment, we can-\nnot assess their use over time. However, we do have up to five years \nof data for general customers, and for these users we do not find \na discernible increase in LPG cylinder purchases over this interval \n(Fig. 5). The median refill rates and distribution of refills remain \nnearly constant over this period. There is a slight decrease in the \ntime between refills as consumers consume more, but not enough \nto result in an increase in a number of cylinders purchased per year \n(see Supplementary Fig. 7 in Supplementary Note 6).\nThe largely unchanging distribution of annual consump-\ntion for general consumers masks the varying patterns of annual \nconsumption when individual consumers are tracked (Fig. 6). \nCounter-intuitively, only a minority of consumers increased their \nconsumption (number of LPG cylinders purchased annually) in \ntheir second year of use. This suggests that policy interventions \nto encourage regular use should also target general consumers. \nRoughly 75% of consumers have purchases that stay the same or \nfluctuate by 1\u20132 cylinder a year (see Supplementary Figs. 8 and 9).\nNotably, the steady refill rate is despite the fact that macro-eco-\nnomic conditions have improved in the region during this period. \nThe net price of LPG was mostly stable (see Supplementary Fig. 10 \nin Supplementary Note 7) while both the Indian39 and Karnataka \nstate40 economies experienced >5% gross domestic product growth \nin the 2013\u20132018 period.\nWhile we are not able to directly assess multi-year consump-\ntion trends for PMUY customers, the data on general customers  \n1\n10\n100\n(Pre-national PMUY)\nGeneral\nconsumer\n(Pre-local PMUY)\nGeneral\nconsumer\n(Post-local PMUY)\nGeneral\nconsumer \nPMUY\nconsumer\nConsumer group category\nMonth-to-month growth in number of LPG consumers (%)\nFig. 1 | Monthly LPG consumer growth. Monthly PMUY and general consumer enrolment growth rates from 2016 (pre-national PMUY) to December 2018 \n(post-local PMUY) in Koppal. A logarithmic y axis is used to capture month-to-month growth rates in consumer enrolment in the form of box plots. The \nfirst three box plots (from left to right) refer to monthly average growth rates for general (non-PMUY) consumers before PMUY was launched nationally, \nbefore PMUY was launched in Koppal and after PMUY was launched in Koppal, respectively. The last box plot presents the growth rate for PMUY \nconsumers after the Koppal launch in June 2017. The range of the box represents the interquartile range (IQR) of the distribution, the horizontal line inside \nthe box is the median value and the Tukey whiskers are extended to the farthest points of the distribution that are not outliers (no more than 1.5\u2009\u00d7\u2009IQR from \nthe edge of the box). The outliers are denoted by dots.\n0\n50\n100\n150\n0\n5,000\n10,000\n15,000\nJun. 2017\nDec. 2017\nJun. 2018\nDec. 2018\nPurchase month year\nTotal daily refill sales (n) to PMUY consumers\nTotal PMUY registered consumers (n)\nDaily refill sales\nCumulative registered consumers\nSmoothed daily refill sales\nFig. 2 | PMUY consumer growth and daily refill sales. Cumulative daily \nPMUY consumer enrolment (green line) and daily refill sales across all \nthree distributors (red dots) between July 2017 and December 2018. The \ntrend in daily refill sales data is depicted by the orange line (smoothed trend \nline). PMUY consumer enrolment began in June 2017, but refill purchase \nbegan in August 2017.\nNature Energy | VOL 4 | SEPTEMBER 2019 | 806\u2013814 | www.nature.com/natureenergy\n808\n\nAnalysis\nNaTUre EnerGY\nare indicative of what the trends may look like over time for PMUY. \nGiven the very low base year consumption level that PMUY con-\nsumers are starting from, it is unclear what impact experience \nwill have on increasing their refill rates substantially in the near \nfuture. If they follow the pattern of the general consumers, however,  \nthey will probably remain at low levels of consumption absent addi-\ntional interventions.\nNotably, the first year median consumption levels have been \ndeclining (Fig. 5). Those who enrolled as LPG consumers earlier \n(for example, in 2014) used more LPG to cook during their first year \nas consumers than later enrollees (for example, 2017). This may \nmean that older consumers were either more motivated or wealthier \nthan later adopters.\nDeterminants of LPG consumption\nWe ran several linear regression models to determine what factors \nbest explain monthly normalized (per 1,000 registered consumers) \nLPG sales by distributor and what drives the observed fluctuations \nin sales patterns (see Supplementary Note 8 and Supplementary  \nFig. 11). We find that a standard deviation (s.d.) increase in distribu-\ntor (up-front) price reduces normalized monthly sales by 0.27 s.d. \nWe also see a comparable effect of seasonal factors in explaining the \nvariation in refill sales. A shift from \u2018no cropping\u2019 season to cropping \nor harvest season increases monthly refill sales by 0.22 or 0.26\u2009s.d. \nrespectively, everything else remaining equal. Refill rates in summer, \nwhen agricultural activity is limited, are ~10% lower than rates dur-\ning cropping and harvest seasons when people are busy with agricul-\ntural work. During the cropping season, people are busy preparing \nfields, planting and weeding. They place a high value on their time \nand meals tend to be quick, which is conducive to cooking with LPG. \nLater, during the harvest season, people have cash in hand, making \nLPG purchases easier33. In contrast, during the summer, people have \nmore time and less income. The climate is hot and dry, so people \nhave easy access to dry biomass as well as crop residues from after the \nLPG as rare/irregular fuel\nLPG as secondary fuel\nLPG as primary fuel\nLPG as exclusive fuel\n0\n200\n400\n600\n800\n1,000\nJan. 2016\nJul. 2016\nJan. 2017\nJul. 2017\nJan. 2018\nJul. 2018\nDec. 2018\nSales month year\nRefills per 1,000 registered consumers\nRural (General), Koppal\nUrban (General + PMUY) + Rural (General + PMUY), India\nRural (PMUY), Koppal\nFig. 3 | Monthly refill sales trends. Normalized monthly refill sales to \nPMUY and general consumers from 2016 to 2018 in rural Koppal. To \nestimate monthly sales per 1,000 consumers, refill numbers per household \nare multiplied by 1,000 (consumers) and then divided by 12 (months). \nAs installation cylinders purchased during enrolment typically take one \nmonth to consume even if LPG is used exclusively for cooking, the effective \nnumber of registered consumers is considered using a lag of one month. \nExclusive, primary, secondary and rare use are defined as 10\u201312, 5\u20139, 2\u20134 \nand 0\u20131 cylinders purchased in one year by an average rural household, \nrespectively20 (see Supplementary Note 5 and Supplementary Table 3 for \nthe derivation of the categories). All-India (General + PMUY) pre-PMUY \naverage purchase was 7.3 cylinders per household in 2015\u2013201615 (blue \ndashed line).\nUjjwala\nPMUY = 5,848 consumers; general = 7,585 consumers\nLPG cylinders purchased in first year\nGeneral\n1\n2\n3\n4\n5\n6\n7\n8\n9\n10\n11\n12\nConsumer category\nFig. 4 | LPG consumption in the first customer year. Histogram plot of the distribution of general (n\u2009=\u20097,585) and PMUY (n\u2009=\u20095,848) consumers by \nthe amount of LPG consumption (including installation cylinder) in their first year as customers. Only consumers with at least one year of customer \nexperience as of 31 December 2018 and whose full purchase history since first (installation) cylinder purchase is known are considered. The red, orange, \ngreen and blue areas represent rare, secondary, primary and exclusive LPG usage, respectively. The red square and circle represent the median and mean \nvalues of the distribution, respectively. The black square, circle, triangle and diamond indicate the 5th, 25th, 75th and 95th quartile of the distributions, \nrespectively. All LPG consumers have to purchase one cylinder in the first year during enrolment, while subsequent refill cylinder purchase varies.\nNature Energy | VOL 4 | SEPTEMBER 2019 | 806\u2013814 | www.nature.com/natureenergy\n809\n\nAnalysis\nNaTUre EnerGY\nharvest period. During the cropping season, which coincides with \nthe monsoon, wood is wet and crop residues have been consumed34. \nNotably, there are three active distributors in Koppal and the regres-\nsion results indicate that customers procuring from one distributor \npurchase fewer cylinders than those procuring LPG from the other \ntwo, other predictors remaining constant. On the basis of several \ninterviews, we found that the worst performing distributor (D3) pro-\nvides home delivery to all villages in its coverage area, while the other \ntwo do not, which appears counter-intuitive as home delivery should \nfacilitate purchases23,28. Whether the comparatively higher normalized  \nmonthly sales for D1 and D2 is due to greater sales acumen of the \ndistributors or better socio-economic/behavioural perspectives of \nthe consumers under the catchment area of these distributors cannot \nbe determined, given the limitations of the data.\nDiscussion\nWe should note the limitations of the study. We rely on mul-\ntiple datasets. Some are national in scope but have limited detail.  \n6\n6\n6.1\n6.6\n6.4\n6\n6\n6\n6\n6\n4.7\n4.6\n4.8\n4\n4\n4\n6.4\n5.9\n6.2\n6.4\n6\n6\n6\n6\n4.6\n4.2\n4\n4\n2015 enrolment (n = 1,498)\n2016 enrolment (n = 3,073)\n2013 enrolment (n = 87)\n2014 enrolment (n = 680)\n1\n2\n3\n4\n5\n1\n2\n3\n4\n5\n0\n2\n4\n6\n8\n10\n12\n0\n2\n4\n6\n8\n10\n12\nYears of LPG ownership/experience\nTotal LPG cylinders purchased annually\nExperience (years)\n1\n2\n3\n4\n5\nFig. 5 | LPG consumption distribution as a function of experience. Box plot of annual LPG cylinder sales by years of experience. Only those who enrolled \nbefore 2017 (consumers for at least two years as on 31 December 2018) and whose full purchase history since first (installation) cylinder purchase is \nknown are considered. Only when consumers have completed an entire year, that year\u2019s purchase data are considered. The mean value is indicated by a \nblack dot. The mean and median values are printed above and below the box plots respectively. The range of the box represents the IQR of the distribution, \nthe horizontal line inside the box is the median value and the Tukey whiskers are extended to the farthest points of the distribution that are not outliers (no \nmore than 1.5\u2009\u00d7\u2009IQR from the edge of the box). The outliers are denoted by dots.\n16%\n7%\n7%\n7%\n22%\n5%\n6%\n13%\n18%\n17%\n7%\n3%\n4%\n28%\n3%\n4%\n11%\n24%\n13%\n8%\n4%\n3%\n21%\n6%\n8%\n15%\n23%\n16%\n6%\n3%\n2%\n26%\n4%\n5%\n13%\n25%\n2015 enrolment (n =  1,498 )\n2016 enrolment (n =  3,073 )\n2013 enrolment (n = 87 )\n2014 enrolment (n =  680 )\n<\u22123\n\u22123\n\u22122\n\u22121\n0\n1\n2\n3\n>3\n<\u22123\n\u22123\n\u22122\n\u22121\n0\n1\n2\n3\n>3\n0%\n10%\n20%\n30%\n0%\n10%\n20%\n30%\nRelative change in annual LPG cylinder purchase\nPercentage of consumers\nRelative change in LPG purchase\nPositive\nNo Change\nNegative\nIn year 2 compared to year 1\nFig. 6 | Relative change in LPG consumption with experience. Relative change in annual LPG consumption in the second year compared to the first year \nby general consumers based on year of enrolment. Only those who enrolled before 2017 (consumers for at least two years as of 31 December 2018) \nand whose full purchase history since first (installation) cylinder purchase is known are considered. Purchase data are considered only when consumers \ncompleted an entire year as customers.\nNature Energy | VOL 4 | SEPTEMBER 2019 | 806\u2013814 | www.nature.com/natureenergy\n810\n\nAnalysis\nNaTUre EnerGY\nAnother is more detailed, covering 25,000 LPG users, but lacks \nsocio-demographic details and represents only a single district. \nThe findings drawn from the last dataset, which include com-\nparisons of refill rates between PMUY and general consumers as \nwell as seasonal variations in consumption, though instructive, \nshould not be generalized beyond the study area. In addition, we \nare looking only at trends after the initial 18 months of the pro-\ngramme. Energy transitions take time and it would be premature \nto pass judgement on PMUY at this stage37, although it is a good \ntime to take stock and suggest course corrections. Despite these \nlimitations, there are a number of key implications of this work for \npolicy and research.\nFirst, overcoming the initial capital cost barrier for PMUY \nbeneficiaries in our study site has clearly not resulted in the use \nof LPG by all consumers. This concern extends to the overall \nprogramme as 24% of PMUY beneficiaries nationally purchased \nno refills during the first year (A. Jindal, personal communica-\ntion). On the basis of their auditors\u2019 recommendations, the OMCs \nare securing more than\u20093\u2009billion rupees in preparation for loan \ndefaults. This has a high opportunity cost in a developing country \nsuch as India41. These results raise questions regarding whether \nPMUY should have been more selective in its targeting of price \ndiscounts and loans for enrolment as taxpayers are now subsidiz-\ning LPG connections for households who are not willing or able \nto buy refills. These households are therefore likely to default on \ntheir loans and remain exclusive biomass users for longer and at \na higher risk from exposure to household air pollution than those \nthat use LPG even for a portion of their cooking. However, from \na health perspective, individual household gains will come only \nwith widespread adoption as household reductions in pollution \nconcentrations and exposure are dependent on both in-house \nand community-level emissions reductions42. Despite issues in \ntargeting and refills, PMUY has clearly encouraged an unprec-\nedented number of households along the path to clean cooking. \nIt, therefore, marks a significant departure from a market-based \nBAU approach and addresses the up-front cost burden that has \nprevented the poorest households from beginning the clean cook-\ning transition process in the past at a scale to potentially have \ncommunity-level effects.\nSecond, the low refill rates (and a large fraction of non-returning \ncustomers) suggest that more effective incentives are needed for \nPMUY beneficiaries to become frequent LPG users and benefit \nfrom lower household air pollution exposure. One option would be \nto offer vouchers with seasonal discounts during summer months \nwhen demand drops. Other options might include behaviour change \ncommunications and strategies18 to \u2018nudge\u201943 poorer consumers to \nsubstitute solid fuels with LPG more often. Such policies could also \nbe beneficial in raising usage rates for non-PMUY customers. Our \ndata indicate that the general customers, while consuming more \nLPG per year than PMUY customers, do have low refill rates com-\npared to the levels needed to attain the goals of a cooking energy \ntransition. With a median refill rate of four cylinders per year for \ngeneral consumers, the government\u2019s over-arching goal of creating \n\u2018smokeless\u2019 villages cannot be attained by creating new incentives \njust for PMUY consumers.\nFinally, the use of this novel dataset provides new insights  \ninto the clean cooking energy transition, which would have been \ndifficult to reliably obtain through standard tools such as self-\nreported surveys and would have been prohibitively expensive \nto achieve via thermal sensors at this scale44. Datasets consisting \nof distributor-level transactions are invaluable sources of infor-\nmation that, properly anonymized, should be made accessible to \nresearchers and policy analysts. This not only promotes trans-\nparency but could facilitate objective programme evaluation and \npotential design changes to this programme to improve targeting \nand effectiveness.\nMethods\nNovel dataset of electronic LPG sales records. For this paper, we capitalize on \nthe OMCs\u2019 robust information technology platform, which electronically stores \nenrolment and sales records for all registered LPG consumers. We use this source \nto extract two datasets at different resolutions.\nFor the first, we use a national database of LPG purchase by PMUY \nbeneficiaries aggregated at the state level. These data were made available by a \nsenior official from the Ministry of Petroleum and Natural Gas, Government of \nIndia. It contains data on 30 million PMUY beneficiaries, aggregated by state, \nwho have completed at least 1 year as LPG consumers (based on their date of \nenrolment). For regression analysis to explain the state-wise variation in LPG \nconsumption, refer to Supplementary Note 9 and Supplementary Fig. 12. For \nfurther details of the regression analyses and validity checks of the underlying \nassumptions for linear models, see Supplementary Code 1 hosted on Figshare45.\nThe second dataset comprises sub-district level data gathered from three \nIndian Oil Corporation Ltd (IOCL) LPG distributors in Koppal district of \nKarnataka state. These were conveniently located near another ongoing project \nsite. The three distributors D1, D2 and D3 serve 25,000 domestic consumers \nfrom around 120 villages across four taluks. While the data gathered include \nboth commercial and domestic consumers, we analysed only data for domestic \nconsumers in this research.\nThis second dataset comprises micro-records for individual customers (for full \nanalysis, see Supplementary Code 2 hosted on Figshare46) based on our merging of \ntwo separate databases from within the IOCL data management system platforms \n(Spandan/Indsoft). These were accessed from the distributor\u2019s terminals with \nspecial permission. Spandan includes a log of every LPG cylinder delivery and \ncritical information such as date of purchase, type of purchase (installation or refill \ncylinder), equipment category (single cylinder/double cylinder), distributor price, \nconsumer name and consumer ID (a unique 16-digit identity number). Indosoft \ncontains the consumer\u2019s home address, enrolment date, consumer category \n(PMUY or general) and so on, which can be linked to the first database by the \nunique consumer ID. Rural Koppal is primarily a dryland region, where rain-fed \nagriculture is practised. We compare relevant key socio-economic indicators of \nrural Koppal with rural India to examine how closely the study area resembles the \nrest of rural India (see Supplementary Table 2), without making any judgement on \nwhether data from rural Koppal can in any way be treated as representative of the \nrest of rural India.\nDescription of sub-district database. Out of 30,303 (domestic, commercial and \ninstitutional) consumers in the dataset, we identify 28,075 domestic consumers \n(of which 15,071 are PMUY beneficiaries) and a corresponding 208,694 LPG \ncylinder sales records across three distributors as of 31 December 2018. For D1 and \nD3, we have data since they started operations in December 2013 and November \n2015, respectively. For D2, who started operations in 2012, data are available only \nsince January 2016. We have used the cumulative number of total consumers as \nof 1 January 2016 and the number of new enrolments as a consumer on that day \nto estimate the total number of consumers as of 31 December 2015 for D2 (the \nbaseline). Hence, the full records of LPG consumers in the study area are available \nfrom January 2016 to December 2018. As some of the customers with D3 were \ntransferred from other locations when they started, we have some consumers \nwhose \u2018full view\u2019 purchase history (all transactions from the date of enrolment as \nLPG consumer up to 31 December 2018) is unknown. We have full records for all \nPMUY consumers and 9,695 general (non-PMUY) consumers.\nThese 36 months of data are equally split into pre-PMUY (January 2016\u2013June \n2017) and post-PMUY (July 2017\u2013December 2018) periods. This is because \nPMUY was launched in Karnataka state on 17 June 2017 (D2 and D3 started \nenrolment in July 2017), while PMUY\u2019s national launch was on 1 May 2016 in \nUttar Pradesh state.\nThe individual cylinder sales transactions are formally known as \u2018delivery\u2019 logs \nin the IOCL system as LPG cylinders are supposed to be home-delivered. However, \nas is the case with many rural distributors, only D3 undertakes home delivery (in a \npartial manner) despite it being \u2018highly recommended\u2019 by OMCs. We subsequently \nrefer to these \u2018delivery\u2019 transactions as purchase or sales transactions as we \ndiscuss the results from the consumers\u2019 or distributors\u2019 perspective, respectively. \nInformation on domestic cylinder subsidies, which vary on a monthly basis and \nby the distributor, were verified manually using the OMC guidelines issued every \nmonth. D1 and D3 also charge a Re\u200910\u201315 premium (as their outlets are farther \nfrom the OMC bottling plant, so additional transportation cost is incurred), and \ntheir consumers get an additional subsidy premium of an equivalent amount, \nwhich keeps the net price uniform for all consumers, irrespective of the distributor \nthat they are associated with. We combine this information with the datasets to get \na complete picture of individual purchase transactions, including information on \nwho purchased, when at what net (and distributor) price and the number of days \nbetween transactions.\nOn the basis of the date of the transaction, we also broadly designate an \nindividual transaction using two different seasonal classifications\u2014climatic and \nagricultural. The standard seasonal classification used by India\u2019s meteorological \ndepartment is pre-monsoon (March\u2013May), monsoon (southwest monsoon)  \n(June\u2013September), autumn (post monsoon) (October\u2013December) and winter \nNature Energy | VOL 4 | SEPTEMBER 2019 | 806\u2013814 | www.nature.com/natureenergy\n811\n\nAnalysis\nNaTUre EnerGY\n(January, February)47. For Koppal, we have slightly modified this as summer \n(March\u2013May), monsoon (June\u2013August), autumn (September\u2013November) and \nwinter (December\u2013February) based on local temperature and rainfall trends. \nWe also consider three local agricultural seasons\u2014cropping (June\u2013September), \nharvest (October\u2013January) and no agricultural activity (February\u2013May)  \nbased on interviews with local stakeholders. A vast majority of the rural \npopulation is directly or indirectly dependent on agriculture or allied activity  \nfor their livelihoods.\nData cleaning and grouping. All data processing, analysis and plot generation \nwere carried out in the R platform48. Details of the code and the packages used \nare provided in Supplementary Codes 1 (National\u2014state-level variance) and 2 \n(Sub-district\u2014Koppal) hosted on Figshare45,46. The raw data collected in MS Excel \nformat from the distributor terminals were first processed by the field staff of a \npartner organization\u2014SAMUHA\u2014in accordance with the Institutional Review \nBoard guidelines of the University of British Columbia. In addition to deleting \nnames, and other personal identifiers (different from the unique 16-digit number \nprovided by OMCs), the first 10 characters of the address were scrubbed. The \nanonymized dataset was then transmitted online to the research team. We used a \nseries of codes in R to join datasets and format them for further analysis.\nWe primarily use two types of datasets created for this study from the available \ndata. First, delivery transactions of the domestic consumers in the form of a \ndata frame, which were then summarized into daily and monthly sales records \nboth by distributor and in total (see Supplementary Code 2 hosted on Figshare46 \nfor details). Second, we generate a customer database by summarizing all of the \npurchase transactions related to an individual consumer to get each consumer\u2019s \nindividual records. This includes their enrolment date, and annual purchase \nrecords of LPG cylinders, where the annual cycle begins from the date of first \ndelivery of the installation cylinder for each individual consumer. Once the \ncustomer database was created, we grouped these consumer records by category \n(PMUY and general), equipment (single or double), by years of experience as a \nconsumer (2 indicating the consumer is in the second year of LPG usage) and \nby year of enrolment. For all these groups, we also sub-grouped the data by the \ndistributor. In addition, we manually collected data on the subsidy amount that \nis credited to a consumer\u2019s bank account by each distributor for every month. We \napplied those data to the transaction database to calculate the up-front (distributor \nprice) and the net (up-front price minus subsidy) price.\nWe also categorized the data into four categories of users to indicate the \nposition/role of LPG in the cooking fuel mix of each consumer: an occasional \n(rare) use of LPG, secondary use of LPG, the primary use of LPG, and exclusive \nuse of LPG as cooking fuel. This is estimated on the basis of how much useful \ncooking energy is needed for a typical family of five members, and therefore how \nmuch LPG is needed to cater to the needs of individual consumers translated into \nmonthly cylinder purchases, on average (see Supplementary Note 5).\nAs the Spandan platform contains data for both domestic and commercial \nconsumers, for our study, we used three filters to select individual transactions \nfor domestic consumers after consultation with OMC officials and accounting for \ntypos (detailed explanation is available within the R code notes in Supplementary \nCode 2 hosted on Figshare46), an issue also raised by India\u2019s federal auditor49. The \ndata were thus also cleaned to adjust for these errors in the following manner. First, \nwe checked if there was a replication of sales transaction records with identical \nconsumer, delivery date and type of refill (installation/refill sales). We rejected \n110 transactions out of a total 214,037 transactions. Second, to select domestic \nconsumers out of the total list, which included commercial and institutional \nsubsidized consumers (schools), we followed a two-step verification process \nby focusing on two critical columns\u2014equipment and scheme associated with \na transaction. We applied a filter for the type of \u2018equipment\u2019\u201414.2\u2009kg cylinder, \nwhich is used only by domestic consumers. While the Government of India has \nintroduced 5\u2009kg cylinders in recent months14, there were no consumers in the \navailable data window who purchased these small-size cylinders. Next, we applied \na filter for four unique scheme codes (\u2018DNSC\u2019, \u2018DSC\u2019, \u2018DSCBANK\u2019, \u2018DSCDBTL\u2019) \nthat are currently or were previously associated with domestic consumers. Even \nif the codes changed for a domestic consumer\u2019s records due to policy shifts (DSC \nto DNSC when the Direct Benefit Transfer scheme was launched), or errors, if \nthey were domestic, we included them. Generally, the base equipment code (14.2 \nindicates 14.2\u2009kg LPG) and domestic codes (starting with D) are less error-prone as \nthey are linked to payment. Entries linked to payment are processed more carefully \nby the operators. We shortlisted 28,414 unique domestic consumers in this process \nfrom the selected transactions. Third, we identified that the \u2018category\u2019 column was \ncritical for analysis purposes. When the \u2018category\u2019 for consumers was missing, we \ncould not distinguish PMUY or general consumer groups for further analysis. \nHence, we dropped a further 339 consumers with \u2018null\u2019 category. Fourth, we \nremoved all consumer transactions if the consumer status was either \u2018cancelled\u2019 or \n\u2018suspended\u2019, generally for failure to provide documentation or because it was found \nfraudulent during an inspection on a later date.\nOnce clean \u2018domestic\u2019 transactions were identified, we grouped the purchase \nby individual consumers and further filtered the data to eliminate consumers who \nhad purchased more than 12 cylinders in a year. Field interviews indicate that the \nlikelihood of a domestic consumer in the study area paying for a non-subsidized \nLPG cylinder for household purpose is very low. Thus, higher (>12) usage by \ndomestic consumers could indicate diversion for commercial purpose. We also \nexcluded \u2018discontinued\u2019 consumers from the analysis (that is, those who ceased to \nbe LPG consumers).\nFor analysis, which requires data since the first purchase (that is, installation \ncylinder purchase during enrolment) to track the effect of experience, we selected \nonly consumers whose enrolment date was either available or could be reasonably \nestimated, if these enrolment dates were within our data window. For example, we \ndo not have transaction data for distributor D2 before January 2016, although it \nwas operating before this. In this case, we selected only those consumers of D2 who \nenrolled on or after 1 January 2016.\nData analysis. PMUY and LPG consumer growth in low-income rural settings. We \ncalculated the compounded monthly growth rate and month-on-month growth \nrates in consumer numbers for both general and PMUY consumers. If there \nwere differences in growth rates of general consumers in the pre and post-PMUY \nperiod, the pre-PMUY growth was considered as the BAU rate to estimate when \nthe LPG customer base would have expanded to current levels without programme \nimplementation. The difference between this and the actual numbers of enrolled \ncustomers was then taken to be a measure of the impact of the programme. We \nalso considered multiple pre-PMUY periods: before national launch, and before \nKoppal launch of PMUY. We also plotted the daily total refill sales and daily \ncumulative consumer number for July 2017 to December 2018 to visually assess the \ngap between enrolment and consumption. The smoothened daily refill sales trend \nline is created with a smoothing function using gam in the R platform. It depicts \na polynomial spline with 30 degrees of freedom and a span of 0.3 using the \u2018bs\u2019 \nfunction in the \u2018splines\u2019 package48 (see Supplementary Code 2 hosted on Figshare46 \nfor details).\nLPG use in PMUY beneficiaries versus general consumers. We compared the LPG \ndemand of PMUY consumers with general consumers in two ways. First, we \nnormalized the monthly refill sales by dividing the refills sold in a month by the \ntotal number of registered general or PMUY consumers. This shows seasonal \nvariances in demand for LPG, as well as how new enrolments and monthly \n(distributor and net) price changes impact the average refill trends. For registered \nconsumers, we use the concept of \u2018effective\u2019 registered consumers for a given \nmonth, which involved using a one-month \u2018lag\u2019 adjustment. As rural consumers \nwho enrol in a given month buy an installation (first) cylinder with the stove kit, \nthey would not be able to fully consume their installation cylinder, return the \ncylinder and then purchase a refill within the same enrolment month, irrespective \nof the season or price. As these new enrollees would not possibly buy refills in the \nsame month that they enrol, even as exclusive LPG users, including them in the \ndenominator (number of consumers) does not provide an accurate picture of the \nactual demand for LPG refills. Hence, we consider the previous month\u2019s consumer \nnumber as the denominator. This is particularly important as there are some \nmonths of very high consumer enrolments.\nSecond, we also compared the population distribution of annual LPG cylinder \npurchases of PMUY and general consumers. When examining the distribution of \nindividual consumer\u2019s purchases, we used all available (including pre-2016) data \nto capture the first year of purchase data of pre-PMUY consumers. As no PMUY \ncustomers have completed a second full year since enrolment, we compared first \nyear mean and median purchases to first year purchases of general consumers to \navoid capturing the effect of experience. We also compared the distributions for \nthese two categories of consumers across key percentiles (5th, 25th, 50th, 75th, \n95th) to better describe how much one distribution would need to shift to match \nthe other one50.\nLPG use increase over time and effect of experience. We tested for an experience \neffect by comparing the annual LPG purchase for all consumers who are in their \nnth year of use over the years 1, 2,\u2026, n (n\u2009>\u20092). As there are no PMUY consumers \nwho have completed two or more complete years, this analysis is limited to only \ngeneral consumers. We examined both changes for a cohort over time as well as \nthe initial range of LPG purchases in the first year of each cohort. In addition, we \nalso examined whether there are any differences in use over time by comparing \nthe median number of days between two consecutive LPG cylinder purchases over \ntime. This last indicator has a finer resolution (365 days versus 12 cylinders), which \nenabled us to detect the marginal effects of experience if any. Box plots were used \nto display the comparison of the distribution of consumption by experience.\nDeterminants of LPG consumption. We used linear regression models to explore \nthe effect of an increase in a number of consumers, prices (distributor price, bank \nsubsidy and net price) and season on LPG cylinder refill purchases, in addition to \nthe distributor and the number of consumers. As price changes on a monthly basis, \nwe used aggregated normalized monthly refill sales as the response variable.\nWe converted the two seasonal categorical variables and the distributor \ndetails as factors using the base \u2018is.factor\u2019 function in R48. We then ran stepwise \nregression using packages \u2018olsrr\u201951 (based on P value) and \u2018MASS\u201952 (based on the \nAkaike information criterion) for auto-selection of models. Forward selection is \na strategy that starts with no predictors in the model, iteratively adds the most \nNature Energy | VOL 4 | SEPTEMBER 2019 | 806\u2013814 | www.nature.com/natureenergy\n812\n\nAnalysis\nNaTUre EnerGY\ncontributive predictors and stops adding predictor variables when no further \nimprovements are statistically significant53. We also perform backward selection as \nwell as simultaneous forward\u2013backward selection51,52. The output models were also \nsubjected to rigorous testing to see whether the five key underlying assumptions of \na linear regression model held true. We check for a linear relationship, multivariate \nnormality, homoscedasticity, statistically insignificant auto-correlation and low \nmulticollinearity (for details, see Supplementary Note 8 and Supplementary Code \n2 hosted on Figshare46). We also reported scaled coefficients (\u2018beta\u2019) to avoid \ndismissing an effect as \u2018small\u2019 when it is just the units of measure that are small, \nespecially when we have very different units such as a number of consumers and \nprice included in the model.\nIn addition, we have also tested whether there is an effect of a loan deferment \nscheme for PMUY consumers that incentivizes LPG purchase (Supplementary \nNote 10).\nLimitations. Unlike surveys, which routinely collect demographic and \nsocioeconomic data at the individual and household level, this dataset has limited \ninformation on these aspects. Moreover, while the implicit assumption is that \npurchase by domestic consumers equals consumption for residential cooking \nneeds, this may not always hold true. LPG for domestic consumers is subsidized \n(for the first 12 cylinders in a year), but may sometimes be diverted to commercial \nconsumers who avoid paying for non-subsidized commercial cylinders54. While \nreforms preceding the launch of PMUY have reduced this \u2018black-marketing\u2019, field \nobservation suggests that subsidized cylinders are still used in roadside restaurants \nand tea shops in villages. Moreover, the purchase of a cylinder does not necessarily \nimply that the last purchased cylinder has been entirely consumed within the data \nwindow for consumption analysis.\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nThe authors do not have ownership over the LPG sales data downloaded from the \nIndian Oil Corporation system. Researchers who seek access to the raw data to \nreplicate this study or to perform other analyses should contact the Ministry of \nPetroleum and Natural Gas and/or IOCL.\nCode availability\nAll the code and plots that were generated for this research, irrespective of whether \nthey were used in the paper, are available via Figshare at https://doi.org/10.6084/\nm9.figshare.7961069 (Supplementary Code 1) and https://doi.org/10.6084/\nm9.figshare.7961066 (Supplementary Code 2). For state-wise data analysis for \nPMUY consumer use, refer to Supplementary Code 145. For Koppal sub-district \nLPG sales data-based analysis, refer to Supplementary Code 246.\nReceived: 9 January 2019; Accepted: 5 June 2019;  \nPublished online: 15 July 2019\nReferences\n\t1.\t Smith, K. R. in Making of New India: Transformation Under Modi \nGovernment (eds Debroy, B., Ganguly, A. & Desai, K.) 401\u2013410 (Dr\u00a0Syama \nPrasad Mookerjee Research Foundation and Wisdom Tree, 2018).\n\t2.\t Anenberg, S. C. et\u00a0al. Cleaner cooking solutions to achieve health, climate, \nand economic cobenefits. Environ. Sci. Technol. 47, 3944\u20133952 (2013).\n\t3.\t International Energy Agency, International Renewable Energy Agency, United \nNations Statistics Division, World Bank & World Health Organization. \nTracking SDG7- The Energy Progress Report (International Bank for \nReconstruction and Development/The World Bank, 2019).\n\t4.\t Global Burden of Disease (GBD) Study 2017 (IHME, 2018).\n\t5.\t Household Air Pollution and Noncommunicable Disease: Summary for Policy \nMakers (Health Effects Institute, 2018).\n\t6.\t Bailis, R., Drigo, R., Ghilardi, A. & Masera, O. The carbon footprint of \ntraditional woodfuels. Nat. Clim. Change 5, 266\u2013272 (2015).\n\t7.\t Dutta, S., Kooijman, A. & Cecelski, E. Energy Access and Gender: Getting the \nRight Balance (The World Bank, 2017).\n\t8.\t Putti, V. R., Tsan, M., Mehta, S. & Kammila, S. The State of the Global Clean \nand Improved Cooking Sector (Energy Sector Management Assistance \nProgram, Global Alliance for Clean Cookstoves, The World Bank, 2015).\n\t9.\t Barua, S. K. & Agarwalla, S. K. Lighting up Lives through Cooking Gas and \nTransforming Society (Indian Institute of Management Ahmedabad, 2018).\n\t10.\tAnnual Report 2016\u20132017 (Ministry of Petroleum and Natural Gas, \nGovernment of India, 2017).\n\t11.\tShri Dharmendra Pradhan hands over 7Croreth LPG connection under \nPMUY; Milestone reached in just 34 months. Press Information Bureau  \n(8 March 2019); pib.nic.in/Pressreleaseshare.aspx?PRID=1568341\n\t12.\tGould, C. F. & Urpelainen, J. LPG as a clean cooking fuel: adoption, use, and \nimpact in rural india. Energy Policy 122, 395\u2013408 (2018).\n\t13.\tWorld Energy Outlook 2018 (International Energy Agency, 2018).\n\t14.\tJindal, A. in The Ujjwala Saga - Unending Happiness & Health 60\u201365 \n(Ministry of Petroleum and Natural Gas, Government of India, 2019).\n\t15.\tDabadge, A., Sreenivas, A. & Josey, A. What has the Pradhan Mantri Ujjwala \nYojana achieved so far? Econ. Political Wkly 53, 69\u201375 (2018).\n\t16.\tJohnson, M. A. & Chiang, R. A. Quantitative guidance for stove usage and \nperformance to achieve health and environmental targets. Environ. Health \nPerspect. 123, 820\u2013826 (2015).\n\t17.\tShankar, A. et\u00a0al. Maximizing the benefits of improved cookstoves: moving \nfrom acquisition to correct and consistent use. Glob. Health Sci. Pract. 2, \n268\u2013274 (2014).\n\t18.\tKar, A. & Zerriffi, H. From cookstove acquisition to cooking transition: \nframing the behavioural aspects of cookstove interventions. Energy Res. Soc. \nSci. 42, 22\u201333 (2018).\n\t19.\tRuiz-Mercado, I., Masera, O., Zamora, H. & Smith, K. R. Adoption  \nand sustained use of improved cookstoves. Energy Policy 39,  \n7557\u20137566 (2011).\n\t20.\tKar, A., Singh, D., Pachauri, S., Bailis, R. & Zerriffi, H. in The Ujjwala Saga \n- Unending Happiness & Health 16\u201321 (Ministry of Petroleum and Natural \nGas, Government of India, 2019).\n\t21.\tSmith, K. R. & Dutta, K. \u201cCooking with Gas\u201d. Energy Sustain. Dev. 15, \n115\u2013116 (2011).\n\t22.\tAssessment Report: Primary Survey on Household Cooking Fuel Usage and \nWillingness to Convert to LPG (CRISIL, 2016).\n\t23.\tJain, A. et\u00a0al. Access to Clean Cooking Energy and Electricity: Survey of States \n2018 (Council on Energy, Environment and Water, 2018).\n\t24.\t573 villages in Manipur identified for conversion to \u2018Smokeless Villages\u2019. The \nAssam Tribune http://www.assamtribune.com/scripts/detailsnew.\nasp?id=apr1518/oth051 (2018).\n\t25.\tLPG Connections under PM Ujjwala Yojana is Providing Smoke Free Kitchens \nto Individuals (MyGov India, 2018).\n\t26.\tPradhan Mantri Ujjwala Yojana: A Giant Step Towards Better Life For All. \nPress Information Bureau http://pib.nic.in/newsite/printrelease.\naspx?relid=148971 (2016). \n\t27.\tRuiz-Mercado, I. & Masera, O. Patterns of stove use in the context of \nfuel\u2013device stacking: rationale and implications. EcoHealth 12,  \n42\u201356 (2015).\n\t28.\tGiri, A. & Aadil, A. Pradhan Mantri Ujjwala Yojana: A demand-side \ndiagnostic study of LPG refills (Policy Brief) (MicroSave, 2018).\n\t29.\tLewis, J. J. & Pattanayak, S. K. Who adopts improved fuels and cookstoves? A \nsystematic review. Environ. Health Perspect. 120, 637\u2013645 (2012).\n\t30.\tMiller, G. & Mobarak, A. M. Gender Differences in Preferences, Intra-\nhousehold Externalities, and Low Demand for Improved Cookstoves (National \nBureau of Economic Research, 2013).\n\t31.\tRogers, E. M. Diffusion of Innovations (Free Press, 2003).\n\t32.\tVenkatesh, V., Thong, J. Y. & Xu, X. Consumer acceptance and use of \ninformation technology: extending the unified theory of acceptance and use \nof technology. Manag. Inf. Syst. Q. 36, 157\u2013178 (2012).\n\t33.\tSingh, S. K. The Kaleidoscope of Cooking (Deutsche Gesellschaft f\u00fcr \nInternational Zusammenarbeit, 2014).\n\t34.\tRajakutty, S. & Kojima, M. Promoting Clean Household Fuels Among the Rural \nPoor: Evaluation of the Deepam Scheme in Andhra Pradesh 65 (The World \nBank, 2002).\n\t35.\tKhandelwal, M. et\u00a0al. Why have improved cook-stove initiatives in india \nfailed? World Dev. 92, 13\u201327 (2017).\n\t36.\tStanistreet, D. et\u00a0al. The role of mixed methods in improved cookstove \nresearch. J. Health Commun. 20, 84\u201393 (2015).\n\t37.\tSmith, K. R. & Jain, A. in Energizing India: Fuelling a Billion Lives (ed Mitra, \nS.) 48\u201372 (Rupa, 2019).\n\t38.\tAhmad, N., Sharma, S. & Singh, D. A. K. Pradhan Mantri Ujjwala Yojana \n(PMUY) step towards social inclusion in India. Int. J. Trend Res. Dev. 5,  \n4 (2018).\n\t39.\tAnand, N. India\u2019s GDP growth slows down to pre-Narendra Modi days. \nQuartz India https://qz.com/india/1294390/achhe-din-indias-gdp-growth-is-\nback-to-pre-narendra-modi-days/ (2018).\n\t40.\tStates of Growth 2.0: The Scorecard, and the Workout on How Each State Got \nto Where It Has (CRISIL, 2019).\n\t41.\tKumar, S. OMCs make provisions for PMUY losses. The Financial Express \nhttps://www.financialexpress.com/industry/omcs-make-provisions-for-pmuy-\nlosses/1264022/ (2018).\n\t42.\tWHO Guidelines for Indoor Air Quality: Household Fuel Combustion (World \nHealth Organization, 2014).\n\t43.\tRaihani, N. J. Nudge politics: efficacy and ethics. Front. Psychol. 4, 972 (2013).\n\t44.\tGraham, E. A. et\u00a0al. Laboratory demonstration and field verification of a \nwireless cookstove sensing system (WiCS) for determining cooking duration \nand fuel consumption. Energy Sustain. Dev. 23, 59\u201367 (2014).\n\t45.\tKar, A., Pachauri, S., Bailis, R. & Zerriffi, H. Using sales data to assess \ncooking gas adoption and the impact of India\u2019s Ujjwala programme in rural \nKarnataka: Supplementary Code 1 (National) Figshare https://doi.org/10.6084/\nm9.figshare.7961069 (2019).\nNature Energy | VOL 4 | SEPTEMBER 2019 | 806\u2013814 | www.nature.com/natureenergy\n813\n\nAnalysis\nNaTUre EnerGY\n\t46.\tKar, A., Pachauri, S., Bailis, R. & Zerriffi, H. Using sales data to assess \ncooking gas adoption and the impact of India\u2019s Ujjwala programme in rural \nKarnataka: Supplementary Code 2 (Sub-district) Figshare https://doi.\norg/10.6084/m9.figshare.7961066 (2019).\n\t47.\tFrequently Asked Questions (India Meteorological Department, 2018).\n\t48.\tR Core Team R: A Language and Environment for Statistical Computing  \n(R Foundation for Statistical Computing, 2018).\n\t49.\tReport of the Comptroller and Auditor General of India on Implementation of \nPAHAL (DBTL) Scheme (Pratyaksh Hanstantrit Labh Yojana) (Comptroller \nand Auditor General of India, 2016).\n\t50.\tRousselet, G. A., Pernet, C. R. & Wilcox, R. R. Beyond differences in  \nmeans: robust graphical methods to compare two groups in neuroscience. \nEur. J. Neurosci. 46, 1738\u20131748 (2017).\n\t51.\tolsrr: Tools for Building OLS Regression Models v.0.5.2 (Aravind Hebbali, \n2018).\n\t52.\tVenables, W. N. & Ripley, B. D. Modern Applied Statistics with S. 4th edn \n(Springer, 2002).\n\t53.\tKassambara, A. Stepwise Regression Essentials in R (STHDA, 2018);  \nhttp://www.sthda.com/english/articles/37-model-selection-essentials-in-r/154-\nstepwise-regression-essentials-in-r/\n\t54.\tBarnwal, P. Curbing Leakage in Public Programs with Direct Benefit Transfers \n(Michigan State University, 2016).\nAcknowledgements\nThis article was developed under Assistance Agreement No. 83542102 awarded by the \nUS Environmental Protection Agency (EPA) to R.B. (sub-award to H.Z.). The EPA has \nnot formally reviewed it. The views expressed in this document are solely those of the \nauthors and do not necessarily reflect those of the Agency. EPA does not endorse any \nproducts or commercial services mentioned in this publication. Part of the research \nwas developed by A.K. during his time in the Young Scientists Summer Program at the \nInternational Institute for Applied Systems Analysis, Laxenburg, Austria, with financial \nsupport from the German National Member Organization. A.K. and H.Z. acknowledge \nsupport from the Wall Solutions Initiative provided by the Peter Wall Institute for \nAdvanced Studies, the \u2018Collaborative Research and Training Experience- Atmospheric \nAerosol Program\u2019 (CREATE-AAP) at the University of British Columbia and the Clean \nCooking Alliance (United Nations Foundation). Access to data and interviews with \nofficials from the Ministry of Petroleum and Natural Gas, Government of India, IOCL, \nHindustan Petroleum Corporation Ltd and Bharat Petroleum Corporation Ltd was \nindispensable to this study. The authors are also indebted to the staff of the partner non-\ngovernmental organization SAMUHA for assistance with data collection. The authors \nalso thank M. Brauer and S. Mehta for their feedback on the early drafts of the paper.\nAuthor contributions\nA.K., S.P. and H.Z. designed the study. A.K. performed the data cleaning and coding. \nA.K., S.P., R.B. and H.Z. contributed to the writing of the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-019-0429-8.\nReprints and permissions information is available at www.nature.com/reprints.\nCorrespondence and requests for materials should be addressed to A.K.\nPublisher\u2019s note: Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2019\nNature Energy | VOL 4 | SEPTEMBER 2019 | 806\u2013814 | www.nature.com/natureenergy\n814\n\n1\nnature research  |  reporting summary\nOctober 2018\nCorresponding author(s):\nAbhishek Kar\nLast updated by author(s): Jun 4, 2019\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nThe sub-district level 'delivery' (purchase, as most homes did not get home delivery) transactions for 25,000 consumers were \ndownloaded from the three distributor terminals of the 'Centralized Ind-Soft' system of the Indian Oil Corporation Ltd. with special \npermission. The national level data for 30 million consumers at state-wise aggregated level was provided by the Ministry of Petroleum \nand Natural Gas, Govt. of India via email. \nData analysis\nWe used open source R platform and open source R packages as detailed in the code (Supplementary Codes 1 &2 ). For the maps, we use \nopen source QGis\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. \nWe strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nFor the purpose of full transparency, we are making available all the codes written and the plots that were generated for this database, irrespective of whether it \nwas used in the paper in Supplementary Codes 1,and 2. The data that support the findings of this study are available from [MoPNG and IOCL] but restrictions apply \nto the availability of these data, which were used under special permission from Government of India for the current study, and so are not publicly available. \nResearchers who seek the raw data to replicate this study or to do other analyses should contact MoPNG and/or IOCL. [Note to reviewers: The data and all codes \nassociated with the data analysis are available during the period of review for limited purpose- to cross-check the underlying data analysis and suggest additional \nanalysis to improve the quality of the paper. \n\n2\nnature research  |  reporting summary\nOctober 2018\nhttps://cp.sync.com/dl/61f786390/ujsuvkq2-xsrjzsi7-72a6gzbx-d5az72jq \nThe data available to the reviewers should not be disseminated further and should be deleted once the review is complete.]\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThere are two datasets- one national level and one sub-district level. The data analysis is primarily quantitative, however, the authors \nhave included in the discussion section some inputs received from stakeholders. \nResearch sample\nThere are two sets of research samples to analyze the PMUY program. First, for the national level data, all consumers who completed \none year as the project (PMUY) beneficiary. Second, for the sub-district level data, we chose three distributors of IOCL for convenience \npurpose (detailed in Methods), and then included all domestic consumers (both PMUY and general) under their catchment area. All these \nconsumers resided in rural areas. The PMUY beneficiaries were all women and the general (non-PMUY) consumers were a mix of male \nand female. \nSampling strategy\nThe distributors were selected based on convenience sampling. But there was no other sampling strategy as we included 100% of all \ndomestic consumers. \nData collection\nThe sub-district data was collected on monthly/ bi-monthly basis  by field staff from the partner NGO- SAMUHA- from the IOCL \ncentralized data management system (distributor terminal) in .csv format. The data was then manually cleaned by the IT dept. of \nSAMUHA as per standard written protocol to erase all names and other personal identifying information such as bank account/ ID card \nnumbers. In case of address, the first 10 characters are removed. Only the sixteen digit unique IOCL consumer code has been retained as \nprimary key for join previous batches of data with future batches. The cleaned database with only IOCL consumer ID were then emailed \nto the first author for analysis as per University's Institutional Review Board guidelines.\nTiming\nThe data collection from the IOCL software (distributor terminal) happened in batches throughout 2017 and 2019 (up to January second \nweek). So, we have all purchase transaction data till 10 October 2018. The MoPNG data on national level state-wise PMUY refills was \nshared with us via email on 4th December 2018.\nData exclusions\nFor the national level data, all consumers who completed one year as the project (PMUY) beneficiary were included. We collected data \nfrom distributor, then sub-selected all consumers who are domestic households (use 14.2 kg cylinder and have bank account linked for \nsubsidy transfer).  We made some assumptions in the exclusion process (clearly noted in the R codes) primarily because the database is \nprone to manual data entry error (see, Indian federal auditor- CAG report discussed and referenced in the Methods). Also, for some \nanalysis, we only considered consumer data for the first year, to avoid experience effect, if any.\nNon-participation\nNot applicable\nRandomization\nFor the national level data, all consumers who completed one year as the project (PMUY) beneficiary were included. For the sub-district \nlevel data, we chose three distributors of IOCL for convenience purpose (detailed in Methods), and then included all domestic consumers \nunder their catchment area.   \nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology\nAnimals and other organisms\nHuman research participants\nClinical data\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\n\n\n    Scientific Research Findings:", "answer": "We estimate that an average rural family needs to purchase five 14.2 kg cylinders annually to meet half of their cooking needs. We find that just 7% of PMUY beneficiaries in Koppal district in Karnataka, India, purchased five or more cylinders annually, suggesting that the beneficiaries seldom use LPG. The general (nonPMUY) consumers in this region use on average two times more LPG cylinders than PMUY beneficiaries. Yet, only 45% of general consumers use five or more cylinders per year, even after several years of experience with LPG. We also find that LPG consumers are sensitive to price and seasonality: LPG cylinder refill rates are lower in the summer when agricultural activity is limited and cash is scarce. These findings suggest the need for additional measures to promote regular LPG use for all rural populations. Although the findings come from a single district in Southern India, they may also apply to other areas with similar socio-economic conditions. A more expansive evaluation of PMUY would help inform the design of targeted incentives to transform infrequent users to regular users.", "id": 27}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Analysis\nhttps://doi.org/10.1038/s41560-019-0430-2\n1Department of Urban Planning, University of California, Los Angeles, CA, USA. 2Institute of the Environment and Sustainability, University of California, \nLos Angeles, CA, USA. *e-mail: rdeepak@ioes.ucla.edu\nA\nt the beginning of the new millennium, energy insecurity, \nglobal climate change and stagnant rural economies led to \npolicies supporting domestic biofuels as a renewable alter-\nnative fuel in more than 60 countries worldwide1. As a consequence, \nglobal production of ethanol and biodiesel combined almost qua-\ndrupled (from about 35 billion litres to 135 billion litres) in the short \nspan from 2005 to 2016 (ref. 2). However, these policies had two \nmajor flaws. First, appropriation of edible crops for biofuel (mainly \ncorn and sugarcane for ethanol, and soybean, canola and palm for \nbiodiesel) was an important factor responsible for food price infla-\ntion alongside other factors such as rising income that drove rapid \ngrowth in food demand (especially meat demand), rising energy \nprices, adverse weather shocks, currency fluctuations and trade \npolicies1,3\u20136, the consequences of which were particularly severe for \npoorer households in developing countries7. Second, these crops \nrequired intensive use of land, water, nitrogen and other farm \nchemicals, which meant low, and in the worst case uncertain, net \nenvironmental benefits8\u201312.\nBeing widely available and replenishable, wastes and biomass \nresidues from agricultural, dairy, forestry and household activi-\nties seem to contain the basic attributes of a sustainable energy \nresource, in stark contrast to bioenergy from food crops13\u201315. The US \nDepartment of Energy 2016 Billion-Ton Study estimates an annual \navailability of 233 million tonnes (Mt) of dry waste16. To put this \nin perspective, the approximately 60 billion litres of corn ethanol \nproduced in the United States in 2017 required about 150\u2009Mt of corn \n(assuming a yield of 402\u2009l ethanol per Mt). Furthermore, wastes and \nbiomass residues can be used to derive a number of alternative \nenergy products, including electricity along with heat, biomethane \n(or renewable natural gas), ethanol, renewable diesel or bio jet fuel, \neach through various conversion pathways, which currently are at \ndifferent stages of technical and economic maturity14,15,17\u201321. Beyond \nenergy production and mitigation of climate change, efficient use \nof wastes and residues is integral to the achievement of sustain-\nable development22, and to redesigning our economies to minimize \nmaterial and energy throughput, that is, towards becoming a circu-\nlar economy23,24. However, at the same time, sustainable use of this \nresource hinges on overcoming some challenges. The collection, \ntransport and storage of biomass feedstocks are costly and could \naccount for over 50% of total cost in the supply chain of bioenergy \nproducts25. The composition of wastes also varies from one location \nto another, and their processing requires substantial energy inputs. \nIn addition, national-scale policies tend to ignore local trade-offs, \nleading to suboptimal use of scarce resources26. Harnessing the full \nenergetic and environmental potential of this resource, therefore, \nrequires a holistic assessment of alternative competing pathways \nto their utilization taking into account the spatial distribution of \neach specific type of waste and the local conditions under which \nthe wastes will be processed. The majority of previous life-cycle \nassessment (LCA) studies have focused on either a smaller number \nof waste types14,27\u201335, certain types of bioenergy product15,19,20,36\u201338 or \ncertain conversion technologies29,31,37,39\u201345. Comparing the effective-\nness and environmental impacts of all feasible conversion pathways \nfor all types of waste from a systems perspective is necessary for \npolicies that address the best use of wastes and biomass residues.\nHere, we determine the net energy gain and the global warming \npotential (GWP) of energy recovery from waste; which pathways \nsimultaneously maximize renewable energy production, net energy \ngain and climate benefits for each type of waste and how this var-\nies given the spatial distribution of their availability (specifically, in \nthe contiguous United States); and what are the aggregate energy \nand climate benefits when all available wastes and biomass resi-\ndues across the contiguous United States are dedicated for a specific \npolicy objective such as maximizing renewable energy production, \nmaximizing net energy gain or maximizing climate benefits. These \nquestions are aimed at deriving both general insights on the optimal \nuse of wastes and biomass residues, and also illustrating their overall \nclimate-change mitigation potential in the context of a large country, \nspecifically the United States. To this end, we quantify life-cycle GHG \nemissions and net energy gain for 15 conversion pathways (detailed \nLife-cycle energy and climate benefits of energy \nrecovery from wastes and biomass residues in the \nUnited States\nBo Liu1 and Deepak Rajagopal\u200a \u200a2*\nAgricultural and forestry residues, animal manure and municipal solid waste are replenishable and widely available. However, \nharnessing these heterogeneous and diffuse resources for energy requires a holistic assessment of alternative conversion path-\nways, taking into account spatial factors. Here, we analyse, from a life-cycle assessment perspective, the potential renewable \nenergy production, net energy gain and greenhouse gas (GHG) emission reduction for each distinct type of waste feedstock \nunder different conversion technology pathways. The utilization of all available wastes and residues in the contiguous United \nStates can generate 3.1\u20133.8 exajoules (EJ) of renewable energy, but only deliver 2.4\u20133.2\u2009EJ of net energy gain, and displace \n103\u2013178 million tonnes of CO2-equivalent GHG emissions. For any given waste feedstock, looking across all US counties where \nit is available, except in rare instances, no single conversion pathway simultaneously maximizes renewable energy production, \nnet energy gain and GHG mitigation.\nNature Energy | VOL 4 | AUGUST 2019 | 700\u2013708 | www.nature.com/natureenergy\n700\n\nAnalysis\nNature Energy\nTable 1 | Description and attributes of conversion pathways\nConversion \npathway\nLabel\nDescription\nFeedstock feasibility\nEnergy input\nEnergy output (main)\nEnergy output \n(coproducts)\nDisplaced \nproducts\nRefs.\nCHP\nE1\nThermal combustion through biomass CHP plants\nAll\nElectricity, heat, diesel\nElectricity\nHeat\nState power grids, \nnatural gas-based \nheat\n15, 39,40\nGasification + \nCHP\nE2\nSyngas is produced through gasification and is then \ncombusted in gas engines to produce electricity and heat\nAll\nElectricity, heat\nElectricity\nHeat\nState power grids, \nnatural-gas-based \nheat\n15, 33,61\nIntegrated \ngasification \ncombined cycle\nE3\nElectricity generation through combined gas and steam \nturbines with no heat recovery\nAll\nElectricity, heat\nElectricity\nN/A\nState power grids\n15, 33,50\nAnaerobic \ndigestion + CHP\nE4\nBiogas is produced through anaerobic digestion and is \nthen combusted in gas engines to produce electricity \nand heat\nAnimal manure, MSW \n(except plastics, rubber \nand leather, and textiles)\nNatural gas, diesel\nElectricity\nHeat\nState power grids, \nnatural-gas-based \nheat\n28, 29, 40, 43,50\nGasification\nM1\nSyngas is produced through gasification and is then \nupgraded and purified to produce methane.\nAll\nElectricity, heat\nMethane\nN/A\nNatural gas\n15,61\nAnaerobic \ndigestion\nM2\nBiogas is produced through anaerobic digestion and is \nthen upgraded and compressed for pipeline transmission\nAnimal manure, MSW \n(except plastics, rubber \nand leather, and textiles)\nElectricity, heat, diesel\nMethane\nN/A\nNatural gas\n28, 40, 43,50\nEnzymatic \nhydrolysis + \nfermentation\nEth1\nEthanol production through pretreatment, enzymatic \nhydrolysis and fermentation\nAg. and forest residues, \nconstruction and \ndemolition (CD) waste, \nMSW wood, paper, yard \ntrimmings\nNatural gas, diesel\nEthanol\nElectricity\nPetroleum-based \ngasoline, state \npower grids\n36, 50,62\nGasification \nFischer\u2013Tropsch \n(FT) synthesis\nRd1\nGasification to decompose biomass into syngas, and \nFT synthesis to convert syngas into liquid fuels with the \npresence of catalysts; excess steam is used for electricity \ngeneration\nAg. and forest residues, \nCD waste, MSW wood, \npaper, plastics, yard \ntrimmings\nElectricity\nRenewable diesel\nRenewable gasoline, bio jet \nfuel, methane, electricity\nPetroleum-based \ndiesel, gasoline \nand jet fuel, \nnatural gas, state \npower grids\n20, 34, 44,50\nPyrolysis + \nhydroprocessing\nRd2\nThermochemical conversion of a feedstock into bio-oil, \nbio-char and pyrolysis gas; integrated with hydrocracking \nand hydrotreatment processes for liquid fuel production\nAg. and forest residues, \nCD waste, food waste, \nMSW wood, paper, \nplastics, yard trimmings\nElectricity, natural gas\nRenewable diesel\nRenewable gasoline\nPetroleum-\nbased diesel and \ngasoline\n35,37\nAlcohol-to-jet \n(ethanol)\nBj1\nBio jet fuel production with ethanol as the intermediate \nproduct\nAg. and forest residues, \nCD waste, MSW wood, \npaper, yard trimmings\nHydrogen, electricity\nBio jet fuel\nRenewable diesel, \nrenewable gasoline\nPetroleum-based \ndiesel, gasoline \nand jet fuel\n50\nSugar-to-jet \n(fermentation)\nBj2\nSugar is separated from waste feedstock and is \nthen converted into hydrocarbon or hydrocarbon \nintermediates through fermentation\nAg. and forest residues, \nCD waste, MSW wood, \npaper, yard trimmings\nHydrogen\nBio jet fuel\nN/A\nPetroleum-based \njet fuel\n50\nPyrolysis in\u00a0situ\nBj3\nFeedstock is dried, ground and then converted to a \nmixture of bio-oil, gas and char at high temperature \n(above 500\u2009\u00b0C). The conversion is continued by \nhydrodeoxygenating the bio-oil with hydrogen, which is \nproduced through steam methane reforming (SMR) of \nprocess off-gases\nForest residues, CD waste, \nMSW wood, paper, yard \ntrimmings\nElectricity\nBio jet fuel\nRenewable diesel, \nrenewable gasoline\nPetroleum-based \ndiesel, gasoline \nand jet fuel\n19, 20,45\nPyrolysis ex\u00a0situ\nBj4\nSame process as Bj3 except that hydrogen is produced \nfrom SMR of natural gas\nForest residues, CD waste, \nMSW wood, paper, yard \ntrimmings\nHydrogen\nBio jet fuel\nRenewable diesel, \nrenewable gasoline\nPetroleum-based \ndiesel, gasoline \nand jet fuel\n19, 20,45\nHTL in\u00a0situ\nBj5\nWet feedstock is converted into biocrude at a \ntemperature of 250\u2013550\u2009\u00b0C (with water as a medium), \nand is then hydrodeoxygenated with hydrogen, which \nis produced through SMR of process off-gases and also \nanaerobic digestion of wastewater\nForest residues, CD waste, \nMSW wood, paper, yard \ntrimmings\nElectricity\nBio jet fuel\nRenewable diesel, \nrenewable gasoline\nPetroleum-based \ndiesel, gasoline \nand jet fuel\n19, 20,45\nHTL ex\u00a0situ\nBj6\nSame process as Bj5 except that hydrogen is produced \nfrom SMR of natural gas\nForest residues, CD waste, \nMSW wood, paper, yard \ntrimmings\nElectricity, hydrogen\nBio jet fuel\nRenewable diesel, renewable \ngasoline\nPetroleum-based \ndiesel, gasoline and \njet fuel\n19, 20,45\nNature Energy | VOL 4 | AUGUST 2019 | 700\u2013708 | www.nature.com/natureenergy\n701\n\nAnalysis\nNature Energy\ndescription in Table 1) and 29 waste feedstocks with spatially explicit \nestimates of waste potential for the United States. We find that the \nsource of electricity consumed during processing and the environ-\nmental footprint of the displaced products are key in determining \nthe best use of wastes and biomass residues. The utilization of all \navailable wastes and residues in the contiguous United States can \ngenerate 3.1\u20133.8\u2009EJ of renewable energy, but deliver only 2.4\u20133.2\u2009EJ \nof net energy gain, and displace 103\u2013178\u2009Mt of CO2-equivalent \n(MtCO2e) GHG emissions. For any given waste feedstock, looking \nacross all US counties where it is available, except in rare instances, \nno single conversion pathway simultaneously maximizes renewable \nenergy production, net energy gain and GHG mitigation.\nAg. residues\nAnimal manure\nForest residues\nMSW\nE1\nE2\nE3\nM1\nEth1\nRd1\nBj1\nBj2\nE1\nE2\nE3\nE4\nM1\nM2\nE1\nE2\nE3\nM1\nEth1\nRd1\nRd2\nBj1\nBj2\nBj3\nBj4\nBj5\nBj6\nE1\nE2\nE3\nE4\nM1\nM2\nEth1\nRd1\nRd2\nBj1\nBj2\nBj3\nBj4\nBj5\nBj6\n\u22125\n0\n5\n10\n15\nGJ per Mg ww\nEnergy production/main product\nEnergy production/coproduct(s)\nNet energy\nAg. residues\nAnimal manure\nForest residues\nMSW\nE1\nE2\nE3\nM1\nEth1\nRd1\nBj1\nBj2\nE1\nE2\nE3\nE4\nM1\nM2\nE1\nE2\nE3\nM1\nEth1\nRd1\nRd2\nBj1\nBj2\nBj3\nBj4\nBj5\nBj6\nE1\nE2\nE3\nE4\nM1\nM2\nEth1\nRd1\nRd2\nBj1\nBj2\nBj3\nBj4\nBj5\nBj6\n0\n5\n10\n15\n20\nAg. residues\nAnimal manure\nForest residues\nMSW\nE1\nE2\nE3\nM1\nEth1\nRd1\nBj1\nBj2\nBAU1\nE1\nE2\nE3\nE4\nM1\nM2\nBAU2\nE1\nE2\nE3\nM1\nEth1\nRd1\nRd2\nBj1\nBj2\nBj3\nBj4\nBj5\nBj6\nBAU3\nE1\nE2\nE3\nE4\nM1\nM2\nEth1\nRd1\nRd2\nBj1\nBj2\nBj3\nBj4\nBj5\nBj6\nBAU4\nBAU5\n\u22121\n0\n1\n2\ntCO2e per Mg ww\nCollection\nTransport to facility\nProcessing\nTransmission & distribution\nEnd use\nDisplacement/main product\nDisplacement/coproduct(s)\nNet GWP\nCurrent management practices\nAg. residues\nAnimal manure\nForest residues\nMSW\nE1\nE2\nE3\nM1\nEth1\nRd1\nBj1\nBj2\nBAU1\nE1\nE2\nE3\nE4\nM1\nM2\nBAU2\nE1\nE2\nE3\nM1\nEth1\nRd1\nRd2\nBj1\nBj2\nBj3\nBj4\nBj5\nBj6\nBAU3\nE1\nE2\nE3\nE4\nM1\nM2\nEth1\nRd1\nRd2\nBj1\nBj2\nBj3\nBj4\nBj5\nBj6\nBAU4\nBAU5\n\u22121\n0\n1\n2\ntCO2e per Mg ww\nCollection\nTransport to facility\nProcessing\nTransmission & distribution\nEnd use\nBiogenic CO2\nDisplacement/main product\nDisplacement/coproduct(s)\nNet GWP\nCurrent management practices\na\nb\nc\nd\nFig. 1 | Energy, net energy and emissions from waste biomass utilization in the United States. a, Energy production and net energy by waste type and \nconversion pathway. b, Energy return on investment by waste type and conversion pathway (the horizontal line refers to an energy return on investment of \n1). c, Life-cycle emissions when biogenic CO2 is excluded. d, Life-cycle emissions when biogenic CO2 is included. ww, wet weight. Electricity pathways: E1\u2014\nCHP, E2\u2014gasification + CHP, E3\u2014integrated gasification combined cycle, E4\u2014anaerobic digestion + CHP. Methane pathways: M1\u2014gasification, M2\u2014\nanaerobic digestion. Ethanol pathway: Eth1\u2014enzymatic hydrolysis + fermentation. Renewable-diesel pathways: Rd1\u2014gasification + FT synthesis, Rd2\u2014\npyrolysis + hydroprocessing. Bio jet fuel pathways: Bj1\u2014alcohol-to-jet (ethanol), Bj2\u2014sugar-to-jet (fermentation), Bj3\u2014pyrolysis (in\u00a0situ), Bj4\u2014pyrolysis \n(ex\u00a0situ), Bj5\u2014HTL (in\u00a0situ), Bj6\u2014HTL (ex\u00a0situ). Business-as-usual (BAU) practices: BAU1\u2014left on field (agricultural residues), BAU2\u2014direct land \napplication (animal manure), BAU3\u2014burning on site (forest residues), BAU4\u2014landfilling without methane flaring or capture (MSW), BAU5\u2014landfilling \nwith 75% of methane capture and use for on-site CHP (MSW). See Table 1 for additional details of conversion pathways.\nNature Energy | VOL 4 | AUGUST 2019 | 700\u2013708 | www.nature.com/natureenergy\n702\n\nAnalysis\nNature Energy\nTechnical comparison of conversion pathways. We first estimate \nthe renewable energy production, net energy gain and GHG foot-\nprint of different conversion pathways on a per unit wet weight basis \nfor various types of waste. Feedstock-level results are depicted in \nSupplementary Figs. 5\u20137. Methods explains how we first calculate \nthese for each distinct waste biomass source at the US county level, \nand subsequently compute a mass-weighted average for each of the \nfour broad categories of wastes at the national level. The renew-\nable energy yield across conversion pathways ranges from 0.2 to \n13.1 gigajoules (GJ) per megagram (Mg) of waste, while net energy \ngain ranges from \u22122.4 to 11.6\u2009GJ\u2009Mg\u22121 (Fig. 1a,b). It is clear that \nthe energy value of coproducts is critical to achieving positive net \nenergy for a number of conversion pathways and waste feedstocks. \nExcept for animal-manure-related pathways, all conversion path-\nways result in positive net energy gains and considerable energy \nreturn on investment. For animal manure, only anaerobic digestion \n(M2) yields positive net energy, and its energy return on investment \nis only slightly greater than 1. The net GWP across the pathways \nranges from \u22120.9 to 0.7\u2009tCO2e\u2009Mg\u22121 (Fig. 1d). As with the impor-\ntance of coproducts in net energy gain, emissions avoided by the \nresulting coproduct(s) displacing a substitute accounts for a sub-\nstantial portion of the climate benefits for most pathways.\nLooking into each broad waste category, for agricultural and for-\nest residues, combined heat and power generation (CHP) offers both \nthe greatest net energy gain and the greatest climate benefits. For \nmunicipal solid waste (MSW), CHP offers the highest net energy gain \nwhile anaerobic digestion returns more climate benefits than other \npathways. When compared with current management practices, all \nconversion pathways result in climate benefits for agricultural resi-\ndues. For animal manure, only anaerobic digestion producing either \nmethane (M2) or electricity and heat (E4) yields climate benefits. \nThis corresponds to previous studies, which indicate that anaerobic \ndigestion is the optimal conversion pathway for animal manure15,27,28. \nAlthough some pathways appear not to contribute to climate-change \nmitigation (that is, result in positive net GWP), all conversion path-\nways for forest residues yield smaller net GWP relative to burning \nthem on site. When compared with landfilling without any methane \nflaring or capture, all conversion pathways for MSW result in smaller \nnegative effects on the climate. However, landfilling with methane \ncapture and on-site CHP would greatly reduce the GHG emissions \nof landfilling and become more attractive than renewable-diesel-\nrelated conversion pathways (Fig. 1c,d).\nBreakdown of GHG emission sources. Disaggregating the con-\ntribution to total GHG emissions from the different stages in the \nproduction chain shows that emissions during the processing stage, \nwhich requires electricity and heat input, and credits for avoided \nemissions attributable to displaced products, are key determinants \nof GHG emissions for most conversion pathways (Fig. 1). This is \ngenerally in line with results from a number of recent studies, such \nas refs. 15,20,34. For agricultural residues, current management prac-\ntice (that is, left and decayed on field) entail no GWP due to the \nfact that the GWPbio index for annual crops is zero. While the same \nGWPbio index applies to animal feed, methane and N2O emissions \nfrom animal farm operations contribute to total emissions from \ndirect land application of manure. For MSW, the major sources of \nnon-biogenic carbon are contained in plastics, rubber and leather, \nand textiles. For non-electricity pathways, non-biogenic carbon in \nMSW feedstocks would be transferred into energy products and \neventually be emitted into the atmosphere as CO2 during end use. \nThis explains a large amount of emissions during the end-use stage \nfor these pathways. For electricity-related pathways (E1\u2013E4), non-\nbiogenic carbon would be emitted as CO2 during the processing \nphase. For other types of MSW feedstock, biogenic carbon would \nbe emitted as biogenic CO2 in various phases. Thus, we treated bio-\ngenic CO2 as a separate source of GHG emissions.\nSensitivity analysis of emission estimates. Given that electric-\nity consumption during biomass processing is the major source of \nenergy inputs and emissions across most conversion pathways, a \nsensitivity analysis on the emission intensities of state power grids \nwas conducted. Note that, even though biomass processing requires a \nsubstantial amount of heat energy, it is typically derived from natural \ngas, whose emission intensity is much less variable across regions rel-\native to the emission intensity of electricity. Results show that cleaner \npower grids in general would yield less climate benefit for electric-\nity pathways and more climate benefit for non-electricity pathways \n(Fig. 2). For cleaner power grids, electricity-related pathways would \non one hand result in lower emissions during the processing stage, \nbut on the other hand lead to less climate benefit from the displace-\nment of grid electricity. For the majority of non-electricity pathways, \nelectricity is only an input, so cleaner power grids would result in \nlower emissions during the processing stage and the overall life cycle. \nFor instance, whereas converting agricultural and forest residues \ninto electricity through CHP (E1) and into biomethane through \ngasification (M1) appear equally beneficial under current condi-\ntions, M1 becomes more beneficial when power grids are cleaner. \nAnother sensitivity analysis on transportation distance was also con-\nducted (Supplementary Fig. 8). However, a distance ranging from \n25 to 150\u2009km negligibly affects results on GHG emissions. Thus, we \nAg. residues\nAnimal manure\nForest residues\nMSW\nE1\nE2\nE3\nM1\nEth1\nRd1\nBj1\nBj2\nE1\nE2\nE3\nE4\nM1\nM2\nE1\nE2\nE3\nM1\nEth1\nRd1\nRd2\nBj1\nBj2\nBj3\nBj4\nBj5\nBj6\nE1\nE2\nE3\nE4\nM1\nM2\nEth1\nRd1\nRd2\nBj1\nBj2\nBj3\nBj4\nBj5\nBj6\n\u22121.5\n\u22121.0\n\u22120.5\n0\n0.5\n1.0\ntCO2e per Mg ww\nCleaner power (\u221250% carbon intensity)\nCurrent state power grids\nFossil rollback (+50% carbon intensity)\nFig. 2 | Sensitivity analysis of emission estimates. Electricity pathways: E1\u2014CHP, E2\u2014gasification + CHP, E3\u2014integrated gasification combined cycle, \nE4\u2014anaerobic digestion + CHP. Methane pathways: M1\u2014gasification, M2\u2014anaerobic digestion. Ethanol pathway: Eth1\u2014enzymatic hydrolysis + \nfermentation. Renewable diesel pathways: Rd1\u2014gasification + FT synthesis, Rd2\u2014pyrolysis + hydroprocessing. Bio jet fuel pathways: Bj1\u2014alcohol-to-jet \n(ethanol), Bj2\u2014sugar-to-jet (fermentation), Bj3\u2014pyrolysis (in\u00a0situ), Bj4\u2014pyrolysis (ex\u00a0situ), Bj5\u2014HTL (in\u00a0situ), Bj6\u2014HTL (ex\u00a0situ).\nNature Energy | VOL 4 | AUGUST 2019 | 700\u2013708 | www.nature.com/natureenergy\n703\n\nAnalysis\nNature Energy\nassumed 150\u2009km as the transportation distance in order to provide \nconservative estimates for net energy gain as well as GHG emissions.\nMaximizing aggregate energy and climate benefits. We next \ndescribe the maximum energy and climate benefits achievable at a \nnational scale through optimal utilization of waste biomass gener-\nated in each county within the United States taking into account \nspatial variation in the electricity mix. As noted earlier, about \n233\u2009Mt of dry waste resources is available annually in the contigu-\nous United States16. The spatial distribution of this total resource \nbase is depicted in Supplementary Figs. 1 and 2. Approximately 25% \nof this total is concentrated in 115 counties, 50% is in 374 coun-\nties and 75% is in 884 counties (Supplementary Figs. 1 and 2 and \nSupplementary Note 1). Agricultural states in the Pacific West, the \nMidwest and the South in general stand out with more agricultural \nresidues than other regions. Counties in the Mountain West and the \nSouth are endowed with substantial forest residues. The availability \nof animal manure corresponds to livestock and poultry production, \nwhich is concentrated in California and the Midwest. The avail-\nability of MSW is concentrated in densely populated regions such \nas Southern California, Florida and parts of the Northeast. Overall, \nhowever, some of the largest metropolitan areas stand out in terms \nof the availability of total waste resources.\nSearching for the conversion pathway that is optimal with respect \nto all three criteria\u2014renewable energy, net energy and GWP\u2014we \nfind that, except in rare instances, no single pathway exists for any \ngiven type of waste across all US counties and states (Table 2). Across \ndifferent types of agricultural residue, CHP (E1) consistently stands \nout with respect to all three objectives for a substantial fraction of \ncounties and states. For animal manure, no single pathway satis-\nfies all three objectives. For forest residues and municipal wastes, \noptimal conversion pathways that satisfy all three objectives vary by \nspecific waste feedstocks. The percentage of locations where there is \na single optimal pathway varies substantially.\nSince a single pathway that achieves all three objectives for any \ngiven waste feedstock across locations is lacking, there is a need to \nconsider three distinct scenarios of optimal use of biomass wastes\u2014\nmaximum energy production (MEP), maximum net energy (MNE) \nand maximum emission reduction (MER). For each county in the \nUnited States, we first select the conversion pathway for each type \nTable 2 | Synergies between renewable energy, net energy and GWP at the county and state levels\nWaste type\nFeedstock\nTotal number \nof counties \nwith feedstock \navailable\nAll three criteria aligned\nTotal number \nof states with \nfeedstock \navailable\nAll three criteria aligned\nOptimal \npathway\nNumber of \ncounties\nPercentage \n(%)\nNumber of \nstates\nPercentage \n(%)\nAg. residues\nBarley straw\n136\n52\n38\n14\n5\n36\nE1\nCitrus residues\n118\n53\n45\n9\n3\n33\nE1\nCorn stover\n1,276\n793\n62\n36\n22\n61\nE1\nCotton gin trash\n815\n329\n40\n17\n6\n35\nE1\nCotton residues\n796\n305\n38\n17\n6\n35\nE1\nNon-citrus residues\n1,686\n795\n47\n48\n20\n42\nE1\nOat straw\n12\n4\n33\n2\n1\n50\nE1\nRice hulls\n144\n77\n53\n6\n3\n50\nE1\nRice straw\n148\n80\n54\n6\n3\n50\nE1\nSorghum stubble\n191\n161\n84\n9\n6\n67\nE1\nSugarcane bagasse\n29\n11\n38\n3\n2\n67\nE1\nSugarcane trash\n29\n11\n38\n3\n2\n67\nE1\nTree-nut residues\n620\n234\n38\n40\n14\n35\nE1\nWheat straw\n696\n207\n30\n32\n11\n34\nE1\nAnimal manure\nHogs, 1,000+ head\n934\n0\n0\n37\n0\n0\n\u2014\nMilk cows, 500+ head\n639\n0\n0\n44\n0\n0\n\u2014\nForest residues\nPrimary mill residues\n488\n178\n36\n44\n12\n27\nE1\nSecondary mill \nresidues\n2,418\n590\n24\n49\n11\n22\nE1\nOther forest residues\n1,256\n588\n47\n35\n15\n43\nE1\nOther forest thinnings\n304\n96\n32\n11\n5\n45\nE1\nMSW\nCD waste\n3,109\n0\n0\n49\n0\n0\n\u2014\nFood waste\n2,792\n0\n0\n48\n0\n0\n\u2014\nMSW wood\n3,109\n2,487\n80\n49\n39\n80\nBj5\nPaper and paperboard\n3,109\n0\n0\n49\n0\n0\n\u2014\nPlastics\n3,109\n0\n0\n49\n0\n0\n\u2014\nRubber and leather\n3,109\n0\n0\n49\n0\n0\n\u2014\nTextiles\n3,109\n0\n0\n49\n0\n0\n\u2014\nYard trimmings\n3,066\n0\n0\n49\n0\n0\n\u2014\nOther MSW\n3,109\n0\n0\n49\n0\n0\n\u2014\nNote: E1\u2014CHP; Bj5\u2014HTL (in\u00a0situ).\nNature Energy | VOL 4 | AUGUST 2019 | 700\u2013708 | www.nature.com/natureenergy\n704\n\nAnalysis\nNature Energy\nof waste under each of the three scenarios. The national results \nare the aggregation of county-level results. The calculations are \ndescribed in Methods and results are depicted in Table 3 and Fig. \n3. Scenario results suggest that there is substantial benefit from uti-\nlizing wastes and biomass residues to either displace energy pro-\nduction or reduce GHG emissions or both. As one would expect, \nMEP results in the highest potential of renewable energy produc-\ntion, which totals 3.8\u2009EJ\u20143.7% of total US energy demand in 2016 \n(ref. 46), and MER results in the highest potential of emissions \nreduction, 178\u2009MtCO2e\u20142.7% of total US GHG emissions in 2016 \n(ref. 47). The MNE scenario indicates the highest potential of net \nenergy as well as a moderate amount of emissions reduction (75% of \nMER). A breakdown of scenario results by waste feedstock reveals \nthe preferred conversion pathways under each of the three scenar-\nios (Supplementary Table 4). CHP (E1) is the preferred option for \nagricultural resides under both the MEP and MNE scenarios, while \neither CHP (E1) or gasification (M1) may maximize GHG emis-\nsion reduction depending on specific feedstock. For dairy manure, \nCHP (E1) is the preferred option that maximizes renewable energy \nproduction, but anaerobic digestion to biomethane (M2) maximizes \nboth the net energy gains and climate benefits. For forest residues, \nCHP (E1) results in the largest amount of renewable energy and \nnet energy gain, while either hydrothermal liquefaction (HTL) with \nin\u00a0situ hydrogen production (Bj5) or gasification (M1) maximizes \nGHG emission reduction. In contrast to other categories of wastes, \noptimal use of MSW feedstocks would require a greater number of \nconversion technology pathways depending on specific feedstock. \nNon-biogenic carbon in MSW is concentrated in three feedstocks\u2014\nplastics, rubber and leather, and textiles. Thus, the non-biogenic \ncarbon is immediately emitted into the atmosphere when process-\ning these feedstocks instead of being stored in landfills. While the \ninclusion of biogenic CO2 reduces net GWP for forest residues and \nMSW (Fig. 1c,d), it does not change the ranking of conversion path-\nways under the three scenarios.\nThe county-level distribution of renewable energy production, \nnet energy gain and its associated climate benefits also indicates \nthat most counties would lose a relatively small amount of energy \nproduction potential from MEP to MER, while most coun-\nties would see a greater increase in terms of emission reduction  \n(Fig. 3). Maximizing energy production would result in negative \nnet energy in 125 counties and emission increase in 532 counties \n(Fig. 3b,c). Therefore, maximizing either net energy or emission \nreduction would lead to better utilization of wastes and residues \nrelative to maximizing renewable energy. Given that the terms \nrenewable energy and clean energy tend to often be used inter-\nchangeably by policy makers, this analysis shows that there exist \npotential trade-offs between different criteria relevant to sustain-\nable development.\nConclusions\nMaximizing the benefits of waste conversion requires attention to \nfirst the life-cycle implications of different technology pathways, \nsecond the spatial distribution of waste feedstocks, and third the \nlocal conditions under which waste feedstocks will be processed. \nThe policy insight that emerges from this analysis is that national \nmandates such as the US Renewable Fuel Standard might not \nmaximize even renewable energy production let alone environ-\nmental benefits. Likewise, renewable portfolio standards, a widely \nemployed policy in the electricity sector, could lead to suboptimal \nuse of waste biomass. In the literature, bioenergy and biofuel poli-\ncies have been analysed mainly from the perspective of climate-\nchange mitigation, food security or cost, but this analysis shows that \nTable 3 | Total renewable energy production, net energy gain \nand GWP across scenarios\nPolicy \nscenarios\nRenewable \nenergy \nproduction\nNet \nenergy \ngain\nGWP\nEJ\nIndex \n(%)\nEJ\nIndex \n(%)\nMtCO2e\nIndex \n(%)\nMEP\n3.8\n100\n2.9\n89\n\u2212103\n58\nMNE\n3.7\n96\n3.2\n100\n\u2212133\n75\nMER\n3.1\n81\n2.4\n76\n\u2212178\n100\nTJ\n<100\n100\u2013500\n500\u20131,000\n1,000\u20132,000\n2,000\u20135,000\n5,000\u201310,000\n10,000\u201320,000\n20,000\u201340,000\n>40,000\nRenewable energy production\u2014MEP scenario\na\nTJ\n<\u20131,000\n\u20131,000\u20130\n0\u2013500\n500\u20131,000 \n1,000\u20132,000\n2,000\u20135,000\n5,000\u201310,000\n10,000\u201320,000\n>20,000\nTJ\n<\u20131,000\n\u20131,000\u20130\n0\u2013500\n500\u20131,000 \n1,000\u20132,000\n2,000\u20135,000\n5,000\u201310,000\n10,000\u201320,000\n>20,000\nTJ\n<\u20131,000\n\u20131,000\u20130\n0\u2013500\n500\u20131,000 \n1,000\u20132,000\n2,000\u20135,000\n5,000\u201310,000\n10,000\u201320,000\n>20,000\nNet energy\u2014MEP scenario\nb\nktCO2e\n<\u22121,000\n\u22121,000 to \u2212500\n\u2212500 to \u2212200\n\u2212200 to \u2212100\n\u2212100 to \u221250\n\u221250 to 0\n0 to 50\n50 to 200\n>200\nktCO2e\n<\u22121,000\n\u22121,000 to \u2212500\n\u2212500 to \u2212200\n\u2212200 to \u2212100\n\u2212100 to \u221250\n\u221250 to 0\n0 to 50\n50 to 200\n>200\nktCO2e\n<\u22121,000\n\u22121,000 to \u2212500\n\u2212500 to \u2212200\n\u2212200 to \u2212100\n\u2212100 to \u221250\n\u221250 to 0\n0 to 50\n50 to 200\n>200\nGWP (w. biogenic CO2)\u2014MEP scenario\nc\nTJ\n<100\n100\u2212500\n500\u22121,000\n1,000\u22122,000\n2,000\u22125,000\n5,000\u221210,000\n10,000\u221220,000\n20,000\u221240,000\n>40,000\nRenewable energy production\u2014MNE scenario\nd\nNet energy\u2014MNE scenario\ne\nGWP (w. biogenic CO2)\u2014MNE scenario\nf\nTJ\n<100\n100\u2212500\n500\u22121,000\n1,000\u22122,000\n2,000\u22125,000\n5,000\u221210,000\n10,000\u221220,000\n20,000\u221240,000\n>40,000\nRenewable energy production\u2014MER scenario\ng\nNet energy\u2014MER scenario\nh\nGWP (w. biogenic CO2)\u2014MER scenario\ni\nFig. 3 | County-level renewable energy production, net energy and emissions. a\u2013c, The MEP scenario. d\u2013f, The MNE scenario. g\u2013i, The MER scenario.\nNature Energy | VOL 4 | AUGUST 2019 | 700\u2013708 | www.nature.com/natureenergy\n705\n\nAnalysis\nNature Energy\nthey also do not optimize energy production. From a methodologi-\ncal perspective, this analysis illustrates the value of combining LCA \nwith spatial analytical techniques for multicriterion assessment of \nalternative conversion pathways and the identification of hotspots \nfor the refinement of existing energy policies. Indexing volumetric \ntargets and mandates as well as financial subsidies for renewable \nenergy to life-cycle emission-based performance measures will lead \nto more sustainable use of wastes and biomass residues.\nThis study is a first step towards using a common system bound-\nary for a consistent comparison of a large variety of waste conver-\nsion technologies from the twin perspectives of net energy gain and \nclimate benefits. Incorporating non-GHG environmental consider-\nations including air quality impacts and freshwater use and water \nquality impacts, as well as an assessment of the levelized life-cycle \ncost of energy for the different pathways, are two important direc-\ntions for future research.\nMethods\nAn overview of conversion technology pathways. The 15 conversion technology \npathways included in this study can be categorized into five groups: electricity \npathways (E1\u2013E4), methane pathways (M1, M2), an ethanol pathway (Eth1), \nrenewable diesel pathways (Rd1, Rd2) and bio jet fuel pathways (Bj1\u2013Bj6). Details \nof the conversion pathways including process description, feedstock feasibility, \nenergy inputs and outputs, coproducts, displaced products and references are \npresented in Table 1.\nApproach to energy and emission accounting. We conducted an LCA to estimate \nthe energy balances and GHG emissions associated with the conversion of a given \nfeedstock to the final energy product(s) in each county. The different phases \nof the life cycle that are accounted for include collection of waste, transport to \nthe conversion facility, processing (including pretreatment), transmission and \ndistribution, and end use (Supplementary Note 2 and Supplementary Fig. 3). \nThornley et\u00a0al. showed that different functional units would result in varying \noutcomes when comparing alternative uses of biomass, and the functional unit \nshould correspond to \u201cthe actual nature of the research questions\u201d48. Since this \nstudy mainly focuses on the optimal use of wastes, the functional unit of this LCA \nis thus 1\u2009Mg of wet waste.\nEnergy and emissions from collection and transport of feedstock are estimated \non the basis of this activity requiring heavy-duty diesel trucks. Feedstock-specific \ntechnology data (including lower heating values, moisture content, non-biogenic \ncarbon content, energy inputs and outputs by the conversion pathway) were \ncollected from the literature to calculate energy and emission flows in each phase \nas well as the overall net energy gains15,39,42,49\u201352. Table 1 shows additional data \nsources. Losses during transmission and distribution were taken into account.50 \nEmissions associated with the provision of energy inputs were based on life-cycle \nemission intensities of electricity generation and other fossil-based fuel production \n(heat, natural gas, diesel, hydrogen)53\u201355. Emission intensities of the production of \nelectricity and fossil-based fuels vary geographically, and variation across states \nin such emission intensities was taken into account (Supplementary Note 3 and \nSupplementary Table 1). Life-cycle GHG emission intensities of state power grids \nwere estimated by multiplying a state\u2019s generation mix from the Emissions & \nGeneration Resource Integrated Database with life-cycle GHG emission intensities \nof respective electricity generation technologies from the LCA Harmonization \nproject56,57. The GWP for non-CO2 GHG is based on IPCC Fifth Assessment \nReport 100-year conversion factors58.\nComparing the burdens associated with converting a given feedstock to \ndifferent end products does not, however, paint a complete picture of the benefits \nof choosing one conversion pathway over another. The ultimate environmental \nbenefit of any given pathway is also a function of the process(es) or product(s) \nthat it displaces. For instance, if conversion of manure to renewable natural gas \nfor pipeline injection entails more GHG emissions relative to conversion to biogas \nfor on-site power generation, it is plausible that the former is more beneficial \nif electricity from biogas displaces clean electricity while renewable natural gas \ndisplaces diesel used in trucks or displaces fossil natural gas. Supplementary \nFig. 4 illustrates a simple schematic representation of this concept. Posen et\u00a0al. \nillustrate this idea in the context of converting cellulosic biomass to ethanol and \ndisplacing gasoline vis-\u00e0-vis producing bioethylene and displacing fossil-fuel-\nderived ethylene59. For the handling of coproducts, we chose the displacement \nmethod over allocation methods based on energy or economics for the following \nreasons. First, the International Standards Organization advocates the use of the \ndisplacement method60 and it has been adopted as the default method in many \nLCA models and in biofuel regulation development in the United States. Second, \nmany pathways yield a number of different types of energy product\u2014electricity, \nheat, methane and/or liquid fuels. The conventional products to be displaced can \neasily be defined. Third, the distinction between main product and coproducts in \nthis study is mainly to categorize the pathways into five groups. We intended to \nexamine the conversion pathways from a systems perspective, that is, all types of \nenergy product through each conversion pathway instead of the main products \nonly. The displacement method represents the idea of system expansion and is \nmore suitable for our analysis. Fourth, the characteristics (utility, energy form \nand so on) of electricity are different from those of other types of energy product, \nand so are those of each other type of energy product. Allocation simply based \non energy content may result in distorted results. In addition, the price ratios \nfor an economic allocation may be challenging, as some of the energy products \nfrom waste conversion may be non-commoditized and the prices may fluctuate \nand vary greatly by geographic location. Net GHG emissions were calculated by \nsubtracting displaced emissions from the life-cycle emissions of each conversion \npathway. Biogenic CO2 emissions are included throughout life cycles. The GWP \nof biogenic CO2 emissions was estimated by multiplying the GWPbio indices by \nbiogenic CO2 emissions. Additional details on the method and data sources for \nbiogenic CO2 emissions are listed in Supplementary Note 4 and Supplementary \nTable 2. Thus, the net GWP of a given feedstock converted through a given \npathway is equivalent to the sum of net GHG emissions and the GWP of biogenic \nemissions. Emissions and energy related to material use (such as enzymes and \ncatalysts) are not included in the analysis.\nThe basic county-level calculations we performed in order to assess the \npotentials of energy production and life-cycle GWP are the following:\n\u2211\n=\n\u2212\nEP\nWW\n(EO\n(1 TD ))\n(1)\ni j c\ni c\nk\ni j k\nk\n, ,\n,\n, ,\n\u2211\n=\n\u2212\n+\n+\n(\n)\nE\nE\nD\nNE\nEP\nWW\nEI\n(2)\ni j c\ni j c\ni c\nl\ni j l\ni\n, ,\n, ,\n,\n, ,\ncollection,\ntransport\n1\n\u2211\n=\n+\n+\n+\n+\n+\n+\n\u2212\nE\nE\nD\nW\nD\nGWP\nWW (\n)\nEmissI\n(EI\nEmissI ))\nEmiss\nEmissI\nEmiss\nGWP\nEP\nEmissI\n(3)\ni j c\ni c\ni\nc\nl\ni j l\nl c\ni j k\nc\ni\ni j k\nm c\n, ,\n,\ncollection,\ntransport\n1\ndiesel,\n, ,\n,\nprocess\n, ,\ndiesel,\n2\nenduse\nbioCO2\n, ,\n,\n=\nGWP\nGWP\nEmiss\n(4)\ni\ni\ni\nbioCO2\nbio,\nbioCO2,\nwhere EPi,j,c is the renewable energy production (MJ) of feedstock i through \nconversion pathway j in county c; WWi,c the wet weight (kg) of feedstock i in \ncounty c; EOi,j,k energy output k (MJ\u2009kg\u22121) of feedstock i through conversion \npathway j; TDk the transmission and distribution loss of energy output k (6.5% \nassumed for electricity, 20% for heat and 2% for methane), NEi,j,c the net energy \n(MJ) of feedstock i through conversion pathway j in county c; EIi,j,l energy input \nl (MJ\u2009kg\u22121) of feedstock i through conversion pathway j; GWPi,j,c the net GWP \n(gCO2e) of feedstock i through conversion pathway j in county c; Ecollection,i the \nenergy consumption rate (MJ\u2009kg\u22121) of collecting feedstock i; Etransport the energy \nconsumption rate (MJ\u2009kg\u22121\u2009km\u22121) of transporting feedstock i to the conversion \nfacility, D1 the transport distance (km) from the temporary storage or collection \nsite to the conversion facility (150\u2009km assumed), EmissIdiesel,c the life-cycle GHG \nemission intensity (gCO2e\u2009MJ\u22121) of petroleum-based diesel in county c, EmissIl,c \nthe life-cycle GHG emission intensity (gCO2e\u2009MJ\u22121) of energy input l in county c, \nEmissprocess the direct GHG emissions (excluding biogenic CO2) during processing, \nWi,j,k the physical weight (kg) of energy output k of feedstock i through conversion \npathway j; D2 the transport distance (km) for distribution (150\u2009km assumed), \nEmissenduse the direct GHG emissions (excluding biogenic CO2) during end use, \nGWPi\nbioCO2 the GWP (gCO2e) of biogenic carbon in feedstock i, EPi,j,k the energy \nproduction (MJ) of output k of feedstock i through conversion pathway j; EmissIm,c \nthe life-cycle GHG emission intensity (gCO2e\u2009MJ\u22121) of energy product m (which \noutput k can substitute) in county c, GWPbio,i the biogenic CO2 global warming \nindex with full impulse response functions for feedstock i and Emissbioco2,i the \nbiogenic CO2 emissions of feedstock i.\nFor the comparison of conversion pathways, county-level results were first \naggregated to the national level and by feedstock. Weighted averages (by weight) of \nresults by feedstock in each of the four broader categories of waste resources were \ncalculated for the comparison by waste type (as shown in Fig. 1). For the current \nmanagement of wastes and residues (that is, BAU practices in Fig. 1), we used the \nsame emission accounting method and life-cycle framework to estimate the GWP \n(Supplementary Note 5 and Supplementary Table 3).\nSensitivity analysis. A sensitivity analysis of net GHG emissions was conducted \nto explore the impacts of the emission intensity of current state power grids and \ntransportation distance. For the sensitivity analysis on electricity, two additional \nelectricity generation scenarios were constructed: \u2018cleaner power\u2019\u2014assuming \na 50% reduction in emission intensity of power grids in all states; and \u2018fossil \nrollback\u2019\u2014assuming a 50% increase in emission intensity of power grids in all \nstates. In addition, a range of 25\u2013150\u2009km was examined to test the sensitivity to \ntransportation distance.\nNature Energy | VOL 4 | AUGUST 2019 | 700\u2013708 | www.nature.com/natureenergy\n706\n\nAnalysis\nNature Energy\nTechnical availability of waste resources. County-level waste availability data \nwere obtained from the base-year estimates under the reference scenario in the US \nDepartment of Energy\u2019s BT16. BT16 estimates the biophysical potential, spatial \ndistribution, economic constraints and environmental impacts associated with \nexisting and potential biomass resources16. Waste resources included in this study \ncomprise four types of waste: agricultural residues (14 feedstocks, including both \nprimary and secondary agricultural residues as defined in BT16), animal manure \n(2 feedstocks), forest residues (4 feedstocks) and MSW (9 feedstocks, including \nfood waste). Technical availability was defined as the maximum potential of waste \nresources without taking into account feedstock costs. BT16 reports dry weight of \nwaste feedstocks, and wet weight was calculated with moisture content for a more \nprecise estimation of energy consumption and emissions during the collection and \ntransport stages.\nScenario analysis. To explore the optimal utilization of waste biomass resources, \nwe developed three alternative scenarios: MEP, MNE and MER. For all scenarios, \nthe optimal conversion pathway for each feedstock was selected on the basis of the \nmaximum value of energy or emission reduction. Under each scenario, the county-\nlevel results were then added up to obtain the potentials for total renewable energy \nproduction, net energy and emission reduction at the national level.\nData availability\nThe data that support the findings of this study are available at https://github.com/\nlabyseson/Waste-LCA\nCode availability\nCodes for energy and emission accounting as well as data visualization are available \nat https://github.com/labyseson/Waste-LCA\nReceived: 30 November 2018; Accepted: 6 June 2019;  \nPublished online: 22 July 2019\nReferences\n\t1.\t The State of Food and Agriculture 2008. Biofuels: Prospects, Risks and \nOpportunities (FAO, 2008).\n\t2.\t Renewables 2017: Global Status Report (REN21, 2017).\n\t3.\t de Gorter, H., Drabik, D. & Just, D. R. How biofuels policies affect the level \nof grains and oilseed prices: theory, models and evidence. Glob. Food Secur. 2, \n82\u201388 (2013).\n\t4.\t To, H. & Grafton, R. Q. Oil prices, biofuels production and food security: \npast trends and future challenges. Food Secur. 7, 323\u2013336 (2015).\n\t5.\t Tadasse, G., Algieri, B., Kalkuhl, M. & Von Braun, J. in Food Price Volatility \nand its Implications for Food Security and Policy 59\u201382 (Springer, 2016).\n\t6.\t Hochman, G., Rajagopal, D., Timilsina, G. & Zilberman, D. Quantifying the \ncauses of the global food commodity price crisis. Biomass Bioenerg. 68, \n106\u2013114 (2014).\n\t7.\t Runge, C. F. & Senauer, B. How biofuels could starve the poor. Foreign Aff. \n86, 41\u201353 (2007).\n\t8.\t Lambin, E. F. & Meyfroidt, P. Global land use change, economic globalization, \nand the looming land scarcity. Proc. Natl Acad. Sci. USA 108, 3465\u20133472 (2011).\n\t9.\t Melillo, J. M. et\u00a0al. Indirect emissions from biofuels: how important? Science \n326, 1397\u20131399 (2009).\n\t10.\tFarrell, A. E. et\u00a0al. Ethanol can contribute to energy and environmental goals. \nScience 311, 506\u2013508 (2006).\n\t11.\tCrutzen, P. J., Mosier, A. R., Smith, K. A. & Winiwarter, W. in Paul J. \nCrutzen: a Pioneer on Atmospheric Chemistry and Climate Change in the \nAnthropocene (eds Crutzen, P. J. & Brauch, H. G.) 227\u2013238 (Springer, 2016).\n\t12.\tRajagopal, D. & Zilberman, D. Environmental, economic and policy aspects \nof biofuels. Found. Trends Microecon. 4, 353\u2013468 (2008).\n\t13.\tWhalen, J. et\u00a0al. Sustainable biofuel production from forestry, agricultural and \nwaste biomass feedstocks. Appl. Energy 198, 281\u2013283 (2017).\n\t14.\tCampbell, J. E. & Block, E. Land-use and alternative bioenergy pathways for \nwaste biomass. Environ. Sci. Technol. 44, 8665\u20138669 (2010).\n\t15.\tTonini, D., Hamelin, L., Alvarado-Morales, M. & Astrup, T. F. GHG emission \nfactors for bioelectricity, biomethane, and bioethanol quantified for 24 \nbiomass substrates with consequential life-cycle assessment. Bioresour. \nTechnol. 208, 123\u2013133 (2016).\n\t16.\t2016 Billion-Ton Report: Advancing Domestic Resources for a Thriving \nBioeconomy, Volume 1: Economic Availability of Feedstocks ORNL/TM-\n2016/160 (US Department of Energy, 2016).\n\t17.\tCarreras-Sospedra, M., Williams, R. & Dabdub, D. Assessment of the \nemissions and air quality impacts of biomass and biogas use in California.  \nJ. Air Waste Manag. Assoc. 66, 134\u2013150 (2016).\n\t18.\tWang, W. & Tao, L. Bio-jet fuel conversion technologies. Renew. Sustain. \nEnergy Rev. 53, 801\u2013822 (2016).\n\t19.\tde Jong, S. et\u00a0al. The feasibility of short\u2010term production strategies for \nrenewable jet fuels\u2014a comprehensive techno\u2010economic comparison. Biofuels \nBioprod. Bioref. 9, 778\u2013800 (2015).\n\t20.\tde Jong, S. et\u00a0al. Life-cycle analysis of greenhouse gas emissions from \nrenewable jet fuel production. Biotechnol. Biofuels 10, 64 (2017).\n\t21.\tStaples, M. D., Malina, R. & Barrett, S. R. The limits of bioenergy for \nmitigating global life-cycle greenhouse gas emissions from fossil fuels.  \nNat. Energy 2, 16202 (2017).\n\t22.\tSustainable Development Goals: 17 Goals to Transform Our World.  \nUnited Nations https://www.un.org/sustainabledevelopment (2018).\n\t23.\tGeissdoerfer, M., Savaget, P., Bocken, N. M. & Hultink, E. J. The Circular \nEconomy\u2014a new sustainability paradigm? J. Clean. Prod. 143,  \n757\u2013768 (2017).\n\t24.\tStahel, W. R. The circular economy. Nat. News 531, 435 (2016).\n\t25.\tLiu, W. et\u00a0al. Economic and life cycle assessments of biomass utilization for \nbioenergy products. Biofuels Bioprod. Bioref. 11, 633\u2013647 (2017).\n\t26.\tLaurent, A. & Espinosa, N. Environmental impacts of electricity  \ngeneration at global, regional and national scales in 1980\u20132011: what  \ncan we learn for future energy planning? Energy Environ. Sci. 8,  \n689\u2013701 (2015).\n\t27.\tAguirre-Villegas, H. A. & Larson, R. A. Evaluating greenhouse gas emissions \nfrom dairy manure management practices using survey data and life cycle \ntools. J. Clean. Prod. 143, 169\u2013179 (2017).\n\t28.\tAguirre\u2010Villegas, H. A., Larson, R. & Reinemann, D. J. From waste\u2010to\u2010worth: \nenergy, emissions, and nutrient implications of manure processing pathways. \nBiofuels Bioprod. Bioref. 8, 770\u2013793 (2014).\n\t29.\tBanks, C. J., Chesshire, M., Heaven, S. & Arnold, R. Anaerobic digestion of \nsource-segregated domestic food waste: performance assessment by mass and \nenergy balance. Bioresour. Technol. 102, 612\u2013620 (2011).\n\t30.\tBroun, R. & Sattler, M. A comparison of greenhouse gas emissions and \npotential electricity recovery from conventional and bioreactor landfills.  \nJ. Clean. Prod. 112, 2664\u20132673 (2016).\n\t31.\tMacias-Corral, M. et\u00a0al. Anaerobic digestion of municipal solid waste and \nagricultural waste and the effect of co-digestion with dairy cow manure. \nBioresour. Technol. 99, 8288\u20138293 (2008).\n\t32.\tMorris, J. Recycle, bury, or burn wood waste biomass? LCA answer depends \non carbon accounting, emissions controls, displaced fuels, and impact costs.  \nJ. Ind. Ecol. 21, 844\u2013856 (2017).\n\t33.\tNuss, P., Gardner, K. H. & Jambeck, J. R. Comparative life cycle assessment \n(LCA) of construction and demolition (C&D) derived biomass and US \nNortheast forest residuals gasification for electricity production. Environ. Sci. \nTechnol. 47, 3463\u20133471 (2013).\n\t34.\tPressley, P. N. et\u00a0al. Municipal solid waste conversion to transportation fuels: \na life-cycle estimation of global warming potential and energy consumption. \nJ. Clean. Prod. 70, 145\u2013153 (2014).\n\t35.\tWang, H., Wang, L. & Shahbazi, A. Life cycle assessment of fast pyrolysis \nof municipal solid waste in North Carolina of USA. J. Clean. Prod. 87, \n511\u2013519 (2015).\n\t36.\tAnex, R. P. et\u00a0al. Techno-economic comparison of biomass-to-transportation \nfuels via pyrolysis, gasification, and biochemical pathways. Fuel 89,  \nS35 (2010).\n\t37.\tIribarren, D., Peters, J. F. & Dufour, J. Life cycle assessment of transportation \nfuels from biomass pyrolysis. Fuel 97, 812\u2013821 (2012).\n\t38.\tBaral, A. & Malins, C. Assessing the Climate Mitigation Potential of Biofuels \nDerived from Residues and Wastes in the European Context (International \nCouncil on Clean Transportation, 2014).\n\t39.\tAstrup, T., M\u00f8ller, J. & Fruergaard, T. Incineration and co-combustion of \nwaste: accounting of greenhouse gases and global warming contributions. \nWaste Manag. Res. 27, 789\u2013799 (2009).\n\t40.\tFruergaard, T. & Astrup, T. Optimal utilization of waste-to-energy in an LCA \nperspective. Waste Manag. 31, 572\u2013582 (2011).\n\t41.\tGabra, M., Pettersson, E., Backman, R. & Kjellstr\u00f6m, B. Evaluation of cyclone \ngasifier performance for gasification of sugar cane residue\u2014Part 1: \ngasification of bagasse. Biomass Bioenerg. 21, 351\u2013369 (2001).\n\t42.\tGabra, M., Pettersson, E., Backman, R. & Kjellstr\u00f6m, B. Evaluation of cyclone \ngasifier performance for gasification of sugar cane residue\u2014Part 2: \ngasification of cane trash. Biomass Bioenerg. 21, 371\u2013380 (2001).\n\t43.\tM\u00f8ller, J., Boldrin, A. & Christensen, T. H. Anaerobic digestion and digestate \nuse: accounting of greenhouse gases and global warming contribution.  \nWaste Manag. Res. 27, 813\u2013824 (2009).\n\t44.\tSwanson, R. M., Platon, A., Satrio, J. A. & Brown, R. C. Techno-economic \nanalysis of biomass-to-liquids production based on gasification. Fuel 89,  \nS19 (2010).\n\t45.\tTews, I. J. et\u00a0al. Biomass Direct Liquefaction Options: TechnoEconomic and Life \nCycle Assessment (Pacific Northwest National Laboratory, 2014).\n\t46.\tJuly 2017 Monthly Energy Review (US Energy Information Administration, \n2017).\n\t47.\tInventory of U.S. Greenhouse Gas Emissions and Sinks: 1990-2016  \n(US Environmental Protection Agency (EPA), 2018).\n\t48.\tThornley, P., Gilbert, P., Shackley, S. & Hammond, J. Maximizing the \ngreenhouse gas reductions from biomass: the role of life cycle assessment. \nBiomass Bioenerg. 81, 35\u201343 (2015).\nNature Energy | VOL 4 | AUGUST 2019 | 700\u2013708 | www.nature.com/natureenergy\n707\n\nAnalysis\nNature Energy\n\t49.\tPhyllis2 Database for Biomass and Waste (Energy Research Centre of the \nNetherlands, 2017); https://phyllis.nl/\n\t50.\tThe Greenhouse Gases, Regulated Emissions, and Energy Use in Transportation \n(GREET) Model GREET_1_2016 (Argonne National Laboratory, 2016).\n\t51.\tWilliams, R. B., Jenkins, B. M. & Kaffka, S. An Assessment of Biomass \nResources in California, 2013 (California Biomass Collaborative, University of \nCalifornia, Davis, 2015).\n\t52.\tWaste Reduction Model (WARM) Tool User\u2019s Guide version 14 (US EPA, 2016).\n\t53.\tCooney, G. et\u00a0al. Updating the US life cycle GHG petroleum baseline to 2014 \nwith projections to 2040 using open-source engineering-based models. \nEnviron. Sci. Technol. 51, 977\u2013987 (2016).\n\t54.\tLee, D., Elgowainy, A. & Dai, Q. Life cycle greenhouse gas emissions of \nhydrogen fuel production from chlor-alkali processes in the United States. \nAppl. Energy 217, 467\u2013479 (2018).\n\t55.\tEcoinvent Database version 3 (Ecoinvent Centre, 2015); https://www.\necoinvent.org/database/database.html\n\t56.\tEmissions and Generation Resource Integrated Database (eGRID2016) (US \nEPA, 2018); https://www.epa.gov/energy/emissions-generation-resource-\nintegrated-database-egrid\n\t57.\tThe Life Cycle Assessment (LCA) Harmonization Project OpenEI Database \n(National Renewable Energy Laboratory, 2012); https://openei.org/apps/LCA/\n\t58.\tEdenhofer, O. et\u00a0al. (eds) Renewable Energy Sources and Climate Change \nMitigation: Special Report of the Intergovernmental Panel on Climate Change \n(Cambridge Univ. Press, 2011).\n\t59.\tPosen, I. D., Griffin, W. M., Matthews, H. S. & Azevedo, I. L. Changing the \nrenewable fuel standard to a renewable material standard: bioethylene case \nstudy. Environ. Sci. Technol. 49, 93\u2013102 (2014).\n\t60.\tEnvironmental Management\u2014Life Cycle Assessment\u2014Principles and Framework \nISO 14040:2006 (International Organization for Standardization, 2006).\n\t61.\tSikarwar, V. S. et\u00a0al. An overview of advances in biomass gasification.  \nEnergy Environ. Sci. 9, 2939\u20132977 (2016).\n\t62.\tMu, D., Seager, T., Rao, P. S. & Zhao, F. Comparative life cycle assessment of \nlignocellulosic ethanol production: biochemical versus thermochemical \nconversion. Environ. Manag. 46, 565\u2013578 (2010).\nAcknowledgements\nThis study would not have been possible without financial support from the UCLA \nGrand Challenges\u2014Sustainable LA programme.\nAuthor contributions\nD.R. conceived and designed the study, guided data collection, modelling and analysis \nand co-wrote the manuscript. B.L. contributed to the study design, collected the data, \nconducted the modelling and analysis and co-wrote the manuscript.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-019-0430-2.\nReprints and permissions information is available at www.nature.com/reprints.\nCorrespondence and requests for materials should be addressed to D.R.\nPublisher\u2019s note: Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2019\nNature Energy | VOL 4 | AUGUST 2019 | 700\u2013708 | www.nature.com/natureenergy\n708\n\n\n    Scientific Research Findings:", "answer": "We find that the type and quantity of energy consumed during biomass processing and the environmental footprint of the displaced products are key to determining the most environmentally beneficial use of any given biomass resource. We estimate that the utilization of all available wastes and biomass residues in the contiguous US can generate 3.1\u20133.8 EJ of renewable energy but deliver only 2.4\u20133.2 EJ of net energy gain if energy generation is maximized, and displace 103\u2013178 million tonne CO2e of Greenhouse Gas (GHG) emissions. For any given waste feedstock, looking across all US counties where it is available, no single conversion pathway simultaneously maximizes renewable energy production, net energy gain and GHG mitigation except in rare instances. Maximizing the energetic and environmental benefits of waste conversion requires a life cycle assessment based of different technology pathways taking into consideration the spatial distribution of biomass resources and local conditions (such as the electricity mix).", "id": 28}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-019-0507-y\n1The Ohio State University School of Environment and Natural Resources, Columbus, OH, USA. 2University of Southern California Sol Price School of Public \nPolicy, Los Angeles, CA, USA. 3Present address: Australian National University School of Regulation and Global Governance, Australian Capital Territory, \nCanberra, Australia. *e-mail: Lee.White@anu.edu.au\nT\no mitigate climate change, electricity grids need to integrate \nlarge shares of renewable generation. Some renewables, \nsuch as solar, are variable and cannot be generated accord-\ning to market needs, which creates challenges in matching supply \nwith demand. Demand-side response (DSR) measures are gain-\ning prominence as a way to align demand with non-dispatchable \nsupply. The residential sector accounts for 30\u201340% of electricity \nconsumption across the Organisation for Economic Co-operation \nand Development countries, which makes it a prime target for \nDSR1. Decision makers, such as the California Public Utilities \nCommission, are enacting policies that require default enrolment in \nDSR programmes2, which can yield participation rates that exceed \n80% (ref. 3); other entities may follow4. Thus, DSR is poised to soon \nreach millions of households, which highlights the need to under-\nstand whether the costs and benefits of DSR are distributed evenly \nacross sociodemographic groups.\nDSR measures typically use price signals to attempt to shift \ndemand away from high-demand \u2018on-peak\u2019 times. Static time-of-use \n(TOU) rates are a common DSR measure that aim to shift electricity \nuse away from on-peak times using a fixed rate schedule with more \nexpensive on-peak times. For households that already struggle with \nelectricity bills, this can be detrimental5\u20138. Households suffering \nfrom energy poverty are forced to make trade-offs between paying \nfor electricity bills versus other necessities, such as food and medi-\ncine5,6,9,10. TOU and other forms of DSR may worsen this trade-off \npressure, often termed \u2018the heat or eat dilemma\u2019.\nThe term \u2018energy poverty\u2019 broadly refers to a confluence of fac-\ntors that result in the inability to maintain a dwelling at a comfort-\nable and healthy temperature, failure of which is associated with \nincreased mortality and morbidity, and having to make decisions \nbetween paying for electricity or other necessities such as food7,9\u201311. \nEnergy poverty is often considered synonymous with \u2018fuel pov-\nerty\u201912, and assesses the same sets of concerns addressed by some \ndefinitions of energy insecurity13.\nEnergy poverty can be considered a state of being in which \nhouseholds face an inability to meet both energy and other costs \nnecessary to live a decent life8. In contrast, energy vulnerability is \ndynamic, with energy-vulnerable households characterized as those \nthat \u201cface a combination of more intense and non-negotiable energy \nneeds as well as a lack of social and/or financial capital\u201d.8 Energy-\nvulnerable households have a limited capacity to adapt to changing \ncircumstances, such as TOU assignment8.\nFurther, energy-vulnerable groups face energy injustices: proce-\ndural injustice in limited access to information, policy participation \nand legal rights; distributional injustice in inequalities in income, \nenergy prices and housing efficiency; and injustice in recognition, \nwhich is a lack of recognition of the differential needs of energy-\nvulnerable groups, and an unequal accordance of respect14. Walker \nand Day14 provide in-depth discussions of these topics. Below, we \ndraw on energy poverty and energy justice literature to define vul-\nnerability indicators.\nLow-income households face pressure to curtail energy costs, \noften with negative impacts8. For instance, during winter months \nwith high heating bills, low-income households curtail energy use \nto thermally uncomfortable levels6. Electricity shut-offs that result \nfrom utility debt can exacerbate both physical and mental health \nconditions5, and low incomes are linked to a higher likelihood of \nmortality during extreme heat events15\u201318. Lower incomes are linked \nto distributional injustices, which include higher likelihood of liv-\ning in inefficient buildings with poor insulation and less efficient \nappliances, which means these homes are more expensive to heat \n(or cool) than others5,14,19\u201321.\nElderly people are at risk of recognition injustices. They require \na narrower band of temperatures for health22 and suffer exacerbated \nmortality during both extreme heat and extreme cold if unable to \nmaintain the appropriate temperatures23\u201326. As a case in point, heat \nwaves in Italy and France in 2003 were associated with higher mor-\ntality rates for elderly individuals17,18. Elderly people also experience \nHealth and financial impacts of demand-side  \nresponse measures differ across sociodemographic \ngroups\nLee V. White\u200a \u200a1,2,3* and Nicole D. Sintov1\nDemand-side response (DSR) measures, which facilitate the integration of high shares of intermittent renewable generation \ninto electric grids, are gaining prominence. DSR measures, such as time-of-use (TOU) rates, charge higher rates during high-\ndemand \u2018on-peak\u2019 times. These rates may disproportionately impact the energy bills and health of vulnerable households, \ndefined as those who face greater energy needs combined with greater social and financial pressures. Here we examine 7,487 \nhouseholds that took part in a randomized control TOU pilot in the southwestern United States. We found that assignment to \nTOU rather than control disproportionately increases bills for households with elderly and disabled occupants, and predicts \nworse health outcomes for households with disabled and ethnic minority occupants than those for non-vulnerable counter-\nparts. These results suggest that vulnerable groups should be considered separately in DSR rate design, and future pilots \nshould seek to determine which designs most effectively avoid exacerbating existing energy injustices or creating new ones.\nNature Energy | VOL 5 | January 2020 | 50\u201360 | www.nature.com/natureenergy\n50\n\nArticles\nNaTURe EneRgy\ngreater difficulty paying for food during seasons with high heating \nor cooling needs, compared to the general population27.\nFamilies with young children experience heightened pressures \nand thus are at risk of recognition injustices. Households with chil-\ndren bear additional costs associated with ensuring that children \nare well fed and healthy5, and thereby face greater competing (child-\ncare) expenses and hence more challenges to meet their energy \nneeds. Children living in energy poverty are more likely to face food \nscarcity and to have health and developmental issues, compared to \nthose not in energy poverty28. Young children are also at higher risk \nof morbidity in extreme heat events25,26.\nPeople with disabilities face higher rates of energy poverty10. \nCompared to the general population, people with disabilities may \nneed more energy to realize a range of essential capabilities24. Illness \nor disability can limit freedom of movement, which raises energy \ncosts due to people being at home more8. Although previous work \nhas not found disability to predict mortality during heat events29, \nfinancial pressure to curtail electricity use may contribute to poorer \nhealth outcomes. The energy needs of those with disabilities vary \ngreatly depending on the individual and disability, but are likely \nto involve a higher energy use. Yet, the needs of individuals with \ndisabilities are often systematically disregarded by decision makers \n(recognition and procedural injustices14).\nFinally, racial and ethnic minorities face procedural injustices in \nthe form of discrimination in areas such as housing, employment \nand credit30, in addition to a lack of informed consent for energy \nprojects, lack of representation in the decision-making and lack of \naccess to information14,31. Distributional injustices arise from these \nprocedural injustices, and racial minorities are more likely to live \nin inefficient housing that necessitates higher energy bills to con-\ntrol indoor temperatures compared to the non-minority counter-\nparts32,33. Racial and ethnic minorities may also face greater health \nimpacts tied to inability to cool homes; the likelihood of death dur-\ning an extreme heat event in the United States is linked to sociode-\nmographic vulnerability, defined partly by ethnic minority and \nLatino immigrant status15,34.\nSome argue that low-income households can save money on static \nTOU rates because their existing patterns place most of their use away \nfrom on-peak times35. However, work that examined critical peak \npricing found that, at baseline, low-income and elderly households \ntend to have a lower-than-average use during on-peak times, whereas \nhouseholds with chronically ill members tend to have a higher on-\npeak use36. Beyond existing use patterns, demand flexibility is con-\nsidered key for households to be able to take financial advantage of \nTOU37. Vulnerable households may face constraints that limit the \nflexibility of electricity use timing, such as poorly insulated homes \nthat prevent the retention of comfortable temperatures if heating or \ncooling systems are turned off5,32,33,38. Some DSR trials found that \nvulnerable (low-income, young children, elderly and/or chronically \nill) households can load shift on par with or to a greater extent than \nnon-vulnerable households36,39\u201341. However, other work found that \nvulnerable (low-income and young children) households load shift \nless and/or have a higher demand during on-peak times and limited \nflexibility compared to non-vulnerable households39,42\u201344. Overall, it is \nunclear how TOU rates will impact vulnerable households in terms \nof bill changes, and responses will probably differ by group. There is a \nrisk that those with higher and less flexible energy needs, such as the \nelderly or those with a disability, will face bill increases.\nEnergy poverty has been associated with a range of nega-\ntive health outcomes, particularly regarding respiratory health45. \nAlthough much work examines the links between energy poverty \nand discomfort, illness and mortality in cold climates11,46\u201348, less \nwork has examined these links related to extreme heat. Prior work \nidentified that greater thermal discomfort associated with energy \npoverty is tied to an increased likelihood of negative impacts on \nboth physical and mental health among households in cold cli-\nmates45,49. The vulnerable groups that we focus on have been found \nto suffer worse outcomes during extreme heat events15\u201318,25,26,34; this \nmay be linked to the inability to access sufficient cooling, but has \nnot yet been examined in the context of energy poverty.\nHere we evaluate the cost and health impacts of TOU among vul-\nnerable (that is, low income, elderly, disability, young children and \nracial/ethnic minority; Table 1 gives the operational definitions) \nversus non-vulnerable households that took part in a randomized \ncontrol TOU pilot in the southwestern United States. We found \nthat, although all households on TOU face bill increases relative to \ncontrols, those vulnerable on the elderly and disability indicators \nface greater bill increases on TOU versus control than their non-\nvulnerable counterparts. Conversely, low-income and Hispanic \nhouseholds face relatively smaller bill increases on TOU versus con-\ntrol than their non-vulnerable counterparts. Households vulnerable \non low-income and disability indicators face worse health outcomes \nregardless of the rate. Relative to their non-vulnerable counterparts, \nhouseholds vulnerable on disability and Hispanic indicators face an \nincreased likelihood of negative health outcomes when assigned to \nTOU, and low-income households face increased discomfort. These \nresults suggest the need to consider vulnerable groups separately, \nand the importance of a careful rate design.\nEffect of TOU on electricity bills\nAll analyses were performed using STATA MP 14.2. The pilot \nincluded two TOU rates and a non-TOU control group. Compared \nTable 1 | Definition of vulnerability on each indicator used in the quantitative analyses\nLow income\nEnrolled in electric utility financial aid programme; some eligible households probably did not complete the enrolment process, \nso this indicator may be imperfect\nElderly\nSomeone over the age of 65 resides in the household\nYoung children\nSomeone under the age of 6 resides in the household\nDisability\nSomeone in the household has a disability or serious medical condition that requires the \u2018home to be cool in the summer\u2019, or \nthat requires them to \u2018use more energy for medical equipment\u2019; although this does not cover the full range of reasons for which a \nperson with a disability may need to use additional energy, available data preclude such a nuanced examination\nHispanic\nRespondent identified as Hispanic. Respondents that identified as non-Hispanic white are considered non-vulnerable. Other \nhouseholds (for example, those that identified as Asian American) could not be considered non-vulnerable in terms of ethnicity/\nrace. Thus, all analyses using the Hispanic vulnerability indicator use a subsample that comprises only Hispanic and non-\nHispanic white participants, and the Hispanic indicator is not included as a control in other analyses\nAfrican American\nRespondent identified as African American. All analyses using the African American vulnerability indicator use a subsample that \ncomprises only African American and non-Hispanic white respondents, and the African American indicator is not included as a \ncontrol in other analyses\nSupplementary Table 1 gives descriptive statistics. The survey question text is provided in Methods.\nNature Energy | VOL 5 | January 2020 | 50\u201360 | www.nature.com/natureenergy\n51\n\nArticles\nNaTURe EneRgy\nto TOU Rate 1 (TOU1), TOU Rate 2 (TOU2) had higher cents per \nkilowatt hour costs on-peak and fewer hours on-peak (Methods).\nThe pilot happened during summer in a hot climate in the south-\nwestern United States; bills were expected to be driven by cooling \nneeds. We used a difference-in-difference-in-differences (triple dif-\nference) approach to examine whether assignment to TOU results \nin greater electricity bill increases for vulnerable versus non-vulner-\nable households. Supplementary Table 2 presents the mean monthly \nbill amounts for all the study groups across the baseline and pilot \nperiods. Each model compares the control to either TOU1 (Table 2) \nor TOU2 (Table 3). There are six models for each rate, one for each \nvulnerability indicator. The mean monthly summer bill amount is \nthe dependent variable, and independent variables are the full set of \ninteraction terms and main effects required for a triple difference \nmodel (Methods).\nBoth TOU rates resulted in bill increases for all participants \n(P\u2009=\u20090.000; Fig. 1, Tables 2 and 3 and Supplementary Note 1). The \ntriple difference term TOU\u00d7Vulnerable\u00d7Pilot in each model gives \nthe estimated effect of the TOU assignment for vulnerable indi-\nviduals during the pilot year (Methods). As expected, households \nvulnerable on the disability indicator assigned to TOU1 (P\u2009=\u20090.011) \nor TOU2 (P\u2009=\u20090.022) and households vulnerable on the elderly \nindicator assigned to TOU2 (P\u2009=\u20090.001) saw greater baseline-to-\npilot-year bill increases compared to non-vulnerable counterparts. \nContrary to expectations, for households vulnerable on the low-\nincome (P\u2009=\u20090.012) and Hispanic indicators (P\u2009=\u20090.014), assignment \nto TOU2 versus control is associated with a smaller increase in \nbills relative to non-vulnerable households. Other groups (African \nAmerican and young children) saw no difference in TOU assign-\nment impacts versus their non-vulnerable counterparts. The \nremaining model terms primarily serve as controls, and are dis-\ncussed in Supplementary Note 1.\nOn-peak energy use reduction\nIn Tables 2 and 3, R2 is consistently <0.10, which suggests that fac-\ntors not included in the model contribute to bill variation. Regional \nfixed effects analysis confirms that changes in on-peak use predict \nbill changes (Supplementary Note 2 and Supplementary Tables 1\u20133).  \nIn a separate triple difference analysis (parallel to the billing analysis \n(Methods)), we found that households vulnerable on the disability \nindicator saw a smaller decrease in on-peak use from baseline to \npilot year when on TOU1 versus control, compared to their non-vul-\nnerable counterparts; no differences were observed for other groups \n(Supplementary Tables 7 and 8; the mean on-peak use reported \nby group and time period is given in Supplementary Table 6).  \nAdditionally, examining reported behavioural efforts to curtail \non-peak air conditioning (AC) use (Table 4), low-income, young \nchildren, Hispanic and African American households reported a \ngreater curtailment compared to their non-vulnerable counter-\nparts, whereas households with elderly members reported less cur-\ntailment; no differences were observed for households with versus \nwithout a disability (Wilcoxon rank sum tests (Methods)).\nTable 2 | Triple difference estimators examining mean monthly bills (US$) for TOU1, by vulnerability group\n(1)\n(2)\n(3)\n(4)\n(5)\n(6)\nLow income\nElderly\nYoung children\nDisability\nHispanic\nAfrican American\nTOU1\u00d7Vulnerable\u00d7Pilot\n1.11 (2.53)\n3.69 (2.72)\n0.72 (3.89)\n8.65* (3.38)\n\u22123.79 (3.29)\n\u22124.43 (5.94)\nTOU1\u00d7Pilot\n13.35*** (1.88)\n12.59*** (1.71)\n13.84*** (1.43)\n12.42*** (1.48)\n15.83*** (1.86)\n15.83*** (1.86)\nVulnerable\u00d7Pilot\n\u22125.18** (1.59)\n0.36 (1.69)\n1.15 (2.52)\n1.25 (2.03)\n2.69 (2.03)\n1.76 (3.98)\nVulnerable\u00d7TOU1\n\u22125.72 (5.28)\n\u22122.40 (6.02)\n3.33 (7.98)\n3.94 (7.57)\n0.02 (7.15)\n2.02 (11.76)\nTOU1\n\u22123.81 (4.06)\n\u22123.34 (3.73)\n\u22124.53 (3.19)\n\u22124.71 (3.23)\n\u22122.76 (4.24)\n\u22122.76 (4.24)\nPilot\n6.18*** (1.18)\n3.94*** (1.06)\n3.92*** (0.88)\n3.85*** (0.93)\n3.32** (1.14)\n3.32** (1.14)\nVulnerable\n\u221252.72*** (3.40)\n\u22123.69 (3.78)\n5.14 (4.95)\n5.11 (4.73)\n\u221223.33*** (4.42)\n\u221234.36*** (7.39)\nR2\n0.08\n0.00\n0.00\n0.01\n0.01\n0.01\nn\n9,738\n9,738\n9,738\n9,738\n7,364\n5,862\nStandard errors in parentheses. *P\u2009<\u20090.05; **P\u2009<\u20090.01; ***P\u2009<\u20090.001.\nTable 3 | Triple difference estimators examining mean monthly bills (US$) for TOU2, by vulnerability group\n(1)\n(2)\n(3)\n(4)\n(5)\n(6)\nLow income\nElderly\nYoung children\nDisability\nHispanic\nAfrican American\nTOU2\u00d7Vulnerable\u00d7Pilot\n\u22126.21* (2.46)\n8.76*** (2.63)\n5.31 (4.18)\n7.49* (3.27)\n\u22127.70* (3.12)\n\u22126.27 (5.81)\nTOU2\u00d7Pilot\n26.45*** (1.92)\n19.96*** (1.65)\n22.98*** (1.36)\n22.18*** (1.43)\n27.92*** (1.84)\n27.92*** (1.84)\nVulnerable\u00d7Pilot\n\u22125.18** (1.59)\n0.36 (1.69)\n1.15 (2.52)\n1.25 (2.03)\n2.69 (2.03)\n1.76 (3.98)\nVulnerable\u00d7TOU2\n3.36 (4.83)\n3.06 (5.34)\n2.58 (7.17)\n4.48 (6.63)\n\u22123.51 (6.21)\n6.99 (10.15)\nTOU2\n\u22128.03* (3.81)\n\u22129.26** (3.44)\n\u22128.22** (2.86)\n\u22128.88** (2.93)\n\u22126.51 (3.83)\n\u22126.51 (3.83)\nPilot\n6.18*** (1.18)\n3.94*** (1.06)\n3.92*** (0.88)\n3.85*** (0.93)\n3.32** (1.14)\n3.32** (1.14)\nVulnerable\n\u221252.72*** (3.40)\n\u22123.69 (3.78)\n5.14 (4.95)\n5.11 (4.73)\n\u221223.33*** (4.42)\n\u221234.36*** (7.39)\nR2\n0.08\n0.01\n0.01\n0.01\n0.02\n0.02\nn\n10,966\n10,966\n10,966\n10,966\n8,414\n6,700\nStandard errors in parentheses. *P\u2009<\u20090.05; **P\u2009<\u20090.01; ***P\u2009<\u20090.001.\nNature Energy | VOL 5 | January 2020 | 50\u201360 | www.nature.com/natureenergy\n52\n\nArticles\nNaTURe EneRgy\nHealth impacts of TOU assignment\nWe tested whether vulnerable versus non-vulnerable house-\nholds report a higher likelihood of seeking medical attention for  \nheat-related reasons, using regional fixed effects regression \ngrouped by climate zone (Methods and Supplementary Tables 9  \nand 10). The dependent variable is the likelihood of seeking  \n\u20135\n0\n5\n10\n15\n20\n25\n30\n35\n40\nControl\nTOU1\nTOU2\nMean change in mean monthly bills ($US)\nRate assignment\na\n\u20135\n0\n5\n10\n15\n20\n25\n30\n35\n40\nControl\nTOU1\nTOU2\nMean change in mean monthly bills ($US)\nRate assignment\nb\n\u20135\n0\n5\n10\n15\n20\n25\n30\n35\n40\nControl\nTOU1\nTOU2\nMean change in mean monthly bills ($US)\nRate assignment\nc\n\u20135\n0\n5\n10\n15\n20\n25\n30\n35\n40\nControl\nTOU1\nTOU2\nMean change in mean monthly bills ($US)\nRate assignment\nd\n\u20135\n0\n5\n10\n15\n20\n25\n30\n35\n40\nControl\nTOU1\nTOU2\nMean change in mean monthly bills ($US)\nRate assignment\ne\n\u20135\n0\n5\n10\n15\n20\n25\n30\n35\n40\nControl\nTOU1\nTOU2\nMean change in mean monthly bills ($US)\nRate assignment\nf\nNon-vulnerable\nVulnerable\nFig. 1 | Mean change in mean monthly summer bills, by vulnerability group and rate assignment. Change in mean monthly summer bills (pilot minus \nbaseline) for the control, TOU1 and TOU2 rate groups for each vulnerability indicator with standard error bars. a, Low income (control, n\u2009=\u20092,865; TOU1, \nn\u2009=\u20092,004; TOU2, n\u2009=\u20092,618). b, Elderly (control, n\u2009=\u20092,865; TOU1, n\u2009=\u20092,004; TOU2, n\u2009=\u20092,618). c, Young children (control, n\u2009=\u20092,865; TOU1, n\u2009=\u20092,004; \nTOU2, n\u2009=\u20092,618). d, Disability (control, n\u2009=\u20092,865; TOU1, n\u2009=\u20092,004, TOU2, n\u2009=\u20092,618). e, Hispanic (control, n\u2009=\u20092,202; TOU1, n\u2009=\u20091,480; TOU2, n\u2009=\u20092,005). \nf, African American (control, n\u2009=\u20091,762; TOU1, n\u2009=\u20091,169; TOU2, n\u2009=\u20091,588). Solid boxes denote a positive triple difference term that indicates a greater \nincrease in bills for vulnerable versus non-vulnerable groups; dashed boxes denote a negative triple difference term that indicates a smaller increase in bills \nfor vulnerable versus non-vulnerable groups (Tables 2 and 3).\nTable 4 | Reported AC curtailment by vulnerable households versus non-vulnerable counterparts\nVulnerability present (mean curtailment)\nVulnerability absent (mean curtailment)\nz score\nP\nda\nLow income\n3.30\n3.10\n\u22124.85\n0.000\n0.16\nElderly\n3.09\n3.24\n3.73\n0.000\n0.12\nYoung children\n3.33\n3.16\n\u22122.80\n0.005\n0.13\nDisability\n3.12\n3.20\n1.72\n0.086\n0.06\nHispanic\n3.29\n3.02\n\u22125.54\n0.000\n0.22\nAfrican American\n3.45\n3.02\n\u22125.27\n0.000\n0.35\nMain sample, n\u2009=\u20094,129; Hispanic subsample, n\u2009=\u20093,108; African American subsample, n\u2009=\u20092,459). Bold font shows the group that made greater efforts to curtail for each behaviour. aCohen\u2019s d effect size.\nNature Energy | VOL 5 | January 2020 | 50\u201360 | www.nature.com/natureenergy\n53\n\nArticles\nNaTURe EneRgy\nmedical attention. Independent variables in all models are rate \nassignment and vulnerability indicators for low income, elderly, \nyoung children and disability. Hispanic and African American \nindicators appear only in models that use subsamples. Interaction \nterms Vulnerable\u00d7TOU are introduced individually in subsequent \nmodels to test group-specific effects of TOU assignment. The \nfrequency of reported discomfort (expected precursor to more \nsevere problems), presence of AC (expected to impact indoor \ntemperature), change in on-peak use (expected to impact indoor \ntemperature) and home ownership (expected to impact indoor \ntemperature via a greater control by homeowners over efficiency \nmeasures, such as insulation) are included as control variables in \nall models. A subset of the models is presented in Table 5, which \ncomprises all the main effects models and models with significant \nvulnerability interaction terms at P\u2009<\u20090.05.\nMain effects indicate that households vulnerable on low-\nincome (P\u2009=\u20090.000), disability (P\u2009=\u20090.000) and Hispanic indicators \n(0.001\u2009<\u2009P\u2009<\u20090.011) are more likely to seek medical attention for \nheat-related reasons (Table 5). TOU assignment alone does not pre-\ndict the likelihood of seeking medical attention. A greater frequency \nof discomfort predicts a higher likelihood of needing medical atten-\ntion for heat-related reasons in all models (P\u2009=\u20090.000), whereas home \nownership predicts a lower likelihood of needing medical attention \nin models (1), (2), (3) and (4) (0.011\u2009<\u2009P\u2009<\u20090.032) and a reduction in \non-peak use predicts a higher likelihood of needing medical atten-\ntion in models (7), (8) and (9) (0.011\u2009<\u2009P\u2009<\u20090.014) (Table 5).\nConsidering the interaction terms, TOU versus control assign-\nment significantly alters the likelihood of seeking medical attention \namong households with either young children (P\u2009=\u20090.045) or dis-\nabled members assigned to TOU1 (P\u2009=\u20090.030), and Hispanic house-\nholds assigned to TOU2 (P\u2009=\u20090.032). For significant interaction \nterms, we performed post hoc tests with the conservative Scheff\u00e9\u2019s \nadjustment applied to significance testing of the contrast between \npairwise comparisons (Table 6). Among families with young chil-\ndren, assignment to TOU1 versus control correlates with a lower \nlikelihood of needing medical attention. For households vulnerable \non the disability indicator, assignment to TOU1 is associated with \na higher likelihood of seeking medical attention relative to non-\nvulnerable households on TOU1 (and non-vulnerable households \nassigned to control rate). Hispanic households assigned to TOU2 \nface higher a likelihood of needing medical attention than non-His-\npanic white households on TOU2.\nDiscomfort in vulnerable versus non-vulnerable households\nWe tested whether vulnerable versus non-vulnerable households \nreport more frequently experiencing discomfort due to homes \nbeing too hot, using regional fixed effects regression grouped by \nclimate zone (Methods and Supplementary Tables 11 and 12). The \ndependent variable is the frequency of discomfort experienced \nwhile trying to save money on electricity. Independent variables are \nrate assignment, vulnerability indicators and Vulnerability\u00d7TOU \ninteraction terms. The models examine the interaction terms for \neach vulnerable group separately. The presence of AC, change in \non-peak use and home ownership are included as control variables \nin all the models.\nRegardless of the rate assignment, low-income (P\u2009=\u20090.000), dis-\nability (P\u2009=\u20090.000) and Hispanic indicators (P\u2009=\u20090.024, Hispanic for \nTOU1 models only) predict more frequent discomfort, whereas \nthe elderly indicator (P\u2009<\u20090.042) predicts less frequent discomfort. \nTwo interaction terms are significant: Low income\u00d7TOU1 and \nAfrican American\u00d7TOU1. For these terms, we performed post hoc \ntests with Scheff\u00e9\u2019s adjustment (Table 7). Low-income households \nassigned to TOU face a higher discomfort than their non-vulnera-\nble counterparts assigned to TOU (this is also true for the control \ngroup). No significant differences were observed for the African \nAmerican\u00d7TOU interaction term.\nDiscussion\nThe results suggest distributional, procedural and recognition injus-\ntices that differ across groups, which highlights the importance of \nconsidering specific subpopulations in the design and rollout of DSR \nprogrammes. The greater cost increases experienced by households \nvulnerable on the disability and elderly indicators assigned to TOU, \nrelative to their non-vulnerable counterparts, suggest recognition \ninjustices14. Cost increases faced by these households probably stem \nfrom inability to shift use times, as evidenced in our findings that \nhouseholds vulnerable on the disability indicator reduced on-peak \nuse less than their non-vulnerable counterparts, and households \nvulnerable on both disability and elderly indicators reported less AC \ncurtailment compared to their non-vulnerable counterparts. These \ngroups may be constrained in load shifting due to being home-\nbound and having a greater reliance on energy for medical equip-\nment, temperature control and completing daily tasks8,23\u201326.\nHouseholds vulnerable on low-income and disability indicators \nface worse health and comfort outcomes than the outcomes faced \nby non-vulnerable counterparts, regardless of rate assignment, \nwhich probably stems from ongoing procedural and distributional \ninjustices. However, TOU appears to widen the difference in health \noutcomes between those vulnerable on the disability indicator and \ntheir non-vulnerable counterparts, which suggests recognition \ninjustices given that this group already struggles to keep homes cool \nunder current distributional injustices.\nTOU similarly correlates with worse health outcomes for \nHispanic households, relative to their non-vulnerable counterparts. \nHispanic households reported greater curtailment of AC compared \nto non-Hispanic white households, and it is possible that this con-\ntributed to negative health outcomes. Hispanic groups are more \nlikely to experience heat distress in extreme heat events15,34, which \nraises concerns that further distributional injustice could worsen \nthe differentials in mortality rates. Future research should evaluate \nwhether this outcome is linked to inefficient homes that limit the \nability to keep cool33.\nOur finding that TOU1 only shows a differential cost effect \nfor those with disabilities, whereas TOU2 shows differential cost \neffects for those vulnerable on the low-income, elderly, disability \nand Hispanic indicators, suggests that the design of TOU rates is \nimportant in predicting outcomes for energy-vulnerable popula-\ntions. Specifically, the length and kilowatt hour expense of on-peak \ntimes appear to play an important role in the group-differentiated \nfinancial impacts of TOU.\nGiven that DSR rates will probably be needed to integrate higher \nshares of non-dispatchable generation, it is important for pilots to \ncontinue trialling multiple rate designs and evaluating the impacts \non vulnerable populations, with the goal of identifying rate designs \nin each locale that meet energy integration needs without worsen-\ning or creating energy injustices. The TOU rates examined here \nincreased electricity bills for all groups. By definition, vulnerable \ngroups are less able to bear cost increases than their non-vulner-\nable counterparts8, which suggests that switching to the TOU rates \nconsidered in this study probably increased hardships such as the \n\u2018heat or eat\u2019 dilemma. However, compared to their non-vulnera-\nble counterparts, only two groups (disability and elderly) experi-\nenced greater cost increases, whereas two groups (low income and \nHispanic) experienced lower cost increases.\nTOU rate design should aim to be cost neutral, and studies of \nother TOU rates have found evidence that some rate designs can, \nindeed, achieve cost neutrality across the general population50. If \nthe TOU rate in our study had achieved cost neutrality across the \ngeneral population, rather than causing increases across the board, \nit is possible that only some of the vulnerable groups examined \n(elderly and disability) would have been worse off, whereas some \n(low income and Hispanic) may have been better off. More exten-\nsive examination of potential rate designs is needed to understand if \nNature Energy | VOL 5 | January 2020 | 50\u201360 | www.nature.com/natureenergy\n54\n\nArticles\nNaTURe EneRgy\nTable 5 | Regional fixed effects logit model grouped by climate zone, predicting likelihood of a household member needing medical attention\n(1)\n(2)\n(3)\n(4)\n(5)\n(6)\n(7)\n(8)\n(9)\nMain effects, \nTOU1\nYoung children in \ninteraction, TOU1\nDisability in \ninteraction, TOU1\nMain effects \n(Hispanic), TOU1\nMain effects \n(African \nAmerican), TOU1\nMain effects, \nTOU2\nMain effects \n(Hispanic), TOU2\nHispanic in \ninteraction, TOU2\nMain effects \n(African \nAmerican), TOU2\nVulnerable\u00d7TOU\n\u22120.71* (0.35)\n0.60* (0.28)\n0.38* (0.18)\nAssigned to TOU\n\u22120.07 (0.14)\n0.06 (0.19)\n\u22120.38** (0.12)\n0.02 (0.21)\n0.05 (0.26)\n\u22120.06 (0.10)\n\u22120.02 (0.15)\n\u22120.20 (0.15)\n\u22120.29* (0.14)\nLow income\n0.57*** (0.14)\n0.57*** (0.14)\n0.59*** (0.14)\n0.58*** (0.16)\n0.69*** (0.17)\n0.70*** (0.14)\n0.82*** (0.18)\n0.81*** (0.18)\n0.99*** (0.26)\nElderly\n0.26 (0.16)\n0.27 (0.16)\n0.26 (0.16)\n0.22 (0.19)\n0.24 (0.23)\n0.09 (0.09)\n0.17 (0.09)\n0.17 (0.09)\n0.35 (0.22)\nYoung children\n0.33 (0.19)\n0.59** (0.21)\n0.33 (0.18)\n0.18 (0.24)\n\u22120.12* (0.05)\n0.47 (0.25)\n0.36 (0.35)\n0.37 (0.35)\n0.11 (0.38)\nDisability\n1.58*** (0.16)\n1.59*** (0.16)\n1.36*** (0.15)\n1.72*** (0.19)\n1.69*** (0.18)\n1.40*** (0.12)\n1.42*** (0.19)\n1.41*** (0.19)\n1.17*** (0.16)\nHispanic\n0.36 (0.14)\n0.49** (0.15)\n0.32 (0.21)\nAfrican American\n0.42 (0.23)\n0.30 (0.18)\nPresence of AC\n0.14 (0.21)\n0.14 (0.21)\n0.13 (0.22)\n0.43* (0.22)\n0.62 (0.76)\n0.18 (0.39)\n0.01 (0.55)\n0.01 (0.56)\n0.40 (0.60)\nFrequency of discomfort \ntrying to save money\n0.80*** (0.05)\n0.80*** (0.05)\n0.80*** (0.05)\n0.76*** (0.06)\n0.82*** (0.05)\n0.85*** (0.06)\n0.92*** (0.05)\n0.92*** (0.06)\n0.88*** (0.10)\nChange in on-peak use \n(kWh, daily average)\n0.06 (0.04)\n0.06 (0.04)\n0.06 (0.04)\n0.05 (0.03)\n0.08 (0.05)\n0.07+ (0.04)\n0.11* (0.04)\n0.11* (0.04)\n0.15* (0.06)\nHome owned\n\u22120.35* (0.15)\n\u22120.35* (0.15)\n\u22120.33* (0.16)\n\u22120.38* (0.15)\n\u22120.36 (0.30)\n\u22120.31 (0.22)\n\u22120.25 (0.29)\n\u22120.25 (0.28)\n\u22120.30 (0.45)\nPseudo R2\n0.20\n0.20\n0.20\n0.21\n0.22\n0.19\n0.22\n0.22\n0.20\nn\n4,869\n4,869\n4,869\n3,682\n2,931\n5,483\n4,207\n4,207\n3297\nStandard errors in parentheses, clustered by climate zone. *P\u2009<\u20090.05; **P\u2009<\u20090.01; ***P\u2009<\u20090.001.\nNature Energy | VOL 5 | January 2020 | 50\u201360 | www.nature.com/natureenergy\n55\n\nArticles\nNaTURe EneRgy\nthis would be borne out in practice. Given that those vulnerable on \ndisability and elderly indicators have a greater need for affordable \nenergy compared to the general population, policy or rate design \ninterventions should ensure that energy costs are low enough for \nthese groups to maintain their health on TOU or other DSR rates \nand still be able to afford other necessities24.\nAs opt-out DSR programmes spread, it is important that the \ncosts of each DSR rate relative to those of other offered rates are \nclearly communicated and that the opt-out procedures be made \nclear to all, particularly vulnerable groups who are at risk of unfa-\nvourable outcomes on DSR. To aid households in evaluating bene-\nfits and burdens of competing rates, it is important to communicate \ncost information in a way that minimizes cognitive burden51,52. \nUsing heuristics to address common misperceptions may improve \nhousehold understanding of energy use53. Given prior findings that \nhouseholds often base decisions regarding TOU enrolment on per-\nceived financial savings, but misperceive the extent of actual finan-\ncial savings54, further testing of an ideal choice architecture and \ninformation presentation for DSR is critical to facilitate all house-\nholds making informed decisions about their electricity rates.\nMore broadly, our findings regarding the worse health and \ncomfort outcomes for households vulnerable on low-income and \ndisability indicators regardless of rate assignment suggest that \nenergy-vulnerable groups in hot climates globally should be the \nfocus of future research. These findings also suggest a need for pol-\nicy intervention to support more affordable cooling, regardless of \nfuture DSR rollout. Cooling centres may help reduce discomfort, \nbut often operate only during weekday business hours, so are dis-\nruptive to family routines and provide only part-time relief55. Thus, \nwe recommend other measures, such as improving building and \nappliance energy efficiency, and carefully designed rates. Efficiency \nimprovement programmes can offer large cost savings and reduce \nemissions, as well as decrease discomfort56,57. Future research should \ndirectly consider the extent to which housing energy efficiency lim-\nits the ability to control bills on DSR rates such as TOU, with a view \nto informing the design of complementary policies to address dis-\ntributional injustices.\nOur results should be viewed in the light of several limitations. \nFirst, our sample comprised individuals who opted into the pilot. \nThus, our sample may be less risk averse than the general popula-\ntion58,59, or may have a greater expectation of monetary savings on \nTOU60. Second, the results may not generalize to all populations. \nSurvey completers had higher mean baseline use and larger mean \nbill increases than partial completers. Control group members were \nmore likely to complete the survey than those assigned to TOU1. \nSeveral vulnerable groups (low income, elderly, disability, Hispanic \nand African American) were less likely to complete the survey com-\npared to non-vulnerable counterparts. Possibly survey completers \nhad fewer time pressures, which suggests conservative estimates \nof the impact on vulnerable groups. Third, our indicator for low-\nincome households relied on enrolment in a utility programme, \nand thus probably underestimates the low-income households \nTable 6 | Post hoc tests of Vulnerable\u00d7TOU interaction terms on the need for medical attention\nControl, vulnerable versus\nTOU, non-vulnerable versus\nTOU, vulnerable versus\nYoung children\u00d7TOU1\n Control, non-vulnerable\n0.59* (0.21)\n0.06 (0.19)\n\u22120.6 (0.25)\n Control, vulnerable\n\u22120.53* (0.16)\n\u22120.65* (0.21)\n TOU, non-vulnerable\n\u22120.12 (0.30)\nDisability\u00d7TOU\n Control, non-vulnerable\n1.38* (0.16)\n\u22120.08 (0.11)\n1.34* (0.20)\n Control, vulnerable\n\u22121.45* (0.10)\n\u22120.03 (0.19)\n TOU, non-vulnerable\n1.43* (0.17)\nHispanic\u00d7TOU2\n Control, non-vulnerable\n0.32 (0.21)\n\u22120.10 (0.15)\n0.50* (0.15)\n Control, vulnerable\n\u22120.52 (0.27)\n0.18 (0.21)\n TOU, non-vulnerable\n0.70* (0.14)\nPairwise comparison, contrast. Standard error in parentheses. *Scheff\u00e9 test significant at the 95% level.\nTable 7 | Post hoc tests of Vulnerable\u00d7TOU interaction terms on discomfort\nControl, vulnerable versus\nTOU, non-vulnerable versus\nTOU, vulnerable versus\nLow income\u00d7TOU1\n Control, non-vulnerable\n0.32* (0.48)\n0.06 (0.02)\n0.28* (0.05)\n Control, vulnerable\n\u22120.26* (0.04)\n\u22120.03 (0.04)\n TOU, non-vulnerable\n0.23* (0.04)\nAfrican American\u00d7TOU\n Control, non-vulnerable\n\u22120.17 (0.10)\n\u22120.02 (0.03)\n0.08 (0.12)\n Control, vulnerable\n0.15 (0.12)\n0.26 (0.09)\n TOU, non-vulnerable\n0.11 (0.13)\nPairwise comparison, contrast. Standard error in parentheses. *Scheff\u00e9 test significant at the 95% level.\nNature Energy | VOL 5 | January 2020 | 50\u201360 | www.nature.com/natureenergy\n56\n\nArticles\nNaTURe EneRgy\nsampled and has an imperfect separation between vulnerable and \nnon-vulnerable households; the associated coefficients should be \ninterpreted with caution. Fourth, we examined vulnerability indica-\ntors in isolation, and hence the results do not capture differential \nimpacts faced by those bearing a double burden, such as being both \nelderly and low income.\nTOU raised the cost of energy for all households in our study, \nbut some vulnerable households (elderly and disability) face greater \nbill increases on TOU compared to their non-vulnerable counter-\nparts. Households vulnerable on low-income and disability indica-\ntors also face more discomfort and more heat-related medical issues \nregardless of rate assignment, which raises general concerns about \nthe health impacts of energy poverty in hot climates. TOU widens \nthe discomfort gap (low income) and increases the likelihood of \nseeking medical attention (disability and Hispanic) of some vulner-\nable groups relative to non-vulnerable counterparts on TOU. Rate \ndesign plays an important role in the impact of price-based DSR \nmeasures on vulnerable households, and future pilots should con-\ntinue to examine multiple potential rate designs to determine which \ndesigns most effectively avoid exacerbating existing energy injus-\ntices or creating new ones.\nMethods\nEthics statement. The University of Southern California\u2019s University Park \nInstitutional Review Board reviewed and approved this research, and granted a \nwaiver of informed consent.\nParticipants. Data are from households that participated in a pilot programme \nadministered by a southwestern US electric utility. The utility sent invitations by \ndirect mail and email soliciting opt in to the 2016 TOU pilot to roughly 197,000 \nhouseholds, 14% of which opted in. Some households that accepted the offer were \nnot enrolled because they were ineligible (for example, were already participating \nin a special rate programme). The utility randomly assigned 21,534 households \nto either TOU1 (n\u2009=\u20094,709), TOU2 (n\u2009=\u20096,365), TOU3 (n\u2009=\u20093,746) or the control \ngroup that opted in to TOU but was not placed on a TOU rate (n\u2009=\u20096,714). Low-\nincome households and those with elderly members were deliberately oversampled \nfor TOU2. TOU3 was not fully rolled out by the start of the study period, due to \nadditional complexities unique to the rate, and so could not be included in the \npresent study; we consider only TOU1, TOU2 and the control group.\nTOU pilot. Households assigned to a TOU rate were shifted to this new rate in \nJune or July of 2016 and remained on these rates for a full year. The period of the \npilot covered by this study occurred in the summer, specifically the months July\u2013\nSeptember 2016. Owing to the geographical region, the weather would have been \nwarm to hot and mostly lacking precipitation for the majority of the sample.\nAfter the rate assignment, participants received information letters. Those \nin the control group received a welcome letter informing them that they would \nremain on their current rate. TOU participants received a letter containing \ninformation on their TOU rate plan and bill protection (if customers paid more on \nTOU at the end of the 12-month pilot than they would have under their previous \nplan, the utility would credit back the difference after the pilot ended). They also \nreceived TOU time-period stickers, conservation-reminder stickers and door \nhangers with recommended seasonal thermostat settings.\nTOU1 and TOU2 on-peak times covered different hours depending on \nweekend versus weekday. On-peak hours were in the evening. Cost per kilowatt \nhour varied depending on the rate and season. Summer rates per kilowatt hour \nfor TOU1 were c23 for super off-peak, c27.61 for off-peak and c34.51 for on-peak. \nSummer rates per kilowatt hour for TOU2 were c17.33 for super off-peak, c29.3  \nfor off-peak and c53.26 for on-peak. TOU1 had six hours on-peak from 14:00 to \n20:00 for summer weekdays and no on-peak times at weekends. TOU2 had three \nhours on-peak from 17:00 to 20:00 for summer weekdays and no on-peak times  \nat weekends.\nHouseholds could earn up to US$200 as compensation for their participation, \ngiven as bill credits; US$100 at enrolment and US$50 after completing each of two \nsurveys. The second survey, which we do not have data for due to the timing of our \ndata request from the utility, was administered in the summer of 2017.\nSurvey. Customers were first surveyed between October and December 2016. At \nthis point, TOU participants had 3\u20136 months\u2019 experience with TOU rates, solely or \nprimarily in the summer. Survey response rates were 82% overall, out of the 18,747 \nhouseholds that remained in the pilot by December 2016 after being enrolled and \nnot being dropped out of the pilot due to relocating, ineligibility or choosing to \nleave. For the full sample examined in this manuscript (n\u2009=\u20097,487), 85% responded \nby email, 11% by mail and 4% by phone.\nElectricity use data. Our participants experienced the TOU pilot during the \nsummer months. Thus, the analyses use hourly electricity use data for each \nhousehold only during the summer months (July, August and September) for \nthe baseline years 2014 and 2015, and for the pilot year of 2016. Summer hourly \nconsumption data were used to form the \u2018on-peak use\u2019 variable (kilowatt hours, \ndaily mean).\nIn forming the \u2018on-peak use\u2019 variable, we took into account the different lengths \nof on-peak time, that is, TOU1 on-peak sums the use during the 6\u2009h on-peak, and \nTOU2 on-peak sums the use during the 3\u2009h on-peak. Control group households \nwere all coded for both hypothetical TOU1 and TOU2 rate structures, and on-peak \nuse was defined as use that happened during the hours designated as on-peak \nby that rate. In each analysis, the full control group is compared to either TOU1 \nor TOU2, using the corresponding on-peak hour coding for control groups to \ngenerate the change in on-peak use variable used for analysis. \u2018Change in on-peak \nuse\u2019 between the baseline and TOU pilot years is taken as the TOU pilot daily \non-peak mean use minus the mean baseline on-peak use. The following equation \ndescribes the change in on-peak use calculation: \u0394U\u2009=\u2009[U2016\u2009\u2013\u2009(U2014\u2009+\u2009U2015)/2], \nwhere U represents the mean daily on-peak use in kilowatt hours in each year.\nElectricity bills. As for use, the billing data for each household were examined only \nfor July, August and September for the baseline years 2014 and 2015, and for the \npilot year 2016. Bill amounts used for the analysis were actual bills that customers \nreceived. That is, they reflect the true amount customers paid each month. All the \nhouseholds kept for analysis had billing data for at least two of the three summer \nmonths. Mean monthly bill size was taken as the mean of bills across July, August \nand September for a given year. A baseline bill variable was created by taking \nthe mean of the 2014 and 2015 bills, and the pilot bill was the 2016 mean bill. A \n\u2018change in bills\u2019 variable was created by subtracting the baseline bills from the pilot \nbills using the equation: \u0394B\u2009=\u2009[B2016\u2009\u2013\u2009(B2014\u2009+\u2009B2015)/2], where B represents mean bill \namounts in US dollars each year.\nClimate zone. Climate zones are defined by a government agency, and are matched \nto each household based on the household\u2019s location. The government agency \nbases climate zone boundaries on the household energy use expected for heating \nand cooling, local temperature and local weather, among other factors. There are \neight climate zones covered by households in our sample.\nSurvey and vulnerability measures. The survey assessed a number of demographic \ncharacteristics, which included respondent age and ethnicity/race, and household \nmember disabilities. It also assessed homeownership status and educational \nattainment, the former of which is included in regression models due to the \nexpectations that homeowners would more easily be able to upgrade appliances \nand building insulation to increase energy efficiency. Homeownership was coded \nas a dichotomous variable with 1 indicating homeownership and 0 indicating \notherwise.\nWe defined vulnerability indicators as follows:\n\u2022 \nLow income. Enrolment in an electric utility financial aid programme (that \ngives households discounts on electricity bills if income falls below certain \nlimits based on the number of household members) serves as an indicator of \nlow income, with those enrolled in a financial aid programme coded 1 and \nothers coded 0. Enrolment status was provided by the utility.\n\u2022 \nElderly. The survey assessed respondent birth year with an open entry. It \nadditionally asked: \u201cHow many people in each of the following age categories \nlived in your home this summer, not including yourself?\u201d, with response \ncategories including \u201cBetween 65 and 74 years old\u201d, \u201cBetween 75 and 84 years \nold\u201d and \u201c85 years or older\u201d. Responses to this question and birth year were \naggregated to determine whether anyone over 65 resided in the household. If \nsomeone over 65 years old lived in the household, that household was coded 1. \nOtherwise, it was coded 0.\n\u2022 \nYoung children. Respondents were asked \u201cHow many people in each of the \nfollowing age categories lived in your home this summer, not including your-\nself?\u201d with the youngest category being \u201cUnder 6 years old\u201d. Households that \nhad at least one member under 6 years old were coded \u20181\u2019 for young children, \notherwise households were coded 0.\n\u2022 \nDisability. If someone answered yes to either \u201cDoes anyone in your household \nhave a disability or serious medical condition that requires your home to be \ncool in the summer?\u201d or \u201cDoes anyone in your household have a disability or \nserious medical condition that requires them to use more energy for medical \nequipment?\u201d, that household was coded 1. Otherwise, it was coded 0.\n\u2022 \nHispanic. Race and ethnicity were assessed using the question \u201cWhich cat-\negories describe you?\u201d, in response to which households could mark as many \noptions as they chose to out of the list provided. If a respondent identified as \n\u201cHispanic, Latino, or Spanish origin\u201d, their household was coded 1 for the  \nHispanic indicator. If a household marked \u201cWhite\u201d as a category that \ndescribed them, and did not mark any other category (including, but not  \nlimited to, \u201cHispanic, Latino, or Spanish origin\u201d), then they were coded as 0. \nThat is, households were only coded 0 if they identified as white alone.  \nOther households (for example, those that identified as African American, \nAsian American, American Indian and so on) could not be considered  \nNature Energy | VOL 5 | January 2020 | 50\u201360 | www.nature.com/natureenergy\n57\n\nArticles\nNaTURe EneRgy\nnon-vulnerable in terms of ethnicity/race, so were treated as missing. Thus, all \nanalyses using the Hispanic vulnerability indicator use a subsample (n\u2009=\u20094,925), \nand the Hispanic indicator is not included as a control in other analyses.\n\u2022 \nAfrican American. Coding follows the procedure described for the Hispanic \nindicator. If the respondent identified as \u201cAfrican American\u201d, their household was \ncoded 1 for the African American indicator. If a household marked \u201cWhite\u201d as a \ncategory that described them, and did not mark any other category (including, \nbut not limited to, \u201cAfrican American\u201d), then they were coded as 0. All analyses \nusing the African American vulnerability indicator use a subsample (n\u2009=\u20093,970), \nand the African American indicator is not included as a control in other analyses.\nAC curtailment and ownership. AC curtailment was assessed by asking respondents, \n\u201cSince the beginning of this summer, how often, if at all, did you take the \nfollowing actions to reduce your household\u2019s electricity use in the afternoons and \nevenings?\u2014Turned off air conditioning\u201d on a 5-point Likert scale where 1\u2009=\u2009Never \nand 5\u2009=\u2009Always, with an additional option of 6\u2009=\u2009Not applicable. We retained those \nwho answered \u201cNot applicable\u201d within the sample, but coded these respondents as \nmissing when conducting analyses that make use of the curtailment scale.\nThe survey assessed both the reported behavioural curtailment of AC and the \npresence of AC technology. Households were considered to have AC if they had \ncentral AC, window AC, evaporative coolers or heat pumps (which are capable of \nproviding AC). Some households gave conflicting answers to their curtailment of \nAC and their ownership of AC technology; for example, some households rated \ntheir curtailment of AC rather than selecting \u201cNot applicable\u201d, but indicated that \nthey did not own AC technology. Households were coded 0 for AC ownership if \nthey chose \u201cNot applicable\u201d for AC curtailment and also indicated they owned \nno AC technology, and were coded 1 for AC ownership if they rated their AC \ncurtailment (that is, did not choose \u201cNot applicable\u201d) and additionally indicated \nthat they owned a form of AC technology. Households that gave conflicting \nanswers were dropped (n\u2009=\u2009850).\nMeasures for discomfort and medical needs. Discomfort was assessed with responses \nto the question \u201cSince June 2016, how often, if ever, were you or any members of \nyour household uncomfortably hot inside your home because you were trying to \nsave money on your electricity bill?\u201d on a 5-point Likert scale with 1\u2009=\u2009Never and \n5\u2009=\u2009Always.\nThe need for medical attention was assessed with responses to the question \n\u201cSince June 2016, about how many times, if ever, did you or any members of your \nhousehold need medical attention because it was too hot inside your home?\u201d, \nwith respondents able to choose between options of \u201cNever\u201d, \u201c1\u201d, \u201c2\u201d, \u2026 to \u201cmore \nthan 10\u201d. A dichotomous variable was created, with respondents coded 1 if they \nanswered that they had needed medical attention at least once, and 0 otherwise.\nDropped participants. By December 2016, 2,787 households had dropped out of \nthe pilot due to relocating, ineligibility or choosing to drop out, which left 18,747 \nhouseholds enrolled. Of these, 16,181 households responded to the survey. Before \nreceipt by the authors, the utility removed respondents who answered 5.4% or \nless of the survey items. Respondents were also removed if they provided the \nsame rating for all items across any of the three multi-item measures in the survey \n(for example, if a participant gave ratings of \u20184\u2019 to all the items in one multi-item \nquestion). Additionally, respondents were removed if they selected all the items in \na \u2018select-all-that-apply\u2019 question in which some categories were mutually exclusive, \nfor example, if when asked \u201cWhat kept you from shifting use in the evening\u201d \nrespondents selected both \u201cNothing keeps me from shifting my use\u201d and \u201cMy \nschedule doesn\u2019t allow me to reduce my usage\u201d.\nThis yielded a sample of n\u2009=\u200916,073 households, with 5,198 in the control \ngroup, 3,522 on TOU1 and 4,593 on TOU2; the 2,760 households on TOU3 \nwere not used, which left an initial sample of n\u2009=\u200913,313 households. Additional \nhouseholds were then dropped for the reasons below.\nFirst, households that were missing electricity use data on any days in July, \nAugust and September in 2014, 2015 or 2016 were dropped. A total of n\u2009=\u20091,339 \nhouseholds were dropped due to incomplete use data; the missing values were \npredominantly clustered across several days or weeks at the beginning of the \nrecorded period, which indicates that either no account was established for that \naddress (that is, residents moved in during the study period) or the house did not \nhave a smart meter at the beginning of the time period.\nSecond, billing outliers were removed. Customers with baseline or pilot period \nuse below the 1st percentile or above the 99th percentile were dropped from the \nsample (n\u2009=\u2009341 households).\nThird, households with incomplete survey data were dropped. Households \nwere dropped if they had not answered the AC curtailment question, if they had \nnot answered the question assessing the presence of AC, if they had not answered \neach vulnerability indicator (excepting the race/ethnicity indicators, which used \nsubsamples), if they had not answered both the discomfort and the medical \nattention questions and if they had not included an answer to homeownership \n(n\u2009=\u20093,296).\nFinally, households that gave conflicting answers to their curtailment of AC \nand their ownership of AC technology were dropped (n\u2009=\u2009850). Our final sample \ncomprised 7,487 respondents.\nDropout analyses. The utility with which we partnered provided data only for \nthose who at least partially completed the survey, so we are unable to evaluate \npresurvey opt outs. We used a two-tailed t-test to understand whether there was \na systematic difference regarding the change in bills from baseline to pilot year \nbetween households that completed the survey and those that did not complete \nthe survey, restricting the sample only to those who had complete billing and \nuse data (n\u2009=\u200911,633); among those with incomplete billing and use data, this \nincompleteness was due to factors beyond the households\u2019 control, such as the \ninstallation of smart meters. We found that survey completers (n\u2009=\u20097,487) saw \nlarger bill increases (baseline to pilot, simple difference) than non-completers \n(n\u2009=\u20094,146); Mcomplete\u2009=\u200916.09, Mnon-complete\u2009=\u200913.35, P\u2009=\u20090.003, d\u2009=\u20090.06. We additionally \nused a two-tailed t-test to examine the systematic difference in change in on-peak \nuse from baseline to pilot (n\u2009=\u200911,633) for both TOU1 peak-time survey completers \n(TOU1\u2009+\u2009control, n\u2009=\u20094,869) versus non-completers (TOU1\u2009+\u2009control, n\u2009=\u20092,688) \nand TOU2 peak-time survey completers (TOU2\u2009+\u2009control, n\u2009=\u20095,483) versus \nnon-completers (TOU2\u2009+\u2009control, n\u2009=\u20093,107), and found no difference between \ncompleters and non-completers in usage reduction (Mcomplete,TOU1\u2009=\u2009\u22120.11, Mnon-\ncomplete,TOU1\u2009=\u2009\u22120.11, P\u2009=\u20090.91, d\u2009=\u20090.002; Mcomplete,TOU2\u2009=\u2009\u22120.11, Mnon-complete,TOU2\u2009=\u2009\u22120.09, \nP\u2009=\u20090.56, d\u2009=\u20090.01). We finally examined differences in baseline use (n\u2009=\u200911,633) \nbetween those who completed the survey (n\u2009=\u20097,487) and those who did not \n(n\u2009=\u20094,146), and found that completers had a higher daily baseline average use \n(Mcomplete\u2009=\u200924.98, Mnon-complete\u2009=\u200922.43, P\u2009=\u20090.000, d\u2009=\u20090.18). In summary, those who \ncompleted the survey had higher baseline use and larger bill increases during the \npilot than the non-completers. The effect size of the baseline difference in daily \naverage use between the survey completers and non-completers is large, whereas \nthe effect size difference in billing is small.\nWe further considered dropout across different conditions due to survey \nmissing data: of 4,514 in the control group, 37% were non-completers, compared \nto 34% of the 3,043 on TOU1 and 36% of the 4,076 on TOU2. A simple logit model \nwas used to estimate the correlation of group assignment with the likelihood of \nhaving incomplete data. We first considered all three rate types using dummy \nvariables (with the control as baseline), and found that, compared to the control, \nTOU1 is associated with a higher likelihood of incomplete data (P\u2009<\u20090.034), but \nTOU2 is not (P\u2009<\u20090.464). We then considered only TOU1 versus TOU2, with \nTOU1 as the baseline, and found no difference in association with non-completion \n(P\u2009<\u20090.155). Finally, we used a simple logit to consider whether vulnerability \npredicts non-completion of the survey. We found that low-income (P\u2009<\u20090.001), \nelderly (P\u2009<\u20090.001), disability (P\u2009<\u20090.001), Hispanic (P\u2009<\u20090.001) and African \nAmerican (P\u2009<\u20090.011) groups were all less likely to complete the survey compared \nto the respective non-vulnerable counterparts.\nDifference-in-difference-in-differences analyses. The difference-in-difference-in-\ndifferences billing analyses are described by the equation:\nBst \u00bc \u03b20 \u00fe \u03b21TreatR1s \u00fe \u03b22Postt\n\u00fe\u03b23Vulnerables \u00fe \u03b24 Treats \u00b4 Postt\n\u00f0\n\u00de\n\u00fe\u03b25 Vulnerables \u00b4 Postt\n\u00f0\n\u00de \u00fe \u03b26 Vulnerables \u00b4 Treats\n\u00f0\n\u00de\n\u00fe\u03b27 Vulnerables \u00b4 Treats \u00b4 Postt\n\u00f0\n\u00de \u00fe \u03b5st\nwhere Bst is the mean monthly bill amount (US dollars), TreatR1s is a dichotomous \nvariable set to 1 if the household was on TOU1 and 0 if the household was in the \ncontrol group, Postt is a dichotomous variable set to 1 if the year was 2016 and 0 \nfor the baseline indicator (the mean for the 2014 and 2015 mean monthly bills \namount is taken to form the baseline indicator), Vulnerables is the indicator for \nvulnerability set to 1 if the household is vulnerable on a given indicator and 0 if \nit is not vulnerable on that indicator (see the descriptions above for the coding \nof vulnerable groups). These terms are included as controls for each individual \nindicator. Subscript s refers to a term that differs across subjects, but is constant \nover time for a given subject. Subscript t refers to a term that changes over time, \nbut is constant across subjects at any given point in time. Terms with sub-script \nst vary across both subjects and time. Treats\u00d7Postt controls for the effect on \nbills due to the assignment to the TOU rate during the pilot year, and takes a \nvalue of 1 for households that were on TOU in the pilot year and 0 for all others; \nVulnerables\u00d7Postt controls for differences experienced during the pilot year by \nvulnerable groups regardless of the rate assignment, and takes a value of 1 for \nvulnerable households during the pilot year and 0 for all others; Vulnerables\u00d7Treats \ncontrols for the differences of vulnerable groups assigned to the TOU condition \nregardless of whether the pilot had begun or not, and takes a value of 1 for \nvulnerable households assigned to a TOU rate and 0 for all others. The term of \ninterest is Vulnerables\u00d7Treats\u00d7Postt, which gives the effect of the TOU assignment \nduring the pilot year on vulnerable groups; vulnerable households assigned to a \nTOU rate have a value of 1 for this term during the pilot year; all other groups, and \nother time periods, take a value of 0. Errors are clustered at the household level. \u03b5st \nrefers to the idiosyncratic error term.\nThe same form is used to examine TOU2, with TreatR2s (a dichotomous variable \nset to 1 if the household was on TOU2 and 0 if the household was in the control \ngroup) substituted for TreatR1s. Likewise, this form is used to examine the on-peak \nusage differences between groups and time periods, with the dependent variable \nbeing on-peak use (kilowatt hour, mean) instead of the monthly bill amount.\nNature Energy | VOL 5 | January 2020 | 50\u201360 | www.nature.com/natureenergy\n58\n\nArticles\nNaTURe EneRgy\nRegional fixed effects regression. We used regional fixed effects analyses to \nconsider the impacts of predictor variables on billing, discomfort and the need for \nmedical attention. By using regional fixed effects grouped by climate zone, these \nmodels control for unobserved time-invariant differences between households, \nfor those differences associated with living in different climate zones. That is, the \npredictions of respective dependent variables in each regional fixed effects analysis \nare estimated conditional on the climate zone, and each climate zone is associated \nwith a unique intercept.\nAC curtailment analysis. We used a subsample of n\u2009=\u20094,129 households that were \nassigned to TOU (not control) and did not answer \u201cNot applicable\u201d when asked \nabout AC curtailment (that is, we only considered households that have access to \nAC in their homes). Only households assigned to TOU were examined to restrict \nthe comparison to households with the incentive to curtail during peak hours. For \nthis subsample, we examined whether the vulnerable or non-vulnerable groups \nreported more frequent curtailment in the form of AC use. We used Wilcoxon rank \nsum tests (due to non-normal distributions) to test for differences in the reported \nfrequency of evening AC curtailment among vulnerable versus non-vulnerable \nhouseholds. Table 4 reports the means, z scores with associated P tests and Cohen\u2019s \nd effect sizes (d). Effect sizes indicate the importance of mean differences; a large \neffect size is >0.8, a medium one >0.5 and a small one >0.2 (ref. 61). The results \nprovide a richer description of how households responded to TOU rates, and are \ncorrelational rather than causal.\nOnline content. Any methods, additional references, Nature Research \nreporting summaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contributions and \ncompeting interests; and statements of data and code availability are available at \nhttps://doi.org/10.1038/s41560-019-0507-y.\nReporting summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nThe processed or aggregated data that support the plots within this paper and other \nfindings of this study are available from the corresponding author upon reasonable \nrequest. Authors signed a non-disclosure agreement with the utility that provided \nthe data analysed in this paper, and under this agreement are unable to make the \nraw data publicly available. Source data for Fig. 1 are provided with the paper.\nReceived: 20 February 2019; Accepted: 24 October 2019;  \nPublished online: 16 December 2019\nReferences\n\t1.\t Electricity Information 2017 (International Energy Agency/OECD, 2017).\n\t2.\t Residential Rate Reform / R.12-06-013 (California Public Utilities Commission, \naccessed 12 January 2018); www.cpuc.ca.gov/General.aspx?id=12154\n\t3.\t Todd, A., Cappers, P. & Goldman, C. Residential Customer Enrollment in \nTime-based Rate and Enabling Technology Programs: Smart Grid Investment \nGrant Consumer Behavior Study Analysis Report LBNL-6247E (Lawrence \nBerkeley National Laboratory, 2013).\n\t4.\t Sperling, D. & Eggert, A. California\u2019s climate and energy policy for \ntransportation. Energy Strateg. Rev. 5, 88\u201394 (2014).\n\t5.\t Hern\u00e1ndez, D. Understanding \u2018energy insecurity\u2019 and why it matters to \nhealth. Soc. Sci. Med. 167, 1\u201310 (2016).\n\t6.\t Anderson, W., White, V. & Finney, A. Coping with low incomes and cold \nhomes. Energy Policy 49, 40\u201352 (2012).\n\t7.\t Anderson, W., White, V. & Finney, A. \u2018You Just Have to Get By\u2019: Coping with \nLow Incomes and Cold Homes (Centre for Sustainable Energy, 2010).\n\t8.\t Middlemiss, L. & Gillard, R. Fuel poverty from the bottom-up: characterising \nhousehold energy vulnerability through the lived experience of the fuel poor. \nEnergy Res. Soc. Sci. 6, 146\u2013154 (2015).\n\t9.\t Bouzarovski, S., Petrova, S. & Sarlamanov, R. Energy poverty policies in the \nEU: a critical perspective. Energy Policy 49, 76\u201382 (2012).\n\t10.\tSnell, C., Bevan, M. & Thomson, H. Justice, fuel poverty and disabled people \nin England. Energy Res. Soc. Sci. 10, 123\u2013132 (2015).\n\t11.\tHealy, J. D. Excess winter mortality in Europe: a cross country analysis \nidentifying key risk factors. J. Epidemiol. Commun. Health 57, 784\u2013789 (2003).\n\t12.\tBoardman, B. Fixing Fuel Poverty: Challenges and Solutions (Earthscan, 2010).\n\t13.\tColton, R. Measuring the Outcomes of Low-Income Energy Assistance \nPrograms through a Home Energy Insecurity Scale (US Department of Health \nand Human Services, 2003).\n\t14.\tWalker, G. & Day, R. Fuel poverty as injustice: integrating distribution, \nrecognition and procedure in the struggle for affordable warmth. Energy \nPolicy 49, 69\u201375 (2012).\n\t15.\tHarlan, S. L., Declet-Barreto, J. H., Stefanov, W. L. & Petitti, D. B. \nNeighborhood effects on heat deaths: social and environmental predictors of \nvulnerability in Maricopa County, Arizona. Environ. Health Perspect. 121, \n197\u2013204 (2013).\n\t16.\tSakka, A., Santamouris, M., Livada, I., Nicol, F. & Wilson, M. On the thermal \nperformance of low income housing during heat waves. Energy Build. 49, \n69\u201377 (2012).\n\t17.\tMichelozzi, P. et\u00a0al. The impact of the summer 2003 heat waves on mortality \nin four Italian cities. Eur. Surveillance 10, 11\u201312 (2005).\n\t18.\tPoumad\u00e8re, M., Mays, C., Le Mer, S. & Blong, R. The 2003 heat wave  \nin France: dangerous climate change here and now. Risk Anal. 25,  \n1483\u20131494 (2005).\n\t19.\tCayla, J.-M., Maizi, N. & Marchand, C. The role of income in energy \nconsumption behaviour: evidence from French households data. Energy Policy \n39, 7874\u20137883 (2011).\n\t20.\tGillard, R., Snell, C. & Bevan, M. Advancing an energy justice perspective of \nfuel poverty: household vulnerability and domestic retrofit policy in the \nUnited Kingdom. Energy Res. Soc. Sci. 29, 53\u201361 (2017).\n\t21.\tWalker, G. & Day, R. Necessary energy uses and a minimum standard of \nliving in the United Kingdom: energy justice or escalating expectations? \nEnergy Res. Soc. Sci. 18, 129\u2013138 (2016).\n\t22.\tOrmandy, D. & Ezratty, V. Health and thermal comfort: from WHO guidance \nto housing strategies. Energy Policy 49, 116\u2013121 (2012).\n\t23.\tVandentorren, S. et\u00a0al. August 2003 heat wave in France: risk factors for death \nof elderly people living at home. Eur. J. Public Health 16, 583\u2013591 (2006).\n\t24.\tDay, R., Walker, G. & Simcock, N. Conceptualising energy use and energy \npoverty using a capabilities framework. Energy Policy 93, 255\u2013264 (2016).\n\t25.\tBasu, R. High ambient temperature and mortality: a review of epidemiologic \nstudies from 2001 to 2008. Environ. Health 8, 40 (2009).\n\t26.\tKnowlton, K. et\u00a0al. The 2006 California heat wave: impacts on hospitalizations \nand emergency department visits. Environ. Health Perspect. 117, 61\u201367 (2009).\n\t27.\tNord, M. & Kantor, L. S. Seasonal variation in food insecurity is associated \nwith heating and cooling costs among low-income elderly Americans. J. Nutr. \n136, 2939\u20132944 (2006).\n\t28.\tCook, J. T. et\u00a0al. A brief indicator of household energy security: associations \nwith food security, child health, and child development in US infants and \ntoddlers. Pediatrics 122, e867\u2013e875 (2008).\n\t29.\tCurriero, F. C. et\u00a0al. Temperature and mortality in 11 cities of the eastern \nUnited States. Am. J. Epidemiol. 155, 80\u201387 (2002).\n\t30.\tPager, D. & Shepherd, H. The sociology of discrimination: racial \ndiscrimination in employment, housing, credit, and consumer markets.  \nAnnu. Rev. Sociol. 34, 181\u2013209 (2008).\n\t31.\tSovacool, B. K. & Dworkin, M. H. Energy justice: conceptual insights and \npractical applications. Appl. Energy 142, 435\u2013444 (2015).\n\t32.\tBednar, D. J., Reames, T. G. & Keoleian, G. A. The intersection of energy and \njustice: modeling the spatial, racial/ethnic and socioeconomic patterns of \nurban residential heating consumption and efficiency in Detroit, Michigan. \nEnergy Build. 143, 25\u201334 (2017).\n\t33.\tReames, T. G. Targeting energy justice: exploring spatial, racial/ethnic and \nsocioeconomic disparities in urban residential heating energy efficiency. \nEnergy Policy 97, 549\u2013558 (2016).\n\t34.\tUejio, C. K. et\u00a0al. Intra-urban societal vulnerability to extreme heat: the role \nof heat exposure and the built environment, socioeconomics, and \nneighborhood stability. Health Place 17, 498\u2013507 (2011).\n\t35.\tFaruqui, A. & Palmer, J. Dynamic pricing of electricity and its discontents. \nSSRN Electron. J. https://doi.org/10.2139/ssrn.1908963 (2011).\n\t36.\tCappers, P., Spurlock, C. A., Todd, A. & Jin, L. Are vulnerable customers any \ndifferent than their peers when exposed to critical peak pricing: evidence \nfrom the US. Energy Policy 123, 421\u2013432 (2018).\n\t37.\tTrain, K. E., McFadden, D. L. & Goett, A. A. Consumer attitudes and \nvoluntary rate schedules for public utilities. Rev. Econ. Stat. 69, 383 (1987).\n\t38.\tHarrison, C. & Popke, J. \u2018Because you got to have heat\u2019: the networked \nassemblage of energy poverty in eastern North Carolina. Ann. Assoc. Am. \nGeogr. 101, 949\u2013961 (2011).\n\t39.\tFaruqui, A., Sergici, S. & Palmer, J. The Impact of Dynamic Pricing on Low \nIncome Customers (Institute of Electric Efficiency, 2010).\n\t40.\tCappers, P., Spurlock, C. A., Todd, A. & Jin, L. Experiences of Vulnerable \nResidential Customer Subpopulations with Critical Peak Pricing Report \nLBNL-1006294 (Lawrence Berkeley National Laboratory, 2016).\n\t41.\tCommission for Energy Regulation. Electricity Smart Metering Customer \nBehaviour Trials (CBT) Findings Report Information Paper CER11080a  \n(The Commission for Energy Regulation, 2011).\n\t42.\tFaruqui, A. & George, S. Quantifying customer response to dynamic pricing. \nElectr. J. 18, 53\u201363 (2005).\n\t43.\tSchofield, J. et\u00a0al. Residential Consumer Responsiveness to Time-varying Pricing \nReport A2 Low Carbon London LCNF Project (Imperial College  \nLondon, 2014).\n\t44.\tNicholls, L. & Strengers, Y. Peak demand and the \u2018family peak\u2019 period in \nAustralia: understanding practice (in)flexibility in households with children. \nEnergy Res. Soc. Sci. 9, 116\u2013124 (2015).\nNature Energy | VOL 5 | January 2020 | 50\u201360 | www.nature.com/natureenergy\n59\n\nArticles\nNaTURe EneRgy\n\t45.\tLiddell, C. & Morris, C. Fuel poverty and human health: a review of recent \nevidence. Energy Policy 38, 2987\u20132997 (2010).\n\t46.\tHealy, J. D. & Clinch, J. P. Quantifying the severity of fuel poverty, its \nrelationship with poor housing and reasons for non-investment in energy-\nsaving measures in Ireland. Energy Policy 32, 207\u2013220 (2004).\n\t47.\tO\u2019Sullivan, K. C., Howden-Chapman, P. L. & Fougere, G. M. Fuel poverty, \npolicy, and equity in New Zealand: the promise of prepayment metering. \nEnergy Res. Soc. Sci. 7, 99\u2013107 (2015).\n\t48.\tThomson, H. & Snell, C. Quantifying the prevalence of fuel poverty across \nthe European Union. Energy Policy 52, 563\u2013572 (2013).\n\t49.\tLiddell, C. & Guiney, C. Living in a cold and damp home: frameworks  \nfor understanding impacts on mental well-being. Public Health 129,  \n191\u2013199 (2015).\n\t50.\tRowlands, I. H. & Furst, I. M. The cost impacts of a mandatory move to \ntime-of-use pricing on residential customers: an Ontario (Canada) case-study. \nEnergy Effic. 4, 571\u2013585 (2011).\n\t51.\tJohnson, E. J. et\u00a0al. Beyond nudges: tools of a choice architecture. Mark. Lett. \n23, 487\u2013504 (2012).\n\t52.\tPeters, E., Hibbard, J., Slovic, P. & Dieckmann, N. Numeracy skill and the \ncommunication, comprehension, and use of risk\u2013benefit information. Health \nAff. 26, 741\u2013748 (2007).\n\t53.\tMarghetis, T., Attari, S. Z. & Landy, D. Simple interventions can correct \nmisperceptions of home energy use. Nat. Energy 4, 874\u2013881 (2019).\n\t54.\tWhite, L. V. & Sintov, N. D. Inaccurate consumer perceptions of monetary \nsavings in a demand-side response programme predict programme \nacceptance. Nat. Energy 3, 1101\u20131108 (2018).\n\t55.\tBerisha, V. et\u00a0al. Assessing adaptation strategies for extreme heat: a public \nhealth evaluation of cooling centers in Maricopa County, Arizona. Weather. \nClim. Soc. 9, 71\u201380 (2017).\n\t56.\tSovacool, B. K. Fuel poverty, affordability, and energy justice in England: \npolicy insights from the Warm Front Program. Energy 93, 361\u2013371 (2015).\n\t57.\tHern\u00e1ndez, D. & Phillips, D. Benefit or burden? Perceptions of energy \nefficiency efforts among low-income housing residents in New York City. \nEnergy Res. Soc. Sci. 8, 52\u201359 (2015).\n\t58.\tQiu, Y., Colson, G. & Wetzstein, M. E. Risk preference and adverse selection \nfor participation in time-of-use electricity pricing programs. Resour. Energy \nEcon. 47, 126\u2013142 (2017).\n\t59.\tNicolson, M., Huebner, G. & Shipworth, D. Are consumers willing to switch \nto smart time of use electricity tariffs? The importance of loss-aversion and \nelectric vehicle ownership. Energy Res. Soc. Sci. 23, 82\u201396 (2017).\n\t60.\tMostafa Baladi, S., Herriges, J. A. & Sweeney, T. J. Residential response to \nvoluntary time-of-use electricity rates. Resour. Energy Econ. 20, 225\u2013244 (1998).\n\t61.\tCohen, J. Statistical Power Analysis for the Behavioral Sciences  \n(Lawrence Earlbaum Associates, 1988).\nAcknowledgements\nThe authors thank their utility partner for furnishing the data, D. Mazmanian for \nextensive advice and J. McPartlan for the considerable time invested in data management.\nAuthor contributions\nBoth authors conceived the paper and designed the research. L.W. designed the analysis \nmethods, performed the analyses and wrote and revised the paper. N.S. reviewed several \ndrafts and made revisions.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-019-0507-y.\nCorrespondence and requests for materials should be addressed to L.V.W.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2019\nNature Energy | VOL 5 | January 2020 | 50\u201360 | www.nature.com/natureenergy\n60\n\n1\nnature research  |  reporting summary\nOctober 2018\nCorresponding author(s):\nLee White\nLast updated by author(s): Oct 17, 2019\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nNo software was used by the authors for data collection (the data were provided by partner)\nData analysis\nSTATA MP 14.2\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. \nWe strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nThe processed or aggregated data that support the plots within this paper and other findings of this study are available from the corresponding author upon \nreasonable request. Authors signed a non-disclosure agreement with the utility that provided the data analysed in this paper, and under this agreement are unable \nto make the raw data publicly available. \n\n2\nnature research  |  reporting summary\nOctober 2018\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThe study is quantitative\nResearch sample\nThe study used an existing dataset provided by a utility in the US that has asked to remain anonymous. The sample is all from a single US \nstate. The sample contained opt-in participants, and may not be completely representative of the general population.\nSampling strategy\nThe sample included individuals who opted in to a utility pilot program. These individuals were then randomized to one of four \nconditions, including a control group.\nData collection\nSome of the data is usage and billing data provided directly by the utility. The rest of the data is survey data collected by the utility; for \nthe full sample examined in this manuscript (n = 7487), 85% responded by email, 11% by mail, and 4% by phone.\nTiming\nSurvey data was collected October-December 2016; usage and billing data are recorded from 2014-2016.\nData exclusions\nParticipants in one of three utility treatment groups were not included due to that treatment having unusual features. Participants were \nalso excluded if they had incomplete billing, usage, and survey data, leaving 7487 of the 16181 that had remained in the pilot program \nand completed the survey.\nNon-participation\n2787 households dropped out of the pilot due to relocating, ineligibility, or choosing to drop out, leaving 18747 households enrolled. Of \nthese, 16181 households responded to the survey. \nRandomization\nParticipants were randomly allocated to a treatment or control condition\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology\nAnimals and other organisms\nHuman research participants\nClinical data\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nHuman research participants\nPolicy information about studies involving human research participants\nPopulation characteristics\nSee above\nRecruitment\nParticipants opted-in to a utility pilot. Thus, the sample may contain participants who are more favorable to this type of rate \nprogram than the general population. \nEthics oversight\nThe University of Southern California\u2019s University Park Institutional Review Board reviewed and approved this research, and \ngranted a waiver of informed consent. \nNote that full information on the approval of the study protocol must also be provided in the manuscript.\n\n\n    Scientific Research Findings:", "answer": "Elderly people and those with disabilities had greater bill increases when moved to time-of-use rates (versus staying on existing rates), compared to the equivalent increase seen in their non-vulnerable counterparts assigned to time-of-use rates. Low-income and Hispanic households had lower bill increases compared to non-vulnerable counterparts. Hispanic households and those with disabilities experienced worse health outcomes on time-of-use rates, while households with young children experienced better health outcomes. Otherwise, no health and bill differences were observed between vulnerable and non-vulnerable groups. These results suggest that time-of-use rates may increase hardships faced by some groups already more likely to face energy poverty, but impacts vary by sociodemographic group and rate design. This study was conducted during the summer in the south-western United States, and we expect these results to extend to similar contexts \u2014 hot climates during summer. Findings may not generalize to different rate designs or climate conditions.", "id": 29}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-019-0460-9\nInstitute for Economy and the Environment, University of St.Gallen, St.Gallen, Switzerland. *e-mail: adrian.rinscheid@unisg.ch\nE\nighty per cent of the world\u2019s coal reserves must stay in the ground \nto reach the target of limiting global warming to well below 2\u2009\u00b0C \ncompared with pre-industrial levels1. In 2008, climate scientists \nhad called for a complete divestment from coal-fired electricity by \n20302, a proposition reiterated by a recent roadmap for rapid decar-\nbonization3. Yet despite the strong growth of renewable energies, coal \nstill accounted for 28% of the world\u2019s primary energy supply in 20174. \nAs Pfeiffer et\u00a0al.5 point out, coal-fired power plants \u201cwill need to be \nunderutilized, retired early, or retrofitted [\u2026] or\u2014in short\u2014stranded\u201d \nif countries are serious about reaching the targets set in Paris.\nAs private markets will not spur such necessary developments \non their own, government policies play an important role in phas-\ning out coal6,7. However, in democratic countries, such policies may \nface public opposition. While several studies suggest that the global \nenergy transition (that is, shifting from non-renewable to renewable \nenergies) is likely to lead to net job creation in most economies8, \nthe closure of coal mines and coal-fired power plants may lead to \ntemporary and regionally concentrated job losses. The anticipation \nof negative employment effects could lead to opposition by those \nworking in the sector9. Moreover, the coal industry has become \nan identity-shaping symbol deeply engrained in the culture of \nsome communities and countries, such as in the German region of \nLusatia10, Silesia in Poland11 or Appalachia in the United States12. \nOpposition to a phase-out of coal is also likely to be fuelled by the \neconomic actors that have to bear parts of the costs, such as utilities \nwhose business models depend on coal, and labour unions repre-\nsenting coal workers. Given their desire to be re-elected, democratic \ngovernments may be responsive to these concerns.\nAt the same time, recent assessments such as the \nIntergovernmental Panel on Climate Change\u2019s Special Report \n\u2018Global Warming of 1.5\u2009\u00b0C\u2019 emphasize the urgency of ambitious \naction to prevent irreversible climate change13. While some jurisdic-\ntions have committed to phasing out coal, such as Canada and the \nUnited Kingdom who launched the Powering Past Coal Alliance in \nNovember 2017, many others are not delivering policies ambitious \nenough to meet the climate challenge. Germany, the largest energy \nconsumer in the European Union, the world\u2019s largest producer of \nlignite and one of the top-ten coal-burning countries in the world4, \nhas recently started to organize its departure from coal.\nGermany has relied heavily on coal for power generation for a \nlong time. In the 1950s, more than 500,000 people were employed \nin the sector, contributing to the German Wirtschaftswunder after \nWorld War II14. While the oil crisis in 1973\u20131974 aided the devel-\nopment of nuclear power, domestic coal was viewed as important \nto ensure energy security, and utilities were thus required to burn \na quota of domestic hard coal15. As operating costs in the coal \nindustry started to outpace market revenues in the 1960s, federal \nand state governments introduced subsidies for hard-coal mining. \nThese amounted to more than \u20ac320 billion until they were phased \nout in 201816.\nIn the 1970s, coal-fired energy generation started to become \nmore controversial. The environmental movement and (later) the \nnewly established Green Party opposed coal-fired power plants and \nopen-pit mining but were even more concerned with nuclear power, \nlargely avoiding simultaneous contestation on two fronts14. The coal \nindustry nurtured strong ties to the country\u2019s two largest parties, the \nChristian Democrats (CDU/CSU) and the Social Democrats (SPD), \nwith the SPD in particular aiming to keep coal mines open as long \nas possible14. In the wake of the 1986 Chernobyl nuclear accident, \nGermany adopted a policy framework in 1990 for the promotion of \nrenewable energies17, which fundamentally changed electricity mar-\nkets. The share of power generated by renewables soared from 3.6% \nto 35.2% between 1990 and 201818.\nWhile Germany has been successful in transitioning from \nnuclear to renewable energy, 35.3% of electricity generation still \ncame from coal in 2018 (Fig. 1). Policy support for renewables is \nan important element in decarbonizing the energy sector19, but it is \nan open question whether layering support schemes for sustainable \ntechnologies on top of the existing institutions20 without address-\ning the legacy of fossil fuels is enough to \u201ceffectively lead energy \nsystems out of carbon lock-in\u201d21. Given only gently decreasing emis-\nsions (Fig. 1), the question of a coal phase-out gained prominence in \nthe aftermath of the adoption of the Paris Agreement. In November \n2016, the German government adopted the Climate Action Plan \n2050, outlining measures to achieve the country\u2019s climate targets. \nThis plan failed to develop a phase-out strategy, and in 2018 the task \nwas delegated to the Coal Commission, an expert commission on \ngrowth, structural change and employment consisting of a variety of \nGermany\u2019s decision to phase out coal by 2038 \nlags behind citizens\u2019 timing preferences\nAdrian Rinscheid\u200a \u200a* and Rolf W\u00fcstenhagen\nCoal-fired power generation is the single most important source of carbon dioxide emissions in many countries, including \nGermany. A government commission recently proposed to phase out coal by 2038, which implies that the country will miss its \n2020 climate target. On the basis of a choice experiment that assessed 31,744 hypothetical policy scenarios in a representative \nsample of German voters, we show that voters prefer a phase-out by 2025. They would uphold their support for greater climate \nambition up to an additional cost to society of \u20ac8.5 billion. Voters in Rhineland and Lusatia, the country\u2019s main coal regions, \nalso support an earlier phase-out, but to a lesser extent than voters in other regions. By demonstrating that political decision-\nmakers are more reluctant than voters in overcoming energy path dependence, our analysis calls for further research to explain \nthe influence of particular stakeholders in slowing energy transitions.\nNature Energy | VOL 4 | OCTOBER 2019 | 856\u2013863 | www.nature.com/natureenergy\n856\n\nArticles\nNATuRe EneRGy\nstakeholders including industry associations, labour unions, state-\nlevel governments, environmental non-governmental organizations \nand independent scientists appointed by the federal government. In \nearly 2019, this group recommended to phase out coal-fired power \ngeneration by 2038, proposing an array of measures to support the \ncoal regions in restructuring their economies.\nIn light of recent concerns about populist backlash against cli-\nmate policy22,23, some observers consider the proposed timeline a \nreasonable compromise between public acceptance and climate \nchange mitigation. Although the compromise has been praised for \nrepresenting a broad societal consensus, given the urgency of ambi-\ntious climate action3,13, some members of the Coal Commission \nhave criticized the plan as not ambitious enough (ref. 24, pages 118 \nand 119) to deliver on Germany\u2019s climate policy targets.\nGiven these concerns and the view expressed in the German \nClimate Action Plan that public support is a central precondition \nfor successful implementation of climate policies25, we investigate \nwhether the recommendations of the Coal Commission are in line \nwith voters\u2019 preferences, particularly regarding the temporal dimen-\nsion. First, we investigate how citizens\u2019 support for a coal phase-out \nis affected by different timelines and other features of a phase-out \nand examine the moderating influences of political orientation and \nclimate change-related beliefs. Second, we explore the preferences \nof citizens living in Germany\u2019s two largest coal regions, Rhineland \nand Lusatia. Our analysis suggests that compared with the recom-\nmendations of the Coal Commission, a more ambitious timeline for \nphasing out coal would actually have been better aligned with citi-\nzens\u2019 preferences.\nEffects of phase-out design on public support\nWith data from a large-scale choice experiment, we examine how \npublic preferences for a coal phase-out in Germany are affected \nby different proposed phase-out timelines and compare citizens\u2019 \npreferences with the recommendations of the expert commis-\nsion. We also investigate the role of other policy attributes (cost, \neffects on jobs and supporting measures for the transformation of \nthe coal regions; see Table 1 and Methods for details). Our analysis  \nis based on an online survey administered to a nonprobability but \nrepresentative sample of 2,161 Germans who are eligible to vote \n(see Supplementary Table 1). The choice experiment involved a \nrating task whereby respondents were exposed to eight consecu-\ntive pairs of hypothetical policy scenarios to phase out coal. In the \nscenarios, the attribute levels of the phase-out policy were varied \n70\n60\n50\n40\n30\nShare of coal in German electricity mix (%)\nGreenhouse gas emissons from energy sector\n(in million tons of CO2e)\n20\n10\n0\n1990\n1995\n2000\n2005\n2010\n2015\n420\n360\n300\n240\n180\n120\n60\nLignite\nHard coal\n0\nYear\nFig. 1 | Share of coal in German electricity mix and energy-related greenhouse gas emissions, 1990\u20132018. Based on data from AG Energiebilanzen18 and \nthe German Environment Agency64.\nTable 1 | Policy attributes and levels for the choice experiment\nPolicy attributes\nAttribute levels\nEnd date of the \nphase-out\n\u2219By 2025\n\u2219By 2030\n\u2219By 2040\n\u2219By 2100\nAnnual cost \nper two-person \nhousehold (overall \ncosts for the \neconomy)\n\u2219\u20ac0\n\u2219\u20ac6 (\u20ac250 million)\n\u2219\u20ac12 (\u20ac500 million)\n\u2219\u20ac18 (\u20ac750 million)\nNumber of lost jobs \nin the coal industry\n\u2219\u20135,000\n\u2219\u201310,000\n\u2219\u201315,000\n\u2219\u201320,000\nNumber of newly \ncreated jobs\n\u22195,000\n\u221910,000\n\u221915,000\n\u221920,000\nMeasures for \nstructural change\n\u2219Investment in expansion of renewable energies\n\u2219Investment in regional funding programmes for new \nbusinesses (for example start-up funding)\n\u2219Investment in modern infrastructure (electric \nvehicles, digitalization)\n\u2219Investment in research and development\n\u2219Mixture of further training and early retirement for \ncoal industry employees\nNature Energy | VOL 4 | OCTOBER 2019 | 856\u2013863 | www.nature.com/natureenergy\n857\n\nArticles\nNATuRe EneRGy\nrandomly. Participants were asked to rate these scenarios on a \nscale from 1 (\u2018very poor\u2019) to 7 (\u2018very good\u2019). To ease interpretation \nof the marginal effects shown in Figs. 2, 3 and 4, the rating scale \nwas dichotomized, using the median value of 4 as a cut-off value.  \nThe resulting dependent variable \u2018Phase-out Support\u2019 is hence \ncoded 0 for cases where a respondent rated a proposal as poor \nto neutral (1 to 4) and 1 for cases where (s)he rated it as positive  \n(5 to 7). The fully randomized design allows us to estimate the \ncausal effects of multiple treatment components simultaneously \nusing simple linear regression26. The subsequent analyses are based \non a sample of 1,984 Germans evaluating 31,744 policy scenar-\nios. This sample is cleaned of respondents who failed to correctly \nanswer an attention check implemented in the choice experiment. \nHowever, all results discussed in the paper remain substantively \nthe same when replicating the analyses to include the inattentive \nrespondents (see Supplementary Tables 2, 4, 6, 8 and 10).\nFigure 2 shows the marginal effects associated with each attri-\nbute level based on regression analysis (Supplementary Table 2), \nusing the dichotomized rating outcome as a dependent variable \nand standard errors grouped at the level of the individual (clustered \nstandard errors). For the timing attribute, we take 2040 as the ref-\nerence category, as this level most closely matches the 2038 time-\nline ultimately recommended by the Coal Commission. We find \nthat policy scenarios with 2025 as an end date have a significantly \nhigher probability of being supported than policies with later end \ndates. Postponing the phase-out to 2040 leads to a decrease in policy \nsupport of 10.7 percentage points, and postponing it to 2100\u2014as \nreflected in the G7\u2019s statement to phase out fossil fuels by the end \nof the century\u2014leads to a decrease in policy support of 15.3 per-\ncentage points, compared with the 2040 baseline. As is apparent in \nthe data, Germans are also sensitive to the cost of a coal phase-out. \nEvery increase in annual cost of \u20ac10 per household (or about \u20ac400 \nmillion per year for the German economy as a whole) decreases \npublic support by about seven percentage points. With regard to \nemployment effects, people prefer scenarios with lower job losses \nover scenarios with higher job losses, but they value newly created \njobs slightly higher than lost old jobs. For instance, while the sce-\nnario with 20,000 lost jobs decreases phase-out support by 9.2 per-\ncentage points compared with a scenario with only 5,000 lost jobs, \ncreating the same number of new jobs increases phase-out support \nby 12.2 percentage points. The type of supportive measures for the \nlocal economy is the least important of the five attributes. Among \nthe design options offered here, the preferred attribute level is an \nexpansion of renewable energies.\nPartisan differences and gateway beliefs\nAs the discussions of the Coal Commission showed, the main ques-\ntion about phasing out coal is not if, but when, the phase-out is \ngoing to happen. Hence, the following analyses focus specifically on \nthe question of timing. While Fig. 2 indicates that the timeline does \nindeed have a considerable effect on citizens\u2019 preferences, there may \nbe differences between population subgroups. In particular, it has \nbeen suggested that party identification structures people\u2019s energy \npolicy preferences27,28. Germany\u2019s party elites represent opposing \nviews on the coal phase-out, ranging from the Greens\u2019 position for \nan early phase-out to the conservative parties tending to defend the \nstatus quo29. In the context of the 2017 federal elections, the partisan \ndivide on the topic became highly visible, and the question of tim-\ning was one of the reasons why the negotiations for a government \ncoalition of the Christian Democrats, the Liberal Democrats (FDP) \nand the Green Party failed in November 201730. Figure 3a shows \nthat there is some variation among partisans with regard to the \nstrength of their timing preferences. Unsurprisingly, Green Party \nsupporters show the strongest preference for an early phase-out in \n2025. What may be more surprising is that supporters of almost all \nother parties also prefer 2025 over 2040. The only exception is the \nrelatively small subsample supporting the Bavarian arm of the CSU, \nwhere the preference for 2025 is not significant. In contrast to public \nstatements by their party leaders, FDP and Green Party voters have \nsimilar views on this issue. For all respondents, phasing out in 2100 \nis the least preferred timeline, although supporters of the right-wing \npopulist party Alternative f\u00fcr Deutschland (AfD) are comparatively \nmore positive about such a late phase-out date than supporters of \nall other parties. In light of other surveys investigating public atti-\ntudes on the German energy transition (either surveys with a broad \nfocus31 or those with a specific focus on the coal phase-out32), the \nmuted differences across different partisans actually reflect a recur-\nring pattern. See Supplementary Table 3 for the supporting regres-\nsion analyses.\nWe also expected beliefs about climate change to influence citi-\nzens\u2019 support of different phase-out timing options. We assessed \nclimate change-related beliefs by asking respondents to estimate \nthe share of global climate scientists who think that the rise in the \natmospheric CO2 concentration since the mid-twentieth century is \nprimarily due to human activities. Perceived scientific consensus \nabout the anthropogenic nature of current climate change functions \nas a \u2018gateway belief\u2019 that influences several other attitudes related to \nclimate change and energy33\u201336. While quantifications show that the \nconsensus is shared by 90\u2013100% of publishing climate scientists37, \nTraining and early retirement\nResearch and development\nInvestment in modern infrastructure\nSupport for new businesses\nExpansion of renewables\nSupportive measures\n 20,000\n15,000\n10,000\n5,000\nNew jobs\n\u221220,000\n\u221215,000\n\u221210,000\n\u22125,000\nLost jobs\n\u20ac 18\n\u20ac 12\n\u20ac 6\n\u20ac 0\nAnnual cost per household\n2100\n2040\n2030\n2025\nEnd date\n\u20130.2\n0\n0.2\nPhase-out support (binary indicator)\nFig. 2 | Average effects of policy attributes on respondents\u2019 preferences \nfor a coal phase-out. Each dot represents an AMCE of randomly assigned \nattribute levels on the probability of supporting a given policy scenario \nrelative to the reference scenario, all else being equal. The horizontal bars \nrepresent the 95% confidence intervals. Dots without bars represent the \nreference level for each policy attribute. There were n\u2009=\u20091,984 respondents \nand 31,744 policy scenarios.\nNature Energy | VOL 4 | OCTOBER 2019 | 856\u2013863 | www.nature.com/natureenergy\n858\n\nArticles\nNATuRe EneRGy\na recent study conducted in the United States highlights that only \n15% of US citizens are aware of this high level of consensus38. In our \nGerman sample, the mean estimate of consensus is 66% (s.d.\u2009=\u200922.9), \nand 18.3% of respondents estimate the consensus to be 90% or \nhigher. Figure 3b shows that perceived consensus strongly moder-\nates the effect of phase-out timelines on preferences. Respondents \nwho think the consensus is below 50% are indifferent to whether \nthe proposed end date is 2025, 2030 or 2040, but their support still \ndecreases if the proposed end date is 2100. The closer respondents\u2019 \nclimate-related beliefs approximate the true level of scientific con-\nsensus, the more pronounced their preference for an earlier phase-\nout. Respondents who (accurately) estimate the consensus to be \n90% or higher prefer a 2025 phase-out date by more than 40 per-\ncentage points over a phase-out date of 2100. See Supplementary \nTable 5 for the supporting regression analyses.\nTies to coal industry weaken support for early phase-out\nTo explore the influence of social embeddedness on preferences for \na coal phase-out, we rely on two additional samples of residents of \nthe two main coal regions, Rhineland (n\u2009=\u2009533) and Lusatia (n\u2009=\u2009501), \nwho took the same survey. Within these independently collected \nregional samples, we further investigate whether the preferences of \npeople having direct ties to the coal industry, for example, through \nacquaintances or by being employed in the sector, differ from those of \nother respondents in the region. The results (Fig. 4 and Supplementary \nTables 7 and 9 for supporting regression analyses) suggest that peo-\nple in the coal regions have less pronounced preferences for an early \nphase-out than respondents in the nationwide sample. However, \nthere are some differences between the two regions. Phasing out \ncoal by 2025 or 2030 instead of 2040 has significantly higher sup-\nport in Rhineland, while respondents in the eastern German region \nof Lusatia tend to support a phase-out in 2030. Even here, later phase-\nout dates are significantly less preferred. An analysis of respondents \nwith strong (red symbols in Fig. 4) and weak (blue symbols) social \nties to the coal industry suggests that in both regions, people with \nstrong ties are indifferent to whether the proposed phase-out date \nis 2025, 2030 or 2040, as the confidence intervals around the point \nestimates for 2025 and 2030 include the dotted reference line.\nConclusion\nAddressing climate change effectively and rapidly requires not only \ninvesting in new energy technologies but also divesting from car-\nbon-intensive energy infrastructures39\u201343. Our study is among the \nfirst to investigate citizens\u2019 views on the second part of this equa-\ntion. Using a large-scale survey, we assessed German voters\u2019 prefer-\nences for different policy design options to phase out coal. We found \nthat the average respondent consistently prefers a more ambitious \ntimeline. All else being equal, the preference was to phase out coal \nby 2025, which contrasts with the Coal Commission\u2019s proposal to \nphase out coal by 2038. A particular strength of our methodologi-\ncal approach is that the choice experiments allow us to scrutinize \nrespondents\u2019 timing preferences in relation to possible trade-offs \nbetween the different attributes of an accelerated phase-out, such as \nhigher cost. By comparing preferences across attributes, we find that \nsupport for an accelerated phase-out is upheld up to an additional \ncost to society of \u20ac8.5 billion (see Supplementary Fig. 1).\nThe acceptance of policy proposals is also sensitive to the \nemployment effects of the energy transition. Cost matters, as do job \nlosses. If delaying the phase-out from 2025 to 2030 would result in \nhalving job losses from 20,000 to 10,000, voters would\u2014with all else \nbeing equal\u2014 accept the later phase-out. Our analysis also shows \nthat the creation of new jobs matters even more than the loss of \nold jobs. Policymakers aiming to find support for ambitious climate \npolicies are therefore well advised to make credible claims about \nhow these policies will lead to new employment opportunities in \nlow-carbon industries.\nOur results also shed light on the similarities and differences \namong population segments. With respect to party identifica-\ntion, preferences for earlier over later phase-out dates are wide-\nspread among almost the entire political spectrum. Even voters \nin Germany\u2019s two largest coal mining regions share\u2014to a large \nextent\u2014the preference for an earlier over a later phase-out. The \nonly notable exception are citizens with strong ties to the coal \nindustry, who have no significant preference for a 2025 phase-out \nover one that happens in 2030 or 2040. Similarly, voters in the east-\nern German region of Lusatia slightly prefer 2030 over 2025 as the \nphase-out date. Moreover, knowledge about the scientific consen-\nsus on anthropogenic climate change is an important predictor \nof supporting an ambitious phase-out. Slightly less than one-fifth \nof respondents are aware that more than 90% of climate scientists \nagree that climate change is human-made. These well-informed \nrespondents have a stronger preference for phasing out coal in 2025 \nthan those who (erroneously) believe that no such consensus exists.\nIn light of our findings, the German Coal Commission\u2019s pro-\nposal to phase out coal by 2038 does not appear to correspond \nwell with voter preferences. This might be an indication that  \nPhase-out support (binary indicator)\nEnd date\n\u20130.4\n\u20130.3\n\u20130.2\n\u20130.1\n0.1\n0.2\n0.3\n2100\n2040\n2030\n2025\na\nb\nDie Gr\u00fcnen (n = 138)\nSPD (n = 325)\nDie Linke (n = 160)\nCDU (n = 321)\nCSU (n = 85)\nFDP (n = 81)\nAfD (n = 129)\nPhase-out support (binary indicator)\nEnd date\n\u20130.3\n\u20130.2\n\u20130.1\n0.1\n0.2\n0.3\n2100\n2040\n2030\n2025\n0\u201349% perceived c. (n = 368)\n50\u201369% perceived c. (n = 513)\n70\u201389% perceived c. (n = 768)\n90\u2013100% perceived c. (n = 390)\n0\n0\nFig. 3 | Average effects of the timing attribute on respondents\u2019 preference for a coal phase-out. a,b, Symbols represent AMCEs for the timing attribute \n(base 2040), conditional on party identification (a) and perceived scientific consensus (\u2018perceived c.\u2019) (b) on the anthropogenic nature of current climate \nchange. The horizontal bars represent the 95% confidence intervals.\nNature Energy | VOL 4 | OCTOBER 2019 | 856\u2013863 | www.nature.com/natureenergy\n859\n\nArticles\nNATuRe EneRGy\ncommission members over-estimated voters\u2019 conservatism, as \npolitical elites have been shown to do frequently44,45. However, even \nassuming that the commission members gave constituents in coal-\nmining regions precedence over voters in other parts of the coun-\ntry would not explain why such a late date was chosen, as even in \nthose regions respondents preferred phase-out dates between 2025 \n(western Germany) and 2030 (eastern Germany). An alternative \nexplanation for this mismatch is that voter preferences simply did \nnot play a decisive role in the consultations of the commission. As \nFig. 5 illustrates, citizen voices (represented by non-governmental \norganizations, for example) were under-represented among the 28 \ncommission members. Moreover, most commission members were \ninsulated from re-election pressures, and some might have empha-\nsized short-term economic interests, such as the Confederation of \nGerman Employers\u2019 Associations (BGA) or the trade union repre-\nsenting workers in mining, chemicals and energy (IGBCE). While a \ndetailed analysis of the decision-making dynamics within the com-\nmission is beyond this paper\u2019s scope, the strong representation of \nincumbent interests within the commission highlights an important \ninstitutional barrier against overcoming energy path dependence. \nFurther work in this area could investigate the ability of corporatist \nstyles of decision-making to reform today\u2019s carbon-intensive energy \nsystems46. To successfully manage \u201cthe next phase of the energy \ntransition\u201d47, which implies making established technologies and \ninfrastructures redundant, we need to enhance our understand-\ning of incumbents\u2019 survival strategies, including their corporate \npolitical activity aimed at slowing down the transition. Moreover, \ngiven the prevalence of particular stakeholders who stress job losses \nrather than new opportunities, the nexus between employment con-\nsiderations and the political feasibility of decarbonization measures \nneeds more scholarly attention8,48. Energy transition researchers and \nmodellers would benefit from engaging with political scientists and \nsociologists to unveil the interests and activities of various actors \nwho are shaping energy policies. Policymakers trying to develop \nambitious climate change mitigation policies should be encouraged \nto find ways of being exposed to a balanced view of the risks and \nopportunities of the energy transition. Our results suggest that in \na democratic setting, such action could be rewarded in future elec-\ntions by voters.\nMethods\nChoice experiment rationale and design. To investigate voters\u2019 policy preferences, \nwe conducted a choice experiment. Choice experiments were developed in \nmarketing research to investigate the importance of different product design \nfeatures in determining purchasing preferences. The idea is to put respondents \nin a hypothetical yet realistic choice situation in which they are confronted with \nbundles of relevant product attributes. By observing stated preferences with regard \nto the presented alternatives, it is possible to examine the relevance of certain \nproduct attributes and their characteristics to individual choices.\nPolitical scientists have adopted the method to gauge citizens\u2019 preferences \nwith regard to different policy proposals or scenarios26,49. Analytically, the design \nfeatures of a policy are similar to product attributes, which is why the method \nprovides a powerful approach to simultaneously estimate the individual effects of \nseveral attributes of a policy proposal on voter preferences50. Choice experiments \nrequire decision-makers to make trade-offs between different policy attributes \nwhen evaluating various multidimensional alternatives. As a consequence, they \ncan mitigate the problem of social desirability bias in public opinion research on \nenvironmental matters26. In our case, using choice experiments may reduce the \nlikelihood of overestimating voters\u2019 appetite for an ambitious phase-out of coal.\nAt the beginning of the choice experiment, respondents were familiarized \nwith five attributes of a potential policy to phase out coal: the timescale of the \nphase-out, estimated costs, effects on employment in terms of layoffs and newly \ncreated jobs, and supporting measures for the transformation of the coal regions. \nWe selected these five attributes on the basis of the following considerations. In \n2017 the German Advisory Council on the Environment (SRU), an expert advisory \npanel to the federal government, recommended a staged approach in which the \ncoal-fired power plants with highest emissions would be disconnected from the \ngrid as early as 202051. The most efficient power plants would be successively shut \ndown in the 2030s, and the phase-out would be completed by 2040 at the latest. \nPhase-out support (binary indicator)\nEnd date\n\u20130.2\n\u20130.1\n0\n0.1\n0.2\n2100\n2040\n2030\n2025\nLusatia, total (n = 473)\nWorking in coal industry\nor having acquaintances (n = 258)\nNot working in coal industry,\nno acquaintances (n = 215)\nPhase-out support (binary indicator)\nEnd date\n\u20130.2\n\u20130.1\n0\n0.1\n0.2\n2100\n2040\n2030\n2025\na\nb\nRhineland, total (n = 491)\nWorking in coal industry\nor having acquaintances (n = 191)\nNot working in coal industry,\nno acquaintances (n = 300)\nFig. 4 | Average effects of the timing attribute on respondents\u2019 preference for a coal phase-out in Rhineland and Lusatia. a,b, Symbols represent AMCEs \nfor the timing attribute (base 2040) for the (a) Rhineland and (b) Lusatia samples, excluding inattentive respondents. The analyses also differentiate \nbetween respondents employed by or with acquaintances in the coal industry and those without strong coal industry ties. The horizontal bars represent \nthe 95% confidence intervals.\n4\n3\n6\n4\n6\nScientists\nNon-governmental organizations\nIndustry associations\nLabour unions\nOther politicians\nOther businesses\nOther civil society\nSubnational governments and\nadministration\n1\n3\n1\nFig. 5 | Composition of the Commission on Growth, Structural Change and \nEmployment. Numbers indicate number of members of the Coal Commission \nentitled to vote per category based on the list of commission members24.\nNature Energy | VOL 4 | OCTOBER 2019 | 856\u2013863 | www.nature.com/natureenergy\n860\n\nArticles\nNATuRe EneRGy\nThe SRU stresses the climate\u2013political necessity of immediately starting the phase-\nout to achieve appropriate implementation of the Paris climate targets in Germany. \nOther studies reach similar conclusions. Depending on the ambition of the first \nstages of the phase-out, studies show that it is technically feasible to accomplish \na coal phase-out by 203552 or as soon as 203053. If the 1.5\u2009\u00b0C target of the Paris \nclimate agreement is taken as the reference point, however, the phase-out of coal \nin Germany must already occur by around 202554. A different time horizon for \nthe phase-out of coal was adopted by the G7 in 2015 when they decided to end \nthe use of fossil fuels by the end of the century55. Hence, reflecting these different \nscenarios, our choice experiment uses 2025, 2030, 2040 and 2100 as attribute levels.\nArguments about costs, the second attribute of our choice experiment, play a \nlarge role in the public debate on phasing out coal. During the negotiations for a \ngovernment coalition consisting of the CDU/CSU, the FDP and the Green Party \n(the so-called Jamaica coalition) in November 2017, a number of energy-intensive \nfirms publicly warned that a phase-out of coal could mean a rise in electricity \nprices by up to 30% (ref. 56). While the exact effect of reducing coal-fired power \ngeneration on the electricity market depends on a variety of factors, including \ndemand response, growth in renewable power generation and cross-border trade, \nit seems plausible that changing the demand\u2013supply balance could have an effect \non prices. A study commissioned by the trade union ver.di indicates that an earlier \nphase-out would entail considerable employment-related costs57. Various proposals \nfor financing the phase-out of coal have been articulated, such as a levy on the \nelectricity price or a structural transformation fund for the German coal regions58. \nTo make the costs a relevant choice consideration for individual respondents, we \nmade the assumption that they would be passed on to consumers. In the choice \ntasks, we presented the attribute as cost per household or overall cost to the \neconomy, respectively, leaving open the concrete financing mechanism through \nwhich those costs would be incurred by consumers. In the instructions to the \nchoice experiment, we mentioned electricity price increases as one such possibility, \nwhich does not preclude other financing mechanisms that would ultimately also \naffect consumers, such as CO2 taxes or emissions trading. As existing investigations \nnecessarily work with a number of assumptions, it is difficult to find clear reference \npoints for plausible cost scenarios. The 2016 study by Ecke, however, offers some \nguidance57. This study estimates that the annual cost of a phase-out until 2040 will \namount to \u20ac499 million if industry and large customers were exempt from the \ncorresponding levy. Projected onto electricity prices, this corresponds to \u20ac0.0014 \nper kilowatt hour, implying an annual electricity price increase of \u20ac4.20 for a \ntypical two-person household with a consumption of 3,000 kWh. To leave space \nfor other factors not taken into account in estimates made to date, we defined \ncost levels of \u20ac6, \u20ac12 and \u20ac18 (annual, per two-person household). As a reference \ncategory, we assumed no costs.\nThe way that cost of a coal phase-out is presented to respondents might \ninfluence the weight they assign to this attribute. To take the possibility of such \ncost-framing effects into account, 50% of respondents (randomly assigned) \nreceived the same information in a different format. In addition to the electricity \nprice increase per household, they were also informed of the corresponding overall \ncosts for the economy. Projecting the costs per household onto the entire economy \nleads to cost levels of \u20ac250 million, \u20ac500 million and \u20ac750 million (see Table 1). As \nit turns out, however, the way the costs were presented in the choice experiment \ndid not influence respondents\u2019 responses. Hence, for all analyses reported in the \npaper, we pooled the data of the respective subsamples.\nAlong with the timescale and the costs, employment effects are an important \nconsideration in planning a coal phase-out. More than 20,000 people are currently \nemployed in the coal industry51. Nearly 70% of those working in the lignite mining \nsector are already over 46 years old and therefore will reach retirement age in the \nmid-2030s. Taking existing early retirement programmes into account, around \n5,000 to 7,500 people remain for whom new job perspectives would have to be \nfound with a phase-out of coal by 2040 at the latest (ref. 51, page 25). In the case \nof an earlier phase-out, or if jobs that are only indirectly dependent on the coal \nindustry are considered, this number rises.\nTo get a complete picture of the effects on employment, the number of newly \ncreated jobs, especially in the renewable sector, must be considered along with the \nnumber of jobs lost51. According to the German government, 330,000 jobs had \nbeen created in the country\u2019s renewable energy sector as of 201559. Phasing out \ncoal could lead to the creation of new jobs not only in the renewable sector but \nalso in other sectors of the economy. Moreover, after closure of the current open-\ncast mines, jobs in restoring the destroyed landscapes will be created or remain \nin place for longer periods of time. Some studies therefore conclude that the net \nemployment effects of phasing out coal could indeed be positive51,60. To account for \nboth positive and negative employment effects, we included two separate attributes, \nranging from 5,000 to 20,000 jobs each.\nFinally, as evidenced by the final report of the German Coal Commission, a \npolicy to phase out coal would need to entail specific measures for supporting \nthe transformation of the regional economy. For example, such a policy might \nprovide financing for early retirement and retraining programmes for coal industry \nemployees51 and/or prioritize deployment of renewable energy in the coal regions. \nOther conceivable measures include investment in modern infrastructure (for \nexample public transport, electric vehicles and digitalization), incentives for the \ncreation of new businesses (such as start-up funding), and public investment \nin research and development. Opinion surveys31 on phasing out coal have not \ninvestigated whether the public has a pronounced preference for particular \nmeasures in supporting the structural transformation, and whether this attribute is \nmore important than others.\nTable 1 provides a summary of the policy attributes and levels. All attributes \nand their levels were briefly explained to study participants before the choice \nexperiment started. The choice experiment itself consisted of eight successive \nrounds. In each round, participants were presented with two policy scenarios for \na phase-out of coal in which the levels of attributes were randomly varied both \nwithin and across the binary comparisons. To prevent order effects, the order in \nwhich the attributes appeared in the description of scenarios was randomized \nacross respondents but fixed for each respondent. At the end of each round, \nparticipants had to evaluate the scenarios using two different scales. First,  \nthey were asked to indicate which of the two scenarios they preferred  \n(forced-choice outcome). Second, participants were asked to provide a more \ndetailed evaluation of the two scenarios, using a scale from 1 (\u2018very poor\u2019) to 7 \n(\u2018very good\u2019) (rating outcome).\nMeasurement of moderators. Party identification was measured with the two-\nstep approach used in the Socio-Economic Panel, which has been conducted since \n1984 by the German Institute for Economic Research61. In a first step, respondents \nwere asked whether they hold a preference for a specific party. If they replied in the \naffirmative (which was the case for 1,275 respondents, or 64% of the main sample), \nthey were then asked which of the seven parties represented in the lower house of \nthe German parliament they identify with, or whether they preferred another party.\nTo measure perceived scientific consensus, respondents were asked to indicate \nthe percentage of climate scientists worldwide who think that the increased \nconcentration of carbon dioxide in the atmosphere since the middle of the \n20th century is primarily due to human activity. They used a slider to choose a \npercentage between 0 and 100%.\nStrong ties with the coal industry were measured by asking whether \nrespondents themselves or someone they know works in a coal mine or a coal-fired \npower station or has done so in the past.\nData analysis. The fully randomized design allows us to simultaneously estimate \nthe causal effects of multiple treatment components based on simple linear \nregression26. Hence, average marginal component effects (AMCEs) were calculated \nusing a simple linear regression estimator with standard errors clustered by \nrespondent, using Stata 14.2 (by StataCorp, see https://www.stata.com/stata14/). \nThe dependent variable is based on the rating scale, and the models include sets of \ndummy variables for the values of all attribute levels.\nTo ease interpretation of the results, we dichotomized the obtained data with \nthe rating scale using the median (which is 4) as the cut-off value. The resulting \ndependent variable \u2018Phase-out Support\u2019 is hence coded 0 for cases where a \nrespondent rated a proposal as poor to neutral (1 to 4) and 1 for cases where (s)he \nrated a proposal as positive (5 to 7). The rationale for using the rating outcome as \nthe dependent variable (instead of the forced-choice outcome) is that it may allow \nfor a more fine-grained assessment of preferences. In the first task (forced choice), \nrespondents had to choose one out of two scenarios in each of eight rounds. \nHowever, the comparison of scenarios may include instances where respondents \nhave either strong preferences for or against both proposals\u2014a situation that \ncannot be meaningfully ascertained by a forced-choice outcome. By contrast, in \nthe rating task respondents could appraise both scenarios independently and on a \nmore fine-grained scale. Nevertheless, replicating the analyses based on the forced-\nchoice outcome leads to substantively the same results (see Supplementary Fig. 2).\nSamples. The choice experiment was implemented in an online survey, which was \nfielded between December 2017 and January 2018. Study participants were drawn \nfrom the opt-in online consumer panel operated by Kantar/Lightspeed, which \nincludes more than 230,000 registered individuals in Germany62. From this panel, \na nonprobability but representative sample of 2,161 Germans entitled to vote at \nnational elections was drawn using an algorithm to match the census population \nas closely as possible on age, gender and household income. Supplementary Table \n1 shows that the sample matches the German population well in terms of age and \ngender. With regard to income, both low-income and high-income households are \nunder-represented. However, given the fact that we also allowed respondents to \nprovide no answer, the deviations appear to be relatively small overall.\nThe two additional regional samples for Rhineland (n\u2009=\u2009533) and Lusatia \n(n\u2009=\u2009501) were drawn from the same consumer panel. As the two coal regions \ndo not by themselves constitute administrative units, the target population for \neach region was defined on the basis of postal codes covering all towns and \nmunicipalities that border the open-pit coal mines. The final lists include 53 postal \ncodes for the Rhineland and 92 postal codes for Lusatia. It is difficult to assess the \nrepresentativeness of the regional samples, as no data comprising the distribution \nof socio-demographic variables for exactly these regions are readily available. \nCompared to the German population as a whole, the two regional samples show \nsome deviations with regard to gender and age. The distribution of income varies \nbetween both samples: while the Rhineland sample includes more high-income \nindividuals than the German sample, the Lusatia sample includes higher shares \nNature Energy | VOL 4 | OCTOBER 2019 | 856\u2013863 | www.nature.com/natureenergy\n861\n\nArticles\nNATuRe EneRGy\nof low-income individuals. This is in line with the different economic conditions \nbetween western and eastern Germany.\nTo identify random responders, we implemented a short attention test \nimmediately after the choice experiment. All analyses shown in the paper are \nbased solely on the responses of all participants who passed this test. Hence, the \nfinal samples consist of 1,984 (Germany), 491 (Rhineland) and 473 (Lusatia) \nrespondents (see Supplementary Table 1). As can be inferred from Supplementary \nTable 1, 247 participants failed to answer the attention test correctly across samples.\nEthics. We have complied with all relevant ethical regulations and guidelines for \nstudy procedures set forth by the Ethics Committee of the University of St.Gallen. \nThe survey for this study was fielded by Kantar/Lightspeed, and all respondents \nwere first informed about the nature of the study before being asked to consent.\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nReplication data for the study are available in the Harvard Dataverse with the \nidentifier https://doi.org/10.7910/DVN/TEFCBL63.\nCode availability\nReplication code for the study is available in the Harvard Dataverse with the \nidentifier https://doi.org/10.7910/DVN/TEFCBL63.\nReceived: 16 April 2019; Accepted: 5 August 2019;  \nPublished online: 16 September 2019\nReferences\n\t1.\t McGlade, C. & Ekins, P. The geographical distribution of fossil fuels unused \nwhen limiting global warming to 2 \u00b0C. Nature 517, 187\u2013190 (2015).\n\t2.\t Hansen, J. et\u00a0al. Target atmospheric CO2: where should humanity aim? Open \nAtmos. Sci. J. 2, 217\u2013231 (2008).\n\t3.\t Rockstr\u00f6m, J. et\u00a0al. A roadmap for rapid decarbonization. Science 355, \n1269\u20131271 (2017).\n\t4.\t BP Statistical Review of World Energy (BP, 2018).\n\t5.\t Pfeiffer, A., Hepburn, C., Vogt-Schilb, A. & Caldecott, B. Committed \nemissions from existing and planned power plants and asset stranding \nrequired to meet the Paris Agreement. Environ. Res. Lett. 13, 054019 (2018).\n\t6.\t Fri, R. W. & Savitz, M. L. Rethinking energy innovation and social science. \nEnergy Res. Soc. Sci. 1, 183\u2013187 (2014).\n\t7.\t Rogge, K. S., Kern, F. & Howlett, M. Conceptual and empirical advances in \nanalysing policy mixes for energy transitions. Energy Res. Soc. Sci. 33,  \n1\u201310 (2017).\n\t8.\t Montt, G. et\u00a0al. Does climate action destroy jobs? An assessment of the \nemployment implications of the 2-degree goal. Int. Labour Rev. 157,  \n519\u2013556 (2018).\n\t9.\t Tvinnereim, E. & Ivarsflaten, E. Fossil fuels, employment, and support for \nclimate policies. Energy Policy 96, 364\u2013371 (2016).\n\t10.\tMorton, T. & M\u00fcller, K. Lusatia and the coal conundrum: the lived \nexperience of the German Energiewende. Energy Policy 99, 277\u2013287 (2016).\n\t11.\tKuchler, M. & Bridge, G. Down the black hole: sustaining national \nsocio-technical imaginaries of coal in Poland. Energy Res. Soc. Sci. 41, \n136\u2013147 (2018).\n\t12.\tCarley, S., Evans, T. P. & Konisky, D. M. Adaptation, culture, and the energy \ntransition in American coal country. Energy Res. Soc. Sci. 37, 133\u2013139 (2017).\n\t13.\tIPCC: Summary for Policymakers. In Global Warming of 1.5\u2009\u00b0C. (eds \nMasson-Delmotte, V. et\u00a0al.) (World Meteorological Organization, 2018).\n\t14.\tRenn, O. & Marshall, J. P. Coal, nuclear and renewable energy policies in \nGermany: from the 1950s to the \u201cEnergiewende\u201d. Energy Policy 99,  \n224\u2013232 (2016).\n\t15.\tLauber, V. & Jacobsson, S. In The Triple Challenge for Europe: Economic \nDevelopment, Climate Change, and Governance (eds Fagerberg, J., Laestadius, \nS. & Martin, B. R.) 173\u2013203 (Oxford Univ. Press, 2015).\n\t16.\tWronski, R. & Fiedler, S. Was Strom wirklich kostet: Vergleich der staatlichen \nF\u00f6rderungen und gesamtgesellschaftlichen Kosten von konventionellen und \nerneuerbaren Energien (Forum \u00d6kologisch-Soziale Marktwirtschaft e.V., \n2017); https://www.greenpeace-energy.de/fileadmin/docs/publikationen/\nStudien/2017-10-Was_Strom_wirklich_kostet_lang.pdf\n\t17.\tW\u00fcstenhagen, R. & Bilharz, M. Green energy market development in \nGermany: effective public policy and emerging customer demand. Energy \nPolicy 34, 1681\u20131696 (2006).\n\t18.\tEnergiebilanzen, A.G. Bruttostromerzeugung in Deutschland ab 1990 nach \nEnergietr\u00e4gern. http://www.ag-energiebilanzen.de/index.php?article_\nid=29&fileName=20161216_brd_stromerzeugung1990-2016.pdf (2019).\n\t19.\tIPCC Renewable Energy Sources and Climate Change Mitigation (Cambridge \nUniv. Press, 2011).\n\t20.\tLaird, F. N. in Germany\u2019s Energy Transition: A Comparative Perspective (eds \nHager, C. & Stefes, C.) 111\u2013131 (Palgrave Macmillan, 2016).\n\t21.\tMeckling, J., Kelsey, N., Biber, E. & Zysman, J. Winning coalitions for climate \npolicy. Science 349, 1170\u20131171 (2015).\n\t22.\tFraune, C. & Knodt, M. Sustainable energy transformations in an age of \npopulism, post-truth politics, and local resistance. Energy Res. Soc. Sci. 43, \n1\u20137 (2018).\n\t23.\tLockwood, M. Right-wing populism and the climate change agenda: \nexploring the linkages. Env. Polit. 27, 712\u2013732 (2018).\n\t24.\tBundesministerium f\u00fcr Wirtschaft und Energie Abschlussbericht, Kommission \n\u201cWachstum, Strukturwandel und Besch\u00e4ftigung\u201d (Bundesministerium f\u00fcr \nWirtschaft und Energie, 2019).\n\t25.\tDeutsche Bundesregierung Klimaschutzplan 2050: Klimaschutzpolitische \nGrunds\u00e4tze und Ziele der Bundesregierung (Bundesministerium f\u00fcr Umwelt, \nNaturschutz und nukleare Sicherheit, 2016).\n\t26.\tHainmueller, J., Hopkins, D. J. & Yamamoto, T. Causal inference in conjoint \nanalysis: understanding multidimensional choices via stated preference \nexperiments. Polit. Anal. 22, 1\u201330 (2013).\n\t27.\tDharshing, S., Hille, S. L. & W\u00fcstenhagen, R. The influence of political \norientation on the strength and temporal persistence of policy framing \neffects. Ecol. Econ. 142, 295\u2013305 (2017).\n\t28.\tStokes, L. C. & Warshaw, C. Renewable energy policy design and framing \ninfluence public support in the United States. Nat. Energy 2, 17107 (2017).\n\t29.\tLeipprand, A. & Flachsland, C. Regime destabilization in energy transitions: \nthe German debate on the future of coal. Energy Res. Soc. Sci. 40,  \n190\u2013204 (2018).\n\t30.\tHuneke, F. & Niggemeier, M.-L. Jamaica and the coal. Energy BrainPool \nhttps://blog.energybrainpool.com/en/jamaica-and-the-coal-energy-market-2-\n0-passes-functioning-test/ (2018).\n\t31.\tSetton, D., Matuschke, I. & Renn, O. Soziales Nachhaltigkeitsbarometer der \nEnergiewende 2017 (Institute for Advanced Sustainability Studies e.V., 2017).\n\t32.\tMehrheit w\u00fcnscht sich m\u00f6glichst schnellen Braunkohle-Ausstieg. Infratest \nDimap https://www.infratest-dimap.de/umfragen-analysen/bundesweit/\numfragen/aktuell/mehrheit-wuenscht-sich-moeglichst-schnellen-braunkohle-\nausstieg/ (2019).\n\t33.\tCook, J., Lewandowsky, S. & Ecker, U. K. H. Neutralizing misinformation \nthrough inoculation: exposing misleading argumentation techniques reduces \ntheir influence. PLoS ONE 12, e0175799 (2017).\n\t34.\tDing, D., Maibach, E. W., Zhao, X., Roser-Renouf, C. & Leiserowitz, A. \nSupport for climate policy and societal action are linked to perceptions about \nscientific agreement. Nat. Clim. Change 1, 462\u2013466 (2011).\n\t35.\tLewandowsky, S., Oberauer, K. & Gignac, G. E. NASA faked the moon \nlanding-therefore, (climate) science is a hoax: an anatomy of the motivated \nrejection of science. Psychol. Sci. 24, 622\u2013633 (2013).\n\t36.\tvan\u00a0der Linden, S., Maibach, E. & Leiserowitz, A. Improving public \nengagement with climate change: five \u201cbest practice\u201d insights from \npsychological science. Perspect. Psychol. Sci. 10, 758\u2013763 (2015).\n\t37.\tCook, J. et\u00a0al. Consensus on consensus: a synthesis of consensus estimates on \nhuman-caused global warming. Environ. Res. Lett. 11, 048002 (2016).\n\t38.\tLeiserowitz, A. et\u00a0al. Climate Change in the American Mind: March 2018  \n(Yale Univ. and George Mason Univ., 2018).\n\t39.\tDavidson, D. J. Exnovating for a renewable energy transition. Nat. Energy 4, \n254\u2013256 (2019).\n\t40.\tGeels, F. W., Sovacool, B. K., Schwanen, T. & Sorrell, S. Sociotechnical \ntransitions for deep decarbonization. Science 357, 1242\u20131244 (2017).\n\t41.\tHeyen, D. A., Hermwille, L. & Wehnert, T. Out of the comfort zone! \nGoverning the exnovation of unsustainable technologies and practices. Gaia \n26, 326\u2013331 (2017).\n\t42.\tKivimaa, P. & Kern, F. Creative destruction or mere niche support? \nInnovation policy mixes for sustainability transitions. Res. Policy 45,  \n205\u2013217 (2016).\n\t43.\tTurnheim, B. & Geels, F. W. Regime destabilisation as the flipside of  \nenergy transitions: lessons from the history of the British coal industry, \n1913\u20131997. Energy Policy 50, 35\u201349 (2012).\n\t44.\tBroockman, D. E. & Skovron, C. Bias in perceptions of public opinion among \nAmerican political elites. Am. Polit. Sci. Rev. 112, 542\u2013563 (2018).\n\t45.\tHertel-Fernandez, A., Mildenberger, M. & Stokes, L. C. Legislative staff and \nrepresentation in Congress. Am. Polit. Sci. Rev. 113, 1\u201318 (2019).\n\t46.\tBreetz, H., Mildenberger, M. & Stokes, L. The political logics of clean energy \ntransitions. Bus. Polit. 20, 492\u2013522 (2018).\n\t47.\tMarkard, J. The next phase of the energy transition and its implications for \nresearch and policy. Nat. Energy 3, 628\u2013633 (2018).\n\t48.\tVona, F. Job losses and political acceptability of climate policies: why the \n\u2018job-killing\u2019 argument is so persistent and how to overturn it. Clim. Policy 19, \n524\u2013532 (2019).\n\t49.\tBechtel, M. M. & Scheve, K. F. Mass support for global climate agreements \ndepends on institutional design. Proc. Natl Acad. Sci. USA 110,  \n13763\u201313768 (2013).\nNature Energy | VOL 4 | OCTOBER 2019 | 856\u2013863 | www.nature.com/natureenergy\n862\n\nArticles\nNATuRe EneRGy\n\t50.\tGampfer, R., Bernauer, T. & Kachi, A. Obtaining public support for \nnorth-south climate funding: evidence from conjoint experiments in donor \ncountries. Glob. Environ. Change 29, 118\u2013126 (2014).\n\t51.\tKohleausstieg jetzt einleiten (Sachverst\u00e4ndigenrat f\u00fcr Umweltfragen, 2017).\n\t52.\tMatthes, D. F. C. et\u00a0al. Zukunft Stromsystem: Kohleausstieg 2035 - Vom Ziel \nher denken (WWF Germany, 2017).\n\t53.\tPietroni, A., Fernahl, A., Perez Linkenheil, C., Niggemaier, M. & Huneke, F. \nKlimaschutz durch Kohleausstieg: Wie ein Ausstieg aus der Kohle Deutschlands \nKlimaziele erreichbar macht, ohne die Versorgungssicherheit zu gef\u00e4hrden \n(energy brainpool, 2017).\n\t54.\tH\u00f6hne, N., Kuramochi, T., Sterl, S. & R\u00f6schel, L. Was bedeutet das Pariser \nAbkommen fu\u0308r den Klimaschutz in Deutschland? (NewClimate Institute, \n2016).\n\t55.\tConnolly, K. G7 leaders agree phase out fossil fuel use end of century.  \nThe Guardian https://www.theguardian.com/world/2015/jun/08/g7-leaders-\nagree-phase-out-fossil-fuel-use-end-of-century (2015).\n\t56.\tStratmann, K., Sigmund, T., Flauger, J. & Kersting, S. Streit um Enegiewende: \nKohleausstieg entzweit die Wirtschaft. Handelsblatt https://www.handelsblatt.\ncom/unternehmen/industrie/streit-um-enegiewende-kohleausstieg-entzweit-\ndie-wirtschaft-/20550070.html (2017).\n\t57.\tEcke, J. Gutachten: Sozialvertr\u00e4gliche Ausgestaltung eines Kohlekonsenses  \n(ver.di - Vereinte Dienstleistungsgewerkschaft / enervis energy advisors \nGmbH, 2016).\n\t58.\tElf Eckpunkte fu\u0308r einen Kohlekonsens. Konzept zur schrittweisen \nDekarbonisierung des deutschen Stromsektors (Langfassung) Version 1.2 \n(Agora Energiewende, 2016).\n\t59.\tO\u2019Sullivan, M., Edler, D. & Lehr, U. Bruttobesch\u00e4ftigung durch erneuerbare \nEnergien in Deutschland und verringerte fossile Brennstoffimporte durch \nerneuerbare Energien und Energieeffizienz (Bundesministerium f\u00fcr Wirtschaft \nund Energie, 2016); https://www.bmwi.de/Redaktion/DE/Downloads/S-T/\nbruttobeschaeftigung-erneuerbare-energien-monitioringbericht-2015.\npdf?__blob=publicationFile&v=11\n\t60.\tDehnen, N., Mattes, A. & Traber, T. Die Besch\u00e4ftigungseffekte der \nEnergiewende: Eine Expertise f\u00fcr den Bundesverband WindEnergie e.V. und die \nDeutsche Messe AG (Deutsches Institut f\u00fcr Wirtschaftsforschung, 2015).\n\t61.\tDIW Glossar: Parteibindung. Deutsches Institut f\u00fcr Wirtschaftsforschung \nhttps://www.diw.de/de/diw_01.c.413409.de/presse/diw_glossar/parteibindung.\nhtml (2018).\n\t62.\tGlobal Panel Book (Lightspeed, 2016); http://www.lightspeedresearch.com/\nwp-content/uploads/2016/09/Lightspeed_PanelBook_Q4_2016.pdf\n\t63.\tRinscheid, A. Replication data for: Germany\u2019s decision to phase out coal by \n2038 lags behind citizens\u2019 timing preferences. Harvard Dataverse https://doi.\norg/10.7910/DVN/TEFCBL (2019).\n\t64.\tNationale Inventarberichte zum Deutschen Treibhausgasinventar 1990 bis 2017 \n(Stand 01/2019) (Umweltbundesamt, 2019).\nAcknowledgements\nWe acknowledge support by the Andlinger Center for Energy and the Environment \nat Princeton University, the Swiss Center of Competence for Energy Research SCCER \nCREST, the Swiss National Science Foundation (grant no. P1SGP1_174939) and \nGreenpeace Germany, who funded data collection. Design of the research project and \ndata analysis was the sole responsibility of the authors.\nAuthor contributions\nA.R. designed the study and analysed the data. A.R. and R.W. wrote the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-019-0460-9.\nReprints and permissions information is available at www.nature.com/reprints.\nCorrespondence and requests for materials should be addressed to A.R.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2019\nNature Energy | VOL 4 | OCTOBER 2019 | 856\u2013863 | www.nature.com/natureenergy\n863\n\n1\nnature research  |  reporting summary\nOctober 2018\nCorresponding author(s):\nAdrian Rinscheid\nLast updated by author(s): Jul 26, 2019\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nSawtooth.\nData analysis\nStata 14.2. \nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. \nWe strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nReplication data and code for the study are available in the Harvard Dataverse with the identifier https://doi.org/10.7910/DVN/TEFCBL \nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\n\n2\nnature research  |  reporting summary\nOctober 2018\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThe study uses quantitative data based on a survey with an embedded choice experiment.\nResearch sample\nThe main sample consists of 2,161 Germans entitled to vote at national elections. This sample is a nonprobability but representative \nsample, matching  the German population well in terms of age and gender. Two additional samples were fielded in the coal regions \nRhineland (n = 533) and Lusatia (n = 501). It is difficult to assess the demographic representativeness of the additional samples, as no \ndata comprising the distribution of socio-demographic variables for exactly these regions are readily available. Supplementary Table 1 \ngives more details on how the 3 samples compare with the German population in terms of distribution of sex, age, and household \nincome. All samples were collected through online sampling, based on the panel operated by Kantar/Lightspeed, which includes 230,000 \nindividuals in Germany.\nSampling strategy\nRespondents were recruited online through the survey firm Kantar/Lightspeed. We used quotas to ensure representativeness of the \ngeneral population. Hard quotas were used on sex (male/female) and age (five groups). Soft quotas were used on household income \n(deciles).\nData collection\nThe surveys were completed online individually by respondents.\nTiming\n18th December 2017 to 8th January 2018.\nData exclusions\nWe excluded the data from respondents who failed to respond correctly to an attention test implemented during the choice experiment \nfrom the analyses shown in the paper. The rationale is that respondents failing the attention test may have provided random responses, \nhence adding random noise to the data. This exclusion citerion was pre-established and reduced the sample size as follows: main sample \n(Germany): 177 exclusions; Rhineland sample: 42 exclusions; Lusatia sample: 28 exclusions.  \nAs a robustness check, we re-ran all analyses with the full samples, including inattentive respondents. None of the results described in \nthe paper change substantially if inattentive individuals are included in the analyses (for the full results, see Supplementary Tables 2, 4, 6, \n8 and 10).\nNon-participation\nThere were no drop-outs.\nRandomization\nFor each participant, the choice experiment consisted of eight successive rounds with two pairs of policy scenarios. The levels of \nattributes were randomly varied both within and across the binary comparisons. To prevent order effects, the order in which the \nattributes appeared was randomized across respondents but fixed for each respondent. \nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology\nAnimals and other organisms\nHuman research participants\nClinical data\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nHuman research participants\nPolicy information about studies involving human research participants\nPopulation characteristics\nSee above.\nRecruitment\nSubjects were recruited through the online survey panel operated by Kantar/Lightspeed. This opt-in sample is likely to under-\nrepresent individuals who have little or no internet access.\nEthics oversight\nEthics Committee of the University of St.Gallen\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\n\n\n    Scientific Research Findings:", "answer": "We find that German voters prefer a coal phase-out by 2025, whereas policies aiming at a later date find less support. Voters would uphold their support for a phase-out by 2025 even in light of additional household costs, although every \u20ac10 increase in annual household cost decreases voter support by about seven percentage points. Voters in the two coal-producing regions affected by the phase-out (Rhineland and Lusatia) also prefer an earlier phaseout. Individuals who are aware of the broad scientific consensus that climate change is caused by humans have particularly strong preferences for an earlier phase-out. In contrast to the government commission\u2019s proposal, the more ambitious timeline preferred by voters would increase the chances of achieving the goals of the Paris Agreement in Germany. While our study suggests that policymakers may underestimate public willingness to support an expedited coal phase-out, we did not explicitly investigate the role of other stakeholders in the policy process. The study focused on Germany and as such does not allow generalization to other countries.", "id": 30}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-019-0476-1\n1Department of Industrial Engineering and Management Sciences, Northwestern University, Evanston, IL, USA. 2Federal Energy Regulatory Commission, \nWashington, DC, USA. *e-mail: jacobmays@cornell.edu\nS\nince the inception of liberalized electricity markets in the \nlate 1990s, merchant investment into new generation capac-\nity in the United States has been dominated by gas-fired units. \nThe simplest explanation for this trend points to marked improve-\nments in the efficiency of combined-cycle plants along with, more \nrecently, low natural gas prices brought about by the shale revo-\nlution. Without attempting to compare the magnitude of the two \neffects, we propose a second explanation: distortions brought about \nby capacity markets in the absence of more complete risk trading.\nThree alternative paradigms govern investments in generation \ncapacity: vertical integration, energy-only markets and installed \ncapacity (ICAP) markets1. In traditional vertically integrated sys-\ntems, investors are guaranteed a rate of return by a regulatory pro-\ncess, which effectively shifts most risk to ratepayers and creates \nthe incentive for excess investment. One goal behind deregulation \nwas to encourage a more efficient allocation of this risk2. The theo-\nretically ideal competitive market provides the basis for a second \nparadigm, energy-only markets. Under the energy-only design, \ngenerators earn all their revenue through the sale of energy and \nancillary services. Prices rise substantially during times of scarcity, \nwhich allows generators to recover their fixed costs. In practice, a \nlack of price-responsive demand, combined with inconsistencies \nbetween the market clearing process and actions taken by opera-\ntors on the grounds of reliability, hampered the formation of effi-\ncient scarcity prices3,4. An administrative response is to introduce an \noperating reserves demand curve that reflects the probability that \nthe system operator will need to take emergency actions (for exam-\nple, voltage reductions or rolling blackouts) to prevent a cascad-\ning failure5. As the probability of such actions approaches one, the \nprice grows to an estimated value of the lost load, which is typically \ntwo orders of magnitude larger than average prices (for example, \nUS$9,000\u2009MWh\u20131 for the Electric Reliability Council of Texas). With \nthe exception of the Electric Reliability Council of Texas, wholesale \nmarket operators in the United States opted for the third paradigm. \nUnder the ICAP design, load-serving entities are required to either \nsupply or procure an administratively determined level of capacity. \nEnergy prices in these markets are capped at a much lower level \n(for example, US$2,000\u2009MWh\u20131 or lower), and revenue from the \nsale of energy and ancillary services is supplemented by payments \nfor capacity. The intent of these capacity payments is to balance the \n\u2018missing money\u2019 that results from the cap and other price-suppress-\ning actions to restore the outcomes that would be achieved in an \nideal energy-only market6\u20138.\nWe are not the first to suggest that capacity markets bring about \na distortion of the technology mix. Indeed, this assertion is com-\nmon enough that the overview in Cramton et\u00a0al.7 deems it \u2018a deep \nconfusion concerning capacity markets\u2019 that \u2018takes many forms and \ncontains contradictory views\u2019. In this context, our contribution lies \nin formalizing a mechanism by which distortion could arise.\nTo illustrate this possibility, we developed a stochastic equilib-\nrium model of the type first described in Ehrenmann and Smeers9, \nas well as an algorithm to solve large-scale instances of the problem. \nThe fuel neutrality of capacity markets relative to an ideal energy-\nonly market relies on the assumption embedded in the classic opti-\nmization framework that the cost of capital is exogenous to the \nmarket design. The model in this article relaxes that assumption, \nwhich enables us to take into account the stabilizing effect that the \nintroduction of capacity mechanisms has on generator revenues. \nOur numerical examples describe a simple system with three gener-\nation technologies: a baseload technology with high capital costs but \nlow operating costs (for example, nuclear), a renewable technology \nwith a similar cost profile but variable output (for example, wind) \nand a peaker technology with low capital costs but high operating \ncosts (for example, gas). Results on these test systems suggest that \nan increased reliance on existing capacity market structures without \nthe emergence of other forms of risk trading is likely to work against \ntechnologies with low operating costs, a category that includes all of \nthe most scalable forms of low-carbon generation.\nResource adequacy and risk trading\nSeparate from restoring the missing money required for proper \ninvestment incentives, advocates for capacity markets point to the \nbenefits that the ICAP design provides in reducing risk. By replac-\ning highly volatile scarcity prices with a more consistent monthly \nor annual payment, capacity mechanisms provide a financial hedge \nfor both generators and loads. In the traditional analysis, capacity \nAsymmetric risk and fuel neutrality in electricity \ncapacity markets\nJacob Mays\u200a \u200a1,2*, David P. Morton\u200a \u200a1 and Richard P. O\u2019Neill2\nIn many liberalized electricity markets, power generators can receive payments for maintaining capacity through capacity \nmarkets. These payments help stabilize generator revenues, making investment in capacity more attractive for risk-averse \ninvestors when other outlets for risk trading are limited. Here we develop a heuristic algorithm to solve large-scale stochastic \nequilibrium models describing a competitive market with incomplete risk trading. Introduction of a capacity mechanism has \nan asymmetric effect on the risk profile of different generation technologies, tilting the resource mix towards those with lower \nfixed costs and higher operating costs. One implication of this result is that current market structures may be ill-suited to \nfinancing low-carbon resources, the most scalable of which have high fixed costs and near-zero operating costs. Development \nof new risk trading mechanisms to replace or complement current capacity obligations could lead to more efficient outcomes.\nNature Energy | VOL 4 | NOVEMBER 2019 | 948\u2013956 | www.nature.com/natureenergy\n948\n\nArticles\nNaTure Energy\npayments have the impact of a call option with a strike price equal \nto the price cap for the market2,7,10. Loads make an upfront payment \nto generators, in exchange for which the effective price they see in \nthe spot market does not exceed the price cap. As energy markets \ndo not produce an accurate underlying price in times of scarcity, \nreal-world capacity markets do not precisely replicate an option. \nWe discuss some of the implementation challenges this causes in \nSupplementary Note 1. In the ideal case, however, incentives on \nthe margin are maintained: loads must pay high spot prices if they \nconsume more than they have procured in advance, and generators \nhave to buy back their position if they are unable to perform.\nIt has long been acknowledged that risk aversion could lead to a \nsuboptimal investment in generation capacity9,11,12. However, prior \nstudies do not demonstrate any consistent technological shift. In \nenergy-only markets, peaking units face substantial risk due to \ntheir dependence on the occurrence of high prices in a small num-\nber of hours. At the same time, several authors emphasized the \nnatural hedge enjoyed by gas generators, which often set the price \nof electricity13\u201316. The ability to trade risk is a key element of this \ndiscussion. With complete risk trading, along with the standard \nassumptions of competitive markets, an equilibrium would achieve \nthe (risk-aware) social optimum17\u201320. The connection between equi-\nlibrium and optimum implies that the impact of risk aversion on the \ngeneration mix can be in any direction depending on the types of \nrisk that most concern market participants. This theoretical result, \nhowever, depends on the presence of \u2018complete\u2019 markets, which \ninclude a security that corresponds to every possible future state of \nthe world. In this context, our focus is less on incompleteness than \non asymmetry: the introduction of a capacity market amounts to a \ntrade against a subset of future world states that peaking plants are \nparticularly concerned about. In the PJM interconnection (PJM), \nfor example, a unit with a marginal cost of US$10\u2009MWh\u20131 could \nexpect to earn 17% of its operating profits from the capacity market, \nwhereas a unit with a US$100\u2009MWh\u20131 marginal cost could expect \n90% (Supplementary Note 2). Along these lines, our results suggest \na different interpretation of the natural hedge of gas. By itself, the \nhedge does not confer an advantage. It is only when combined with \na capacity mechanism, which removes risk related to the frequency \nof scarcity events, that the hedge becomes an advantage.\nThe size of this advantage depends on the availability of other \noutlets for risk trading. In practice, the evidence suggests that mar-\nkets in risk are far from complete21. A natural resolution to the \nproblem occurs if investors in generation are able to sign long-term \ncontracts with loads, who are exposed to price risk in the oppo-\nsite direction. With the support of state and corporate procurement \nprogrammes, a robust market has developed for renewable genera-\ntors that seek long-term power purchase agreements22. In areas with \ncompetitive retail markets, however, contracts rarely exceed two \nyears outside these programs23. Small retailers are unable to com-\nmit to long-term contracts without risking customer defection in \nperiods when wholesale prices drop below the contracted rate11. \nThe phenomenon of insufficient demand-side interest in hedging \nis seen as characteristic to commodity markets beyond electricity24, \nespecially in the presence of transaction costs25.\nIf they are unable to secure a power purchase agreement, mer-\nchant generators are often able to hedge risk on timescales of five \nto seven years through contracts with financial traders able to bear \nsome of the risk26. However, these bilateral deals cannot match the \nliquidity of centralized capacity markets, which leads some to pro-\npose formalized exchanges27,28. Owing to the proprietary nature of \nthese arrangements, it is a challenge to assess the liquidity in this \nmarket. Nevertheless, the available evidence suggests some endo-\ngeneity between cash flows and the cost of capital. Among pub-\nlicly available ratings, debt issued for projects in the energy-only \nElectric Reliability Council of Texas market receive ratings of B \nand CCC, whereas projects in PJM typically receive a higher BB  \nrating29. Separate evidence comes from renewable support schemes \nin Europe. Although some programmes (for example, feed-in tariffs) \nlead to predictable revenue streams, others (for example, quotas) \nresult in payments that vary with market conditions. The relative \nease of financing stable revenues can lead to a greater deployment \nfor the same amount of government support30,31.\nEven if the size of the effect is small at present, insufficient risk \ntrading may have large consequences for long-term market design. \nIn general, long-run models predict that the incorporation of solar \nand wind will lead to lower average electricity prices but greater vol-\natility, which prompted many to consider how capacity mechanisms \nmay need to evolve to guarantee resource adequacy1,32\u201335.\nTwo-generator example\nTo illustrate the impact that incomplete markets in risk may have \non the capacity mix that arise in equilibrium, we developed two \nsimple test systems. The first features two technologies: a baseload \nresource with a high capital cost but a low and certain operating \ncost, and a peaking resource with a low capital cost but a high and \nuncertain operating cost. The second example, developed in the \nnext section, adds a renewable resource with variable availability. \nIn the context of current US markets, these three technologies can \nbe thought of as existing nuclear plants, many of which are con-\nsidering retirement in the near term, and new gas plants and wind \nfarms, which are being constructed at a rapid rate. With that said, \nthe parameters were chosen to best illustrate the impact of risk trad-\ning, rather than to recreate or predict market outcomes. In keep-\ning with US market design principles, the focus is on economic \nefficiency, with no explicit environmental goals within the market. \nSupport mechanisms for carbon-free generation are assumed to be \nreflected through the investment and energy costs associated with \neach technology. Unlike most US markets, we did not implement a \nmandatory capacity market: resource decisions are driven directly \nby the value of load expressed by consumers. Instead, we mim-\nicked a capacity market through a call option with a strike price of \nUS$1,000\u2009MWh\u20131. If the cost of capital were exogenous to the mar-\nket design, this would have no effect on the resulting capacity mix7,8. \nWe also considered a future settling at US$50\u2009MWh\u20131.\nFor the experiments, the nominal demand is based on the load \nduration curve of PJM in 2017, during which demand ranged from \n58 to 146\u2009GW. The first example focuses on two generation technol-\nogies, baseload and peaker, and two sources of uncertainty, demand \nand the fuel cost of the peaker. For both technologies, 90% of ICAP \nis available in every time block. Fuel prices and demand shifters are \nmodelled as random variables with 100 equally likely year-long sce-\nnarios for the dispatch problem. We considered two contracts: a call \noption with a strike price of US$1,000\u2009MWh\u20131 and a future settling \nat US$50\u2009MWh\u20131. The probabilistic model is discussed in Methods, \nand additional details on the scenario assumptions are given in \nSupplementary Note 2 and Supplementary Tables 1 and 2.\nTo characterize the risk attitude of market participants, we used \na weighted sum of the expected value of surplus, with weight \u03b2, and \nconditional value at risk (CVaR) of the \u03b1-level tail of surplus, with \nweight 1\u2009\u2212\u2009\u03b2. CVaR is a coherent risk measure that lends itself to \ninclusion in a mathematical optimization model36,37. Lower values \nof \u03b1 and \u03b2 imply a greater risk aversion, with either \u03b1\u2009=\u20091 or \u03b2\u2009=\u20091 \nimplying a risk-neutral agent. In real-world contexts, market opera-\ntors have limited ability to ascertain or influence the risk preferences \nof market participants. Accordingly, our results do not focus on the \neffect of risk aversion itself, but instead on the availability of risk \ntrades, a phenomenon that market operators and regulators have \na greater potential to address. Along these lines, our priority when \nchoosing \u03b1 and \u03b2 was not necessarily to match the risk preferences \nof real-world market participants, but to offer a clear illustration of \nthe impact of incomplete markets. Tests with a number of values for \nthese parameters yielded the same directional results.\nNature Energy | VOL 4 | NOVEMBER 2019 | 948\u2013956 | www.nature.com/natureenergy\n949\n\nArticles\nNaTure Energy\nUsing risk parameters \u03b1\u2009=\u2009\u03b2\u2009=\u20090.5 for all the market partici-\npants, the socially optimal mix (that is, the complete trading case) \nincludes a baseload capacity of 89.7\u2009GW and a peaking capacity of \n77.6\u2009GW. We tested cases with no trading, each contract individu-\nally, and both contracts available, and compared the results to this \nsocial optimum. Table 1 shows the capacity mix that arises in each \ncase, the financial trades made by investors in each technology and \nthe decrease in consumer surplus relative to the complete trading \ncase. The Methods section describes the associated mathematical \nmodel, solution algorithm and metric measuring proximity of the \nalgorithm\u2019s solution to an equilibrium.\nIncomplete trading leads to a lower ICAP and significant degra-\ndation of surplus. To achieve an equilibrium close to the complete \ntrading ideal, both contracts must be available.\nThe most striking result in Table 1 occurs in the Option Only \ncase, in which the baseload technology is absent from the equilib-\nrium mix. This observation is particularly important because, as \ndiscussed above, the options contract is designed to have the same \nfinancial impact as the capacity markets currently utilized or under \ndiscussion in many real-world markets. A hint as to the mechanism \nat work is provided by the trades made in the Both Contracts case. \nThe peaker technology prefers to sell options, selling 78.4\u2009GW of \noptions contracts against only 1.2\u2009GW of futures. The baseload tech-\nnology exhibits the opposite preference, selling 80.7\u2009GW of futures \nbut no options. Tables 2 and 3 show the reason for these trading \npreferences. Table 2 shows the payouts for the futures contract in \nthe Both Contracts case. The 50 bold cells in the table represent \nthe worst-case realizations of demand and fuel cost for the basel-\noad technology in the No Trading case. As \u03b1\u2009=\u20090.5 in this example, \nthese 50 scenarios are the worst-case scenarios identified by the \nCVaR calculation for the baseload investors. Analogously, Table 3 \nshows the payouts for the option contract in the Both Contracts case \nalong with the worst-case combinations of demand and fuel costs \nfor the peaker technology. The difference in the bold cells between \nthe two tables reflects the risk exposure of the two technologies: \nalthough investors in the peaking technology are only concerned \nabout lower-than-expected demand, investors in the baseload tech-\nnology are concerned about both lower-than-expected demand and \nlower-than-expected fuel cost. As we discuss in Methods section, \nemploying CVaR is equivalent to modifying the nominal probability \ndistribution under a risk-neutral objective function. In this numeri-\ncal example, the modified probability mass function for each agent \ncorresponds to placing probability 3\n200\nI\n on the bold scenarios and \n1\n200\nI\n on the regular scenarios in the respective tables (in place of the \nnominal 1\n100\nI\n on all scenarios).\nTo improve their overall risk profile, investors in each technol-\nogy choose to sell contracts with payouts that align with favourable \nunderlying outcomes. Higher payouts for the futures contract cor-\nrespond to non-bold cells in Table 2, whereas higher payouts for the \noptions contract correspond to non-bold cells in Table 3. In cases \nfor which only one trade is available, the equilibrium shifts towards \nthe resource with a risk profile that is better balanced by that trade. \nAccordingly, moving from the No Trading case to the Future Only \ncase increases the baseload capacity from 76.1 to 89.2\u2009GW. The \neffect is more drastic in moving from the No Trading case to the \nOption Only case, which brings the peaking capacity from 88.9 to \n166.7\u2009GW and eliminates the baseload technology from the mix.\nAn additional interesting feature of Table 1 is that, under the \nOption Only case, investors in peaking technology sell a volume of \noptions that exceeds the ICAP. As the assumed availability is 0.9, \nit is clear that these units cannot physically back up the financial \ntrade. Real-world capacity mechanisms typically require physical \nassets. From a liquidity standpoint, a related issue is that the two-\nstage model developed in this article enables generators to hedge \nover their entire operating life, whereas real-world capacity com-\nmitments are often much shorter (for example, one year in PJM).\nIt is a mistake to object that a short delivery period eliminates \nthe financial hedging property of capacity markets: more important \nthan the cash flow for the agreed period is the volatility of future \ncash flows. The option payouts in Table 3, which reflect the distri-\nbution of annual revenues during scarcity periods that might be  \nexpected in an energy-only setting, range from US$0 to 138\u2009kW\u20131\u2009yr\u20131,  \nwith an average of US$41\u2009kW\u20131\u2009yr\u20131 and a coefficient of variation of  \n1.12. The presence of a capacity market reduces this volatility sub-\nstantially. In PJM, for example, capacity market prices have ranged \nfrom US$28 to 60\u2009kW\u20131\u2009yr\u20131 in the six years since the introduction \nof Capacity Performance38, with an average of US$47\u2009kW\u2009yr\u20131 and a \ncoefficient of variation of 0.24. That said, both the physical back-up \nrequirement and the short-term nature of real-world commitments \nmake the level of hedging achieved in Table 1 unlikely. Similar to \nthe strategy in de Maere d\u2019Aertrycke et\u00a0al.39, we mimicked the lack \nof liquidity by limiting the allowed volume of options trades to a \npercentage of ICAP for each technology and recalculated the equi-\nlibrium mix. As shown in Fig. 1, the amount of baseload capacity in \nequilibrium declines as liquidity in the options market increases, to \nreach zero when trade volumes equal 60% of ICAP.\nThree-generator example\nIn the second example, we added a variable generation resource with  \nfour scenarios to govern its availability. The availability scenarios for \nthe variable technology are distinguished in how they are coupled with  \nthe load duration curve. Although each scenario has an average avail-\nability of 37.5% over time, the four are temporally correlated with a  \nfixed load in the following manner: strongly positive, weakly positive, \nweakly negative and strongly negative. Availability for the baseload  \nand peaker generators remains at 0.9. Scenarios for fuel and demand  \nare retained from the two-generator example, which results in a total  \nof 400 equally likely year-long second-stage scenarios. Investment cost  \nand risk parameters were updated to ensure that all three technologies  \nare represented in both the socially optimal mix and the No Trading \nequilibrium. Additional details on these assumptions are given in  \nSupplementary Note 4. In this example, the socially optimal mix \nTable 1 | Equilibria in two-generator example\nNo trading\nOption \nonly\nFuture \nonly\nBoth \ncontracts\nCapacity (GW)\n\u2002Baseload\n76.1\n0.0\n89.2\n89.6\n\u2002Peaker\n88.9\n166.7\n76.8\n77.7\n\u2002Total\n165.1\n166.7\n166.0\n167.3\nTrade volume (GW)\n\u2002Baseload\n\u2003Future\n80.3\n80.7\n\u2003Option\n0.0\n0.0\n\u2002Peaker\n\u2003Future\n31.3\n1.2\n\u2003Option\n167.1\n78.4\nProximity to \nequilibrium (%)\n0.000\n0.000\n0.000\n0.002\nChange in surplus \n(US$\u2009billion\u2009yr\u20131)\n\u221212.3\n\u22126.8\n\u22122.1\n\u22120.0\nFour financial trading cases are shown in order of increasing surplus. The No Trading case leads \nto underinvestment in capacity. The capacity mix tilts towards the peaker technology in the \nOption Only case and towards the baseload technology in the Future Only case. A mix close to \nthe complete trading ideal is achieved in the Both Contracts case. Risk parameters \u03b1\u2009=\u2009\u03b2\u2009=\u20090.5 are \nassumed for the consumer and both generators.\nNature Energy | VOL 4 | NOVEMBER 2019 | 948\u2013956 | www.nature.com/natureenergy\n950\n\nArticles\nNaTure Energy\nincludes a baseload capacity of 31.3\u2009GW, peaking capacity of 104.9\u2009GW  \nand variable capacity of 139.2\u2009GW.\nIn addition to the option and futures contracts used in the pre-\nvious example, we define a unit contingent contract that matches \nthe availability of the variable generator40. The interaction between \navailability and demand in determining the occurrence of scar-\ncity conditions means it is difficult to determine in advance which \ncombinations of random variable realizations will represent the \nworst-case scenarios for market participants. However, the intent \nis to structure one contract that matches the risk profile of each \ntechnology: futures to the baseload technology, options to the peak-\ning technology and unit contingent to the variable technology. The \nequilibrium results are split into two tables, oriented in order of \nincreasing surplus. Table 4 shows the capacity mix that arises when \nat most one trade is available. Comparing the bold cells within each \nrow, it can be seen that introducing any of the trades individually \nresults in more capacity of the corresponding technology. The  \nmarket share of the baseload unit collapses after the introduction of the  \nunit contingent contract. As in the two-generator example, intro-\nduction of the option contract (analogous to current capacity mar-\nkets) results in a complete exit of the baseload technology.\nTable 5 shows the results for the cases in which two or three \ncontracts are available. Here the bold cells demonstrate the effect \nof introducing the third contract. In all three pairwise comparisons, \nthe addition of the third contract shifts the capacity mix towards the \ncorresponding technology.\nAlthough the directional pattern holds when moving from one to \ntwo available trades, the magnitudes are in some instances weaker. \nIn the most extreme example, the Future case and the Future\u2009+\u2009Unit \ncase result in precisely the same equilibrium. With that capacity \nmix, the worst-case outcomes for the consumer involve a low output \nfor the variable technology. Accordingly, the consumer is reluctant \nto purchase unit contingent contracts and none are traded.\nTable 2 | Alignment of futures payout to risk faced by baseload technology\nDemand\nFuel cost\n1\n2\n3\n4\n5\n6\n7\n8\n9\n10\n\u20021\n\u2212225\n\u2212194\n\u2212162\n\u2212131\n\u2212107\n\u221277\n\u221247\n\u221216\n14\n45\n\u20022\n\u2212221\n\u2212189\n\u2212157\n\u2212125\n\u221299\n\u221268\n\u221237\n\u22125\n26\n57\n\u20023\n\u2212218\n\u2212185\n\u2212151\n\u2212118\n\u221292\n\u221260\n\u221228\n4\n37\n69\n\u20024\n\u2212214\n\u2212179\n\u2212145\n\u2212111\n\u221285\n\u221252\n\u221219\n14\n48\n81\n\u20025\n\u2212192\n\u2212161\n\u2212130\n\u221299\n\u221275\n\u221243\n\u22129\n26\n60\n94\n\u20026\n\u2212148\n\u2212116\n\u221284\n\u221252\n\u221228\n4\n35\n66\n97\n128\n\u20027\n\u2212119\n\u221284\n\u221248\n\u221212\n19\n52\n84\n116\n148\n180\n\u20028\n\u2212101\n\u221265\n\u221230\n6\n41\n69\n105\n140\n176\n211\n\u20029\n\u221273\n\u221238\n\u22123\n32\n67\n97\n132\n168\n203\n239\n\u200210\n\u221241\n\u22125\n31\n67\n102\n132\n168\n203\n238\n273\nEach cell corresponds to one combination of demand (increasing by row) and fuel cost (increasing by column). The values show the US dollar payout per kilowatt of a futures contract in that demand and \nfuel cost scenario in the Both Contracts case. Bold indicates the 50 worst-case scenarios for the baseload technology in the No Trading case. Alignment between the payout structure and risk profile leads \ninvestors in baseload generation to prefer futures to options when both are available.\nTable 3 | Alignment of option payout to risk faced by peaker \ntechnology\nDemand\nFuel cost\n1\n2\n3\n4\n5\n6\n7\n8\n9\n10\n\u20021\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n\u20022\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n\u20023\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n\u20024\n0\n0\n0\n0\n0\n0\n0\n0\n0\n0\n\u20025\n14\n10\n6\n2\n0\n0\n0\n0\n0\n0\n\u20026\n51\n47\n43\n39\n36\n35\n31\n27\n23\n19\n\u20027\n76\n75\n74\n73\n72\n71\n67\n63\n59\n55\n\u20028\n89\n87\n85\n83\n81\n81\n80\n79\n78\n77\n\u20029\n111\n108\n105\n102\n99\n99\n97\n95\n93\n91\n\u200210\n138\n135\n132\n129\n126\n126\n123\n120\n117\n114\nEach cell corresponds to one combination of demand (increasing by row) and fuel cost (increasing \nby column). The values show the US dollar payout per kilowatt of an options contract in that \ndemand and fuel cost scenario in the Both Contracts case. Bold indicates the 50 worst-case \nscenarios for the peaker technology in the No Trading case. Alignment between the payout \nstructure and risk profile leads investors in peaking generation to prefer options to futures when \nboth are available.\n0\n20\n40\n60\n80\n100\n0\n40\n80\n120\n160\n200\nBaseload\nTotal\nMaximum option trade volume (% of installed capacity)\nEquilbrium capacity (GW)\nFig. 1 | Decline of baseload capacity with increased options trading. Total \nICAP grows slightly with the availability of options contracts. As these \ncontracts are better suited to the risk profile of the peaker technology, \nbaseload capacity falls to zero.\nNature Energy | VOL 4 | NOVEMBER 2019 | 948\u2013956 | www.nature.com/natureenergy\n951\n\nArticles\nNaTure Energy\nThe trades that emerge in the All Contracts case echo the results \nof the two-generator example. Each technology prefers contracts of \nthe trade adapted to its risk profile: investors in the baseload tech-\nnology sell 38.5\u2009GW of futures, investors in the peaker sell 96.7\u2009GW \nof options and investors in the variable technology sell 91.8\u2009GW of \nunit contingent contracts. Unlike the two-generator example, in \nwhich the Both Contracts case reached an equilibrium close to the \nsocial optimum, there remains a US$0.6\u2009billion\u2009yr\u20131 gap between the \nAll Contracts case and the complete trading solution. Additional \ncontracts are required to bridge this divide.\nConclusion\nThis article argues that the introduction of a capacity market will \ntilt the technology mix that arises in equilibrium towards resources \nwith higher operating costs. The syllogism proceeds as follows. The \nfinancial impact of a capacity mechanism is to replace volatile scar-\ncity prices with more regular revenues. The higher a unit\u2019s operating \ncosts, the more its cost recovery in an energy-only market relies on \nscarcity prices. Therefore, the introduction of a capacity market has \na stronger impact on the risk profile of technologies with higher \noperating costs.\nThe magnitude of this effect in real-world markets is difficult \nto evaluate due to its dependence on the risk tolerance of market \nparticipants, their evaluation of potential future scenarios and the \navailability of other outlets for risk trading. However, this article\u2019s \nnumerical experiments reveal the potential for substantial shifts, \nas both examples exhibit a significant increase in capacity of the \npeaking technology after the introduction of an options contract. \nThe struggles of US nuclear units, which have been unable to find \ncounterparties for long-term contracts without state support despite \ntheir potential role as a hedge against future increases in natural \ngas and carbon prices, suggests that the mechanism demonstrated \nin this article may be at work in current markets. This result is \nparticularly important in the context of efforts to reduce carbon  \nTable 4 | Equilibria in the three-generator example for cases \nwith 0 and 1 contract\nNo trading\nUnit only\nOption \nonly\nFuture \nonly\nCapacity (GW)\n\u2002Baseload\n38.7\n1.9\n0.0\n70.9\n\u2002Peaker\n97.6\n121.3\n131.0\n79.9\n\u2002Variable\n118.1\n195.1\n167.7\n48.9\nTrade volume (GW)\n\u2002Baseload\n\u2002Future\n63.8\n\u2002Option\n0.0\n\u2002Unit\n1.9\n\u2002Peaker\n\u2002Future\n52.3\n\u2002Option\n132.0\n\u2002Unit\n35.8\n\u2002Variable\n\u2002Future\n17.4\n\u2002Option\n8.1\n\u2002Unit\n195.1\nProximity to \nequilibrium (%)\n0.012\n0.003\n0.001\n0.004\nChange in surplus \n(US$\u2009billion\u2009yr\u20131)\n\u221217.6\n\u221213.1\n\u22126.0\n\u22123.6\nFour financial trading cases are shown in order of increasing surplus. Bold cells facilitate pairwise \ncomparison within each row. Compared to the No Trading case, introducing a single contract leads \nto more capacity of the corresponding technology.\nTable 5 | Equilibria in the three-generator example for cases with 2 and 3 contracts\nFuture\u2009+\u2009Unit\nOption\u2009+\u2009Unit\nOption\u2009+\u2009Future\nAll Contracts\nCapacity (GW)\n\u2002Baseload\n70.9\n0.0\n68.2\n42.5\n\u2002Peaker\n79.9\n126.9\n81.3\n96.6\n\u2002Variable\n48.9\n195.7\n56.8\n120.4\nTrade volume (GW)\n\u2002Baseload\n\u2002Future\n63.8\n61.4\n38.5\n\u2002Option\n0.0\n0.0\n1.5\n\u2002Unit\n0.0\n0.0\n0.0\n\u2002Peaker\n\u2002Future\n52.3\n10.2\n0.9\n\u2002Option\n138.1\n67.9\n96.7\n\u2002Unit\n0.0\n14.3\n0.0\n\u2002Variable\n\u2002Future\n17.4\n21.5\n12.8\n\u2002Option\n2.3\n0.0\n0.0\n\u2002Unit\n0.0\n199.1\n91.8\nProximity to equilibrium (%)\n0.006%\n0.003%\n0.144%\n0.004%\nChange in surplus (US$\u2009billion\u2009yr\u20131)\n\u22123.6\n\u22121.0\n\u22120.9\n\u22120.6\nFour financial trading cases are shown in order of increasing surplus. Bold cells facilitate pairwise comparison within each row. Compared to the All Contracts case, removing a single contract leads to a \nlower capacity of the corresponding technology.\nNature Energy | VOL 4 | NOVEMBER 2019 | 948\u2013956 | www.nature.com/natureenergy\n952\n\nArticles\nNaTure Energy\nemissions. The majority of energy in low-carbon systems is likely to \nbe provided by some combination of hydroelectric, nuclear, wind \nand solar resources, all of which are characterized by high capital \ncosts and low operating costs. Accordingly, capacity markets as cur-\nrently structured may work against efforts to decarbonize.\nTo achieve an efficient capacity mix requires the ability to share \nrisk. Integrated resource planning, which can be made versa-\ntile enough to incorporate whatever sources of risk are viewed to \nbe most salient, is one way to do so. A second, modelled in this \narticle, is an energy-only market accompanied by the robust trade \nof instruments adapted to the risk profile of customers and vari-\nous technologies. Neither of these strategies is without challenges. \nCapacity markets can offer a partial solution to the problem of risk \nsharing. The results of this study, however, reveal the consequences \nthat choices made in the design of these markets may have for the \nresource mix that arises in liberalized electricity systems.\nMethods\nModelling framework. Here we formulated a multiple-agent variant of a two-stage \nstochastic program for capacity expansion. In the first stage, risk-averse investment \nand financial trading decisions are made by a number of agents. Spot market prices \nare determined through a perfectly competitive dispatch over time blocks, t 2 T\nI\n,  \nperformed by the system operator in the second stage. The ultimate source of all \ngenerator revenue in the model is energy sales; financial contracts settle based on \nthe prices realized in the second stage. Instead of writing the optimality conditions \nfor each of the subproblems and reformulating as a complementarity problem9,41,42, \nwe anticipate the algorithm that we describe later in this section and focus on the \nproblems solved by each agent.\nOur modelling framework builds most directly on the approaches of de Maere \nd\u2019Aertrycke and co-workers41,]42, which describe a risk-averse capacity equilibrium \nwith incomplete markets. We make two modifications of note, which we \ndescribe in detail after introducing the model. Incomplete trading has significant \nramifications, both computationally and conceptually. The computational issue \nis that, instead of a comparatively simple optimization problem, the problem \nis formulated using a construct called a multiple optimization problem with \nequilibrium constraints43. Although modest-sized instances can be solved by the \nPATH solver44, numerical tests on larger instances can fail to converge. To help \naddress this issue, we proposed a heuristic decomposition algorithm similar \nin spirit to that of H\u00f6schle et\u00a0al.45, which allows the identification of equilibria \nthrough a series of convex quadratic programs. Although the test systems used \nin this article are kept simple, this decomposition approach allows an easier \nextension to a richer set of technologies and risks. The conceptual concern is that \nthe uniqueness of the equilibrium solutions cannot, in general, be guaranteed. We \nrevisit this topic after introducing the algorithm. The optimization models are \nimplemented in AMPL46 and solved using CPLEX47.\nIn the three-generator example, uncertainty comes from three elements.  \nThe random variable D+, with scenarios indexed by s 2 S\nI\n, reflects an upward \ndemand shift in all the time blocks. The random variables Agt, with scenarios \nindexed by r 2 R\nI\n, denotes the availability of each generator in a given time  \nblock. The random variables CEN\ng\nI\n and D\u2212, with scenarios indexed by f 2 F\nI\n,  \ncomprise generator fuel costs as well as the downward demand shift perfectly \ncorrelated with higher prices. These three sources of randomness are assumed  \nto be independent and to have finite support. In the two-generator example, \nrandomness is modelled in the same way except that it does not include  \nthe random variables Agt. Each market participant, a 2 A\nI\n, uses a coherent  \nrisk measure \u03c1a in its decision problem. As shown in Artzner et\u00a0al.36, this \ncorresponds to choosing the worst-case distribution in a risk set, which takes  \nthe form of a convex subset of probability measures on F \u00b4 R \u00b4 S\nI\n. The risk  \nmeasures we specifically chose are convex combinations of expected value \nand CVaR of the surplus (Philpott et\u00a0al.18 and de Maere d\u2019Aertrycke et\u00a0al.39 give \nadditional examples of this construction). The CVaR portion of the calculation \nidentifies the worst (100\u03b1) percentage of second-stage outcomes for the market \nparticipant in question, and calculates the average surplus only among those \nscenarios. Thus, starting from a shared nominal distribution, each participant \nperforms a risk-averse optimization that implicitly assigns a higher probability  \nto the combinations of fuel, availability and demand scenarios most harmful  \nto their interest. The difference in expected profit when evaluated using this \nadjusted distribution versus the nominal distribution corresponds to the risk \npremium required by investors. This endogenous representation of the risk \npremium stands in contrast to the classical capacity expansion framework, in \nwhich investment cost is annualized according to an exogenous cost of capital.\nDispatch. The economic dispatch (ED) problem was formulated as a convex \nquadratic program, the outcome of which depends on the realization of the \nrandom variables that govern demand, availability and fuel costs. The problem \nsplits operations into simple time blocks of varying length, which reflect the \ndifferent levels of load seen throughout the year without explicitly representing \noperational considerations, such as unit commitment decisions, which in general \npreclude equilibria, or ramping constraints, which can exacerbate the issue \nof multiple equilibria. In general, the inclusion of these constraints leads to a \ngreater need for capacity, but their implications for the resource mix are not well \nunderstood48.\nNotation. \nSets:\ng 2 G\nI\n, set of all generation technologies\nt 2 T\nI\n, set of time blocks\nf 2 F\nI\n, set of scenarios for fuel prices\nr 2 R\nI\n, set of scenarios for generator availability profiles\nDfix\nt\nI\n, set of scenarios for demand shifters\nParameters:\nB, value of non-price-responsive load (US$\u2009MWh\u20131)\nLt, length of time block t (h)\nDfix\nt\nI\n, baseline level of non-price-responsive load in time block t (MW)\nDres\nt\nI\n, amount of price-responsive demand, which bids at a value declining \nlinearly from B to 0 (MW)\nScenario-specific parameters (that is, realizations of random variables):\nAgrt, availability of generator g in time block t under availability profile scenario \nr (%)\nD\u001f\nf\nI\n, marginal cost for generator g under fuel price scenario f (US$\u2009MWh\u20131)\nD\u001f\nf\nI\n, downward shift in demand in all time blocks under fuel price scenario f \n(MW)\ndfix\nt\nI\n, upward shift in demand in all time blocks under demand shifter scenario \ns (MW)\nVariables:\nygt, production by technology g in time block t (MW)\ndfix\nt\nI\n, amount of non-price-responsive demand cleared in time block t (MW)\ndfix\nt\nI\n, amount of responsive demand cleared in time block t (MW)\nCapacity expansion decision (fixed in dispatch model):\nxg, quantity installed of generation technology g (MW)\nFormulation. The ED problem, with Hfrs denoting the total surplus given scenarios f \nfor fuel price, r for generator availability and s for demand, is stated as:\n\u00f0ED\u00defrs\nHfrs \u00bc\nmaximize\ny;d\nX\nt2T\nLtB dfix\nt\n\u00fe dres\nt\n\u001a \u00f0dres\nt \u00de2=\u00f02Dres\nt \u00de\n\u001e\n\ue001\n\u001f\nX\nt2T\nX\ng2G\nLtCEN\nfg ygt\n\u00f01\u00de\nsubject to :\ndfix\nt\n\u00fe dres\nt\n\u00fe D\u00fe\ns \u001e D\u001e\nf \u00bc\nX\ng2G\nygt\n8t 2 T\n\u00f02\u00de\n0\u2264ygt \u2264Agrtxg\n8g 2 G; t 2 T\n\u00f03\u00de\n0\u2264dfix\nt \u2264Dfix\nt\n8t 2 T\n\u00f04\u00de\n0\u2264dres\nt \u2264Dres\nt\n8t 2 T\n\u00f05\u00de\nEquation (1) includes a linear term for the value of non-price-responsive load, a \nquadratic expression for the value of price-responsive load and a linear term for \nthe cost of producing power. The responsive load is the first departure from the \nmodel in de Maere d\u2019Aertrycke et\u00a0al.41. Its presence does not carry any philosophical \nsignificance: it is introduced for computational convenience, to help avoid the \ninevitability of multiple optimal dual solutions in the linear version of the model \ndiscussed in Abada et\u00a0al.42. The power balance constraint in equation (2) splits \ndemand into four segments: dfix\nt\nI\n, dres\nt\nI\n, D\u001f\nf\nI\n and D\u001f\nf\nI\n. The second, more meaningful \ndeparture from de Maere d\u2019Aertrycke et\u00a0al.41 is the absence of the terms D\u001f\nf\nI\nand D\u001f\nf\nI\n \nfrom the objective function. This choice implies that consumers do not experience  \nany gain or loss of surplus as a result of these exogenous shocks to demand, which has \nimportant consequences when comparing our numerical results to those of de Maere \nd\u2019Aertrycke et\u00a0al.39, which are based on the model in de Maere d\u2019Aertrycke et\u00a0al.41.  \nConceptually, the question is which events in the sample space F \u00b4 R \u00b4 S\nI\n should \ncustomers be most concerned about. Upward shocks to demand could come \nfrom, for example, a higher-than-expected economic growth that leads to more \nproductive uses of electricity. In this case, consumers might not be concerned about \nhigh-demand scenarios, as shortages would be balanced by an increase in surplus \nfrom consumption. An upward shock could also come, however, from an increased \nuse of air conditioning due to higher summer temperatures, which consumers \nmight experience as neutral or even a loss. By excluding these shocks, we ensure \nthat consumers concentrate on the high-demand scenarios in which shortages \noccur more frequently. This definition of consumer surplus appears subsequently \nin the consumer model (CON). The optimal values of the ED model across all the \nscenarios define the social surplus (SOC) used in the complete trading model.\nNature Energy | VOL 4 | NOVEMBER 2019 | 948\u2013956 | www.nature.com/natureenergy\n953\n\nArticles\nNaTure Energy\nContracts. In conjunction with the capacity investment decisions made in the \nfirst stage, investors can sign contracts with consumers that settle based on the \nenergy price in the second stage. Assume a small set of contracts \u03b7k\nfrs\nI\n is available to \ntrade. Let \u03bbfrst be the price of energy in time block t given fuel price f, profile r and \ndemand shifter s, that is, the dual variable that corresponds to the power balance \nconstraint in equation (2) divided by the length Lt, and let \u03b7k\nfrs\nI\n represent the payout  \nof contract k 2 K\nI\n given this scenario. A call option gives the purchaser the right to \nbuy a unit of energy for a predetermined price if the spot price exceeds that level. \nThus, if contract k indexes a call option that covers all the time blocks with strike \nprice \u03bbk, we can calculate:\n\u03b7k\nfrs \u00bc\nX\nt2T\nLt maxf0; \u03bbfrst \u001c \u03bbkg\n\u00f06\u00de\nIf k instead indexes a futures contract that covers all time blocks at price \u03bbk, the \npayout is calculated as:\n\u03b7k\nfrs \u00bc\nX\nt2T\nLt\u00f0\u03bbfrst \u001c \u03bbk\u00de\n\u00f07\u00de\nWe also define a unit contingent contract that tracks the availability of generator g, \na construct often used for variable technologies. When k indexes a unit contingent \ncontract for generator g, the payout is calculated as:\n\u03b7k\nfrs \u00bc\nX\nt2T\nAgrtLt\u00f0\u03bbfrst \u001c \u03bbk\u00de\n\u00f08\u00de\nMarket participants. Each market participant solves a convex quadratic program, \nchoosing a quantity of financial contracts that maximizes its risk-adjusted surplus. \nGenerator capacities, outputs of the dispatch model for each scenario and the \nprices and payouts of contracts are fixed parameters in these models. We define \nnotation here that can be shared between the consumer and generator models, that \nis, among all market participants.\nNotation. \nSets:\na 2 A\nI\n, set of market participants (generators and consumers)\nvk\na;\u001fvk\na\nI\n, set of contracts\nParameters:\n\u03b1a, tail probability at which CVaR is evaluated by market participant a, \n0\u2009<\u2009\u03b1a\u2009\u2264\u20091\n\u03b2a, weight given to the expected value in the risk measure for market \nparticipant a, 0\u2009\u2264\u2009\u03b2a\u2009\u2264\u20091\n\u03b3, weight on the term that penalizes imbalance in the contract volumes between \nmarket participants\nvk\na;\u001fvk\na\nI\n, minimum and maximum volume of contract k to be purchased or sold by \nmarket participant a (MW)\npfrs, nominal probability of scenario (f,r,s)\n\u03b7k\nfrs\nI\n, investment cost for generator g annualized at risk-free rate (US$\u2009MW\u20131)\nProvisional parameters (that is, values calculated by other agents):\n\u03bbfrst, price of energy in time block t under scenario (f,r,s) (US$\u2009MWh\u20131)\n\u03c0fgrst, operating profit for technology g in block t under scenario (f,r,s) \n(US$\u2009MW\u20131)\n\u03b7k\nfrs\nI\n, payout of contract k under scenario (f,r,s) (US$\u2009MW\u20131)\n\u03d5k, price of contract k incurred in the first stage (US$\u2009MW\u20131)\nVariables:\nvk\na\nI\n, volume of contract k purchased or sold by market participant a (MW)\nVaRa, value at risk for market participant a (US$)\nua\nfrs\nI\n, surplus for market participant a under scenario (f,r,s) (US$)\nc 2 A\nI\n, auxiliary variable used in the calculation of value at risk (VaR) (US$)\nConsumer model. We distinguished consumer and generator agents by using c 2 A\nI\n \nand g 2 A\nI\n in place of the generic agent a, and we assume a single consumer; that \nis, in terms of our notation, g 2 G\nI\n. Given the assumption of perfect competition, \nthe resulting decisions are equivalent to those of a larger number of small, identical \nload serving entities. Defining \u03c1a to be the risk measure associated with market \nparticipant a, the consumer\u2019s problem is stated as follows:\n\u00f0CON\u00de\n\u03c1c \u00bc\nmaximize\nvc;uc;uc\u00fe;VaRc\n\u00f01 \u001d \u03b2c\u00de\nVaRc \u001d 1=\u03b1c\nX\nf 2F\nX\nr2R\nX\ns2S\npfrsuc\u00fe\nfrs\n0\n@\n1\nA\n\u00fe\u03b2c\nX\nf 2F\nX\nr2R\nX\ns2S\npfrsuc\nfrs\n0\n@\n1\nA \u001d \u03b3=2\nX\nk2K\nX\na2A\nvk\na\n \n!2\n\u00f09\u00de\nsubject to\nuc\nfrs \u00bc \u001e\nX\nk2K\nvk\nc \u03d5k \u001e \u03b7k\nfrs\n\ue010\n\ue011\n\u00fe\nX\nt2T\nLtB dfix\nt\n\u00fe dres\nt\n\u001d \u00f0dres\nt \u00de2=\u00f02Dres\nt \u00de\n\u001e\n\ue001\n\u001f\nX\nt2T\nLt\u03bbfrst dfix\nt\n\u00fe dres\nt\n\u00fe D\u00fe\ns \u001f D\u001f\nf\n\ue010\n\ue011\n8f 2 F; r 2 R; s 2 S\n\u00f010\u00de\nVaRc \u001f uc\nfrs \u2264uc\u00fe\nfrs\n8f 2 F; r 2 R; s 2 S\n\u00f011\u00de\n0\u2264uc\u00fe\nfrs\n8f 2 F; r 2 R; s 2 S\n\u00f012\u00de\nvk\nc \u2264vk\nc \u2264\u001fvk\nc\n8k 2 K\n\u00f013\u00de\nThe consumer maximizes a convex combination of CVaR and expected value \nof surplus, subtracting a proximal term that penalizes imbalances between the \ncontracts bought and sold by the market participants. In equilibrium, this third \nterm of equation (9) must equal zero. Although similar in appearance, the inclusion \nof this proximal term does not technically yield an augmented Lagrangian, because \nthe market clearing condition for financial contracts is not part of the consumer\u2019s \ntrue objective function. Constraint (10) calculates the surplus for the consumer \nin every scenario that results from the purchase of contracts in the first stage \nand energy in the second. Constraints (11) and (12) dictate the value of auxiliary \nvariables used in the CVaR calculation. Constraint (13) dictates a minimum and \nmaximum volume of each contract. Although we discuss specific choices of these \nparameters in Supplementary Note 3, an algorithmic advantage of including these \nconstraints on trading is to guarantee that all the subproblems are bounded. Note \nthat vc\nk is a decision variable in model (CON), but the contract volumes for the \ngenerators, which appear in the objective function\u2019s final term, enter model (CON) \nas provisional parameters.\nGenerator model. We modelled a single investor in each generation technology \ng 2 G\nI\n. As in the case of the consumer, investment decisions can equivalently be \nrepresented as a large number of identical firms. The investor\u2019s problem is stated as:\n\u00f0GEN\u00deg\n\u03c1g \u00bc\nmaximize\nvg;ug;ug\u00fe;VaRg\n\u00f01 \u001d \u03b2g\u00de\nVaRg \u001d 1=\u03b1g\nX\nf 2F\nX\nr2R\nX\ns2S\npfrsug\u00fe\nfrs\n0\n@\n1\nA\n\u00fe\u03b2g\nX\nf 2F\nX\nr2R\nX\ns2S\npfrsug\nfrs\n0\n@\n1\nA \u001d \u03b3=2\nX\nk2K\nX\na2A\nvk\na\n \n!2\n\u00f014\u00de\nsubject to\nug\nfrs \u00bc \u001eCINV\ng\nxg \u001e\nX\nk2K\nvk\ng \u03d5k \u001e \u03b7k\nfrs\n\ue010\n\ue011\n\u00fe\nX\nt2T\n\u03c0fgrstxg\n8f 2 F; r 2 R; s 2 S\n\u00f015\u00de\nVaRg \u001f ug\nfrs \u2264ug\u00fe\nfrs\n8f 2 F; r 2 R; s 2 S\n\u00f016\u00de\n0\u2264ug\u00fe\nfrs\n8f 2 F; r 2 R; s 2 S\n\u00f017\u00de\nvk\ng \u2264vk\ng \u2264\u001fvk\ng\n8k 2 K\n\u00f018\u00de\nThe consumer and generator models differ only in the calculation of scenario \nsurplus in equations (10) and (15). For simplicity, we employed the operating profit \n\u03c0fgrst, calculated as the dual variable to the maximum generation constraint in the \ndispatch model, equation (3), multiplied by availability Agrt.\nDecomposition approach. With the assumption of perfect competition, investors \nin technology g will build new capacity until the risk measure \u03c1g\u2009=\u20090 in equilibrium. \nThus, the challenge is to identify values for the ICAP x and contract prices \u03d5 that \nbalance the market in all financial contracts and result in zero risk-adjusted profit \nfor all generation technologies, given the prices and operating profits that arise \nfrom the dispatch model. We refer to this equilibrium problem as (EQ).\nNature Energy | VOL 4 | NOVEMBER 2019 | 948\u2013956 | www.nature.com/natureenergy\n954\n\nArticles\nNaTure Energy\nAlgorithm 1 shows the solution approach that we employ. The algorithm \nchooses a capacity mix, finds the implied dispatch solutions, identifies the prices \nthat balance the markets for financial contracts (to tolerance \u03b5) and then updates \nthe capacity mix based on profitability. With capacity decisions made in the outer \nloop, one interpretation of the algorithm is that it simulates the process of entry \nand exit in competitive markets. The use of a proximal term in the objective \nfunction for the market participants, as well as the sequential updating, invites \ncomparison to the alternating direction method of multipliers49,50, an avenue \npursued on a model similar to ours in H\u00f6schle et\u00a0al.45. As in H\u00f6schle et\u00a0al.45, \nconvergence is not guaranteed, but was achieved in all the numerical tests after \nlimited experimentation with different step sizes \u03c3 and \u03bc. We compare our \napproach to that of H\u00f6schle et\u00a0al.45 in Supplementary Note 5.\nAlgorithm 1. \nInput: An instance of (EQ) defined by models (ED), (CON), and (GEN).\nOutput: Near-equilibrium solution to (EQ)\n\u2003 Define \u03c3, \u03bc, \u03b4, \u03b5\u2009>\u20090; let \u03c1a \u00bc 0 : 8a 2 A\nI\n; initialize x, \u03d5\n\u2003 loop\n\u2003 \u2003 \u03bbfrst; \u03c0fgst; \u03b7k\nfrs : 8\u00f0f ; r; s\u00de 2 F \u00b4 R \u00b4 S; k 2 K\nI\n\u2003 \u2003 solve (ED)frs; update \u03bbfrst; \u03c0fgst; \u03b7k\nfrs : 8\u00f0f ; r; s\u00de 2 F \u00b4 R \u00b4 S; k 2 K\nI\n\u2003 \u2003 solve (CON)\n\u2003 \u2003 solve maxk2K j P\na2A\nvk\naj>\u03b5\nI\n\u2003 \u2003 while \u03d5k  \u03d5k \u00fe \u03c3 P\na2A\nvk\na\n8k 2 K\nI\n do\n\u2003 \u2003 \u2003 \u03d5k  \u03d5k \u00fe \u03c3 P\na2A\nvk\na\n8k 2 K\nI\n\u2003 \u2003 \u2003 solve (CON)\n\u2003 \u2003 \u2003 solve \u00f0GEN\u00deg\n8g 2 G\nI\n\u2003 \u2003 end while\n\u2003 \u2003 if maxg2G j\u03c1gj<\u03b4\nI\n then\n\u2003 \u2003 \u2003 return x and \u03d5\n\u2003 \u2003 end if\n\u2003 end loop\nThe potential for multiple equilibria poses a challenge for the interpretation of \nthe numerical results51. In an effort to avoid this possibility, we omit intertemporal \nconstraints from model (ED) and maintain a constant merit order through all the \nscenarios. Under these assumptions and with no financial trading, uniqueness can \nbe shown for the model in Abada et\u00a0al.42. In both numerical examples, capacity is \ninitialized at the same, socially optimal starting point for all the cases within each \nexample, with the goal to avoid a spurious result. Contract prices are initialized at \ntheir expected payouts using the nominal probability distribution. Starting from \nalternative points in ad hoc tests did not uncover alternative equilibria.\nComplete trading. For comparison, we also constructed a model that assumes \ncomplete trading. In this setting, an equilibrium can be identified through a \nrisk-averse optimization problem without the need for equilibrium constraints, \nusing the intersection of the risk sets of all market participants17\u201320. This can be \nrepresented as a single large-scale optimization problem that comprises investment \ndecisions and the dispatch model for each scenario. To be concise, we employed \nas a variable Hfrs to indicate the surplus that arises from the dispatch using the \nchosen amount of capacity x. Replacing this variable with the objective function \nfrom (ED)frs and inserting the constraints in (ED)frs for each scenario recovers the \ncomplete model. With subscript i denoting societal risk preferences, the complete \ntrading model is written as:\n\u00f0SOC\u00de\n\u03c1i \u00bc\nmaximize\nx;y;d;H;ui;ui\u00fe;VaRi\n\u00f01 \u001d \u03b2i\u00de VaRi \u001d 1=\u03b1i\nX\nf 2F\nX\nr2R\nX\ns2S\npfrsui\u00fe\nfrs\n2\n4\n3\n5\n\u00fe\u03b2i\nX\nf 2F\nX\nr2R\nX\ns2S\npfrsui\nfrs\n2\n4\n3\n5\n\u00f019\u00de\nsubject to\nui\nfrs \u00bc \u001e\nX\ng2G\nCINV\ng\nxg \u00fe Hfrs\n8f 2 F; r 2 R; s 2 S\n\u00f020\u00de\nVaRi \u001f ui\nfrs \u2264ui\u00fe\nfrs\n8f 2 F; r 2 R; s 2 S\n\u00f021\u00de\n0\u2264ui\u00fe\nfrs\n8f 2 F; r 2 R; s 2 S\n\u00f022\u00de\nConvergence. As is characteristic of the alternating direction method of \nmultipliers, the algorithm exhibits slow convergence near the equilibrium. For \ncases in which the termination criterion \u03b4 in Algorithm 1 is not met, we report the \nsolution found at iteration 50,000 of the outer loop. As a measure of proximity to \nequilibrium, we compute the value:\nmaxfg2G:xg > 0g\n\u03c1g\nCINV\ng\n\u001f\u001f\u001f\u001f\u001f\n\u001f\u001f\u001f\u001f\u001f\n\u00f023\u00de\nfor each case, where \u03c1g is the risk measure calculated by problem (GEN)g.\nData availability\nThe code and data used for numerical tests in this study are available in a public \nrepository (https://doi.org/10.5281/zenodo.3242844).\nReceived: 24 January 2019; Accepted: 6 September 2019;  \nPublished online: 28 October 2019\nReferences\n\t1.\t Bushnell, J., Flagg, M. & Mansur, E. Capacity Markets at a Crossroads \nWorking Paper 278 (Energy Institute at Haas, 2017).\n\t2.\t Oren, S. S. Generation adequacy via call options obligations: safe passage to \nthe promised land. Electricity J. 18, 28\u201342 (2005).\n\t3.\t Joskow, P. in The New Energy Paradigm (ed. Helm, D.) Ch. 4 (Oxford Univ. \nPress, 2007).\n\t4.\t Cramton, P. & Stoft, S. The Convergence of Market Designs for Adequate \nGenerating Capacity with Special Attention to the CAISO's Resource Adequacy \nProblem (MIT Center for Energy and Environmental Policy Research, 2006); \nhttps://dspace.mit.edu/handle/1721.1/45053\n\t5.\t Hogan, W. W. Electricity scarcity pricing through operating reserves. Econ. \nEnergy Environ. Policy 2, 65\u201386 (2013).\n\t6.\t Joskow, P. & Tirole, J. Reliability and competitive electricity markets. RAND J. \nEcon. 38, 60\u201384 (2007).\n\t7.\t Cramton, P., Ockenfels, A. & Stoft, S. Capacity market fundamentals. Econ. \nEnergy Environ. Policy 2, 27\u201346 (2013).\n\t8.\t Bothwell, C. & Hobbs, B. F. Crediting wind and solar renewables in electricity \ncapacity markets: the effects of alternative definitions upon market efficiency. \nEnergy J. 38(SI), 173\u2013188 (2017).\n\t9.\t Ehrenmann, A. & Smeers, Y. Generation capacity expansion in a risky \nenvironment: a stochastic equilibrium analysis. Oper. Res. 59,  \n1332\u20131346 (2011).\n\t10.\tVazquez, C., Rivier, M. & Perez-Arriaga, I. J. A market approach to long-term \nsecurity of supply. IEEE Trans. Power Syst. 17, 349\u2013357 (2002).\n\t11.\tNeuhoff, K. & De Vries, L. Insufficient incentives for investment in electricity \ngenerations. Util. Policy 12, 253\u2013267 (2004).\n\t12.\tSchiro, D. A., Hobbs, B. F. & Pang, J. S. Perfectly competitive capacity \nexpansion games with risk-averse participants. Comput. Optim. Appl. 65, \n511\u2013539 (2016).\n\t13.\tRoques, F. A., Nuttall, W. J., Newberry, D. M., de Neufville, R. & Connors, S. \nNuclear power: a hedge against uncertain gas and carbon prices? Energy J. 27, \n1\u201324 (2006).\n\t14.\tRoques, F. A., Newbery, D. M. & Nuttall, W. J. Fuel mix diversification \nincentives in liberalized electricity markets: a mean\u2013variance portfolio theory \napproach. Energy Econ. 30, 1831\u20131849 (2008).\n\t15.\tGross, R., Blyth, W. & Heptonstall, P. Risks, revenues and investment in \nelectricity generation: why policy needs to look beyond costs. Energy Econ. \n32, 796\u2013804 (2010).\n\t16.\tGuo, X., Beskos, A. & Siddiqui, A. The natural hedge of a gas-fired power \nplant. Comput. Manag. Sci. 13, 63\u201386 (2016).\n\t17.\tEhrenmann, A. & Smeers, Y. in Stochastic Optimization Methods in  \nFinance and Energy: New Financial Products and Energy Market Strategies \n(eds. Bertocchi, M., Consigli, G. & Dempster, M. A. H.) 273\u2013310  \n(Springer, 2011).\n\t18.\tPhilpott, A., Ferris, M. & Wets, R. Equilibrium, uncertainty  \nand risk in hydrothermal electricity systems. Math. Program. 157,  \n483\u2013513 (2016).\n\t19.\tFerris, M. & Philpott, A. Dynamic Risked Equilibrium (2018); www.\noptimization-online.org/DB_HTML/2018/04/6577.html\n\t20.\tRalph, D. & Smeers, Y. Risk trading and endogenous probabilities in \ninvestment equilibria. SIAM J. Optim. 25, 2589\u20132611 (2015).\n\t21.\tNewbery, D. Missing money and missing markets: reliability, capacity \nauctions and interconnectors. Energy Policy 94, 401\u2013410 (2016).\n\t22.\tO\u2019Shaughnessy, E., Heeter, J. & Sauer, J. Status and Trends in the US Voluntary \nGreen Power Market (2017 Data) (National Renewable Energy Laboratory, \n2018); www.nrel.gov/docs/fy19osti/72204.pdf\n\t23.\tJoskow, P. Capacity payments in imperfect electricity markets: need and \ndesign. Util. Policy 16, 159\u2013170 (2008).\n\t24.\tCheng, I. H. & Xiong, W. Financialization of commodity markets. Annu. Rev. \nFinanc. Econ. 6, 419\u2013441 (2014).\n\t25.\tHirshleifer, D. Hedging pressure and futures price movements in a general \nequilibrium model. Econometrica 58, 411\u2013428 (1990).\nNature Energy | VOL 4 | NOVEMBER 2019 | 948\u2013956 | www.nature.com/natureenergy\n955\n\nArticles\nNaTure Energy\n\t26.\tResource Investment in Competitive Markets (PJM Interconnection, 2016); \nwww.pjm.com//media/library/reports-notices/special-reports/20160505-\nresource-investment-in-competitive-markets-paper.ashx\n\t27.\tPierpont, B. & Nelson, D. Markets for Low Carbon, Low Cost Electricity \nSystems (Climate Policy Initiative, 2017); https://climatepolicyinitiative.org/\npublication/markets-low-carbon-low-cost-electricity-systems-working-paper\n\t28.\tGimon, E. On Market Designs for a Future with a High Penetration of Variable \nRenewable Generation (Energy Innovation LLC, 2017); https://\namericaspowerplan.com/wp-content/uploads/2017/10/On-Market-Designs-\nfor-a-Future-with-a-High-Penetration-of-Renew.pdf\n\t29.\tMacdonald, B. & Du, M. Project Finance Industry Report Card (S&P Global \nRatings, 2017); https://www.spratings.com/documents/20184/1634005/SPGR_\nProject+Finance+Industry+Report+Card-052217/1a91054b-709b-49cb-9473-\n3f0f3affbc2a\n\t30.\tButler, L. & Neuhoff, K. Comparison of feed-in tariff, quota and auction \nmechanisms to support wind power development. Renew. Energy 33, \n1854\u20131867 (2008).\n\t31.\tPolzin, F., Egli, F., Steffen, B. & Schmidt, T. S. How do policies mobilize \nprivate finance for renewable energy? A systematic review with an investor \nperspective. Appl. Energy 236, 1249\u20131268 (2019).\n\t32.\tLevin, T. & Botterud, A. Electricity market design for generator revenue \nsufficiency with increased variable generation. Energy Policy 87, 392\u2013406 (2015).\n\t33.\tFrew, B. A. et\u00a0al. Revenue Sufficiency and Reliability in a Zero Marginal Cost \nFuture (National Renewable Energy Laboratory, 2016); www.nrel.gov/docs/\nfy17osti/66935.pdf\n\t34.\tMilligan, M. et\u00a0al. Wholesale electricity market design with increasing levels \nof renewable generation: revenue sufficiency and long-term reliability. \nElectricity J. 29, 26\u201338 (2016).\n\t35.\tJoskow, P. L. Challenges for wholesale electricity markets with intermittent \nrenewable generation at scale: the US experience. Oxf. Rev. Economic Policy \n35, 291\u2013331 (2019).\n\t36.\tArtzner, P., Delbaen, F., Eber, J. M. & Heath, D. Coherent measures of risk. \nMath. Finance 9, 203\u2013228 (1999).\n\t37.\tRockafellar, R. T. & Uryasev, S. Optimization of conditional value-at-risk.  \nJ. Risk 2, 21\u201341 (2000).\n\t38.\tStrengthening Reliability: An Analysis of Capacity Performance (PJM \nInterconnection, 2018); https://www.pjm.com/-/media/library/reports-notices/\ncapacity-performance/20180620-capacity-performance-analysis.ashx (2018).\n\t39.\tde Maere d\u2019Aertrycke, G., Ehrenmann, A. & Smeers, Y. Investment with \nincomplete markets for risk: the need for long-term contracts. Energy Policy \n105, 571\u2013583 (2017).\n\t40.\tRenewable Energy Certificates Annex to the EEI Master Power Purchase and Sale \nAgreement (Edison Electric Institute, 2010); www.eei.org/resourcesandmedia/ \nmastercontract/Documents/EEI%20RECs%20Annex%20v1.pdf\n\t41.\tD\u2019Aertrycke, G., Ehrenmann, A., Ralph, D. & Smeers, Y. Risk Trading in \nCapacity Equilibrium Models Cambridge Working Paper Economics: 1757 \n(University of Cambridge, 2017).\n\t42.\tAbada, I., de Maere d\u2019Aertrycke, G. & Smeers, Y. On the multiplicity of \nsolutions in generation capacity investment models with incomplete markets: a \nrisk-averse stochastic equilibrium approach. Math. Program. 165, 5\u201369 (2017).\n\t43.\tBritz, W., Ferris, M. & Kuhn, A. Modeling water allocating institutions based \non multiple optimization problems with equilibrium constraints. Environ. \nModel. Softw. 46, 196\u2013207 (2013).\n\t44.\tFerris, M. C. & Munson, T. S. Complementarity problems in GAMS and the \nPATH solver. J. Economic Dyn. Control 24, 165\u2013188 (2000).\n\t45.\tH\u00f6schle, H., Le Cadre, H., Smeers, Y., Papavasiliou, A. & Belmans, R. An \nADMM-based method for computing risk-averse equilibrium in capacity \nmarkets. IEEE Trans. Power Syst. 33, 4819\u20134830 (2018).\n\t46.\tFourer, R., Gay, D. M. & Kernighan, B. W. in Algorithms and Model \nFormulations in Mathematical Programming (eds. Wallace, S. W.) 150\u2013151 \n(Springer, 1989).\n\t47.\tIBM ILOG CPLEX optimizer (IBM, 2015).\n\t48.\tSchwele, A., Kazempour, J. & Pinson, P. Do unit commitment constraints \naffect generation expansion planning? A scalable stochastic model. Energy \nSyst. https://doi.org/10.1007/s12667-018-00321-z (2019).\n\t49.\tGlowinski, R. & Marroco, A. Sur l\u2019approximation, par \u00e9l\u00e9ments finis  \nd\u2019ordre un, et la r\u00e9solution, par p\u00e9nalisation-dualit\u00e9 d\u2019une classe  \nde probl\u00e8mes de Dirichlet non lin\u00e9aires. ESAIM: Math. Model. Num. 9,  \n41\u201376 (1975).\n\t50.\tGabay, D. & Mercier, B. A dual algorithm for the solution of nonlinear \nvariational problems via finite element approximation. Computers Math. Appl. \n2, 17\u201340 (1976).\n\t51.\tG\u00e9rard, H., Lecl\u00e8re, V. & Philpott, A. On risk averse competitive equilibrium. \nOper. Res. Lett. 46, 19\u201326 (2018).\nAcknowledgements\nThis research was supported, in part, by Northwestern University\u2019s Center for \nOptimization and Statistical Learning. The views expressed in this article are not \nnecessarily those of the Federal Energy Regulatory Commission.\nAuthor contributions\nJ.M. planned and performed the analysis. D.P.M. reviewed the manuscript, in particular \nwith reference to the models and algorithm. R.P.O. reviewed the manuscript, in \nparticular with reference to the electricity market design.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-019-0476-1.\nCorrespondence and requests for materials should be addressed to J.M.\nReprints and permissions information is available at www.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2019\nNature Energy | VOL 4 | NOVEMBER 2019 | 948\u2013956 | www.nature.com/natureenergy\n956\n\n\n    Scientific Research Findings:", "answer": "We find that from the perspective of investors, the financial impact of resource adequacy mechanisms is to replace highly volatile energy market revenues with relatively stable payments for capacity. This hedging property reduces the risk of investing in new generation, enabling developers to secure financing at lower cost. However, the quality of the hedge for each generation technology depends on how well the design of the mechanism aligns with its particular risk profile. Our findings suggest that the structure of current capacity mechanisms inadvertently favours generation resources with low capital costs and high operating costs, such as gas- and oil-fired peaking plants, over technologies with the opposite cost structure, like solar, wind and nuclear. Accordingly, current mechanisms may work against efforts to decarbonize.", "id": 31}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-018-0277-y\nEnergy Politics Group, Department of Humanities, Social and Political Sciences, ETH Zurich, Z\u00fcrich, Switzerland. *e-mail: florian.egli@gess.ethz.ch;  \nbjarne.steffen@gess.ethz.ch; tobiasschmidt@ethz.ch\nK\neeping climate change within safe limits and achieving the \ngoals of the Paris Agreement require fast and ample redirec-\ntion of financial flows towards low-carbon technologies1\u20133. As \napproximately two-thirds of global greenhouse gas emissions stem \nfrom the energy sector4, the rapid deployment of low-carbon energy \ntechnologies, such as renewable energy technologies (RETs), is cru-\ncial for emissions reductions5. Solar photovoltaics (PVs) and wind \nwill probably play central roles in this transition6. Importantly, as \nRETs are more capital intensive than fossil fuel technologies, large \nportions of their life-cycle cost are incurred upfront and need to \nbe financed7,8. Extant literature has established the adverse effect of \nhigh costs of capital on the levelized costs of electricity (LCOEs) for \nRETs9, CO2 abatement cost7,10 and RET deployment in integrated \nassessment models11. Consequently, high costs of capital are consid-\nered major obstacles to RET deployment12,13. By the same logic, low \ncosts of capital can contribute to the observed cost reductions for \nsolar PV and wind energy14\u201316. Financial markets thus have a con-\nstraining or enabling role in the low-carbon energy transition17\u201320.\nWhile individual investors know their cost of capital (CoC), typi-\ncally this information remains unavailable to researchers21,22, espe-\ncially concerning developments over time. This paper addresses this \ngap, by analysing the German solar PV and onshore wind power \nfinancing market, which has a particularly long investment his-\ntory23. We exploit the fact that utility-scale renewable energy invest-\nments in Germany are almost exclusively realized in project finance \nstructures24 (see Supplementary Note 1 for a description), in which \nthe costs of capital reveal unbiased information about the under-\nlying investment projects and technologies25. We proceed in four \nsteps. First, using newly compiled project data, we depict the CoC \nand its components and analyse the changes over 18 years. Second, \nwe use qualitative insights from in-depth interviews with 41 invest-\nment professionals to identify the drivers of the observed changes \nin financing conditions. Third, we quantify an experience effect \nwithin the renewable energy finance industry, leading to lower costs \nof capital. Fourth, we quantify the effect of the observed changes in \ncosts of capital on LCOEs. The methods are structured along the \nsame four steps. We find that the CoC declined by 69% for solar \nPV and by 58% for wind onshore projects between the early period \nof the RET finance industry (2000\u20132005) and 2017. For both tech-\nnologies, the cost of debt decreased more than the cost of equity. \nFocusing on the cost of debt, we identify and estimate a financing \nexperience curve. For each doubling of cumulative investment, the \ndebt margins (see Supplementary Table 1 for definitions of financial \nterms) decreased by 11% for both technologies. During the same \ntime, we observe a decline in the general interest rate resulting in \nlower costs of capital that had a substantial effect on the economic \nattractiveness of RETs. Finally, we estimate that 41% of total solar \nPV LCOE reductions and 40% of wind onshore LCOE reductions \nbetween 2000\u20132005 and 2017 were due to lower financing costs. \nThese result from three effects: lower capital expenditures (CAPEX) \nto be financed (strongest effect for solar PV), lower general interest \nrate (strongest effect for wind onshore), and financing experience. \nWe conclude with implications for researchers and policymakers.\nChanges in financing conditions\nIn the first step, we analyse the temporal dynamics of the CoC and \nits components (see Methods). We compile data on the financing \nconditions of 133 representative utility-scale renewable energy proj-\nects, undertaken between 2000 and 2017, to establish the temporal \ndynamics of costs of capital for solar PV and wind onshore. The \nproject data are provided by leading renewable energy investors, \ncovering lead arrangers responsible for 85% of the solar PV and 80% \nof the wind onshore investment sums between 2000 and 2017 (see \nSupplementary Fig. 1). Figure 1 displays the cost of debt, the cost \nof equity and the CoC for all projects in our data set. Both solar \nPV and wind onshore projects experienced substantial decreases \nin costs of capital. While some variance in CoC is normal due to \nslightly different project conditions, the data show a clear decrease \nin the lower bound for cost of debt and cost of equity over time. The \nlower bound of cost of debt dropped from around 5% to less than \n0.5% for both technologies. Lower bound equity returns fell from \naround 10% to below 4%.\nFigure 2 draws on the same data as Fig. 1 to calculate the average \nacross projects and compares the early period of the RET finance \nindustry (2000\u20132005) to 2017. It first shows that the cost of debt \ndecreased more than the cost of equity in relative terms and that \nA dynamic analysis of financing conditions for \nrenewable energy technologies\nFlorian\u00a0Egli\u200a \u200a*, Bjarne\u00a0Steffen\u200a \u200a* and Tobias\u00a0S.\u00a0Schmidt\u200a \u200a*\nRenewable energy technologies often face high upfront costs, making financing conditions highly relevant. Thus far, the dynam-\nics of financing conditions are poorly understood. Here, we provide empirical data covering 133 representative utility-scale \nphotovoltaic and onshore wind projects in Germany over the last 18 years. These data reveal that financing conditions have \nstrongly improved. As drivers, we identify macroeconomic conditions (general interest rate) and experience effects within the \nrenewable energy finance industry. For the latter, we estimate experience rates. These two effects contribute 5% (photovoltaic) \nand 24% (wind) to the observed reductions in levelized costs of electricity (LCOEs). Our results imply that extant studies may \noverestimate technological learning and that increases in the general interest rate may increase renewable energies\u2019 LCOEs, \ncasting doubt on the efficacy of plans to phase out policy support.\nNature Energy | VOL 3 | DECEMBER 2018 | 1084\u20131092 | www.nature.com/natureenergy\n1084\n\nArticles\nNATuRE EnERGy\ndecreases in both components were more pronounced for solar PV \nthan for wind onshore. However, the project CoC also depends on \nthe leverage and the corporate tax rate. Leverage denotes the share \nof debt of the total investment sum (see Methods). As equity bears \nthe first project losses, a higher leverage is an indication for lower \nproject risk. For both technologies, the leverage increased, reaching \nover 80% debt financing in 2017 (see Supplementary Fig. 2). During \nthis period, the German corporate tax rate decreased from 41% to \n30%, resulting in relatively higher costs of debt as interest rate pay-\nments are deductible from taxable revenues.\nFigure 2c,d summarizes the resulting after-tax CoC. The CoC in \n2017 was in the range of 1.6% (solar PV) to 1.9% (wind onshore), \ncorresponding to a low-risk corporate bond of a financial service \nfirm (BB+\u200b to BBB)26. Stated differently, CoC declined by over two-\nthirds (3.5 percentage points) for solar PV projects and more than \nhalf (2.6 percentage points) for wind onshore projects. While the \nCoC for solar PV projects in 2000\u20132005 was higher than for wind \nonshore projects, the former had a lower CoC than the latter in \n2017. Similar trends were observed for additional financial indica-\ntors, such as loan tenors and debt service coverage ratios (DSCRs; \nsee Supplementary Fig. 2). Over our study period, the duration of \nthe feed-in tariff stayed constant at 20 years. Banks offering longer \nloan tenors is therefore an indication of higher confidence in the \nproject. The DSCR is a measure of project cash flows available to \n0\n5\n10\n15\n2005\n2000\n2015\n2010\n2020\nCost of capital (%)\n0\n5\n10\n15\n2010\n2005\n2000\n2015\n2020\nSolar PV\nWind onshore\nCost of equity\nCost of debt\nCost of capital\na\nYear\nb\nYear\nCost of equity\nCost of debt\nCost of capital\nFig. 1 | CoC over time. a,b, Cost of debt, cost of equity and CoC in Germany for 43 solar PV (a) and 78 wind onshore (b) projects between 2000 and 2017 \n(N\u2009=\u200b\u2009121). We show CoC numbers only for projects where cost of debt and cost of equity, as well as capital structure (leverage), are known (29 solar PV \nand 26 wind onshore projects).\n9.3\n4.8\n5.5\n1.5\n0\n2\n4\n6\n8\n10\n\u201349%\n\u201372%\nCost of equity\nCost of debt\n2.3\n2.8\n2.2\n2.5\n1.2\n0.6\n6\n0\n2\n4\n0.9\n5.1\n1.6\n\u201369%\nAvg 2000\u20132005\nAvg 2000\u20132005\n2017\n2017\nchange\nchange\n9.5\n5.4\n4.8\n1.6\n0\n2\n4\n6\n8\n10\n\u201343%\n\u201367%\n2.1\n2.4\n1.4\n2.3\n1.1\n1.0\n4\n0\n6\n2\n1.9\n0.9\n4.5\n\u201358%\nUnleveraged,\nno tax\nconsidered\nLeveraged,\nafter-tax\nSolar PV\nWind onshore\na\nb\nc\nd\nLev. = 0.7\nTax = 41%\nLev. = 0.87\nTax = 30%\nLev. = 0.75\nTax = 41%\nLev. = 0.82\nTax = 30%\nTax\nTax\n%\n%\n%\n%\nFig. 2 | Components and dynamics of CoC. a,b, Changes in unleveraged cost of debt and cost of equity for solar PV (a) and wind onshore (b) projects.  \nc,d, Changes in leveraged (after-tax) CoC for solar PV (c) and wind onshore (d) projects. The tax effect was due to a decrease in the corporate tax rate \nthat led to a smaller cost reduction from tax-deductible debt interest payments.\nNature Energy | VOL 3 | DECEMBER 2018 | 1084\u20131092 | www.nature.com/natureenergy\n1085\n\nArticles\nNATuRE EnERGy\npay debt obligations, namely the principal repayment and interest \nrate payments. Lower DSCRs can thus be interpreted as an addi-\ntional indication for lower project risk.\nDrivers of change\nIn the second step, we use qualitative interviews with investment \nprofessionals (N\u2009=\u200b\u200941, see Supplementary Table 3) to inductively \nreveal and understand the underlying drivers of the observed \nchanges in financing conditions27. Averaging more than ten years of \nrenewable energy investment experience, the interviewees demon-\nstrate an in-depth understanding of the market dynamics over time. \nFrom these interviews, we distil drivers of CoC reductions on three \nnested levels: the macroeconomic environment (economy), the \nrenewable energy sector and the renewable energy finance indus-\ntry (see Methods). The latter two are related to experience gained \nthrough deployment and financing of RET. Figure 3 illustrates the \nmain drivers on the three identified levels.\nOn the economy level, expansive monetary policies in the after-\nmath of the 2008\u20132009 financial crisis resulted in low refinancing \ncosts for banks, which decreased the CoC of the economy28. The \nlarge supply of capital increased the pressure on bank fees and even-\ntually lowered them too. At the same time, extensive bank lending \ntended to lead to overconfident credit issuance, thereby increasing \ndefault rates29,30. The extensive lending made the evaluation of com-\npanies\u2019 credit eligibility more difficult and thereby increased the \ninvestment attractiveness of projects with predictable cash flows, \nsuch as RET assets in project finance structures.\nOn the renewable energy sector level, technology deployment \nhad a favourable impact on financing conditions. As more renewable  \nenergy projects were undertaken, technologies became more \nmature (that is, more reliable). In parallel, the availability of data \non technology performance made an assessment of this increasing  \nreliability possible, while financial data showed low default rates. \nTogether with a higher confidence in partners for construction \nand operation of RET assets, as these companies had increasingly \nestablished track records, these developments provided impetus \nfor investment professionals to convince their boards to invest in \nrenewable energy assets. According to our qualitative results, the \ndeployment effect was more pronounced for solar PV than for  \nwind onshore projects because of wind turbines\u2019 larger operational \nrisk due to their design complexity31 and moving parts. Moreover, \nwind resource availability is more difficult to predict than solar \nirradiation. Partly as a result of these factors, the CoC decreased  \nfaster for solar PV projects than for wind onshore projects. Finally, \nstable and reliable RET support policies were a prerequisite for \nRET investment in Germany in the past23. However, gradually, \nsome RET projects are being partly exposed to market prices32,33, \nwhich is reflected by few shorter loan tenors and higher DSCRs (see \nSupplementary Fig. 2).\nOn the level of the renewable energy finance industry, investors \nbenefited from growing RET markets and subsequent learning-by-\ndoing (for example, better risk assessment)34. Larger markets allowed \nbanks to form in-house project finance teams specialized in RETs. \nThe knowledge and data that these teams accumulated allowed for \na more accurate technology assessment. Consequently, project risks \ndeclined. For example, as the market had accumulated experience \non historical wind speeds, investors shifted from calculating project \nreturns on wind resource estimations with 90% certainty (90th per-\ncentile of the distribution, p90) to trusting the median (p50). While \nthe observed increases in loan tenors and decreases in DSCRs (see \nSupplementary Fig. 2) confirm lower project risk, we see two diver-\ngent trends in project leverage. On the one hand, investors advanced \nto higher leverages to increase returns on equity; on the other hand, \nsome investors started to accept lower leverages to place their equity \nin a market environment with few renewable energy investment \nopportunities on offer.\nMoreover, the investor ecosystem matured and competition \nincreased. In a maturing investment market, institutional investors \n(for example, insurers and pension funds) started to perceive renew-\nable energy project finance as an attractive asset class. Institutional \ninvestors usually demand lower returns and larger project sizes than \nsmaller early-stage investors35. The capital inflow from the new \ngroup of institutional investors hence created an incentive to build \nlarger projects and increased competition for projects, which gener-\nally compressed debt margins further. While lower margins lead to \nlower LCOEs and are thus potentially conducive to RET deploy-\nment, some investors fear that the increasing capital inflow could \ncreate an asset bubble with financing conditions that would no lon-\nger reflect project risks. Last, the use of standardized deal structures \nfacilitated the investment process and contributed to more efficient \nfinancing markets with lower margins.\nFinancing experience rates\nThe third step of our analysis focuses on the effects that are related \nto experience with deployment and financing of RETs (see Fig. 3). \nThe innovation literature has identified a roughly constant percent-\nage unit cost decline\u2014the experience or learning rate\u2014with each \ndoubling of cumulative production (Wright\u2019s law, see Methods)36. \nThis experience effect is a well-known characteristic of RET invest-\nment cost14\u201316,37,38. Our results from the second step demonstrate that \nexperience matters for the renewable energy finance industry. We \nhence propose a financial experience effect analogous to Wright\u2019s \nlaw39 and estimate a corresponding experience rate.\nTo identify the experience rate, we focus on debt, because this \nis where most cost reductions have occurred (compare Fig. 2). We \nEconomy\nRenewable\nenergy\nsector\nRenewable\nenergy\nfinance\nindustry\nDrivers of changes in financing conditions\nCapital markets: low-cost liquidity, few\ninvestment alternatives, low return\nexpectations\nBanks: low-cost refinancing, low bank\nfees, preference for project finance\nAvailability of performance data:\naccumulated operation experience of\nRET assets\nTechnology reliability: proven track\nrecord of technology, low default rates of\nprojects\nSupport policies: regulatory\nenvironment; for example, introduction of\nexposure to market risks\nLearning by doing: in-house RET\nknowledge, better risk assessment and\ndue diligence processes\nInvestment ecosystem: standardized\ninvestment structures, frame contracts,\npartner networks\nMarket entry of investors: new investor\ntypes (for example, large banks, insurers,\npension funds), increasing investor competition\nLevel\nDrivers related to experience with deployment and financing of RET\nDrivers related to general economic development\n1\n2\n1\n2\nFig. 3 | Drivers of changes in financing conditions in a nested hierarchy. The \ngeneral economic environment led to more favourable financing conditions \n(for all sectors) over the period of our study. All drivers in the renewable \nenergy sector and the renewable energy finance industry contributed to \nmore favourable financing conditions (in the renewable energy sector) over \nthe period of our study, with the exception of changes to support policies, \nwhich potentially introduce new uncertainties into the RET sector.\nNature Energy | VOL 3 | DECEMBER 2018 | 1084\u20131092 | www.nature.com/natureenergy\n1086\n\nArticles\nNATuRE EnERGy\nanalyse three debt indicators that reflect investment safety mar-\ngins, namely, the debt margin, DSCR and loan tenors. For riskier \nprojects, investors demand higher debt margins as compensation, \na higher DSCR to create a buffer in case of cash flow complica-\ntions, and a lower loan tenor to reduce the risk exposure to a shorter \nperiod. More experienced investors should be able to judge invest-\nment projects more accurately, thereby reducing the required safety \nmargins and generating an empirically observable experience effect. \nFigure 4 shows the experience rates for the three variables. We find \nan experience rate of 11% on the debt margins of both technologies. \nWe also detect experience rates of 13% for the DSCR of solar PV \nprojects and of 17% for the DSCR of wind onshore projects (see \nMethods). Regarding the loan tenors, we find an experience rate \nof \u2212\u200b3%, that is, increasing loan tenors with increasing experience. \nHowever, this finding is insignificant for wind onshore projects. \nIn sum, the third step of our analysis establishes the statistical sig-\nnificance of the experience effect in renewable energy financing, as \nfound qualitatively in the second step. Increased RET deployment \ncontributes to better financing conditions.\nIn the following, we compare the experience effect with the exog-\nenous effect (economy level) from changes in general interest rates. \nThe cost of debt of a RET project can be decomposed into two ele-\nments covering the baseline country risk and project specific risk40. \nFigure 5 shows the yields of a 10-year German government bond \n(the best proxy for baseline country risk) and the estimated debt \nmargins (the best proxy for project specific risk). While the bond \nyields are driven by monetary policy and exogenous to renewable \nenergy deployment, the debt margins reflect dynamics related to \nexperience with deployment and financing of RET.\nThree observations can be made in Fig. 5. First, the change in debt \nmargins seems small compared with government bond yields but is \neconomically substantial. While government bond yields decreased \nby 5 percentage points, debt margins have declined by 1.5 percentage \npoints for solar PV projects and 1 percentage point for wind onshore \nprojects between 2000 and 2017. For comparison, this experience-\ndriven decrease corresponds to a change in the corporate ratings of a \nfinancial service firm from B+\u200b to AAA for solar PV or from BBB to \nAAA for wind onshore26. Second, Fig. 5 reveals different dynamics \nbetween the two technologies. Due to larger increases in cumulative \ninvestment for solar PV, its debt margin decreased more than was the \ncase for wind onshore projects. As a relatively novel technology, solar \nPV projects were perceived riskier and thus charged with a higher \ndebt margin in 2000. In 2017, investors no longer make a difference \nand charge almost identical margins. This catch-up of solar PV con-\nfirms the pattern shown in Fig. 2 and the qualitative findings from \nthe previous section. Third, debt margins are higher than the base-\nline country risk rate in 2017, largely as a result of exceptionally low \ngovernment bond yields due to the expansive monetary policy after \nthe financial crisis. Considering the observed trend towards higher \nleverages and the concurrently increasing importance of the cost of \ndebt, this finding points out that changes in the general interest rate \nlevel potentially have a large impact on the CoC for RETs.\nImpact on LCOE\nIn the fourth and final step, we calculate the LCOE in the early \nperiod of the RET finance industry (2000\u20132005) and in 2017 (see \nMethods).\n2000\n2005\n2010\n2015\n2020\n5\n4\n0\n1\n2\n6\n3\nYear\n0.3%\n2.1%\nBond yield or margin (%)\n2.5%\n5.3%\n1.0%\n1.1%\nSolar PV debt margin\nWind onshore debt margin\nGeneral interest rate level\n(10-year German government bond)\nFig. 5 | Changes in the general interest rate level versus debt margins. \nDebt margins are predicted values using the estimated experience rate from \nFig. 4 and global investment data from 2000 to 2017 (see Methods). A data \nvalidity check regarding the decomposition of the cost of debt into debt \nmargin and government bond yield is provided in Supplementary Fig. 4.\n0.1\n1.0\n10.0\n1\n100\n10,000\n0.01\n0.1\n1.0\n1\n100\n10,000\n1\n10\n100\n1\n100\n10,000\n0.01\n0.1\n1.0\n1\n100\n10,000\n1\n10\n100\n1\n100\n10,000\n0.1\n1.0\n10.0\n1\n100\n10,000\na\nb\nc\nd\ne\nf\nER = 11 \u00b1 8%\nER = 11 \u00b1 7%\nER = 13 \u00b1 10%\nER = 17 \u00b1 8%\nER = \u20133 \u00b1 2%\nER = \u20133 \u00b1 3%\nWind onshore\nSolar PV\nDebt margin (%)\nDSCR\nLoan tenor (years)\nCumulative world investment\n(US$ billion)\nCumulative world investment\n(US$ billion)\nFig. 4 | Experience rates for risk metrics including the 95% confidence \ninterval. a,b, Debt margin experience rate (ER) for solar PV projects \n(N\u2009=\u200b\u200927) (a) and for wind onshore projects (N\u2009=\u200b\u200922) (b). c,d, DSCR \nexperience rate for solar PV projects (N\u2009=\u200b\u200935) (c) and for wind onshore \nprojects (N\u2009=\u200b\u200936) (d). e,f, Loan tenor experience rate for solar PV projects \n(N\u2009=\u200b\u200936) (e) and for wind onshore projects (N\u2009=\u200b\u200934) (f). All axes are \nlogarithmic, and fits are linear. All linear fits are significant at the 5% \nlevel or below, except for the wind onshore loan tenors (f). The data \nare from German projects between 2000 and 2017. Between 2000 and \n2017, global cumulated solar PV investment doubled eight times, and \nwind onshore investment just short of six times. The results are robust to \nincluding investor fixed effects (for example, controlling for different sizes \nof investors), choosing Europe as the relevant scope for experience (that \nis using European instead of global investment data) and using alternative \ndata to measure investment (see Supplementary Tables 5\u20137).\nNature Energy | VOL 3 | DECEMBER 2018 | 1084\u20131092 | www.nature.com/natureenergy\n1087\n\nArticles\nNATuRE EnERGy\nFigure 6 shows that the LCOE declined for both technolo-\ngies, bringing both technologies into the generation cost ranges \nfor fossil fuel-fired power plants, estimated to be between US$50 \nand US$170 for G20 countries in 201741. Around 60% of this \ndecline is due to lower technology cost (CAPEX) with the remain-\ning 40% due to lower financing cost. Three effects contribute to \nthe change in financing costs. First, the initial investment to be \nfinanced (CAPEX) decreased, which lowers the financing cost. \nSecond, the general interest rate decreased. Third, an experience \neffect led to the compression of financing margins. The three \neffects differ in importance between the two studied technologies. \nThe large reduction in solar PV CAPEX during the period of our \nstudy (see Supplementary Table 4) led to lower financing costs, \nwhich contributed roughly one-third (36%) of LCOE reductions. \nConversely, onshore wind CAPEX stayed relatively constant (see \nSupplementary Table 4), increasing the relative importance of the \ngeneral interest rate effect, which contributed one-fifth (20%) of \nLCOE reductions. Thus, the channels through which financing \ncosts contribute to lowering LCOEs vary according to the relative \nreductions in CAPEX. As solar PV and wind onshore are becom-\ning mature technologies and future CAPEX reductions become less \nlikely, the relative importance of the general interest rate and expe-\nrience effects will increase.\nDiscussion\nThis paper compiles a project-level data set for financing conditions \nand makes three contributions. First, it identifies the drivers of the \nchanges in financing conditions. Second, it estimates an experience \neffect for financing conditions and compares it with the changes \nin the general interest rate. Third, it demonstrates the effect of the \nchanges in financing conditions on the LCOE.\nFor researchers, our results suggest that the dynamics of \nfinancing conditions should receive more attention in models \nthat include investments in low-carbon technologies. In failing to \naccount for these dynamics, researchers could overestimate the \ntechnology learning effect by attributing the full LCOE change \nto reductions in capital and operating expenditures. Accounting  \nfor different channels of LCOE reductions via financing costs may \nbe particularly important\u2014especially as the increasing use of auc-\ntions makes data on generation costs readily available, increas-\ning the use of LCOE learning curves41. To include sensitivity  \nanalyses regarding the dynamics of financing conditions in models,  \nfurther research should help improve the understanding of  \nthe processes that affect renewable energy financing condi-\ntions. While we have separated three effects contributing to  \nfinancing costs\u2019 LCOE effect, their dynamics are yet to be fully  \nunderstood, opening up avenues for future research. For example, \nit is not evident whether deployment and the associated reduc-\ntions in investment costs would have been as large as observed \nwithout reductions in the CoC. The accumulation of experi-\nence in the finance industry, the excess availability of capital and  \nthe reductions in investment costs all depend on each other  \nand together constitute the impact of financing costs on  \nthe LCOE. Future research also should investigate to what extent \nthis paper\u2019s conclusions are applicable to other regions and  \nother technologies.\nFor policymakers, our findings stress the importance of \npolicies that are conducive to favourable financing conditions \nfor RETs. First, our results suggest an important co-benefit of \ndeployment policies: the acceleration of technological change \nby allowing the finance industry to experiment and learn. RET \ninvestments are long-term, and the finance industry typically \nstruggles to assess long-term risks of new technologies without \na track record34,42. For instance, green state investment banks  \ncan be an instrument to accelerate learning in the finance indus-\ntry, helping investors assess projects and build confidence in \nnew technologies43. Second, our results indicate that a large RET \nfinancing market and a high degree of competition between inves-\ntors were crucial in creating more favourable financing conditions \nfor RETs. Therefore, policies should try to crowd-in a broad spec-\ntrum of investors. Third, our findings point out that policymakers  \nshould be vigilant in responding to changes in monetary poli-\ncies that have an impact on RET costs. As RET generation costs \napproach grid parity, policymakers in some countries consider \nphasing out fixed remuneration schemes for RETs. While some \nhave argued that achieving high RET shares requires de-risking \npolicies in any case44, our results stress the particular impor-\ntance of policy intervention (for example, RET support or carbon  \npricing) given the likelihood of an imminent increase in interest \nrates. Ending policies might be premature and put climate change \ntargets at risk. Policymakers also could evaluate new approaches, \nsuch as green monetary policies, to ensure attractive financ-\ning conditions for RETs and other low-carbon technologies in  \nthe future45.\n313\n187\n8\n0\n100\n200\n300\n400\n500\n1%\n59%\n36%\nLCOE (US$ MWh\u20131)\n4%\n59\n2017\n500\n67\nWind onshore\nSolar PV\na\n41%\nchange in\nfinancing cost\n85\n28\n54\n7\n0\n20\n40\n60\n80\n100\n120\nLCOE (US$ MWh\u20131)\n4%\n60%\n20%\n17%\n2017\n62\n113\n40% \nchange in\nfinancing cost\nGeneral interest rate effect\nExperience effect\nLower capital expenditures\nChange in financing cost from:\nCapital and operating expenditures\nFinancing expenditures\nLCOE components\nb\nAvg 2000\u20132005\nAvg 2000\u20132005\nFig. 6 | Historical impact of changes in financing costs on LCOE. a,b, LCOE for solar PV (a) and wind onshore (b). The percentages indicate the \ncontributions of the respective parts to the change in LCOE. We parameterize the LCOE model using data for Germany (see Supplementary Table 4). \nSensitivities are provided in Supplementary Fig. 5. The numbers do not always add up due to rounding.\nNature Energy | VOL 3 | DECEMBER 2018 | 1084\u20131092 | www.nature.com/natureenergy\n1088\n\nArticles\nNATuRE EnERGy\nMethods\nCase selection. The case selection includes three dimensions: technology, country \nand project type. First, we focus on solar PV and wind onshore technologies, the \nmost deployed non-hydro RETs. In 2016, solar PV and wind onshore technologies \naccounted for a global capacity of 291\u2009GW and 452\u2009GW, respectively (for example, \ncompared with 14\u2009GW for wind offshore generation)46. Second, we focus on \nGermany, one of the earliest markets to adopt these technologies. Germany \nadded the most solar PV capacity in 13 of the 17 years analysed, and the most \nwind onshore capacity in 8 of the 17 years analysed46. Our sample period begins \nin 2000, when Germany enacted its landmark legislation on renewable energy \nsources (EEG), with a feed-in tariff that triggered large-scale renewable energy \ninvestments23. The feed-in tariff was never changed retroactively. The German \nelectricity market has been liberalized since 199823, and the vast majority of \ninvestment in RET was private47. Third, we restrict the analysis to project finance \nstructures, exploiting the fact that 96% of large solar PV projects and 88% of large \nwind onshore projects in Germany between 2000 and 2015 were undertaken using \nproject finance24.\nData collection. We contacted leading investors directly to assemble two sets \nof data: quantitative data on the financing conditions of reference projects, and \nqualitative data on the drivers of changes in financing conditions. The former is \nused for steps one, three and four of the paper, and the latter is used for step two. \nAll investor interviews were conducted between September 2017 and January \n2018, following the Chatham House Rule, which states that \u2018participants are free \nto use the information received, but neither the identity nor the affiliation of \nthe speaker(s) [\u2026\u200b] may be revealed\u201948. The interviews were conducted in person \nor over the phone by one to three researchers who took individual notes. All \ninterviews were recorded and transcribed verbatim.\nWe use theoretical sampling to include the most revelatory interviewees and \nbalance our sample to represent various perspectives from the finance industry27,49. \nThe sampling took place in three stages. First, we searched for publicly available \naddresses of senior investment professionals working at large debt and equity \ninvestment firms, using the Bloomberg New Energy Finance database50. Second, \nwe used the contact network of a private renewable energy finance industry \npartner in the INNOPATHS research consortium, Allianz Climate Solutions, to \nreach out to relevant market actors. Third, we employed snowball sampling by \nasking key contacts from our network to refer us to relevant actors and teams, and \nthen continued to ask for references following each contact with an investment \nprofessional. The resulting sample is well balanced among different kinds of \nfinancial actor and includes 17 debt providers (13 commercial banks and 4 \ninvestment banks), 16 equity providers, 7 public actors (4 public utilities and 3 \npublic investment banks) and 1 former researcher (see Supplementary Table 3 for \nthe full interviewee sample). The sampled financial actors were lead arrangers in \n81% of solar PV capacity additions and 85% of the solar PV investment sum, and in \n49% of onshore wind capacity additions and 80% of the onshore wind investment \nsum, between 2000 and 2017 (see Supplementary Fig. 1). Thus, our sample covers \nthe relevant actors in a balanced manner and is relevant in size to elicit financing \nconditions that are representative of the German investment market. Reflecting the \ninternational nature of the renewable energy finance industry, the investors in our \nsample are based in Germany, Switzerland, the UK, the Netherlands, France, Italy, \nLuxembourg and Norway (see Supplementary Table 3).\nQuantitative data. To ensure comparable data on project financing conditions, we \ndefined a reference project with an investment sum of \u20ac\u200b20 million, using standard \ntechnology from established manufacturers (polycrystalline modules without a \ntracker for solar PV projects and 1.5\u20132\u2009MW turbines on a standard foundation \nfor wind onshore projects). While the relatively small standard deviation of the \ndata (see Supplementary Table 2) indicates a good comparability of projects \nacross investors, we control for investor differences (for example, investor size) by \nincluding investor fixed effects in the estimations of experience rates (see below).\nWe asked the investment professionals to provide information on the all-in \nCoC, cost of debt, cost of equity, debt margin, DSCR, leverage (that is, project \ncapital structure) and loan duration (tenor) for any such reference project that they \nhad financed (which is the case for 37 interviewees) or had advised on (which is \nthe case for 4 interviewees; see Supplementary Table 3) between 2000 and 2017 (see \nFig. 1 and Supplementary Fig. 2). For CoC components (cost of debt, cost of equity \nand leverage), the interviewees were free to indicate ranges instead of absolute \nvalues, in which case, we take the average by project. Wherever possible, the debt \nproviders indicate not only the all-in cost of debt, but also the debt margins. For \nprojects with available information on debt margins, we calculate the all-in cost \nof debt as the sum of the baseline rate (10-year government bond51) and the debt \nmargin. This approach yields all-in cost of debt data comparable to where debt \nproviders revealed all-in costs of debt (see Supplementary Fig. 4). We also screen \npublicly available onshore wind park investment prospectuses\u2014mainly from civic-\nowned assets (German B\u00fcrgerwindparks)\u2014between 2000 and 2017 for the data on \nthe cost of debt. We do not consider this source for the cost of equity data because \ninvestment prospectuses often offer overly optimistic equity returns ex ante. On the \nother hand, the cost of debt figures reflect the rates offered by banks. The resulting \ndata set consists of 48 solar PV and 85 wind onshore projects. The number of \nobservations, means, standard deviations, and minimums and maximums for all \nvariables are described in Supplementary Table 2.\nTo estimate experience rates, we use investment data from the United Nations \nEnvironment Programme52, which is available from 2004 onward. For the years \nbefore 2004, we take global investment costs per megawatt for solar PV and \nwind onshore projects50 and multiply these figures with global capacity from the \nInternational Renewable Energy Agency (IRENA)46. For 2000 and 2001, we used \nthe solar PV investment costs from 2002 because ref.\u200946 provides no data. For 2017, \nwe extrapolate the changes from the previous year.\nQualitative data. To develop the drivers of changes in financing conditions, we \napply an interview case study design with two stages of data collection53. First, \nopen exploratory interviews (N\u2009=\u200b\u20098) were conducted to gain early insights on \nthe dynamics and drivers of changes in financing conditions and to define the \nstructure for the second phase of the interviews. Second, we conducted 33 semi-\nstructured interviews with employees from debt and equity investment firms \nwho had significant experience in the renewable energy finance industry (23 of \nthese interviewees are the same individuals who provided the quantitative data \nmentioned above). Note that we contacted three investment professionals from the \nexploratory interviews again for the semi-structured interviews and the collection \nof project financing conditions data.\nIf more than one researcher conducted an interview (N\u2009=\u200b\u200915), one of them \nsummarized it using the recording, transcript and notes. If only one researcher \nconducted the interview (N\u2009=\u200b\u200926), the resulting summary was cross-checked by \nanother researcher. This procedure ensures accurate and consistent recording, \nexpands the scope of insights and enhances confidence in the findings53. Following \nEisenhardt\u2019s approach53, we continued holding interviews until no additional \ninsights were observed.\nChanges in financing conditions. In the first step of the paper, we show the data as \nelicited without tax considerations and we calculate the after-tax CoC because the \nGerman corporate tax rate was cut 4 times, from an initial 52% in 2000 to 30% in \n2008, and remained at that level until 201754 (see Supplementary Fig. 3). Equations \n(1) and (2) define the CoC without considering tax and the after-tax CoC.\n=\n+\nK E\nV\nK D\nV\nCoCnotaxconsidered\n(1)\nE\nD\n-\n=\n+\n\u2212\nK E\nV\nK D\nV\nT\nAfter tax CoC\n(1\n)\n(2)\nE\nD\nIn these equations, E and D denote equity and debt investment, respectively; V \nsignifies the total investment sum; KE and KD refer to cost of equity and cost of \ndebt, respectively; and T represents the corporate tax rate. The leverage L is equal \nto D/V. To analyse the changes over time, we use the average during the 2000\u20132005 \nperiod as the starting point due to limited data availability in the early years. As \ncosts of capital decreased already between 2000 and 2005, this approach yields a \nconservative estimate for the changes over time.\nTaking the derivatives of equation (2) yields equations (3)\u2013(6), below. \nEquations (3) and (4) show that the changes in the cost of equity and debt affect \nthe CoC, depending on leverage and corporate tax rate. Equation (5) shows that \nthe effect of increasing leverage depends on the difference between KD and KE. \nMore precisely, if (1\u2009\u2212\u200b\u2009T)KD\u2009<\u200b\u2009KE holds, the CoC decreases with increasing project \nleverage. Typically, this condition holds in reality (see characteristics of project \nfinance above). Equation (6) illustrates that a change in the tax rate affects costs of \ncapital in the opposite direction; that is, a decrease in the tax rate increases costs of \ncapital.\n\u2202\n\u2202\n=\n\u2212\nK\nL\nCoC\n(1\n)\n(3)\nE\n\u2202\n\u2202\n=\n\u2212\nK\nT L\nCoC\n(1\n)\n(4)\nD\n\u2202\n\u2202\n=\n\u2212\n\u2212\nL\nT K\nK\nCoC\n(1\n)\n(5)\nD\nE\n\u2202\n\u2202\n= \u2212\nT\nLK\nCoC\n(6)\nD\nFigure 2c,d represents this fact with a grey upward bar for tax changes, indicating \nthe higher CoC due to the lower corporate tax rate.\nDrivers of change. In the second step, we use qualitative data to establish the \ndrivers behind the changes in financing conditions. The interviews were semi-\nstructured in the sense that the interviewees were free to name and explain the \nNature Energy | VOL 3 | DECEMBER 2018 | 1084\u20131092 | www.nature.com/natureenergy\n1089\n\nArticles\nNATuRE EnERGy\nmain drivers that led to the changes in financing conditions, but the conversations \nfollowed a pre-determined set of topics. At the end of each interview, we asked \nthe interviewee whether crucial points were missing. This feedback was included \niteratively in the first few interviews. Key statements were summarized by two \nresearchers after each interview. Once the interview summaries were completed \n(see data collection above), we loosely followed the \u2018grounded theory\u2019 approach55 \nby comparing incident (that is, statement) to incident to iteratively create common \npatterns and drivers. We constantly compared new incidents with emerging drivers \n(recursive cycling among different investor interviewees)27. Two researchers \nconducted this \u2018constant comparison\u2019, verifying drivers and ensuring their \naccuracy56. As a result, we identified eight drivers, which we categorized in a nested \nhierarchy of three levels: economy, renewable energy sector and renewable energy \nfinance industry.\nFinancing experience rates. In the third step, we apply a one-factor experience \ncurve, following Wright\u2019s law39, and adapt it to financial indicators. Applying a \none-factor experience curve may lead to estimates that are biased upwards due to \nan omitted variable bias57. The most commonly cited omitted factor is research \nand development (R&D) spending38,58. However, service industries, such as the \nfinance industry, typically do not use R&D departments, or even the term R&D. \nInstead, innovation activities are organized in project-based teams59. Perhaps as \na consequence, some empirical evidence even points to a negative effect of R&D \nspending on service innovation60. Finally, the evidence of our interviews points to \nfactors such as track records, improved processes or market competition as drivers \nof the experience effect. Quantifying these factors individually is impossible, which \nis why we choose to use a one-factor experience curve and discuss the components \nqualitatively in step two. For each of the financial indicators (that is, debt margin, \nDSCR and loan tenor), we define experience curves as follows:\n\uf8eb\n\uf8ed\n\uf8ec\uf8ec\uf8ec\n\uf8f6\n\uf8f8\n\uf8f7\uf8f7\uf8f7\uf8f7\n=\n\u2212\nI\nI\nI\nI\nDebtMargin( )\nDebtMargin( )\n(7)\nt\nt\nb\n0\n0\n1\n\uf8eb\n\uf8ed\n\uf8ec\uf8ec\uf8ec\n\uf8f6\n\uf8f8\n\uf8f7\uf8f7\uf8f7\uf8f7\n=\n\u2212\nI\nI\nI\nI\nDSCR( )\nDSCR( )\n(8)\nt\nt\nb\n0\n0\n2\n\uf8eb\n\uf8ed\n\uf8ec\uf8ec\uf8ec\n\uf8f6\n\uf8f8\n\uf8f7\uf8f7\uf8f7\uf8f7\n=\n\u2212\nI\nI\nI\nI\nTenor( )\nTenor( )\n(9)\nt\nt\nb\n0\n0\n3\nIn equation (7), DebtMargin denotes the debt margin in percentage points. In \nequation (8), DSCR signifies the transformed DSCR. We transform the elicited \nDSCR values by subtracting 1 because the DSCR has a natural lower bound of 1. \nAs we are taking the log in the next stage, we transform one value in our sample \nfrom 0 to 0.01. In equation (9), Tenor represents loan tenor duration in years. In \nall three equations, I refers to the cumulative world investment volume in billions \nof US dollars, and b1\u20133 signifies the experience parameter for each variable. In each \nof equations (7)\u2013(9), I0 denotes the first investment, and It represents cumulative \ninvestment at time t.\nWe define an individual experience rate, ER\u2009=\u200b\u20091\u2009\u2212\u200b\u20092\u2212b, for each variable \nof interest and quantify it by estimating equations (7)\u2013(9) separately for both \ntechnologies i, using an ordinary least squares regression according to equation (10):\n\u03b2\n\u03b2\n\u03b5\n=\n+\n+\nV\nI\nln(D\n)\nln( )\n(10)\nit\ni\ni\nit\nit\n0\n1\nin which t denotes the year, DV denotes the dependent variable (see equations (7)\u2013(9)),  \nand, again, I signifies cumulative world investment in billions of US dollars.\nAs mentioned previously, a potential caveat concerning the data is the \nheterogeneity of investors. Thus, we apply investor fixed effects in a robustness \ncheck, which does not change the results (see Supplementary Tables 5 and 6). The \nchoice of the independent variable in the specification of the experience rate also \nis subject to some debate in extant literature. Depending on the technology and \napplication, the relevant geographical scope to accumulate experience changes61, \nwhich affects the empirical identification of experience rates58. While the evidence \nindicates global experience effects for RETs because innovation benefits cannot \nbe kept locally34, this argument should hold even more for the finance industry\u2014\nespecially as large investors usually are active internationally. Our choice of \ncumulative global investment is driven by exploratory investor interviews, which \npoint out that the financing of large project finance deals is international and \nincreasingly global, so global investment figures appear to be the most relevant. \nHowever, because our investor sample is Europe-based (see Supplementary Table \n3), we test for a European specification of the experience effect by using cumulative \nEuropean investment. We do so by using capacity data for Europe from IRENA46 \nand investment cost data for Germany from the Bloomberg New Energy Finance \ndatabase50 (three-year moving average). The results remain very similar for solar \nPV, but estimates for the wind onshore experience rate become larger (see columns \n3 and 4 in Supplementary Tables 5 and 6). Finally, we conduct a robustness check \nwith alternative investment data sources, using global data on investment  \ncost per megawatt50 (three-year moving average) and IRENA data on global \ncapacity additions46. The results do not change (see columns 5 and 6 in \nSupplementary Tables 5 and 6). We always use robust standard errors to allow  \nfor heteroscedastic residuals (for example, decreasing variance of the error \nterm with decreasing debt margins because the market is becoming more \ncompetitive). Along most specifications, the results remain very similar. In cases \nwhen they change, we report a conservative experience effect by using global \ncumulative investment (that is, typically equal or close to the lowest value across \nspecifications). We report the range of the estimated experience rates across all \nspecifications in Supplementary Table 7.\nImpact on LCOE. In the fourth step, we calibrate an LCOE model according to \nequation (11) to quantify the effect of the observed changes in financing costs on \nlifetime RET generation costs. We calculate the LCOE for both technologies i (solar \nPV and wind onshore) and the two points in time, t (t\u2009=\u200b\u20091 in 2000\u20132005; t\u2009=\u200b\u20092 in \n2017), as displayed in Fig. 6.\n=\n+ \u2211\n\u2211\n\u03c4\n\u03c4\n\u03c4\n\u03c4\n=\n=\n+\n=\n=\n+\n\u03c4\n\u03c4\n\u03c4\n\u03c4\nC\nLCOE\n(11)\nit\nit\nC\n1\n20\n(1\nCoC )\n1\n20\nFLH\n(1\nCoC )\nit\nit\nit\nit\nCit denotes the initial investment cost (CAPEX) per megawatt at \u03c4\u2009=\u200b\u20090, Cit\u03c4 \nrepresents the operation and maintenance costs (OPEX) per megawatt per year \nfrom \u03c4\u2009=\u200b\u20091 to \u03c4\u2009=\u200b\u200920 (constant) and FLHit\u03c4 signifies the full load hours of the asset \nper year from \u03c4\u2009=\u200b\u20091 to \u03c4\u2009=\u200b\u200920 (constant). Our discount rate CoCit is the technology- \nand time-specific CoC.\nOn the OPEX, we assume 2% annual inflation. We parametrize the LCOE \nmodel by using real data for full load hours, investment cost (US$\u2009MW\u22121) and \noperation and maintenance cost (US$\u2009MW\u22121\u2009yr\u22121) in Germany, and the CoC from \nour project database (see Supplementary Table 4).\nFor both points in time t, we estimate a baseline with 0% CoC. We separate this \nbaseline into CAPEX, represented by the first term of equation (12) and an OPEX \ncomponent represented by the second term.\n=\n\u2211\n+\n\u2211\n\u2211\n\u03c4\n\u03c4\n\u03c4\n\u03c4\n\u03c4\n\u03c4\n\u03c4\n\u03c4\n\u03c4\n=\n=\n=\n=\n=\n=\n=\nC\nC\nLCOE\nFLH\nFLH\n(12)\nit\nit\nit\nit\nit\n,CoC 0\n1\n20\n1\n20\n1\n20\nWe then estimate the same model with the observed CoC from our data rit and \ndefine the change to the baseline as the financing expenditures \u03b4it of the LCOE (see \nequation (13)). Note that rit depends on the project leverage and tax rate according \nto equation (2).\n\u03b4 =\n\u2212\n=\n=\nLCOE\nLCOE\n(13)\nit\nit\nr\nit\n,CoC\n,CoC 0\ni\ni\nAs a result, we obtain three LCOE components (CAPEX, OPEX and financing \nexpenditures) for both technologies at both points in time, which allows us to \ndisplay the changes in each component over time. We define the change in the \nfinancing expenditures, \u03b4it, as \u2206\u200bi following equation (14). Note that in Fig. 6, \u2206\u200bi is \ndenoted \u2018change in financing cost\u2019.\n\u0394\n\u03b4\n\u03b4\n=\n\u2212\n=\n=\n(14)\ni\ni t\ni t\n,\n1\n,\n2\nWe disentangle three effects that contribute to the change in financing cost, namely \nexperience effect \u0394i\nEXP, general interest rate effect \u0394i\nINT and the effect resulting \nfrom lower CAPEX to be financed \u0394i\nCAPEX. The sum of the three effects equals the \ntotal change in financing cost by definition as shown in equation (15).\n\u0394\n\u0394\n\u0394\n\u0394\n=\n+\n+\n(15)\ni\ni\ni\ni\nEXP\nINT\nCAPEX\nWe start with the last term and define the effect resulting from lower CAPEX as \nthe hypothetical LCOE change with constant CoC (part 1 of equation 16) minus \nthe \u2018pure\u2019 CAPEX and OPEX changes (identical to the LCOE at CoC\u2009=\u20090). In doing \nso, we define a counterfactual scenario of identical technological change (that \nis, lower capital expenditure), absent changes in financing conditions. Given the \nmutually reinforcing mechanism of financing conditions and technological change \n(for example, it is not clear that the capital expenditure would have decreased, \nabsent improvements in financing conditions), this approach might overestimate \nthe part of change attributed to \u0394i\nCAPEX. As a consequence, equation (15) provides \nconservative estimates of the other two effects:\n\u0394\n=\n\u2212\n\u2212\n\u2212\n=\n=\n=\n=\n=\n=\n=\n=\nLCOE\nLCOE\n(LCOE\nLCOE\n)\n(16)\ni\ni t\nt\ni t\nt\ni t\ni t\nCAPEX\n,\n1,CoC[\n1]\n,\n2,CoC[\n1]\n,\n1,CoC 0\n,\n2,CoC 0\nTo separate the remaining part of the change in financing cost into experience \neffect and general interest rate effect, we use the share of the debt margin of the \ntotal change in cost of debt (\u03c6). In equation (17), d denotes the difference between \nthe value in 2017 and the value in 2000\u20132005, and GenIntRate represents the \nNature Energy | VOL 3 | DECEMBER 2018 | 1084\u20131092 | www.nature.com/natureenergy\n1090\n\nArticles\nNATuRE EnERGy\ngeneral interest rate. Note that this is computing the share of the changes displayed \nin Fig. 5.\n\u03c6\n=\n+\nd(DebtMargin )\nd(DebtMargin\nGenIntRate)\n(17)\ni\ni\ni\nDEBT\nWe assume that a similar relation holds for the equity side and stipulate \n\u03c6\n\u03c6\n\u03c6\n=\n=\ni\ni\ni\nDEBT\nEQUITY\n. Combining equations (15), (16) and (17), we now can \nidentify the experience effect \u0394i\nEXP and the general interest rate effect \u0394i\nINT, which \nare shown in equations (18) and (19).\n\u0394\n\u03c6 \u0394\n\u0394\n=\n\u2212\n(\n)\n(18)\ni\ni\ni\ni\nEXP\nCAPEX\n\u0394\n\u03c6\n\u0394\n\u0394\n=\n\u2212\n\u2212\n(1\n) (\n)\n(19)\ni\ni\ni\ni\nINT\nCAPEX\nEthics statement. No ethics approval needed. The methodology used in this paper \ndoes not require institutional ethical approval according to the guidelines set out \nby ETH Zurich. Informed consent was obtained from all the interviewees.\nNo signature requirement for informed consent. In advance of participating in the \ninterview, respondents were provided with an information sheet describing the \ntype of questions they would be asked. The information sheet also emphasized \nthe anonymity of data and their right to withdraw from the study. Choosing to \nparticipate in the interview beyond that point was interpreted as informed consent.\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nThe data displayed in Figs. 1, 2 and 4 and used for calculations in Figs. 5 and 6 are \navailable upon reasonable request to T.S.S.\nReceived: 4 April 2018; Accepted: 1 October 2018;  \nPublished online: 5 November 2018\nReferences\n\t1.\t IPCC: Summary for Policymakers. In Climate Change 2014: Mitigation of \nClimate Change (eds Edenhofer, O. et al.) (Cambridge University Press, 2014).\n\t2.\t Climate Finance in 2013-14 and the USD 100 Billion Goal (OECD &  \nCPI, 2015).\n\t3.\t Polzin, F. Mobilizing private finance for low-carbon innovation \u2013 A \nsystematic review of barriers and solutions. Renew. Sustain. Energy Rev. 77, \n525\u2013535 (2017).\n\t4.\t Perspectives for the Energy Transition \u2013 Investment Needs for a Low-Carbon \nEnergy System (OECD/IEA & IRENA, 2017).\n\t5.\t Williams, J. H. et al. 2050: the pivotal role of electricity. Science 335,  \n53\u201360 (2012).\n\t6.\t Chu, S. & Majumdar, A. Opportunities and challenges for a sustainable \nenergy future. Nature 488, 294\u2013303 (2012).\n\t7.\t Schmidt, T. S. Low-carbon investment risks and de-risking. Nat. Clim. \nChange 4, 237\u2013239 (2014).\n\t8.\t Hirth, L. & Steckel, J. C. The role of capital costs in decarbonizing the \nelectricity sector. Environ. Res. Lett. 11, 114010 (2016).\n\t9.\t 2017 Annual Technology Baseline (National Renewable Energy  \nLaboratory, 2017).\n\t10.\tOndraczek, J., Komendantova, N. & Patt, A. WACC the dog: The effect of \nfinancing costs on the levelized cost of solar PV power. Renew. Energy 75, \n888\u2013898 (2015).\n\t11.\tIyer, G. C. et al. Improved representation of investment decisions in \nassessments of CO2 mitigation. Nat. Clim. Change 5, 436\u2013440 (2015).\n\t12.\tCreutzig, F. et al. The underestimated potential of solar energy to mitigate \nclimate change. Nat. Energy 2, 17140 (2017).\n\t13.\tShrimali, G., Nelson, D., Goel, S., Konda, C. & Kumar, R. Renewable \ndeployment in India: Financing costs and implications for policy. Energy \nPolicy 62, 28\u201343 (2013).\n\t14.\tTrancik, J. et. al. Technology Improvement and Emissions Reductions as \nMutually Reinforcing Efforts: Observations from the Global Development of \nSolar and Wind Energy (MIT, 2015).\n\t15.\tNemet, G. F. Beyond the learning curve: factors influencing cost reductions \nin photovoltaics. Energy Policy 34, 3218\u20133232 (2006).\n\t16.\tBolinger, M. & Wiser, R. Understanding wind turbine price trends in the US \nover the past decade. Energy Policy 42, 628\u2013641 (2012).\n\t17.\tHall, S., Foxon, T. J. & Bolton, R. Investing in low-carbon transitions: energy \nfinance as an adaptive market. Clim. Policy 17, 280\u2013298 (2017).\n\t18.\tBrunnschweiler, C. N. Finance for renewable energy: An empirical analysis of \ndeveloping and transition economies. Environ. Dev. Econ. 15, 241\u2013274 (2010).\n\t19.\tBest, R. Switching towards coal or renewable energy? The effects of financial \ncapital on energy transitions. Energy Econ. 63, 75\u201383 (2017).\n\t20.\tBest, R. & Burke, P. J. Adoption of solar and wind energy: The roles of carbon \npricing and aggregate policy support. Energy Policy 118, 404\u2013417 (2018).\n\t21.\tThe Impact of Risks in Renewable Energy Investments and the Role of Smart \nPolicies (Ecofys, 2016).\n\t22.\tMapping the Cost of Capital for Wind and Solar Energy in South Eastern \nEuropean Member States (Ecofys, 2017).\n\t23.\tJacobsson, S. & Lauber, V. The politics and policy of energy system \ntransformation - explaining the German diffusion of renewable energy \ntechnology. Energy Policy 34, 256\u2013276 (2006).\n\t24.\tSteffen, B. The importance of project finance for renewable energy projects. \nEnergy Econ. 69, 280\u2013294 (2018).\n\t25.\tGatti, S. Project Finance in Theory and Practice: Designing, Structuring, and \nFinancing Private and Public Projects (Academic Press, Waltham, 2013).\n\t26.\tDamodaran, A. Equity risk premiums (ERP): determinants, estimation and \nimplications - The 2016 Edition. SSRN Electron. J. https://doi.org/10.2139/\nssrn.2742186 (2016).\n\t27.\tEisenhardt, K. M. & Graebner, M. E. Theory building from cases: \nopportunities and challenges. Acad. Manag. J. 50, 25\u201332 (2007).\n\t28.\tJim\u00e9nez, G., Ongena, S., Peydr\u00f3, J. L. & Saurina, J. Credit supply and \nmonetary policy: Identifying the bank balance-sheet channel with loan \napplications. Am. Econ. Rev. 102, 2301\u20132326 (2012).\n\t29.\tBorio, C. & Zhu, H. Capital regulation, risk-taking and monetary policy:  \nA missing link in the transmission mechanism? J. Financ. Stab. 8,  \n236\u2013251 (2012).\n\t30.\tAltunbas, Y., Gambacorta, L. & Marques-Ibanez, D. Does monetary policy \naffect bank risk? Int. J. Cent. Bank 10, 95\u2013135 (2014).\n\t31.\tHuenteler, J., Schmidt, T. S., Ossenbrink, J. & Hoffmann, V. H. Technology \nlife-cycles in the energy sector - Technological characteristics and  \nthe role of deployment for innovation. Technol. Forecast. Soc. Change 104, \n102\u2013121 (2016).\n\t32.\tPahle, M. & Schweizerhof, H. Time for tough love: towards gradual risk \ntransfer to renewables in Germany. Econ. Energy Environ. Policy 5,  \n117\u2013134 (2016).\n\t33.\tMay, N. G. & Neuhoff, K. Financing Power: Impacts of Energy Policies in \nChanging Regulatory Environments DIW Berlin Discuss. Pap. No. 1684 \n(German Institute for Economic Research, Berlin, Germany, 2017).\n\t34.\tNeuhoff, K. Large-scale deployment of renewables for electricity generation. \nOxford Rev. Econ. Policy 21, 88\u2013110 (2005).\n\t35.\tNelson, D. in Renewable Energy Finance: Powering the Future (ed. Donovan, \nC. W.) 273\u2013305 (Imperial College Press, London, 2015).\n\t36.\tYelle, L. E. The learning curve: historical review and comprehensive survey. \nDecis. Sci. 10, 302\u2013328 (1979).\n\t37.\tQiu, Y. & Anadon, L. D. The price of wind power in China during its \nexpansion: Technology adoption, learning-by-doing, economies of scale, and \nmanufacturing localization. Energy Econ. 34, 772\u2013785 (2012).\n\t38.\tRubin, E. S., Azevedo, I. M. L., Jaramillo, P. & Yeh, S. A review of learning \nrates for electricity supply technologies. Energy Policy 86, 198\u2013218 (2015).\n\t39.\tWright, T. P. Factors affecting the cost of airplanes. J. Aeronaut. Sci. 3, \n122\u2013128 (1936).\n\t40.\tElton, E. J., Gruber, M. J., Agrawal, D. & Mann, C. Explaining the rate spread \non corporate bonds. J. Finance 56, 247\u2013277 (2001).\n\t41.\tRenewable Power Generation Costs in 2017 (IRENA, 2018).\n\t42.\tLehmann, P. & S\u00f6derholm, P. Can technology-specific deployment policies be \ncost-effective? The case of renewable energy support schemes. Environ. \nResour. Econ. 71, 475\u2013505 (2018).\n\t43.\tGeddes, A., Schmidt, T. S. & Steffen, B. The multiple roles of state investment \nbanks in low-carbon energy finance: An analysis of Australia, the UK and \nGermany. Energy Policy 115, 158\u2013170 (2018).\n\t44.\tTietjen, O., Pahle, M. & Fuss, S. Investment risks in power generation:  \nA comparison of fossil fuel and renewable energy dominated markets. Energy \nEcon. 58, 174\u2013185 (2016).\n\t45.\tCampiglio, E. et al. Climate change challenges for central banks and financial \nregulators. Nat. Clim. Change 8, 462\u2013468 (2018).\n\t46.\tRenewable Capacity Statistics 2017 (IRENA, 2017).\n\t47.\tDefinition und Marktanalyse von B\u00fcrgerenergie in Deutschland (Trend \nResearch, 2013).\n\t48.\tChatham House Rule (Chatham House, 2002).\n\t49.\tHoppmann, J., Peters, M., Schneider, M. & Hoffmann, V. H. The two  \nfaces of market support\u2014How deployment policies affect technological \nexploration and exploitation in the solar photovoltaic industry. Res. Policy 42, \n989\u20131003 (2013).\n\t50.\tAsset Database (BNEF, 2017).\n\t51.\tLong-Term Interest Rate Statistics for EU Member States (European Central \nBank, 2018).\n\t52.\tGlobal Trends in Renewable Energy Investment 2017 (Frankfurt School-UNEP \n& BNEF, 2017).\n\t53.\tEisenhardt, K. M. Building theories from case study research. Acad. Manag. \nRev. 14, 532\u2013550 (1989).\nNature Energy | VOL 3 | DECEMBER 2018 | 1084\u20131092 | www.nature.com/natureenergy\n1091\n\nArticles\nNATuRE EnERGy\nexpressed and arguments employed herein do not necessarily reflect the official views \nof the Swiss Government. This work was conducted as part of the European Union\u2019s \nHorizon 2020 research and innovation programme project INNOPATHS under grant \nagreement no. 730403.\nAuthor contributions\nT.S.S., B.S. and F.E. developed the research idea. F.E., B.S. and T.S.S. conducted the \ninvestor interviews, collected, analysed and interpreted the data, and wrote the \nmanuscript. T.S.S. secured project funding.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-018-0277-y.\nReprints and permissions information is available at www.nature.com/reprints.\nCorrespondence and requests for materials should be addressed to F.E. or B.S.  \nor T.S.S.\nPublisher\u2019s note: Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2018\n\t54.\tStatutory Corporate Income Tax Rate (OECD, 2018).\n\t55.\tWalker, D. & Myrick, F. Grounded theory: An exploration of process and \nprocedure. Qual. Health Res. 16, 547\u2013559 (2006).\n\t56.\tGlaser, B. G. Emergence vs Forcing: Basics of Grounded Theory Analysis \n(Sociology Press, Mill Valley, CA, USA, 1992).\n\t57.\tNordhaus, W. D. The perils of the learning model for modeling endogenous \ntechnological change. Energy J. 35, 1\u201314 (2014).\n\t58.\tLindman, \u00c5. & S\u00f6derholm, P. Wind power learning rates: A conceptual \nreview and meta-analysis. Energy Econ. 34, 754\u2013761 (2012).\n\t59.\tMiles, I. Patterns of innovation in service industries. IBM Syst. J. 47,  \n115\u2013128 (2008).\n\t60.\tElche-Hotelano, D. Sources of knowledge, investments and appropriability as \ndeterminants of innovation: An empirical study in service firms. Innov. \nManag. Policy Pract. 13, 224\u2013239 (2011).\n\t61.\tHuenteler, J., Niebuhr, C. & Schmidt, T. S. The effect of local and global \nlearning on the cost of renewable energy in developing countries. J. Clean. \nProd. 128, 6\u201321 (2016).\nAcknowledgements\nThe authors thank M. J\u00e4ger, M. Pahle, F. Polzin, L. Reile and O. Tietjen from the \nINNOPATHS project, participants of the 2017 oikos Finance Academy at the University \nof Zurich, participants of the 41st IAEE International Conference in Groningen (2018) \nand members of ETH Zurich\u2019s Energy Politics Group for helpful comments on earlier \ndrafts of the paper. This work was supported by the Swiss State Secretariat for Education, \nResearch and Innovation (SERI) under contract number 16.0222. The opinions \nNature Energy | VOL 3 | DECEMBER 2018 | 1084\u20131092 | www.nature.com/natureenergy\n1092\n\n1\nnature research  |  reporting summary\nApril 2018\nCorresponding author(s):\nFlorian Egli, Bjarne Steffen, Tobias S. Schmidt\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistical parameters\nWhen statistical analyses are reported, confirm that the following items are present in the relevant location (e.g. figure legend, table legend, main \ntext, or Methods section).\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nAn indication of whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistics including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) AND \nvariation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nClearly defined error bars \nState explicitly what error bars represent (e.g. SD, SE, CI)\nOur web collection on statistics for biologists may be useful.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nNo software was used for data collection\nData analysis\nStandard statistical software suite STATA (version StataSE 14) was used\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers \nupon request. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nFigures 1, 4 and 5 present new raw data, which is available from the corresponding author on reasonable request.\n\n2\nnature research  |  reporting summary\nApril 2018\nField-specific reporting\nPlease select the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/authors/policies/ReportingSummary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nMixed-method study, consisting of (semi-)structured interviews, descriptive statistics and regression analysis\nResearch sample\nThe sample consists of 41 renewable energy investment professionals with experience in the German market. The resulting sample is \nwell balanced  among  different  kinds  of  financial  actors  and includes  17  debt  providers  (13 commercial  banks  and  4  investment  \nbanks),  16  equity  providers,  7  public  actors  (4  public utilities and 3 public investment banks) and 1 former researcher. The sample \nincludes 20 investors based in Germany, 10 in Switzerland, 4 in the UK, 3 in the Netherlands, 1 in Italy, 1 in France, 1 in Norway and 1 in \nLuxemburg. 38 of the respondents were male, 3 were female. The sampled financial \nactors were lead arrangers in 81% of solar PV capacity additions and 85% of the solar PV \ninvestment sum, and in 49% of onshore wind capacity additions and 80% of the onshore wind \ninvestment sum, between 2000 and 2017.. The sample therefore enables to elicit representative information on German solar PV and \nwind onshore projects over our study period.\nSampling strategy\nWe use theoretical sampling to include the most revelatory investors and balance our sample to represent various perspectives from the \nfinancing industry. The sampling took place in three stages. First, we searched for publicly available addresses of senior investment \nmanagers working in large debt and equity investment firms, using the Bloomberg New Energy Finance (BNEF) database. Second, we \nused the contact network of a private renewable energy financing industry partner in the INNOPATHS research consortium, Allianz \nClimate Solutions (ACS), to reach out to relevant market actors. Third, we employed snowball sampling by asking exploratory \ninterviewees to refer us to relevant actors and teams and continuing to ask for references at the end of each semi-structured interview. \nAt the end of each interview, we asked the interviewee whether crucial points were missing. This feedback was included iteratively in the \nfirst few interviews. Following Eisenhardt\u2019s approach, we continued holding interviews until no additional insights were observed.\nData collection\nThe interviews were conducted in person or over the phone by one to three researchers, who took individual notes. All interviews were \nrecorded and transcribed verbatim. If more than one researcher conducted the interview (N = 15), one of them summarised it using the \nrecording, the transcript and the notes. If only one researcher conducted the interview (N = 26), the resulting summary was cross-\nchecked by another researcher. This procedure ensures accurate and consistent recording, expands the scope of the insights and \nenhances confidence in the findings.\nTiming\nWe conducted  all  interviews  between  September  2017  and  January  2018.\nData exclusions\nNo data was excluded from the study.\nNon-participation\nOf a total of 55 requests, 44 replied positively (3 transferred the request internally, but no interview could be arranged), 10 did not \nrespond in spite of at least one reminder/follow-up and 1 declined participation stating no interest in participating in research projects.\nRandomization\nTheoretical sampling was chosen instead of randomization to include the most revelatory cases.\nReporting for specific materials, systems and methods\nMaterials & experimental systems\nn/a Involved in the study\nUnique biological materials\nAntibodies\nEukaryotic cell lines\nPalaeontology\nAnimals and other organisms\nHuman research participants\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\n\n3\nnature research  |  reporting summary\nApril 2018\nHuman research participants\nPolicy information about studies involving human research participants\nPopulation characteristics\nSee above.\nRecruitment\nSee above.\n\n\n    Scientific Research Findings:", "answer": "We find that financing conditions improved greatly for solar photovoltaics (PV) and onshore wind energy in Germany between the introduction of the feed-in tariff in 2000 and 2017. During this period, the cost of capital decreased from 5.1% to 1.6% for solar PV and from 4.5% to 1.9% for onshore wind. These reductions stem from two effects. First, a strong reduction in the general interest rate level, following the financial crisis. Second, a reduction of debt margins, due to increased investment experience with renewable energy technologies.", "id": 32}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-018-0282-1\n1ETH Zurich, Zurich, Switzerland. 2University of Bonn, Bonn, Germany. 3University of Bamberg, Bamberg, Germany. 4University of St. Gallen, St. Gallen, \nSwitzerland. *e-mail: vtiefenbeck@ethz.ch\nI\nndividuals\u2019 choices and behaviour are a key lever influencing \nenergy consumption, along with the technical energy efficiency \nof the products and infrastructure used1. To tackle environmen-\ntal challenges, it is important to put people at the centre of energy \nresearch, and to empirically validate how to promote sustainable \ndecision-making among individual consumers. Energy consump-\ntion is a low-involvement topic for most people; many consumers \nare unaware of the energy efficiency of their homes and devices2, or \nunderestimate the long-term benefits of potential investments into \nenergy efficiency upgrades3.\nAs digitization advances, it becomes increasingly feasible to \nmonitor the energy consumption of households, specific appliances \nor activities in real time4. As a result, digitally enabled behavioural \ninterventions can be deployed at population scale and become more \npowerful through personalization and context specificity. Beyond \nthat, it becomes increasingly possible to systematically evaluate the \nimpact of behavioural interventions with large and diverse samples \nof participants. The availability of high-resolution consumption \ndata enables more and more personalized and flexible interven-\ntions4. These developments may open up new avenues towards \nmore powerful digital strategies for behaviour change.\nYet, while early feedback intervention studies in which people \nwere provided with information about their energy consumption \nreported large savings effects of 5\u201315%5,6 with small convenience \nsamples, spurring the large-scale roll-out of smart meters in many \ncountries, those savings have not materialized in larger field tri-\nals7\u20139. The most widespread form of feedback intervention is \u2018home \nenergy reports\u2019: periodic mailings that compare the electricity use \nof individual households with similar homes in the neighbour-\nhood, thus tapping into social norms3. Deployed at population level \n(households can opt out, but few do), those programmes typically \nyield electricity savings of 2%3,10. Other programmes use digital \ntechnologies, delivering feedback on electricity use via web portals \nor in-home displays; studies with large opt-in samples report elec-\ntricity savings in the range of 1\u20135%8,11\u201313\u2014far less than the savings \nreported by earlier studies with smaller samples and a higher degree \nof involvement from study administrators6,14.\nEarly studies were subject to several methodological issues \nthat compromised the internal and external validity of the results, \noverestimating the savings potential of these feedback interven-\ntions7,8,13. For instance, a meta-analysis of 156 field trials on energy \nconservation found substantially smaller savings effects of 1.99% \nfor high-quality studies with adequate controls, compared to stud-\nies without such controls (9.57%)13. Likewise, a meta-analysis of 33 \nfield trials on in-home displays found weighted mean conservation \neffects of 2.61% for high-quality (\u2018class A\u2019) studies using representa-\ntive sampling techniques, compared to 8.21% for \u2018class C\u2019 studies \ncharacterized by small samples of volunteers and a high degree of \ninvolvement from study administrators8.\nAlthough randomized controlled trials eliminate most threats to \nthe internal validity of studies15\u201317, the external validity of the results \nmay still be compromised if the people who choose to participate in a \nstudy differ from the study\u2019s intended population7. The vast majority \nof feedback programmes on energy consumption use opt-in recruit-\nment strategies, where participants actively register to take part in \nthose programmes7,9. There is increasing evidence that individuals \nwho sign up for energy efficiency studies or demand-side manage-\nment programmes are indeed different from the general population: \nparticipation rates are higher among households with high levels of \neducation and income18,19, among more altruistic and environmen-\ntally concerned individuals18 and among those with a higher interest \nand expertise in energy topics than the general population9,20. Most \nbehavioural programmes do not even provide information about \nthe number of households initially contacted, and those that do, \nreport participation rates in the range of 4\u20138%6,20\u201323. These numbers \nhave raised concerns that the results of opt-in studies may be largely \nbiased by an already-motivated subgroup of the general population \n(\u2018energy enthusiasts\u2019 or \u2018positive greens\u2019), who represent only a small \nfraction of the population9,24. The response of these volunteers to \nthe treatments may not be very indicative for the response of the \nReal-time feedback promotes energy conservation \nin the absence of volunteer selection bias and \nmonetary incentives\nVerena\u00a0Tiefenbeck\u200a \u200a1,2*, Anselma\u00a0W\u00f6rner1, Samuel\u00a0Sch\u00f6b3, Elgar\u00a0Fleisch1,4 and Thorsten\u00a0Staake3\nFeedback interventions have proved to be effective at promoting energy conservation behaviour, and digital technologies have \nthe potential to make interventions more powerful and scalable. In particular, real-time feedback on a specific, energy-inten-\nsive activity may induce considerable behaviour change and savings. Yet the majority of feedback studies that report large \neffects are conducted with opt-in samples of individuals who volunteer to participate. Here we show that real-time feedback \non resource consumption during showering induces substantial energy conservation in an uninformed sample of guests at 6 \nhotels (265 rooms, N\u2009=\u200919,596 observations). The treatment effects are large (11.4% reduction in energy use), indicating that \nthe real-time feedback induced substantial energy conservation among participants who did not opt in, and in a context where \nparticipants were not financially responsible for energy costs. We thus provide empirical evidence for real-time feedback as a \nscalable and cost-efficient policy instrument for fostering resource conservation among the broader public.\nCorrected: Publisher Correction; Publisher Correction\nNature Energy | VOL 4 | JANUARY 2019 | 35\u201341 | www.nature.com/natureenergy\n35\n\nArticles\nNature Energy\ngeneral population (volunteer selection bias). On the one hand, it is \nconceivable that those individuals are already more aware of effec-\ntive energy conservation measures and have already taken action \nbefore the intervention, making it more difficult for them to realize \nadditional savings in those studies25. On the other hand, it is likely \nthat they are particularly open and receptive to these interventions, \nthus inflating estimates of intervention effectiveness26.\nOne commonality that the majority of larger energy feedback tri-\nals share is that they provide aggregate consumption information \nat the household level. This makes it difficult for the individual to \nestablish a link between the current action and its impact on energy \nconsumption27. The results of a recent randomized controlled field \ntrial25 suggest that real-time feedback on a specific, energy-inten-\nsive activity may induce much larger savings. In a two-month study \nwith an opt-in sample of 697 Swiss households, the treatment group \nreceived real-time feedback on the environmental impact of specific, \nenergy-intensive activity (showering), while they could directly take \naction. The intervention yielded large and stable energy savings of \n22% on the target behaviour over the duration of the study. At the \nhousehold level, this reduction led to much larger conservation \ngains\u2014also in absolute terms\u2014than aggregate feedback on energy \nuse among the same pool of households. From a technology and \ncost perspective, the large-scale roll-out of focused real-time inter-\nventions is increasingly feasible25. Yet, given the decline in the effect \nsize and the resulting wave of disillusionment once smart meter-\ning trials with aggregate feedback moved from small convenience \nsamples to a broader population, the key question is whether the \npromising large savings effects of activity-specific real-time feed-\nback will also materialize among individuals who do not self-select \ninto a research study.\nAnother controversial issue is that the communication strategies \nof most energy-conservation programmes focus on the financial \nbenefits for the consumer as incentives for behaviour change28\u201331. \nFrom a standard economics perspective, this approach makes sense, \nas rational consumers should respond to monetary incentives in \ntheir resource-consumption decisions32\u201334. Consequently, mon-\netary incentives play a key role in demand-side management35,36; \nthey have the potential to break established consumer patterns \nand to initiate the development of new patterns of behaviour by \nmaking an alternative behaviour more attractive37,38. However, in \nmany contexts, financial motives are not a viable strategy to pro-\nmote energy conservation: employees, tenants whose rents include \nutilities or hotel guests who do not pay the marginal cost of their \nenergy consumption. While the provision of large, persuasive \nmonetary benefits does not scale well to the wider population, \nmonetary incentives may also crowd out the intrinsic motivation \nfor pro-social behaviour28,39,40 and generate adverse effects41: as \nindividuals tend to internalize the logic of reward systems easily, \nmonetary incentives can lead to the deterioration of morals and \nreduce intrinsic motivation40,42.\nHere we evaluate whether the large savings effects from digi-\ntal activity-specific feedback25 are also realistic in settings where a \nvolunteer selection bias can be ruled out, and where study subjects \nhave no financial incentives for resource conservation. We provide \nactivity-specific feedback on resource consumption to uninformed \nhotel guests during a habitual resource-intensive activity: shower-\ning. We find that even in this setting, the digital behavioural inter-\nvention creates large and significant conservation effects of 11.4% or \n0.215\u2009kWh per shower. Given that most people take a daily shower, \nscaling up this kind of intervention could produce substantial \nenergy (and water) savings. More importantly, the results suggest \nthat activity-specific real-time feedback\u2014and possibly other digital \ninterventions\u2014has the potential to transform behavioural interven-\ntions into a highly relevant policy instrument for fostering energy \nconservation and behaviour change at the population level3,13,43.\nEffect of feedback among uninformed hotel guests\nWe conducted a natural field experiment in the context of an \nenergy-intensive habitual activity: showering. In a randomized con-\ntrolled trial, guests at six Swiss hotels (see Table 1) encountered a \nsmart shower meter fitted to the shower in the bathroom of their \nhotel room. The devices measured the energy and water consump-\ntion of every shower taken, and displayed feedback on each ongoing \nshower in real time. We equipped a total of 265 rooms with these \ndevices and collected a data set with 19,596 observations (water \nextractions, after pre-processing) from February to April 2016.\nThe shower meters were installed between the shower hose and \nthe shower head, and included a small screen that activated to dis-\nplay feedback as soon as the water was turned on (Fig. 1). In the \ntreatment condition (60% of rooms, randomly assigned), smart \nshower meters displayed real-time feedback on the resource con-\nsumption of the ongoing shower: total water consumption in litres \n(one decimal), total energy use in (kilo)watt hours, a dynamic rating \nof the current energy-efficiency class (A\u2013G) and a four-stage anima-\ntion of a polar bear standing on a gradually melting ice floe with \nstage transitions at predefined energy-use thresholds (see Methods \nfor details). This is the same intervention with the same device  \nand display elements as the treatment group of the opt-in household  \nsample in Tiefenbeck et al.25. The remaining 40% of rooms were ran-\ndomly assigned to the control condition: while guests in the treat-\nment rooms received real-time feedback on their resource use from \nthe beginning of the shower (along with the current water tempera-\nture), control-group devices displayed only water temperature.\nHotel guests exposed to real-time feedback consumed signifi-\ncantly less energy per shower than the control group (Fig. 2b). The \ntreatment effect of our intervention is large and significant: guests \nTable 1 | Overview of the participating hotels\nHotel\nCategory\nNo. of \nparticipating \nrooms\nNo. of \nobservations\nHotel 1\nBusiness, four-star\n96\n7,923\nHotel 2\nBusiness, four-star\n67\n6,123\nHotel 3\nBusiness, four-star\n43\n2,789\nHotel 4\nBusiness, three-star\n11\n1,494\nHotel 5\nTourism, four-star\n42\n814\nHotel 6\nTourism\n10\n453\nTotal\n269\n19,596\na\nb\nc\nFig. 1 | Smart shower meter. a, The smart shower meter for displaying \nreal-time feedback on resource consumption to hotel guests was installed \nbetween the shower head and the shower hose. b, Two snapshots of the \ntreatment group\u2019s display. c, The control group\u2019s display. Credit: Amphiro \nAG (b,c)\nNature Energy | VOL 4 | JANUARY 2019 | 35\u201341 | www.nature.com/natureenergy\n36\n\nArticles\nNature Energy\nin the treatment group used on average 0.215\u2009kWh less energy per  \nshower than the control-group mean of 1.883\u2009kWh (Table 2,  \ncolumn 1). This represents a reduction of 11.4% (t(19,596)\u2009=\u2009\u22124.88,  \nP\u2009<\u20090.001). Controlling for flow rate (column 2), the effect  \nis still highly significant, with a reduction of 0.188\u2009kWh \n(t(8,882)\u2009=\u20093.59, P\u2009<\u20090.001), or 10.0%. To determine whether  \nsubsampling for observations in which flow rate is available biases  \nthis results, we included a third model specification for this sub-\nsample but without controlling for flow rate (column 3). The \ntreatment effect is significant in all three models (see Methods \nfor details on the regression analyses) and large (ranging between \n10.0% and 13.2%).\nThe results illustrate that activity-specific real-time feedback \ncan be an efficient measure to foster energy conservation, not only \namong a volunteer sample, but also among a random, uninformed \nsample of individuals.\nAdditional analyses\nAs an alternative functional form, we estimated a log-linear regres-\nsion model. The results (reported in Table 3) are consistent with the \nresults of the non-transformed version reported above and show a \nstrong and significant treatment effect of the real-time feedback.\nTo further corroborate the reported results, we ran the same \nmodels with varying filter thresholds, reducing the data pre-pro-\ncessing to an absolute minimum, with very similar results: if we \nremove only observations deviating over five standard deviations \nfrom mean energy or water consumption, and mean average tem-\nperature, we obtain a sample of 25,490 out of the initial 25,647 \nobservations. Running model (1) on this sample yields a slightly \nsmaller, but still highly significant treatment effect of \u22120.182\u2009kWh \n(s.e.m. 0.044, P\u2009<\u20090.001).\nTo get an understanding of the effects of the six hotels with their \ndifferent infrastructure and setting, we also computed a fixed-\neffects model with dummy variables for the individual hotels. The \nresults are presented in Table 4 and show that the treatment effect \nis highly significant, albeit slightly smaller than in models (1)\u2013(3). \nOnly in hotel 5, the energy use per shower differs significantly from \nthe other hotels, which may be due to different infrastructure (for \nexample, more low-flow shower heads) or guest characteristics. \nOtherwise, the impact on energy use per shower is very similar \nbetween the different hotels. Regardless of the model specification, \nthe treatment effect is large and significant; thus, non-self-selected \nparticipants also respond to real-time feedback in the complete \nabsence of monetary incentives.\nFurthermore, we conducted a cost\u2013benefit analysis for installing \nthe metering device in the hotels\u2019 showers based on the treatment \neffect estimated in model (1). We assumed a retail price of 40\u2009CHF \nfor the smart shower meter and fuel cost for water heating of 0.128 \nCHF\u2009kWh\u22121 and water cost of 3.8 CHF\u2009m\u22123, as in Tiefenbeck et al.25. \nIf we extrapolate from the treatment effect of 0.21\u2009kWh and 3.56\u2009l \nper shower and assume on average 1.2 showers per day per room, as \nobserved during the period of our study, this results in an amortiza-\ntion time of 2.2\u2009years.\nComparisons with a volunteer-household sample\nIn line with the earlier findings on volunteer selection bias7\u20139, the \ntreatment effect in the previous study with a volunteer-household \nsample18 (0.592\u2009kWh, or 22%) was larger than the effect we observe \nin the hotel setting (0.215\u2009kWh, or 11.4%). However, it is impor-\ntant to note that we do not seek to quantify the self-selection effect. \nOur hotel setting and the study context in Tiefenbeck et al.25 dif-\nfer in multiple aspects other than the two key variables of interest. \nFirst, our study examines the behavioural response to feedback in \nthe short term (one or a few nights spent per guest at the hotel). \nConsequently, any attempt to compare the two studies would need \nto focus on the short-term behaviour of the household volunteer \n0\n1\n2\n3\n4\n5\nMean energy use per shower (kWh)\nControl\nTreatment\nStudy group\nFig. 2 | Effect of consumption feedback. Group-wise distribution of \nenergy use per shower in hotel rooms with energy consumption feedback \n(treatment group) and the control group, shown as boxplots (N\u2009=\u200919,596). \nThe line in the middle of the box represents the median, and the diamond \nrepresents the mean energy use. The box spans the first quartile to the \nthird quartile, and the whiskers extend up to 1.5 times the interquartile \nrange from the top or bottom of the box.\nTable 2 | Main treatment effect\nEnergy use per shower (kWh)\nConsumption feedback (treatment\u2009=\u20091, control\u2009=\u20090)\n\u22120.215*** (0.044)\n\u22120.188*** (0.050)\n\u22120.252*** (0.056)\nFlow rate (mean-centred, min\u22121)\n\u2013\n0.098*** (0.010)\n\u2013\nConstant\n1.884*** (0.032)\n1.881*** (0.036)\n1.902*** (0.039)\nObservations\n19,596\n8,824\n8,824\nR2\n0.008\n0.047\n0.011\nStandard errors are in parentheses, adjusted for clustering at the room level; *** indicates significance at the 0.1% level.\nNature Energy | VOL 4 | JANUARY 2019 | 35\u201341 | www.nature.com/natureenergy\n37\n\nArticles\nNature Energy\nsample. Indeed, an analysis of the first three showers only in the \nvolunteer sample also yielded a smaller treatment effect (17.8% \nor 0.46\u2009kWh) than in the full two-month evaluation. Robustness \nchecks using the first two, four or five (instead of three) showers \nwere conducted, with very similar results. Second, individuals may \nreact differently to feedback in their familiar environment at home \nversus in a hotel room, or guests may perceive the mere presence of \nthe shower meter as a signal that the hotel management cares about \nenvironmental issues and pays attention to how much energy and \nwater their guests are using.\nAnother remarkable difference relates to participants\u2019 energy \nuse per shower in the absence of feedback, measured in the respec-\ntive control groups. Hotel guests in the control group consumed \n28% less energy per shower, namely M\u2009=\u20091.88\u2009kWh, s.d.\u2009=\u20091.25\u2009kWh, \nthan the control group in the household setting, with a mean of \nM\u2009=\u20092.62\u2009kWh, s.d.\u2009=\u20091.67\u2009kWh (ref.\u200925), t(20,236)\u2009=\u2009\u221211.1, P\u2009<\u20090.001. \nThis difference is noteworthy for two reasons. First, according to \nstandard economic theory, one would expect individuals to take \nlonger showers at a hotel than at home, as they do pay a marginal \ncost for every kilowatt hour of consumed energy. Yet, the results \nsuggest that the hotel guests did not exploit the zero marginal cost \nof consumption. We attribute lower consumption in the control \ngroup largely to differences in the technical infrastructure between \nthe hotels and households: we observed lower flow rates in the hotel \nrooms, probably caused by a higher share of water-saving shower \nheads installed in the hotel rooms. Second, this difference may \nalso partially explain the smaller treatment effect in the hotels. The \nstudy in the household setting had revealed a strong positive inter-\naction between the treatment effect and baseline consumption: in \nthe household study, a 1-kWh increase in baseline consumption led \nto a 0.32\u2009kWh increase in the savings effect25. To put it simply, it \nis far easier to cut a 20-min shower short by a few minutes (and \nkilowatt hours) than to realize substantial reductions on a 1-min \nshower. The control-group mean of energy consumption per \nshower in our hotel sample is 1.88\u2009kWh, compared to 2.62\u2009kWh in \nthe household sample (0.74\u2009kWh difference). Interpreting the con-\ntrol-group mean as a proxy for baseline consumption, an increase \nin baseline consumption by 0.74\u2009kWh would increase the savings \neffect by 0.24\u2009kWh, which almost exactly matches the difference \nin the observed savings effect. Thus, if we control for the lower \nconsumption at the hotels in the absence of feedback, the savings \neffects among the hotel guests and among the volunteer household \nsample are in fact comparably large.\nDiscussion\nIn the case of aggregate feedback, most energy efficiency studies \nyielded much smaller savings effects once those interventions were \nevaluated with large, non-self-selected samples7\u20139. In other words, \nthose interventions resonated much less with broader, non-self-\nselected audiences than with those individuals who had opted to \nparticipate. By contrast, with a highly significant treatment effect of \n11.4% among uninformed hotel guests, this study provides empirical  \nevidence that activity-specific real-time feedback can induce sub-\nstantial behaviour change among a broader population, even in a \nsetting without monetary incentives for resource conservation. \nThus, providing real-time feedback on a specific energy-intensive \nactivity may generate large and persistent savings effects25 not only \namong the small percentage of the population who tend to opt into \nenergy efficiency studies11,20,21,23; the results indicate that the inter-\nvention successfully induces substantial behaviour change and \nresource conservation among broader audiences.\nRegarding the cost-effectiveness of the intervention in the hotel \ncontext studied, on the basis of the savings effects observed, the \ndevice pays itself off in a hotel within 2.2 years on average, which \nis a very low amortization time as compared to other energy effi-\nciency investments2,44. Thus, the results suggest that even in settings \nwhere third parties pay for the marginal cost of resource consump-\ntion, activity-specific feedback can be a cost-effective and scalable \nstrategy to foster energy conservation.\nDespite our best efforts, there are limitations to our study. While \nwe were able to measure effects for 100% of the hotel guests, they \nmay not be representative of the general population. Although we \ndiversified our sample by including different types of hotel, with dif-\nferent comfort categories, room rates and primary target custom-\ners (business versus tourism), and in different locations, additional \nstudies in other settings and other countries would be valuable. \nFurthermore, despite our efforts to limit differences in potential \nHawthorne effects by displaying real-time water temperature on the \ncontrol-group devices, it is conceivable that the treatment with real-\ntime feedback on resource consumption draws more attention to the \nfact that the smart shower meter measures data than does real-time \ninformation on water temperature, conveying a stronger feeling of \nbeing monitored among its users. Moreover, due to the short dura-\ntion of guests\u2019 hotel stays, we are not able to examine effects over \nTable 3 | Main treatment effect with log transformation of the dependent variable\nEnergy use per shower (kWh)\nConsumption feedback (treatment\u2009=\u20091, control\u2009=\u20090)\n\u22120.124*** (0.024)\n\u22120.112*** (0.028)\n\u22120.148*** (0.032)\nFlow rate (mean-centred, l\u2009min\u22121)\n\u2013\n0.056*** (0.006)\n\u2013\nConstant\n0.413*** (0.017)\n0.410*** (0.020)\n0.422*** (0.022)\nObservations\n19,596\n8,824\n8,824\nR2\n0.008\n0.047\n0.011\nStandard errors are in parentheses, adjusted for clustering at the room level; *** indicates significance at the 0.1% level.\nTable 4 | Treatment effect of consumption feedback and fixed \neffects for different hotels\nEnergy use per shower (kWh)\nConsumption feedback (treatment\u2009=\u20091,  \ncontrol\u2009=\u20090)\n\u22120.197*** (0.041)\nHotel 2\n0.037 (0.051)\nHotel 3\n\u22120.072 (0.091)\nHotel 4\n0.117 (0.090)\nHotel 5\n\u22120.361*** (0.058)\nHotel 6\n\u22120.191 (0.127)\nConstant (Hotel 1)\n1.918*** (0.045)\nObservations\n19,596\nR2\n0.021\nStandard errors are in parentheses, adjusted for clustering per room; *** indicates significance at \nthe 0.1% level.\nNature Energy | VOL 4 | JANUARY 2019 | 35\u201341 | www.nature.com/natureenergy\n38\n\nArticles\nNature Energy\ntime. While several studies (lasting between 2 and 16 months25,45) \ndocument the mid-term effect stability of activity-specific real-time \nfeedback with opt-in samples, further research needs to investigate \nwhether the large savings effects also persist over time among non-\nself-selected participants.\nHere, we provide robust empirical evidence that activity-spe-\ncific real-time feedback can induce substantial behaviour change \nand resource conservation\u2014even for a sample of individuals who \nneither volunteered to participate in an environmental study, nor \nreaped financial benefits from energy conservation. Given the \ndebate on volunteer self-selection7 and the dwindling treatment \neffects of other feedback interventions once they are deployed \namong broader samples7\u20139,12, this empirical validation is critical \nto provide solid recommendations for the design of future energy \nconservation programmes46. Information technology increasingly \nmakes it possible to monitor behaviour in real time, to provide \nindividuals with feedback on their ongoing activities and to col-\nlect granular data on the real-world impact of interventions from \nmillions of individuals in the field4 at rapidly declining costs. The \nresults of this study highlight the potential of digital interventions \nto transform behaviour in energy-intensive activities, which can be \nimplemented and monitored at the population level.\nMethods\nExperimental set-up. We conducted a natural field experiment in which we \ntargeted showering as an example of a resource-intensive, low-involvement \nactivity. Participants were not recruited as individuals; instead, we collaborated \nwith six hotels that allowed us to conduct the experiments in their rooms without \ninforming the guests upfront about the experiment. Shower data were collected in \nbatches after several weeks and without a time stamp, which guaranteed complete \nanonymity of the guests\u2019 identity. The study was approved by the Internal Review \nBoard of the University of Bamberg. Similar natural field experiments in hotel \ncontexts have been conducted to investigate the impact of other behavioural \ninterventions such as commitment strategies47 or social comparisons48,49.\nOverall, guests staying in 265 different hotel rooms took part in our \nexperiment. Hotel guests encountered smart shower meters as part of their rooms\u2019 \nbathroom equipment. The guests who stayed in rooms assigned to the treatment \ngroup received real-time feedback on how much energy and water they consumed \nover the course of their shower (details below). This activity-specific consumption \nfeedback was displayed by a shower meter that had previously been used in framed \nfield experiments in private households25. Room assignment was randomized over \nfloor levels and room categories to minimize confounding factors from differences \nin infrastructure (for example, water pressure) and did not change throughout \nthe study. Approximately 40% of the rooms were assigned to the control group. \nThey serve as the reference group to calculate the treatment effect. In those rooms, \nthe same device was installed, but it displayed only water temperature. While the \ntemperature reading does not convey information about the resource use and \nremains relatively static over the course of a shower, it indicates that the device \nmeasures data, thus reducing potential differences between the treatment and \ncontrol group for Hawthorne effects to occur50,51.\nAs the smart shower meters are powered by the water flow via a small internal \ngenerator, the screen displaying the feedback switches on as soon as water flows \nthrough the device and remains active for up to three minutes after the end of a \nshower. Thus, short interruptions to the water flow (for instance while soaping) \nstill result in a single shower being recorded. Before the device switches off, it stores \nthe final data in its internal memory, which was read out at the end of the study.\nThe experiment took place in six different hotels in Switzerland, recruited on \nthe basis of existing contacts. Data were collected between February and April \n2016. Four of the hotels focus on business customers and the other two on private \ntourists. The categorization into business and tourism was defined on the basis of \ninformation provided by the hotels\u2019 management. Of course, we cannot rule out \nthat business hotels also had guests on private holidays, or that the tourism hotels \nhosted some business guests during the course of the study. Depending on their \nsize, the participating hotels allowed us to install shower meters in 10 to 96 of their \nguest rooms, respectively.\nDisplay content in the treatment group. Most feedback devices display a bundle \nof elements rather than a single numeric metric52 to put the measurement data into \ncontext; frequently used elements include historic comparisons, peer comparisons, \nanalogies and energy savings tips. Likewise, the smart shower meters in rooms \nassigned to the treatment condition displayed water consumption in litres (one \ndecimal), energy use in (kilo)watt hours, current water temperature, a dynamic \nrating of the current energy-efficiency class (A\u2013G) and a four-stage animation \nof a polar bear standing on a melting ice floe with stage transitions at predefined \nenergy use thresholds. The energy consumption displayed on the screen represents \nthe lower bound of the energy used (without losses), and is calculated using the \nstandard engineering formula for heat energy (Q\u2009=\u2009m\u2009\u00d7cp\u2009\u00d7\u2009\u2206T, with heat energy \nQ, mass of water m, heat capacity cp, and \u2206T being the difference between the \nmeasured water temperature of the ongoing shower and the average cold-water \ntemperature). In the analysis of energy savings, we take into account the same \naverage heating efficiency and losses as in Tiefenbeck et al.25. The energy-efficiency \nclass displayed was inspired by the (static) energy-efficiency class scale indicated \non household appliances in Europe. The smart shower meter dynamically indicates \nthe energy efficiency of the current shower based on the energy use in the ongoing \nshower, starting in energy efficiency class A and progressing to B, C and so on at \npredefined kilowatt hour thresholds; the thresholds were defined on the basis of \nthe distribution of energy use per shower in a pilot study. The four stages of the \npolar bear animation are tied to the energy-efficiency class, and change with the \ntransitions from B to C, D to E and E to F, respectively. While the polar bear may \nbe an eye-catching and memorable display element, it does not seem to drive the \nsavings effects. A related study specifically examined the effect of variations of the \ndesign choices of the feedback elements; the results indicate that if the polar bear \nanimation makes any difference, it reduces rather than increases the effectiveness \nof the display52.\nData. For each water extraction, the smart shower meter recorded energy and \nwater consumption, average water temperature, interruptions and the duration. In \naddition, in 168 of the rooms, the average flow rate per shower was also measured. \nUsing the data stored on the device, energy consumption can be converted to \nwater consumption and vice versa. Given the high correlation between water and \nenergy consumption per shower (0.989)25, the choice of the unit of analysis does \nnot change the results in any meaningful way. This article focuses on resource \nconsumption in units of energy in kilowatt hours.\nThe raw data set included observations of 25,647 measured showers from 269 \nhotel rooms at 6 different hotels (see Data availability). In a first pre-processing \nstep, the data were cleaned by removing outliers from malfunctioning devices; to \nthis end, observations that deviated by over 3 standard deviations from the mean  \nof the energy consumed or water volume per shower were removed from the  \nsample\u2014that is, only observations in the interval [ \u0304\n\u0304\n\u2212\u00d7 . .\n+ \u00d7 . .\nx\nx\n3\ns d ,\n3\ns d\n] were retained. Furthermore, we removed data points that most likely did not \nrepresent showers\u2014for example water extractions of volumes below 6.5\u2009litres \nand observations deviating over 2 standard deviations from average temperature, \nwhich probably represent cleaning or other procedures. A member of the research \nteam accompanied cleaning personnel at one hotel for several hours to gather \ninformation on cleaning practices to identify water extractions for cleaning. The \nspecific choice of 6.5\u2009litres was based on this assessment; we conducted robustness \nchecks in which we changed this threshold to other values (5\u2009litres or 10\u2009litres), \nwhich generated very similar results.\nAfter this pre-processing step, the final data set included 19,596 showers \nfrom 265 hotel rooms (11,384 observations in the treatment group and 8,218 in \nthe control group). The average flow rate per shower could be measured only for \n168 rooms and 8,824 observations, so only these data points are included in the \nestimation of models (2) and (3) in Table 2.\nSince the study is a natural field experiment with uninformed participants, we \nwere not able to collect socio-demographic data about the guests who stayed in the \nrooms with the smart shower meters during the study.\nData analysis. We used the data points observed in the control group to quantify \nthe energy use per shower in the participating hotels without feedback. To \nestimate the treatment effect of the feedback intervention, we used a simple linear \nregression model and also computed its log-linear transformation:\n\u03b2\n\u03b2\n\u03f5\n=\n+\n+\ny\nx\n(1)\ni\ni\ni\n0\n1\n\u03b2\n\u03b2\n\u03f5\n=\n+\n+\ny\nx\nln\n(2)\ni\ni\ni\n0\n1\nwhere the dependent variable yi is the energy consumption in shower i. The \nvariable xi is binary, indicating treatment (=1) or no treatment (=0), and thus \ncoefficient \u03b21 estimates the treatment effect. The intercept \u03b20 represents the control-\ngroup mean in this model, as xi\u2009=\u20090 for observations in the control group. The \nresults of this analysis are reported in Table 2, column 1, and in Table 3, column 1, \nwith the natural logarithm of the dependent variable.\nFor rooms in which the smart shower meter also measured the flow rate, we \nestimated an additional model that controls for the centred flow rate in litres fi:\n\u03b2\n\u03b2\n\u03b2\n\u03f5\n=\n+\n+\n+\ny\nx\nf\n(3)\ni\ni\ni\ni\n0\n1\n2\n\u03b2\n\u03b2\n\u03b2\n\u03f5\n=\n+\n+\n+\ny\nx\nf\nln\n(4)\ni\ni\ni\ni\n0\n1\n2\nThe results for model (3), which includes the flow rate in the regression, are \nreported in Table 2, column 2, and for model (4) in Table 3, column 2, with log-linear  \nNature Energy | VOL 4 | JANUARY 2019 | 35\u201341 | www.nature.com/natureenergy\n39\n\nArticles\nNature Energy\ntransformation. In both model specifications, standard errors were clustered at \nthe room level to account for infrastructural influences. Two-sided t-tests were \nconducted to test whether the coefficients were significantly different from zero.\nTo get an understanding of the effects of the six hotels with their different \ninfrastructure and setting, we also computed a fixed-effects model with dummy \nvariables for the individual hotels. In this model, the constant represents the \nestimates for the largest hotel (hotel 1) and dummy variables are included for the \nother hotels. Results are reported in Table 4.\n\u03b2\n\u03b2\n\u03b1\n\u03b1\n\u03b1\n\u03b1\n\u03b1\n\u03f5\n=\n+\n+\n+\n+\n+\n+\n+\ny\nx\nh\nh\nh\nh\nh\n(5)\ni\ni\ni\ni\ni\ni\ni\ni\n0\n1\n2 2\n3 3\n4 4\n5 5\n6 6\nCode availability. The Stata code used to generate results reported in this article is \navailable on request from the corresponding author.\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nThe data that support the findings of this study are available at https://doi.\norg/10.6084/m9.figshare.6984323.v1.\nReceived: 27 March 2018; Accepted: 9 October 2018;  \nPublished online: 19 November 2018\nReferences\n\t1.\t Sovacool, B. K. Diversity: energy studies need social science. Nature 511, \n529\u2013530 (2014).\n\t2.\t Allcott, H. & Greenstone, M. Is there an energy efficiency gap? J. Econ. \nPerspect. 26, 3\u201328 (2012).\n\t3.\t Allcott, H. & Mullainathan, S. Behavior and energy policy. Science 327, \n1204\u20131205 (2010).\n\t4.\t Tiefenbeck, V. Bring behaviour into the digital transformation. Nat. Energy 2, \n17085 (2017).\n\t5.\t Darby, S. The Effectiveness of Feedback on Energy Consumption - A Review for \nDEFRA of the Literature on Metering Billing and Direct Displays \n(Environmental Change Institute, University of Oxford, Oxford, 2006).\n\t6.\t Ehrhardt-Martinez, K. et al. Advanced Metering Initiatives and Residential \nFeedback Programs: A Meta-Review for Household Electricity-Saving \nOpportunities (Washington, DC, American Council for an Energy-Efficient \nEconomy, 2010).\n\t7.\t Davis, A. L., Krishnamurti, T., Fischhoff, B. & Bruine de Bruin, W. Setting a \nstandard for electricity pilot studies. Energy Policy 62, 401\u2013409 (2013).\n\t8.\t McKerracher, C. & Torriti, J. Energy consumption feedback in perspective: \nintegrating Australian data to meta-analyses on in-home displays. Energy \nEffic. 6, 387\u2013405 (2013).\n\t9.\t Kelly, J. & Knottenbelt, W. Does disaggregated electricity feedback reduce \ndomestic electricity consumption? A systematic review of the literature. 1605, \n00962v2 (2016).\n\t10.\tAllcott, H. & Mullainathan, S. External Validity and Partner Selection Bias \nWorking Paper No. 18373 (NBER, 2012).\n\t11.\tSchleich, J., Klobasa, M., G\u00f6tz, S. & Brunner, M. Effects of feedback on \nresidential electricity demand: findings from a field trial in Austria. Energy \nPolicy 61, 1097\u20131106 (2013).\n\t12.\tBuchanan, K., Russo, R. & Anderson, B. The question of energy reduction: \nThe problem(s) with feedback. Energy Policy 77, 89\u201396 (2015).\n\t13.\tDelmas, M. A., Fischlein, M. & Asensio, O. I. Information strategies and \nenergy conservation behavior: A meta-analysis of experimental studies from \n1975 to 2012. Energy Policy 61, 729\u2013739 (2013).\n\t14.\tAbrahamse, W., Steg, L., Vlek, C. & Rothengatter, T. A review of intervention \nstudies aimed at household energy conservation. J. Environ. Psychol. 25, \n273\u2013291 (2005).\n\t15.\tHaynes, L., Service, O., Goldacre, B. & Torgerson, D. Test, Learn, Adapt: \nDeveloping Public Policy with Randomised Controlled Trials (London Cabinet \nOffice Behavioural Insights Team, 2012).\n\t16.\tVine, E., Sullivan, M., Lutzenhiser, L., Blumstein, C. & Miller, B. \nExperimentation and the evaluation of energy efficiency programs. Energy \nEffic. 7, 627\u2013640 (2014).\n\t17.\tCampbell, D. T. Reform as experiments. Am. Psychol. 24, 409\u2013429 (1969).\n\t18.\tClark, C. F., Kotchen, M. J. & Moore, M. R. Internal and external influences \non pro-environmental behavior: Participation in a green electricity program. \nJ. Environ. Psychol. 23, 237\u2013246 (2003).\n\t19.\tSulyma, I., Tiedemann, K., Pedersen, M., Rebman, M. & Yu, M. Experimental \nevidence: a residential time of use pilot. In Proc. ACEEE Summer Study \nEnergy Effic. Build. 292\u2013304 (ACEEE, 2008).\n\t20.\tBaladi, S. M., Herriges, J. A. & Sweeney, T. J. Residential response to \nvoluntary time-of-use electricity rates. Resour. Energy Econ. 20,  \n225\u2013244 (1998).\n\t21.\tHerter, K., Wood, V. & Blozis, S. The effects of combining dynamic pricing, \nAC load control, and real-time energy feedback: SMUD\u2019S 2011 Residential \nSummer Solutions Study. Energy Effic. 6, 641\u2013653 (2013).\n\t22.\tSchleich, J., Klobasa, M., G\u00f6lz, S. & Brunner, M. Effects of feedback on \nresidential electricity demand-findings from a field trial in Austria. Energy \nPolicy 61, 1097\u20131106 (2013).\n\t23.\tLossin, F., Kozlovskiy, I., Sodenkamp, M., & Staake, T. Incentives to go green: \nan empirical investigation of monetary and symbolic rewards to motivate \nenergy savings. In Proc. Eur. Conf. Inf. Syst. 1\u201316 (2016).\n\t24.\tA Framework for Pro-Environmental Behaviours (Department for \nEnvironment Food and Rural Affairs, 2008).\n\t25.\tTiefenbeck, V. et al. Overcoming salience bias: how real-time feedback fosters \nresource conservation. Manag. Sci. 64, 983\u20131476 (2016).\n\t26.\tFrederiks, E. R., Stenner, K., Hobman, E. V. & Fischle, M. Evaluating energy \nbehavior change programs using randomized controlled trials: Best practice \nguidelines for policymakers. Energy Res. Soc. Sci. 22, 147\u2013164 (2016).\n\t27.\tFaruqui, A., Sergici, S. & Sharif, A. The impact of informational feedback on \nenergy consumption-A survey of the experimental evidence. Energy 35, \n1598\u20131608 (2010).\n\t28.\tSchwartz, D., Bruine de Bruin, W., Fischhoff, B. & Lave, L. Advertising energy \nsaving programs: The potential environmental cost of emphasizing monetary \nsavings. J. Exp. Psychol. Appl. 21, 158\u2013166 (2015).\n\t29.\tAllcott, H. & Sweeney, R. L. The role of sales agents in information \ndisclosure: evidence from a field experiment. Manage. Sci. 63, 21\u201339 (2017).\n\t30.\tEnergy Efficiency Campaign: \u2018Deutschland Macht\u2019s Effizient\u2019 (German Federal \nMinistry for Economic Affairs and Energy, 2016); http://www.deutschland-\nmachts-effizient.de/\n\t31.\tCustomer Incentives for Energy Efficiency Through Program Offerings (US \nEnvironmental Protection Agency, 2010).\n\t32.\tLandry, C. E., Lange, A., List, J. A., Price, M. K. & Rupp, N. G. Toward an \nunderstanding of the economics of charity: evidence from a field experiment. \nQ. J. Econ. 121, 747\u2013782 (2006).\n\t33.\tKarlan, D. & List, J. A. Does price matter in charitable giving? Evidence from \na large-scale natural field experiment. Am. Econ. Rev. 97, 1774\u20131793 (2007).\n\t34.\tGoette, L., Stutzer, A. & Frey, B. M. Prosocial motivation and blood \ndonations: a survey of the empirical literature. Transfus. Med. Hemotherapy \n37, 149\u2013154 (2010).\n\t35.\tOlmstead, S. M. & Stavins, R. N. Comparing price and nonprice approaches \nto urban water conservation. Water Resour. Res. 45, 1\u201310 (2009).\n\t36.\tBorenstein, S. The long-run efficiency of real time electricity pricing. Energy J. \n26, 93\u2013116 (2005).\n\t37.\tMaki, A., Burns, R. J., Ha, L. & Rothman, A. J. Paying people to protect the \nenvironment: a meta-analysis of financial incentive interventions to promote \nproenvironmental behaviors. J. Environ. Psychol. 47, 242\u2013255 (2016).\n\t38.\tBamberg, S. Is a residential relocation a good opportunity to change people\u2019s \ntravel behavior?: Results from a theory-driven intervention study. Environ. \nBehav. 38, 820\u2013840 (2006).\n\t39.\tFrey, B. S. & Oberholzer-Gee, F. The cost of price incentives: an  \nempirical analysis of motivation crowding-out. Am. Econ. Rev. 87,  \n746\u2013755 (1997).\n\t40.\tSandel, M. J. What Money Can\u2019t Buy: The Moral Limits of Markets \n(Macmillan, London, 2012).\n\t41.\tGneezy, U. & Rustichini, A. A fine is a price. J. Legal Stud. 29, 1\u201317 (2000).\n\t42.\tTh\u00f8gersen, J. Monetary incentives and environmental concern. Effects of a \ndifferentiated garbage fee. J. Consum. Policy 17, 407\u2013442 (1994).\n\t43.\tKarlin, B., Zinger, J. F. & Ford, R. The effects of feedback on energy \nconservation: a meta-analysis. Psychol. Bull. 141, 1205\u20131227 (2015).\n\t44.\tSchopfer, S., Tiefenbeck, V. & Staake, T. Economic assessment of  \nphotovoltaic battery systems based on household load profiles. Appl. Energy \n223, 229\u2013248 (2018).\n\t45.\tTiefenbeck, V., Tasic, V., Sch\u00f6b, S. & Staake, T. Long-lasting effects or \nshort-term spark? On the persistence of behaviour change induced by \nreal-time feedback on resource consumption. In Proc. Eur. Conf. Inf. Syst. \n1\u201317 (2016).\n\t46.\tI\u2019m Not Surprised. Nat. Energy 2, 17101 (2017).\n\t47.\tBaca-Motes, K., Brown, A., Gneezy, A., Keenan, E. A. & Nelson, L. D. \nCommitment and behavior change: evidence from the field. J. Consum. Res. \n39, 1070\u20131084 (2013).\n\t48.\tGoldstein, N. J., Cialdini, R. B. & Griskevicius, V. A room with a viewpoint: \nusing social norms to motivate environmental conservation in hotels.  \nJ. Consum. Res. 35, 472\u2013482 (2008).\n\t49.\tSchultz, P. W., Khazian, A. M. & Zaleski, A. C. Using normative social \ninfluence to promote conservation among hotel guests. Soc. Influ. 3,  \n4\u201323 (2008).\n\t50.\tSchwartz, D. et al. The Hawthorne effect and energy awareness. Proc. Natl \nAcad. Sci. USA 110, 15242\u201315246 (2013).\n\t51.\tTiefenbeck, V. On the magnitude and persistence of the Hawthorne effect \n- evidence from four field studies. In Proc. European Conference on Behaviour \nand Energy Efficiency 1\u20136 (2016).\nNature Energy | VOL 4 | JANUARY 2019 | 35\u201341 | www.nature.com/natureenergy\n40\n\nArticles\nNature Energy\n\t52.\tAbleitner, L., Sch\u00f6b, S., Tiefenbeck, V. & Fridgen, G. Real-world impact of \ninformation systems: the effect of seemingly small design choices. In Proc. \nWork. Inf. Technol. Syst. 1\u201316 (2017).\nAcknowledgements\nWe thank S. H\u00e4cki and T. Bachmann from Swiss Mobiliar insurance for their great \nefforts in reaching out to hotels and managing the on-site study implementation and \ndata collection. We would also like to express our gratitude to the management of the six \nSwiss hotels for the opportunity to run the study. Funding for this work was provided \nby the MTEC foundation of ETH Zurich (data analysis) as well as by Swiss Mobiliar \ninsurance (hardware, deployment).\nAuthor contributions\nV.T. and T.S. designed the study. S.S. wrote the software of the study devices and enabled \nthe technical side of the data collection. V.T. oversaw the study implementation. A.W. \nand V.T. analysed the data. V.T. and A.W. drafted the manuscript; T.S. and E.F. edited the \nmanuscript. V.T., T.S. and E.F. secured funding for the study.\nCompeting interests\nV.T., A.W. and E.F. declare no competing financial interests. T.S. and S.S. are  \nco-founders of and hold shares in Amphiro AG, the SME that manufactures the smart \nshower meters. T.S. and S.S. were not involved in the data analysis, hotel selection or \nroom assignment.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-018-0282-1.\nReprints and permissions information is available at www.nature.com/reprints.\nCorrespondence and requests for materials should be addressed to V.T.\nPublisher\u2019s note: Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2018\nNature Energy | VOL 4 | JANUARY 2019 | 35\u201341 | www.nature.com/natureenergy\n41\n\n1\nnature research  |  reporting summary\nMarch 2018\nCorresponding author(s):\nVerena Tiefenbeck\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistical parameters\nWhen statistical analyses are reported, confirm that the following items are present in the relevant location (e.g. figure legend, table legend, main \ntext, or Methods section).\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nAn indication of whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistics including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) AND \nvariation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nClearly defined error bars \nState explicitly what error bars represent (e.g. SD, SE, CI)\nOur web collection on statistics for biologists may be useful.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nCustom software for the shower meter Amphiro was used to measure resource consumption in the showers in 265 rooms in 6 different \nhotels and for the data read-out\nData analysis\nStata SE 14.1 was used to preprocess the resource consumption data, to conduct regression analyses, and plot the presented graphs\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers \nupon request. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nThe data that support the findings of this study is available on figshare.com: https://doi.org/10.6084/m9.figshare.6984323.v1.\n\n2\nnature research  |  reporting summary\nMarch 2018\nField-specific reporting\nPlease select the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\nFor a reference copy of the document with all sections, see nature.com/authors/policies/ReportingSummary-flat.pdf\nBehavioural & social sciences\nStudy design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nOur study is a randomized controlled trial at six hotels, to examine resource consumption during showering in a natural field experiment \nin a setting of complete absence of monetary incentives among a sample of uninformed participants who did not self-select into the \nstudy. We collected quantitative data from 19,602 showers (energy and water consumption, temperature, flowrate).\nResearch sample\nThe study sample contains data on resource consumption in the shower from the guests of 265 rooms of 6 different tourist and business \nhotels in Switzerland. Due to the the hotels' privacy regulations we were not able to collect any demographical data, but this sampling \nstrategy had the advantage that our sample did not self-select into the study. This point is discussed in the article. \nSampling strategy\nWe ran the study as a natural field experiment, thus guests encountered the measurement and feedback device as part of their \nbathroom equipment. We conducted powertests based on the data of a similar previous study in households (Tiefenbeck et al. (2016), \nManagement Science). \nData collection\nA smart shower meter was used to collect data on resource consumption from every water extraction during the study period. The \ndevice called \"Amphiro\" is depicted in the article. The data is stored on the device and was read out by the researchers at the end of the \nstudy. \nTiming\nData collection took place from February to April 2016.\nData exclusions\nThe data was cleaned by removing outliers from malfunctioning devices; to this end, observations that deviated by over 3 standard \ndeviations from the mean of the energy consumed or water volume per shower were removed.  Furthermore, we excluded data points \nthat most likely did not represent showers \u2013 e.g. water extractions of volumes below 6.5 litres and observations deviating over 2 \nstandard deviations from average temperature, which probably represent cleaning or other procedures .\nNon-participation\nWe are not aware of any hotel guests that asked to have the device removed, so to the best of our knowledge, data on all guests of the \nparticipating rooms during the period of the study are present in our raw dataset.\nRandomization\nAt each hotel, we randomly allocated rooms to the two experimental conditions; the randomization was done by floor level to account \nfor heterogeneity in the flow rates in the different floors of the buildings. \n\n\n    Scientific Research Findings:", "answer": "Hotel guests who received real-time feedback on the energy consumed while showering used 11.4% (0.21 kWh) less energy per shower than hotel guests in a control group. The results are important for two reasons. First, they indicate that real-time feedback on a specific activity can induce large behaviour change and resource savings among the broader population and not only among the kind of people who volunteer to participate in energy conservation studies. Second, participants had no financial motives for behaviour change (hotel guests pay a fixed room rate). Even in this setting, the digital behavioural intervention created large conservation effects.\nTwo potential caveats remain. First, we do not know how representative the behaviour of hotel guests is for the tested behaviour compared to the general population. Second, we cannot measure long-term behaviour in a hotel setting. However, in other experiments that we conducted with households over several months, conservation effects from real-time feedback were stable.", "id": 33}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41560-018-0105-4\n\u00a9 2018 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.\n1The Energy Institute at the Johannes Kepler University Linz, Linz, Austria. 2Center for Secure Energy Informatics, Salzburg University of Applied Sciences, \nPuch/Salzburg, Austria. *e-mail: reichl@energieinstitut-linz.at\nR\nesidential electricity prices are made up of a number of compo-\nnents, including network tariffs, taxes and surcharges, such as a \nrenewables surcharge, a usage surcharge and an energy charge. \nIn this Article, we focus on network tariffs, which are the source \nfor recovering the capital and operational expenditures of provid-\ning transmission and distribution of electricity, as well as the grid \noperators\u2019 rate of return on grid investments, and are thereby the \nmain source for financing grid infrastructure.\nAs power grids are considered natural monopolies, the distribu-\ntion of these costs between consumers is not achieved autonomously \nthrough market forces1. In many modern economies, defining the \ncost shares for providing transmission and distribution of electricity \nbetween the consumers is assigned to an authorized entity, for which \nmost European Union nations have established a dedicated regula-\ntory authority since the electricity market liberalization in 2003. \nCommonly used and currently discussed network tariffs represent a \ncombination of volumetric energy charges (charging customers for \nthe amount of consumed electricity; \u20ac\u200b per kWh), fixed charges per \ncustomer (independent of their energy consumption but possibly \nvarying over households based on their contracted capacity, but not \nrelying on actual load measurements; \u20ac\u200b per household per year) and \npeak-demand charges (based on the actual measured capacity; \u20ac\u200b per \nkW peak load).\nThe costs of electricity networks are mainly determined by their \ncapacity, the maximum amount of energy that the grid is dimen-\nsioned to stand at any given point in time. Despite this, volumetric \ntariffs, which do not directly reflect the nature of these costs, are \nstill widely applied. In the past, when load profiles of residential \nusers were approximately homothetic, the application of tariffs with \na dominant volumetric charge was well justified. Nowadays, there \nis increasing diversity in daily load profiles, and part of this devel-\nopment is due to the increase of distributed generation, the advent \nof low-capacity storage (for example, in-home batteries for stor-\ning photovoltaic-produced electricity), the arising prospect of an \nincreasing number of electric vehicles, and the vision of house-to-\nhouse electricity trading to balance the overproduction from own-\ngeneration without the need (of higher levels) of the power grid2\u20135.\nThe deployment of renewable energy in the residential sector has \nincreasingly severe repercussions on electricity grids; for a growing \nshare of consumers, the connection to the public grid will largely \nserve as a backup option, rather than being the primary source for \ntheir electricity acquisition6,7. For such consumers, the volumes of \nelectricity consumed from the grid will be subordinate and likewise \nwill their contribution to the financing of the grid be low in the case \nof volumetric tariffs8\u201311. These trends will inevitably lead to a real-\nlocation of the burdens of grid-cost recovery12\u201314.\nConsidering that these innovations are more likely to happen \nfirst among a subgroup of the population owning single-family \ndwellings (as most of these innovations require property rights for \ninstallation), a significant social imbalance induced from shifting \nthe burdens of financing the grid towards lower-income classes may \narise15,16. This can hamper the public acceptance of these innova-\ntions. Moreover, some even envision a possible \u2018death spiral sce-\nnario\u201917\u201320, where higher network tariffs will be charged to poorer \ncustomers, which eventually threatens to collapse the whole elec-\ntricity supply system. Other recent studies (see for example refs 21,22) \nconsider such worries overblown, but still call for a timely and care-\nful revision of tariffs to avoid free-riding behaviour.\nSeveral options have been proposed to deal with this issue, \namong them minimum network charges per household irrespective \nof actual grid utilization, increased fixed charges and peak-demand \ncharges23\u201326. These latter charges have long been used in commercial \nand industrial network tariffs27, but are a novel development in the \nresidential electricity market.\nPolicymakers have to strike a proper balance between different \nobjectives when designing network tariffs, for example, they should \nbe easy to understand, fair, cost reflective, encourage energy effi-\nciency and send the right signals to maximize the economic effi-\nciency of power grids28. Therefore, difficult trade-offs have to be \nmade that will directly influence the monthly bills of the consum-\nExploring the impact of network tariffs on \nhousehold electricity expenditures using load \nprofiles and socio-economic characteristics\nValeriya Azarova\u200a \u200a1, Dominik Engel\u200a \u200a2, Cornelia Ferner2, Andrea Kollmann1 and Johannes Reichl1*\nGrowing self-generation and storage are expected to cause significant changes in residential electricity utilization patterns. \nCommonly applied volumetric network tariffs may induce imbalance between different groups of households and their respec-\ntive contribution to recovering the operating costs of the grid. Understanding consumer behaviour and appliance usage together \nwith socio-economic factors can help regulatory authorities to adapt network tariffs to new circumstances in a fair way. Here, \nwe assess the effects of 11 network tariff scenarios on household budgets using real load profiles from 765 households. Thus \nwe explore the possibly disruptive impact of applying peak-load-based tariffs on the budgets of households when they have \nbeen mainly charged for consumed volumes before. Our analysis estimates the change in household network expenditure for \ndifferent combinations of energy, peak and fixed charges, and can help to design tariffs that recover the costs needed for the \nsustainable operation of the grid.\nNature Energy | VOL 3 | APRIL 2018 | 317\u2013325 | www.nature.com/natureenergy\n317\n\n\u00a9 2018 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.\nArticles\nNaTUre EnerGy\ners and network tariff decisions should be informed by empirical \nresearch that defines who these consumers are and how their bills \nwill change.\nWith the advent of smart metering, recent literature has recon-\nsidered volumetric network tariffs and discussed measured, \ncapacity-oriented schemes to address the issues outlined above \n(for Australia see, for example, ref. 29, for the United Kingdom \nsee, for example, ref. 30 and for the United States see, for example,  \nrefs 27,31\u201334). The goal of this Article is to assess different network tariff  \nschemes with respect to their effects on the budgets of individual \nhouseholds conditional on their socio-economic backgrounds. Our \nanalysis aims at contributing to the ongoing tariff debate by pre-\nsenting the effects of alternative tariff schemes. We use a unique \nreal-world dataset of 765 Austrian households, whose electricity \nconsumption was metered for a one-year period, and for whom we \nhave detailed socio-demographic data at the individual level. Our \nanalysis shows that alternative tariff schemes, in particular changing \nfrom a volume-based scheme to a scheme that recovers a substantial \nportion of network costs through measured peak-demand charges, \nmay induce substantially increased electricity expenditures to a cer-\ntain share of households. Further analyses provide evidence that \nthese increasing expenditures cannot sufficiently be explained by \nthe possession of electric appliances, but are due to noticeable and \nsystematic differences in the electricity consumption patterns along \nthe politically critical dimensions of, among others, the households\u2019 \nincome situation and number of children.\nHousehold survey data\nHousehold-level data was gathered in two surveys among residen-\ntial electricity customers, conducted between April 2010 and March \n2011 in the region of Upper Austria, Austria. In the first survey, we \ncontacted more than 10,000 households via mail and asked them \nto allow us to collect their 15-minute electricity load profiles. We \nrecruited 973 households for participation, all of which gave their \ndistribution grid operator written permission to send us the house-\nholds\u2019 individual 15-minute electricity load profiles measured by \nsmart meter, for the full period of our survey. In total, we collected \n35,040 electricity load values for every household. Throughout the \nregistration process, households provided information about the \nnumber of people in the household, the type (apartment, single-\nfamily house, semi-detached house) and size (in m2) of their dwell-\ning, the technologies used for warm water and heat preparation \n(electricity, gas, district heating, heat pumps, biomass, oil), as well as \ntheir endowment with specific electric appliances that have a high \npower demand (swimming pool, fish tank, water bed, sauna, home \ncinema).\nIn the second survey, the same households were offered \u20ac\u200b10 if \nthey provided us with additional information about their socio-\neconomic characteristics (income and composition of the house-\nhold) and further information about their electric appliances. The \nfinal dataset includes 765 observations (which we refer to as \u2018full \nsample\u2019), for 406 of which we have additional information about \nthe household\u2019s income (henceforth called \u2018subsample\u2019). Details are \nprovided in Supplementary Note 1 as well as Supplementary Tables \n1 and 2.\nScenarios of network tariffs\nEver since Bonbright examined the principles of designing tariffs \nfor recovering the costs of natural monopolies35 in 1961, the topic \nhas been intensely researched for public utilities in general, as well \nas for electricity networks specifically (see, for example, refs 29,36\u201338). \nAs these sources establish, among others, a comprehensive devel-\nopment of network tariffs has to consider the dimensions \u2018system \nsustainability\u2019, \u2018economic efficiency\u2019 and \u2018distributive justice\u2019. In our \nstudy, tariff scenarios are exclusively designed to cover the range \nof potential network tariffs based on the candidate tariff compo-\nnents discussed above, to allow comparisons of the effects of these \nschemes on household expenditures. Technically, we follow the ulti-\nmate quantitative paradigm for designing network tariffs, which is \nto first determine the overall quantity of costs that shall be recov-\nered and then to define a distribution key by putting weights on \nthe candidate tariff components. According to this practice, and \nby treating our sample as if it was a tariff zone on its own, we first \nassess the sum of network charges to be paid by our full sample in \nthe Austrian tariff scheme as it was in force in 2016, which we refer \nto as \u2018reference scenario\u2019 henceforth, resulting in total charges of \u20ac\u200b\n136,209.10.\nTariff scenarios in our study recover this sum by putting differ-\nent weights on the three components: energy charges in \u20ac\u200b per kWh, \nfixed charges in \u20ac\u200b per year and peak charges in \u20ac\u200b per kW peak load. \nThese different weights are given in Table 1. Details on the calcula-\ntion of the respective tariffs are provided in Methods. Among the \nselected 11 alternative tariff schemes, 8 scenarios include a peak-\ndemand charge. The three scenarios not relying on measured loads \n(f100, f50/e50 and e100) are similar to the network tariffs currently \napplied in Europe. As an example, scenario f100 is similar to the \nscheme applied in the Netherlands in 2016, where all customers \nfaced the same network charges irrespective of their actual con-\nsumption patterns39. An overview of the volumetric share of tariffs \nas applied in some European countries in 2016 is given in ref. 40.\nFor our analysis, we quantify the annual network expenditures of \neach household under the tariff scenarios. We define our quantity of \ninterest as the percentage by which the network costs differ between \nthe reference scenario and the respective alternative scenario. This \nallows interpretation of the results in direct relation to the house-\nholds\u2019 expenditures in the status quo, that is, whether they would \nface increasing or decreasing costs under a certain alternative tariff \nscheme.\nTariff impact on household network expenses\nWe first assess the impact of the tariff scenarios from Table 1 on \nhousehold annual network expenditures (see Methods for details). \nFor the interpretation of our results, we point to the ex-post nature \nof our investigation: each of the tariff scenarios sends a specific sig-\nnal to households, for example, pa100 and pm100 provide incentives \nfor avoiding peak loads while they do not penalize high electricity \nconsumption. Ideally, households would analyse their consumption \npatterns with respect to the applied tariff scheme and adapt their \nbehaviour to minimize costs under certain boundary conditions41,42. \nAs the data exploited in our study were already collected in 2010\u2013\n2011, households were not provided with such signals.\nFigure 1 shows box plots of the change (in %) of the annual net-\nwork expenditures under the 11 tariff scenarios compared with \nthe reference scenario for the 765 households (for analysis of the \nsubsample see Supplementary Fig. 1). Several outcomes are evi-\ndent: first, the most significant change in households\u2019 network \nexpenditures are calculated for scenarios with a dominant share of \nfixed or peak-load charges and thereby deviate significantly from \nthe reference scenario. The interpretation works vice versa, that is, \none would observe similar changes in household network costs if, \nsay, f100 was currently applied and was substituted for the reference \nscenario.\nSecond, the box plots reveal that for some households the \nincreased costs in certain tariff scenarios are very high compared \nwith reference levels. For illustration, we marked the two house-\nholds that experience the highest increase in network expenditures \nunder peak-based scenarios by open and filled triangles. These are \nsingle-person households that consume moderate volumes of elec-\ntricity (that is, 1,805 and 1,604\u2009kWh in total during the observation \nperiod) and their current network costs in the reference scenario \nare low compared with the mean in our sample. At the same time, \nthese households produce massive peak loads (see Supplementary \nNature Energy | VOL 3 | APRIL 2018 | 317\u2013325 | www.nature.com/natureenergy\n318\n\n\u00a9 2018 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.\nArticles\nNaTUre EnerGy\nFigs. 2 and 3). Consequently, under tariff schemes charging exclu-\nsively for measured peak demand, such as pa100, they have to pay \nup to \u20ac\u200b435.40 and \u20ac\u200b471.60 more, translating to +\u200b422% and +\u200b499%, \nrespectively. In contrast to their sensitivity to tariffs, which empha-\nsize measured peak demand, the same two households face only \naround 80% higher charges in scenario f100. When neglecting peak \ndemand and applying scenario e100, which is determined by the \nvolume of consumed energy only, the opposite effect occurs and \nthese households would actually pay around 10% less than in the \nreference scenario.\nEven when the households marked by open and filled trian-\ngles are considered as outliers in our sample, these cases are still \nobserved in a relatively small sample of 765 households, suggest-\ning that a relevant number of households may face significant addi-\ntional burdens when/if household peak-load-based charges are \nintroduced. We check how many households in the sample also tend \nto have increasing costs under some of the scenarios while experi-\nence cost savings in others. We find that 321 households face lower \ncosts in scenario e100 and higher costs in scenario pa100, which \nmeans that nearly 40% of the households consume relatively mod-\nerate volumes of energy in total, but at the same time frequently \nproduce significant peak loads. This example demonstrates that a \ndifferent weighting of the volumetric, peak and fixed components \ncan have strongly diametrical effects on the network expenditures \nof individual households.\nWhile we do not think that this result shall prevent the appli-\ncation of such innovative network tariffs, before their introduction \na careful impact assessment appears necessary, which may be fol-\nlowed by mechanisms balancing hardship cases during the transi-\ntion period.\nNetwork expenditures and socio-economic characteristics\nIn this section, we assess which household characteristics aid in \nexplaining the impact the tariff scenarios have on households\u2019 net-\nwork expenditures. Regression results of the full sample with seven \nexplanatory variables and a constant are shown in Table 2 (for main \nresults excluding outliers see Supplementary Table 4), while results \n\u201350\nScenario\n0\n50\n100\n150\n200\n250\n300\n350\n400\n450\n500\nDifference to currently applied tariff (%)\nf100\nf50/e50\ne100\npa100\npa50/e50\nf50/pa50\nf/pa/e\npm100\npm50/e50\nf50/pm50\nf/pm/e\nFig. 1 | Change in annual network expenditure under different tariff \nscenarios. The full sample of 765 households is analysed in each of the \nscenarios.The bottom of each box is the 25th percentile, the horizontal \nline in the middle of each box represents the median and the top of each \nbox is the 75th percentile. Vertical lines outside of the box (whiskers) \nend at the 10th and 90th percentiles. The dots are considered as outliers. \nFor illustration, the two households with the highest increase in network \nexpenditures under scenarios including a peak component are marked by \nopen and filled triangles.\nTable 1 | Residential network tariff scenarios applied to the full \nsample of 765 households\nScenario\nDescription \n(overall \nnetwork costs \nare recovered \nthrough)\nFixed charge \n(\u20ac per \nhousehold \nper year)\nEnergy \ncharge \n(\u20ac per \nkWh)\nPeak/\ncapacity \ncharge  \n(\u20ac per kW \npeak load)\nReference\nTariff as applied in \nAustria in 2016\n24.60\n0.043\n\u2013\nf100\n100% flat tariff\n178.05\n\u2013\n\u2013\npa100\n100% peak charge, \nbased on the \naverage of the  \n12 monthly \nmeasured  \npeak loads\n\u2013\n\u2013\n39.07\npm100\n100% peak charge, \nbased on the one \nmaximum load\n\u2013\n\u2013\n29.59\ne100\n100% energy \ncharge, only based \non consumed \nvolume\n\u2013\n0.050\n\u2013\nf50/e50\n50% from fixed \ncharges and 50% \nfrom consumed \nvolume\n89.02\n0.025\n\u2013\nf50/pa50\n50% from fixed \ncharges and 50% \nfrom peak charges \n(average of the  \n12 monthly peaks)\n89.02\n\u2013\n19.53\nf50/pm50\n50% from fixed \ncharges and 50% \nfrom peak charges \n(one maximum \nload)\n89.02\n\u2013\n14.79\npa50/e50\n50% from peak \ncharges (average \nof the 12 monthly \npeaks) charge \nand 50% from \nconsumed volume\n\u2013\n0.025\n19.53\npm50/e50\n50% from peak \ncharges (one \nmaximum load) \nand 50% from \nconsumed volume\n\u2013\n0.025\n14.79\nf/pa/ea\n14% from fixed \ncharges, 43% \nfrom consumed \nvolume and 43% \nfrom peak charges \n(average of the 12 \nmonthly peaks)\n24.60\n0.022\n16.83\nf/pm/ea\n14% from fixed \ncharge, 43%  \nfrom consumed \nvolume and  \n43% from peak \ncharges (one \nmaximum load)\n24.60\n0.022\n12.75\naThe fixed charge is taken from the reference scenario, while the remaining portion is split into \nequal quantities, see Methods for details.\nNature Energy | VOL 3 | APRIL 2018 | 317\u2013325 | www.nature.com/natureenergy\n319\n\n\u00a9 2018 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.\nArticles\nNaTUre EnerGy\nof the regression exploiting the subsample, including informa-\ntion about the households\u2019 income, household type and electricity \nconsuming amenities, are presented in Table 3. The results of both \nregressions show similar impact on household expenditures for \nnetwork tariffs, with only minor changes in the magnitude of the \ncoefficients and their significance, which we interpret as evidence \nfor the robustness of our results. The applied statistical model is pre-\nsented in Methods.\nThe regression results shown in Table 2 suggest that the number \nof residents of a household (Nr_persons), the living space (Square), \nwhether the dwelling is in a rural or an urban environment (Rural) \nor whether the dwelling is a single-family house (House), are associ-\nated with lower network costs under scenarios with a fixed charge \ncompared with the reference scenario, and higher costs in the fully \nenergy-based scenario e100. This is explained by the fact that these \nhouseholds (ceteris paribus) consume higher volumes of electricity \nand thereby benefit from tariffs, which put only subordinate weight \non the number of consumed units. Interestingly, almost the same \nhousehold characteristics (with exception of a rural versus urban \nenvironment and living space) are related to lower network costs \nin scenarios emphasizing measured peak demand (pa100, pm100, \npm50/e50, and pa50/e50), providing evidence that households with \nmore residents and households situated in a single-family house \ntend to produce fewer peak loads, ceteris paribus.\nThe inclusion of the household amenities in the regressions is \nessential even though it is not the focus of our analysis, as they would \nlikely suffer from omitted variable bias otherwise, as these are corre-\nlated with the policy-relevant variables. The signs of their effects are \nas expected conditional on which tariff component the respective \nscenario emphasizes. Tumble dryer and PC increase households\u2019 \ncosts in the fully energy-based scenario e100, while flow heaters \nbecome increasingly costly in the scenarios charging for peak loads. \nPool owners benefit from peak or fixed tariffs compared with the \nstatus quo, and households with a sauna have a significant disadvan-\ntage in tariffs charging for measured peak demand. Further analysis \nof whether the amenities are related to high energy volumes or peak \nloads are given in Supplementary Note 3 and Supplementary Tables \n5 and 6.\nFrom a policy perspective, it is important to notice that most \nof the parameters that significantly contribute to lower network \ncharges under the respective alternative scenario (that is, house-\nholds living in single-family houses, having larger living spaces \nand swimming pool ownership) are usually associated with higher \nincome levels. As the sum of collected revenues from all households \ntogether is required to remain unchanged under any new tariff \nscheme, a reduction of the financial contribution of higher-income \nhouseholds would automatically mean an increase of burdens for \nlower-income households compared with the situation under the \nreference scenario. As such distributional effects are problematic \nfrom a public choice perspective, it is important to identify whether \ncertain tariff scenarios are actually associated with shifting burdens \ntowards households with lower incomes.\nMost importantly, the regression output in Table 3 provides evi-\ndence for the distributional effects when the tariff scheme is changed \nfrom emphasizing the volumetric component as in the reference \nscenario, towards a stronger weighting of peak demand. A house-\nTable 2 | Effects of household characteristics and amenities on the relative difference of their network costs in our full sample\nf100\nf50/e50\ne100\npa100\npa50/\ne50\nf50/pa50\nf/pa/e\npm100\npm50/\ne50\nf50/pm50\nf/pm/e\nNr_persons\n\u2212\u200b\n15.643***\n\u2212\u200b6.564*** 2.515***\n\u2212\u200b6.275***\n\u2212\u200b\n1.880***\n\u2212\u200b\n10.959***\n\u2212\u200b3.782***\n\u2212\u200b7.351***\n\u2212\u200b2.418***\n\u2212\u200b11.497***\n\u2212\u200b4.245***\n(1.354)\n(0.568)\n(0.218)\n(1.312)\n(0.626)\n(1.096)\n(0.605)\n(1.383)\n(0.656)\n(1.143)\n(0.639)\nSquare\n\u2212\u200b0.173***\n\u2212\u200b0.073***\n0.028***\n\u2212\u200b0.085**\n\u2212\u200b0.029\n\u2212\u200b0.129***\n\u2212\u200b\n0.049***\n\u2212\u200b0.050\n\u2212\u200b0.011\n\u2212\u200b0.112***\n\u2212\u200b0.034*\n(0.042)\n(0.018)\n(0.007)\n(0.041)\n(0.019)\n(0.034)\n(0.019)\n(0.043)\n(0.020)\n(0.035)\n(0.020)\nDummy_\nsquare\n\u2212\u200b8.391**\n\u2212\u200b3.521**\n1.349**\n\u2212\u200b3.664\n\u2212\u200b1.158\n\u2212\u200b6.028**\n\u2212\u200b2.157\n\u2212\u200b3.384\n\u2212\u200b1.018\n\u2212\u200b5.888*\n\u2212\u200b2.036\n(3.641)\n(1.528)\n(0.585)\n(3.528)\n(1.683)\n(2.948)\n(1.626)\n(3.720)\n(1.765)\n(3.074)\n(1.718)\nPool\n\u2212\u200b\n23.022***\n\u2212\u200b9.661***\n3.701***\n\u2212\u200b12.790**\n\u2212\u200b4.545*\n\u2212\u200b17.906*** \u2212\u200b7.098***\n\u2212\u200b15.572***\n\u2212\u200b5.935**\n\u2212\u200b19.297*** \u2212\u200b8.296***\n(5.860)\n(2.459)\n(0.942)\n(5.678)\n(2.709)\n(4.745)\n(2.617)\n(5.987)\n(2.841)\n(4.947)\n(2.765)\nSauna\n\u2212\u200b2.238\n\u2212\u200b0.939\n0.360\n9.864**\n5.112**\n3.813\n4.096**\n16.102***\n8.231***\n6.932*\n6.784***\n(4.430)\n(1.859)\n(0.712)\n(4.292)\n(2.048)\n(3.587)\n(1.978)\n(4.526)\n(2.148)\n(3.740)\n(2.090)\nSolarium\n\u2212\u200b12.620\n\u2212\u200b5.295\n2.029\n9.739\n5.884\n\u2212\u200b1.440\n3.327\n9.265\n5.647\n\u2212\u200b1.677\n3.123\n(10.398)\n(4.363)\n(1.672)\n(10.075)\n(4.806)\n(8.420)\n(4.643)\n(10.624)\n(5.041)\n(8.779)\n(4.905)\nRural\n\u2212\u200b18.723*** \u2212\u200b7.856***\n3.010***\n\u2212\u200b7.644\n\u2212\u200b2.317\n\u2212\u200b13.183***\n\u2212\u200b4.584**\n\u2212\u200b6.262\n\u2212\u200b1.626\n\u2212\u200b12.493*** \u2212\u200b3.988*\n(4.827)\n(2.025)\n(0.776)\n(4.677)\n(2.231)\n(3.909)\n(2.155)\n(4.932)\n(2.340)\n(4.075)\n(2.277)\nHouse\n\u2212\u200b15.212***\n\u2212\u200b6.383***\n2.445***\n\u2212\u200b\n18.694***\n\u2212\u200b8.125***\n\u2212\u200b16.953*** \u2212\u200b9.104***\n\u2212\u200b19.303***\n\u2212\u200b8.429***\n\u2212\u200b17.258***\n\u2212\u200b9.366***\n(3.994)\n(1.676)\n(0.642)\n(3.870)\n(1.846)\n(3.235)\n(1.784)\n(4.081)\n(1.936)\n(3.372)\n(1.884)\nConstant\n108.497*** 45.665***\n\u2212\u200b\n17.168***\n52.072***\n17.452***\n80.285***\n30.031***\n49.352***\n16.092***\n78.925***\n28.859***\n(5.875)\n(2.465)\n(0.944)\n(5.692)\n(2.715)\n(4.757)\n(2.623)\n(6.002)\n(2.848)\n(4.960)\n(2.771)\nObservations 763\nR2\n0.373\n0.373\n0.373\n0.164\n0.106\n0.347\n0.217\n0.152\n0.102\n0.326\n0.199\nAll values are given as percentage changes between the reference scenario and the stated alternative scenario. Standard errors are given in parentheses. * <\nP\n0.10, ** <\nP\n0.05, *** <\nP\n0.01.\nNature Energy | VOL 3 | APRIL 2018 | 317\u2013325 | www.nature.com/natureenergy\n320\n\n\u00a9 2018 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.\nArticles\nNaTUre EnerGy\nTable 3 | Effects of household characteristics and amenities on the relative difference of their network costs in our subsample\nf100\nf50/e50\ne100\npa100\npa50/e50 f50/pa50\nf/pa/e\npm100\npm50/e50 f50/pm50\nf/pm/e\nLogincome\n\u2212\u200b7.526\n\u2212\u200b3.183\n1.160\n\u2212\u200b12.400** \u2212\u200b5.620**\n\u2212\u200b9.963**\n\u2212\u200b5.875**\n\u2212\u200b12.187**\n\u2212\u200b5.514**\n\u2212\u200b9.857**\n\u2212\u200b5.782**\n(5.014)\n(2.121)\n(0.773)\n(5.124)\n(2.435)\n(4.235)\n(2.373)\n(5.551)\n(2.646)\n(4.417)\n(2.556)\nDummy_income 6.012\n2.543\n\u2212\u200b0.927\n\u2212\u200b2.266\n\u2212\u200b1.596\n1.873\n\u2212\u200b0.580\n\u2212\u200b2.719\n\u2212\u200b1.823\n1.646\n\u2212\u200b0.777\n(4.418)\n(1.868)\n(0.681)\n(4.514)\n(2.145)\n(3.731)\n(2.090)\n(4.890)\n(2.331)\n(3.891)\n(2.252)\nSquare\n\u2212\u200b0.223***\n\u2212\u200b0.094*** 0.034***\n\u2212\u200b0.087\n\u2212\u200b0.026\n\u2212\u200b0.155***\n\u2212\u200b0.052** \u2212\u200b0.050\n\u2212\u200b0.008\n\u2212\u200b0.136***\n\u2212\u200b0.036\n(0.055)\n(0.023)\n(0.008)\n(0.056)\n(0.027)\n(0.046)\n(0.026)\n(0.061)\n(0.029)\n(0.048)\n(0.028)\nDummy_square \u2212\u200b3.457\n\u2212\u200b1.462\n0.533\n\u2212\u200b4.063\n\u2212\u200b1.765\n\u2212\u200b3.760\n\u2212\u200b1.991\n\u2212\u200b2.447\n\u2212\u200b0.957\n\u2212\u200b2.952\n\u2212\u200b1.291\n(4.229)\n(1.788)\n(0.652)\n(4.321)\n(2.053)\n(3.572)\n(2.001)\n(4.681)\n(2.232)\n(3.724)\n(2.156)\nRural\n\u2212\u200b17.282***\n\u2212\u200b7.309*** 2.663***\n\u2212\u200b6.530\n\u2212\u200b1.933\n\u2212\u200b11.906**\n\u2212\u200b3.983\n\u2212\u200b5.243\n\u2212\u200b1.290\n\u2212\u200b11.263**\n\u2212\u200b3.425\n(5.934)\n(2.510)\n(0.915)\n(6.064)\n(2.882)\n(5.012)\n(2.808)\n(6.569)\n(3.132)\n(5.227)\n(3.025)\nHouse\n\u2212\u200b18.469*** \u2212\u200b7.811***\n2.846***\n\u2212\u200b10.319*\n\u2212\u200b3.736\n\u2212\u200b14.394*** \u2212\u200b5.704** \u2212\u200b11.259*\n\u2212\u200b4.206\n\u2212\u200b14.864*** \u2212\u200b6.111**\n(5.257)\n(2.223)\n(0.810)\n(5.372)\n(2.553)\n(4.440)\n(2.487)\n(5.819)\n(2.774)\n(4.630)\n(2.680)\nNr_persons\n\u2212\u200b10.824*** \u2212\u200b4.578*** 1.668***\n\u2212\u200b4.959*\n\u2212\u200b1.645\n\u2212\u200b7.892***\n\u2212\u200b2.871**\n\u2212\u200b6.442**\n\u2212\u200b2.387\n\u2212\u200b8.633***\n\u2212\u200b3.514**\n(2.816)\n(1.191)\n(0.434)\n(2.878)\n(1.368)\n(2.379)\n(1.332)\n(3.118)\n(1.486)\n(2.480)\n(1.436)\nDryer\n\u2212\u200b10.605*** \u2212\u200b4.485*** 1.634***\n\u2212\u200b0.213\n0.710\n\u2212\u200b5.409\n\u2212\u200b0.801\n\u2212\u200b2.266\n\u2212\u200b0.316\n\u2212\u200b6.435*\n\u2212\u200b1.690\n(3.933)\n(1.663)\n(0.606)\n(4.019)\n(1.910)\n(3.322)\n(1.861)\n(4.354)\n(2.076)\n(3.464)\n(2.005)\nDishwasher\n\u2212\u200b5.062\n\u2212\u200b2.141\n0.780\n2.821\n1.801\n\u2212\u200b1.120\n0.884\n\u2212\u200b0.302\n0.239\n\u2212\u200b2.682\n\u2212\u200b0.469\n(5.754)\n(2.433)\n(0.887)\n(5.879)\n(2.794)\n(4.860)\n(2.722)\n(6.369)\n(3.036)\n(5.068)\n(2.933)\nPool\n\u2212\u200b6.495\n\u2212\u200b2.747\n1.001\n\u2212\u200b15.068** \u2212\u200b7.034**\n\u2212\u200b10.781**\n\u2212\u200b6.962**\n\u2212\u200b15.695**\n\u2212\u200b7.347**\n\u2212\u200b11.095**\n\u2212\u200b7.233**\n(6.230)\n(2.635)\n(0.960)\n(6.367)\n(3.025)\n(5.262)\n(2.948)\n(6.897)\n(3.288)\n(5.488)\n(3.176)\nSauna\n3.280\n1.387\n\u2212\u200b0.506\n14.893***\n7.193***\n9.086**\n6.671***\n24.390***\n11.942***\n13.835***\n10.786***\n(5.184)\n(2.193)\n(0.799)\n(5.297)\n(2.517)\n(4.379)\n(2.453)\n(5.739)\n(2.736)\n(4.566)\n(2.643)\nFlowheater\n0.539\n0.228\n\u2212\u200b0.083\n9.293*\n4.605*\n4.916\n4.062*\n11.837**\n5.877**\n6.188\n5.164**\n(4.842)\n(2.048)\n(0.746)\n(4.948)\n(2.351)\n(4.090)\n(2.291)\n(5.361)\n(2.555)\n(4.265)\n(2.468)\nBoiler\n\u2212\u200b1.252\n\u2212\u200b0.529\n0.193\n0.238\n0.215\n\u2212\u200b0.507\n0.019\n\u2212\u200b1.092\n\u2212\u200b0.450\n\u2212\u200b1.172\n\u2212\u200b0.557\n(4.011)\n(1.696)\n(0.618)\n(4.099)\n(1.948)\n(3.388)\n(1.898)\n(4.440)\n(2.117)\n(3.533)\n(2.045)\nPC\n\u2212\u200b18.551***\n\u2212\u200b7.846*** 2.859***\n\u2212\u200b6.555\n\u2212\u200b1.848\n\u2212\u200b12.553*** \u2212\u200b4.078*\n\u2212\u200b7.860\n\u2212\u200b2.500\n\u2212\u200b13.205*** \u2212\u200b4.644*\n(4.876)\n(2.062)\n(0.752)\n(4.983)\n(2.368)\n(4.119)\n(2.307)\n(5.398)\n(2.573)\n(4.295)\n(2.486)\nHouseholdtype2 \u2212\u200b7.891\n\u2212\u200b3.337\n1.216\n\u2212\u200b4.129\n\u2212\u200b1.456\n\u2212\u200b6.010\n\u2212\u200b2.316\n\u2212\u200b4.803\n\u2212\u200b1.793\n\u2212\u200b6.347\n\u2212\u200b2.608\n(13.331)\n(5.638)\n(2.055)\n(13.623)\n(6.474)\n(11.260)\n(6.308)\n(14.758)\n(7.035)\n(11.742)\n(6.796)\nHouseholdtype3 \u2212\u200b27.576*** \u2212\u200b11.663*** 4.250***\n\u2212\u200b14.860** \u2212\u200b5.305\n\u2212\u200b21.218***\n\u2212\u200b8.279**\n\u2212\u200b13.619*\n\u2212\u200b4.684\n\u2212\u200b20.597*** \u2212\u200b7.741**\n(7.120)\n(3.012)\n(1.097)\n(7.276)\n(3.458)\n(6.014)\n(3.369)\n(7.882)\n(3.758)\n(6.272)\n(3.630)\nHouseholdtype4 \u2212\u200b24.570*** \u2212\u200b10.392*** 3.787***\n\u2212\u200b0.861\n1.463\n\u2212\u200b12.716**\n\u2212\u200b2.014\n\u2212\u200b1.216\n1.285\n\u2212\u200b12.893**\n\u2212\u200b2.167\n(6.719)\n(2.842)\n(1.036)\n(6.866)\n(3.263)\n(5.675)\n(3.179)\n(7.438)\n(3.546)\n(5.918)\n(3.425)\nHouseholdtype5 \u2212\u200b11.895\n\u2212\u200b5.031\n1.833\n\u2212\u200b0.521\n0.656\n\u2212\u200b6.208\n\u2212\u200b1.020\n1.329\n1.581\n\u2212\u200b5.283\n\u2212\u200b0.218\n(10.098)\n(4.271)\n(1.556)\n(10.319)\n(4.904)\n(8.529)\n(4.778)\n(11.179)\n(5.329)\n(8.894)\n(5.148)\nHouseholdtype6 \u2212\u200b20.376**\n\u2212\u200b8.618**\n3.140**\n\u2212\u200b5.744\n\u2212\u200b1.302\n\u2212\u200b13.061*\n\u2212\u200b3.849\n\u2212\u200b3.351\n\u2212\u200b0.106\n\u2212\u200b11.864\n\u2212\u200b2.812\n(9.071)\n(3.837)\n(1.398)\n(9.270)\n(4.405)\n(7.662)\n(4.292)\n(10.042)\n(4.787)\n(7.990)\n(4.624)\nHouseholdtype7 \u2212\u200b20.485\n\u2212\u200b8.664\n3.157\n\u2212\u200b5.315\n\u2212\u200b1.079\n\u2212\u200b12.900\n\u2212\u200b3.670\n\u2212\u200b9.601\n\u2212\u200b3.222\n\u2212\u200b15.043\n\u2212\u200b5.527\n(20.805)\n(8.799)\n(3.206)\n(21.260)\n(10.103)\n(17.572)\n(9.843)\n(23.031)\n(10.979)\n(18.325)\n(10.606)\nHouseholdtype8 16.867\n7.134\n\u2212\u200b2.599\n2.876\n0.138\n9.871\n2.372\n5.545\n1.473\n11.206\n3.528\n(16.390)\n(6.932)\n(2.526)\n(16.748)\n(7.959)\n(13.843)\n(7.755)\n(18.144)\n(8.649)\n(14.436)\n(8.355)\nHouseholdtype9 \u2212\u200b29.454*** \u2212\u200b12.457*** 4.539***\n\u2212\u200b15.589\n\u2212\u200b5.525\n\u2212\u200b22.522*** \u2212\u200b8.720*\n\u2212\u200b14.585\n\u2212\u200b5.023\n\u2212\u200b22.020**\n\u2212\u200b8.285\n(10.246)\n(4.333)\n(1.579)\n(10.470)\n(4.975)\n(8.654)\n(4.848)\n(11.342)\n(5.407)\n(9.024)\n(5.223)\nConstant\n198.240*** 83.846*** \u2212\u200b30.552*** 143.948*** 56.698*** 171.096***\n75.599*** 143.489*** 56.468***\n170.866***\n75.400***\n(36.683)\n(15.515)\n(5.653)\n(37.485)\n(17.813)\n(30.983)\n(17.356)\n(40.609)\n(19.359)\n(32.310)\n(18.700)\nObservations\n404\nR2\n0.534\n0.534\n0.534\n0.204\n0.138\n0.452\n0.270\n0.203\n0.147\n0.437\n0.260\nAll values are given as percentage changes between the reference scenario and the stated alternative scenario. See Supplementary Table 2 for definitions of household types. Standard errors are given in \nparentheses. * <\nP\n0.10, ** <\nP\n0.05, *** <\nP\n0.01.\nNature Energy | VOL 3 | APRIL 2018 | 317\u2013325 | www.nature.com/natureenergy\n321\n\n\u00a9 2018 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.\nArticles\nNaTUre EnerGy\nhold\u2019s income level is expressed in two ways in our sample, that is, \ndirectly through the income variable and indirectly through the \nhousehold\u2019s amenities. The regression output provides evidence that \nthe log of the monthly net income (Logincome) of a household is \nsignificant (at different levels) for all scenarios introducing a charge \nfor measured peak demand, but it is not significant for the remain-\ning three scenarios. A respective negative coefficient of income \nmeans that households with a higher income are associated with \nlower peak loads compared with households with the same charac-\nteristics but less income, ceteris paribus. As our regression controls \nfor a number of household characteristics and appliances, this effect \ncannot stem from the disproportionate equipment of wealthier \nhouseholds with these observed amenities per se, but instead could \nbe from (1) households with lower incomes having load-intensive \namenities not observed in our sample, which higher-income house-\nholds do not have, (2) among those households having the same \namenities, wealthier households having more modern and thereby \nless load-intensive amenities, or (3) wealthier households utiliz-\ning their amenities in a less load-intensive way, for example, using \ntheir sauna differently. The question which of these options plays a \nrole in the frequency and extent of peak loads is important from a \npolicy perspective, considering that tariffs with a charge for mea-\nsured peaks seem to favour higher income levels. In the case that \noption (1) and/or (2) are relevant, policy could aid lower-income \nhouseholds in identifying these load-intensive amenities and pro-\nvide support for substituting them, where possible. To investigate \nsuch a potential relationship, in Supplementary Table 3 we regress \nthe frequency of peak loads and the annual energy consumption on \nhousehold characteristics and amenities again, and extend the set of \nexplanatory variables by an interaction term for households with an \nincome below the median and the respective amenities. We find evi-\ndence that electricity consumption patterns of two amenities differ \nbetween the income groups: in below-median-income households, \npools produce fewer peak loads, while flow heaters are respon-\nsible for about 1,000 additional kWh in this group. The higher \nenergy consumption of flow heaters in lower-income households  \nmay actually point to the need for respective policy measures (see \nSupplementary Note 2 for more details).\nWith respect to the composition of the households, we observe \nthat compared with single households, households made up of a \ncouple with children (Householdtype5/6) or a couple without chil-\ndren (Householdtype3/4) have an advantage when tariffs with a \nsubstantial fixed charge are introduced, and experience disadvan-\ntages under a 100% energy-based tariff. Further evidence on the \nsystematically different consumption patterns of different house-\nhold types are provided in Figs. 2 and 3. Figure 2 compares the load \nprofiles of these household groups on winter Saturdays and sum-\nmer workdays: households with children (Householdtype2/5/6/7) \nhave substantially higher loads than those without children, and \nhigher-income households consume more electricity. While the \ndifference between households with and without children may, to \nsome extent, be rooted in the higher average number of residents, \nthe additional consumption during almost all times of the day sup-\nports the regression findings of energy-based tariffs being in favour \nof childless households in absolute terms. Figure 3 compares the \npercentage of households exceeding certain load thresholds at dif-\nferent times of a day at least once during the observation period: \nagain households with children are more likely to produce high \npeak loads than those without children, and higher-income house-\nholds are also more likely to produce significant peaks. However, \ndespite higher-income households being above the lower-income \nhouseholds in both the average energy consumption (Fig. 2) and \nthe likelihood of exceeding a certain load threshold (Fig. 3), the \njoint estimation suggests that peak-load-based tariffs are more \nfavourable for high-income households than tariffs based mainly \non energy charges (ceteris paribus).\nSummarizing the results of our statistical analysis, we find that \nthe living situation of a household and its electricity consuming \namenities as well as its income level and the number of children \nseem to play a role in whether a change from the reference scenario \nto a tariff scheme charging for peak demand is associated with ben-\nefits or additional burdens. Our results indicate that, ceteris paribus, \n1.0\n0.8\na\nb\nd\nc\nAverage demand (kW)\nAverage demand (kW)\nAverage demand (kW)\nAverage demand (kW)\n0.6\n0.4\n0.2\n0\n1.0\n0.8\n0.6\n0.4\n0.2\n0\n1.0\n0.8\n0.6\n0.4\n0.2\n0\n1.0\n0.8\n0.6\n0.4\n0.2\n0\n00:00 03:00 06:00 09:00\nTime of day\nTime of day\nTime of day\nTime of day\n12:00 15:00 18:00 21:00 00:00\n00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00\n00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00\n00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00\nFull sample\nHigher-income\nhouseholds\nLower-income\nhouseholds\nH0\nFull sample\nHouseholds with \nchildren\nHouseholds without \nchildren\nH0\nFig. 2 | Comparison of the load profiles of different types of households in our subsample with respective standardized load profile H0. Load profile H0 \nis used by Austrian utilities for forecasting and accounting household electricity consumption when no data from load metering is available43. a,b, Average \nload on winter Saturdays (a) and summer workdays (b) for high-income households (above median monthly net household income of \u20ac\u200b2,043; n\u2009=\u200b\u2009209), \nlow-income households (below-median income; n\u2009=\u200b\u2009197), all households in the respective sample and the corresponding standard load profile H0.  \nc,d, Yearly average load on Saturdays during winter (c) and summer workdays (d) of households with (n\u2009=\u200b\u200991) and without children (n\u2009=\u200b\u2009315),  \ncompared with all households in the sample households and the corresponding standard load profile H0.\nNature Energy | VOL 3 | APRIL 2018 | 317\u2013325 | www.nature.com/natureenergy\n322\n\n\u00a9 2018 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.\nArticles\nNaTUre EnerGy\nhouseholds with higher income are better off when tariffs charging \nfor measured peak demand are introduced. Among others, ref. 44 \nshowed that photovoltaic panels in combination with home stor-\nage appliances reduce the peak loads of households significantly, \nwhile at the same time also reduce the volumes of electricity pur-\nchased via the grid. Higher-income households are likely to install \nphotovoltaic panels and related peak-load-reducing equipment at \nhigher rates than lower-income households45,46. Shifting the burdens \nof recovering network costs towards lower-income households may \ntherefore become an even more pressing issue in the future.\nDiscussion\nThe ongoing transformation of the electricity system calls for the \nreconsideration of the current network tariff schemes. Our analysis \naims to provide evidence on the potential magnitude of such tar-\niff changes, and identify whether potentially increasing costs are \nequally distributed among the population or are more pronounced \nfor specific groups. Considering that the availability of residential \nload profiles is relatively new to policymakers and scientists, and \nthat the lack of respective socio-economic background information \nassociated with these load profiles is hampering comprehensive \ninvestigations, the knowledge gained in this study provides substan-\ntial input for the ongoing debate about developing and implement-\ning new tariff schemes.\nInvestigating data on 765 households in Austria, we find that the \nchange in network charges, depending on the scenarios applied, can \n(in extreme cases) reach a decrease of 50% or an increase of 500% \ncompared with the status quo. This demonstrates that some of the \ntested tariff scenarios may have a disruptive impact on some house-\nholds\u2019 budgets if implemented from one accounting period to the \nnext.\nWe find it important to highlight the potentially low predict-\nability for households of their annual network costs under tariffs \nemphasizing peak charges (per measured kW). Considering, for \nexample, tariff scenario pm100, where household network charges \nare defined by the highest load during one out of 35,040 quarters of \nan hour, significantly increased network costs from one year to the \nother can arise from one unusually coincidental use of appliances. \nThe high level of sensitivity in household electricity costs to small \nlapses on the part of the household when measured peak-demand \ncharges are used makes it necessary to apply such tariffs thought-\nfully, following careful research. The need for a delicate handling of \nthe roll-out of measured peak-demand tariff schemes is particularly \nurgent when such schemes are not accompanied by support mecha-\nnisms, such as extended transition periods allowing households to \nadapt to the new price signals.\nUnder volumetric network tariffs, every reduction of the units \nof consumed energy results in an under-recovery of network costs. \nThis can either be compensated by increasing the price for the unit \nof consumed energy, or by decoupling the revenues from network \ntariffs from consumed volumes. Looking at the socio-demographic \ncharacteristics of the households in the sample, we see that tariffs \ncombining measured peak-demand and volumetric components \ncould provide a new balance for the distribution of network costs\u2014\nas these tariffs are cost reflective and, due to the peak-load charge, \nthey signal the consumer to decrease overall consumption while not \npenalizing any specific group of consumers. An additional fixed \ncomponent can account for costs invariant to consumption pat-\nterns, such as charges for metering itself. We find that such tariffs \ncould provide a solid response to the increase of prosumers, while \navoiding the shifting of burdens towards households not yet ready \nfor taking this step.\nThis analysis is limited to data on Austrian consumers, so further \nresearch with data on more households, including data on photo-\nvoltaic ownership and home storage installations could improve our \nunderstanding of the tariffs\u2019 effects on both consumers and prosumers.\nMethods\nStatistical methods. The quantity of interest in our analyses is defined as the \npercentage by which the network costs differ between the reference scenario and \nthe alternative scenarios, and we refer to this quantity by:\n\u0394 =\n\u2212\n\u00d7\nC\nC\nC\n100\n(1)\ni j\ni j\ni r\ni r\n,\n,\n,\n,\nwhere Ci,r is the annual network costs of household i in the reference \nscenario, while Ci,j stands for i\u2032\u200bs costs in the jth alternative scenario, and \n\u2208\n\u2215\n\u2215\n\u2215\n\u2215\n\u2215\nj\n(f100, e100, f50 pa50, f50 e50, f50 pm50, pa100, pm100, f pa e,\n\u2215\n\u2215\n\u2215\n\u2215\npa50 e50, pm50 e50, f pm e). A negative sign of \u0394i,j therefore indicates a cost \nHouseholds exceeding\n4 kW at least once (%)\nHouseholds exceeding\n4 kW at least once (%)\n60\na\nc\nd\nb\n40\n20\n0\nHouseholds exceeding\n7 kW at least once (%)\nHouseholds exceeding\n7 kW at least once (%)\n10\n8\n6\n4\n2\n0\n10\n8\n6\n4\n2\n0\n60\n40\n20\n0\n00:00 03:00 06:00 09:00 12:00\nTime of day\nTime of day\nTime of day\nTime of day\n15:00 18:00 21:00 00:00\n00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00\n00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00\n00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00\nAll households\nHigher-income\nhouseholds\nLower-income\nhouseholds\nAll households\nHouseholds with \nchildren\nHouseholds without\nchildren\nFig. 3 | Percentage of households exceeding certain load thresholds at a certain time of the day at least once during the observation period. a,b, \nPercentage of high-income households (above median monthly net household income of \u20ac\u200b2,043; n\u2009=\u200b\u2009209), low-income households (below-median \nincome; n\u2009=\u200b\u2009197) and all households in the respective sample exceeding 4\u2009kW (a) and 7\u2009kW (b) during the respective time of the day. c,d, Percentage of \nhouseholds with (n\u2009=\u200b\u200991) and without children (n\u2009=\u200b\u2009315), compared with all households, exceeding 4\u2009kW (c) and 7\u2009kW (d) during the respective time of the \nday. Thresholds correspond to the available capacity contracts in the tariff zone of our sample; the minimum contract for households is for 4\u2009kW, the next \nhigher contract is for 7\u2009kW. In practice, actual demand limitations of household connections are significantly higher than contracted.\nNature Energy | VOL 3 | APRIL 2018 | 317\u2013325 | www.nature.com/natureenergy\n323\n\n\u00a9 2018 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.\nArticles\nNaTUre EnerGy\nreduction under the alternative scenario j compared with today\u2019s regulatory \npractice, while a positive sign points to increased costs for household i.\nTo investigate which household characteristics are associated with cost savings \nor incremental costs under the different alternative scenarios, we estimate \u0394i,j with \na linear regression model. Thereby we consider that \u0394i,j is a function of household-\nlevel characteristics \n\u2026\n\u2026 \u2026\nx\nx\nx\n,\n,\n,\n,\n,\ni\nk i\nk i\n1,\n,\n, , such that:\n\u0394\n\u03b2\n\u03b2\n\u03b5\n=\n+ \u2026 +\n+\nx\nx\n(2)\ni j\ni\nj\nk i\nk j\ni j\n,\n1,\n1,\n,\n,\n,\nwhere \u03b2k,j holds the incremental average percentage points by which the network \ncosts change when alternative scenario j is applied instead of the reference scenario, \nwhen household characteristic k increases by one unit. \u03b5i,j references the error term. \nAll regressions presented in this study estimate equation (2).\nConsidering that the alternative scenarios as used in our study are correlated, \nfor example, all \u2018f\u2019 scenarios rely on fixed charges, simultaneous estimation of all \n11 alternative scenarios suggests itself. We pool the 11 resulting equations (one per \nalternative scenario) and estimate the resulting system by the seemingly unrelated \nregression model47. As we rely on exactly the same set of household characteristics \nfor explaining the deterministic portion in equation (2), results are identical to \nthe ordinary least squares model. Whenever a variable is included in one of the \nregressions for which missing values have been imputed, a respective dummy \nvariable is included to test whether the imputation has significantly affected the \nanalyses. None of these dummies\u2019 coefficients has a significant opposing effect \ncompared with the main variable\u2019s coefficient, so we conclude that our analyses do \nnot suffer from bias caused by the imputations.\nScenarios. To cover the range of potential network tariff schemes, we designed a \ntotal number of 11 scenarios.\nFirst, we designed one respective tariff scenario recovering the network costs \nthrough only one of the three components\u2014volume, fixed charge and measured \npeak load (average and maximum).\nScenario f100 is a 100% fixed charge. Thereby, f100 represents a flat charge for \nall households, which is \u20ac\u200b136,209.10\u2009/\u2009765\u2009households\u2009=\u200b\u2009\u20ac\u200b178.05\u2009per household per \nyear for the full sample in our study.\nScenario pa100 represents a scheme charging for measured peak load only. In \nthis scenario, the definition of kW peak load follows the Austrian tariff structure \nin 201648, where a so-called smart meter tariff was included for testing only (in the \nresidential sector). There, kW peak load as relevant for billing is not defined as  \nthe one maximum load out of the 35,040 metered load values during one year  \nper Austrian meter. Peak load as relevant for setting a household\u2019s peak charge  \nis defined as the average of the 12 monthly peak loads during the respective  \nyear. Scenario pa100 for the full sample analysis therefore sets  \n\u20ac\u200b136,209.10\u2009/\u20093,485.59\u2009kW\u2009total\u2009=\u200b\u2009\u20ac\u200b39.07\u2009per\u2009kW of billing relevant peak load, \nwhere kW total is the sum over all 765 corresponding peak-load values.\nScenario pm100 also represents a scheme charging for peak demand only, but \ninstead of averaging, the highest of the 12 monthly peaks is applied, such that it sets \n\u20ac\u200b136,209.10\u2009/\u20094,603.3\u2009kW\u2009total\u2009=\u200b\u2009\u20ac\u200b29.59\u2009per kW of billing relevant peak load, where \nkW total is the sum over all 765 corresponding peak-load values.\nScenario e100 is a fully volumetric tariff and includes only a payment  \nper unit of consumed energy. Thereby it is \u20ac\u200b136,209.10\u2009/\u20092,691,272\u2009kWh\u2009total\u2009=\u200b\u2009 \n\u20ac\u200b0.0506\u2009per\u2009kWh consumed during the respective year, where kWh total is  \nthe aggregated electricity consumption of all 765 households.\nNext, we define five scenarios each representing a hybrid of two of the \ncandidate tariff components, fixed, peak-load and volumetric charges:\nScenario f50/e50 puts 50% of the weight on the fixed charge and 50% on  \nthe consumed volume. Thereby, \u20ac\u200b136,209.10\u2009\u00d7\u200b\u20090.5\u2009=\u200b\u2009\u20ac\u200b68,104.55 are recovered  \nby the fixed component, and exactly the same amount comes from the  \nvolumetric component. The fixed component of f50/e50 is therefore given by  \n\u20ac\u200b68,104.55\u2009/\u2009765\u2009households\u2009=\u200b\u2009\u20ac\u200b89.02\u2009per household per year. The volumetric \ncomponent of scenario f50/e50 results from distributing the respective  \nquantity of \u20ac\u200b68,104.55 among the total number of consumed kWh, that is,  \n\u20ac\u200b68,104.55\u2009/\u20092,691,272\u2009kWh\u2009total\u2009=\u200b\u2009\u20ac\u200b0.025\u2009per\u2009kWh consumed.\nScenario f50/pa50 represents a scheme with 50% of the costs recovered  \nfrom a fixed charge and 50% from a measured peak charge. The fixed charge  \nper household is the same as in the f50/e50, the peak charge is defined as  \n\u20ac\u200b68,104.55\u2009/\u20093,485.59\u2009kW\u2009total\u2009=\u200b\u2009\u20ac\u200b19.53\u2009per\u2009kW of billing relevant peak load,  \nwhere kW total is the sum over all 765 corresponding peak-load values  \n(average definition).\nScenario f50/pm50 the peak charges are calculated as \u20ac\u200b\n68,104.55\u2009/\u20094.603.3\u2009kW\u2009total\u2009=\u200b\u2009\u20ac\u200b14.79\u2009per kW of billing relevant peak load (based on \nthe one maximum peak definition).\nScenarios pa50/e50 and pm50/e50 are each a combination where 50% of weight \nis put on the consumed volume of energy and 50% of the measured peak demand, \naverage or maximum definition, respectively.\nIn addition, we also test two scenarios combining all three candidate tariff \ncomponents:\nScenario f/pa/e imposes a fixed charged of \u20ac\u200b24.60 as found in the Austrian \ntariff as applied in 2016, and splits the remaining quantity in equal portions. \nRevenues collected from the fixed charge therefore result in a total of  \n\u20ac\u200b24.60\u2009\u00d7\u200b\u2009765\u2009=\u200b\u2009\u20ac\u200b18,819 for our full sample, which we subtract from the \u20ac\u200b136,209.10 \nto calculate the shares that have to be recovered by peak and energy components: \n(\u20ac\u200b136,209.10\u2009\u2212\u200b\u2009\u20ac\u200b18,819)\u2009\u00d7\u200b\u20090.5\u2009=\u200b\u2009\u20ac\u200b58,695.05. The peak component is calculated as \u20ac\u200b\n58,695.05\u2009/\u20093,485.59\u2009kW\u2009total\u2009=\u200b\u2009\u20ac\u200b16.83\u2009per kW peak and the energy component is \nequal to \u20ac\u200b58,695.05\u2009/\u20092,691,272\u2009kWh\u2009total\u2009=\u200b\u2009\u20ac\u200b0.022 for each kWh consumed.\nScenario f/pm/e fixed and energy charges are the same as in the previous \nscenario and the peak charge is calculated with the maximum peak definition \ninstead of the averaged definition.\nWe compute the network charges for every household under each of the \nscenarios for the full sample as well as for the subsample of 406 households, for \nwhich the sum of network charges accumulates to \u20ac\u200b74,794.28 in the reference \nscenario.\nEthics statement. The survey data were collected by Energy Institute at the \nJohannes Kepler University Linz, following high European Union standards of data \nprotection and voluntary study participation. The methodology used in this paper \ndoes not require institutional ethical approval according to the guidelines set out \nby the Energy Institute at the Johannes Kepler University Linz. Confidentiality and \nanonymity of participants was ensured and informed written consent was obtained \nfrom all the interviewees.\nData availability. The data that support the plots within this paper and other \nfindings of this study are available from the corresponding author upon reasonable \nrequest.\nReceived: 5 June 2017; Accepted: 26 January 2018;  \nPublished online: 12 March 2018\nReferences\n\t1.\t P\u00e9rez-Arriaga, I. J. & Smeers, Y. in Transport Pricing of Electricity Networks \nCh. 7 (Springer, Boston, MA, 2003).\n\t2.\t Jenkins, J. D. & P\u00e9rez-Arriaga, I. J. Improved regulatory approaches for the \nremuneration of electricity distribution utilities with high penetrations of \ndistributed energy resources. Energy J. 38, 63\u201391 (2017).\n\t3.\t Schreiber, M., Wainstein, M. E., Hochloff, P. & Dargaville, R. Flexible \nelectricity tariffs: power and energy price signals designed for a smarter grid. \nEnergy 93, 2568\u20132581 (2015).\n\t4.\t Linvill, C., Shenot, J. & Lazar, J. Designing Distributed Generation Tariffs Well: \nFair Compensation in a Time of Transition (Regulatory Assistance Project, \n2013).\n\t5.\t Rodr\u00edguez Ortega, M. P., P\u00e9rez-Arriaga, J. I., Abbad, J. R. & Gonz\u00e1lez, J. P. \nDistribution network tariffs: A closed question? Energy Policy 36,  \n1712\u20131725 (2008).\n\t6.\t Mountain, B. The rise of distributed generation and storage in South \nAustralia. Energy Spectr. 549, 17\u201319 (2016).\n\t7.\t Bronski, P. et al. The Economics of Grid Defection (Rocky Mountain Institute, \n2014); http://go.nature.com/ZXC65t\n\t8.\t Nijhuis, M., Gibescu, M. & Cobben, J. F. G. Analysis of reflectivity and \npredictability of electricity network tariff structures for household consumers. \nEnergy Policy 109, 631\u2013641 (2017).\n\t9.\t Schill, W. P., Zerrahn, A. & Kunz, F. Prosumage of solar electricity: pros, \ncons, and the system perspective. Econ. Energy Environ. Policy 6,  \n7\u201331 (2017).\n\t10.\tPicciariello, A., Vergara, C., Reneses, J., Fr\u00edas, P. & S\u00f6der, L. Electricity \ndistribution tariffs and distributed generation: quantifying cross-subsidies \nfrom consumers to prosumers. Util. Policy 37, 23\u201333 (2015).\n\t11.\tCossent, R., G\u00f3mez, T. & Fr\u00edas, P. Towards a future with large penetration of \ndistributed generation: Is the current regulation of electricity distribution \nready? Regulatory recommendations under a European perspective.  \nEnergy Policy 37, 1145\u20131155 (2009).\n\t12.\tJohnson, E., Beppler, R., Blackburn, C., Staver, B., Brown, M. & Matisoff, D. \nPeak shifting and cross-class subsidization: the impacts of solar PV on \nchanges in electricity costs. Energy Policy 106, 436\u2013444 (2017).\n\t13.\tEid, C., Guill\u00e9n, J. R., Mar\u00edn, P. F. & Hakvoort, R. The economic effect of \nelectricity net-metering with solar PV: consequences for network cost \nrecovery, cross subsidies and policy objectives. Energy Policy 75,  \n244\u2013254 (2014).\n\t14.\tP\u00e9rez-Arriaga, I. J., Schwene, S., Ruester, S., Batlle, C. & Glachan, J. M. From \nDistribution Networks to Smart Distribution Systems: Rethinking the Regulation \nof European Electricity DSOs Final Report (THINK, 2013).\n\t15.\tChapman, A. J., McLellan, B. & Tezuka, T. Residential solar PV policy:  \nan analysis of impacts, successes and failures in the Australian case.  \nRenew. Energy 86, 1265\u20131279 (2016).\n\t16.\tMacintosh, A. & Wilkinson, D. Searching for public benefits in solar \nsubsidies: a case study on the Australian government\u2019s residential photovoltaic \nrebate program. Energy Policy 39, 3199\u20133209 (2011).\n\t17.\tCastaneda, M., Jimenez, M., Zapata, S., Franco, C. J. & Dyner, I. Myths and \nfacts of the utility death spiral. Energy Policy 110, 105\u2013116 (2017).\nNature Energy | VOL 3 | APRIL 2018 | 317\u2013325 | www.nature.com/natureenergy\n324\n\n\u00a9 2018 Macmillan Publishers Limited, part of Springer Nature. All rights reserved.\nArticles\nNaTUre EnerGy\n\t18.\tLaws, N. D., Epps, B. P., Peterson, S. O., Laser, M. S. & Wanjiru, G. K. On the \nutility death spiral and the impact of utility rate structures on the adoption of \nresidential solar photovoltaics and energy storage. Appl. Energy 185, 627\u2013641 \n(2017).\n\t19.\tKind, P. Disruptive Challenges: Financial Implications and Strategic responses \nto a Changing Retail Electric Business (Edison Electric Institute, 2013).\n\t20.\tSeverance, C. A. A Practical, affordable (and least business risk) plan to \nachieve \u201c80% clean electricity\u201d by 2035. Electr. J. 24, 8\u201326 (2011).\n\t21.\tMuaafa, M., Adjali, I., Bean, P., Fuentes, R., Kimbrough, S. O. & Murphy, F. \nH. Can adoption of rooftop solar panels trigger a utility death spiral? A tale \nof two US cities. Energy Res. Social. Sci. 34, 154\u2013162 (2017).\n\t22.\tCostello, K. W. & Hemphill, R. C. Electric utilities\u2019 \u2018death spiral\u2019: hyperbole or \nreality? Distribution pricing: theoretical principles and practical approaches. \nElectr. J. 27, 7\u201326 (2014).\n\t23.\tPassey, R., Haghdadi, N., Bruce, A. & MacGill, I. Designing more cost \nreflective electricity network tariffs with demand charges. Energy Policy 109, \n642\u2013649 (2017).\n\t24.\tBorenstein, S. The economics of fixed cost recovery by utilities. Electr. J. 29, \n5\u201312 (2016).\n\t25.\tMcLaren, J., Davidson, C., Miller, J. & Bird, L. Impact of rate design \nalternatives on residential solar customer bills: increased fixed charges, \nminimum bills and demand-based rates. Electr. J. 28, 43\u201358 (2015).\n\t26.\tJargstorf, J., Kessels, K. & Belmans, R. Capacity-based grid fees for residential \ncustomers. In 10th International Conference on the European Energy Market \n(EEM) 1\u20138 (IEEE, 2013); https://doi.org/10.1109/EEM.2013.6607294\n\t27.\tHledik, R. Rediscovering residential demand charges. Electr. J. 27,  \n82\u201396 (2014).\n\t28.\tReneses, J. & Rodr\u00edguez Ortega, M. P. Distribution pricing: theoretical \nprinciples and practical approaches. IET Gener. Transm. Dis. 8, 1645\u20131655 \n(2014).\n\t29.\tSimshauser, P. Distribution network prices and solar PV: resolving rate \ninstability and wealth transfers through demand tariffs. Energy Econ. 54, \n108\u2013122 (2016).\n\t30.\tStrielkowski, W., \u0160treimikien\u0117, D. & Bilan, Y. Network charging and \nresidential tariffs: a case of household photovoltaics in the United Kingdom. \nRenew. Sustain. Energy Rev. 77, 461\u2013473 (2017).\n\t31.\tHledik, R. & Greenstein, G. The distributional impacts of residential demand \ncharges. Electr. J. 29, 33\u201341 (2016).\n\t32.\tRubin, S. Moving toward demand-based residential rates. Electr. J. 28,  \n63\u201371 (2015).\n\t33.\tBlank, L. & Gegax, D. Residential winners and losers behind the energy \nversus customer charge debate. Electr. J. 27, 31\u201339 (2014).\n\t34.\tFaruqui, A. The ethics of dynamic pricing. Electr. J. 23, 13\u201327 (2010).\n\t35.\tBonright, J.C. Principles of Public Utility Rates (Columbia University Press, \nNew York, 1961).\n\t36.\tBrown, T., Faruqui, A. & Grausz, L. Efficient tariff structures for distribution \nnetwork services. Econ. Anal. Policy 48, 139\u2013149 (2015).\n\t37.\tPicciariello, A., Reneses, J., Fr\u00edas, P. & S\u00f6der, L. Distributed generation and \ndistribution pricing: Why do we need new tariff design methodologies?  \nElectr. Power Syst. Res. 119, 370\u2013376 (2015).\n\t38.\tSakhrani, V. & Parsons, J. Electricity Network Tariff Architectures: A \nComparison of Four OECD Countries No. 10\u2013008 (Center for Energy and \nEnvironmental Policy Research, 2010).\n\t39.\tAdapting Distribution Network Tariffs to a Decentralised Energy Future (EDSO \nfor Smart Grids, 2015); http://www.edsoforsmartgrids.eu/wp-content/\nuploads/151014_Adapting-distribution-network-tariffs-to-a-decentralised-\nenergy-future_final.pdf\n\t40.\tNetwork Tariffs (EURELECTRIC, 2016); http://www.eurelectric.org/\nmedia/268408/network_tariffs__position_paper_final_as-2016-030-0149-01-e.pdf\n\t41.\tKlaassen, E. A. M., Kobus, C. B. A., Frunt, J. & Slootweg, J. G. Responsiveness \nof residential electricity demand to dynamic tariffs: experiences from a large \nfield test in Netherlands. Appl. Energy 183, 1065\u20131074 (2016).\n\t42.\tAyres, I., Raseman, S. & Shih, A. Evidence from two large field experiments \nthat peer comparison feedback can reduce residential energy usage. J. Law \nEcon. Organ. 29, 992\u20131022 (2013).\n\t43.\tZ\u00e4hlwerte, Datenformate und standardisierte Lastprofile, Sonstige Marktregeln \nStrom\u2014Version 3.3, Kapitel 6, 1\u201373 (Energie-Control Austria, 2011).\n\t44.\tFares, R. L. & Webber, M. E. The impacts of storing solar energy in the home \nto reduce reliance on the utility. Nat. Energy 2, 17001 (2017).\n\t45.\tSommerfeld, J., Buys, L., Mengersen, K. & Vine, D. Influence of demographic \nvariables on uptake of domestic solar photovoltaic technology. Renew. Sustain. \nEnergy Rev. 67, 315\u2013323 (2017).\n\t46.\tDe Groote, O., Pepermans, G. & Verboven, F. Heterogeneity in the adoption \nof photovoltaic systems in Flanders. Energy Econ. 59, 45\u201357 (2016).\n\t47.\tZellner, A. An efficient method of estimating seemingly unrelated regressions \nand tests for aggregation bias. J. Am. Stat. Assoc. 57, 348\u2013368 (1962).\n\t48.\tVerordnung der Regulierungskommission der E-Control, mit der die Entgelte f\u00fcr \ndie Systemnutzung bestimmt werden (Systemnutzungsentgelte-Verordnung \n2012 in der Fassung der Novelle 2016) SNE-VO 2012 idF Novelle 2016, 1\u201312 \n(Energie-Control Austria, 2016); https://www.e-control.at/\ndocuments/20903/388512/SNE-2012-idF-Novelle-2016-konsolidiert.pdf/\ncef56a01-20b8-4b7c-8cb3-9429a6a56271\nAcknowledgements\nThis research was done as a part of PEAKapp project that has received funding under the \nEuropean Union\u2019s Horizon 2020 research and innovation programme, grant agreement \nNo. 695945 (http://www.peakapp.eu/). A.K. received additional funding from the \nAustrian Ministry for Transport, Innovation and Technology (No. 848114). D.E. and C.F. \ngratefully acknowledge the financial support by the Austrian Federal Ministry of Science, \nResearch and Economy, the Austrian National Foundation for Research, Technology \nand Development and the Federal State of Salzburg. This research was enriched through \ndiscussions about international network regulation with B. Mountain, R. Muruais,  \nJ. Cohen and D. Pezenka.\nAuthor contributions\nA.K. and J.R. were primarily responsible for the creation and implementation of the \nsurvey instrument. D.E. and C.F. were primarily responsible for the creation and \nmanagement of the dataset, including load profiles. V.A. and J.R. mainly contributed  \nto data analysis. All authors contributed to the writing of the paper with V.A., A.K. and \nJ.R. as the primary authors.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-018-0105-4.\nReprints and permissions information is available at www.nature.com/reprints.\nCorrespondence and requests for materials should be addressed to J.R.\nPublisher\u2019s note: Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\nNature Energy | VOL 3 | APRIL 2018 | 317\u2013325 | www.nature.com/natureenergy\n325\n\n\n    Scientific Research Findings:", "answer": "We modelled the effects of 11 network tariff scenarios on household budgets using real load profiles from 765 households collected between April 2010 and March 2011 in Austria. We found that, for tariffs emphasizing peak charges, the predictability of annual network costs for households is potentially low, because the costs are driven by only a small number of peak consumption values. Moreover, the impact of peak charges on household budgets can be disruptive. More specifically, we found up to a 500% increase in network charges for some households under the extreme tariff scheme tested. Although the potential size of the effects is dramatic, one might expect that households would adapt to the new tariff schemes after a while. We were unable to test such a response as we only used historic data where no alternative tariffs with price signals were applied.", "id": 34}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Analysis\nhttps://doi.org/10.1038/s41560-018-0256-3\n1Laboratory for Energy and Nanoscience, Khalifa University of Science and Technology, Abu Dhabi, United Arab Emirates. 2Dubai Electricity and Water \nAuthority, Dubai, United Arab Emirates. 3Arctic Renewable Energy Center (ARC), Department of Physics and Technology, University of Troms\u00f8,  \nTroms\u00f8, Norway. *e-mail: matteo.chiesa@ku.ac.ae\nA\nfter decades on the fringes, solar energy has arrived as a \nmajor player in the electricity generation sector. While just \nfive years ago, conventional wisdom grouped solar with the \nimpractical and subsidy-dependent \u2018alternative\u2019 energy technolo-\ngies, today every few months there are announcements of new large-\nscale solar projects around the globe promising electricity prices \nequal to or lower than what can be achieved with fossil fuels1,2, as \nshown in Fig. 1a. While the well-publicized steep learning curve of \nphotovoltaic (PV) modules3 has played a large role in bringing us to \nthis point, other factors are now beginning to drive price reductions \nas modules have dropped to just over a third of total system prices, \nwith the remainder represented by inverters, sun trackers (if used) \nand other components for balance of system, as well as the cost of \nlabour and various \u2018soft costs\u2019 related to regulation, taxes, developer \nexpenses and grid connection, accounted for in Fig. 1b based on data \nfrom the United States4,5. Particularly noteworthy and illustrative of \nthis trend have been a number of announcements of low-priced \nutility-scale PV projects from the Middle East, especially the United \nArab Emirates and Saudi Arabia6\u20139, where power purchase agree-\nment (PPA) auction bids below 3\u00a2\u2009kWh\u22121 have become the norm. \nIf sustainable, these cheap projects would represent the achieve-\nment of the US SunShot pricing goals for 203010 more than a decade \nahead of schedule. As these targets represent the expectation that \nunsubsidized solar overtakes fossil fuels on cost, this development \nwould have significant implications for the world\u2019s energy systems. It \nwould make solar the economic favourite for new-generation capac-\nity, accelerating the transition to a renewable-based energy future.\nHowever, the speed at which prices have fallen has prompted \nsuspicion that these prices are attained only via hidden government \nsubsidies; that they are not intended to make money at all but are \na consequence of a \u2018race to the bottom\u2019 driven by the PPA auction \nmodel; or are \u2018loss leaders\u2019, strategic moves by large forward-look-\ning companies seeking to establish a foothold in a young industry \npoised to grow11\u201313. Despite these concerns, low-priced PV projects \nhave spread across the world in the intervening two years14\u201316.\nUnderstanding the extent to which these prices are real, sus-\ntainable and reproducible is critical for setting energy strategies to \nencourage the continued rapid PV deployment that will be needed \nover the next decade and beyond to achieve energy system decar-\nbonization at the rate required by climate agreement targets17. In this \nanalysis, we draw on a range of publicly available data on PV system \ncosts and performance, along with local knowledge of the region, \nwith the aim of answering the questions: how did electricity from \nPV get so cheap so fast? Can these ultralow prices be sustained and \nextended to other markets? And how much lower could they get?\nReasons for the decline in PV electricity prices\nThe recent trend in solar electricity prices has been much more \naggressive than that of hardware prices due to a number of factors. \nOne part of the answer is better exploitation of the solar resource\u2014\nthe world\u2019s sunniest areas are increasingly being developed as PV \nprojects become less dependent on state subsidies. Other plant opti-\nmizations to maximize energy yield have become widely adopted, \nincluding single-axis tracking and high inverter-loading ratios (d.c. \narray capacity to a.c. inverter capacity) of ~1.3 for both tracking and \nnon-tracking systems. These implementations, coupled with fall-\ning hardware prices and tax incentives, were credited with bringing \nthe levelized cost of solar electricity in sunny regions of the United \nStates to around 5\u00a2\u2009kWh\u22121 in 201518.\nEven in this context of rapid pricing declines, the arrival of the \nUnited Arab Emirates, and Saudi Arabia soon afterwards, as leaders \nin cheap solar electricity surprised observers with the sheer scale of \nthe price decrease in a region with little historical interest in renew-\nable energy. A closer look at the details of these projects begins to \nreveal the interplay of global and local conditions that lead to the \nobserved prices. The projects appear to have realized large reduc-\ntions in both capital and operating expenses by leveraging on a \nnumber of factors. These include forward-bidding of expected lower \nfuture hardware prices, which up to this point have been borne  \nout; low construction and operation and maintenance (O&M) \nlabour costs that are typical of the Gulf region; scaling up the plants \nfrom the 10\u2013100\u2009MW level that was prevalent until recently to the \ngigawatt scale; extended PPA term to 25 years (relative to the then-\nstandard 20 years18); and favourable financial terms. Moreover,  \nEvaluating the factors that led to low-priced solar \nelectricity projects in the Middle East\nHarry\u00a0Apostoleris1, Sgouris\u00a0Sgouridis\u200a \u200a1, Marco\u00a0Stefancich2 and Matteo\u00a0Chiesa1,3*\nThe past few years have seen the rise of large-scale, low-priced solar energy projects around the world. Oil-producing coun-\ntries in the Middle East, in particularly the United Arab Emirates and Saudi Arabia, have become unexpected leaders in this \nmovement with record-low power purchase agreement prices, below 3\u00a2\u2009kWh\u22121, for a number of new photovoltaic installations, \nbeating the cost of fossil fuel generation. In this Analysis, we bring together technical, economic and financial information from \nglobal and local sources to study whether these prices can be replicated elsewhere and further reduced. We find that hardware \ncosts, cost of labour, favourable cost of capital, low taxes and low, but positive, profit margins contribute to the reduction in \ncosts. Reduced hardware prices contributed the most and also led to further reduction in cost of capital. We demonstrate how \nsimilar costs can be and have been achieved in other markets.\nNature Energy | VOL 3 | DECEMBER 2018 | 1109\u20131114 | www.nature.com/natureenergy\n1109\n\nAnalysis\nNature Energy\nit appears that developers did not have to account for land costs that \ncould be around US$5,000 per acre in the United States19. The verti-\ncally integrated state-owned utilities are also involved as partners, \nexpected to take majority ownership of the venture on completion. \nAs such, they are believed to take on the costs for grid interconnec-\ntion and access roads as infrastructure development projects13,20,21.\nProject financing packages can be fairly complex but can be \noutlined as featuring high debt-to-equity ratios (70\u201380%), low \ninterest rates (reported starting values for interest range between \n120 and 180 basis points over the London Interbank rate (LIBOR)) \nand significant opportunities for refinancing over the project life-\ntime21. The establishment of these terms by the consortium of local \nbanks that financed Dubai\u2019s 200\u2009MW project in 2015 certainly \nprovided encouragement for international consortia to offer com-\nparable packages for the subsequent, larger projects. In contrast, \nutility-scale PV projects being developed at the same time in the \nUnited States could expect debt ratios in the range of 40\u201360%, with \ninterest rates of 4\u20135%, and return on equity expectations around \n10%22,23. In Table 1 we list the key financial parameters for the four \nrecord-setting projects in the United Arab Emirates and Saudi \nArabia, as thoroughly as they can be reconstructed from a collec-\ntion of public announcements.\nThe path to 3\u00a2\u2009kWh\u22121\nTo investigate the system configurations and cost structures \nrequired to reach 3\u00a2\u2009kWh\u22121, we first establish a baseline system price \nand LCOE breakdown for 2016, just before Dubai\u2019s Mohammed bin \nRashid Al Maktoum (MBR) phase III announcement, based on a \nnumber of sources5,23\u201326, shown in Fig. 2.\nCapEx depends heavily on the price of hardware, which accounts \nfor the majority of the total installed cost, with construction labour \nbeing the second-largest contributor. CapEx represents about 50% \nof the LCOE in the baseline model, with operation and maintenance \n(O&M) and financing costs representing the balance. The baseline \nfinancial model assumes a debt fraction of 60% with an interest \nrate of 5% and a return on equity of 10%, for a total weighted aver-\nage cost of capital of 7%. The CapEx includes a profit margin for \nthe project developer, while the construction contractor\u2019s profit is \nimplicitly included in the contractor markup. The LCOE represents \nthe cost to the project developer of building and operating the sys-\ntem, and of satisfying the debt and equity financers of the project \nwith their required returns, and hence represents a minimum sus-\ntainable price for the sale of electricity. The baseline assumptions \nare listed in the first column of Table 2.\nWe subsequently chart a reasonable path to a sub-3\u00a2\u2009kWh\u22121 \nLCOE by applying a series of successive cost reductions, each  \njustified on the basis of a collection of publicly available industry \nstatistics and local knowledge. The full set of model inputs is listed \nin Table 2. The key assumptions are listed below.\nContinued future declines in the cost of hardware\nSince in this analysis we are actually looking backwards to projec-\ntions that were made in 2016, we consider what the project\u2019s hard-\nware cost would be, assuming that the developers guessed future \npricing trends more or less correctly. At the time of the current \nwriting, both the Dubai and Abu Dhabi projects are in the middle \nof their construction timeline, with hardware procurement under-\nway. Therefore, we use current prices as a reasonable proxy for the \naverage cost of hardware for the systems. Additional discussion of \nhardware pricing projects can be found in Supplementary Note 1, \nSupplementary Table 1 and Supplementary Fig. 1.\nCost of labour is very low due to the connection between labour \nmarkets in the Gulf and the low-cost Indian subcontinent. With \nlocal contractors assuming most construction duties, and reported \nwages for construction work and even some skilled trades reported \nas less than US$5\u2009h\u22121 by local sources27,28, we believe that a reduction \nin labour costs of 50% relative to the US benchmark in refs\u20094,25 is a \nreasonable and perhaps even a conservative estimate (further details \nprovided in the Supplementary Note 2 and Supplementary Table 2). \nThis low cost of labour is also the primary driver of reductions in \na\nb\n2.5\n0\n0.5\n1.0\n1.5\n3.5\n4.5\n2.0\n3.0\n4.0\n5.0\n2010\n2011\n2012\n2013\nYear\nDate\n2014\n2015\n2016\n2017\nSystem or component price\n(US$ Wdc\u20131)\nInstalled system price\nHard cost\nModule\nSouth Africa\nFrance\nPeru\nIndia\nJordan\nZambia\nBrazil\nUSA\nDubai\nGermany\nMexico\nChile\nArgentina\nMorocco\n200\nChina\nJan. 2010\n50\n100\n150\n250\n300\n350\nJan. 2011\nJan. 2012\nJan. 2013\nJan. 2014\nJan. 2015\nJan. 2016\nJan. 2017\nElectricity price (US$ MWh\u20131)\n0\nJan. 2018\nSaudi Arabia\nAbu Dhabi\nFig. 1 | The reduction of PV system and electricity prices around the world. a, Solar auction history of PPA prices (source: International Renewable Energy \nAgency1). b, Module price evolution, hardware price and utility-scale system price\u2014as a bar chart based on ref.\u20094; numbers for components and total \nsystem price are from ref.\u20095 (except 2017).\nNature Energy | VOL 3 | DECEMBER 2018 | 1109\u20131114 | www.nature.com/natureenergy\n1110\n\nAnalysis\nNature Energy\noperation and maintenance costs, which we reduce to US$10,000 \nMW\u22121 yr\u22121, based on local assessments by industry insiders29, esca-\nlating over the plant lifetime to US$15,000 to account for increased \nservice requirements and component replacements.\nFinancial parameters are based on the public reports summa-\nrized in Table 1 and include a debt fraction of 80% and an inter-\nest rate of 3.5%, with a return on equity of 10%. Some observers \nhave speculated that money was essentially given out for free to \nthe project developers, to achieve headline-grabbing pricing mile-\nstones13. Our analysis shows that such extreme assumptions are not \nnecessary to achieve the reported prices under today\u2019s conditions. \nFurthermore, while early projects were financed principally by local \nbanks, encouraging some doubt as to the model\u2019s global sustainabil-\nity, the growing role of international financial institutions suggests \nthat confidence in solar energy in the broader financial sector may \nbe growing to the point where similarly favourable financial terms \nmay be reached in other markets. In addition, the return on equity \ncan be maintained at a globally competitive 10%.\nTax treatment\nWhile the baseline includes a 5% sales tax on hardware, the low-\ncost model removes this in keeping with the tax-free conditions of \nthe United Arab Emirates at the time these bids were offered. The \n2018 introduction of a 5% value-added tax may add this cost com-\nponent back in future projects. The LCOE model does not include \nincome taxes, which often will primarily be felt (due to the narrow \nprofit margins of these projects) through the effective reduction of \nthe cost of debt, in countries where debt interest payments are tax-\ndeductible. Hence, the net impact of income taxes is often to reduce \nthe effective cost of the project.\nOverhead and profit margins\nThe CapEx includes two profit margins: one for the developers, \nand one included in the contractor overhead. To estimate the over-\nhead\u2009+\u200b\u2009profit component for the large plant in the low-cost model, \nwe extrapolate the assumption that contractor overhead can be \nreduced to some degree (in percentage terms) with plant size25. \nTable 1 | Financial parameters for low-priced PV projects\nMBR phase II\nMBR phase III\nSweihan\nSakaka\nDevelopers\nACWA Power  \n(Saudi Arabia)\u2009+\u200b\u2009TSK (Spain)\nMasdar (United Arab Emirates)\u2009+\u200b\u2009 \nALJ (Saudi Arabia)\u2009+\u200b\u2009EDF (France)\nMarubeni (Japan)\u2009+\u200b\u2009Jinko \nSolar (China)\nACWA Power\nDate of close\nJuly 2015\nJune 2017\nMay 2017\nFebruary 2018\nTotal debt in US$ million\n34433,a\n65521\n65035\n260\nDebt fraction\n86%\n70%\n75%\n80% (estimated)\nDebt interest rate (basis \npoints over LIBOR)\n180 (average)\n175 (starting)\n120 (starting)\nUnknown\nDebt interest rate \n(calculated)\n2.6%b\n3.6%\n2.9%\nN/A\nPPA price (\u00a2\u2009kWh\u22121)\n5.84\n2.99\n2.94\n2.34\nFrom left: 200\u2009MW phase 2 of Dubai\u2019s MBR solar park; 800\u2009MW phase 3; Abu Dhabi\u2019s 1,177\u2009MW Sweihan solar park; and the inaugural plant of the Sakaka solar park in northern Saudi Arabia. Values are \nbased on reporting by IJGlobal except where otherwise noted. aSources differ, giving a range of total project costs between US$326 million34 and US$400 million33, with some specifying a 4% debt  \ninterest rate33. The tabulated parameters are those that we judge most likely to be accurate on the basis of some educated guesswork and back-calculation. bFinancing for MBR phase II was provided  \nby a consortium of local banks. Subsequent projects were funded by international consortia, dominated by institutions from the developing partners\u2019 home countries but including institutions from  \nother countries.\nEPC hard\nConstruction labour\nPermitting\nSales tax\nLand and prep.\nContingency\nEPC soft\nGrid tie\n1.0\n0\n0.2\n0.4\n0.6\n1.4\n0.8\n1.2\n1.6\n2.0\na\nb\nSystem price (US$ W\u20131)\nLCOE (\u00d710\u20132 US$ kWh\u20131)\n1.8\nHardware (via EPC contractor)\nConstruction labour\n(via EPC contractor)\nConstruction, developer-borne\nDebt service\nSales tax\nEPC (markup)\nOperation\nReturn on equity\n5\n0\n1\n2\n3\n7\n4\n6\n8\nDeveloper profit\nFig. 2 | Bottom-up model for PV costs. a,b, Capital expenditures in \n\u2212\nS\nU $ Wa c. .\n1  (a) and levelized cost of electricity in US$\u2009kWh\u22121 (b) in the baseline model \nof a 100\u2009MW plant with a 20-year PPA, based on 2016 market data for component costs, free of income tax and incentives. Cost components are \nsummarized and attributed in Table 1.\nNature Energy | VOL 3 | DECEMBER 2018 | 1109\u20131114 | www.nature.com/natureenergy\n1111\n\nAnalysis\nNature Energy\nThe developer profit margin is typically low4, and can be reduced \nfurther for a large project. As a further justification of our scaling \nassumptions, we note that in several of these projects, the developer \neither scaled up the plant to achieve the target price, or offered to \nbuild a larger plant at a lower PPA price, giving some indication of \nthe gains that developers see in economies of scale. The assumption \nof fairly slim profit margins can be justified by their large size as \nwell as by the fact that these are \u2018trophy\u2019 projects in a hot emerging \nmarket that will form a cornerstone of their developers\u2019 reputation. \nHowever it is worth noting that we assume profit margins that are \nslim, not negative\u2014we do not see a reason to consider these proj-\nects as \u2018loss leaders\u2019 that sell below cost purely for the sake of estab-\nlishing their developer in the market.\nImplementing the full set of assumptions laid out in the second \ncolumn of Table 2, we find that LCOE below 3\u00a2\u2009kWh\u22121 is achieved \nnaturally without further manipulation as shown in Fig. 3. This \nreduction is driven by the combined impact of low cost of financ-\ning and reduced capital expenditures, shown in Fig. 4. We note that \nthere is the possibility of relaxing some of the assumptions while \nstill achieving the price target. For example, we could remove the \nassumption that the utility bears the costs of interconnection, which \nhas never been publicly confirmed, and still reach the target price; \nor, we could account for the fact that PV hardware may in general be \nsubjected to some form of import duties (say 5% applicable to half \nof the CapEx, which would boost the LCOE to 2.91\u00a2\u2009kWh\u22121 in our \nmodel). If, on the other hand, we are to tighten these assumptions, \nthe 2.34\u00a2\u2009kWh\u22121 price of the Sakaka project can be reproduced, as \ndescribed in Supplementary Note 3 and Supplementary Fig. 3.\nDiscussion, outlook and lessons\nWe have shown here that, at least up to the level of about 3\u00a2\u2009kWh\u22121, \nrecently observed low solar electricity prices can be profitably real-\nized under reasonable assumptions for the region. The key ques-\ntions remaining are whether these conditions are sustainable, and \nhow these low prices may be replicated in other markets.\nThe contribution of declining hardware prices, while not  \nsolely responsible for the record bids, is nonetheless important. It is  \ntherefore worthwhile to comment briefly on how sustainable the \ncurrent situation of declining module prices specifically seems to \nbe, although a deep study is beyond the scope of this analysis. We \npoint, by way of reassurance, to the recent decision in China to roll \nback some support for its solar industry, which appears to have actu-\nally caused module prices to decline further due to falling domestic \ndemand (see more in Supplementary Discussion). Therefore, fears \nof a sudden spike in module prices that would threaten near-term \ninvestment in PV capacity seem, for now, unlikely to materialize.\nThere seem to be two areas where direct actions by government \nhave made a significant difference. The first is in the removal of \nsome indirect costs, which may be criticized as not applicable to \nother markets. As a counterargument, we note the recent spate of \nsub-3\u00a2\u2009kWh\u22121 PPAs in the southwestern United States, which seem \nto have achieved many of the same savings by building along the \ntransmission infrastructure of soon-to-be decommissioned coal-\nfired plants. Solar projects in the United States do not claim to be \nunsubsidized, as they benefit from the 30% income tax credit; how-\never, at least one project, outside of Las Vegas, NV, appears to come \nin just below the 3\u00a2\u2009kWh\u22121 line even when the impact of the tax \ncredit is removed16. It is also worth noting that these projects will \nbe subject to the recent 30% tariff on imported modules, suggest-\ning that this move (probably mitigated by the continually falling \nmodule prices) is having limited impact on PV development in the \nUnited States. The second area is in the encouragement of favourable  \nfinancing deals for developers, which involves both the involvement \nof state-connected companies and banks in the projects, and (more \nwidely applicable) the clear commitment of the government to  \nlong-term renewable energy development, which certainly creates \na sense of stability in the industry. The low PPAs in the United \nStates suggest that the financing environment in the United States \nhas reached a similar level of favourability to solar that is found in \nthe Gulf, since\u2014as we have shown\u2014low-cost financing with high \ndebt fractions is an essential ingredient of low PPA prices. However, \nanother hotspot for low PPA prices has been Chile14, where PV proj-\nects can benefit first of all from the world\u2019s best solar resource in \nthe Atacama Desert30,31, but where the financial environment is seen \nto be less stable. This may have been mitigated in the latest round \nof bidding by the prominence of the Saudi project developer ALJ, \nwhich has won a number of projects across South America, and \ncan reasonably be speculated to have eased the financial aspect via \naccess to more freely flowing Gulf capital.\nConclusions\nThe rapid pace of development of solar energy, especially with \nregard to pricing, can make it difficult to keep up with the true state \nof the industry. Separating reality from hype requires a hard look \nat the most aggressive claims in terms of costs, risks and returns\u2014\nthe factors that really matter to developers, banks, and providers \nand consumers of electricity. Solar costs vary widely by region, \nand fully explaining this variation requires an understanding of \nlocal conditions, which we have sought to provide as context for \nthe recent record-low solar electricity prices and bids in the United \nArab Emirates and Saudi Arabia, and their subsequent extensions to \nTable 2 | Model inputs for low-cost PV project reconstruction\nBaseline\n3 \u00a2\u2009kWh\u22121 in the \nUnited Arab Emirates\nModule cost (\n\u2212\nS\nW\nU $\nd c. .\n1 )\n0.6424\n0.28a\nInverter cost (\n\u2212\nS\nU $ Wa c. .\n1 )\n0.12\n0.084\nTracker\u2009+\u200b\u2009BOS cost \n(\n\u2212\nS\nW\nU $\nd c. .\n1 )\n0.26\n0.22 (\u2212\u200b15% from \n201836)\nSales tax (%)\n5\n0\nConstruction labour cost \n(\n\u2212\nS\nW\nU $\nd c. .\n1 )\n0.16\n0.07 (50% of \nbenchmark 20174)\nEPC markup (%)\n825\n5b\nLand (US$ per acre)\n5,00019\n0\nLand prep. (US$ per acre)\n10,00019\n10,000\nEnvironmental permitting \n(US$)\n250,00025\n0\nInterconnection cost \n(\n\u2212\nS\nW\nU $\nd c. .\n1 )\n0.0325\n0\nTransmission line (US$)\n6,000,00025\n0 (borne by utility)\nContingency (%)\n3\n3\nDeveloper margin (%)\n1\n0.5\nOperation cost \n(US$\u2009MW\u22121\u2009yr\u22121)\n20,000, linear \nescalation to 25,000 \nover 20 years18\n10,000\u2212\u200b15,000 over \n25 years\nDebt fraction (%)\n6023\n80\nDebt interest (%)\n523\n3.5\nReturn on equity (%)\n1023\n10\nPPA term (years)\n2018,37\n25\nY1 yield (\n\u2212\nWh MW\nG\na c. .\n1 )\n2.836\n2.836\nCapEx (\n\u2212\nS\nW\nU $\na c. .\n1 )\n1.93\n0.95\nLCOE (\u00a2\u2009kWh\u22121)\n7.38\n2.85\nCost figures from ref.\u200924 unless otherwise noted. BoS, balance of system. EPC, engineering\u2013\nprocurement\u2013construction. aPVInsights module spot price, 25 June 2018. bMarked down by \nextrapolating the model of ref.\u200925.\nNature Energy | VOL 3 | DECEMBER 2018 | 1109\u20131114 | www.nature.com/natureenergy\n1112\n\nAnalysis\nNature Energy\nother regions. We have demonstrated here that unsubsidized prices \nbelow 3\u00a2\u2009kWh\u22121 are attainable when the right combination of local \nconditions is realized, and that the minimum for solar prices should \nbe expected to continue dropping (with some slowing expected as \ninterest rates rise) with hardware costs, barring some significant \nchange in the financial sector\u2019s perception of solar\u2019s profitability. \nWhile certain costs do seem to have been mitigated in the case of the  \nGulf projects, which could be argued to be a subsidy, the success of \nprojects in other regions at achieving comparable savings indicates \nthat these costs can be significantly reduced by skilful engineer-\ning, without state intervention. We wish to emphasize that govern-\nment policy remains an important element to remove barriers to \nPV deployment, chief among which is the need for access to cheap \nfinancing. As global interest rates are expected to rise, it becomes \nmore critical for solar projects to be able to prove themselves as low \nrisk to achieve favourable financing packages. The clear commit-\nment to renewable energy as state policy (as well as the involvement \nof large public or semi-public companies in the development of PV \nprojects) was certainly a factor in providing the necessary degree \nof confidence that allowed banks to overcome their historical aver-\nsion to renewable energy projects32 and begin lending at low rates \nin the Gulf. Such a mentality of strategic energy planning, whereby \ngovernments intervene not to subsidize but merely to enable and \nencourage investment in technologies that are believed to be of \nvalue to the society, is an essential element of low electricity prices \nand may provide guidance regarding how to continue the spread of \nlow solar prices into other markets.\nMethods\nBottom-up cost and performance modelling. In the investigation of recent low-\npriced PV bids and PPAs, public announcements regarding, for example, financing \nterms, provide a window into some of the mechanisms by which these projects \nprobably achieved their low prices. However, significant uncertainty remains, not \nthe least of which surrounds the actual costs of these projects, for which there \nare conflicting reports21,33,34. Furthermore this \u2018top-down\u2019 look gives little insight \ninto the relative impact of different cost-cutting mechanisms, and therefore \nprovides little grounding for assessing whether these prices can be transferred to \nother markets; and as many of the details are proprietary, it would in fact not be \nfeasible to attempt an exact reconstruction of any particular project. To circumvent \nthis challenge, we adopted a bottom-up modelling approach, quantifying cost \ncomponents based on a range of industry reports and analysis to build up a model \nof how these prices could be achieved under realistic global and local conditions. \n5\n0\n1\n2\n3\n7\n4\nLCOE (\u00d710\u20132 US$ kWh\u20131)\n6\n8\nHardware \n(via EPC contractor)\nConstruction labor \n(via EPC contractor)\nConstruction, \ndeveloper-borne\nDebt service\nSales tax\nEPC (markup)\nOperation\nReturn on equity\nBaseline\n7.38 \u00a2 kWh\u20131\nScale up\n7.02 \u00a2 kWh\u20131\nExtend PPA\nterm 6.51 \u00a2 kWh\u20131\nReduce hardware\nprices 4.50 \u00a2 kWh\u20131\nReduce construction\nlabour cost 4.10 \u00a2 kWh\u20131 \nReduce O&M\n3.72 \u00a2 kWh\u20131\nLower cost of\nfinancing 3.12 \u00a2 kWh\u20131\nLand, tax, fees\n2.85 \u00a2 kWh\u20131\nFig. 3 | Modelled LCOE for a sample plant under Abu Dhabi\u2019s weather conditions. Sales tax of 5% applies to hardware cost; EPC markup covers overhead \nand profit for the construction firm; developer-borne construction expenses include permitting and land costs.\n1.0\n0\n0.2\n0.4\n0.6\n1.4\n0.8\n1.2\n1.6\n2.0\n1.8\nBaseline\nUS$1.93 W\na.c.\u20131\n100-->800MW\nUS$1.82 W\na.c.\u20131\nReduce\nhardware cost\nUS$1.18 W\na.c.\u20131\nConstruction labour\ncost reduction\nUS$1.05 W\na.c.\u20131\nRemove\nsales tax\nUS$1.01 W\na.c.\u20131\nFree land\nUS$0.99 W\na.c.\u20131\nPermitting fees\nUS$0.94 W\na.c.\u20131\nEPC hard\nPermitting\nConstruction labour\nContingency\nLand and prep.\nEPC soft\nSales tax\nGrid tie\nDeveloper profit\nModule tariff\nSystem price (US$ W\u20131)\nFig. 4 | Modelled CapEx for a sample plant based on assumed costs for UAE large-scale solar. For large-scale projects, hardware represents the bulk of \nCapEx, followed by labour.\nNature Energy | VOL 3 | DECEMBER 2018 | 1109\u20131114 | www.nature.com/natureenergy\n1113\n\nAnalysis\nNature Energy\nThis allows us to look at the relative impact of the many factors that contribute to \nthe ultimate cost of producing electricity. Performance modelling was performed \nin System Advisor Model (SAM) for a plant designed to match the known \nspecifications of Dubai\u2019s MBR solar park, phase 3, the first of the sub-3\u00a2\u2009kWh\u22121 \nprojects. The \u2018baseline\u2019 model scaled down the MBR model to 100MWac while \nretaining the same assumptions about hardware costs. LCOE was calculated in a \nspreadsheet (available on request) as:\n =\n+ \u2211\n\u2211\n=\n+\n=\n+\nC\nLCOE\ni\nN\nO\nr\ni\nN\nE\nr\n1 (1\n)\n1 (1\n)\ni\ni\ni\ni\nusing the annual energy yield calculated by SAM and discounting by the weighted \naverage cost of capital. Performance modelling of the UAE plant (3\u00a2\u2009kWh\u22121) \nuses International Weather for Energy Calculations weather data for Abu Dhabi; \nmodelling for Sakaka (see Supplementary Note 1) used Photovoltaic Geographical \nInformation System weather data (http://re.jrc.ec.europa.eu/pvgis/, retrieved June \n2018) for Al Jouf airport (coordinates 23.694\u00b0\u2009N, 40.088\u00b0\u2009E, a few kilometres from \nthe solar park site). The model was implemented in SAM 2014.1.14, based on \nthe known or presumed specifications of Dubai\u2019s MBR phase 3, including large \n(310\u2009Wp) CS6X modules from Canadian Solar, 5000\u2009W inverters from GE, an \ninverter loading ratio of 1.3 and single-axis tracking.\nData availability\nThe data that support the plots within this paper and other findings of this study \nare available from the corresponding author upon reasonable request.\nReceived: 6 March 2018; Accepted: 29 August 2018;  \nPublished online: 8 October 2018\nReferences\n\t1.\t Renewable Energy Auctions: Analysing 2016 (International Renewable Energy \nAgency, 2017).\n\t2.\t Annual Energy Outlook 2018 (US Energy Information Administration, 2017).\n\t3.\t Energy Darwinism: The Evolution of the Energy Industry (CitiGPS, 2013).\n\t4.\t Fu, R., Feldman, D. J., Margolis, R. M., Woodhouse, M. A. & Ardani, K. B. \nUS Solar Photovoltaic System Cost Benchmark: Q12017 (National Renewable \nEnergy Laboratory, 2017).\n\t5.\t Bolinger, M, Seel, J. & LaCommare, K. H. Utility-Scale Solar 2016: An \nEmpirical Analysis of Project Cost, Performance, and Pricing Trends in the \nUnited States (Lawrence Berkeley National Laboratory, 2017).\n\t6.\t 800MW Third Phase of Mohammed bin Rashid Al Maktoum Solar Park Reaches \nFinancial Close (Dubai Electricity and Water Authority, 15 June 2017).\n\t7.\t Graves, L. Abu Dhabi plant to produce region\u2019s cheapest electricity from solar. \nThe National (1 March 2017).\n\t8.\t Dipaola, A. Saudi Arabia gets cheapest bids for solar power in auction. \nBloomberg Markets (3 October 2017).\n\t9.\t Kenning, T. ACWA Power wins 300\u2009MW Saudi solar project. PV Tech  \n(6 February 2018).\n\t10.\tThe SunShot Initiative\u2019s 2030 Goal: 3\u00a2 per Kilowatt Hour for Solar Electricity \n(US Department of Energy, 2016).\n\t11.\tSix-cent energy is not the new normal. PV Magazine (20 January 2015).\n\t12.\tGraves, L. Fuel for thought: there\u2019s nothing flowery about the renewables \nbidding bubble. The National (9 April 2017).\n\t13.\tWeaver, J. F. How is Saudi Arabia setting solar pricing records? Is it \nsustainable \u2013 repeatable? Electrek (8 October 2017).\n\t14.\tBellini, E. Chile\u2019s auction concludes with average price of $32.5/MWh  \nPV Magazine (3 November 2017)\n\t15.\tDeign, J. Mexican Solar Sets a Record Low Price for Latin America (Greentech \nMedia, 29 November 2017).\n\t16.\tSpector, J. Nevada\u2019s 2.3-Cent Bid Beats Arizona\u2019s Record-Low Solar PPA Price \n(Greentech Media, 12 June 2018).\n\t17.\tSgouris, S., Denes, C. & Ugo, B. The sower\u2019s way: quantifying the narrowing \nnet-energy pathways to a global energy transition. Environ. Res. Lett. 11, \n094009 (2016).\n\t18.\tBolinger, M., Weaver, S. & Zuboy, J. Is $50/MWh solar for real? Falling \nproject prices and rising capacity factors drive utility\u2010scale PV toward \neconomic competitiveness. Prog. Photovoltaics (2015).\n\t19.\tGoodrich, A., James, T. & Woodhouse, M. Residential, Commercial, and \nUtility-Scale Photovoltaic (PV) System Prices in the United States: Current \nDrivers and Cost-Reduction Opportunities (National Renewable Energy \nLaboratory, 2012).\n\t20.\tDEWA\u2019s rock-bottom solar bids: the real story (Solar GCC Alliance,  \n13 January 2015); http://www.solargcc.com/dewas-rock-bottom-solar-bids-\nthe-real-story/\n\t21.\tBintcliffe, J. Dubai attracts lowest solar tariff, again. Global Project Finance \nand Infrastructure Journal (11 May 2016).\n\t22.\tFeldman, D, Lowder, T. & Schwabe, P. P. V. Project Finance in the United \nStates, 2016 NREL/BR-6A20-66991 (National Renewable Energy Laboratory, \n2016); http://www.nrel.gov/docs/fy16osti/66991.pdf\n\t23.\tUtility-Scale Solar Photovoltaic Power Plants: A Project Developer\u2019s Guide \n(International Finance Corporation, 2015).\n\t24.\tFu, R. et al. US Solar Photovoltaic System Cost Benchmark Q12016 NREL/\nPR-6A20-66532 (National Renewable Energy Laboratory, 2016).\n\t25.\tFu, R. et al. Economic competitiveness of US utility-scale photovoltaics \nsystems in 2015: regional cost modeling of installed cost ($/W) and LCOE  \n($/kWh). In Proc. 2015 IEEE 42nd Photovoltaic Specialist Conference (PVSC) \n1\u201311 (IEEE).\n\t26.\tTrube, J., Fischer, M. & Metz, A. International Technology Roadmap for \nPhotovoltaic Materials, Processes and Products Relevant for Strategic Decisions \nReport 0947-076\u00d7\u200b (Wiley-V, 2016).\n\t27.\tFahy, A. & Bouyamourn, M. Wages in Abu Dhabi\u2019s construction sector jump \n17 per cent in 12 months. The National (17 August 2015).\n\t28.\tBouyamourn, A. Construction workers returning to India as UAE living costs \nsoar and wages rise back home. The National (9 March 2015).\n\t29.\tTuomiranta, A., Abdul Aziz, M. & Ghedira, H. Zoning Study for Deployment \nof Photovoltaic Power Stations in the UAE (2017).\n\t30.\t\u0160\u00fari, M., Cebecauer, T. & Skoczek, A. SolarGIS: Solar data and online \napplications for PV planning and performance assessment. In Proc. 26th Eur. \nPhotovoltaics Solar Energy Conf. (2011). \n\t31.\tMolina, A., Falvey, M. & Rondanelli, R. A solar radiation database for Chile. \nSci. Rep. 7, 14823 (2017).\n\t32.\tScott, A. Bank financing falls short of UAE\u2019s green economy ambitions.  \nThe National (31 May 2015).\n\t33.\tSaudi\u2019s ACWA Power gets $344 mln loan for Dubai solar project. Rueters  \n(1 March 2015).\n\t34.\tBellini, E. Chile\u2019s auction concludes with average price of $32.5/MWh.  \nPV Magazine (3 November 2017).\n\t35.\tCarvalho, S. Abu Dhabi closes $872 financing on world\u2019s larges solar power \nplant. Rueters (24 May 2017).\n\t36.\tMunsel, M. 10 Trends That Will Shape the Global Solar Market in 2018 \n(Greentech Media, 31 January 2018).\n\t37.\tMark, B. & Joachim, S. Utility-Scale Solar 2015: An Empirical Analysis of \nProject Cost, Performance, and Pricing Trends in the United States (2016).\nAcknowledgements\nWe note the valuable contribution of our colleagues with whom we have discussed this \nwork, in particular I. Almansouri and A. Tuomiranta. We thank M. Kissamitaki for \ndesigning the figures.\nAuthor contributions\nM.C. conceived and proposed this analysis, which was further elaborated in collaboration \nwith S.S. and H.A. H.A. carried out the analysis and drafted the manuscript in consultation \nwith M.C. and M.S. S.S. provided insights into the financial aspects of the projects and \nM.S. provided a utility company perspective on the project development process.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41560-018-0256-3.\nReprints and permissions information is available at www.nature.com/reprints.\nCorrespondence and requests for materials should be addressed to M.C.\nPublisher\u2019s note: Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2018\nNature Energy | VOL 3 | DECEMBER 2018 | 1109\u20131114 | www.nature.com/natureenergy\n1114\n\n\n    Scientific Research Findings:", "answer": "We performed a detailed cost breakdown of large-scale solar energy projects in the United Arab Emirates. Many factors contributed to the low price, but two stand out as having the most impact: the plummeting cost of solar panels and the reduced costs of financing. Solar panel prices have fallen by a factor of 6 since the beginning of this decade. The cost of building a solar power plant is now similar to the cost of other technologies. The cost of solar panels represents only one-third of the total construction costs. We found that solar power projects were financed by large loans covering 70\u201380% of the project costs, with low interest rates. The project developers were often large, state-connected companies, meaning that the projects were implicitly underwritten by the national government. It seems that these companies may have been willing to accept a somewhat lower return on their investment than purely private developers. However, we showed that these projects are still profitable when all of the costs are accounted for. It is important to note that our work attempts to reconstruct these projects based on publicly available information, and does not utilize proprietary information from the project developers themselves; the study should not be considered the authoritative opinion on exactly how any particular project achieved its results, but rather a general and transparent analysis on how a project could achieve a particular electricity price, given a set of local conditions that are more or less known.", "id": 35}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change\nnature climate change\nhttps://doi.org/10.1038/s41558-025-02283-4\nAnalysis\nFossil fuel subsidy reforms have become \nmore fragile\n \nPaasha Mahdavi\u2009\n\u200a\u20091\u2009\n, Michael L. Ross\u2009\n\u200a\u20092\u2009\n & Evelyn Simoni2\nSince the mid-2010s, many governments have pledged to reduce their \nsubsidies for fossil fuels. Yet, it is unclear whether these reforms have been \nimplemented, with prior studies showing conflicting results. Here we collect \noriginal monthly data on the 21 countries with the largest gasoline subsidies \nin the 2003\u20132015 period and evaluate their reform efforts from 2016 to \n2023. Since 2016, there has been an increase in the frequency and ambition \nof subsidy reforms but a drop in their durability: just 30% of the reforms \nsurvived for 12\u2009months, and only 9% survived for 36\u2009months. Subsidies rose \nfor 12 countries in our sample and were virtually unchanged in the other \n9. This pattern calls into question the effectiveness of recent strategies for \nreducing fossil fuel subsidies.\nOne of the most cost-effective ways to reduce carbon emissions is by \neliminating fossil fuel subsidies1,2. These subsidies boost the consump-\ntion of fossil fuels and, hence, slow the transition to renewable energy3; \nincrease road traffic, air pollution and road fatalities, all of which dam-\nage public health4,5; and can have a crippling impact on government \nrevenues6.\nYet, they are also difficult to abolish. Many intergovernmental \norganizations\u2014including the World Bank, the International Monetary \nFund (IMF) and the International Energy Agency\u2014have urged gov-\nernments to eliminate these subsidies. At the 2021 United Nations \nClimate Change Conference (COP26) in Glasgow, Scotland, govern-\nments pledged to phase down \u2018inefficient fossil fuel subsidies\u2019. But \npolitical leaders are often wary of reform, realizing that higher fuel \nprices are unpopular and could lead to protests that could threaten \ntheir government7,8.\nIn the early 2010s, international organizations stepped up their \nadvocacy work for fossil fuel subsidy reform, sponsoring new initia-\ntives, reports and conferences9. The IMF began to emphasize subsidy \nreform in its policy recommendations10. Subsidy reforms were directly \nincorporated into the Sustainable Development Goals (SDG 12.c) in \nSeptember 2015, and 15 countries included fossil fuel subsidy reforms \nin their Intended Nationally Determined Contributions for the Decem-\nber 2015 Paris conference11. It is unclear whether these initiatives led \nto a reduction in subsidies. One study found that, from 2015 to 2020, \n41 countries reformed consumer fossil fuel subsidies12; yet, according \nto the International Energy Agency, from 2015 to 2022 these subsidies \nrose from US$183 billion to US$343 billion (ref. 13).\nThe reform of fossil fuel subsidies is difficult to study in part \nbecause they are hard to measure. In many countries, subsidies rise \nand fall over time owing to a mixture of market forces and government \npolicies that can be hard to disentangle. Governments often provide \nunreliable information about their own policies: they may announce \nreforms but fail to implement them or, conversely, adopt reforms \nquietly in an effort to avoid protests.\nTo better measure government policy reforms, we adopt a novel \napproach. Like other scholars, we focus on consumer subsidies for \ngasoline, which can be calculated using the widely accepted price gap \nmethod14. We begin by calculating the average implicit gasoline tax or \nsubsidy for all countries from 2003 to 2015, using the Ross\u2013Hazlett\u2013\nMahdavi dataset15. We then take as our sample those countries that, \non average, were net subsidizers during this 13-year period. These 21 \ncountries account for 97% of all gasoline subsidies in the 2003\u20132015 \nperiod. They constitute a \u2018most likely\u2019 sample: if progress has been \nmade since 2016 in subsidy reform, we are most likely to find sub-\nstantial reforms in these 21 countries. Although they are all fossil fuel \nexporters, they vary widely in economic and social development. \nThey also represent many locations across the Global South, including \nSub-Saharan Africa, Latin America, the Middle East, Southeast Asia \nand the Caspian Basin (Table 1).\nWe then leverage a valuable feature common to these 21 countries. \nAll of them used government mandates to set fuel prices and kept them \nfixed for months or years at a time. Hence, whenever the government \nraised the nominal per-litre price by at least 10%\u2014effectively reducing \nthe subsidy for their citizens\u2014we code it as a reform. We also measure a \nReceived: 24 September 2024\nAccepted: 18 February 2025\nPublished online: xx xx xxxx\n Check for updates\n1Department of Political Science, Bren School of Environmental Science and Management (by courtesy), University of California, Santa Barbara, CA, USA. \n2Department of Political Science, University of California, Los Angeles, CA, USA. \n\u2009e-mail: paasha@ucsb.edu; mlross@polisci.ucla.edu\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nOur study makes three contributions to research on climate policy \nand politics.\nFirst, it documents the remarkable frequency with which fossil fuel \nsubsidy reforms have failed. The 21 countries in our sample enacted \n132 reforms during the 2000\u20132023 period. All but nine of them (6.8%) \neventually failed. We disaggregated these failures by type and period \nto better understand the obstacles to reform.\nSecond, our study helps to resolve conflicting interpretations of \nthe recent trend in fossil fuel subsidies. Some studies describe pro-\ngress in rolling back fossil fuel subsidies18, while others show that the \ntotal value of these subsidies has increased over the same period5,13. \nOur analysis resolves this paradox by demonstrating that the rising \nfrequency of reforms after 2015 was offset by a commensurate decline \nin reform duration.\nFinally, our study sheds light on the disappointing record of sub-\nsidy reforms, especially since 2016. By December 2023, no country \nin our sample had considerably lower subsidies than in January 2016. \nWe are cautious about extrapolating our finding that there has been \nlittle net progress to other types of climate policy: fuel subsidies are \nexceptionally hard for governments to change, especially in oil- and \ngas-producing countries. We believe that our findings should cause \npolicy advocates to look for new strategies that could make fossil fuel \nsubsidy reform more durable and, hence, more effective.\nResults\nWe track two varieties of subsidy reform. The most common type is \nprice reform, which we define as an increase in the nominal price of \ngasoline by 10% or more in local currency units in a single month. For \nrobustness, we also track an alternative measure that indicates a nomi-\nnal increase of 25% or more over 3\u2009months (Extended Data Fig. 5). Price \nreform ends (or \u2018fails\u2019) if the subsidy returns to its pre-reform level, \nwhich can happen through two mechanisms: the government may \nexplicitly reverse the price increase, a process we call backtracking; \nor the government may maintain the post-reform price, but its value \ndisappears through inflation, a falling exchange rate or rising global \noil prices. We refer to the latter process as erosion.\nThe second type is fixity reform, which occurs when a government \nstops using fixed prices and subsidies altogether, allowing gasoline \nprices to fluctuate in response to changing oil prices for at least three \nconsecutive months. A fixity reform can fail in two ways: if the govern-\nment reimposes fixed prices for three or more consecutive months; or if \nthe government interferes with the market price to restore a substantial \nsubsidy (which we define as a subsidy exceeding 20 cents per litre in \nconstant 2015 US dollars) for three consecutive months.\nIncidence of reform\nFrom 2000 to 2015, the countries in our sample adopted an average of \n0.24 reforms per year per country; from 2016 to 2023, the reform rate \nrose to 0.38 per year, an increase of more than 50% (Table 2). There were \nmore reforms in 2015 and in 2016 than in any other year.\nsecond type of reform, when a government transitions from fixed prices \nto floating, unsubsidized, market-based prices. We refer to increases in \nthe fixed price as price reforms and shifts from fixed to floating prices \nas fixity reforms. Because earlier studies suggest that many reforms are \ntransitory16, we measure both the initiation of reforms and their dura-\ntion. To measure changes, we compare the frequency and duration of \nreforms from the January 2000 to December 2015 period, leading up \nto the 2015 Paris conference, with the January 2016 to December 2023 \nperiod following the end of the conference. While the final text of the \nParis Agreement did not mention subsidy reform, the December 2015 \nconference was a landmark in the global effort to promote climate \npolicies17 and provides a useful cut point for our analysis. Our core \nresults are similar if we use December 2014 instead (Extended Data \nFig. 6). They are also unchanged if we drop any country in our sample \n(Extended Data Figs. 7 and 8).\nWe have three broad findings.\nFirst, since 2016 there has been a notable increase in the frequency \nof reforms. From 2000 to 2015, the countries in our sample collectively \ninitiated an average of 0.24 price reforms per country year; from 2016 \nto 2023, they initiated about 0.38 reforms per country year, a 58% \nincrease. There was also a sharp rise in fixity reforms, which we regard \nas more ambitious than price reforms.\nSecond, reforms were already fragile in the pre-2016 period, and \ngrew even more fragile after 2015. From 2000 to 2015, only 45% of all \nsubsidy reforms survived at least 12\u2009months and only 22% survived for \n36\u2009months; from 2016 to 2023, 30% survived for 12\u2009months and only 9% \nlasted 36\u2009months. Ambitious reforms were especially fragile.\nFinally, subsidies per litre of gasoline rose substantially in 12 coun-\ntries in our sample and changed little in the other 9. On average, the \nvalue of the per-litre subsidy in our sample rose 65.1% over these 8\u2009years, \nequivalent to an annual increase of 6.5%.\nTable 1 | Changes in net taxes from 2016 to 2023\nTaxes Q1\u20132016\nTaxes Q4\u20132023\nDifference\nAngola\n0.52\n\u22120.30\n0.82\nIran\n\u22120.07\n\u22120.55\n0.48\nIndonesia\n0.13\n\u22120.24\n0.37\nSudan\n0.42\n0.07\n0.36\nIraq\n\u22120.02\n\u22120.31\n0.30\nLibya\n\u22120.29\n\u22120.56\n0.27\nMalaysia\n0.01\n\u22120.24\n0.26\nEgypt\n\u22120.07\n\u22120.30\n0.23\nAlgeria\n\u22120.13\n\u22120.33\n0.20\nBahrain\n\u22120.11\n\u22120.29\n0.19\nAzerbaijan\n0.04\n\u22120.13\n0.17\nKuwait\n\u22120.20\n\u22120.37\n0.17\nQatar\n\u22120.04\n\u22120.13\n0.09\nOman\n\u22120.06\n\u22120.12\n0.06\nNigeria\n0.06\n0.01\n0.05\nEcuador\n\u22120.05\n\u22120.09\n0.04\nTrinidad and Tobago\n0.08\n0.10\n\u22120.02\nUnited Arab Emirates\n0.03\n0.06\n\u22120.03\nMyanmar\n0.05\n0.10\n\u22120.05\nSaudi Arabia\n\u22120.16\n\u22120.10\n\u22120.06\nAverage level of tax for the first quarter of 2016 and the final quarter of 2023, in real US \ndollars per litre. Negative values reflect subsidies; positive values reflect taxes. The difference \nshows the increase in the subsidy over this period (conversely, the decline in the tax). Note: \nVenezuela is excluded due to the unavailability of reliable exchange rate data in IMF records \nfrom 2018 onwards.\nTable 2 | Summary statistics of reform onset and durability\n2000\u20132015\n2016\u20132023\nRate of price reform onset (per country per year)\n0.22\n0.25\nRate of fixity reform onset (per country per year)\n0.02\n0.13\nReforms lasting at least 12 months (% of total \nreforms in the period)\n44.74\n30.36\nReforms lasting at least 36 months (% of total \nreforms in the period)\n22.37\n8.93\nFailure rate from backsliding (% of total reforms \nin the period)\n0.09\n0.34\nFailure rate from erosion (% of total reforms in \nthe period)\n0.87\n0.55\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nDuration of reform\nSubsidy reforms tend to be short-lived, and reforms failed more \nquickly after 2016. From 2000 to 2015, 45% of all reforms survived \nfor 12\u2009months, and 22% survived for at least 36\u2009months; from 2016 to \n2023, only 30% survived for 12\u2009months, and 9% survived for at least \n36\u2009months (Fig. 1).\nDepth of reform\nOver the full 2000\u20132023 period, 80% of all reforms were price reforms \nand 20% were fixity reforms. This ratio changed over time. From 2000 to \n2015, we found an average of just 0.024 new fixity reforms per country \nyear. From 2016 to 2023, however, they occurred at a rate of 0.13 new \nreforms per country year, a fivefold increase. Fixity reforms accounted \nfor most of the overall rise in the frequency of reforms after Paris.\nThe advantages of fixity reforms are self-evident: floating prices are \nnot automatically eroded by inflation, currency depreciation or global \noil shocks. In theory, fixity reforms should be more durable than price \nreforms. In practice, governments found it politically difficult to allow \nprices to float, especially when world oil prices rose too far. As a result, fix-\nity reforms were no more durable than price reforms. While fixity reforms \nwere more likely than price reforms to survive in their first year, they were \nless likely to survive after 12\u2009months. The mean duration of price reforms \nwas 1.8\u2009years, while the mean duration of fixity reforms was 1.5\u2009years.\nBy December 2023, when our data end, three countries still had \nfloating prices in place: Myanmar, which is the poorest country in our \nsample; United Arab Emirates, which is the second wealthiest; and \nNigeria, which falls near the middle. We discuss these cases below.\nSources of failure\nErosion was the most common source of failure during the 2000\u20132023 \nperiod, bringing about the end of 47% of reforms within the first year \nand 61% within 3\u2009years. Backtracking brought about the end of 14% of \nall reforms in the first year and 19% within 3\u2009years.\n0\n5\n10\n0\n01 02 03 04 05 06 07 08 09 10\n11\n12 13 14 15 16\n17\n18 19 20 21 22 23\nPre-2016 \u2212 fixity\nPre-2016 \u2212 price\nPost-2016 \u2212 fixity\nPost-2016 \u2212 price\na   Reform onsets, 2000\u20132023\n0\n5\n10\n15\n0\n01 02 03 04 05 06 07 08 09 10\n11\n12 13 14 15 16\n17\n18 19 20 21 22 23\nPre-2016 \u2212 fixity\nPre-2016 \u2212 price\nPost-2016 \u2212 fixity\nPost-2016 \u2212 price\nb    Reform failures, 2000\u20132023\n0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since reform onset\nSurvival probability\nPre-2016\nPost-2016\nc    Duration of reforms, pre/post-2016\n0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since reform onset\nSurvival probability\nPrice reform\nFixity reform\nd    Duration of price and fixity reforms\nFig. 1 | Incidence and duration of fuel subsidy reforms, 2000\u20132023. a, Change \nover time in the onset of price and fixity subsidy reforms, summed across all \ncountries in a given year. b, Change over time in the failure of price and fixity \nreforms, summed across all countries in a given year. c, Kaplan\u2013Meier survival \nrate of all subsidy reforms before (dashed orange) and after (solid blue) \nDecember 2015. d, Kaplan\u2013Meier survival rate of fixity (solid blue) and price \n(dashed orange) subsidy reforms.\n0\n0.0001\n0.0002\n0.0003\n2003 2005 2007 2009 2011\n2013\n2015\n2017\n2019\n2021 2023\nBacksliding\nErosion\nFig. 2 | Reform failure due to backsliding or erosion. Overlapping density plots \nfor backsliding and erosion over the 2000\u20132023 period, with the y axis showing \nkernel density estimates. Backsliding refers to reversions of the nominal price to \nthe pre-reform nominal price. Erosion refers to reversions of the real price to pre-\nreform real prices due to inflation, currency depreciation or global oil price change.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nAfter 2016, backtracking became more common. In the 2000\u20132015 \nperiod, just 9.6% of all reform failures were caused by backtracking; \nafter 2016, it rose to 38% of all failures (Fig. 2).\nSubsidy level trends\nPer-litre gasoline subsidies rose (or taxes fell) in 12 of the 21 countries \nfrom the beginning of 2016 to December 2023. In the remaining nine, \nthere was little change (Table 1).\nFigure 3 shows the trajectories of each country\u2019s gasoline tax \nor subsidy per litre for the full 2000\u20132023 period, along with the \nunweighted mean. The distance above or below the horizontal line at \n0.0 represents the size of the tax or subsidy. The parallel fluctuations \nacross the sample reflect the prevalence of price fixity, such that move-\nments are primarily due to global oil price shifts. Subsidies increased \ngradually in almost all countries from 2000 to 2008. Since then, they \nhave fluctuated but generally remained large, except for brief periods \nin 2008\u20132009, 2014\u20132015 and 2020.\nDiscussion\nOur study finds that since 2016 there has been a rise in number of gaso-\nline subsidy reforms but a decline in their durability. As a result, from \nthe beginning of 2016 to the end of 2023, per-litre subsidies grew for \n12 countries in our sample and changed little in the other 9.\nWe also document a sharp rise around 2015 in fixity reforms, which \nhave the potential to be more far reaching and sustainable. These fixity \nreforms, however, have been just as fragile as price reforms.\nDespite their ambition, when push came to shove\u2014sometimes \nquite literally in the form of protests\u2014leaders chose to restore subsi-\ndies. This occurred most often through the erosion of price reforms \nby devaluation or inflation, but also through backsliding that returned \nprices to pre-reform levels in the face of public opposition. The price \nshock that followed the Russian invasion of Ukraine was particularly \ndamaging, causing the reversal of fixity reforms in all but one case.\nFor advocates of fossil fuel subsidy reform, our results are sober-\ning. Several studies report a broad campaign to encourage fossil fuel \nsubsidy reforms in the mid-2010s, thanks to initiatives backed by \nthe Organisation for Economic Co-operation and Development, the \nInternational Energy Agency, the World Bank, the IMF and the United \nNational Environmental Programme, and ultimately articulated in \nsection 12.c of the Sustainable Development Goals10,11. Yet, our findings \nsuggest that these initiatives generated little progress towards remov-\ning gasoline subsidies in the countries where reforms were needed.\nThe poor reform records of these 21 countries should not be entirely \nsurprising when taking politics into account. Recent research suggests \nthat fuel subsidies remain popular in many countries, where citizens \ngenerally want to see them raised, not reduced19; that when citizens learn \nthat their government is providing them with large gasoline subsidies, \nthey tend to view the government more favourably20; that when subsidies \nare removed, citizens often rise in protest7; and that successful reforms \ntend to be idiosyncratic and largely driven by local, country-specific \nfactors16,21,22. Citizens are less likely to oppose subsidy reform if they are \ngiven credible promises of increased public spending on health, labour \nand other social outcomes23,24, yet this credibility can itself be fragile.\nAlthough fiscally irresponsible and economically damaging, fuel \nsubsidies continue to have political advantages for leaders. According \nto the theory of the \u2018rentier state\u2019, when undemocratic governments are \nequipped with revenue derived from oil wealth-constituting \u2018rents\u2019 for \nstate coffers, they distribute these rents in exchange for public support \nfor the regime25,26. This stands in contrast with governments that lack \nresource wealth and are compelled to recognize the democratic rights \nof citizens in order to collect revenues from taxes.\nIn the early 2000s, fossil fuel subsidies became a common way for \ngovernments to distribute oil wealth, especially when governments \nlacked the administrative capacity to effectively deliver public goods \nsuch as schools, health care and public infrastructure27. Subsidies tend \nto benefit middle- and upper-class car-owning urbanites who are often a \ncritical source of government support28,29. Experimental evidence finds \nsupport for tenets of this theory, demonstrating that wealthy citizens \nperceive higher levels of government performance when they receive \nlarger fuel subsidies20.\nWe see two possible explanations for the simultaneous increase \nin the frequency and drop in the durability of reforms since 2016. The \nfirst is the volatility of oil prices, which can drive cycles of reform and \nbacksliding: falling prices can lower the barriers to reform, while rising \nprices, if passed on to consumers, can ignite protests. But this expla-\nnation does not fit with the evidence: oil prices were more volatile in \nthe 2000\u20132015 period than in 2016\u20132023 (Extended Data Fig. 9 and \nExtended Data Table 1).\nThe other explanation is political: an increase in global pressure \nfor price reform may have caused governments in the Global South to \nreluctantly initiate reform, even when domestic support was lacking. \nThis could bring about a rise in the incidence of short-lived reforms, \nended by government backsliding2. This explanation is consistent with \nour finding that backsliding became much more common after 2015.\n\u22121.0\n\u22120.5\n0\n0.5\n1.0\n1.5\n0\n01\n02\n03\n04\n05\n06\n07\n08\n09\n10\n11\n12\n13\n14\n15\n16\n17\n18\n19\n20\n21\n22\n23\nFig. 3 | Gasoline taxes and subsidies, 2000\u20132023. The grey lines represent \na given country\u2019s per-litre tax or subsidy on regular unleaded gasoline in \nconstant 2015 dollars. The red line represents the unweighted mean value across \ncountries. The sample includes all countries that maintained net subsidies from \nJanuary 2003 to December 2015. The vertical green line shows the timing of the \nParis meeting in December 2015.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nThere were only three exceptions to the overall pattern of reform \nfailure. None of them, however, offers a straightforward model for other \ncountries: the United Arab Emirates has maintained fixity reforms, but \nthe burden of higher gasoline prices largely falls on non-citizens, who \noutnumber citizens 9 to 1; Myanmar has maintained fixity reforms dur-\ning a crippling civil war and trade sanctions, while offering subsidized \ngasoline to military-aligned companies30; and since May 2023, Nigeria \nhas adopted a mixed system of reforms and subsidies that has created \nnationwide fuel shortages31,32.\nOur findings underscore the extraordinary challenge of fossil fuel \nsubsidy reform. We see two plausible policy responses.\nThe first would be for advocates to redouble their efforts, look-\ning for novel strategies to build public support and guard reforms \nagainst pressures created by inflation, currency devaluation and oil \nprice fluctuations.\nThe second would be to look for less unpopular ways to reduce \nthe cost of subsidies. Market pricing may be the first-best solution, \nbut it is often politically out of reach. Alternatively, governments \ncould adopt policies that reduce the demand for subsidized fuel, \nfor example, with regulatory measures to improve fuel efficiency, \npromote electric vehicles and invest in public transit. If successful, \npro-renewables policy mixes could reduce emissions33 while foster-\ning new coalitions of political support that could ward off future \npolitical setbacks34,35.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-025-02283-4.\nReferences\n1.\t\nDavis, L. W. The economic cost of global fuel subsidies. Am. Econ. \nRev. 104, 581\u201385 (2014).\n2.\t\nRentschler, J. & Bazilian, M. Reforming fossil fuel subsidies: \ndrivers, barriers and the state of progress. Clim. Policy 17, 891\u2013914 \n(2017).\n3.\t\nCharap, M. J., da Silva, M. A. R. & Rodriguez, M. P. C. Energy \nSubsidies and Energy Consumption: A Cross-Country Analysis. \nWorking Paper No.13/112 (International Monetary Fund, 2013).\n4.\t\nErickson, P. et al. Why fossil fuel producer subsidies matter. \nNature 578, E1\u2013E4 (2020).\n5.\t\nBlack, M. S., Liu, A. A., Parry, I. W. & Vernon, N. IMF Fossil Fuel \nSubsidies Data: 2023 Update (International Monetary Fund,  \n2023).\n6.\t\nSteadman, S., Gen\u00e7s\u00fc, I., Mustapha, S., Colenbrander, S. & \nTyson, J. Indebted: How to Support Countries Heavily Reliant on \nOil and Gas Revenues to Secure Long-Term Prosperity (Overseas \nDevelopment Institute, 2023).\n7.\t\nNatalini, D., Bravo, G. & Newman, E. Fuel riots: definition, evidence \nand policy implications for a new type of energy-related conflict. \nEnergy Policy 147, 111885 (2020).\n8.\t\nvon Uexkull, N., R\u00f8d, E. G. & Svensson, I. Fueling protest? Climate \nchange mitigation, fuel prices and protest onset. World Dev. 177, \n106536 (2024).\n9.\t\nVan de Graaf, T. & Blondeel, M. in The Politics of Fossil Fuel \nSubsidies and Their Reform (eds Skovgaard, J. & van Asselt, H.)  \nCh. 1 (Cambridge Univ. Press, 2018).\n10.\t Skovgaard, J. The Economisation of Climate Change: How the G20, \nthe OECD and the IMF Address Fossil Fuel Subsidies and Climate \nFinance (Cambridge Univ. Press, 2021).\n11.\t\nElliott, C., Bernstein, S. & Hoffmann, M. Credibility dilemmas \nunder the Paris Agreement: explaining fossil fuel subsidy reform \nreferences in INDCs. Int. Environ. Agreem. 22, 735\u2013759 (2022).\n12.\t Sanchez, L., Wooders, P., Mostafa, M. & Bechauf, R. 53 Ways to \nReform Fossil Fuel Consumer Subsidies and Pricing (International \nInstitute for Sustainable Development, 2020).\n13.\t Muta, T. & Erdogan, M. The Global Energy Crisis Pushed Fossil Fuel \nConsumption Subsidies to an All-Time High in 2022 (International \nEnergy Agency, 2023).\n14.\t Kojima, M. & Koplow, D. Fossil Fuel Subsidies: Approaches and \nValuation. Policy Research Working Paper 7220 (World Bank,  \n2015).\n15.\t Ross, M. L., Hazlett, C. & Mahdavi, P. Global progress and \nbacksliding on gasoline taxes and subsidies. Nat. Energy 2, 16201 \n(2017).\n16.\t Martinez-Alvarez, C. B., Hazlett, C., Mahdavi, P. & Ross, M. L. \nPolitical leadership has limited impact on fossil fuel taxes and \nsubsidies. Proc. Natl Acad. Sci. USA 119, e2208024119 (2022).\n17.\t Van de Graaf, T. Is OPEC dead? Oil exporters, the Paris Agreement \nand the transition to a post-carbon world. Energy Res. Social Sci. \n23, 182\u2013188 (2017).\n18.\t van den Bergh, J. C. & Savin, I. Political leadership, climate \npolicy, and renewable energy. Proc. Natl Acad. Sci. USA 120, \ne2301291120 (2023).\n19.\t Vieites, Y., Andretti, B., Weiss, M., Jacob, J. & Hallack, M. Effectively \ncommunicating the removal of fossil energy subsidies: evidence \nfrom Latin America. Glob. Environ. Change 81, 102690 (2023).\n20.\t Kubinec, R. & Milner, H. V. Taxes in the time of revolution: an \nexperimental test of the rentier state during Algeria\u2019s Hirak. World \nPolitics 76, 294\u2013333 (2024).\n21.\t Inchauste, G. & Victor, D. G. The Political Economy of Energy \nSubsidy Reform (The World Bank, 2017).\n22.\t Skovgaard, J. & van Asselt, H. The Politics of Fossil Fuel Subsidies \nand Their Reform (Cambridge Univ. Press, 2018).\n23.\t Kyle, J. Local corruption and popular support for fuel subsidy \nreform in Indonesia. Comp. Polit. Stud. 51, 1472\u20131503 (2018).\n24.\t Harring, N., J\u00f6nsson, E., Matti, S., Mundaca, G. & Jagers, S. C. \nCross-national analysis of attitudes towards fossil fuel subsidy \nremoval. Nat. Clim. Change 13, 244\u2013249 (2023).\n25.\t Beblawi, H. & Luciani, G. The Rentier State (Routledge, 1990).\n26.\t Ross, M. L. The Oil Curse: How Petroleum Wealth Shapes the \nDevelopment of Nations (Princeton Univ. Press, 2012).\n27.\t Hertog, S. The political economy of distribution in the Middle \nEast: is there scope for a new social contract? Dev. Policy 88, \n88\u2013113 (2017).\n28.\t Bates, R. H. Markets and States in Tropical Africa: The Political Basis \nof Agricultural Policies (Univ. California Press, 1981).\n29.\t Kim, S. E. & Urpelainen, J. Democracy, autocracy and the urban \nbias: evidence from petroleum subsidies. Polit. Stud. 64, 552\u2013572 \n(2016).\n30.\t Myint, M., Seong, J. & Aliyev, S. Energy Sector: Myanmar \nInfrastructure Monitoring. Policy Research Working Paper 170200 \n(The World Bank, 2022).\n31.\t Anyaogu, I. Nigerian gasoline prices soar as shortages worsen  \ncost of living. Reuters (2024); https://www.reuters.com/world/ \nafrica/nigerian-gasoline-prices-soar-shortages-worsen-cost- \nliving-crisis-2024-04-30/\n32.\t Nigeria: Post-financing Assessment Discussions (International \nMonetary Fund, 2024).\n33.\t Axsen, J., Pl\u00f6tz, P. & Wolinetz, M. Crafting strong, integrated \npolicy mixes for deep CO2 mitigation in road transport. Nat. Clim. \nChange 10, 809\u2013818 (2020).\n34.\t Breetz, H., Mildenberger, M. & Stokes, L. The political logics of \nclean energy transitions. Bus. Politics 20, 492\u2013522 (2018).\n35.\t Beiser-McGrath, L. F., Bernauer, T. & Prakash, A. Command and \ncontrol or market-based instruments? Public support for policies \nto address vehicular pollution in Beijing and New Delhi. Environ. \nPolitics 32, 586\u2013618 (2023).\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nPublisher\u2019s note Springer Nature remains neutral with  \nregard to jurisdictional claims in published maps and  \ninstitutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2025\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nMethods\nMeasurement of fuel subsidies\nMost research on fossil fuel subsidies is concerned with estimating the \ndamages they cause and the benefits of subsidy reform1,2,36. Our study \nbuilds on two smaller bodies of positive research.\nThe first is made up of cross-national quantitative studies of gaso-\nline taxes and subsidies\u2014the most readily measured type of fossil fuel \nsubsidy. In general, these studies demonstrate that gasoline taxes \nand subsidies are closely associated with slowly changing economic \nfactors, especially income per capita and fossil fuel wealth. The role of \npolitics is more uncertain: some studies find a link between fuel taxes \nand political institutions29,37,38, whereas others do not39.\nThese studies have important limitations: they only cover the \nperiod until 2015 or 2016; they look simultaneously at taxes and subsi-\ndies, and generally make no distinction between the political challenges \nof raising taxes and the political challenges of removing subsidies; and \nthey cannot make strong claims about causal effects.\nThe second set consists of qualitative case studies, which offer \nin-depth analyses of specific instances of reform12,21,22,40. While their \nemphases vary, in general they suggest that success is more likely to \nfollow a well-designed communications strategy, targeted compen-\nsation for affected communities and a gradual, phased approach to \nimplementation2. One advantage of these studies is their capacity to \ncover many types of fossil fuel subsidies and describe more subtle, \ncontext-specific influences. Their disadvantages include a tendency to \nfocus on cases of success (that is, selection on the dependent variable) \nand less attention to the durability of reforms.\nOur analysis has several innovations. It is the first quantitative \nanalysis to cover the post-Paris period, allowing us to provide an \nupper-bound estimate of Paris\u2019s effects; it measures both the inci-\ndence of reforms and their duration; it distinguishes between price \nreforms and fixity reforms, charting the growing popularity (but lim-\nited success) of the latter; and it offers a more detailed analysis of how \nreforms fail.\nMeasurement approach\nMost countries tax the consumption of fossil fuels rather than sub-\nsidize it. Among those that subsidize, some target their subsidies \ntowards low-income, vulnerable populations and use them for rela-\ntively brief periods; others subsidize most or all of their citizens over \nmany years. Our study is concerned with the latter group of countries, \nwhich are typically the focus of international efforts to reduce fossil \nfuel subsidies.\nWe follow convention and define a subsidy as the difference \nbetween the price paid by consumers and the cost of bringing the fuel \nto market14. To measure the size of these subsidies, we use the price \ngap method, which compares the observed retail price of a fuel in each \ncountry with a global benchmark price, which represents the supply \ncost. Anytime the retail price falls below the benchmark price (for exam-\nple, the cost of supplying the fuel), it denotes a subsidy for that period.\nTo identify the countries that were net subsidizers in the pre Paris \nera, we use the ref. 15 dataset of monthly gasoline taxes and subsidies, \nwhich covers 157 countries. We include in the sample all countries \nwhose median gasoline price for the 2003\u20132015 period was below \nthe median benchmark for the same period, meaning they were net \nsubsidizers during this initial period. This yields 22 countries, one of \nwhich (Yemen) we drop due to missing data in the period after its civil \nwar began.\nData and sample\nWe used a combination of government documents, oil and gas industry \nreporting, and media reports to extend the original 2000\u20132015 data to \nDecember 2023 for our sample. Data are missing for 173 of the 8,568 \ncountry-months owing to periods of political disruption for Myanmar \nand Venezuela.\nThe countries in our sample are located in different regions\u2014Latin \nAmerica, the Middle East, North Africa, Central Asia, Sub-Saharan Africa \nand Southeast Asia\u2014yet share two important characteristics: they are \nexporters of oil or fossil gas; and they tend to set fuel prices by govern-\nment fiat, protecting them from global price fluctuations. The use of \nfixed gasoline prices creates a link between the global oil price and the \nsize of the subsidy: when global prices rise, so do the benefits accruing \nto local consumers who continue to enjoy the fixed price and, hence, \nare receiving a larger per-unit subsidy.\nSubsidy reforms\nWe analyse two varieties of subsidy reform. The most common type is \nprice reform, when a country raises the nominal price by 10% or more in \nlocal currency units in a single month. For robustness, we use an alter-\nnative measure of price reform, when governments raise the nominal \nprice by 25% or more over a 3-month period. Among the 21 countries \nin our sample, all but one (Oman) initiated at least one episode of price \nreform between 2000 and 2023. Three countries (Iran, Sudan and \nAngola) had ten or more price reforms (Extended Data Fig. 3 and 4).\nPrice reform ends (or fails) if the subsidy returns to its pre- \nreform level, which can happen through two mechanisms: the gov-\nernment may lower the nominal price to its pre-reform level, the \nprocess we call backtracking; or the government may maintain the \npost-reform price, but its value disappears through inflation, a fall-\ning exchange rate or rising global oil prices. We refer to the latter \nprocess as erosion.\nThe second type is fixity reform, which occurs when a government \nstops using fixed prices and subsidies altogether, allowing gasoline \nprices to fluctuate in response to changing oil prices for at least three \nconsecutive months. A fixity reform can fail in two ways: the \u2018fixity\u2019 \ncomponent can fail if the government reimposes fixed prices for three \nor more consecutive months; or the \u2018no subsidies\u2019 component can \nfail if the government interferes with the market price to restore a \nsubstantial subsidy (which we define as a subsidy exceeding 20 cents \nper litre in constant 2015 US dollars for three consecutive months). \nAll 21 countries had government-fixed gasoline prices for most of the \n2000\u20132023 period, but ten of them shifted one or more times to float-\ning prices (Extended Data Fig. 2).\nIn Extended Data Fig. 1, we illustrate our measures using the case \nof Nigeria.\nData availability\nThe fossil fuel subsidy data supporting the findings of this study are \navailable via the Harvard DataVerse at https://dataverse.harvard.edu/\ndataset.xhtml?persistentId=doi:10.7910/DVN/80JO6T.\nCode availability\nThe code to replicate all findings of this study is available via Code \nOcean at https://doi.org/10.24433/CO.8539725.v1.\nReferences\n36.\t Parry, I., Black, M. S. & Vernon, N. Still Not Getting Energy Prices \nRight: A Global and Country Update of Fossil Fuel Subsidies \n(International Monetary Fund, 2021).\n37.\t Cheon, A., Urpelainen, J. & Lackner, M. Why do governments \nsubsidize gasoline consumption? An empirical analysis of  \nglobal gasoline prices, 2002\u20132009. Energy Policy 56, 382\u2013390 \n(2013).\n38.\t Fails, M. D. What types of political regimes subsidize fuel \nconsumption? Extr. Ind. Soc. 9, 101037 (2022).\n39.\t Mahdavi, P., Martinez-Alvarez, C. B. & Ross, M. L. Why do govern\u00ad\nments tax or subsidize fossil fuels? J. Politics https://doi.org/10.1086/ \n719272 (2022).\n40.\t Clements, B. J. et al. Energy Subsidy Reform: Lessons and \nImplications (International Monetary Fund, 2013).\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nAcknowledgements\nWe thank J. Cameron for her valuable research assistance.  \nWe received helpful feedback following presentations at  \nthe University of California Santa Barbara, the University of \nCalifornia Berkeley, the University of Notre Dame, the US Agency \nfor International Development, the 2023 annual meeting  \nof the American Political Science Association and the 2023 \nmeeting of the International Political Science Association in \nBuenos Aires.\nAuthor contributions\nP.M. and M.L.R. conceptualized the study and oversaw data \ncollection. E.S. performed the data analysis and interpretation of the \nresults. M.L.R. led the manuscript writing. All authors contributed \nequally to reviewing and editing the manuscript and approved the \nfinal version.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at  \nhttps://doi.org/10.1038/s41558-025-02283-4.\nCorrespondence and requests for materials should be addressed to \nPaasha Mahdavi or Michael L. Ross.\nPeer review information Nature Climate Change thanks  \nFederica Genovese, Jakob Skovgaard and Thijs Van de Graaf for their \ncontribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\n0\n200\n400\n600\n00\n02\n04\n06\n08\n10\n12\n14\n16\n18\n20\n22\n0.2\n0.4\n0.6\n0.8\n00\n02\n04\n06\n08\n10\n12\n14\n16\n18\n20\n22\n0.4\n0.6\n0.8\n00\n02\n04\n06\n08\n10\n12\n14\n16\n18\n20\n22\n\u22120.50\n\u22120.25\n0.00\n0.25\n00\n02\n04\n06\n08\n10\n12\n14\n16\n18\n20\n22\n00\n02\n04\n06\n08\n10\n12\n14\n16\n18\n20\n22\n00\n02\n04\n06\n08\n10\n12\n14\n16\n18\n20\n22\nExtended Data Fig. 1 | Combined figures of the case of Nigeria. Panels a, b, and \nc show the price paid by consumers over time. a, The price in current Nigerian \nNaira. b, The price denominated in current US dollars, which reveals the effects \nof changes in the value of the currency. c, The price in constant 2015 US dollars, \nwhich accounts for the additional effects of inflation. d, The value of the per-liter \nimplicit tax or subsidy measured with the price-gap method. e, The onset of  \nprice reforms. f, The periods of fixity reform for which Nigeria implemented \nfloating prices.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nExtended Data Fig. 2 | Periods of Fixity Reform.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nExtended Data Fig. 3 | Increases of Monthly Prices Above 10 percent.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nExtended Data Fig. 4 | Increases of Quarterly Prices Above 25 percent.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nExtended Data Fig. 5 | Survival Estimates for Price Reform 25 percent, 2000-2015 vs 2016-2023.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nExtended Data Fig. 6 | Sensitivity Analysis: Kaplan-Meier survival rate of all subsidy reforms before (dashed orange) and after (solid blue) December 2014.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nExtended Data Fig. 7 | Leave Out Analysis. Pre and Post 2016.\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPre Paris\nPost Paris\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nExtended Data Fig. 8 | Leave Out Analysis. Fixity versus Price Reforms.\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n0.0\n0.1\n0.2\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.9\n1.0\n0\n1\n2\n3\n4\n5\nYears since Reform Onset\nSurvival Probability\nPrice Reform\nFixity Reform\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nExtended Data Fig. 9 | Daily oil price volatility, 2000-2023. Data from the U.S. Energy Information Administration: \u2018Europe Brent Spot Price FOB\u2019.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-025-02283-4\nExtended Data Table 1 | Comparison of Oil Volatility in 2000-2015 versus 2016-2023\n\n\n    Scientific Research Findings:", "answer": "We find that since 2016, and in comparison to the 2000\u20132015 period, there has been a rise in the frequency and ambition of subsidy reforms but a drop in their durability: 70% of all reforms failed in the first year, and 91% failed within 3 years. The size of subsidies rose for 12 countries in our sample and were virtually unchanged for the other 9. One-third of all failures were caused by explicit policy reversals, often in response to popular protests; for the other two-thirds, the reforms were erased by inflation, currency depreciation and rising oil prices. The most ambitious reforms were the most likely to fail. We do not know if other types of climate policy are also failing, but note that fuel subsidies are highly visible to citizens and their removal is often unpopular. Our findings imply that efforts to trim gasoline subsidies have largely been unsuccessful.", "id": 36}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 15 | April 2025 | 375\u2013384\n375\nnature climate change\nhttps://doi.org/10.1038/s41558-025-02291-4\nArticle\nNovel marine-climate interventions \nhampered by low consensus and  \ngovernance preparedness\n \nEmily M. Ogier\u2009\n\u200a\u20091,2\u2009\n, Gretta T. Pecl\u2009\n\u200a\u20091,2, Terry Hughes3, Sarah Lawless3,4, \nCayne Layton\u2009\n\u200a\u20091,2, Kirsty L. Nash\u2009\n\u200a\u20092,3 & Tiffany H. Morrison3,5,6\nNovel marine-climate interventions are now being rapidly implemented to \naddress both the causes and consequences of warming oceans. However, \nthe governance implications of proposed upscaling of such interventions \nare uncertain. We conduct a survey of 332 intervention practitioners, \nrevealing five types and 17 sub-types of interventions proposed or \ndeployed in 37 marine systems globally. Most (71%) report marine-climate \ninterventions aimed at supporting species and ecosystem adaptation, \nwith 29% aimed primarily at climate mitigation and societal adaptation. \nPerceptions of climate benefits vary widely, with low consensus across \npractitioners on the climate goals of specific interventions. Intervention \ndecision-making also remains focused on technical feasibility to meet \nminimum permitting requirements, with limited appraisal and management \nof broader ecological, cultural and social risks and benefits of intervention. \nPractitioners also warn that many marine-climate interventions are currently \nbeing tested and deployed in an under-regulated pseudo-scientific bubble.\nThe need to act in response to projected future ocean warming and \nocean change has risen on the global scientific and political agenda1,2. \nDirect climate-driven changes in marine social\u2013ecological systems \nrange from ecological effects, such as changes in species distribution, \nabundance or community biodiversity3, to social, cultural and eco-\nnomic impacts on marine-dependent societies, such as reduced food \nand livelihood security4. These changes (for example, 3% per decade \nloss in fisheries replenishment and >US$800 million in direct economic \nloss from individual marine heatwave events) are now being observed \nat unprecedented scale and intensity in marine systems5\u20137.\nSuch rapidly changing conditions form a clear and urgent mandate \nfor novel interventions to sustain marine ecosystems and dependent \nsocieties2,8\u201310. Marine interventions can be understood as deliber-\nate, planned actions in marine systems to achieve desired outcomes \nor goals. Historically, marine interventions have been designed to \nconserve and restore species or local ecological communities and the \necosystem services they generate11,12 or improve coastal community \nwell-being through securing rights and strengthening livelihoods13,14. \nInterventions to substantively contribute to climate goals such as \nmitigation or adaptation have been either secondary or absent.\nToday, oceans are on the front line of new planned climate actions. \nThese interventions are novel both in their deployment of new and \noften untested technologies (for example, genomics, altering ocean \nbiogeochemistry, rights-based frameworks) and in new oceanic, cli-\nmatic and social conditions. In pursuing climate mitigation, the ocean \nis now a frontier for both clean energy creation (for example, off-\nshore wind energy) and carbon removal (for example, ocean alkalinity \nenhancement)15,16 required to meet the Paris Agreement17. In enabling \nclimate adaptation, high-profile interventions include protecting \nclimate refugia to conserve specific marine ecosystems18, climate \nReceived: 30 July 2024\nAccepted: 20 February 2025\nPublished online: 3 April 2025\n Check for updates\n1Institute of Marine and Antarctic Studies, University of Tasmania, Hobart, Tasmania, Australia. 2Centre for Marine Socioecology, University of Tasmania, \nHobart, Tasmania, Australia. 3College of Science and Engineering, James Cook University, Townsville, Queensland, Australia. 4Australian Institute of \nMarine Science, Townsville, Queensland, Australia. 5School of Geography, Earth and Atmospheric Sciences, University of Melbourne, Parkville, Victoria, \nAustralia. 6Environmental Policy Group, Wageningen University Research, Wageningen, the Netherlands. \n\u2009e-mail: Emily.Ogier@utas.edu.au\n\nNature Climate Change | Volume 15 | April 2025 | 375\u2013384\n376\nArticle\nhttps://doi.org/10.1038/s41558-025-02291-4\nTo understand the challenges and opportunities for responsible \ngovernance of interventions, our analysis extended these responsible \nresearch and innovation frameworks by matching common operational \nmarine governance arrangements with intervention risks. Examples \nof such governance arrangements include frameworks for assessing \nfeasibility, risk and impact46,47, mechanisms for public appraisal48,49, \noperational oversight30,50 and marine-climate policy leadership33,51,52. \nCommon risks of marine-climate interventions include ineffective-\nness relative to desired goal, social\u2013ecological harm, negligence in \naddressing harm, public distrust of interventions and opportunity \ncost in intervening (Table 1 and Supplementary Table 1).\nOur typology of interventions and framework of governance \npreparedness enabled us to identify, compare and synthesize key \nchallenges and opportunities in ensuring responsible governance \nof interventions. In doing so, our intended audiences are policy-\nmakers, financiers and scientists engaged in intervention design, \ndecision-making, investment and implementation. Such insights are \nurgently needed by these groups to facilitate the strategic selection, \ndeployment and ongoing oversight of appropriate interventions at \nthe scale required to sustain marine estates and coastal communities \nthroughout the changing climate.\nResults\nFrom our survey sample of 332, we found that the emerging global \ncommunity of intervention practitioners was dominated by interven-\ntion scientists (58%), followed by intervention policymakers (14%) and \nnon-governmental organization (NGO) practitioners (14%). Repre-\nsentatives of Traditional Owners and First Nations and of local com-\nmunity or industry sectors accounted for 1% and 3%, respectively, of \nour sample. Respondents\u2019 engagement in marine-climate interventions \nranged from involvement in research development, deployment and \nmonitoring (63%) to programme design, project management and site \nimplementation (13%), policy development, regulation and oversight \n(10%), funding (4%) and other (9%).\nNot-for-profit organizations, government agencies and science \norganizations were identified as leaders of intervention best practice \nby survey respondents (38%, 31% and 30% respectively, n\u2009=\u200982).\nIntervention types and awareness\nWe found an array of marine-climate interventions being proposed, \ntrialled and deployed globally (Table 2). Practitioners\u2019 awareness of \nnovel marine-climate interventions varied substantially across the \ntypes and sub-types identified (Table 2). Overall, practitioners were \nmost aware of interventions concerned with marine bioengineering \nand coastal and marine restoration. Awareness was highest for artificial \nhabitat manipulation (79%, n\u2009=\u2009332), a sub-type of marine bioengineer-\ning intervention, followed by regrowing targeted underwater species \nadaptation planning for coastal communities (for example, the Samoa \nOcean Strategy 2020\u20132030 (ref. 19)), adaptation of marine manage-\nment (for example, adaptive fisheries management plans20) and adap-\ntation of specific marine ecosystem processes (for example, breeding \nthermally tolerant marine species genotypes21).\nHowever, systematic and comparative understanding of \nmarine-climate intervention development and deployment remains \ncritically low22. In particular, there has been limited empirical investi-\ngation of the \u2018pacing problem\u201923 whereby innovation outpaces govern-\nance preparedness to anticipate and responsibly manage risk across \nthe range of novel marine-climate interventions currently active or \nunder consideration22,24\u201327. Governance preparedness involves public \nand private institutions and actors engaged in processes of respon-\nsible rule, steerage and guidance28. Any lag in the responsiveness of \ngovernance regimes is problematic because the rapid emergence and \nplanned upscaling of novel marine-climate interventions29 presents \nan array of risks for marine ecosystems30,31 and coastal societies and \nrightsholders22,32,33 at local, regional and climate system and climate \npolicy scales (Table 1). Implicated governance action arenas include \nmarine and coastal conservation, tenure and rights of local commu-\nnities, small-scale fishers and Indigenous peoples, ocean economy \ndevelopment and decarbonization.\nTo track the extent to which governance arrangements are keep-\ning pace with novel marine-climate interventions, we surveyed 332 \npractitioners. We used an online questionnaire to survey the emerging \nglobal community of intervention practitioners to ascertain what types \nof interventions are being planned or deployed, how interventions are \nbeing designed, their geographic distribution and stage of develop-\nment, types of climate goals and benefits pursued and arrangements to \nresponsibly govern intervention. On the basis of our results, we devel-\noped a typology of major types and sub-types of novel marine-climate \ninterventions and cross-referenced them against recent authoritative \nstudies (for example, reviews by the National Academies of Sciences, \nEngineering, and Medicine10,34).\nIn ascertaining the degree to which responsible governance \narrangements were present, we adapted and extended existing frame-\nworks for responsible research and innovation to the governance realm. \nTo date, the need for responsible research and innovation (as set out \nby refs. 31,35) has been met largely by developing codes of conduct for \nspecific types of experimental research (for example, scientific codes \nof conduct for research on marine carbon dioxide removal30,36\u201338). \nWhile such scientific codes are necessary39\u201342, they have limited remit \nor powers beyond experimental-scale research on single types of inter-\nventions. The additional governance necessary to facilitate and steer \ndeployment of innovations at scale\u2014to both ensure no undue harm \nand deliver ecological and social benefits (that is, \u2018responsible govern-\nance\u2019)\u2014is less well understood43\u201345.\nTable 1 | Governance challenges of marine-climate intervention risks\nIntervention risk\nOutcome\nScale of outcome\nIllustrative stories\nIneffectiveness\nFailure to achieve stated marine-climate mitigation or \nadaptation goal\nLocal/community scale of deployment\n30,37,46,91,92\nHarm\nUnintended cultural, social or ecological harm\nLocal/community scale of deployment\n93\u201395\nNegligence\nFurther harm due to responsibility gaps in amelioration or \nrestitution of intervention harms\nLocal/community scale of deployment\n49,96\nDistrust\nPublic distrust/rejection of intervention\nLocal/community scale of deployment\n32\nGlobal/system-wide\n33,97\nOpportunity cost\nHigh opportunity cost/crowding out of other critical collective \nactions for marine-climate mitigation or adaptation\nLocal/community scale of deployment\n9,98\nGlobal/system-wide\n56,83,84,99\nScale refers to scale at which outcome is directly observable.\n\nNature Climate Change | Volume 15 | April 2025 | 375\u2013384\n377\nArticle\nhttps://doi.org/10.1038/s41558-025-02291-4\n(for example, coral and kelp), a sub-type of coastal and marine restora-\ntion intervention (76%). Other sub-types of interventions with elevated \nlevels of awareness included regrowing targeted coastal species (for \nexample, mangroves), assisted evolution of marine species (for exam-\nple, coral hybridization) and natural stabilization of reefs and coasts \n(66%, 66% and 62%, respectively). By contrast, awareness of marine \ngeoengineering interventions was low, with the notable exception of \ninterventions to shade and cool water and habitat (40%). Awareness of \nmarine socio-institutional capacity-building interventions was excep-\ntionally low (4%) (Table 2).\nIntervention scientists and intervention policymakers had the \ngreatest breadth of awareness of different types of interventions (5.1 \nand 5.0 types, respectively, on average), while NGO practitioners, \nrepresentatives of Traditional Owners and First Nations and of local \ncommunity or industry sectors were aware of relatively fewer (3.9, \n4.3, 3.3 and 4.8 types, respectively). Due to the emergent nature of the \nintervention community and likely standard error, the results of our \nsurvey offer an initial exploratory analysis of the range of intervention \npractitioners, types and locations.\nGeographical distribution and stage of development\nInterventions were occurring in multiple regions globally, noticeably \nclustered in locations that are warming faster than the global average \n(that is, marine hotspot locations; see ref. 53). Respondents reported \nin detail, interventions (n\u2009=\u2009309) that are distributed across 37 differ-\nent specific marine or coastal locations and in most oceans and major \nseas (Fig. 1a).\nOceans and seas where reported intervention activity was greatest \nwere Australia\u2019s tropical waters (16%), Australia\u2019s and New Zealand\u2019s \ntemperate waters (15%), the North Pacific (15%) and the wider Caribbean \n(10%). Almost all of the interventions reported as active in Australia\u2019s \ntropical waters were occurring in the Great Barrier Reef region (92%). \nThese interventions were predominantly to support coral reef restora-\ntion, for example, through re-seeding coral, breeding of heat-resistant \ncoral symbionts and coral reef habitat restoration and creation (6%, 4% \nand 2% of all reported interventions). Multiple types of interventions \nactive within the same ocean region were reported almost without \nexception, with 96% of the interventions occurring in the same ocean \nregion as at least one other type of intervention (Fig. 1b).\nIn terms of development, the majority of interventions identi-\nfied were at pilot or full implementation stage (46% and 38%, respec-\ntively, n\u2009=\u2009207 interventions; Fig. 1c) while 16% were at concept stage. \nDevelopment was most progressed for marine bioengineering and \ncoastal and marine restoration interventions. Specific interventions \nreported as having the highest level of technical readiness and devel-\nopment included artificial manipulation of habitats and regrowing \nof targeted coastal species (53% and 65% at implementation stage, \nrespectively; Fig. 1c).\nStated climate goals\nClimate goals pursued through marine intervention included both \nmitigation and adaption, alongside non-climate goals (that is, bio-\ndiversity protection). The most stated climate goal was to increase \nthe biophysical adaptation or resilience of local marine ecosystems \nto climate-driven changes (57% of interventions, n\u2009=\u2009211; Fig. 2). This \nclimate goal was being pursued across all five intervention types, most \ncommonly through coastal and marine restoration (for example, kelp \nforest and seagrass bed restoration), followed by marine bioengineer-\ning (for example, assisted evolution of coral). Notably, biophysical \nadaptation and resilience was also being pursued through marine \nsocio-institutional capacity building (for example, development of \nclimate-adaptive fisheries management regimes). However, the goal \nTable 2 | Types and levels of awareness of novel marine-climate interventions (n\u2009=\u2009332 respondents)\nType\nDetail\nSub-type\nAwareness (%)\nMarine geoengineering\nManipulation of the oceanic and atmospheric climate to \nincrease uptake and removal of atmospheric carbon or \nmitigate direct heating effects\nShading and cooling water and habitat\n40\nOcean fertilization\n3\nOcean alkalinity enhancement\n2\nArtificial upwelling and downwelling\n2\nMarine bioengineering\nManipulation of marine evolutionary processes and \necosystem function and condition\nArtificial habitat manipulation\n79\nAssisted evolution of marine species\n66\nAssisted migration and colonization of marine \nspecies\n34\nControlling climate-exacerbated destructive \nspecies\n2\nCoastal and marine restoration\nRepairing a climate-impacted catchment-to-marine \necosystem or population\nRegrowing targeted underwater species\n76\nRegrowing targeted coastal species\n66\nNatural stabilization of reefs and coasts\n62\nCatchment habitat restoration\n1\nMarine social\u2013institutional \ncapacity building\nEnabling communities and organizations to make \nmarine-climate decisions and redress climate impacts\nAnticipatory marine-climate science\n1\nClimate-resilient marine protected area \nmanagement\n1\nCoastal adaptation community planning\n1\nClimate-adaptive fisheries management\n1\nBiological marine carbon \ndioxide removal\nCreation or restoration of carbon sinks from natural \nmarine resources\nAquaculture for carbon sequestration\n56\nFive broad intervention types were apparent: marine geoengineering, marine bioengineering, coastal and marine restoration, marine social\u2013institutional capacity building and biological \nmarine carbon dioxide removal. Within these five broad types, 17 sub-types were apparent on the basis of their treatment and primary goal (that is, restorative in the case of coastal and \nmarine restoration; adaptive in the case of marine bioengineering) and their focal sub-system (that is, catchment-to-ocean in the case of coastal and marine restoration; air\u2013ocean exchange \nprocesses in the case of marine geoengineering). See Supplementary Table 2, for detailed typology.\n\nNature Climate Change | Volume 15 | April 2025 | 375\u2013384\n378\nArticle\nhttps://doi.org/10.1038/s41558-025-02291-4\nof social adaptation and resilience to climate change was reported \nfor only 3% of interventions. For 5% of the interventions described, no \nclimate goal was identified despite the survey design, which focused on \ninterventions in the context of climate-driven change in oceans (Fig. 2).\nClimate mitigation was also being pursued across all interven-\ntion types, with carbon removal (rather than emissions avoidance, \nfor example) the second-most stated goal (27%; Fig. 2). Mitigation  \ninterventions ranged from those designed to intervene in carbon \ncycles via marine geoengineering (for example, ocean alkalinity \nenhancement to increase air\u2013ocean carbon exchange) to those work-\ning on biological mechanisms of carbon sequestration (that is, aqua-\nculture for carbon sequestration). Some respondents also reported \nmarine bioengineering (for example, heat-resistant kelp breeding \nprogrammes) and coastal and marine restoration (for example, sea-\ngrass meadow restoration) as aiming for carbon sequestration as a \nsecondary goal.\nFor some interventions, there was substantial variation in the \nclimate goals identified. For example, for artificial habitat manipula-\ntion, stated climate goals included increasing biophysical adapta-\ntion and resilience (58%), carbon removal (19%), no climate goal (19%) \nand increasing social adaptation and resilience (2%; Supplementary \nTable 3). Similarly, for regrowing coastal and underwater species, the \nvariety of stated climate goals included increasing biophysical adapta-\ntion and resilience (63%), carbon removal (27%), no climate goal (5%), \nincreasing social adaptation and resilience (3%) and raising climate \nawareness (1%; Supplementary Table 2).\nc\nCoastal adaptation community planning\nClimate-resilient marine protected area management\nClimate-adaptive aquaculture development\nClimate-adaptive fisheries management\nAnticipatory marine-climate science\nCatchment habitat restoration\nNatural stabilization of reefs and coasts\nRegrowing targeted coastal species\nRegrowing targeted underwater species\nArtificial manipulation of habitats\nAssisted evolution of marine species\nAssisted migration and colonization of marine species\nControlling climate-exacerbated destructive species\nShading and cooling water and habitats\nOcean alkalinity enhancement\nOcean fertilization\nAquaculture for carbon sequestration\nBiological marine carbon\ndioxide removal\n0\n6\n11\n17\nGlobal\nNorth Pacific\nSouth Pacific\nAustralia\n(tropical)\nAustralia and\nNew Zealand\n(temperate)\nSouth Atlantic\nMediterranean\nRed Sea\nIndian Ocean\nMarine social\u2013institutional\ncapacity building\n0\n6\n11\n17\nGlobal\nNorth Pacific\nSouth Pacific\nAustralia\n(tropical)\nAustralia and\nNew Zealand\n(temperate)\nSouth Atlantic\nMediterranean\nRed Sea\nIndian Ocean\nCoastal and marine\nrestoration\n0\n6\n11\n17\nGlobal\nNorth Pacific\nSouth Pacific\nAustralia\n(tropical)\nAustralia and\nNew Zealand\n(temperate)\nSouth Atlantic\nMediterranean\nRed Sea\nIndian Ocean\nMarine\nbioengineering\n0\n6\n11\n17\nGlobal\nNorth Pacific\nSouth Pacific\nAustralia\n(tropical)\nAustralia and\nNew Zealand\n(temperate)\nSouth Atlantic\nMediterranean\nRed Sea\nIndian Ocean\nMarine\ngeoengineering\n0\n6\n11\n17\nGlobal\nNorth Pacific\nSouth Pacific\nAustralia\n(tropical)\nAustralia and\nNew Zealand\n(temperate)\nWider Caribbean \nNorth Atlantic\nWider Caribbean \nNorth Atlantic\nWider Caribbean \nNorth Atlantic\nWider Caribbean \nNorth Atlantic\nWider Caribbean \nNorth Atlantic\nSouth Atlantic\nMediterranean\nRed Sea\nIndian Ocean\nb\nType of intervention:\nMarine geoengineering\nMarine bioengineering\nCoastal and marine restoration\nMarine social\u2212institutional capacity building\nBiological marine carbon dioxide removal\nGlobal\nn = 227\nAustralia and New Zealand\n(temperate)\nn = 46\nAustralia (tropical)\nn = 48\nIndian Ocean\nn = 12\nSouth Atlantic\nn = 3\nRed Sea\nn = 1\nMediterranean\nn = 11\nNorth Atlantic\nn = 22\nWider Caribbean\nn = 31\nSouth Pacific\nn = 7\nNorth Pacific\nn = 46\na\nCoastal adaptation community planning\nClimate-resilient marine protected area management\nClimate-adaptive aquaculture development\nClimate-adaptive fisheries management\nAnticipatory marine-climate science\nCatchment habitat restoration\nNatural stabilization of reefs and coasts\nRegrowing targeted coastal species\nRegrowing targeted underwater species\nArtificial manipulation of habitats\nAssisted evolution of marine species\nAssisted migration and colonization of marine species\nControlling climate-exacerbated destructive species\nShading and cooling water and habitats\nOcean alkalinity enhancement\nOcean fertilization\nAquaculture for carbon sequestration\n0\n12\n24\nConcept\nImplementation\nPilot\n0\n12\n24\nConcept\nImplementation\nPilot\n0\n12\n24\nConcept\nImplementation\nPilot\n0\n12\n24\nConcept\nImplementation\nPilot\n0\n12\n24\nConcept\nImplementation\nPilot\nBiological marine carbon\ndioxide removal\nMarine social\u2013institutional\ncapacity building\nCoastal and marine\nrestoration\nMarine\nbioengineering\nMarine\ngeoengineering\nFig. 1 | Global distribution and development of marine-climate interventions. a,b, Global distribution (n = 309 interventions) across ocean regions by major types \n(a) and sub-types (b). c, Stages of development (n = 207 interventions) by marine-climate intervention sub-type. Panel a generated using rnaturalearth v.1.0.1.9000.\n\nNature Climate Change | Volume 15 | April 2025 | 375\u2013384\n379\nArticle\nhttps://doi.org/10.1038/s41558-025-02291-4\nGovernance of intervention risk\nGaps in the use of available governance arrangements were notable \n(Fig. 3) and included the low use of data co-produced with Indigenous \nPeoples and Local Communities (11% of interventions identified, \nn\u2009=\u2009233), ethics assessment (14%), accountability and transparency \nmechanisms (15%), social impact mitigation measures (15%) and mecha-\nnisms to recognize and address unintended social impact (21%), and \nstrategic leadership capacity (19% of interventions) (Supplementary \nTable 4).\nPreparedness to responsibly govern specific intervention risks \n(Table 1 and Supplementary Tables 5 and 10) was varied. The risk of \nineffectiveness was the risk most frequently addressed by applying \navailable government arrangements. Levels of use across the available \narrangements ranged between 57 and 70% of all interventions, from \nuse of multiple data types (70% of interventions) and data sources for \nfeasibility assessment (65%) to multiple forms of assessing implementa-\ntion risk (59%; Supplementary Table 4).\nBy contrast, levels of use of available arrangements to govern the \nrisk of unintended harms ranged between 32 and 50% of interventions. \nLevels of use of available arrangements to govern risks of public distrust \nin interventions were even lower (between 28 and 50%), and for risks \nof negligence in addressing intervention effects, the range was lower \nagain (between 28 and 35%). Concerningly, arrangements to govern the \nrisk of opportunity cost in pursuing a given intervention were present \nin only 19% of interventions, implying that most interventions are not \nbeing assessed against one another (Supplementary Table 4).\nUnsurprisingly, marine social\u2013institutional capacity building was \nthe intervention type where responsible governance arrangements \nwere most frequently applied across all risks (44% mean level of pres-\nence). Arrangements for governing risks of ineffectiveness, harms, \nnegligence, distrust and opportunity cost were present at rates of 63%, \n50%, 50%, 32% and 30%, respectively. By contrast, marine bioengineer-\ning was the type of intervention for which responsible governance \narrangements were least often in place (33% mean level of presence \nacross all risks; Supplementary Table 6).\nIn reporting on the risks and opportunities of an intervention, \npractitioners (n\u2009=\u2009130) held divergent positions on the necessity of \nfurther responsible governance arrangements. While 23% of practition-\ners called for increased levels of governmental appraisal, planning, \ncoordination and regulation, a contrasting 16% called for reduction \nof such arrangements, which they perceived to be hindering rates of \nimplementation and upscaling. Practitioners requesting strengthened \ngovernance asked for increased rigour in technical feasibility, risk and \nimpact assessment (25%), increased policy and community support, \nincluding funding (22%), reduced scientific uncertainty combined \nwith increased science coordination (7%), increased climate mitigation \nand mechanisms for addressing other underlying stressors (4%) and \ngreater inclusion of Indigenous Peoples and Local Communities (3%).\nDiscussion\nAn open question in marine-climate research is what the pro-\nposed upscaling of novel marine-climate interventions means for \nclimate action and long-term well-being of marine systems and \nmarine-dependent people. Most governance arrangements in place \nare limited to formal risk assessments and regulatory and permitting \nprocesses27,54,55 based on retrospective understandings and technolo-\ngies operating under high levels of uncertainty56. The observed low \nlevel of governance preparedness to responsibly govern the risks posed \nby novel and experimental marine-climate interventions indicates \nthat the pacing problem is indeed present. Responsible governance \nregimes are needed to avoid risks of maladaptation and the potentially \nhigh opportunity cost of marine-climate interventions. Fortunately, \ngaps in responsible governance constitute a resolvable problem where \npublic-interest actors have principles to guide them and, increasingly, \nthe operational arrangements and practices to mandate and use35,36. \nOur analysis of practitioner observations highlights multiple reasons \nfor and opportunities to address this gap.\nFirst, marine-climate intervention remains science driven with \nlimited involvement of public institutions or communities and only \n14% of practitioners in our sample working within government. By \ncontrast, the biophysical science sector\u2019s major role in novel interven-\ntion development and in standard-setting for intervention research \nis likely to explain the emphasis on technical feasibility assessment \nas the main form of intervention appraisal. The relative absence of \nstrong public-interest actors and processes, combined with the lim-\nited competencies of conventional science organizations39\u201342,57, may \nalso explain the low levels of use of responsible governance arrange-\nments (Fig. 3). Not surprisingly, organizations in the NGO and research \nMarine geoengineering\nMarine bioengineering\nCoastal and marine\nrestoration\nMarine social\u2013institutional\ncapacity building\nBiological carbon dioxide\nremoval\nCarbon emissions\navoidance\nCarbon ofsetting\nCarbon removal\nSocial adaptation\nand resilience\nNo climate goal\nClimate awareness\nBiophysical adaptation\nand resilience\nFig. 2 | Climate-related goals of marine interventions identified by survey respondents (n\u2009=\u2009211 interventions). Width of strand indicates the number of interventions \nfor which goal was identified. More than one goal could be identified for each intervention described. See Supplementary Table 3 for a detailed list of goals.\n\nNature Climate Change | Volume 15 | April 2025 | 375\u2013384\n380\nArticle\nhttps://doi.org/10.1038/s41558-025-02291-4\nsectors were recognized as best-practice leaders of interventions by \n68% of the survey respondents. Actors within these networks clearly \nwield considerable potential influence and include \u2018impact inves-\ntors\u201958,59, ocean philanthropic funders60\u201362, speculators attracted by \nthe establishment of carbon markets63 and scientists whose formative \nrole in experimentation and intervention innovation renders them \nde facto governors35,64. Engagement of public actors in the formative \nstages of marine-climate intervention development is critical because \nof risks of sociocultural harms and the potential high opportunity \ncost of upscaling. Opportunities to do so include adoption of proven \nparticipatory and deliberative early public engagement frameworks \n(see refs. 65,66) and the use of bioethical assessment frameworks67 by \nfunders and regulators.\nSecond, interventions using bioengineering or restorative \ntechniques have the highest level of awareness and were the types \nof marine-climate interventions with the highest rates of implemen-\ntation and widest geographic distribution across our sample of 322 \npractitioners. While other types of marine-climate interventions were \nreported (for example, marine geoengineering, biological marine car-\nbon dioxide removal, marine social\u2013institutional capacity building), \nawareness was lower, which is likely to be at least partly explained by \ntheir more limited geographical distribution and their development \nIntervention\nrisk\nResponsible\ninnovation\ndimension\nResponsible\ngovernance\narrangement\nType\nof\nintervention\nTechnical feasibility assessment\nMultiple forms of assessment of implementation risk\nMultiple data sources for feasibility assessment\nMultiple data types for feasibility assessment\nEthics assessment\nConsideration of social risks and impacts\nAssessment of negative impacts\nAcceptability to stakeholders considered in feasibility assessment\nUse of public consultation data in feasibility assessment\nUse of data co-produced with IPLC in risk assessment\nFormal consideration of trade-ofs between risks and benefits\nForm(s) of formal oversight in addition to regulations\nSocial impact oversight mechanisms\nEnvironmental impact oversight mechanisms \nAccountability and transparency oversight mechanisms \nBiophysical impact mitigation measures in place\nSocial impact mitigation measures in place\nStrategic capacity to allow progress while constraining risk\nStakeholder consultation and/or public survey data in risk assessment\nDeliberation opportunities for IPLC rights and/or interest holders\nDeliberation opportunities for IPLC rights and/or interest holders\nAssessment of cumulative impacts\nMonitoring for unintended negative impacts\nUse of robust social data to assess impact\nHarm\nInefectiveness\nAnticipation\nInclusion and\nanticipation\nDistrust\nNegligence\nOpportunity cost\nReflexivity\nProportion\nGeoengineering\nBioengineering\nCoastal and marine\nrestoration\nSocial\u2013institutional\ncapacity building\nBiological carbon\ndioxide removal\n0\n0.50\n1.00\nResponsiveness\nInclusion and\nreflexivity\nFig. 3 | Use of responsible governance arrangements to manage anticipated risks of novel marine-climate interventions (n\u2009=\u2009233 interventions). See \nSupplementary Tables 1 and 5 for definitions and data. IPLC, Indigenous Peoples and Local Communities.\n\nNature Climate Change | Volume 15 | April 2025 | 375\u2013384\n381\nArticle\nhttps://doi.org/10.1038/s41558-025-02291-4\nbeing more commonly at the concept or pilot stage. Funders and pub-\nlic policy decision-makers charged with large-scale climate action \nface choices between available marine-climate interventions. The \ninformation asymmetry between intervention types observed in our \nstudy may limit the possibility of decision-makers being able to make \ninformed choices among the full range of viable interventions. Increas-\ning marine-climate action literacy among decisions-makers is a criti-\ncal means of addressing this challenge, starting with sharing of more \naccessible science-based information about climate action options  \n(for example, Reef Adapt tool (https://www.reefadapt.org)).\nThird, despite the low awareness of interventions to build social \nand institutional capacity for climate adaptation and mitigation \ndetected in our results, these intervention types warrant additional \nconsideration. The intervention examples reported by survey respond-\nents highlighted important ways forward, including restitution and \nformalization of marine and coastal tenure for Indigenous Peoples \nand Local Communities, partnerships in climate intervention science \nprogrammes and coastal adaptation planning with local communities. \nOur results also revealed that practitioners are increasingly recognizing \nthe need for social\u2013institutional capacity to reduce intervention risks \nof sociocultural harms, negligence and distrust. Social\u2013institutional \ncapacity can be built through rights-based frameworks (see refs. 68\u201371 \nand the Turning Tides facility (https://turningtidesfacility.org)) and \nregional and community-led programmes for coastal climate adapta-\ntion planning (see ref. 72) and by enabling local rule-making for climate \naction (see ref. 35). Such institutions support a priori recognition of \nequity, cultural rights and interests before conceptualization and \nassessment of an intervention. By coupling regional and local-scale \nsocio-institutional capacity-building interventions with local bioengi-\nneering, geoengineering and restorative interventions, such initiatives \ncould be \u2018bright spots\u2019 in which social, cultural and biophysical goals \nare mutually recognized and pursued (see refs. 14,20,72,73).\nFourth, marine-climate interventions are occurring in all major \nocean regions, and within each of these regions more than one specific \nintervention sub-type is under way. For example, in the North Atlantic \nand in tropical waters adjacent to Australia, the range of interventions \ndescribed includes those in all intervention types. This co-occurrence \nmay present additional under-recognized governance challenges for \npolicy- and decision-makers who are faced with both prioritizing across \ninterventions and managing their cumulative and synergistic impacts \nwithin a single marine region. We found that formal consideration of \ntrade-offs between intervention risks and benefits and assessment of \ncumulative impacts are not widely occurring governance practices \n(reported as occurring for only 31 and 32% of interventions, respec-\ntively). The results suggest that many decision-makers are materially \nunderprepared for these strategic challenges, which extend beyond \nmanaging the risks posed by single interventions22,30,32. Tools such as \nanticipatory social and cumulative impact assessment frameworks74,75 \nare available to support decision-makers to integrate their assessment \nand management of discrete climate actions into broader strategic \nassessment and planning at the marine estate or regional communi-\nties\u2019 level (for example, the IPCC\u2019s shared socioeconomic pathways76, \nwhich could be adapted for marine regions).\nFifth, our results highlighted that systematic and comparative \nassessment of marine-climate interventions continues to be confounded \nby a lack of clarity and low consensus in stated climate-related interven-\ntion goals. In some cases, interventions such as seaweed afforestation \nwere reported to be pursuing both climate mitigation and adaptation \ngoals simultaneously. We observed frequent use of the term \u2018resilience\u2019 \nby many practitioners in self-reporting the ecological and social goals \nand benefits of their chosen intervention in lieu of providing more \nintervention-specific detail to which they were invited. Resilience has \nbeen widely critiqued for being conceptually vague, ignoring power \nand politics, and being operationally weak77,78. This lack of clarity on \nresilience obfuscates efforts to assess effectiveness of interventions \nand fails to deal with power asymmetries and inequity in pursing cli-\nmate actions79. Rectifying this lack of climate goal precision (and there-\nfore accountability) will depend on funders and public-interest actors \ndemanding uptake of principles and codes of practice (see High-Quality \nBlue Carbon Principles80) and, increasingly, standards for monitoring, \nreporting and evaluation of specific intervention effects30,37,81,82.\nFinally, practitioners themselves are among those who query the \nrigour of technical feasibility assessment and evaluation of impact of \ninterventions against intended and claimed benefits and co-benefits. \nOne-quarter of survey respondents raised these concerns. Indeed, \nemerging social science suggests that entrepreneurial hype com-\nbined with an absence of strong monitoring, evaluation and reporting \nrequirements can produce perverse outcomes whereby speculative \ninterventions are prioritized over effective ones8,63,83\u201385. Our results \nunderscore that claims of multiple conservation and climate goals and \nco-benefits combined with low levels of technical feasibility assessment \n(50% of reported interventions) and very low levels of accountability \nand oversight (16%) and strategic capacity to steer innovation and man-\nage marine-climate intervention risk (18%) increase the likelihood of \npoor choices, contributing to both unintended negative consequences \nand missed opportunities in climate mitigation and adaption. The prac-\ntitioner concerns are therefore valid, while at the same, it is important \nto acknowledge that carbon removal and climate adaptation goals \nmay be feasible in some cases for specific species and ecosystems86.\nIn conclusion, our global survey and subsequent statistical analysis \nrevealed that, broadly, governance of novel marine-climate interven-\ntions is occurring in both a \u2018scientific bubble\u2019 and an \u2018institutional \nvoid\u201987,88. Future research could incorporate inferential analyses to \nexplore distinctions across practitioner groups, regions and jurisdic-\ntions and interactions among factors affecting practitioner awareness, \nintervention impacts and operational governance practices in place. \nSuch analysis would support further development of responsibly \ngoverned and context-appropriate marine-climate actions.\nOvercoming the pacing problem through timely uptake and for-\nmalization of available or emergent responsible governance practices \n(Fig. 3) requires continued scientific leadership to ensure that technical \ndesign, assessment and monitoring of marine-climate interventions \nis adopted and sufficiently rigorous. At the same time, the bioethical, \nanticipatory and reflexive requirements of responsible governance \ndemand that many other actors are better engaged alongside these \npractitioner communities89\u2014to build legal and institutional capacities \nand to ensure that governance of multiple marine-climate interventions \nis underpinned by climate action and justice principles90.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-025-02291-4.\nReferences\n1.\t\nIPCC Climate Change 2021: The Physical Science Basis (eds \nMasson-Delmotte, V. et al.) (Cambridge Univ. Press, 2021);  \nhttps://doi.org/10.1017/9781009157896\n2.\t\nIPCC: Summary for Policymakers. In Climate Change 2022: \nImpacts, Adaptation and Vulnerability (eds P\u00f6rtner, H.-O. et al.) \n(Cambridge Univ. Press, 2022); https://doi.org/10.1017/ \n9781009325844.001\n3.\t\nPecl, G. T. et al. Biodiversity redistribution under climate change: \nimpacts on ecosystems and human well-being. Science 355, \neaai9214 (2017).\n4.\t\nBarnett, J. Global environmental change II: political economies of \nvulnerability to climate change. Prog. Hum. Geogr. 44, 1172\u20131184 \n(2020).\n\nNature Climate Change | Volume 15 | April 2025 | 375\u2013384\n382\nArticle\nhttps://doi.org/10.1038/s41558-025-02291-4\n5.\t\nHughes, T. P. et al. Coral reefs in the Anthropocene. Nature 546, \n82\u201390 (2017).\n6.\t\nSmith, K. E. et al. Socioeconomic impacts of marine heatwaves: \nglobal issues and opportunities. Science 374, eabj3593 (2021).\n7.\t\nBindoff, N. L. et al. in Special Report on the Ocean and Cryosphere \nin a Changing Climate (eds P\u00f6rtner, H.-O. et al.) 447\u2013587 \n(Cambridge Univ. Press, 2019).\n8.\t\nMorrison, T. H. et al. Save reefs to rescue all ecosystems. Nature \n573, 333\u2013336 (2019).\n9.\t\nSovacool, B. K. et al. Coral reefs, cloud forests and radical climate \ninterventions in Australia's Wet Tropics and Great Barrier Reef. \nPLOS Clim. 2, e0000221 (2023).\n10.\t A Research Strategy for Ocean-Based Carbon Dioxide Removal and \nSequestration (NASEM, 2002) https://doi.org/10.17226/26278\n11.\t\nProber, S. M. et al. Shifting the conservation paradigm: a synthesis \nof options for renovating nature under climate change. Ecol. \nMonogr. 89, e01333 (2019).\n12.\t Saunders, M. I. et al. Bright spots in coastal marine ecosystem \nrestoration. Curr. Biol. 30, R1500\u2013R1510 (2020).\n13.\t Sebastian, P., Sparks, L. D., Resolute, P. & Prasetijo, R. Connecting \ncommunities to coral reefs: a socio-ecological perspective on \ncoral restoration programs in a remote marine protected area.  \nJ. Coast. Conserv. 28, 39 (2024).\n14.\t Baker, D. M. et al. Improving human well-being outcomes in \nmarine protected areas through futures thinking. One Earth 6, \n1286\u20131290 (2023).\n15.\t Gattuso, J.-P. et al. Ocean solutions to address climate change \nand its effects on marine ecosystems. Front. Mar. Sci. 5, 18  \n(2018).\n16.\t Gattuso, J.-P. et al. The potential for ocean-based climate action: \nnegative emissions technologies and beyond. Front. Clim. 2, \n575716 (2021).\n17.\t Brun, A. Conference diplomacy: the making of the Paris \nagreement. Pol. Govern. 4, 115\u2013123 (2016).\n18.\t Morrison, T. H. et al. Advancing coral reef governance into the \nanthropocene. One Earth 2, 64\u201374 (2020).\n19.\t Samoa Ocean Strategy 2020\u20132030, Integrated Management for a \nHealthy and Abundant Future of Samoa\u2019s Ocean (Government of \nSamoa, 2020).\n20.\t Bell, J. D. et al. Pathways to sustaining tuna-dependent Pacific \nIsland economies during climate change. Nat. Sustain. 4, \n900\u2013910 (2021).\n21.\t Anthony, K. et al. New interventions are needed to save coral \nreefs. Nat. Ecol. Evol. 1, 1420\u20131422 (2017).\n22.\t Cooley, S. R. et al. Sociotechnical considerations about ocean \ncarbon dioxide removal. Annu. Rev. Mar. Sci. 15, 41\u201366 (2022).\n23.\t Marchant, G. E. in The Growing Gap Between Emerging \nTechnologies and Legal-Ethical Oversight: The Pacing Problem \n(eds Marchant, G. E. et al.) 199\u2013205 (Springer, 2011); https://doi.org/ \n10.1007/978-94-007-1356-7_13\n24.\t Cherry, T. L. et al. Climate cooperation in the shadow of solar \ngeoengineering: an experimental investigation of the moral \nhazard conjecture. Environ. Polit. 32, 362\u2013370 (2023).\n25.\t Pamplany, A., Gordijn, B. & Brereton, P. The ethics of \ngeoengineering: a literature review. Sci. Eng. Ethics 26,  \n3069\u20133119 (2020).\n26.\t Dalby, S. Ecological, security, and change in the Anthropocene. \nBrown J. World Aff. 13, 155\u2013164 (2007).\n27.\t Webb, R. M. International Governance of Ocean-Based \nCarbon Dioxide Removal: Recent Developments and Future \nDirections (Sabin Center for Climate Change Law, 2024); \nhttps://scholarship.law.columbia.edu/cgi/viewcontent.\ncgi?article=1217&context=sabin_climate_change\n28.\t Colebatch, H. K. Making sense of governance. Policy Soc. 33, \n307\u2013316 (2014).\n29.\t Vozzo, M. L. et al. From experiment to intervention: a case study of \nscaling up marine eco-engineering from research to application. \nEnviron. Sci. Policy 158, 103800 (2024).\n30.\t Boyd, P. et al. Operational monitoring of open-ocean carbon \ndioxide removal deployments: detection, attribution, and \ndetermination of side effects. Oceanography https://doi.org/ \n10.5670/oceanog.2023.s1.2 (2023).\n31.\t Stilgoe, J. et al. Developing a framework for responsible \ninnovation. Res. Policy 42, 1568\u20131580 (2013).\n32.\t Nawaz, S. et al. Public evaluations of four approaches to \nocean-based carbon dioxide removal. Clim. Policy 23, 379\u2013394 \n(2023).\n33.\t Cox, E. et al. Public attitudes and emotions toward novel carbon \nremoval methods in alternative sociotechnical scenarios. Environ. \nRes. Lett. 19, 084026 (2024).\n34.\t A Research Review of Interventions to Increase the Persistence and \nResilience of Coral Reefs (NASEM, 2019); https://doi.org/ \n10.17226/25279\n35.\t Macnaghten, P. The Making of Responsible Innovation (Cambridge \nUniv. Press, 2020); https://doi.org/10.1017/9781108871044\n36.\t Boettcher, M. et al. A Code of Conduct for Marine Carbon \nDioxide Removal Research (Aspen Institute, 2023); https://www.\naspeninstitute.org/publications/a-code-of-conduct-for-marine-ca\nrbon-dioxide-removal-research/\n37.\t Hurd, C. L. et al. Forensic carbon accounting: assessing the role of \nseaweeds for carbon sequestration. J. Phycol. 58, 347\u2013363 (2022).\n38.\t Rosen, P. et al. Guidance for the Potential Application of \nMarine Carbon Dioxide Removal (mCDR) in US National Marine \nSanctuaries (NOAA, 2023); https://doi.org/10.25923/chkd-fd21\n39.\t Lenk, H. Ethics of responsibilities distributions in a technological \nculture. AI Soc. 32, 219\u2013231 (2017).\n40.\t Carrier, M. How to conceive of science for the benefit of society: \nprospects of responsible research and innovation. Synthese 198, \n4749\u20134768 (2019).\n41.\t Carrier, M. & Irzik, G. Responsible research and innovation: \ncoming to grips with an ambitious concept. Synthese 198, \n4627\u20134633 (2019).\n42.\t Low, S. & Buck, H. J. The practice of responsible research and \ninnovation in \u2018climate engineering\u2019. Wiley Interdiscip. Rev. Clim. \nChange 11, e644 (2020).\n43.\t Voegtlin, C. & Scherer, A. G. Responsible innovation and the \ninnovation of responsibility: governing sustainable development \nin a globalized world. J. Bus. Ethics 143, 227\u2013243 (2017).\n44.\t Smolka, M. & B\u00f6schen, S. Responsible innovation ecosystem \ngovernance: socio-technical integration research for \nsystems-level capacity building. J. Responsible Innov. 10, 2207937 \n(2023).\n45.\t Brent, K. et al. From informal to formal governance of solar \nradiation management. Clim. Policy https://doi.org/10.1080/14693\n062.2024.2430688 (2024).\n46.\t Boyd, P. W. et al. Potential negative effects of ocean afforestation \non offshore ecosystems. Nat. Ecol. Evol. 6, 675\u2013683 (2022).\n47.\t Kenny, I. et al. Aligning social and ecological goals for successful \nmarine restoration. Biol. Conserv. 288, 110357 (2023).\n48.\t Fritz, L. et al. Public engagement for inclusive and sustainable \ngovernance of climate interventions. Nat. Commun. 15, 4168 \n(2024).\n49.\t Ram\u00edrez\u2010Monsalve, P. et al. Marine restoration governance \narrangements: issues of legitimacy. Environ. Policy Gov. 32, \n122\u2013134 (2022).\n50.\t Fidelman, P. et al. Regulatory implications of coral reef restoration \nand adaptation under a changing climate. Environ. Sci. Policy 100, \n221\u2013229 (2019).\n51.\t Hansen-Magnusson, H. Making polar and ocean governance \nfuture-proof. Polit. Gov. 10, 60\u201369 (2022).\n\nNature Climate Change | Volume 15 | April 2025 | 375\u2013384\n383\nArticle\nhttps://doi.org/10.1038/s41558-025-02291-4\n52.\t Sovacool, B. K. et al. Risk\u2013risk governance in a low\u2010carbon future: \nexploring institutional, technological, and behavioral tradeoffs in \nclimate geoengineering pathways. Risk Anal. https://doi.org/ \n10.1111/risa.13932 (2022).\n53.\t Hobday, A. J. & Pecl, G. T. Identification of global marine hotspots: \nsentinels for change and vanguards for adaptation action. Rev. \nFish Biol. Fish. 24, 415\u2013425 (2014).\n54.\t Bell\u2010James, J., Foster, R. & Shumway, N. The permitting process \nfor marine and coastal restoration: a barrier to achieving global \nrestoration targets? Conserv. Sci. Pract. 5, e13050 (2023).\n55.\t Shumway, N. et al. Policy solutions to facilitate restoration in \ncoastal marine environments. Mar. Policy 134, 104789 (2021).\n56.\t Morrison, T. H. et al. Overcoming lock-in of science-policy \nresponses to reef heating. Mar. Policy 170, 106380 (2024).\n57.\t Bellamy, R. et al. \u2018Opening up\u2019 geoengineering appraisal: \nmulti-criteria mapping of options for tackling climate change. \nGlob. Environ. Change 23, 926\u2013937 (2013).\n58.\t Hockerts, K. et al. Defining and conceptualizing impact investing: \nattractive nuisance or catalyst? J. Bus. Ethics 179, 937\u2013950 (2022).\n59.\t Sumaila, U. R. et al. Financing a sustainable ocean economy. Nat. \nCommun. 12, 3259 (2021).\n60.\t Blackwatters, J. E. et al. More than funders: the roles of \nphilanthropic foundations in marine conservation governance. \nConserv. Sci Pract. https://doi.org/10.1111/csp2.12829 (2022).\n61.\t Holmes, G. Biodiversity for billionaires: capitalism, conservation \nand the role of philanthropy in saving/selling nature. Dev. Change \n43, 185\u2013203 (2012).\n62.\t Lewis, F. et al. Funding Trends 2023: Tracking the State of Global \nOcean Funding (Our Shared Seas, 2023); https://oursharedseas.\ncom/funding/\n63.\t Boettcher, M. et al. Navigating potential hype and opportunity in \ngoverning marine carbon removal. Front. Clim. 3, 664456 (2021).\n64.\t Schadeberg, A. et al. Science governs the future of the \nmesopelagic zone. NPJ Ocean Sustain. 2, 2 (2023).\n65.\t Bellamy, R. et al. Deliberative mapping of options for tackling \nclimate change: Citizens and specialists \u2018open up\u2019 appraisal of \ngeoengineering. Public Underst. Sci. 25, 269\u2013286 (2016).\n66.\t Ainscough, J. & Willis, R. Embedding deliberation: guiding the use \nof deliberative mini-publics in climate policy-making. Clim. Policy \nhttps://doi.org/10.1080/14693062.2024.2303337 (2024).\n67.\t Macpherson, C. C. Why bioethics should address climate change \nand how it might do so in Bioethical Insights into Values and \nPolicies (ed. Macpherson, C.) 199\u2013216 (Springer, 2016);  \nhttps://doi.org/10.1007/978-3-319-26167-6_15\n68.\t Robinson, C. J. et al. Indigenous benefits and carbon offset \nschemes: an Australian case study. Environ. Sci. Policy 56, 129\u2013134 \n(2016).\n69.\t Reyes-Garc\u00eda, V. et al. Recognizing Indigenous peoples\u2019 and local \ncommunities\u2019 rights and agency in the post-2020 Biodiversity \nAgenda. Ambio 51, 84\u201392 (2022).\n70.\t Costanza, R. et al. Common asset trusts to effectively steward \nnatural capital and ecosystem services at multiple scales.  \nJ. Environ. Manage. 280, 111801 (2021).\n71.\t Fischer, M. et al. Empowering her guardians to nurture our ocean\u2019s \nfuture. Rev. Fish Biol. Fish. 32, 271\u2013296 (2022).\n72.\t Cinner, J. E. et al. Building adaptive capacity to climate change \nin tropical coastal communities. Nat. Clim. Change 12, 117\u2013123 \n(2018).\n73.\t Spalding, A. K. et al. Engaging the tropical majority to make ocean \ngovernance and science more equitable and effective. NPJ Ocean \nSustain. 2, 8 (2023).\n74.\t Muiderman, K. et al. Four approaches to anticipatory climate \ngovernance: different conceptions of the future and implications \nfor the present. Wiley Interdiscip. Rev. Clim. Change 11, e673 \n(2020).\n75.\t Dooley, K. et al. Carbon-dioxide removal and biodiversity: a threat \nidentification framework. Glob. Policy 12, 34\u201344 (2021).\n76.\t Riahi, K. et al. The Shared Socioeconomic Pathways and their \nenergy, land use, and greenhouse gas emissions implications: an \noverview. Glob. Environ. Change 42, 153\u2013168 (2017).\n77.\t Brown, K. Global environmental change I: a social turn for \nresilience? Prog. Hum. Geogr. 38, 107\u2013117 (2014).\n78.\t Olsson, L. et al. Why resilience is unappealing to social science: \ntheoretical and empirical investigations of the scientific use of \nresilience. Sci. Adv. 1, e1400217 (2015).\n79.\t Mikulewicz, M. Thwarting adaptation\u2019s potential? A critique of \nresilience and climate-resilient development. Geoforum 104, \n267\u2013282 (2019).\n80.\t High-Quality Blue Carbon Principles and Guidance: A Triple-Benefit \nInvestment for People, Nature and Climate (Meridian Institute, \n2022); https://merid.org/wp-content/uploads/2022/11/HQBC-PG_\nFINAL_11.8.2022.pdf\n81.\t Bayraktarov, E. et al. The cost and feasibility of marine coastal \nrestoration. Ecol. Appl. 26, 1055\u20131074 (2016).\n82.\t Bayraktarov, E. et al. Motivations, success, and cost of coral reef \nrestoration. Restor. Ecol. 27, 981\u2013991 (2019).\n83.\t Sovacool, B. K. Reckless or righteous? Reviewing the \nsociotechnical benefits and risks of climate change \ngeoengineering. Energy Strategy Rev. 35, 100656 (2021).\n84.\t Morrison, T. H. et al. Radical interventions for climate-impacted \nsystems. Nat. Clim. Change 12, 1100\u20131106 (2022).\n85.\t Low, S. et al. Undone science in climate interventions: contrasting \nand contesting anticipatory assessments by expert networks. \nEnviron. Sci. Policy 137, 249\u2013270 (2022).\n86.\t Jakovac, C. C. et al. Costs and carbon benefits of mangrove \nconservation and restoration: a global analysis. Ecol. Econ. 176, \n106758 (2020).\n87.\t Hajer, M. Policy without polity? Policy analysis and the \ninstitutional void. Policy Sci. 36, 175\u2013195 (2003).\n88.\t Leong, C. Hajer\u2019s institutional void and legitimacy without polity. \nPolicy Sci. 50, 573\u2013583 (2017).\n89.\t Morrison, T. H. Evolving polycentric governance of the Great \nBarrier Reef. Proc. Natl Acad. Sci. USA 114, E3013\u2013E3021  \n(2017).\n90.\t Kooiman, J. & Jentoft, S. Meta-governance: values, norms and \nprinciples, and the making of hard choices. Public Adm. 87, \n818\u2013836 (2009).\n91.\t Bach, L. T. et al. Testing the climate intervention potential of \nocean afforestation using the Great Atlantic Sargassum Belt. Nat. \nCommun. 12, 2556 (2021).\n92.\t Morganti, T. et al. Monitoring marine carbon dioxide removal: \nquantitative analysis of indicators for carbon removed and \nenvironmental side-effects. ESS Open Archive (2025).\n93.\t Eriksen, S. et al. Adaptation interventions and their effect \non vulnerability in developing countries: Help, hindrance or \nirrelevance? World Dev. 141, 105383 (2021).\n94.\t Sultana, F. The unbearable heaviness of climate coloniality. Polit. \nGeogr. 99, 102638 (2022).\n95.\t Bower, E. R. et al. Enabling pathways for sustainable livelihoods \nin planned relocation. Nat. Clim. Change https://doi.org/10.1038/\ns41558-023-01753-x (2023).\n96.\t Levin, L. A. et al. Deep-sea impacts of climate interventions. \nScience 379, 978\u2013981 (2023).\n97.\t Satterfield, T. et al. Exploring public acceptability of direct air \ncarbon capture with storage: climate urgency, moral hazards and \nperceptions of the \u2018whole versus the parts\u2019. Climatic Change 176, \n14 (2023).\n98.\t Hughes, T. P., Baird, A. H., Morrison, T. H. & Torda, G. Principles for \ncoral reef restoration in the Anthropocene. One Earth 6, 656\u2013665 \n(2023).\n\nNature Climate Change | Volume 15 | April 2025 | 375\u2013384\n384\nArticle\nhttps://doi.org/10.1038/s41558-025-02291-4\n99.\t Hart, P. S. et al. Moral hazard or not? The effects of learning about \ncarbon dioxide removal on perceptions of climate mitigation in \nthe United States. Energy Res. Soc. Sci. 89, 102656 (2022).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this licence, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2025\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-025-02291-4\nMethods\nSurvey instrument design\nObservations and attitudes towards novel marine-climate interven-\ntions were surveyed using an online survey\u2013questionnaire targeting \nactive practitioners. Questions were designed to capture observational \ndata describing current arrangements for governing interventions and \npositional attitudinal data concerning perceived benefits and costs, \ngaps in governance, risks, and missed and emerging opportunities \n(see survey\u2013questionnaire in Supplementary Table 8). Types of inter-\nventions used to design response options were identified from recent \nauthoritative reviews15,34,84. The questionnaire used a mix of selected \nchoice questions, ratings and open-ended text response questions. The \nquestionnaire was delivered using Qualtricsxm online survey software.\nThe survey instrument and specific questions were pre-tested \nby members of the research group, revised and then formally piloted \nthrough a soft launch of the survey in October 2022 with members of \nthe study\u2019s technical advisory committee.\nEthics statement\nThe study was approved by the James Cook University Human Research \nEthics Committee (approval number H88845) in accordance with the \nAustralian National Statement on Ethical Conduct in Human Research, \n2007 (Updated 2018)100, which addresses matters of harm and benefit \nto study participants. The recruitment method and survey instrument \nwere designed to provide participants with information about the \nstudy and obtain informed consent before participation through an \ninitial screening question.\nParticipant selection and recruitment\nActors engaged in novel marine-climate interventions constitute an \nemerging group. The survey was designed to target participants in this \nbroad group via their participation in professional networks associated \nwith interventions. Participant selection was therefore opportunistic, \nand stratified sampling of specific sub-groups, such as Traditional \nOwners and First Nations, was not pursued101. Strategies to limit sample \nbias included translating the survey\u2013questionnaire into six languages: \nEnglish, French, Portuguese, Spanish, Japanese and Chinese (Simpli-\nfied). The survey was distributed to professional networks engaged in \na wide range of marine-climate interventions, including marine and \ncoastal restoration, marine conservation, community and small-scale \nfisher development, marine carbon dioxide removal and other forms of \ngeoengineering, seafood afforestation, solar radiation management, \nand coastal and marine community adaptation planning. Professional \npractitioner networks assisting with survey distribution in the later \nstages of recruitment also included practitioners from the NGO sec-\ntors, local community representatives and Traditional Owners and \nFirst Nations.\nProfessional communities approached were invited to respond on \nthe basis of their engagement with \u2018new and emerging\u2019 marine-climate \ninterventions. The survey instrument was designed to collect both \ndescriptive and attitudinal data about the broad array of observed \ninterventions as well as those specific interventions for which respond-\nents had subject-matter expertise and direct professional experience. \nExamples of interventions were provided in recruitment materials to \nassist in clarifying the term, and these included assisted evolution, \ncloud brightening, seaweed farming, coral propagation and trans-\nlocation. These examples were not exhaustive of a broader range \nof interventions and may have had limited recruitment of potential \nrespondents engaged in other types of interventions.\nPractitioners were recruited for survey through professional \nopen online networks and using published affiliations information. \nRecruitment methods included general broadcast emails using profes-\nsional international email lists, published professional email addresses \nand social media posts in six languages via Twitter using professional \naccounts. Multiple phases of recruitment occurred in early October \n2022, mid November 2022, late January 2023 and early March 2023. \nThe same distribution methods were used with revised recruitment \nand advertising messages reflecting the stage of the survey and the \ntime left to participate.\nData collection\nThe online survey\u2013questionnaire was launched on 31 October 2022 \nand remained open until 15 March 2023. Three hundred and thirty-two \nresponses met the criteria for level of question completion and were \nretained for analysis. These responses included those undertaken in \nfive non-English languages\u2014Chinese, Japanese, French, Portuguese and \nSpanish\u2014which accounted for 18% of the final sample. These responses \nwere translated into English by native speakers with marine expertise \nbefore analysis.\nSurvey data were not treated to any weighting to adjust for the \nexpected population because the survey population was an emerg-\ning specialist group, and population characteristics were not estab-\nlished. Response rates to the survey by sub-group are therefore not \nreported. Representativeness of the survey data was therefore sub-\nject to sample bias although recruitment methods were adjusted to \ntarget non-English speakers in five other languages and practitioners \nin non-scientific networks. A degree of sample bias was accepted as \nan expected limitation of the study due to the nature of the emerging \ngroup being surveyed and the online survey\u2013questionnaire instru-\nment used101. No identifying data were collected from respondents \nalthough in some cases participants provided personal identifying \ndata in response to open-text questions.\nData analysis\nTwo units of analysis were used to examine the data. Data on respond-\nents\u2019 role, interaction with interventions and general awareness of \ninterventions were treated as data about the respondents, while data \nin response to survey question six onwards were treated as data about \nthe intervention the respondent was asked to identify as the one with \nwhich they were most familiar.\nSelected choice data where respondents answered by selecting \nfrom a pre-determined set of options were analysed using basic descrip-\ntive statistics (frequency counts). These included questions to identify \nawareness of interventions, involvement in active intervention plan-\nning and deployment, the stage of development of the intervention \nrespondents were most familiar with, the types of actors and organi-\nzations engaged in their development and the presence or absence of \nspecific governance arrangements. Open-ended text responses were \nanalysed using thematic content analysis102\u2013104 to code data and thereby \nconvert the qualitative data into quantitative data. To increase coding \nreliability105, first-pass coding frameworks were reviewed and tested by \nother members of the project team before finalizing and then under-\ntaking the thematic analysis. Basic descriptive statistics (frequency \ncounts) were then used to analyse the coded survey data by theme.\nBoth selected choice data and open-ended text data describ-\ning the types and sub-types of interventions respondents (n\u2009=\u2009332 \nrespondents) were aware of and were most familiar with (n\u2009=\u2009240 inter-\nventions) were initially analysed using a coding framework based on \nthe initial typology we developed from a review of published studies \n(Supplementary Table 2). Inductive thematic coding of open-text \ndata describing interventions was undertaken using NVivo 20 quali-\ntative research software, and initial type and sub-type codes were \nsubsequently adjusted in response to survey data thematic codes \n(Supplementary Table 9). These final codes were checked against the \nchoice response the respondents selected for verification and against \nmore recent authoritative reviews (for example, review by the National \nAcademies of Sciences, Engineering, and Medicine10) for salience.\nOpen-ended text data about the location of the identified inter-\nventions were analysed using a coding framework developed in Nvivo \nand organized by oceans and seas described in response data text. \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-025-02291-4\nFrequency analysis was then undertaken by intervention type and \nsub-type using the following identified oceans and seas: North Atlantic, \nSouth Atlantic, North Pacific, South Pacific, Red Sea, Mediterranean \nSea, Indian Ocean, Wider Caribbean, Australia (tropical), and Australia \nand New Zealand (temperate).\nOpen-ended text data in which respondents identified the climate \ngoals of the intervention they were most familiar with were thematically \ncoded using a coding framework developed from the ten intervention \nclimate benefits identified by review of the literature (Supplementary \nTables 3 and 10). Frequency analysis was then undertaken by climate \nbenefit code, and data were analysed for distributions of climate goals \nby intervention type. Open-ended text data in which respondents iden-\ntified the major risks and opportunities of the identified intervention \nwere thematically coded using a coding framework developed induc-\ntively from the data (Supplementary Table 7). Frequency analysis was \nthen undertaken by type of risk or missed opportunity.\nTo determine the extent of responsible governance of new \nmarine-climate interventions, we developed a methodological heu-\nristic by extending Stilgoe31 and Macnaghten\u2019s35 framework for respon-\nsible innovation (based on anticipation, inclusion, responsiveness, \nreflexivity), which they apply to emerging scientific innovations and in \nparticular to climate geoengineering innovations. We matched these \nfour dimensions to the new categories of marine-climate intervention \nrisk (Table 1), which we developed from the literature. In extending the \nresponsible innovation framework, we defined responsible govern-\nance as rules, guidance and steerage overseen by governing actors \nto prevent intervention risk. We identified specific examples of such \ngovernance arrangements through a review of the literature on instru-\nments and processes for governing technical feasibility, cumulative risk \nand impact assessment, including public deliberation in intervention \ndevelopment and approval and anticipatory climate governance. These \nspecific governance arrangements were then matched to relevant \nsurvey response data fields (Supplementary Tables 1 and 5). Indication \nof the specific governance arrangements described in survey ques-\ntions was analysed for each intervention type using basic descriptive \nstatistics (Supplementary Tables 4 and 6).\nReporting summary\nFurther information on research design is available in the Nature \nPortfolio Reporting Summary linked to this article.\nData availability\nThe raw dataset generated by the survey research is not available due \nto restrictions to protect study participant privacy and to limit reuse to \nstudies of a similar nature, in accordance with the National Statement \non Ethical Conduct in Human Research, 2007 (Updated 2018)100 (James \nCook University Human Research Ethics Committee approval number \nH88845). The processed and de-identified dataset is, however, available \nfrom the authors on reasonable request for studies of a similar nature.\nReferences\n100.\tNational Statement on Ethical Conduct in Human Research 2007 \n(Updated 2018) (The National Health and Medical Research \nCouncil, the Australian Research Council and Universities \nAustralia, 2018).\n101.\t McRobert, C. J. et al. A multi-modal recruitment strategy using \nsocial media and internet-mediated methods to recruit a \nmultidisciplinary, international sample of clinicians to an online \nresearch study. PLoS ONE 13, e0200184 (2018).\n102.\tNaeem, M. et al. A step-by-step process of thematic analysis to \ndevelop a conceptual model in qualitative research. Int. J. Qual. \nMethods 22, 16094069231205788 (2023).\n103.\tLin, A. C. Bridging positivist and interpretivist approaches to \nqualitative methods. Policy Stud. J. 26, 162\u2013180 (1998).\n104.\tMiles, M. B. & Huberman, M. Qualitative Data Analysis: A Methods \nSourcebook (Sage, 1994).\n105.\tBraun, V. & Clarke, V. Toward good practice in thematic analysis: \navoiding common problems and be(com)ing a knowing \nresearcher. Int. J. Transgend. Health 24, 1\u20136 (2023).\nAcknowledgements\nThe authors acknowledge the translators of the survey\u2013questionnaire \nand responses: S. Kwong, M. Angelini, J. Morias, V. Huertas, M. Tatsumi, \nH. Mera and C. Appert. C. Cullen-Knox assisted in data analyses. \nIllustration and figure design work was undertaken by V. H. Martin. \nFunding for this study was received from the Australian Government \nvia the Australia Research Council Discovery Projects grant scheme \nfor the project, Novel governance for marine ecosystems in rapid \ntransition (DP220103921; T.H.M., T.H., G.T.P. and E.M.O.). The authors \nthank attendees of the international Governing Changing Oceans \nworkshops, where early versions of this manuscript were presented \nand discussed, in particular P. Cohen. The workshops were funded \nthrough The Nature Conservancy\u2019s Science for Nature and People \nPartnership (SNAPP 054: Governing Changing Oceans; T.H.M., G.T.P \nand P. Cohen). They were supported in kind by UTAS, IMAS, Centre \nfor Marine Socioecology, University of Melbourne, Wageningen \nUniversity, Packard Foundation, Conservation International, Australian \nInstitute of Marine Science, the ARC Centre of Excellence for Coral \nReef Studies and WorldFish-CGIAR.\nAuthor contributions\nE.M.O. led the literature review, survey design, data collection, survey \ndata analysis and manuscript preparation. G.T.P. contributed to \nsurvey design, survey data analysis and manuscript preparation. T.H. \ncontributed to survey design and manuscript preparation. S.L. and \nC.L. contributed to survey data analysis and manuscript preparation. \nK.L.N. contributed to the literature review, survey data analysis and \nvisualization. T.H.M. co-led the literature review, survey design and \ndata collection and contributed to the survey data analysis and \nmanuscript preparation.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41558-025-02291-4.\nCorrespondence and requests for materials should be addressed to \nEmily M. Ogier.\nPeer review information Nature Climate Change thanks Christopher L. \nCummings, Yuwan Malakar and Yoshitaka Ota for their contribution to \nthe peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\n1\nnature portfolio  |  reporting summary\nApril 2023\nCorresponding author(s):\nEmily M. Ogier\nLast updated by author(s): 06/02/2025\nReporting Summary\nNature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nQualtricsxm online survey program (v. 09.2022)\nData analysis\nQualtricsxm online survey program (v. 07.2023)\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and \nreviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A description of any restrictions on data availability \n- For clinical datasets or third party data, please ensure that the statement adheres to our policy \n \nThe raw dataset generated by the survey research is not available due to restrictions to protect study participant privacy, in accordance with the Australian National \nStatement on Ethical Conduct in Human Research, 2007 (Update 2018) (approval number H88845). The processed and de-identified dataset is subject to controlled \naccess, whereby access can be permitted for \u2018extended\u2019 use of data in future research projects that are either (i) an extension of, or closely related to, the original \n\n2\nnature portfolio  |  reporting summary\nApril 2023\nproject; or (ii) in the same general area of research (for example, social research), in accordance with the Australian National Statement on Ethical Conduct in \nHuman Research, 2007 (Update 2018). The authors will respond to requests made via email to make data available within 6 weeks.\nResearch involving human participants, their data, or biological material\nPolicy information about studies with human participants or human data. See also policy information about sex, gender (identity/presentation), \nand sexual orientation and race, ethnicity and racism.\nReporting on sex and gender\nNot applicable\nReporting on race, ethnicity, or \nother socially relevant \ngroupings\nGroupings of survey participants was reported on the basis of professional engagement with novel marine-climate \ninterventions. Groupings used in reporting results were: scientists; policymakers; Non-Government Organisation \npractitioners; representatives of Traditional Owners and First Nations; representatives of local community or industry \nsectors; representatives of industry organisations.  \nPopulation characteristics\nPractitioners engaged in novel marine-climate interventions constitute a specialist group. The survey was designed to target \nparticipants in this group via their participation in professional networks associated with interventions. Participation selection \nwas therefore opportunistic and stratified sampling was not pursued. However, the survey-questionnaire was made available \nin six languages: English, French, Portuguese, Spanish, Japanese, Chinese (Simplified). \nRecruitment\nThis specialist group was recruited for survey through professional open online networks and using published affiliations \ninformation. The survey instrument was designed to collect both descriptive and attitudinal data about the broad array of \nobserved interventions as well as those specific interventions for which respondents had subject matter expertise and direct \nprofessional experience. \n \nRecruitment methods included general broadcast emails using professional international email lists and social media posts in \nsix languages via Twitter using professional accounts. Professional communities approached were invited to respond based \non their engagement with \u201cnew and emerging\u201d marine-climate interventions. Examples of interventions were provided in \nrecruitment materials to assist in clarifying the term, and these included assisted evolution, cloud brightening, seaweed \nfarming, coral propagation and translocation. These examples were not exhaustive of a broader range of interventions and \nmay have had limited recruitment of potential respondents engaged in other types of interventions. Representativeness of \nthe survey data was therefore subject to sample bias but this was accepted as an expected limitation due to the nature of the \nspecialist group being targeted and the online survey-questionnaire instrument used.\nEthics oversight\nJame Cook University Human Research Ethics Committee\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nObservations and attitudes to novel marine-climate interventions were surveyed using an online survey-questionnaire targeting \nactive practitioners. Questions were designed using mixed methods to capture quantitative and qualitative observational data \ndescribing the state-of-play of governance of interventions and positional attitudinal data concerning perceived benefits and costs, \ngaps in governance, risks, and missed and emerging opportunities. The questionnaire used a mix of selected choice questions, ratings \nand open-ended text response.\nResearch sample\nThe research sample were practitioners professionally engaged in novel marine-climate interventions, inclusive of funders, research \nscientists, representatives of First Nations, industry or community groups partnering in interventions, and government agency staff \ninvolved in policy development or regulation. The research sample was not representative. Rationale for the chosen research sample \nwas based on the emergence of this professional community and its unknown population size. \nSampling strategy\nOpportunistic sampling was used. No statistical methods were used to pre-determine the sample size because the total population of \npractitioners being targeted for the research was not known (i.e., as an emergent practitioner community). As no sample sizes were \nchosen, the sample size obtained (n=332) was deemed to be sufficient as thematic saturation was observed in the qualitative \nresponses to open-ended questions.\nData collection\nData was collected via the questionnaire, which was delivered using Qualtricsxm online survey software (v. 09.2022). The \nquestionnaire was completed at the participants' convenience and no researcher was present. Researchers involved in collecting the \nstudy data were aware of the study hypothesis. No experimental conditions were used in the research.\n\n3\nnature portfolio  |  reporting summary\nApril 2023\nTiming\nThe online survey was conducted continuously between 31/10/2022 - 15/03/2023.\nData exclusions\nTwo data exclusions were applied, as follows: \n- Data which was personally identifying (three instances). One instance of this occurred where a participant entered a written text \nresponse to an open-ended question which named another researcher. Two additional instances included the respondents providing \nspecific identifying details of their professional roles. \n- Data entered in response to open-text questions which did not meet data inclusion criteria (79 instances). Specifically, data were \nexcluded where the entries were not recognisable as text and thematic coding was not possible.\nNon-participation\nThe number of participants who did not proceed once informed consent had been indicated via the online survey screening question \nwas 46. This group was not included in the sample for analysis.\nRandomization\nParticipants were not allocated into experimental groups. Reported groups were based on self-selected choice options provided in \nthe questionnaire for type of practitioner. \nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology and archaeology\nAnimals and other organisms\nClinical data\nDual use research of concern\nPlants\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nNovel plant genotypes\nN/A\nSeed stocks\nN/A\nAuthentication\nN/A\nPlants\n\n\n    Scientific Research Findings:", "answer": "Our study identifies a wide diversity of marine\u2013climate interventions proposed or already deployed in 37 marine systems. Multiple types of intervention co-occur in all major ocean basins. Most practitioners (71%) report interventions aimed at supporting marine species and ecosystem adaptation, while 29% report interventions aimed at climate mitigation and societal adaptation. Perceptions of climate outcomes vary widely, with low consensus on intended and realized climate benefits of interventions. The practitioner community is science-dominated with limited involvement of public institutions and communities. Arrangements for responsibly governing intervention risks are seldom observed, indicating the pacing problem is indeed present. Intervention assessment and approval are narrowly focused on technical feasibility to meet minimum permitting requirements, with limited assessment of cumulative impacts, public deliberation and management of broader ecological, cultural and social risks and benefits. Identified policy gaps provide clear directions to improve governance readiness for marine\u2013climate interventions.", "id": 37}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 15 | February 2025 | 147\u2013152\n147\nnature climate change\nhttps://doi.org/10.1038/s41558-024-02237-2\nArticle\nMitigation policies interactions delay the \nachievement of carbon neutrality in China\n \nYu Liu\u2009\n\u200a\u20091,2,9\u2009\n, Mingxi Du\u2009\n\u200a\u20093,9\u2009\n, Lingyu Yang\u2009\n\u200a\u20094,5,9, Qi Cui\u2009\n\u200a\u20096\u2009\n, Yawen Liu7, \nXinbei Li\u2009\n\u200a\u20094,5, Nenggao Zhu4,5, Ying Li8, Chen Jiang4,5, Peng Zhou\u2009\n\u200a\u20096,  \nQiuyu Liu3 & Canfei He\u2009\n\u200a\u20091\nThe achievement of China\u2019s carbon neutrality is crucial for the 1.5\u2009\u00b0C target \nof the Paris Agreement and must involve the implementation of various \nmitigation policies. However, these efforts are hindered by poor knowledge \nof the interactions between policies. Here we use a dynamic computable \ngeneral equilibrium model of China (CEEGE model) and create a policy \nportfolio area of 1,295 scenarios covering four major mitigation strategies \n(carbon pricing, energy efficiency, renewable energy and electrification \nof end uses). When the interactions between mitigation policies are \nconsidered, the percentage of scenarios in which the carbon neutrality \ntarget is reached by 2060 decreases by 84%, with the years in which these \nscenarios are achieved being delayed by 5\u20136\u2009years. Only the combinations \nwith renewable energy and electrification of end uses generate synergetic \neffects on both economic and mitigation impacts. Our work can inform the \nformulation of more efficient mitigation policy portfolios by emphasizing \npolicy interactions.\nTo ensure that the 1.5\u2009\u00b0C target of the Paris Agreement is met, effec-\ntive climate policies must be implemented at the national level1\u20134. As \nthe largest CO2 emitter and the second-largest economy in the world, \nChina has achieved its goal of a 45% reduction in carbon intensity rela-\ntive to 2005 levels in 20204. Demonstrating increased ambition, China \nannounced the following dual carbon targets in December 2020: peak-\ning CO2 emissions before 2030 and reaching carbon neutrality by 2060. \nVarious studies have proposed a range of mitigation policies to address \nthis challenge; these include support for non-fossil energy deployment, \nthe electrification of end-use sectors, energy efficiency improvement, \nnegative emission technologies and market-based measures (such as \ncarbon pricing), as well as other industry-specific policies3,5\u20137.\nTo meet its dual carbon targets, China needs to implement these \npolicies in tandem8. On the basis of the portfolios of mitigation poli-\ncies, China is projected to successfully achieve its dual carbon targets, \nwith cumulative gross domestic product (GDP) losses of ~1.7\u20135.7%  \n(refs. 9\u201311). However, simply aggregating the (seemingly ideal) effects \nof multiple policies does not accurately reflect the actual policy out-\ncome, as the trade-offs and synergies between policies can weaken or \nstrengthen their effects12,13. Combining and sequencing policy imple-\nmentation can largely impact the efficiency, stringency and externality \nof mitigation policies14\u201316. Nevertheless, most studies on achieving \nChina\u2019s dual carbon targets have focused on the effectiveness and eco-\nnomic costs of mitigation policies17\u201319, while the interactions between \nmitigation policies and their impacts on attaining China\u2019s carbon \nneutrality have been overlooked.\nRecent studies have measured the trade-off and synergetic effects \nthrough comparisons of the results of the simultaneous implementa-\ntion of multiple policies with the aggregation of single-implemented \npolicies12,20. Nevertheless, these studies focused mostly on the \nReceived: 19 February 2024\nAccepted: 11 December 2024\nPublished online: 31 January 2025\n Check for updates\n1College of Urban and Environmental Sciences, Peking University, Beijing, China. 2Institute of Carbon Neutrality, Peking University, Beijing, China.  \n3School of Public Policy and Administration, Xi\u2019an Jiaotong University, Xi\u2019an, China. 4School of Public Policy and Management, University of Chinese \nAcademy of Sciences, Beijing, China. 5Institutes of Science and Development, Chinese Academy of Sciences, Beijing, China. 6School of Economics  \nand Management, China University of Petroleum, Qingdao, China. 7Digital Economy Laboratory, University of International Business and Economics, \nBeijing, China. 8Chinese Academy of Personnel Science, Beijing, China. 9These authors contributed equally: Yu Liu, Mingxi Du, Lingyu Yang.  \n\u2009e-mail: yu.liu@pku.edu.cn; dumingxi28@xjtu.edu.cn; cuiqi@upc.edu.cn\n\nNature Climate Change | Volume 15 | February 2025 | 147\u2013152\n148\nArticle\nhttps://doi.org/10.1038/s41558-024-02237-2\nSuccessful cases with and without policy \ninteractions\nTo examine the effects of policy interactions, we compare the simula-\ntion results of mitigation scenarios under two assumptions: an actual \nsimultaneous-implementation assumption (ASIA), in which the \nsimulation results are obtained by implementing mitigation policies \nsimultaneously, and an idealized simple-superposition assumption \n(ISSA), in which the simulation results are obtained by aggregating the \nresults of separately implemented mitigation policies (Supplemen-\ntary Section 4.4). The baseline scenario consists of China\u2019s nationally \ndetermined contribution target submitted in 2015, with the goals of \npeaking carbon dioxide emissions before 2030 and reducing carbon \ndioxide emissions per unit of GDP by 60% from 2005 levels. The energy \nand emission development assumptions are designed with reference \nto the National Bureau of Statistics of China (NBSC; 2022)28 and the \nInternational Energy Agency (IEA; 2022)29 STEP (stated policies sce-\nnario) scenarios. The economic and social development assumptions \nmainly refer to the NBSC (2022)30 and the United Nations (2022)31.  \nA detailed description of the scenario design (including macroeconom-\nics, society, energy, emissions, carbon sequestration and mitigation \nstrategies) for China from 2020 to 2060 can be found in Methods and \nSupplementary Section 3.\nSpecifically, the results from our simulation indicate that achiev-\ning carbon neutrality in China would be an even greater challenge under \nASIA than what is perceived under ISSA (Fig. 1). While carbon neutral-\nity could be achieved in 23.5% of the scenarios under ISSA, it could be \nachieved in only 3.7% of the scenarios under ASIA. This suggests that \nbecause of policy interactions, the percentage of scenarios in which \nthe carbon neutrality target could be achieved by 2060 decreases \nsubstantially by 84%. Furthermore, the timeline of achieving carbon \nneutrality under ASIA is expected to be delayed by ~5\u20136\u2009years compared \nwith that under ISSA. For scenarios in which the carbon neutrality target \nis successfully achieved, the average year achieving carbon neutrality \nis 2054 under ISSA but 2059 under ASIA. These findings highlight the \nneed for China to carefully consider and optimize its mitigation strate-\ngies, as previous research has suggested3,32, particularly in terms of the \npotential mitigation effects when multiple policies are implemented \nin combination.\nAmong all the scenarios, the percentages of policy combinations \nthat lead to the achievement of the carbon neutrality target differ \ninteractions between mitigation policies in certain industries, includ-\ning transport, construction and agriculture21\u201324. Since sectors are inter-\nlinked through the supply chain, the interactions between mitigation \npolicies in one sector may influence the effects of mitigation policies in \nother sectors, which has been overlooked in the literature. Moreover, \nthe impact of policy interactions on the timeframe of achieving the \ncarbon neutrality target needs to be further explored25.\nThus, this study aims to address the current insufficient under-\nstanding of the interactions between mitigation policies in achieving \nChina\u2019s carbon neutrality target by developing a dynamic computable \ngeneral equilibrium model of China (CEEGE model), which follows \nthe framework of the CHINAGEM model provided by the Centre of \nPolicy Studies at Victoria University26,27, which is based on China\u2019s \ninput\u2012output table in 2017 with detailed energy sectors. To achieve \nthis goal, we establish a comprehensive policy portfolio area towards \n2060 that consists of 1,295 scenarios based on four different types of \naggregated mitigation strategies (carbon pricing (C), energy efficiency \n(A), renewable energy (R) and electrification of end uses (E)). Various \ngrades of policy intensity and portfolios are specified to reduce the \nuncertainty of future mitigation policies. For each policy portfolio, \nwe examine the trade-offs and synergies between mitigation policies \non mitigation effects, the economic costs of mitigation and efficiency \n(GDP loss per unit carbon reduction) by comparing the results with or \nwithout the policy interaction assumption.\nOur results indicate that the interactions between mitigation poli-\ncies make attaining China\u2019s carbon neutrality more challenging than \nexpected, if this expectation refers to a simple aggregation of the direct \neffects of each mitigation policy implemented separately12,20. The per-\ncentage of scenarios achieving carbon neutrality by 2060 decreases sub-\nstantially by 84%, and the attainment of carbon neutrality under these \nscenarios is postponed by 5\u20136\u2009years when the interactions between \nmitigation policies are considered. While synergetic effects expand the \npotential space of policies with complementary mechanisms, trade-off \neffects compress the space of policies with competing mechanisms. \nAmong the analysed policies, only combinations of renewable energy \nand electrification of end uses are complementary and show no com-\npeting mechanisms compressing the space of policies. In contrast, \ncombinations of carbon pricing and renewable energy have the greatest \ntrade-off effect. Thus, the findings highlight the need for careful policy \ncombination to minimize carbon emissions and economic losses.\n2020\n2025\n2030\n2035\n2040\n2045\n2050\n2055\n\u20132\n0\n2\n4\n6\n8\n10\n12\n2060\n2055\n2050\n2045\n2040\n2035\n2030\n2025\n2020\n\u20132\n0\n2\n4\n6\n8\n10\n12\nCO2 emissions (Gt)\nASIA\nISSA\n2050\n2052\n2054\n2056\n2058\n2060\n2062\nYear\nASIA\nYear\nISSA\na\nb\nFig. 1 | Emission pathways in China from 2020 to 2060. a, CO2 emission under \nASIA and ISSA (GtCO2). The divergent colour from light to dark reflects the \nincreasing number of policies combined. The black line represents emissions \nunder the baseline scenario. The white dot represents residual CO2 emissions \nallowed (2\u2009GtCO2) to achieve the carbon neutrality target considering the carbon \nsequestration capacity of carbon capture and storage (CCS) technology and \nforest systems. The sample size is 1,295, including all policy scenarios.  \nb, Distribution of years for the scenarios in which the carbon neutrality target \nis achieved. The width of the density map shows the frequency of the data \ndistribution and the box plot shows the first quantile, mean value and third \nquantile of all the results. The sample size is 48, including policy scenarios \nachieving carbon neutrality under ASIA.\n\nNature Climate Change | Volume 15 | February 2025 | 147\u2013152\n149\nArticle\nhttps://doi.org/10.1038/s41558-024-02237-2\nconsiderably under ASIA and ISSA (Fig. 2). To test the impact of each \npolicy combination, two mitigation strategies are combined and \nmaintained at their highest levels, which represent those prioritized, \nwhereas the other two strategies are randomly combined with varying \nintensities. Under ISSA, residual CO2 emissions in 2060 for all mitiga-\ntion policy scenarios range from \u22120.8 to 2.8\u2009GtCO2, with the proportion \nof mitigation policy scenarios achieving carbon neutrality ranging from \n88% (A5R5CxEx and C5R5AxEx) to 100% (A5E5CxRx and C5E5AxRx). However, \nunder ASIA, residual CO2 emissions rise to 1.4\u20133.8\u2009GtCO2, with the \nproportion of scenarios in which carbon neutrality is achieved being \nreduced to a range of 28% (C5A5RxEx) to 72% (R5E5CxAx).\nThe ISSA results suggest that the most effective policy combina-\ntions for achieving the carbon neutrality target are dominated by the \nhighest levels of A and E, or C and E, with an average attainment of carbon \nneutrality of 2053\u20132054 (Fig. 2a). However, the most effective policy \ncombination under ASIA is that of R and E, with an average timeline of \nreaching carbon neutrality by 2059. The electrification of end uses is a \nkey factor for attaining carbon neutrality, as the most potent portfolios \nmust contain E policies under both AISA and ISSA. Furthermore, in terms \nof the timeline for achieving carbon neutrality, all scenarios under ASIA \nare projected to achieve the target later than those under ISSA. The aver-\nage economic costs under ISSA and ASIA differ considerably (Fig. 2b). \nThe results reveal that, under ISSA, the average cumulative economic \ncost from 2020 to 2060 is ~3.8% of GDP, which is much greater than the \nresults (~2.9%) of the scenarios under ASIA, with the best economic \nperformance observed in the combination of the highest level of A \nand R. Nevertheless, under ASIA, the best policy combination from the \neconomic side changes to the highest level of R and E. R policy clearly \nplays a crucial role in determining the economic cost.\nTrade-offs and synergies between mitigation \npolicies\nOur findings highlight the notable synergetic effect of the policy com-\nbination of renewable energy and electrification of end uses (R and E), \nwhich results in a 5\u201320% increase in carbon reduction and a \u22125% to \u221225% \ndecrease in cumulative GDP cost by 2060 under ASIA compared with \nISSA (Fig. 3). The E policy supports the substitution of electricity for \nfossil fuels by encouraging end-use sectors to adjust energy consump-\ntion preferences and increase electricity demand; in contrast, the R \npolicy increases the proportion of green electricity by reducing the \ncost of renewable electricity generation in the overall electricity mix. \nOn the economic cost side, the R policy promotes low-cost renewable \nelectricity and mitigates the economic cost. Even though the E policy \nincreases the cost of electricity production due to its expanded scale, \nthe R policy helps mitigate the economic impact. Thus, when R and E \npolicies are combined, the policy space of each is expanded to deliver \nresults that are stronger than those from individual implementation, \nhighlighting their potential mutual complementarity.\nResidual CO2 emissions (Gt)\nEconomic cost (%)\nA5R5\nCxEx\n88%\n88%\n92% 96% 100% 100%\n64% 60%\n28%\n72%\n32%\n28%\n73%\n68%\n22%\n58%\n4%\n33% 33% 29%\n94%\n83% 86%\nC5R5\nAxEx\nR5E5\nCxAx\nC5A5\nRxEx\nA5E5\nCxRx\nC5E5\nAxRx\nC5E5\nAxRx\nA5E5\nCxRx\nC5A5\nRxEx\nR5E5\nCxAx\nC5R5\nAxEx\nA5R5\nCxEx\nA5R5\nCxEx\nC5R5\nAxEx\nR5E5\nCxAx\nC5A5\nRxEx\nA5E5\nCxRx\nC5E5\nAxRx\nC5E5\nAxRx\nA5E5\nCxRx\nC5A5\nRxEx\nR5E5\nCxAx\nC5R5\nAxEx\nA5R5\nCxEx\n2053\n2053\n2054\n2052\n2053\n2054\n2059\n2059\n2059\n2059\n2059\n2059\nISSA\nASIA\nISSA\nASIA\na\nb\n4\n3\n2\n1\n0\n\u20131\n\u20131.5\n\u20132.0\n\u20132.5\n\u20133.0\n\u20133.5\n\u20134.0\n\u20134.5\n\u20135.0\nFig. 2 | China\u2019s residual emissions and cumulative economic cost in 2060.  \na, Residual emissions of scenarios with two maximum policy intensities under \nISSA and ASIA (GtCO2). The box plot shows the first, second (the median value) \nand third quantiles of all the results, where the maximum and minimum values \nare denoted with whiskers. The divergent colour from light to dark reflects the \nincreasing policy intensity of the scenarios. The dashed line represents the \nresidual CO2 emissions that reach the carbon neutrality target, which is set below \n2\u2009GtCO2 considering the carbon sequestration capacity from CCS technology \nand forest systems. The column chart represents the percentage of scenarios \nthat reach the carbon neutrality target by 2060, and the text in the bar chart \nrepresents the average year for policy combinations to achieve the target. \nThe sample size is 150, including policy scenarios with two maximum policy \nintensities. For example, A5R5CxEx indicates the scenarios with the highest A \nand R policy levels, which represents those that are prioritized. b, Cumulative \neconomic cost of mitigation (%) for scenarios achieving the carbon neutrality \ntarget with two maximum policy intensities. The dashed lines represent the \naverage economic costs under ISSA and ASIA. The column chart represents the \nproportion of scenarios with economic costs below the value indicated by the \ndashed line. The box plot shows the first, second (the median value) and third \nquantiles of all the results, where the maximum and minimum values are denoted \nwith whiskers. The sample size is 150, including policy scenarios with two \nmaximum policy intensities.\n\nNature Climate Change | Volume 15 | February 2025 | 147\u2013152\n150\nArticle\nhttps://doi.org/10.1038/s41558-024-02237-2\nBy contrast, the carbon pricing and renewable energy (C and R) \npolicy combination is the only one that has trade-off effects from both \nmitigation and economic perspectives. The carbon reduction attains \na further 4\u201314% drop, whereas the cumulative GDP cost increases by \nan extra 2\u201312% under ASIA compared with that under ISSA by 2060. \nFrom a mitigation standpoint, the mechanism of the C policy is to \nincrease the cost of fossil energy to encourage the shift to renewable \nenergy, whereas the mechanism of the R policy is to increase the share \nof renewable energy by reducing the cost of renewable energy. These \nmechanisms have high similarity and consistency, which leads to each \npolicy having a reduced policy space when they are implemented \ntogether. The same applies to the economic side, as the benefit of \nthe R policy on the economic cost is also inhibited by the simultane-\nous implementation of the C policy. In summary, when policies with \ncompeting mechanisms, such as C and R, are implemented together, \ntrade-offs are likely to arise, as the potential policy space is compressed \nby the policies involved.\nThe other policy combinations have a trade-off effect in terms \nof carbon mitigation but a synergetic effect on the economic cost. \nCompared with ISSA, C and A, A and R, C and E and A and E under ASIA \ndecrease additional carbon reduction by an average of \u221212%, \u22129%, \u22124% \nand \u22125%, respectively, by 2060. The corresponding additional GDP \nlosses are reduced by an average of \u22124%, \u22123%, \u22127% and \u22128%, respectively. \nThe most substantial improvement in mitigation efficiency is observed \nin the R and E combination, followed by A and E and C and E. The largest \ndecrease in efficiency is seen in the C and R combination, followed by C \nand A and A and R. The same trends are evident when more mitigation \npolicies are combined (Supplementary Section 5); specifically, the \nsimultaneous implementation of policies with R and E at high levels \nhas notable positive impacts on all dimensions.\nIn summary, synergetic effects occur when mitigation policies \nwith complementary mechanisms are jointly implemented, expanding \ntheir potential policy space. Conversely, trade-off effects occur when \nmitigation policies with competing mechanisms are jointly imple-\nmented, compressing their policy space. Beyond mitigation efficiency, \nmitigation policies also affect household income, consumption and \nutility levels through changes in electricity prices (Supplementary  \nSection 5). A comparison of the results of ISSA and ASIA reveals that the \nR and E scenario still has the most notable positive effect on increas-\ning the cumulative household consumption scale and utility levels \nby 2060, which is the opposite for the C and R scenario. A detailed \ndiscussion of the household and sectoral results can be found in Sup-\nplementary Section 5.\nPolicy implications for improving mitigation \nefficiency\nThe Chinese government should explore the synergetic effect of \nmitigation policies, which could be largely achieved by constructing \nlarge-scale renewable energy generation and gradually expanding the \nshare of electricity in final energy consumption. The following sugges-\ntions can be used when designing the alignment of the implementation \npace and intensity of policies.\nAccelerate the development and commercialization of energy \nstorage technologies\nEnergy storage addresses the intermittency of renewable energy and \ngrid load shortages, promoting renewable energy penetration and \ndistributed grids. Compared with the baseline scenario, energy stor-\nage can increase emission reductions at a cost to the economy (Sup-\nplementary Section 5). Documents from the Chinese government33,34 \noutline goals for energy storage technology, with new energy storage \ntechnologies fully market-oriented by 2030. However, China\u2019s finan-\ncial and tax support for energy storage is still in the developmental \nstage and is mainly advisory, lacking comprehensive and long-term \nincentives, in contrast to the situation in developed countries such \nas the United States and Germany35. Additionally, the levelized cost \nin China among various types of storage facilities ranges from 0.12 to \n0.27\u2009US$\u2009kWh\u22121 (refs. 36,37). Among them, lithium-ion batteries have \nrelatively low costs, while hydrogen energy storage tends to be more \nexpensive. Thus, differentiated subsidy policies should be considered \non the basis of the type of technology and the application scenario of \nstorage energy facilities.\nEmphasize the electrification targets for end-use sectors\nAlthough the Chinese government has issued guidelines related to \nend-use electrification38, the measures proposed at the sectoral level \nare mostly framework oriented and lack detailed policies for effective \nimplementation (Supplementary Section 2). An overall electrification \ntarget for end-use sectors, such as transport and construction, are \nrelatively underdeveloped39,40. Future policies should include explicit \ntargets for end-use sectors to increase electrification. According to our \nresults and related studies41\u201343, total electricity consumption needs to \naccount for no less than 70% of end-use energy consumption to achieve \ncarbon neutrality by 2060. By sector, manufacturing sectors with a high \nlevel of electrification, such as textiles and equipment manufacturing, \nneed to maintain an electrification rate of no less than 80% and for the \nconstruction and transport sectors, the share of electricity should \nexceed 70% and 50%, respectively.\nSpecify long-term emission abatement strategies\nAmong the 94 government documents collected35 and reviewed in this \nstudy, 59 focus on goals for 2025, 25 focus on goals for 2030 and 18 \nfocus on goals for 2035 (Supplementary Sections 2 and 4). Notably, only \ntwo documents44,45 propose long-term development goals for 2060, \nrepresenting 4.9% of the total policies proposed after 2020. In contrast, \nFrance, the United Kingdom and the European Union have introduced \nfour, nine and nine policies, respectively, that target 2050, accounting \n\u201325\n\u201320\n\u201315\n\u201310\n\u20135\n0\n5\n10\n15\n20\n25\n\u201325\n\u201320\n\u201315\n\u201310\n\u20135\n0\n5\n10\n15\n20\n25\nRelative change of economic cost (%)\nRelative change of carbon reduction (%)\nR and E\nA and E\nC and E\nA and R\nC and A\nC and R\nFig. 3 | Carbon reduction and economic cost of policy combinations under \nASIA compared with ISSA (%). The colour of the dots denotes the intensity of \nthe first policy in the policy combinations (darker colours represent greater \nintensity) and the size of the dots denotes the intensity of the second policy \n(larger sizes represent greater intensity). The sample size is 150, including all \ndual-policy scenarios that implement two policies together.\n\nNature Climate Change | Volume 15 | February 2025 | 147\u2013152\n151\nArticle\nhttps://doi.org/10.1038/s41558-024-02237-2\nfor 14.3%, 18.8% and 23.1% of all policies proposed after 2020, respec-\ntively35. Thus, the Chinese government must enhance the planning and \nformulation of emission reduction policies from a long-term perspec-\ntive. For example, according to the simulation results presented here, \nChina\u2019s emissions need to decrease by more than 25% and 55% in 2040 \nand 2050, respectively, compared with current emissions to ensure the \nachievement of the carbon neutrality target by 2060.\nIn addition to a global perspective, carbon mitigation efforts \ndepend not only on the policies adopted but also on the policies \nachieved46. To achieve the climate target in a just, orderly and equi-\ntable manner, rather than blindly pursuing more stringent carbon \nreduction policies, which could lead to escalating economic losses, \ndeveloping countries should focus more on optimizing the combina-\ntion and sequence of carbon reduction policies16,47. This approach not \nonly enhances the effectiveness of carbon reduction but also minimizes \nthe economic cost of carbon reduction efforts. Additionally, since \npolicy interactions can have implications beyond national boundaries \nunder the need to advance national and global climate policy48, policy \ninteractions must also be considered for international cooperation \namong countries to reduce carbon emissions.\nIn summary, this study assesses the effectiveness of various miti-\ngation strategies to achieve China\u2019s carbon neutrality target with the \nCEEGE model. Model robustness is tested through uncertainty analysis \nacross five dimensions: baseline scenarios, mitigation policies, model \nparameters, energy storage technology and carbon sequestration \ncapacity (Methods and Supplementary Section 6). Additionally, two \npoints can be further explored. First, this study focuses on mature \nand foreseeable renewable energy and storage technologies, exclud-\ning emerging technologies such as green hydrogen, because of the \nuncertain changing trend in production costs49,50 and data limita-\ntions for modelling. Second, although 1,295 scenarios with varying \npolicy intensities are created here, policy combinations in grades of \npolicy intensities at different timelines have not been considered, \nwhich could be an important direction for future research to analyse  \npolicy sequencing.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-024-02237-2.\nReferences\n1.\t\nWei, Y. et al. Self-preservation strategy for approaching global \nwarming targets in the post-Paris agreement era. Nat. Commun. \n11, 1624 (2020).\n2.\t\nParry, I., Mylonas, V. & Vernon, N. Mitigation policies for the Paris \nagreement: an assessment for G20 countries. J. Assoc. Environ. \nResour. Econ. 8, 797\u2013823 (2021).\n3.\t\nWei, Y. M. et al. Policy and management of carbon peaking  \nand carbon neutrality: a literature review. Engineering 14,  \n52\u201363 (2022).\n4.\t\nLiu, Z. et al. Challenges and opportunities for carbon neutrality in \nChina. Nat. Rev. Earth Environ. 3, 141\u2013155 (2022).\n5.\t\nWeng, Y., Cai, W. & Wang, C. Evaluating the use of BECCS and \nafforestation under China\u2019s carbon-neutral target for 2060.  \nAppl. Energy 299, 117263 (2021).\n6.\t\nZheng, Z., Hu, G., Mu, X. & Li, K. From low carbon to carbon \nneutrality: a bibliometric analysis of the status, evolution and \ndevelopment trend. J. Environ. Manag. 322, 116087 (2022).\n7.\t\nJiang, H. D., Xue, M. M., Liang, Q. M., Masui, T. & Ren, Z. Y. How \ndo demand-side policies contribute to the electrification and \ndecarburization of private transportation in China? A CGE-based \nanalysis. Technol. Forecast. Soc. Change 175, 121322 (2022).\n8.\t\nProject Team on the Strategy and Pathway for Peaking Carbon \nEmissions and Carbon Neutrality. Analysis of a peaked carbon \nemission pathway in china toward carbon neutrality. Engineering \n7, 1673\u20131677 (2021).\n9.\t\nDuan, H. et al. Assessing China\u2019s efforts to pursue the 1.5\u2009\u00b0C \nwarming limit. Science 372, 378\u2013385 (2021).\n10.\t Zhang, S. & Chen, W. Assessing the energy transition in China \ntowards carbon neutrality with a probabilistic framework.  \nNat. Commun. 13, 87 (2022).\n11.\t\nJiang, H. D., Pradhan, B. K., Dong, K., Yu, Y. Y. & Liang, Q. M. An \neconomy-wide impacts of multiple mitigation pathways toward \ncarbon neutrality in China: a CGE-based analysis. Energy Econ. \n129, 107220 (2024).\n12.\t Vigui\u00e9, V. & Hallegatte, S. Trade-offs and synergies in urban \nclimate policies. Nat. Clim. Change 2, 334\u2013337 (2012).\n13.\t Pe\u00f1asco, C., Anad\u00f3n, L. D. & Verdolini, E. Systematic review of the \noutcomes and trade-offs of ten types of decarbonization policy \ninstruments. Nat. Clim. Change 11, 257\u2013265 (2021).\n14.\t Bertram, C. et al. Complementing carbon prices with technology \npolicies to keep climate targets within reach. Nat. Clim. Change 5, \n235\u2013239 (2015).\n15.\t Pahle, M. et al. Sequencing to ratchet up climate policy \nstringency. Nat. Clim. Change 8, 861\u2013867 (2018).\n16.\t Meckling, J., Sterner, T. & Wagner, G. Policy sequencing toward \ndecarbonization. Nat. Energy 2, 918\u2013922 (2017).\n17.\t Chen, X. & Lin, B. Q. Towards carbon neutrality by implementing \ncarbon emissions trading scheme: policy evaluation in China. \nEnergy Policy 157, 112510 (2021).\n18.\t Liang, S., Lin, X. Y., Liu, X. X. & Pan, H. R. The pathway to China\u2019s \ncarbon neutrality based on an endogenous technology CGE \nModel. Int. J. Environ. Res. Public Health 19, 6251 (2022).\n19.\t Yang, X., Nielsen, C. P., Song, S. & McElroy, M. B. Breaking the \nhard-to-abate bottleneck in China\u2019s path to carbon neutrality with \nclean hydrogen. Nat. Energy 7, 955\u2013965 (2022).\n20.\t Xie, Y. S. et al. More synergies or more trade-offs? The interaction \namong multiple assessment indicators in sustainable urban \ndevelopment in Guangzhou, China. J. Environ. Manag. 324, \n116324 (2022).\n21.\t Xing, J. et al. Integrated crop\u2013livestock\u2013bioenergy system brings \nco-benefits and trade-offs in mitigating the environmental \nimpacts of Chinese agriculture. Nat. Food 3, 1052\u20131064 (2022).\n22.\t Zhang, R. & Hanaoka, T. Cross-cutting scenarios and strategies \nfor designing decarbonization pathways in the transport sector \ntoward carbon neutrality. Nat. Commun. 13, 3629 (2022).\n23.\t Lam, A. & Mercure, J. Which policy mixes are best for decarbonising \npassenger cars? Simulating interactions among taxes, subsidies \nand regulations for the United Kingdom, the United States, Japan, \nChina, and India. Energy Res. Soc. Sci. 75, 101951 (2021).\n24.\t Axsen, J., Pl\u00f6tz, P. & Wolinetz, M. Crafting strong, integrated \npolicy mixes for deep CO2 mitigation in road transport.  \nNat. Clim. Change 10, 809\u2013818 (2020).\n25.\t Sun, L., Cui, H. & Ge, Q. Will China achieve its 2060 carbon \nneutral commitment from the provincial perspective?  \nAdv. Clim. Change Res. 13, 169\u2013178 (2022).\n26.\t Mai, Y., Dixon, P. & Rimmer, M. CHINAGEM: A Monash-Styled \nDynamic CGE Model of China (Centre of Policy Studies/IMPACT \nCentre, 2010).\n27.\t Dixon, P., Koopman, R. & Rimmer, M. The MONASH style of \ncomputable general equilibrium modeling: a framework for \npractical policy analysis. Handb. Comput. Gen. Equip. Model 1, \n23\u2013103 (2013).\n28.\t China Energy Statistical Yearbook (China Statistics Press, 2022).\n29.\t World Energy Outlook 2022 (IEA, 2022).\n30.\t China Statistical Yearbook (China Statistics Press, 2022).\n31.\t World Population Prospects (United Nations, 2022).\n\nNature Climate Change | Volume 15 | February 2025 | 147\u2013152\n152\nArticle\nhttps://doi.org/10.1038/s41558-024-02237-2\n32.\t Wang, C., Guan, D. & Cai, W. Grand challenges cannot be treated \nin isolation. One Earth 1, 24\u201326 (2019).\n33.\t The 14th Five-Year Plan for New Energy Storage Development \nImplementation Plan (NDRC, 2022); www.gov.cn/zhengce/\nzhengceku/2022-03/22/content_5680417.htm\n34.\t Medium and Long-Term Plan for the Development of Pumped \nStorage Industry (2021\u20132035) (NEA, 2021); www.nea.gov.cn/ \n2021-09/09/c_1310177087.htm\n35.\t NewClimate Institute, Wageningen University and Research & PBL \nNetherlands Environmental Assessment Agency. Climate policy \ndatabase. Zenodo https://doi.org/10.5281/zenodo.10869734 (2023).\n36.\t Wen, J., Liu, N., Pei, J., Ruochen, X. & Dawei, L. Life cycle cost \nanalysis for energy storage technology. Thermal Power Gener. 50, \n24\u201329 (2021).\n37.\t Zhu, Y. et al. Comparative techno-economic evaluation of energy \nstorage technology: a multi-time scales scenario-based study in \nChina. J. Energy Storage 89, 111800 (2024).\n38.\t Guidelines on Further Promoting the use of Electricity to Replace \nOther Energy Sources (NDRC, 2022); www.jnxh.org.cn/html/xwdt/\ntzgg/250.html\n39.\t Nascimento, L. & H\u00f6hne, N. Expanding climate policy adoption \nimproves national mitigation efforts. npj Clim. Action 2, 12 (2023).\n40.\t Nascimento, L. et al. Twenty years of climate policy: G20 \ncoverage and gaps. Clim. Policy 22, 158\u2013174 (2021).\n41.\t He, J. et al. Towards carbon neutrality: a study on China\u2019s \nlong-term low-carbon transition pathways and strategies.  \nEnviron. Sci. Ecotechnol. 9, 100134 (2022).\n42.\t Jiang, K., He, C., Zhu, S., Xiang, P. & Chen, S. Transport scenarios \nfor China and the role of electric vehicles under global 2\u2009\u00b0C/1.5\u2009\u00b0C \ntargets. Energy Econ. 103, 105172 (2021).\n43.\t Yu, S. et al. Synthesis Report 2022 on China\u2019s Carbon Neutrality: \nElectrification in China\u2019s Carbon Neutrality Pathways  \n(Energy Foundation China, 2022); www.efchina.org/Reports-en/\nreport-lceg-20221104-en\n44.\t Working Guidance for Carbon Dioxide Peaking and Carbon \nNeutrality (The State Council of the People\u2019s Republic of  \nChina, 2021); www.gov.cn/gongbao/content/2021/content_ \n5649728.htm\n45.\t Notice on Printing and Distributing the \u2018Guiding Opinions on \nPromoting the High-Quality Development of Central Enterprises \nand Doing a Good Job in Carbon Neutralization\u2018 (SASAC, 2021); \nwww.sasac.gov.cn/n2588035/c22499825/content.html\n46.\t Fransen, T. et al. Taking stock of the implementation gap in \nclimate policy. Nat. Clim. Change 13, 752\u2013755 (2023).\n47.\t Dai, M. et al. Country-specific net-zero strategies of the pulp and \npaper industry. Nature 626, 327\u2013334 (2024).\n48.\t Nascimento, L. et al. Climate policy in 2023. Nat. Rev. Earth \nEnviron. 5, 255\u2013257 (2024).\n49.\t Zhen, Z., Ou, X., Wang, Y. & Zhou, S. Assessing transition pathways \nof hydrogen production in China with a probabilistic framework. \nEnviron. Sci. Technol. 58, 13263\u201313272 (2024).\n50.\t Risco-Bravo, A., Varela, C., Bartels, J. & Zondervan, E. From green \nhydrogen to electricity: a review on recent advances, challenges, \nand opportunities on power-to-hydrogen-to-power systems. \nRenew. Sust. Energ. Rev. 189, 113930 (2024).\nPublisher\u2019s note Springer Nature remains neutral with  \nregard to jurisdictional claims in published maps and  \ninstitutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2025\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02237-2\nMethods\nWe use a dynamic computable equilibrium model of China (CEEGE \nmodel) and establish a policy range containing 1,295 scenarios that \ncover four mitigation strategies with five intensity grades and their \nportfolios. It can depict the cost-neutrality mechanisms of mitigation \npolicies, capture dynamic sectoral changes through the upstream\u2013\ndownstream industrial chain and adjust both quantity and price at \nequilibrium simultaneously. Three indicators are used to measure \nthe trade-offs and synergies between policy couples in terms of \ntheir carbon-mitigating effects, economic costs of mitigation and \ncarbon-mitigating efficiency. The research framework is presented \nin Supplementary Section 1.\nCEEGE model\nThe CEEGE model is a dynamic computable equilibrium model of \nChina, following the framework of the CHINAGEM model provided \nby the Centre of Policy Studies at Victoria University26,27. The CEEGE \nmodel is based on the neoclassical economics theory and assumes \nthat the market is fully competitive and that the returns to scale of \nproduction are constant. It contains six economic agents (production, \ninvestment, consumption, government, foreign and inventory) and \nthree primary factors (labour, capital and land) and is solved with the \nGEMPACK software51. In this model, a system of linearized equations is \nestablished to describe the behaviours of agents in response to price \nchanges and determines the equilibrium price and quantity by equat-\ning the demand and supply for all goods and factors. The model can be \nused to capture the direct and indirect effects of exogenous changes in \nthe economy and assess the impact mechanism across the economy. \nHence, it provides a valuable tool for various policy-oriented studies \nrelated to carbon mitigation52,53. The modules of production, house-\nhold consumption, governmental consumption, investment, export \ndemand, carbon accounting and pricing, the dynamic module, the \nequilibrium mechanism, macro-economic closure and energy storage \nare presented in Supplementary Section 4.\nScenario design\nWe establish a baseline scenario for the 2017\u20132060 period and 1,295 \nmitigation policy scenarios considering different intensities and port-\nfolios of four major mitigation strategies (Supplementary Section 3).  \nThe baseline scenario reflects future economic growth, energy con-\nsumption and carbon emissions under the current mitigation poli-\ncies. The historical economic growth and energy consumption data \nfor 2017\u20132021 are calibrated according to the NBSC28,30 and British \nPetroleum Statistical Review54. For 2022 to 2060, the projections of \nGDP growth, population, labour and industrial structure refer mainly \nto existing studies31,55,56. The energy consumption towards 2050 is \ncalibrated by endogenizing energy efficiency improvement rates \naccording to the projection of the IEA29 and extrapolated to the year \n2060. In 2060, the annual GDP and population growth rate will decline \nto 2.5% and \u22120.6%, respectively; energy consumption will increase to \n6,054\u2009Mtce and carbon emissions will decrease to 7.9\u2009GtCO2.\nTo establish mitigation policy scenarios, we extensively review the \nexisting mitigation policies (Supplementary Section 2) and summarize \nthe four most important mitigation strategies, including energy effi-\nciency improvement, renewable energy development, electrification \nof end uses and carbon pricing. Energy efficiency policy (A) increases \nthe energy efficiency improvement rate of industries and residents; \nthe economic cost of energy efficiency policy is burdened by industries \nand residents according to their consumption of energy products. \nRenewable energy policy (R) refers to the decrease in generation and \naccommodation costs of solar power, onshore wind power, offshore \nwind power and bioenergy power due to technological progress; the \ndevelopment of renewable energy sectors will crowd out investment \nin fossil energy sectors, so the declining economic costs in the renew-\nable energy sector will be accompanied by rising costs in the fossil  \nenergy sector. The electrification of end-uses policy (E) prompts \nindustries and residents to replace fossil fuels with electricity in their \nterminal energy consumption while keeping total terminal energy \nconsumption unchanged. The carbon pricing policy (C) is modelled \nas a tax on carbon emissions emitted by industries and residents and \nthe carbon tax revenue is recycled to residents via lump-sum transfer.\nThen, on the basis of a systematic literature review of the above- \nmentioned policies, we specify five grades for the intensity of each \npolicy (Supplementary Section 3). The first and fifth grades correspond \nto the lowest and highest intensities of each policy. Given five grades for \nthe four mitigation strategies, 20 single-policy scenarios are obtained. \nCombining these mitigation policies with different grades, we obtain \n150 dual-policy scenarios which implement two policies together,  \n500 triple-policy scenarios and 625 quad-policy scenarios. Finally,  \n1,295 mitigation policy scenarios are established to provide a relatively \nmore comprehensive consideration of the uncertainty in the intensities \nand portfolios of future mitigation policies.\nUncertainty discussions\nThis paper discusses the uncertainty of the simulation results from five \naspects: the baseline scenario, mitigation policies, model parameters, \nenergy storage technology and carbon sequestration capacity. First, \nChina\u2019s carbon dioxide emission pathway may present varying predic-\ntions under the baseline scenario (Supplementary Section 3) and the \nstrictness of achieving the mitigation target directly impacts emission \nreduction and economic costs. Thus, we set two scenarios for acceler-\nating and delaying the emission reduction progress in the baseline to \ntest this uncertainty. Second, different intensities of mitigation policies \nmay lead to a nonlinear changing trend or turning point for reduced \nemissions and economic losses. To test this uncertainty, we created \n1,295 scenarios by combining different policies with different intensi-\nties on the basis of an extensive literature review (Supplementary Sec-\ntion 3.2). We also developed two additional policies with more stringent \nintensity (grade 6) and weaker intensity (grade 0). Third, the simulation \nresults for achieving the carbon neutrality target may be affected by the \nsubstitution elasticity between different energy products in the model. \nTo address this uncertainty, the substitution elasticity between energy \nproducts was increased or reduced by 25% to test the robustness of the \nresults. Fourth, energy storage technology can additionally support \nthe growth of high proportions of renewable energy and increase the \neconomic cost. In this paper, we set two scenarios with high-cost and \nlow-cost storage energy development to test the robustness of the \nsimulation results. Finally, enhancing the carbon sequestration capac-\nity is crucial for achieving carbon neutrality goals, as the percentage \nof scenarios in which the carbon neutrality target is achieved by 2060 \nwill decrease substantially from 11.1% to 0.23%, with a decrease in the \ncarbon sequestration capacity from 2.5\u2009GtCO2 to 1.5\u2009GtCO2 (Supple-\nmentary Fig. 22). In this study, we assumed that the capacity is limited \nto a maximum of 2\u2009GtCO2 from CCS technology and forest systems by \n2060. On the basis of the five categories of uncertainty analysis above, \nwe find that the trade-offs and synergies between mitigation policy \npairs remain robust: the simultaneous implementation of policies with \nR and E at high levels has notable positive impacts on all dimensions, \nand the policy combination with C and R has notable negative impacts \non all dimensions. See Supplementary Section 6 for details about the \nuncertainty analysis.\nReporting summary\nFurther information on research design is available in the Nature \nPortfolio Reporting Summary linked to this article.\nData availability\nAll data supporting the conclusions of this work are provided in Sup-\nplementary Section 7 and are available via Zenodo at https://doi.org/ \n10.5281/zenodo.14325940 (ref. 57).\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02237-2\nCode availability\nThe code of the dynamic computable equilibrium model of China \n(CEEGE model) is available via Zenodo at https://doi.org/10.5281/\nzenodo.14325940 (ref. 57).\nReferences\n51.\t Harrison, W. & Pearson, K. Computing solutions for large  \ngeneral equilibrium models using GEMPACK. Comput. Econ. 9, \n83\u2013127 (1996).\n52.\t Hertel, T. Global applied general equilibrium analysis using the \nglobal trade analysis project framework. Handb. Comput. Gen. \nEquip. Model 1, 815\u2013876 (2013).\n53.\t Anderson, K. & Wittwer, G. in Policy Analysis and Modeling of the \nGlobal Economy (ed. Maskus, K.) 249\u2013278 (World Scientific, 2021).\n54.\t Statistical Review of World Energy (British Petroleum, 2021);  \nwww.bp.com/en/global/corporate/energy-economics/statistical- \nreview-of-world-energy.html\n55.\t Zhang, X. et al. Energy\u2013economic transformation pathways  \nand policies for carbon neutrality targets. Manag. World 38, \n35\u201366 (2022).\n56.\t Real GDP Long-Term Forecast (Indicator) (OECD, 2023); www.\noecd.org/en/data/indicators/real-gdp-long-term-forecast.html\n57.\t Liu, Y. et al. Code scripts for \u2018Mitigation policies interactions  \ndelay the achievement of carbon neutrality in China\u2019. Zenodo \nhttps://doi.org/10.5281/zenodo.14325940 (2024).\nAcknowledgements\nYu Liu was supported by the National Natural Science Foundation of \nChina (grant nos. 72125010, 72243011 and 71974186), The Fundamental \nResearch Funds for the Central Universities, Peking University and \nHigh-performance Computing Platform of Peking University. M.D. is \nsupported by the National Natural Science Foundation of China  \n(grant no. 72304222) and Young Talent Program of Xi\u2019an Jiaotong \nUniversity (grant no. GG6J007). Q.C. was supported by National Natural \nScience Foundation of China (grant no. 72373163) and Beijing Natural \nScience Foundation (grant no. 9222016), Fundamental Research \nFunds for the Central Universities from China Petroleum University \n(grant no. 23CX06039A) and Taishan Scholar Program of Shandong \nProvince. Yawen Liu was supported by the China Postdoctoral Science \nFoundation (grant no. 2023M730590) and the Fundamental Research \nFund for the Central Universities in UIBE (grant no. 22QD30).\nAuthor contributions\nYu Liu, M.D. and L.Y. conceived the research. Yu Liu, M.D., L.Y. and Q.C. \ndesigned the research. Q.C., L.Y., Yu Liu, M.D. and Yawen Liu collected \nthe data and conducted the model simulations with inputs from X.L., \nN.Z., Y. Li and C.J. M.D., Q.C., L.Y. and Yu Liu interpreted the results and \nwrote the paper with inputs from Yawen Liu, X.L., N.Z., Y. Li, C.J., P.Z., \nQ.L. and C.H. All authors discussed the results and contributed to \nrevising the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41558-024-02237-2.\nCorrespondence and requests for materials should be addressed to \nYu Liu, Mingxi Du or Qi Cui.\nPeer review information Nature Climate Change thanks Sheng Zhou \nand the other, anonymous, reviewer(s) for their contribution to the \npeer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\n1\nnature portfolio  |  reporting summary\nMarch 2021\nCorresponding author(s):\nYu Liu, Mingxi Du and Qi Cui\nLast updated by author(s): Dec 8, 2024\nReporting Summary\nNature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nEconomic data and emission data are based on the China's input-output table of the year 2017 and China Energy Statistical Yearbook(2022).\nData analysis\nThe CEEGE model (China Energy-Economic Computable General Equilibrium Model)  is used to estimate the ERR, ECT and EER indicator; Excel \n( Microsoft 365MSO) and Origin (version 2019) are used to do calculations and draw figures.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and \nreviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A description of any restrictions on data availability \n- For clinical datasets or third party data, please ensure that the statement adheres to our policy \n \nAll data supporting the conclusions of this work are provided at Supplementary Information and https://doi.org/10.5281/zenodo.14325940.\n\n2\nnature portfolio  |  reporting summary\nMarch 2021\nHuman research participants\nPolicy information about studies involving human research participants and Sex and Gender in Research. \nReporting on sex and gender\nThis information has not been collected. This information is not relevant here.\nPopulation characteristics\nThis information has not been collected. This information is not relevant here.\nRecruitment\nThis information has not been collected. This information is not relevant here.\nEthics oversight\nThis information has not been collected. This information is not relevant here.\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nEcological, evolutionary & environmental sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nHere we use a dynamic computable general equilibrium model, create a policy portfolio area of 1,295 scenarios covering four major \nmitigation strategies, and examined trade-offs and synergies between mitigation policies on mitigation effects, and economic costs \nby comparing the results between the ASIA and ISSA assumptions.\nResearch sample\n159 production sectors and 157 products based on the China Input-output table  and the China Electric Power Yearbook are involved \nin this study.\nSampling strategy\nFor this study, the \u2018electricity\u2019 sector in the original input-output table is split into nine new sectors with different power sources, \nincluding eight generation sectors and one sector of \u2018power transmission and distribution\u2019. The original \u2018crude oil and gas sector\u2019 is \nsplit into three new sectors: \u2018crude oil\u2019, \u2018conventional gas\u2019, and \u2018LNG gas\u2019.\nData collection\nEconomic data and emission data are based on the China's input-output table of the year 2017 and China Energy Statistical \nYearbook(2022). The BAU energy and economic development data are mainly from China statistical yearbook(2022), IEA(2022), and \nUnited Nation(2022). The policy scenario setting data are from the policy documents of China\u2019s State Council and relevant literature.\nTiming and spatial scale\nThe actual economic data and emission data is from 2017 to 2022, and the estimated data is from 2023 and 2060 under the \npredictions of relevant research and multiple mitigation policy scenarios.\nData exclusions\nNo data were excluded.\nReproducibility\nSource data are provided with this paper. All data used here are cited in the text or provided in the supplementary data and source \ndata; All computer codes generated during this study are provided with this paper.\nRandomization\nAll sectors for China are involved in this study. So randomization is not relevant here.\nBlinding\nThis study is based on existing data. So blinding is not relevant here.\nDid the study involve field work?\nYes\nNo\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \n\n3\nnature portfolio  |  reporting summary\nMarch 2021\nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology and archaeology\nAnimals and other organisms\nClinical data\nDual use research of concern\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\n\n\n    Scientific Research Findings:", "answer": "We find that interactions between mitigation policies in China would reduce the percentage of scenarios achieving carbon neutrality by 2060 by 84%, and delay the years in which these scenarios are achieved by 5\u20136 years. This decline is driven by how different policies influence each other\u2019s implementation space: complementary policies expand it, while competitive policies constrain it. Among all mitigation polices, the combination of carbon pricing and renewable energy exhibits trade-off effects on both mitigation and economic outcomes. Conversely, the combination of renewable energy and the electrification of end uses demonstrates synergistic effects, benefiting both economic and mitigation impacts. These findings imply that rather than merely increasing policy intensity to achieve carbon neutrality, policymakers should prudently design mitigation portfolios to maximize synergies and minimize trade-offs. Promoting the joint implementation of renewable energy and electrification polices is an effective measure to reduce carbon emissions in China.", "id": 38}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 14 | December 2024 | 1261\u20131267\n1261\nnature climate change\nhttps://doi.org/10.1038/s41558-024-02162-4\nArticle\nAssessing the impacts of fertility and \nretirement policies on China\u2019s carbon \nemissions\nLing Tang\u2009\n\u200a\u20091, Junai Yang\u2009\n\u200a\u20092, Jiali Zheng\u2009\n\u200a\u20093, Xinlu Sun\u2009\n\u200a\u20094, Lu Cheng5, \nKehan He\u2009\n\u200a\u20096, Ling Li\u2009\n\u200a\u20097, Jinkai Li\u2009\n\u200a\u20098\u2009\n, Wenjia Cai\u2009\n\u200a\u20099, Shouyang Wang\u2009\n\u200a\u20091,2,10, \nPaul Drummond\u2009\n\u200a\u200911 & Zhifu Mi\u2009\n\u200a\u20094\u2009\nThe gradual adjustment of fertility and retirement policies in China has \nsocial benefits in terms of coping with population aging. However, the \nenvironmental consequences of these policies remain ambiguous. Here we \ncompile environmentally extended multiregional input\u2013output tables to \nestimate household carbon footprints for different population age groups \nin China. Subsequently, we estimate the age-sex-specific population under \ndifferent fertility policies up to 2060 and assess the impacts of fertility and \nretirement policies on household carbon footprints. We find that Chinese \nyoung people have relatively higher household carbon footprints than their \nolder counterparts due to differences in income by age group. Relaxing \nfertility policies and delaying retirement age are associated with an increase \nin population (and labour supply) and thus increases in household carbon \nfootprints, with the majority of these increases from the fertility side. \nThese results may help policymakers understand interactions among those \nmeasures targeting population aging and climate action.\nMitigating climate change and coping with population aging are  \nboth critical goals for China in achieving sustainable development1,2. \nAs the world\u2019s largest carbon emitter, China aims to have a carbon \nemissions peak before 2030 and achieve carbon neutrality by 20603. \nCurrently, China is turning towards more sustainable development, \nwith the deceleration of China\u2019s annual carbon emissions growth from \n10% (2000\u20132010) to 2% (2010\u20132020)4. However, China remains an \nimportant driver of global carbon emissions due to its large popula-\ntion and growing household consumption over the past 20\u2009years. To \nbetter explore the drivers of carbon emissions, the household carbon \nfootprint (the sum of direct and indirect carbon emissions of house-\nhold consumption along the supply chain) has received increasing \nattention recently1,5,6. In addition, China is one of the most populous \ncountries in the world, with a population that is nearing its peak and \naging rapidly7. In 2020, China\u2019s total fertility rate was only 1.3 births \nper woman, which is far below the replacement level (2.1) needed for \na stable population8. It is projected that China\u2019s population will peak \nat 1.45 billion in 2029 (with a range of 1.42 to 1.48 billion from 2025 to \n2035)2,9, after which contraction is expected. At the same time, China \nis aging rapidly, with the proportion aged 65\u2009years and above doubling \nfrom 7% in 2000 to 14% in 202010.\nChina has implemented a national strategy to address popula-\ntion aging, including relaxing fertility policies and delaying retire-\nment age. In the 1970s, a one-child policy was introduced to curb \nReceived: 2 August 2022\nAccepted: 18 September 2024\nPublished online: 11 October 2024\n Check for updates\n1School of Economics and Management, University of Chinese Academy of Sciences, Beijing, China. 2Academy of Mathematics and Systems Science, \nChinese Academy of Sciences, Beijing, China. 3The School of Management, Xi\u2019an Jiaotong University, Xi\u2019an, China. 4The Bartlett School of Sustainable \nConstruction, University College London, London, UK. 5School of Ecology and Environment, Renmin University of China, Beijing, China. 6Institute for \nClimate and Carbon Neutrality, The University of Hong Kong, Hong Kong SAR, China. 7International School of Economics and Management, Capital \nUniversity of Economics and Business, Beijing, China. 8School of Economics, Beijing Institute of Technology, Beijing, China. 9Department of Earth System \nScience, Tsinghua University, Beijing, China. 10School of Entrepreneurship and Management, ShanghaiTech University, Shanghai, China. 11Institute for \nSustainable Resources, University College London, London, UK. \n\u2009e-mail: lijinkai@sina.com; z.mi@ucl.ac.uk\n\nNature Climate Change | Volume 14 | December 2024 | 1261\u20131267\n1262\nArticle\nhttps://doi.org/10.1038/s41558-024-02162-4\nthose of developed countries, where older people are estimated to \nhave higher carbon footprints5,22. The difference in carbon footprint \ndistribution by age, between developed and developing countries, is \ndue mainly to the difference in wealth and income across age groups. \nIn wealthier developed countries, older people tend to be wealthier \nthan younger people and thus can afford a higher level of consump-\ntion and tend to have higher carbon footprints23. In developing \ncountries (for example, China), young people have higher incomes \nthan older people (by 57%, according to our individual data), associ-\nated with higher consumption and carbon footprints (by 69% and  \n77%, respectively).\nAfter examining the expenditure categories in greater detail, it \nis evident that the top two contributors to the average carbon foot-\nprint of all age groups are consumption related to residence and that \nrelated to transport (Fig. 1b and Supplementary Note 1). There are \nsome meaningful differences across age groups not only absolutely \nbut also proportionally (Supplementary Fig. 1), representing their \ndifferences in lifestyle choices and life stage24. For residence, the \nyoung people have the highest carbon footprint (1.08\u2009tCO2 per capita \nin 2017) and contribute to the largest share of total residence-related \nfootprint (46%), the majority of which is from renting or purchasing \na house18 and using electronic devices25; the older people have the \nhighest proportional share of residence-related carbon footprint (41%) \nas they might be accustomed to using traditional energy-intensive \ndevices for heating and cooking (such as Kang and stove)26 and spend \nmuch of their time at home (and thus have large household energy \nconsumption)27. For transport, the young people\u2019s transport-related \nfootprint is the highest both absolutely (accounting for 50% of the \ntotal transport-related footprint by all groups) and proportionally \n(accounting for 25% of their own total footprint), largely from com-\nmuting to work28 and from a few big trips each year (for example, from \ntheir workplace to their hometown)29. Moreover, the absolute and pro-\nportional per capita carbon footprints related to clothing, goods and \ntransport have decreased gradually with age; however, health-related \ncarbon footprint has increased with age, as has education-related \ncarbon footprint until the individual is in their 30\u2009s to 40\u2009s, after which \nit decreases (Fig. 1b). We further explore how unevenly the per capita \ncarbon footprint is distributed among different age groups using the \nTheil index. The higher the index value is, the greater the inequality \nin terms of the distribution between age groups. In 2017, the Theil \nindex for clothing, transport and education-related carbon footprint \nwas the highest, at 0.06, three times the average value of expenditure \ncategories (0.02) (Supplementary Table 3).\nThe preceding patterns also generally hold in all Chinese prov-\ninces: young people have a relatively higher per capita carbon footprint \nthan that of older people (by 1.21 to 2.93 times), and consumption pat-\nterns vary over the life course (for example, the young age group has \nlarger clothing-, goods- and transport-related carbon footprints, and \nthe middle-aged group generates most of the education-related carbon \nfootprint). Regarding the Theil index, eastern provinces (for example, \nGuangdong and Hainan), central provinces (for example, Anhui and \nHunan) and southwestern provinces (for example, Guangxi and Chong-\nqing) have higher values than northwest provinces (for example, Inner \nMongolia and Gansu) (Fig. 1c and Supplementary Table 5).\nBetween 2012 and 2017, China\u2019s average per capita carbon foot-\nprint increased by 17%, from 2.00\u2009tCO2 in 2012 to 2.34\u2009tCO2 in 2017. In \nparticular, young people experienced larger increases (30%) than did \nmiddle-aged (12%) and older (8%) people during this period, meaning \nthat the difference in carbon footprints across age groups grew (with \nan increase in the Theil index from 0.01 to 0.02; Supplementary Tables 3 \nand 4). At the provincial level, the average per capita carbon footprint \nand Theil index increased in most provinces, mainly because of the \ngrowing carbon footprint of the young generation that ranges from a \n10% increase (Yunnan) to a 211% increase (Ningxia) from 2012 to 2017 \n(Supplementary Tables 5 and 6).\npopulation growth and alleviate severe poverty in China2. Following \nthe introduction of this policy, the fertility rate decreased\u2014the total \nfertility rate declined sharply from approximately 5.8 in 1970 to 2.8 \nin 1979 and was thought to be approximately 1.6 in 2010\u2014resulting \nin a rapidly aging population7,11. In October 2015, China\u2019s one-child \npolicy was replaced with a two-child policy to counter this trend12. \nThe two-child policy has had a positive effect on the birth rate: more \nthan 10 million babies were born as a second child in China dur-\ning 2013\u20132017, and the proportion of newborns who were second \nchildren in new births increased from 30% in 2013 to 50% in 20178. \nHowever, a continuous fall in the number of women of childbearing \nage and a gradual decline in the effect of the two-child policy resulted \nin a drop in the number of new births during the period 2017\u201320208. \nIn May 2021, in an attempt to tackle demographic challenges, China \nfurther relaxed its fertility policy with a three-child policy, allowing \nall couples to have up to three children13. In addition, many sup-\nportive measures have been implemented to address housing and \neducational costs, aiming to ease the financial burden of raising \nchildren14. However, the retirement age in China, 60\u2009years for men \nand an average of 52.5\u2009years for women (50\u2009years for women workers \nand 55\u2009years for women cadres)2, is among the lowest in the world: \nthe official retirement age for most developed countries is 65\u2009years \nor even higher15,16. According to the Outline of the 14th Five-Year Plan \n(2021\u20132025), China called for the extension of the statutory retire-\nment age in a gradual, flexible and differentiated manner to reduce \nthe negative impacts of population aging17. Changing fertility and \nretirement policies are likely to have great effects on the population \nage structure and potentially influence household consumption and \ncarbon footprints.\nMany studies have estimated the effect of population aging on \ncarbon emissions in China, finding that population aging may reduce18, \nincrease19,20 or have nonlinear effects on carbon emissions21. However, \nno study has assessed the impacts of policies that address population \naging\u2014including fertility (particularly the three-child policy) and \nretirement policies\u2014on carbon emissions or household carbon foot-\nprints. Thus, we aim to address this gap in the literature. In this Article, \nwe first investigate age-based household carbon footprints in China \nand its provinces by compiling a global multiregional input\u2013output \n(MRIO) table and employing a large-scale household survey. We fur-\nther estimate the age distribution of the population in China and its \nprovinces up to 2060 by using a cohort-component method and then \nassess the impacts of fertility and retirement policies on household \ncarbon footprints.\nAge-based household carbon footprint\nThe total and per capita household carbon footprint varies greatly \nacross China\u2019s provinces. Eastern provinces (which have large popu-\nlations) tend to have higher total carbon footprints (particularly in \nShandong, Guangdong and Jiangsu; Fig. 1a and Supplementary Table 1). \nNorthwestern provinces (with high carbon intensity) and eastern \nprovinces (with high household consumption) tend to have higher \nper capita carbon footprints. For example, Ningxia (a northwestern \nprovince) had the highest per capita carbon footprint (6.68\u2009tons CO2 \n(tCO2) in 2017, six times that of Sichuan (a southwestern province) \nat 1.05\u2009tCO2 (Fig. 1c). Taking China as a whole, its per capita carbon \nfootprint is much lower than that of developed countries. Specifically, \nthe Chinese per capita carbon footprint was 2.34\u2009tCO2 in 2017, approxi-\nmately one-sixth of that in the United States (13.37\u2009tCO2) and one-third \nof that in Japan (6.29\u2009tCO2) and the United Kingdom (6.03\u2009tCO2), but \nsimilar to that in Mexico (2.31\u2009tCO2) and almost three times that in \nIndia (0.78\u2009tCO2) (Fig. 1b).\nIn China, carbon footprint is inversely correlated with age. Young \nChinese people (<30\u2009years) have relatively higher household car-\nbon footprints than those of middle-aged (30\u201359\u2009years) and older \n(\u226560\u2009years) groups. The observed results are quite different from \n\nNature Climate Change | Volume 14 | December 2024 | 1261\u20131267\n1263\nArticle\nhttps://doi.org/10.1038/s41558-024-02162-4\n3.64\u201319.24\n19.25\u2013101.19\n101.20\u2013194.40\n194.41\u2013345.57\nN\n0\n500 km\nTotal carbon footprint\n(MtCO2)\n<30\n30\u201339 40\u201349 50\u201359\n\u226560\nAverage\nAge group\nProvince\nRegion\nUnited \nStates\nJapan United \nKingdom\nRussia\nEU27\nMexico Brazil\nIndia\n0\n2\n4\n6\n12\n14\n16\nOthers\nHousehold direct\nGoods\nTransport\nEducation\nFood\nClothing\nResidence\nChina\nHealth\n2.92\n2.58\n1.97\n1.87\n1.65\n2.34\n13.37\n6.29\n6.03\n5.55\n5.46\n2.31\n1.68\n0.78\na\nb\nc\nXinjiang\nBeijing\nInner Mongolia\nShanghai\nNingxia\nTianjin\nShanxi\nLiaoning\nShandong\nQinghai\nHeilongjiang\nHebei\nGuangdong\nZhejiang\nJiangsu\nGuizhou\nJilin\nShaanxi\nGansu\nHainan\nChongqing\nHenan\nAnhui\nFujian\nHubei\nHunan\nJiangxi\nGuangxi\nYunnan\nTibet\nSichuan\n0\n2\n4\n6\n8\n10\nTheil index\n<30\n30\u201339\n40\u201349\n50\u201359\n\u226560\n0\n0.03\n0.06\n0.09\n0.12\n0.15\nAverage\nTheil index\nPer capita carbon footprint (tCO2)\nPer capita carbon \nfootprint (tCO2)\nPer capita carbon footprint (tCO2)\nNo data\n0.01\u20131.85\n1.86\u20132.56\n2.57\u20134.44\n4.45\u20136.68\nFig. 1 | Household carbon footprint in 2017. a, Total and per capita household \ncarbon footprints for 31 of China\u2019s provinces. The cut-out of islands is the South \nChina Sea Islands. The data for the basemap were derived from the Resource  \nand Environment Data Cloud Platform (https://www.resdc.cn/Default.aspx). \nb, Per capita household carbon footprints for eight expenditure categories for \nChina\u2019s different age groups and for international comparisons. We calculate  \nthe per capita household carbon footprints of the United States, Japan,  \nUnited Kingdom and other countries using data from the EXIOBASE database \n(https://www.exiobase.eu/index.php)36. c, Per capita household carbon \nfootprints and the Theil index for 31 of China\u2019s provinces. Provinces are sorted \naccording to their average per capita carbon footprint.\n\nNature Climate Change | Volume 14 | December 2024 | 1261\u20131267\n1264\nArticle\nhttps://doi.org/10.1038/s41558-024-02162-4\nFertility and retirement policies\nTo evaluate the impacts of fertility and retirement policies on China\u2019s \nhousehold carbon footprints, we first estimate the population of China \nand its 31 provinces up to 2060 by age (0\u2013100+) and sex (male and \nfemale) under different fertility policies: previous two-child policy, \nthe latest three-child policy and the assumed \u2018replacement-level\u2019 pol-\nicy (with fertility rate reaching the replacement level of 2.1 (ref. 30)). \nThen we explore the potential effect of these fertility policies and their \ncombination with retirement delay policies on the household carbon \nfootprints (Supplementary Data 1\u20136). We use retirement age as the \nthreshold to classify older people and assume that such a retirement \ndelay policy affects only the population age structure31.\nWe find that the preceding two kinds of policies will both pose \na challenge to carbon emissions mitigation. As for fertility policies, \nthey affect the population mainly in terms of size and structure, \nand thus affect the carbon footprints. In specific, our results show \nthat the Chinese population will peak in 2023 (1.41 billion), 2030 \n(1.41 billion) and 2040 (1.44 billion) under the two-child, three-child \nand replacement-level policies, respectively (Fig. 2a). From 2017 \nto 2060, the total population will decrease from 1.40 billion to 1.15 \nbillion (two-child policy), 1.30 billion (three-child policy) and 1.39 \n(replacement-level policy), which means the population differences \nare 12\u201320% under different policies (Fig. 2a). The mean population age \nof a person will increase from 38\u2009years to 51\u2009years (two-child policy), \n47\u2009years (three-child policy) and 45\u2009years (replacement-level policy), \nthus, the proportion of older people will increase from 17% to 42% \n(two-child policy), 37% (three-child policy) and 35% (replacement- \nlevel policy) (Fig. 2b\u2013e). Due to relaxing fertility policies, there is an \n8\u201312% increase in per capita footprints (the blue, yellow and red solid \ncurves in Fig. 3 (China)), and the total footprints in China are likely to \nbe 21\u201335% higher.\nThe preceding effects also hold at the provincial level, but the \nextent of the impact varies (Supplementary Note 2). The provinces \nwith a higher Theil index are more sensitive to changes in fertility \n2020\n2025\n2030\n2035\n2040\n2045\n2050\n2055\n2060\n0.8\n1.0\n1.2\n1.4\n1.6\nTwo-child policy\nThree-child policy\nReplacement-level policy\nPopulation (billion)\n1.41\n1.41\n1.44\na\n5\u20139\n15\u201319\n25\u201329\n35\u201339\n45\u201349\n55\u201359\n65\u201369\n75\u201379\n85\u201389\n95+\n0\n0.01\n0.02\n0.03\n0.04\n0.05\nAge group\nAge group\nPercentage of population\n0.01\n0.02 0.03 0.04 0.05\n5\u20139\n15\u201319\n25\u201329\n35\u201339\n45\u201349\n55\u201359\n65\u201369\n75\u201379\n85\u201389\n95+\n0\n0.01\n0.02\n0.03\n0.04\n0.05\nPercentage of population\n0.01\n0.02 0.03 0.04 0.05\n5\u20139\n15\u201319\n25\u201329\n35\u201339\n45\u201349\n55\u201359\n65\u201369\n75\u201379\n85\u201389\n95+\n0.01\n0.02\n0.03\n0.04\n0.05\nAge group\nAge group\nYear\n5\u20139\n15\u201319\n25\u201329\n35\u201339\n45\u201349\n55\u201359\n65\u201369\n75\u201379\n85\u201389\n95+\n0.01\n0.02\n0.03\n0.04\n0.05\n0\n0.01\n0.02 0.03 0.04 0.05\nb\nc\ne\nd\n0\n0.01\n0.02 0.03 0.04 0.05\nMale\nMale median\nMale mean\nFemale median\nFemale mean\nFemale\nFig. 2 | Population changes and population age structure in China under \ndifferent fertility policies. a, Population changes at the national level from 2017 \nto 2060 under the two-child policy, three-child policy and replacement-level \npolicy. b, Population pyramid for males and females by age in 2017 (two-child \npolicy). c\u2013e, Population pyramids for males and females in 2060 under the  \ntwo-child policy (c), three-child policy (d) and replacement-level policy (e).\n\nNature Climate Change | Volume 14 | December 2024 | 1261\u20131267\n1265\nArticle\nhttps://doi.org/10.1038/s41558-024-02162-4\npolicies (in terms of larger changes in per capita carbon footprints; \nSupplementary Fig. 2). For example, in Inner Mongolia, which has \nthe highest Theil index in 2060, changing fertility policies are pro-\njected to increase its average per capita carbon footprint by 18\u201328% \n(Fig. 3, Inner Mongolia). By comparison, in Guizhou, which has the \nlowest Theil index in 2060, changing fertility policies are projected \nto increase its average per capita carbon footprint by only 4\u20135% \n(Fig. 3, Guizhou).\nFertility policies in combination with retirement delay tend to \nfurther increase the carbon footprint in China. Notably, most of the car-\nbon footprint increase comes from relaxing fertility policies (increas-\ning total (per capita) carbon footprint by 21\u201335% (8\u201312%) for 2060), \n0\n0.4\n0.8\n1.2\n1.6\nFootprint change\nFootprint change\nFootprint change\nFootprint change\nFootprint change\nFootprint change\nFootprint change\nFootprint change\nChina\nGuizhou\nBeijing\nHeilongjiang\n1.0\n1.1\n1.2\n1.3\nImpact of policy\nImpact of policy\nImpact of policy\nImpact of policy\nImpact of policy\nImpact of policy\nImpact of policy\nImpact of policy\n0\n0.4\n0.8\n1.2\n1.6\nYunnan\nTianjin\nShanghai\nZhejiang\n1.0\n1.1\n1.2\n1.3\nSichuan\nShanxi\nJiangxi\n1.0\n1.1\n1.2\n1.3\n0\n0.4\n0.8\n1.2\n1.6\nQinghai\nLiaoning\nChongqing\nGansu\n1.0\n1.1\n1.2\n1.3\n0\n0.4\n0.8\n1.2\n1.6\nNingxia\nXinjiang\nHebei\nHubei\n1.0\n1.1\n1.2\n1.3\n0\n0.4\n0.8\n1.2\n1.6\nShaanxi\nJiangsu\nHunan\n1.0\n1.1\n1.2\n1.3\n0\n0.4\n0.8\n1.2\n1.6\nGuangdong\nGuangxi\nJilin\n1.0\n1.1\n1.2\n1.3\n2020\n2030\n2040\n2050\n2060\n0\n0.4\n0.8\n1.2\n1.6\nFujian\nYear\nYear\nYear\nYear\n2020\n2030\n2040\n2050\n2060\nHainan\n2020\n2030\n2040\n2050\n2060\nShandong\n2020\n2030\n2040\n2050\n2060\nInner Mongolia\n1.0\n1.1\n1.2\n1.3\n0\n0.4\n0.8\n1.2\n1.6\nTibet\nOlder people\nMiddle-aged people\nTwo-child policy\nThree-child policy\nReplacement-level policy\nReplacement-level and retirement delay policies\nThree-child and retirement delay policies\nTwo-child and retirement delay policies\nYoung people\nHenan\nAnhui\nFig. 3 | The impacts of fertility and retirement policies on carbon footprints \nin China and its provinces. Change in total household carbon footprint under \nthe two-child policy compared with the 2017 level (left y axis); change in per \ncapita household carbon footprint under different policies compared with the \ntwo-child policy (right y axis). Provinces are sorted according to the value of the \nTheil index in 2060, from the lowest value of the Theil index, in Guizhou, to the \nhighest value, in Inner Mongolia.\n\nNature Climate Change | Volume 14 | December 2024 | 1261\u20131267\n1266\nArticle\nhttps://doi.org/10.1038/s41558-024-02162-4\nwhile delaying retirement policy has far smaller impacts (by only 2\u20133% \n(2\u20133%)). The fertility policies in combination with retirement delay are \nprojected to have greater impacts on those provinces with a higher \nTheil index, which is similar to the impacts of fertility policy alone (Sup-\nplementary Fig. 3). Moreover, when focusing only on the impact of the \nretirement delay policy, we found that the impact tends to be greater \nin provinces with large discrepancies in per capita carbon footprints \nbetween middle-aged and older people (Supplementary Fig. 4). For \nexample, in Inner Mongolia, which has the highest discrepancies in per \ncapita carbon footprints between middle-aged and older people, the \nretirement delay policy is projected to increase its average per capita \ncarbon footprint by approximately 5% (Fig. 3, Inner Mongolia). By \ncomparison, in Yunnan, which has the lowest discrepancies in per capita \ncarbon footprints between these two age groups, the retirement delay \npolicy is projected to increase its average per capita carbon footprint \nby less than 0.10% (Fig. 3, Yunnan).\nDiscussion\nOur results show that Chinese young people have relatively higher per \ncapita household carbon footprints compared with older people. The \nbig driver behind this headline result might be differences in income, \nwhich leads to differences in household consumption and then car-\nbon footprints (Supplementary Note 3); The results differ from those \nof existing research on developed countries, which have concluded \nthat older people tend to have higher per capita carbon footprints \ncompared with their younger counterparts. Such a distinctive pattern \nis due mainly to the difference in income and consumption of China\u2019s \nolder people from other developed countries23.\nOur analysis highlights residence and transport as the two largest \ncontributors to carbon footprint, and there is variability in them across \nage groups. Notably, the assumptions about how these two factors \nmight evolve as the population ages will have a meaningful impact \non our projection results, and we conducted an uncertainty analysis \nassuming that both residence- and transport-related carbon foot-\nprints will still be a feature of today\u2019s young group as they age under \ndifferent assumptions (follow Chinese forebears\u2019 or Western peers\u2019 \npatterns; Supplementary Note 4). In particular, the residence-related \ncarbon footprint from the young group will remain high as they move \ninto the next age cohort if following Chinese older people\u2019s specific \npatterns (for example, spending more time at home27), or will even \ngrow if following Western peers (with more electronics and devices \nplugged in at the residence32). In addition to these proportional dif-\nferences, our overall results for household carbon footprints and \nthe impacts of fertility and retirement policies do not change much  \n(Supplementary Figs. 14\u201316).\nAs for policies, our result shows that relaxing fertility policies \nand delaying retirement age will boost the population (and labour \nsupply) and then lead to increases in total and per capita household \ncarbon footprints, most of which come from the fertility side. We do \nnot interpret the result to imply that such policies should be avoided to \nreduce environmental pressure33. Rather, our result provides evidence \nof interactions between the policies targeting population aging and \nclimate change, highlighting the importance of synergizing these two \ntypes of policies. Although fertility and retirement policies may pose a \nchallenge to China\u2019s carbon emissions mitigation, these policies (par-\nticularly those for retirement delay) can lower the dependency ratio \nand thus improve the demographic dividend (Supplementary Note 2).\nIn addition, we find that the provinces with large discrepancies \nin carbon footprints across age groups are more sensitive to changes \nin fertility and retirement policies. This result therefore highlights \nthe potential of emissions mitigation through reducing the discrep-\nancy in carbon footprints across age groups. Although consumption \npatterns and lifestyles are different across age groups due to their \nvarious requirements over the life course, the discrepancy in carbon \nfootprints between age groups can be narrowed by reducing income \nand consumption inequality and encouraging greener consumption. \nSpecifically, we suggest that the greatest potential leverage from life-\nstyle changes will result from the targeting of young people by promot-\ning green consumption (such as adopting public transportation such \nas buses, subways and shared bikes as well as purchasing high-quality \nand long-lasting goods6).\nChina\u2019s Nationally Determined Contributions acknowledges \nthe difficulty of achieving carbon neutrality by 2060. Therefore, it \nis worth exploring what might happen if this target is not reached. \nAccording to related research and official plans34,35, we consider a set \nof policy scenarios assuming that China achieves a 60%, 70%, 80%, 90% \nor even 100% reduction in carbon intensity (Supplementary Note 5). \nOur results show that there are fewer emissions in the future with a \nmore aggressive target, and achieving 60%, 70% and 80% reductions \nfrom 2017 to 2060 (instead of the 90% target) is estimated to nearly \nquadruple, triple and double, respectively, the overall emissions \nin all scenarios (Supplementary Fig. 17). Furthermore, there is an \ninteraction between the policy proposals and emissions targets\u2014the \neffects of fertility and retirement policies will somewhat aggravate \nthe difficulty of achieving carbon neutrality by 2060, and such policy \neffects will increase if the official target is not reached. Nevertheless, \nour major findings regarding structural patterns of carbon footprints \n(across age groups) and changing trends (due to different fertility \nand retirement policies) do not change very much (Supplementary \nFig. 17). This verifies the accuracy of our estimates, which might rely \nlittle on this official target.\nOverall, this study provides evidence of interactions between \nclimate actions and demographic policies in China. We find that relax-\ning fertility policies and delaying retirement age are associated with \nan increase in total and per capita household carbon footprints. Our \nresults add to the literature on climate change and population, which \nhas typically evaluated the effect of demographic structure on emis-\nsions without considering the independent effect of the population \npolicy (especially in China) that contributes to bringing about the \nchange in demographic structure in the first place. Our results also \noffer insights for developing countries undergoing economic and \ndemographic transformation for more sustainable development.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-024-02162-4.\nReferences\n1.\t\nMi, Z. et al. Economic development and converging household \ncarbon footprints in China. Nat. Sustain. 3, 529\u2013537 (2020).\n2.\t\nZeng, Y. & Hesketh, T. The effects of China\u2019s universal two-child \npolicy. Lancet 388, 1930\u20131938 (2016).\n3.\t\nXi Jinping\u2019s speech at the General Debate of the Seventy- \nFifth United Nations General Assembly. Xinhua News Agency \n(22 September 2020); http://www.gov.cn/xinwen/2020-09/22/\ncontent_5546169.htm\n4.\t\nZheng, J. et al. The slowdown in China\u2019s carbon emissions \ngrowth in the new phase of economic development. One Earth 1, \n240\u2013253 (2019).\n5.\t\nChristis, M., Breemersch, K., Vercalsteren, A. & Dils, E. A \ndetailed household carbon footprint analysis using expenditure \naccounts\u2014case of Flanders (Belgium). J. Clean. Prod. 228, \n1167\u20131175 (2019).\n6.\t\nWiedenhofer, D. et al. Unequal household carbon footprints in \nChina. Nat. Clim. Change 7, 75\u201380 (2017).\n7.\t\nPeng, X. China\u2019s demographic history and future challenges. \nScience 333, 581\u2013587 (2011).\n\nNature Climate Change | Volume 14 | December 2024 | 1261\u20131267\n1267\nArticle\nhttps://doi.org/10.1038/s41558-024-02162-4\n8.\t\nOffice of the Leading Group of the State Council for the Seventh \nNational Population Census. Major Figures on 2020 Population \nCensus of China (China Statistics Press, 2021).\n9.\t\nHuang, J. et al. Effect of fertility policy changes on the population \nstructure and economy of China: from the perspective of the \nshared socioeconomic pathways. Earths Future 7, 250\u2013265 (2019).\n10.\t National Bureau of Statistics. China Statistical Yearbook 2021 \n(China Statistics Press, 2021).\n11.\t\nChen, Y. et al. Provincial and gridded population projection for \nChina under shared socioeconomic pathways from 2010 to 2100. \nSci. Data 7, 83 (2020).\n12.\t China drops one-child policy and allows all couples to have two. \nXinhua News Agency (29 October 2015); http://english.www.gov.\ncn/news/top_news/2015/10/29/content_281475222969126.htm\n13.\t The Chinese Communist Party Politburo meeting chaired by \nPresident Xi Jinping. Xinhuanet (31 May 2021); http://www.\nxinhuanet.com/2021-05/31/c_1127513015.htm\n14.\t China releases decision on third-child policy, supporting \nmeasures. Xinhua News Agency (20 July 2021) http://english.\nwww.gov.cn/policies/latestreleases/202107/20/content_\nWS60f6c308c6d0df57f98dd491.html\n15.\t M\u00e4cken, J. Work stress among older employees in Germany: \neffects on health and retirement age. PLoS ONE 14, e0211487 \n(2019).\n16.\t De Wind, A. et al. Pathways through which health influences early \nretirement: a qualitative study. BMC Public Health 13, 292 (2013).\n17.\t Outline of the 14th Five-Year Plan (2021\u20132025) for National \nEconomic and Social Development and Vision 2035 of the People\u2019s \nRepublic of China (China State Council, 2021); http://www.gov.cn/ \nxinwen/2021-03/13/content_5592681.htm\n18.\t Zhang, Z., Cui, Y. & Zhang, Z. Unequal age-based household \ncarbon footprint in China. Clim. Policy 23, 577\u2013592 (2022).\n19.\t Zhang, C. & Tan, Z. The relationships between population factors \nand China\u2019s carbon emissions: does population aging matter? \nRenew. Sustain. Energy Rev. 65, 1018\u20131025 (2016).\n20.\t Fan, J., Zhou, L., Zhang, Y., Shao, S. & Ma, M. How does \npopulation aging affect household carbon emissions? Evidence \nfrom Chinese urban and rural areas. Energy Econ. 100,  \n105356 (2021).\n21.\t Yang, T. & Wang, Q. The nonlinear effect of population aging on \ncarbon emission\u2014empirical analysis of ten selected provinces in \nChina. Sci. Total Environ. 740, 140057 (2020).\n22.\t Long, Y. et al. Unequal age-based household emission and its \nmonthly variation embodied in energy consumption\u2014a cases \nstudy of Tokyo, Japan. Appl. Energy 247, 350\u2013362 (2019).\n23.\t Zheng, H. et al. Ageing society in developed countries challenges \ncarbon mitigation. Nat. Clim. Change 12, 241\u2013248 (2022).\n24.\t Zagheni, E. The leverage of demographic dynamics on carbon \ndioxide emissions: does age structure matter? Demography 48, \n371\u2013399 (2011).\n25.\t Yu, M., Meng, B. & Li, R. Analysis of China\u2019s urban household \nindirect carbon emissions drivers under the background of \npopulation aging. Struct. Change Econ. Dyn. 60, 114\u2013125 (2022).\n26.\t Li, S., Wang, Q. & Li, R. How aging impacts environmental \nsustainability\u2014insights from the effects of social consumption \nand labor supply. Humanit. Soc. Sci. Commun. 11, 387 (2024).\n27.\t Li, K., Li, H., Wang, Y., Yang, Z. & Liang, S. Household carbon \nfootprints of age groups in China and socioeconomic influencing \nfactors. Sci. Total Environ. 923, 171402 (2024).\n28.\t Huang, Y., Shigetomi, Y., Chapman, A. & Matsumoto, K. \nUncovering household carbon footprint drivers in an aging, \nshrinking society. Energies 12, 3745 (2019).\n29.\t Czepkiewicz, M. et al. Urban structural and socioeconomic \neffects on local, national and international travel patterns and \ngreenhouse gas emissions of young adults. J. Transp. Geogr. 68, \n130\u2013141 (2018).\n30.\t World Fertility Report: 2003 (UN Population Division, 2003); \nhttps://www.un.org/en/development/desa/population/\npublications/pdf/fertility/worldFertilityReport2003.pdf\n31.\t Zeng, Y. & Wang, Z. A policy analysis on challenges and \nopportunities of population/household aging in China.  \nJ. Popul. Ageing 7, 255\u2013281 (2014).\n32.\t Liddle, B. Consumption-driven environmental impact and age \nstructure change in OECD countries: a cointegration-STIRPAT \nanalysis. Demogr. Res. 24, 749\u2013770 (2013).\n33.\t O\u2019Neill, B. C. et al. The effect of education on determinants of \nclimate change risks. Nat. Sustain. 3, 520\u2013528 (2020).\n34.\t Yu, B. et al. Approaching national climate targets in China \nconsidering the challenge of regional inequality. Nat. Commun. \n14, 8342 (2023).\n35.\t Institute of Climate Change and Sustainable Development \nof Tsinghua University et al. China\u2019s Long-Term Low-Carbon \nDevelopment Strategies and Pathways: Comprehensive Report \n(Springer, 2022).\n36.\t Stadler, K. et al. EXIOBASE 3: developing a time series of detailed \nenvironmentally extended multi-regional input\u2013output tables.  \nJ. Ind. Ecol. 22, 502\u2013515 (2018).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this licence, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2024\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02162-4\nMethods\nHousehold carbon footprint and MRIO analysis\nCarbon footprints measure the greenhouse gas emissions generated \nin the value chains connected with the products consumed in the form \nof final demand1. In this study, we consider CO2 emissions and focus \non household consumption in China. Household carbon footprints \ncome from household consumption activities, including direct energy \nuse (direct carbon footprints), for example, during cooking, heating \nand driving, and the consumption of goods and services, which are \nproduced by using energy as intermediate inputs (indirect carbon \nfootprints).\nIndirect household carbon footprints are calculated on the \nbasis of input\u2013output analysis. Wassily Leontief developed the \ntheoretical framework of the input\u2013output analysis in the late \n1930s37. The fundamental linear equation of the MRIO model can be  \nexpressed as:\n\u239b\n\u239c\n\u239c\n\u239c\n\u239c\n\u239d\nx1\nx2\n\u22ee\nxr\n\u239e\n\u239f\n\u239f\n\u239f\n\u239f\n\u23a0\n=\n\u239b\n\u239c\n\u239c\n\u239c\n\u239c\n\u239d\nA1,1 A1,2 \u22efA1,s\nA2,1 A2,2 \u22efA2,s\n\u22ee\n\u22ee\n\u22f1\n\u22ee\nAr,1 Ar,2 \u22efAr,s\n\u239e\n\u239f\n\u239f\n\u239f\n\u239f\n\u23a0\n\u239b\n\u239c\n\u239c\n\u239c\n\u239c\n\u239d\nx1\nx2\n\u22ee\nxr\n\u239e\n\u239f\n\u239f\n\u239f\n\u239f\n\u23a0\n+\n\u239b\n\u239c\n\u239c\n\u239c\n\u239c\n\u239d\n\u2211s \u2211t y1,s\nt\n\u2211s \u2211t y2,s\nt\n\u22ee\n\u2211s \u2211t yr,s\nt\n\u239e\n\u239f\n\u239f\n\u239f\n\u239f\n\u23a0\n,\n(1)\nwhere xr denotes the total output for each sector in province r; Ar,s is \nthe technical coefficient matrix, which reflects the input requirement \nby sector in province r to produce one unit of output of the sector in \nprovince s; and yr,s\nt  is the final demand vector of category t, including \nhousehold consumption (t\u2009=\u20091), government consumption (t\u2009=\u20092), capi-\ntal investment (t\u2009=\u20093) and exports (t\u2009=\u20094). Equation (1) can also be abbre-\nviated as follows:\nx = Ax + y,\n(2)\nwhere x, A and y are the block matrix or vector in equation (1); then we \ncan obtain the following:\nx = (I \u2212A)\n\u22121y,\n(3)\nwhere I is the identity matrix, and (I \u2212A)\n\u22121 is the Leontief inverse matrix.\nIndirect carbon footprints are calculated by introducing carbon \nintensity (carbon emissions per unit of economic output) by sector:\ne = f(I \u2212A)\n\u22121 \u0302y,\n(4)\nwhere f is the carbon intensity vector, and the carbon emissions used \nto produce f are from the China Emission Accounts and Datasets \n(CEADs). Sector-specific indirect carbon footprints es\nt consumed by \nfinal demand t in province s can be calculated as follows:\nes\nt = f(I \u2212A)\n\u22121 \u0302yr,s\nt .\n(5)\nNotably, there are 45 sectors in the CEADs (regarding emissions \ndata) and 42 sectors in the Chinese MRIO tables38. Due to data availabil-\nity, we mapped 45 sectors to 42 sectors to calculate household carbon \nfootprints and then aggregated to eight expenditure categories for \nfurther analyses: food, clothing, residence, goods, transport, educa-\ntion, health and others6. Specifically, goods include household facili-\nties and durables; transport contains transport and communications; \neducation refers to education, culture and entertainment39.\nFor direct household carbon footprints, the emissions data are \nobtained from the CEADs, where energy-related emissions are listed \nseparately. We allocate the emissions from the energy use of coal and \nnatural gas to the direct household carbon footprints of residence, and \noil emissions are for the category of transport1,40.\nFinally, the total carbon footprint from household consumption \ncan be combined as follows:\nces\nl.t = \u2211\nr\ner,s\nl,t + de\ns\nl,t,\n(6)\nwhere ces\nl,t and de\ns\nl,t are the total and direct carbon footprints, respec-\ntively, of expenditure categories l in province r for household con-\nsumption (t\u2009=\u20091); er,s\nl,t  is the indirect household carbon footprint of \nexpenditure categories l in province r caused by household consump-\ntion in province s.\nTracing household carbon footprints to specific age groups\nIn this section, we trace the household carbon footprints to various \nage groups according to their expenditure in terms of consuming \nproducts. The consumption data used in this study are obtained from \na large-scale household survey (China Family Panel Studies (CFPS))41, \nand the age-based population data are from the China Provincial Sta-\ntistical Yearbooks. Notably, consumption data in a household survey \nare usually collected at the household level and need to be allocated to \nper capita consumption for further age-based analysis. In this study, \nwe use the Organisation for Economic Co-operation and Develop-\nment modified equivalence scale to distinguish children from adults \n(including head of household and other adults) in calculating per \ncapita consumption42,43, rather than simply assuming equal weights \non all household members as has typically been done in previous stud-\nies. In particular, the head of household is weighted by 1, each addi-\ntional adult aged 14 years and older is weighted by 0.5, and each child \nbelow 14 years is weighted by 0.3. It is worth exploring the sensitivity \nof our results to weight choice; thus we also run our model with a per \ncapita calculation and weight-adjusted calculation (Supplementary \nNote 6), and the associated results show that using a different weight \nchoice would change the structural pattern of carbon footprints \nacross age groups but would not change the national average or total  \n(Supplementary Fig. 18).\nThe introduction of the CFPS dataset allows for the downscaling \nof household consumption into age cohorts:\nyr,s\nl,q = c s\nl,q \u00d7 ps\nq \u00d7\nyr,s\nl\ny s\nl\n,\n(7)\nwhere yr,s\nl,q represents the household consumption of sector l in province \nr caused by age group q in province s; c s\nl,q represents the per capita \nhousehold consumption, which is calculated by using the Organisation \nfor Economic Co-operation and Development modified equivalence \nscale; ps\nq denotes the population, derived from the China Provincial \nStatistical Yearbooks; yr,s\nl /y s\nl  denotes the proportion of household \nconsumption for each sector that is finally produced in province r and \nconsumed in province s (obtained from the MRIO model). Thus, the \nage-based indirect household carbon footprint can be obtained by the \nfollowing:\ner,s\nl,q = k r,s\nl\n\u00d7 yr,s\nl,q,\n(8)\nwhere er,s\nl,q is the indirect household carbon footprint, and k r,s\nl  denotes \nthe indirect carbon emissions per unit of household consumption:\nk r,s\nl\n=\ner,s\nl\nyr,s\nl\n.\n(9)\nFor households, the direct carbon footprint of province s can be \nsplit into age group q as follows:\nde\ns\nl,q = \u03b8 s\nl,q \u00d7 de\ns\nl ,\n(10)\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02162-4\nwhere de\ns\nl,q denotes the direct carbon footprints from the household \nconsumption for the products of sector l by age group q in province s; \n\u03b8 s\nl,q is the proportion of household consumption by age group q in all \nage groups for sector l in province s:\n\u03b8 s\nl,q =\nc s\nl,q \u00d7 ps\nq\n\u2211q(c s\nl,q \u00d7 ps\nq) .\n(11)\nFinally, the total household carbon footprint by age group can be \ncalculated by combining indirect and direct carbon footprints:\nces\nl,q = \u2211\nr\ner,s\nl,q + de\ns\nl,q,\n(12)\nwhere ces\nl,q represents the province- and sector-specific total household \ncarbon footprints, and the associated per capita carbon footprints can \nbe further calculated by dividing them by the population.\nTheil index\nIn this study, we use the Theil index to measure how unevenly house-\nhold carbon footprints are distributed across age groups44. The Theil \nindex ranges from zero to one, with a higher value indicating greater \ninequality of distribution across age groups. The Theil index is calcu-\nlated by the following:\nT = \u2211\nq\ndq ln (\ndq\nwq\n) ,\n(13)\nwhere dq is the household carbon footprint share of age group q on the \ntotal and wq is a weighting variable (share of population) for age group \nq, which can be calculated as follows:\ndq =\nceq\n\u2211qceq\n,\n(14)\n\u03c9q =\nPq\n\u2211qPq\n.\n(15)\nProjection\nWe conduct projection similarly on the basis of the input\u2013output analy-\nsis, introducing the year index n to the original form of equation (4):\nes,v\nl.q,n = f s\nl,n(I\u2212\u2212\u2212A)\n\u22121 \u0302y s,v\nl,q,n,\n(16)\nwhere es\nl,q,n is the total household carbon footprints for the year n under \nthe policy scenario v, f s\nl,n is the carbon intensity, (((I\u2212\u2212\u2212A)))\n\u2212\u2212\u22121 is the Leontief \ninverse matrix, and y s,v\nl,n,q is the household consumption.\nCarbon intensity. Carbon intensity is modelled following China\u2019s \nLong-Term Low-Carbon Development Strategies and Pathways: Com-\nprehensive Report35, as the planned annual changing rates listed in \nSupplementary Table 12. Using these planned annual changes (as well \nas the linear interpolations around them), we can project future carbon \nintensities on the basis of the 2017 baseline (Supplementary Note 5).\nHousehold consumption. The household consumption is calculated \nby multiplying per capita household consumption with population. We \nmodel per capita household consumption under the assumption that \nyoung people would become much like their forebears in terms of fol-\nlowing the income effect of consumption and concomitant emissions:\nlog( yq,l) = aq,l + bq,l log(mq),\n(17)\nwhere y indicates per capita consumption, and m denotes per capita \nincome, which is projected on the basis of the long-term gross domestic \nproduct (GDP) forecast under the shared socioeconomic pathways \nmiddle-of-the-road scenario45,46. The coefficient b is the income elas-\nticity of consumption (Supplementary Note 3), which is estimated \npositive in all cases on the basis of our individual data and thus suggests \nthat increases in per capita incomes result in greater consumption \n(Supplementary Table 8). We project that the level of consumption \nfrom all age groups will become higher along economic development, \nwhile the structure of expenditure otherwise will not greatly change \n(Supplementary Fig. 8).\nPopulation. We estimate the age-sex-specific population from 2020 \nto 2060 using a cohort-component method\u2014a dominant method of \npopulation projection47. The basic demographic balancing equation is \nas follows:\nP(N ) = P(0) + B(0, N ) \u2212D(0, N ) + G(0, N ),\n(18)\nwhere P(N) is the population at a given time, P(0) is the population at \nthe start of the interval, and B(0, N), D(0, N) and G(0, N) are the num-\nber of live births, number of deaths and net migration, respectively,  \nduring the interval.\nThe cohort-component method of population projection \nextends the preceding demographic balancing equation to estimate \nage-sex-specific populations on the basis of a series of related factors, \nsuch as age-specific fertility rates, sex ratio at birth, age-sex-specific \nmortality rates and net migration (Supplementary Note 7).\nThe impacts of fertility and retirement policies\nIn this study, we explore the potential effects of fertility policy and its \ncombination with retirement policy on the household carbon footprint. \nSpecifically, there are three fertility policies: the two-child policy, the \nthree-child policy and the assumed replacement-level policy. As for \nthe retirement policy, we assume that the retirement age (as well as \nthe threshold for older people in this study) is extended linearly from \nthe Chinese current level (60\u2009years for men and 55\u2009years for women2) \nin 2020 to the age prevailing in developed countries (65\u2009years for both \nmen and women) in 2050, and remains constant afterwards (Supple-\nmentary Table 7)31. The fertility and retirement policies would affect \nthe population and economy, thereby impacting the household con-\nsumption and carbon footprint in equation (16).\nThe impact on population. The fertility and retirement policies affect \nthe population in terms of size and structure, respectively. In particular, \nwe project the population under different fertility policies with differ-\nent fertility rates following the cohort-component method. As for the \nretirement policy, we use retirement age as the threshold to classify \nolder people, such that the retirement policy affects the population \nage structure31.\nThe impact on the economy. Fertility and retirement policies would \nhave impacts on the supply of labour and the potential levels of the \neconomy48:\n\u0394GDPs,v\nn = \u2211\nq\n\u2211\ng\n\u0394labourrates,v\nq,g,n \u00d7 ratios,\n(19)\nwhere \u0394GDP\ns,v\nn  is the changes in GDP under policy v (where GDP gener-\nally reflects the potential level of economy and can be used as the \nproxy for income45,46), \u0394labourrate\ns,v\nq,g,n is the changes in labour force \nparticipation rate for gender g, and ratio\ns is the ratio of labour remu-\nneration to GDP.\nFor fertility policies, we project the population with different \nfertility rates, then estimate the associated labour force participation \nrates by the ratio of working population to total population, and finally \ncompute the changes of labour force participation rate across policies.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02162-4\nFor retirement policies, the change in labour force participation \nrate is the difference between the counterfactual labour force partici-\npation curve (under the new retirement policy) and the actual labour \nforce participation curve (under the current retirement policy). The \nactual curve is estimated by the ratio of working population to total \npopulation, using the data derived from 2020 Chinese Census. For the \ncounterfactual curve, the ordinary least square method is employed \nto quantify the impact of the new retirement age on labour force par-\nticipation rate48:\nlabourrates\nq,g = \u03b20 + \u03b21D(ages\nq,g) + \u03b22ages\nq,g + \u03b5s\nq,g,\n(20)\nwhere labourrate\ns\nq,g represents the labour force participation rate, ages\nq,g \ndenotes the lower bound for the corresponding age group (for exam-\nple, age\u2009=\u200950 for the 50\u201354 age group), and the dummy variable D(ages\nq,g) \nindicates whether the corresponding age group reaches the retirement \nage (D\u2009=\u20091) or not (D\u2009=\u20090). The coefficient \u03b21 measures the extent to which \nthe new retirement age affects the labour force participation rate. In \nbuilding the counterfactual labour force curve, the labour force par-\nticipation rate changes with age, that is, following the trend of actual \nlabour force participation curve below the current retirement age, \ndropping with the magnitude of \u03b21 at the current retirement age, and \nreturning to and holding the actual trend at and above, respectively, \nthe delated new retirement age.\nThe causality of policy to carbon footprint. Supplementary Fig. 6 \nillustrates a causal chain from policy to carbon footprint. In specific, \nthe fertility and retirement policies affect the population (in terms of \nsize and structure, via equation (18)) and the supply of labour (in terms \nof labour force participation rate, via equation (20)), then affect the \neconomy (in terms of GDP and household income, via equation (19)) \nand per capita consumption (via equation (17)), and result in changes \nin household carbon footprints (via equation (16)).\nData description\nIn this study, the China MRIO tables for 2012 and 2017 are compiled \nusing a gravity model based on the single regional input\u2013output tables \nfor Chinese provinces; the detailed information can be found in our \nprevious work38,49. The carbon emissions inventory can be sourced from \nthe CEADs, the household consumption expenditure data between age \ngroups are obtained from the CFPS dataset41, and the age-sex-specific \ndemographic data in China and its 31 provinces are based on the 2020 \nChinese Census. Moreover, the data in the MRIO tables and the house-\nhold expenditure data in the CFPS are all calculated on the basis of \n2012 prices50.\nThe CFPS, developed by Peking University, is a nearly nationwide \nand comprehensive social survey and aims to serve research needs \nregarding various current social phenomena in China41. The main vari-\nables used in this study are as follows: (1) the size of the household (one, \ntwo, three or more persons), (2) the number of children (<14\u2009years) and \nadults (\u226514\u2009years), (3) the geographic classification of the household \n(31 provinces), (4) the expenditures of eight expenditure categories \n(food, clothing, residence, goods, transport, education, health and \nothers) of the household, (5) household income, (6) the head of the \nhousehold (the person who is in charge of the household) and (7) the \nages of all household members.\nUncertainties and limitations\nThere are several uncertainties and limitations in the calculations of \nthis study. First, the economic data (such as those on national accounts \nand interregional trade) and carbon emissions inventories are the \nmain uncertainties of this study. Previous research reported that the \nuncertainty of consumption-based carbon accounts at the national \nlevel is in the range of 5\u201315% and 2\u201316%51. Moreover, the MRIO analysis \nhas also been validated by our previous calculations1,38. Second, the \ninput\u2013output analysis enables us to estimate the carbon footprints \nfor the \u2018average\u2019 products, and it is often criticized for using too many \nsector aggregations52. Due to data availability, we have to use only 42 \ncategories, of which \u2018transport\u2019 is one (with no distinction between \nprivate, public and so on) and so is \u2018food\u2019 (with no detailed information \non diet shift), and if provincial input\u2013output tables for more than 42 \nsectors are available for China, we will improve our method by using \nmore expenditure categories to avoid the bias due to using aggregate \nanalyses. Third, due to data availability, we assume that the footprint \nintensities (carbon emissions per unit of consumption expenditure, \nin equation (8)) on a category are the same across the age groups, \nwhich introduces uncertainty; capturing the differences in footprint \nintensity between age groups would improve our projection. Fourth, \nwe consider only CO2 emissions, and including other greenhouse gas \nemissions is an important issue for the future research if related data \nfor Chinese provinces are available. Fifth, most shared socioeconomic \npathways data are provided in GDP terms (which are net of trade), and \nlong-term scenarios of consumption can be substantially impacted by \nchanges in trade terms; thus, improving the projections (particularly \nfor future income and consumption using appropriate proxies) is an \nimportant direction to improve our work45. Sixth, we keep interme-\ndiate technology constant in the input\u2013output matrix due to data \navailability, and capturing the structural change in the economy over \na long time horizon is an important direction to improve our research53. \nSeventh, the assumption for future carbon intensity has impacts on \nresults (Supplementary Note 5), and introducing defensible long-term \nplans (if available for China) is an important direction to improve our \nmethod54. Finally, there is tremendous uncertainty about both the \nfuture of fertility and retirement in China (Supplementary Note 8), and \nmuch work is needed to adequately set out such uncertainty (as well as \nthe linkages to emissions) for future research.\nData availability\nAll the source data used in this study are publicly available and open \naccess. The carbon emissions inventory in China is sourced from the \nCEADs (http://www.ceads.net/), the household survey is sourced \nfrom CPFS (https://opendata.pku.edu.cn/dataverse/CFPS), and the \nage-sex-specific demographic data in China are sourced from the \n2020 Chinese Census (https://www.stats.gov.cn/sj/pcsj/rkpc/7rp/\nindexch.htm).\nCode availability\nThe codes developed for the analyses and to generate results are avail-\nable from the corresponding author on reasonable request.\nReferences\n37.\t Leontief, W. W. Quantitative input and output relations in the econo\u00ad\nmic systems of the United States. Rev. Econ. Stat. 18, 105\u2013125 (1936).\n38.\t He, K., Mi, Z., Zhang, J., Li, J. & Coffman, D. The polarizing trend of \nregional CO2 emissions in China and its implications. Environ. Sci. \nTechnol. 57, 4406\u20134414 (2023).\n39.\t Wang, Q. et al. Distributional impact of carbon pricing in Chinese \nprovinces. Energy Econ. 81, 327\u2013340 (2019).\n40.\t Zhao, H. et al. Inequality of household consumption and air \npollution-related deaths in China. Nat. Commun. 10, 4337 (2019).\n41.\t Xie, Y. & Hu, J. An introduction to the China Family Panel Studies \n(CFPS). Chin. Sociol. Rev. 47, 3\u201329 (2014).\n42.\t What Are Equivalence Scales? (OECD, 2011).\n43.\t Dudel, C., Garbuszus, J. M. & Schmied, J. Assessing differences  \nin household needs: a comparison of approaches for the \nestimation of equivalence scales using German expenditure data. \nEmpir. Econ. 60, 1629\u20131659 (2021).\n44.\t Bianco, V., Proskuryakova, L. & Starodubtseva, A. Energy \ninequality in the Eurasian Economic Union. Renew. Sust.  \nEnergy Rev. 146, 111155 (2021).\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02162-4\n45.\t Oswald, Y., Owen, A. & Steinberger, J. K. Large inequality in \ninternational and intranational energy footprints between \nincome groups and across consumption categories. Nat. Energy \n5, 231\u2013239 (2020).\n46.\t Yu, B., Wei, Y., Kei, G. & Matsuoka, Y. Future scenarios for energy \nconsumption and carbon emissions due to demographic \ntransitions in Chinese households. Nat. Energy 3, 109\u2013118 (2018).\n47.\t Raftery, A. & \u0160ev\u010d\u00edkov\u00e1, H. Probabilistic population forecasting: \nshort to very long-term. Int. J. Forecast. 39, 73\u201397 (2023).\n48.\t Barrell, R., Kirby, S. & Orazgani, A. The Macroeconomic Impact \nFrom Extending Working Lives (WP95) (Department for Work and \nPensions, 2011).\n49.\t Mi, Z. et al. Chinese CO2 emission flows have reversed since the \nglobal financial crisis. Nat. Commun. 8, 1712 (2017).\n50.\t Liu, Q. & Peng, Z. China\u2019s Input\u2013Output Tables in Comparable \nPrices 1992\u20132005. (China Statistics Press, 2010).\n51.\t Rodrigues, J., Moran, D., Wood, R. & Behrens, P. Uncertainty of \nconsumption-based carbon accounts. Environ. Sci. Technol. 52, \n7577\u20137586 (2018).\n52.\t Guan, Y. et al. Burden of the global energy price crisis on \nhouseholds. Nat. Energy 8, 304\u2013316 (2023).\n53.\t Sun, Y. et al. Global supply chains amplify economic costs of \nfuture extreme heat risk. Nature 627, 797\u2013804 (2024).\n54.\t Shan, Y. et al. Impacts of COVID-19 and fiscal stimuli on global \nemissions and the Paris Agreement. Nat. Clim. Change 11, \n200\u2013206 (2021).\nAcknowledgements\nThis work was supported by grants from the National Natural  \nScience Foundation of China (71971007 to L.T., 72374144 to  \nL.L. and 71988101 to S.W.), the Beijing Natural Science Foundation \n(JQ21033 to L.T.), the National Social Science Fund of China Key \nResearch Project (23VRC063 to L.T.), the Postdoctoral Fellowship \nProgram of CPSF (GZC20241864 to J.Y.) and China Scholarship \nCouncil (201906020094 to J.Y.).\nAuthor contributions\nL.T., J.L. and Z.M. designed the study. J.Y., X.S. and L.C. collected \ndata and performed calculations. L.T., J.Y., J.Z. and L.L. prepared \nthe manuscript. All authors (L.T., J.Y., J.Z., X.S., L.C., K.H., L.L., J.L., \nW.C., S.W., P.D. and Z.M.) participated in performing the analysis and \ncontributed to writing the manuscript. L.T. and Z.M. coordinated and \nsupervised the project.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s41558-024-02162-4.\nCorrespondence and requests for materials should be addressed to \nJinkai Li or Zhifu Mi.\nPeer review information Nature Climate Change thanks Jared Starr \nand the other, anonymous, reviewer(s) for their contribution to the \npeer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\n\n    Scientific Research Findings:", "answer": "We find that younger people in China tend to have higher household carbon footprints due to greater income and consumption, a pattern that contrasts with developed countries where older people typically have higher carbon footprints. Relaxing fertility policies and delaying retirement age are projected to increase household carbon footprints in China, primarily through boosting the population and labour. These impacts of the policy changes seem to be greater in regions where disparities in income and consumption among different age groups are larger. These findings do not imply that such policies targeting population ageing should be avoided to alleviate environmental pressure. Instead, they provide evidence of the interactions between policies targeting population ageing and climate change, and highlight the importance of synergizing policies with different targets.", "id": 39}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 14 | September 2024 | 936\u2013942\n936\nnature climate change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nRemoving development incentives in risky \nareas promotes climate adaptation\nHannah Druckenmiller\u2009\n\u200a\u20091,2,3,6\u2009\n, Yanjun (Penny) Liao\u2009\n\u200a\u20092,6, Sophie Pesek\u2009\n\u200a\u20094, \nMargaret Walls\u2009\n\u200a\u20092 & Shan Zhang5\nAs natural disasters grow in frequency and intensity with climate \nchange, limiting the populations and properties in harm\u2019s way will be \nkey to adaptation. This study evaluates one approach to discouraging \ndevelopment in risky areas\u2014eliminating public incentives for development, \nsuch as infrastructure investments, disaster assistance and federal flood \ninsurance. Using machine learning and matching techniques, we examine \nthe Coastal Barrier Resources System (CBRS), a set of lands where these \nfederal incentives have been removed. We find that the policy leads to lower \ndevelopment densities inside designated areas, increases development in \nneighbouring areas, reduces flood damages and alters local demographics. \nOur results suggest that the CBRS generates substantial savings for the \nfederal government by reducing flood claims in the National Flood Insurance \nProgram, while increasing the property tax base in coastal counties.\nLimiting exposure to climate risks is an important aspect of adapta-\ntion to climate change, particularly in coastal areas. As sea levels rise, \ntidal flooding worsens, and coastal storms become more frequent and \nsevere, limiting the number of people and properties in harm\u2019s way will \nbe key to managing climate damages.\nIn the United States, state and local governments are primarily \nresponsible for land-use and zoning decisions, but federal policies also \nplay important roles in shaping development. Federal investments in \nroads, utilities and other infrastructure lay the groundwork for popu-\nlation growth. Other government programmes, such as the National \nFlood Insurance Program (NFIP), which offers flood insurance at sub-\nsidized rates in most locations, and disaster assistance programmes, \nwhich provide funding for disaster recovery, also affect location deci-\nsions. By partially shifting disaster costs from property owners to the \ngovernment, these programmes reduce the financial disincentives to \ndevelopment in risky areas.\nWhether withdrawing some of these financial incentives would \ncurb development, lower the costs of disasters and help communities \nprepare for climate change is unclear. Many factors affect development \ndecisions and federal incentives are only some of the factors at play. \nEmpirical research into this question is limited because few policy \nexperiments exist where a clear comparison can be made of \u2018treatment\u2019 \nsettings, where incentives for development have been removed, and \n\u2018control\u2019 settings, similar areas where such incentives remain.\nOne such experiment does exist, however. The 1982 Coastal Barrier \nResources Act (CBRA) designated certain areas along the Atlantic and \nGulf coasts as a Coastal Barrier Resources System (CBRS). In these areas, \nmost new federal expenditures and financial assistance are prohibited. \nThis includes the prohibition of federal funding for new infrastructure, \ndisaster relief and flood insurance through the NFIP, among others. The \nlaw intends to transfer the full cost of private development in these areas \nfrom taxpayers to property owners. Besides removing federal incentives, \nCBRS designations do not otherwise prohibit development. Because the \npolicy is both less restrictive and less costly than most land-use regula-\ntions, it may offer an attractive option for managing development in \nareas at high risk of natural disasters and climate change.\nIn this study, we leverage the CBRA to study the long-term eco-\nnomic impacts of removing federal incentives in high-risk areas and \nassess its efficacy as a land conservation and climate adaptation strat-\negy. We not only ask how effective the CBRA has been at discouraging \ndevelopment within designated areas, but also assess the spillover \neffects of the policy on neighbouring areas.\nReceived: 8 September 2023\nAccepted: 5 July 2024\nPublished online: 5 August 2024\n Check for updates\n1Division of Humanities and Social Sciences, California Institute of Technology, Pasadena, CA, USA. 2Resources for the Future, Washington, DC, USA. \n3National Bureau of Economic Research, Cambridge, MA, USA. 4Energy and Resources Group, University of California, Berkeley, CA, USA. 5Department  \nof Economics, Old Dominion University, Norfolk, VA, USA. 6These authors contributed equally: Hannah Druckenmiller, Yanjun (Penny) Liao.  \n\u2009e-mail: hdruck@caltech.edu\n\nNature Climate Change | Volume 14 | September 2024 | 936\u2013942\n937\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nby building footprints), a 41% reduction relative to control areas \n(Fig. 3a). Using an alternative measure of development, we estimate \nthat CBRS units have 0.044, or 83%, fewer buildings per acre than con-\ntrol areas. The magnitude of these effects highlights the central role of \nfederal incentives in shaping development patterns along vulnerable \ncoastlines. Our results pass the standard synthetic control placebo \ntest (Extended Data Fig. 5) and are robust to the use of an alternative \nresearch design (Suppmentary Section C).\nLocal officials may be concerned that by curbing development \nlevels, CBRS designations may adversely affect local property revenues. \nUsing property-level data from Zillow, we do not find evidence of such \nan effect. We find that CBRS designations increase mean sales prices \nand total assessed value per acre within designated areas, although \nthe estimates are not statistically significant. Still, the positive effect \non prices is consistent with prior evidence that, despite increasing \ncosts for landowners, coastal land-use regulations can increase local \nproperty values by restricting supply and generating nature-based \namenities30. A higher assessed value per acre indicates that the lower \ndevelopment densities are offset by higher average values per property. \nWe find no evidence of systematic differences in the characteristics of \nproperties in CBRS and control areas.\nFinally, we examine the impact on local demographics. Because \nCBRS designations transfer the costs of development and disasters \nto state and local governments and property owners, the policy may \nattract homeowners who are more able to bear these costs. Using \ndata from the American Community Survey, we find that CBRS des-\nignations increase median household income by US$15,000, or 19%, \nrelative to control areas, and increase the rent-to-income ratio by 2 \np.p., or 6%. Thus, CBRS areas tend to attract more affluent residents \nand have become less affordable for renters. Indeed, the CBRA raises \nthe barrier to entry in designated areas by, for example, necessitat-\ning self-insurance against floods31,32. Other land-use regulations, \nsuch as the California coastal boundary zone, have led to similar \ndemographic shifts29.\nThe occupancy rates within CBRS areas are 14 p.p. lower than \nthose in control areas, suggesting a greater prevalence of second \nhomes or vacation rentals, consistent with policy resulting in greater \nnatural amenities. Finally, CBRS designations changed the racial \nmakeup of residents. While CBRS areas already contained more white \npeople than average coastal areas before designation, we estimate \nthat the policy increased the share of white people by 3% (2.9 p.p.) \nand reduced the share of Black people by 29% (\u22121.8 p.p.) relative to \nthe control group.\nA recent study employed a spatial regression discontinuity design \nto compare long-term changes in development densities between \nCBRS units and areas just outside their boundaries in five states1 (an \nupdate of the authors\u2019 earlier study that used cross-sectional data and \nan ordinary least squares regression design2). The authors find lower \ndevelopment rates by more than 75% in CBRS areas. However, the \nregression discontinuity approach, which uses neighbouring areas for \ncomparison with the CBRS, is not appropriate if the CBRS boundaries \nare based on development levels (Extended Data Fig. 1) and does not \nallow for the estimation of spillover effects, which our results show \nare important. Our study provides a comprehensive assessment of \nthe policy\u2019s impact on flood damages, property markets and develop-\nment, sociodemographic outcomes, and local government finances.\nWe develop an alternative method for estimating the causal \neffects of CBRS designations. We construct a control group to com-\npare with the CBRS treatment group using a spatial machine learning \ntechnique known as regionalization in combination with a synthetic \ncontrols research design. This procedure allows us to mimic the pro-\ncess by which natural resource planners determined CBRS boundaries \nbased on geomorphic and development features (Extended Data \nFig. 2), thus identifying a set of coastal areas that could have been \nselected for CBRS designation in 1982 but were not. To illustrate our \napproach, we show one example of a CBRS treatment area and con-\nstructed control area, overlaid with parcel-level data on the value of \nproperties (Fig. 1).\nOur analysis addresses several open questions about land-based \nclimate adaptation. First, it has long been suggested that federal incen-\ntives play a role in encouraging development in risky areas3, yet quanti-\ntative research on this question is limited4\u20138. We evaluate whether the \nremoval of these incentives on coastal lands has been a cost-effective \nadaptation strategy. Second, by examining the spillover effects of \nthe policy on surrounding lands, we provide estimates of how natural \ninfrastructure affects coastal property values and flood damages, add-\ning to a large literature on the hazard protection and amenity value of \nnatural lands9\u201319. Third, our analysis sheds new light on how removing \nfederal incentives affects local government finances by providing the \nfirst estimates of how the CBRA affects property tax revenues. Finally, \nwe show that CBRS designations led to demographic changes, adding \nto the literature on residential sorting behaviours in response to envi-\nronmental conditions and land-use regulations20\u201329.\nIdentification of control regions\nWe measure the effects of removing federal financial incentives for \ndevelopment by comparing outcomes in CBRS units and control areas \ntoday. This approach relies on finding control areas that were statisti-\ncally indistinguishable from treatment areas at the time of designation \nin 1982, while measuring divergence in outcomes over the four decades \nsince, based on treatment status (Methods).\nOur spatial machine learning and matching procedure identifies \na set of control regions that closely resemble CBRS treatment areas \nbefore the policy is implemented (Table 1, Extended Data Table 1 and \nExtended Data Fig. 3). Treatment and control areas share common \npre-trends in development densities (both inside the designated \nunits and in spillover areas) and are balanced across a wide array of \npre-treatment observable characteristics, including measures of land \ncover, elevation, infrastructure density, proximity to urban centres, \nincome and flood risk. To illustrate the output of our procedure, we \nplot examples of treatment and control units (Fig. 2) and map their \ngeographic distribution (Extended Data Fig. 4).\nThe effect of removing federal incentives within \nthe CBRS\nWe find that the CBRS has been effective at curbing development within \ndesignated areas (Table 2). CBRS designations see 0.12 percentage \npoints (p.p.) less built-up surface area (the percent of land covered \nTreatment \nControl\nHigher\nLower\nAssessed value\nUnit\nSpillover \narea\nEach dot is\na property\nFig. 1 | Overview of research design. Left: an example CBRS (treatment) area, \nwith its spillover area (2 km buffer). Right: same, but for a control area. Properties \nfrom ZTRAX, which comprises one set of outcomes studied here, are overlaid on \nthe maps as points. The colour of the points corresponds to the total assessed \nvalue of the home.\n\nNature Climate Change | Volume 14 | September 2024 | 936\u2013942\n938\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nHeterogeneous effects on development\nThe efficacy of CBRS designations at detering development may \ndepend on surrounding natural and social systems, including state and \nlocal policies. Here, we explore where CBRS designations are more and \nless effective by estimating unit-specific treatment effects (Methods).\nThere is a wide distribution of individual treatment effects on \nbuilt-up surface area (Fig. 3b). Nearly 80% of the unit-level estimates \nare negative, suggesting that CBRS designations prevent development \nin most cases. However, the range of estimates is large. Comparing the \ncharacteristics of the most and least effective units (Extended Data \nTable 2), we find that more effective units tend to have lower built-up \nareas at the time of designation and be larger in size. Natural resource \nplanners ought to consider these characteristics when making new \nCBRS designations. However, we find no significant differences in \npre-treatment land cover, elevation, distance to coast or proximity to \nurban centres. Additionally, the most effective and least effective units \nare equally likely to be located on barrier islands or capes, suggesting \nCBRS designations on these land forms do not have systematically dif-\nferent effects on development from those on the mainland.\nThe average treatment effect is negative in 10 out of 13 states \n(Extended Data Table 3), with greater than 50% reductions in built-up \narea attributable to the policy in Delaware, Georgia, Maine, North \nCarolina, Rhode Island, South Carolina, Texas and Virginia. We estimate \npositive average treatment effects in Alabama, Florida and New York. \nThese heterogeneous effects could reflect differences in state and local \npolicies, some of which reinforce CBRS designations by withdrawing \nlocal funding in these areas, while others counteract CBRS designations \nby offering increased subsidies to compensate for the withdrawal of \nfederal incentives. While we find no evidence of systematic differences \nin pro-environment and pro-development leanings across effective and \nnon-effective units (Extended Data Table 2), prior case studies find the \nCBRA can be less effective in areas with very high development pres-\nsure33 and powerful growth coalitions34.\nThe spillover effects of removing subsidies on \nnearby areas\nWe next examine whether removing federal subsidies within CBRS \ndesignations has an impact on nearby, but untreated, areas. More \nopen space and natural lands in CBRS units might create benefits such \nas amenity values and flood protective services. These benefits, in \nturn, might encourage development and demographic changes in \nnearby areas.\nWe estimate the impact of CBRS designation on nearby areas as a \nfunction of distance to the unit boundary (Fig. 4). We find that removing \nfederal subsidies causes denser development, higher property sales \nprices and higher assessed values per acre in neighbouring areas. The \neffect on development densities is increasing in distance from the unit. \nWithin 1,000 m of CBRS units, we estimate an additional 0.03, or 4% \nmore, buildings per acre. Between 1,000 m and 2,000 m, the effect size \nincreases to 0.15 (+20%) and 0.19 (+47%) buildings per acre. The larger \nspillover effects at greater distances can be explained by the CBRS \nwithdrawing infrastructure investment inside treated units: as trans-\nportation and utilities are associated with network effects, we would \nexpect smaller increases in development closer to CBRS boundaries \ndue to limited supply of essential infrastructure.\nIn contrast, average sales prices and assessed values per acre are \nhighest closer to the CBRS units and decline with distance. Properties in \nthe 0\u2013500 m band, for example, sell at a US$377,000 premium, or 77% \nmore than in control areas. Notably, the lower development densities \nin built structures within CBRS units are more than offset by increased \ndevelopment in neighbouring areas, indicating that the price increases \nTable 1 | Covariate balance across CBRS and synthetic \ncontrol\nCBRS\nControl\nSMD\n(1)\n(2)\n(3)\nDesignated area\nWetland share\n0.36\n0.35\n0.01\nBarren share\n0.23\n0.22\n0.03\nTree cover share\n0.04\n0.04\n\u22120.01\nElevation (m)\n1.60\n1.78\n\u22120.02\nPre-1982 buildings\n3.22\n3.42\n\u22120.00\nCommunity distance urban \npopulation (millions)\n1.54\n1.67\n\u22120.04\nMedian household income (thousand \nUS$)\n34.12\n35.49\n\u22120.08\nNorth Atlantic share\n0.46\n0.50\n\u22120.08\nSouth Atlantic share\n0.31\n0.24\n0.15\nGulf Coast share\n0.24\n0.26\n\u22120.05\nBuilt-up surface 1960 (%)\n0.03\n0.03\n\u22120.00\nBuilt-up surface 1970 (%)\n0.03\n0.03\n\u22120.00\nBuilt-up surface 1980 (%)\n0.04\n0.04\n\u22120.00\nSpillover area\nFlood zone A share\n0.27\n0.27\n\u22120.01\nFlood zone V share\n0.28\n0.27\n0.07\nBuilt-up surface 1960 (%)\n0.19\n0.21\n\u22120.02\nBuilt-up surface 1970 (%)\n0.27\n0.29\n\u22120.02\nBuilt-up surface 1980 (%)\n0.38\n0.39\n\u22120.01\nWe assess the success of our procedure in identifying control areas by comparing the mean \ncharacteristics of CBRS units (column (1)) and the synthetic control (column (2)). Column (3) \nshows the standardized mean difference (SMD) between treatment and control areas using \nsynthetic control weights.\nCBRS treatment unit\nMatched counterfactual\nLand cover, 1985\n10 km\nDeveloped\nCropland\nWater\nWetland\nGrass/shrub\nForest\nBarren\nUnit boundary\nFig. 2 | Example treatment and control areas. Here we show three CBRS units \n(top panel) above counterfactual areas resulting from the three-to-one match \n(bottom panel). Counterfactual units are constructed using a spatial clustering \nalgorithm based on land cover, elevation and distance to coast. Note that the \nadjacent comparisons shown here are not directly used in estimation; we simply \nprovide them to develop the reader\u2019s intuition for how the regionalization \nalgorithm functions. Instead, the matched regions serve as the donor pool of \ncounterfactual units used in the synthetic controls approach.\n\nNature Climate Change | Volume 14 | September 2024 | 936\u2013942\n939\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nin spillover areas cannot be explained by a supply contraction. Higher \nproperty values near CBRS lands are consistent with a large hedonic \nliterature that shows that coastal amenities such as beach access and \nhazard protection affect real estate markets13\u201318.\nHigher assessed value per acre in spillover areas results from both \ndenser development (more houses per acre) and higher values per \nproperty. Assessed values per acre are US$186,000 higher, or about \n50% larger than the control group, in the 0\u2013500 m distance band. The \nmagnitude of this effect decreases gradually with distance.\nWe next examine the impacts of the CBRS on flood damages in \nspillover areas. We find economically large and statistically significant \nnegative impacts on overall damages from flooding, as measured by \nflood insurance claims per acre (Fig. 4d). Annual insurance claims are \nUS$420 to US$760 less per acre (40\u201364% lower than in control areas). \nNotably, treatment and control spillover areas have nearly identical \nshares of land in floodplains (Table 1) and distributions of properties\u2019 \ndistances to the shoreline (Extended Data Fig. 6), suggesting that the \ndifferences in flood damages is not caused by geographic differences.\nThe per-acre measure of flood damages is influenced by devel-\nopment densities in flood zones and flood damages per property. \nConsistent with the increased development in spillover areas, there \nare more buildings per acre in high-risk flood zones of CBRS spillover \nareas as compared with control spillover areas (Fig. 4f). However, \nflood claims are US$19\u201327 lower per US$1,000 of coverage in treat-\nment areas (Fig. 4e), representing a 14\u201325% reduction in flood dam-\nages accounting for flood insurance uptake. In other words, the CBRS \nprovides flood protection at the property level. These flood protec-\ntion services are probably generated by more natural lands inside the \nunits acting as barriers that dissipate and absorb flood waters. Indeed, \nprevious work establishes a link between wetlands and reductions in \nflood damages9\u201312.\nA back-of-the-envelope calculation shows that the original system \nof CBRS units generates US$389 million per year in savings for the NFIP. \nThis figure represents approximately 19% of average annual NFIP claims \nin Atlantic and Gulf coast counties over the period 2009\u20132021. The origi-\nnal units make up only 0.46% of land areas in these counties. If we assume \nthe CBRS units added later along the Gulf and Atlantic coasts generate \nsimilar benefits, the total saving in the current system (excluding the \nGreat Lakes) would be US$930 million per year in annual NFIP claims \ngenerated from removing federal development subsidies on only 1% of \nlands in these counties. This finding complements two existing studies \nof the savings CBRA generates in federal post-disaster assistance35,36.\nTable 2 | The direct effect of the CBRS within designated areas\nOutcome\nEstimate\nStandard error\nP value\nControl mean\nRelative effect (%)\nN\nDevelopment densities\nBuilt-up area (%)\n\u22120.1195\n0.071\n0.0927\n0.2881\n\u221241\n928\nBuildings per acre\n\u22120.0439\n0.0088\n8.7 \u00d7 10\u22127\n0.0528\n\u221283\n460\nAssessed values\nTotal assessed value (US$ per acre)\n24,141\n31,799\n0.4481\n78,518\n+31\n460\nProperty characteristics\nSales price (US$)\n107,373\n294,247\n0.716\n1,032,754\n+10\n106\nLot size (acres)\n1.24\n2.75\n0.654\n4.97\n+25\n98\nSquare footage\n\u2212132\n814\n0.872\n4,120\n\u22123\n91\nBedrooms\n0.04\n0.22\n0.863\n3.31\n+1\n83\nPopulation demographics\nWhite (%)\n2.88\n1.19\n0.0158\n87.79\n+3\n453\nBlack (%)\n\u22121.81\n0.93\n0.0526\n6.29\n\u221229\n453\nMedian household income (US$)\n15,105\n3,462\n1.6 \u00d7 10\u22125\n78,105\n+19\n434\nMedian rent as percent of income\n1.97\n1.11\n0.0759\n31.07\n+6\n346\nOccupied housing units (%)\n\u221213.58\n2.16\n7.1\u201310\n69.29\n\u221220\n452\nThe direct effects are estimated as weighted differences of outcomes inside CBRS and control units. The estimation uses synthetic control weights constructed from pre-trends in development \ndensities and pre-treatment measures of land use, geography and sociodemographic variables. Test statistics are based on a two-sided t-test. The relative effect is calculated as the percent \nchange from outcomes in the absence of treatment, as measured by the control group mean.\nNon-efective\nEfective\nDistribution of individual treatment efects \n\u2212100\n\u221250\n0\n50\n100\n0\n5\n10\n15\nRelative efect (% change)\n\u22121.0\n\u22120.5\n0\n0.5\n1.0\nEfect on built-up surface (p.p.)\nYear\nYear\n0\n10\n20\n30\n40\n50\nDevelopment densities in CBRS and control areas \n1960\n1980\n2000\n0\n0.1\n0.2\n0.3\n0.4\n0.5\n1960\n1980\n2000\n\u22120.2\n\u22120.1\n0\n0.1\n0.2\nFrequency\nFrequency\nTreatment \u2013 control\nBuilt-up surface (%)\na\nb\nFig. 3 | Effect of CBRS designations on development densities. a, Trends in \nbuilt-up surface area in treatment and control areas: left is the percent of land \narea covered in built-up surface in treatment (solid blue line) and control (dashed \ngrey line) areas every 10 years between 1960 and 2010; right is the difference in \nbuilt-up surface over time (treatment minus control). b, The distribution of unit-\nspecific treatment effects: left is the estimate of the effect of CBRS designation on \nbuilt-up surface, measured in p.p.; right is the relative effect (top coded at 100%).\n\nNature Climate Change | Volume 14 | September 2024 | 936\u2013942\n940\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nFinally, we estimate the spillover effect of CBRS designations on \ndemographics (Fig 4g\u2013i). The policy is associated with a large increase \nin household income (over US$10,000), a decline in the share of occu-\npied housing units (5\u20138 p.p.), and an increase in the share of white \npeople (1.5\u20132.5 p.p.). These effects largely mirror the demographic \neffects of the policy within the designations. Given the above evidence \nthat CBRS lands generate natural amenities and protection services, \nthese results are consistent with past findings on residential sorting in \nresponse to environmental amenities20,37.\nIn the spillover areas, unlike in the CBRS, federal flood insurance \nand disaster aid are still available. Thus the equity implications of \nthe policy are murky. The environmental benefits appear to flow \nto wealthier households in both the CBRS and spillover areas. But \nin the CBRS, those households also bear most of the costs of devel-\nopment because federal infrastructure spending, disaster aid and \nsubsidized flood insurance are unavailable. The same is not true in \nthe spillover areas.\nFiscal impacts on counties\nWe calculate the effect of CBRS designations on property tax revenues \nby combining our total assessed value estimates in both CBRS units and \ntheir spillover areas with current average county property tax rates. \nWe find no evidence of a change in property tax revenues within CBRS \ndesignations. However, we find a US$911 million per year increase in \nrevenues in spillover areas. This figure represents 2.5% of total local \nproperty tax revenues in Atlantic and Gulf coast states.\nOur findings are informative for coastal communities caught \nbetween, on one hand, increasing disaster costs, and on the other, the \nfiscal implications of limiting development. Our calculation suggests \nthat there is not necessarily a hard trade-off between the two objectives. \nRather, we show that curbing development in environmentally sensitive \nareas can increase the property tax base by increasing development \ndensities and property values in nearby locations.\nDiscussion\nCoastal communities are centres of economic development and home \nto critical natural resources but face substantial threats from climate \nchange and human development. This paper investigates the efficacy \nof one policy approach to limiting populations in harm\u2019s way\u2014with-\ndrawing the availability of federal financial assistance in risky areas. \nWe show that, in the case of the CBRA, this approach has been highly \neffective at limiting development. Moreover, the resulting conservation \nof natural lands generates environmental services in surrounding com-\nmunities, increasing property values and reducing flood damages. In \ncombination with analyses of the savings in federal disaster assistance \nexpenditures35, our findings provide a comprehensive assessment \nuseful for policymakers.\nMethodologically, we develop a new approach for causal infer-\nence, using spatial machine learning and matching to construct a \ncomparable control group by mimicking the original CBRA designation \nprocess. This approach tackles the central problem of non-random \nassignment of treatment present in a broad class of place-based policies \n\u22120.1\n0\n0.1\n0.2\n0.3\n500\n1,000\n1,500\nDistance to unit boundary (m)\nDistance to unit boundary (m)\nDistance to unit boundary (m)\nDistance to unit boundary (m)\nDistance to unit boundary (m)\nDistance to unit boundary (m)\nDistance to unit boundary (m)\nDistance to unit boundary (m)\nDistance to unit boundary (m)\n2,000\n0\n250\n500\n500\n1,000\n1,500\n2,000\n0\n100\n200\n300\n500\n1,000\n1,500\n2,000\n\u22121,500\n\u22121,000\n\u2212500\n0\n500\n1,000\n1,500\n2,000\n\u221240\n\u221220\n0\n500\n1,000\n1,500\n2,000\n\u22120.05\n0\n0.05\n0.10\n0.15\n500\n1,000\n1,500\n2,000\n0\n5,000\n10,000\n15,000\n20,000\n25,000\n500\n1,000\n1,500\n2,000\n\u22120.10\n\u22120.05\n0\n500\n1,000\n1,500\n2,000\n0\n0.02\n0.04\n500\n1,000\n1,500\n2,000\nMedian houshold\nincome (US$)\nDemographics\nOccupied units (%)\nWhite (%)\nNFIP claims\nper acre (US$)\nFlood damages\nNFIP claims per\nUS$1,000 coverage (US$)\nBuildings per\nacre in SFHA\nBuildings per acre\nDevelopment and property values \nAverage sales\nprice (US$1,000)\nAssessed value per\nacre (US$1,000)\na\nb\nc\nd\ne\nf\ng\nh\ni\nFig. 4 | Spillover effects of the CBRS in neighbouring communities. a\u2013i, Three \neconomic development outcomes (a\u2013c), three flood insurance outcomes (d\u2013f) \nand three demographic outcomes (g\u2013i) are shown. Dots are point estimates for \neach 500 m band and the corresponding 90% (95%) confidence intervals are \nshown in dark (light) blue. Confidence intervals are based on robust standard \nerrors. The sample size is N\u2009=\u20091,930 for all panels except b, where N\u2009=\u2009852. \nSFHA stands for \u201cSpecial Flood Hazard Area\u201d, the term used by the US Federal \nManagement Agency to describe areas with high flood risk.\n\nNature Climate Change | Volume 14 | September 2024 | 936\u2013942\n941\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nand could be applied to other settings where it is notoriously difficult \nto establish causal effects.\nOur results indicate that removing federal development incen-\ntives can be a cost-effective option for preventing over-development \nin risky areas while generating co-benefits. Still, programmes like the \nCBRS are designed to pre-empt development in risky areas, not assist \nin managed retreat. In areas where strategic relocation of people and \ncapital is deemed necessary, other policy interventions are likely to \nbe required.\nOur findings have the potential to inform a number of ongoing pol-\nicy discussions. The Strengthening Coastal Communities Act (S.5185) \nwould add approximately 292,000 acres to the CBRS. Our estimates \ncould serve as inputs into the cost\u2013benefit analysis of this proposal. \nMore generally, our estimates can inform state- and local-level deci-\nsions about zoning practices and government support for infrastruc-\nture development and repair in high-risk coastal areas. Notably, our \nresults apply to risky areas beyond just coastlines; a similar approach \nto managing development could be taken in inland floodplains or in \nwildfire-prone areas38.\nOur study has important limitations. First, although we believe \nour empirical approach represents a step forward, this is ultimately a \nretrospective, quasi-experimental analysis. Second, our estimates do \nnot capture the full range of benefits and costs associated with CBRS \ndesignation. Additional costs may include distortions to the spatial \nallocation of economic growth39, while benefits may include wildlife \nhabitat, water filtration or recreation opportunities. Third, we are not \nable to observe state and local policies that may either counteract or \nreinforce CBRS designations. We find heterogeneous treatment effects \nacross states but are unable to attribute those differences to specific \npolicies. Our estimates should therefore be interpreted as the net \neffect of CBRS designations, inclusive of state and local responses, to \nthe federal policy.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-024-02082-3.\nReferences\n1.\t\nBranham, J., Kaza, N., BenDor, T. K., Salvesen, D. & Onda, K. \nRemoving federal subsidies from high-hazard coastal areas slows \ndevelopment. Front. Ecol. Environ. 20, 500\u2013506 (2022).\n2.\t\nOnda, K., Branham, J., BenDor, T. K., Kaza, N. & Salvesen, D. Does \nremoval of federal subsidies discourage urban development? An \nevaluation of the US Coastal Barrier Resources Act. PloS ONE 15, \ne0233888 (2020).\n3.\t\nBagstad, K. J., Stapleton, K. & D\u2019Agostino, J. R. Taxes, subsidies, \nand insurance as drivers of United States coastal development. \nEcological Economics 63, 285\u2013298 (2007).\n4.\t\nCordes, J. J. & Yezer, A. M. J. In harm\u2019s way: does federal spending \non beach enhancement and protection induce excessive \ndevelopment in coastal areas? Land Econ. 74, 128\u2013145 (1998).\n5.\t\nKousky, C. & Olmstead, S. M. Induced Development in Risky \nLocations: Fire Suppression and Land Use in the American West \n(Resources for the Future, 2010).\n6.\t\nPeralta, A. & Scott, J. B. Moving to flood plains: the unintended \nconsequences of the National Flood Insurance Program on \npopulation flows. In Proc. Environmental Risk, Justice and \nAmenities in Housing Markets Vol. 19 https://aamperalta.github.io/ \nfiles/floods.pdf (American Economic Association, 2019).\n7.\t\nCraig, R. K. Coastal adaptation, government-subsidized \ninsurance, and perverse incentives to stay. Climatic Change 152, \n215\u2013226 (2019).\n8.\t\nBrowne, M. J., Dehring, C. A., Eckles, D. L. & Lastrapes, W. D. Does \nNational Flood Insurance Program participation induce housing \ndevelopment? J. Risk Insur. 86, 835\u2013859 (2019).\n9.\t\nCostanza, R. et al. The value of coastal wetlands for hurricane \nprotection. Ambio 37, 241\u2013248 (2008).\n10.\t Narayan, S. et al. The value of coastal wetlands for flood damage \nreduction in the northeastern USA. Sci. Rep 7, 9463 (2017).\n11.\t\nSun, F. & Carson, R. T. Coastal wetlands reduce property damage \nduring tropical cyclones. Proc. Natl Acad. Sci. USA 117, 5719\u20135725 \n(2020).\n12.\t Taylor, C. A. & Druckenmiller, H. Wetlands, flooding, and the \nClean Water Act. Am. Econ. Rev. 112, 1334\u20131363 (2022).\n13.\t Beltran Hernandez, A., Maddison, D. & Elliot, R. Is flood risk \ncapitalised into property values? Ecol. Econ. 146, 668\u2013685 (2018).\n14.\t Hino, M., Field, C. B. & Mach, K. J. Managed retreat as a response \nto natural hazard risk. Nat. Clim. Change 7, 364\u2013370 (2017).\n15.\t Below, S., Beracha, E. & Skiba, H. Land erosion and coastal home \nvalues. J. Real Estate Res. 37, 499\u2013536 (2015).\n16.\t Zhang, Y., Hwang, S. N. & Lindell, M. K. Hazard proximity or risk \nperception? Evaluating effects of natural and technological \nhazards on housing values. Environ. Behav. 42, 597\u2013624 (2010).\n17.\t Landry, C. E., Turner, D. & Allen, T. Hedonic property prices and \ncoastal beach width. Appl. Econ. Perspect. Policy 44, 1373\u20131392 \n(2022).\n18.\t Bin, O., Crawford, T. W., Kruse, J. B. & Landry, C. E. Viewscapes and \nflood hazard: coastal housing market response to amenities and \nrisk. Land Econ. 84, 434\u2013448 (2008).\n19.\t Kuwayama, Y., Olmstead, S. & Zheng, J. A more comprehensive \nestimate of the value of water quality. J. Public Econ. 207, 104600 \n(2022).\n20.\t Tiebout, C. M. A pure theory of local expenditures. J. Polit. Econ. \n64, 416\u201324 (1956).\n21.\t Banzhaf, S., Ma, L. & Timmins, C. Environmental justice: the \neconomics of race, place, and pollution. J. Econ. Perspect. 33, \n185\u2013208 (2019).\n22.\t Bakkensen, L. A. & Ma, L. Sorting over flood risk and implications \nfor policy reform. J. Environ. Econ. Manag. 104, 102362 (2020).\n23.\t Glaeser, E. L. & Gyourko, J. The Impact of Zoning on Housing \nAffordability NBER Working Paper 8835 (NBER, 2002).\n24.\t Sieg, H., Smith, K. V., Banzhaf, S. H. & Walsh, R. Estimating \nthe general equilibrium benefits of large changes in spatially \ndelineated public goods. Int. Econ. Rev. 45, 1047\u20131077 (2004).\n25.\t Banzhaf, S. H. & McCormick, E. in The Political Economy of \nEnvironmental Justice Ch. 2 (Stanford University Press, 2012).\n26.\t Gamper-Rabindran, S. & Timmins, C. Does cleanup of hazardous \nwaste sites raise housing values? Evidence of spatially localized \nbenefits. J. Environ. Econ. Manag. 65, 345\u201360 (2013).\n27.\t Taylor, L. O., Phaneuf, D. J. & Li, X. Disentangling property value \nimpacts of environmental contamination from locally undesirable \nland uses: implications for measuring post-cleanup stigma. J. \nUrban Econ. 93, 85\u201398 (2017).\n28.\t Haninger, K., Ma, L. & Timmins, C. Does cleanup of hazardous \nwaste sites raise housing values? Evidence of spatially localized \nbenefits. J. Assoc. Environ. Resour. Econ. 4, 197\u2013241 (2017).\n29.\t Kahn, M. E., Vaughn, R. & Zasloff, J. The housing market effects \nof discrete land use regulations: evidence from the California \ncoastal boundary zone. J. Hous. Econ. 19, 269\u2013279 (2010).\n30.\t Severen, C. & Plantinga, A. J. Land-use regulations, property \nvalues, and rents: decomposing the effects of the California \nCoastal Act. J. Urban Econ. 107, 65\u201378 (2018).\n31.\t Ehrlich, I. & Becker, G. S. Market insurance, self-insurance, and \nself-protection. J. Polit. Econ. 80, 623\u2013648 (1972).\n32.\t Cicchetti, C. J. & Dubin, J. A. A microeconometric analysis of \nrisk aversion and the decision to self-insure. J. Polit. Econ. 102, \n169\u2013186 (1994).\n\nNature Climate Change | Volume 14 | September 2024 | 936\u2013942\n942\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\n33.\t Salvesen, D. The Coastal Barrier Resources Act: has it  \ndiscouraged coastal development? Coast. Manag. 33,  \n181\u2013195 (2005).\n34.\t Branham, J., Salvesen, D., Kaza, N. & BenDor, T. K. A  \nwrench in the machine: how subsidy removal alters the \npolitics of coastal development. J. Am. Plann. Assoc. 90, \n18\u201329 (2022).\n35.\t Coburn, A. S. & Whitehead, J. C. An analysis of federal \nexpenditures related to the Coastal Barrier Resources Act (CBRA) \nof 1982. J. Coast. Res. 35, 1358\u20131361 (2019).\n36.\t The Coastal Barrier Resources Act: Harnessing the Power of Market \nForces to Conserve America\u2019s Coasts and Save Taxpayers\u2019 Money \n(US Fish & Wildlife Service, 2002).\n37.\t Banzhaf, S. H. & Walsh, R. P. Segregation and tiebout sorting: \nthe link between place-based investments and neighborhood \ntipping. J. Urban Econ. 74, 83\u201398 (2013).\n38.\t Walls, M., Wibbenmeyer, M. & Kousky, C. Does the Coastal Barrier \nResources Act Provide a Policy Template to Address Wildfire Risk? \n(2019); https://www.resources.org/common-resources/does- \ncoastal-barrier-resources-act-provide-policy-template-address- \nwildfire-risk/\n39.\t Hsieh, C.-T. & Moretti, E. Housing constraints and spatial \nmisallocation. Am. Econ. J. Macroecon. 11, 1\u201339 (2019).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this licence, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2024\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nMethods\nThe goal of our empirical strategy is to estimate the causal effect of \nCBRS designations. The core challenge is identifying a set of appro-\npriate counterfactual units to serve as \u2018control\u2019 areas for CBRS \u2018treat-\nment\u2019 areas. The CBRA encourages the conservation of biologically \nrich, underdeveloped and hurricane-prone coastal areas. Therefore, \nwe cannot simply compare our outcomes of interest in CBRS units to \nthose in all other coastal areas; we must identify comparable areas that \ncould have been selected for CBRS designation but were not. We do so \nusing a novel procedure designed to mimic the process by which natural \nresource planners defined CBRS boundaries. Intuitively, our method \nfor finding counterfactual treatment areas relies on finding locations \nthat are indistinguishable (to the algorithm) from CBRS lands at the \ntime of designation40,41.\nSelection process and criteria for CBRS designations\nWe first describe the selection process for CBRS designations, as our \nempirical approach relies on replicating it. The CBRS designation pro-\ncess is described in detail in the 1982 Federal Register and in refs. 42,43, \nwhere the US Fish & Wildlife Service (FWS) establishes a set of \u2018defini-\ntions and delineation criteria\u2019. CBRS designations were then based upon \nthe application of these criteria to on-the-ground situations.\nThe first criteria for CBRS designations is that the land should be a \n\u2018coastal barrier\u2019, a class of low coastal land forms that protect landward \nareas from tidal, wave or wind energies. For the purposes of CBRS des-\nignations, the definition of coastal barriers also includes all associated \naquatic habitats such as adjacent wetlands, marshes, estuaries, inlets \nand nearshore waters. In addition to meeting this geological definition, \nthe CBRS requires coastal barriers to be \u2018undeveloped\u2019 in order to be \nincluded in the system. Specifically, the delineation criteria state that \nan area should be considered \u201conly if there are few manmade structures \non the barrier or any portion thereof and these structures and man\u2019s \nactivities on the barrier do not significantly impede geomorphic and \necological processes\u201d42.\nAccording to the Federal Register, Reports to Congress and conver-\nsations with programme officers at the FWS, natural resource planners \nexamined US Geological Survey (USGS) topographic maps and aerial \nphotographs to make the original CBRS designations based on just two \ncriteria: (1) adherence to the geologic definition of a coastal barrier, \nwhich depends on elevation, land cover and location relative to the \nshoreline; and (2) development levels.\nCounterfactual construction\nWe now describe our procedure for identifying plausible control areas: \nareas that could have been selected for CBRS designation in 1982 based \non the selection criteria, but were not.\nCBRS boundaries do not follow traditional administrative bounda-\nries; they were hand drawn to follow geomorphic and development \nfeatures44. Our first step is to trace out potential counterfactual areas \nusing an automated procedure that closely resembles this process. \nImportantly, we can observe close proxies for the information the \nplanners had available at the time of designation\u2014aerial photographs \nand topographic maps (Extended Data Fig. 2). We begin with 300 m \nresolution gridded data on historical land cover, development levels, \nelevation and distance to coast. Each cell of the raster represents a dis-\ntinct observation that will be grouped into a region. We only consider \ngrid cells within 2 km of the coast. We exclude any cell that is 100% \nwater, within a CBRS unit (including both original units designated \nin 1982 and all units designated since then) or an otherwise protected \narea, and all grid cells within 2 km of a CBRS unit (to avoid selecting \ncontrol areas that may be \u2018treated\u2019 by spillover effects of CBRS units).\nWe then apply regionalization to group these pixels into spatially \ncontiguous areas that share similar attributes45. Regionalization is one \ntype of clustering\u2014a machine learning technique that sorts observa-\ntions into groups without any prior idea about what the groups are. \nClusters are delineated so that members of a group should be more \nsimilar to one another than they are to members of a different group. \nFor example, observations in one group may have consistently high \nscores on some traits but low scores on others. Regionalization is an \napplication of clustering to spatial data that can be used to provide \ninsights into the geographic structure of complex multivariate spatial \ndata. A \u2018region\u2019 is similar to a cluster, in the sense that all members of \na region have been grouped together, but a region also describes a \nclear geographic area. That is, regionalization uses the same logic as \nstandard clustering techniques, but also requires connectivity: two \ncandidates can only be grouped together in the same region if a path \nexists from one member to another member that never leaves the \nregion45 (see Supplementary Section B.2 for details).\nRegionalization groups all coastal pixels into spatial clusters \n(\u2018regions\u2019). We limit our sample to only those regions that would have \nmet the basic requirements for inclusion into the CBRS by conducting \na three-to-one propensity score match between the regions and CBRS \nunits based on land cover, development levels and elevation. The algo-\nrithm effectively acts as a natural resource planner would have\u2014tracing \nout low-elevation coastal lands that had high shares of wetlands and \nbeaches (barren), while avoiding highly developed areas. To illustrate \nthe results of this procedure, we show examples of CBRS areas beside \nmatched counterfactuals (Fig. 2). Notably, the adjacent comparisons \nshown here are not directly used in estimation; we simply provide them \nto develop the reader\u2019s intuition for how the regionalization algorithm \nfunctions. Instead, the matched regions serve as the donor pool for \ncounterfactual units.\nNext we apply multi-unit synthetic controls, a technique designed \nto reduce selection bias in observational studies by weighting the \ncontrol group to better match the treatment group prior to the inter-\nvention46\u201349. We match on time trends in development densities, both \ninside the designated units and in spillover areas, for the two decades \npreceding the policy intervention (1960, 1970 and 1980). We also match \non a suite of covariates that we can observe around the time of designa-\ntion in 1982, including measures of land cover, elevation, infrastructure \ndensity, proximity to urban centres, income and flood risk in spillover \nareas. This allows us to determine a set of weights that, on aggregate, \nbalances pre-trends in development densities and pre-treatment char-\nacteristics across CBRS units and the control group. We can then use \nthe synthetic control to estimate what would have happened to the \nCBRS units if they had not been treated by the policy.\nOur empirical strategy requires the CBRS and control areas to \nhave been comparable at the time of designation, while allowing for, \nand measuring, divergence in outcomes over the four decades since, \nbased on treatment status. We therefore limit our sample to the original \nset of CBRS units designated in 1982, and all data used in the process \nof generating counterfactuals comes from as close to this year as pos-\nsible (and no later than the year 1990). Data sources and processing \ntechniques are described in detail in Supplementary Section B.1.\nTo illustrate the output of our procedure, we plot the locations of \nall CBRS and counterfactual units included in our sample (Extended \nData Fig. 4). We include control areas in all 18 Atlantic and Gulf Coast \nstates, although four of these states (Maryland, New Hampshire, New \nJersey and Virginia) did not have any units in the original CBRS. Because \nsome states had a limited pool of undeveloped coastal areas in 1982, \nour procedure does not require the distribution of control units to \nmatch the distribution of treatment units across states. However, we \ndo balance treatment and control units across three broad regions\u2014the \nNorth Atlantic, South Atlantic and Gulf Coast\u2014to avoid making inap-\npropriate comparisons between, for example, coastal areas in Maine \nand Mississippi.\nWe assess the success of our procedure for constructing coun-\nterfactual areas by comparing the mean characteristics of the CBRS \ntreatment group and synthetic control (Table 1). Encouragingly, our \nmachine learning and matching procedure brings the observable \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\ncharacteristics of CBRS treatment and counterfactual areas into align-\nment. Column (3) shows that standardized mean difference (SMD) \nbetween CBRS and counterfactual areas is at or below 0.1 (the com-\nmonly used threshold for accessing balance) across the 18 covariates.\nAdditional considerations for causal identification\nOur approach to identification requires the assumption that not all \nareas that meet the CBRS delineation criteria were designated as part \nof the original system in 1982. This assumption is supported by the \nfact that additional areas have been added to the CBRS over the past \n40 years, with the most recent addition proposed in December 202250. \nAccording to a 1988 Report to Congress51 and conversations with pro-\ngramme officers at the FWS, not all eligible areas were designated as \npart of the original system for two primary reasons. First, USGS quad-\nrangles used in the inventory process did not show sufficient detail to \nidentify all qualifying undeveloped coastal barrier areas. Second, the \nscientific definition of what qualifies as a \u2018coastal barrier\u2019 has evolved \nover time. We argue that both of these factors are plausibly exogenous \nto the housing market outcomes we analyse.\nOne additional concern is that local politics might have affected \nwhich areas were designated as part of the CBRS. In particular, if local \ngovernments were concerned about the negative impacts of CBRS \ndesignation on property tax revenues, they may have objected to \nCBRS designations in their jurisdiction. Reassuringly, no units were \nremoved from the original proposed set of designations after the public \ncomment period42, reflecting limited ability of localities to affect the \ndesignation process. Additionally, we empirically test for the presence \nof this type of selection in two ways. First, we test for differences in local \nreliance on property tax revenues. Second, we test for differences in \nthe voting records of local members of Congress on environmental \nissues, as measured by the League of Conservation Voters National \nEnvironmental Scorecard, which records the voting records of all \nmembers of Congress on major environmental legislation. We find no \nmeaningful differences in either of these measures across treatment \nand control areas: local reliance on property tax revenues is 35.2% in \ntreatment areas and 35.9% in control areas (P value on difference = 0.61) \nand environmentally friendly voting records (League of Conservation \nVoters score) are 57.0 in treatment areas and 56.8 in control areas (P \nvalue on difference = 0.96). These results are inconsistent with local \npolitics causing selection away from pro-development and towards \npro-environment areas.\nConsideration of development in the designation process\nIt is worth emphasizing the consideration of development levels in the \nCBRS designation process. The \u2018any portion thereof\u2019 language in the \ndefinition of an \u2018undeveloped\u2019 coastal barrier is key because it means \nthat the statutory definition does not require an entire coastal barrier \nto be designated as a CBRS unit; the FWS had the authority to include \nonly the portions of the barrier that were underdeveloped. In fact, the \nboundary was often intentionally placed to exclude developed areas. \nThe Federal Register explains, \u201cthe side boundary is placed immediately \nadjacent to the cluster of structures or the area with a full complement \nof infrastructure indicating the end of the developed portion of the \ncoastal barrier\u201d42. We show an example of this practice for a unit in \nNorth Carolina, where developed areas were excluded from the CBRS \ndesignation to satisfy the definition of an \u2018undeveloped\u2019 coastal barrier \n(Extended Data Fig. 1a).\nThe central role of development levels in the delineation process \ncalls into question whether the causal effect of CBRS designation can be \nidentified through a spatial regression discontinuity design, as in ref. 1.  \nThe core assumption of a spatial regression discontinuity\u2014that land \nlocated just within and just outside the boundary can be assumed to be \nsimilar in all ways except for assignment to treatment\u2014is problematic \nin this setting because natural resource planners hand-selected CBRS \nboundaries to avoid already developed areas. To see this, we calculate \nthe mean share of developed land just inside and just outside of CBRS \nboundaries in 1985 using data from the USGS\u2019s Land Change Moni-\ntoring, Assessment, and Projection (LCMAP) product52. We find that \naround the time of designation in 1985, the share of developed land \nwas already 9.6 p.p. (95% confidence interval = 6.5 to 12.8) higher just \noutside the boundary than just inside the boundary, representing a \nmore than doubling in development levels at the boundary. We define \n\u2018just\u2019 inside and outside using 100 m buffers, following ref. 1. We find \nthat there is a discrete jump in the share of developed land even as one \napproaches the boundary (Extended Data Fig. 1b).\nOur approach to identification, as described above, explicitly takes \ninto account the consideration of development levels in delineating \nCBRS boundaries and is designed to be able to recover the spillover \neffects of the policy on neighbouring areas. Indeed, we find that our \nprocedure does well in replicating the consideration of development \nlevels in drawing CBRS boundaries. Even before the policy was enacted, \ndevelopment levels were 33% higher just outside than just inside CBRS \nboundaries (within 100 m). We find a similar pattern among our syn-\nthetic control areas, where development levels are 24% higher just \noutside than just inside the boundary.\nOutcomes\nWe examine two sets of outcomes related to the removal of develop-\nment subsidies: direct effects within CBRS units and spillover effects \nin neighbouring communities. All outcomes are measured using the \nmost recent data available (2010 onwards) such that we capture the \nlong-term effect of CBRS designation. That is, we compare outcomes \ntoday in treatment and control units under the assumption that these \ntwo groups were comparable at the time of designation in 1982, and \ntest whether outcomes have diverged over the past four decades based \non treatment status.\nDirect effects include impacts on development levels, land values, \nand composition of the housing stock and population. We measure \ndevelopment levels using two different data sources. The first measure \nis the share of built-up surface area from the HISDAC-US Building Foot-\nprint Area (BUFA) database (https://dataverse.harvard.edu/dataverse/ \nhisdacus). Assembled from property assessment records, BUFA is a \ngridded data product that estimates the sum of building areas at 250 m \nspatial resolution every 10 years from 1900 to 2010. The key advantage \nof this dataset is that we can observe the evolution of built-up area over \ntime, including in the decades before CBRS designation. Our second \nmeasure of development densities is the number of structures per acre, \ncalculated from Microsoft Maps\u2019 Building Footprint database (https:// \ngithub.com/microsoft/USBuildingFootprints). This dataset provides \napproximately 130 million computer-generated building footprints \nderived from satellite imagery for the entire United States. These data \noffer higher spatial resolution than HISDAC-US, but only for a single \nsnapshot in time during the post-period.\nWe measure land values using property sales prices and assess-\nment values from the Zillow Transactions and Assessment Database \n(ZTRAX) (https://www.zillow.com/research/ztrax/). The ZTRAX data-\nset also provides information on the composition of the housing stock \n(for example, lot size, year built, square footage, number of bedrooms). \nBecause outcome data from the ZTRAX Database (property sales prices \nand characteristics) are incomplete\u2014only about half of treatment and \ncontrol units have one or more parcels in the database\u2014we calculate \na separate set of synthetic control weights for the ZTRAX outcomes \n(Extended Data Fig. 3 and Extended Data Table 1). Missingness in \nthe ZTRAX database is described in Supplementary Section B.1 and \nexplored at length in ref. 53.\nWe evaluate the composition of the population using census \nblock-group-level 5-year estimate (2016\u20132020) from the American \nCommunity Survey. Building footprints and ZTRAX data points are \nassigned to CBRS units by intersecting the geocoded polygons and \npoints with CBRS boundaries. To calculate census observations for \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\neach CBRS area, we aggregate the values from all census block groups \nthat intersect a given CBRS unit using population weights calculated \nfrom high-resolution (1 km) gridded population data.\nWe also examine the effects of CBRS areas on neighbouring com-\nmunities. To do so, we draw a 2 km buffer around each CBRS unit and \ncounterfactual area. When constructing spillover areas, we exclude \nany area that is in a current CBRS designation or otherwise protected \narea. For treatment spillover areas, if a geography falls within the 2 km \nof multiple CBRS designations, we assign it only to the closest CBRS \nunit. For counterfactual spillover areas, we exclude any area that is in \na spillover area (within 2 km) of a treated unit.\nIn addition to the outcomes described above, we measure the \nimpact of CBRS designation on flood damages in order to test whether \npreserving natural lands provides protective services from flooding. \nWe measure flood damages using flood insurance claims from the NFIP, \nthe dominant insurer for flooding in the United States. Notably, home-\nowners within CBRS areas are not eligible to participate in the NFIP, so \nwe do not look at the direct effect of CBRS designation on NFIP claims. \nBecause flooding is an infrequent event, we average annual flood claims \nover the years 2009 to 2020. The most granular geographic identifier \navailable in the NFIP claims and policies data at the time of analysis was \nthe property census tract. To estimate the value of NFIP claims in each \nCBRS spillover area, we aggregate the values from all census tracts that \nintersect a given spillover area, weighting by the number of buildings \nlocated in the Special Flood Hazard Areas designated by the US Federal \nEmergency Management Agency (the high-flood-risk areas where flood \ninsurance is required for properties with federally backed mortgages). \nThe intuition behind this approach is that the distribution of NFIP \npolicies and claims over geographic space will probably resemble the \ndistribution of structures in areas with high flood risk.\nWe provide summary statistics for our outcome measures (Supple-\nmentary Table B.1). Column 1 shows the mean values in counterfactual \nareas and column 2 shows the mean values in CBRS units. All means are \nweighted by the synthetic control weights used in estimation.\nEstimating equations\nOnce we have constructed the synthetic control, we estimate \nthe effect of CBRS designations on built-up surface area using a \ndifference-in-differences framework. We use the weighted regression\nBit = \u03b2(Ti \u00d7 POSTt) + \u03bbi + \u03b4POSTt + \u03b3Ri + \u03f5it\n(1)\nwhere i indexes the CBRS or control area and t indexes the time period. \nB denotes the percent of land area covered in built-up surface, T is an \nindicator for whether the area was treated by the CBRS and POST is an \nindicator equal to one in the post-treatment period. The regression also \nincludes unit-level fixed effects (\u03bbi) and a control for whether the unit \nis located on a barrier island or cape (R). We weight each observation \nby the synthetic control weight. The error term \u03b5 captures unexplained \nvariations. The treatment effect of interest, \u03b2, captures any systematic \ndifferences in built-up surface area caused by CBRS designation.\nFor all other outcomes, where outcome measures are not available \nin the pre-treatment period, we estimate the direct effect of CBRS des-\nignation using a simple weighted regression that compares outcomes \nin treatment areas today with outcomes in control areas today. The \nestimating equation is\nYi = \u03b1 + \u03b2Ti + \u03b3Ri + \u03f5i\n(2)\nwhere all variables are defined as in equation (1), except here Y cor-\nresponds to one of our other outcome variables. The regression is \nweighted by the synthetic control weights. The treatment effect of \ninterest, \u03b2, captures any systematic differences between the outcomes \nin CBRS and control areas today, under the identifying assumption that \nthese areas would have been comparable in the absence of treatment.\nTo estimate the spillover effects of CBRS designations on neigh-\nbouring communities, we turn to a spatial lag model. This allows us  \nto capture heterogeneity in spillover effects by distance to the unit. \nThe estimating equation is:\nYi,b =\n4\n\u2211\nb=1\n\u03b2b1[B = b] \u00d7 Ti + \u03b3Ri + \u03bbbXi,b + \u03f5i,b\n(3)\nwhere the indicator 1[B\u2009=\u2009b] is equal to one if the observation falls within \ndistance band b relative to the CBRS or counterfactual unit boundary. \nWe use four distance bands, in increments of 500 m, out to a total \ndistance of 2 km. Because we match on built-up surface area in overall \nspillover areas (not by distance band), we include distance-band level \ncontrols for 1985 land cover, flood risk and share protected area (Xi,b). \nThe regression is weighted using the synthetic control weights.\nHeterogeneous treatment effects\nWe next explore where CBRS designations are more and less effec-\ntive by identifying individual counterfactual and treatment effects. \nThis represents a departure from the main analysis, where we pool all \ntreated units and find a set of synthetic control weights that balances \nthe pre-trends in development densities and mean characteristics of \ntreatment and control, on aggregate. We do not employ individual \ntreatment effects in the main analysis because it is not possible to \nidentify an appropriate counterfactual for each individual CBRS unit; \nsome units look very different from all other coastal areas in the United \nStates. For this part of the analysis, we use a paired-down set of match-\ning variables that only includes pre-trends in development densities \nand pre-treatment land cover, elevation, proximity to urban centres \nand region. We identify synthetic controls that match well along all \nof these dimensions for 90 treatments units (55% of the full sample).\nWe evaluate heterogeneity in the effect of the CBRS on devel-\nopment densities in three ways. First, we report the distribution of \ntreatment effects across the sample of 90 units for which we are able \nto identify individual treatment effects. Second, we compute average \ntreatment effects by state. Third, we use a classification analysis to \ncompare the average characteristics of the most and least affected \nunits using two-sample t-tests. We define the least affected units as \nthose with a positive estimate of the individual treatment effect and \nthe most affected units as those with treatment effects, measured in \nrelative terms, in the most negative quartile (more than 67% reduction \nin development densities).\nImpacts on local property tax revenues\nWe calculate the approximate impact of CBRS designation on county \nrevenues from property taxes using the equation:\nRc = tc \u00d7 (\u03b2AV\nd Ac,d +\n4\n\u2211\nb=1\n\u03b2AV\nb Ac,b)\n(4)\nwhere Rc denotes the estimated effect of CBRS designation on property \ntax revenue in county c. The parameter \u03b2AV\nd  is our estimated effect of \nCBRS designation on total assessed value per acre within the CBRS unit \nand \u03b2AV\nb  is the estimated effect on assessed value within each distance \nband b of the CBRS unit. We multiply these estimated effects by the \ntotal acreage within the unit, Ac,d, and in each distance band, Ac,b, respec-\ntively. Summing these values gives us an estimate of the effect of CBRS \ndesignation on total assessed value within the county (the term in \nparentheses). Notably, this estimate captures the direct effects within \nCBRS units and the spillover effects on neighbouring properties, both \nin number and value of developed properties. To estimate the effect \non property tax revenue, we then multiply by the county tax rate, tc. \nThis tax rate is derived from the Lincoln Institute of Land Policy Prop-\nerty Tax Database (https://www.lincolninst.edu/data/significant- \nfeatures-property-tax/access-database/). The nominal tax rate is \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\ncalculated by averaging municipal and school district taxes and adding \nthem to base county tax rates.\nImpacts on NFIP payouts\nTo calculate the impact of CBRS designations on NFIP payouts in spillo-\nver areas, we multiply our point estimates of NFIP claims paid per acre \n(Fig. 4) by distance band with the acres of land in each band. We then \nsum over all CBRS units to calculate the total impact on NFIP payouts. \nFor savings within the CBRS units, we calculate the average claims per \nacre across all counterfactual areas and multiply it by CBRS land areas.\nRobustness to alternative research design\nIn Supplementary Section C, we present results based on overlap \nweighting under a propensity score matching framework as an alter-\nnative approach and robustness check.\nData availability\nAll raw data used in this study are publicly available except the \nproperty-level information on home values and characteristics from \nthe Zillow Transaction and Assessment Database (ZTRAX), which we \naccessed through a licence available to researchers. Instructions for \naccessing raw data are provided in Supplementary Section B.1. The pro-\ncessed datasets are available at https://github.com/hdruckenmiller/cbra.  \nSource data are provided with this paper.\nCode availability\nThe analysis code is available via GitHub at https://github.com/hdruck \nenmiller/cbra. The analysis code and scripts are available via Zenodo \nat https://doi.org/10.5281/zenodo.12199232 (ref. 54).\nReferences\n40.\t Englander, G. Information and spillovers from targeting policy in \nPeru\u2019s anchoveta fishery. Am. Econ. J. Econ. Policy 15, 390\u2013427 (2023).\n41.\t Pollmann, M. Causal inference for spatial treatments. Preprint at \narXiv https://arxiv.org/abs/2011.00373 (2020).\n42.\t Federal Flood Insurance Prohibition for Undeveloped Coastal \nBarriers; Proposed Identification and Submission of Report to \nCongress, (US Department of the Interior, 1982); https://www. \ngovinfo.gov/content/pkg/FR-1982-08-16/pdf/FR-1982-08-16.pdf\n43.\t Undeveloped Coastal Barriers: Report to Congress (US Department \nof the Interior, 1982); https://www.fws.gov/sites/default/files/ \ndocuments/Undeveloped-Coastal-Barriers-Report-1982.pdf\n44.\t John H. Chafee Coastal Barrier Resources System (US Fish & \nWildlife Service, 2022); https://www.fws.gov/glossary/john-h- \nchafee-coastal-barrier-resources-system\n45.\t Guo, D. Regionalization with dynamically constrained agglome\u00ad\nrative clustering and partitioning (redcap). Int. J. Geogr. Inf. Sci. \n22, 801\u2013823 (2008).\n46.\t Abadie, A., Diamond, A. & Hainmueller, J. Synthetic control methods \nfor comparative case studies: estimating the effect of California\u2019s \nTobacco Control Program. J. Am. Stat. Assoc. 105, 493\u2013505 (2010).\n47.\t Abadie, A. Using synthetic controls: feasibility, data requirements, \nand methodological aspects. J. Econ. Lit. 59, 391\u2013425 (2021).\n48.\t Ben-Michael, E., Feller, A. & Rothstein, J. Synthetic controls with \nstaggered adoption. J. R. Stat. Soc. B 84, 351\u2013381 (2022).\n49.\t Arkhangelsky, D., Athey, S., Hirshberg, D. A., Imbens, G. W. & \nWager, S. Synthetic difference-in-differences. Am. Econ. Rev. 111, \n4088\u20134118 (2021).\n50.\t S.5185 Strengthening Coastal Communities Act of 2022 (US \nCongress, 2022); https://www.congress.gov/bill/117th-congress/\nsenate-bill/5185/text\n51.\t 1988 Report to Congress: Coastal Barrier Resources System  \n(US Department of the Interior, 1988); https://www.fws.gov/media/\n1988-report-congress-coastal-barrier-resources-system\n52.\t Land Change Monitoring, Assessment, and Projection (USGS, 2022); \nhttps://www.usgs.gov/special-topics/lcmap\n53.\t Nolte, C. et al. Data practices for studying the impacts of \nenvironmental amenities and hazards with nationwide property \ndata. Land Econ. 100, 200\u2013221 (2023).\n54.\t Druckenmiller, H., Liao, Y. (P.), Pesek, S., Walls, M. & Zhang, S. \nScripts and code for \u2018Removing development incentives in risky  \nareas promotes climate adaptation\u2019. Zenodo https://doi.org/10.5281/ \nzenodo.12199232 (2024).\n55.\t Onslow Beach Complex L05 (2 of 2) and Topsail Unit L06 (1 of 2) \n(John H. Chafee Coastal Barrier Resources System, US Fish & \nWildlife Service, 2018); https://www.fws.gov/media/onslow- \nbeach-complex-l05-2-2-topsail-unit-l06-1-2\n56.\t Hyannis Quadrangle: 7.5 Minute Series (Topographic) (USGS, \n1986); https://prd-tnm.s3.amazonaws.com/StagedProducts/ \nMaps/HistoricalTopo/GeoTIFF/MA/MA_New%20Bedford_353168_ \n1986_100000_geo.tif\n57.\t Hyannis Quadrangle: 7.5 Minute Series (Orthophotoquad) (USGS, \n1977); https://prd-tnm.s3.amazonaws.com/StagedProducts/Maps/\nHistoricalTopo/GeoTIFF/MA/MA_Hyannis_351037_1974_25000_\ngeo.tif\nAcknowledgements\nThis research was funded by the Lincoln Institute of Land Policy (H.D., \nY.L. and M.W.). We thank T. BenDor, P. Mulder, A. Reilly, C. Taylor and  \nD. Wright for helpful comments and suggestions. Property assessment \nand transaction data were provided by Zillow through the Zillow \nTransaction and Assessment Dataset (ZTRAX). More information on \naccessing the data can be found at https://www.zillow.com/research/ \nztrax. The results and opinions are those of the author(s) and do not \nreflect the position of Zillow Group.\nAuthor contributions\nH.D. and Y.L. conceived the study and contributed equally as first \nauthors. H.D., Y.L. and M.W. designed the analysis. H.D., Y.L., S.P. and \nS.Z. constructed and analysed the data. All authors contributed to \nwriting and reviewing the manuscript.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at  \nhttps://doi.org/10.1038/s41558-024-02082-3.\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41558-024-02082-3.\nCorrespondence and requests for materials should be addressed to \nHannah Druckenmiller.\nPeer review information Nature Climate Change thanks Allan Beltran \nand the other, anonymous, reviewer(s) for their contribution to the \npeer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nExtended Data Fig. 1 | CBRS boundaries depend on development levels. \n(a) Example of CBRS unit boundary for Topsail Unit (L06) in North Carolina55. \nDeveloped areas are excluded from the system unit to meet the definition of an \n\u2018underdeveloped\u2019 coastal barrier. (b) Share of developed land within 100m of \nCBRS boundaries in 1985 using LCMAP. Left side of the plot is just inside CBRS \nboundaries and right side is just outside CBRS boundaries. Points show the \naverage share developed in 20m increments approaching the boundary.  \nShaded areas show 95% confidence intervals. Panel a adapted with permission \nfrom ref. 55, US Fish and Wildlife Service.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nExtended Data Fig. 2 | Comparison of maps used by CBRS planners and our \nmodel. Example of input data for the Sandy Neck CBRS Unit (C09). The top row \nshows the sources of information used by CBRS planners in making the original \ndesignations, USGS quadrangles56 (top left) and aerial photographs57 (top right). \nThe bottom row shows two of the inputs into our model, land cover maps from \nLCMAP (bottom left) and built-up surface area from HISDAC-US (bottom right). \nThe CBRS unit is outlined in yellow. Panels adapted with permission from:  \ntop left, ref. 56, US Geological Survey; top right, ref. 57, US Geological Survey.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nExtended Data Fig. 3 | Trends in development densities in CBRS treatment \nand control areas, compared across full and ZTRAX samples. Top panel shows \ntrends in built-up surface area in treatment and control areas in the full sample, \nwhere the left plot shows the percent of land area covered in built-up surface in \ntreatment (blue) and control (dashed black) areas every 10 years between 1960 \nand 2010 and the right plot shows the difference (treatment minus control). \nSame but in the subsample for of units with at least one ZTRAX transaction.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nExtended Data Fig. 4 | Geographic distribution of treatment and control \nunits. Panel a maps CBRS treatment units (blue) and constructed controls \n(green). All treatment units receive a weight of one. The shading of the control \n(green) points corresponds to the synthetic control weights, with lower weights \n(less relevant observations) indicated by lighter shades and higher weights (more \nrelevant observations) indicated by darker shades. Panel b shows the distribution \nof treatment and control units across states (left) and regions (right).\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nExtended Data Fig. 5 | Placebo test for the effect of CBRS designations on \ndevelopment densities. Left panel shows the trajectories of development densities \nin true treated areas (blue) and the synthetic control (dashed black). Right panel \nshows the difference in between treatment and control areas in blue. We then repeat \nour estimation procedure 100 times, but each time, we create a placebo treatment \ngroup by dropping the true CBRS units from the sample and randomly assigning \n50 control units to have treatment status. The difference between the placebo \ntreatment and controls are shown by black lines (one for each run).\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nExtended Data Fig. 6 | Average distance to the coast in spillovers of CBRS \nand controls. We calculated the weighted average distance to the coast for all \nproperties within the 2km spillover areas of CBRS lands and control areas. We \nthen categorized these average distances into deciles. We conducted weighted \naverage t-tests for the distance from the coast between CBRS areas and controls \nin each decile and found no statistically significant differences. This exercise \nsuggests that the lower flood damages observed in CBRS spillover areas as \ncompared to control spillover areas cannot be explained by differences in \ndistance to coast.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nExtended Data Table 1 | Covariate balance across CBRS and synthetic control in ZTRAX subsample\nWe assess the success of our procedure for identifying control areas by comparing the mean characteristics of CBRS units (column 1) and the synthetic control (column 2). Column 3 shows \nthe standardized mean difference (SMD) between treatment and control areas using synthetic control weights. This table is analogous to Table 1, but in the subsample of units that have at \nleast one ZTRAX transaction.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nExtended Data Table 2 | Characteristics of effective and noneffective CBRS units\nThis table compares the mean characteristics of effective (column 1, N = 23) and noneffective (column 2, N=19) CBRS units using individual treatment effects. We define effective units as those \nwith treatment effects, measured in relative terms, in the most negative quartile (greater than 67% reduction in development densities). We define noneffective units as those with a positive \ntreatment effect estimate. Column 4 shows the difference in means and column 3 reports the p-value on a two-sample, two-sided t-test of whether the difference in means is distinguishable \nfrom zero.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-02082-3\nExtended Data Table 3 | Average effect of CBRS designations on development densities, by state\nWe estimate individual treatment effects for the subsample of 90 CBRS designations for which it is possible to identify an individual synthetic controls. Column (1) shows the average \ntreatment effects by state. Column (2) shows the relative effect, calculated as the percent change from outcomes in the absence of treatment, as measured by the control group mean. \nColumn (3) shows the number of CBRS designations used to calculate the state-level averages.\n\n\n    Scientific Research Findings:", "answer": "We evaluate the impact of a policy that eliminated federal infrastructure investments, flood insurance and disaster assistance in high-risk coastal zones in the USA. We find that the policy lowered development densities (buildings per acre) by 83% in designated areas but raised them by 37% in neighbouring communities, suggesting that the approach can reallocate growth away from the most at-risk areas. The policy also generated co-benefits in neighbouring areas, increasing property values and providing flood protection services by conserving natural lands. Our results suggest that withdrawing development incentives in risky areas can reduce disaster exposure and damages, lower government expenditures and increase the property tax base. A similar approach could be effective in other high-risk areas such as inland floodplains or wildfire zones and could be replicated in other countries, with the caveat that our findings are based on a US federal policy focused on reducing coastal flood risk.", "id": 40}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 14 | August 2024 | 876\u2013882\n876\nnature climate change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nOpportunities to strengthen Africa\u2019s efforts \nto track national-level climate adaptation\nAndreea C. Nowak\u2009\n\u200a\u20091,2\u2009\n, Lucy Njuguna\u2009\n\u200a\u20093, Julian Ramirez-Villegas\u2009\n\u200a\u20091,2, \nPytrik Reidsma2,6, Krystal Crumpler\u2009\n\u200a\u20094 & Todd S. Rosenstock\u2009\n\u200a\u20095\nTracking progress towards the Global Goal on Adaptation requires \ndocumentation of countries\u2019 intentions, against which future progress \ncan be measured. The extent to which existing national policy documents \nprovide adequate baselines is unclear. We evaluated the adequacy \nof African Nationally Determined Contributions (NDCs) (N\u2009=\u200953) and \nNational Adaptation Plans (NAPs) (N\u2009=\u200915) against three criteria\u2014coverage, \nconsistency and robustness\u2014mapped to the adaptation cycle. Fifty-three \npercent of NAPs and 8% of NDCs cover all elements needed for providing \nsufficient baselines for tracking adaptation progress. Only 40% and 9% of the \nNAPs and NDCs, respectively, provide consistent links between climate risk \nassessment, planning, implementation and tracking. No document provided \nfully robust indicators to operationalize tracking. Notable efforts towards \nadequacy exist, especially in NAPs. The findings illustrate continental-scale \nadvances and shortcomings for tracking progress, and emphasize \nopportunities in upcoming NDC revisions and NAP processes to enhance \ntheir coverage, consistency and robustness for future adaptation tracking.\nTracking adaptation implementation and effectiveness is needed \nto enhance financial and technical support for climate action1\u20133 and \nmitigate risks of maladaptation4. The Paris Agreement established \nthe Global Goal on Adaptation (GGA), aiming to enhance resilience, \nincrease adaptative capacity and reduce vulnerability to climate \nchange5. It also introduced the Global Stocktake (GST) to assess pro-\ngress towards these objectives. The Glasgow\u2013Sharm el-Sheikh work \nprogramme (2022\u20132023) developed the GGA framework6. The first \nGST2, syntheses of adaptation science7 and government reports8\u201311 \nhighlight limited evidence and ability to document adaptation progress \nand called for continued development of methods to track progress. \nCapitalizing on this momentum and on the substantial work on adap-\ntation tracking principles and frameworks3,12\u201317, the UAE-Bel\u00e9m Work \nProgram, established by the United Nations Framework Convention \non Climate Change (UNFCCC) in 2023, seeks to define indicators for \nassessing the GGA and the targets agreed upon in the framework.\nCountries use various instruments to report planned adapta-\ntion targets, actions and support needed to the UNFCCC. Nationally \nDetermined Contributions (NDCs)\u2014pledges for national climate \naction\u2014are political documents that outline country priorities, needs \nand committements18. National Adaptation Plans (NAPs) provide \ndetails on implementation including goals, objectives and actions, \nand support the operationalization of the adaptation components out-\nlined in NDCs9,19,20. NAPs are increasingly accompanied by monitoring \nand evaluation (M&E) systems to track implementation9. Adaptation \nprogress can then be intentionally reported through National Com-\nmunications, Adaptation Communications (Adcoms) and/or Biennial \nTransparency Reports (BTRs)18. While the first BTRs are due at the end \nof 2024, stand-alone Adcoms deliver partial information on countries\u2019 \nadaptation actions and programmes but less on their results8.\nTracking adaptation requires setting a baseline of countries\u2019 com-\nmitments, against which progress can be measured13,14. However, it \nis unclear to what extent NDCs and NAPs can deliver this1. Existing \napproaches for assessing adaptation plans offer insights into their \nadequacy for providing meaningful baselines of intentions. Plan \nquality and adaptation tracking studies highlight the importance of \nReceived: 16 November 2023\nAccepted: 3 June 2024\nPublished online: 19 July 2024\n Check for updates\n1Climate Action, Bioversity International, Rome, Italy. 2Plant Production Systems Group, Wageningen University and Research, Wageningen,  \nthe Netherlands. 3Climate Action, International Center for Tropical Agriculture (CIAT), Nairobi, Kenya. 4Food and Agriculture Organization of the  \nUnited Nations, Rome, Italy. 5Climate Action, Bioversity International, Montpellier, France. 6Deceased: Pytrik Reidsma. \n\u2009e-mail: a.nowak@cgiar.org\n\nNature Climate Change | Volume 14 | August 2024 | 876\u2013882\n877\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nWe use a broad definition of adequacy to refer to the inclusion of a set \nof elements and characteristics needed for enabling adaptation track-\ning. This is distinct from the strict IPCC definition, which refers to the \nsufficiency of adaptation solutions to avoid or manage risks31, or the \nGST definition of adequacy, which refers to policies\u2019 ability to capture \nnational needs, considering the nature and severity of risks17. We focus \non Africa because of its general high levels of vulnerability to climate \nchange and the urgent need to adapt32. We apply three criteria derived \nfrom the plan quality and adaptation tracking literature: coverage, \nconsistency and robustness, organized around the adaptation cycle \n(Box 1, Fig. 1 and Methods). We apply a pragmatic 0\u20131 scoring system \nto assess performance of the documents against each criterion and \ncompute an adequacy index on the basis of equally weighted aggrega-\ntion of criteria scores. The criteria allow us to: (1) inventory countries\u2019 \nintentions, establishing a baseline for future assessments of progress \nand (2) identify opportunities for enhancing second-generation NDCs \nand improving NAPs to enable future adaptation tracking efforts.\nResults\nWe developed a protocol (Supplementary Tables 2\u20134) and reviewed \n53 NDCs and 15 NAPs available as of 30 September 2022 from which \nwe extracted more than 7,000 data points on countries\u2019 intentions \n\u2018the adaptation cycle\u201915,21\u201324. The adaptation cycle sets out four ideal \nstages of adaptation including: (1) climate risk and impact assessment,  \n(2) planning, (3) implementation and (4) monitoring, evaluation and \nlearning (MEL). The cycle underpins adaptation planning and tracking \nsupport under the UNFCCC, particularly in the context of NAPs, the \nGGA and the GST2,25,26.\nEffective plans follow the steps of the adaptation cycle, iterate as \nnew information becomes available and facilitate implementation by \nensuring meaningfulness of adaptation efforts to context8,21,27. Plan \nquality is determined by the depth and breadth of its contents, which \nshould cover key adaptation cycle elements and ensure consistency \nacross them21,28\u201330, including providing a robust basis for tracking its \nimplementation9. Key factors found to affect plan quality include policy \nand economic aspects (for example, ability to prioritize action, fund-\ning), technical capacity (for example, risk assessments, existence of \nM&E systems) and legitimacy (for example, stakeholder integration)27. \nWith few exceptions, plan quality characteristics described above \nhave been used in qualitative case studies within developed or urban \ncontexts, with limited proof of scalability for large-scale assessments \nto identify broad trends.\nHere we evaluate African NDC and NAPs to determine their  \nadequacy for informing adaptation tracking (Supplementary Table 1). \nBox 1\nConcepts and criteria for assessing policy adequacy for enabling \nadaptation tracking\nAdaptation cycle. Describes an iterative adaptation process used in \npolicies and projects. The framework used by the UNFCCC to support \nnational adaptation processes25,26 and the GGA dimensional targets \nagreed upon at COP28 (ref. 6) includes four key components: climate \nrisk and impact assessment; adaptation planning; implementation; \nand monitoring, evaluation and learning (MEL).\nApplication of the cycle. Policy documents describing intentions, \nsuch as NDCs and NAPs, do not provide information on implemented \nadaptation, but rather on planned goals and objectives and priority \nactions for implementation. Hence, we adjust the second, third and \nfourth components of the cycle to the context of NDCs and NAPs, \nas described below. Specifically, planning (Component 2) refers to \ndesired effects of adaptation; actions (Component 3) refer to planned \nrather than implemented actions, which cover measures, projects \nand programmes (hereafter \u2018actions\u2019) and MEL (Component 4) refers \nto indicators for tracking adaptation.\nComponent 1: Climate risk and impact assessment. We refer to the \nassessment of \u2018hazards and systems at risk\u2019 or impacted by climate \nchange, which provides the rationale for adaptation.\nComponent 2: Planning. We refer to the adaptation \u2018goals\u2019 and \n\u2018objectives\u2019 that countries plan for. These describe the country\u2019s \nadaptation vision and desired outcomes.\nComponent 3: Implementation. We refer to the \u2018actions\u2019 or measures \nto be implemented, including projects, programmes and initiatives \naimed to respond to risks and to achieve objectives and goals.\nComponent 4: MEL. We refer to \u2018indicators\u2019 embedded in existing or \nforthcoming (that is, in development) adaptation MEL systems, as \ntools to support assessment of implementation and effectiveness \nof actions. We limit the scope to indicators to inform the UAE-Bel\u00e9m \nWork Program and technical resources of M&E systems, rather than \nfocusing on broader organizational structures, legal frameworks, \ngovernance, and human, financial and technical resources9,40. Hence, \nwe do not address the aspect of \u2018learning\u2019 in MEL.\nAdequacy. In our assessment, adequacy describes the extent to \nwhich NDCs and NAPs can enable a basis for adaptation tracking. \nAdequacy is determined by aggregating insights (or scores) on the \ncoverage, consistency and robustness of documents, as described \nbelow and in Methods.\nCriterion 1: Coverage. Refers to the inclusion of a set of elements \nlinked to the adaptation cycle components15. These elements refer to: \n\u2018climate hazards\u2019 and \u2018systems at risk\u201941 underlying the \u2018climate risk and \nimpact assessment\u2019 component; \u2018goals\u2019 and \u2018objectives\u2019, describing \nthe adaptation vision15,29 and underlying the \u2018planning\u2019 component; \n\u2018actions\u2019, describing strategies and/or measures to respond to risks10,15 \nand referring to the \u2018implementation\u2019 component; and \u2018indicators\u2019, \nrepresenting tools to measure adaptation36,42 and underlying the MEL \ncomponent.\nCriterion 2: Consistency. Describes the internal coherence of policies, \nmanifested through intentional linkages established between the key \ncomponents of the adaptation cycle, namely \u2018climate risk and impact \nassessment, planning, implementation and MEL\u201921,28,30,43. Consistency \nenables context-fit tracking and meaningful insights on progress and \neffectiveness.\nCriterion 3: Robustness. Describes the quality of indicator sets included. \nRobustness is determined by the inclusion of indicator sets that meet \nSMART+ criteria30,44\u201346: specific, measurable, with an assigned data \nsource, relevant, time-bound and with a measurable target, and that \nmeasure distinct aspects of adaptation across the cycle, such as climate \nparameters30, which determine the nature and extent of adaptation \nactions required; inputs needed for adaptation efforts36,44,47; adaptation \nprocesses46,47; outputs of actions36,44,46,47; or outcomes36,44,47,48.\n\nNature Climate Change | Volume 14 | August 2024 | 876\u2013882\n878\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nfor adaptation. These include information on climate risk and impact \nassessment, planning, implementation and MEL (Box 1). Despite vari-\nation among countries and documents, our results show that most \nNAPs and NDCs provide only a fraction of the information required to \nenable adaptation tracking. Examination of the three criteria\u2014cover-\nage, consistency and robustness\u2014identified pervasive gaps and oppor-\ntunities for informing ongoing discussions on the GGA framework and \nopportunities to enhance depth and breadth of future NDCs and NAPs.\nAdequacy\nThe adequacy of NDCs and NAPs for informing adaptation tracking \nvaried greatly among countries (Fig. 2a\u2013c and Extended Data Table 1). \nGenerally, NDCs had lower adequacy scores (minimum\u2013maximum \nrange of 0\u20131) than NAPs (P\u2009<\u20090.001), except for the Democratic Republic \nof the Congo (NDC\u2009=\u20090.8, NAP\u2009=\u20090.7) and Sierra Leone (0.9, 0.4), probably \nbecause these two NAPs were in the initial stages of development at \nthe time of the analysis (Extended Data Tables 2 and 3). NDC adequacy \nscores ranged between 0.2 and 0.9 (median\u2009=\u20090.39, s.d.\u2009=\u20090.19). Only 11% \n(6) of NDCs had scores in the upper quartile (>0.75), including NDCs \nfrom Angola, the Democratic Republic of the Congo, Ethiopia, Sierra \nLeone, Burundi (2021) and Uganda (2022). Lower quartile scores (<0.25) \nwere first submissions (between 2016\u20132018) (n\u2009=\u20094), updated first sub-\nmissions (2016\u20132021) (n\u2009=\u20094) and second submissions (2021\u20132022) \n(n\u2009=\u20092). Low-income countries (Methods), representing 38% of the \ncountries in the dataset, had higher scores for their NDCs compared \nwith countries in middle- and high-income groups together (P\u2009<\u20090.05, \nN\u2009=\u200953) (Extended Data Tables 1 and 3). Neither adaptation funding nor \ngovernance influenced NDC adequacy (P\u2009>\u20090.05).\nFifteen (15) countries had both an NDC and a NAP at the time of \nour analysis. Sixty-seven percent of the NAPs (8) had adequacy scores \nof more than 0.75. High-scoring NAPs were submitted in or after 2021, \nexcept for Burkina Faso and Cameroon NAPs which were submitted \nin 2015 and Ethiopia\u2019s NAP submitted in 2019. In some cases, despite a \ntime lag between NAP and NDC releases (2016 and 2022, respectively, \nin Sudan, 2022 and 2021 in Chad, and 2022 in the Central African Repub-\nlic), adequacy scores remained low (0.4) across all policy documents of \nthese three countries. This potentially indicates limited complementa-\nrities or limited observable learning between preparations of the two \ndocuments. Adaptation funding levels did not affect NAP adequacy \nscores (P\u2009>\u20090.05), yet governance readiness\u2014an indicator of institutional \npreparedness\u2014had a significant positive effect (\u03b2\u2009=\u20091.5182, s.e.\u2009=\u20090.4921, \nP\u2009=\u20090.0104), explaining a significant proportion of the variance in the \nNAP adequacy score (F(1,11)\u2009=\u20099.52, adjusted R2\u2009=\u20090.415, P\u2009=\u20090.010).\nCoverage\nAll NDCs and NAPs included at least half of the six elements used to \nassess coverage (Fig. 2d). However, coverage was more complete in \nNAPs than in NDCs (P\u2009<\u20090.001, N\u2009=\u200915 and N\u2009=\u200953, respectively), with \nmean scores of 0.7 and 0.9, respectively (Extended Data Table 1). \nOnly four (4) or less than 10% of the NDCs, but more than half (8) of \nthe NAPs, included information on all six elements: Angola, Burundi, \nSierra Leone and Uganda (NDC); and Benin, Burkina Faso, Cameroon, \nEthiopia, Liberia, Madagascar, South Africa and Togo (NAP). Elements \nmost featured, in descending order, were climate hazards and systems \nat risk (all NDCs and NAPs), adaptation actions (95% and 100% of NDCs \nand NAPs, respectively), objectives (70%, 93%), goals (47%, 87%) and \nindicators (23%, 67%) (Fig. 3a,b). We provide detailed summaries of \nobservations by element type, country and document in Extended \nData Table 4.\nConsistency\nMost documents provided evidence of only half of the linkages defining \nconsistency (mean\u2009=\u20090.53, s.d.\u2009=\u20090.28). NAPs registered higher consist-\nency scores than NDCs (P\u2009<\u20090.001, N\u2009=\u200915 and N\u2009=\u200953, respectively) with \na narrower spread (Fig. 2e). Eleven documents were fully consistent, \nindicated by maximum consistency scores. These included 5 NDCs \n(less than 10%), that is, Burundi, the Democratic Republic of the Congo, \nGuinea, Sierra Leone and Uganda, and 6 NAPs (less than 50%), including \nBenin, Burkina Faso, Cameroon, Liberia, Madagascar and South Africa. \nMost often, countries link climate risk and impact assessment and \naction implementation (87% and 100% of NDCs and NAPs, respectively) \n(Fig. 4a,b). Less often, climate risk and impact assessment intentionally \nlink to planning (72%, 87%) or planning to implementation (68%, 87%). \nMEL is the least consistent component across the adaptation cycle, \npartly due to indicators being featured less overall (see \u2018Coverage\u2019 \nsection). Fewer documents provided linkages between climate risk \nand impact assessment and MEL (23%, 60%), suggesting a potential \ndisconnect between assessments of climate risks and measurements \nof the impacts of adaptation on risk reduction. Less frequent linkages \nwere also observed between planning and MEL (17%, 67%) and imple-\nmentation and MEL (15%, 40%), as few documents included indicators \nexplicitly linked to planned goals, objectives or actions (Extended \nData Table 5).\nRobustness\nTwenty-two documents (about two-thirds of our sample) featured indi-\ncators, that is, 10 NAPs and 12 NDCs (examples in Table 1). However, none \nCriterion 1: Coverage\nAre key elements of the adaptation \npolicy cycle covered, providing a \ncomprehensive baseline of adaptation \nintentions at national level?\nComponent 1:\nClimate risk\nand impacts\nassessment\nComponent 2:\nAdaptation\nplanning\nComponent 3:\nAdaptation\nimplementation\nComponent 4:\nMonitoring,\nevaluation and\nlearning\na\nb\nCriterion 2: Consistency\nAre adaptation cycle components\nconsistently linked to one another,\nproviding the basis for a context-fit\nadaptation tracking?\nCriterion 3: Robustness\nAre there sets of indicators that meet \nSMART+ criteria and serve diverse \nmonitoring functions, providing a \nrobust basis for operational tracking?\nSMART+ \ndesign\nDiverse\nfunctions\nI\nComponent 4\nComponent 2\nComponent 3\nComponent 1\nH\nS\nG\nI\nA\nO\nComponent 3\nComponent 1\nComponent 4\nComponent 2\nc\nFig. 1 | Conceptual framework for assessing adequacy of NDCs and NAPs. a\u2013c, The iterative adaptation policy cycle (a) applied to NDCs and NAPs, which describes \nkey foundational components of adaptation. This informs (b) criteria and questions for assessing policy adequacy, visualized in c.\n\nNature Climate Change | Volume 14 | August 2024 | 876\u2013882\n879\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nmet all characteristics of robustness (Fig. 2f). Overall, NDCs had lower \nrobustness scores than NAPs (P\u2009<\u20090.001, N\u2009=\u200912 and N\u2009=\u200910, respectively). \nOn average, NDCs met 2 out of 11 characteristics (s.d.\u2009=\u20092.9, N\u2009=\u200912) and \nNAPs met 5 (s.d.\u2009=\u20093.5, N\u2009=\u200910) (Extended Data Table 1). Often, gaps in \nrobustness were linked to indicators without assigned data sources \n(observed in 83% of NDCs and 79% of NAPs) or without timeframes \n(50%, 80%) (Extended Data Table 6). Nominally, the largest gap was \nobserved for indicators associated with climate parameters. However, \nmonitoring of climate parameters is undertaken on the basis of inter-\nnational scientific standards whose details are typically not included in \npolicy documents. SMART+ characteristics met, in descending order, \nrefer to relevance (R) to context (all NDCs and NAPs), specificity (S) \nand measurability (M) (92%, 100%) and targets (83%, 90%). M&E func-\ntion characteristics most common were, in descending order, outputs \nmeasurement (83% and 100% of the NDCs and NAPs, respectively), \noutcomes (75% and 80%), processes and inputs (58%, 100%). NDCs of \nEthiopia and Rwanda and Madagascar\u2019s NAP had highest robustness \nscores (0.8). In addition, we found 37 documents that set objective or \naction-level targets without identifying indicators, indicating potential \nentry points for future indicator development (Extended Data Table 7).\nDiscussion\nOur analysis of African NDCs and NAPs indicates that they generally lack \nsufficient information to enable adaptation tracking. The core issue \nis their partial coverage of the adaptation cycle and the inconsistency \namong components, leading to an incomplete and at times unclear \narticulation of what needs to be tracked and how. Of particular concern, \nrelatively few documents specify indicators for tracking adaptation \n(23% and 67% of the NDCs and NAPs, respectively). Even in cases where \nindicators have been identified, shortfalls in quality call into question \ntheir utility in enabling meaningful insights into adaptation. These \nresults underscore the challenges that African governments face in \nassessing and reporting national progress on adaptation, and reveal \nspecific opportunities for African governments to target in the near \nterm, as they develop adaptation plans, revise NDCs and define the \nspecificities of the GGA framework.\nWe found that NAPs provide a more adequate basis for adaptation \ntracking than NDCs (Fig. 2, and Extended Data Tables 1 and 3). The rela-\ntive adequacy of NAPs is anticipated as they tend to be comprehensive \nand operational9, often the result of multiyear, multistakeholder pro-\ncesses embedded in domestic policies33, and are backed by substantial \ntechnical and financial support9,34. The gap in adequacy between NAPs \nand NDCs underscores the crucial role of NAPs in operationalizing \nadaptation components of NDCs, including articulating detailed plans \nfor assessing progress9,19. However, only a fraction of African countries \nhave formulated NAPs so far. The target towards ensuring that all coun-\ntries have NAPs by 2030, established in the GGA framework6, provides \na welcome political imperative for focusing on NAPs development and \nimplementation as a pathway towards improving adaptation tracking \ninfrastructure.\nAccelerating this process for African countries represents a criti-\ncal opportunity in the coming years. Enabling cross- and intra-country \nlearning would be one approach to catalyse enhancement. While net-\nworks such as the Adaptation Forum can provide a space for knowledge \nexchange, concrete learning modalities remain to be determined. \nOur dataset indicates countries that have more adequate NAPs and \ncountries with high scores on individual criteria. These can serve as \nnoteworthy exemplars. We were also able to identify countries with \nhigh-scoring NDCs and no published NAPs, which illustrate the poten-\ntial of the former to advance NAP planning processes, in addition \nto their observed political and commitment functions11,18. While the \npotential for NDCs to provide adequate reference points for tracking \n0\n0.2\n0.4\n0.6\n0.8\n1.0\na\n0\n0.2\n0.4\n0.6\n0.8\n1.0\nb\nc\nd\ne\nf\n0\n0.2\n0.4\n0.6\n0.8\n1.0\nScore\nNDC\nNAP\nFig. 2 | Adequacy of African NDCs and NAPs for informing adaptation \ntracking. a,b, Adequacy scores for NDCs (N\u2009=\u200953) (a) and NAPs (N\u2009=\u200915) (b).  \nc\u2013f, Summary of scores for adequacy (c) and by criteria of coverage (d), \nconsistency (e) and robustness (f). Adequacy scores represent aggregated \nweighted scores across coverage, consistency and robustness criteria. Each violin \nrepresents the density of scores (y axis) within a criteria group and document \ntype (x axis), with wider areas indicating higher density (that is, higher number \nof documents with that score). The violin plot is overlaid with a swarm plot, \nwhere individual data points represent a document and are separated randomly \n(jittered) to reduce overplotting and enhance visualization of data distribution. \nScores for each criterion are scaled to 0\u20131 for comparison (Methods). Results \nfrom two-sided Wilcoxon rank-sum tests with continuity correction show \nsignificant differences in adequacy, coverage (P\u2009<\u20090.001), consistency and \nrobustness (P\u2009<\u20090.005) scores between NDCs and NAPs, with NDC scores being \nsignificantly lower than NAP scores. We provide detailed results of statistical \ntests in Extended Data Table 3.\n\nNature Climate Change | Volume 14 | August 2024 | 876\u2013882\n880\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nis limited to a few African countries, this positive deviance presents \na compelling opportunity for further exploration and exploitation, \nas countries progress towards developing second-generation NDCs \nby 2025. Moreover, existing voluntary guidelines on communicating \nadaptation information under the UNFCCC35 can inform NAP and NDC \nalignment efforts, deemed pivotal for consistent national planning.\nThe first GST2 and recent research suggest that climate plans \nimprove over time9. Updated NDCs have been enhanced to include addi-\ntional adaptation targets and indicators compared with their original \nversions. Similarly, NAPs have increasingly become more consistent \nand operational. In our continental analysis, however, the evidence was \nvaried. Updated NDCs tended to have higher adequacy scores than the \nfirst submissions, which were mainly rebranded Intended Nationally \nDetermined Contributions (INDCs) developed before the enactment \nof the Paris Agreement11; yet the improvements, with scores ranging \nbetween 0.2 and 0.4, still miss many of the critical aspects for effec-\ntive tracking. Moreover, the evaluation of NAPs shows that those with \ntop quartile scores were released between 2015 and 2022, suggesting \nlittle advancement over time at the continental level. It is critical to \nnote that these results do not reflect an individual country\u2019s progress \nover time. An important limitation of our approach was the inability \nto track document content over time, as earlier versions of NDCs were \nnot available in the UNFCCC Registry. This underscores an opportunity \nfor enhanced transparency and systematic archiving of NDCs and \nNAPs to effectively monitor and assess the evolution of adaptation \nambitions and baselines for adaptation tracking. Moreover, given  \nthe stark variation in the level of detail captured in the NDC updates,  \nthe second-generation NDCs due in 2025 provide an opportune  \njuncture for taking stock and drawing lessons from NDC evolution.\nThe development of more robust indicators represents another \nopportunity to strengthen the basis of tracking. Activity-based indi-\ncators, which represent 84% of all indicators mapped, are needed for \nunderstanding progress in implementation; however, complementary \noutcome indicators are needed to track effectiveness36,37 and facilitate \nenhanced result reporting in future Adcom8. Further development of \neffectiveness indicators can be inspired by existing examples in NDCs \nand NAPs (Table 1) but also by existing objective-level targets (Extended \nData Table 7), which can facilitate the identification of outcome-based \nindicators that align with established national priorities. Similarly, the \nadaptation cycle framework highlights the importance of identifying \nand tracking climate risk indicators to determine whether actions \neffectively reduce climatic risks. While such indicators are probably \nembedded in specialized institutional structures for weather and \nclimate observation, they need to be acknowledged and integrated \nin planning processes to ensure consistency with national priorities \nand enable effective adaptation tracking. Lastly, improvements in \nindicator quality along the SMART+ criteria, particularly specificity, \ndata sources, targets and timeframes, are warranted to ensure mean-\ningful and practical indicators. Our analysis indicates that countries \nseeking to operationalize tracking are typically already moving in  \nthe direction of contextually relevant and measurable indicators. \nThis sets a valuable precedent. The recent decision of the IPCC on the \nSeventh Assessment Cycle38 and the UAE-Bel\u00e9m Work Program offer \nmomentum for enhanced guidance on how governments can increase \ncoherence between nationally relevant indicators and the GGA.\nThis research applied a parsimonious approach to evaluating \nNDCs and NAPs adequacy for enabling adaptation tracking across all \nAfrican countries. However, coverage, consistency and robustness \nrepresent only a fraction of the many facets of plan quality, and the \ndata used to consider potential levers of change were coarse. Thus, \nwe were only able to uncover some potential features that determine \ndifferences in plan adequacy, such as governance aspects and income \nstatus. However, the approach allowed us to investigate how 53 African \ncountries intend to adapt and track efforts, in contrast to the many frag-\nmented cases usually investigated. Intercountry comparative studies \ncan help identify broader trends and new opportunities for accelerat-\ning support, which is typically not possible with deeper case studies on \nindividual countries, cities or landscapes. Furthermore, our approach \n0\n20\n40\n60\n80\n100\nHazards\nSystems at risk\nGoals\nObjectives\nActions\nIndicators\na\nHazards\nSystems at risk\nGoals\nObjectives\nActions\nIndicators\nDocuments (%)\nb\nFig. 3 | Extent of coverage of adaptation elements in policy documents.  \na,b, NDCs (N\u2009=\u200953) (a) and NAPs (N\u2009=\u200915) (b). Percentages calculated on the basis  \nof mentions of the six adaptation elements in each document.\na\nAssess\nPlan\nImplement\nMEL\nAssess\n72\n87\n23\nPlan\n72\n68\n17\nImplement\n87\n68\n15\nM&E\n23\n17\n15\nb\nAssess\nPlan\nImplement\nMEL\nAssess\n87\n100\n60\nPlan\n87\n87\n67\nImplement\n100\n87\n40\nM&E\n60\n67\n40\n15\n100\nDocuments (%)\nFig. 4 | Consistency across adaptation cycle components. a,b, NDCs (N\u2009=\u200953) (a) \nand NAPs (N\u2009=\u200915) (b). Values represent percentages of documents establishing \na linkage between the adaptation cycle components (abbreviated for figure \nformatting): assessment (referring to climate risk and impact assessment), \nplanning, implementation and MEL.\n\nNature Climate Change | Volume 14 | August 2024 | 876\u2013882\n881\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nonly looks at adaptation intentions as baselines for future tracking. \nIntentions and \u2018good\u2019 plans are insufficient; they are only a precondi-\ntion but not a guarantee for effectiveness9,21,27,39. This underscores the \ncritical role of examining complementary documents to understand \nadaptation progress and effectiveness, such as Adcoms and BTRs, as \nwell as collecting field data to assess whether and how intentions and \nplans translate into results on the ground.\nLeveraging existing planning processes and policies would spare \ngovernments from the burden of establishing new institutional struc-\ntures and data systems for the sole purpose of adaptation tracking. This \nwould also allow them to track and report on adaptation aspects that \nare already prioritized in national policies, thus enhancing linkages \nbetween adaptation tracking at national and global scales. Existing \nNAPs and NDCs reflect a pragmatic, early-stage approach to gov-\nernments\u2019 engagement with adaptation tracking; they initiate the \nconversation on what could be tracked and how. Our assessment dem-\nonstrates the opportunities to build on African NDCs and NAPs, with \nclear steps forward to enable future tracking and reporting processes.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-024-02054-7.\nReferences\n1.\t\nGarschagen, M. et al. in Climate Change 2022: Impacts, \nAdaptation and Vulnerability (eds P\u00f6rtner, H.-O. et al.) 2610\u20132613 \n(Cambridge Univ. Press, 2023).\n2.\t\nTechnical Dialogue of the First Global Stocktake. Synthesis \nReport by the Co-Facilitators on the Technical Dialogue (UNFCCC \nSecretariat, 2023).\n3.\t\nFisher, S. Much ado about nothing? Why adaptation measurement \nmatters. Clim. Dev. 16, 161\u2013167 (2023).\n4.\t\nReckien, D. et al. Navigating the continuum between adaptation \nand maladaptation. Nat. Clim. Change 13, 907\u2013918 (2023).\n5.\t\nThe Paris Agreement (United Nations, 2015).\n6.\t\nUNFCCC. in Report of the Conference of the Parties serving \nas the meeting of the Parties to the Paris Agreement on its fifth \nsession, held in the United Arab Emirates from 30 November to 13 \nDecember 2023 22\u201328 (United Nations, 2024).\n7.\t\nBerrang-Ford, L. et al. A systematic global stocktake of evidence \non human adaptation to climate change. Nat. Clim. Change 11, \n989\u20131000 (2021).\n8.\t\nAdaptation Gap Report 2023: Underfinanced. Underprepared. \nInadequate Investment and Planning on Climate Adaptation  \nLeaves World Exposed (United Nations Environment Programme, \n2023).\n9.\t\nLeiter, T. Do governments track the implementation of national \nclimate change adaptation plans? An evidence-based global \nstocktake of monitoring and evaluation systems. Environ. Sci. \nPolicy 125, 179\u2013188 (2021).\n10.\t Lesnikowski, A., Ford, J., Biesbroek, R., Berrang-Ford, L. & \nHeymann, S. J. National-level progress on adaptation. Nat. Clim. \nChange 6, 261\u2013264 (2016).\n11.\t\nLeinaweaver, J. & Thomson, R. The elusive governance of climate \nchange: nationally determined contributions as commitments \nand negotiating positions. Glob. Environ. Polit. 21, 73\u201398 (2021).\n12.\t Adaptation Committee. Approaches to Reviewing the Overall \nProgress Made in Achieving the Global Goal on Adaptation \n(UNFCCC, 2021).\n13.\t Ford, J. D., Berrang-Ford, L., Lesnikowski, A., Barrera, M. & \nHeymann, S. J. How to track adaptation to climate change: a \ntypology of approaches for national-level application. Ecol. Soc. \nhttps://www.jstor.org/stable/26269369 (2013).\n14.\t Ford, J. D. & Berrang-Ford, L. The 4Cs of adaptation tracking: \nconsistency, comparability, comprehensiveness, coherency. \nMitig. Adapt. Strateg. Glob. Change 21, 839\u2013859 (2016).\n15.\t Berrang-Ford, L. et al. Tracking global climate change adaptation \namong governments. Nat. Clim. Change 9, 440\u2013449 (2019).\n16.\t Singh, C. et al. Interrogating \u2018effectiveness\u2019 in climate change \nadaptation: 11 guiding principles for adaptation research and \npractice. Clim. Dev. 14, 650\u2013664 (2022).\n17.\t Magnan, A. K. et al. Status of global coastal adaptation. Nat. Clim. \nChange 13, 1213\u20131221 (2023).\n18.\t Leiter, T. Nationally determined contributions (NDCs) as a \ngovernance instrument \u2013 accounting for politics, negotiation \nprogress, and related mechanisms under the Paris Agreement. \nEnviron. Polit. 33, 552\u2013557 (2024).\nTable 1 | Robustness of adaptation indicators\nIndicator examples by main function\nCriteria\nS\nM\nA\nR\nT\n+\nClimate parameter\n Change in annual temperaturea\nx\nx\nx\n Extreme precipitation eventsa\nx\nx\nx\n Number of hot daysa\nx\nx\nx\nInput\n \u0007Number of gender-sensitive \ninstitutional analyses carried outb\nx\nx\nx\nx\n \u0007Number of value-chain climate risk \nassessments completedc\nx\nx\nx\n \u0007Accurate data on exposure to climate in \nhigh-risk areas by 2030d\nx\nx\nx\nx\nProcess\n \u0007Detailed spatial plans for all districts by \n2025 and 2030d\nx\nx\nx\nx\n \u0007Number of SNRM plans integrating \nadaptation (\u2026)c\nx\nx\nx\n \u0007Number of development plans \nintegrating adaptatione\nx\nx\nx\nx\nx\nOutput\n Area (ha) under irrigation by 2030f\nx\nx\nx\nx\nx\n \u0007Number of climate-resilient crop \nvarieties developedd\nx\nx\nx\nx\n \u0007Rate of women farmers benefiting from \ntechnical (\u2026) support (%)g\nx\nx\nx\nx\nOutcome\n \u0007Rate of recovery and restoration of the \nfertility of degraded soilsh\nx\nx\nx\n \u0007Percentage reduction of crop and \nanimal disease cases (30% reduction \nby 2030)i\nx\nx\nx\nx\nx\n Increase in yield per hectare (%)c\nx\nx\nx\nExamples extracted from a subset of agriculture indicators, organized by their monitoring \nand evaluation function (that is, climate parameter, input, process, output and outcome) \nand assessed by the SMART+ characteristics: specific (S), measurable (M), assigned Data \n(A), relevant (R), and time-bound (T) and target (+). One indicator can simultaneously meet \nseveral SMART+ characteristics but only one function. Climate parameter indicators are \nsourced from the entire set of indicators, as none of the agriculture sector ones included \nclimate parameters. \u2018x\u2019 indicates that the indicator meets the respective characteristic. A full \nassessment of each indicator in the dataset is available in Supplementary Table 5. aAngola \nNDC. bDemocratic Republic of Congo NAP. cEthiopia NAP. dRwanda NDC. eDemocratic \nRepublic of Congo NDC. fUganda NDC. gBenin NDC. hBurkina Faso NAP. iEthiopia NDC. SNRM, \nsustainable management of natural resources (the acronym was not spelled out in the \nformulation of the indicator, but in the body of the document).\n\nNature Climate Change | Volume 14 | August 2024 | 876\u2013882\n882\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\n19.\t Smithers, R. et al. The Role of the NAP Process in Translating NDC \nAdaptation Goals into Action: Linking NAP Processes and NDCs \n(Deutsche Gesellschaft f\u00fcr Internationale Zusammenarbeit (GIZ), \n2017).\n20.\t National Adaptation Plans. Technical Guidelines for the National \nAdaptation Plan Process (UNFCCC, 2012).\n21.\t Reckien, D. et al. Quality of urban climate adaptation plans over \ntime. NPJ Urban Sustain. 3, 13 (2023).\n22.\t Swart, R. et al. Reframing climate services to support municipal \nand regional planning. Clim. Serv. 22, 100227 (2021).\n23.\t Park, S. E. et al. Informing adaptation responses to climate \nchange through theories of transformation. Glob. Environ. Change \n22, 115\u2013126 (2012).\n24.\t Scott, H. & Moloney, S. Completing the climate change \nadaptation planning cycle: monitoring and evaluation by local \ngovernment in Australia. J. Environ. Plan. Manage. 65, 650\u2013674 \n(2022).\n25.\t Adaptation Committee. 25 Years of Adaptation Under the UNFCCC \n(UNFCCC, 2019).\n26.\t Adaptation Committee. Enhancing Coherent Action on Adaptation \n2012\u20132015. 2015 Overview Report (UNFCCC, 2015).\n27.\t Olazabal, M., Galarraga, I., Ford, J., Sainz De Murieta, E. & \nLesnikowski, A. Are local climate adaptation policies credible?  \nA conceptual and operational assessment framework. Int. J. \nUrban Sustain. Dev. 11, 277\u2013296 (2019).\n28.\t Dupuis, J. & Biesbroek, R. Comparing apples and oranges: the \ndependent variable problem in comparing and evaluating climate \nchange adaptation policies. Glob. Environ. Change 23, 1476\u20131487 \n(2013).\n29.\t Woodruff, S. C. & Stults, M. Numerous strategies but limited \nimplementation guidance in US local adaptation plans. Nat. Clim. \nChange 6, 796\u2013802 (2016).\n30.\t Goonesekera, S. M. & Olazabal, M. Climate adaptation indicators \nand metrics: state of local policy practice. Ecol. Indic. 145, 109657 \n(2022).\n31.\t Begum, A. et al. in Climate Change 2022: Impacts, Adaptation and \nVulnerability (eds P\u00f6rtner, H.-O. et al.) 121\u2013196 (Cambridge Univ. \nPress, 2023).\n32.\t Trisos, C. H. et al. in Climate Change 2022: Impacts, Adaptation \nand Vulnerability (eds P\u00f6rtner, H.-O. et al.) 1285\u20131455 (Cambridge \nUniv. Press, 2023).\n33.\t Leiter, T. Too little, too slow? Climate adaptation at the United \nNations climate change negotiations since the adoption of the \nParis Agreement. Carbon Clim. Law Rev. 16, 243\u2013258 (2022).\n34.\t In-country NAP Support Programs (NAP Global Network, 2024).\n35.\t Adaptation Committee. Draft Supplementary Guidance for \nVoluntary Use by Parties in Communicating Information in \nAccordance with the Possible Elements of an Adaptation \nCommunication (UNFCCC, 2022).\n36.\t Leiter, T. et al. Adaptation Metrics. Current Landscape and Evolving \nPractices (Global Commission on Adaptation, 2019).\n37.\t Craft, B. & Fisher, S. Measuring the adaptation goal in the global \nstocktake of the Paris Agreement. Clim. Policy 18, 1203\u20131209 \n(2018).\n38.\t Planning for the Seventh Assessment Cycle. Options for the \nProgramme of Work in the Seventh Assessment Cycle (IPCC, 2024).\n39.\t Berke, P. & Godschalk, D. Searching for the Good Plan. A meta- \nanalysis of plan quality studies. J. Plan. Lit. 23, 227\u2013240 (2009).\n40.\t Njuguna, L., Biesbroek, R., Crane, T. A., Dewulf, A. & Tam\u00e1s, P. Do \ngovernment knowledge production and use systems matter for \nglobal climate change adaptation tracking? Insights from Eastern \nAfrica. Reg. Environ. Change 23, 85 (2023).\n41.\t IPCC Climate Change 2022: Impacts, Adaptation and Vulnerability \n(eds M\u00f6ller, V. R. et al.) 2897\u20132930 (Cambridge Univ. Press, 2022).\n42.\t Leiter, T. & Pringle, P. in Adaptation Metrics: Perspectives on \nMeasuring, Aggregating and Comparing Adaptation Results  \n(eds Christiansen, L. et al.) 29\u201347 (UNEP DTU Partnership, 2018).\n43.\t Dekens, J. Using Climate Risk Assessment to Measure Adaptation \nSuccess at the National Level: Preliminary Lessons from 12 \nCountries (NAP Global Network, 2024).\n44.\t Hale, T. N. et al. Sub- and non-state climate action: a framework \nto assess progress, implementation and impact. Clim. Policy 21, \n406\u2013420 (2021).\n45.\t Leagnavar, P., Bours, D. & McGinn, C. Good Practice Study on \nPrinciples for Indicator Development, Selection, and Use in \nClimate Change Adaptation Monitoring and Evaluation (Global \nEnvironment Facility, 2015).\n46.\t Klostermann, J. et al. Towards a framework to assess, compare \nand develop monitoring and evaluation of climate change \nadaptation in Europe. Mitig. Adapt. Strateg. Glob. Change 23, \n187\u2013209 (2018).\n47.\t Makinen, K. et al. Indicators for Adaptation to Climate Change \nat National Level \u2013 Lessons from Emerging Practice in Europe. \nTechnical Paper 2018/3 (European Topic Centre on Climate \nChange impacts, Vulnerability and Adaptation (ETC/CCA), 2018).\n48.\t Donatti, C. I., Harvey, C. A., Hole, D., Panfil, S. N. & Schurman, H. \nIndicators to measure the climate change adaptation outcomes \nof ecosystem-based adaptation. Clim. Change 158, 413\u2013433 \n(2020).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this licence, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2024\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nMethods\nOur analysis of NDCs and NAPs focuses on the potential of national \nplanning and international reporting documents to provide a basis \nfor adaptation tracking. We explored three questions: (1) Do NDCs and \nNAPs provide information covering the adaptation cycle to enable a \ncomprehensive understanding of adaptation intentions? (2) Do they \nprovide consistent information across the adaptation cycle to facilitate \ncontext-meaningful tracking? (3) Do they provide robust indicators to \nensure operational tracking? These questions guided data extraction \nand analysis. The resulting dataset represents the most comprehensive \noverview available so far on African countries\u2019 intentions for adapta-\ntion. We developed a detailed protocol available at protocolexchange \n(https://doi.org/10.21203/rs.3.pex-2399/v1)49, informed by the Global \nAdaptation Mapping Initiative (GAMI) protocols7 (see Data collection \nand Data analysis below).\nAnalytical framework\nWe used the adaptation cycle as the guiding framework to assess \nadequacy of NAPs and NDCs (Fig. 1 and Box 1). Each component of \nthe cycle provides insights for understanding progress on adapta-\ntion2. Interpretations of the cycle vary across the literature, involving \nbetween four and eight components covering different aspects of \nadaptation progress. For example, ref. 23 used four cycle components: \nproblem, adaptation vision, implementation and MEL, to assess drivers \nof incremental and transformative change in Australia\u2019s wine sector. \nAn assessment of European urban plans determined plan quality using \nfive components: impact, risk and vulnerability assessments, goals, \nmeasures, implementation tools and processes, and M&E21. Others \nused a six-phase interpretation of the cycle: groundwork preparations, \nrisk assessment, option identification, assessment, implementation \nand M&E, to identify opportunities for enhancing climate services \nin urban and peri-urban Europe22. An assessment of local adaptation \nplans used an eight-component cycle to describe local M&E systems: \ngroundwork, assessments of current and future situation, objectives, \nstrategies, option assessment, prioritization, implementation and \nM&E24. All these approaches emphasize a logical sequence and link-\nages between the components and a cyclical sequence from climate \nrisk assessments to M&E, and aim to provide an ideal framework for \ncomprehensive policies21. We chose the four-component adaptation \ncycle for its practicality and alignment with the UNFCCC framework \nto guide adaptation support2,25,26 as well as GGA and GST processes. \nBecause NAPs and NDCs outline plans for adaptation, we adjusted defi-\nnitions of the cycle components to match information types delivered \nby such documents (Box 1).\nWe established three criteria to assess adequacy, understood \nas the potential of NDCs and NAPs to enable the basis for adaptation \ntracking: coverage, consistency and robustness (Fig. 1). The criteria, \ndrawn from the plan quality and adaptation tracking scholarship, are \nnot exhaustive. Instead, they are designed to: (1) bridge theoretical \ndiscussions with practical application for a continental-scale analysis, \n(2) deliver insights into the breadth (that is, coverage) and quality (that \nis, consistency, robustness) of the policies\u2019 content and (3) provide a \nbaseline for how adaptation is planned, aimed at being implemented \nand tracked across countries.\nCoverage. This criterion describes inclusion of key elements of the \nadaptation cycle, which describe the \u2018why\u2019, \u2018what\u2019, \u2018how\u2019 and \u2018so what\u2019 \nof adaptation. This approach builds on an existing framework to track \nadaptation among governments15, which suggests that a comprehen-\nsive understanding of adaptation progress rests on an assessment \nof the vulnerability context, goals, actions and results. In our assess-\nment, policies with adequate coverage include information on six  \ncore elements mapped to the adaptation cycle (Supplementary \nTable 2). Hazards, systems at risk, goals, objectives and actions pro-\nvide context to tracking and inform the development of adequate and \nmeaningful indicators30. Together, these six elements deliver a baseline \nunderstanding of intentions for adaptation.\nConsistency. This criterion looks at alignment between adaptation \ncycle components. Policies need to establish clear linkages between \nclimate risk and impact assessments, planning of goals and objectives, \nactions to implement, and MEL, to facilitate context-fit tracking and \nmeaningful insights on progress21,30. Intentional linkages maximize \neffectiveness of plans in reducing risks and vulnerabilities28. Our assess-\nment of consistency was pragmatic and did not examine the extent \nto which sets of actions sufficiently address risks. A document was \nconsidered fully consistent if it provided evidence of intentional link-\nages across all four components of the adaptation cycle. For instance, \nactions to switch to drought-resistant crops or strengthen early warn-\ning systems for hydroclimatic risks in coastal areas indicate linkages \nbetween assessment and planning (Supplementary Table 3). While such \nan approach is inherently subjective, it provides a practical, preliminary \napproximation of how policies can deliver meaningful information \nfor tracking.\nRobustness. This criterion focuses on the quality, rather than the \nsheer presence, of sets of adaptation indicators as an entry point for \noperational tracking9,30,46. First, robustness refers to the SMART design \ncharacteristics45,46, which allow determination of whether indicators \nare specific (S), measurable (M), with an assigned data source (A), rele\u00ad\nvant (R) for the adaptation context, time-bound (T) and with a target \n(+). The target characteristic is not typically included in SMART assess-\nments, yet it has been acknowledged as critical for gauging benchmarks \nfor progress in implementation and effectiveness30,44. Considering the \narray of best practices for designing performance indicators45, our \nchoice of SMART+ characteristics is pragmatic; these are well known \nand widely used in the development and adaptation community.\nSecond, robustness refers to the \u2018function\u2019 of the set of indica-\ntors featured in the document as entry points for capturing progress \nacross the adaptation cycle. We distinguished between indicators \nused for measuring: climate parameters, which define a baseline of \nclimate conditions and inform the actions required to address risks30,47; \ninputs, such as human, financial or technical resources required for \ndesigning and implementing adaptation actions36,44,47; process, which \nindicate progress in designing adaptation policy processes and insti-\ntutions46,47; outputs, referring to products or immediate results from \nactivities36,44,46,47; and outcomes, describing effects of adaptation, \nincluding changes in behaviour, environmental, social and/or eco-\nnomic conditions36,44,47,48. One indicator can only have one of these five \nfunctions; hence, documents cover this function diversity through the \nindicator sets they propose. This approach to robustness, while not \nexhaustive, allowed us to assess the quality (or depth) of information, \nwhich can enable effective tracking. This is particularly relevant in the \ncontext of the newly established UAE-Bel\u00e9m Work Program, which \nseeks to provide guidance on identifying and measuring indicators.\nData sources\nWe assessed countries\u2019 adaptation intentions and therefore focused on \nNAPs and NDCs. This allowed us to explore whether and how existing \nplans and international reporting efforts can offer an adequate basis \nfor tracking. We reviewed all NAPs (N\u2009=\u200915) and NDCs with an adapta-\ntion component (N\u2009=\u200953) available on NAP Central (https://napcentral.\norg/)50 and NDC Registry (https://unfccc.int/NDCREG)51 and published \nbefore 30 September 2022 (Supplementary Table 1). They offer com-\nparative insights into adaptation priorities, needs and commitments \nacross the continent, while acknowledging the strengths and opportu-\nnities of national adaptation planning processes to contribute to future \ntracking efforts. NDCs and NAPs serve distinct, yet complementary \nroles within a country. NDCs are high-level strategic documents, also \nused to signal negotiation positions on key global agendas, such as \n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nadaptation finance or loss and damage compensation11,18. In contrast, \nNAPs are embedded in domestic policy processes, providing detailed \nimplementation plans that contribute to the Paris Agreement objec-\ntives of adaptation33; they can help to operationalize adaptation com-\nponents of NDCs and provide plans to track implementation, including \ndetailed M&E systems that measure action progress in implementation \nand achievement of goals and objectives9.\nRecent analyses found that stand-alone Adcom can provide infor-\nmation on implemented adaptation8. By 30 September 2022, only eight \nAfrican countries had submitted a stand-alone Adcom52. Their infre-\nquent submission, especially by least developed countries, limits oppor-\ntunities for continental comparisons of adaptation progress. Seven \ncountries including Angola, Burundi, Kenya, Mauritius, Mauritania,  \nSouth Africa and Sudan had integrated Adcoms in their NDCs and they \nwere considered in our analysis. Exclusion of stand-alone Adcoms from \nour analysis was a practical decision, yet it represents a limitation of \nour study. However, our findings remain highly relevant by provid-\ning comparable information on adaptation priorities at continental \nscale, representing starting points for countries\u2019 future tracking and \nreporting efforts, including development of future BTRs and Adcoms.\nInitial document screening\nWe first reviewed a random sample of 20 documents for their structure \nand content type. This allowed us to assess the applicability of the \nframework criteria to the source documents, to fine-tune definitions \nand establish clear rules in the Protocol. For instance, we broadened \nour initial definition of actions to include on-the-ground initiatives and \nrelevant policies and programmes, recognizing that NDCs and NAPs \nfocus mainly on planned adaptation53. We also refined our approach to \nrecording information on targets, which are not mapped as stand-alone \nelements but linked to actions, objectives or indicators.\nData collection\nData were collected manually by a team of three over the course of \n5\u2009months. Documents in French (n\u2009=\u200923) and Spanish (n\u2009=\u20091) were trans-\nlated into English using open-source software (Google Translate). \nExtraction was guided by the Protocol informed by GAMI codebooks7, \nwhich outlined criteria for information inclusion, definitions and pro-\ncedures to help minimize interpretation differences across the extrac-\ntors. Data extracted were cross-checked by the team lead member \n(A.N.), who randomly revisited 75% of source documents to validate \nand enhance the accuracy of the information extracted. In addition, \nthe team conducted regular reconciliation meetings to discuss and \nresolve any discrepancies in extraction, ensuring consistent adherence \nto Protocol guidelines and criteria49. Data were recorded in two separate \ndatabases54. The main database, \u2018Adaptation Elements\u2019, captures data \non \u2018coverage\u2019 and \u2018robustness\u2019, where each row represents a unique \nreference to an adaptation element. The secondary database, \u2018Adapta-\ntion Elements Linkages\u2019, captures data on document \u2018consistency\u2019, with \neach row representing a document evaluated for consistency.\nData types\nFor coverage, we recorded explicit mentions of hazards, systems at \nrisk, adaptation goals, objectives, actions and indicators, as individual \nobservations in the database54. For each of these elements, we collected \nadditional information on targets, understood as desired benchmark \nvalues for future reference44, whenever available in the documents. \nVariable definitions are available in Supplementary Table 2. For consist-\nency, we searched through text, figures and tables for evidence of link-\nages between adaptation cycle components: climate risk and impact \nassessment, planning, implementation, MEL. We documented explicit \nconnections between sets of climate risk and impact assessment ele-\nments (that is, hazards, systems at risk), sets of planning elements (that \nis, goals, objectives), sets of implementation elements (that is, actions) \nand MEL elements (that is, indicators), rather than mapping linkages \nbetween unique observations. We only included connections that were \nclear in the source documents (rather than assumed), because they \nprovide a clear indication of intentional and deliberate articulation \nof adaptation elements (Supplementary Table 3). For robustness, we \nrelied on data collected to assess coverage, hence no additional data \nwere required.\nOther data\nWe used additional data to quickly investigate possible explainers of \nadequacy that could serve as foundations for future investigations. \nHence, we looked at a selection of relevant variables that were high-\nlighted in the literature and for which data were available to enable \ncross-country comparisons. Specifically, we used the World Bank \n\u2018Country classification by income level\u2019 (considering gross national \nincome (GNI) per capita)55, which categorizes countries into four \nincome groups: low, lower-middle, upper-middle and high-income. \nWe expected low-income countries to have fewer financial resources to \ninvest in comprehensive planning processes and therefore lower docu-\nment adequacy scores. In addition, we examined \u2018governance\u2019 scores \nthat form the Notre Dame Global Adaptation Index (ND-GAIN) index56; \nthese measure institutional quality driving adaptation, such as regula-\ntory quality, rule of law, political stability and control of corruption. We \nconsidered 5-year average governance scores, using the publication \ndate of NDCs and NAPs as a benchmark to identify the relevant 5\u2009years \nfor each country. The hypothesis was that higher governance scores \nindicate higher institutional readiness to adapt and to make effective \nuse of adaptation investment for developing adequate plans. In addi-\ntion, we incorporated data on \u2018adaptation finance\u2019 compiled by the \nClimate Policy Initiative57, which capture finance received by a country \nfor climate adaptation activities from public, private and blended \nfinance sources. As finance data were averages of 2019 and 2020, we \nused them under the assumption that countries that submitted NAPs \nand NDCs before that timeframe (2015\u20132019) or after it (2020\u20132022) \nreceived comparable levels of financial support for adaptation. We \nexpected higher levels of climate finance to drive higher adequacy \nscores for NDCs and NAPs.\nData analysis\nAll analyses were performed in R58 and focused on document type, given \nthe distinct functions of NDCs and NAPs for adaptation9,11,18.\nCoverage scores were determined on the basis of documents\u2019 \ninclusion of the six adaptation elements: hazards, systems at risk, \ngoals, objectives, actions and indicators. First, we summarized the \n\u2018Adaptation Elements\u201954 dataset by country and document, using binary \nvariables to indicate the presence (1) or absence (0) of an element in \nthe document. This practical approach allowed us to ensure consist-\nent evaluation across documents, reduce scoring ambiguity, improve \ninter-rater reliability and streamline data analysis. For each document\u2013\ncountry combination, occurrences were summed, yielding a score with \na min-max range of 0\u20136. The coverage score per document was then \ndetermined by dividing the total score by the maximum value (min-max \nrange 0\u20131). Total counts of observations per element, document and \ncountry were also computed and are included in Extended Data Table 4.\nConsistency scores were determined by the number of linkages \nidentified across the adaptation cycle in each document. We sum-\nmarized the \u2018Adaptation Linkages\u201954 dataset into a table indicating \nthe presence (1) or absence (0) of linkages between: climate risk and \nimpact assessment and planning, climate risk and impact assessment \nand implementation, climate risk and impact assessment and MEL, \nplanning and implementation, planning and MEL, and implementation \nand MEL. On the basis of this, the total number of linkages per docu-\nment was determined, with min-max range of 0\u20136. The consistency \nscore (min-max 0\u20131) was obtained by dividing the total number of link-\nages by the maximum value; higher coverage scores indicate a higher \nnumber of linkages. The summarized data were also used to generate \n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\na 4\u2009\u00d7\u20094 adjacency matrix and heat maps summarizing how often (that \nis, number of countries) each pair of components co-occurred in the \ndataset. We did not assess the quality of the linkages, such as whether \nthe assumptions underlying the connections were scientifically sound. \nThis represents a limitation of the current analysis but can be explored \nin future works that seek to test the validity and reliability of these link-\nages, providing evidence of results and outcomes.\nRobustness scores were determined on the basis of documents\u2019 \ninclusion of sets of indicators that satisfied a combination of SMART+ \ndesign and M&E function characteristics. First, this required an \nexpert-based analysis of the indicators using deductive content meth-\nods. Content analyses based on expert assessments are time consuming \nand inherently subjective, driven by value judgments, yet they provide \nimportant complementary sources of information, allowing distillation \nof nuanced insights in the data17,59,60. To mitigate bias in categorization \nand ensure reproducibility, coding definitions and instructions were \ngrounded in established literature and expert knowledge and are \ndetailed in the Protocol (Supplementary Tables 4 and 5). In addition, the \nProtocol includes examples to ensure a uniform understanding of the \ndefinitions and instructions. Each indicator was assessed by two other \nindividuals to ensure coding reliability and consistency. Reconciliation \nmeetings were organized to address coding discrepancies.\nWe used binary values to classify indicators that met (1) or not (0) \nthe 11 quality traits defining robustness: specific (S), measurable (M), \nassigned data source (A), relevant (R), time-bound (T), with target (+) \nand used to measure climate parameters, inputs, processes, outputs \nor outcomes. One indicator can satisfy multiple or all 6 SMART+ char-\nacteristics but can only meet 1 of 5 M&E functions. As our analysis of \nrobustness focused on the document level, we first counted the number \nof SMART+ characteristics met by each indicator (SMART+i) and then \naveraged it across all indicators to calculate a mean number of SMART+ \ncharacteristics per document (SMART+d). Similarly, for the M&E func-\ntion, we averaged the number of characteristics met by all indicators \nin a document (M&Ed). The robustness score was then computed by \nsumming SMART+d and M&Ed and then dividing it by the maximum \nvalue possible (11). This yielded a score with min-max range of 0\u20131, with \nhighest values indicating higher number of robustness characteristics \nmet by that document.\nAdequacy scores for each document were calculated by construct-\ning an index that aggregated the weighted scores for coverage, consist-\nency and robustness. In the absence of substantive evidence in the \nliterature regarding the relative importance of one criterion over the \nother, we assigned equal weights of 1/3. This meant that each criterion \ncontributed equally to the adequacy score. The resulting adequacy \nscore had values with min-max range of 0\u20131.\nTo determine the statistical significance of differences in scores \nbetween groups (that is, NDCs vs NAPs and low-income countries vs others) \n(Extended Data Table 3), we conducted Wilcoxon rank-sum tests, which \nallowed us to assess whether there were significant disparities in adequacy \nscores between groups, without relying on assumptions of normality. A \ncontinuity correction was applied to the test to adjust for ties. For country \nincome ranking, we grouped low- and lower-middle-income countries \ninto the \u2018lower\u2019 group, while upper-middle- and high-income-countries \nwere grouped in the \u2018other\u2019 category.\nLastly, we built linear regression models using the \u2018lm\u2019 function \nin R to investigate the relationship between adequacy scores for each \ndocument type (dependent variable), adaptation finance and govern-\nance (independent variables). The Democratic Republic of the Congo \nand South Sudan did not have data on governance, hence the models \nwere based on 51 NDCs and 13 NAPs. The following equations were used \nin the linear models:\nA-NAP = \u03b20 + \u03b21 \u00d7 Governance + \u03f5\n(1)\nA-NAP = \u03b20 + \u03b21 \u00d7 Finance + \u03f5\n(2)\nA-NDC = \u03b20 + \u03b21 \u00d7 Governance + \u03f5\n(3)\nA-NDC = \u03b20 + \u03b21 \u00d7 Finance + \u03f5\n(4)\nwhere A-NAP is the NAP adequacy score; A-NDC is the NDC adequacy \nscore; \u03b20 is the baseline adequacy score for NAP/NDC; \u03b21 is the coeffi-\ncient for the independent variable, indicating the change in adequacy \nscore for a one-unit increase in governance/finance; governance and \nfinance are independent variables, representing the values for the \ngovernance readiness and finance variable, respectively; and \u03f5 is the \nerror term, representing the random variation in adequacy score that \nis not explained by the independent variables.\nReporting summary\nFurther information on research design is available in the Nature \nPortfolio Reporting Summary linked to this article.\nData availability\nData are available on Harvard Dataverse at https://doi.org/10.7910/\nDVN/VK3CP9 (ref. 54). Source data are provided with this paper.\nCode availability\nThe supporting code for the results is available on GitHub at  \nhttps://doi.org/10.5281/zenodo.11237938 ref. 61.\nReferences\n49.\t Nowak, A. C., Njuguna, L. & Crumpler, K. How can governments \nengage in adaptation tracking? A protocol for assessing national \nadaptation policies. protocolexchange https://doi.org/10.21203/\nrs.3.pex-2399/v1 (2023).\n50.\t NAP Central https://napcentral.org/ (UNFCCC, 2023).\n51.\t Nationally Determined Contributions Registry https://unfccc.int/\nNDCREG (UNFCCC, 2023).\n52.\t Adaptation Communications Registry https://unfccc.int/ACR \n(UNFCCC, 2024).\n53.\t Lesnikowski, A., Ford, J. D., Berrang-Ford, L., Barrera, M. & \nHeymann, J. How are we adapting to climate change? A global \nassessment. Mitig. Adapt. Strateg. Glob. Change 20, 277\u2013293 \n(2015).\n54.\t Nowak, A., Njuguna, L. & Crumpler, K. Adaptation elements \nin African Nationally Determined Contributions (NDCs) and \nNational Adaptation Plans (NAPs). Harvard Dataverse https://doi.\norg/10.7910/DVN/VK3CP9 (2023).\n55.\t World Bank Country and Lending Groups \u2013 World Bank Data Help \nDesk (The World Bank Group, 2024).\n56.\t Country Index Technical Report (Univ. of Notre Dame Global \nAdaptation Initiative, 2023).\n57.\t Meattle, C. et al. Landscape of Climate Finance in Africa (Climate \nPolicy Initiative, 2022).\n58.\t R Core Team. R: A Language and Environment for Statistical \nComputing http://www.R-project.org/ (R Foundation for Statistical \nComputing, 2023).\n59.\t Erlingsson, C. & Brysiewicz, P. A hands-on guide to doing content \nanalysis. Afr. J. Emerg. Med. 7, 93\u201399 (2017).\n60.\t Krippendorff, K. Content Analysis: An Introduction to Its \nMethodology (Sage, 2004).\n61.\t Nowak, A. Adaptation-tracking-africa. GitHub https://doi.org/ \n10.5281/zenodo.11237938 (2024).\nAcknowledgements\nThe World Bank funded \u2018Accelerating Impacts of CGIAR Climate \nResearch\u2019 (AICCRA) Project supported this work (A.C.N., T.S.R.) and it \nwas implemented by the Alliance of Bioversity and CIAT with partners. \nWe thank G. Wamukoya for direction and consultation.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nAuthor contributions\nA.C.N., L.N., T.S.R., J.R.-V. and P.R. designed the study. A.C.N., L.N. \nand K.C. collected and curated the data. A.C.N. analysed the data, \nprepared and finalized the paper. L.N., T.S.R., J.R.-V., P.R. and K.C. \nedited the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at  \nhttps://doi.org/10.1038/s41558-024-02054-7.\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s41558-024-02054-7.\nCorrespondence and requests for materials should be addressed to \nAndreea C. Nowak.\nPeer review information Nature Climate Change thanks Timo Leiter, \nAlexandre Magnan, Diana Reckien and the other, anonymous, \nreviewer(s) for their contribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nExtended Data Table 1 | Descriptive statistics\nDescriptive statistics from analysis of African NDCs (N\u2009=\u200953) and NAPs (N\u2009=\u200915). \n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nExtended Data Table 2 | Scores by country and document\nScores for adequacy (index), coverage, consistency and robustness for African NDCs (N\u2009=\u200953) and NAPs (N\u2009=\u200915). Countries\u2019 names are presented in descending ordered by overall adequacy \nscores. \u201cDate\u201d refers to the publication date of the document on the UNFCCC web page at the time of the analysis. \u201cPublication\u201d version refers to the label of the document on the UNFCCC \nweb page. Calculation of adequacy index, coverage, consistency, and robustness criteria are provided in the Methods section of the article. \n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nExtended Data Table 3 | Comparisons between NDCs and NAPs: statistical tests\nWe used two-sided Wilcoxon rank-sum tests with continuity correction to compare adequacy, consistency, coverage, and robustness scores between NDCs and NAPs. The data were not \nnormally distributed; therefore, non-parametric tests were chosen. The p-values indicate the significance of the observed differences between NDCs and NAPs, with asterisks denoting levels \nof significance: ***p\u2009<\u20090.001: highly significant, **p\u2009<\u20090.01: very significant; *p\u2009<\u20090.05: significant. Sample size for NDCs is N\u2009=\u200953 and for NAPs is N\u2009=\u200915 for all comparisons in Set 1. In Set 3, we only \nuse sample size for NDCs (N\u2009=\u200953). \n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nExtended Data Table 4 | Coverage of adaptation elements (total counts, %)\nTotal counts refers to the frequency of mentions for each adaptation element in NDCs (N\u2009=\u200953) and NAPs (N\u2009=\u200915). For hazards and systems at risk, the absolute number represents the number of \nunique hazard and systems at risk categories identified in a document (15 and 15, respectively). \n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nExtended Data Table 5 | Consistency between adaptation cycle components\nConsistency is indicated by the number of linkages between adaptation cycle components for each NDC (N\u2009=\u200953) and NAP (N\u2009=\u200915). Total linkages represents the sum across linkages between \nall six combinations of components. Values of 1 indicate presence of a linkage, empty cells indicate absence. Adaptation cycle components abbreviated for table format. \n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nExtended Data Table 6 | Robustness of documents\nNDCs (N\u2009=\u200953) and NAPs (N\u2009=\u200915) that fulfill criteria of robustness, by robustness characteristics. \u201cRob.\u201d refers to the number of robustness characteristics fulfilled by the document, by averaging \ninformation for each country-document combination. \u201cCl\u201d-\u201cOc\u201d indicate number of indicators for each country-document combination fulfilling function characteristics, namely climate, \ninput, processes, outputs and outcomes, respectively; \u201cS\u201d-\u201cT\u2009+\u2009\u201d indicate number of indicators for each country-document combination fulfilling SMART design characteristics, namely \nspecific, measurable, with assigned data sources, relevant, time-bound and with a target, respectively. \n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-02054-7\nExtended Data Table 7 | Adaptation targets in NDCs and NAPs with no indicators\nSubset of NDCs (N\u2009=\u200934) and NAPs (N\u2009=\u20093) that do not identify indicators, but include targets linked to goals, objectives, and actions. Target n refers to the total number of targets that link to \na goal, objective or an action identified in the document. Target % indicates the proportion of targets from all observations extracted from the document. \u201cGoal\u201d-\u201cAction\u201d indicate numbers \nof targets linked to a goal, objective, or action, respectively. \u201cMeasurable\u201d-\u201cMeasurable & Timebound\u201d indicates the proportion of targets that are measurable, timebound, or measurable & \ntimebound, respectively calculated from Targets n. Empty cells indicate no target meeting the specific characteristic. \n\n1\nnature portfolio  |  reporting summary\nMarch 2021\nCorresponding author(s):\nAndreea Nowak\nLast updated by author(s): Nov 23, 2023\nReporting Summary\nNature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nData was collected manually from policy documents (Nationally Determined Contributions and National Adaptation Plans), using a protocol \navailable on Protocol Exchange (https://doi.org/10.21203/rs.3.pex-2399/v1). Data was collected in two Microsoft Excel databases, one \nfocusing on information on adaptation elements and another with information on linkages between adaptation elements. All data is available \non Harvard Dataverse (https://doi.org/10.7910/DVN/VK3CP9). The list of policy documents reviewed is available in Supplementary \nInformation Note 1. \nData analysis\nData analysis was performed in R (version 4.2.3). The codes used for data analysis are available as R Markdown files on a public GitHub \nrepository (https://github.com/andreeanowak/adaptation-tracking-africa). Coding rules were established in the Research Protocol available \non Protocol Exchange (https://doi.org/10.21203/rs.3.pex-2399/v1) and also incldued in the Supplementary Information of the manuscript.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and \nreviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.\n\n2\nnature portfolio  |  reporting summary\nMarch 2021\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A description of any restrictions on data availability \n- For clinical datasets or third party data, please ensure that the statement adheres to our policy \n \nData are available on Harvard Dataverse (https://doi.org/10.7910/DVN/VK3CP9) under a CC0 1.0 License. These datasets will allow reproducing the entire analysis \nconducted. The supporting code for the results is available on Github (https://github.com/andreeanowak/adaptation-tracking-africa) under a CC0 1.0 License. The \nResearch Protocol is available  on Protocol Exchange (https://doi.org/10.21203/rs.3.pex-2399/v1), licensed under a CC BY 4.0 License. No restrictions on data \navailability exist.\nHuman research participants\nPolicy information about studies involving human research participants and Sex and Gender in Research. \nReporting on sex and gender\nDoes not apply\nPopulation characteristics\nSee above\nRecruitment\nSee above\nEthics oversight\nSee above\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nEcological, evolutionary & environmental sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nWe reviewed Nationally Determined Contributions (NDCs) and National Adaptation Plans (NAPs) of African countries to identify \nopportunities for informing efforts to track climate change adaptation at national scales, across the continent. Most information \ncollected was qualitative. We developed a data collection and analysis protocol to ensure consistency and comprehensiveness of \ndata and findings. Some quantitative information, specifically on quantified adaptation targets, was also collected. The study protocol \nis published on Protocol Exchange (https://doi.org/10.21203/rs.3.pex-2399/v1)\nResearch sample\nWe analyzed 68 documents (n=68), of which 53 Nationally Determined Contributions (NDCs) and 15 National Adaptation Plans (NAPs) \nSampling strategy\nWe selected these n=68 documents based on their online availability before the pre-established cut off date of 30 September 2022. \nThese documents cover all African countries, except for Libya, which did not publish any NDC or NAP by the cut-off date.\nData collection\nData were collected by a theme of three. All NDCs and NAPs were downloaded from the UNFCCC repositories (NAP Central and NDC \nRegistry). Data were recorded in two separate databases,  one focusing on information on adaptation elements and another with \ninformation on linkages between adaptation elements. All data are available on Harvard Dataverse (https://doi.org/10.7910/DVN/\nVK3CP9).The Protocol explaining data extraction procedures is available on Protocol Exchange (https://doi.org/10.21203/\nrs.3.pex-2399/v1)\nTiming and spatial scale\nData were collected between August 2022 and December 2022. Data cover eight years (2015-2022, depending on the publishing \ndate of the document) and 53 African countries. \nData exclusions\nWe excluded two NDCs (updated versions of Egypt and Equatorial Guinea) and four NAPs (of Cabo Verde, Mozambique, Niger, and \nZambia) because they were published after the pre-established cut-off date of 30 September 2022\nReproducibility\nAll code and data are freely available on Dataverse and GitHub repositories. See data and code availability statements on the \nManuscript\n\n3\nnature portfolio  |  reporting summary\nMarch 2021\nRandomization\nDoes not apply\nBlinding\nDoes not apply\nDid the study involve field work?\nYes\nNo\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology and archaeology\nAnimals and other organisms\nClinical data\nDual use research of concern\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nAntibodies\nAntibodies used\nDoes not apply\nValidation\nDoes not apply\nEukaryotic cell lines\nPolicy information about cell lines and Sex and Gender in Research\nCell line source(s)\nDoes not apply\nAuthentication\nDoes not apply\nMycoplasma contamination\nDoes not apply\nCommonly misidentified lines\n(See ICLAC register)\nDoes not apply\nPalaeontology and Archaeology\nSpecimen provenance\nDoes not apply\nSpecimen deposition\nDoes not apply\nDating methods\nDoes not apply\nTick this box to confirm that the raw and calibrated dates are available in the paper or in Supplementary Information.\nEthics oversight\nNo ethical approval or guidance was required, as this study was a review of policy documents\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nAnimals and other research organisms\nPolicy information about studies involving animals; ARRIVE guidelines recommended for reporting animal research, and Sex and Gender in \nResearch\nLaboratory animals\nThe study did not involve laboratory animals\n\n4\nnature portfolio  |  reporting summary\nMarch 2021\nWild animals\nThe study did not involve wild animals\nReporting on sex\nThis information was not collected\nField-collected samples\nThe study did not involve samples collected from the field\nEthics oversight\nNo ethical approval or guidance was required, as this study was a review of policy documents\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nClinical data\nPolicy information about clinical studies\nAll manuscripts should comply with the ICMJE guidelines for publication of clinical research and a completed CONSORT checklist must be included with all submissions.\nClinical trial registration\nDoes not apply\nStudy protocol\nDoes not apply\nData collection\nDoes not apply\nOutcomes\nDoes not apply\nDual use research of concern\nPolicy information about dual use research of concern\nHazards\nCould the accidental, deliberate or reckless misuse of agents or technologies generated in the work, or the application of information presented \nin the manuscript, pose a threat to:\nNo\nYes\nPublic health\nNational security\nCrops and/or livestock\nEcosystems\nAny other significant area\nExperiments of concern\nDoes the work involve any of these experiments of concern:\nNo\nYes\nDemonstrate how to render a vaccine ineffective\nConfer resistance to therapeutically useful antibiotics or antiviral agents\nEnhance the virulence of a pathogen or render a nonpathogen virulent\nIncrease transmissibility of a pathogen\nAlter the host range of a pathogen\nEnable evasion of diagnostic/detection modalities\nEnable the weaponization of a biological agent or toxin\nAny other potentially harmful combination of experiments and agents\nChIP-seq\nData deposition\nConfirm that both raw and final processed data have been deposited in a public database such as GEO.\nConfirm that you have deposited or provided access to graph files (e.g. BED files) for the called peaks.\nData access links \nMay remain private before publication.\nDoes not apply\nFiles in database submission\nDoes not apply\n\n5\nnature portfolio  |  reporting summary\nMarch 2021\nGenome browser session \n(e.g. UCSC)\nDoes not apply\nMethodology\nReplicates\nDoes not apply\nSequencing depth\nDoes not apply\nAntibodies\nDoes not apply\nPeak calling parameters\nDoes not apply\nData quality\nDoes not apply\nSoftware\nDoes not apply\nFlow Cytometry\nPlots\nConfirm that:\nThe axis labels state the marker and fluorochrome used (e.g. CD4-FITC).\nThe axis scales are clearly visible. Include numbers along axes only for bottom left plot of group (a 'group' is an analysis of identical markers).\nAll plots are contour plots with outliers or pseudocolor plots.\nA numerical value for number of cells or percentage (with statistics) is provided.\nMethodology\nSample preparation\nDoes not apply\nInstrument\nDoes not apply\nSoftware\nDoes not apply\nCell population abundance\nDoes not apply\nGating strategy\nDoes not apply\nTick this box to confirm that a figure exemplifying the gating strategy is provided in the Supplementary Information.\nMagnetic resonance imaging\nExperimental design\nDesign type\nDoes not apply\nDesign specifications\nDoes not apply\nBehavioral performance measures\nDoes not apply\nAcquisition\nImaging type(s)\nDoes not apply\nField strength\nDoes not apply\nSequence & imaging parameters\nDoes not apply\nArea of acquisition\nDoes not apply\nDiffusion MRI\nUsed\nNot used\nPreprocessing\nPreprocessing software\nDoes not apply\n\n6\nnature portfolio  |  reporting summary\nMarch 2021\nNormalization\nDoes not apply\nNormalization template\nDoes not apply\nNoise and artifact removal\nDoes not apply\nVolume censoring\nDoes not apply \nStatistical modeling & inference\nModel type and settings\nDoes not apply\nEffect(s) tested\nDoes not apply\nSpecify type of analysis:\nWhole brain\nROI-based\nBoth\nStatistic type for inference\n(See Eklund et al. 2016)\nDoes not apply\nCorrection\nDoes not apply\nModels & analysis\nn/a Involved in the study\nFunctional and/or effective connectivity\nGraph analysis\nMultivariate modeling or predictive analysis\n\n\n    Scientific Research Findings:", "answer": "Despite variation among countries and documents, most African NAPs and NDCs provide only a fraction of the information fundamental for adaptation tracking. Only eight NAPs and four NDCs covered information on all key elements of the adaptation cycle: climate hazards, systems at risk, goals, objectives, actions and indicators. Six NAPs and five NDCs presented fully consistent narratives, articulating linkages between risk assessments, adaptation planning, implementation and tracking. Ten NAPs and five NDCs included adaptation indicators, yet no indicator set was fully robust. This calls into question the utility of existing indicators in tracking adaptation progress. NAPs provide a more adequate basis for tracking compared with NDCs, underscoring the importance of accelerating NAP development and implementation as a pathway towards improving tracking infrastructure. Examining the coverage, consistency and robustness of these documents provides a benchmark for policy adequacy and reveals opportunities for advancing adaptation tracking.", "id": 41}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 14 | June 2024 | 644\u2013651\n644\nnature climate change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-01984-6\nThe carbon dioxide removal gap\nWilliam F. Lamb\u2009\n\u200a\u20091,2\u2009\n, Thomas Gasser3, Rosa M. Roman-Cuesta4, \nGiacomo Grassi\u2009\n\u200a\u20094, Matthew J. Gidden\u2009\n\u200a\u20093, Carter M. Powis5, Oliver Geden\u2009\n\u200a\u20096, \nGregory Nemet\u2009\n\u200a\u20097, Yoga Pratama\u2009\n\u200a\u20093, Keywan Riahi\u2009\n\u200a\u20093, Stephen M. Smith\u2009\n\u200a\u20095, \nJan Steinhauser\u2009\n\u200a\u20093, Naomi E. Vaughan\u2009\n\u200a\u20098,9, Harry B. Smith\u2009\n\u200a\u20098,9 & \nJan C. Minx\u2009\n\u200a\u20091,2\nRapid emissions reductions, including reductions in deforestation-based \nland emissions, are the dominant source of global climate mitigation \npotential in the coming decades. However, carbon dioxide removal (CDR) \nwill also have an important role to play. Despite this, it remains unclear \nwhether current national proposals for CDR align with temperature targets. \nHere we show the \u2018CDR gap\u2019, that is, CDR efforts proposed by countries fall \nshort of those in integrated assessment model scenarios that limit warming \nto 1.5\u2009\u00b0C. However, the most ambitious proposals for CDR are close to levels \nin a low-energy demand scenario with the most-limited CDR scaling and \naggressive near-term emissions reductions. Further, we observe that many \ncountries propose to expand land-based removals, but none yet commit to \nsubstantively scaling novel methods such as bioenergy carbon capture and \nstorage, biochar or direct air carbon capture and storage.\nCDR can support climate mitigation in three ways1,2. First, in the \nshort-term, it can reduce net emissions. While many CDR methods \nare costly and technologically immature, afforestation and land-based \nremovals already make a contribution today. Second, in the mid-term, \nCDR can counterbalance residual emissions in \u2018hard-to-abate\u2019 sectors, \nallowing countries to reach their stated net-zero CO2 or greenhouse gas \n(GHG) emissions objectives. And third, in the long-term, CDR could \nbe used to reach net-negative emissions. This could compensate for \nhistorical emissions and allow global temperature exceedance to be \nreversed (however, it would not avoid the impacts associated with an \novershoot of 1.5\u2009\u00b0C, such as biodiversity loss and sea level rise)3.\nYet despite the apparent importance of CDR, there are few dedi-\ncated efforts to track real-world deployments, commitments, policies \nor related developments in the sector2,4. By contrast, tracking is widely \navailable for emissions reductions5\u20137. In particular, none have evaluated \nthe removal component of the \u2018emissions gap\u2019\u2014a science-policy device \nfor assessing progress towards the Paris Agreement temperature goal, \npublished each year in the Emissions Gap Report7 and supported by an \nunderlying evidence base8\u201310. So far, the emissions gap has been formu-\nlated in terms of net GHG emissions, with no distinction having been \nmade between gross emissions and removals (Fig. 1). This simplifies the \nassessment to a single aggregated gap and recognizes certain empirical \nrealities: most countries do not distinguish emissions and removals \nin their targets, and integrated assessment model (IAM) reporting \nhas tended to combine emissions and removals on managed land as \na single net indicator. However, there are several compelling reasons \nwhy CDR should be distinguished in the gap analysis.\nIn the first instance, this is a simple transparency issue. As many \ncountries have pledged net-zero targets, an assessment of their implied \nemissions and removals will provide a better understanding of how \ncountries want to achieve these goals11. In turn, this opens a space for \ncritical reflection on the fairness and ambition of proposed reductions, \nlevels of residual emissions and potential overdependence on CDR12\u201316. \nA second reason is that emissions and removals are fundamentally dif-\nferent categories, involving different technologies, implementation \noptions and risks, with varying policy and governance requirements \nReceived: 11 July 2023\nAccepted: 19 March 2024\nPublished online: 3 May 2024\n Check for updates\n1Mercator Research Institute on Global Commons and Climate Change (MCC), Berlin, Germany. 2Priestley International Centre for Climate, University of \nLeeds, Leeds, UK. 3International Institute for Applied Systems Analysis (IIASA), Laxenburg, Austria. 4Joint Research Centre, European Commission, Ispra, \nItaly. 5School of Enterprise and the Environment, University of Oxford, Oxford, UK. 6German Institute for International and Security Affairs (SWP), Berlin, \nGermany. 7University of Wisconsin-Madison, Madison, WI, US. 8School of Environmental Sciences, University of East Anglia, Norwich, UK. 9Tyndall Centre \nfor Climate Change Research, University of East Anglia, Norwich, UK. \n\u2009e-mail: lamb@mcc-berlin.net\n\nNature Climate Change | Volume 14 | June 2024 | 644\u2013651\n645\nAnalysis\nhttps://doi.org/10.1038/s41558-024-01984-6\nTo estimate the CDR gap, we first organize our analysis around \ntwo categories of CDR that differ in terms of scale, technology readi-\nness and permanence: \u2018conventional CDR on land\u2019 and \u2018novel CDR\u2019. \nThe former consists of methods conventionally defined as removals \nin the land use, land-use change and forestry (LULUCF) sector (for  \nexample, afforestation, restoration). Novel CDR comprises all other \nCDR methods, such as biochar, direct air carbon capture and storage  \n(DACCS) or bioenergy carbon capture and storage (BECCS). (In  \nMethods, we further explain our definitions, including the notable \nexclusion of removals driven by indirect anthropogenic effects). \nWhereas conventional CDR on land methods are already widely \nadopted and integrated into national climate pledges, novel CDR \nmethods remain at an early stage of adoption and policy integration2. \nStudies are now beginning to report total current CDR deployments \nfollowing these definitions18,19, which we estimate as approximately \n3\u2009GtCO2\u2009yr\u22121, of which 99.9% is from conventional CDR on land (Fig. 2)19.\nTo estimate proposed levels of CDR upscaling by countries, we \ndraw from documents submitted to the United Nations Framework \nConvention on Climate Change (UNFCCC): the nationally determined \ncontributions (NDCs) and the long-term strategies (also known as \nincluding critical issues such as permanence and land use17. Finally, \nwhile CDR makes a trivial contribution to climate change mitigation \ntoday (Fig. 2), according to scenarios, it could become the dominant \nresponse in the second half of the twenty-first century2. In some coun-\ntries with large existing land-based removals, it could become the \ndominant response much sooner.\nIn this Article, we provide a conceptualization and quantification \nof the \u2018CDR gap\u2019\u2014the gap between levels of CDR that are proposed by \ngovernments and levels of CDR in IAM scenarios that limit warming \nto 1.5\u2009\u00b0C. Importantly, our evaluation introduces further normativity \ninto the assessment of global mitigation pledges by making a judge-\nment regarding the appropriate division of effort between emissions \nreductions and removals. Concretely, this judgment manifests in the \nscenarios we choose as a point of comparison to national proposals, \nincluding the specific amounts and types of CDR they implement, \nas well as their rates of emissions reductions. However, rather than \nobscure this choice by comparing against broad scenario ranges, we \ninstead select individual scenarios and aim to discuss and justify our \nparticular choices, further opening the discourse on \u2018how much CDR \nis needed\u2019 to meet the Paris Agreement.\nNet GHG emissions\nGross GHG emissions\nGross CO2 removals\nRange of net GHG emissions in the NDCs\na An assessment of the emissions gap combining emissions and removals\nb An assessment of the emissions gap separating emissions and removals\nEmissions: Non-CO2 GHGs\nEmissions: Fossil CO2\nEmissions: Managed land\nRemovals: Conventional CDR on land\nRemovals: Novel CDR\n2050\nThe emissions gap\nRange of gross CO2\n \nremovals in the NDCs\nThe CDR gap\n2010\n0\n2010\n2030\n2100\n2030\n2050\n2100\nThe emissions gap\nRange of gross GHG emissions in the NDCs\nGHG emissions\n0\nGHG emissions\nScenario pathway to 1.5 \u00b0C or 2 \u00b0C\nFig. 1 | Combined versus separate assessments of the emissions and CDR gap. a,b, A stylized scenario pathway that reaches net-zero CO2 and GHG emissions under \ncombined (a) versus separate (b) assessment. Typically, the gap would be assessed against a scenario range and median level, rather than a single scenario.\n\nNature Climate Change | Volume 14 | June 2024 | 644\u2013651\n646\nAnalysis\nhttps://doi.org/10.1038/s41558-024-01984-6\nthe long-term low-emissions development strategies). These give \ninsight into levels of CDR in 2030 and 2050, compared to historical \ninventory-based reporting. There are currently no strict requirements \nfor reporting CDR in either of these documents, so several assumptions \nmust be made to extract this information where it is implicit in national \ntargets (see Methods).\nTo benchmark levels of CDR proposed by countries, we use the \ncompilation of IAM scenarios vetted by the IPCC 6th Assessment \n(AR6)1,20. While novel CDR (such as BECCS) is reported in the AR6 \nscenario database, conventional CDR on land is only inconsistently \nreported as afforestation and instead tends to be combined with emis-\nsions as a net LULUCF flux. We therefore use a novel re-analysis of \nthe IPCC database using the OSCAR model that extracts the removal \ncomponent of the LULUCF flux in each scenario corresponding to our \ndefinition of conventional CDR on land (see Methods)21.\nCDR in national mitigation pledges\nOur NDC assessment finds that countries\u2019 conventional CDR on land \nwill change from \u22123.0\u2009GtCO2\u2009yr\u22121 for the period 2011\u20132020 (that is, the \nremovals reported in GHG inventories once the indirect effects are fac-\ntored out in this study; see Methods) to approximately \u22123.1\u2009GtCO2\u2009yr\u22121 \n(unconditional pledges) or about \u22123.5\u2009GtCO2\u2009yr\u22121 (conditional pledges) \nin 2030. While some countries include novel CDR in their qualitative \ndescription of mitigation efforts towards the 2030 pledges and a few \nprovide initial quantifications (for example, Korea, Canada, Norway), \nthese are currently not possible to distinguish from avoided emissions \n(for example, fossil-based CCS). We therefore estimate zero commit-\nments towards novel CDR by 2030, with no change from current levels \nof approximately 2\u2009MtCO2\u2009yr\u22121.\nIn the case of the long-term strategies, there is a general acknowl-\nedgement that CDR is needed to realize national net-zero targets22. \nIndeed, most countries include at least a qualitative description of \nhow this type of mitigation effort would be achieved. However, only \n40 countries have outlined scenarios in their long-term strategies \nthat depict quantifiable levels of CDR by 2050 (28 if EU countries are \ncombined as one). If we assume that all other countries sustain their \ncurrent levels of removals, proposed CDR as reflected in the long-term \nstrategies range between \u22124.6 and \u22125.0\u2009GtCO2\u2009yr\u22121 in 2050, the major-\nity of which is conventional CDR on land (85% and 81%, respectively).\nCDR in mitigation scenarios\nIn scenarios that limit warming to below 2\u2009\u00b0C (see Methods for scenario \ndefinitions), gross emissions reductions are the dominant mitigation \nresponse in the coming three decades. Between 2020 and 2050, emis-\nsions are reduced by 62% (46\u201375%). Subsequently, CDR becomes the \nmain mitigation strategy in the second half of the twenty-first century, \nwith scenarios cumulating 670 (450\u20131100)\u2009GtCO2 of removals by 2100. \nNovel CDR tends to continuously scale up in scenarios throughout the \ntwenty-first century and accounts for over half of cumulative remov-\nals by 2100. By contrast, conventional CDR on land starts from a high \nbaseline but quickly reaches saturation by the mid-century due to land \narea constraints for afforestation/restoration.\nScenarios vary considerably in their levels and types of CDR \ndeployment, depending on how policy choices, technology avail-\nability and socio-economic developments shape the speed and depth \nof gross emissions reductions (Table 1). We therefore highlight three \n\u2018focus scenarios\u2019 that depict different emission reduction and CDR \npathways to hold warming below 1.5\u2009\u00b0C:\nBECCS\nDACCS\nOther\n60\n0\n0\n\u201300005\n\u20130.0010\n\u20130.0015\n\u20130.0020\n\u20130.0025\n\u20131\n\u20132\n\u20133\nGt CO2e\nGlobal total greenhouse gas emissions and removals\nEmissions: Non-CO2 GHGS\nEmissions: Fossil CO2\nEmissions: Managed land\nRemovals: Conventional CDR on land\nRemovals: Novel CDR\n40\n20\n0\nFig. 2 | Current global CDR versus emissions. Updated from ref. 19 (see Methods) with additional emissions data from ref. 58, using global warming potentials with \na 100\u2009year time horizon from the IPCC 6th Assessment Report59. Emissions data for 2019 are plotted, while LULUCF removals are the 2011\u20132020 annual average, and \nnovel CDR removals are an estimate for 2020.\nTable 1 | Reasons why CDR deployments vary in scenarios\nReasons why scenarios deploy more \nCDR\nReasons why scenarios deploy less \nCDR\n\u2022 \u0007Emissions reductions are \ndelayed.46,47\n\u2022 \u0007A wider portfolio of CDR methods \nare available, lowering their \ncosts relative to deep emissions \nreductions.29,48,49\n\u2022 \u0007The portfolio of mitigation \ntechnologies that can lower \nresidual emissions at the point of \nnet-zero CO2 is more limited (such \nas CCS for industrial processes).50\n\u2022 \u0007A more-stringent temperature \ntarget is applied, lowering the \navailable carbon budget.50\n\u2022 \u0007The scenario is permitted to \ninitially exceed a warming level \nand compensate for this with net \nnegative emissions later in the \ncentury.29\n\u2022 \u0007A temperature target is chosen that \nhas already been exceeded, such \nas 1\u2009\u00b0C.51\n\u2022 \u0007For scenarios that use a full-century \ncarbon budget rather than a peak \nbudget52, values assumed for \neconomic discount rates can push \nmitigation further into the future.53\n\u2022 \u0007Emissions reductions are faster and \nimplemented without delay.46,47\n\u2022 \u0007A wider portfolio of (demand-side) \nmitigation options are available, with \nlower costs relative to CDR.23,54\n\u2022 \u0007A wider portfolio of mitigation \ntechnologies that can lower residual \nemissions at the point of net-zero \nCO2 is available (such as CCS for \nindustrial processes).50\n\u2022 \u0007A less-stringent temperature target \nis applied, increasing the available \ncarbon budget.50\n\u2022 \u0007Assumptions differ strongly \nabout different limitations to \nCDR deployment, including both \ntechnological progress55 as well \nas social and environmental \nsustainability.24,25,56,57. Scenarios may \nlimit the speed or total quantity of \ndeployment on the basis of some or \nall of these considerations.\n\nNature Climate Change | Volume 14 | June 2024 | 644\u2013651\n647\nAnalysis\nhttps://doi.org/10.1038/s41558-024-01984-6\n\u2022 Focus on demand reduction \u2013 a scenario that reduces global \nenergy demand through efficiency and sufficiency measures, with \na low long-term dependency on CDR23. Annual removals in 2050 \nare \u22124.8\u2009GtCO2, entirely from conventional CDR on land.\n\u2022 Focus on renewables \u2013 a scenario that rapidly implements a \nsupply-side transformation towards renewable energy24. Annual \nremovals in 2050 are \u22127.6\u2009GtCO2, including a small contribution \nfrom novel CDR (\u22120.91\u2009GtCO2).\n\u2022 Focus on carbon removal \u2013 a scenario with rapid near-term emis-\nsions reductions but a subsequent incomplete phase out of fossil \nfuels, leading to higher residual emissions at net-zero CO2. Annual \nremovals in 2050 are \u22129.8\u2009GtCO2, with a large contribution from \nnovel CDR (\u22123.5\u2009GtCO2).\nThe first two of these focus scenarios feature CDR levels at the \nlower end of the range in below 2\u2009\u00b0C scenarios (see Methods), while the \nlatter sits just above the median (see Table 2). Scenarios at the upper \nend of the below 2\u2009\u00b0C range (95th percentile) feature CDR deployments \nof \u221214\u2009GtCO2\u2009yr\u22121 in 2050\u2014levels that probably encounter feasibility \nconstraints in terms of scale-up and bioenergy resource availability25. \nAs all three scenarios hold warming below 1.5\u2009\u00b0C with no or limited \novershoot, they mainly differ in CDR upscaling due to the speed of their \nnear-term reductions and the quantity and type of residual emission \nthat they need to compensate to reach net-zero CO2 (Fig. 3 and Sup-\nplementary Table 1). We include 2\u2009\u00b0C (for example, C3) pathways in \nthe overall scenario range (Fig. 4 and Table 2) but do not select them as \nfocus scenarios, which would highlight both lower CDR requirements \nand lower gross emissions reductions, but also higher climate impacts.\nThe CDR gap\nAcross both categories of removals, a CDR gap already emerges by \n2030 (Table 2). Compared with 2011\u20132020, the conditional NDCs would \nexpand CDR by \u22120.5\u2009GtCO2\u2009yr\u22121 in 2030. This contrasts with an increase \nof \u22121\u2009GtCO2\u2009yr\u22121 in 2030 in the Focus on demand reduction scenario, \nwhich has the lowest CDR requirements. The CDR gap in 2050 is then \nstrongly determined by the chosen scenario benchmark. Compared \nwith 2020, additional CDR in 2050 implied by the upper estimate \nof the long-term mitigation strategies (from 28 countries including \nthe EU, assuming all others sustain current removals) would sum to \n\u22121.9\u2009GtCO2\u2009yr\u22121. This approaches levels in the Focus on demand reduc-\ntion scenario (an additional \u22122.3\u2009GtCO2\u2009yr\u22121), but falls short by multiple \ngigatons compared with the other focus scenarios. The most ambitious \nof current CDR plans are therefore close to a conservative level of CDR \nscaling, albeit one that would need to be coupled with deep, near-term \nemissions reductions.\nThe gap in conventional CDR on land\nNeither the NDCs in 2030 nor the long-term strategies in 2050 propose \nlevels of conventional CDR on land sufficient to meet those projected \nin scenarios (Table 2 and Fig. 4). However, our analysis only captures \ncountries with quantifiable scenarios, which represent about 38% of \ncurrent conventional CDR on land removals. These countries plan \nto increase removals by \u22120.8 to \u22121.0\u2009GtCO2\u2009yr\u22121, when adjusting the \nlong-term strategies to remove \u2018indirect anthropogenic effects\u2019 (see \nMethods). By contrast, the focus scenarios increase conventional CDR \non land by an additional \u22122.3\u2009GtCO2\u2009yr\u22121 (Focus on demand reduction) \nto \u22124.1\u2009GtCO2\u2009yr\u22121 (Focus on renewables).\nOur analysis assumes that all other countries without quantifiable \nscenarios (accounting for 62% of current conventional CDR on land) \ncan sustain their existing removals. This includes China, India and the \nDemocratic Republic of the Congo, which all have substantial forest \nconservation and restoration potentials26 and could be instrumental \nin closing the gap in conventional CDR on land.\nThe gap in novel CDR\nNo country transparently includes novel CDR as a distinct portion \nof their pledged mitigation efforts by 2030. By contrast, below 2\u2009\u00b0C \nscenarios already implement \u22120.06\u2009GtCO2\u2009yr\u22121 of additional novel CDR \nby 2030.\nLooking forward to 2050, many countries mention novel CDR \nin their long-term strategies and some quantify it in their illustra-\ntive national scenarios. At the upper estimate, approximately \n\u22120.96\u2009GtCO2\u2009yr\u22121 of additional novel CDR can be inferred from \nthese scenarios, largely driven by the US (\u22120.5\u2009GtCO2\u2009yr\u22121), Canada \n(\u22120.23\u2009GtCO2\u2009yr\u22121) and the EU (\u22120.08\u2009GtCO2\u2009yr\u22121). This compares to \nthe \u22120.91\u2009GtCO2\u2009yr\u22121 of (global) additional novel CDR in the Focus on \nrenewables scenario and the \u22123.5\u2009GtCO2\u2009yr\u22121 in the Focus on carbon \nremovals scenario. There is no gap in novel CDR compared with the \nFocus on demand reduction scenario, which avoids scaling up novel \nCDR entirely (but does, however, scale up conventional CDR on land).\nOur analysis assumes that countries without quantifiable sce-\nnarios do not currently plan to implement novel CDR. This includes \nChina, Norway and Saudi Arabia, which are all developing technology \nroadmaps towards novel CDR and could contribute to closing the gap.\nDiscussion\nOur initial quantification of the CDR gap highlights that countries \nalso lack progress in this domain of climate mitigation. While some \nare planning to scale CDR to meet the temperature goal of the Paris \nAgreement, together they fall short by hundreds of megatons in 2030 \nand by hundreds of megatons to multiple gigatons in 2050, depending \nTable 2 | Scaling of CDR to 2030 and 2050 in scenarios, NDCs and long-term strategies (GtCO2\u2009yr\u22121)\nAdditional total CDR from 2020 \n(GtCO2\u2009yr\u22121)\nAdditional conventional CDR on \nland from 2020 (GtCO2\u2009yr\u22121)\nAdditional novel CDR from \n2020 (GtCO2\u2009yr\u22121)\nGross GHG emissions \nreductions from 2020 (%)\n2030\n2050\n2030\n2050\n2030\n2050\n2030\n2050\nBelow 2\u2009\u00b0C scenarios\n\u22121.1\n(0.01 to \u22123.4)\n\u22124.5\n(0.92 to \u221211)\n\u22120.85\n(0.014 to \u22123)\n\u22122.3\n(2.5 to \u22126)\n\u22120.06\n(0 to \u22121.1)\n\u22122.4\n(\u22120.5 to \u22129.1)\n25\n(4.2\u201350)\n62\n(46\u201375)\nFocus on demand reduction\n\u22121\n\u22122.3\n\u22121\n\u22122.3\n0\n0\n51\n78\nFocus on renewables\n\u22122.9\n\u22125.1\n\u22122.7\n\u22124.1\n\u22120.14\n-0.91\n39\n80\nFocus on carbon removal\n\u22121.6\n\u22127.4\n\u22120.66\n\u22124.0\n\u22120.95\n\u22123.5\n40\n77\nNDCs*\n(\u22120.05 to \u22120.53)\nNA\n(\u22120.05 to \u22120.53)\nNA\n0\nNA\nNA\nNA\nLong-term mitigation strategies\nNA\n(\u22121.5 to \u22121.9)**\nNA\n(\u22120.8 to \u22121.0)**\nNA\n(0.7\u20130.96)*\nNA\nNA\nBelow 2\u2009\u00b0C scenarios refer to categories C1 and C3 in the AR6 scenario database. For these categories, the median and 5\u201395th percentiles are reported. In the lower range of some scenarios, \nconventional CDR on land decreases compared with 2020, which gives rise to negative numbers. The analysis of the NDCs (*) was complemented by other official reports containing \ninformation on the country\u2019s mitigation targets (for example, National Communications, Biannual Updated Reports, REDD+ documents, national mitigation strategies). The additional CDR in \nthe long-term mitigation strategies (**) assumes that countries without a quantifiable strategy preserve their current levels of conventional CDR on land. One hundred eleven NDCs (that is, \nexcluding small island states, city states and countries with no land use fluxes) and all long-term strategies up to November 2023 (COP28) were considered for the analysis.\n\nNature Climate Change | Volume 14 | June 2024 | 644\u2013651\n648\nAnalysis\nhttps://doi.org/10.1038/s41558-024-01984-6\non the benchmarked scenario. The importance of planning for CDR \nat scale in 2050 is therefore not currently reflected at the policy level, \neven under assumptions of rapid and sustained emissions reductions \nin the short term. However, three important caveats should be noted \nin this Analysis.\nFirst, although most countries have committed to net-zero targets, \nthey still provide little information on what role CDR will play in reach-\ning them. Within the NDCs, ambiguities and a lack of transparency lead \nto wide ranging assessments of not only the land-use flux and implied \nremovals, but also overall emissions levels27,28. These problems are even \nmore apparent with the long-term strategies, which lack any common \nreporting structure and where underlying scenarios are illustrative \nrather than formal commitments13. As of COP28, only 68 countries \n(42 when excluding EU countries) have actually submitted a long-term \nstrategy. Further, not all pledges have an associated climate law in their \nhome jurisdictions10.\nNevertheless, the NDCs and long-term strategies are among the \nfew reference points available for evaluating national CDR proposals, \nand they are the only documents that can be feasibly analysed and \naggregated for a global assessment. It is therefore critical that future \niterations of these documents contain the required transparency \nfor evaluating national targets based on both gross emissions  \nand removals.\nSecond, IAMs have a prominent role in shaping climate miti-\ngation policy advice and have been subject to several criticisms.  \nDiscussions have focused on whether sustainable levels of bioenergy \nuse are exceeded in scenarios, whether CDR tends to substitute for \nshort-term emissions reductions, and whether the full scope of low \ndemand, low CDR, or \u2018degrowth\u2019 scenarios has yet been explored29\u201332. \nIn addition, IAMs have mainly modelled afforestation, BECCS and \nDACCS, while other methods have been scarcely explored29. By drawing \nfrom scenario evidence, this CDR gap assessment is similarly exposed \nto such criticisms.\nIn this assessment, we take a pragmatic approach and recognize \nthat IAM scenarios provide the best current evidence available to bench-\nmark country proposals for CDR. We also select specific focus scenarios \na\nb\nEmissions: Non-CO2 GHGs\n60\nFocus on demand reduction\nFocus on renewables\nFocus on carbon removal\n50\n40\n30\n20\nAnnual GHG emissions and\nremovals (GtCO2yr\u22121)\nAnnual GHG emissions and\nremovals (GtCO2yr\u22121)\n10\n0\n\u201310\n60\n50\n40\n30\n20\n10\n0\n\u201310\n\u201320\n2020\n2060\n2100\n2020\n2060\n2100\n2020\n2060\n2100\nEmissions: Fossil CO2 (other)\nEmissions: Fossil CO2\nEmissions: Fossil CO2 (transportation)\nEmissions: Fossil CO2 (buildings)\nEmissions: Fossil CO2 (industry)\nEmissions: Fossil CO2 (energy supply)\nEmissions: Managed land\nRemovals: Conventional CDR on land\nRemovals: Novel CDR\nNet GHG emissions\n2020\nAt time of net-zero CO2\nFocus on\ndemand\nreduction\n(2052)\nFocus on\nrenewables\n(2067)\nFocus on\ncarbon\nremoval\n(2047)\nFocus on\ndemand\nreduction\n(2098)\nFocus on\nrenewables\n(2089)\nFocus on\ncarbon\nremoval\n(2077)\nAt time of net-zero GHG\nNet CO2 emissions\nFig. 3 | The three focus scenarios. a, The emissions and removals pathways of each scenario in the twenty-first century. b, The residual gross GHG emissions and \nremovals of each scenario at the point of net-zero CO2 and net-zero GHG emissions. The error bar in b depicts the median and interquartile range (n\u2009=\u2009189) of gross \nemissions and removals in scenarios, sourced from refs. 20,21.\n\nNature Climate Change | Volume 14 | June 2024 | 644\u2013651\n649\nAnalysis\nhttps://doi.org/10.1038/s41558-024-01984-6\nto increase the transparency in a set of possible CDR futures and their \nunderlying determinants, but orient our selection to scenarios at the \nlower end of CDR requirements. Other selections are possible and can \nbe made using the supplementary data file of this Analysis. Alternative \napproaches for benchmarking CDR levels should also be explored, for \ninstance by assessing the residual emissions associated with bottom-up \nenergy and material requirements for meeting human needs33. One area \nof needed improvement is the separation of gross LULUCF emissions \nand removals in scenario reporting\u2014information that we have sourced \nhere from a re-analysis of the AR6 scenario database21.\nFinally, a recurring concern in the literature is that including CDR \nin mitigation discussions may deter near-term emissions reductions34. \nStates, corporations or other interest groups seeking an excuse for \ndoing very little may exploit the fact that CDR can compensate for \nemissions, overplaying the quantity of removals that may be achieved \nat some (later) point in time. Indeed, a variety of claims and discursive \nstrategies beyond CDR are used to excuse or delay climate action, \nwhich may help political actors resolve the tension between powerful \nincumbent fossil interests and increasing domestic or international \ncalls for climate action35\u201337. Given the commercial stakes at play, sci-\nentists therefore face enormous challenges in facilitating a nuanced \ndialogue on CDR.\nThe assessment we provide of the CDR gap contributes to this \ndialogue by asking \u2018how much is needed?\u2019 and \u2018what are countries \nplanning?\u2019. We believe it is important to situate such questions in the \nscientific literature and provide a space to critically reflect on them. \nHowever, we acknowledge that this will not prevent interest groups \nfrom exploiting the integration of CDR in the climate debate. We there-\nfore plainly state: our assessment of CDR in no way underplays the \nneed for rapid, immediate and deep emissions reductions across all \nsectors, including a rapid decrease in fossil fuel use and the halting of \ndeforestation. Indeed, our Analysis reinforces this fact, as the longer \nsuch reductions are delayed, the higher future CDR requirements are \nand the wider the CDR gap becomes.\nConventional and novel CDR (GtCO2 yr\u22121)\nThe extent of future carbon dioxide removal depends on the scenario by which climate goals are met\nCurrent and proposed levels of carbon dioxide removal are insuficient to meet the Paris temperature goal\n0\na\nb\n\u20134\n\u20138\n\u201312\n0\n\u20132.5\n\u20135.0\n\u20137.5\n\u201310.0\n0\n\u20132.5\n\u20135.0\n\u20137.5\n\u201310.0\n2010\nCurrent CDR\nNDCs (unconditional)\nLong-term strategies (low)\nLong-term strategies (high)\nNDCs (conditional)\nFocus on renewables\nFocus on renewables\nFocus on carbon removals\nFocus on carbon removals\nFocus on demand reduction\nFocus on demand reduction\n2020\n2030\n2040\n2050\nConventional CDR\non land (GtCO2 yr\u22121) \nNovel CDR\n(GtCO2 yr\u22121) \nBelow 2 \u00b0C scenarios\nBelow 2 \u00b0C scenarios\n2011\u2013\n2020\n2030\n2050\nCurrent CDR\nCDR in Paris-relevant\nscenarios\nFocus on\nrenewables\n(39% GHG reduction by 2030)\nFocus on carbon\nremovals\n(40% GHG reduction by 2030)\nFocus on demand\nreduction\n(51% GHG reduction by 2030)\nFig. 4 | The carbon dioxide removal gap. a, Current levels of CDR and levels \nin Paris-relevant scenarios up to 2050. The orange shaded areas depict the \n5th\u201395th (light orange) and 25th\u201375th (dark orange) percentiles of IPCC C1 and \nC3 scenarios that limit warming to below 2\u2009\u00b0C. The orange lines depict three \nfocus pathways that limit warming to 1.5\u2009\u00b0C, alongside the gross greenhouse gas \nemissions reductions required by 2030 for each. b, Levels of current, proposed \nand scenario-based CDR, split by conventional CDR on land and novel CDR in \n2020, 2030 and 2050. Pink bars depict proposed CDR levels in the NDCs and \nthe long-term mitigation strategies. Orange bars depict CDR levels in the three \nfocus scenarios, as well as the overall scenario medians (centre line) and ranges \n(5th\u201395th (light orange) and 25th\u201375th (dark orange) percentiles).\n\nNature Climate Change | Volume 14 | June 2024 | 644\u2013651\n650\nAnalysis\nhttps://doi.org/10.1038/s41558-024-01984-6\nThere are varying challenges to closing the CDR gap. While conven-\ntional CDR on land is already well integrated into climate governance, \nexperience has highlighted difficulties in monitoring, reporting and \nverification38\u201340. An overdependence on land-based removals brings \nrisks for land availability, food production and ownership rights12. On \nthe other hand, if designed well, they can be integrated with sustainable \ndevelopment and biodiversity objectives41. In addition, forest carbon \nis vulnerable to reversal, and expectations that regional sinks can be \npreserved in the coming decades have been challenged, highlighting \nthe importance of policies that promote sustainable management, pre-\nvent illegal removals and limit the impact of natural disturbances42\u201344.\nRegarding novel CDR, there is little existing capacity and rates of \npotential scaleup are very high, both in the long-term strategies (up to \n0.96\u2009GtCO2\u2009yr\u22121, or 470 times current levels) and in below 2\u2009\u00b0C scenarios \n(up to 2.4\u2009GtCO2\u2009yr\u22121, or 1,200 times current levels, but with some sce-\nnarios at or near 0). Although technology adoption and scaleup rates \nhave been impressive in several analogous historical cases45, novel \nCDR methods such as BECCS may face headwinds due to high capital \ncosts, a dependency on state support and other factors. In our view, \nnear-term policies to support these methods in their formative phase \nare urgently needed, without which it is difficult to conceive of any \ngigaton-scale contribution from novel CDR in 2050 and beyond. In \naddition, regulatory action that robustly defines, monitors, reports \nand verifies novel CDR is lagging. Importantly, enhanced emissions \nreductions are needed to reduce our dependence on dramatically \nscaling up these nascent CDR technologies.\nTo what extent is the CDR gap due to inadequate proposals by \ncountries, versus a failure to specify them in the first place? Our analy-\nsis of the long-term strategies covers 28 countries (including the EU), \nsumming to 38% of current removals. Due to this limitation, we assume \nthat all other countries can sustain their current conventional CDR on \nland. This is a generous assumption, given how difficult it will be to sus-\ntain such removals amid mounting climate impacts42\u201344. On the other \nhand, we may underestimate proposals for novel CDR where national \npolicymaking is in its infancy (even though countries would have little \nincentive to develop concrete plans but exclude these from their com-\nmunicated targets). Given these uncertainties, it remains important \nto continuously track new developments and update estimates of the \nCDR gap as national policies and targets are refined.\nCDR entails many challenges for designing policy, supporting \ninnovation and ensuring sustainable, equitable and durable remov-\nals. Our Analysis shows that scenarios meeting the Paris temperature \ngoal imply a very rapid scaleup of CDR and that governments are not \nplanning for this. A twofold strategy that limits our dependence on \nCDR through rapid and deep emissions reductions, yet aggressively \nsupports and scales CDR implementation is not a contradiction but a \nnecessary pathway towards successful climate policy.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-024-01984-6.\nReferences\n1.\t\nIPCC: Summary for Policymakers. In Climate Change 2022: \nMitigation of Climate Change (eds Shukla, P. R. et al.) (Cambridge \nUniv. Press, 2022).\n2.\t\nSmith, S. M. et al. The State of Carbon Dioxide Removal 1st edn \n(The State of Carbon Dioxide Removal, 2023); https://www.\nstateofcdr.org\n3.\t\nBabiker, M. et al. in Climate Change 2022: Mitigation of Climate \nChange (eds Shukla, P. R. et al.) Ch. 12 (Cambridge Univ. Press, \n2022).\n4.\t\nTracking Clean Energy Progress (IEA, 2022); https://www.iea.org/\ntopics/tracking-clean-energy-progress\n5.\t\nClimate Action Tracker (New Climate Institute and Climate \nAnalytics, 2023); https://climateactiontracker.org/\n6.\t\nBoehm, S. et al. State of Climate Action 2022 (WRI, 2022);  \nhttps://doi.org/10.46830/wrirpt.22.00028\n7.\t\nEmissions Gap Report 2022: The Closing Window \u2014 Climate Crisis \nCalls for Rapid Transformation of Societies (UNEP, 2022);  \nhttps://www.unep.org/emissions-gap-report-2022\n8.\t\nden Elzen, M. G. J. et al. Updated nationally determined \ncontributions collectively raise ambition levels but need \nstrengthening further to keep Paris goals within reach.  \nMitig. Adapt. Strateg. Glob. Change 27, 33 (2022).\n9.\t\nMeinshausen, M. et al. Realization of Paris Agreement pledges \nmay limit warming just below 2\u2009\u00b0C. Nature 604, 304\u2013309 (2022).\n10.\t Rogelj, J. et al. Credibility gap in net-zero climate targets leaves \nworld at high risk. Science 380, 1014\u20131016 (2023).\n11.\t\nRogelj, J., Geden, O., Cowie, A. & Reisinger, A. Net-zero emissions \ntargets are vague: three ways to fix. Nature 591, 365\u2013368 (2021).\n12.\t Dooley, K. et al. The Land Gap Report 2022 (Climate Resource, \n2022); https://www.landgap.org/\n13.\t Smith, H. B., Vaughan, N. E. & Forster, J. Long-term national \nclimate strategies bet on forests and soils to reach net-zero. \nCommun. Earth Environ. 3, 305 (2022).\n14.\t Buck, H. J., Carton, W., Lund, J. F. & Markusson, N. Why residual \nemissions matter right now. Nat. Clim. Change 13, 351\u2013358 (2023).\n15.\t Lund, J. F., Markusson, N., Carton, W. & Buck, H. J. Net zero and the \nunexplored politics of residual emissions. Energy Res. Soc. Sci. \n98, 103035 (2023).\n16.\t McLaren, D. P., Tyfield, D. P., Willis, R., Szerszynski, B. & Markusson, \nN. O. Beyond \u2018net-zero\u2019: a case for separate targets for emissions \nreduction and negative emissions. Front. Clim. 1, 4 (2019).\n17.\t Carton, W., Lund, J. F. & Dooley, K. Undoing equivalence: \nrethinking carbon accounting for just carbon removal.  \nFront. Clim. 3, 664130 (2021).\n18.\t Friedlingstein, P. et al. Global Carbon Budget 2023. Earth Syst. \nSci. Data 15, 5301\u20135369 (2023).\n19.\t Powis, C. M., Smith, S. M., Minx, J. C. & Gasser, T. Quantifying \nglobal carbon dioxide removal deployment. Environ. Res. Lett. \nhttps://doi.org/10.1088/1748-9326/acb450 (2023).\n20.\t Byers, E. et al. AR6 scenarios database hosted by IIASA. Zenodo \nhttps://doi.org/10.5281/zenodo.5886911 (2022).\n21.\t Gidden, M. J. et al. Aligning climate scenarios to emissions \ninventories shifts global benchmarks. Nature 624, 102\u2013108 (2023).\n22.\t Buylova, A., Fridahl, M., Nasiritousi, N. & Reischl, G. Cancel \n(out) emissions? The envisaged role of carbon dioxide removal \ntechnologies in long-term national climate strategies. Front. Clim. \n3, 675499 (2021).\n23.\t Grubler, A. et al. A low energy demand scenario for meeting the \n1.5\u2009\u00b0C target and sustainable development goals without negative \nemissions technologies. Nat. Energy 3, 515\u2013527 (2018).\n24.\t Soergel, B. et al. A sustainable development pathway for climate \naction within the UN 2030 Agenda. Nat. Clim. Change 11, 656\u2013664 \n(2021).\n25.\t Fuss, S. et al. Negative emissions\u2014Part 2: costs, potentials and \nside effects. Environ. Res. Lett. 13, 063002 (2018).\n26.\t Mo, L. et al. Integrated global assessment of the natural forest \ncarbon potential. Nature 624, 92\u2013101 (2023).\n27.\t Fyson, C. L. & Jeffery, M. L. Ambiguity in the land use component \nof mitigation contributions toward the Paris Agreement goals. \nEarths Future 7, 873\u2013891 (2019).\n28.\t Benveniste, H., Boucher, O., Guivarch, C., Treut, H. L. & Criqui, P. \nImpacts of nationally determined contributions on 2030 global \ngreenhouse gas emissions: uncertainty analysis and distribution \nof emissions. Environ. Res. Lett. 13, 014022 (2018).\n\nNature Climate Change | Volume 14 | June 2024 | 644\u2013651\n651\nAnalysis\nhttps://doi.org/10.1038/s41558-024-01984-6\n29.\t Strefler, J. et al. Carbon dioxide removal technologies are not \nborn equal. Environ. Res. Lett. 16, 074021 (2021).\n30.\t Fuss, S. et al. Betting on negative emissions. Nat. Clim. Change \nhttps://doi.org/10.1038/nclimate2392 (2014).\n31.\t Key\u00dfer, L. T. & Lenzen, M. 1.5\u2009\u00b0C degrowth scenarios suggest the \nneed for new mitigation pathways. Nat. Commun. 12, 2676 (2021).\n32.\t Geden, O. Climate advisers must maintain integrity. Nature 521, \n27\u201328 (2015).\n33.\t Bergman, A. & Rinberg, A. in CDR Primer (eds. Wilcox, J. et al.)  \nCh. 1 (Hemlock Printers, 2021).\n34.\t Carton, W., Hougaard, I., Markusson, N. & Lund, J. F. Is carbon \nremoval delaying emission reductions? WIREs Clim. Change \nhttps://doi.org/10.1002/wcc.826 (2023).\n35.\t Moe, E. & R\u00f8ttereng, J.-K. S. The post-carbon society: rethinking \nthe international governance of negative emissions. Energy Res. \nSoc. Sci. 44, 199\u2013208 (2018).\n36.\t Lamb, W. F. et al. Discourses of climate delay. Glob. Sustain. 3, e17 \n(2020).\n37.\t Painter, J. et al. Climate delay discourses present in global \nmainstream television coverage of the IPCC\u2019s 2021 report. \nCommun. Earth Environ. 4, 118 (2023).\n38.\t Grassi, G. et al. The key role of forests in meeting climate targets \nrequires science for credible mitigation. Nat. Clim. Change 7, \n220\u2013226 (2017).\n39.\t Grassi, G. et al. Carbon fluxes from land 2000\u20132020: bringing \nclarity to countries\u2019 reporting. Earth Syst. Sci. Data 14, 4643\u20134666 \n(2022).\n40.\t Giebink, C. L. et al. The policy and ecology of forest-based \nclimate mitigation: challenges, needs, and opportunities.  \nPlant Soil 479, 25\u201352 (2022).\n41.\t IPCC: Summary for Policymakers. In Climate Change and Land: \nAn IPCC Special Report on Climate Change, Desertification, Land \nDegradation, Sustainable Land Management, Food Security, and \nGreenhouse Gas Fluxes in Terrestrial Ecosystems (eds Shukla, P. R. \net al.) (Cambridge University Press, 2019).\n42.\t Kraxner, F. & Nordstr\u00f6m, E.-M. in The Future Use of Nordic Forests \n(eds Westholm, E. et al.) 63\u201381 (Springer, 2015).\n43.\t Hyyrynen, M., Ollikainen, M. & Sepp\u00e4l\u00e4, J. European forest \nsinks and climate targets: past trends, main drivers, and future \nforecasts. Eur. J. Forest Res. https://doi.org/10.1007/s10342-023-\n01587-4 (2023).\n44.\t Korosuo, A. et al. The role of forests in the EU climate policy: are \nwe on the right track? Carbon Balance Manage. 18, 15 (2023).\n45.\t Nemet, G., Greene, J., M\u00fcller-Hansen, F. & Minx, J. C. Dataset on \nthe adoption of historical technologies informs the scale-up of \nemerging carbon dioxide removal measures. Commun. Earth \nEnviron. 4, 397 (2023).\n46.\t Strefler, J. et al. Between Scylla and Charybdis: Delayed \nmitigation narrows the passage between large-scale CDR and \nhigh costs. Environ. Res. Lett. 13, 044015 (2018).\n47.\t Pr\u00fctz, R., Strefler, J., Rogelj, J. & Fuss, S. Understanding the carbon \ndioxide removal range in 1.5\u2009\u00b0C compatible and high overshoot \npathways. Environ. Res. Commun. 5, 041005 (2023).\n48.\t Realmonte, G. et al. An inter-model assessment of the role of \ndirect air capture in deep mitigation pathways. Nat. Commun. 10, \n3277 (2019).\n49.\t Fuhrman, J. et al. Diverse carbon dioxide removal approaches \ncould reduce impacts on the energy\u2013water\u2013land system.  \nNat. Clim. Change https://doi.org/10.1038/s41558-023-01604-9 \n(2023).\n50.\t Luderer, G. et al. Residual fossil CO2 emissions in 1.5\u20132\u2009\u00b0C \npathways. Nat. Clim. Change 8, 626\u2013633 (2018).\n51.\t Breyer, C. et al. Proposing a 1.0\u2009\u00b0C climate target for a safer future. \nPLoS Clim. 2, e0000234 (2023).\n52.\t Riahi, K. et al. Cost and attainability of meeting stringent climate \ntargets without overshoot. Nat. Clim. Change 11, 1063\u20131069 \n(2021).\n53.\t Emmerling, J. et al. The role of the discount rate for emission \npathways and negative emissions. Environ. Res. Lett. 14, 104008 \n(2019).\n54.\t van Vuuren, D. P. et al. Alternative pathways to the 1.5\u2009\u00b0C target \nreduce the need for negative emission technologies. Nat. Clim. \nChange 8, 391\u2013397 (2018).\n55.\t Fasihi, M., Efimova, O. & Breyer, C. Techno-economic assessment \nof CO2 direct air capture plants. J. Clean. Prod. 224, 957\u2013980 \n(2019).\n56.\t Andreoni, P., Emmerling, J. & Tavoni, M. Inequality repercussions \nof financing negative emissions. Nat. Clim. Change 14, 48\u201354 \n(2024).\n57.\t Fuhrman, J. et al. Food\u2013energy\u2013water implications of negative \nemissions technologies in a +1.5\u2009\u00b0C future. Nat. Clim. Change 10, \n920\u2013927 (2020).\n58.\t Crippa, M. et al. CO2 Emissions of All World Countries \u2013 2022 \nReport (European Commission, 2022); https://edgar.jrc.ec.europa.\neu/dataset_ghg70\n59.\t Forster, P. et al. in Climate Change 2021: The Physical Science Basis \n(eds Masson-Delmotte, V. et al.) Ch. 7 (Cambridge Univ. Press, \n2021).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2024\n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-01984-6\nMethods\nFollowing the IPCC and State of CDR reports, we defined CDR as \u201cHuman \nactivities capturing CO2 from the atmosphere and storing it durably in \ngeological, land or ocean reservoirs, or in products. This includes \nhuman enhancement of natural removal processes, but excludes \nnatural uptake not caused directly by human activities.\u201d1,2 Important \ncharacteristics of this definition are its unambiguous inclusion of both \nconventional land-based sinks and emerging CDR methods, as well as \nrequirements for durability and direct human intervention19.\nA wide array of CDR technologies have been developed, tested or \nare in practice today60. In this Analysis, we follow ref. 2 and categorize \nafforestation, reforestation, forest management, soil carbon sequestra-\ntion, wetland restoration and durable harvested wood products as \u2018con-\nventional CDR on land\u2019. \u2018Novel CDR\u2019 comprises all other CDR methods,  \nsuch as biochar as well as those that store carbon in the lithosphere \nincluding direct air carbon capture and storage (DACCS), bioenergy \ncarbon capture and storage (BECCS) and enhanced weathering.\nDirect versus indirect anthropogenic CDR\nWhereas novel CDR methods are solely the result of direct human inter-\nvention, land can remove CO2 from the atmosphere through a combina-\ntion of direct anthropogenic effects (such as land-use change, forest \nharvest and regrowth), indirect anthropogenic effects (such as fertiliza-\ntion because of elevated atmospheric CO2) and natural effects (such as \nclimate variability). These effects are impossible to disentangle through \nobservations but can be partitioned using earth system models61.  \nThe different treatment of indirect anthropogenic effects and of man-\naged land concepts are the main reasons for the major discrepancy \nbetween national inventories and global bookkeeping models used \nin the IPCC assessment reports62,63.\nTo keep consistency with the IPCC definition of CDR, we con-\nsidered CDR on land as only the net direct human-induced removal \ncomponent occurring in managed areas of forests and soils. (Note: \ndeforestation is human induced but is categorized as emissions, not \nCDR, and is therefore excluded). Defining CDR in this way orients poli-\ncymakers towards addressing those activities under their direct control \n(for example, forest and soil management practices) and avoids claims \non CDR that result from global factors outside their direct control (for \nexample, the CO2-fertilization effect).\nTo evaluate current conventional CDR on land on this basis, we \nstarted from the latest compilation of national LULUCF inventories39, \nconsidering all negative fluxes from forest land and other land uses as \nremovals. A global ratio of direct to indirect anthropogenic removals \nderived from ref. 19 was then applied to the forest land fluxes to remove \nthe indirect component. The resulting global and national levels of \ncurrent conventional CDR on land were then taken as the baseline for \nany changes observed in the NDCs and long-term strategies (described \nbelow). Where these documents describe an increase in conventional \nCDR on land compared with the baseline inventory, we considered \nthis increase as representing direct removals only. Where a decrease \nis described in the long-term strategies, we preserved the current ratio \nof direct to indirect removals. The final analysis considered direct \nanthropogenic removals only, as shown in Supplementary Fig. 2.\nCDR in national 2030 mitigation pledges\nSeveral assumptions need to be taken to extract CDR levels from NDCs. \nFirst, with the aim of identifying quantifiable conventional CDR on land \nand considering the frequent lack of LULUCF information in the NDCs, \nwe gathered as much official information as possible up to a cut-off \ndate of November 2023 (that is, COP28). This included not only NDCs, \nwhich we prioritize, but also other relevant national submissions to the  \nUNFCCC where mitigation targets and information on activities and flux  \ndisaggregation are usually included, such as long-term mid-century \nstrategies (for example, USA, Chile), National Communications (for \nexample, China, Japan, New Zealand), Biennial Update Reports (for \nexample, Peru) and Forest Reference Emission Levels (for example, \nMalaysia, Peru, Mexico). Whenever available, we also considered other \nnational documents, such as climate strategies (for example, Norway, \nChile, Thailand, the Philippines, Mexico, Peru), GHG projections (for \nexample, Brazil) or assessments of national targets (for example, India). \nWe prioritized documents by ranking countries according to their \ncontribution to global emissions and removals, using the PRIMAP \nHist-CR database64 and ref. 39. We searched for this information in \n111 of 195 countries reporting under the UNFCCC framework, exclud-\ning small island states, city states and countries with no or very low \nland-use fluxes.\nSecond, we followed different strategies to extract information \nfrom these documents and estimate the specific contribution of \nLULUCF removals to national pledges, depending on the level of trans-\nparency and information available for each country. As summarized in \nSupplementary Fig. 1, countries can be categorized into three groups:\n\u2022 \nGroup A: countries with the least amount of information  \nregarding their headline mitigation target and the contribution \nof LULUCF. For these countries, we assumed that removals in \n2030 remained consistent with the historic trend (2011\u20132020).  \n(n\u2009=\u200925, historic inventory-based gross removals (2011\u20132020)\u2009=\u2009 \n\u22120.38\u2009GtCO2\u2009yr\u22121, no additional CDR for 2030)\n\u2022 \nGroup B1: countries with a specified LULUCF target in 2030 but \nno information regarding the contribution of removals. We \nscaled the LULUCF target to the historic ratio of emissions and \nremovals (2011\u20132020). (n\u2009=\u200955, historic inventory-based gross \nremovals\u2009=\u2009\u22123.45\u2009GtCO2\u2009yr\u22121, additional conditional CDR for \n2030\u2009=\u2009\u22120.22\u2009GtCO2\u2009yr\u22121)\n\u2022 \nGroup B2: countries with a specified LULUCF target in 2030 and \nwith information on the specific contribution of removals. We \ndirectly reported these removals in our analysis. (n\u2009=\u200931, historic \ninventory-based gross removals (2011\u20132020)\u2009=\u2009\u22123.87\u2009GtCO2\u2009yr\u22121, \nadditional conditional CDR for 2030\u2009=\u2009\u22120.33\u2009GtCO2\u2009yr\u22121).\nIt is relevant to note that the national extra removals (that is, CDR) \nare presented here as the difference between committed removals in \n2030 (un/conditional) and countries\u2019 average removals for the previous \ndecade (2011\u20132020). This approach offers high temporal coherence \nbetween countries\u2019 emissions and mitigation commitments in 2030.\nHistorical averages of removals were based on an update (July \n2023) of a compiled database of national GHG inventories obtained \nfrom UNFCCC submissions39. Emissions were calculated as the sum of \nall positive GHG fluxes detailed in ref. 39 (that is, forest, deforestation, \norganic soils and other), while removals are the sum of all negative \nfluxes. Most emissions come from deforestation and organic soils, \nwhile most removals come from forests. The category \u2018other\u2019 is either a \nremoval or an emission, depending on the country as it includes other \nnon-forest land uses (croplands, grasslands, wetlands, settlements). \nSince not all countries contribute similarly to global mitigation targets, \nbelow we provide more insights for several key countries, with addi-\ntional examples in Supplementary Information Section 1.\nBrazil: There are several possible scenarios for Brazil\u2019s LULUCF \ncommitments in 2030, but none of them are described in their latest \nNDC (2022). One scenario is presented in the Low Carbon Agriculture \nProgramme (ABC and ABC+), but their targets are considered obsolete \n(year 2014). We therefore used Brazil\u2019s national mitigation projections \nand mitigation options for 2030 and 2050 published by the Minis-\ntry of Science and Technology in 201765. This official report includes \nland-use net emissions for its business-as-usual (BAU) trajectory (2030) \n(298\u2009MtCO2e) and two commitment scenarios based on two difference \nprices for mitigation investment (270 and 189 MtCO2e as uncondi-\ntional and conditional LULUCF emissions in 2030). Historical removals \n(2011\u20132020) are about \u2212400\u2009MtCO2\u2009yr\u22121, while the extra removals (that \nis, CDR) under conditional commitments that we estimated in 2030 \n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-01984-6\nare about \u221245\u2009MtCO2\u2009yr\u22121. Brazil is a Category B1 country (numerical \nLULUCF target with unspecified removals).\nIndonesia: Its 2022 NDC submission is highly informative on \nLULUCF quantitative targets (\u2212500 and \u2212729\u2009MtCO2e unconditional and \nconditional committed emissions) and projected BAU (714\u2009MtCO2e). \nThe NDC also describes a list of mitigation activities disaggregated \nbetween emission avoidance and removals, to support their claims. \nHistorical removals (2011\u20132020) are about \u2212370\u2009MtCO2\u2009yr\u22121, while \nthe extra removals (that is, CDR) under conditional commitments we \nestimated in 2030 are about \u2212165\u2009MtCO2\u2009yr\u22121. Indonesia is a Category \nB2 country (numerical LULUCF target with specified removals).\nChina: The latest NDC includes the target to \u2018increase forest stock \nvolume by around 6 billion cubic metres in 2030 from the 2005 level\u2019, \nwhich is not easily translated into a CO2 sink value. However, LULUCF \ntargets are better covered in the Third Biennial Update Report 3 (2018), \nwhere forest sink projections are specified for 2030 as BAU (2030) \n\u2212410\u2009MtCO2e (range: 390\u2013430\u2009MtCO2e), and two forest sink targets are \npresented under two scenarios of action that preserve the same com-\nmitment for forests sinks (\u2212495\u2009MtCO2e) (range: 470\u2013520\u2009MtCO2e). To \nallow the comparison of the target with the LULUCF historical trend, \nforest sink targets were then complemented by the average sink of \nother non-forest land uses for the period 2011\u20132020, raising the com-\nmitted sink to \u2212806\u2009MtCO2e\u2009yr\u22121 for 2030. Due to China\u2019s current large \nsink of about \u22121,135\u2009MtCO2e (average 2011\u20132020), their LULUCF targets \nfor 2030 translated into a weakening of removals, that is, an increase \nin net emissions (~326\u2009MtCO2\u2009yr\u22121), which markedly reduced the global \nLULUCF sink commitments. China is a Category B2 country (numerical \nLULUCF target with specified removals).\nCDR in national 2050 mitigation pledges\nTo calculate CDR in national 2050 mitigation pledges, we relied on \ninformation in the long-term strategies as analysed in refs. 13,66, read-\ning all submissions up to November 2023 (that is, COP28). We identi-\nfied the subset of these that have quantified scenarios describing how \nthey will reach their stated climate objective (for example, net-zero \nGHG emissions). Often, these scenarios are depicted in a figure or a \ntable, where the contribution of novel or conventional CDR on land is \nincluded as a portion of the mitigation effort. If the long-term strategy \ndid not include such quantitative material, we assumed that any cur-\nrent removals (that is, from conventional CDR on land) are sustained \nuntil 2050. As in the NDCs, most countries describe the total LULUCF \nflux in their scenarios, rather than providing a breakdown of emissions \nand removals in this sector. We counted the entirety of these fluxes in \n2050 as removals. In other words, we assumed zero deforestation. This \nassumption is consistent with the text and framing of the long-term \nstrategies. For example, no countries describe deforestation in their \nscenarios and several of them, such as Cambodia and Colombia, explic-\nitly pledge zero deforestation. However, we acknowledge that it is a \nsimplification.\nIn the case of the European Union, we discarded all member-state \ndocuments and instead relied on modelling studies performed by the \nEuropean Commission describing EU-wide pathways to net zero by the \nmid-century67. While these were published before the United Kingdom \nformally left the European Union, we continued to include the United \nKingdom\u2019s long-term strategy separately.\nScenario selection and re-analysis\nOur selection of IAM scenarios drew from the latest IPCC AR6 vetted  \nscenario database20. We used the C1 and C3 scenario categories, which \ntogether are referred to as \u2018below 2\u2009\u00b0C scenarios\u2019 in the main manu-\nscript. These scenarios could be considered as those most relevant \nto, but not necessarily all consistent with, the Paris Agreement tem-\nperature goal.\nWe used the scenario re-analysis provided in ref. 21 that splits emis-\nsions and removals in the land-use sector. Their analysis was conducted \nby running the OSCAR bookkeeping model using variables reported in \nthe AR6 scenario database, including forest land area, cropland area \nand forestry activity, to evaluate the direct anthropogenic removals on \nmanaged land. These scenario projections followed and extended the \nexperimental setup used for the 2021 Global Carbon Budget68.\nData availability\nThe data for this Analysis are available via Zenodo at https://doi.org/ \n10.5281/zenodo.10821849 (ref. 69). All raw and processed data are freely \naccessible, except for complete national-level CDR estimates in 2030 \n(that is, from the NDCs and other national documents) which will be \nmade available upon reasonable request. Source data are provided \nwith this paper.\nCode availability\nThe code for this Analysis is available via Zenodo at https://doi.org/ \n10.5281/zenodo.10821849 (ref. 69).\nReferences\n60.\t Minx, J. C. et al. Negative emissions\u2014Part 1: research landscape \nand synthesis. Environ. Res. Lett. 13, 063001 (2018).\n61.\t Gasser, T. & Ciais, P. A theoretical framework for the net \nland-to-atmosphere CO2 flux and its implications in the definition \nof \u2018emissions from land-use change\u2019. Earth Syst. Dyn. 4, 171\u2013186 \n(2013).\n62.\t Grassi, G. et al. Critical adjustment of land mitigation pathways \nfor assessing countries\u2019 climate progress. Nat. Clim. Change 11, 14 \n(2021).\n63.\t Grassi, G. et al. Harmonising the land-use flux estimates of global \nmodels and national inventories for 2000\u20132020. Earth Syst.  \nSci. Data 15, 1093\u20131114, 2023.\n64.\t G\u00fctschow, J. & Pfl\u00fcger, M. The PRIMAP-hist national historical \nemissions time series (1750\u20132021) v2.4.2. Zenodo https://doi.org/ \n10.5281/zenodo.7727475 (2023).\n65.\t Mitigation Paths and Policy Instruments to Reach Brazilian Goals in \nthe Paris Agreement (MCIT, 2017); https://www.gov.br/mcti/pt-br/\nacompanhe-o-mcti/sirene/publicacoes/acordo-de-paris-e-ndc/\narquivos/pdf/trajetoriasebookb_final.pdf\n66.\t Smith, H., Vaughan, N. E. & Forster, J. Navigating Net Zero: \nAnalysing Residual Emissions in Long-Term National Climate \nStrategies. Preprint at https://doi.org/10.2139/ssrn.4678157 \n(2024).\n67.\t In-Depth Analysis in Support on the COM(2018) 773 (European \nCommission, 2018); https://climate.ec.europa.eu/system/\nfiles/2019-08/long-term_analysis_in_depth_analysis_figures_ \n20190722_en.pdf\n68.\t Friedlingstein, P. et al. Global carbon budget 2021. Earth Syst.  \nSci. Data 14, 1917\u20132005 (2022).\n69.\t Lamb, W. The carbon dioxide removal gap dataset (version 1) \n[Data set]. Zenodo https://doi.org/10.5281/zenodo.10821849 \n(2024).\nAcknowledgements\nThis work was supported by the European Union ERC-2020-SyG \n\u2018GENIE\u2019 (951542) grant (W.F.L., J.C.M., G.N., T.G., M.J.G., Y.P., J.S., \nK.R.); the UK Natural Environment Research Council \u2018CO2RE Hub\u2019 \n(NE/V013106/1) grant (S.M.S.); the European Union Horizon 2020 \n\u2018ESM2025\u2019 (101003536) and \u2018RESCUE\u2019 (101056939) grants (T.G.); the \nGerman Federal Ministry of Education and Research \u2018CDRSynTra\u2019 \n(01LS2101A) (J.C.M., O.G.) and \u2018ASMASYS\u2019 (01LS2101A) grants (O.G.).\nAuthor contributions\nW.F.L., G.N., S.M.S., O.G., K.R. and J.C.M conceived the idea for  \nthe paper. W.F.L., T.G., R.M.R.-C., G.G., M.J.G., C.M.P., Y.P., J.S., N.E.V. \nand H.S. contributed to data gathering and the analysis. W.F.L.  \n\nNature Climate Change\nAnalysis\nhttps://doi.org/10.1038/s41558-024-01984-6\nwrote the paper. All authors contributed to drafting, reviewing and \nediting the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s41558-024-01984-6.\nCorrespondence and requests for materials should be addressed to \nWilliam F. Lamb.\nPeer review information Nature Climate Change thanks Andres \nClarens, Wim Carton and the other, anonymous, reviewer(s) for their \ncontribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\n\n    Scientific Research Findings:", "answer": "In our study we found that compared with 2020, the most ambitious national proposals for CDR imply an additional 0.5 GtCO2yr\u20131 of removals by 2030, and 1.9 GtCO2yr\u20131 by 2050. Compared with CDR scaling in Paris Agreement-consistent scenarios, we found that these national CDR proposals tend to fall short by hundreds of megatonnes of carbon dioxide in 2030 to several gigatonnes of carbon dioxide in 2050, highlighting a \u2018CDR gap\u2019. However, we find that the most ambitious proposals do come close to levels in a low-energy-demand scenario where CDR requirements are minimized, suggesting that if countries pledge more ambitious emissions reductions consistent with these scenarios, the CDR gap will be closed. As levels of reporting vary, our evaluation of proposed CDR does assume that a number of countries simply maintain their current levels of (conventional) removals. In addition, it remains unknown to what extent firm CDR policies will follow these proposals.", "id": 42}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 14 | May 2024 | 476\u2013481\n476\nnature climate change\nArticle\nhttps://doi.org/10.1038/s41558-024-01972-w\nModel-based financial regulations impair the \ntransition to net-zero carbon emissions\nMatteo Gasparini\u2009\n\u200a\u20091,2\u2009\n, Matthew C. Ives\u2009\n\u200a\u20091,2, Ben Carr3,4, Sophie Fry\u2009\n\u200a\u20095  \n& Eric Beinhocker\u2009\n\u200a\u20092,5\nInvestments via the financial system are essential for fostering the green \ntransition. However, the role of existing financial regulations in influencing \ninvestment decisions is understudied. Here we analyse data from the \nEuropean Banking Authority to show that existing financial accounting \nframeworks might inadvertently be creating disincentives for investments \nin low-carbon assets. We find that differences in the provision coverage ratio \nindicate that banks must account for nearly double the loan loss provisions \nfor lending to low-carbon sectors as compared with high-carbon sectors. \nThis bias is probably the result of basing risk estimates on historical data. \nWe show that the average historical financial risk of the oil and gas sector \nhas been consistently estimated to be lower than that of renewable energy. \nThese results indicate that this bias could be present in other model-based \nregulations, such as capital requirements, and possibly impact the ability of \nbanks to fund green investments.\nThe urgency of climate change has not always been matched by the \npace of action by governments. However, increasing concerns about \nclimate-induced financial instability and stranded assets1\u201319 have led \nsome academics and financial regulators to advance a set of possible \npolicy changes to help catalyse the green transition20\u201328. While vari-\nous policies aimed at assessing climate-related financial risks\u2014which \ncould possibly indirectly stimulate the net-zero carbon transition\u2014have \nbecome widespread in recent years (for example, climate stress testing, \nclimate-related risk disclosure)29\u201332, financial policies aimed at directly \nfostering green investments have not always gained traction among \npolicymakers (for example, differentiated capital requirements). Yet,  \na largely neglected question in this literature and among policymakers \nis whether existing financial regulations could be negatively contribut-\ning to the net-zero carbon transition.\nThis paper assesses whether widely used model-based risk \nregulations might create disincentives for financial institutions to \ndivest their portfolios from high-carbon assets. Such financial regu-\nlations have extensively required banks to use statistical models for \nassessing firms\u2018 and investments\u2018 financial risk for various purposes  \n(for example, financial stability). For example, capital requirements \n(for example, Basel III/IV) aim to force banks to hold higher capital \nbuffers for investments that are \u2018estimated\u2019 to be riskier. Accounting \nrules (for example, IFRS9) appraise the \u2018fair value\u2019 of outstanding loans \non banks\u2019 balance sheets, reducing their net value by the amount of \n\u2018estimated\u2019 expected losses. These regulatory frameworks affect key \nmetrics of financial institutions, which ultimately influence manage-\nment incentives and resource allocation33\u201336.\nWe focus on financial accounting rules, which are a key driver of \nthe profitability of banks, and leverage model-based estimates of risk. \nA key measure in this framework is loan loss reserves (LLR), which is \nan allowance for potential future losses from outstanding loans. Due \nto the structure of double-entry accounting, LLR are liabilities which \nnet the valuation of assets by the amount of their expected losses. Any \nchange in LLR results in loan loss provision (LLP) charges, which are a \n\u2018present\u2019 cost of the \u2018future expected\u2019 credit losses from outstanding \nloans (ECL). When there is any change in these model-based estimates \nof risk, banks are expected to account for any estimated financial losses \nbefore they occur. In turn, differentials in any of these estimates may \ninfluence banks\u2019 profitability, management behaviour and resource \nallocation (see Supplementary Information 4).\nReceived: 26 July 2023\nAccepted: 1 March 2024\nPublished online: 2 April 2024\n Check for updates\n1Smith School of Enterprise and The Environment, University of Oxford, Oxford, UK. 2Institute for New Economic Thinking, University of Oxford, Oxford, \nUK. 3Grantham Research Institute, London School of Economics, London, UK. 4Bloomberg L.P., Enterprise Products, London, UK. 5Blavatnik School of \nGovernment, University of Oxford, Oxford, UK. \n\u2009e-mail: matteo.gasparini@ouce.ox.ac.uk\n\nNature Climate Change | Volume 14 | May 2024 | 476\u2013481\n477\nArticle\nhttps://doi.org/10.1038/s41558-024-01972-w\nSimulation of a divestment strategy\nDue to the model-based risk estimates of PCR required by the account-\ning regulation, the performance of financial institutions would be sub-\nstantially impacted if they were to swiftly shift their portfolio away from \nhigh-carbon to other investments. Our modelling shows that if banks \nhad to stop lending to firms in high-carbon sectors and lend only to \nlow-carbon ones, the portfolio average PCR would need to increase by \nmore than 100 basis points (1%) across most institutions in the European \nbanking sector (Fig. 1). This effect is consistent for most banks in our \nsample and across various nations, except for a few institutions with \nlow PCR for high-carbon assets. Banks in countries with the largest \ndifference in PCR between high-carbon and low-carbon assets would \nbe hit substantially more according to our analysis. Most financial \ninstitutions would be affected by this shift regardless of their size, but in \nline with our empirical observations, banks in the smaller size quartile \nwould be impacted more than others (2.35% increase compared with \n0.9% simple average).\nWe estimate that a shift in investments away from high-carbon to \nlow-carbon assets would require a loan-weighted average increase of \n35% of LLR for banks in the European Union (Fig. 2). This result is con-\nsistent after controlling for bank size and country of headquarter. The \ndecision to divest from high-carbon assets could lead to more than dou-\nbling of provisions for some banks in our sample and could thus have \nmaterial impacts on the bank\u2019s stock market valuations. The increase \nin LLR will not only depend on the difference between the estimated \nexpected loss from lending to low-carbon and high-carbon activi-\nties, but also on the share of high-carbon loans. The higher the share \nof current outstanding loans towards high-carbon firms, the more \npronounced the impact on LLR given a certain level of difference in  \nPCR. This relationship further exacerbates the potential impact of a \nTo examine the impact of such financial accounting rules, we \nuse data from the European Banking Authority (EBA) transparency \nexercise, which provides the amount of LLR and outstanding loans of \nsupervised banks in the European Union by economic sector (defined \nas Nomenclature of Economic Activities (NACE) rev2 level 1). We com-\nbine these data with the results of the EBA risk assessment exercise, \nwhich reports the average exposure towards climate policy relevant \nsectors (CPRS)5 within each NACE level 1 section, to classify sectors \nas \u2018high carbon\u2019 or \u2018low carbon\u2019. We classify sectors with a share of \nCPRS higher than 95% as high carbon and provide a set of robustness \nanalyses. We are particularly interested in the ratio of LLR over the \nvalue of outstanding loans, which is a proxy of banks\u2019 estimates of \nexpected credit losses. This measure is often called provision cover-\nage ratio (PCR).\nOur empirical analysis allows us to observe that model-based \nestimates of risk are lower for high-carbon sectors compared with \nlow-carbon ones. We then provide an assessment of the implications \nof this observation for some key financial metrics of banks if they had \nto divest from high-carbon assets. Specifically, we utilize the account-\ning relationships among some of these metrics to show that an active \ndivestment from high-carbon assets could be costly for banks. We \nargue that this, in turn, could create perverse incentives impairing \nthe shift of financial resources from high-carbon to low-carbon assets, \npossibly including much needed investments in renewable energy. \nFinally, we provide some possible explanations as to why some of \nthese models may lead to estimates that are negatively correlated with \ncarbon emissions.\nResults\nOur analysis shows that in 2021, the average PCR of banks in the EU was \nsubstantially lower for high-carbon (1.8%) than low-carbon sectors \n(3.4%), as reported in Table 1. Such a difference has substantial implica-\ntions for banks\u2019 return on capital and profitability, and therefore heavily \ninfluences management incentives and behaviours. Our analysis shows \nthat this result is consistent for banks of different portfolio sizes and \nacross countries of the banks\u2019 headquarters, with the only exception \nbeing Italy. Looking at the results by the size of banks, this effect is \nexacerbated for smaller financial institutions in absolute terms, but \nin relative terms, there is no correlation between the difference in PCR \nand bank size. This finding is also consistent across countries, regard-\nless of the large variation in terms of absolute PCR between Nordic and \nSouthern/ Eastern European regions.\nThese results emerge from banks\u2019 statistical models based on his-\ntorical information as required by the accounting framework. Standard \nbackward-looking risk models can show a high-carbon portfolio to be \nrelatively low risk, even if there is a possibility of a rapid transition to \ngreen energy (see Discussion). Although it is arguably difficult to take \nan objective stance on the correct estimate of risk for these investments \non a forward-looking basis, our analysis is sufficient to show that the \nstructure of model-based risk frameworks may have an unintended side \neffect that is potentially in conflict with the purpose of the regulations \nor other societal goals. By affecting financial institutions\u2019 incentives, \nmodel-based financial regulations may create perverse outcomes pos-\nsibly leading to more investments in polluting activities.\nSimulating the effect of a divestment from high-carbon activities \nand a re-investment in low-carbon sectors allows us to better under-\nstand the effects of such action on banks\u2019 financial metrics and the \nlinked management incentives, which ultimately affect behaviours \nand resources allocation. Specifically, as indicated by the account-\ning rules, we assume that if a bank had to divest from high-carbon \nsectors and re-invest the proceeds in low-carbon sectors, the PCR of \nsuch investments would need to increase to the higher level of the \nlatter (Table 1). This would in turn lead to a higher level of loan loss \nprovisions and higher costs due to the structure of the accounting  \nrules (see Methods).\nTable 1 | PCR for high-carbon and low-carbon investments \nfor European banks\nPCR \nlow-carbon \nsectors (%)\nPCR \nhigh-carbon \nsectors (%)\nNumber of \nbanks\nTotal sample\n3.40\n1.80\n59\nLoan \nbook size \n(quartile)\n0\u201325\n7.29\n3.28\n14\n25\u201350\n3.68\n2.01\n15\n50\u201375\n2.82\n1.13\n15\n75\u2013100\n3.20\n1.94\n15\nCountry\nAustria\n2.98\n2.12\n3\nBelgium\n3.70\n2.27\n2\nDenmark\n1.92\n0.60\n3\nFinland\n1.46\n1.17\n2\nFrance\n3.24\n1.98\n9\nGermany\n2.08\n1.00\n12\nGreece\n12.49\n6.52\n4\nHungary\n4.85\n3.53\n1\nIreland\n5.02\n4.95\n2\nItaly\n4.75\n4.87\n7\nNetherlands\n2.42\n1.04\n4\nPortugal\n6.73\n3.25\n1\nSpain\n3.48\n2.42\n5\nSweden\n0.71\n0.44\n4\nExposure-weighted average PCR for sectors classified as high carbon and low carbon for 59 \nof the largest European banks participating in the EBA transparency exercise, representing \n93% of total banking exposure as of June 2021. PCR defined as the ratio of LLR over value \nof outstanding loans. The table reports the breakdown by bank size (quartile of total loan \noutstanding) and country of the bank headquarter.\n\nNature Climate Change | Volume 14 | May 2024 | 476\u2013481\n478\nArticle\nhttps://doi.org/10.1038/s41558-024-01972-w\ndivestment for banks more exposed to high-carbon sectors, creating \ngreater hysteresis in investing in high-carbon sectors and contributing \nfurther to the build-up of risk in assets that could become stranded in \na green transition.\nThe increased PCR, LLR and the resulting LLP charges driven by \na potential divestment strategy could weigh substantially on banks\u2019 \nnet profits. An increase in LLR not only impacts the liability side of \nthe balance sheet, but also the income statement through decreased \nprofits. To simulate this effect, we take the absolute increase in loan \nloss provisions, and we compare it to each bank\u2019s cumulative profits \nfrom 2016 to 2020. We select 5\u2009years of profits to smooth possible bad \nyears or extraordinary items in the financial reporting and to provide \na stable baseline for our counterfactual analysis.\nWe estimate that for some banks, the transition could cost as much \nas 5\u2009years of profits over the divestment horizon and, on an outstanding \nloan weighted average basis, 15% of the previous 5\u2009years of profits due \nto a large increase in LLR (Fig. 3). The total sum of banks\u2019 lost profits due \nto the increase in provisions following a divestment from high-carbon \n6\n2\n\u20134\n\u20132\n4\n\u20136\n8\n10\n12\nBanca Popolare di Sondrio\nHamburg Commercial Bank AG\nEurobank\nSkandinaviska Enskilda Banken\nHSBC Continental Europe\nAlpha Bank\nAareal Bank AG\nABN AMRO Bank N.V.\nBanco Comercial Portugu\u00eas, SA\nBNG Bank N.V.\nJyske Bank A/S\nAIB Group plc\nNordea Bank Abp\nICCREA BANCA S.P.A.\nLa Banque Postale\nRaifeisen bankenen\nGroupe BPCE\nSoci\u00e9t\u00e9 g\u00e9n\u00e9rale\nCassa Centrale\nLandesbank Baden-W\u00fcrttemberg\nErste Group Bank AG\nDekaBank Deutsche\nPiraeus Financial Holdings\nBanco Santander, S.A.\nBayerische Landesbank\nCommerzbank\nDZ BANK\nCaixaBank, S.A.\nCo\u00f6peratieve Rabobank U.A.\nOTP-csoport\nNorddeutsche Landesbank\nBankinter, S.A.\nOP Osuuskunta\nSwedbank -group\nCr\u00e9dit Mutuel\nBNP Paribas\nNykredit Realkredit A/S\nSvenska Handelsbanken\nUniCredit S.p.A.\nDanske Bank A/S\nGroupe Cr\u00e9dit Agricole\nDeutsche Bank\nKBC Groep\nBBVA\nBpifrance\nING Groep N.V.\nKommuninvest -group\nBanca Monte dei Paschi\nDeutsche Pfandbriefbank AG\nLandesbank\nBanco de Sabadell, S.A.\nRaifeisen Bank International AG\nBank of Ireland Group plc\nBanco BPM SpA\nIntesa Sanpaolo S.p.A.\nBelfius Bank\nDeutscher sparkassen\nNational Bank of Greece, S.A.\nRCI Banque\nChange in coverage ratio (%)\nEU banks\n\u00d8 100bps\nChange in coverage ratio (%)\nGross exposure (EURm)\na\n\u20136\n\u20134\n\u20132\n0\n2\n4\n6\n8\n10\n12\n0\n50,000\n100,000\n150,000\n200,000\n250,000\n300,000\n350,000\n400,000\n450,000\nGroupe Cr\u00e9dit Agricole\nBanco Santander, S.A.\nGroupe BPCE\nCr\u00e9dit Mutuel\nUniCredit S.p.A.\nIntesa Sanpaolo S.p.A.\nING Groep N.V.\nSoci\u00e9t\u00e9 g\u00e9n\u00e9rale\nCo\u00f6peratieve Rabobank U.A.\nDeutsche Bank\nBBVA\nCaixaBank, S.A.\nNordea Bank Abp\nSvenska Handelsbanken\nDanske Bank A/S\nBayerische Landesbank\nCommerzbank\nSkandinaviska Enskilda Banken\nDZ BANK\nErste Group Bank AG\nLandesbank Baden-W\u00fcrttemberg\nKBC Groep\nLandesbank\nABN AMRO Bank N.V.\nBanco BPM SpA\nBanco de Sabadell, S.A.\nNykredit Realkredit A/S\nSwedbank -group\nBNG Bank N.V.\nRaifeisen Bank International AG\nICCREA BANCA S.P.A.\nNorddeutsche Landesbank\nBpifrance\nDeutscher sparkassen\nBanca Monte dei Paschi\nOP Osuuskunta\nBelfius Bank\nLa Banque Postale\nBNP Paribas\nJyske Bank A/S\nHSBC Continental Europe\nBank of Ireland Group plc\nDeutsche Pfandbriefbank AG\nAareal Bank AG\nAIB Group plc\nCassa Centrale\nAlpha Bank\nEurobank\nBankinter, S.A.\nPiraeus Financial Holdings\nRaifeisen bankenen\nBanco Comercial Portugu\u00eas, SA\nHamburg Commercial Bank AG\nBanca Popolare di Sondrio\nNational Bank of Greece, S.A.\nOTP-csoport\nRCI Banque\nDekaBank Deutsche\nKommuninvest -group\nb\nEU banks\n0\nGermany\nSweden\nAustria\nBelgium\nDenmark\nFinland\nFrance\nGreece\nHungary\nIreland\nItaly\nNetherlands\nPortugal\nSpain\nGross exposure (EURm)\nFig. 1 | Change in PCR for 59 largest European banks. Absolute percentage \nchange in PCR following a divestment from high-carbon assets and \ncorresponding re-investment in low-carbon assets, maintaining a constant \nlevel of outstanding loans by bank. Colours represent the country of bank\u2019s \nheadquarters. The change in PCR represents the difference between PCR \nrequired for low-carbon as opposed to high-carbon assets, for each bank in our \nsample. Horizontal line represents the average in basis points (bps). a, Banks \nranked by absolute change in PCR. b, The same information ranked by gross loan \nexposure (largest to smallest from left to right).\n50\n60\n70\n80\n90\n100\n0\n10\n20\n30\n40\n100\n\u2013100\n\u201350\n0\n50\n150\n200\n250\n300\nDenmark\nGermany\nSweden\nBelgium\nFrance\nFinland\nNetherlands\nAustria\nGreece\nIreland\nSpain\nPortugal\nHungary\nItaly\nGross exposure (\u20ac100 billion)\nRelative increase in Loan Loss Reservest (%)\nShare of high\u2013carbon sectors (%)\nFig. 2 | Relative increase in LLR. Relative increase in LLR following a divestment \nfrom high-carbon assets and corresponding re-investment in low-carbon assets, \nmaintaining a constant level of outstanding loans. Horizontal axis represents \ncurrent share of high-carbon sector outstanding loans (June 2021). Bubbles \nrepresent banks in our sample, colour coded on the basis of the country of \nheadquarters. Bubble size represents the total value of outstanding loans. \nRelative increase in LLR represents the absolute increase in LLR over the level of \nLLR as of June 2021. Results are gross-exposure weighted. Horizontal line is the \nweighted average by gross exposure across banks (35%).\n\nNature Climate Change | Volume 14 | May 2024 | 476\u2013481\n479\nArticle\nhttps://doi.org/10.1038/s41558-024-01972-w\nassets could be of the order of \u20ac28 billion (considering the 59 largest \nbanks in the European Union). This is only a rough estimate as it does \nnot account for (1) how such divestment could affect other investments \nin a network of interconnected banks (indirect effects), (2) whether \nsufficient alternative investment opportunities are available to the \nbanks or (3) the impact on the costs and prices of alternative energy \ngeneration options resulting from changes in the investments in those \noptions. However, this figure is useful to assess the materiality of our \nfindings. The European Central Bank (ECB) estimates that the impact \nof physical risk and transition risk could be around \u20ac17 billion and \u20ac53 \nbillion, respectively, in a short-term scenario for the 41 largest banks \nin the European Union.\nAlthough there are a few instances of banks that experience \nhigher profits due to their lower estimate of risk for low-carbon \nthan high-carbon sectors, our results show consistently that most \nbanks\u2019 profits would be negatively impacted by a divestment from \nhigh-carbon assets. Our findings are also robust to the classification \nof specific sectors as high carbon. It is the prevalence of the lowest \nPCR among the high-carbon sectors, in general terms, that drives our \nkey results. We found that relabelling some selected sectors between \nhigh-carbon and low-carbon clusters does not alter the main outcome \nof our study, although the magnitude of the impact can change (Sup-\nplementary Information 1). This sensitivity test provides us with confi-\ndence that sectors with particularly low (high) levels of PCR among the \nhigh-carbon (low-carbon) sectors are not driving our results.\nWe then simulate the impact of allocating each sector partially \nto the low-carbon and to the high-carbon cluster depending on their \nmedian share of CPRS found among banks in the European Union tak-\ning part in the EBA risk assessment exercise. This robustness analysis \nsimulates a partial divestment of only the high-carbon portion of invest-\nments in each NACE level 1 and allows us to better investigate the het-\nerogeneity of high-carbon/low-carbon sectors within each NACE level 1 \nsection. This is because the underlying CPRS classification leverages a \nmuch more granular sectoral classification (NACE level 4), which better \ncaptures whether economic activities are high carbon or low carbon. \nOnce again, we find that our main results persist. Moreover, our results \nare robust after controlling for different time periods. If we use quar-\nterly average levels from March 2020 to June 2022 (maximum temporal \ndepth of the data), the impacts are similar (100% increase in PCR, 33% \nincrease in provisions, 14% impact on previous 5-years profits).\nThe robustness of our results highlights that our findings are not \na function of the specific high-carbon/low-carbon classification used \nbut driven by a lower average estimated risk for high-carbon sectors \ncompared with low-carbon ones. As long as the structure of the regula-\ntion foresees that (1) losses are costs that are accounted for as expected \ncosts as opposed to incurred costs and (2) provision coverage ratios are \nproportional to model-based estimates of risk, then divesting from an \n\u2018estimated\u2019 low-risk asset and re-investing in an \u2018estimated\u2019 high-risk \nasset mechanically leads to higher costs in the income statement. \nIndeed, despite not being able to use carbon emissions data directly, \nin our Discussion and Methods, we provide strong evidence for a nega-\ntive correlation between CPRS/emission intensity of assets and risk esti-\nmates (Supplementary Fig. 1). This in turn leads to a confirmation of our \nconclusion that there probably exists an implicit incentive structure that \nmight inadvertently favour assets involved in high-carbon activities.\nDiscussion\nThe bias shown towards high-carbon assets identified in this paper \nprobably emerges from the backward-looking nature of risk estimates. \nThat is, it is the outcome of using models that rely on the historical \nrelationship between a firm\u2019s financial performance and past risk as \na predictor of future risk. As discussed in the literature and by policy-\nmakers, such models are useful but may not be well suited to capturing \nuncertain macro-economic outcomes when there are structural breaks \nor non-marginal changes in the system, such as the clean energy transi-\ntion. In these risk-based models, the creditworthiness of firms is often \n\u2013220\n60\n\u201360\n\u2013140\n\u201320\n\u2013100\n20\n\u2013180\nCr\u00e9dit Mutuel\nBanca Popolare di Sondrio\nAIB Group plc\nJyske Bank A/S\nKBC Groep\nBNP Paribas\nSkandinaviska Enskilda Banken\nSwedbank -group\nOP Osuuskunta\nBpifrance\nBNG Bank N.V.\nAlpha Bank\nDZ BANK\nAareal Bank AG\nBankinter, S.A.\nBanco Santander, S.A.\nBayerische Landesbank\nGroupe BPCE\nABN AMRO Bank N.V.\nBanco Comercial Portugu\u00eas, SA\nGroupe Cr\u00e9dit Agricole\nCo\u00f6peratieve Rabobank U.A.\nIntesa Sanpaolo S.p.A.\nBelfius Bank\nSoci\u00e9t\u00e9 g\u00e9n\u00e9rale\nErste Group Bank AG\nOTP-csoport\nLandesbank\nDekaBank Deutsche\nBanco de Sabadell, S.A.\nCaixaBank, S.A.\nKommuninvest -group\nLa Banque Postale\nNykredit Realkredit A/S\nRaifeisen Bank International AG\nDeutsche Pfandbriefbank AG\nRCI Banque\nSvenska Handelsbanken\nDanske Bank A/S\nCassa Centrale\nING Groep N.V.\nBBVA\nRaifeisen bankenen\nNordea Bank Abp\nBanco BPM SpA\nICCREA BANCA S.P.A.\nLandesbank\nBank of Ireland Group plc\nShare of net profits lost (%)\nEU banks\nFig. 3 | Impact on average 5\u2009years net profits. Impact on net profits following \na divestment from high-carbon assets and corresponding re-investment in low-\ncarbon assets, maintaining a constant level of outstanding loans by banks. Bars \nrepresent the share of cumulative 2016\u20132020 profits lost due to the required \nincrease in LLR. The impact represents the ratio of absolute increase in LLR over \nthe cumulative profits between 2016 and 2020. The horizontal line represents the \naverage loss of profits (\u221215%).\n\nNature Climate Change | Volume 14 | May 2024 | 476\u2013481\n480\nArticle\nhttps://doi.org/10.1038/s41558-024-01972-w\nestimated through financial ratios measuring profitability (for example, \nEarnings Before Interest, Taxes (EBIT)/Revenue), solvency (for example, \nDebt/Asset, Interest/EBITDA (Earnings Before Interest, Taxes, Deprecia-\ntions and Amortizations)) and liquidity (for example, short-term debt/\nworking capital). If these ratios have been historically favourable for \nhigh-carbon firms, as previous research has highlighted37, risk models \nwill probably produce favourable outcomes for this type of investment. \nThis phenomenon might arguably limit investments in green assets if \ntheir past risk estimates have been relatively high.\nTo illustrate this, we use a simple analysis based on a dataset of \n228 oil and gas, and 235 renewable energy firms worldwide and finan-\ncial information between 2010 and 2021, retrieved from Bloomberg \n(Supplementary Information 2). We use this dataset as a representa-\ntive sample of some of the most relevant sectors in the high-carbon \nand low-carbon clusters. We construct some financial ratios that are \ncommonly used in risk assessment to investigate the origins of risk \nestimate differentials. We then contrast them to infer the likely relative \nmagnitude between these two important sectors in the net-zero carbon \ntransition. The average share of interest expenses over EBITDA for \nthe period 2010\u20132021 is lower for oil and gas (16%) than for renewable \nenergy firms (32%), and the average debt over asset ratio is lower for oil \nand gas (31%) than for renewable energy (42%) (Fig. 4a). Similarly, the \noutcome of one such model retrieved from Bloomberg shows consist-\nently higher average estimates of risk (expressed in terms of probability \nof default) for renewable energy than for oil and gas between 2010 and \n2021 (Fig. 4b). This highlights how historically, investing in the former \nmight have been less risky compared with investing in the latter, due \nto the higher solvency and lower indebtedness.\nThese ratios have been a good proxy of the historical creditwor-\nthiness of firms and have been used extensively by financial analysts. \nHowever, problems arise if these historical metrics are not representa-\ntive of the future, following a change in the probability distribution \nof losses38. For example, we estimate that if there were an increase \nin the average global level of carbon tax enforced on Scope 1 and 2 \nemissions to US$100 (or climate policies with an equivalent shadow \ncarbon price), the ratio of interest expenses over EBITDA for oil and \ngas firms might increase substantially above the ratio of renewable \nenergy companies (from 16% to 46% against 32% for renewable energy). \nSimilarly, a partial write-off of oil reserves valuations in the balance \nsheet of oil and gas companies of US$20 per barrel might result in an \nincrease in the debt to asset ratio of these firms, much higher than \nthe average value observed among renewable energy companies \n(from 16% to 86% against 32% for renewable energy). In such case, \nfinancial ratios and the resulting risk estimates might become lower \nfor renewable energy investments. A more forward-looking frame-\nwork which includes scenario analyses that consider climate-related \nrisks might be better suited to capturing such unprecedented  \nemerging risks.\nIn conclusion, our results suggest that model-based financial \nregulations, and in particular accounting rules, might disincentivize \nbanks from divesting from high-carbon sectors by directly impact-\ning their profitability. This side effect of the rules might impair the \ntransition towards net-zero carbon emissions and in turn contribute \nto increasing the build-up of transition risk in the financial system. \nOur comparison of financial ratios between oil and gas and renewable \nenergy firms indicates that this effect might penalize investments in \nclean energy. Current financial accounting practices might uninten-\ntionally hinder the shift of funds required for the green transition, \nespecially in Europe where these investments are often provided by \nthe banking sector. While the desire to promote a green transition \nmay be based on broader social objectives that lie beyond the remit \nof financial regulators, the deeper problem for regulators is that this \ntransition could represent a potential source of systemic risk. Broader \nresearch is needed to determine whether the existing regulations suf-\nficiently account for any such emerging sources of systemic risks that \nmight accompany the green transition. More research is also needed \nto shed light on whether this bias might be present in other similar \nmodel-based frameworks (for example, capital requirements). Finally, \nregulators and investors should investigate risk models that include \nforward-looking assessments of climate and energy transition risk to \nensure that those risks are appropriately incorporated in decisions and \nto remove any inadvertent bias.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-024-01972-w.\nReferences\n1.\t\nSemieniuk, G. et al. Stranded fossil-fuel assets translate to major \nlosses for investors in advanced economies. Nat. Clim. Change 12, \n532\u2013538 (2022).\nAfter $100 CO2\ncarbon tax\nAverage 2010\u20132021\nAfter $20 per barrel\noil reserves write-of\nAverage 2010\u20132021\n16%\n32%\n32%\n46%\n31%\n42%\n86%\n42%\nOil and gas\nRenewable energy\nTotal borrowing/total assets\nInterest expenses/EBITDA\n2010\n15\n12\n11\n13\n12%\n17\n14\n16\n18\n19\n20\n2021\n0%\n2%\n4%\n6%\n8%\n10%\nProbability of default oil and gas (5yr)\nProbability of default renewable energy (5yr)\na\nb\nYears\nFig. 4 | Comparison of financial ratios between oil and gas and renewable \nenergy industries. a, Average of 5 years interest expenses over EBITDA and total \nborrowing over total assets of 228 oil and gas and 235 renewable energy firms \nin our sample between 2010 and 2021. Left: simulation of the impact of US$100 \ncarbon tax on EBITDA expressed in terms of average interest expenses over \nEBITDA ratio. Right: impact of US$20 per barrel write-off of oil reserves on total \nassets expressed in terms of average total borrowing over total assets. b, Average \nof Bloomberg 5 years Probability of Default (PD) estimate through time for the \ncompanies in the sample.\n\nNature Climate Change | Volume 14 | May 2024 | 476\u2013481\n481\nArticle\nhttps://doi.org/10.1038/s41558-024-01972-w\n2.\t\nLamperti, F., Bosetti, V., Roventini, A. & Tavoni, M. The public costs \nof climate-induced financial instability. Nat. Clim. Change 9, \n829\u2013833 (2019).\n3.\t\nMercure, J. F. et al. Macroeconomic impact of stranded fossil fuel \nassets. Nat. Clim. Change 8, 588\u2013593 (2018).\n4.\t\nDietz, S., Bowen, A., Dixon, C. & Gradwell, P. Climate value at  \nrisk\u2019 of global financial assets. Nat. Clim. Change 6, 676\u2013679 \n(2016).\n5.\t\nBattiston, S., Mandel, A., Monasterolo, I., Sch\u00fctze, F. & Visentin, G. \nA climate stress-test of the financial system. Nat. Clim. Change 7, \n283\u2013288 (2017).\n6.\t\nWeyzig, F., Kuepper, B., van Gelder, J. W. & van Tilburg, R. The  \nPrice of Doing Too Little Too Late (Green European Foundation, \n2014).\n7.\t\nD\u2019Orazio, P., Hertel, T. & Kasbrink, F. No Need to Worry? Estimating \nthe Exposure of the German Banking Sector to Climate-related \nTransition Risks Ruhr Economic Papers, No. 946 (Leibniz-Institut \nf\u00fcr Wirtschaftsforschung, 2022).\n8.\t\nBattiston, S., Guth, M., Monasterolo, I., Neudorfer, B. & Pointner, \nW. Austrian Banks\u2019 Exposure to Climate-related Transition Risk \nFinancial Stability Report 40 (Oesterreichische Nationalbank, \n2020).\n9.\t\nCampiglio, E. & van der Ploeg, R. Macrofinancial risks of the \ntransition to a low-carbon economy. Rev. Environ. Econ. Policy 16, \n173\u2013195 (2022).\n10.\t Campiglio, E., Monnin, P. & von Jagow, A. Climate Risks in Financial \nAssets (Council on Economic Policies, 2019).\n11.\t\nClimate-Related Financial Risks: A Survey on Current Initiatives \n(BIS, 2020).\n12.\t Di Virgilio, S., Faiella, I., Mistretta, A. & Narizzano, S. Assessing \nCredit Risk Sensitivity to Climate and Energy Shocks (Bank of Italy, \n2023).\n13.\t Caldecott, B., Clark, A., Koskelo, K., Mulholland, E. & Hickey, C. \nStranded assets: environmental drivers, societal challenges, and \nsupervisory responses. Annu. Rev. Environ. Resour. 46, 417\u2013447 \n(2021).\n14.\t Alogoskoufis, S. et al. ECB\u2019s Economy-Wide Climate Stress Test \n(European Central Bank, 2021).\n15.\t Monasterolo, I. Climate change and the financial system. Annu. \nRev. Resour. Econ. 12, 299\u2013320 (2020).\n16.\t Roncoroni, A., Battiston, S., Escobar-Farf\u00e1n, L. O. L. & \nMartinez-Jaramillo, S. Climate risk and financial stability in the \nnetwork of banks and investment funds. J. Financ. Stabil. 52, \n100870 (2021).\n17.\t Sen, S. & von Schickfus, M. T. Climate policy, stranded assets, and \ninvestors\u2019 expectations. J. Environ. Econ. Manage. 100, 102277 \n(2020).\n18.\t Vermeulen, R. et al. The heat is on: a framework for measuring \nfinancial stress under disruptive energy transition scenarios. Ecol. \nEcon. 190, 107205 (2021).\n19.\t Gasparini, M., Baer, M. & Ives, M. C. A re-evaluation of the financial \nrisks of the net zero transition. SSRN https://doi.org/10.2139/\nssrn.4254054 (2023).\n20.\t Campiglio, E. et al. Climate change challenges for central banks \nand financial regulators. Nat. Clim. Change 8, 462\u2013468 (2018).\n21.\t Dafermos, Y. & Nikolaidi, M. Greening Capital Requirements \n(Centre for Sustainable Finance, Grantham Research Institute on \nClimate Change and the Environment, 2022).\n22.\t Philipponnat, T. Breaking the Climate\u2013Finance Doom Loop \n(Finance Watch, 2020).\n23.\t Bolton, P., Despres, M., Pereira da Silva, L. A., Samama, F. \n& Svartzman, R. The Green Swan (Bank for International \nSettlements, 2020).\n24.\t Campiglio, E. Beyond carbon pricing: the role of banking and \nmonetary policy in financing the transition to a low-carbon \neconomy. Ecol. Econ. 121, 220\u2013230 (2016).\n25.\t Alessi, L., Di Girolamo, F., Pagano, A. & Petracco Giudici, \nM. Accounting for Climate Transition Risk in Banks\u2019 Capital \nRequirements (European Commission, 2022).\n26.\t Dafermos, Y. & Nikolaidi, M. How can green differentiated capital \nrequirements affect climate risks? A dynamic macrofinancial \nanalysis. J. Financ. Stabil. 54, 100871 (2021).\n27.\t Diluiso, F., Annicchiarico, B., Kalkuhl, M. & Minx, J. C. Climate \nactions and macro-financial stability: the role of central banks.  \nJ. Environ. Econ. Manage. 110, 102548 (2021).\n28.\t Dunz, N., Naqvi, A. & Monasterolo, I. Climate sentiments, \ntransition risk, and financial stability in a stock-flow consistent \nmodel. J. Financ. Stabil. 54, 100872 (2021).\n29.\t Ameli, N., Kothari, S. & Grubb, M. Misplaced expectations from \nclimate disclosure initiatives. Nat. Clim. Change 11, 917\u2013924 \n(2021).\n30.\t Ameli, N., Drummond, P., Bisaro, A., Grubb, M. & Chenet, H. \nClimate finance and disclosure for institutional investors: why \ntransparency is not enough. Clim. Change 160, 565\u2013589 (2020).\n31.\t Goldstein, A., Turner, W. R., Gladstone, J. & Hole, D. G. The private \nsector\u2019s climate change risk and adaptation blind spots. Nat. Clim. \nChange 9, 18\u201325 (2019).\n32.\t Edwards, I., Yapp, K., Mackay, S. & Mackey, B. Climate-related \nfinancial disclosures in the public sector. Nat. Clim. Change 10, \n588\u2013591 (2020).\n33.\t Gropp, R., Mosk, T., Ongena, S. & Wix, C. Banks response to higher \ncapital requirements: evidence from a quasi-natural experiment. \nRev. Financ. Stud. 32, 266\u2013299 (2019).\n34.\t Glancy, D. & Kurtzman, R. How do capital requirements affect loan \nrates? Evidence from high volatility commercial real estate. Rev. \nCorp. Financ. Stud. 11, 88\u2013127 (2022).\n35.\t Beatty, A. & Liao, S. Financial accounting in the banking industry: \na review of the empirical literature. J. Account. Econ. 58, 339\u2013383 \n(2014).\n36.\t Aymanns, C., Caccioli, F., Farmer, J. D. & Tan, V. W. C. Taming the \nBasel leverage cycle. J. Financ. Stabil. 27, 263\u2013277 (2016).\n37.\t Schmidt, T. Low-carbon investment risks and de-risking. Nat. Clim. \nChange 4, 237\u2013239 (2014).\n38.\t Holscher, M., Ignell, D., Lewis, M. & Stiroh, K. Climate Change and \nthe Role of Regulatory Capital: A Stylized Framework for Policy \nAssessment Finance and Economics Discussion Series 2022-068 \n(Board of Governors of the Federal Reserve System, 2022).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this licence, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2024\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-01972-w\nMethods\nData\nWe used data from the 2021 EBA transparency exercise, which provides \nportfolio-level information of banks\u2019 gross exposure and accumulated \nprovisions (LLR) by NACE sector level 1 at the end of June 2021. We used the  \nmost recent data, but with additional robustness analysis, ensured \nthat the results do not change using different years (the reader should \nnote that due to the structure of this modelling, the provision cover-\nage ratios oscillate with time in level but the relative difference across \nsectors is generally preserved). NACE is a standard classification of \nsectors in the European Union. It has various levels of granularity from 1 \n(least granular) to 4 (most granular), and the EBA transparency exercise \nrelies on this classification. The exercise is an annual data collection to \nfoster transparency and to complement banks\u2019 own disclosures. The \ndata published includes 111 EU banks across 25 countries and provides \ninformation regarding banks\u2019 assets, liabilities, loan loss provisions and \nother financial information for each bank.\nWe used the legal entity identifier (LEI) code in the EBA dataset to \ncomplement this information with the historical net profit data from \nBloomberg. The data identifiers were matched with each LEI code in \nour sample through manual research on the Bloomberg terminal. We \nstarted from the largest 60 banks in our sample representing 95% of the \ntotal banking exposure, but we excluded one bank because its name and \nLEI code were missing, which did not allow us to retrieve their income \ninformation. This bank represents ~2% of total EU banking assets. \nAfter this manipulation, our dataset covered more than 93% of total \nbanking loans in the European Union and provided us with LLR, total \nlending amount for all NACE sectors (level 1) and cumulative net profits \nfrom 2016 to 2021 for the largest 59 banks in the EU. A summary of the \nsector-level statistics is reported in Supplementary Table 1.\nHigh-carbon sectors classification\nWe added to this dataset the information necessary to classify sectors \nas high carbon (that is, sectors with high levels of emission intensity). \nSpecifically, we complemented the data with the results of the EBA \nRisk Assessment exercise, which provides median values of CPRS as \ndefined in ref. 5 within each NACE level 1. CPRS is a classification used \nto assess the exposure of investments to transition risks, including \ncarbon taxation, and is a proxy for the level of carbon emissions asso-\nciated with an investment. The exercise was carried out by the EBA \nand a sample of 29 volunteer banks from 10 countries representing \n50% of the total EU banking assets, with the objective of obtaining a \npreliminary quantification of the exposure of banks to climate-related \nrisks, particularly focusing on transition risk. The data annex provided \n(publicly available) discloses the share of CPRS sectors in each NACE \nlevel 1 section according to banks\u2019 classification of their own clients in \nCPRS. This information is particularly useful because it allows us to have \na more granular labelling of low-carbon and high-carbon sectors than \nthe NACE level 1 (which would not be sufficient to address the heteroge-\nneity of some sectors). The CPRS rely on NACE level 4, which provides \na better discrimination between climate-sensitive sectors and others \n(additional information provided in Supplementary Information 2).\nThe bank-level information on total gross loan amount and LLR \nby NACE code were grouped into high-carbon and low-carbon sectors. \nWe defined sectors as \u2018high-carbon\u2019 if they had a median share of CPRS \nhigher than 95%, as reported by banks in the EBA Risk Assessment exer-\ncise. This gave us the following high-carbon sectors and their respective \ncodes: A - Agriculture, forestry and fishing; B - Mining and quarrying; \nD - Electricity, gas, steam and air conditioning supply; E - Water supply, \nsewerage, waste management; H - Transport and storage; and L - Real \nestate activities. We acknowledge that our approach has limitations, \nbut we extensively tested the robustness of our results to a change in \nthe methodology used to classify low-carbon and high-carbon sec-\ntors (Supplementary Information 1). Moreover, we compared our \nclassification to more granular data reporting emission intensity to \nprovide transparency about their level of correlation. It should be \nnoted that the banks participating in the climate risk exercise did not \ninclude Sweden, Denmark and Norway, but results do not change if \nthose countries are excluded due to their relatively low materiality in \nthe overall EU banking system.\nData availability prevented us from assigning carbon emissions \nto loans directly. However, the CPRS classification we used is highly \ncorrelated with GHG emissions intensity (Supplementary Fig. 1). The \nEBA Risk Assessment provides a breakdown of emission intensity by \npercentiles for CPRS and non-CPRS. They use individual firms\u2019 GHG \nemissions from the data provider Trucost (representing 30% of total \nbanks\u2019 loan amount) and a proxy based on the average GHG emission \nintensity at NACE rev2 level 4 for the remaining loan amount. Each bank \nloan is classified in percentiles of emission intensity in a range from \nvery low to very high (more details could be found in the EBA 2020 Risk \nAssessment Report, Table 19). We used these data to test the correlation \nbetween the share of loan amount in CPRS/non-CPRS and its emission \nintensity. There is a clear correlation between the share of loan amount \nof CPRS and the clusters of emission intensity (Supplementary Fig. 1). \nAround 85% of the loan amount classified as having \u2018very high\u2019 emis-\nsion intensity are in CPRS. At the opposite end of the spectrum, only \n8% of the loan amount of CPRS are in the \u2018very low\u2019 emission intensity \nbucket. The correlation between the share of loan amount in CPRS \n(non-CPRS) and its emission intensity is therefore strongly positive \n(negative) and around 90% (\u221290%). In Supplementary Information 1, \nwe show that this correlation is very unlikely to change with different \nclassifications using a set of robustness analyses.\nSimulation of a divestment strategy\nUsing the data available, we could provide an estimate of the poten-\ntial impact of a divestment from high-carbon assets on EU banks\u2019 \nfinancials. The primary assumption in this estimation is that the total \namount of loans of each bank is left unvaried. In other words, the simu-\nlation assumes that banks shift their lending portfolio directly from \nhigh-carbon to low-carbon investments. We also assume that sufficient \nlow-carbon investments are available for these transactions. The label-\nling in our data allowed us to calculate the average risk estimate (PCR) \nof low-carbon and high-carbon sectors for all banks in our sample. We \nmade use of the accounting relationship between provisions cover-\nage ratio, LLP charges and net profits to assess the impact of a divest-\nment from high-carbon assets on these metrics (all else being equal). \nImportantly, we did not rely on an explicit economic model, but on the \naccounting relationship among these metrics. In turn, our results were \ngenerated by the structure of the regulation as long as a bank divests \nfrom a low-PCR asset and re-invests in a high-PCR asset.\nIt should be noted that LLP changes are only the direct effect of \nthis divestment on bank\u2019s net profit changes at the time they make \nthe investment. This is an expected loss, not necessarily a loss that \nwill occur in the future. More specifically, three conditions need to be \nsatisfied to generate an increase in costs from a divestment by banks:\n1.\t Losses are costs that must be accounted for as \u2018expected\u2019 as \nopposed to \u2018incurred\u2019. That is, financial firms must account \nfor any change in the portfolio expected losses, not the actual \nincurred losses;\n2.\t Provision coverage ratios must be equal to model-based \nestimates of \u2018expected losses\u2019. That is, expected losses are \nproportional to measures of risk;\n3.\t Risk estimates of the asset in which a bank is divesting are \nlower than the asset in which it is making a new investment;\nConditions (1) and (2) are provided by the structure of the regula-\ntion and replicated in the stylized analysis proposed in this paper (see \nSupplementary Information 4). Evidence supporting condition (3) \nis provided in our empirical analysis and further corroborated in the \nanalyses described in the Discussion and Supplementary Information 2.  \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-01972-w\nIn particular, from the three conditions above, it emerges that the \nresults of the simulation are grounded on PCR differentials. For this \nreason, we paid particular attention to demonstrating a negative cor-\nrelation between high-carbon sectors and risk measures.\nMore formally, we defined the PCR as the LLR (or accumulated \nprovisions in EBA terminology) divided by the gross exposure for \nthe high-carbon and low-carbon sectors i for each bank j. The PCR \nrepresents the expected credit loss (of non-default counterparties) \nand the corresponding loan loss provisions which banks must allocate \nto lending activities in each sector. This measure is assumed to be the \nmodel-based output from each institution risk model, in line with the \naccounting regulation:\nPCRi, j =\nLoan loss reservesi, j\nGross exposurei, j\n.\n(1)\nWe then calculated the change in the level of LLR following a divest-\nment from high-carbon assets. This was performed by assuming that \nall low-carbon loans replacing the high-carbon ones would require the \naverage PCR of existing low-carbon assets. In other words, a divestment \nfrom low-PCR assets and re-investment in high-PCR assets would lead \nto an increase in the overall average PCR. More formally, the increase/\ndecrease in provision for bank j is defined as follows:\nLoan loss provision chargesj = \u0394Loan loss reservesj\n= (PCRlow\u2212carbon, j \u2212PCRhigh\u2212carbon, j) \u00d7 Gross exposurehigh\u2212carbon, j\n.\n(2)\nThis result provides the expected increase or decrease in provi-\nsions if a bank had to shift the totality of its assets from high-carbon \nto low-carbon investments. This relationship is an accounting identity \ndefined by the framework. The impact of additional loan loss provisions \non a particular bank\u2019s income statement is considered an LLP \u2018charge\u2019 \n(that is, additional cost) with direct effect on their net profit. In par-\nticular, the increase in provisions (that is, the LLP charges) is directly \ndeducted from net profit, being an additional cost for the bank in the \nfiscal year of the divestment. This in turn provides a direct estimate \nof the change in net profits following a divestment from high-carbon \nassets. More formally:\nNet profitj,t+1 = Net profitj,t \u2212Loan loss provision chargesj,\n(3)\nwhere j refers to each bank in our sample, t is the starting point period \nand t\u2009+\u20091 is the period post divestments. Importantly, to simulate the \neffect of the divestment, we assumed it to occur entirely in one fiscal \nyear. This divestment would probably be spread across multiple years, \nbut frontloading the entire impact allows us to better investigate the \nimplicit incentive structure created by the regulation. This simple \napproach allowed us to simulate what would be the impact of a divest-\nment from high-carbon assets on banks\u2019 balance sheets and income \nstatements, testing the hypothesis that a potential divestment strategy \nmight be costly, disincentivizing banks from taking such action.\nData availability\nThe data used for the analyses and the results have been deposited in \nZenodo at https://doi.org/10.5281/zenodo.10632853 (ref. 39).\nLoan data, including loan loss reserves and gross exposure, were \nextracted from the EBA transparency exercise website available at \nthis link (https://www.eba.europa.eu/risk-analysis-and-data/eu-wide- \ntransparency-exercise). Financial information, including banks\u2019 prof-\nits, and oil and gas and renewable energy companies\u2019 financials were \nretrieved from the dataset Bloomberg and can be shared only with \nBloomberg\u2019s permission. Source data are provided with this paper.\nReferences\n39.\t Gasparini, M. Dataset paper model-based financial regulations \nimpair the transition to net zero carbon emissions. Zenodo  \nhttps://doi.org/10.5281/zenodo.10632853 (2024).\nAcknowledgements\nThe project has received no external funding. We thank P. Tufano \n(Harvard Business School), C. Hepburn (University of Oxford), J. Stock \n(Harvard University) and R. Barker (University of Oxford, ISSB) for \nfeedback on this research; the discussants and participants at the \nEAERE conference and various people at the European Central Bank \n(ECB); the International Sustainability Standards Board (ISSB), and the \nInternational Monetary Fund (IMF) for feedback on this research.\nThis paper only reflects the views of the authors. Any organizations \nthey may be affiliated with do not accept any liability for opinions \nexpressed in it.\nAuthor contributions\nM.G. contributed to conceptualizing the study, carrying out the \nanalyses and writing of the paper. M.C.I., B.C., S.F. and E.B. contributed \nto the writing of the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains \nsupplementary material available at  \nhttps://doi.org/10.1038/s41558-024-01972-w.\nCorrespondence and requests for materials should be addressed to \nMatteo Gasparini.\nPeer review information Nature Climate Change thanks the \nanonymous reviewers for their contribution to the peer review of this \nwork.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\n\n    Scientific Research Findings:", "answer": "We find that under the current regulations, if 59 of the largest banks in the European Union (EU) were to divest from high-carbon sectors and reinvest in other activities, they would record, on average, losses equivalent to about 15% of their previous 5 years\u2019 profits. We show that this is due to the increase in loan loss provisions required to cover the higher estimated risk of low-carbon-intensity activities, compared with high-carbon-intensity activities. We show that the average estimate of risk (expressed in terms of the ratio between loan loss reserves and outstanding loans) among EU banks is lower for carbon-intensive activities as opposed to low-carbon activities (1.8% and 3.4%, respectively, in 2021). This is likely to be due to the backward-looking structure of model-based risk estimates that fails to adequately incorporate recent policy changes, the declining costs of low-carbon technologies and other ongoing factors. We argue that this creates disincentives for banks to invest in new low-carbon assets and exposes them to future risks from high-carbon assets.", "id": 43}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 14 | April 2024 | 353\u2013356\n353\nnature climate change\nhttps://doi.org/10.1038/s41558-024-01952-0\nArticle\nGlobal corporate tax competition challenges \nclimate change mitigation\nYuwan Duan\u2009\n\u200a\u20091, Zengkai Zhang\u2009\n\u200a\u20092\u2009\n, Yuze Li\u2009\n\u200a\u20093, Shouyang Wang\u2009\n\u200a\u20094,5,6, \nCuihong Yang4,5 & Yi Lu\u2009\n\u200a\u20097\nMany countries have cut their corporate tax rates in the past decades to \nattract foreign investment. To prevent this, a global minimum tax policy was \napproved by OECD countries in 2021. Global changes in corporate tax rates \ncould reshape production and investment networks while impacting welfare \nand global emission patterns. Here we develop a theoretical multi-country \nmulti-industry general equilibrium model and show that global corporate \ntax competition during 2005\u20132016 would increase global carbon emissions \nand shift more emissions to developing economies. Implementing a \nglobal minimum tax rate of 15% would reduce global carbon emissions and \neffectively decrease the developing economies\u2019 emissions. The results \nhighlight that corporate tax policies should be coordinated with climate \nregulations.\nOver the past two decades, an increasing number of countries have \nbeen engaging in a \u2018race to the bottom\u2019 on corporate tax rates. \nSome countries have even implemented a zero tax rate to attract \ninternational capital inflows. Multinational enterprises (MNEs) may \ntake advantage of differing tax regimes between jurisdictions and \neffectively avoid paying taxes. To prevent this, a global minimum \ntax, the second pillar of the overall global tax reform agreement, \nwas approved by over 130 countries and jurisdictions in October \n2021. The reform is designed to ensure that MNEs are subject to a \nminimum tax rate of 15% in every country of operation starting in \n2023. As MNEs play a central role in shaping countries\u2019 production \npatterns, the global race to the bottom in corporate tax and the global \nminimum tax reform may reshape the global production network, \nadversely impacting economic and environmental development. Cor-\nporate tax-cutting may even offset the burden of climate regulations \nand challenge CO2 mitigation. For instance, developing countries,  \nwhich usually have greater carbon intensity and lower corporate  \ntax rates, tend to be tax havens1. Consequently, the global tax race \nhas led to more production moving to developing countries and \nreshaping the global production network, which could increase \nglobal carbon emissions.\nA regression analysis of corporate tax rate cuts and changes in \ncarbon emissions at the country level suggests a negative correla-\ntion between them. This analysis controlled for population, gross \ndomestic product (GDP) per capita and energy use per unit of GDP, \nand encompasses 56 countries over a 10-year period (Supplementary \nInformation 1). Results suggest that corporate tax competition could \nhinder climate cooperation and present challenges to climate change \nmitigation efforts2. Corporate tax competition influences the geo-\ngraphical distribution of production and global emissions by influenc-\ning foreign investment flows and trade flows, which are the two main \nchannels of international carbon transfer. Intuitively, a corporate tax \ncut in one country would increase the net profits of firms located in \nthis country and attract more multinational production (MP) activi-\nties, which may further stimulate more exports. Meanwhile, changes \nin corporate tax rates also affect government revenue, which in turn \ncan have an impact on a country\u2019s consumption patterns and overall \noutput. The production allocation would further change global and \nregional CO2 emissions. To comprehensively quantify the emission \neffect of the global corporate tax changes, we build a multi-country \nmulti-industry general equilibrium model by incorporating MNEs, \ninternational trade flows and corporate tax. We calibrate the model \nReceived: 8 January 2023\nAccepted: 6 February 2024\nPublished online: 28 March 2024\n Check for updates\n1School of International Trade and Economics, Central University of Finance and Economics, Beijing, China. 2State Key Laboratory of Marine \nEnvironmental Science, College of the Environment and Ecology, Xiamen University, Xiamen, Fujian, China. 3Questrom School of Business, Boston \nUniversity, Boston, MA, USA. 4Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing, China. 5School of Economics \nand Management, University of Chinese Academy of Sciences, Beijing, China. 6School of Entrepreneurship and Management, ShanghaiTech University, \nShanghai, China. 7School of Economics and Management, Tsinghua University, Beijing, China. \n\u2009e-mail: zengkaizhang@xmu.edu.cn\n\nNature Climate Change | Volume 14 | April 2024 | 353\u2013356\n354\nArticle\nhttps://doi.org/10.1038/s41558-024-01952-0\nFigure 1 shows the change in regional CO2 emissions caused by global \ncorporate tax competition.\nThere is a significant negative correlation between changes in CO2 \nemission (in percentages) and corporate tax cuts. Intuitively, corporate \ntax cuts would lead to an increase in the economy\u2019s output and subse-\nquently result in higher local CO2 emissions. The simulation results \nindicate that China experienced the most substantial increase in emis-\nsions in absolute terms, with emissions rising by 0.7%. The corporate \ntax rate in China has decreased from 33% in 2005 to 25% in 2016, which \nhas attracted more global production to move to China and thereby led \nto an increase in industrial output and emissions. Similarly, emerging \neconomies, such as India and Vietnam, also experienced corporate \ntax cuts and increasing CO2 emissions. The CO2 emissions of emerging \ncountries are sensitive to changes in corporate tax rates. Meanwhile, \nsome developed economies, such as Japan, Italy and Germany, also \ncut their corporate tax sharply. Germany\u2019s CO2 emissions increased \nby 1.8%, as its corporate tax decreased from 38.3% in 2005 to 29.7% in \n2016. In contrast, the United States\u2019s corporate tax remained consistent \nover the studied period. As a result, it experienced a decrease in both \noutput and emission (decreased by 1.2%).\nWelfare is a composite measure encompassing both real income \nand the disutility arising from global carbon emissions, following the \nframework outlined in ref. 22. Real income is measured as the national \nincome divided by the consumption price. We find that the global \ncorporate tax competition from 2005 to 2016 resulted in a 0.029% \nincrease in the weighted average of real income of different economies \nby GDP. The corporate tax cut has reduced production distortions, real-\nlocated production to the regions with high marginal labor products \nand improved global economic efficiency, which brought real income \ngains. Meanwhile, the global corporate tax cut raised global carbon \nemissions, negatively impacting national welfare. The changes in real \nincome notably outweighed the disutility brought about by carbon \nemission, and the weighted average welfare of different regions by \nGDP in the benchmark year increased by 0.011%.\nThe emission impact mechanism of global tax \ncompetition\nCorporate tax competition alters regional production in two ways: first, \nby exerting an influence on final consumption via its impact on govern-\nment revenue and wage income; second, by engendering a reconfigu-\nration of the global production chain through its effects on MP and \ntrade patterns. We decompose the emission effect into scale effect and \nstructural effect using equation (18). The scale effect captures the emis-\nsion changes attributable to shifts in countries\u2019 consumptions, while \nthe structural effect reflects the emission changes stemming from the \nchanges in MP and trade patterns across economies (the share of an \neconomy\u2019s production in each of other economies\u2019 total expenditures). \nFigure 2 depicts our decomposition results.\nThe simulation results show that for most regions, the scale effect \noverwhelmingly dominates their respective emission changes. To \nillustrate, the scale and structural effect increased China\u2019s territorial \nCO2 emissions by 0.6% and 0.1%, respectively. This implies that the \nreconfiguration of the global production structure triggered by global \ntax competition led to a marginal 0.1% rise in China\u2019s emissions, while \nchanges in income patterns across countries caused a more substantial \n0.6% increase. Nonetheless, in economies characterized by steady or \nescalating tax rates, such as the United States, the scale effect precipi-\ntated a reduction in territorial emissions. International MP and trade \nactivities generate a spillover effect across regions in terms of both \nproduction and emissions. Consequently, in certain economies such as \nMexico, even though corporate tax rates decreased from 2005 to 2016, \nthe scale effect still increases its emissions. This could be attributed \nto the fact that the increase in import partners\u2019 income stimulates \nproduction in Mexico through trade connections, thereby leading to \nincreased production and emissions.\nand quantify the economic and emission impacts of two global corpo-\nrate tax shocks: the global corporate tax competition and the global \nminimum tax policy approved by OECD (Organisation for Economic \nCooperation and Development) countries in 2021.\nThere are three main strands of literature relevant to our study. \nThe first strand of literature analyses the causal relationship between \nenvironmental policy and foreign investment theoretically3,4, or empiri-\ncally using reduced-form regressions5\u201310. Most of these studies focus on \npollution rather than climate change; moreover, they only capture the \npartial equilibrium effect of direct foreign investment while ignoring \nthe general equilibrium effect. More importantly, these studies have \nyet to investigate the effect of international tax competition on global \nemissions. The second strand of studies investigates the welfare conse-\nquences of the interaction between trade and multinational production \nin general equilibrium frameworks11\u201315. However, these studies do not \ninclude environmental issues. The third strand of studies focuses on \nthe intersection of trade and the environment using a general or partial \nequilibrium framework. However, these studies fail to consider the role \nof multinational production16\u201321.\nThe emission and welfare effect of corporate tax \ncompetition\nTo quantify the emission and welfare effect of global tax competition, \nwe first calibrate our general equilibrium model to 2016, which we \ndenote as our benchmark equilibrium. On the basis of the benchmark \nequilibrium, we conduct our first counterfactual by forcing all econo-\nmies\u2019 corporate tax changes as they actually did in 2005\u20122016 while \nkeeping other exogenous variables constant. The corporate tax in \nmost countries decreased notably from 2005 to 2016 (Supplementary \nInformation 2), indicative of intense tax competition. We found that the \nglobal corporate tax competition from 2005 to 2016 increased global \nemissions by 128.7\u2009Mt, which is equivalent to 0.4% of global CO2 emis-\nsions from fuel combustion in 2016 (32,141\u2009Mt) based on International \nEnergy Agency (IEA) data. Emissions in developing countries increased \nby 118.5\u2009Mt, but only by 10.2\u2009Mt in developed countries. Global corpo-\nrate tax competition shifts more emissions to developing economies. \nCanada\nChina\nGermany\nIndia\nJapan\nUSA\nRest of world\n10 Mt\n30 Mt\n50 Mt\n\u20133\n\u20132\n\u20131\n0\n1\n2\n3\n\u201312\n\u20138\n\u20134\n0\n4\nChange in CO2 emissions (%)\nChange in corporate tax rates (%)\nr = \u20130.82\nP < 0.01\nFig. 1 | The emission impact of corporate tax competition over the period \n1995\u20132016 at the regional level. In this paper, the changes in tax rates are \nmeasured in relative terms, specifically by computing the ratio of the new tax \nrates incremented by one, divided by the old tax rates incremented by one. The \ny axis measures the changes in CO2 emission (in percentages) caused by global \ncorporate tax competition. The size of each circle denotes the volume of the \nabsolute change in regional CO2 emissions. We employ the Pearson correlation \ncoefficient (r) to assess the degree of correlation, and the P value is two-sided.\n\nNature Climate Change | Volume 14 | April 2024 | 353\u2013356\n355\nArticle\nhttps://doi.org/10.1038/s41558-024-01952-0\nEconomies that have implemented tax cuts, such as Canada and \nGermany, tend to have a positive structural effect (Fig. 2). Intuitively, \neconomies with larger tax cuts attracted more MP production and \nenhanced exports, thus raising territorial emissions. Supplementary \nFigs. 1 and 2 (in Supplementary Information 3) present the negative \ncorrelations between the changes in MP production and changes in \nexports against corporate tax cuts at the country level. Supplementary \nInformation 4 further presents the changes in trade flows and MP flows \nat the bilateral level caused by the global corporate tax competition. \nEconomies that had implemented sharp tax cuts experienced a notable \nincrease in exports because the tax cuts increased their business profits \nand stimulated more trade flows from this economy. Meanwhile, a \ncountry\u2019s corporate tax decrease would confer a comparative advan-\ntage and make it more appealing to foreign businesses. For instance, \nthe MP flows from the United States to China increased by 10.0%, while \nMP flows to the United States decreased sharply. The above shows that \nboth the MP and trade elements are essential for accurately quantify-\ning the emission effects of corporate tax changes (Supplementary \nInformation 5).\nScenario analysis of 15% global minimum \ncorporate tax\nIn 2021, G20 countries proposed a global minimum corporate tax of at \nleast 15% for MNEs. We quantify the potential welfare and environmental \neffects of this policy by forcing a 15% corporate tax rate on economies \nwhose original corporate tax is less than 15% based on the benchmark \nequilibrium in 2016. The simulation results show that global carbon \nemissions would exhibit a reduction of 45.0\u2009Mt (0.14% in relation to the \ntotal emissions observed in 2016) in this counterfactual. This implies \nthat the global minimum corporate tax would contribute to climate \nchange mitigation. However, the increase in corporate tax rate would \nintroduce additional distortion to global production, and lead to a \nreduction in regional real wage (Supplementary Information 6) and \nglobal welfare by 0.002%. The impact of both the emission and wel-\nfare effects resulting from the minimum corporate tax rate policy is \nnotably more modest compared with that of global tax competition. \nThis discrepancy primarily arises from the relatively minor alterations \nin tax rates in the minimum tax policy.\nTable 1 summarizes the change in regional emissions, as well as the \nscale effect and structural effect of this exercise for the major econo-\nmies. Corporate taxes in economies such as China and Russia have \nincreased since their original effective corporate tax rates are lower \nthan 15%. Therefore, the global minimum tax would decrease business \nprofits in these economies, shrink their production and decrease their \nemissions. CO2 emissions in China and the EU decreased by 0.30% \nand 0.28%, respectively (Table 1). In contrast, the economies whose \noriginal effective corporate tax rates are equal to or exceed 15% could \npotentially reap competition advantages from the minimum tax policy. \nThis could manifest through an enhanced capacity to attract more \nproduction and annex a larger portion of their competitor\u2019s market \nshare, which consequently resulted in more emissions. For instance, the \nCO2 emissions in the United States and Brazil would increase by 0.06% \nand 0.05%, respectively. We not only evaluate the change in regional \nemissions but also evaluate the change in bilateral carbon transfer \n(Supplementary information 7). The results show that the minimum \ntax policy could alleviate the pollution-haven effect of international \nproduction fragmentation.\nDifferent from the effect of global tax competition, the structural \neffect plays a larger role than the scale effect for ~60% of the econo-\nmies in this exercise. This indicates the heightened influence of the \nreconfiguration of the global production structure on emissions. The \ndifference primarily stems from the more pronounced dispersion of \ntax rate changes across countries in this exercise than that in the exer-\ncise on global tax competition. The coefficient of variation for tax rate \nchanges for the two exercises are 1.24 and 0.82, respectively. A more \npronounced disparity in tax rate changes among countries would lead \nto larger shifts in competitive advantage and consequently result in \na relatively greater transformation of production structures. At the \neconomic level, we observed a negative correlation both between the \nstructural effect and changes in corporate tax rates, as well as between \nthe scale effect and changes in corporate tax rates (Supplementary \ninformation 8). An increase in a country\u2019s corporate tax rate would \ntrigger relocation of production activities from the country, thereby \nleading to a decrease in its output, emissions and income levels. This \ndecline in income would in turn result in reduced demand for products, \nconsequently amplifying the reduction in both output and emissions.\nDiscussion\nWe found that global corporate tax competition increased global car-\nbon emissions and the introduction of a global minimum 15% tax rate \ncould mitigate the negative impact of such competition on climate \nchange mitigation. However, the impact is somewhat limited. The 15% \nminimum corporate tax rate, while a positive step, remains relatively \nlow and should be seen as an initial effort to address the issue of global \ncorporate tax competition. Furthermore, the minimum corporate \ntax primarily addresses regional disparities in global corporate tax \ncompetition and does not account for sectoral variations in carbon \nintensity. On the basis of our findings, we suggest that corporate tax \nadjustment should be coordinated with the implementation of cli-\nmate policies. For example, a higher minimum corporate tax could be \nimposed on fossil-fuel companies, thereby contributing to the phasing \nout of fossil-fuel subsidies. Sectors with higher emission intensity have \nexpanded their scales more due to global tax competition (Supple-\nmentary Information 9). Meanwhile, the increased revenues resulting \nfrom these minimum tax reforms should be directed towards initia-\ntives aligned with low greenhouse gas emissions and climate-resilient \ndevelopment, such as financing renewable energy development to \naccelerate the transition to cleaner energy sources.\nThe present study mainly focuses on the impact of tax change \non regional and global carbon emissions through reshaping global \nproduction and investment patterns. However, it is worth noting \nthat these policies may impact regional emissions through other \nchannels. For instance, the global corporate tax rate would reduce \ngovernment revenue23 and further decrease governments\u2019 financial \ninvestment in low-carbon technology, which may negatively impact \nclimate change mitigation, especially in developing economies. Mean-\nwhile, cross-country trade and investment flows may transfer clean \nproduction technology across countries24,25. Therefore, when tax \nchanges impact trade and investment flow, production technology \nUK\nCanada\nChina\nGermany\nJapan\nIndia\nRest of world\nUSA\n10Mt\n30Mt\n50Mt\n\u20131.0\n\u20130.5\n0\n0.5\n1.0\n1.5\n\u20131.0\n0\n1.0\n2.0\nStructure efect (%)\nScale efect\n(%)\nFig. 2 | The change in regional emissions through the scale and structure \neffects.  The size of each circle denotes the volume of the absolute change in \nregional CO2 emissions.\n\nNature Climate Change | Volume 14 | April 2024 | 353\u2013356\n356\nArticle\nhttps://doi.org/10.1038/s41558-024-01952-0\nin each country may also change and thus, impact global and regional \nemissions. Future studies could consider these channels. In addi-\ntion, there are also different carbon accounting principles, such as \nproduction-based and consumption-based accounting principles26. \nThis study mainly focuses on production-based emission changes. \nSince a change in global tax rates could seriously impact cross-country \ntrade and investment patterns, consumption-based emissions may  \ndiffer from production-based emissions. Future studies should  \nexplore this.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-024-01952-0.\nReferences\n1.\t\nMadi\u00e8s, T., Tarola, O. & Taugourdeau, E. Tax haven, pollution \nhaven or both? Int. Tax Public Finance 29, 1527\u20131560 (2022).\n2.\t\nIovino, L., Martin, T. & Sauvagnat, J. Corporate taxation and \ncarbon emissions. SSRN https://doi.org/10.2139/ssrn.3880057 \n(2021).\n3.\t\nAbe, K. & Zhao, L. Endogenous international joint ventures and \nthe environment. J. Int. Econ. 67, 221\u2013240 (2005).\n4.\t\nCopeland, B. R. & Taylor, S. M. Trade and the Environment. \n Theory and Evidence (Princeton Univ. Press, 2003).\n5.\t\nCai, X., Lu, Y., Wu, M. & Yu, L. Does environmental regulation \ndrive away inbound foreign direct investment? Evidence from a \nquasi-natural experiment in China. J. Dev. Econ 123, 73\u201385 (2016).\n6.\t\nHanna, R. US environmental regulation and FDI: evidence from a \npanel of US-based multinational firms. Am. Econ. J. Appl. Econ. 2, \n158\u2013189 (2010).\n7.\t\nDean, J. M., Lovely, M. E. & Wang, H. Are foreign investors \nattracted to weak environmental regulations? Evaluating the \nevidence from China. J. Dev. Econ. 90, 1\u201313 (2009).\n8.\t\nChung, S. Environmental regulation and foreign direct \ninvestment: evidence from South Korea. J. Dev. Econ. 108, \n222\u2013236 (2014).\n9.\t\nBruca, A., Javorcik, B. & Love, I. Good for the environment, good \nfor business: foreign acquisitions and energy intensity. J. Int. Econ. \n21, 103247 (2019).\n10.\t Eskeland, G. S. & Harrison, A. E. Moving to greener pastures? \nMultinationals and the pollution haven hypothesis. J. Dev. Econ. \n70, 1\u201323 (2003).\n11.\t\nRamondo, N. & Rodr\u00edguez-Clare, A. Trade, multinational \nproduction, and the gains from openness. J. Polit. Econ. 121, \n273\u2013322 (2013).\n12.\t Arkolakis, C., Ramondo, N., Rodriguez-Clare, A. & Yeaple, S. \nInnovation and production in the global economy. Am. Econ. Rev. \n108, 2128\u20132173 (2018).\n13.\t Alviarez, V. Multinational production and comparative advantage. \nJ. Int. Econ. 119, 1\u201354 (2019).\n14.\t Wang, Z. Multinational production and corporate taxes: a \nquantitative assessment. J. Int. Econ. 126, 103353 (2020).\n15.\t Tintelnot, F. Global production with export platforms. Q. J. Econ. \n132, 157\u2013209 (2017).\n16.\t Elliott, J. et al. Trade and carbon taxes. Am. Econ. Rev. 100, \n465\u2013469 (2010).\n17.\t Copeland, B. R., Shapiro, J. S. & Taylor, M. S. in Handbook of \nInternational Economics Vol. 5 (eds Gopinath, G. et al.) 61\u2013146 \n(Elsevier, 2022).\n18.\t Larch, M. & Wanner, J. Carbon tariffs: an analysis of the trade, \nwelfare, and emission effects. J. Int. Econ. 109, 195\u2013213 (2017).\n19.\t Shapiro, J. S. & Walker, R. Why is pollution from U.S. \nmanufacturing declining? The roles of environmental regulation, \nproductivity, and trade. Am. Econ. Rev. 108, 3814\u20133854 (2018).\n20.\t Farrokhi, F. & Lashkaripour, A. Can Trade Policy Mitigate Climate \nChange? STEG Working Paper (STEG, 2022).\n21.\t Duan, Y., Ji, T., Lu, Y. & Wang, S. Environmental regulations and \ninternational trade: a quantitative economic analysis of world \npollution emissions. J. Public. Econ. 203, 104521 (2022).\n22.\t Shapiro, J. S. Trade costs, CO2, and the environment.  \nAm. Econ. J. Econ. Policy 8, 220\u2013254 (2016).\n23.\t Bierbrauer, F., Brett, C. & Weymark, J. A. Strategic nonlinear \nincome tax competition with perfect labor mobility.  \nGames Econ. Behav. 82, 292\u2013311 (2013).\n24.\t Liu, Z. Foreign direct investment and technology spillovers: \ntheory and evidence. J. Dev. Econ. 85, 176\u2013193 (2008).\n25.\t Demena, B. A. & van Bergeijk, P. A. A meta-analysis of FDI and \nproductivity spillovers in developing countries. J. Econ. Surv. 31, \n546\u2013571 (2017).\n26.\t Peters, G. P. & Hertwich, E. G. CO2 embodied in international trade \nwith implications for global climate policy. Environ. Sci. Technol. \n42, 1401\u20131407 (2008).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner)  \nholds exclusive rights to this article under a publishing agreement \nwith the author(s) or other rightsholder(s); author self-archiving  \nof the accepted manuscript version of this article is solely  \ngoverned by the terms of such publishing agreement and  \napplicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2024\nTable 1 | Changes in regional emissions under the 15% global \nminimum corporate tax\nChange in CO2 (%)\nStructure effect (%)\nScale effect (%)\nAustralia\n0.05\n0.14\n\u22120.09\nBrazil\n0.05\n0.10\n\u22120.04\nCanada\n\u22121.63\n\u22121.01\n\u22120.61\nChina\n\u22120.30\n\u22120.10\n\u22120.19\nEU\n\u22120.28\n\u22120.07\n\u22120.21\nIndia\n0.00\n0.04\n\u22120.04\nIndonesia\n\u22120.04\n0.01\n\u22120.06\nJapan\n\u22120.13\n0.03\n\u22120.16\nKorea\n\u22120.07\n0.03\n\u22120.11\nMexico\n0.03\n0.04\n\u22120.01\nNorway\n0.09\n0.16\n\u22120.07\nRussia\n\u22120.36\n\u22120.10\n\u22120.27\nSwitzerland\n\u22120.91\n\u22120.40\n\u22120.51\nTurkey\n\u22120.02\n0.03\n\u22120.06\nUnited States\n0.06\n0.12\n\u22120.06\nThe EU countries are presented as a union in Table 1.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-01952-0\nMethods\nGeneral equilibrium model incorporating trade and MP\nWe build a multicountry multi-industry general equilibrium model \nby incorporating MNEs, international trade flows and corporate tax \nto quantify the emission effect of corporate tax changes27. Specifi-\ncally, we extended the quantitative trade and MP model developed \nby refs. 11,12, by incorporating corporate tax and linking production \nto environmental emissions. In the model, a firm from country i can \nserve country n by building an affiliate (MP) in country l and shipping \ngoods to country n (Trade). This entails MP costs associated with the \npair {i, l} and trade costs associated with the pair {l, n}. Given these \ncosts between countries, a firm in any country chooses how much of \nits products will serve market n, and via which production location \nit will serve market n to maximize profit. Then, a corporate tax cut in \ncountry B would increase the profits of the production located in B and \nattract more MP activities and stimulate more trade flows from B. The \nexpansion of production in B would directly increase the CO2 emissions \nthrough a scale effect. It would also increase labor demand and lift the \nwage level in B, as labor is a necessary input in production. However, the \nwage change in B would further reshape MP flows to B and trade flows \nfrom B by influencing marginal production costs. Meanwhile, tax cuts \nin B would impact output and emissions in B and other countries. For \nexample, the tax cut in B will decrease the income of its competitors \nby taking up their market share. In other words, corporate tax cuts in \none country would impact both MP and trade flows across countries \nand reshape global production and emission networks.\nConsider a world economy comprising i\u2009=\u20091, \u2026, N countries, j\u2009=\u20091, \n\u2026, J sectors and one factor of production, labor; sector j in country i has \na fixed number M j\ni  of firms, each of which produce one product variety \nindexed by \u03c9 j \u2208[0, 1] using labor under monopolistic competition. We \nassumed that each product has no input\u2013output linkage with others.\nHouseholds. In each country, there are a measure of Ln representative \nhouseholds at a wage rate wn that maximize utility by consuming final \ngoods from different sectors Q j\nn with welfare function of\nUn =\nj\n\u220f\nj=1\nQ j\nn\n\u03b1 j\nn\n1\n[1 + (\n\u2211j\u2211i\u2211le j\nil\n\u03bcn\n)\n2\n]\n(1)\nwhere \u2211j\u03b1 j\nn = 1. The first bracketed term represents the utility from \nconsuming goods and the second bracketed term in the equation \nrepresents the disutility from carbon emissions. e j\nil is the CO2 emissions \nemitted by firms of the sector j from country i located in country l. the \nparameter \u03bcn dictates the social cost of CO2 emissions. Supplementary \nInformation 10 provides a more detailed discussion on how we incor-\nporated environmental issues and sector heterogeneity into the model.\nWith Cobb\u2013Douglas preferences, the consumption price index in \ncountry n is given by Pn =\nj\n\u220f\nj=1\n(\nP j\nn\n\u03b1 j\nn )\n\u03b1 j\nn, where P j\nn is the consumption price \nof sector j goods in country n. The production technology of the final \ngoods Q j\nn is given by Q j\nn = [\u222b\u03c9 jq j\nn (\u03c9 j)\n\u03c3\u22121\n\u03c3 d\u03c9 j]\n\u03c3\n\u03c3\u22121\n, where q j\nn (\u03c9 j) is country \nn\u2019s quantity consumption of goods \u03c9 j from the lowest-cost supplier.\nProduction and emissions. Since labor is the only input in produc-\ntion, the labor used in producing goods in country n and sector j is a \nfunction of output:\ny j\nil(\u03c9 j) = \u0303z j\nil (\u03c9 j) l\nj\nl (\u03c9 j) ,\n(2)\nwhere l j\nl (\u03c9 j) is labor input and \u0303z j\nil (\u03c9 j) is production efficiency. This \nsuggests that the marginal cost of each good in country l is wl, the wage \nlevel in country n. The production generates carbon emissions e j\nil (\u03c9 j) \nas a by-product. A firm allocates a portion \u03b6 j\nl  of the output y j\nil (\u03c9 j) to \nemission abatement activities to reduce its tax payment. On the basis \nof Copeland and Taylor (2001)31, we assumed the pollution abatement \ntechnology to be e j\nil (\u03c9 j) = (1 \u2212\u03b6 j\nl )\n1\n\u03b2 j\nl y j\nil (\u03c9 j). The net production after \nabatement investment is, q j\nil (\u03c9 j) = z j\nil (\u03c9 j) [l j\nl (\u03c9 j)]\n(1\u2212\u03b2 j\nl )\ne j\nil (\u03c9 j)\n\u03b2 j\nl , where \nz j\nil (\u03c9 j) = [ \u0303z j\nil (\u03c9 j)]\n1\u2212\u03b2 j\nl  and z j\nil (\u03c9 j) are drawn independently across firms \nfrom a Fr\u00e9chet distribution with parameters \u03b8 j: Pr (z j\nil < x) \n= exp (\u2212T j\nilx\u2212\u03b8 j). \u03b8 j governs the dispersion of productivity for the affili-\nates located in different product locations. The firm\u2019s problem implies \nthat the cost of unit input bundle is c j\nil = \u03b2 j\nl\n\u2212\u03b2\nj\nl (1 \u2212\u03b2 j\nl )\n\u2212(1\u2212\u03b2 j\nl )\n(\u03c1l)\u03b2 j\nl (wl)(1\u2212\u03b2 j\nl ) \nand \u03b2 j\nl =\n\u03c1le j\nil\nc j\nilq j\nil\n, where wl is the wage and \u03c1l is the emission tax in country l. \nFollowing ref. 12, the average productivity is a combination of produc-\ntivities in parent country and in host country, that is, we have T j\nil = T j\ni T j\nl .\nInternational trade costs, prices and profits. Following refs. 12,14, \neach firm originating in country i (parent country) chooses the produc-\ntion location l (host country) to serve the country n market (consump-\ntion country). Each firm from country i that produce in l aces an \nidiosyncratic productivity shock z j\nil (\u03c9 j) and incurs an iceberg multi-\nnational production (MP) cost r j\nil\u2009\u2265\u20091, with r j\nii = 1. Firms that export from \nl to n incurs an iceberg trade cost \u03c4 j\nln\u2009\u2265\u20091, with \u03c4 j\nnn = 1. The unit cost for \nthis firm is:\nc j\niln (\u03c9 j) =\nc j\nilr j\nil\u03c4 j\nln\nz j\nil (\u03c9 j)\n,\n(3)\nThe standard results for monopolistic competition imply that the \npre-tax profits of firm \u03c9j  from country i serving market n from its plant \nin country l can be expressed as:\n\u03c0 j\niln (\u03c9 j) = 1\n\u03c3 [ \u0303\u03c3c j\niln (\u03c9 j)]\n1\u2212\u03c3\nX j\nnP j\nn\n\u03c3\u22121\nwith \u0303\u03c3 =\n\u03c3\n\u03c3 \u22121 .\n(4)\nHere, X j\nn = P j\nnQ j\nn is the total expenditure on sector j goods. Con-\nsider a firm in sector j with its headquarters located in country i. The \nprofits made by its affiliate in country l are subject to the local cor-\nporate income tax tl before it can be paid out as dividend. The firm \nwill choose its production location to serve market n by maximizing \nits profits, or to put it differently, by minimizing its tax-adjusted  \nunit cost:\nl j\nin(\u03c9 j) = arg max\nk=1\u2026..N{(1 \u2212tk)\u03c0 j\nikn(\u03c9 j)} = arg min\nk=1\u2026..N {(1 \u2212tk)\n1\n1\u2212\u03c3\n\u03be j\nikn\nz j\nik(\u03c9 j)\n} ,\n(5)\nwhere \u03be j\nikn = c j\nikr j\nik\u03c4 j\nkn. Equation (5) indicates that a firm decides its pro-\nduction sites on the basis of the effective cost shifter of taxation \n((1 \u2212tl)\n1\n1\u2212\u03c3), together with production-site-specific productivities \n(z j\nik (\u03c9 j)), wages in the host countries (wl), and trade friction (\u03c4 j\nln) and \nMP friction (r j\nil).\nExpenditure shares. We solved the firm\u2019s choices of production sites \nand destination markets, and then aggregated these decisions to coun-\ntry level. According to the property of the Fr\u00e9chet distribution, the \nprobability that each firm in sector j from i serving market n will choose \nlocation l for production is:\n\u03c6 j\niln =\nX j\niln\nX j\nin\n=\nT j\nil[(1 \u2212tl)\n1\n1\u2212\u03c3 \u03be j\niln]\n\u2212\u03b8 j\n\u2211k T j\nik [(1 \u2212tk)\n1\n1\u2212\u03c3 \u03be j\nikn]\n\u2212\u03b8 j .\n(6)\nThe price of goods of firms originating from i arriving at country \nn is:\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-01952-0\nP j\nin =\n\u0303\u03c3\u03b4 j[\u2211\nl\nT j\nil(\u03be j\niln)\n\u2212\u03b8 j\n]\n\u22121\n\u03b8 j\n,\n(7)\nwhere \u03b4 j = \u0393 (1 +\n1\u2212\u03c3\n\u03b8 j )\n1\n1\u2212\u03c3  is a constant. According to the constant  \nelasticity of substitution (CES) production technology in each country, \nthe share of total expenditure in country n devoted to goods produced \nby firms from i is:\n\u03bbjE\nin =\nX j\nin\nX j\nn\n=\nM j\ni (\u03d5 j\nin)\n1\u2212\u03c3\n\u2211i\u2032Mj\ni\u2032(\u03d5 j\ni\u2032n)\n1\u2212\u03c3 ,\n(8)\nwhere \u03d5 j\nin = [\u2211lT j\nil (\u03be j\niln)\n\u2212\u03b8 j\n]\n\u22121\n\u03b8 j\n represents the average unit cost of goods \nfrom parent country i to market n. We can also obtain the expression \nfor aggregate price in country n, P j\nn = (\u2211i M j\ni P j\nin\n1\u2212\u03c3\n)\n1\n1\u2212\u03c3.\nDenoting X j\niln as the expenditure of sector j country n on goods \nproduced in country l by firms from country i, and the corresponding \ntrade share is given by:\n\u03bb j\niln =\nX j\niln\nX j\nn\n= \u03c6 j\niln\u03bbjE\nin =\nM j\ni (\u03d5 j\nin)\n1\u2212\u03c3\n\u2211i\u2032 Mj\ni\u2032(\u03d5 j\ni\u2032n)\n1\u2212\u03c3\nT j\nil[(1 \u2212tl)\n1\n1\u2212\u03c3 \u03be j\niln]\n\u2212\u03b8 j\n\u2211k T j\nik [(1 \u2212tk)\n1\n1\u2212\u03c3 \u03be j\nikn]\n\u2212\u03b8 j .\n(9)\nThe share of goods produced in country l originating from country \ni, that is, the MP share, is given by:\n\u03bbjM\nil =\n\u2211n X j\niln\nY j\nl\n=\n\u2211n \u03bbjE\nin\u03c6 j\nilnX j\nn\nY j\nl\n,\n(10)\nwhere Y j\nl = \u2211i \u2211n Xj\niln  is the total output in country l. The share of \nexpenditure of country n on goods produced from country l, that is, \nthe trade share, is given by:\n\u03bbjT\nln = \u2211\ni\n\u03bb j\niln.\n(11)\nFor country k, total expenditure on goods j is the sum of expendi-\nture by households (\u03b1 j\nk Ik). The total income (Ik) includes the wage \nrevenue (wkLk), the trade deficit (Dk), the corporate tax revenue (\u2227k), \nthe emission tax revenue (\u03b3k) and the profits from MP firms (\u03c0k). The \nmarket clearance condition implies that:\nX j\nk = \u03b1 j\nkIk,\n(12)\nwhere Ik = wkLk + Dk +\u2227k + \u03b3k + \u03c0k  and \u2227k = \u2211j\u2227j\nk , \u03c0k = \u2211j\u03c0 j\nk , \u03b3k = \u2211j\u03b3 j\nk . \nFrom the production technology, we have \u2227j\nk =\ntk\n\u03c3 \u2211i\u2211nX j\nikn =\ntk\n\u03c3 Y j\nk , \n\u03c0 j\nk = \u2211l \u2211n (1 \u2212tl)\n1\n\u03c3 X j\nkln and \u03b3 j\nk = \u03c1ke j\nik =\n\u03c3\u22121\n\u03c3 \u03b2 j\nk Y j\nk. The labor market clear-\nance condition implies that:\nwkLk = \u03c3 \u22121\n\u03c3\n\u2211j [(1 \u2212\u03b2 j\nk ) Y j\nk ] .\n(13)\nThe total emission in country l is given by:\ne j\nl = \u2211i\u2211je j\nil = \u2211j\n\u03c3 \u22121\n\u03c3\u03c1l\n\u03b2 j\nl Y j\nl .\n(14)\nDefinition\nGiven local fundamentals {Li, T j\nil, M j\ni }, spatial frictions {\u03c4 j\nln, \u03b3 j\nil} and param-\neters {\u03c3, \u03b8 j, \u03c1l,\u03b1 j\nl , \u03b2 j\nl }, a competitive equilibrium under corporate tax \nstructure {tl} and emission tax { \u03c1l} is a wage vector {wl} that satisfy \nequilibrium condition (12) and (13), where the related variables are \ndefined by equations (3) and (6)\u2013(11) for all countries. Note that once \nwe solved wl for each country, other endogenous variables such as X j\ni , \n\u03bb j\niln are immediately obtained by using equations (6)\u2013(11) and (14).\nFollowing ref. 14, we used Exact-hat algebra to solve the equilib-\nrium to reduce the burden of parameter calculation. For each variable \nx in the original equilibrium, we denoted x\u2032 as its counterfactual value \nand \u0302x =\nx\u2032\nx  as the relative changes. Then the equilibrium conditions in \nrelative changes satisfy the following equations:\nThe change in trade share \u2236\u0302\u03bb j\niln =\n( \u0302\u03d5 j\nin)\n(1\u2212\u03c3)\n\u2211i\u2032 \u03bbjE\ni\u2032n( \u0302\u03d5 j\ni\u2032n)\n(1\u2212\u03c3)\n\u0302\u03c6 j\niln\n(15)\nThe product market clearance condition \u2236X j\nk\n\u2032\n= \u03b1 j\nk I j\nk\n\u2032\n(16)\nThe labor market clearance condition \u2236\u02c6wl =\n\u2211j \u2211i \u2211n (1 \u2212\u03b2 j\nl )Xj\u2032\niln\n\u2211j \u2211i \u2211n (1 \u2212\u03b2 j\nl )X j\niln\n(17)\nwhere\n\u0302\u03be j\niln = (\u02c6\u03c1l)\n\u03b2 j\nl (\u02c6wl)\n1\u2212\u03b2 j\nl , (\u02c6\n1 \u2212tl) =\n1\u2212tl\n\u2032\n1\u2212tl , \u0302\u03b6 j\niln = (\u02c6\n1 \u2212tl)\n1\n1\u2212\u03c3 \u0302\u03be j\niln,\n\u02c6\u03c6 j\niln =\n( \u0302\u03b6 j\niln)\n\u2212\u03b8 j\n\u2211l\u2032 [( \u0302\u03b6 j\nil\u2032n)\n\u2212\u03b8 j\n\u03c6 j\nil\u2032n]\n.\nand \u0302\u03d5 j\nin = [\u2211l \u0302\u03be j\u2212\u03b8 j\niln\nT j\nil(\u03be j\niln)\n\u2212\u03b8 j\n\u2211l\u2032 T j\nil(\u03be j\niln)\n\u2212\u03b8 j ]\n\u22121\n\u03b8 j\n.\nNote that {Li, T j\nil, \u03c4 j\nln, \u03b3 j\nil} are absent from these hat change expres-\nsions. This simplifies the estimation of the model. Given a vector of \ncorporate tax shock, we can solve the relative change of each variable \nin the new equilibrium compared to the benchmark equilibrium using \nthe above equations, and then calculate the relative changes in welfare \nand carbon emissions.\nAs previously mentioned, corporate tax competition would impact \nMP and trade activities across countries, reshaping global produc-\ntion networks and emission networks. To provide a more intuitive \nunderstanding of how the corporate tax rate changes impact global \nemissions, we decomposed the emission effect of the corporate tax \nchanges into scale effect and structural effect:\nd ln el =\n1\nel \u2211j [\u03c5 j\nl \u2211i\u2211n (X j\nilnd ln X j\nn)]\n+\n1\nel \u2211j [\u03c5 j\nl \u2211i\u2211n (X j\nilnd ln\nX j\niln\nX j\nn )] ,\n(18)\nwhere \u03c5 j\nl = e j\nl /Y j\nl  is the emission intensity, that is, emission directly emit-\nted per output of sector j good in country l. The first term in equation \n(18) reflects the scale effect, representing the emission changes attribut-\nable to shifts in countries\u2019 consumption patterns. The second term in \nequation (18) is the structural effect, which reflects the emission changes \nresulting from shifts in an economy\u2019s production share within each of \nthe other economies\u2019 total expenditures. Consequently, it underscores \nthe role of changes in MP and trade patterns across economies.\nData and calibration\nTo solve the equilibrium, we calibrated the model to the year 2016, \nwhich is the most recent year for the ICIO\u2013AMNE (Inter-Country Input\u2013\nOutput, Activity of Multinational Enterprises) tables (https://www.\noecd.org/sti/ind/analytical-amne-database.htm). We imputed bilateral \ntrade and MP shares and consumption preference \u03b1 j\nl  from the ICIO\u2013\nAMNE 2016 data constructed by the OECD28, and set \u03b8\u2009=\u20094.5 and \u03c3\u2009=\u20094 \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-024-01952-0\nfollowing ref. 12. The effective corporate tax rates for the different \neconomies were obtained from the World Bank (http://www.doingbusi-\nness.org/). The changes in the corporate tax rates of the different \nregions were obtained from KPMG (https://home.kpmg/xx/en/home.\nhtml). The country sectoral-level emission data were from ref. 29. The \ncost- minimizing problem faced by firms implies that \u03b2 j\nl =\n\u03c1le j\nil\nc j\nilq j\nil\n= \u03c1l\u03c5 j\nl , \nand therefore \u03b2 j\nl  indicates the share of carbon revenue in output at the \ncountry and sector level. The national carbon revenue in GDP is there-\nfore \n\u03c1l\u2211jej\nl\nGDPl . To estimate \u03b2 j\nl , we first derived the share of the energy and \ncarbon tax revenue in the GDP at the national level, utilizing the data \nfrom the OECD database. These data, combined with national emis-\nsions, yield the emission tax rate \u03c1l at the national level. Subsequently, \nmultiplying this tax rate by the country-sector-specific emission inten-\nsity generates an estimate of \u03b2j\nl (Supplementary Information 11).\nWe calibrated the value of the social cost of CO2 emissions param-\neter \u03bcn following the method in ref. 21. We chose the values \u03bcn, so that \na 1-ton increase in CO2 emissions decreases global GDP by US$29, the \nmost recent estimate of the marginal social cost of CO2 emissions by \nan interagency panel of the US government30. We used the wage level \nof the first country as the numeraire. Note that the choice of numeraire \nwould not impact the changes in the ratio of two nominal variables \n(Supplementary Information 12).\nGiven these, we relied on MATLAB to solve the equilibrium. For a \ngiven change in corporate tax vector \u0302t ={ \u0302t1, \u2026 , \u0302tN}, we first guessed a \nvector of changes in wages \u0302w ={\n\u0302w1, \u2026 ,\n\u0302wN} (we have \u0302w1\u2009=\u20091 since the wage \nlevel of the first country was taken as the numeraire), and then calcu-\nlated the changes in trade shares and expenditures by using equations \n(15) and (16). Finally, substituting the corresponding variables into \nequation (17), we verified whether the labor market clearance condition \nheld. If not, we adjusted our guess of \u0302w until the equilibrium condition \n(13) was finally obtained. A robustness check of the results is provided \nin Supplementary Information 12.\nReferences\n27.\t Tian, K., Zhang, Y., Li, Y., Ming, X. & Wang, S. Regional trade \nagreement burdens global carbon emissions mitigation. Nat. \nCommun. https://doi.org/10.1038/s41467-022-28004-5 (2022).\n28.\t Cadestin, C. et al. Multinational Enterprises and Global Value \nChains: The OECD Analytical AMNE Database (OECD Publishing, \n2018).\n29.\t Zhang, Z., Guan, D., Wang, R., Meng, J. & Du, H. Embodied  \ncarbon emissions in the supply chains of multinational \nenterprises. Nat. Clim. Change https://doi.org/10.1038/s41558-020-\n0895-9 (2020).\n30.\t Interagency Working Group on Social Cost of Greenhouse Gases. \nTechnical Support Document: Technical Update of the Social Cost \nof Carbon for Regulatory Impact Analysis Under Executive Order \n12866 (United States Government, 2016).\n31.\t Copeland, B. R. & Taylor, M. S. International Trade and the \nEnvironment: A Framework for Analysis. Working Paper Series \n8540 (National Bureau of Economic Research, 2001).\nAcknowledgements\nWe gratefully acknowledge the financial support from the National \nNatural Science Foundation of China (No. 72394404 to Z.Z.), the \nNational Social Science Foundation of China (No. 22CJY019 to Y.D.), \nthe National Natural Science Foundation of China (No. 71988101 to \nW.S., No. 72261147471 to Y.D., Nos. 71834004 and 71974141 to Z.Z.), \nthe major program of the National Social Science Fund of China \n(No. 22&ZD086 to Y.D., No. 19ZDA062 to Y.C., No. 20&ZD079 to Y.L.). \nSpecial thanks to Longfei Cai and Yuan Gao from Central University  \nof Finance and Economics in China for the excellent research  \nassistant work.\nAuthor contributions\nY.D. and Z.Z. designed the research. Y.D. built the general equilibrium \nmodel and wrote the methods. Z.Z. created figures and drafted \nthe initial manuscript. Y. Li. collected the raw data and revised the \nmanuscript. S.W., C.Y. and Y. Lu. commented on the results and \ndiscussion. All authors contributed to writing the manuscript and \ndiscussed the results at all stages.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains \nsupplementary material available at  \nhttps://doi.org/10.1038/s41558-024-01952-0.\nCorrespondence and requests for materials should be addressed to \nZengkai Zhang.\nPeer review information Nature Climate Change thanks Luis Lopez and \nthe other, anonymous, reviewer(s) for their contribution to the peer \nreview of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\n\n    Scientific Research Findings:", "answer": "We find that the global corporate tax competition from 2005 to 2016 resulted in an increase in global emissions by reshaping the geographical distribution of production through an impact on trade and investment. This leads to a shift of more emissions towards developing economies, with a notable rise of 118.5 MtCO2emissions in developing countries compared with a modest rise of 10.2 MtCO2in developed countries. To address the global corporate tax competition, more than 130 countries and jurisdictions approved a global minimum tax rate of 15% in October 2021. We find that the global minimum corporate tax would contribute to climate change mitigation, albeit to a modest extent. Our study may underestimate the environmental effects of global corporate tax competition as we do not consider its negative impact on low-carbon technology.", "id": 44}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 14 | January 2024 | 48\u201354\n48\nnature climate change\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\nInequality repercussions of financing \nnegative emissions\nPietro Andreoni\u2009\n\u200a\u20091,2\u2009\n, Johannes Emmerling\u2009\n\u200a\u20091 & Massimo Tavoni1,2\nNegative emissions technologies are attracting the interest of investors \nin the race to make them effective and profitable. When deployed at scale, \nthey will be financed through public funds, reducing the fiscal space for \na socially inclusive climate transition. Moreover, if the private sector \nowns negative emissions technologies, potentially large profits would \ndisproportionally benefit investors and equity holders. Here we quantify \nthe inequality repercussions of direct air capture of CO2 in a 1.5\u2009\u00b0C scenario, \nusing a regional integrated assessment model that features within-country \nincome heterogeneity. We find that, under a single carbon market, financing \nnegative emissions technologies could double the increase in income \ninequality of climate policy. The effects are highest around the time of net \nzero and in scenarios with carbon budget overshoot. We identify the drivers \nof the inequality increase and discuss policy provisions to mitigate the \nequity concerns of CO2 removal strategies.\nIn 2022 alone, foundations linked to private donors such as Google \nand Facebook pledged almost US$1 billion to advance the research and \ndevelopment of negative emissions technologies (NETs)1. In a world \nof insufficient climate mitigation efforts, these technologies can help \nthe decline of CO2 emissions, offset hard-to-abate emissions and aid \nrecovery from temperature overshoot2.\nAlthough they are criticized for being currently immature or \nspeculative3,4, as well as for the possibility of crowding out emission \nreductions5\u20137, there is consensus that NETs will be needed at scale to \nachieve the Paris Agreement targets. Therefore, research and devel-\nopment on these technologies is necessary to reduce the uncertainty \nabout their cost, potential and scalability, and to build an effective and \nefficient portfolio of options to reach net-zero or net-negative emis-\nsions. However, leaving this effort exclusively to the private sector \ncould pose additional challenges.\nOnce they mature, the cost-minimizing solution to incentivize NETs \nis deemed to be their integration into a single carbon market alongside \ntraditional mitigation strategies such as renewables8. In this paper, we \npoint out that this \u2018first-best\u2019 setting could entail major drawbacks: if the \ncarbon price is a tax from the perspective of carbon emitters, it would \nbe seen as a subsidy for carbon removal companies. The carbon prices \nnecessary to reach 1.5\u2009\u00b0C or well below 2\u2009\u00b0C are likely to exceed several \nhundred dollars per ton of CO2 (ref. 9). Meanwhile, carbon removal \ncosts might become as low as one to two hundred dollars per ton of \nCO2 avoided3. Therefore, unless NETs are assumed unbounded and \nthus behave like a perfect backstop technology (setting the price for \nthe carbon market), the profit margin for the companies owning NETs \ncould become excessively large in such an institutional set-up.\nThe public would pay for these windfall profits: up to net-zero \nemissions, NET revenues would be paid for by the carbon tax rev-\nenues, reducing other recycling options. After that, unless intertem-\nporal mechanisms such as carbon funds have been put into place10, \ngovernments will finance NETs by increasing taxes, issuing debt or \ncutting spending11.\nFinancing these potentially large profits with public funds suggests \nthat NETs can substantially influence the income distribution via two \nchannels. First, remunerating negative emissions reduces the amount \nof carbon tax revenues potentially available to tackle inequality and \nalleviate poverty12 (revenue drying effect). Second, if these technolo-\ngies are privately owned and are allowed to reap windfall profits due to \nincreasing carbon prices, their financing will disproportionately benefit \nthe high end of the income distribution that holds stocks and capital \n(ownership effect). Since the carbon tax is expected to be mostly regres-\nsive12\u201314 and equity ownership is concentrated at the top of the income \ndistribution, both mechanisms can increase inequality and erode the \nsupport for the ambitious climate goals of the Paris Agreement.\nReceived: 4 August 2022\nAccepted: 18 October 2023\nPublished online: 30 November 2023\n Check for updates\n1RFF-CMCC European Institute on Economics and the Environment, Fondazione Centro Euromediterraneo sui Cambiamenti Climatici, Milan, Italy. \n2Politecnico di Milano, Milan, Italy. \n\u2009e-mail: pietro.andreoni@eiee.org\n\nNature Climate Change | Volume 14 | January 2024 | 48\u201354\n49\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\nredistributed through neutral transfers to households (pink area). \nWhen NETs payments exceed carbon tax revenues, the government \nmust raise income taxes (dark blue area) to balance the budget.\nSince we do not consider the cost of public funds nor potential \naggregate growth effects due to variations in the income distribution, \nthese fluxes have no aggregate impact on the economy or emissions. \nHowever, they still influence welfare distribution because each flux \nimpacts different income deciles differently (Fig. 1b). Emission reduc-\ntion costs and the carbon tax are assumed to be regressive for rich \ncountries and progressive for developing countries, following empiri-\ncal evidence12. We do not consider dynamics such as employment \neffects and stranded asset ownership21 that could have distributional \nimplications22 both within and across societies. Costs and revenues of \nNETs are assumed to fall mainly on the wealthiest two deciles of each \ncountry, which own most stocks and capital. For better data availability, \nwe use wealth distribution as a proxy for equity ownership. In our main \nscenarios, we assume that all capital is owned exclusively by each coun-\ntry\u2019s citizens. Finally, the redistribution of carbon tax revenues offsets \nthe regressivity of the carbon tax. The income taxes raised after net \nzero are progressive to reflect the structure of most taxation systems \n(see the section \u2018Calibration of elasticities\u2019 in Methods and Extended \nData Fig. 1 for details).\nScenario design\nTo quantify the inequality implications of NET ownership, we run sce-\nnarios consistent with 1.5\u2009\u00b0C with 33% likelihood, imposing a carbon \nbudget of 700\u2009GtCO2 from 2019 to 21002 achieved through a global \nuniform carbon tax (see Supplementary Annex G for alternative speci-\nfications of the carbon tax and different carbon budgets). To assess the \nimplications of scenario design, we consider a case without overshoot \nof the carbon budget where emissions remain at net zero in exchange \nSeveral studies have found that an equitable allocation of CO2 \nremoval can shift part of the mitigation burden to countries with higher \nhistorical responsibility15\u201317. However, no existing literature considers \nthe potential repercussions of financing negative emissions at scale \nin shaping future inequality within societies. To explore and quan-\ntify these dynamics, we use an open-source integrated assessment \nmodel18. This model has high regional detail, featuring 57 interacting \ncountries/regions, and allows for a granular representation of income \ndynamics between countries. We extend it by differentiating income \ngroups within countries19 and introducing direct air capture (DAC) as \na representative NET (see Methods for details). DAC is an example of a \nNET which has catalysed notable financial interest; it is considered to \nhave considerable deployment potential and lower environmental and \nsocial trade-offs than biological NETs such as bioenergy with carbon \ncapture and storage (BECCS)3,20.\nWe quantify the inequality implications of financing negative \nemissions in 1.5\u2009\u00b0C scenarios under the assumptions of global carbon \ntaxation, a single carbon market for emission reductions and removal, \nand private ownership of carbon removal companies. We quantify the \nwithin-country and global inequality implications under a broad set of \nscenarios and assumptions.\nModelling NETs and their distributional \nimplications\nWe distinguish the economic implications of climate policies with six \ncomponents (Fig. 1a). Total mitigation costs (black dotted line) are \ncomposed of the sum of emission reduction costs (orange area) and \nNET costs, which include investment and operational costs (green area).\nThe remaining components sum to zero for each timestep and \ncountry. The government receives the carbon tax revenues (brown \narea), which are used to subsidize NETs at the carbon price or are \nNET costs (%)\n\u22125.0\n\u22122.5\n0\n2.5\na\nb\n2020\n2040\n2060\n2080\n2100\nCosts and gains,\n global aggregate (% GDP)\n0\n10\n20\n30\n40\nD1\nD2\nD3\nD4\nD5\nD6\nD7\nD8\nD9\nD10\nDecile\nYear\nPercentage of flow falling on decile D,\n country median \nCarbon tax, emission reduction costs and redistribution\nIncome taxes\nNETs profits\nSource\nCarbon tax payments\nCost of NETs\nEmission reduction costs\nIncome taxes\nNET revenues\nRedistribution\nFig. 1 | Characterization of aggregate and distributional effects of financial \nflows. a, Global (aggregated for all countries and deciles) financial flows and \nclimate costs in percentage of GDP, in a scenario with a 700\u2009GtCO2 carbon budget \nwith budget overshoot. The black dotted line represents global net costs\u2014that \nis, the sum of emission reduction costs (orange area) and NET costs (green \narea). The other fluxes sum to zero at each point in time. Income taxes (dark blue \narea) can only be raised to finance negative emissions after budget overshoot. \nCarbon tax revenues (brown area) may be used to pay for NET revenues (light \nblue area), and the excess is recycled back into the economy (pink area). b, Share \nof resources (median across countries) accruing to each income decile per type \nof flow according to the calibration of carbon tax, equity ownership and income \ntaxes\u2019 regressivity, in the year 2075. The black dashed line represents the median \nacross countries of income distribution across deciles. If the flow is a cost from \nthe perspective of the households (carbon taxes, emission reduction costs, \nincome taxes), it is regressive if it falls above the quantile line for low deciles and \nbelow for high deciles. If the flow is a gain from the perspective of the households \n(revenue recycling, NET revenues), it is regressive if it falls above the quantile \nline for high decile and below for low deciles. The net regressivity of a flow (for \nexample, financing NET revenues with carbon taxes or income taxes) can be \nvisualized as proportional to the distance between the two lines (light and dark \nblue or light blue and brown, respectively).\n\nNature Climate Change | Volume 14 | January 2024 | 48\u201354\n50\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\nfor faster short-term reductions (Fig. 2). This leads to higher carbon \nprices and policy costs before net zero and lower thereafter, confirming \nrecent evidence23,24. It also leads to a smaller temperature overshoot, \nwhich has been shown to reduce climate risks23. In this paper, we limit \nthe scope to mitigation and do not consider the climate damages: while \nhighly uncertain, they will only exacerbate our main findings given \ntheir adverse distributional consequences19.\nNegative emissions increase within-country \ninequality\nWe identify a clear connection between carbon removals and inequality \n(Fig. 3). Regions that deploy more NETs experience a higher increase in \ninequality (Fig. 3a as measured by the Gini index and see Supplemen-\ntary Annex L for alternative inequality metrics). The financing of NETs \nprimarily drives this effect. Using a Shapley\u2013Owen decomposition, we \ncan isolate the contribution of each economic flux to the total inequality \nvariation (bars in Fig. 3a): the percentage of inequality increase due to \nNET financing increases with brackets. It accounts for more than 70% \nof the total for the countries highly relying on carbon removal. Most \nof the total NET contribution to inequality is due to the ownership \neffect (filled bar in Fig. 3a). Partly, this is because we do not consider \nprogressive schemes for carbon tax recycling that would exacerbate \nthe revenue drying effect. We distinguish three phases in the evolution \nof inequality variation (Fig. 3b). In the beginning, inequality decreases \nsince the costs of NET technologies are initially higher than the carbon \nprice, and firms do not make profits (inset chart in Fig. 3b). Afterwards, \nas companies start making profits, inequality increases and peaks by the \ntime emissions reach net zero. Finally, in the latter part of the century, \nthe inequality increase is reduced: technology reaches maturity, its \ncosts do not fall anymore and the carbon prices level off after climate \nneutrality (Fig. 2). Furthermore, after net zero, negative emissions are \nfinanced by raising progressive income taxes instead of the carbon \ntax. In principle, it would be possible to redistribute tax revenues and \nraise income taxes to finance negative emissions even before net zero, \nwhich would reduce adverse distributional effects but could hinder \neconomic growth. It is worth noting that if government subsidies were \nto cover the early losses of NET companies, which is observed in the \ncurrent policy environment, then the early\u2014albeit small\u2014reduction in \ninequality would not happen, thus exacerbating the inequality reper-\ncussion of NET financing.\nThese results depend on the interaction between the amount of \ncarbon removed, the geographical distribution of CO2 removal, the \ncarbon price, NET costs and assumptions about wealth concentration. \nWe developed a simplified analytical model to identify the underly-\ning drivers (see the section \u2018A simple analytical model for NETs\u2019 in \nMethods, Supplementary Annex A for validation and derivation and \nAnnexe B for a sensitivity analysis on these drivers). We find that three \nfactors drive the distributional implications of NETs resulting from the \nownership effect:\n\t(a)\tThe regressivity of financing carbon removal, expressed in the \nmodel by the elasticities \u03b7wealth and \u03b7tax. The more concentrated \nthe capital (high \u03b7wealth) and the more regressive the taxes used \nto finance NETs (low \u03b7tax), the more negative emissions will cause \nan increase in inequality. In the standard set-up, the elasticities \nof wealth and taxes are constant, while the carbon tax elastic-\nity varies with income per capita. In the real world, capital con-\ncentration and taxation schemes\u2019 progressivity vary across \ncountries. However, wealth tends to be more concentrated than \nincome and carbon taxes (see the section \u2018Calibration of elas-\nticities\u2019 in Methods and Supplementary Annex C for a sensitivity \nanalysis on wealth concentration and income taxes).\n\t(b)\tThe share of NET revenues over GDP in each country. The larger \nthis term, the higher the inequality due to NETs. This factor de-\npends mainly on negative emissions potential. If this potential \nis concentrated in small economies, NET revenues could ac-\ncount for several percentage points of their GDP, exacerbating \nthe distributional effects. Therefore, these countries might be \nunwilling or unable to perform large amounts of removal with-\nout international financial transfers. If properly designed, such \ninternational transfers could increase feasibility of the carbon \nremoval effort and reduce between-country inequality (see Sup-\nplementary Annex E for a scenario including international trans-\nfers related to CO2 removal effort). However, these transfers \nwould still benefit the higher end of the income distribution in \nthe receiving countries, exacerbating internal inequality.\n\t(c)\tThe profit margin for a ton of carbon removed, expressed as the \nratio between NET revenues and costs. The higher this term, \nthe higher the inequality increase due to NETs. In our model, \nthe profit margin peaks at 80\u201390% at the time of net zero, when \ncarbon prices are the highest and technology has attained ma-\nUS$111 tCO2\n\u20131\nUS$952 tCO2\n\u20131\nUS$192 tCO2\n\u20131\nUS$1,031 tCO2\n\u20131\nWith overshoot\nWithout overshoot\n2020\n2040\n2060\nYear\nYear\n2080\n2100\n2020\n2040\n2060\n2080\n2100\n\u221210\n0\n10\n20\n30\nEmissions and removal (GtCO2 yr\u20131)\nCarbon removal\nIndustrial emissions\nLand use change\nFig. 2 | Emissions and carbon prices under a 1.5\u2009\u00b0C scenario with and without \novershoot. The solid black lines represent net emissions (sources \u2212 sinks) and \ndotted lines the average of carbon prices across regions, weighted by GDP. \nCarbon prices in 2030 and at the peak are highlighted. Cumulative CO2 removal \nin the century equals 332\u2009GtCO2 in the overshoot scenario and 127\u2009GtCO2 in the \npeak budget scenario. In the overshoot scenario, net-negative emissions from \nNETs account for 256\u2009GtCO2 through the century (compatible with the [0,680] \nnet-negative emission range of C1\u2013C3 scenarios of the IPCC Sixth Assessment \nReport).\n\nNature Climate Change | Volume 14 | January 2024 | 48\u201354\n51\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\nturity. We explore scenarios with more pessimistic technologi-\ncal change in the last section of the results, the consequences of \npolicy provisions to reduce the profit margin by considering a \nprofit cap for NETs in Supplementary Annex F and the role of dif-\nferent carbon pricing schemes in Supplementary Annex G.\nThe role of overshoot for inequality\nWe found that large-scale negative emissions increase within-country \ninequality. We now ask whether this result varies across scenarios with \nand without overshoot. In the \u2018no overshoot\u2019 scenario, each region is \nnot allowed to have net-negative emissions. This results in a global \ncumulative carbon removal of 126\u2009GtCO2, less than half of the amount \nsequestered in the scenario with overshoot. On the other hand, the no \novershoot scenario has higher emission reductions in the first part of \nthe century.\nWe find that the no overshoot scenario has higher inequality \nbefore net zero as emission reductions, which are regressive, are \nfrontloaded (Fig. 4). This increase is more accentuated for regions \nwith high removal since more negative emissions are deployed in the \n2030s and 2040s24 and these are financed with a higher carbon price \n(Fig. 2). However, negative emissions are lower in the second half of \nthe century if the budget is not overshot. Consequently, the inequality \nimplications of negative emissions are reduced, especially for highly \nimpacted countries (fourth bracket).\nThus, promoting net-negative emissions by overshooting emis-\nsions and temperature leads to an intertemporal inequality trade-off \nbetween regressive mitigation in the short term and regressive NET \nownership in the long term. For the countries that are more impacted by \nthe inequality repercussions of negative emissions, Fig. 4 indicates that \na smaller overshoot is preferable. Furthermore, while not estimated in \nthis study, climate damages avoided with lower temperature overshoot \nwould be likely to have progressive effects and support the rationale \nfor reducing overshoot.\nThe intertemporal trade-off we identified partly rests on the \nassumption that emission reduction costs are, for most countries, \nregressive, in line with the literature19. However, for regions with high \nNET deployment, most of the increase in inequality between the over-\nshoot and no overshoot scenarios in the first part of the century is \n22%\n50%\n63%\n72%\nRevenue drying\nOwnership\nAbatement\n0\n0.5\n1.0\n0\n20\n40\n60\nCarbon removed \n(% abated + removed)\nYear\nYear\nInequality increase, net-zero year\n (\u2206 Gini points) \nCarbon sequestered (GtCO2)\n10\n20\n30\n40\n50\na\nBreakeven year\n0\n0.25\n0.50\n0.75\n1.00\n2020\n2040\n2060\n2080\n2100\nChange in Gini index (points)\nCarbon removed \n(% abated + removed)\nLow\nMedium\nHigh\nVery high\n\u2212100\n\u221250\n0\n50\n100\n2040 2070 2100\nProfit margin (%)\nb\nFig. 3 | Within-country inequality variation due to climate policy in a 1.5\u2009\u00b0C \nscenario with overshoot. a, Inequality increase expressed as difference in \nGini index relative to the baseline in 2070 against cumulative carbon removed \nas share of total mitigation effort (carbon removed plus abated). Each point \nrepresents a country/region. Bubble size indicates the amount of carbon \nremoved until 2100. Countries/regions are clustered into four brackets using an \nequal frequency method according to their share of removal over total mitigation \neffort: Low\u2009=\u2009(0%, 3.5%); Medium\u2009= (3.5%, 6%); High\u2009= (8%, 18%); Very high\u2009= (18%, \n65%). The map shows which countries belong to each bracket. Bars show, for \neach bracket, the median contribution of emission reduction costs (unfilled \nbar), the ownership effect (filled bar) and the revenue drying effect (striped bar). \nThe percentage of the total increase due to NETs is highlighted (that is, sum of \nownership and revenue drying effects). The dashed grey tendency line shows \nthe positive correlation among the plotted variables. b, Time trend of inequality \nincreases per countries/regions, clustered as in a. The small time chart represents \nthe median profit margin for NET companies over time starting from 2040. \nBreakeven year is when the median profit margin becomes positive. Shaded areas \nrepresent confidence intervals (33rd and 66th percentiles).\n\u22120.50\n\u22120.25\n0\n0.25\n2020\n2040\n2060\nYear\n2080\n2100\nInequality increase (\u2206 Gini points)\nCarbon removed \n(% abated + removed)\nLow\nMedium\nHigh\nVery high\nFig. 4 | Inequality difference (variation in the Gini index) in the scenario \nwithout budget overshoot relative to a scenario with budget overshoot. \nCountries/regions are clustered by their share of cumulative removal over \ntotal mitigation effort in the overshoot scenario (that is, the brackets and  \nthe countries belonging to each bracket are the same as in Fig. 3). As in  \nFig. 3, bars represent the median inequality increase per bracket in selected \nyears (2050, 2075, 2100) and filled bars represent the median contribution  \nto the inequality increase due to NETs (sum of ownership and revenue  \ndrying effects). Shaded areas represent confidence intervals (33rd and  \n66th percentiles).\n\nNature Climate Change | Volume 14 | January 2024 | 48\u201354\n52\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\ndue to NETs financing. Therefore, the results are less dependent on \nassumptions about the emission reduction costs.\nGlobal inequality repercussions\nGlobal inequality is affected by negative emissions not only through \nthe within-country effect we have analysed so far but also through \nbetween-country inequality, since the geographical distribution of \nnegative emissions varies across regions. If rich countries take up most \nof the carbon removal effort and bear the associated costs, this could \ncompensate for the inequality impacts seen so far, with ramifications \nfor global equity.\nWe quantify the between-country effects of different interna-\ntional distributions of removal effort and compare them to the within- \ncountry effect. In our model, the costs of NETs and the international  \ndistribution of the removal effort drive between-country inequality. \nSince both dimensions are highly uncertain, we consider two cost \nprofiles for the NET (low costs and high costs) and two international \ndistributions of removal potential (Global North and Global South) \nto compare the within- and between-country effects of NETs (Fig. 5, \nopaque and transparent areas). We group the inequality contributions \ninto NET costs (green areas) and NET financing (blue areas), which \ninclude the sum of the carbon tax, income taxes, NET revenues and \ntransfers. This last factor affects within- but not between-country \ninequality because we assume it does not change the aggregate GDP.\nAcross all combinations of NET costs and potential (Fig. 5 and \nSupplementary Annex L for alternative inequality metrics), the \nwithin-country effects (sum of opaque blue and opaque green areas) \nare dominant compared to between-country ones (transparent green \narea). In our central specification of DAC potential and costs (Global \nNorth, low costs), the between-country effect of CO2 removal is an order \nof magnitude lower than the within-country one around the time of \nnet zero. At the end of the century, the effect is more relevant because \nthe total cost of removal peaks, but it is insufficient to compensate \nfor within-country regressivity. An allocation of removal potential \ntowards the Global South consistently exacerbates global inequal-\nity, unless trade of carbon removal is considered. In this case, poorer \ncountries could benefit from selling negative emissions permits to \nwealthy countries (see Supplementary Annex E for a simulation of \ninternational transfers). On the contrary, relaxing the assumption that \ncapital is owned exclusively by each country\u2019s citizens would increase \nbetween-country inequality because most financial centres are located \nin the Global North (see Supplementary Annex D for a scenario includ-\ning international ownership of capital).\nAssuming higher costs of negative emissions amplifies their \nbetween-country distributional effects. At the same time, it reduces \nthe within-country regressivity because it decreases the profit margin \nfor NET companies. Towards the end of the century, under a progres-\nsive allocation of removal potential (Global North, high costs) the two \neffects roughly counterbalance (bottom-right panel of Fig. 5). These \nresults indicate that a progressive international distribution of removal \neffort, or a properly designed set of international transfers (Supple-\nmentary Annex E), can indeed contribute to reducing global inequality. \nHowever, especially under the assumption of cheap NETs and around \nnet zero, our results suggest that the within-country inequality increase \nis more relevant in shaping the global income distribution.\nDiscussion on policy instruments and key \nassumptions\nThe results described so far rest upon two main policy assumptions: \nNETs are privately owned and a single carbon market exists for emission \nreduction and removal. While these assumptions align with the current \npolicy landscape, alternative policy choices can mitigate the inequal-\nity implications of NETs. First, the regressive nature of financing NETs \ncould be alleviated by altering either the ownership structure of NETs \ncompanies or the taxation system. Taxes levied to finance NET revenues \ncould be tailored to target the same income brackets as those who own \nthe negative emissions industry. Alternatively, the public sector could \nparticipate in or own the NET companies (see Supplementary Annex C \nfor sensitivity on NET ownership and taxation systems). Second, the \nmarket structure could be redesigned to lower the profit margin. For \ninstance, separating markets for emission reduction and CO2 removal \ncould reduce the carbon price applicable to NET companies and thus \ntheir profit margin. Alternatively, the maximum profit allowed for \nthese companies could be capped even within a single carbon market \n(see Supplementary Annex F for scenarios with profit caps for NETs).\nAll these policy design options face hurdles and limitations. Pro-\ngressive taxes would alleviate the inequality increase after net zero, \nbut with limited effect until NET revenues are paid for with carbon \ntax revenues. Separated targets for emission reduction and removal \nmight lead to under- or overprovision of negative emissions. Similarly, \nTotal NETs efect\n2075\n2100\nLow costs\nHigh costs\nGlobal North\nGlobal South\nGlobal North\nGlobal South\n\u22120.2\n0\n0.2\n0.4\n0.6\n0.8\n\u22120.2\n0\n0.2\n0.4\n0.6\n0.8\nInequality driver\nNETs costs\nNETs transfers\nInequality contribution\nBetween\nWithin\nFig. 5 | Global inequality variation due to NET deployment (excluding \nemission reduction effects on inequality) as the difference in the mean log \ndeviation index relative to the baseline, in selected years and scenarios. The \ninequality contribution of NETs is divided into four factors: between-countries \neffect of NET costs (green area, transparent), within-countries effect of NET \ncosts (green area, opaque), within-countries effect of NET financing (blue area, \ntransparent) and between-countries effect of NET financing (blue area, opaque). \nThe last factor is always null because NET financing does not affect GDP but only \nits distribution. NET financing (blue areas) groups the algebraic sum of carbon \ntax payments, redistribution of residual carbon tax revenues, NET revenues and \nincome taxes. Dots indicate the net inequality contribution of CO2 removal. The \ncontribution of the total inequality increase due to each group is decomposed \nvia a Shapley\u2013Owen decomposition of the mean log deviation index. For each \ngroup, the Theil index is calculated in its within- (transparent areas) and between- \n(opaque areas) country contributions. The Theil index is an entropy measure \noften used to measure inequality. Unlike the Gini index, it allows to disaggregate \nthe within- and between-country contributions of total inequality variation. The \nShapley\u2013Owen decomposition allows us to determine how much each income \nflow contributes to the total variation of inequality. In the \u2018low costs\u2019 scenario, \nour representative technology reaches an average cost of around US$150 per \ntCO2 avoided in 2100, while in the \u2018high costs\u2019 scenario the cost is around US$250 \nper tCO2 avoided in 2100. In the \u2018Global North\u2019 scenario, the upper bound \nconstraint on negative emissions is distributed regionally following our central \nspecification case\u2014that is, geological storage capacity. In the \u2018Global South\u2019 \nscenario, the constraint is allocated with an equal per capita rule\u2014that is, more \npopulous countries perform more CO2 removal.\n\nNature Climate Change | Volume 14 | January 2024 | 48\u201354\n53\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\nprofit or price caps would distort the market, incentivize strategic \nbehaviour from the companies, might undermine long-term incentives \nfor cost reduction and would reduce the attractivity of investments in \nNETs. Nonetheless, especially around and after net-zero emissions, the \nseverity of the distributional dynamics we highlighted suggests that \nthe drawbacks of a fully integrated carbon market justify preventive \nintervention. At the same time, alternative policy provisions should \nprovide adequate attractiveness and profitability for investments in \nNETs (see Supplementary Annex F for a quantification of this trade-off) \nand should strive to maintain overall policy efficiency. Therefore, the \neconomic and political economy implications of the different policy \noptions sketched here should be carefully explored.\nRegarding the broad mitigation policy context, relaxing the cli-\nmate target reduces the distributional concerns of financing NETs \nbecause of the contextual reduction of negative emissions and carbon \nprices (see Supplementary Annex G for a 1,300\u2009GtCO2 carbon budget). \nHowever, higher temperature increases climate change impacts and \ntheir adverse distributional implications, which are not considered \nin this setting.\nThe inequality consequences of financing NETs described in \nthis paper depend on policy design but also on the characteristics of \nthe technology. Here we focus on DAC, but other negative emission \nstrategies exist. For example, considering BECCS would affect some \nof our implicit assumptions about the business model, value chain \nand regional potential (Box 1). Similarly, other technologies such as \nenhanced rock weathering or ocean alkalinization come with their own \nspecificities. To fully understand how different technologies shape \nthe distributional implications of a diversified CO2 removal portfolio, \na more detailed approach is necessary. Nonetheless, the dynamics \ndescribed so far would at least partly hold true for other technologies \nso long as privately owned NET companies make large profits from \nthe carbon market.\nConclusions\nThe framework we have developed\u2014incorporating technological and \nsocial dynamics into a climate-economy model\u2014allows us to highlight \nthe adverse distributional implications of financing negative emissions \nin the race to meet the Paris Agreement targets. Although robust to \ndifferent specifications, our results rest upon assumptions that, while \ncommon in the climate mitigation literature (for example, a single \ncarbon market) or reasonable given the current policy environment \n(for example, private ownership of NET companies), depend on policy \nand ethical choices: who should finance, within and across societies, \nthe provision of a global public good such as CO2 removal, who should \nown the means to provide this public good, and how to design policies \nwhich avoid unintended and detrimental consequences? Our findings \nsuggest that, when confronted with these questions, policymakers \ndesigning institutions for effective and fair CO2 removal should focus \non their distributional implications: policy provisions that affect the \nownership structure and the market for NETs can mitigate the inequal-\nity concerns described here but require political capital and a sound \nunderstanding of the repercussion of these crucial climate strategies. \nMethodologically, the paper shows that, when evaluating decarboni-\nzation strategies, there is a need for models that are able to capture \nthe interactions of technology and finance with societal goals such as \ninequality control.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-023-01870-7.\nReferences\n1.\t\nRathi, A. Stripe, Alphabet and others to spend nearly $1 billion  \non carbon removal. BNN Bloomberg www.bnnbloomberg.ca/ \nstripe-alphabet-and-others-to-spend-nearly-1-billion-on-carbon- \nremoval-1.1751098 (2022).\n2.\t\nIPCC Climate Change 2022: Mitigation of Climate Change  \n(eds Shukla, P. R. et al.) (Cambridge Univ. Press, 2022).\n3.\t\nFuss, S. et al. Negative emissions\u2014part 2: costs, potentials and \nside effects. Environ. Res. Lett. 13, 063002 (2018).\n4.\t\nSmith, P. et al. Biophysical and economic limits to negative CO2 \nemissions. Nat. Clim. Change 6, 42\u201350 (2016).\n5.\t\nMcLaren, D. Quantifying the potential scale of mitigation \ndeterrence from greenhouse gas removal techniques. Climatic \nChange 162, 2411\u20132428 (2020).\n6.\t\nGrant, N., Hawkes, A., Mittal, S. & Gambhir, A. Confronting \nmitigation deterrence in low-carbon scenarios. Environ. Res. Lett. \n16, 064099 (2021).\n7.\t\nMarkusson, N., McLaren, D. & Tyfield, D. Towards a cultural \npolitical economy of mitigation deterrence by negative emissions \ntechnologies (NETs). Glob. Sustain. 1, e10 (2018).\n8.\t\nRickels, W., Proel\u00df, A., Geden, O., Burhenne, J. & Fridahl, M. \nIntegrating carbon dioxide removal into European emissions trading. \nFront. Clim. https://doi.org/10.3389/fclim.2021.690023 (2021).\n9.\t\nIPCC: Summary for Policymakers. In Climate Change 2022: \nMitigation of Climate Change (eds Shukla, P. R. et al.) (Cambridge \nUniv. Press, 2022).\n10.\t Bednar, J. et al. Operationalizing the net-negative carbon \neconomy. Nature 596, 377\u2013383 (2021).\n11.\t\nBednar, J., Obersteiner, M. & Wagner, F. On the financial viability \nof negative emissions. Nat. Commun. 10, 1783 (2019).\n12.\t Budolfson, M. et al. Climate action with revenue recycling has \nbenefits for poverty, inequality and well-being. Nat. Clim. Change \n11, 1111\u20131116 (2021).\n13.\t Steckel, J. C. et al. Distributional impacts of carbon pricing in \ndeveloping Asia. Nat. Sustain 4, 1005\u20131014 (2021).\n14.\t Ohlendorf, N., Jakob, M., Minx, J. C., Schr\u00f6der, C. & Steckel, J. C. \nDistributional impacts of carbon pricing: a meta-analysis. Environ. \nResour. Econ. 78, 1\u201342 (2021).\nBox 1\nOther NETs could entail \ndifferent distributional \nimplications\nUnlike DAC, the supply chain of BECCS involves biomass, the \nsale of which could transfer part of the revenues to farmers with \npositive distributional effects. This could reduce the regressivity of \nfinancing carbon removal as explained in the \u2018Negative emissions \nincrease within-country inequality\u2019 section. Furthermore, the \naverage net cost of removal with BECCS is generally assessed to \nbe lower than with DAC3, therefore increasing the profit margin at \nthe same carbon price and exacerbating inequality. Finally, one of \nthe main constraints for BECCS deployment is biomass availability, \nwhich would move the international distribution of removal effort \nto the Global South. Depending on whether international transfers \nare considered, this could have positive or detrimental global \ninequality implications. However, we showed in Fig. 5 that different \ninternational distributions of removal effort tend to have a smaller \nimpact on global inequality relative to ownership effect, especially \nfor technologies with a lower average cost, as is the case for BECCS.\n\nNature Climate Change | Volume 14 | January 2024 | 48\u201354\n54\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\n15.\t Pozo, C., Gal\u00e1n-Mart\u00edn, \u00c1., Reiner, D. M., Mac Dowell, N. & \nGuill\u00e9n-Gos\u00e1lbez, G. Equity in allocating carbon dioxide removal \nquotas. Nat. Clim. Change 10, 640\u2013646 (2020).\n16.\t Lee, K. S. B., Fyson, C. & Schleussner, C.-F. Fair distributions of  \ncarbon dioxide removal obligations and implications for effective  \nnational net-zero targets. Environ. Res. Lett. https://doi.org/10.1088/ \n1748-9326/ac1970 (2021).\n17.\t Fyson, C. L., Baur, S., Gidden, M. & Schleussner, C.-F. Fair-share \ncarbon dioxide removal increases major emitter responsibility. \nNat. Clim. Change 10, 836\u2013841 (2020).\n18.\t Gazzotti, P. et al. Persistent inequality in economically optimal \nclimate policies. Nat. Commun. https://doi.org/10.1038/s41467-\n021-23613-y (2021).\n19.\t Dennig, F., Budolfson, M. B., Fleurbaey, M., Siebert, A. & Socolow, \nR. H. Inequality, climate impacts on the future poor, and carbon \nprices. Proc. Natl Acad. Sci. USA 112, 15827\u201315832 (2015).\n20.\t Minx, J. et al. Negative emissions\u2014part 1: research landscape and \nsynthesis. Environ. Res. Lett. 13, 063001 (2018).\n21.\t Semieniuk, G. et al. Stranded fossil-fuel assets translate to major \nlosses for investors in advanced economies. Nat. Clim. Change 12, \n532\u2013538 (2022).\n22.\t Semieniuk, G. et al. Potential pension fund losses should not \ndeter high-income countries from bold climate action. Joule 7, \n1383\u20131387 (2023).\n23.\t Drouet, L. et al. Net zero-emission pathways reduce the physical \nand economic risks of climate change. Nat. Clim. Change 11, \n1070\u20131076 (2021).\n24.\t Riahi, K. et al. Cost and attainability of meeting stringent climate \ntargets without overshoot. Nat. Clim. Change 11, 1063\u20131069 \n(2021).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2023\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\nMethods\nThe baseline model\nRICE50+ is an open-source highly regionalized integrated assessment \nmodel18. It represents 57 different countries and regions, each char-\nacterized by a marginal abatement cost curve function calibrated on \ndetailed-process models and by projections about socioeconomic \ndrivers and baseline emission intensity of GDP. Abatement costs vary \nby region, such that fully abating emissions can cost hundreds to thou-\nsands of dollars per ton of CO2, and are subject to exogenous technolog-\nical learning. The model solves a problem of discounted consumption \nmaximization with a time horizon until 2300 and five-year timestep \nperiods. The model allows solutions with multiple possible coalitions, \nincluding the singleton coalitions (Nash solution) and the grand coali-\ntion (cooperative). Each planner solves an intertemporal maximization \nproblem of discounted consumption and the utility function explicitly \nincludes inequality aversion other than intertemporal25. Different \nbaseline projections of socioeconomic determinants are represented \nwith the shared socioeconomic pathways (SSPs)26. The model is open \nsource and available at https://github.com/witch-team/RICE50xmodel \nand a description and presentation of the model is available at https://\ndoi.org/10.18174/sesmo.18038 (ref. 27).\nImplementation of the income distribution\nTo model the distributional effects of NETs, we introduce a decile-based \nrepresentation of household heterogeneity, conceptually following \nDennig et al.19.\nAs in the standard DICE model, in our model final income Y is cal-\nculated for each timestep t and region n by subtracting from the gross \nincome YGROSS that emerges from a Cobb\u2013Douglas production function \n(calibrated to match exogenous GDP projections) the cost of emission \nreductions and carbon removal:\nYGROSS (t, n) = tfp (t, n) \u00d7 K (t)\n\u03b1 \u00d7 pop (t, n)\n1\u2212\u03b1\n(1)\nwhere tfp is the total factor productivity, calibrated to match exog-\nenous GDP projections from different socioeconomic projections \n(SSPs) assuming an optimal level of savings rate (S). Population (pop) \nis exogenous and also follows SSP projections, while capital (K) at each \ntimestep is given by (\u2202k is the depreciation rate of capital)\nK (t + 1, n) = K (t, n) \u00d7 (1 \u2212\u2202k) + S (t, n) \u00d7 Y(t, n)\n(2)\nThe final output Y is derived by subtracting emission reduction \ncosts (COSTER) and NET costs (COSTNET):\nY (t, n) = YGROSS (t, n) \u2212COSTER (t, n) \u2212COSTNET (t, n)\n(3)\nTo assign these costs to each decile d, we implement weights w\u03b7 \nthat are driven by an elasticity \u03b7:\nw\u03b7 (t, n, d) =\nq0(t, n, d)\n\u03b7\n\u2211dq0(t, n, d)\n\u03b7\n(4)\nwhere q0(t,n,d) are the baseline quantiles, obtained from a disaggrega-\ntion of Gini index projections to 2100 from Rao et al.28, as described in \nref. 29. The definition of w\u03b7 gives that \u2211\n10\nd=1 w\u03b7 (t, n, d) = 1.\nThe elasticities \u03b7 defined as such imply that if \u03b7\u2009=\u20090 the costs \nare distributed equally per capita and therefore will benefit the \nrichest deciles, and if \u03b7\u2009=\u20091 the cost is assigned distribution neutrally \nacross the deciles. Therefore, the higher the elasticity, the more \nprogressive the distribution of the cost. If the flow is positive (that \nis, it describes gains, such as the recycling of the carbon tax or NET \nrevenues) the opposite holds, and a lower elasticity indicates more \nprogressive flows.\nThe final income for each decile YDIST is described by\nYDIST (t, n, d) = YGROSS (t, n) \u00d7 q0 (t, n, d)\n\u2212(COSTABATE (t, n) + CTX (t, n))\n\u00d7wctax(t, n, d) + (REVNET (t, n) \u2212COSTNET (t, n))\n\u00d7wnet (t, n, d) + TRANSFER (t, n) \u00d7 wredist\n(t, n, d) \u2212GENTAX(t, n) \u00d7 wtax (t, n, d)\n(5)\nwith each term defined as follows. CPRICE is the carbon price  \nin the region, EIND the emissions from the industry and energy sector, \nENEG the carbon removed with NETs (nature-based solutions are not paid \nfor since land use change is exogenous in the model) and\nCTX (t, n) = CPRICE (t, n) \u00d7 EIND(t, n)\nREVNET (t, n) = CPRICE (t, n) \u00d7 ENEG(t, n)\nGENTAX (t, n) = max (0, REVNET (t, n) \u2212CTX (t, n))\nTRANSFER (t, n) = max (0, CTX (t, n) \u2212REVNET (t, n))\nSince we do not consider the marginal cost of raising public funds \nnor intertemporal and international transfers, the algebraic sum across \ndeciles of carbon tax revenues, subsidies to NETs, income taxes and \ntransfers must for each period and region equal zero:\nGENTAX (t, n) + CTX (t, n) \u2212REVNET (t, n) \u2212TRANSFER (t, n) = 0\nThis is equivalent to modelling an implicit government tasked with \nbalancing the difference in its budget due to climate policy. Because \n\u2211\n10\nd=1 w\u03b7 (t, n, d) = 1, it holds also that\n\u2211dREVNET (t, n) \u00d7 wnet (t, n, d) + TRANSFER\n(t, n) \u00d7 wredist (t, n, d) \u2212GENTAX (t, n)\n\u00d7wtax (t, n, d) \u2212CTX (t, n) \u00d7 wctax (t, n, d) = 0\nfrom which it follows that Y (t, n) = \u2211dYDIST (t, n, d), providing the \nconsistency between the aggregate income equation (equation (3)) \nand the income equation with resolved inequality (equation (5)).\nA simple analytical model for NETs\nWe derive an analytical reduced-form model that only considers \nthe direct contribution of NETs to within-country inequality, to \nanalyse the determinants of the behaviour of the numerical model. \nThis model excludes the revenue drying effect\u2014that is, the fact \nthat carbon tax revenues used for financing NETs could have been \notherwise redistributed progressively. From the income equation \n(dependency from time and region is dropped for readability, sub-\nscript d indicates the variable or parameter that depends on the \ndecile), considering only the terms linked to NET revenues and \nneglecting the contributions of abatement costs and the carbon \ntax redistribution, we get\nYd = Y0,d + REVNET \u00d7 (wnet\nd\n\u2212wtax\nd ) \u2212COSTNET \u00d7 wnet\nd\n(6)\nNormalizing for total income Y0 and rearranging variables (see \nSupplementary Annex A for the analytical proof and validation)\nqd \u00d7 gdploss = qd,0 + revgdp \u00d7 (profmargin \u00d7 wnet\nd\n\u2212wtax\nd )\n(7)\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\nwith the newly defined parameters defined as follows:\n\u2022 \nprofmargin =\nENEG\u00d7CPRICE\u2212COSTNET\nENEG\u00d7CPRICE\n is the profit margin of NETs\n\u2022 \nrevgdp =\nENEG\u00d7CPRICE\nY0\n is the revenue of NETs relative to baseline GDP\n\u2022 \nwnet\nd  is the equity share of quantile d, from wnet\nd\n=\nqd,0\n\u03b7net\n\u2211iqd,0\n\u03b7net\n\u2022 \nwtax\nd  is the share of tax paid by quantile d, from wtax\nd\n=\nqd,0\n\u03b7tax\n\u2211iqd,0\n\u03b7tax\n\u2022 \nqd,0 is the baseline share of income accruing to quantile d\n\u2022 \nqd is the final share of income accruing to quantile d\n\u2022 \ngdploss =\nY0\u2212COSTNET\nY0\n is the aggregate GDP loss due to NET costs\nTo exemplify the behaviour of this model, let us assume a repre-\nsentation by just quantiles of the economy, where the bottom 50% \nhold 20% of national income and the top 50% hold the remaining 80%, \nthat is, q1,0\u2009=\u20090.2 and q2,0\u2009=\u20090.8. Let us assume a carbon price of US$500 \nper tCO2 and an average cost for NETs of US$100 per tCO2 of carbon \navoided, that is, a profit margin of 0.8. Furthermore, let us assume \n\u03b7net\u2009=\u20091.8 and \u03b7tax\u2009=\u20090.8 (the median elasticity of the carbon tax in our \ncalibration), and that NET revenues cover 5% of GDP of the nation. \nUnder this assumption, the following will hold:\nq1 = 0.2 + 0.05 ((1 \u22121\n5 ) \u00d7 0.076 \u22120.248) = 0.8\nq2 = 0.8 + 0.05 ((1 \u22121\n5 ) \u00d7 0.924 \u22120.752) = 0.19\nNow with q1\u2009+\u2009q2\u2009=\u20090.99 because the cost of NETs is lost from the \neconomy, gdploss\u2009=\u20090.01. Renormalizing, we get q1\u2009=\u20090.81, q2\u2009=\u20090.19, and \ncalculating the difference in the Gini index before and after the NET \neffect gives\n\u0394Gini = Gini (0.81, 0.19) \u2212Gini (0.2, 0.8) = 0.01\nThe net effect of NET deployment under these assumptions causes \nan inequality increase of 1 point of the Gini index.\nCalibration of elasticities\nAbatement cost and carbon tax elasticities are calibrated as in Budolf-\nson et al.12, in which they perform a meta-study on the distributional \neffect of carbon taxation and regress the result against GDP, finding a \nweak negative correlation. To complete our model, two more calibrated \nparameters are needed: the income-to-capital-ownership elasticity and \nthe income-to-taxation elasticity.\nTo estimate the first one, since data on capital ownership are avail-\nable for only a small subset of countries and years, we use the wealth \nincome distribution as a proxy of capital ownership. We use a panel \nof data of 37 countries from 2000 to 2020 at the decile level from the \npublicly available WID database (https://wid.world/). These are mostly \nOECD countries for which both pre- and post-tax income distribution \nis available. We filter the years before 2000 to get a large but relatively \nhomogeneous sample of the characteristics of the global economy. We \nconsider post-tax income and wealth distributions. We estimate the \nelasticity by fitting the following model for each country and year with \nnonlinear least square methods (nls function in stats package R), where \nq\u00b0,d represents the quantiles on wealth and income (pre-tax income is \nspecified with the \u2018pre\u2019 subscript):\nqwealth,d =\nqincome,d\n\u03b7wealth\n\u2211i qincome,d\n\u03b7wealth d {1 \u223610}\n(8)\nWe assume perfect coincidence between the deciles of income \nand the deciles of wealth; that is, a household in the nth decile for \nthe income distribution will belong to the same decile of the wealth \ndistribution.\nThe elasticity of the taxation system is extrapolated in a similar \nmanner. First, we build the percentage of total government revenues \nfrom taxes per percentile, calculated as\ntd =\nqincome\u2212pre,d \u2212qincome,d\n\u2211i qincome\u2212pre,d \u2212qincome,d\nd {1 \u223610}\n(9)\nThis quantity is not a common metric of taxation progressivity \nbut represents, for a dollar of taxes, how much on average is taken from \neach percentile of the income distribution. We then relate this quantity \nas before, estimating the following model for each country and year:\ntd =\nqincome,d\n\u03b7tax\n\u2211i qincome,d\n\u03b7tax d {1 \u223610}\n(10)\nThe resulting estimated \u03b7tax and \u03b7wealth are shown in Extended Data \nFig. 1. The median of both values is relatively stable and we choose a \nvalue of 1.8 for \u03b7wealth and of 1.4 for \u03b7tax.\nScenarios\nWe implement the global constraint on cumulative CO2 emissions \nvia a global uniform carbon tax, which increases exponentially at a \ngrowth rate of 5% per year. In the first periods, the trajectory of the tax \nis smoothed out from the exponential path to avoid abrupt changes in \nthe pricing profile. To mimic the existence of hard-to-abate sectors and \nemissions, the maximum rate of abatement via conventional mitigation \nMIUmax is capped at 0.975 for all countries, corresponding to 97.5% of \nbaseline emissions and around 2\u2009GtCO2 in 2100 globally. Non-CO2 gases \nare not priced and do not contribute to the global budget.\nTo decrease the computational complexity of the problem and \nto avoid implicit transfers across regions via mitigation effort shift-\ning, each region solves as an independent optimization and conver-\ngence is achieved by solving iteratively all regions. Nonetheless, the \nsuperimposed shape of the global carbon tax produces a de facto \ncooperative scenario even though each regional optimization prob-\nlem is solved independently, in the sense that implicit cooperation \nis assumed in negotiating (and abiding to) the carbon tax level and \ngrowth across countries.\nAbatement is determined at each iteration by the carbon tax tra-\njectory and saving rates are fixed to the long-term convergence value \nof 25%. Therefore, the only free variable in the model is investment in \nnegative emissions. At each iteration, the starting level of the carbon \ntax is shifted via a bisection algorithm until the budget is reached and \nall free variables are stable across iterations.\nWithin the optimization problem at each iteration, the carbon \ntax is fixed and constrains the emission control rate at the equivalent \npoint in the marginal abatement cost curve MACt,n (\u00b0), such that, for \neach region n and period t\nMACt,n (MIU(t, n)) = CPRICE (t, n)\n(11)\nwith the carbon price, CPRICE, defined as\nCPRICE (t, n) = min (ctax(t), CPRICEmax (t, n))\nwhere CPRICEmax(t,n) = MACt,n(MIUmax(t,n)) and ctax is the exogenous \ncarbon tax fitted iteratively. This correction is to avoid, in the late part \nof the century, that the carbon price stays below the maximum sensible \nmarginal cost in each region.\nThe carbon price so defined also regulates the amount of NETs \nin the system thanks to a subtractive term that modifies the income \nequation (equation (3)) into:\nY (t, n) = YGROSS (t, n) \u2212COSTER (t, n) \u2212COSTNET (t, n) \u2212CPRICE (t, n)\n\u00d7 (E (t, n) \u2212Eit\u22121 (t, n))\n(12)\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\nwhere E represents total emissions and Eit\u22121 are the emissions in the \nprevious model iteration. Total emissions E are written as\nE (t, n) = EIND(t, n) \u2212Eneg (t, n) \u2212Eland (t, n)\n(13)\nwith EIND(t,n)\u2009=\u2009Ebaseline(t,n)\u2009\u00d7\u2009(1\u2009\u2212\u2009MIU(t,n)), and baseline emissions \nEbaseline calibrated from SSP projections and exogenous land use emis-\nsions Eland.\nSince MIU is fixed at each model iteration by the carbon tax, the \nonly free decision variable in the model is the investment in the repre-\nsentative NETs technology that regulates the amount of gross negative \nemissions.\nEquation (12) therefore ensures that, given the carbon price at \neach iteration of the model, negative emissions will be deployed if \nthe marginal cost of deployment (including storage) is lower than \nthe carbon price and no other constraint is reached. At convergence, \nthe difference between emissions across the past two iteration is \nbelow a certain threshold of tolerance E(t,n)\u2009\u2212\u2009Eit\u22121(t,n) < eps and the \nsolution stabilizes, while the term \u2212CPRICE(t,n)\u2009\u00d7\u2009(E(t,n)\u2009\u2212\u2009Eit\u22121(t,n)) \nin equation (12) goes to approximately zero in the income equation \n(equation (3)), therefore producing no effect in the final solution \nof the model.\nDAC as a representative technology for NETs\nIn the standard release of RICE50+, negative emissions are allowed, \nas in the original DICE model, because the decision variable MIU(t,n) \nthat controls the percentage of baseline emissions abated can reach \nhigher values than 100% (specifically 120%, meaning that each region \ncan remove at most 20% of its baseline emissions in each given year). \nSince CPRICE\u2009=\u2009MACt,n(MIU(t,n)), this removal happens at the top of \nthe marginal abatement cost curve, for carbon prices that can exceed \nUS$1,000 per tCO2. However, most proposed NETs4 are projected to \nbe competitive at much lower costs, ranging from US$100 per tCO2 \nof BECCS to US$200 per tCO2 of DAC. While there is large uncertainty \nsurrounding these costs, this means that such technologies would \nbecome competitive at lower levels of the marginal abatement costs \ncurve, de facto modifying its shape. Moreover, embedding negative \nemissions into a single MAC does not allow the explicit representation \nof offsetting that is necessary in our set-up to model the distributional \nimplications of financing NETs even before net-zero emissions.\nTherefore, we explicitly model a representative technology for \nNETs. We characterize it as an investable technology in the model, such \nthat each region can choose to reduce emissions with either standard \nabatement or via carbon removal. Since we explicitly model NETs, MIU \nis capped below 100%, meaning the investable technology is the only \nmeans to reach net-negative emissions.\nWe choose to model our representative NET technology as DAC. \nDAC is modelled as in Realmonte et al.30 as an investable technology \nwith depreciating capital, and emissions captured must be stored \ngeologically. Total costs are divided into investment and variable \ncosts (which implicitly include fuel and operation and maintenance \ncosts), as well as storage costs. Total costs shrink with time due to learn-\ning by doing. The storage cost depends on the storage type (aquifer, \nexhausted oil and gas field, for instance) and each storage type has a \nregional cumulative capacity limit. No leakage is considered from the \ngeological sites. See Supplementary Annex I for the implications of \nthis representation on the shape of the marginal cost curve for abate-\nment and removal.\nThe DAC module in our model is characterized by the following \nequations and parameters:\nCapital ENEG[GtCO2 per year] (since we assume full utilization, capac-\nity corresponds to the carbon removed each year) in the next period \nt\u2009+\u20091 depends on the residual stock of capital net of depreciation \u2202 times \nthe new capacity in the period expressed as the investments \nINET[T US$ per year] over the unitary investment cost:\nENEG (t + 1, n) = ENEG (t, n) \u00d7 (1 \u2212\u2202tstep)\n+\nINET(t,n)\ntotcost(t)\u00d7lifetime\u00d7fraccapex \u00d7 tstep\n(14)\nThe total cost per ton of carbon avoided, totcost, is time depend-\nent, because the technology is characterized by learning by doing and \ncosts decrease with the global cumulative capacity installed:\ntotcost (t) = max (totcostfloor, totcost0 \u00d7\n\u2211n,t\u2217={t0\u2236t}ENEG(t\u2217, n)\n\u2211nENEG(t0, n)\n\u2212learn\n)\n(15)\nwhere totcost0 = 453 US$ per tCO2 is the initial investment total unitary \ncost of DAC for carbon removed, totcostfloor = 100 US$ per tCO2 is the \nfloor cost, and the learning rate, learn, is set at 13.6% in the central \nspecification case and at 6% for the high cost DAC scenario. Therefore, \nthe average cost of the technology in the low cost scenarios starts from \n450 US$ per tCO2 in 2020 to 160, 110 and 100 US$ per tCO2 in 2050, 2075 \nand 2100, respectively. In the high cost scenarios, it starts from \n450 US$ per tCO2 in 2020 to 300, 250 and 235 US$ per tCO2 in 2050, 2075 \nand 2100, respectively.\nTo obtain the unitary investment cost to install a plant with removal \ncapacity of 1\u2009GtCO2 per year, the total cost, totcost, is multiplied by the \nlifetime of the plant (assumed to be 20 years) and by the fraction of \ntotal cost accruing to capital costs, which we assume to be 40% of \ntotal costs31. Under this cost structure, the investment costs are paid \nupfront and therefore tend to increase the average cost profile (relative \nto the numbers shown above) when the technology is ramping up and \ndecrease it to the variable costs (for DAC, mainly energy expenditures) \nwhen the technology is scaling down and no new investments are made.\nThe total costs for NETs COSTNET[\nT$\nyear ] that enter the income equa-\ntion (equation (3)) are as follows:\nCOSTNET = INET + ENEG (t, n) \u00d7 (1 \u2212fraccapex)\n\u00d7totcost(t) + \u2211storstorcost(stor) \u00d7 ESTOR(stor, t, n)\n(16)\nHere, the second term identifies operating expenditures for NETs \nand the last term represents storage costs. ESTOR is the yearly carbon \ndioxide stored underground in each type of storage type, stor (saline \naquifers, onshore and offshore exhausted oil and gas fields), each of \nwhich possesses a characteristic storage cost, storcost.\nThe yearly carbon dioxide stored is subject to the following condi-\ntions, which ensure that total yearly carbon store matches removals and \nthat the carbon stored in each storage type does not exceed the storage \ncapacity STORCAP, which depends on regional capacity:\nENEG (t, n) = \u2211\nstor\nESTOR(stor, t, n)\n\u2211\nt\u2217={t0\u2236t}\nESTOR (stor, t\u2217, n) < STORCAP (stor, n)\nTwo additional constraints are considered. First, to avoid exces-\nsive amount of removal in a model that does not include constraints, \nmaximum yearly global capacity of DAC is capped to 30\u2009GtCO2 per \nyear of removal. The way this constraint is distributed across regions \nvaries with the scenario analysed: in our central specification case, \neach regional constraint on yearly removal is set to mimic the storage \ncapacity, such that\nENEG (t, n) <\n\u2211storSTORCAP (stor, n)\n\u2211stor,nSTORCAP (stor, n) \u00d7 maxneg\nThe Global South scenario distributes the constraint according \nto 2020 population levels:\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\nENEG (t, n) <\npop(2020, n)\n\u2211npop(2020, n) \u00d7 maxneg\nSince our model does not explicitly consider the cost of energy and \nthe increase in investments in clean energy technologies necessary to \ncover the demand of large-scale DAC systems, we add an additional con-\nstraint on the global maximum yearly removal, maxneg, that depends \non the carbon budget (that is, the temperature target in 2100) with \na logistic function, such that the total amount of carbon captured is \nconsistent with projections from the ENGAGE scenarios of the WITCH \nmodel23, an energy-detailed integrated assessment model. The maxi-\nmum amount of global yearly NETs is therefore given by:\nmaxneg = 21.3 \u00d7 (1 + e0.00631\u00d7(carbonbudget\u22121069))\nFinally, NET deployment is subject to a market growth constraint \nto mimic the characteristic time of scale-up of the technology. This \nconstraint is written as\nINET (t + 1, n)\ntotcost (t + 1) \u00d7 lifetime \u00d7 fraccapex\n\u2264\nINET (t, n)\ntotcost (t) \u00d7 lifetime \u00d7 fraccapex\n\u00d7(1 + mktgwt)\ntstep + tstep \u00d7 mktgwt,0\n(17)\nwhere mktgwt = 6% per year is the maximum allowed market growth and \nmktgwt,0 is the initial NET capital in 2015.\nData availability\nAll scenario runs and code used to produce the figures and data \nanalysis are available in open source at https://doi.org/10.5281/\nzenodo.8397488.\nCode availability\nThe model used to produce the scenarios is available in open source \nat https://github.com/witch-team/RICE50xmodel. To run the model, \nan active GAMS licence with CONOPT3 is required.\nReferences\n25.\t Berger, L. & Emmerling, J. Welfare as equity equivalents. J. Econ. \nSurv. 34, 727\u2013752 (2020).\n26.\t Riahi, K. et al. The Shared Socioeconomic Pathways and their \nenergy, land use, and greenhouse gas emissions implications: an \noverview. Glob. Environ. Change 42, 153\u2013168 (2017).\n27.\t Gazzotti, P. RICE50+: DICE model at country and regional level. \nSocio-Environ. Syst. Model. 4, 18038 (2022).\n28.\t Rao, N. D., Sauer, P., Gidden, M. & Riahi, K. Income inequality \nprojections for the shared socioeconomic pathways (SSPs). \nFutures 105, 27\u201339 (2019).\n29.\t Emmerling, J. et al. Global inequality consequences of climate \npolicies when accounting for avoided climate impacts. Cell Rep. \nSustain. (in the press).\n30.\t Realmonte, G. et al. An inter-model assessment of the role of \ndirect air capture in deep mitigation pathways. Nat. Commun. 10, \n3277 (2019).\n31.\t Shayegh, S., Bosetti, V. & Tavoni, M. Future prospects of direct \nair capture technologies: insights from an expert elicitation \nsurvey. Front. Clim. https://doi.org/10.3389/fclim.2021.630893 \n(2021).\nAcknowledgements\nThe authors have received no external funding for this work.\nAuthor contributions\nAll the authors designed and conceptualized the research questions \nand scenarios. P.A. and J.E. performed the model advancements.  \nP.A. ran the scenarios, analysed the data and produced the figures. \nP.A. and M.T. wrote the first draft. All the authors extensively \ncontributed to the final version of the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at  \nhttps://doi.org/10.1038/s41558-023-01870-7.\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41558-023-01870-7.\nCorrespondence and requests for materials should be addressed to \nPietro Andreoni.\nPeer review information Nature Climate Change thanks Carl-Friedrich \nSchleussner and the other, anonymous, reviewer(s) for their \ncontribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01870-7\n1.78\n1.8\n1.79\n1.77\n1.75\n1.75\n1.71\n1.71\n1.72\n1.74\n1.73\n1.75\n1.72\n1.72\n1.72\n1.72\n1.73\n1.72\n1.72\n1.75\n1.77\n1.77\n1.4\n1.42\n1.41\n1.4\n1.4\n1.4\n1.41\n1.42\n1.39\n1.36\n1.37\n1.34\n1.37\n1.38\n1.37\n1.38\n1.4\n1.42\n1.43\n1.44\n1.38\n1.38\nN=35\nN=36\nN=35\nN=35\nN=35\nN=35\nN=36\nN=35\nN=36\nN=37\nN=37\nN=37\nN=37\nN=36\nN=37\nN=36\nN=37\nN=37\nN=37\nN=37\nN=37\nN=37\n1\n2\n2000\n2005\n2010\n2015\n2020\nElasticity\nExtended Data Fig. 1 | Calibration of income taxes and wealth elasticities. Income taxes elasticities (red) and wealth-to-income elasticity (blue) for each year \nbetween 2000 and 2020. Boxplots highlight median, 25th and 75th percentiles, points represent outliers. The number of observations and the median of the values is \nhighlighted in text for each year.\n\n\n    Scientific Research Findings:", "answer": "We find that privately owned NETs integrated into a carbon market designed for emission reductions increases economic inequality. Within each country, three factors drive the severity of the inequality increase: (1) the profit margin of NET companies; (2) the concentration of ownership of NET companies towards the top of the income distribution; and (3) the amount of negative emissions in the market. Therefore, small economies with high carbon-removal potential, concentrated equity ownership and expensive mitigation options are particularly susceptible to the inequality risk. In our analysis, we model a single NET (direct air capture) and global climate policy in line with the Paris Agreement. However, our results generalize to other negative emission options and to other climate policies (for example, net-zero pledges) whenever the above considerations apply and the technology value chain mostly benefits the rich.", "id": 45}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 13 | July 2023 | 679\u2013684\n679\nnature climate change\nhttps://doi.org/10.1038/s41558-023-01710-8\nArticle\nGlobal benefits of the international diffusion \nof carbon pricing policies\nManuel Linsenmeier\u2009\n\u200a\u20091,2\u2009\n, Adil Mohommad\u2009\n\u200a\u20093 & Gregor Schwerhoff\u2009\n\u200a\u20093\nCarbon pricing policies are essential for mitigating climate change, but the \nglobal benefits of leadership and the international diffusion of these policies \nare not well understood. Here we provide robust and statistically significant \nevidence showing that the adoption of carbon pricing in one country can \nexplain the subsequent adoption of carbon pricing in other countries. For a \nneighbouring country, diffusion increases the probability of policy adoption \non average by several percentage points. Translating these empirical \nestimates with Monte Carlo simulations into global reductions in emissions \nthrough policy diffusion suggests that for many countries, decreases in \nemissions as a result of diffusion could be larger than domestic emission \nreductions. These results support the adoption of stringent climate  \npolicies, especially in countries in which climate change mitigation  \nmight be considered as not very important because of relatively low levels  \nof domestic emissions.\nDespite the need for more stringent climate policies to achieve the Paris \nclimate targets, many countries seem reluctant to ratchet up their miti-\ngation efforts. This might be partly because the costs of mitigation are \nincurred domestically and immediately, whereas most of its benefits will \nbe reaped globally and in the future. In addition, more ambitious climate \naction might be hindered by concerns about the limited effectiveness \nof nations\u2019 domestic efforts to reduce greenhouse gas (GHG) emissions \nif other countries do not do make similar changes. This consideration \nis especially pertinent in relatively small countries. Indeed, in 2021 the \nsmallest 90% of emitters contributed only about 20% of global GHG \nemissions. Such a narrow perspective, however, neglects the fact that \ninternational leadership in climate change mitigation can yield sub-\nstantial benefits beyond domestic emission reductions1,2. For exam-\nple, stringent climate policies at home can support the international \ndiffusion of technological innovations that reduce mitigation costs in \nother countries3,4. Furthermore, domestic climate policies can show \nthe political feasibility and certain benefits of carbon pricing5, and they \ncan create incentives related to trade6 and diplomacy7 that can nudge \nother countries to adopt the same or similar policies. This latter process \nwhereby adoption of a policy in one country increases the probability \nof adoption in other countries is usually referred to as policy diffusion8.\nResults from qualitative studies provide ample evidence for climate  \npolicy diffusion. For example, evidence has been described for strong \nmutual influences among the world\u2019s first adopters of carbon pricing \npolicies in Scandinavia in the 1980s (ref. 9). According to one study10, \nthe subsequent adoption of carbon pricing by other countries can \nbe at least partially explained by the emulation of earlier policies and \nlearning from previous experiences. International diffusion has also \nbeen actively promoted by early adopters themselves and through \nmultilateral initiatives such as the World Bank\u2019s Partnership for Market \nReadiness (PMR)11. Furthermore, several case studies of carbon pric-\ning policies report empirical evidence for international diffusion\u2014for \nexample, for California12, Kazakhstan13 and China14\u2014and the influence \nof multilateral initiatives has been acknowledged for carbon pricing \npolicies in Latin America15. Earlier work also examined the diffusion of \nsupport for carbon pricing at the subnational level and between firms \nor organizations16,17. Several quantitative studies also provide evidence \nin support of an international diffusion of climate policies6,18\u201322.\nIn this study, we empirically examine the international diffusion of \nclimate policies from 1988 to 2021 and quantify its global benefits. The \nanalysis focuses on carbon pricing policies, which can be considered \nthe most salient and possibly the most stringent policies for climate \nchange mitigation. We first construct a global dataset of carbon pric-\ning policies, countries\u2019 characteristics and linkages between countries \nrelated to geography, trade and international organizations. We then \nestimate Cox proportional hazard models that include spatial lags \nReceived: 3 November 2022\nAccepted: 22 May 2023\nPublished online: 3 July 2023\n Check for updates\n1Climate School, Columbia University, New York, NY, USA. 2Department of Geography and Environment, London School of Economics and Political \nScience, London, UK. 3Research Department, International Monetary Fund, Washington, DC, USA. \n\u2009e-mail: mpl2157@columbia.edu\n\nNature Climate Change | Volume 13 | July 2023 | 679\u2013684\n680\nArticle\nhttps://doi.org/10.1038/s41558-023-01710-8\nMexico were followed by other pricing policies in the same region. For \nLatin America, the role of international organizations has been empha-\nsized15. In Asia, early carbon pricing policies in Japan were followed by \npricing policies in China and South Korea.\nMotivated by these findings, we next conduct an econometric \nanalysis to more systematically identify whether the adoption of car-\nbon pricing in one country affected the probability of its subsequent \nadoption elsewhere. To do so, we estimate a Cox proportional hazard \nmodel (equation (1) in Methods) with several country characteristics \nas explanatory variables. We use Lasso regressions and a detailed \nexamination of multicollinearity for the selection of variables (Methods  \nand Supplementary Tables 4\u20136). On the basis of a statistical test  \nusing Schoenfeld residuals23 we cannot reject the null hypothesis \nof proportional hazards for models with the preferred five or more \ncontrol variables. To model international diffusion, we also include \na spatial lag of prior carbon pricing adoption in other countries. For \nthis variable, we use several alternative metrics of the proximity of \ncountries (Methods). We find the best model fits for a metric that \ncombines the gross domestic product (GDP) of countries with the \ngeographical distances between them in the spirit of gravity models \nof international trade, and for a metric based on joint membership in \ninternational organizations (Supplementary Table 7). We then multiply \nof policy adoption (Methods). The spatial lags are constructed using \nalternative metrics of the proximity of countries. Possible concerns \nabout causality are addressed with a series of robustness tests and a \nplacebo test. In the last part, we use our empirical estimates to calculate \nthe expected emission reductions due to policy diffusion using Monte \nCarlo simulations. We consider these indirect reductions in emissions \nas a proxy for the international leverage of a country\u2019s domestic climate \npolicy and examine its variation across countries. We also use these \nsimulations to quantify the global coverage of carbon pricing policies \nthat can be expected because of policy diffusion.\nResults\nWe first examine our dataset of carbon pricing policies from 1988 to 2021 \nto identify possible patterns of policy diffusion. Visual inspection of the \nrelative timing of policy adoption shows that carbon pricing was often \nintroduced successively by geographically close countries (Fig. 1a).  \nIn Europe, for example, the earliest carbon pricing policy in Finland \nwas followed by similar policies in Scandinavian countries, the Baltics \nand other parts of Europe. Qualitative work on the role of diffusion in \nthis context highlighted the importance of the pioneering adoption in \nFinland, which was soon \u201cemulated by its Nordic neighbors\u201d9. Similarly, \nin the Americas, relatively early carbon pricing policies in Canada and \n1990\n1995\n2000\n2005\n2010\n2015\n2020\nYear\n0\n5%\n10%\n15%\n20%\nShare of countries with a carbon pricing policy\nTax\nETS\n2015\n2005\n2015\n2013\n2011\n1991\n2019\n2012\n2015\n2013\n2012\n2008\n2018\n2017\n2017\n2014\n2019\n1990\na\nb\nc\n...\nAFG\nPAK\nTJK\nTKM\nUZB\nIRN\nKGZ\nNPL\nAZE\nBTN\nARE\nQAT\nBGD\nKWT\nMDV\nOMN\nIND\nARM\nBHR\nIRQ\nSYR\nMMR\nSAU\nYEM\nJOR\nLBN\nLKA\nTUR\nISR\nERI\nTHA\nDJI\nEGY\nSOM\nSDN\nCYP\nLBY\nETH\nGRC\nCHN\nMYS\nIDN\nCAF\nCOD\nUGA\nSSD\nKEN\n\u2013100\n\u201360\n\u201340\n\u201320\n0\n20\n40\n60\n\u201350\n0\n50\nLongitude (degrees)\nYear of adoption\nLatitude (degrees)\n100\n150\n1988\n1990\n1992\n1994\n1996\n1998\n2000\n2002\n2004\n2006\n2008\n2010\n2012\n2014\n2016\n2018\n2020\n2022\nFig. 1 | Adoption of carbon pricing policies over time and around the world. \na, Map showing the adoption of the first carbon pricing policy for every country. \nHashes indicate countries in which the first policy was a subnational policy. See \nSupplementary Fig. 2 for a more detailed map of Europe (blue box). b, Adoption \nof carbon taxes and ETSs over time. c, Proximity of countries on the basis of \nphysical distance, GDP and joint membership in international organizations. \nArrows indicate the three strongest influences on every country; position in chart \napproximates average distances. Countries are indicated by ISO three-letter \ncountry codes. See Supplementary Fig. 3 for the complete version of c. Base map \nadapted from World Bank Official Boundaries under a Creative Commons licence \nCC BY 4.0.\n\nNature Climate Change | Volume 13 | July 2023 | 679\u2013684\n681\nArticle\nhttps://doi.org/10.1038/s41558-023-01710-8\nthese two metrics to create a hybrid metric for the empirical analysis \nand simulations. Furthermore, in our main specification we focus on \nthe first policy in every country and we consider all members of the \nEuropean Union (EU) emissions trading system (ETS) without a prior \ncarbon tax together as one country that adopted its first carbon pricing \npolicy in 2005 (Methods and Supplementary Fig. 2).\nOur empirical analysis yields robust statistical evidence for an \ninternational diffusion of carbon pricing policies (Table 1, column \n1). The magnitude of the estimated coefficient of policy diffusion is \nsubstantial. For example, according to our main estimates (Table 1, \ncolumn 1), prior adoption of carbon pricing by Canada increases the \nprobability of adoption in the USA by a factor of about 1.78, or by 78% \n(95% confidence interval (CI) of 35\u2013134%). In Germany, prior adoption \nby France increases the probability by 20% (10\u201331%), whereas in China, \nprior adoption by Japan increases it by 25% (12\u201339%). For comparison, \nin the USA, prior adoption by China increases the probability of adop-\ntion by 8.5% and in Germany, prior adoption by Japan by about 1.9%.\nIn our main specification, we consider carbon taxes and ETSs as \ntwo alternative designs of the same policy. This is informed by previous \nfindings that there are no systematic differences between countries \nthat chose either of the two designs24. Furthermore, we consider it \nlikely that in many cases the decision to adopt carbon pricing was made \nbefore the choice of instrument design, as in the case of the EU ETS9. \nConsistent with this idea, we find stronger evidence for policy diffusion \nif we consider ETSs and taxes as the same policy than if we distinguish \nbetween them (Table 1). In additional analysis, we find suggestive \nevidence that carbon pricing policies with a higher stringency (higher \neconomy-wide average carbon prices, taking into account sectoral \nprices and emissions) exert a stronger influence on subsequent adop-\ntion elsewhere (Supplementary Table 8, column 5).\nWe conduct several robustness checks (Methods). This includes \nthree ways of treating members of the EU ETS (Supplementary Fig. 2  \nand Supplementary Table 8, columns 1\u20133), dropping subnational \npolicies (Supplementary Table 8, column 4), adding control variables \n(Supplementary Table 9, columns 1 and 2), changing the imputation \nmethod (Supplementary Table 9, column 3) and stratifying the model \n(Supplementary Table 9, column 4). We find that our results are overall \nvery robust. An additional placebo test does not show evidence of spuri-\nous diffusion25. Furthermore, we find the best model fit for a lag time of \n1\u20132 years (Supplementary Table 11). Additional evidence suggests that \nthe marginal effect of a new policy decreased with the total number of \nexisting policies (Supplementary Table 10 and Supplementary Fig. 5a). \nWe use this insight on \u2018saturation\u2019 as motivation to estimate a non-linear \nmodel that we then use for all simulations (Supplementary Fig. 5b).\nOverall, the results of the empirical analysis suggest that between \n1988 and 2021, carbon pricing policies diffused internationally. We \nnext examine how much this diffusion could contribute to reduc-\ntions of GHG emissions globally. Specifically, we use our estimated \nmodel to quantify the emission reductions that can be attributed to the \nadoption of carbon pricing in a given country, distinguishing between \ndirect (domestic) emissions reduction and indirect (foreign) emission \nreductions due to diffusion. All results are based on the empirical \nestimates from the econometric analysis, but we assume a constant \nbaseline hazard, which means that all differences in the probability \nof policy adoption between countries can be attributed to the spatial \nlag and country characteristics. Given the probabilistic nature of our \nmodel, we conduct Monte Carlo simulations. All simulations start in \n2022 from the carbon pricing policies adopted by the end of 2021. For \nevery country without a carbon price, we conduct 30,000 simulations \nin which this country adopts carbon pricing in 2022. We then compare \nthe results with results from counterfactual simulations in which this \ncountry does not adopt carbon pricing in 2022. This comparison allows \nus to attribute policy adoption in other countries to the diffusion of \none specific policy.\nWe find that indirect emission reductions are as large as or even \nlarger than direct emission reductions in most countries. Specifically, \nfrom 2022\u20132050, about 70% of countries (97 out of 138) have larger \nindirect than direct cumulative emission reductions (Fig. 2a). Further-\nmore, we find that indirect emission reductions are far more equally \ndistributed across countries than direct emission reductions (Fig. 2b). \nThis result also suggests that the total reductions in emissions from \npolicy adoption and diffusion are more equally distributed than only \ndirect domestic emission reductions.\nFor simplicity, we assume that carbon pricing policies reduce  \nGHG emissions by the same rate r\u2009=\u20091% per year in all countries relative \nto a situation without a carbon pricing policy. This rate is conservative \ncompared to known reductions in emissions in existing ETSs and esti-\nmated reductions for carbon taxes (Methods). In a sensitivity analysis, \nwe vary the value of this parameter between 0.1% and 10% and find that \nthis changes the number of countries with larger indirect than direct \nemission reductions by only a few percentage points (Supplementary \nFig. 7). We do not find evidence that later adopters tended to adopt \nsystematically more or less stringent polices than earlier adopters (Sup-\nplementary Fig. 6). Furthermore, we find similar results for alternative \nproximity metrics (Supplementary Fig. 8).\nTable 1 | Results of empirical analysis of policy adoption \n1988\u20132021 with Cox proportional hazard models\nPolicy\nCarbon price\nTax\nETS\nColumn\n1\n2\n3\nSpatial lag of carbon pricing\n0.9152***\n0.3423\n0.8619***\n(\u22120.2251)\n(\u22120.2885)\n(\u22120.1816)\nlog real GDP per capita PPP\n0.4782\n0.5221\n\u22120.4602\n(\u22120.4226)\n(\u22120.388)\n(\u22121.0017)\nGovernment effectiveness\n\u22120.2327\n\u22120.0758\n0.5461\n(\u22120.4439)\n(\u22120.3869)\n(\u22121.177)\nRegulatory quality\n1.4777***\n0.6952*\n1.7684**\n(\u22120.4559)\n(\u22120.3969)\n(\u22120.8692)\nReserves of oil\n0.0093\n0.0192\n\u22120.0492\n(\u22120.1408)\n(\u22120.1917)\n(\u22120.3266)\nGovernment expenditure\n0.3199\n0.0399\n0.014\n(\u22120.2221)\n(\u22120.441)\n(\u22120.5959)\nGov. expenditure for welfare\n0.1914\n0.4893\n0.9689**\n(\u22120.2184)\n(\u22120.373)\n(\u22120.4885)\nDemocracy index\n0.6016\n0.7486\n\u22120.3912\n(\u22120.5236)\n(\u22120.645)\n(\u22120.8534)\nEmission intensity\n0.3478**\n0.3392***\n0.1653\n(\u22120.1393)\n(\u22120.1189)\n(\u22120.2152)\nGrowth rate of debt to GDP ratio\n0.2975*\n0.4281***\n0.3717*\n(\u22120.1785)\n(\u22120.1558)\n(\u22120.2175)\nAIC\n173.4\n219.3\n81.8\nlog-likelihood\n\u221276.7\n\u221299.7\n\u221230.9\nNumber of observations\n5,322\n6,061\n5,295\nCountries\n167\n188\n159\nPolicies\n25\n26\n11\nColumn 1 shows the main specification. Columns 2 and 3 show results for only carbon taxes \nand only ETSs, respectively. Results are based on the gravity-IO proximity metric. Results for \nother metrics are shown in Supplementary Table 7. Results excluding subnational policies \nand for different ways of dealing with Europe are shown in Supplementary Table 8. See also \nadditional robustness tests in Supplementary Table 9. AIC, Akaike information criterion; \nPPP, purchasing power parity. Standard errors clustered by country are in parentheses. \nSignificance is based on a two-sided t-test: *P\u2009<\u20090.1, **P\u2009<\u20090.05, ***P\u2009<\u20090.01.\n\nNature Climate Change | Volume 13 | July 2023 | 679\u2013684\n682\nArticle\nhttps://doi.org/10.1038/s41558-023-01710-8\nWe find that indirect emission reductions are largest throughout \nthe Middle East and in South Asia and Southeast Asia (Fig. 2a,c). The \nproximity metrics suggest that countries with large indirect emission \nreductions tend to be relatively centrally located, are members of  \nsimilar international organizations to large emitters and are  \nalso located physically close to countries with relatively large emis-\nsions and no carbon pricing policies as of the end of 2021. To further  \ninvestigate these determinants of indirect emission reductions, we \ncalculate for every country its \u2018network centrality\u2019 on the basis of the \nproximities and GHG emissions of all countries (equations (9\u201311)). \nWe find that network centrality can explain about 13% of the variation \nin indirect emission reductions across countries, which increases to \n43% if we take into account the emissions of other countries, and to \n58% if we better account for the \u2018cascading nature\u2019 of policy diffusion \n(equation (11)).\nIn the last part of the analysis, we examine how diffusion affects \nthe future geographical coverage of carbon pricing policies. To \nthis aim, we conduct similar Monte Carlo simulations starting in  \n2022 and continuing until 2100. Both scenarios start from exist-\ning policies. In the first scenario, we use our empirically estimated  \nrelationship between the spatial lag of carbon pricing and the prob-\nability of policy adoption. In the second scenario, we set international \npolicy diffusion to zero. All other parameter values are chosen as in the \nprevious exercise.\nWe find that policy diffusion substantially increases the geographical  \ncoverage of carbon pricing over the time period 2022\u20132050 (Fig. 3).  \nIn our simulations, by 2050 carbon pricing policies will be in place in \nabout 50% of countries, 21 percentage points more than without diffu-\nsion. By 2100, the difference increases to more than 30 percentage points \n(Fig. 3a). In a sensitivity analysis, we multiply the baseline hazard and the \ndiffusion term in the model with factors between 0.5 and 2. Both para\u00ad\nmeters have a positive effect on the number of countries with a carbon  \nprice (Fig. 3b). Similarly, the effect of diffusion increases with either of \nthe two parameters. For example, as the factor of the baseline hazard is \nincreased from 1 to 2, the additional global coverage of carbon pricing \ndue to diffusion by 2050 increases from 21 to 32 percentage points.\nThese results add a key detail to the global benefits of interna-\ntional policy diffusion. From the perspective of an individual country, \npolicy diffusion can add substantial global emission reductions to \ndomestic emission reductions. From a global perspective, however, \na high baseline probability of policy adoption and/or a strength of \nmutual influences (both in historical perspective) are required to reach, \nfor example, 80% coverage by 2050. Because most of the countries \nthat adopt carbon pricing late in our simulations have relatively low \ndomestic emissions, the projected global coverage of pricing policies \nis generally higher in terms of global GHG emissions than in terms \nof countries, and the results are less sensitive to the two parameters \n(Supplementary Fig. 11).\na\nb\nc\n10\n\u20133\n10\n\u20132\n10\n\u20131\n10\n0\n10\n1\n10\n\u20133\n10\n\u20134\n10\n\u20135\n10\n\u20132\n10\n\u20131\n10\n0\n10\n1\nDirect emission reductions (Gt CO2eq)\n10\n\u20132\n10\n\u20131\nIndirect emission reductions (Gt CO2eq)\nEmission reductions (Gt CO2eq)\n0\n10\n20\n30\n40\nNumber of countries\nDirect\nIndirect\n\u2013100\n\u201350\n0\n50\nLongitude (degrees)\n100\n150\n\u201360\n\u201340\n\u201320\n0\n20\n40\n60\nLatitude (degrees)\nIndirect emission reductions (Gt)\n0.25\n0.20\n0.15\n0.10\n0.05\n0\nFig. 2 | Direct and indirect emission reductions from carbon pricing policies \non the basis of Monte Carlo simulations of future policy adoption. All panels \nshow cumulative emission reductions from simulated policy adoption in \n2022\u20132050. a, Scatter plot of direct and indirect emission reductions. Dashed \nblack line indicates where indirect emission reductions are larger than direct \nemission reductions. G20 economies are shown in blue. Countries with a carbon \nprice by the end of 2021 are omitted. b, Histogram of direct and indirect emission \nreductions. c. Map of indirect emission reductions. Countries with a carbon \npricing policy by the end of 2021 are shown in dark grey. CO2eq denotes CO2 \nequivalent. Base map adapted from World Bank Official Boundaries under a \nCreative Commons license CC BY 4.0.\n\nNature Climate Change | Volume 13 | July 2023 | 679\u2013684\n683\nArticle\nhttps://doi.org/10.1038/s41558-023-01710-8\nDiscussion\nThe main contribution of this paper is the quantification of reductions \nin GHG emissions that can be attributed to the international diffusion \nof carbon pricing policies. These indirect emission reductions can be \ninterpreted as a quantitative measure of the international leverage of \na country in terms of global GHG emission reductions due to future \ndiffusion of its policy. Overall, our results suggest that the magnitude \nof indirect emission reductions can be substantial. With our empiri-\ncally estimated parameters, future indirect emission reductions will \nbe larger than domestic emission reductions in 63% of countries that \ndid not have a carbon pricing policy in place by the end of 2021. This \nevidence for large positive spillovers of domestic climate policy adop-\ntion provides additional support for the adoption of stringent climate \npolicies, especially in countries in which climate policies might so far \nhave been deemed to have relatively little importance because of a \nrelatively small domestic economy.\nOur results speak to a long-standing debate about free-riding in \nglobal environmental policy26. According to this theory, unilateral \nclimate policy weakens the incentives of other countries to implement \nclimate policy themselves. In this paper, we show that free-riding is \nnot the only possible reaction of countries to unilateral climate policy  \nin the case of carbon pricing policies. Instead, we find that these policies \ndiffuse internationally. We consider this as evidence consistent with the \nidea that leadership in climate policy can send a credible signal about \nthe willingness to cooperate to other countries, supporting the forma-\ntion of global climate coalitions that have been proposed27,28. Similarly, \nex-ante modelling studies have repeatedly predicted carbon leakage29 \nas a consequence of leadership. Empirical studies, however, find that \nleakage is a very minor or no concern30\u201332. Similar to other studies that \nexamine the evidence for free-riding in climate policy33, our results \nprovide a possible (partial) explanation for this lack of leakage.\nWe find large indirect emission reductions especially for countries \nin close proximity to large emitters without a carbon pricing policy. \nThis includes several countries on the Arabian peninsula and in South \nand Southeast Asia. Our analysis of network centrality as a determi-\nnant of countries\u2019 international leverage suggests that the \u2018cascading\u2019 \nnature of policy diffusion is important to explain some of these results. \nNotably, we find that countries have become more similar over the last \n30 years in terms of their international leverage, suggesting that the \nbenefits from policy diffusion have also become globally more equally \ndistributed (Supplementary Fig. 10).\nThe results for emission reductions from the simulations for a specific \ncountry should, however, not be considered as precise estimates because \nof some necessarily neglected heterogeneity. Specifically, some of the \nempirically estimated parameters are likely to differ between countries \nwith large uncertainty, including the effectiveness of future carbon pricing \npolicies. To assess the robustness of our results, we conduct a sensitivity \nanalysis in which we change this and other parameters. Furthermore, \nowing to the construction of the scenarios, indirect emission reductions \nsimulated for a specific pioneering country are not additive with those \nsimulated for another pioneering country. This means that our estimates \nfor individual countries can in some sense be considered as upper bounds.\nTheories of policy diffusion propose several mechanisms through \nwhich the adoption of a policy in one jurisdiction can influence the \nadoption of the same or a similar policy elsewhere. These mechanisms \nare often referred to as learning, competition, emulation and coer-\ncion25,34\u201339. Previous literature on climate policies has especially focused \non emulation and learning10,11. Somewhat consistent with this, we do \nnot find evidence that shared export markets have been important for \ninternational diffusion, which suggests that the role of international \ncompetition is limited. Furthermore, our results suggest that interna-\ntional organizations contributed to policy diffusion, owing possibly to \ninternational coordination and exchanges of information\u2014consistent \nwith the role of emulation and learning\u2014and possibly to compliance \nwith international norms10. We consider this to be an encouraging \nfinding for present and future attempts to increase the geographical \ncoverage of carbon pricing policies through international relations.\nPrevious qualitative research also suggests that certain design \nattributes of carbon pricing policies have diffused because of emula-\ntion13,40. Future research might examine the overall relevance of the \ndiffusion of policy design using similar quantitative methods. For \nexample, we consider it plausible that international diffusion also \nmatters for ratcheting up the stringency of existing climate policies; \nfor example, increases in carbon prices.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \n0.50\n0.75\n1.00\n1.25\n1.50\n1.75\n2.00\nBaseline hazard (relative to estimate)\n0.50\n0.75\n1.00\n1.25\n1.50\n1.75\n2.00\nDifusion parameter (relative to estimate)\n31\n(+4)\n35\n(+8)\n39\n(+12)\n43\n(+16)\n46\n(+19)\n49\n(+22)\n52\n(+25)\n34\n(+6)\n40\n(+12)\n45\n(+17)\n49\n(+22)\n54\n(+26)\n57\n(+29)\n37\n(+8)\n44\n(+15)\n50\n(+21)\n55\n(+26)\n60\n(+31)\n39\n(+10)\n48\n(+18)\n55\n(+25)\n42\n(+11)\n51\n(+21)\n58\n(+28)\n44\n(+13)\n54\n(+23)\n47\n(+14)\n57\n(+25)\n60\n(+32)\n63\n(+34)\n66\n(+37)\n60\n(+30)\n64\n(+34)\n68\n(+38)\n70\n(+41)\n64\n(+33)\n68\n(+37)\n71\n(+41)\n74\n(+43)\n62\n(+30)\n67\n(+35)\n71\n(+39)\n74\n(+43)\n77\n(+45)\n65\n(+32)\n70\n(+37)\n73\n(+41)\n77\n(+44)\n79\n(+47)\na\nb\n+21%\nGlobal coverage of carbon pricing policies (% countries)\nwith difusion\nversus without difusion\n2000\n2020\n2040\n2060\n2080\n2100\nYear\n0%\n10%\n20%\n30%\n40%\n50%\n60%\n70%\nShare of countries with carbon price\n1989\u20132021\n2022\u20132100, with difusion\n2022\u20132100, without difusion\nFig. 3 | Future global coverage of carbon pricing policies from Monte Carlo \nsimulations of future policy adoption. All panels show the future number of \ncountries with carbon pricing policies as a share of all countries for simulations \nstarting in 2022 from existing policies by end of 2021. a, Time series of future \npolicy adoption for parameter values set to empirical estimates, for simulations \nwith and without diffusion. b, Share of countries with carbon pricing policies \nby 2050 for different parameter values of the baseline hazard and the diffusion \ncoefficient. Red square indicates empirical estimates. See Supplementary Fig. 11 \nfor results in global GHG emissions.\n\nNature Climate Change | Volume 13 | July 2023 | 679\u2013684\n684\nArticle\nhttps://doi.org/10.1038/s41558-023-01710-8\nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-023-01710-8.\nReferences\n1.\t\nSchwerhoff, G. The economics of leadership in climate change \nmitigation. Clim. Policy 16, 196\u2013214 (2016).\n2.\t\nH\u00f6hne, N., Fekete, H., den Elzen, M. G., Hof, A. F. & Kuramochi, \nT. Assessing the ambition of post-2020 climate targets: a \ncomprehensive framework. Clim. Policy 18, 425\u2013441 (2018).\n3.\t\nDechezlepr\u00eatre, A., Glachant, M., Ha\u0161c\u01d0, C. I., Johnstone, N. & \nM\u00e9ni\u00e8re, Y. Invention and transfer of climate change\u2013mitigation \ntechnologies: a global analysis. Rev. Environ. Econ. Policy 5, \n109\u2013130 (2011).\n4.\t\nBarrett, P. Can International Technological Diffusion Substitute  \nfor Coordinated Global Policies to Mitigate Climate Change?  \nIMF Working Paper https://doi.org/10.5089/9781513585765.001  \n(IMF, 2021).\n5.\t\nGoulder, L. H. Climate change policy\u2019s interactions with the tax \nsystem. Energy Econ. 40, S3\u2013S11 (2013).\n6.\t\nSteinebach, Y., Fern\u00e1ndez-i-Mar\u00edn, X. & Aschenbrenner, C. Who \nputs a price on carbon, why and how? A global empirical analysis \nof carbon pricing policies. Clim. Policy 21, 277\u2013289 (2021).\n7.\t\nKammerer, M. & Namhata, C. What drives the adoption of climate \nchange mitigation policy? A dynamic network approach to policy \ndiffusion. Policy Sci. 51, 477\u2013513 (2018).\n8.\t\nWalker, J. L. The diffusion of innovations among the American \nStates. Am. Polit. Sci. Rev. 63, 880\u2013899 (1969).\n9.\t\nHarrison, K. The comparative politics of carbon taxation.  \nAnnu. Rev. Law Soc. Sci. 6, 507\u2013529 (2010).\n10.\t Thisted, E. V. & Thisted, R. V. The diffusion of carbon taxes and \nemission trading schemes: the emerging norm of carbon pricing. \nEnviron. Polit. 29, 804\u2013824 (2020).\n11.\t\nBiedenkopf, K., M\u00fcller, P., Slominski, P. & Wettestad, J. A global \nturn to greenhouse gas emissions trading? Experiments, actors, \nand diffusion. Glob. Environ. Polit. 17, 1\u201311 (2017).\n12.\t Bang, G., Victor, D. G. & Andresen, S. California\u2019s cap-and-trade \nsystem: diffusion and lessons. Glob. Environ. Polit. 17, 12\u201330 (2017).\n13.\t Gulbrandsen, L. H., Sammut, F. & Wettestad, J. Emissions trading \nand policy diffusion: complex EU ETS emulation in Kazakhstan. \nGlob. Environ. Polit. 17, 115\u2013133 (2017).\n14.\t Heggelund, G., Stensdal, I., Duan, M. & Wettestad, J. China\u2019s \ndevelop\u00adment of ETS as a GHG mitigating policy tool: a case of policy \ndiffusion or domestic drivers? Rev. Policy Res. 36, 168\u2013194 (2019).\n15.\t Ryan, D. & Micozzi, M. The politics of climate policy innovation: the \ncase of the Argentine carbon tax. Environ. Polit. 30, 1155\u20131173 (2021).\n16.\t Meckling, J. Carbon Coalitions: Business, Climate Politics, and the \nRise of Emissions Trading (The MIT Press, 2011).\n17.\t Paterson, M., Hoffmann, M., Betsill, M. & Bernstein, S. The \nmicro foundations of policy diffusion toward complex global \ngovernance: an analysis of the transnational carbon emission \ntrading network. Comp. Polit. Stud. 47, 420\u2013449 (2014).\n18.\t Baldwin, E., Carley, S. & Nicholson-Crotty, S. Why do countries \nemulate each others\u2019 policies? A global study of renewable \nenergy policy diffusion. World Dev. 120, 29\u201345 (2019).\n19.\t Fankhauser, S., Gennaioli, C. & Collins, M. Do international factors \ninfluence the passage of climate change legislation? Clim. Policy \n16, 318\u2013331 (2016).\n20.\t Sauquet, A. Exploring the nature of inter-country interactions in \nthe process of ratifying international environmental agreements: \nthe case of the Kyoto Protocol. Public Choice 159, 141\u2013158 (2014).\n21.\t Dolphin, G. G. & Pollitt, M. G. The International Diffusion of Climate \nPolicy: Theory and Evidence. RFF Working Paper https://www.rff. \norg/publications/working-papers/the-international-diffusion- \nof-climate-policy-theory-and-evidence (2021).\n22.\t Raghoo, P. & Shah, K. U. A global empirical analysis on the \ndiffusion & innovation of carbon pricing policies. J. Clean. Prod. \n362, 132329 (2022).\n23.\t Grambsch, P. M. & Therneau, T. M. Proportional hazards tests and \ndiagnostics based on weighted residuals. Biometrika 81, 515\u2013526 \n(1994).\n24.\t Skovgaard, J., Ferrari, S. S. & Knagg\u00e5rd, \u00c5. Mapping and clustering \nthe adoption of carbon pricing policies: what polities price \ncarbon and why? Clim. Policy 19, 1173\u20131185 (2019).\n25.\t Braun, D. & Gilardi, F. Taking \u2018Galton\u2019s problem\u2019 seriously: towards \na theory of policy diffusion. J. Theor. Polit. 18, 298\u2013322 (2006).\n26.\t Hoel, M. Global environmental problems: Tthe effects of \nunilateral actions taken by one country. J. Environ. Econ. Manag. \n20, 55\u201370 (1991).\n27.\t Nordhaus, W. & Chen, X. A sharper image? Estimates of the \nprecision of nighttime lights as a proxy for economic statistics.  \nJ. Econ. Geogr. 15, 217\u2013246 (2015).\n28.\t Parry, I. W. H., Black, S. & Roaf, J. Proposal for an International \nCarbon Price Floor among Large Emitters. IMF Staff Climate Note \nNo. 2021/001 (IMF, 2021).\n29.\t B\u00f6hringer, C., Fischer, C., Rosendahl, K. E. & Rutherford, T. F. \nPotential impacts and challenges of border carbon adjustments. \nNat. Clim. Change 12, 22\u201329 (2022).\n30.\t Naegele, H. & Zaklan, A. Does the EU ETS cause carbon leakage \nin European manufacturing? J. Environ. Econ. Manag. 93, 125\u2013147 \n(2019).\n31.\t Dechezlepr\u00eatre, A., Gennaioli, C., Martin, R., Mu\u00fbls, M. & Stoerk, T. \nSearching for carbon leaks in multinational companies. J. Environ. \nEcon. Manag. 112, 102601 (2022).\n32.\t Eskander, S. M. S. U. & Fankhauser, S. The impact of climate \nlegislation on trade-related carbon emissions 1996\u20132018. Environ. \nResource Econ. 85, 167\u2013194 (2023).\n33.\t Schmidt, R. C., Drupp, M., Nesje, F. & Hoegen, H. Testing the free- \nrider hypothesis in climate policy. Preprint at https://doi.org/ \n10.48550/arXiv.2211.06209 (2022).\n34.\t Simmons, B. A., Dobbin, F. & Garrett, G. Introduction: the \ninternational diffusion of Liberalism. Int. Organ. 60, 781\u2013810 \n(2006).\n35.\t Shipan, C. R. & Volden, C. The mechanisms of policy diffusion. \nAm. J. Polit. Sci. 52, 840\u2013857 (2008).\n36.\t Volden, C., Ting, M. M. & Carpenter, D. P. A formal model of \nlearning and policy diffusion. Am. Polit. Sci. Rev. 102, 319\u2013332 \n(2008).\n37.\t Shipan, C. R. & Volden, C. Policy diffusion: seven lessons for \nscholars and practitioners. Public Adm. Rev. 72, 788\u2013796 (2012).\n38.\t Sugiyama, N. B. The diffusion of conditional cash transfer \nprograms in the Americas. Glob. Soc. Policy 11, 250\u2013278 (2011).\n39.\t Jordan, A. & Huitema, D. Innovations in climate policy: the politics \nof invention, diffusion, and evaluation. Environ. Polit. 23, 715\u2013734 \n(2014).\n40.\t Wettestad, J. & Gulbrandsen, L. H. The Evolution of Carbon \nMarkets: Design and Diffusion (Transforming Environmental Politics \nand Policy) (Routledge, 2019).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2023\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01710-8\nMethods\nEmpirical analysis of policy diffusion\nWe use econometric models to identify policy diffusion in the data \non past policy adoption. To do so, we estimate a model that relates \nadoption of a policy in a country i at time t to the adoption of the \nsame policy in other countries j\u2009=\u20091, \u2026, Nc, j !\u2009=\u2009i prior to time t (with Nc  \nbeing the number of countries in the sample). This is a common  \nempirical strategy to identify policy diffusion and has been used in \nthe literature on climate policy7,20,21,41. Technically, the model accounts \nfor the mutual influences between countries with spatial lags, which  \nare calculated as a weighted average of prior policy adoption in all  \nother countries. We use alternative weighting schemes based on geo-\ngraphical proximity, trade and international institutional linkages, \nwhich we consider as representing some of the alternative diffusion \nmechanisms cited in the main text.\nThe choice of our model is informed by two characteristics of \nour dependent variable. The first characteristic is that any possible \nfuture policy adoption is unobserved, which means that our dependent \nvariable is right-censored. Specifically, at the time of analysis, policy \nadoption is only recorded in the Carbon Pricing Dashboard of the World \nBank up until April 2022, which means that 2021 is the most recent year \nin our sample. The second characteristic is that our dependent variable \nis binary, taking on only values 0 or 1. Both these characteristics are \ncommon in survival analysis, which is also referred to as event history \nanalysis, and can be addressed with proportional hazard models.\nWe thus follow previous work on policy diffusion and model policy \ndiffusion with semi-parametric Cox proportional hazard models20,21,38,41. \nAs compared to parametric proportional hazard models, the Cox model \ndoes not require an assumption about a specific functional form of the \nsurvival function, and the results can therefore be considered more \nrobust to model mis-specification42. Formally, we estimate models of \nthe general form\nh (t, Xi,t, Wi,t) = h0 (t) exp (Xi,t\u22121\u03b2X) exp (Wi,t\u22121\u03b2W)\n= h0 (t) exp (Xi,t\u22121\u03b2X + Wi,t\u22121\u03b2W)\n(1)\nThe hazard function h(.) of a country i in year t represents the \nprobability that the policy is adopted by that country in that year con-\nditional on it not yet being implemented at time t\u2009\u2212\u20091. This hazard rate \nis composed of a baseline hazard rate h0(t) and a second partial hazard \nterm that includes the time-dependent matrices Xi,t\u22121 and Wi,t\u22121. In the \nCox model, the functional form of the baseline hazard is not prescribed \na-priori and not necessarily smooth, but estimated on the basis of  \nthe patterns of policy adoption in the data. For robustness, we also  \nestimate a stratified version of the model with different baseline  \nhazards h0,k(t) whereby the six continents are indexed by k.\nFor both the left-hand side and the right-hand side of equation \n(1) we model policy adoption Yi,t as a binary variable that takes on the \nvalue 1 for all years t, t\u2009+\u20091, \u2026,T if a policy has been adopted before or \nin year t. To account for autocorrelation, we cluster standard errors at \nthe level of individual countries.\nThe model is estimated from panel data on countries\u2019 adoption of \nclimate policies by maximizing a likelihood function. Unbiasedness of \nthe estimated coefficients relies on the proportional hazard assump-\ntion. This assumption is satisfied if conditional on all explanatory vari-\nables the hazard ratio of two countries is constant over time. We address \npossible violations of this assumption with our set of control variables \nand with stratification, and conduct statistical tests of Schoenfeld \nresiduals23. The control variables are discussed further below.\nThe matrix Xi,t\u22121 accounts for possible domestic influences in  \ncountry i in year t\u2009\u2212\u20091. All explanatory variables are lagged by one year \nto address concerns about reverse causality.\nThe matrix Wi,t\u22121 is a weighted average of policies Yj,t\u22121 \u2208 {0,1} \nadopted in other countries j\u2009=\u20091, \u2026, Nc, i !\u2009=\u2009j at time t\u2009\u2212\u20091, also referred \nto as a spatial lag. Other countries are weighted on the basis of a  \ncertain metric of proximity between countries. In mathematical terms, \nwe calculate\nWi,t =\n\u2211\nNc\nj=1,j\u2260iwi,j,tYj,t\n\u2211\nNc\nj=1,j\u2260iwi,j,t\n(2)\nwhere the weight wi,j,t quantifies how much country j influences country \ni in year t on the basis of a specific metric. These weights wi,j,t are con-\nstructed from several alternative data sources. For trade, we use data on \nannual bilateral trade flows from the International Monetary Fund (IMF) \nand calculate the export share xi,j,t and import share mi,j,t for every pair \nof countries in the data (i, j) and every year t. We then use these shares \nas weights; that is, wi,j,t = xi,j,t and wi,j,t = mi,j,t for exports and imports, \nrespectively. Note that the weights are generally not symmetric for a \npair of countries; that is, wi,j,t !\u2009=\u2009wj,i,t.\nFor indirect trade links, we compare the vectors of export shares \nof every pair of countries (i, j) for every time step t, xi,k,t and xj,k,t, and \ncalculate the L1 norm of the difference between the two vectors:\nwi,j,t =\n\u2211k\u2209{i,j} ||xi,k,t \u2212xj,k,t||\nNc \u22122\n(3)\nFor geographical proximity, we calculate the distance between \ncentroids of countries and use the inverse distance di,j as weight:\nwi,j =\n1\ndi,j\n(4)\nFurthermore, we construct an additional metric that is based \non geographical proximity but also takes the size of countries into \naccount. This is motivated by the hypothesis that policies in larger \neconomies have a stronger effect on policy adoption elsewhere. The \nsize of countries is expressed by the GDP of a country. In mathematical \nterms, we define another set of weights\nwi,j,t =\nlog GDPj,t\ndi,j\n(5)\nwhere di,j is again the distance between countries. A country is therefore \nconsidered more influential for domestic policy adoption the closer it \nis in space and the larger its economy is. This metric is closely related \nto gravity models of international trade that make similar assumptions \nabout the factors that determine trade between countries43.\nWe also consider shared membership in international organiza-\ntions as important for policy diffusion. For this metric, we use data on \nmembership of individual countries in international organizations \nfrom the Correlates of War database. For the sample of countries and \nyears of our data, the dataset includes 431 international organizations \nwith at least one member. For the weights w we calculate for every pair \nof countries how many memberships in international organisations are \nshared. That is, we divide the number of international organizations in \nwhich both countries are members by the number of organizations in \nwhich any of the two or both countries are members:\nwi,j,t =\n\u2211k{i \u2208Ok,t \u2227j \u2208Ok,t}\n\u2211k{i \u2208Ok,t \u2228j \u2208Ok,t}\n(6)\nwith the sets of member countries of the international organisation k \nin year t denoted as Ok,t.\nThe datasets on trade and memberships in international organiza-\ntions do not cover all years for all countries. Specifically, the dataset on \ntrade tends to cover only the more recent years, whereas the dataset on \ninternational organizations covers only the years up to 2014. To keep a \nconsistent sample throughout the empirical analysis without making \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01710-8\nassumptions about time trends or relationships between variables, we \nfill missing values by keeping values constant at the beginning and at \nthe end of our sample period. The datasets overlap for 188 countries, \nwhich is our main sample of countries in the analysis. The largest coun-\ntries not included in the sample are Venezuela and North Korea. A map \nof countries can be found in Supplementary Fig. 1.\nFor domestic control variables, we consider a large number of \npossible variables. Informed by the literature (for example, refs. 6,21, \n22,44\u201349), they include GDP per capita, the growth rate of GDP per \ncapita, government debt as a share of GDP, emissions of CO2 per GDP, \nthe service and the industry shares of GDP, the import and export \nshares of GDP, reserves of fossil fuels, a variety of governance indica-\ntors from the World Bank such as government effectiveness, control \nof corruption and regulatory quality, air quality, government expendi-\nture for welfare as share of GDP, public belief in climate change, and \nindices of democracy. In total, we collect data from 9 different sources  \n(Supplementary Table 1) for 21 variables (Supplementary Table 3).\nMany of these variables have missing values, for earlier years \nand for certain countries (Supplementary Table 2). This results in a \ndilemma. On the one hand, we want to consider as many domestic \ninfluences as possible to avoid omitted variable biases. On the other \nhand, for the analysis of diffusion it is important to have a relatively \ncomplete set of countries owing to the geographical dimension of the \nphenomenon. As a way out of this dilemma, we use iterative multiple \nimputation to fill the missing values. To limit the extent of extrapola-\ntion across countries, we first use multiple imputation to fill missing \nvalues for countries for which at least one observation of that vari-\nable exists. We also include the year as a possible predictor to model \ncountry-specific time trends. In the second step, we fill missing values \nof all other variables and countries. Our only criterion to consider a \ncountry for our analysis is thus the availability of at least one observa-\ntion of real GDP provided from the World Bank, which is the case for \n211 countries, including all 188 countries for which we also have data \nfor our spatial lags (Supplementary Fig. 1). As a robustness test, we also \nconstruct a dataset in which we only keep values constant for every \ncountry and variable, without any other imputation. We find that with \nthis method we can construct a set of 145 countries for which data on 17 \nvariables are available. Reassuringly, we find that our main results are \nrobust to this alternative method (Supplementary Table 9, column 3).\nThe choice of which of the domestic control variables to include in \nour model represents a trade-off. We do not want to exclude important \nvariables to avoid omitted variable biases, but including too many vari-\nables, many of which are highly correlated, leads to multicollinearity. \nTo find a good trade-off, we use a two-step procedure. In the first set, we \nuse Lasso regression in combination with 10-fold crossvalidation, the \nlatter of which addresses concerns of over-fitting, to identify a set of \nimportant predictors. This yields a set of ten most important variables \n(Supplementary Table 4). In addition, to gain further insights into the \nrelative importance of these variables, we also estimate a Lasso model \nwith higher penalty parameters \u03b1.\nIn the second step, we examine multicollinearity for these ten vari-\nables using the variance inflation factor (VIF). We find that for a model \nwith all ten variables, the typical upper limit for VIF of 10 is exceeded \nby one variable (Supplementary Table 5). We hence stepwise drop vari-\nables from the model until the upper limit is satisfied. At every step, we \nfocus on the variable with the highest VIF and examine its correlation \nwith all other variables. We then drop the variable itself or the most \nstrongly correlated variable depending on which of the two variables \nis considered as more important by the additional Lasso regressions \nin Supplementary Table 4. We thus stepwise drop the industry share \nof GDP.\nThis yields our preferred model specification with nine control \nvari\u00adables for domestic influences on climate policy: GDP per capita, \ngovernment effectiveness, regulatory quality, reserves of oil, govern\u00ad\nment expenditure, government expenditure for welfare (health, \neducation and social protection), a democracy index, emission inten-\nsity of the economy, and the growth rate of the debt to GDP ratio. The \ninfluence of all the remaining variables is examined in robustness \ntests. The results of these robustness tests are reassuring, as our main \nestimates for the spatial lag of carbon pricing are barely affected by \nany of the domestic influences (Supplementary Table 9).\nModelling the effect of policy diffusion on GHG emissions\nIn the second step of the analysis, we use our empirical estimates to \ncalculate the expected reductions in CO2 emissions that can be causally \nattributed to policy diffusion. For this purpose, we use the estimated \ncoefficients of all control variables and the spatial lag and feed them \ninto Monte Carlo simulations of policy adoption and policy diffusion \nusing the model in equation (1).\nWe construct counterfactual scenarios that allow us to quantify \nthe emission reductions that can be attributed to diffusion. For every \ncountry i, we compare a scenario A in which country i adopts carbon \npricing in year t with a scenario B in which country i does not do so. For \neach of the two scenarios, we calculate the hazard rate of policy adop-\ntion at time t\u2009+\u20091 for all other countries j =\u2009i based on equation (1). The \ndifference between the hazard rates of the two scenarios A and B can \nthen be considered the additional hazard of policy adoption in country \nj that can be attributed to policy diffusion from country i.\nThe Monte Carlo simulations are based on equation (1). We assume \nthe actually implemented carbon pricing policies for the year t\u2009=\u20092021 \nand let the simulations run from 2022 onwards. That is, we simulate the \nadoption and diffusion of climate policies from 2022 to 2050. To do so, \nat every time step 2022\u2009\u2264\u2009t\u2009\u2264\u20092050 we update the spatial lag Wj,t of every \ncountry, calculate its hazard of policy adoption and use this hazard to \ndraw from a probability distribution to determine whether the country \nadopts or does not adopt the policy at this time step.\nWe conduct 30,000 simulations for every country for scenario B \nand 100,000 simulations for scenario A, which is the counterfactual \nof scenario B for all countries. The simulations of scenario B result for \nevery country i in one matrix of probabilities of policy adoption of \ncountry j in year t, PB\ni,j,t with \u2211\n2050\nt=2022PB\ni,j,t \u22641\u2200i, j. The simulations of sce-\nnario A result in another matrix PA\nj,t that again satisfies \u2211\n2050\nt=2022PA\nj,t \u22641\u2200j. \nBecause there is no difference between the counterfactuals, this matrix \nPA\nj,t is the same for all countries i. On the basis of these probabilities, for \nevery country i we subsequently calculate the expected direct emission \nreductions and the expected indirect emission reductions due to  \npolicy diffusion. To map the probabilities of policy adoption onto GHG \nemissions, we assume that a carbon pricing policy reduces emissions \nby the same percentage r\u2009=\u20091% per year in all countries. A similar assump-\ntion, namely that climate policies and carbon pricing policies are  \nsimilarly effective across countries, has been made in the literature \nbefore our study50,51. The assumed value of 1% per year is slightly  \nconservative relative to estimated emission reductions from carbon \npricing policies 2003\u20132016 of about 3% per year51. Existing ETSs with \ngradually tighter caps on emission permits also allow for a comparison \nof this number. For example, in the EU ETS, between 2013 and 2020  \nthe number of permits was reduced by 1.74 % per year. In California, \nover the same period the cap was decreased by between 2% and 3.3% \nper year.\nBecause we use the same value for the parameter r for direct and \nindirect emission reductions from policy adoption, our results with \nregard to their relative sizes are relatively robust to changes in the \nparameter. We confirm this with a sensitivity analysis in which we \nvary the rate between 0.5% and 10% per year, showing that this rate \ndoes not substantially affect our comparison of indirect and direct \nemission reductions (Supplementary Fig. 7). Reassuringly, we do not \nfind clear time trends in the data on economy-wide average carbon \nprice among past policies, which suggests that there has not been a \nsystematic difference in stringency between followers and leaders \n(Supplementary Fig. 6).\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01710-8\nFormally, for every country i we calculate the direct emission \nreductions from 2022\u20132050 of implementing the policy in year 2022 as\n\u0302R\ndirect\ni,2050 = \u2211\n2050\nt=2022 [Ei,t \u2212Ei,2022 (1 \u2212r)\nt\u22122022]\n(7)\nwhere r is the effectiveness of carbon pricing as in the section above. \nFor indirect emission reductions that can be attributed to policy dif-\nfusion from country i to other countries, we use the probabilities of \npolicy adoption PA\nj,t and PB\ni,j,t of the scenarios A and B respectively. In \nmathematical terms, we take the difference between the expected \nemission reductions between the two scenarios:\n\u0302R\nindirect\ni,2050\n= \u2211j\u2260i[\u2211\n2050\n\u03be=2022[ (PB\ni,j,\u03be \u2212PA\nj,\u03be) \u22c5\n\u2211\n2050\nt=\u03be [Ej,t \u2212Ej,\u03be (1 \u2212r)\nt\u2212\u03be ]]]\n(8)\nThe indirect emission reductions are influenced by a country\u2019s \nproximity to other countries and the emissions and existing car-\nbon pricing policies of those other countries. To understand the \nimportance of these different influences, we use common metrics to \nquantify the centrality of a node in a network and adjust them for our \npurposes. Our first measure of centrality is the closeness centrality \nfor directed graphs:\nCentrality Ai = \u2211\nNc\nj=1,j\u2260iwj,i\n(9)\nOur second measure also takes the emissions E of other countries \ninto account:\nCentrality Bi = \u2211\nNc\nj=1,j\u2260iwj,iEj\n(10)\nOur third measure takes the emissions of other countries into \naccount on which a country has an indirect influence (through a third \ncountry). For this third metric we calculate\nCentrality Ci = \u2211\nNc\nj=1,j\u2260iwj,i (\u2211\nNc\nk=1,k\u2209{i,j}wk,jEk)\n(11)\nData\nWe use data on carbon pricing from the Carbon Pricing Dashboard of \nthe World Bank. The dataset includes pricing policies at the national and  \nsubnational level (Supplementary Table 2). We assign subnational  \npricing schemes to the corresponding countries and focus on the first \ncarbon pricing policy in every country. For a robustness test, we ignore \nsubnational pricing policies. Furthermore, for another two robustness \ntests we keep only either carbon tax or ETS policies in the sample. For  \nthe analysis of price levels, we combine this dataset with the World Carbon \nPricing Database52. Data on GHG emissions are from a previous report53.  \nFor the explanatory variables we use 9 different sources (Supplementary \nTable 2) for 21 raw variables (Supplementary Table 3). We use iterative  \nmultiple imputation to fill missing values (see Methods). Descriptive  \nstatistics of all covariates are shown in Supplementary Table 3. Our main \nsample covers 188 countries from 1988\u20132021 (Supplementary Fig. 1).\nReporting summary\nFurther information on research design is available in the Nature Port-\nfolio Reporting Summary linked to this article.\nData availability\nAll data are publicly available and were obtained from the following \nsources: Carbon Pricing Dashboard of the World Bank (https://car-\nbonpricingdashboard.worldbank.org/); World Carbon Pricing Data-\nbase (https://github.com/g-dolphin/WorldCarbonPricingDatabase); \nWorld Development Indicators of the World Bank (WDI) (https://\ndatabank.worldbank.org/source/world-development-indicators); \nWorld Governance Indicators (WGI) (https://info.worldbank.\norg/governance/wgi/); GHG emissions from ref. 53 and https://\ndoi.org/10.5281/zenodo.5566761; reserves of fossil fuels from the \nEnergy Intelligence Agency (EIA) (https://www.eia.gov/); Global Debt \nDatabase (GDD) (https://www.imf.org/external/datamapper/data-\nsets/GDD); Government Finance Statistics (GFS) (https://data.imf.\norg/?sk=a0867067-d23c-4ebc-ad23-d3b015045405); Expenditure by \nFunction of Government (COFOG) (https://data.imf.org/?sk=5804c5e1-\n0502-4672-bdcd-671bcdc565a9); Democracy Index (Polity5) (https://\nwww.systemicpeace.org/polityproject.html); and public belief in \nclimate change from Gallup (https://news.gallup.com/poll/117772/\nawareness-opinions-global-warming-vary-worldwide.aspx).\nCode availability\nA replication package is available at https://github.com/mlinzze/\nclimate-policy-diffusion.\nReferences\n41.\t Abel, D. The diffusion of climate policies among German \nmunicipalities. J. Public Policy 41, 111\u2013136 (2021).\n42.\t Lee, E. T. & Wang, J. W. Statistical Methods for Survival Data \nAnalysis 3rd edn (Wiley, 2003).\n43.\t Baier, S. & Standaert, S. Gravity Models and Empirical Trade \n(Oxford Univ. Press, 2020).\n44.\t Stadelmann, M. & Castro, P. Climate policy innovation in the \nSouth\u2014domestic and international determinants of renewable \nenergy policies in developing and emerging countries.  \nGlob. Environ. Change 29, 413\u2013423 (2014).\n45.\t Dolphin, G., Pollitt, M. G. & Newbery, D. M. The political economy \nof carbon pricing: a panel analysis. Oxf. Econ. Pap. 72, 472\u2013500 \n(2019).\n46.\t Best, R. & Zhang, Q. Y. What explains carbon-pricing variation \nbetween countries? Energy Policy 143, 111541 (2020).\n47.\t Levi, S., Flachsland, C. & Jakob, M. Political economy determi\u00ad\nnants of carbon pricing. Glob. Environ. Polit. 20, 128\u2013156 (2020).\n48.\t Linsenmeier, M., Mohommad, A. & Schwerhoff, G. Policy \nsequencing towards carbon pricing among the world\u2019s largest \nemitters. Nat. Clim. Change 12, 1107\u20131110 (2022).\n49.\t Zhou, S., Matisoff, D. C., Kingsley, G. A. & Brown, M. A. \nUnderstanding renewable energy policy adoption and evolution \nin Europe: the impact of coercion, normative emulation, \ncompetition, and learning. Energy Res. Soc. Sci. 51, 1\u201311 (2019).\n50.\t Eskander, S. M. S. U. & Fankhauser, S. Reduction in greenhouse \ngas emissions from national climate legislation. Nat. Clim. Change \n10, 750\u2013756 (2020).\n51.\t Best, R., Burke, P. J. & Jotzo, F. Carbon pricing efficacy: \ncross-country evidence. Environ. Resour. Econ. 77, 69\u201394 (2020).\n52.\t Dolphin, G. & Xiahou, Q. World carbon pricing database: sources \nand methods. Sci. Data 9, 573 (2022).\n53.\t Minx, J. C. et al. A comprehensive and synthetic dataset for \nglobal, regional, and national greenhouse gas emissions by \nsector 1970\u20132018 with an extension to 2019. Earth Syst. Sci. Data \n13, 5213\u20135252 (2021).\nAcknowledgements\nWe are grateful to the following for comments and suggestions (IMF, \nunless otherwise noted): S. S. Aiyar, M. Aufhammer (UC Berkeley),  \nM. Bangalore (LSE), H. Berger, S. Black, W. Chen, S. Fankhauser (Oxford \nUniversity), F. Jaumotte, T. Kliatskova (World Bank), V. Kluyev and  \nV. Thakoor; and participants of internal seminars at IMF and LSE, the \nsummer conference of the Association of Environmental and Resource \nEconomists and the annual conference of the European Association of \nEnvironmental and Resource Economists. M.L. acknowledges financial \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01710-8\nsupport from the UK Economic and Social Research Council (ESRC) \nwith grant number 2300776. For the purpose of open access, the \nauthors have applied a Creative Commons Attribution (CC BY) licence \nto any Author Accepted Manuscript version arising. All remaining \nerrors are our own. The views expressed in this paper are those of the \nauthors and do not necessarily represent the views of the IMF or its \nexecutive board or management.\nAuthor contributions\nM.L., A.M. and G.S. designed the research. M.L. collected the data  \nand conducted the analysis. M.L., A.M. and G.S. wrote and revised  \nthe manuscript.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41558-023-01710-8.\nCorrespondence and requests for materials should be addressed to \nManuel Linsenmeier.\nPeer review information Nature Climate Change thanks  \nSam Fankhauser, Pravesh Raghoo, Jakob Skovgaard and  \nYves Steinebach for their contribution to the peer review  \nof this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\n\n    Scientific Research Findings:", "answer": "We find that a country that introduces carbon pricing increases the probability that connected countries will also introduce carbon pricing. A connection between countries can be related to geography, trade or international organizations. This international diffusion of climate policies can result in large emissions reductions abroad that even exceed the domestic emissions reductions of a policy. Previous research indicates that the positive influence might work through policy learning, technology diffusion or fairness considerations. The results also suggest that countries can introduce carbon pricing without concern for freeriding. Leadership in climate policy contributes to resolving the challenge of climate change at the global level by encouraging others. The paper also points out that international policy diffusion addresses possible concerns about carbon leakage. By making it more likely that other countries introduce carbon pricing, domestic carbon pricing reduces carbon emissions abroad.", "id": 46}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 13 | July 2023 | 685\u2013692\n685\nnature climate change\nhttps://doi.org/10.1038/s41558-023-01697-2\nArticle\nBioenergy-induced land-use-change \nemissions with sectorally fragmented \npolicies\nLeon Merfort\u2009\n\u200a\u20091\u2009\n, Nico Bauer\u2009\n\u200a\u20091, Florian Humpen\u00f6der\u2009\n\u200a\u20091, David Klein1, \nJessica Strefler\u2009\n\u200a\u20091, Alexander Popp\u2009\n\u200a\u20091,2, Gunnar Luderer\u2009\n\u200a\u20091,3  \n& Elmar Kriegler\u2009\n\u200a\u20091,4\nControlling bioenergy-induced land-use-change emissions is key to \nexploiting bioenergy for climate change mitigation. However, the effect \nof different land-use and energy sector policies on specific bioenergy \nemissions has not been studied so far. Using the global integrated \nassessment model REMIND-MAgPIE, we derive a biofuel emission factor (EF) \nfor different policy frameworks. We find that a uniform price on emissions \nfrom both sectors keeps biofuel emissions at 12\u2009kg\u2009CO2\u2009GJ\u22121. However, \nwithout land-use regulation, the EF increases substantially (64\u2009kg\u2009CO2\u2009GJ\u22121 \nover 80 years, 92\u2009kg\u2009CO2\u2009GJ\u22121 over 30 years). We also find that comprehensive \ncoverage (>90%) of carbon-rich land areas worldwide is key to containing \nland-use emissions. Pricing emissions indirectly on the level of bioenergy \nconsumption reduces total emissions by cutting bioenergy demand but \nfails to reduce the average EF. In the absence of comprehensive and timely \nland-use regulation, bioenergy thus may contribute less to climate change \nmitigation than assumed previously.\nTo limit global warming and achieve the Paris climate targets, society \nneeds to bring down global carbon emissions to net zero and strongly \nreduce non-CO2 emissions1. Future cost-effective climate change miti-\ngation strategies often deploy bioenergy at large scales2. The use of \nbiofuels promises a low-carbon alternative to fossil-fuel-based liquids \nas well as the possibility to enable carbon dioxide removal from the \natmosphere using bioenergy with carbon capture and storage (BECCS)3. \nThe combined ability of biofuels to overcome and compensate for \ndecarbonization bottlenecks is a major driver of its future large-scale \ndeployment4\u20136. However, the relevance of bioenergy as a means to \nclimate change mitigation is also controversially discussed7,8, since its \nproduction will be in competition with other land-use (LU) activities \nand will thus increase the already existing pressure on land systems9,10, \naffecting biosphere integrity, biogeochemical flows, freshwater use \nand food prices9,11\u201313. Compared with other renewable energy sources, \nbioenergy has substantially higher specific land requirements14,15, \nand there is the threat that CO2 emissions from direct and indirect \nLU change16,17 (LUC/ILUC) associated with bioenergy production can \nlargely offset abated emissions18.\nA broad range of studies have investigated LUC and ILUC emis-\nsions induced by bioenergy production at different locations. They \nhave identified vastly different emission factors (EFs) ranging from \n0 to 100\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121 (refs. 19\u201322), potentially exceeding the EF of \nfossil diesel (74\u2009kg\u2009CO2\u2009GJ\u22121; ref. 23). This broad uncertainty reflects \nthe heterogeneity of land types22, characterized by stored carbon \ncontent and crop yield rates24,25. Yet, these studies do not reflect the \ninterplay between future global climate policies and the allocation of \nland areas for bioenergy and food production. Given that stringent \nclimate policies are projected to be the main driver for bioenergy \ndemand4, however, it is crucial to link the assessment of EFs to the \nReceived: 9 April 2021\nAccepted: 12 May 2023\nPublished online: 26 June 2023\n Check for updates\n1Potsdam Institute for Climate Impact Research (PIK), Member of the Leibniz Association, Potsdam, Germany. 2Faculty of Organic Agricultural Sciences, \nUniversity of Kassel, Witzenhausen, Germany. 3Global Energy Systems Analysis, Technische Universit\u00e4t Berlin, Berlin, Germany. 4Faculty of Economics \nand Social Sciences, University of Potsdam, Potsdam, Germany. \n\u2009e-mail: leon.merfort@pik-potsdam.de\n\nNature Climate Change | Volume 13 | July 2023 | 685\u2013692\n686\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nfor 2020\u20132100 into account; and second, a time-dependent marginal \n30-year EF (EFmar30(t)) that accounts for incremental CO2 emissions in \nt and relates it to increases of bioenergy over the following 30 years. \nEFmar30 varies over time; therefore, we also report the average over time \n(EFav.mar30), weighted by bioenergy production increment (Methods).\nPolicy design\nAll scenarios reach the same climate target by imposing a uniform car-\nbon price applied on all GHGs in the energy sector at the level necessary \nto comply with the carbon budget. However, they differ with respect \nto assumptions on the coordination of additional LU and energy poli-\ncies (Table 1), implying different carbon price levels. A highly idealized \nscenario with a globally uniform carbon price (UCP) in both sectors \nserves as an aspirational benchmark, representing a world in which \nclimate policies are internationally harmonized and implemented \nin the energy and LU sectors without delay. This policy is contrasted \nwith a second benchmark scenario, where the LU sector lacks any regu-\nlatory scheme for controlling emissions (noLUreg)\u2014that is, there is \nneither a price instrument on any type of LU-based GHG emissions \nnor any widespread land-protection scheme. Both policies assume \nthat bioenergy is carbon neutral in the energy sector. However, while \nLU emissions are controlled in the UCP case, there is a regulation gap \nin the noLUreg scenario.\nTo close the regulation gap, as a first set of alternative policies \nwe explored the effect of different LU regulatory schemes to serve \nas substitutes for full GHG emission regulations. First, we assumed \nreduced GHG price levels on LU emissions of 10\u201350% compared with \nthe energy sector emission prices (LUprice10\u201350%). Despite the tax \ndiscount, all LU-emission sources are covered by the regulation. These \nsectorally differentiated prices can be motivated by concerns over food \nsecurity or high food prices induced by high emission prices. Second, \nwe implemented different direct land protection policies, protecting \nall or 90% of all forests, or only focus areas\u2014namely, primary forests, \nbiodiversity hotspots or last of the wild areas (protX; see \u2018Protected \nareas\u2019 in the Supplementary Information). Third, to assess regionally \nfragmented policies, we assumed that only Organisation for Economic \nCo-operation and Development (OECD) countries implement full \nGHG emission pricing in the LU sector (LUprice-OECD), and to avoid \npotential carbon leakage effects via the bioenergy channel, which can \nbe considerable46,47, we also excluded bioenergy trade (LUprice-OECD_\nnoTrd). In particular, exports from tropical regions with high carbon \nstocks might promote additional LUC.\nContrasted to broad LU policies that are difficult to implement on \na global scale, we examined how much a tax on bioenergy consump-\ntion can reduce LUC emissions. We analysed the effects of different \nbioenergy tax levels overcoming the carbon neutrality assumption by \nattributing EFs to bioenergy (bioTax10\u201350) and also excluded bioen-\nergy trade (bioTax10\u201350_noTrd).\nEx-post EFs\nIn the absence of restrictions on LU emissions (noLUreg), cumulative \n(2020 to 2100) bioenergy-induced LUC emissions increase more than \ntenfold from 44\u2009GtCO2 in the UCP case to 493\u2009GtCO2 (Fig. 1a). This is \nqualitatively similar to but quantitatively more muted than the results \nof Wise et al.18. Since global bioenergy production only doubles to \n236\u2009EJ\u2009yr\u22121, EFav80 increases from 12 to 64\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121, which is only \nslightly smaller than the EF of conventional diesel (Fig. 1b).\nBetween these benchmark scenarios, LU and energy policies lead \nto different consequences for bioenergy and LUC emissions. Applying \n20% of the energy systems\u2019 carbon price to terrestrial GHG emissions \nreduces EFav80 to 33\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121. In contrast, a tax on bioenergy \nconsumption fails to lower specific emissions, and total emissions \ndecline only as a consequence of reduced bioenergy demand. Prohib-\niting bioenergy imports is not effective at all, since the largest part of \nthe biomass is consumed domestically in most regions.\nfuture transformation pathways of the energy system and the climate \npolicy framework. Our study assesses bioenergy EFs under a range of \nalternative climate change mitigation policies and thereby closes this \ngap in the literature.\nMost studies analysing global climate change mitigation pathways \nusing integrated assessment models (IAMs) consider an idealized \nclimate policy framework putting a uniform carbon-equivalent price \non GHG emissions from all sectors, sources and countries8. This policy \nmitigates carbon emissions from the LU sector, including those caused \nby bioenergy production. Under such a highly idealized policy frame-\nwork, bioenergy can be treated as carbon neutral in the energy sector, \nsince the associated emissions are regulated in the LU sector. However, \nin the current real-world situation, energy and LU policies are region-\nally and sectorally fragmented. While many countries have already \nstarted to implement GHG emission pricing in the energy sector, insti-\ntutional capacity building is much less developed in the agricultural \nand forestry sectors26, leaving a regulatory gap of emissions in the LU \nsector27. The implicit assumption in IAMs of institutional feasibility of \nLU mitigation has been criticized28, because the regulatory gap can \nlead to substantial emission leakage from regulated to unregulated \nregions29,30 or sectors18,31,32 involving excess bioenergy production.\nTo date, potential bioenergy-induced LUC/ILUC emissions are \nregulated via energy sector policies (if at all)19, such as by renewable \nfuel mandates (for example, the US Renewable Fuel Standard)33, fuel \ncarbon intensity standards (for example, California\u2019s Low Carbon \nFuel Standard)34 or setting lower bounds on GHG emission savings \nfor biofuels (for example, as part of the European Union\u2019s renewable \nenergy directive35). These approaches attribute highly uncertain EFs \nto different types of biofuels on the basis of representative feedstocks \nand conversion processes. Only few studies have analysed policy frame-\nworks apart from a uniform carbon price36,37. None of them systemati-\ncally compared LU-based regulation schemes (which may be weak or \nfragmented) with energy sector policies regarding their effectiveness \nin reducing total and specific emissions related to bioenergy produc-\ntion and its LU displacement effects. In this study, we investigate the \neffectiveness of differently shaped energy and LU policies on bioenergy \nEFs and ILUC, which is crucial to explore policy options for avoiding \nunsustainable bioenergy use.\nEstimating future bioenergy emissions\nWe applied the IAM framework REMIND-MAgPIE4,38,39 to derive cli-\nmate change mitigation pathways for different policy frameworks. \nAll scenarios are compatible with limiting warming well below 2\u2009\u00b0C \nby setting a carbon budget of 1,000\u2009GtCO2 to total energy-based and \nLU-based CO2 emissions from 2018 to 210040. Key socio-economic \nassumptions on population, gross domestic product, dietary choices \nand energy demand projections that drive the model results reflect \na middle-of-the-road scenario (Shared Socioeconomic Pathway 2)41. \nThe coupling of the energy-system model REMIND42,43 with the LU \nmodel MAgPIE44,45 allows for the analysis of feedback effects between \nbioenergy demand, production and associated LUC emissions, while \naccounting for the latest energy sector developments such as cheap \nrenewables and the increasing viability of electrified transportation, \nwhich determine the substitutability of biomass.\nFor each policy framework, we derived bioenergy EFs by compar-\ning these pathways with a counterfactual future without bioenergy \navailable for decarbonizing the energy sector. We thus isolated LUC \nemissions changes and attributed them to bioenergy production, \nand we derived ex-post EFs (EFex-post) in units of kg CO2 emitted per \nGJ of biofuel produced. Biofuels are the most prominent secondary \nenergy carrier produced from biomass2 (Extended Data Fig. 1). We \nthus focused our analysis on EFs for liquid fuels. Since the choices of \nboth timing and evaluation period have a large influence on the EF, \nwe applied different metrics to determine EFex-post: first, an average \n80-year EF (EFav80) taking cumulative bioenergy and CO2 emissions \n\nNature Climate Change | Volume 13 | July 2023 | 685\u2013692\n687\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nThe impact of bioenergy on LUC emissions in the presence of \nland-protection schemes depends on the precise areas that are removed \nfrom the available land pool. A policy protecting all forests resembles \nthe UCP case to a large extent (EFav80\u2009=\u200924\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121 and emissions \nof 107\u2009GtCO2 for protForest) but leads to much stronger conversion of \nother natural lands (for example, savannahs and shrubland; Extended \nData Fig. 3). Exempting 10% of the forest areas from the protection \nscheme (protForest90%) already doubles EFav80 to 49\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121, \nand targeting only focus areas performs even worse (\u2018Sensitivity analy-\nsis\u2019 in the Supplementary Information). LU emission pricing only in \nOECD countries is ineffective at reducing EFav80 because of the small \nshare of biomass production and carbon leakage via international \nagricultural markets (Supplementary Fig. 16 and \u2018Regional perspective\u2019 \nin the Supplementary Information).\nTiming and evaluation period have a large impact on the EF. The \n30-year EF strongly varies over time (Fig. 1b) and exceeds EFav80, indi-\ncating that biofuel production in the short term causes higher specific \nLU emissions than the combustion of fossil diesel. Interestingly, in \nscenarios with stringent LU mitigation, both the variability in time \nand the difference between the 30- and the 80-year time horizons are \nmuch smaller. Enforcing effective LU emissions regulatory schemes \nthus also substantially reduces uncertainty.\nSpatial allocation of bioenergy production and \nemissions\nThere is a considerable disconnect between the spatial patterns of \nbioenergy production and additional LUC emissions. We find that\u2014\nirrespective of the policy design\u2014a large fraction of LUC emissions \ndoes not originate from the sites of bioenergy cultivation but occurs \nindirectly at formerly forested areas or pasture, where agricultural \nactivity displaced by bioenergy production is moved (Fig. 2). Those \nILUC emissions as well as bioenergy plantations directly replacing \ncarbon-rich ecosystems contribute to high EFs (for example, in the \nnorthern regions of South America for noLUreg; Fig. 2a,c,d). Without \nLU policies globally, more than 85% of the additional emissions induced \nby bioenergy production originate from territories that together gen-\nerate less than 16% of total biomass production (the red and dark red \nwedges in Fig. 2b). By contrast, the main part of the bioenergy (more \nthan three quarters across all policy settings) is being produced with \na direct 80-year EF of less than 37\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121 (half the EF of diesel; \nthe blue wedges in Fig. 2), directly causing less than 10% of the total \nbioenergy-induced emissions if LU regulation policies are absent. \nTherefore, by only accounting for direct LUC emissions within major \nbioenergy-producing regions, only a small fraction of attached emis-\nsions can be traced. Accordingly, the total ILUC emissions related to \nthe total bioenergy production are considerable and vary strongly with \nthe regulatory framework.\nThis leads to two conclusions. First, the high flexibility of the LU \nsector in reallocating agricultural uses makes ILUC emissions hard to \navoid, although the absolute level depends on the underlying global LU \nor energy sector regulatory framework. Second, previous studies ana-\nlysing the direct LUC EF of bioenergy often suggested that increasing \nbioenergy production is linked to increasing EF, implying that limiting \nbioenergy production can also effectively limit EF (for example, Daio-\nglou et al.22). This rests on the assumption that expanding agricultural \narea due to bioenergy use proceeds along the lines of least marginal EF \nof land conversion. However, while such an allocation would be optimal \nfrom a sustainability perspective, it is not the allocation that emerges \nin the LU sector, as the EF is not the main criterion for allocating crop \nland by means of economic choices. Consequentially, a bioenergy tax \nreducing the overall consumption of energy crops does not automati-\ncally lead to sparing areas with high carbon stocks, as the allocation of \nemissions by EFs is not affected by the tax on bioenergy use (compare \nnoLUreg and bioTax10 in Fig. 2b).\nComponents of CO2 emissions and the role of BECCS\nRegarding the composition of total cumulated CO2 emissions, we \nobserved vastly different allocations of the carbon budget for the \nvarying policy assumptions (Fig. 3a). We found that a large fraction of \nTable 1 | Policy design\nScenario\nLU sector policies\nEnergy sector \npoliciesa,b\nPolicy type\nBenchmark policies\nUCP\nCarbon price on all \nGHGsb at the energy \nsector level\nBioenergy \nis treated as \ncarbon neutral\nHarmonized \nenergy and LU \nsector policies\nnoLUreg\nNone\nBioenergy \nis treated as \ncarbon neutral\nNone (no LU \nregulation)\nFragmented policies\nLUprice10\u201350%\nCarbon price on all \nGHGsb, lower than \nthe energy sector \nlevelc\nBioenergy \nis treated as \ncarbon neutral\nBroad land \nprotection\nprotForest\nPrimary and \nsecondary forests \nare protected\nprotForest90%\n90% of all primary \nand secondary \nforests are protected \nin each grid cell\nBioenergy \nis treated as \ncarbon neutral\nNarrow land \nprotection\nprotPrimforest\nPrimary forests are \nprotected\nprotBH\nBiodiversity hotspots \nare protected\nprotLW\nLast of the wild areas \nare protected\nLUprice-OECD\nCarbon price on all \nGHGsb at the energy \nsector level, only in \nOECD countries\nBioenergy \nis treated as \ncarbon neutral\nRegionally \nfragmented \npolicies\nLUprice-OECD_\nnoTrd\nCarbon price on all \nGHGsb at the energy \nsector level, only in \nOECD countries\nBioenergy \nis treated as \ncarbon neutral; \nno bioenergy \ntrade in OECD \ncountries\nbioTax10\u201350\nNone\nBioenergy taxd\nBioenergy \npolicies\nbioTax10\u201350_\nnoTrd\nNone\nBioenergy taxd; \nno bioenergy \ntrade globally\nSensitivity analysise\nUCP_\nbioYield75%\nSame as UCP; reduction of bioenergy \ncrop yields to 75%\nLow yields\nUCP_\nhighcostTC\nSame as UCP; increased costs for \ntechnological change in the LU sector\nThe scenarios are divided into benchmark scenarios and scenarios with fragmented energy \nand LU policies. They are grouped into different policy types. For this study, we classified \nLU policies based on a global GHG price or with full forest protection as \u2018broad\u2019 land \nprotection policies, while all others are defined as \u2018narrow\u2019, since they miss substantial parts \nof carbon-rich areas (\u2018Protected areas\u2019 in the Supplementary Information). aIn all scenarios, \na uniform carbon price is applied to all GHGs in the energy sector. bWhile carbon prices \nare globally uniform from 2050 on, they differ between regions before 2050 for reasons of \ninterregional equity (Methods). cThe carbon price within the LU sector is at a level of 10\u201350% \nof the price in the energy system for scenarios LUprice10\u201350%, respectively. dThe tax level \nis determined by the carbon price multiplied by a predefined, fixed factor given in kg CO2 \nper GJ of primary energy (PE) dry matter (DM) biomass. For example, bioTax20 represents \na policy charging bioenergy as if it had an EF of 20\u2009kg\u2009CO2\u2009GJPE\n\u22121. Due to conversion losses, \n10\u201350\u2009kg\u2009CO2\u2009GJPE\n\u22121 correspond to 24\u2013122\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121, and while the tax level is expressed \nin units of PE, ex-post EFs are expressed in units of biofuel for a better comparability with \nvalues from the literature. eSee \u2018Sensitivity analysis\u2019 in the Supplementary Information for \ndetails and more scenarios.\n\nNature Climate Change | Volume 13 | July 2023 | 685\u2013692\n688\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nbiomass is used in combination with CCS (58\u201395% in 2100), particularly \nwhen total production is reduced by a bioenergy tax (Extended Data \nFig. 6). In the absence of comprehensive LU sector emission regula-\ntions, however, a large fraction of carbon dioxide removals (CDR) from \nBECCS is used to compensate the additional LUC emissions (Fig. 3b). \nWithout LU mitigation, for instance, in 2100 only 15% of cumulated \nCDR from BECCS remain after subtracting bioenergy-induced LUC \nemissions. Before 2050, cumulated bioenergy-induced LUC emissions \neven exceed BECCS savings by far for all policy settings except UCP \n(Extended Data Fig. 5). Note that this only compares LUC emissions \nwith direct CDR but does not account for fossil fuel substitution. While \nthe substitution effect plays an important role in reducing emissions \nfrom fossil fuels, quantifying the avoided emissions attributed to the \nbioenergy part of BECCS is inherently ambiguous, as there are also other \nsubstitution options such as direct electrification in the integrated \nsystems perspective.\nWe furthermore observed that cumulated energy system emis-\nsions need to be reduced in scenarios with high LUC emissions com-\npared with the UCP policy to balance the total budget. The additional \nbiomass is then used to accelerate the phase-out of fossil fuels, par-\nticularly oil (Extended Data Fig. 7).\nIf bioenergy is priced in the energy sector, LUC emissions decrease \nwith increasing tax level, but the reduced demand for bioenergy \nenforces a stronger and faster electrification than in both the UCP \nand the noLUreg scenarios (Extended Data Fig. 8). At the same time, \nthe share of emissions from the transport sector increases due to the \nCumulated LUC emissions (left scale)\nAveraged bioenergy production (right scale)\nUCP\nNone\nBroad land protection\nNarrow land\nprotection\nReg. frag.\nBioenergy policies\nLow yields\n0\n100\n200\n300\n400\n500\na\nb\n0\n100\n200\nLUC emissions (GtCO2)\nBioenergy production (EJ yr\u20131)\n809\n212\n192\n164\n320\n1,516\n165\n301\nEFav80\nEFav.mar30\nUCP\nnoLUreg\nLUprice10%\nLUprice20%\nLUprice30%\nLUprice50%\nprotForest\nprotForest90%\nprotPrimforest\nprotBH\nprotLW\nLUtax-OECD\nLUtax-OECD_noTrd\nbioTax10\nbioTax20\nbioTax30\nbioTax40\nbioTax50\nbioTax10_noTrd\nbioTax20_noTrd\nbioTax30_noTrd\nbioTax40_noTrd\nbioTax50_noTrd\nUCP_bioYield75%\nUCP_highcostTC\n0\n50\n100\n150\nEFex-post (kg CO2 GJbiofuel\n\u20131)\nDiesel*\nNatural \ngas*\nWind**\nHydro**\nSolar\nPV**\nFig. 1 | Bioenergy-induced LUC emissions, bioenergy production and EFs.  \na, Cumulative global bioenergy-induced LUC emissions and bioenergy \nproduction, given as the averaged annual global production, are both evaluated \nfor the period from 2020 to 2100 and shown for different policies. The policies \nare grouped into different policy types as defined in Table 1. The white bars \nindicate cumulative emissions and averaged annual bioenergy production in the \nperiod from 2020 to 2050. b, Ex-post EFs given per unit of biofuel produced for \ndifferent policies. The box plots show the temporal variation of EFmar30(t) for years \nt between 2025 and 2070 (sample size of n\u2009=\u20099; bioTax40 and bioTax40_noTrd are \nexceptions with n\u2009=\u20098, since the year 2070 is excluded; see \u2018Emission factors over \ntime\u2019 in the Supplementary Information). The minimum and maximum of each \nbox confine the interquartile range, the whiskers represent the first and fourth \nquartiles, and the centre line indicates the median value. *The reference EFs for \ndiesel and natural gas are taken from UBA23. **The reference values for electricity \nproduced (here in kg\u2009CO2\u2009GJelectricity\n\u22121) from solar photovoltaic (PV), hydro and \nwind are medium values from AR5 based on life cycle analysis61. For a comparison \nwith 20-year marginal EFs and EFs of N2O, see Extended Data Fig. 2.\n\nNature Climate Change | Volume 13 | July 2023 | 685\u2013692\n689\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nlack of biofuels. The dwindling availability of biomass even leads to \nhigher CO2 prices (Fig. 3c) compared with the already high prices in \nnoLUreg, which makes direct air capture with carbon storage (DACCS) \ncompetitive as a CDR option.\nIt is also worth noting that even a comparatively small carbon price \non LU-sector-based emissions (LUprice10%) is sufficient to abate most \nof the non-bioenergy-related LUC CO2 emissions (from 235 in noLUreg \nto 56\u2009GtCO2). At the same time, the required carbon price is reduced by \n35% from US$291 to US$193 per tCO2 in 2050 (Fig. 3c).\nDiscussion and conclusion\nIn our study, we analysed the impact of bioenergy on LUC emissions in \ndifferent climate policy settings, comparing LU policies controlling land \nallocation with energy policies controlling bioenergy use. We showed \nthat a uniform carbon price across energy and LU sectors substan-\ntially lowers the specific emissions attributed to a unit of bioenergy in \ncomparison with scenarios without broad LU regulations. Regulations \nof bioenergy demand reduce deployment level but fail to induce a \nreduction of the EF, keeping it virtually at a high level and thus resulting \nin substantially higher overall mitigation effort and carbon prices to \nreach climate targets. A bioenergy consumption tax fails to steer LUC \ndecisions towards low-EF areas and cannot prevent the conversion of \nhigher-carbon land; hence, it is not suitable to emulate the uniform car-\nbon price regime across all sectors. Our analysis thus shows that global \nregulation of LUC emissions\u2014while being extremely aspirational\u2014is \nessential for bioenergy to be beneficial for cost-effective mitigation. \nThe comprehensive coverage of all carbon-rich land areas worldwide \nis of the utmost importance, as even a 90% forest protection has only \nlittle effect on the EF. Since the conditions for such an exhaustive policy \nframework to materialize are highly idealized, involving close collabo-\nration between and functioning governance within all countries, this \nfinding implies climate policy sequencing: first, global LU regulation \nneeds to be in place, and only then should large-scale bioenergy be \nconsidered. Pushing ambitious land-based mitigation is compelling \n44\n493\n277\n107\n276\n350\n62\n111\n236\n169\n138\n190\n167\n23\n25\n100\n225\n400\n12\n48\n108\n191\nGlobal cumulative emissions (GtCO2)\nGlobal averaged bioenergy production (EJ yr\u20131)\nGtCO2\nEJ yr\u20131\na\nUCP\nnoLUreg\nLUprice10%\nprotForest\nprotPrimforest\nbioTax10\nbioTax50\nnoLUreg\nnoLUreg\nnoLUreg\nb\nd\nc\nCumulative bioenergy-induced\nLUC emissions (GtCO2 Mha\u20131)\nCumulative bioenergy production\n(EJ Mha\u20131)\n25\n20\n15\n10\n5\n0\n1.0\n0.8\n0.6\n0.4\n0.2\n0\nSpatial LU characteristic\nEFloc    (kg CO2 GJbiofuel\n\u20131)\nILUC\nEFloc       > 74\n37 < EFloc        < 74\nEFloc      < 37 \nNo emissions\nav80\nex-post\nex-post\nex-post\nFig. 2 | Spatial allocation of LUC CO2 emissions and bioenergy production.  \na, Spatially disaggregated bioenergy EFs for the noLUreg scenario. In bright blue \nareas, bioenergy is produced without additional LUC emissions at the place of \nproduction; in dark red areas, only ILUC emissions occur. Other territories are \nclassified by the ratio of bioenergy-induced emissions to bioenergy production \n(EFav80\nloc ). The country map is based on the object wrld_simpl in the R package \nmaptools. b, Total global bioenergy-induced LUC emissions and the global \naveraged annual bioenergy production, colour-coded according to the \nassociated EFs as described in a, for selected scenarios (the other scenarios are \nshown in Extended Data Fig. 4). c,d, The spatial distribution of bioenergy \nproduction (c) and bioenergy-induced emissions (d) for the example of South \nAmerica. In all four panels, the quantities are cumulated over the period between \n2020 and 2100. See the Methods for a description of the analysis of EFav80\nloc  and the \nSupplementary Information for figures of the other policy assumptions, \nincluding maps of baseline (not bioenergy-related) emissions.\n\nNature Climate Change | Volume 13 | July 2023 | 685\u2013692\n690\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nbecause GHG emissions from food production will need to be sub-\nstantially reduced regardless to meet the 1.5\u00b0 or well below 2\u00b0 target48.\nEven though a bioenergy tax is an inferior substitute for LU pol-\nicies, it is still a crucial element in energy policies, since an EF can \nbe directly applied to bioenergy use. To value fossil emissions and \nbioenergy-induced LUC equally in the absence of LU regulations, our \nresults indicate that the required EF (64\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121 over 80 years \nand 92\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121 over 30 years) is at the upper end of values \nidentified in dedicated studies, which range between 0 and 100\u2009kg\u2009 \nCO2\u2009GJbiofuel\n\u22121 on a 30-year time horizon20. Interestingly, uniform regula-\ntion of LU sector emissions reduces not only the EF but also the uncer-\ntainty due to choices of operationalization, such as the time horizon.\nIt is important to point out that the bioenergy crop yields \nreported in our study reflect average yields based on modelling results  \n(Methods). Present-day yields from field experiments reported by  \nLi et al.49 (11.5 and 8.1\u2009t DM per ha per yr for Miscanthus and switchgrass, \nrespectively) exceed current practice50. Globally averaged bioenergy \ncrop yields in the UCP scenario are even higher in 2020 (14.1\u2009t DM per \nha per yr), reaching 19.4\u2009t DM per ha per yr in 2040 and 25.7\u2009t DM per ha \nper yr in 2100 due to technological progress induced by research and \ndevelopment (R&D) (Supplementary Fig. 13). However, bioenergy pro-\nduction is low in our scenarios in 2020 and can be realized in a few highly \nproductive areas (LU optimization in MAgPIE), which explains why the \nbioenergy yields reported here are higher than in field experiments. \nCurrent and future yields are important for the amount of bioenergy \nthat can be realized. We observe that 25% lower initial bioenergy yields \n(UCP_bioYield75%) increase EFav80 from 12 to 14\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121, and less \noptimistic assumptions on R&D efficiency (UCP_highcostTC, affecting \nfuture bioenergy and food crop yields) increase it to 16\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121. \nFurthermore, massive investments in technological progress are still \nrequired to facilitate large-scale bioenergy production even in an ide-\nalized policy framework (Supplementary Figs. 14 and 15). For details \nregarding crop yield assumptions, refer to \u2018Sensitivity analysis\u2019 in the \nSupplementary Information.\nAnother note is that the carbon price only internalizes the carbon \nvalue, omitting other adverse side effects, such as unsustainable fresh-\nwater use, biodiversity loss51 or LUC in low-carbon ecosystems (such \nas savannahs). For instance, about 300\u2009Mha of non-forest natural land \nis converted into cropland by 2100 in UCP, one third of that due to bio-\nenergy (Extended Data Fig. 3). The EF does not capture N2O emissions \nfrom fertilization of bioenergy crops52 or the impact on food prices53 and \nalso ignores benefits from fossil fuel substitution. For example, we find \nthat within the UCP policy framework, agricultural commodity prices \nincrease sharply by 50\u2013120% from 2010 to 2100 (higher values for pes-\nsimistic crop yield assumptions), with half of that being attributed to \nbioenergy (Extended Data Fig. 9). Furthermore, our results depend on \ntechno-economic assumptions of BECCS and DACCS technologies. In par-\nticular, more optimistic assumptions on DACCS might reduce the demand \nfor bioenergy. Lastly, it is noteworthy that the carbon price as an indicator \nof mitigation effort is not equivalent to total mitigation costs of a policy.\nOur study confirms that LU emission pricing is an effective and \nefficient instrument to regulate LUC emissions even under large-scale \nUCP\nNone\nBroad land prot.\nNarrow land\nprot.\nReg. f.\nBioenergy policies\nL. yields\nUCP\nnoLUreg\nLUprice10%\nLUprice20%\nLUprice30%\nLUprice50%\nprotForest\nprotForest90%\nprotPrimforest\nprotBH\nprotLW\nLUtax-OECD\nLUtax-OECD_noTrd\nbioTax10\nbioTax20\nbioTax30\nbioTax40\nbioTax50\nbioTax10_noTrd\nbioTax20_noTrd\nbioTax30_noTrd\nbioTax40_noTrd\nbioTax50_noTrd\nUCP_bioYield75%\nUCP_highcostTC\n\u2013500\n0\n500\n1,000\n1,500\nCO2 emissions (GtCO2)\n0\n25\n50\n75\n100\n\u03b7BECCS (%)\n150\n200\n250\n300\n350\nUCP\nnoLUreg\nLUprice10%\nLUprice20%\nLUprice30%\nLUprice50%\nprotForest\nprotForest90%\nprotPrimforest\nprotBH\nprotLW\nLUtax-OECD\nLUtax-OECD_noTrd\nbioTax10\nbioTax20\nbioTax30\nbioTax40\nbioTax50\nbioTax10_noTrd\nbioTax20_noTrd\nbioTax30_noTrd\nbioTax40_noTrd\nbioTax50_noTrd\nUCP_bioYield75%\nUCP_highcostTC\nCarbon price (US$2005 per tCO2)\nSector\nNon-transport\nTransport\nLUC from food production\nLUC/ILUC from bioenergy\nDirect air capture\nRemoval BECCS residues\nRemoval BECCS energy crops\nPolicy type\nUCP\nNone\nBroad land protection\nNarrow land protection\nRegionally fragmented policies\nBioenergy policies\nLow yields\na\nc\nb\nNet emissions\nFig. 3 | Composition of emissions, BECCS efficiency and carbon prices.  \na, Composition of total cumulative (2020\u20132100) anthropogenic CO2 emissions \nfor different policy assumptions. Energy sector emissions are split into demand-\nside emissions from the transport sector and non-transport-sector emissions \ncomprising supply-side emissions (for example, for electricity production) \nand demand-side emissions from buildings and industry, including process \nemissions. LUC emissions from food production are CO2 LUC emissions not \nrelated to bioenergy. Bioenergy from residues is assumed to be carbon neutral. \nb, The 80-year BECCS efficiency factor \u03b7BECCS indicates how much of the stored \ncarbon is net removal from the atmosphere if bioenergy-induced LUC emissions \nare subtracted. For instance, \u03b7BECCS\u2009=\u200915% for noLUreg implies that only 15% of \nthe CDR savings are effectively removed from the atmosphere, as the remaining \n85% are offset by LUC emissions. The 30-year BECCS efficiency factor is shown in \nExtended Data Fig. 5. c, Energy system GHG prices in the year 2050. After a phase-\nin period, prices are equal across regions from 2050 on.\n\nNature Climate Change | Volume 13 | July 2023 | 685\u2013692\n691\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\ndeployment of bioenergy. However, we also show that those policies \ncannot be emulated by regulation of bioenergy use, raising the ques-\ntion to what extent the LU sector can be effectively regulated to make \nlarge-scale bioenergy use sustainable. The literature points to numer-\nous challenges for regulating LU, ranging from monitoring, reporting \nand verification to the need for huge institutional capacity54\u201356. Effective \nregulation may also be undermined by high bioenergy and food prices \ncreating political pressure to clear land. Moreover, the distributional \nimplications of regulating LU emissions affect land tenure and liveli-\nhoods (for example, small-scale landowners that need to clear land to \neat), raising strong equity and political economy concerns57\u201359.\nHence, the policy challenge is to either close the regulatory gap by \ncomprehensively regulating the LU sector (on a global level, as our analy-\nsis on regional fragmentation shows) and produce biomass at scale, or \nsubstantially reduce bioenergy demand by increasing its price, as bioen-\nergy cannot be treated carbon neutral otherwise. Given the difficulties in \nregulating the LU sector described above, this implies that bioenergy may \nplay a much smaller role in climate change mitigation than suggested by \nmost IAM scenarios2. The main driving force behind this challenge is the \nhuge demand for non-electric energy, particularly transport fuels. Broad \nand deep electrification of end uses would thus lower the pressure on the \nland system and bypass the regulatory gaps in the LU sector60.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author con-\ntributions and competing interests; and statements of data and code \navailability are available at https://doi.org/10.1038/s41558-023-01697-2.\nReferences\n1.\t\nAllen, M. et al. in Special Report on Global Warming of 1.5 \u00b0C  \n(eds Masson-Delmotte, V. et al.) Summary for Policymakers  \n(IPCC, WMO, 2018).\n2.\t\nBauer, N. et al. Global energy sector emission reductions and \nbioenergy use: overview of the bioenergy demand phase of the \nEMF-33 model comparison. Climatic Change 163, 1553\u20131568 (2018).\n3.\t\nKlein, D. et al. The value of bioenergy in low stabilization \nscenarios: an assessment using REMIND-MAgPIE. Climatic \nChange 123, 705\u2013718 (2014).\n4.\t\nBauer, N. et al. Bio-energy and CO2 emission reductions: an \nintegrated land-use and energy sector perspective. Climatic \nChange 163, 1675\u20131693 (2020).\n5.\t\nRose, S. K. et al. Bioenergy in energy transformation and climate \nmanagement. Climatic Change 123, 477\u2013493 (2014).\n6.\t\nKrey, V., Luderer, G., Clarke, L. & Kriegler, E. Getting from here to \nthere\u2014energy technology transformation pathways in the EMF27 \nscenarios. Climatic Change 123, 369\u2013382 (2014).\n7.\t\nIPCC Special Report on Climate Change and Land (Cambridge \nUniv. Press, 2022).\n8.\t\nCreutzig, F. et al. Reconciling top-down and bottom-up  \nmodelling on future bioenergy deployment. Nat. Clim. Change 2, \n320\u2013327 (2012).\n9.\t\nHumpen\u00f6der, F. et al. Large-scale bioenergy production:  \nhow to resolve sustainability trade-offs? Environ. Res. Lett. 13, \n024011 (2018).\n10.\t Popp, A. et al. The economic potential of bioenergy for climate \nchange mitigation with special attention given to implications for \nthe land system. Environ. Res. Lett. 6, 034017 (2011).\n11.\t\nHeck, V., Gerten, D., Lucht, W. & Popp, A. Biomass-based negative \nemissions difficult to reconcile with planetary boundaries.  \nNat. Clim. Change 8, 151\u2013155 (2018).\n12.\t Hof, C. et al. Bioenergy cropland expansion may offset positive \neffects of climate change mitigation for global vertebrate \ndiversity. Proc. Natl Acad. Sci. USA 115, 13294\u201313299 (2018).\n13.\t Schueler, V., Weddige, U., Beringer, T., Gamba, L. & Lamers, \nP. Global biomass potentials under sustainability restrictions \ndefined by the European Renewable Energy Directive 2009/28/\nEC. GCB Bioenergy 5, 652\u2013663 (2013).\n14.\t Searchinger, T. D., Beringer, T. & Strong, A. Does the world have \nlow-carbon bioenergy potential from the dedicated use of land? \nEnergy Policy 110, 434\u2013446 (2017).\n15.\t Luderer, G. et al. Environmental co-benefits and adverse \nside-effects of alternative power sector decarbonization \nstrategies. Nat. Commun. 10, 5229 (2019).\n16.\t Verstegen, J. A. et al. What can and can\u2019t we say about indirect \nland-use change in Brazil using an integrated economic\u2013land-use \nchange model? GCB Bioenergy 8, 561\u2013578 (2016).\n17.\t Arima, E. Y., Richards, P., Walker, R. & Caldas, M. M. Statistical \nconfirmation of indirect land use change in the Brazilian Amazon. \nEnviron. Res. Lett. 6, 024010 (2011).\n18.\t Wise, M. et al. Implications of limiting CO2 concentrations for land \nuse and energy. Science 324, 1183\u20131186 (2009).\n19.\t Daioglou, V. et al. Progress and barriers in understanding and \npreventing indirect land-use change. Biofuels Bioprod. Bioref. 14, \n924\u2013934 (2020).\n20.\t Warner, E., Zhang, Y., Inman, D. & Heath, G. Challenges  \nin the estimation of greenhouse gas emissions from biofuel- \ninduced global land-use change. Biofuels Bioprod. Bioref. 8, \n114\u2013125 (2014).\n21.\t Wicke, B., Verweij, P., Meijl, H., van Vuuren, D. P. & Faaij, A. \nP. Indirect land use change: review of existing models and \nstrategies for mitigation. Biofuels 3, 87\u2013100 (2012).\n22.\t Daioglou, V. et al. Greenhouse gas emission curves for  \nadvanced biofuel supply chains. Nat. Clim. Change 7,  \n920\u2013924 (2017).\n23.\t Juhrich, K. CO2-Emissionsfaktoren f\u00fcr fossile Brennstoffe \n(Umweltbundesamt, 2016); https://www.umweltbundesamt.de/\npublikationen/co2-emissionsfaktoren-fuer-fossile-brennstoffe\n24.\t Wicke, B., Dornburg, V., Junginger, M. & Faaij, A. Different palm oil \nproduction systems for energy purposes and their greenhouse \ngas implications. Biomass Bioenergy 32, 1322\u20131337 (2008).\n25.\t Creutzig, F. et al. Bioenergy and climate change mitigation: an \nassessment. GCB Bioenergy 7, 916\u2013944 (2015).\n26.\t State and Trends of Carbon Pricing 2020 (World Bank, 2020); \nhttps://openknowledge.worldbank.org/handle/10986/33809\n27.\t Eskander, S. M. S. U. & Fankhauser, S. Reduction in greenhouse \ngas emissions from national climate legislation. Nat. Clim. Change \n10, 750\u2013756 (2020).\n28.\t Creutzig, F. Economic and ecological views on climate change \nmitigation with bioenergy and negative emissions. GCB Bioenergy \n8, 4\u201310 (2016).\n29.\t Otto, S. A. C. et al. Impact of fragmented emission reduction \nregimes on the energy market and on CO2 emissions related  \nto land use: a case study with China and the European Union  \nas first movers. Technol. Forecast. Soc. Change 90, 220\u2013229 \n(2015).\n30.\t Gonz\u00e1lez-Eguino, M., Capell\u00e1n-P\u00e9rez, I., Arto, I., Ansuategi, A. \n& Markandya, A. Industrial and terrestrial carbon leakage under \nclimate policy fragmentation. Clim. Policy 17, S148\u2013S169 (2017).\n31.\t Reilly, J. et al. Using land to mitigate climate change: hitting \nthe target, recognizing the trade-offs. Environ. Sci. Technol. 46, \n5672\u20135679 (2012).\n32.\t Popp, A. et al. Land-use protection for climate change mitigation. \nNat. Clim. Change 4, 1095\u20131098 (2014).\n33.\t Lark, T. J. et al. Environmental outcomes of the US Renewable Fuel \nStandard. Proc. Natl Acad. Sci. USA 119, e2101084119 (2022).\n34.\t Yeh, S., Witcover, J., Lade, G. E. & Sperling, D. A review of low \ncarbon fuel policies: principles, program status and future \ndirections. Energy Policy 97, 220\u2013234 (2016).\n\nNature Climate Change | Volume 13 | July 2023 | 685\u2013692\n692\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\n35.\t Directive (EU) 2018/2001 of the European Parliament and of the \nCouncil of 11 December 2018 on the promotion of the use of \nenergy from renewable sources (Text with EEA relevance). Off. J. \nEU L328(61), 82\u2013209 (2018).\n36.\t Calvin, K. et al. Trade-offs of different land and bioenergy policies \non the path to achieving climate targets. Climatic Change 123, \n691\u2013704 (2014).\n37.\t Calvin, K. et al. Near-term limits to mitigation: challenges arising \nfrom contrary mitigation effects from indirect land-use change \nand sulfur emissions. Energy Econ. 42, 233\u2013239 (2014).\n38.\t Kriegler, E. et al. Fossil-fueled development (SSP5): an energy and \nresource intensive scenario for the 21st century. Glob. Environ. \nChange 42, 297\u2013315 (2017).\n39.\t Bauer, N. et al. Quantification of an efficiency\u2013sovereignty \ntrade-off in climate policy. Nature 588, 261\u2013266 (2020).\n40.\t Rogelj, J. et al. in Special Report on Global Warming of 1.5\u2009\u00b0C (eds \nMasson-Delmotte, V. et al.) Ch. 2 (IPCC, WMO, 2018).\n41.\t Riahi, K. et al. The Shared Socioeconomic Pathways and their \nenergy, land use, and greenhouse gas emissions implications: an \noverview. Glob. Environ. Change 42, 153\u2013168 (2017).\n42.\t Baumstark, L. et al. REMIND2.1: transformation and innovation \ndynamics of the energy\u2013economic system within climate and \nsustainability limits. Geosci. Model Dev. Discuss. https://doi.org/ \n10.5194/gmd-2021-85 (2021).\n43.\t Luderer, G. et al. REMIND\u2014REgional Model of INvestments and \nDevelopment. Zenodo https://doi.org/10.5281/zenodo.4091409 \n(2020).\n44.\t Dietrich, J. P. et al. MAgPIE 4\u2014a modular open-source framework \nfor modeling global land systems. Geosci. Model Dev. 12, \n1299\u20131317 (2019).\n45.\t Dietrich, J. P. et al. MAgPIE\u2014an open source land-use modeling \nframework. Zenodo https://doi.org/10.5281/ZENODO.1418752 \n(2020).\n46.\t Daioglou, V. et al. Implications of climate change mitigation \nstrategies on international bioenergy trade. Climatic Change 163, \n1639\u20131658 (2020).\n47.\t Lamers, P., Hoefnagels, R., Junginger, M., Hamelinck, C. & Faaij, A. \nGlobal solid biomass trade for energy by 2020: an assessment of \npotential import streams and supply costs to North-West Europe \nunder different sustainability constraints. GCB Bioenergy 7, \n618\u2013634 (2015).\n48.\t Clark, M. A. et al. Global food system emissions could preclude \nachieving the 1.5\u00b0 and 2\u00b0C climate change targets. Science 370, \n705\u2013708 (2020).\n49.\t Li, W., Ciais, P., Makowski, D. & Peng, S. A global yield dataset \nfor major lignocellulosic bioenergy crops based on field \nmeasurements. Sci. Data 5, 180169 (2018).\n50.\t Fike, J. H. et al. Switchgrass nitrogen response and  \nestimated production costs on diverse sites. GCB Bioenergy 9, \n1526\u20131542 (2017).\n51.\t Lecl\u00e8re, D. et al. Bending the curve of terrestrial biodiversity \nneeds an integrated strategy. Nature https://doi.org/10.1038/\ns41586-020-2705-y (2020).\n52.\t Popp, A. et al. On sustainability of bioenergy production: \nintegrating co-emissions from agricultural intensification. \nBiomass Bioenergy 35, 4770\u20134780 (2011).\n53.\t Lotze-Campen, H. et al. Impacts of increased bioenergy \ndemand on global food markets: an AgMIP economic model \nintercomparison. Agric. Econ. 45, 103\u2013116 (2014).\n54.\t Ochieng, R. M., Visseren-Hamakers, I. J., Arts, B., Brockhaus, M. \n& Herold, M. Institutional effectiveness of REDD+ MRV: countries \nprogress in implementing technical guidelines and good \ngovernance requirements. Environ. Sci. Policy 61, 42\u201352 (2016).\n55.\t Wang, X. et al. Taking account of governance: implications for \nland-use dynamics, food prices, and trade patterns. Ecol. Econ. \n122, 12\u201324 (2016).\n56.\t Waller, L. et al. Contested framings of greenhouse gas removal \nand its feasibility: social and political dimensions. WIREs Clim. \nChange 11, e649 (2020).\n57.\t Stevanovi\u0107, M. et al. Mitigation strategies for greenhouse gas \nemissions from agriculture and land-use change: consequences \nfor food prices. Environ. Sci. Technol. 51, 365\u2013374 (2017).\n58.\t Hasegawa, T. et al. Risk of increased food insecurity under \nstringent global climate change mitigation policy. Nat. Clim. \nChange 8, 699\u2013703 (2018).\n59.\t Fujimori, S. et al. A multi-model assessment of food security \nimplications of climate change mitigation. Nat. Sustain. 2, \n386\u2013396 (2019).\n60.\t Luderer, G. et al. Impact of declining renewable energy costs on \nelectrification in low-emission scenarios. Nat. Energy 7, 32\u201342 (2021).\n61.\t IPCC Climate Change 2014: Mitigation of Climate Change (eds \nEdenhofer, O. et al.) (Cambridge Univ. Press, 2014).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2023\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nMethods\nGeneral\nTo assess the impact of bioenergy and to fully cover feedback effects \nbetween LUC CO2 emissions and bioenergy demand, we used the cou-\npled integrated assessment modelling framework REMIND-MAgPIE.\nREMIND v.2.1.2 is an open-source global multi-regional \nRamsey-type general equilibrium model of economic growth with \na detailed representation of the energy sector, hard-coupled to the \nmacro-economic core42,43. Using optimization methods, it finds a mar-\nket equilibrium while maximizing intertemporal global welfare. Via \ndifferent conversion routes (see Supplementary Table 1 for a list of \ntechnologies converting biomass feedstocks into secondary energy \ncarriers, including conversion efficiencies and carbon capture rates \nfor BECCS technologies), REMIND represents the supply, trade and \nconversion of biomass feedstocks along the value chain to final energy \ncarriers along with relevant GHG emissions and removals. The REMIND \nmodel therefore values the energy and the carbon content of biomass \nfeedstocks given the market conditions and the regulatory framework. \nIn climate change mitigation scenarios, most of the biomass (primarily \nDM lignocellulosic biomass; Extended Data Fig. 10) is converted into \nbio-liquids (Extended Data Fig. 1).\nMAgPIE v.4.2.1 is an open-source global multi-regional partial \nequilibrium model of the LU sector that models LU dynamics spatially \nexplicitly using recursive dynamic optimization44,45. The model covers \ntwo types of modern purpose-grown bioenergy feedstocks\u2014namely, \ngrassy and woody biomass\u2014as well as residues from forestry and agri-\nculture. Both types are lignocellulosic biomass feedstocks that can be \nused for second-generation bioenergy technologies (Supplementary \nTable 1). First-generation feedstocks from sugar/starch or oil crops \nare also available, but we assumed that production may be expanded \nonly until 2030 and is limited afterwards, due to higher environmental \nimpacts than second-generation feedstocks. Since the irrigation of bio-\nenergy crops leads to unsustainable freshwater use9, we only allowed \nfor rain-fed production.\nBoth models are soft-coupled, balancing prices and quantities of \npurpose-grown lignocellulosic bioenergy feedstocks and GHGs4. For \nresidues and first-generation feedstocks, prices and quantities are not \ndirectly coupled but instead exogenously prescribed and harmonized \nbetween REMIND and MAgPIE42. Since we focused our analysis on \nsecond-generation biofuels from purpose-grown feedstocks (residues \nare assumed to be carbon neutral), a potential bioenergy tax is imposed \nonly on and the ex-post evaluation of EFs refers only to purpose-grown \nlignocellulosic biomass. Therefore, in the remainder of the study, \nthe term \u2018bioenergy\u2019 always refers to purpose-grown lignocellulosic \nbiomass, if not stated otherwise. The main policy instrument to meet \na given climate target is a pricing of GHG emissions. GHG prices that are \nby default applied to all types of GHGs from all sectors and sources are \nderived in REMIND and passed to MAgPIE to meet the predefined GHG \nbudget in 2100 of total energy- and LU-sector-based CO2 emissions. All \nscenarios in this study were derived with middle-of-the-road assump-\ntions on socio-economic drivers (Shared Socioeconomic Pathway 2) \nand meet a global CO2 emissions budget of 1,000\u2009GtCO2 from 2018 to \n2100, allowing for a temporary overshoot. This budget is derived by \nsubtracting 100\u2009GtCO2 emissions due to Earth system feedback from \nthe remaining carbon budget of 1,170 given in Rogelj et al.40 (67th per-\ncentile for the 2\u2009\u00b0C target), arriving at 1,070\u2009GtCO2. As safety margin, \nthis value is rounded down to 1,000\u2009GtCO2.\nCarbon prices\nIn the UCP scenario, all types of GHG emissions from the energy and LU \nsectors are charged with a uniform carbon equivalent price PGHG(t, r)  \nin US$ per tCO2 that increases with time t. Prices can differ between \nmodelling regions r before 2050 for reasons of inter-regional equity, \nbut they will eventually converge to globally harmonized prices by \n205038. In the LU sector, the GHG price incentivizes avoiding emissions \nfrom LUC and may trigger afforestation for CDR. A detailed descrip-\ntion of the techno\u2010economic data on reforestation and afforestation \ncan be found in Strefler et al.62. In the energy sector, the GHG price \namplifies the demand for bioenergy as it substitutes fossil fuels and \ngenerates carbon revenues from BECCS technologies. These revenues \nfor negative emissions in combination with biofuel revenues incentiv-\nize bioenergy even in the presence of high feedstock prices or high \nbioenergy tax levels.\nIn scenarios with a partial LU price (LUprice10\u201350%), the price on \nGHG emissions in the LU sector is reduced in every time step and every \nmodelling region to the corresponding fraction of the respective price \nlevel on energy-system-related GHG emissions (for example, to 10%).\nTo be consistent with the narrative of a largely unregulated agri-\nculture, forestry and other LU sector, we assumed that in scenarios \nwithout a price on CO2 emissions from LUC (noLUreg, protForest, \nprotForest90%, protPrimforest, protBH, protCPD, protFF, protLW, \nbioTax and partially LUtax-OECD scenarios), non-CO2 GHGs are also \nexempted from the GHG price in the LU sector. This has the side effect \nthat these scenarios also involve substantially higher non-CO2 GHG \nemissions from agricultural activities than scenarios with a carbon \nprice, in particular CH4 and N2O. As a result, radiative forcing levels \nand the resulting global mean temperature responses differ between \nscenarios, even though cumulative CO2 emissions coincide. How-\never, since agricultural CH4 emissions are not related to bioenergy \nproduction, and N2O emissions from grassy bioenergy production \nare negligible compared with LUC CO2 emissions (Extended Data  \nFig. 2), we omitted the effect of non-CO2 GHG emissions for assessing \nthe impact of bioenergy. Nevertheless, differences between scenarios \nin global mean temperature in 2100 as a result of varying LU-related CH4 \nand N2O emissions (derived with MAGICC v.6; ref. 63) are in the range \nof less than 0.2\u2009K (Supplementary Fig. 18).\nLand protection\nIn scenarios with explicit land-protection schemes (protForest, pro-\ntForest90%, protPrimforest, protBH, protCPD, protFF and protLW), \nwe removed the respective areas from the land pool that is potentially \navailable for any agricultural activities (Supplementary Figs. 3\u20139). In \nprotForest, all primary and secondary forests are protected (3,683\u2009Mha \nin total); in protForest90%, only 90% of the forests are protected in each \nsimulation unit (3,315\u2009Mha in total; Supplementary Fig. 10 shows the \nMAgPIE simulation units); and in protPrimforest, only primary forests \nare protected (1,339\u2009Mha in total). The other land-protection policies \naffect only some focus areas. In protBH, biodiversity hotspots are pro-\ntected (909\u2009Mha); in protCPD, centres of plant diversity are protected \n(651\u2009Mha); in protFF, frontier forests are protected (1,084\u2009Mha); and in \nprotLW, last of the wild areas are protected (3,635\u2009Mha)64.\nAdditionally, in all scenarios, specific land areas are protected or \ndedicated for afforestation according to the Nationally Determined \nContributions targets of the nations that are participating in the Paris \nclimate agreement (Supplementary Fig. 11).\nCrop yields and technological change\nYields of agricultural production (including bioenergy crops) are \ngrowing endogenously, driven by investments into R&D triggering \ntechnological change. The yield rate improvements depend on R&D \nand infrastructure spending according to an elasticity derived from \nempirical estimates65 with the investment-to-yield ratio:\nIY (\u03c4r) = 1.9 \u00d7 10\n3\u03c42.4\nr\n(US$ ha\n\u22121)\nas a function of the LU intensity indicator \u03c4r (developed by Dietrich \net al.66) for each region r. In this way, yield improvements become more \nexpensive with higher LU intensity. For instance, in a region r1 with \n\u03c4r1 = 1, increasing yields by about 1% requires investment costs of \n1.9\u2009\u00d7\u2009103\u2009\u00d7\u200912.4\u2009\u00d7\u20090.01 (US$\u2009ha\u22121) =\u2009US$19\u2009ha\u22121, while in a region r2 with \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nalready high LU intensity \u03c4r2 = 2, increasing yields by about 1% requires \n19\u2009\u00d7\u200922.4(US$\u2009ha\u22121)\u2009=\u2009US$100\u2009ha\u22121. For more details, please refer to \nDietrich et al.65,66.\nYields in MAgPIE are initialized with yields from the hydrology \nand vegetation model LPJmL. The yield rates in LPJmL reflect values \nachievable on the basis of geophysical and hydrological conditions \nand plant physiology. The yields from LPJmL are recalibrated (typically \nreduced) to match regional cropland in 199567, setting the starting \npoint for the optimization in MAgPIE. During the optimization process, \nyields can increase over time as described above. Over the twenty-first \ncentury, R&D-driven improvements can lead to yield rates that exceed \nbase-year values. Since yields for modern herbaceous bioenergy crops \nat the commercial farm level may be lower than what has been found in \nsmall-scale trials68, we added sensitivity scenarios, in which the initial \nbioenergy yields are reduced to 75% or 50% compared with the default \nvalues (UCP_bioYield75 and UCP_bioYield50). Since future yield rates \nand the productivity of the R&D sector are subject to uncertainty, we \nfurther analysed a scenario in which marginal costs for yield-improving \ntechnological change are substantially increased (UCP_highcostTC). \nWithin the parametrization for the investment yield ratio IY(\u03c4r) for this \nscenario, we used the highest estimates for the parameters (2.3 instead \nof 1.9 for the pre-factor and 3.3 instead of 2.4 for the exponent/elastic-\nity). The higher exponent in particular increases the marginal costs \nfor technological change in regions where yields are already high. It is \nimportant to highlight that not only bioenergy yields but also yields \nof food crops and forage are increasing due to technological change. \nPasture yields are not linked to crop yield increases but to an exogenous \npasture management factor. Also, while R&D-induced technological \nchange is the sole driver of yield gains for a fixed physical environment \nand fixed crop type, it is not the only reason why average yields within \nan economic region vary over time. For instance, shifting production \nof a certain crop type to more fertile land increases average yields, \nwhile extending production to, for example, marginal land with lower \nyields reduces them.\nThe sensitivity analysis on scenarios with lower yields can be found \nin the Supplementary Information, section \u2018Sensitivity analysis: Pes-\nsimistic yield assumptions\u2019.\nFood demand and food prices\nFood demand in MAgPIE depends on the income of households. Prices \ninfluence food demand only via the income effect of the price shock, \nwith increasing prices reducing the real income of households. Due \nto the impact of food prices on the real income of households, higher \nfood prices may lead to small reductions in demand. How food demand \ndepends on real income (modelled via demand elasticities of income) \nis described in Bodirsky et al.69 and Bodirsky et al.70. However, the effect \nis very small, and the differences between food demand trajectories \nare almost negligible in the scenarios (Supplementary Fig. 17). This is \nthus equivalent to the assumption of a food-first policy. In particular, \nthe demand for livestock products is not affected by the different \npolicy settings. The potential (unintended) effect that bioenergy \nreduces emissions indirectly due to changes in consumption71 is thus \nnot present in this study. Therefore, the derived EFs only account \nfor changes in livestock production systems rather than changes in \nmeat demand.\nBioenergy tax\nAs explained above, the default policy assumption regarding the pric-\ning of emissions is a uniform carbon price on both energy- and LU-based \nGHG emissions. Emissions related to bioenergy production are thus \nalready penalized directly within the LU sector, so the energy system by \ndefault treats bioenergy as a carbon-neutral energy carrier. In the sce-\nnarios with only energy sector policies (bioTax), we assigned an ex-ante \nEF (EFex-ante) to bioenergy that should reflect potential bioenergy-related \nGHG emissions. EFex-ante represents emissions on a global average and is \nequal for each economic region r and time step t. It directly transforms \ninto a bioenergy tax Tbio(t,r) via the price on GHGs PGHG(t,r):\nTbio (t, r) = EFex-ante \u00d7 PGHG (t, r) (US$ GJ\n\u22121\nPE)\nwhich is applied to every unit of DM biomass\u2014that is, at the level of PE. \nSince the literature and the results of the present study indicate that \nspecific emissions attributed to a unit of bioenergy are highly uncertain \neven on a global average, we explored the effect of different values \nof EFex-ante ranging from 10 to 50\u2009kg\u2009CO2\u2009GJPE\n\u22121, which translates to 24 \nto 122\u2009kg\u2009CO2\u2009GJbiofuel\n\u22121 (41% energy conversion efficiency). It is worth \nnoting that EFex-ante is in general not equal to the actual emissions that \nare eventually attributed to bioenergy, which are derived ex-post from \nour scenarios (EFex-post).\nIn most other publications applying the REMIND model, bioenergy \nis actually charged with a \u2018sustainability tax\u2019 that reduces the demand \nfor bioenergy irrespective of the policy design to reflect uncovered \nexternalities, such as unsustainable water usage, food price increase, \nthe loss of biodiversity and nitrogen losses to the environment38. In \nthe present study, however, we deactivated this tax, since we wanted \nto assess the impact of bioenergy given a certain policy assumption in \nan otherwise uncontrolled market.\nEx-post EF\nDue to ILUC induced by bioenergy production, it is impossible to dis-\nentangle LUC CO2 emissions related to bioenergy production from \nLUC emissions that result from other agricultural activities such as an \nexpansion of cropland or pasture. For each policy setting p, we defined \nand derived metrics with different time aggregation methods to com-\npare the model results. The time aggregation is crucial for the metric \nbecause cropland expansion causes immediate emissions to increase \nbioenergy production afterwards. To evaluate the effect of bioenergy \nproduction on LU emissions, we derived EFex-post by comparing two sce-\nnarios for each p\u2014first, the actual policy scenarios with the availability \nof purpose-grown bioenergy (bioOn) and, second, the counterfactual \nscenario, in which purpose-grown lignocellulosic bioenergy produc-\ntion is not allowed (bioOff). A similar approach has been applied in, for \nexample, Daioglou et al.22 and Pehl et al.72 Using the full scenario output \nas a basis considers feedback effects from reduced afforestation in the \npresence of large-scale bioenergy production (see Supplementary  \nFig. 11 for afforestation potentials in the counterfactual scenarios). \nUsing a counterfactual scenario also ensures that the total additional \nLUC emissions are correctly attributed to total bioenergy without mak-\ning the accounting error that occurs when counterfactual regrowth \nof natural vegetation is ignored in partial analysis (identified by  \nHaberl et al.73). Although the scenario results comprise full information \nabout changes in emissions and bioenergy production, choices about \nthe aggregation over time have substantial influence on the metric.\nFirst, the 80-year average EF comprises all LUC CO2 emissions \nattributed to bioenergy production:\nEF\nav80(p) =\nE80\nbioOn(p) \u2212E80\nbioOff(p)\nB80\nbioOn(p)\n(kg CO2 GJ\n\u22121)\nwhere E80\nbioOn and E80\nbioOff denote the global net cumulative LUC emis-\nsions over the period 2020\u20132100 for the scenario with bioenergy on \nand off, respectively. B80\nbioOn represents the global cumulative amount \nof purpose-grown lignocellulosic biomass produced globally over \nthe same period. The 80-year period equals the period for the carbon \nbudget. Due to the asynchronicity of emissions and bioenergy har-\nvest, a shorter time horizon would generally lead to higher EFs during \nperiods of bioenergy production expansion.\nSecond, we defined the marginal 30-year EF at time t, which relates \nadditional emissions in t to increments of bioenergy production in t \nthat are assumed constant for the coming 30 years, as\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nEF\nmar30(p, t) = EbioOn(p, t) \u2212EbioOff(p, t)\n30 \u00d7 Badd\nbioOn(p, t)\n(kg CO2 GJ\n\u22121)\nwith EbioOn(p,t) and EbioOff(p,t) being the total LUC emissions in the year \nt for the scenario policy assumption p and its respective counterfactual \nscenario without bioenergy. Badd\nbioOn(p, t) is the bioenergy production \nincrement in year t\u2014that is, the difference in total production from the \nprevious year. The EF is thus based on two assumptions. First, we \nassume that the difference in emissions from the counterfactual sce-\nnario in year t is fully attributed to the bioenergy production increment \nin that same year. Second, we assume that bioenergy plantations initi-\nated in the given year will be constantly producing the same amount \nof bioenergy over the next 30 years on the additional cropland.\nThe average EF for the 80-year time horizon matches the time \nframe of the carbon budget. Due to the flexibility of the model to allo-\ncate bioenergy use and emissions optimally over time, bioenergy pro-\nduction is maximized towards the end of the evaluation period, while \nemissions peak much earlier. Deriving an average 30-year EF would \nthus be meaningful only in scenarios that meet an emissions budget \nby 2050. To assess the value of bioenergy in a shorter time frame, given \nour assumptions on the climate target formulation, we thus use the \nmarginal EF, which only depends on bioenergy production increment \nand bioenergy-induced emission in a given year.\nThird, since EF\nmar30 (p, t) varies over time, we derived an averaged \nmarginal EF:\nEF\nav.mar30(p) =\n2070\n\u2211\nt=2025\nEF\nmar30 (p, t) \u00d7B\nadd\nbioOn(p, t)\n\u2211\n2070\nt=2025Badd\nbioOn(p, t)\n(kg CO2 GJ\n\u22121)\napplying the bioenergy production increment addition as weight to \nthe 30-year EF of a given year. Years with the highest increment of \nbioenergy production thus receive the highest weight. Since in the \nscenarios the climate target is met in 2100, due to the 30-year evalu-\nation period only EFs until 2070 are utilized to calculate the average \nEF. Also, in some scenarios, bioenergy production decreases again \nafter 2070\u2014calculating EFmar30 is thus not meaningful in these years. \nFor two scenarios with very low bioenergy production (bioTax40 \nand bioTax40_noTrd), bioenergy production decreases after 2060, \nso EFav.mar30 here only covers the years until 2060 (see Supplementary \nFigs. 1 and 2 for time series of emissions, bioenergy production incre-\nment and EFs).\nEFex-post is usually expressed in terms of CO2 emissions per unit of \nbiofuel, to make it comparable with fossil fuels. Since the thermochemi-\ncal conversion to liquid fuels is subject to substantial conversion losses \n(the energy conversion efficiency for second-generation biofuels \n(Fischer-Tropsch diesel) is only 41%), the EF per unit of PE biomass is \nsmaller, and other energy carriers derived from biomass, in particular \nelectricity or hydrogen, exhibit different EFs due to different energy \nconversion efficiencies (Supplementary Table 1).\nEFs were also evaluated in a spatially disaggregated manner. For \nour study, the LU model MAgPIE was applied using 1,000 distinct \nsimulation units, rMAgPIE, revealing individual patterns of agricultural \nactivities. Each simulation unit represents a cluster of aggregated \n0.5-degree-resolution grid cells with similar properties44,74 (Supple-\nmentary Fig. 10), and for each of them, an 80-year average EF can be \ncalculated individually:\nEF\nav80\nloc\n(p, rMAgPIE) =\nE80\nbioOn (p, rMAgPIE) \u2212E80\nbioOff (p, rMAgPIE)\nB80\nbioOn (p, rMAgPIE)\n(kg CO2 GJ\n\u22121)\nThere are clusters of grid cells without bioenergy production \n(EF\nav80\nloc\n= \u221e) and others for which the difference in emissions from the \ncounterfactual scenario is zero or even marginally negative\u2014that is, a \nsimulation unit rMAgPIE with equal or less emissions than in the scenario \nwithout bioenergy. Here the EF\nav80\nloc  is set to zero.\nSince the EFs are given as the ratio of emissions and bioenergy \nproduction, there is no information on the total volume of each of these \nquantities in the different areas in Fig. 2a. For selected scenarios, the \nspatial allocation of LUC emissions and bioenergy production quanti-\nties is depicted in the section \u2018Spatial land-use characteristics\u2019 of the \nSupplementary Information.\nUsing a counterfactual scenario to derive the EF can lead to a situ-\nation in which additional LUC emissions to this counterfactual scenario \nare rather small or even zero, while baseline LUC emissions from the \ncounterfactual scenario (in the same simulation unit) are already sub-\nstantial. Bioenergy production is then associated with a relatively small \nEF\nav80\nloc , even though the actually occurring emissions are large. However, \nsince these emissions also emerge in the baseline/counterfactual sce-\nnario, in this study they are not attributed to bioenergy. Such areas can \nbe interpreted as \u2018lost\u2019 territories that will suffer LUC irrespective of \nthe bioenergy production, given a distinct policy setting.\nThe 80-year BECCS efficiency factor\nWe defined the efficiency of the CDR potential of BECCS as\n\u03b7BECCS = (1 \u2212\nE80\nbioOn \u2212E80\nbioOff\nCDR\n80\nBECCS,bioOn\n) \u00d7 100%,\nwhere CDR\n80\nBECCS,bioOn  are the total cumulative negative emissions  \nassociated with BECCS from purpose-grown biomass. A scenario with-\nout bioenergy-induced LUC would thus imply an efficiency of 100%, \nwhile in a scenario in which bioenergy-related emissions are equal to \nthe CDR saving via BECCS, the efficiency is 0%. The BECCS efficiency \nfactor shown in Extended Data Fig. 5 uses cumulative emissions and \nsavings for the 30-year time horizon until 2050 instead.\nSince bioenergy from residues is allowed in the counterfactual \nscenarios, we excluded the BECCS emission savings related to residues \nfrom the calculation of \u03b7BECCS. This efficiency factor is derived to relate \nbioenergy-induced LUC emissions to the CDR potential of BECCS. It \ndoes not, however, cover other benefits of bioenergy to the energy \nsystem, particularly the benefits of substituting biofuels for fossil \nfuels. However, quantifying the avoided emissions attributed to the \nbioenergy part of BECCS is inherently ambiguous, as there are also \nother substitution options such as direct electrification in the inte-\ngrated systems perspective. Due to the fixed quantity target on total \nemissions, more or less emissions from the LU sector (including the \nCCS part) will be\u2014by definition of the target\u2014balanced with changes \nin fossil fuel use. Consequently, the partial effect of the substitution \neffect throughout the overall fossil fuel sector is not independent \nof the net emission changes caused by LU and carbon removal of the \nsame quantity of bioenergy. In addition, bioenergy-related emissions \ndo not cover all negative effects, as described in the paragraph on the \nbioenergy tax.\nData availability\nThe model runs and scenario data for this study are archived at Zenodo \nunder a CC-BY-4.0 license75.\nCode availability\nREMIND is open source and available on GitHub. The model version \nused in this study is 2.1.2, which can be downloaded at https://github.\ncom/remindmodel/remind/releases/tag/v2.1.2. MAgPIE is open source \nand available on GitHub. The model version used in this study is 4.2.1, \nwhich can be downloaded at https://github.com/magpiemodel/mag-\npie/releases/tag/v4.2.1. Documentation can be found at https://rse.\npik-potsdam.de/doc/magpie/4.2.1/.\nReferences\n62.\t Strefler, J. et al. Carbon dioxide removal technologies are not \nborn equal. Environ. Res. Lett. 16, 074021 (2021).\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\n63.\t Meinshausen, M., Wigley, T. M. L. & Raper, S. C. B. Emulating \natmosphere\u2013ocean and carbon cycle models with a simpler \nmodel, MAGICC6\u2014part 2: applications. Atmos. Chem. Phys. 11, \n1457\u20131471 (2011).\n64.\t Kreidenweis, U. et al. Pasture intensification is insufficient \nto relieve pressure on conservation priority areas in open \nagricultural markets. Glob. Change Biol. 24, 3199\u20133213 (2018).\n65.\t Dietrich, J. P., Schmitz, C., Lotze-Campen, H., Popp, A. & M\u00fcller, C. \nForecasting technological change in agriculture\u2014an endogenous \nimplementation in a global land use model. Technol. Forecast. \nSoc. Change 81, 236\u2013249 (2014).\n66.\t Dietrich, J. P. et al. Measuring agricultural land-use intensity\u2014a \nglobal analysis using a model-assisted approach. Ecol. Model. \n232, 109\u2013118 (2012).\n67.\t Humpen\u00f6der, F. et al. Investigating afforestation and bioenergy \nCCS as climate change mitigation strategies. Environ. Res. Lett. 9, \n064029 (2014).\n68.\t Searle, S. Y. & Malins, C. J. Will energy crop yields meet \nexpectations? Biomass Bioenergy 65, 3\u201312 (2014).\n69.\t Bodirsky, B. L. et al. Global food demand scenarios for the 21st \ncentury. PLoS ONE 10, e0139201 (2015).\n70.\t Bodirsky, B. L. et al. The ongoing nutrition transition thwarts \nlong-term targets for food security, public health and \nenvironmental protection. Sci. Rep. 10, 19778 (2020).\n71.\t Searchinger, T., Edwards, R., Mulligan, D., Heimlich, R. & Plevin, R. \nDo biofuel policies seek to cut emissions by cutting food? Science \n347, 1420\u20131422 (2015).\n72.\t Pehl, M. et al. Understanding future emissions from low-carbon \npower systems by integration of life-cycle assessment and \nintegrated energy modelling. Nat. Energy 2, 939\u2013945 (2017).\n73.\t Haberl, H. et al. Correcting a fundamental error in greenhouse \ngas accounting related to bioenergy. Energy Policy 45,  \n18\u201323 (2012).\n74.\t Dietrich, J. P., Popp, A. & Lotze-Campen, H. Reducing the loss of \ninformation and gaining accuracy with clustering methods in a \nglobal land-use model. Ecol. Model. 263, 233\u2013243 (2013).\n75.\t Merfort, L. et al. Model run and scenario data for study \n\u2018Bioenergy-induced land-use change emissions with sectorally \nfragmented policies\u2019. Zenodo https://doi.org/10.5281/zenodo. \n7799031 (2023).\n76.\t IPCC Climate Change 2014: Synthesis Report (eds Core Writing \nTeam, Pachauri, R. K. & Meyer L. A.) (IPCC, 2014).\nAcknowledgements\nThe research leading to these results has received funding by the \nDIPOL from the German Federal Ministry of Education and Research \n(BMBF) under grant number 01LA1809A (L.M., N.B. and J.S.). This work \nwas supported by the NAVIGATE project funded by the European \nUnion\u2019s Horizon 2020 research and innovation programme under \ngrant number 821124 (N.B., F.H. and J.S.), by the RESCUE project from \nthe European Union\u2019s Horizon Europe programme under grant number \n101056939 (L.M.) and by the ARIADNE project from BMBF under grant \nnumber 03SFK5A (D.K. and G.L.).\nAuthor contributions\nL.M. performed the model experiments, analysed the scenarios, \nproduced the figures and led the writing of the manuscript. L.M., \nN.B., F.H., G.L. and E.K. designed the study, the scenarios and the \nanalysis. L.M., N.B., F.H., D.K., J.S., A.P., G.L. and E.K. contributed to the \ndevelopment of the models, the presented ideas and the text.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at  \nhttps://doi.org/10.1038/s41558-023-01697-2.\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s41558-023-01697-2.\nCorrespondence and requests for materials should be addressed  \nto Leon Merfort.\nPeer review information Nature Climate Change thanks the \nanonymous reviewers for their contribution to the peer review of  \nthis work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nExtended Data Fig. 1 | Primary energy biomass distribution. The area plots \nshow the share of biomass that is converted into different secondary energy \ncarriers. The main conversion process is the biomass to liquids route via \ngasification and Fischer-Tropsch synthesis of lignocellulosic biomass (in our \nscenarios, all liquids derived from lignocellulosic biomass originate from that \nprocess). The biomass shown here originates from both purpose-grown energy \ncrops and residues (see Extended Data Fig. 10, showing shares of different \nfeedstocks). Please note that some conversion processes have a second energy \ncarrier type as co-product. In particular in the biomass-based Fischer-Tropsch \nprocess electricity is co-produced, which makes up almost all of the electricity \nthat is derived from lignocellulosic biomass.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nExtended Data Fig. 2 | Bioenergy emission factors derived with different \nmetrics. In addition to EFs in Fig. 1a we here explicitly show marginal 30-year EFs \nfor the years 2035 and 2050 in green, 80-year average N2O EFs in bright red, as \nwell as the averaged marginal 20-year EF in orange. Please note that by definition \nthe 20-year marginal EF is simply the 30-year marginal EF multiplied with 1.5, \nsince the accounting period for bioenergy production is divided by that factor. \nThe boxplot shows the temporal variation of EFmar30(t) for years t between 2025 \nand 2070 (sample size of n\u2009=\u20099; bioTax40, bioTax40_noTrd are exceptions with \nn\u2009=\u20098, since the year 2070 is excluded, see \u2018Emission factors over time\u2019 in the SI). \nThe minima and maxima of the box confine the inter-quartile range, the whiskers \nrepresent the 1st and 4th quartile, and the center states the median value. N2O EFs \nare converted to CO2-eq using a global warming potential of 26576 and are \ncalculated equivalently to the 80-year average EF for CO2. It can be seen that the \ndifferent policy assumptions have a rather small effect on the N2O EF with values \nbetween 4.6 and 7.2\u2009kg CO2/GJbiofuel. While the N2O EF can mostly be neglected \nrelative to the high CO2 EF in scenarios where LU policies are missing, for the UCP \nscenario cumulative N2O emissions until 2100 in carbon equivalents are roughly \n40% of the cumulative CO2 emissions, since the EF for CO2 is relatively small.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nExtended Data Fig. 3 | Land cover change. Shown are changes in the land \ncover with respect to 1995 of the four main land types represented in MAgPIE, \n\u2018Forest\u2019, \u2018Other Land\u2019, \u2018Pasture and Rangelands\u2019 and \u2018Cropland\u2019. The dark shading \nrepresents the land cover change that is attributed to bioenergy production \n(\u2018bioen.-induced\u2019) and the light shading represents the land cover change that \nis observed in the respective counterfactual scenario (\u2018cf. baseline\u2019), that is, \nthe fraction that is not attributed to bioenergy. Areas classified as \u2018Other Land\u2019 \ncomprise non-forest natural vegetation (such as savannahs or shrubland), \nabandoned agricultural land and deserts. Some of these areas also store \nsubstantial amounts of carbon.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nExtended Data Fig. 4 | Allocation of LUC CO2 emissions and bioenergy production by land-use characteristic. In Fig. 2b only a selection of scenarios was \npresented. Here we show the allocation of emissions and bioenergy production for all policy assumptions. For a description, please refer to Fig. 2 in the main text.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nExtended Data Fig. 5 | The BECCS efficiency factor in 2050. \u03b7BECCS is an \nindicator of how much of the sequestered carbon is effectively removed from the \natmosphere if bioenergy-induced LUC emissions are subtracted. A negative \nefficiency indicates that bioenergy-induced LUC emissions exceed CDR savings \nby BECCS in the year 2050. For instance, \u03b7BECCS = \u2212200% implies that emissions \nare 200% higher than savings, that is, three times as high. For the UCP policy \nsetting, emissions and savings are just equal in 2050.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nExtended Data Fig. 6 | Share of biomass used with CCS and total production \nof primary energy biomass. The share of biomass that is used in combination \nwith CCS, shown on the left-hand side, is given as a fraction of total primary \nenergy biomass production, that is, including traditionally used biomass and \n1st generation biofuels (which cannot be combined with CCS) and residues. This \ntotal primary energy production is shown on the right-hand side for comparison. \nFor clarity we highlight only a few scenarios with colors, grey lines refer to the \nother scenarios.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nExtended Data Fig. 7 | Liquids by primary energy carrier. Biomass, coal and oil can be converted to liquid fuels and biomass and coal have an option to add CCS \n(E-fuels are not considered in this study). Shown are shares of all liquids that are produced from these primary energy carriers, either with or without CCS included. \nShares are given in the year 2050 and 2100 for different policy settings.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nExtended Data Fig. 8 | Final energy mix. Total final energy is compound of different energy carriers. The shares are shown for different policy settings in the years \n2050 and 2100.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nExtended Data Fig. 9 | Global agricultural price index relative to 2010. \nBlack dots denote the agricultural price index of a given scenario. Orange \nbars represent the index of the respective counterfactual scenario without \nbioenergy production. The green bars indicate the difference between \nscenarios with and without bioenergy, that is, the fraction of the average price \nchange for agricultural goods induced by allowing for bioenergy production. \nThe agricultural price index is defined as the change of averaged prices of all \nagricultural commodities relative to the year 2010 (which has the value 100).  \nAdd\u00edtionally to the scenarios shown in the main text, we here show the  \nadditional scenario from the sensitivity analysis on pessimistic crop yields  \n(UCP_bioYield50), see \u2018Sensitivity analysis: Pessimistic yield assumptions\u2019 in \nthe SI. Please refer to \u2018Food demand and prices\u2019 in the SI for more details on the \nimpact of policy assumptions on the food prices.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01697-2\nExtended Data Fig. 10 | Bioenergy feedstocks to the energy system. The area \nplots show the share of bioenergy from different sources. While today residues \nare almost exclusively used traditionally, that is, using the solid biomass for \ncooking and heating mostly in developing countries, in our scenarios this shifts \nwithin the next 30 years towards modern usage, mostly in form of creating \nsecond generation bio liquids from it (see also Extended Data Fig. 1).\n\n\n    Scientific Research Findings:", "answer": "With an average emission factor (EF) of 92 kgCO2GJ\u20131, we find that the production of modern biofuels, if averaged over a 30-year period, causes land-use-change emissions that are higher than those from burning fossil diesel. If policymakers tax bioenergy according to these average expected emissions, that is, apply a similar carbon price to a litre of biofuels as to a litre of diesel, the total future bioenergy-induced emissions decrease, as the demand is reduced. However, we show that such a policy cannot bring down the high average emissions that are attributed to biofuels. Only strict and globally comprehensive protection of natural land will reduce the EF and hence, only then, will those biofuels that replace fossil fuels effectively reduce CO2emissions.", "id": 47}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 13 | June 2023 | 523\u2013531\n523\nnature climate change\nhttps://doi.org/10.1038/s41558-023-01679-4\nArticle\nParticipating in a climate prediction market \nincreases concern about global warming\nMoran Cerf\u2009\n\u200a\u20091\u2009\n, Sandra C. Matz\u2009\n\u200a\u20091\u2009\n & Malcolm A. MacIver\u2009\n\u200a\u20092,3,4,5,6\u2009\nModifying attitudes and behaviours related to climate change is difficult. \nAttempts to offer information, appeal to values and norms or enact policies \nhave shown limited success. Here we examine whether participation in \na climate prediction market can shift attitudes by having the market act \nas a non-partisan adjudicator and by prompting participants to put their \n\u2018money where their mouth is\u2019. Across two field studies, we show that betting \non climate events alters: (1) participants\u2019 concern about climate change, (2) \nsupport for remedial climate action and (3) knowledge about climate issues. \nWhile the effects were dependent on participants\u2019 betting performance in \nStudy 1, they were independent of betting outcomes in Study 2. Overall, our \nfindings suggest that climate prediction markets could offer a promising \npath to changing people\u2019s climate-related attitudes and behaviour.\nThe combined forces of social media and rise of populism have ampli-\nfied the politicization of knowledge. What is considered true often \ndepends on group membership rather than scientific evidence and \nfacts1\u20136. This politicization is seen in numerous topics, including climate \nchange. Overwhelming scientific evidence suggests that climate change \nis occurring7, is caused by human activity8 and is likely to result in dire \nconsequences9,10. Nonetheless, actions of governments around the \nworld lag behind what climate scientists say is needed. In some cases, \nthis inaction is related to a lack of concern about climate change. For \nexample, in the United States, surveys show that over a third of the \npopulation believes that the seriousness of global warming is exag-\ngerated11, and more than half the population disagrees with the claim \nthat climate change is caused by humans12.\nRaising concern about climate change and support for remedial \naction at the individual and collective level is challenging for numer-\nous reasons. First, it is difficult to attribute a specific climate-related \nincident to a single cause. Second, remedial actions taken by one indi-\nvidual or collective often do not yield visible outcomes. Third, the \ncost of action is immediate whereas the benefits are distributed over \nlong time horizons13. Specifically, while climate change will adversely \nimpact future generations, for most people there is no immediate cost \nto rejecting its occurrence on ideological grounds14. Compounded by \nthe brain\u2019s challenges in thinking about temporally or spatially distant \nevents15\u201317, these factors make it difficult to change sceptics\u2019 views on \nthe topic and garner support for corrective action.\nHowever, acknowledging the role of erroneous beliefs that have no \nimmediate cost offers a potential pathway to shifting people\u2019s climate \nchange-related attitudes and behaviours: devise a mechanism to make \nmaintaining false beliefs costly in the near term. Research shows that \npeople often behave in ways that contradict their stated beliefs when \nmoney is on the line18. For example, climate sceptics publicly deny \nglobal warming but do not invest in geographic regions that will prob-\nably suffer from a rise in sea levels19. Building on this, we suggest using \nclimate prediction markets to shift attitudes towards the scientific \nconsensus by increasing the cost of maintaining false beliefs. Simply, \nwe provide a financial reward (/penalty) for correct (/incorrect) predic-\ntions about soon-to-occur events that are impacted by global warming \nusing a prediction market.\nTraditionally, prediction markets have been implemented to \ncrowd-source estimates about uncertain events in the future20. Those \nmarkets have been shown to accurately predict the outcomes of elec-\ntions21, reproducibility of scientific findings22, spread of disease23 or \naggregation of group choices24. In the context of climate change, pre-\ndiction markets have been suggested as a tool for aggregating views \non policies25 and as a way to provide credible signals about climate \nscience26\u201328. However, there is no empirical evidence supporting the \nReceived: 8 September 2020\nAccepted: 21 April 2023\nPublished online: 8 June 2023\n Check for updates\n1Columbia Business School, Columbia University, New York, NY, USA. 2McCormick School of Engineering, Northwestern University, Evanston, IL, USA. \n3Department of Mechanical Engineering, Northwestern University, Evanston, IL, USA. 4Department of Biomedical Engineering, Northwestern University, \nEvanston, IL, USA. 5Department of Neurobiology, Northwestern University, Evanston, IL, USA. 6Department of Computer Science, Northwestern \nUniversity, Evanston, IL, USA. \n\u2009e-mail: nature@morancerf.com; sm4409@gsb.columbia.edu; maciver@northwestern.edu\n\nNature Climate Change | Volume 13 | June 2023 | 523\u2013531\n524\nArticle\nhttps://doi.org/10.1038/s41558-023-01679-4\ntake a position on topics. Study 1 had 4,737 interactions (stock offers, \ntrades and so on) with an average of 9.5 contracts per day.\nConcern about global warming increases, \nconditional on winning\nWe first tested whether engaging in the climate prediction market had \nan impact on how concerned participants were about climate change. \nSpecifically, we ran linear regressions to predict climate concern in the \npost-survey from experimental condition (0\u2009=\u2009control, 1\u2009=\u2009treatment), \ncontrolling for participants\u2019 concern in the pre-survey. Contrary to our \nexpectations, participating in the climate prediction market did not \nlead to an overall increase in climate concern compared to the control \ngroup (B\u2009=\u2009\u22120.005, SE(B)\u2009=\u20090.015, \u03b2\u2009=\u2009\u22120.001, t\u2009=\u2009\u22120.04, p\u2009=\u2009.976; Fig. 2; \nall results hold when using the difference between pre- and post-survey \nas outcome). B, Unstandardized regression coefficient. SE, Standard \nerror. \u03b2, Standardized regression coefficient. t, statistical coefficients \nof test; p, statistical coefficient of significance.\nHowever, exploratory analyses of the treatment condition \nrevealed an effect conditional on participants\u2019 performance in the \nbetting market. Specifically, we used the robust MM-type estimator30 \nto regress the difference in concern between post- and pre-surveys \n(higher values indicate a shift towards more concern about climate \nchange) onto two indicators of performance: (1) the number of bets \nwon and (2) total earnings. Here betting outcomes significantly and \nconsistently predicted the change in concern (number of bets won: \nB\u2009=\u20090.007, SE(B)\u2009=\u20090.003, t\u2009=\u20092.44, p\u2009=\u20090.017; total earnings: B\u2009=\u20090.01, \nSE(B)\u2009=\u20090.005, t\u2009=\u20092.37, p\u2009=\u20090.021). That is, participants\u2019 concerns about \nclimate change increased if they were accurate in their predictions.\nFinally, we tested whether the impact of the treatment varied \nbetween believers and sceptics. Using the robust MM-type estimator to \nregress the difference in concern between post- and pre-surveys on the \nbinary believer/sceptic variable, we saw a marginally significant effect \n(B\u2009=\u20090.14, SE(B)\u2009=\u20090.08, t\u2009=\u20091.71, p\u2009=\u20090.089) suggesting that the treatment \nwas more effective for believers than sceptics. The moderating effects \nof performance on concern were found to be equally strong for both \nbelievers and sceptics (B\u2009=\u20090.001, SE(B)\u2009=\u20090.01, t\u2009=\u20090.11, p\u2009=\u20090.911).\nTo further explore participants\u2019 engagement with the climate \nprediction market, we tested for differences between believers and \nsceptics in betting outcomes (bets won and total earnings; Supplemen-\ntary Fig. 2) and behaviour (confidence, defined as the distance from \nthe neutral US$0.50/US$0.50 value). Despite believers being among \nthe highest earners in our market (top 11% of earners), both groups did \nnot significantly differ in the number of bets won (B\u2009=\u20090.70, SE(B)\u2009=\u20092.01, \np\u2009=\u20090.728) or the total earnings (B\u2009=\u20091.44, SE(B)\u2009=\u20091.33, p\u2009=\u20090.282). How-\never, the bets of believers indicated marginally higher levels of confi-\ndence (B\u2009=\u20090.48, SE(B)\u2009=\u20090.24, p\u2009=\u20090.053; Supplementary Fig. 6).\nAltogether, Study 1 offers suggestive evidence that prediction \nmarkets can increase concern for climate change under certain condi-\ntions (that is, successful betting). Despite the promising results, Study \n1 also suffers from a number of limitations. First, by virtue of its reliance \non a real-world market resembling the one seen in public exchanges \n(\u2018two-sided\u2019), it was hard to isolate the treatment effects (that is, partici-\npants may have placed bets that did not turn into contracts). Second, \nthe decision to target only individuals with polarized positions made \nobtaining a shift in concern challenging because believers are already \nat a climate concern ceiling, while sceptics are hardest to shift. Third, \nthe size of our participant pool made it impossible to detect small \neffects that are common in behaviour-change research. Fourth, the \nfact that we opted for a passive control group that did not engage in any \nmeaningful task during the prediction period prevented us from test-\ning whether the effect of successful betting on concern was uniquely \nrelated to climate predictions or the result of participants experiencing \npositive outcomes.\nStudy 2 overcomes these limitations by testing the effects in a \ncontrolled experimental setting, with an active control group that \nnotion that betting on climate-related events can shift people\u2019s: (1) \nconcerns about the consequences of climate change, (2) support for \nremedial action at the individual/collective level and (3) knowledge \nabout climate topics.\nClimate prediction markets\nWe introduce climate prediction markets as a novel intervention and \nreport experimental findings on how participating in the markets influ-\nences people\u2019s concern about climate change, support for action and \nclimate knowledge. Our market offers individuals the opportunity \nto bet on future outcomes (that is, \u2018the average temperature in the \nNorthern Hemisphere in the coming month will be higher than that in \nthe equivalent time window over the last decade\u2019) and earn money if \ntheir predictions are proven right.\nWe implemented two different prediction markets across two field \nstudies. In both studies participants engaged in a market where they \ntook positions on future climate events and earned money based on \ntheir prediction accuracy.\nBetting topics were set by the experimenters and were released \nintermittently (between 1 and 3 days apart in Study 1 and daily in  \nStudy 2). The bets reflected both events that dominated the news (that \nis, California wildfires, extreme heat waves) and events that were less \nsalient to the average participant (that is, Antarctic Sea ice extent, \nchange in the Air Quality Index). All bets had a settle date/time and an \nunambiguous source for determining the outcome. We term a particu-\nlar prediction a \u2018bet\u2019. For each bet, participants could decide whether \nthey wanted to make a bet, which position to take (Yes/No) and how \nmuch money to wager.\nWe surveyed participants before (\u2018pre-survey\u2019) and after \n(\u2018post-survey\u2019) the period during which they engaged in the predic-\ntion market (Fig. 1a). In both studies, we compared participants who \nengage in the climate prediction market to a control condition (Study \n1: passive control group, Study 2: active control group that participated \nin a sports and entertainment prediction market). Comparisons with \nthe control group allow us to account for changes that might occur \nnaturally over time (for example, natural variation in the salience of \nclimate disasters that are known to impact people\u2019s attitudes about \nclimate change29).\nStudy 1\nParticipants (n\u2009=\u2009143) were recruited online and screened for climate \nbeliefs and US nationality. Climate belief was defined as agreement \nwith the statement \u2018Global warming refers to the idea that the world\u2019s \naverage temperature has been increasing over the past 150 years and \nmay be increasing more. Do you think that global warming is happen-\ning?\u2019 (Supplementary Table 5 and Supplementary Fig. 4 provide demo-\ngraphic and climate concern breakdowns for the 70 climate believers \nand 73 sceptics included in the study). We screened for individuals \nwith polarized positions by selecting only individuals who answered \n\u2018Yes\u2019/\u2018No\u2019, skipping those who said \u2018Don\u2019t know\u2019. Participants completed \ntwo surveys, one before the beginning of the prediction period and \none at its conclusion. The surveys captured participants\u2019 concerns \nabout climate change, support for climate action, climate knowledge \nand variables such as demographics, political orientation and more \n(Methods and Supplementary Information). Participants within the \ngroups of believers and sceptics were randomly assigned to either the \ncontrol (n\u2009=\u200973) or treatment group (n\u2009=\u200970). Each participant in the \ntreatment group received US$20 to fund bets in the prediction market. \nDuring the prediction period, participants made bets on future events \n(Supplementary Methods show all bets). Because of the double-auction \nstructure of the market (if one participant bet US$0.55 that an event \nwould occur, it only becomes a contract if another participant bets \nUS$0.45 that the event will not occur; Supplementary Methods provide \ndetails on the betting mechanism), not all bets turned into contracts. \nWe analysed bet offers as reflections of participants\u2019 willingness to \n\nNature Climate Change | Volume 13 | June 2023 | 523\u2013531\n525\nArticle\nhttps://doi.org/10.1038/s41558-023-01679-4\nengaged in a non-climate prediction market, using a much larger \nsample size of people who are less extreme in their beliefs about \nclimate change.\nStudy 2\nParticipants (n\u2009=\u20091,005) were recruited online similar to Study 1. Of \nthe total participants, those (n\u2009=\u2009664) who wagered at least US$10 and \nplaced at least 15 bets were included in the analyses. As in Study 1, pre- \nand post-surveys measured participants\u2019 climate concern, support and \nknowledge (Methods). Between surveys, participants were randomly \nassigned to either a climate prediction market (n\u2009=\u2009356, treatment) or a \nsports and entertainment prediction market (n\u2009=\u2009308, control; Methods \nand Supplementary Information provide evidence that the randomi-\nzation was successful). Both prediction markets ran for a period of 35 \ndays during which one new bet was posted daily. Upon logging into the \nprediction market, participants saw an overview of their betting profile \n(amount won thus far, number of bets placed) and were informed about \nthe outcomes of previous bets. Participants then saw the daily bet  \n(Fig. 1c). Participants were asked to decide whether to bet, which posi-\ntion to take and how much money to wager. Each participant received \nUS$20 at the beginning of the study. Overall, participants placed 10,384 \nbets (15.6 bets per person).\nPre-survey\nDemographics \nClimate concerns\nClimate support\nClimate knowledge\na\nControl\nTreatment\nStudy 1: No predictions\nStudy 2: Sports/entertainment prediction market\nStudy 1,2: Climate prediction market\nPost-survey\nClimate concerns\nClimate support\nClimate knowledge\nb\nc\nPercent\n100\n0\nBet won by:*\nAverage amount bet (US$)\n5\n10\n15\n25\n20\nStudy 2\nTime (days)\nEarnings (US$)\n1,000\n500\n\u2013500\n0\n\u20131,000\n4\n8\n12\n16\n20\n24\n28\n32\n36\nStudy 1\nPredicted no\nPredicted yes\nAbstained\nDid not visit site\nAverage bet\nBelievers\nSceptics\nTime (days)\n0\n25\n50\n75\n30\n35\n1\n0.25\n0.50\n0.75\n0\n*\u2018Predicted yes\u2019 bets align with expectations of climate science\nCO2 level (ppm)\nGlobal carbon dioxide levels\nin Earth\u2019s atmosphere\n420\n410\nYear\n400\n390\n380\n2012\n2013\n2014\n2015\n2016\n2017\n2018\n2019\n2020\n2021\n2022\nFig. 1 | Experimental design. a, Participants first answered various questions \nabout their views on climate issues in a pre-survey (Supplementary Tables 2, 4, \n18 and 19 provide survey questions). Afterwards, participants were divided into \ntreatment (climate prediction market) and control (Study 1: no predictions, \nStudy 2: sports/entertainment prediction market) groups. b, Participants in the \nprediction markets made bets continuously (left: Study 1 earnings breakdown \nover time) or daily (right: Study 2; left y axis corresponds to the percent of \nparticipants taking yes/no/abstain positions on each daily bet, and the right \ny axis corresponds to the average daily bet wager). We denote above each bet \nthe prediction that ended up being accurate (note that we attempted to ensure \nthat \u2018Yes\u2019 bets would align with climate science). c, Visualization of the climate \npredictions market in Study 1 (on a dedicated website, titled C-Hedge; left) and \n2 (right). The wager in Study 1 ranged from US$0.50 to US$0.99 with a position \nof less than US$0.50 amounting to switching (that is, US$0.40 \u2018Yes\u2019 is US$0.60 \n\u2018No\u2019). The wager in Study 2 ranged from US$0.01\u2013US$1. Following the prediction \nperiod, both treatment and control participants completed a post-survey \naddressing climate issues and assessments of their overall experience.\n\nNature Climate Change | Volume 13 | June 2023 | 523\u2013531\n526\nArticle\nhttps://doi.org/10.1038/s41558-023-01679-4\nIncrease in concern about global warming\nWe tested whether engaging in the climate prediction market had an \nimpact on how concerned participants were about climate change, \nhow supportive they were of remedial action and how much they knew \nabout climate change. Specifically, we ran a series of linear regressions \nto predict climate concern, support and knowledge in the post-survey \nfrom category of experimental condition (0\u2009=\u2009control, 1\u2009=\u2009treatment) \ncontrolling for the respective concern, support and knowledge in the \npre-survey and including the socio-demographic variables to increase \nthe precision of the estimates (Supplementary Information \u2018Robust-\nness checks\u2019). The treatment group showed significantly higher lev-\nels of concern (B\u2009=\u20090.12, SE(B)\u2009=\u20090.045, \u03b2\u2009=\u20090.08, t\u2009=\u20092.69, p\u2009=\u20090.007;  \nFig. 3a), support for remedial action (B\u2009=\u20090.13, SE(B)\u2009=\u20090.058, \u03b2\u2009=\u20090.09, \nt\u2009=\u20092.19, p\u2009=\u20090.029; Fig. 3b), and knowledge (B\u2009=\u20091.58, SE(B)\u2009=\u20090.22, \n\u03b2\u2009=\u20090.52, t\u2009=\u20097.15, p\u2009<\u20090.001; Fig. 4) in the post-survey compared to \nthe control group.\nGiven that we asked the same question regarding concern and \nsupport in the pre- and post-survey, we could compare participants\u2019 \nscores to understand the underlying mechanisms of the effects. Partici-\npants in the treatment condition showed significantly higher levels of \nconcern in the post-survey than the pre-survey (t(355)\u2009=\u20092.23, p\u2009=\u20090.026; \nFig. 3) while controls did not (t(307)\u2009=\u2009\u22120.93, p\u2009=\u20090.353; paired t tests). \nSimilarly, participants in the treatment condition increased their sup-\nport for remedial action (t(355)\u2009=\u20092.89, p\u2009=\u20090.004) while controls did not \n(t(307)\u2009=\u20090.37, p\u2009=\u20090.712; paired t tests).\nIn addition to testing our main hypotheses, we conducted a series \nof exploratory analyses. First, we tested whether the treatment effect \nwas stronger in certain conditions (that is, as in Study 1, when partici-\npants were successful in their bets). While we did not observe signifi-\ncant interaction effects between the experimental condition and the \nbet winnings for climate concern (B\u2009=\u20090.009, SE(B)\u2009=\u20090.011, \u03b2\u2009=\u20090.04, \nt\u2009=\u20090.75, p\u2009=\u20090.455) or support (B\u2009=\u20090.007, SE(B)\u2009=\u20090.014, \u03b2\u2009=\u20090.04, t\u2009=\u20090.55, \np\u2009=\u20090.583), we found a significant moderation for climate knowledge \n(B\u2009=\u20090.182, SE(B)\u2009=\u20090.053, \u03b2\u2009=\u20090.41, t\u2009=\u20093.44, p\u2009<\u20090.001). Notably, we \nobserved a significant interaction between the treatment and politi-\ncal ideology. The treatment was more effective at increasing support \nfor remedial action among more conservative participants (B\u2009=\u20090.077, \nSE(B)\u2009=\u20090.036, \u03b2\u2009=\u20090.13, t\u2009=\u20092.14, p\u2009=\u20090.033). All treatment effects were \nindependent of initial climate concerns, suggesting that participants \nat all levels of climate concern were equally affected by their involve-\nment in the climate prediction market.\nIn line with findings on motivated reasoning1,6,31, we observed a \nmarginally significant relationship between political ideology and the \npercentage of bets that superficially align with climate change (r\u2009=\u2009\u22120.10, \nt\u2009=\u2009\u22121.90, p\u2009=\u20090.058). Testing for correlations between political ideol-\nogy and outcomes (number of bets won, r\u2009=\u2009\u22120.02, p\u2009=\u20090.526; and total \namount earned, r\u2009=\u2009\u22120.004, p\u2009=\u20090.921) or betting behaviours (total bets \nplaced, r\u2009=\u2009\u22120.01, p\u2009=\u20090.828; and total amount spent, r\u2009=\u20090.05, p\u2009=\u20090.192) \ndid not show any significant correlations (Supplementary Fig. 8). Accord-\ningly, political ideology did not influence participants\u2019 engagement with \nthe markets, confidence in their bets or prediction accuracy.\nImplications of climate prediction markets\nIn line with existing theoretical arguments about the power of climate \nprediction markets32, our findings from two field studies suggest that \nparticipating in such markets can influence people\u2019s attitudes towards \nclimate change. Specifically, we show that participants who bet on \nclimate-related events reported higher levels of concern about climate \nchange, showed higher levels of support for remedial climate action \nand had higher levels of knowledge on climate issues. While the positive \nimpact of our intervention on attitudes was conditional on betting suc-\ncess in Study 1, it was unrelated to earnings in Study 2. This discrepancy \nmight, in part, be explained by the fact that the participants in Study 1 \nwere recruited to be highly polarized in their views on climate change.\nThe effects of our intervention are small, with our experimen-\ntal condition explaining between 1% and 7% of the variance in the \npost-survey responses regarding concern, support and knowledge. \nHowever, we argue that our intervention offers a meaningful tool for \nbehaviour change. Prior work has suggested that when considered at \nscale, small effects can turn into highly impactful outcomes33. Further, \nour intervention results in positive attitude shifts across the entire \npolitical spectrum. Neither political ideology nor people\u2019s prior views \non climate change moderated the effect. The only exception to this lack \nof moderation by political ideology was the shift in climate support in \nStudy 2, where the intervention was stronger among more conserva-\ntive participants. The success of the intervention is promising given \nthat prior works have reported adverse reactive behaviour among cli-\nmate sceptics targeted with attempts to shift their climate views12,34\u201337.  \n4\n2\n6\nClimate concern (post)\n1\n2\n4\n3\n5\n7\n6\nBeliever\nSceptic\n1\n2\n3\n4\n5\n6\n7\nTreatment\n1\n2\n4\n3\n5\n7\n6\nControl\nClimate concern (pre)\nIncreased climate concern\n(relative to pre-survey)\nIncreased climate concern\n(relative to pre-survey)\n\u2206climate concern (post\u2013pre)\nControl\nTreatment\nCondition\n\u22120.10\n0\n0.10\nn = 73\nn = 70\nFig. 2 | Distributions of climate beliefs before and after participating in the \nclimate market. Taking the average of the three climate concern questions \n(Supplementary Fig. 4), we see that among the treatment (t(68)\u2009=\u20090.14, p\u2009=\u20090.890, \ntwo-tailed paired t test; left) and control (t(70)\u2009=\u20090.22, p\u2009=\u20090.830, two-tailed paired \nt test; centre) groups, there is no significant change in concerns following a \nmonth of waiting/betting (markers above the diagonal). Self-described believers \nand sceptics are marked by different symbols. Participants\u2019 group designation \naligns with the reported answer to the survey questions with the majority of \nbelievers scattered at the top-right of the panels (Supplementary Fig. 5). Right \npanel depicts the average difference concern score among treatment and control \nconditions between the pre- and post-surveys. Error bars depict standard errors \n(t(106)\u2009=\u20090.12, p\u2009=\u20090.901, two-tailed independent t test).\n\nNature Climate Change | Volume 13 | June 2023 | 523\u2013531\n527\nArticle\nhttps://doi.org/10.1038/s41558-023-01679-4\nWe propose that even if the participation in a climate prediction market \nis limited, the media accounts about market valuations, the predic-\ntion outcomes and the dissemination of knowledge that is derived \nfrom the markets may yield an increased shift in concerns in the  \nlarger population.\nWhile it is difficult to translate the effects in field studies to \npopulation-level outcomes, there are some metrics that could be \nimpacted by interventions such as ours. For example, if the predic-\ntion market in Study 1 was scaled to 1,000,000 climate believers and \nsceptics, and all believers decided to invest their annual earnings \nfrom, say, US$500 in market money (US$25, on average, if applied \nto our results) into countering climate change, this would result in \nan estimated US$25,000,000 of additional funding for climate solu-\ntions. Note that this amount could quickly increase when considering \nhighly motivated players that might have far greater yields than the \n5% earnings observed for the average believer in Study 1. Similarly, \ngiven that participation in climate markets such as the one in Study 2 \nyields an increase in climate concern, support and knowledge, such an \nintervention among a representative subset of the population could \nyield a shift in attitudes among millions12 of individuals (Box 1 provides \nimplementation details).\nThe majority of previous attempts at getting people to update \ntheir existing position on climate change focused on highlighting sci-\nentific consensus37\u201341, neutralizing partisan conflicts34\u201337 or appealing \nto norms31,42\u201345. The success of a number of those efforts was driven \nprimarily by increasing knowledge and providing information, which, \nin turn, helped shift perspectives. Some of the challenges in previous \nstudies have been attributed to: (1) motivated reasoning (that is, rejec-\ntion of new information that contradicts standing beliefs6), (2) desire \nto signal social identity within a group by clinging to information that \nfosters collective homogeneity46, (3) active efforts to foster uncer-\ntainty about climate science47. Our intervention offers a solution to all \nthree of these challenges by: (1) making motivated reasoning costly, \n(2) anonymizing people\u2019s decisions (thereby protecting their posi-\ntion within a group of climate sceptics, for example) while conveying \naggregated public opinion and (3) creating higher levels of certainty \nby having people actively engage with scientific sources. Addition-\nally, because the change in attitudes is intrinsically driven, it has the \npotential to be less threatening to one\u2019s identity and hence more sus-\ntainable. Together, these features might allow people to engage with \nclimate-related topics in a way that is less polarizing and less prone to \npartisan interpretation4,34\u201337.\nConservative\nLiberal\n1\n2\n3\n4\n5\n6\n7\nIncreased climate concern\n(relative to pre-survey)\nTreatment\na\nb\nControl\nTreatment\nControl\nIncreased climate concern\n(relative to pre-survey)\nIncreased climate support\n(relative to pre-survey)\nIncreased climate support\n(relative to pre-survey)\n2\n4\n6\n2\n4\n6\n2\n4\n6\nClimate concern\nClimate concern (post)\nClimate concern (pre)\n\u2206climate concern (post\u2013pre)\nClimate support\n2\n4\n6\n2\n4\n6\nClimate support (pre)\n2\n4\n6\nClimate support (post)\n\u20130.05\n0\n0.05\n0.10\n*\n0.15\n\u2206climate support (post\u2013pre)\n\u20130.05\n0\n0.05\n0.10\n0.15\nControl Treatment\nCondition\nControl Treatment\nCondition\n+\nn = 356\nn = 308\nn = 356\nn = 308\nFig. 3 | Effects of condition on climate concern and support. a,b, Participants \nconcern (a) and support (b) before (x axis) and after (y axis) the prediction \nmarket. Right column depicts the shifts in concern across control and treatment \nconditions, calculated as the average difference scores across conditions \nbetween the post- and pre-surveys. Error bars depict standard errors (*\u0394Concern: \n(651)\u2009=\u20092.27, p\u2009=\u20090.024, two-sided independent t test, +\u0394Support: t(661)\u2009=\u20091.82, \np\u2009=\u20090.069, two-sided independent t test).\n\nNature Climate Change | Volume 13 | June 2023 | 523\u2013531\n528\nArticle\nhttps://doi.org/10.1038/s41558-023-01679-4\nOur findings contribute to the existing knowledge on behaviour \nchange, both inside and outside of the climate domain. First, they align \nwith existing empirical results on how betting can serve as a tool for \nboosting engagement and behaviour change. For example, participat-\ning in sports prediction markets was shown to drive engagement with \nathletics48, and trading stocks of companies increases people\u2019s con-\nsumption of news related to those companies49. Second, the findings \nspeak to a growing body of work suggesting that reducing the \u2018distance\u2019 \n(psychologically, temporally or spatially) to a problem can lead to shifts \nin attitudes and behaviours37,50,51. For example, people who live closer \nto coastlines, where the effect of climate change is more concrete, \nexpress greater concern about climate change and higher support for \nregulating carbon emissions52, although unlike our intervention, this \nmay not hold for sceptics53. While our intervention does not change \nthe physical setting of participants, continuous engagement with \ntangible climate-related events may reduce the psychological distance \nto climate change and make its impact appear more imminent. Third, \nour findings align with simulations suggesting that participation in \nclimate prediction markets should foster alignment with scientific \nclimate consensus32.\nLimitations\nOur studies had a number of limitations. First, given that these were field \nexperiments involving real-time responses, the results are impacted by \nongoing events (that is, actions taken by other participants, news cycle \nor saliency of climate-related events). In Study 2, for example, climate \nevents dominated the news during our pre-survey period (including a \n50-year record high heatwave in Europe), which probably has impacted \nour baseline climate concern and support levels. This might have made \nit harder to see bigger increases in concern in the post-survey. By virtue \nof their realistic nature, our studies produce findings of high ecologi-\ncal validity and the results might vary depending on when the studies \nare conducted. Future climate prediction market studies should rep-\nlicate our findings across multiple time windows to corroborate the \noutcomes\u2019 generalizability.\nSecond, given that the participants were recruited based on \nlocation and climate beliefs, our results reflect the behaviour of US \nparticipants and not necessarily the world population. Indeed, the \npolarization of US citizens with respect to climate change is larger \nthan in other countries54. This polarization might have made it more \ndifficult to shift concerns with our prediction markets, suggesting that \nour results could be a conservative estimate of the effect size elsewhere.\nThird, we cannot speak to the exact mechanisms of our effects. \nBetting behaviour is the reflection of a complex combination of fac-\ntors, including: (1) participants\u2019 view/knowledge on topics, (2) their \nconfidence, (3) their risk tolerance, (4) their understanding of the \nmarket forces, (5) the amount of time participants have to do research \nand place bets, (6) the availability of funds, (7) the likelihood that \nothers would take the opposite position of a prediction (in Study 1) \nand (8) the available information on the outcomes (that is, more data \nwere available as the settle dates approached, in Study 1) and other \npsychological mechanisms. Future research could investigate these \nmechanisms individually.\nFourth, our limited study duration imposed a stringent cap on \nthe temporal horizon of predictions. This cap aligns poorly with the \nlonger timescale of climate change. We could not, for example, look \ninto notable changes in Earth\u2019s temperature within the time limit. This \nlimitation forced us to generate climate predictions with large spatial \ndomains (that is, multiple cities) or comparison to historical events. The \nuncertain relationship between near-term events and outcomes caus-\nally related to global warming inevitably caused some of our markets to \nreflect weather events rather than climate events. An implementation \nof climate prediction markets on a longer period (that is, years) would \nallow for long-term predictions and understanding of the effect of new \ninformation on these predictions, irrespective of the temporal horizon \n(that is, predictions about the year 2100 can be updated far ahead of \ntheir settle date if new information in, for example, 2025 suggests a \nneed for change of bet values). In fact, when Study 2 concluded, we \nasked participants to make predictions that span years into the future \n(Supplementary Table 19), which could be analysed when they settle \n(data available along with our Supplementary Information).\nFifth, our studies were limited to a financial allotment of US$20 \nper participant, capping motivation and outcomes. Participants were \nlimited to using their allotted amount and, correspondingly, partici-\npants who lost much of their income early were effectively excluded \nfrom further activity (and presumably less engaged with the study). \nThe fact that participants did not invest their own money may have \nchanged their overall motivation compared to prediction markets in \n0\n25\n50\n75\n100\n2.5\n5.0\n7.5\nClimate knowledge (pre)\nCount\nControl (n = 308)\nTreatment (n = 356)\n5\n10\n15\n20\nClimate knowledge (post)\n15.5\n20.5\n5.43 \u00b1 1.49\n5.60 \u00b1 1.53\nt(642) = \u20131.45, p = 0.147, two-sided t test\n10.00 \u00b1 3.20\n8.41 \u00b1 2.54\nt(657) = 7.11, p < 0.001, two-sided t test\n10.5\nFig. 4 | Climate knowledge increases after participating in a climate \nprediction market. Participants\u2019 knowledge was evaluated by comparing \nthe treatment and control groups\u2019 knowledge in the pre-survey (left; n.s.) and \npost-survey (right). Vertical lines are the mean of the corresponding color \ndistribution. n.s., not significant.\n\nNature Climate Change | Volume 13 | June 2023 | 523\u2013531\n529\nArticle\nhttps://doi.org/10.1038/s41558-023-01679-4\npublic exchanges. However, prior work has suggested that using virtual \nmoney may be as effective as real money55. We argue that this limitation \nmay indicate that a real-world prediction market could in fact amplify \nthe outcomes identified.\nTaken together, these limitations suggest that while our work pro-\nvides an initial feasibility test for climate prediction markets, further \nresearch should examine the markets\u2019 ability to shift attitudes persistently \nacross a more diverse set of samples. Specifically, future work should \ninvestigate whether changes in concern, support and knowledge are \nsustained long term and whether continuous participation in climate \nmarkets solidifies those changes. Additionally, further analyses of the \nbets could focus on the positions taken by individuals as dependent \nobservations to test whether certain outcomes affect future attitudes or \nbets (that is, losing multiple bets in sequence leading to less extreme bets).\nFinally, we strongly advocate for replication of our results using \nlarge-scale prediction markets, implemented over a longer period in \nan open, non-experimental setting56. This would allow market forces to \nstrengthen the effects and could lead to widespread attitude change.\nConclusion\nThis study offers empirical evidence for the ability of prediction markets \nto change people\u2019s attitudes about climate change. The engagement \nwith climate prediction markets in a domain that is uniformly quantita-\ntive and less polarizing than politics could not only support existing \nmethods to change climate concerns4,13,51,57 but also act as an ultimate \npolling tool to help scientists, activists and politicians aggregate public \nopinion about trends, policy preferences and future scientific predic-\ntions. It has not escaped our notice that the powerful financial instru-\nment proposed here could be used in other topics of controversy where \nan agreed-upon arbiter of truth could allow individuals to reflect their \nviews through market economics rather than publicly stated opinions.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-023-01679-4.\nReferences\n1.\t\nHart, P. S. & Nisbet, E. C. Boomerang effects in science \ncommunication: how motivated reasoning and identity cues \namplify opinion polarization about climate mitigation policies. \nCommun. Res. 39, 701\u2013723 (2012).\nBox 1\nApplications\nClimate prediction markets can be a useful tool for financial policy \nestimation27, for evaluation of public opinion25 and for aggregation \nof views and signals about the future26\u201328. Importantly, by making \nfalse beliefs regarding the future impacts of climate change \ncostly, climate prediction markets are likely to present a more \naccurate reflection of expectations about climate change. As \nsuch, they might help overcome politically motivated scepticism \nand gradually shift attitudes by highlighting that concern about \nclimate change is more widespread than surveys suggest. Our \nfinding that in various conditions the shift in concern was stronger \namong conservatives is particularly notable in this regard. As \none key attribute of any prediction market is its reliance on an \nindependent adjudicator, the power of prediction markets is that \nparticipants, upon entry, agree on the source they will use to \ndetermine the outcome.\nAn additional advantage of climate prediction markets is that \nthey make it possible to quantify probabilities about future events \nlong before the outcomes are manifested. In Study 1, for example, \nparticipants traded positions with different values far before the \ncontracts\u2019 settle date. This indicates that the additional information \nabout the future manifested itself in people\u2019s present behaviour \nregardless of the ultimate outcome. As an example, outcomes of \nelectoral bets (say, the winner of a presidential election in Brazil) may \nnot be determined for several months. However, new information \nabout current events and policies might change the bet\u2019s value in \nways that capture the crowd\u2019s changing estimate of the probability of \nthe outcome.\nTaken together, these prediction markets\u2019 properties make their \napplication at scale (that is, under a federally regulated authority) \na promising instrument in the arsenal of climate policymakers56. \nAs current direct financial incentives\u2014for example, tax credits \nor government subsidies\u2014produce limited results with respect \nto shifting concerns on climate58, prediction markets could act \nas a complementary tool for climate policy. The introduction of \nlarge-scale prediction markets for climate change could create a new \nsector of the financial information industry where climate attribution \nand prediction modelling would grow from a small academic \nenterprise to one that can help numerous governmental and private \nsector entities plan for the manifold effects of climate change and \nprioritize mitigation efforts.\nImplementing multiple climate prediction markets over the course \nof this work has highlighted the importance of market makers (that \nis, the individuals generating the bets) in the process. For example, \nthe selection of market topics or market launch times may influence \nmarket activity by nudging participants towards engagement with \na specific topic. Similarly, setting markets that are realistic and fair \nrequires effort and knowledge on the topic. Bets that are too extreme \nin any direction make all participants act in unison irrespective of their \npersonal opinions. For example, the thresholds we chose for settling \nthe first set of markets in Study 2 led to almost all bets resolving in \nfavour of predictions that aligned with climate scepticism (Fig. 1b). \nAlthough we do not find evidence that this alignment with climate \nscepticism affected our main outcomes (shift in climate concern and \nsupport), it resulted in a general shift among participants towards \nmore conservative bets in subsequent markets (Supplementary \nInformation \u2018Mid-study survey\u2019 examines this effect). The impact \nthat such choice of markets can have on the outcomes (and \nassociated attitude change) raises concerns about potential market \nmanipulation. While prior work has explicitly suggested market \nmanipulation as a means to subsidize certain positions, encourage \ninformed traders and reward accuracy25, our experience suggests \nthat this is not necessary for shifting attitudes. Put simply, engaging in \nclimate prediction markets without any manipulation yields increase \nin concern about climate change and support for climate action. \nThe unfortunate reality of climate change means that any randomly \nchosen period of time we may have selected for our studies would \nhave had an abundance of salient climate anomalies that yielded an \nincrease in concern about climate change. Given trends in global \nwarming and its causal role in extreme weather events, this will be \neven more true in the future.\n\nNature Climate Change | Volume 13 | June 2023 | 523\u2013531\n530\nArticle\nhttps://doi.org/10.1038/s41558-023-01679-4\n2.\t\nCampbell, T. H. & Kay, A. C. Solution aversion: on the relation \nbetween ideology and motivated disbelief. J. Personal. Soc. \nPsychol. 107, 809 (2014).\n3.\t\nKahan, D. M. et al. The polarizing impact of science literacy and \nnumeracy on perceived climate change risks. Nat. Clim. Change \n2, 732\u2013735 (2012).\n4.\t\nGuilbeault, D., Becker, J. & Centola, D. Social learning and partisan \nbias in the interpretation of climate trends. Proc. Natl Acad. Sci. \nUSA 115, 9714\u20139719 (2018).\n5.\t\nMa, Y., Dixon, G. & Hmielowski, J. D. Psychological reactance from \nreading basic facts on climate change: the role of prior views and \npolitical identification. Environ. Commun. 13, 71\u201386 (2019).\n6.\t\nDruckman, J. N. & McGrath, M. C. The evidence for motivated \nreasoning in climate change preference formation. Nat. Clim. \nChange 9, 111\u2013119 (2019).\n7.\t\nSmerdon, J. Climate Change: The Science of Global Warming and \nour Energy Future (Columbia Univ. Press, 2018).\n8.\t\nRosenzweig, C. et al. Attributing physical and biological impacts \nto anthropogenic climate change. Nature 453, 353\u2013357 (2008).\n9.\t\nCiscar, J.-C. et al. Physical and economic consequences of \nclimate change in Europe. Proc. Natl Acad. Sci. USA 108,  \n2678\u20132683 (2011).\n10.\t Xu, C., Kohler, T. A., Lenton, T. M., Svenning, J.-C. & Scheffer, M. \nFuture of the human climate niche. Proc. Natl Acad. Sci. USA 117, \n11350\u201311355 (2020).\n11.\t\nGallup Environment Report (Gallup, 2022); https://news.gallup.\ncom/poll/1615/environment.aspx\n12.\t Bliuc, A.-M. et al. Public division about climate change rooted in \nconflicting socio-political identities. Nat. Clim. Change 5, 226\u2013229 \n(2015).\n13.\t Weber, E. U. & Stern, P. C. Public understanding of climate change \nin the United States. Am. Psychologist 66, 315 (2011).\n14.\t Caplan, B. in The Myth of the Rational Voter Ch. 5 (Princeton Univ. \nPress, 2011).\n15.\t Frost, R. & McNaughton, N. The neural basis of delay discounting: \na review and preliminary model. Neurosci. Biobehav. Rev. 79, \n48\u201365 (2017).\n16.\t Suddendorf, T., Redshaw, J. & Bulley, A. The Invention of Tomorrow: \nA Natural History of Foresight (Basic Books, 2022).\n17.\t Hunt, L. T. et al. Formalizing planning and information  \nsearch in naturalistic decision-making. Nat Neurosci 24, 1\u201314 \n(2021).\n18.\t Schlenker, W. & Taylor, C. A. Market Expectations About Climate \nChange (National Bureau of Economic Research, 2019).\n19.\t Bernstein, A., Gustafson, M. T. & Lewis, R. Disaster on the horizon: \nthe price effect of sea level rise. J. Financ. Econ. 134, 253\u2013272 \n(2019).\n20.\t Tziralis, G. & Tatsiopoulos, I. Prediction markets: an extended \nliterature review. J. Prediction Mark. 1, 75\u201391 (2007).\n21.\t Berg, J. E., Nelson, F. D. & Rietz, T. A. Prediction market accuracy in \nthe long run. Int. J. Forecast. 24, 285\u2013300 (2008).\n22.\t Dreber, A. et al. Using prediction markets to estimate the \nreproducibility of scientific research. Proc. Natl Acad. Sci. USA 112, \n15343\u201315347 (2015).\n23.\t Polgreen, P. M., Nelson, F. D., Neumann, G. R. & Weinstein, R. A. \nUse of prediction markets to forecast infectious disease activity. \nClin. Infect. Dis. 44, 272\u2013279 (2007).\n24.\t Yeh, P. F. Using prediction markets to enhance US intelligence \ncapabilities. Stud. Intell. 50, 137\u2013149 (2006).\n25.\t Lucas, G. M.Jr & Mormann, F. Betting on climate policy: using \nprediction markets to address global warming. Univ. Calif. Davis \nLaw Rev. 52, 1429\u20131486 (2018).\n26.\t Vandenbergh, M. P., Raimi, K. T. & Gilligan, J. M. Energy and \nclimate change: a climate prediction market. Univ. Calif. Los \nAngeles Law Rev. 61, 61\u201367 (2013).\n27.\t Hsu, S.-L. A prediction market for climate outcomes. Univ. Colo. \nLaw Rev. 83, 179\u2013256 (2011).\n28.\t Sumner, S. & Jackson, A. L. Using prediction markets to guide \nglobal warming policy. In 63rd 153 International Atlantic Economic \nConf. 14\u201318 (Citeseer, 2008).\n29.\t Cerf, M., Greenleaf, E., Meyvis, T. & Morwitz, V. G. Using \nsingle-neuron recording in marketing: opportunities, challenges, \nand an application to fear enhancement in communications.  \nJ. Mark. Res. 52, 530\u2013545 (2015).\n30.\t Koller, M. & Stahel, W. A. Sharpening wald-type inference in \nrobust regression for small samples. Comput. Stat. Data Anal. 55, \n2504\u20132515 (2011).\n31.\t Bayes, R., Druckman, J. N., Goods, A. & Molden, D. C. When and \nhow different motives can drive motivated political reasoning. \nPolitical Psychol. 41, 1031\u20131052 (2020).\n32.\t Nay, J. J., Van der Linden, M. & Gilligan, J. M. Betting and  \nbelief: Prediction markets and attribution of climate change.  \nIn 2016 Winter Simulation Conf. (WSC) 1666\u20131677  \n(IEEE, 2016).\n33.\t G\u00f6tz, F. M., Gosling, S. D. & Rentfrow, P. J. Small effects: the \nindispensable foundation for a cumulative psychological science. \nPerspect. Psychol. Sci. 17, 205\u2013215 (2022).\n34.\t Kerr, J. R. & Wilson, M. S. Changes in perceived scientific \nconsensus shift beliefs about climate change and GM food safety. \nPLoS ONE 13, e0200295 (2018).\n35.\t Van der Linden, S., Leiserowitz, A. & Maibach, E. The gateway \nbelief model: a large-scale replication. J. Environ. Psychol. 62, \n49\u201358 (2019).\n36.\t Van der Linden, S. L., Leiserowitz, A. A., Feinberg, G. D. &  \nMaibach, E. W. The scientific consensus on climate change as a \ngateway belief: experimental evidence. PLoS ONE 10, e0118489 \n(2015).\n37.\t Zhang, B. et al. Experimental effects of climate messages vary \ngeographically. Nat. Clim. Change 8, 370\u2013374 (2018).\n38.\t Cook, J. et al. Quantifying the consensus on anthropogenic global \nwarming in the scientific literature. Environ. Res. Lett. 8, 024024 \n(2013).\n39.\t Howe, L. C., MacInnis, B., Krosnick, J. A., Markowitz, E. M. \n& Socolow, R. Acknowledging uncertainty impacts public \nacceptance of climate scientists\u2019 predictions. Nat. Clim. Change \n9, 863\u2013867 (2019).\n40.\t Ding, D., Maibach, E. W., Zhao, X., Roser-Renouf, C. & Leiserowitz, \nA. Support for climate policy and societal action are linked to \nperceptions about scientific agreement. Nat. Clim. Change 1, \n462\u2013466 (2011).\n41.\t Lewandowsky, S., Gignac, G. E. & Vaughan, S. The pivotal role of \nperceived scientific consensus in acceptance of science. Nat. \nClim. Change 3, 399\u2013404 (2013).\n42.\t Abrahamse, W. & Steg, L. Social influence approaches to \nencourage resource conservation: a meta-analysis. Glob. Environ. \nChange 23, 1773\u20131785 (2013).\n43.\t Allcott, H. Social norms and energy conservation. J. Public Econ. \n95, 1082\u20131095 (2011).\n44.\t Griskevicius, V., Cialdini, R. B. & Goldstein, N. J. Social norms: an \nunderestimated and underemployed lever for managing climate \nchange. Int. J. Sustainability Commun. 3, 5\u201313 (2008).\n45.\t Schultz, P. W., Nolan, J. M., Cialdini, R. B., Goldstein, N. J. & \nGriskevicius, V. The constructive, destructive, and  \nreconstructive power of social norms. Psychol. Sci. 18,  \n429\u2013434 (2007).\n46.\t Kahan, D. M. Climate-science communication and the \nmeasurement problem. Political Psychol. 36, 1\u201343 (2015).\n47.\t Oreskes, N. Merchants of Doubt: How a Handful of Scientists have \nObscured the Truth on Issues from Tobacco Smoke to Climate \nChange (Bloomsbury Publishing, 2015).\n\nNature Climate Change | Volume 13 | June 2023 | 523\u2013531\n531\nArticle\nhttps://doi.org/10.1038/s41558-023-01679-4\n48.\t Nielsen Sports Legal Sports Betting: What it Would Mean for NFL \nTV Partners & Advertisers (American Gaming Association, 2016); \nhttps://www.americangaming.org/sites/default/files/Nielsen_\nNFL_Betting.pdf\n49.\t Cerf, M. & Garcia-Garcia, M. Consumer Neuroscience (MIT Press, \n2017).\n50.\t Sugerman, E. R., Li, Y. & Johnson, E. J. Local warming is real: a \nmeta-analysis of the effect of recent temperature on climate \nchange beliefs. Curr. Opin. Behav. Sci. 42, 121\u2013126 (2021).\n51.\t Zaval, L., Keenan, E. A., Johnson, E. J. & Weber, E. U. How warm \ndays increase belief in global warming. Nat. Clim. Change 4, \n143\u2013147 (2014).\n52.\t Milfont, T. L., Evans, L., Sibley, C. G., Ries, J. & Cunningham, A. \nProximity to coast is linked to climate change belief. PLoS ONE 9, \ne103180 (2014).\n53.\t Osberghaus, D. & Fugger, C. Effects of extreme weather \nexperience on climate change belief. In Ann. Conf. European \nAssociation of Environmental and Resource Economists 26\u201329 \n(2019).\n54.\t Constantino, S. M., Cooperman, A. D., Keohane, R. O. & Weber, \nE. U. Personal hardship narrows the partisan gap in COVID-19 \nand climate change responses. Proc. Natl Acad. Sci. USA 119, \ne2120653119 (2022).\n55.\t Servan-Schreiber, E., Wolfers, J., Pennock, D. M. & Galebach, \nB. Prediction markets: does money matter? Electron. Mark. 14, \n243\u2013251 (2004).\n56.\t Cerf, M., Matz, S. C. & MacIver, M. A. Participating in a climate \nfutures marketincreases support for costly climate policies. Nat. \nClim. Change https://doi.org/10.1038/s41558-023-01677-6 (2023).\n57.\t Albright, E. A. & Crow, D. Beliefs about climate change in the \naftermath of extreme flooding. Climatic Change 155, 1\u201317 (2019).\n58.\t de Young, R. Changing behavior and making it stick: the \nconceptualization and management of conservation behavior. \nEnviron. Behav. 25, 485\u2013505 (1993).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2023, corrected publication 2023\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01679-4\nMethods\nStudy 1\nParticipants. A total of 160 participants were recruited for the study. \nParticipants were recruited online using Prolific Academic and through \nthe Reddit \u2018Climate Change sceptics\u2019 group. Participants were screened \non Prolific using two questions: (1) \u2018What is your nationality?\u2019 (only \nparticipants who answered \u2018United States\u2019 to this question were eligi-\nble to participate) and (2) \u2018Do you believe in climate change?\u2019 (people \ncould answer: \u2018Yes\u2019, \u2018No\u2019, \u2018Don\u2019t know\u2019 or \u2018Not applicable/rather not \nsay\u2019; an equal number of people saying \u2018Yes\u2019 or \u2018No\u2019 were recruited). \nParticipants\u2019 location within the United States varied and spanned areas \nthat are deemed high and low for their support for climate change sci-\nence59 (Supplementary Fig. 1). To ensure that participants beliefs about \nclimate change were consistent with the earlier Prolific screening, we \nadded an additional filter question (\u2018Global warming refers to the idea \nthat the world\u2019s average temperature has been increasing over the past \n150 years and may be increasing more. Do you think that global warm-\ning is happening?\u2019; Answers: \u2018Yes\u2019, \u2018No\u2019 and \u2018Don\u2019t know\u2019). Only people \nwho responded \u2018Yes\u2019 or \u2018No\u2019 were included in the study.\nSeventeen participants were excluded from the analysis, broken \ndown as follows: three were excluded because they did not complete \nthe required surveys, five because they failed an attention check ques-\ntion in either the pre-/post-survey and nine because they did not fulfil \nthe requirement to use the entirety of the allotted US$20. Of the 143 \nremaining participants, 73 were used as controls (35 believers, 38 \nsceptics) and the remaining 70 were used as the treatment group (35 \nbelievers, 35 sceptics). Supplementary Table 5 provides a breakdown \nof all participants\u2019 demographics.\nParticipants in the control group received US$5 for completing \nthe pre-survey and an additional US$5 for completing the post-survey. \nParticipants in the treatment group received the same renumeration \nfor these surveys, along with additional US$20 to spend in the climate \nprediction market (Supplementary Fig. 3 provides bet topics distribu-\ntion). Participants were instructed to use the full amount for climate \npredictions. At the end of the study, participants received their earnings \nin the climate prediction market. Participants who lost all their US$20 \nallotment during the betting period received only a US$10 participation \nfee. In total, participants in the treatment group could earn anything \nfrom US$10 (participation fee) to a maximum of US$650 (participation \nfee and their earning from bet wins).\nExperimental procedure. On the day of the study initiation, partici-\npants received a message instructing them to complete a pre-survey. At \nthe end of the survey, they were given a personalized link to a web-based \nonline climate betting site. Upon logging in to the site, they were pre-\nsented with a number of climate betting markets and could take a posi-\ntion on any number of them (Fig. 1). In addition, participants could \nchoose to trade a position with other participants. The number of avail-\nable markets changed daily based on old markets closing and new mar-\nkets opening. Participants could place multiple bets on the same market \nand could trade continuously before the bet\u2019s settle date and time.\nDuring the betting period, participants could log in to the pre-\ndiction market site whenever they wished, monitor their currently \navailable funds, view the available markets, make bets or trade posi-\ntions. The market mechanism was \u2018double auction\u2019 (Supplementary \nInformation), which required two participants to take opposite bets \nsuch that the sum of two bets was US$1 (that is, if one participant chose \nto wager US$0.60 that a \u2018Yes\u2019 bet will occur, only when another partici-\npant wagered US$0.40 that a \u2018No\u2019 on the outcome would a contract be \ninitiated). If no participant was willing to take the opposite wager, the \noffer remained pending until the participant making the offer chose \nto revoke it. The manifested value of each market at any given moment \nwas that of the last \u2018Yes\u2019 transaction to occur. That is, if a participant \nmade a bet for US$0.82 that the average methane level in October 2018 \nwill be the highest on record and another participant took the opposite \nposition at US$0.12, then all participants saw the current market value \nas US$0.82. Accordingly, the values of markets represented the aggre-\ngated stable amount of money people were agreeing to wager on each \ntopic. Naturally, as the settlement date of markets approached, the \nbets were likely to converge to the probability (0\u2026100) of the correct \noutcome (that is, if the market asked whether the number of disasters \nin a certain location be more than 10 by a certain date, and a few days \nbefore the closing time, a number of disasters already reached 9, the \nlikelihood of a \u2018Yes\u2019 bet was higher). The betting period was initiated \non 9 September 9 2018, and lasted until 11 November 11 2018. When \nthe betting period was complete, participants were instructed to com-\nplete a post-survey. Once participants completed the post-survey, \nthey were paid for their participation in the entire study. The pre- and \npost-surveys included a variety of questions (Supplementary Infor-\nmation provide all questions), but the main focus of the study was the \nsubset of questions pertaining to the concern about climate change.\nTo ensure the site\u2019s robustness to large-scale use and to reduce \nthe risk of technical issues jeopardizing the real-time experiment, we \nran a pilot test of the site for two months before the experiment on a \nsmaller group of participants.\nStudy 2 was similar to Study 1 in its design, with the following \ndeviations: (1) the criteria for exclusion in Study 1 was stricter (that is, \nparticipants were asked to use the full amount of money allotted to \nthem), (2) the betting period in Study 1 was longer and bets were not \nreleased daily but rather intermittently, (3) the participant population \nfor Study 1 was selected such that the pool was more polarized, (4) the \ncontrol group for Study 1 did not participate in an alternative predic-\ntion market, (5) the treatment group\u2019s bets in Study 1 occurred in a \ndouble auction, which pitted the believers and sceptics against each \nother with predictions occurring only when two participants claimed \nopposite sides of a bet such that the sum of the positions was US$1 (Sup-\nplementary Information provides an explanation of the double-auction \nfulfilment method), (6) participants in Study 1 could trade their bets \nin the market as the settle date was approaching based on the value of \nthe trade at the time, (7) participants in Study 1 did not have to take a \nposition on a bet as soon as it appeared on the portal but could choose \nto make a decision to enter as more information became available (the \noption price presumably reflected the information availability and \noutcome certainty), (8) participants in Study 1 could take contrary \npositions on the same bet or hedge their bets with a variety of positions.\nStudy 2\nParticipants. Participants were recruited through an online panel, \nProlific Academic. Our target sample for the start of the study was 1,000 \nparticipants with anticipated attrition rates of approximately 30\u201340% \nover the course of the entire assessment period. To obtain this initial \nsample, we recruited 1,754 participants whose native language was \nEnglish and who currently resided within the United States. We used \nProlific\u2019s representative sample criteria to ensure that the sample was \ngeneralisable. We excluded participants who took less than two minutes \nto complete the survey and who failed an attention check embedded \nin the survey (n\u2009=\u2009134).\nAll participants were asked a series of questions about their con-\ncerns pertaining to climate change, their support for climate action \nand their knowledge on basic climate-related topics (Measures). We \nfurther excluded participants whose answers were at ceiling (by calcu-\nlating the sum across four \u2018Concern\u2019 questions with a score of 1\u20137 each \nand excluding participants with a score of 27 or 28, out of 28; n\u2009=\u2009615; \nSupplementary Information).\nA total of 1,005 participants met all inclusion criteria. Participants \nwere randomly assigned to a treatment (n\u2009=\u2009524) or control condition \n(n\u2009=\u2009481; Supplementary Table 6). Participants in the treatment condi-\ntion were told that they would participate in a four-week-long climate \nprediction market, while participants in the control condition were \ntold about their participation in a sports and entertainment prediction \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01679-4\nmarket (Supplementary Table 6 provides evidence that the treatment \nand control groups randomization assignments did not significantly \ndiffer from one another).\nEach participant was allotted US$20 to use for bets throughout \nthe study. We considered participants\u2019 study records complete if they: \n(1) placed at least 15 bets and spent at least US$10 from their allocated \nwages during the prediction period, (2) completed the post-survey at \nthe end of the study, which included the same climate-related meas-\nures (that is, concern, support and knowledge) as the pre-survey. After \nexcluding participants who did not meet these criteria, we were left with \nan analysis sample consisting of 664 participants (34% overall attri-\ntion rate; 32% in the treatment group, 36% in the control; x2(2)\u2009=\u20091.71, \np\u2009=\u20090.191, n.s.). Participants were compensated with a fixed sum of US$11 \nfor completing the pre- and post-surveys and a variable additional \namount depending on their earnings in the prediction period.\nExperimental procedure. The study consisted of three main building \nblocks: (1) the pre-survey that measured participants\u2019 concern about \nclimate change (climate concern), their support for possible solu-\ntions (climate support) and their knowledge on climate issues (climate \nknowledge) before the intervention, (2) a five-week-long prediction \nmarket and (3) a post-survey that captured climate concern, support \nand knowledge following the intervention. Participants were recruited \nbetween 17 July and 21 July 2022 and completed the pre-survey as part \nof the initial screening procedure. After exclusion, participants were \nrandomly assigned to the five-week climate (treatment) or sports/\nentertainment (control) prediction markets. The betting period started \non 1 August 2022 and concluded on 4 September 2022. The final bet was \nsettled on 6 September 2022. The post-survey was sent to participants \non 8 September 2022 and was closed on 14 September 2022, at which \npoint all participants were paid (Fig. 1).\nEach participant received a unique login identifier that allowed \nthem to use a personalized version of study surveys. Every day at 10:00 \nall participants received an email through the study messaging system \nindicating that a new bet was available on the prediction portal. The \nmessage included a link to the prediction portal.\nUpon receipt of the daily reminder, participants had 20\u2009hours to \nenter the portal (Fig. 1), look at that day\u2019s bet and decide whether to make \na prediction. Once participants logged in to the portal, they were greeted \nwith their personal identification and a summary of their personal study \nmetrics. The metrics were: how many bets they had already placed, how \nmuch money they currently held in their wallet, how much money they \nhad in bets escrow (that is, bets awaiting resolution) and their total earn-\nings up to that point. Below the metrics, participants saw a summary of the \nbets that already materialized and their outcome. Below this information, \nparticipants saw the day\u2019s new bet (that is, \u2018Will the number of wildfires \nin California exceed 5,500 by August 8, 2022?\u2019) alongside the bet\u2019s settle \ndate/time (that is, \u2018August 8, 2022, 23:59 EDT\u2019), the source for deter-\nmining the outcome (that is, \u2018https://www.fire.ca.gov/incidents/2022\u2019) \nand (where possible) a graph of the history of the variable being bet on,  \nshowing the settle date of the current bet with respect to that graph.\nParticipants were then asked to indicate whether they wanted to \nabstain from betting, predict \u2018Yes\u2019 or predict \u2018No\u2019. If participants elected \nto make a Yes/No prediction, they advanced to the next screen where \nthey were asked to select the bet amount. Depending on their level of \nconfidence, participants could bet any amount between US$0.01 and \nUS$1. After participants determined their position and bet wager, they \nwere asked to confirm their decision or restart their decision. Once the \nparticipants confirmed their decision, the bet was locked for the day \nand they were not able to alter their bet.\nMeasures\nClimate concern. We measured people\u2019s concern about climate change \nin both the Pre- and Post-survey using the following four items: 1) \u2018Do  \nyou think that global warming/climate change is happening?\u2019 (Definitely  \nnot\u2014Definitely yes, 2) \u2018Do you think global warming/climate change \nis the result of human activities?\u2019 (Definitely not\u2014Definitely yes), 3) \n\u2018How much risk do you believe global warming/climate change poses \nto humanity\u2019s health, safety and prosperity?\u2019 (None at all\u2014Extremely \nhigh), and 4) \u2018Some people say that global warming/climate change \nis simply a scam. What do you think about this?\u2019 (Strongly disagree\u2014\nstrongly agree; reverse coded). The measure was adopted from work \nby Weber and colleagues54. Responses were recorded on a 7-point scale. \nWith a Cronbach\u2019s alpha of 0.93 in both the Pre- and Post-surveys, the \ninternal consistency of our measure was found to be excellent.\nSupport for climate change solutions. We measured people\u2019s support \nfor climate change solutions in both the pre- and post-survey using the \nfollowing three items: (1) \u2018Addressing global warming/climate change \nshould be a priority of the government\u2019 (strongly disagree\u2014strongly \nagree), (2) \u2018I feel personally responsible to help slow down global warm-\ning/climate change\u2019 (for example, by making changes to my lifestyle or \npaying higher taxes) (strongly disagree\u2014strongly agree) and (3) \u2018Some \npeople say that climate change is real, but that the cost of fixing it today \nmight not be worth the investment (that is, that the cost of fixing it today \nis higher than the cost of the damages caused by it)\u2019 (strongly disa-\ngree\u2014strongly agree). Responses were recorded on a seven-point scale. \nWith a Cronbach\u2019s alpha of 0.84 in the pre- and 0.87 in the post-survey,  \nthe internal consistency of our measure was found to be good.\nClimate knowledge. The questions for climate knowledge differed \nbetween pre- and post-survey. The pre-survey asked ten relatively \ngeneric multiple-choice questions (that is, \u2018How many major layers does \nEarth\u2019s atmosphere have?\u2019 Or \u2018What is the primary effect of greenhouse \ngasses?\u2019. The post-survey, on the other hand, asked a more compre-\nhensive set of questions that were directly related to knowledge about \nclimate change (that is, \u2018When does a tropical disturbance become a \ntropical storm and gains a name?\u2019 or \u2018What percentage of heat from \nglobal warming has the ocean absorbed in the past 40 years?\u2019; Sup-\nplementary Tables 2, 4, 18 and 19 provide all questions).\nSocio-demographic control variables. We collected information \nabout a wide range of participants\u2019 socio-demographic characteristics. \nThese included: age, gender, ethnicity, education, employment status, \nincome, religious beliefs, political ideology and number of children \n(Supplementary Tables 18 and 19).\nEthics statement\nAll participants in Study 1 and Study 2 signed an online consent form \nupon initial engagement with the pre-survey. Study 1 protocols were \napproved by Northwestern University\u2019s Institutional Review Board \n(STU00206273). Study 2 protocols were approved by Columbia Uni-\nversity\u2019s Institutional Review Board (AAAU2501).\nReporting summary\nFurther information on research design is available in the Nature Port-\nfolio Reporting Summary linked to this article.\nData availability\nThe data generated during the work are available at https://doi.org/ \n10.17605/OSF.IO/PH72Y.\nCode availability\nThe codes used for the analyses are available at https://doi.org/ \n10.17605/OSF.IO/PH72Y.\nReferences\n59.\t Marlon, J., Howe, P., Mildenberger, M., Leiserowitz, A. & Wang, \nX. Yale Climate Opinion Maps 2019 (Yale Program on Climate \nChange Communication, 2019).\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01679-4\nAcknowledgements\nWe thank M. Kim and A. Yu for assistance with managing the prediction \nmarket and maintaining the study throughout the betting periods; \nA. Mishra and I. Katz for designing and coding an initial version of \nthe online prediction markets; J. McCoy for discussion on prediction \nmarket utility; B. B. McShane, S. Franconeri, J. N. Druckman and  \nA. Coughlan for comments on the manuscript; E. Weber for support \nwith climate concern assessment; G. Schmidt, K. Hayhoe, D. Horton, \nE. Berlow and K. Marvel for help with identifying climate markets; and \nR. Behnam and members of the US Commodities and Futures Trading \nCommission for discussions on federal implementation of climate \nprediction markets. This work was funded by the Columbia University \nTamer Center for Social Enterprise (M.C. and S.C.M.) and by the \nNorthwestern Institute on Complex Systems (M.C. and M.A.M.).\nAuthor contributions\nAll authors equally designed the research, performed the research, \nanalysed the data and wrote the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41558-023-01679-4.\nCorrespondence and requests for materials should be addressed to \nMoran Cerf, Sandra C. Matz or Malcolm A. MacIver.\nPeer review information Nature Climate Change thanks  \nPaul Stern, Michael Vandenbergh and the other, anonymous, \nreviewer(s) for their contribution to the peer review  \nof this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\n\n    Scientific Research Findings:", "answer": "Participation in a climate prediction market, where individuals make predictions about climate futures and earn/lose money based on their forecasting accuracy, increases concern about global warming, support to mitigate the risks of climate change and knowledge about climate issues. This is true across levels of initial belief in climate change and political ideology. In one study (with a polarized group of climate believers and sceptics), the shift in perspective was contingent on winning (people who made money in their predictions also shifted their beliefs), whereas in a study including participants with more moderate views, the changes occurred independent of winning. This research provides a practical way to increase people\u2019s concern about climate change, as well as a powerful tool for policymakers to poll public opinion about climate issues, test potential future policies, inject private money to a new financial instrument and even use potential market gains towards climate solutions.", "id": 48}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 13 | April 2023 | 351\u2013358\n351\nnature climate change\nhttps://doi.org/10.1038/s41558-022-01592-2\nArticle\nWhy residual emissions matter right now\nHolly Jean Buck\u2009\n\u200a\u20091\u2009\n, Wim Carton\u2009\n\u200a\u20092, Jens Friis Lund3 & Nils Markusson4\nNet-zero targets imply that continuing residual emissions will be balanced \nby carbon dioxide removal. However, residual emissions are typically not \nwell defined, conceptually or quantitatively. We analysed governments\u2019 \nlong-term strategies submitted to the UNFCCC to explore projections of \nresidual emissions, including amounts and sectors. We found substantial \nlevels of residual emissions at net-zero greenhouse gas emissions, on \naverage 18% of current emissions for Annex I countries. The majority of \nstrategies were imprecise about which sectors residual emissions would \noriginate from, and few offered specific projections of how residual \nemissions could be balanced by carbon removal. Our findings indicate the \nneed for a consistent definition of residual emissions, as well as processes \nthat standardize and compare expectations about residual emissions \nacross countries. This is necessary for two reasons: to avoid projections \nof excessive residuals and correspondent unsustainable or unfeasible \ncarbon-removal levels and to send clearer signals about the temporality of \nfossil fuel use.\nNearly three-quarters of the world\u2019s global greenhouse gas emissions \nare covered by a net-zero law, policy or political pledge as of early \n20221. In its simplest form, net zero involves balancing some amount \nof remaining emissions with an equal amount of negative emissions \nthrough carbon dioxide removal. This idea of achieving a \u2018balance \nbetween anthropogenic emissions by sources and removals by sinks\u2019 \nwas enshrined in Article 4.1 of the Paris Agreement and has become \na prominent feature of recent IPCC assessments as well as country \nstrategies. Net-zero targets are driven by science that indicates that \nto limit warming to 1.5\u2009\u00b0C, the world must reach net-zero CO2 emis-\nsions around 2050 and net-zero greenhouse gas emissions later in \nthe century (2095\u20132100 with no or limited overshoot, 2070\u20132075 with \nhigh overshoot)2.\nWith the advent of net zero as a concept, the category of \u2018residual \nemissions\u2019 has emerged to denote emissions that are regarded as hard \nto abate and will need to be compensated via carbon removal. In the \nintegrated modelling literature, residual emissions may be defined as \nthose whose abatement remains uneconomical or technically infeasible \nunder the assumptions of a specific model and mitigation scenario3. \nFrom a governance or territorial standpoint, for example as stated in \nthe city of San Francisco\u2019s climate plan, residual emissions are simply \nthose \u201cthat remain due to limited existing options to eliminate or \nreduce them further\u201d.4 For corporations, residual emissions may be \ndefined in terms of the value chain; there may be emissions outside of \nthe scope of the company\u2019s direct control.\nCountries are currently detailing their strategies for how to reach \nnet-zero goals, which presents an opportunity to understand how \nthey see residual emissions at net zero. Specifically, governments are \nsubmitting long-term low-emissions development strategies (LT-LEDS) \nas invited under Article 4, paragraph 19 of the Paris Agreement. These \nstrategies are intended as an evolving visioning exercise, with emphasis \non process rather than the resulting document5\u20137. The idea was that this \nprocess could inform medium-term nationally determined contribu-\ntion target setting8. Creating LT-LEDS is a highly political process, and \nnations have approached it in different ways, although most have \nemployed both stakeholder engagement and modelling tools to create \npossible pathways.\nSimply reading a plan does not give immediate insight into what \nsort of buy-in the plan has across different internal actors within \nthe government or how involved external stakeholders in different  \nsectors truly are, both of which bear on how seriously the country \nwill be implementing the plan. Nations also have different levels of \nReceived: 17 April 2022\nAccepted: 21 December 2022\nPublished online: 9 March 2023\n Check for updates\n1 Department of Environment and Sustainability, University at Buffalo, Buffalo, NY, USA. 2Centre for Sustainability Studies, Lund University, Lund, Sweden. \n3Department of Food and Resource Economics, University of Copenhagen, Copenhagen, Denmark. 4Lancaster Environment Centre, Lancaster University, \nLancaster, UK. \n\u2009e-mail: hbuck2@buffalo.edu\n\nNature Climate Change | Volume 13 | April 2023 | 351\u2013358\n352\nArticle\nhttps://doi.org/10.1038/s41558-022-01592-2\nBecause projections out to 2050 are generally not yet in updated official \npolicy documents, the LT-LEDS remain the most accessible source of \ninformation on national expectations of amounts of residual emissions \nat mid-century. These countries are the first adopters of both LT-LEDS \nand net-zero targets, and their assessment and actions may set the tone \nfor countries that follow.\nIn what follows, we analyse country LT-LEDS strategies to examine \nfour key questions. (1) How are residual emissions defined? (2) What \namounts are countries projecting? (3) How are residual emissions \ndistributed among sectors? (4) What are the expectations around the \nland sector\u2019s ability to compensate for residual emissions?\nDefinition of residual emissions\nOur analysis of the 50 LT-LEDS shows that there is no consistent defini-\ntion or use of the concept of residual emissions. A majority of LT-LEDS \ndo not explicitly mention the concept of residual emissions, despite \nhaving a net-zero target. Few countries provide an explicit definition \nor elaborate how residual emissions amounts are arrived at, explain \nwhat criteria were used to determine them or specify what greenhouse \ngases make up the residual emissions.\nThe examples in Table 2 illustrate the variance in how countries \ndescribe residual emissions in LT-LEDS. Countries such as Switzerland \nand Norway suggest an absolute limit on abatement options by describ-\ning residual emissions as those that \u2018cannot\u2019 be completely eliminated. \nBy contrast, France and Nepal exemplify a more fluid understanding, \nwhere the need for residual emissions owes to \u2018the current state of \nknowledge\u2019 and with the expectation that technological advance-\nment might change this. Sweden explicitly mentions the ambition to \nminimize residual emissions as much as possible, suggesting at least \nsome political leverage over the amount of residual emissions allowed \nin LT-LEDS. Finally, some countries make explicit reference to economic \nconsiderations in their description of residual emissions.\nWe also examined the approach the countries took to projecting \nresidual emissions. In theory, there are two main ways to estimate the \namount of residual emissions at mid-century. The first is a top-down \napproach that starts with a specified national policy target (such as 85% \nor 90% of emissions from a baseline year) and either simply sets resid-\nual emissions equal to that or uses economy-wide or sector-specific \nmodelling to figure out how to solve for it. The second is a bottom-up \nstakeholder-informed approach that estimates possible reductions \nin each sector then aggregates those sectoral estimates. In principle, \na third approach is also possible\u2014one that begins with negative emis-\nsions, with either a top-down approach that starts with a target sink \ncapacity or a bottom-up approach that estimates the capacity for each \nsource of carbon removals and then projects allowable residual emis-\nsions equal to that amount. However, countries are not at present using \nan approach that leads with negative emissions. In our sample of 50 \nLT-LEDS, around one-third of countries utilized a top-down approach, \nabout 15% used a bottom-up approach, about 10% set residual emissions \nequal to the level of forest sinks and the rest used a combined approach \nor left the approach unspecified.\nAmounts of residual emissions\nThe 18 LT-LEDS in our sample that include Annex I countries with a quan-\ntification of residual emissions together project residuals of 2.2\u2009Gt\u2009yr\u20131 \nin 2050 in their most ambitious scenarios (Fig. 1). This corresponds to \n17.9% of these countries\u2019 current emissions. Together, these countries \nare currently responsible for 18% of global emissions. Should the rest \nof the world make similar projections, the resulting residuals would be \nover 12\u2009Gt\u2009yr\u20131 (if weighted by current emissions). This sets out a need \nfor a substantial carbon-removal effort.\nHowever, this figure of 12\u2009Gt\u2009yr\u20131 probably underestimates the \nglobal residual emissions that countries will be planning for. We say \nthis for three reasons. First, most countries included between two \nand four low-carbon scenarios. For all these countries, we chose the \nplanning capacity\u2014not just scientifically speaking in terms of having \nforecasting tools and data, but in terms of institutional and political \npossibilities to articulate a 2050 goal and explicate what would be \nneeded to achieve it. Costa Rica\u2019s strategy, for example, states plainly \nthat achieving the structural transformation requires new tools in \nterms of making political decisions and analysing what steps will be \nneeded to see them succeed and that traditional approaches based on \noptimization models will not deliver9. It situates the LT-LEDS within a \nbroader development planning process, led by the Ministry of National \nPlanning and Economic Policy. For other countries, the LT-LEDS are not \nso well integrated into planning or sustainable development institu-\ntions. While in this paper we treat the outputs from these processes as \ncomparable, it is important to understand that they are only facets of \na deeply individual set of circumstances and processes.\nThe content of these strategies is more speculative than a defini-\ntive \u2018plan\u2019. Most LT-LEDS present pathways\u2014what-if explorations of \ndifferent scenarios for reaching desired targets\u2014created using a variety \nof methods. These scenarios and quantified projections inform the \nstrategy but are meant to be illustrative of possible futures, not predic-\ntive or prescriptive10. This means that in this paper, when we discuss \na country\u2019s estimation of residual emissions at mid-century, we are \nreferring to the most ambitious scenario they have offered, not their \npreferred target or what they are necessarily planning for. Our sample \nreflects this diversity and is characterized by different approaches to \noffsetting, removal methods and target framing (Table 1).\nWhile most countries submitted LT-LEDS in 2020 or 2021, some \ncountries, such as Germany and Canada, submitted their LT-LEDS a few \nyears ago (in 2016) and have enacted more ambitious policy since the \nfirst iteration of their plans. The Paris Agreement and Katowice Rulebook  \ndo not clearly specify whether LT-LEDS should be continuously \nupdated, although at COP-26 in 2021, countries were encouraged to \nsubmit or update before COP-27. As of mid-2022, 51 long-term strate-\ngies have been submitted; 50 were examined for this Article, of which \n28 include a quantified projection of residual emissions at net zero \n(in all but four cases, this is 2050). These countries are responsible for \nonly about a fifth of current emissions and contain few large emitters. \nTable 1 | Summary of information in the long-term strategies \n(N\u2009=\u200950)\nTarget framing\nYear of net-zero ambition\nNet zero\n31\n2040\n1\nCarbon neutral\n6\n2045\n2\nClimate neutral\n6\n2050\n31\nEmissions reduction\n5\n2060\n1\nReduction versus business as usual\n1\n2065\n1\nOther\n1\nNot specified\n14\nConsiders natural negative emissions \ntechnologies?\nConsiders technological \ncarbon removal techniques?\nYes\n36\nYes\n25\nNo\n4\nNo\n12\nNot specified\n10\nNot specified\n13\nFocus on territorial emissions only?\nUse of offsetting?\nIncludes consumption\n7\nYes\n25\nTerritorial only\n20\nNo\n13\nNot specified\n23\nNot specified\n12\nDefines residual emissions?\nQuantifies residual emissions?\nYes\n25\nYes\n28\nNo or unclear\n25\nNo\n22\n\nNature Climate Change | Volume 13 | April 2023 | 351\u2013358\n353\nArticle\nhttps://doi.org/10.1038/s41558-022-01592-2\nscenario with the smallest number of residual emissions for this calcu-\nlation. Second, most countries do not include international aviation \nand shipping in their projections, both of which are commonly seen \nas hard-to-abate sectors. They could represent substantial sources \nof residual emissions: the International Energy Agency\u2019s Net Zero by \n2050 scenario includes 210\u2009MtCO2 from aviation and 120\u2009MtCO2 from \nshipping, while also making strong assumptions about behavioural \nchange and demand reductions in aviation11. Finally, and crucially, this \ncalculation is derived from projections from wealthy Annex I countries, \nand poorer countries may claim higher shares of residual emissions \nas well as later net-zero dates. This would be in accordance with the \nprinciple of common but differentiated responsibilities and respective \ncapacities12. In other words, extrapolating from the most ambitious \ncurrent projections of the world\u2019s richest countries still gives a baseline \nindication of residual emissions in the double digits.\nExpectations of carbon removal via LULUCF\nWe examined the projected role of land use, land-use change and \nforestry (LULUCF) for the 18 Annex I countries that offer estimations  \nof residual emissions at net zero to understand whether countries \nprojected that this sector would compensate for residual emissions. \nThe plans for future LULUCF vary in their concreteness and detail; some \ninclude several scenarios specifying amounts of future LULUCF while \nothers offer only vague ideas about future mitigation through LULUCF.\nMost countries expect to enhance or maintain the removal capacity  \nof the LULUCF sector (Table 3). For many of the countries that plan \nfor enhanced removals from the LULUCF sector, these removals will \nequal or surpass their expected residual emissions by the point of net \nzero. This is the case for, among others, Finland, Iceland, Hungary,  \nLatvia, Portugal, Slovakia, Spain and Sweden. However, for the biggest \nemitters in the sample, expected LULUCF removals fall far short of \nresiduals. This is the case for Australia, Canada, France, Switzerland, \nthe United Kingdom and the United States. Taken together, these six \ncountries comprise 96% of the total residuals of the sample. As these \ncountries comprise the majority of residuals, their plans will be decisive \nfor the overall amount of residuals that will have to be removed through \nmeans other than the LULUCF sector.\nSources of residual emissions\nOf the countries with quantitative projections of residual emissions,  \n15 Annex I countries provide a quantitative sectoral breakdown, shown \nin Fig. 2. Notably, across these countries, electricity is not responsi-\nble for many residual emissions, aligning with common expectations  \nthat electricity is feasible to decarbonize. Agriculture and industry \nrepresent the largest residual emissions. The prominence of agriculture \nbrings up the question of whether residual emissions are expected  \nto be CO2 or other greenhouse gases, which is unspecified in most \nstrategies. Only the United Kingdom includes aviation in its account-\ning of residual emissions, amounting to nearly half of its total. Notably, \nthese figures are mainly from Organisation for Economic Co-operation \nand Development countries, and many of the non-Annex I countries \nindicated that they would have residual emissions from energy.\nTable 2 | Selected references to residual emissions in long-term strategies\nCountry\nDescription\nExamples of references to residual emissions with varying degrees of certainty\nCosta Rica\n\u201cToday, the great imperative in Costa Rica \u2026 would be to transform the emissions pattern of the economy into a net-zero emissions, or \nnegative emissions (i.e., removals) society, in sectors where it is possible - and very low emissions where it is not possible to reach zero. In \npractice, this means that each sector will be transformed toward zero emissions, yet at different speeds.\u201d9\nSwitzerland\n\u201cThe emission of greenhouse gases cannot be completely eliminated in some sectors. From a current perspective, this includes \nagricultural food production, some industrial processes, such as cement manufacture, and waste incineration. To achieve the net-zero goal, \nthese remaining emissions must be balanced by the use of technologies or processes that remove CO2 from the atmosphere and store it \npermanently.\u201d26\nIceland\n\u201cThe goal of climate neutrality will not be reached without using removals of carbon from the atmosphere to compensate for emissions that \nare unlikely to be eliminated.\u201d27\nJapan\n\u201cDespite the progress in energy efficiency and decarbonization in each sector, there are some sectors where CO2 emissions are unavoidable. \nCO2 from those sectors can be removed by specific measures such as Direct Air Carbon Capture and Storage (DACCS), Bio-Energy with \nCarbon Capture and Storage (BECCS), and forest sink measures.\u201d28\nExamples of residual emissions constrained by current state of technological knowledge\nFrance\nGlossary entry: \u201cNear-total decarbonisation: maximum reduction of greenhouse gas emissions, the residual emissions, which are \nunavoidable according to the current state of knowledge, being mainly due to agriculture, and to a lesser extent to industrial processes, \nwaste, domestic air transport and gas leaks (biogas, hydrogen, fluorinated gases).\u201d12,29\nNepal\n\u201cDue to the limited capacity of current technologies, there are still emissions from energy and IPPU. However, with future technological \nadvancements, this can be avoided and reduced.\u201d13,30\nExamples of residual emissions delimited politically\nSweden\n\u201c[S]ome agricultural emissions are likely to remain even after 2045. These remaining emissions will need to be compensated for with \nsupplementary measures. It is nevertheless essential to work to ensure that these remaining emissions are as small as possible.\u201d31\nUnited Kingdom\n\u201cWe are clear that the purpose of greenhouse gas removals is to balance the residual emissions from sectors that are unlikely to achieve full \ndecarbonisation by 2050, whilst not substituting for ambitious mitigation to achieve net zero. GGRs must not be pursued as a substitute for \ndecisive action across the economy to reduce emissions, often referred to as mitigation deterrence.\u201d32\nExamples of residual emissions defined partly in economic terms\nAustralia\n\u201cAdditional direct emissions reductions could be enabled through a more aggressive approach to technology. Informed by the Technology \nInvestment Roadmap and annual LETS, Australia could focus on bringing down the costs of currently very expensive abatement \nopportunities in hard-to-abate sectors like industry and agriculture.\u201d33\nUnited States\n\u201cIn the three decades to 2050, our emissions from energy production can be brought close to zero, but certain emissions such as non-CO2 \nfrom agriculture will be difficult to decarbonize completely by mid-century \u2026 While mitigation opportunities exist for many sources of \nnon-CO2 GHG emissions, costs and applicability vary. Because it is challenging to eliminate all of these sources, some remaining non-CO2 \nemissions will need to be offset in 2050 by net-negative CO2 emissions.\u201d34\nBold text indicates authors\u2019 emphasis.\n\nNature Climate Change | Volume 13 | April 2023 | 351\u2013358\n354\nArticle\nhttps://doi.org/10.1038/s41558-022-01592-2\nThe projections in country strategies cohere largely with the  \nsectoral breakdown of residual emissions one can find in the litera-\nture, although countries may be projecting larger amounts than in \nthe literature. The International Energy Agency\u2019s Net Zero by 2050 \nscenario describes a largely decarbonized power sector. Out of 1.5\u2009Gt \nof residual emissions in this scenario, 40% is from heavy industries, \nmainly in developing economies (chemicals, steel, cement), and 33% \nis from aviation, shipping and trucks; notably, this scenario is focused \nonly on energy, not land.\nScenario studies analysed in the IPCC Sixth Assessment Report \n(AR6)2 similarly highlight residual emissions from non-electric energy, \nparticularly in transport and industry (2.7.3). The AR6 also presents esti-\nmations of residual GHG emissions at net zero from illustrative mitiga-\ntion pathways (IMPs) (fig. SPM.5). The pathways compatible with below \n1.5\u2009\u00b0C with limited or no overshoot have residuals of 6.79\u2009Gt (\u2018shifting \ndevelopment pathways\u2019, IMP-SP), 8.73\u2009Gt (\u2018low demand\u2019, IMP-LD) and \n11.87\u2009Gt (\u2018high renewables\u2019, IMP-Ren), with half to two-thirds of these \nfrom non-CO2 emissions13. In other words, analysis of net-zero and 1.5\u2009\u00b0C \ncompatible pathways from the scientific literature also anticipates \nthat the majority of residual emissions will be from agriculture, with \nsome residual emissions from industry and transport. Yet estimations \nof total amounts vary widely depending on scenario, and regional \nanalysis is limited.\nDiscussion\nOur analysis of the LT-LEDS submitted to the UNFCCC so far shows  \nthat (1) residual emissions do not have a standard conceptual defi-\nnition; (2) countries\u2019 projected residual emissions are a substantial \npercentage of current emissions, averaging around 18% for Annex I \ncountries in the most ambitious scenarios; (3) while most residual emis-\nsions in ambitious scenarios are indicated to come from agriculture, \nindustry and mobility, few countries specify sectoral breakdowns;  \n(4) for countries analysed, LULUCF sinks by 2050 cannot balance out \nall residual emissions.\nAs countries look towards submitting or updating LT-LEDS in \nadvance of future UNFCCC events, researchers, policymakers and civil \nsociety should work towards standardizing expectations on residual \nemissions. Right now, state and non-state actors alike can self-define, \n0\n100\n200\n300\n400\n500\n600\n0\n10\n20\n30\n40\n50\n60\n70\n80\n0\n20\n40\n60\n80\n100\n120\n0\n100\n200\n300\n400\n500\n600\n700\n800\n0\n10\n20\n30\n40\n50\n60\n0\n100\n200\n300\n400\n500\n0\n10\n20\n30\n40\n50\n60\n70\n80\n0\n1\n2\n3\n4\n5\n0\n2\n4\n6\n8\n10\n12\n0\n0.5\n1.0\n1.5\n2.0\n2.5\n0\n10\n20\n30\n40\n50\n60\n70\n80\n0\n5\n10\n15\n20\n25\n30\n35\n40\n0\n5\n10\n15\n20\n0\n50\n100\n150\n200\n250\n300\n350\n0\n10\n20\n30\n40\n50\n60\n0\n10\n20\n30\n40\n50\n0\n100\n200\n300\n400\n500\n0\n1,000\n2,000\n3,000\n4,000\n5,000\n6,000\n7,000\n8,000\nAustralia\nAustria\nBelgium\nCanada\nFinland\nFrance\nHungary\nIceland\nLatvia\nMalta\nPortugal\nSlovakia\nSlovenia\nSpain\nSweden\nSwitzerland\nUnited Kingdom\nGrey: 2019 emissions\nGreen: projected emissions \nat net zero\n29.9%\n16.8%\n29.6%\n13%\n16.8%\n16%\n18%\n18.4%\n9.2%\n8.7%\n7%\n21.4%\n21.2%\n20.4%\n19%\n17.5%\n15%\nUnited States\n24.5%\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nEmissions (MtCO2e)\nFig. 1 | Residual emissions versus 2019 emissions, Annex I countries. The 2019 emissions are from UNFCCC inventories; total GHG emissions without LULUCF.  \nCO2e, CO2-equivalent.\n\nNature Climate Change | Volume 13 | April 2023 | 351\u2013358\n355\nArticle\nhttps://doi.org/10.1038/s41558-022-01592-2\nand claim, various amounts of residual emissions. The gift of the Paris \nAgreement framework is its flexibility in exactly how countries choose \nto balance sources and sinks of emissions. However, specifying residual \nemissions will mitigate against the risk that governments put things \nthat are expensive or politically inconvenient to abate into the \u2018residual \nbox\u2019, thus increasing the amount of residual emissions\u2014and thereby \ncreating pressures for an even larger carbon-removal infrastructure.\nConcerns about the feasibility, sustainability and societal impacts \nof carbon removal at several gigatons per year14,15 have led to calls to \nmoderate expectations of future carbon removal16. This is because \nterrestrial carbon removal at the scales indicated in this Article would \nrequire vast amounts of land and entail severe risks for food produc-\ntion and/or biosphere functioning17,18 as well as the land rights and \nlivelihoods of rural communities and Indigenous peoples19. While \nsome industrial carbon-removal techniques such as direct air carbon \ncapture and storage have a much smaller direct land footprint, this \napproach comes with large energy requirements20, which could divert \nenergy, and critical minerals and the associated land for renewables, \nfrom other societal needs. Ultimately, the idea that some emissions are \nhard to abate must be examined in light of these risks and challenges \nwith scaling carbon removal.\nMany actors have called for greater clarity in net-zero targets \nand plans, regarding carbon removal but also around pathways in \ngeneral12,21\u201323. Norms are evolving about how to develop net-zero path-\nways, as set forth in the UN Race to Zero campaign or the Science-Based \nTargets Initiative. The latter sets out cross-sector and sector-specific \npathways that include a 90% reduction by 2050, with pathways that \nreach a \u2018low\u2013medium\u2019 global level of carbon removal of 1\u20134\u2009Gt\u2009yr\u20131 in \n205024. This could be an effort that sets global norms around corpo-\nrate residual emissions. While we applaud the business community \nand NGOs for attempting to set norms, we see a much clearer role for \ngovernments in this area, even while acknowledging that governments \nwill face difficulties in this space. There is political advantage in leav-\ning residual emissions strategically ambiguous as governments need \nto accommodate the interests of different sectors and regions. At the \nsame time, both industries and communities can benefit from certainty \nin planning, and better setting out clarity and expectations around \nresidual emissions also has political and economic benefits.\nWe make the following three recommendations for policymakers \ndeveloping long-term strategies. These recommendations are also \nimportant for the researchers and NGOs supporting their work, who \nhave a critical role in supporting international policymaking (Box 1).\nFirst, include clear projections for (1) the amount of residual emis-\nsions, (2) where they originate sectorally and spatially and (3) the types \nof greenhouse gas. Scenarios and the graphical user interfaces used to \nexplore them can be made more user friendly, allowing broader engage-\nment with these key issues in climate policy. Multiscalar datasets linking \nbroader analysis of residual emissions to regional or facility-level data \nwould enable critical debates about infrastructure and enable planning \nfor just transitions.\nSecond, the policy and research communities should suggest \ndefined criteria by which \u2018hard to abate\u2019 should be judged. While sectors  \nsuch as aviation, steel and agriculture are commonly understood as \ndifficult to decarbonize, terms such as difficult, unavoidable, hard \nto abate, impossible to eliminate and so on carry value judgements \nabout what kind of activities a society should or should not engage \nin and what costs are reasonable. This normativity is unavoidable. \nHowever, greater transparency around how emissions come to be \nconsidered residual is critical for the legitimacy of decarbonization \nefforts. Defining criteria would allow for comparison and negotiation \nand the development of international norms on how to determine dif-\nficulty of abatement. This is particularly important given that what is \nhard to abate changes along with technological developments, such \nas green hydrogen and low-carbon aviation. Thus, assumptions and \nnorms around hard-to-abate emissions must be constantly revised.\nThe scientific community has a key role in supporting society in \ndefining these criteria, in terms of both creating tools and producing \nresearch. Researchers can also produce analysis to answer the following \nkey questions. What processes and sectors lack technological options \nfor fully eliminating emissions? Are there technologies that would \nbecome options under different policy scenarios? Where are there \nopportunities for demand-side options to lower residual emissions \nfurther, and what social factors enable and constrain those options? \nThese questions require interdisciplinary research, and governments \nshould support this research, directly funding and coordinating it as \nwell as being receptive to existing efforts and incorporating them into \nprogrammes.\nThird, be explicit about whether residual emissions\u2014and net zero \nas a goal\u2014are a temporary stopgap towards a further state of decar-\nbonization or a state to maintain in perpetuity. Clarity on whether \nresidual emissions are a temporary condition or a permanent state \nis important, both for calibrating expectations for the future of the \nfossil fuel sector and for understanding the intended role for carbon \nremoval. If negative-emission capacity is being used to compensate for \nresidual emissions domestically or in another country, it is not avail-\nable for legacy carbon removal or coping with overshoot. Although \nthe AR62 frames these roles of carbon removal as complementary, \nthey may be in conflict if we assume carbon-removal potential will be \nlimited for social and sustainability reasons. Clarity on the temporal-\nity of residual emissions is also important because strategies such as \nsoil carbon sequestration have apparently high mid-century technical \npotential, but these sinks saturate after ~20 years and require ongoing \nmaintenance14. Land-based sinks already accounted for may saturate \nover time, as may carbon stored in products. Net zero needs to be a \ndurable state22, not something that might be achieved and then be \nlost again. The timing of various carbon-removal strategies needs to \nbe better planned for, and the ability to do so hinges on understanding \nTable 3 | Overview of countries\u2019 residuals, recent and \ncurrent LULUCF35 and long-term LULUCF outlook\nCountry\nResiduals \n(MtCO2e)\n2020 \nLULUCF \n(MtCO2)\nAverage  \n2000\u20132020 \nLULUCF (MtCO2)\nLong-term \nLULUCF \noutlook\nAustralia\n139\n\u221243\n17\nEnhance\nAustria\n13\n\u22125\n\u22127\nAmbiguous\nBelgium\n10\n\u22121\n\u22122\nMaintain or \nenhance\nCanada\n149\n9\n\u22128\nEnhance\nFinland\n9\n\u221217\n\u221222\nMaintain or \nenhance\nFrance\n80\n\u221235\n\u221240\nEnhance\nHungary\n5\n\u22126\n\u22124\nMaintain\nIceland\n1\n6\n6\nEnhance\nLatvia\n4\n\u22123\n\u22126\nEnhance\nMalta\n0\n0\n0\nMaintain\nPortugal\n9\n\u22128\n\u22127\nEnhance\nSlovakia\n7\n\u22126\n\u22127\nMaintain\nSlovenia\n2\n0\n\u22125\nEnhance\nSpain\n29\n\u221238\n\u221239\nMaintain\nSweden\n11\n\u221237\n\u221239\nEnhance\nSwitzerland\n68\n\u22122\n\u22122\nAmbiguous\nUnited Kingdom\n76\n\u22121\n0\nEnhance\nUnited States of \nAmerica\n1,605\n\u2212813\n\u2212818\nEnhance\n\nNature Climate Change | Volume 13 | April 2023 | 351\u2013358\n356\nArticle\nhttps://doi.org/10.1038/s41558-022-01592-2\nwhether net zero is a stopgap or permanent state. While governments \nwill have a challenging time being explicit about this, given their need \nto address multiple domestic actors, the research institutions and \nNGOs working in policy have more flexibility to be explicit about this \nin their analyses and can spell out the implications of treating residual \nemissions as continuing versus temporary.\nResidual emissions need to be openly analysed in both science and \npolitics because the stakes of continuing to treat residual emissions \nas a technocratic matter are high. Large and unsubstantiated claims \non residual emissions will undermine mitigation. Moreover, failing \nto decide and agree on residual emissions, and instead allocating \nthem according to simple market logics, means that more-powerful \n0\n0.2\n0.4\n0.6\n0.8\n1.0\nTotal\nOther GHGs\nInternational\naviation and shipping\nAgriculture\nIndustry\nWaste\nBuildings\nTransport\nElectricity\nEnergy\nUnited States\nUnited Kingdom\nSwitzerland\nSweden\nSpain\nSlovenia\nPortugal\nLatvia\nHungary\nFrance\nFinland\nCanada\nBelgium\nAustria\nAustralia\nFractions of total residuals (CO2e)\nCountry\nFig. 2 | Sectoral breakdowns of residual emissions at mid-century in the most \nambitious scenarios. Data are for Annex I countries that featured projections \nwith quantified sectoral breakdowns. Year depicted is 2050 for all countries \nbesides Sweden, which has projections for 2045 when it reaches net zero. Finland \nhas a target of net zero at 2035 but includes projections for 2050. Note that some \ncountries group electricity and transport into energy, and the United States does \nnot report agriculture but rather CO2 and other GHGs.\nBox 1\nEmerging research areas for international net-zero policy\nInternational policy efforts are needed to solve multiple problems \nthat underlie the net-zero framework. One problem is how residual \nemissions and removals can be matched. Carbon-removal-focused \ninternational cooperation efforts are absent or poorly described in \nLT-LEDS, even though cross-country efforts might be the most cost \neffective36,37. Some countries indicate that they may need to procure \ncarbon removal from abroad (Switzerland, Australia), yet no countries \nindicated that they intended to produce surplus removals for global \nmarkets. The challenge here has typically been read as (1) the need \nto work out issues with market mechanisms, as Article 6 negotiations \nare tackling, and (2) the need for better monitoring, reporting and \nverification to make exchangeable removals credible38\u201342. Both of \nthese are serious challenges.\nHowever, there is another pressing international policy need \nto create safeguards against dynamics where countries expect to \nacquire removals in developing countries, creating rushes\u2014for \nland, terrestrial carbon storage, space for ocean carbon removal, \ngeological sequestration capacity or renewable resources to  \npower carbon-removal technologies, such as direct air capture.\nA second problem is that the evolving carbon marketplaces \nhave no way of making sure that removals are in fact compensating \nfor emissions from sectors and activities that are truly hard to \nabate. Alternative frameworks might have nations with similar \nsocioeconomic capacities striving for the same amount of ambition \nin terms of decarbonizing each sector or dividing residual emissions \naccording to luxury and subsistence emissions43.\nA third policy challenge is that from a climate-justice perspective, \nwealthy countries with historical responsibility, such as the United \nStates, should deliver net-negative emissions sooner to allow  \npoorer countries some net residual emissions post 2050. However, \nif such wealthy countries decide to use their capacity for carbon \nremoval to balance residuals in expensive but possible-to-reduce \nsectors to lower the costs of meeting net-zero goals, this adds \nfurther pressure on other countries. Moreover, the geopolitics of \ncarbon removal are such that some countries have greater capacity \nfor land-based and geologic sinks. Countries with large sinks might \nseek to use them to give competitive advantages to their industrial \nor agricultural sectors, with a risk of less-stringent policies for \ndecarbonizing those sectors. In other words, if carbon removal is a \nnatural resource with finite capacity, the choices a country makes \nin allocating that resource have global-justice dimensions. Thus, \nresidual emissions can be seen as an emerging, important focal \npoint for climate justice and the UNFCCC negotiations, alongside \nemissions reductions goals, loss and damage, and climate finance. \nResearchers have an important role to play in producing a robust \nfoundation for those discussions.\n\nNature Climate Change | Volume 13 | April 2023 | 351\u2013358\n357\nArticle\nhttps://doi.org/10.1038/s41558-022-01592-2\nactors (countries, sectors, companies) will claim remaining residual \nemissions and corresponding negative emissions capacity, leaving \nless-powerful or less-well-organized actors unable to operate or, more \nlikely, to continue to operate illegally. Further, the ambiguity of residual \nemissions\u2014as a temporary measure while zero-carbon technologies \nare developed versus residual emissions as a long-term feature of the \nenergy system\u2014risks not just confusing publics and stakeholders, but \ndecreasing support for net-zero targets more broadly.\nThese questions may seem like far-off matters in a world where \nemissions have not even peaked. But 2050 is not so distant, and the \nscience is clear that fossil fuel production must rapidly be curtailed \nand most fossil fuel reserves must remain unextracted to meet a 1.5\u2009\u00b0C \ntemperature goal25. Publics, investors, planners and other decision \nmakers need greater clarity on the longer-term aims of net zero to \nguide decisions around fossil fuel phaseout as well as what sort of \nremoval efforts to invest in. Future expectations act in the present: our \nexpectations of 2050 inform choices made today. Many actors may see \nnet zero as a temporary state towards a net-negative society, but this \nvision is not yet evident in national strategies.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author con-\ntributions and competing interests; and statements of data and code \navailability are available at https://doi.org/10.1038/s41558-022-01592-2.\nReferences\n1.\t\nClimate Watch (World Resources Institute).\n2.\t\nClimate Change 2022: Mitigation of Climate Change. Contribution \nof Working Group III to the Sixth Assessment Report of the \nIntergovernmental Panel on Climate Change (eds Shukla, P. R. \net al.) (Intergovernmental Panel on Climate Change, 2022).\n3.\t\nLuderer, G. et al. Residual fossil CO2 emissions in 1.5\u20132\u2009\u00b0C \npathways. Nat. Clim. Change 8, 626\u2013633 (2018).\n4.\t\nFocus 2030: A Pathway to Net Zero Emissions (SF Environment, \n2019).\n5.\t\nHans, F., Day, T., R\u00f6ser, F., Emmrich, J. & Hagemann, M. Making \nLong-Term Low GHG Emissions Development Strategies a Reality \n(The 2050 Pathways Platform, 2020).\n6.\t\nWilliams, J. & Waisman, H. 2050 Pathways: A Handbook  \n(The 2050 Pathways Platform, 2017).\n7.\t\nAnastasia, O. Developing Mid-Century Long-Term Low Emission \nDevelopment Strategies (LT-LEDS) (Intergovernmental Panel on \nClimate Change, 2017).\n8.\t\nWaisman, H. et al. A pathway design framework for national low \ngreenhouse gas emission development strategies. Nat. Clim. \nChange 9, 261\u2013268 (2019).\n9.\t\nGovernment of Costa Rica. National Decarbonization Plan. United \nNations Climate Change https://unfccc.int/documents/204474 \n(2018).\n10.\t Ross, K., Schumer, C., Fransen, T., Wang, S. & Elliott, C. Insights \non the first 29 long-term climate strategies submitted to the \nUnited Nations Framework Convention on Climate Change. World \nResour. Inst. https://doi.org/10.46830/wriwp.20.00138 (2021).\n11.\t\nNet Zero by 2050 (IEA, 2021).\n12.\t Mohan, A., Geden, O., Fridahl, M., Buck, H. J. & Peters, G. P. \nUNFCCC must confront the political economy of net-negative \nemissions. One Earth 4, 1348\u20131351 (2021).\n13.\t van der Wijst, K., Byers, E., Riahi, K., Schaeffer, R. & van Vuuren, D. \nData for Figure SPM.5 - Summary for Policymakers of the Working \nGroup III Contribution to the IPCC Sixth Assessment Report  \n(Global Green Growth Institute, 2022).\n14.\t Fuss, S. et al. Negative emissions\u2014part 2: costs, potentials and \nside effects. Environ. Res. Lett. 13, 063002 (2018).\n15.\t Thoni, T. et al. Deployment of negative emissions technologies \nat the national level: a need for holistic feasibility assessments. \nFront. Clim. 2, 590305 (2020).\n16.\t Field Christopher, B. & Mach Katharine, J. Rightsizing carbon \ndioxide removal. Science 356, 706\u2013707 (2017).\n17.\t Boysen, L. R. et al. The limits to global-warming mitigation by \nterrestrial carbon removal. Earths Future 5, 463\u2013474 (2017).\n18.\t Fujimori, S. et al. Land-based climate change mitigation measures \ncan affect agricultural markets and food security. Nat. Food 3, \n110\u2013121 (2022).\n19.\t Dooley, K. & Kartha, S. Land-based negative emissions: risks for \nclimate mitigation and impacts on sustainable development.  \nInt. Environ. Agreem. 18, 79\u201398 (2018).\n20.\t Realmonte, G. et al. An inter-model assessment of the role of \ndirect air capture in deep mitigation pathways. Nat. Commun. 10, \n3277 (2019).\n21.\t Rogelj, J., Geden, O., Cowie, A. & Reisinger, A. Three ways to \nimprove net-zero emissions targets. Nature 591, 365\u2013368  \n(2021).\n22.\t Fankhauser, S. et al. The meaning of net zero and how to get it \nright. Nat. Clim. Change 12, 15\u201321 (2022).\n23.\t Hale, T. et al. Assessing the rapidly-emerging landscape of net \nzero targets. Clim. Policy 22, 18\u201329 (2022).\n24.\t SBTI Corporate Net-Zero Standard Version 1.0 (SBTI, 2021).\n25.\t Welsby, D., Price, J., Pye, S. & Ekins, P. Unextractable fossil fuels in \na 1.5\u2009\u00b0C world. Nature 597, 230\u2013234 (2021).\n26.\t Switzerland\u2019s Long-Term Climate Strategy. United Nations Climate \nChange https://unfccc.int/documents/268092 (The Federal \nCouncil, Government of Switzerland, 2021).\n27.\t On the Path to Climate Neutrality: Iceland\u2019s Long-Term Low \nEmission Development Strategy. United Nations Climate Change \nhttps://unfccc.int/documents/307770 (Government of Iceland \nMinistry of Environment and Natural Resources, 2021).\n28.\t The Long-Term Strategy under the Paris Agreement. United \nNations Climate Change https://unfccc.int/documents/307817 \n(Government of Japan, 2021).\n29.\t National Low-Carbon Strategy: The Ecological and Inclusive \nTransition Towards Carbon Neutrality. United Nations Climate \nChange https://unfccc.int/documents/268346 (Ministry for  \nthe Ecological and Solidary Transition, Government of France, \n2020).\n30.\t Nepal\u2019s Long-Term Strategy for Net-Zero Emissions. United \nNations Climate Change https://unfccc.int/documents/307963 \n(Government of Nepal, 2021).\n31.\t Sweden\u2019s Long-Term Strategy for Reducing Greenhouse Gas \nEmissions. United Nations Climate Change https://unfccc.int/\ndocuments/267243 (Ministry of the Environment, Government of \nSweden, 2020).\n32.\t Net Zero Strategy: Build Back Greener. United Nations Climate \nChange https://unfccc.int/documents/307547 (Government of \nthe United Kingdom, 2021).\n33.\t Australia\u2019s Long-Term Emissions Reduction Plan. United Nations \nClimate Change https://unfccc.int/documents/307803 \n(Australian Government Department of Industry, Science, Energy \nand Resources. Commonwealth of Australia, 2021).\n34.\t The Long-Term Strategy of the United States: Pathways to Net-Zero \nGreenhouse Gas Emissions by 2050. United Nations Climate \nChange https://unfccc.int/documents/308100 (United States \nDepartment of State and the United States Executive Office of the \nPresident, 2021).\n35.\t Grassi, G. et al. Carbon fluxes from land 2000\u20132020: bringing \nclarity on countries\u2019 reporting. Earth Syst. Sci. Data 14, 4643\u20134666 \n(2022).\n36.\t Buylova, A., Fridahl, M., Nasiritousi, N. & Reischl, G. Cancel \n(out) emissions? The envisaged role of carbon dioxide removal \n\nNature Climate Change | Volume 13 | April 2023 | 351\u2013358\n358\nArticle\nhttps://doi.org/10.1038/s41558-022-01592-2\ntechnologies in long-term national climate strategies. Front. Clim. \n3, 675499 (2021).\n37.\t Fajardy, M. & Mac Dowell, N. Recognizing the value of collabo\u00ad\nration in delivering carbon dioxide removal. One Earth 3, 214\u2013225 \n(2020).\n38.\t Arcusa, S. & Sprenkle-Hyppolite, S. Snapshot of the Carbon \nDioxide Removal certification and standards ecosystem  \n(2021\u20132022). Clim. Policy https://doi.org/10.1080/14693062.2022\n.2094308 (2022).\n39.\t Brander, M., Ascui, F., Scott, V. & Tett, S. Carbon accounting  \nfor negative emissions technologies. Clim. Policy 21, 699\u2013717 \n(2021).\n40.\t Honegger, M., Poralla, M., Michaelowa, A. & Ahonen, H.-M. Who is \npaying for carbon dioxide removal? Designing policy instruments \nfor mobilizing negative emissions technologies. Front. Clim. 3, \n672996 (2021).\n41.\t Honegger, M. et al. The ABC of governance principles for carbon \ndioxide removal policy. Front. Clim. 4, 884163 (2022).\n42.\t Mace, M. J., Fyson, C. L., Schaeffer, M. & Hare, W. L. Large\u2010scale \ncarbon dioxide removal to meet the 1.5\u2009\u00b0C limit: key governance \ngaps, challenges and priority responses. Glob. Policy 12, 67\u201381 \n(2021).\n43.\t Shue, H. Subsistence protection and mitigation ambition: \nnecessities, economic and climatic. Br. J. Polit. Int. Relat. 21, \n136914811881907 (2019).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons license, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons license, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons license and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this license, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2023\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01592-2\nMethods\nCountry long-term strategies were downloaded from the UNFCCC \nand were qualitatively coded in a spreadsheet by two independent \ncoders, a research assistant and a member of the research team, for \nthe following information:\n\t(1)\t Type of target (for example, carbon neutrality, net zero or other)\n\t(2)\t Coverage of target (GHGs or CO2)\n\t(3)\t Year of net zero, for countries with net-zero or carbon-neutral \ntargets\n\t(4)\t Whether there is a definition of residual emissions or hard-to- \nabate/remaining emissions and, if so, how it is introduced\n\t(5)\t Whether there is a quantitative projection of residual emissions \nat net zero and, if so, what the amount is\n\t(6)\t Sectoral breakdowns of residual emissions\n\t(7)\t The source and process of generating the projections (which \napproaches were used; whether they appeared to be top-down \nor bottom-up; which particular models were used to generate \nthem)\n\t(8)\t Mentions of public or stakeholder consultation or engagement\nIn a few cases, other government documents or sources were \nalso used for reference, including technical annexes for government \nstrategies.\nPercentages of current country emissions were derived from the \nWorld Resources Institute\u2019s Climate Watch platform at https://www.\nclimatewatchdata.org/ (ref. 1).\nCurrent-year emissions were derived from the 2019 emissions listed \nin UNFCCC inventories for total GHG emissions without LULUCF, at \nhttps://unfccc.int/process-and-meetings/transparency-and-reporting/\ngreenhouse-gas-data/ghg-data-unfccc/ghg-data-from-unfccc.\nRecent and current LULUCF data are from (ref. 35).\nThe coded data was used to generate the tables and figures  \nin the Article. The analysis is straightforward; the work was simply \nin extracting the amounts of residual emissions and sectoral break-\ndowns because these are not presented in a standard form across  \nthe documents, and in some cases they appear in charts but are  \nnot well explicated in the main text of the reports.\nReporting summary\nFurther information on research design is available in the Nature  \nPortfolio Reporting Summary linked to this article.\nData availability\nThe data analysed in the current study are provided in Supplementary \nData 1. The majority of the relevant data was extracted from publicly \navailable documents available from the UNFCCC at https://unfccc.int/\nprocess/the-paris-agreement/long-term-strategies. Percentages of  \ncurrent country emissions were derived from the World Resources  \nInstitute\u2019s Climate Watch platform at https://www.climatewatchdata.\norg. Current-year emissions were derived from the 2019 emissions  \nlisted in UNFCCC inventories for total GHG emissions without  \nLULUCF, at https://unfccc.int/process-and-meetings/transparency- \nand-reporting/greenhouse-gas-data/ghg-data-unfccc/ghg-data- \nfrom-unfccc. Recent and current LULUCF data are from (ref. 35).\nAcknowledgements\nThis work was supported by the Swedish Research Council Formas, \ngrant no. 2018-01686 (H.J.B. and W.C.) and grant no. 2019-01953  \n(all authors). We thank A. Palumbo-Compton for research assistance.\nAuthor contributions\nH.J.B. conceived the idea for the paper and led the analysis and \nwriting. W.C., J.F.L. and N.M. contributed to the analysis and \ndevelopment of the argument. All authors contributed to drafting, \nreviewing and editing the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41558-022-01592-2.\nCorrespondence and requests for materials should be addressed to \nHolly Jean Buck.\nPeer review information Nature Climate Change thanks William Lamb, \nMari\u00ebsse van Sluisveld and Clea Schumer for their contribution to the \npeer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\n1\nnature portfolio  |  reporting summary\nMarch 2021\nCorresponding author(s):\nHolly Buck\nLast updated by author(s): Dec 18, 2022\nReporting Summary\nNature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nNo custom code was used in this analysis.  Microsoft Excel was used to collect data.\nData analysis\nNo custom code was used in this analysis.  Microsoft Excel was used to analyze data.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and \nreviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A description of any restrictions on data availability \n- For clinical datasets or third party data, please ensure that the statement adheres to our policy \n \nThe data analysed in the current study are provided as supplement.  Most of the relevant data was extracted from publicly available documents available from the \nUNFCCC at https://unfccc.int/process/the-paris-agreement/long-term-strategies.  Percentages of current country emissions were derived from the World Resources \nInstitute\u2019s Climate Watch platform, at https://www.climatewatchdata.org.  Current year emissions were derived from the 2019 emissions listed in UNFCCC \ninventories for total GHG emissions without LULUCF, at https://unfccc.int/process-and-meetings/transparency-and-reporting/greenhouse-gas-data/ghg-data-\n\n2\nnature portfolio  |  reporting summary\nMarch 2021\nunfccc/ghg-data-from-unfccc.  Recent and current LULUCF data is from Grassi, G. et al. Carbon fluxes from land 2000-2020: bringing clarity on countries\u2019 reporting. \nEarth Syst Sci Data Discuss 2022, 1\u201349 (2022).\nHuman research participants\nPolicy information about studies involving human research participants and Sex and Gender in Research. \nReporting on sex and gender\nN/A\nPopulation characteristics\nN/A\nRecruitment\nN/A\nEthics oversight\nN/A\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThis study analyzes policy documents.  The data in these documents is both quantitative (i.e. reports of GHG emissions) and \nqualitative (e.g. narrative descriptions of stakeholder engagement strategies).\nResearch sample\nThe data was extracted from long-term low-emissions development strategies submitted to the UNFCCC under the Paris Agreement.  \nThese were used because this is the first body of policy documents to consistently include projections and discussion of mid-century \nclimate goals.\nSampling strategy\nAll the long-term climate strategies were analyzed; most analysis focused on coutnries with net-zero by 2050 goals.\nData collection\nThis was a document analysis with no human subjects.\nTiming\nAnalysis took place between April 2020 and March 2022.\nData exclusions\nNo data were excluded.\nNon-participation\nNo human subjects in this study.\nRandomization\nNo human subjects in this study.\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \n\n3\nnature portfolio  |  reporting summary\nMarch 2021\nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology and archaeology\nAnimals and other organisms\nClinical data\nDual use research of concern\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\n\n\n    Scientific Research Findings:", "answer": "Our results show that most countries with quantified projections are expecting large amounts of residual emissions by 2050, and that substantial carbon removal effort would therefore be needed. We found no consistent definition or use of the concept of residual emissions across the 50 long-term low-emissions development strategies (LT-LEDS) analysed. A majority of the LT-LEDS did not mention the concept of residual emissions at all, despite having a net-zero target. The Annex 1 country LT-LEDS that did quantify residual emissions at net zero projected substantial levels; on average, 18% of current emissions based on the lowest-level projections. Agriculture and industry represented the largest sources of residual emissions in the 15 Annex I countries that included sectoral breakdowns. While some countries plan for land-based removals to compensate for their residuals, the biggest emitters expect land-based removals to fall far short of residuals, indicating a need for technological removal.", "id": 49}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 13 | March 2023 | 244\u2013249\n244\nnature climate change\nArticle\nhttps://doi.org/10.1038/s41558-023-01597-5\nCross-national analysis of attitudes towards \nfossil fuel subsidy removal\nNiklas Harring\u2009\n\u200a\u20091\u2009\n, Erik J\u00f6nsson\u2009\n\u200a\u20091, Simon Matti\u2009\n\u200a\u20092, Gabriela Mundaca\u2009\n\u200a\u20093 & \nSverker C. Jagers1\nIn 2021, governments of 51 countries spent US$697\u2009billion on subsidizing \nfossil fuels. Removing fossil fuel subsidies is crucial not only for reducing \nCO2 emissions and making carbon pricing more effective, but also for \nmaking more valuable use of government funds. Currently, however, \nscientific evidence on the scale and scope of public attitudes towards  \nfossil fuel subsidy-removal policies is lacking, yet it is instrumental for \ngauging political feasibility. Furthermore, previous studies tend to focus  \non carbon pricing in the developed world only. Here we present a \ncomparative analysis of attitudes towards both carbon taxation and fossil \nfuel subsidy removal, focusing on five developing countries across four \ncontinents. It is found that (1) removing fossil fuel subsidies is not more \nundesirable than introducing carbon taxation and (2) the public has \nmore-positive attitudes towards subsidy removal if optimal use of the saved \nfiscal revenues is specified.\nTo reach the CO2-emission reduction targets of the Paris Agreement\u2019s \nNationally Determined Contributions, a growing number of countries \nare considering implementing domestic carbon taxes. These would \nincrease the price on fossil fuels (coal, natural gas, end-use electricity \nand petroleum) to decrease fossil fuel consumption (for example, Coali-\ntion of Finance Ministers). However, and repeatedly recognized during \nboth the 26th UNFCCC Conference of the Parties (COP26) meeting in \nGlasgow and the recently finalized COP27 in Sharm el-Sheikh, many \ncountries currently have policies that keep end-user prices artificially \nlow through subsidies. This encourages increases in both production \nand consumption of fossil fuels and thus effectively counteracts the \nintended objective of carbon pricing. In addition, subsidies repre-\nsent a burden on the governments\u2019 fiscal budgets through deficits \nand revenue losses. The Organisation for Economic Co-operation and \nDevelopment (OECD) found that tax breaks and spending programmes \n(fossil fuel support) in the G20 countries, linked to both the produc-\ntion and use of coal, oil, gas and other petroleum products, had risen \nto US$190\u2009billion in 2021, a level that is higher than in previous years \n(30% higher than in 2020)1. The OECD and International Energy Agency \n(IEA) have also estimated that governments in 51 countries provided \nUS$697.2\u2009billion in fossil fuel subsidies in 2021, doubling the amount \nfrom 20202, an amount that is three times the annual amount needed \nto eradicate global extreme poverty3.\nAll these mentioned costs are, however, only the direct costs of the \nsubsidies themselves. According to the International Monetary Fund \n(IMF), including also indirect costs (the contribution of fossil fuels to \nglobal warming, local air pollution and other externalities, and foregone \nconsumption tax) would increase the figure for annual fossil fuel subsi-\ndies by around US$6\u2009trillion, or 6.5% of global GDP3,4. They also find that \n45% of the benefits from direct fossil fuel subsidies goes to the richest \nquintile, while only 7% goes to the poorest 20% of the population5.\nRemoving subsidies on all fossil fuels simultaneously should be \nthe natural first step to reduce CO2 emissions6 since removing subsi-\ndies only on some fossil fuels will risk increasing the consumption of \nanother, still subsidized, fossil fuel (compare ref. 7). Particularly in devel-\noping countries, increasing fiscal revenues originated from savings \nfrom removed fossil fuel subsidies can be used for welfare-enhancing \nprojects (for example, investments in health care and education) and \nspurring economic growth8 and eradicate the regressivity of the exist-\ning subsidies.\nReceived: 16 March 2022\nAccepted: 6 January 2023\nPublished online: 23 February 2023\n Check for updates\n1Centre for Collective Action Research, Department of Political Science, University of Gothenburg, Gothenburg, Sweden. 2Political Science Unit, \nDepartment of Social Sciences, Technology and Arts, Lule\u00e5 University of Technology, Lule\u00e5, Sweden. 3Department of Geography and Spatial Sciences, \nUniversity of Delaware, Newark, DE, USA. \n\u2009e-mail: niklas.harring@pol.gu.se\n\nNature Climate Change | Volume 13 | March 2023 | 244\u2013249\n245\nArticle\nhttps://doi.org/10.1038/s41558-023-01597-5\nto investments that increase social and economic welfare systems lead \nto more-positive attitudes towards subsidy removal (refs. 26,30,31). At the \noutset, we report that optimal use of savings from subsidy removal has \npositive effects on public attitudes.\nIn our survey, the respondents were also asked about their social \nand economic characteristics, whether they own a fossil fuelled vehicle \nand their views regarding various climate change scenarios (Supple-\nmentary Information). In addition, we empirically analyse the effects \nof these variables on their policy attitudes.\nWe proceed from the well-established hypothesis that an impor-\ntant driver of policy attitudes is the balance of perceived personal \ncosts and benefits of a proposed policy32\u201335. First, we hypothesize that \nacceptance of removing fossil fuel subsidies will be lower than the cor-\nresponding attitude to introducing a carbon tax, as the former indicates \na more visible and direct loss of money for the consumer compared \nwith the indirect workings of a tax.\nH1: The public acceptance of removing subsidies on fossil fuels \nis lower than the public acceptance of introducing a CO2 tax on  \nfossil fuels.\nSecond, a range of studies (for example, refs. 36,37) have demon-\nstrated how individuals display more-positive attitudes towards poli-\ncies directed towards industry rather than towards themselves, in much \nthe same way as people in general tend to prefer less-stringent policies \nover more-coercive ones (for example, ref. 38). This might be due to both \ngeneral beliefs concerning how the proposed policy will have direct \nimplications for personal welfare (personal outcome expectancies)33 \nand distributional preferences driven by the attitude that industry \nrather than individuals should bear the main costs of climate change39. \nAs such, we hypothesize that people dislike policies that imply direct \npersonal costs more than policies aimed towards industry, even if these \nmight indirectly affect consumer prices, and that public acceptance \nof removing fossil fuel subsidies for private consumption therefore is \nlower than for those for industrial use.\nH2: The public acceptance of removing fossil fuel subsidies for \nprivate consumption is lower than the public acceptance of removing \nfossil fuel subsidies for industrial use.\nMoreover, a growing body of research concludes that negative \nattitudes towards price-based climate policy tools can be alleviated \nthrough policy design, in particular revenue recycling, where a price \nincrease is combined with a specified use of the available public \nfunds26,31 (but see ref. 30). Although research is somewhat inconclu-\nsive concerning the attitudinal effects of different forms of revenue \nrecycling (for example, fee-and-dividend solutions, increased invest-\nments in welfare systems (for example, education and health care) and \nusing revenues for specific climate-related projects), we nevertheless \nexpect that transparency in the use of generated public funds will \ntrigger more-positive attitudes compared with non-specified use of \npublic funds19,40. Such additional information aims to prevent peo-\nple\u2019s perception that the subsidy removal is only an increased cost for  \nthe household.\nH3: Compared with non-specified revenue use, the public accept-\nance of removing fossil fuel subsidies for private consumption is higher \nwhen revenue use is specified.\nPublic attitudes towards fossil fuel subsidy \nremoval\nTo determine how different policy designs affect public policy sup-\nport, as well as test our three hypotheses (H1, H2 and H3), we randomly \nassigned the respondents to one of a total of seven groups (Methods). \nOn a 0\u201310 scale, the average support is 6.22 for removing industrial-use \nsubsidies, 6.31 for removing subsidies on private consumption of fossil \nfuels and 6.33 for introducing a carbon tax. Apparently, the differences \nbetween these numbers are small. The statistical testing of the means \n(M) confirms this as well. When t testing the differences between the \nproposal of removing subsidies on private consumption (M\u2009=\u20096.31, \nThese issues have started to be acknowledged also by world lead-\ners, for example, in the Glasgow Climate Pact at the 2021 COP26, which \ncalls for \u201caccelerating efforts toward the phase-out of [\u2026] inefficient \nfossil fuel subsidies, recognizing the need for support toward a just \ntransition\u201d.9 The concept of just transitions implies recognizing and \nattempting to counteract the profound societal impacts that a shift \ntowards a low-carbon future implies, not the least in the form of job \nlosses due to fossil fuel industry decline and uneven distribution of \ncosts and benefits both within and between countries (for example,  \nref. 10). Given that such a shift may have immediate negative conse-\nquences for individuals and groups, the political feasibility of removing \nfossil fuel subsidies in any country highly depends on the degree of \npublic acceptability of such a policy11,12.\nNumerous scholars and policy experts have advocated putting \na price on carbon as a highly cost-effective way to reduce GHG emis-\nsions13,14. Introducing such policies has, however, become a vexing \nproblem for decision makers worldwide. The examples of Australia in \n2015, France in 2018 and Ecuador in 2019 demonstrate how widespread \nthe public\u2019s negative attitudes towards carbon taxes and removal of \nfossil fuel subsidies seem to be, and thus how difficult they are to imple-\nment. Several factors are known to determine policy attitudes, includ-\ning perceptions of fairness, effectiveness, political trust and climate \nconcern15 (compare refs. 16\u201321.) Political feasibility of carbon-pricing \nimplementation and subsidy removal requires that one measures and \nanalyzes how public opposition can be minimized. Such analyses are \nindeed crucial for stakeholders, policymakers and academics involved \nin climate change and policymaking. The empirical analysis of politi-\ncal feasibility, balancing effectiveness and cost efficiency with public \nacceptability, is imperative, especially as policymakers tend to be \nreluctant to introduce policies if levels of public acceptability are low22. \nFrom a theoretical perspective, understanding why certain policies \ngenerate negative perceptions, and the extent to which a design of a \npolicy measure affects mass policy attitudes, is of great interest since it \nspeaks to theories of policy feedback and how policy design can create \nits own constituency of support (for example, refs. 23,24)\nIn light of this, a number of recent experimental survey studies \nhave suggested policies that could make carbon pricing more readily \nacceptable to the public, for example, fee-and-dividend approaches25 \n(feebates), earmarking of tax revenues for necessary investments26 and \neven rhetorical shifts from \u2018tax\u2019 to \u2018fee\u201927. These studies have focused \nmostly on (1) carbon taxation and (2) the developed world. Far fewer \nstudies are concerned with public attitudes towards climate policy in \ndeveloping countries, and even fewer, if at all, with attitudes towards \nthe removal of fossil fuel subsidies as a climate change mitigation \nstrategy15. However, considering the literature focusing on both con-\ntextual drivers of climate policy attitudes (for example, ref. 21) and \ncross-national patterns in carbon pricing (for example, ref. 28), we \ndo not expect that attitudes and attitude formation differ systemati-\ncally between the Global North and Global South. We rather assume \nthat both attitudes and policy are sensitive to a range of complex and \ncountry-specific factors.\nBy using a 1\u2009\u00d7\u20097, pre-registered, factorial-design survey experiment \n(N\u2009=\u20096,636), we make the following contributions to the related litera-\nture. (1) We consider five developing countries (Ecuador, Egypt, India, \nIndonesia and Mexico) that currently subsidize both consumption and \nproduction of fossil fuels. We select these countries because they have \nsome of the highest levels of subsidies on consumption of fossil fuels29. \n(2) We analyse public attitudes in these countries towards (a) the intro-\nduction of a carbon tax and (b) the removal of subsidies on both indus-\ntrial and private consumption of fossil fuels. (3) We examine whether \nand how attitudes towards subsidy removal and carbon taxation may \ndiffer from each other. (4) We compare attitudes towards removal of \nsubsidies on private consumption of fossil fuels with those towards \nremoval of subsidies on fossil fuels for industrial use. (5) We study \nwhether policies that reallocate money spent on fossil fuel subsidies \n\nNature Climate Change | Volume 13 | March 2023 | 244\u2013249\n246\nArticle\nhttps://doi.org/10.1038/s41558-023-01597-5\ns.d.\u2009=\u20092.67) and the introduction of a carbon tax (M\u2009=\u20096.33, s.d.\u2009=\u20092.77), \nwe find no statistically significant differences (t(1,893.16)\u2009=\u2009\u20130.1604, \nP\u2009=\u20090.4363). The first hypothesis is thus rejected. Nor do we find \nany differences between attitudes toward removing subsidies on \nindustrial-use fossil fuels (M\u2009=\u20096.22, s.d.\u2009=\u20092.57) and subsidies on private \nconsumption of fossil fuels (M\u2009=\u20096.31, s.d.\u2009=\u20092.67) (t(1,896.18)\u2009=\u20090.6985, \nP\u2009=\u20090.7575). Hence, we reject H2.\nRevenue recycling and fossil fuel subsidy removal\nAs a next step, we investigate whether people\u2019s attitudes towards \nthe removal of subsidies on private consumption of fossil fuels are \nimpacted when four alternative uses of fiscal revenues saved from such \nremovals are part of the proposed policy. In addition to the proposal to \nremove subsidies on private consumption of fossil fuels, the respond-\nents were randomly assigned five different alternatives for revenue use: \ninvestments to enhance welfare in society (for example, education and \nhealth care), income tax reductions, investments in climate adaptation \nmeasures, cash transfers to the poor and most-affected households \nand no information about revenue use.\nWhen aggregating the groups where revenue use is specified \n(M\u2009=\u20096.49, s.d.\u2009=\u20092.59) and comparing them with the group with unspeci-\nfied revenue use (M\u2009=\u20096.31, s.d.\u2009=\u20092.67), we find a significant differ-\nence (t(1,428.25)\u2009=\u20091.88, P\u2009=\u20090.03). In line with H3, public acceptance \nof removing subsidies for private consumption of fuels is higher \nwhen revenue use is specified, as compared with non-specified rev-\nenue use. Considering the treatment groups separately, we find \nthat private-consumption subsidy removal reaches a higher level of \nacceptance if revenues are directed towards investments in welfare \nsystems (M\u2009=\u20096.59, s.d.\u2009=\u20092.55) compared with a non-specified use \n(t(1,893.68)\u2009=\u20092.36, P\u2009=\u20090.01) or towards investments in climate adap-\ntation (M\u2009=\u20096.62, a.d.\u2009=\u20092.65) compared with non-specified revenue use \n(t(1,887.99)\u2009=\u20092.57, P\u2009=\u20090.01). However, we do not find any statistically \nsignificant differences between a proposal to use fiscal revenues to \nreduce income taxes (M\u2009=\u20096.25, s.d.\u2009=\u20092.48) or to provide cash transfers \nto the poor and most-affected households (M\u2009=\u20096.49, s.d.\u2009=\u20092.66) and \nnon-specified revenue use (M\u2009=\u20096.31, s.d.\u2009=\u20092.67): (t(1,884.34)\u2009=\u2009\u20130.50, \nP\u2009=\u20090.69) and (t(1,894)\u2009=\u20091.53, P\u2009=\u20090.06), respectively. Taken together, \nwe cannot reject our third hypothesis. Attitudes towards removing \nsubsidies can turn more positive when revenue use is specified. How-\never, these results are also dependent on the type of revenue recycling \nproposed. Whereas investments overall drive more-positive attitudes, \nmonetary compensation (either to all or to the most affected) does not.\nCross-national comparison of public attitudes\nWhen we, more exploratorily, consider each of our countries (Ecua-\ndor, Egypt, India, Indonesia and Mexico) individually, we find that the \nattitudes towards fossil fuel subsidy removal are on the same level as \nattitudes towards the introduction of a carbon tax. In the comparison, \nEgypt constitutes an exception, with the least positive attitudes towards \nremoval of fossil fuel subsidies for industrial use (M\u2009=\u20095.4) and private \nconsumption (M\u2009=\u20095.3) compared with averages in the other countries \nof 6.2 for industrial use and 6.3 for private consumption (Fig. 1). Over-\nall, from our results, we can conclude that the resistance towards (or \nacceptance of) the removal of fossil fuel subsidies is on par with the \npublic opinion on introducing taxes on CO2.\nA commitment to use the tax money saved from removing existing \nsubsidies in a way that benefits stakeholders will increase the level of \npublic acceptance. At the country level, we find that the use of revenues \nfor \u2018investment in climate adaptation\u2019 is the most popular alternative \nin both Mexico and Ecuador, while it is the least popular alternative in \nEgypt (Fig. 2). These results indicate that there are potentially impor-\ntant country-specific characteristics that should be considered by \npolicymakers aiming to remove fossil fuel subsidies. In this explora-\ntive part of our study, we do not have any causal claims or hypotheses \nregarding mechanisms. However, factors such as cultural differences, \ntax levels and differences in welfare programmes could potentially \nexplain country variation in support for various uses of revenues.\nDiscussion\nContrary to our expectations, when investigating public opinion on the \nremoval of existing fossil fuel subsidies in five developing countries, \nwe do not find the attitude towards removal of existing subsidies to \n\u20132\n\u20131\n0\nEcuador\nEgypt\nIndia\nIndonesia\nAttitude\nAttitude\nAttitude\nAttitude\nAttitude\nAttitude\nMexico\nAll countries\n1\n2\n\u20132\n\u20131\n0\n1\n2\n\u20132\n\u20131\n0\n1\n2\n\u20132\n\u20131\n0\n1\n2\nTaxes (baseline)\nProduction subsidies\nPrivate consumption subsidies\n\u20132\n\u20131\n0\n1\n2\n\u20132\n\u20131\n0\n1\n2\nFig. 1 | Attitudes towards a tax on CO2 and removal of fossil fuel subsidies worldwide. Estimated average treatment effects. Points indicate the estimated effect; \nlines indicate 95% confidence intervals.\n\nNature Climate Change | Volume 13 | March 2023 | 244\u2013249\n247\nArticle\nhttps://doi.org/10.1038/s41558-023-01597-5\nbe more negative than that towards the introduction of a carbon tax. \nOur study is unique in its focus on fossil fuel subsidies, and there are, \ncurrently, not many studies with which we can compare our findings. \nTherefore, there are reasons to be cautious and not draw any conclu-\nsions regarding the level of support for fossil fuel subsidy removal in \nthese countries. Survey research is always sensitive to certain formula-\ntions and sampling strategies, and we know, from both previous studies \nand a range of real-world examples, that carbon pricing is indeed politi-\ncally challenging and that rising fuel prices have spurred resistance in \nmany countries across the world. However, one way of interpreting \nour results is that the public in fact considers a subsidy-removal policy \nas being equally acceptable (or not acceptable) as the introduction \nof a carbon tax. If this is the case, we should expect real-life sugges-\ntions for subsidy removal to be met with similar public opposition \nor acceptability as we have seen for other carbon-pricing measures. \nFurthermore, another finding from our study is that attitudes can be \naffected (in this case, increasing public acceptance) by combining a \npossible subsidy removal with a revenue-recycling policy. Yet again, \nwe call for more studies to be able to more thoroughly evaluate and \ncalibrate the size and strength of this effect. However, the results so \nfar correspond with previous research on carbon pricing, consistently \nshowing that revenue recycling increases support for such policies25\u201327. \nWe also find that the respondents\u2019 concern for climate change appears \nto be a strong driver of policy attitudes, which has also been previously \nshown to be a strong predictor of climate policy support in the Global \nNorth15, and finally that the impact of revenue recycling varies across \nthe five countries (compare ref. 31).\nThese findings may have important policy implications. First, \nour overall results concerning policy attitudes imply that removing \nsubsidies on fossil fuels may not present much more of a political chal-\nlenge than introducing carbon taxation. More important, by specifying \nalternatives for revenue recycling where public funds currently used \nfor subsidies are instead directed towards other public investments, \nthe level of acceptability may increase. However, the answer to the \nquestion of which specific investments are the most popular seems \nto be determined by national context. This further highlights the need \nfor careful country-specific empirical investigations to determine pre-\nferred options for revenue recycling among the public, before making \npolitical decisions to remove or roll back existing fossil fuel subsidies.\nThe study has other limitations. Although conducted over several \ncontinents, the total number of countries is small, and there are prob-\nably important nuances to be grasped by extending the sample to \nother countries using representative samples. Furthermore, neither \ndifferent levels of subsidy cuts nor any variation in how quickly the \nsubsidies should be removed is specified by the study. From previ-\nous research, however, we can expect that such elements of policy \ndesign do affect policy attitudes. In addition, fuel prices are always \nfluctuating, and the survey was conducted before the notable rise in \nenergy prices caused partly by the conflict in Ukraine. Furthermore, \nthe current experimental design has no control group to benchmark \nthe treatment groups against.\nA venue for future research is to study the degree to which public \nacceptance of various policy instruments is affected by such price fluc-\ntuations. Furthermore, future research should test similar hypotheses \nwhere respondents are provided with more information on how certain \npolicy instruments work. Misunderstanding, or lack of information, \nmight be part of the explanation to the similar support for removing \nsubsidies on fossil fuels and introducing a carbon tax. Developing a \nmore innovative design, including a control group, may also be a future \navenue to consider.\nOur study is one of rather few investigating attitudes towards \nclimate policy instruments in the Global South. As these countries \nare parties of the Paris Agreement and thereby struggle to find ways \nto limit their emissions, there is an increasing need for knowledge on \nattitudes and attitude formation in contexts outside the Global North. \nSimultaneously, there is a need for studies targeting actors\u2019 (citizens, \nconsumers, business and other stakeholders) acceptance of subsidy \nremovals in specific contexts. This need is palpable in both the develop-\ning and developed countries as subsidies on fossil fuel consumption \nand production do exist also within the OECD member states and \n\u20132\n\u20131\n0\nEcuador\nEgypt\nIndia\nIndonesia\nAttitude\nAttitude\nAttitude\nAttitude\nAttitude\nAttitude\nMexico\nAll countries\n1\n2\n\u20132\n\u20131\n0\n1\n2\n\u20132\n\u20131\n0\n1\n2\n\u20132\n\u20131\n0\n1\n2\n\u20132\n\u20131\n0\n1\n2\n\u20132\n\u20131\n0\n1\n2\nRevenue use unspecified (baseline)\nWelfare\nCash transfers to poor and most-afected households\nAdaptation\nGeneral reduction of income tax\nFig. 2 | Support for different proposals of revenue use from the removal of fossil fuel subsidies by country. Estimated average treatment effects. Points indicate \nthe estimated effect; lines indicate 95% confidence intervals.\n\nNature Climate Change | Volume 13 | March 2023 | 244\u2013249\n248\nArticle\nhttps://doi.org/10.1038/s41558-023-01597-5\nsince the formation of policy attitudes is probably driven by a range \nof complex country- and situation-specific factors.\nFinally, since climate change concern is a factor that significantly \naffects policy attitudes, further public and media attention assigned \nto climate change may make subsidy removals more conceivable and \nopen up promising avenues for developing countries to contribute to \nthe global mitigation of climate change. At the same time, fossil fuel \nsubsidy removal frees public funds for investing in social and economic \ndevelopment, which would be of great value and use in many develop-\ning countries.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author con-\ntributions and competing interests; and statements of data and code \navailability are available at https://doi.org/10.1038/s41558-023-01597-5.\nReferences\n1.\t\nUpdate on Recent Progress in Reform of Inefficient Fossil-Fuel  \nSubsidies that Encourage Wasteful Consumption (OECD and IEA,  \n2021); http://www.oecd.org/fossil-fuels/publicationsandfurther \nreading/OECD-IEA-G20-Fossil-Fuel-Subsidies-Reform-Update- \n2021.pdf\n2.\t\nSupport for Fossil Fuels Almost Doubled in 2021, Slowing \nProgress Toward International Climate Goals, According to New \nAnalysis from OECD and IEA (OECD, 2022); https://www.oecd.\norg/newsroom/support-for-fossil-fuels-almost-doubled-in-\n2021-slowing-progress-toward-international-clima\nte-goals-according-to-new-analysis-from-oecd-and-iea.htm\n3.\t\nAlternative Uses of Pre-tax Fossil-Fuel Subsidies per Year  \n(UNDP, 2021).\n4.\t\nParry, I., Black S. & Vernon, N. Still Not Getting Energy Prices Right: \nA Global and Country Update of Fossil Fuel Subsidies (IMF, 2021).\n5.\t\nCoady, D., Flamini, V. & Sears, L. The Unequal Benefits of Fuel \nSubsidies Revisited: Evidence for Developing Countries (IMF, 2015).\n6.\t\nMundaca, G. How much can CO2 emissions be reduced if fossil \nfuel subsidies are removed? Energy Econ. 64, 91\u2013104 (2017).\n7.\t\nFreire-Gonz\u00e1lez, J. & Ho, M. S. Policy strategies to tackle rebound \neffects: a comparative analysis. Ecol. Econ. 193, 107332 (2022).\n8.\t\nMundaca, G. Energy subsidies, public investment and \nendogenous growth. Energy Policy 110, 693\u2013709 (2017).\n9.\t\nGlasgow Climate Pact FCCC/PA/CMA/2021/L.16 (UNFCCC, 2021).\n10.\t Wang, X. & Lo, K. Just transition: a conceptual review. Energy Res. \nSoc. Sci. 82, 102291 (2021).\n11.\t\nZheng, Z., Liu, Z., Liu, C. & Shiwakoti, N. Understanding public \nresponse to a congestion charge: a random-effects ordered logit \napproach. Transp. Res. A 70, 117\u2013134 (2014).\n12.\t Jagers, S. C., Matti, S. & Nordblom, K. The evolution of public  \npolicy attitudes: comparing the mechanisms of policy support \nacross the stages of a policy cycle. J. Public Policy 40, 428\u2013448 \n(2020).\n13.\t Sterner, T. et al. Policy design for the Anthropocene. Nat. Sustain. \n2, 14\u201321 (2019).\n14.\t Effective Carbon Prices (OECD, 2013); https://doi.org/10.1787/9789\n264196964-en\n15.\t Bergquist, M., Nilsson, A., Harring, N. & Jagers, S. C. Meta-analyses \nof fifteen determinants of public opinion about climate change \ntaxes and laws. Nat. Clim. Change 12, 235\u2013240 (2022).\n16.\t Baranzini, A. et al. Carbon pricing in climate policy: seven \nreasons, complementary instruments, and political economy \nconsiderations. WIREs Clim Change 8, e462 (2017).\n17.\t Baranzini, A. & Carattini, S. Effectiveness, earmarking and \nlabeling: testing the acceptability of carbon taxes with survey \ndata. Environ. Econ. Policy Stud. 19, 197\u2013227 (2017).\n18.\t Kallbekken, S. & Aasen, M. The demand for earmarking:  \nresults from a focus group study. Ecol. Econ. 69, 2183\u20132190 \n(2010).\n19.\t Hammar, H. & Jagers, S. C. Can trust in politicians explain \nindividuals\u2019 support for climate policy? The case of CO2 tax.  \nClim. Policy 5, 613\u2013625 (2006).\n20.\t Thalmann, P. The public acceptance of green taxes: 2 million \nvoters express their opinion. Public Choice 119, 179\u2013217 (2004).\n21.\t Drews, S. & van den Bergh, J. C. J. M. What explains public support \nfor climate policies? A review of empirical and experimental \nstudies. Clim. Policy 16, 855\u2013876 (2016).\n22.\t Burstein, P. The impact of public opinion on public policy: a \nreview and an agenda. Polit. Res. Q. 56, 29 (2003).\n23.\t Stimson, J. A. Public Opinion in America: Moods, Cycles, and \nSwings (Westview Press, 1999).\n24.\t Wlezien, C. & Soroka, S. N. Political institutions and the opinion\u2013\npolicy link. West Eur. Polit. 35, 1407\u20131432 (2012).\n25.\t Jagers, S. C., Lachapelle, E., Martinsson, J. & Matti, S. Bridging  \nthe ideological gap? How fairness perceptions mediate the  \neffect of revenue recycling on public support for carbon taxes \nin the United States, Canada and Germany. Rev. Policy Res. 38, \n529\u2013554 (2021).\n26.\t Carattini, S., Kallbekken, S. & Orlov, A. How to win public support \nfor a global carbon tax. Nature 565, 289\u2013291 (2019).\n27.\t L\u00f6fgren, \u00c5. & Nordblom, K. Puzzling tax attitudes and labels.  \nAppl. Econ. Lett. 16, 1809\u20131812 (2009).\n28.\t Mahdavi, P., Martinez-Alvarez, C. B. & Ross, M. L. Why Do \nGovernments Tax or Subsidize Fossil Fuels? CGD Working Paper \n541 (CGD, 2020).\n29.\t Energy Subsidies, Tracking the Impact of Fossil-Fuel Subsidies  \n(IEA, 2021); https://www.iea.org/topics/energy-subsidies\n30.\t Mildenberger, M., Lachapelle, E., Harrison, K. & \nStadelmann-Steffen, I. Limited impacts of carbon tax rebate \nprogrammes on public support for carbon pricing. Nat. Clim. \nChange https://doi.org/10.1038/s41558-021-01268-3 (2022).\n31.\t Beiser-McGrath, L. F. & Bernauer, T. Could revenue recycling \nmake effective carbon taxation politically feasible? Sci. Adv. 5, \neaax3323 (2019).\n32.\t Bernauer, T. & McGrath, L. F. Simple reframing unlikely to  \nboost public support for climate policy. Nat. Clim. Change 6, \n680\u2013683 (2016).\n33.\t Kallbekken, S., Garcia, J. H. & Korneliussen, K. Determinants of \npublic support for transport taxes. Transp. Res. A 58, 67\u201378 (2013).\n34.\t Schuitema, G., Steg, L. & Forward, S. Explaining differences in \nacceptability before and acceptance after the implementation  \nof a congestion charge in Stockholm. Transp. Res. A 44,  \n99\u2013109 (2010).\n35.\t Lubell, M., Zahran, S. & Vedlitz, A. Collective action and citizen \nresponses to global warming. Polit. Behav. 29, 391\u2013413 (2007).\n36.\t Rinscheid, A., Pianta, S. & Weber, E. U. Fast track or slo-mo? Public \nsupport and temporal preferences for phasing out fossil fuel cars \nin the United States. Clim. Policy 20, 30\u201345 (2020).\n37.\t Harring, N., Jagers, S. C. & Matti, S. The significance of political \nculture, economic context and instrument type for climate policy \nsupport: a cross-national study. Clim. Policy 19, 636\u2013650 (2019).\n38.\t Attari, S. Z. et al. Preferences for change: do individuals  \nprefer voluntary actions, soft regulations, or hard regulations  \nto decrease fossil fuel consumption? Ecol. Econ. 68,  \n1701\u20131710 (2009).\n39.\t Cai, B., Cameron, T. A. & Gerdes, G. R. Distributional preferences \nand the incidence of costs and benefits in climate change policy. \nEnviron. Resour. Econ. 46, 429\u2013458 (2010).\n40.\t Kallbekken, S. & S\u00e6len, H. Public acceptance for environmental \ntaxes: self-interest, environmental and distributional concerns. \nEnergy Policy 39, 2966\u20132973 (2011).\n\nNature Climate Change | Volume 13 | March 2023 | 244\u2013249\n249\nArticle\nhttps://doi.org/10.1038/s41558-023-01597-5\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons license, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons license, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons license and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this license, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2023\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-023-01597-5\nMethods\nWe conducted an online survey experiment (carried out through You-\nGov) in five countries. Our sample is based on pooled groups from \nEcuador, Egypt, Mexico, Indonesia and India, which all have substantial \nconsumption- and production-based fossil fuel subsidies. We had \nslightly more than 1,000 respondents in Ecuador and slightly more than \n1,400 in each of the other countries, all of whom were asked about their \nsupport/acceptance of climate policy introduction. We use a 1\u2009\u00d7\u20097 facto-\nrial survey experiment where respondents participating in the study \nwere randomly exposed to different kinds of policy measures (treat-\nments), which they were asked to evaluate (hypotheses pre-registered \nat OSF Registries41). (1) One group gave their opinion about the proposal \nof introducing a carbon tax in their country (as a point of reference \nfor us to compare with the other proposals). (2) One group gave their \nopinion about the proposal of removing the current industrial subsides \nto fossil fuels in their country. (3) One group gave their opinion about \nthe proposal of removing the current private-consumption subsides \nto fossil fuels in their country. Four different groups gave their opinion \nabout the proposal of removing the current consumption subsides on \nfossil fuels in their country plus any of the following additional policies: \n(4) use the surplus funding for general welfare purposes (for example, \nimproved health or education), (5) use the surplus funding to com-\npensate for a general reduction of the income tax, (6) use the surplus \nfunding for climate change adaptation projects (for example, flooding \nprevention) and (7) use the surplus funding for cash transfer to the \npoor most-affected households to keep their welfare levels unchanged.\nFollowing previous research demonstrating how factors at the \nindividual level affect policy attitudes, the study includes both beliefs \n(climate concern) and standard socioeconomic items (age, sex, income, \neducation and urban/rural place of residence).\nSample and respondents\nThe samples are based on quota criteria. That is, the probability for \neach individual who could theoretically be included is not determined \nin advance but is based on their demographic background information, \nsuch as gender, age and region, from population statistics/census from \neach country.\nRespondents participating in the study were randomly exposed \nto different kinds of policy measures (treatments). They did not know \nthe treatment group to which they had been assigned. Subsequent to \nthe question on policy support, they were asked to state their evalu-\native response to the specific policy. The respondents also answered \nsurvey questions regarding their gender, age, educational background, \nhousehold income level, area of residence and climate concern.\nData collection\nData were collected by YouGov. YouGov uses their proprietary panels \nand proprietary sampling technology. YouGov begins by framing quo-\ntas on the basis of the census of the named populations. This frame is \nthe basis on which the sampling software controls the flow of members \ninto each survey. The sampling system will randomly select from each \npanel and allocate to surveys according to the quotas set. Panellists \nreceive an invitation email containing a survey link. When they access \nthe link, the router checks against quotas on all live surveys and allo-\ncates them to a survey they qualify for.\nStatistical analysis\nAll the samples from the different countries were pooled when testing \nH1, H2 and H3. With 1,400 respondents in four countries and 1,000 \nrespondents in one country, the total sample contained 6,600 respond-\nents. These were then divided into seven groups (1,000 respondents \nin each group). To test H1, H2 and H3, we used independent-sample \none-sided t tests and ordinary least-squares regression models with \nrobust standard errors (results reported in Supplementary Informa-\ntion). We used the standard P\u2009<\u2009.05 criterion for determining whether \nthere are differences between the groups. Hypotheses H1, H2 and H3 \nwere supported if the null was rejected, and the estimates are statisti-\ncally significant and have the expected signs and directions for both \nthese tests. To test H3, group 3 was compared with an aggregated group \nbased on group 4, group 5, group 6 and group 7. For the exploratory part \nwhere we investigated the role of individual factors for policy support, \nwe used ordinary least-squares models.\nEthics\nThis study has been reviewed and approved by the legal division of \nLule\u00e5 University. In addition, the survey company (YouGov) has all the \nrequired permits and obtained informed consent from all participants.\nReporting summary\nFurther information on research design is available in the Nature Port-\nfolio Reporting Summary linked to this article.\nData availability\nData for replication are available via the Harvard Dataverse 41: https://\ndoi.org/10.7910/DVN/0SU8CJ\nCode availability\nThe code for the statistical analysis is available via the Harvard Data-\nverse41: https://doi.org/10.7910/DVN/0SU8CJ\nReferences\n41.\t Harring, N., J\u00f6nsson, E. Matti, S., Mundaca, G. & Jagers, S. C. \nReplication data for: cross-national analysis of attitudes towards \nfossil fuel subsidy removal (OSF Registries, 2022, and Harvard \nDataverse, 2023); https://doi.org/10.17605/OSF.IO/89CWY; \nhttps://doi.org/10.7910/DVN/0SU8CJ\nAcknowledgements\nWe are grateful for financial support from FORMAS\u2014the Swedish \nresearch council for Sustainable Development, 2019-00916, 2019-\n02005; the Swedish Research Council, 2016-03058; the Swedish \nEnergy Agency, 2019-006655.\nAuthor contributions\nS.C.J. initiated the study. N.H., E.J., S.M., G.M. and S.C.J. conceptualized \nthe paper and designed the survey experiments and contributed \nto the interpretation of the results. E.J. performed the analyses and \nimplemented the data presentation and visualization with contribution \nfrom N.H., S.M., G.M. and S.C.J. G.M. provided statistics on fossil fuel \nconsumption, production and subsidies. Finally, N.H., E.J., S.M., G.M. \nand S.C.J. wrote the main manuscript.\nFunding\nOpen access funding provided by University of Gothenburg.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s41558-023-01597-5.\nCorrespondence and requests for materials should be addressed  \nto Niklas Harring.\nPeer review information Nature Climate Change thanks Ay\u015fe \nUydurano\u011flu and the other, anonymous, reviewer(s) for their \ncontribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\n1\nnature research  |  reporting summary\nApril 2020\nCorresponding author(s):\nDBPR NCLIM-22030453\nLast updated by author(s): Mar 29, 2022\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nNo software was used for data collection.\nData analysis\nStata version 16.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and \nreviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nData will be available at https://osf.io/89cwy upon publication.\n\n2\nnature research  |  reporting summary\nApril 2020\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThe study uses a experimental 1x7 factorial design where quantitative survey data were used.\nResearch sample\nThe sample consists of 1009 respondents from Ecuador, 1407 respondents from Egypt, 1416 respondents from India, 1402 \nrespondents from Indonesia and 1402 respondents from Mexico. The sample is representative with respect to gender and age.\nSampling strategy\nData were collected by YouGov. YouGov used their proprietary panels and proprietary sampling technology. A power calculation \n(G*Power, Faul et al., 2007) based on effect size Cohen's d = 0.15 suggests a group sample size of 699 to ensure sufficient power \n(two-tailed, \u03b1 = .05, \u03b2-1 = .80). We have group sample sizes of minimum 1 000 respondents.\nData collection\nData were collected through an online survey where the respondents were randomly assigned to the seven different experiment \ngroups.\nTiming\nEcuador: June 30th \u2013 July 14th 2021. Egypt: June 29th \u2013 July 9th 2021. Indonesia: June 29th \u2013 July 9th 2021. India: June 29th \u2013 July \n10th 2021. Mexico: June 29th \u2013 July 10th 2021.\nData exclusions\nNo respondents were excluded except in the explorative regression models, where respondents with missing values on covariates \nwere excluded.\nNon-participation\nNo participants dropped out/declined participation.\nRandomization\nRespondents were randomly assigned to the seven experiment groups.\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology and archaeology\nAnimals and other organisms\nHuman research participants\nClinical data\nDual use research of concern\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nHuman research participants\nPolicy information about studies involving human research participants\nPopulation characteristics\nSee above.\nRecruitment\nParticipants were internet users and given a minor economic incentive to participate via YouGov\nEthics oversight\nApproved by the Legal Division of Lule\u00e5 University of Technology\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\n\n\n    Scientific Research Findings:", "answer": "We find that opposition to removal of existing subsidies is not greater than opposition to possible carbon pricing implementation. Our results indicate that the public in fact considers a subsidy removal policy as being equally desirable (or undesirable) as the introduction of a carbon tax. If so, we should expect suggestions for subsidy removal to be met with similar public opposition or support as has been seen for other carbon pricing measures. There are few studies with which we can compare our findings. Therefore, there are reasons to be cautious in making general conclusions about the extent to which there will be support for the removal of fossil fuel subsidies, at least in the countries we studied. Furthermore, the public in different countries may turn out to have different preferences for how additional revenues from either removing subsidies or collected taxes should be used .", "id": 50}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 13 | January 2023 | 48\u201357\n48\nnature climate change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nCooperative adaptive management of the \nNile River with climate and socio-economic \nuncertainties\nMohammed Basheer\u2009\n\u200a\u20091, Victor Nechifor\u2009\n\u200a\u20092,3, Alvaro Calzadilla\u2009\n\u200a\u20092, \nSolomon Gebrechorkos4, David Pritchard5, Nathan Forsythe\u2009\n\u200a\u20095, \nJose M. Gonzalez\u2009\n\u200a\u20091, Justin Sheffield4, Hayley J. Fowler5,6 & Julien J. Harou\u2009\n\u200a\u20091,7\u2009\nThe uncertainties around the hydrological and socio-economic implications \nof climate change pose a challenge for Nile River system management, \nespecially with rapidly rising demands for river-system-related services and \npolitical tensions between the riparian countries. Cooperative adaptive \nmanagement of the Nile can help alleviate some of these stressors and \ntensions. Here we present a planning framework for adaptive management \nof the Nile infrastructure system, combining climate projections; \nhydrological, river system and economy-wide simulators; and artificial \nintelligence multi-objective design and machine learning algorithms. \nWe demonstrate the utility of the framework by designing a cooperative \nadaptive management policy for the Grand Ethiopian Renaissance Dam that \nbalances the transboundary economic and biophysical interests of Ethiopia, \nSudan and Egypt. This shows that if the three countries compromise \ncooperatively and adaptively in managing the dam, the national-level \neconomic and resilience benefits are substantial, especially under climate \nprojections with the most extreme streamflow changes.\nHuman activities have increased the atmospheric concentration of \nGHGs, leading to warming of the Earth\u2019s land, atmosphere and ocean1\u20134. \nGlobal initiatives such as the Paris Agreement5 and the Sustainable \nDevelopment Goals6 aim to reduce the impacts of climate change by lim-\niting the rise in global temperature and stepping up adaptation efforts. \nHowever, the global mean temperature continued to rise over the past \ntwo decades, from 0.89\u2009\u00b0C in 2001\u20132010 to 1.09\u2009\u00b0C in 2011\u20132020 above \npre-industrial conditions1. The Nile Basin (Fig. 1) faces the threat of cli-\nmate change alongside water scarcity, rapidly rising pressures on water \nresources due to population and economic growth, and a politically \ncomplex transboundary water management system. The Nile Basin is \nlocated in northeastern Africa, occupies around 10% of the continent\u2019s \narea and extends over 11 countries. Supplementary Section 1 provides \nfurther information on the Nile geography.\nSince the beginning of the twentieth century, several large water \ninfrastructure projects have been constructed on the Nile River to \nreduce the spatial and temporal variabilities of the river flow and to facil-\nitate water supply, flood control and hydropower generation. Most of \nthese infrastructures are in Egypt, Sudan and Ethiopia. Currently, most \nof the consumptive usage of the Nile streamflow, measured at Aswan, is \nlocated in Sudan and Egypt7,8. Over the past two decades, Ethiopia has \nbeen increasing its use of the Nile, primarily for hydropower generation. \nThe construction of the Grand Ethiopian Renaissance Dam (GERD) on \nthe Nile near the Ethiopian\u2013Sudanese border (Fig. 1) flared up political \nReceived: 16 May 2022\nAccepted: 10 November 2022\nPublished online: 9 January 2023\n Check for updates\n1Department of Mechanical, Aerospace and Civil Engineering, University of Manchester, Manchester, UK. 2Institute for Sustainable Resources, University \nCollege London, London, UK. 3Joint Research Centre, European Commission, Seville, Spain. 4School of Geography and Environmental Science, \nUniversity of Southampton, Southampton, UK. 5School of Engineering, Newcastle University, Newcastle upon Tyne, UK. 6Tyndall Centre for Climate \nChange Research, Newcastle University, Newcastle upon Tyne, UK. 7Department of Civil, Environmental and Geomatic Engineering, University College \nLondon, London, UK. \n\u2009e-mail: julien.harou@manchester.ac.uk\n\nNature Climate Change | Volume 13 | January 2023 | 48\u201357\n49\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nGERD negotiations was a series of meetings held from November 2019 \nto February 2020, with the United States administration and the World \nBank as observers18. These meetings produced a proposal for the dam\u2019s \ninitial filling and long-term operation (hereon referred to as the Wash-\nington draft proposal)19, but the proposal was unacceptable to Ethiopia, \nwhich opined that the proposal would limit power generation from the \ndam and restrain future development20.\nAlongside this changing political landscape, the quantity and \nintensity of rainfall and streamflow in the Nile Basin have changed \nover the past two decades and are expected to continue to change due \nto climate change. However, the direction and magnitude of future \nchanges to the Nile climate are uncertain, stemming from different \nscenarios, modelling and downscaling choices21,22. In the Nile context, \nthese inconsistencies in climate projections diminish the value of using \none projection or a multi-model ensemble mean in climate adaptation \nplanning22,23.\ntensions between Ethiopia, Sudan and Egypt on Nile water use9\u201312. The \nconstruction of the GERD started in 2011, and when completed, the dam \nwill have a total storage capacity of 74 billion cubic meters (bcm) and an \ninstalled power capacity of 5,150\u2009MW. The total storage capacity of the \nGERD is equivalent to 1.5 times the historical mean annual river flow at \nthe dam location. The dam is expected to result in a range of opportu-\nnities and risks to Sudan and Egypt. Subject to coordination and data \nsharing, hydropower generation, irrigation water supply reliability and \nflood control in Sudan could improve because of the dam13\u201316. Still, the \ndam will probably produce adverse environmental impacts and losses \nto recession agriculture in Sudan17. For Egypt, the GERD is expected to \nreduce hydropower generation and impose irrigation water deficits if \nthere is no coordination on managing multi-year droughts7,12.\nDespite over a decade of negotiations, there is still disagreement \nand political tension between Ethiopia, Sudan and Egypt on the GERD\u2019s \ninitial filling and long-term operation. A milestone in the tripartite \n0\n200\n400\n600\n800\n100\nkm\nBlue Nile\n30\u00b0 N\n26\u00b0 N\n22\u00b0 N\n18\u00b0 N\n14\u00b0 N\n10\u00b0 N\n6\u00b0 N\n2\u00b0 N\n2\u00b0 S\n48\u00b0 E\nHAD\nGERD\nOther large dams\nLake\nNile River\nNile Basin\nBlue Nile Basin\nWhite Nile Basin\nTekeze-Atbara Basin\nNational boundary\nSudd wetlands\n44\u00b0 E\n40\u00b0 E\n36\u00b0 E\n32\u00b0 E\n28\u00b0 E\nEgypt\nN\n24\u00b0 E\nMediterranean Sea\nSudan\nDemocratic\nRepublic of\nthe Congo\nRwanda\nBurundi\nEthiopia\nKenya\nUganda\nTanzania\nWhite Nile\nBahr\nEl-Ghazal\nMain Nile\nTekeze-Atbara\nLake\nTana\nRed Sea\nEritrea\nLake\nVictoria\nSouth\nSudan\nSobat\nBahr\nEl-Jebel\nFig. 1 | The Nile Basin and its major tributaries and dams. The map shows the extents of the basins of the three main tributaries of the Nile: the Blue Nile, the White \nNile and the Tekeze-Atbara.\n\nNature Climate Change | Volume 13 | January 2023 | 48\u201357\n50\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nThe large uncertainties associated with the water resource implica-\ntions of changes in socio-economic and climate systems motivate adap-\ntive infrastructure development plans24\u201329. Several approaches have been \nproposed to enable planning under deep uncertainty26,30; these allow \ndesigning robust and flexible plans that maximize resilience and mini-\nmize investment costs on the basis of, for example, adaptation tipping \npoints31, dynamic adaptive planning32 and dynamic adaptive policy path-\nways24,33. Several recent studies have applied such adaptive methods to \nplanning water resource systems in various contexts25,34\u201338. For instance, a \nreservoir adaptive planning framework has been developed to explicitly \nconsider learning about climate uncertainty over time35, and other stud-\nies have optimized the indicators, actions and/or thresholds in the design \nprocess of adaptive plans for water resource systems33,36,39,40. However, \nnone of the previous studies considered engineering performance \nalongside economy-wide performance in the design process of climate \nadaptation plans for large water infrastructure systems, even though \nthe ultimate goal of building and operating infrastructure is typically \nto stimulate economic development and generate economy-wide gains.\nHere we introduce a planning framework for adaptive management \nof river infrastructure systems to consider both the socio-economic \nand hydrological uncertainties of climate change and policy.  \nThe framework uses climate and socio-economic data from the Coupled \nModel Intercomparison Project 6 (CMIP6)41 to drive integrated hydro-\nlogical, economy-wide and river system simulators of the Nile Basin. \nOur adaptive planning framework uses artificial-intelligence-based \nalgorithms to design efficient adaptive plans for climate change on \nthe basis of thousands of iterations between the algorithms and the \nintegrated simulators. We use the framework to design a coopera-\ntive adaptive management policy for the GERD (for 2020\u20132045) that \nconsiders economic and river system interests of Ethiopia, Sudan and \nEgypt. The results demonstrate that if the GERD is adaptively managed, \nthe economic and resilience benefits to individual countries (especially \nunder extreme climate projections) are substantially larger than with \nless adaptive responses such as the Washington draft proposal.\nAdaptive planning framework for Nile \nmanagement\nIn this study, we introduce an adaptive planning framework for \nmanaging the Nile infrastructure system in the face of climate and \nsocio-economic changes, including four interconnected stages  \n(Fig. 2): (1) selecting plausible climate change projections, (2) simu-\nlating the hydrological implications of the selected projections,  \nClimate\nprojections\nHydrological \nsystem\nEconomy\nClimate\nparameters \nRiver \ninfrastructure \nsystem\nSSPs\nChanges in water\nand electricity\ndemands driven\nby economic\nperformance\nChanges in non-hydro\ngeneration capacity\nbased on investment\nbehaviour\nWater and \nelectricity supplies \nRiver system\nstatus\nMonitor\nAdaptation\ndecisions\nClimate change scenarios\nInterim\nadaptation\nmeasures\nShort-term\nplans\n1\n1\n1\n2\n2\n3\n3\n4\nLong-term\nadaptation\nmeasures\nRadiative forcings\nClimate \npolicies \nPopulation\ngrowth\nprojections\nLabour\ngrowth\nprojections\nSectoral\nproductivity\nprojections\nProjections of municipal\nwater demands driven\nby population growth\nRiver \nstreamflow\nprojections \nProjections of evaporation \nrates from open water bodies \nProjections of agricultural\nevapotranspiration demands \nFig. 2 | Adaptive planning framework for Nile infrastructure management in the face of climate and socio-economic changes. The framework includes four \nstages numbered from 1 to 4.\n\nNature Climate Change | Volume 13 | January 2023 | 48\u201357\n51\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nSSP585 (ACCESS-CM2)\nSSP245 (ACCESS-ESM1-5)\nSSP585 (ACCESS-ESM1-5)\nSSP126 (BCC-CSM2-MR)\nSSP245 (CESM2)\nSSP370 (CESM2)\nSSP585 (CESM2-WACCM)\nSSP126 (GFDL-ESM4)\nSSP245 (GFDL-ESM4)\nSSP245 (GISS-E2-1-G)\nSSP370 (GISS-E2-1-G)\nSSP370 (GISS-E2-1-G)\nSSP585 (GISS-E2-1-G)\nSSP370 (IPSL-CM6A-LR)\nSSP245 (MIROC6)\nSSP245 (MPI-ESM1-2-HR)\nSSP370 (MRI-ESM2-0)\nSSP585 (MRI-ESM2-0)\nSSP126 (UKESM1-0-LL)\nSSP585 (UKESM1-0-LL)\nSSP245 (ACCESS-ESM1-5*)\nSSP585 (ACCESS-ESM1-5*)\nSSP245 (CESM2*)\nSSP370 (CESM2*)\nSSP585 (CESM2-WACCM*)\nSSP245 (MIROC6*)\nSSP126 (UKESM1-0-LL*)\nSSP585 (UKESM1-0-LL*)\nSSP585 (ACCESS-CM2*)\n200\n300\n400\n500\n600\n700\n2050\nAnnual naturalized flow (bcm)\nYear\nSSP585 (UKESM1-0-LL)\nSSP585(ACCESS-ESM1-5*)\nNile Basin\n30\n60\n90\n120\n150\nAnnual naturalized flow (bcm)\nYear\nSSP585 (UKESM1-0-LL)\nSSP585 (ACCESS-ESM1-5*)\nBlue Nile Basin\n200\n250\n300\n350\n400\nAnnual naturalized flow (bcm)\nSSP585 (UKESM1-0-LL)\nSSP585 (ACCESS-ESM1-5*)\n White Nile Basin\n0\n20\n40\n60\n80\n2010\n2020\n2030\n2040\n2010\n2020\n2030\n2040\n2050\n2050\nYear\nYear\n2010\n2020\n2030\n2040\n2010\n2020\n2030\n2040\n2050\nAnnual naturalized flow (bcm)\nSSP585 (UKESM1-0-LL)\nSSP585 (ACCESS-ESM1-5*)\nTekeze-Atbara Basin\n\u201320\n0\n20\n40\n60\nEOC precipitation change over the Nile (%)\n1\n2\n3\n4\n5\n6\na\nb\nc\nd\ne\nf\nEOC temperature change over the Nile (\u00b0C)\nSSP585\n(UKESM1-0-LL)\nSSP585\n(ACCESS-ESM1-5*)\nSSP5 (8.5 W m\u20132)\nSSP3 (7.0 W m\u20132)\nSSP2 (4.5 W m\u20132)\nSSP1 (2.6 W m\u20132)\n\u201311\n0\n20\n40\n60\n80\n95\nNile streamflow change, 2021\u20132050 (%)\n4\n6\n8\nMean annual economic growth, 2021\u20132050 (%)\n0.6\n0.8\n1.0\n1.2\n1.4\n1.6\n1.8\nMean annual population growth, 2021\u20132050 (%)\nEthiopia\nSudan\nEgypt\nEthiopia\nSudan\nEgypt\nEthiopia\nSudan\nEgypt\nEthiopia\nSudan\nEgypt\nSSP5\nSSP3\nSSP2\nSSP1\n20\n35\n50\n65\nMean percentage urban population, 2021\u20132050 (%)\nFig. 3 | Impacts of climate change on the Nile under multiple projections.  \na, Changes in precipitation, naturalized streamflow and temperature. b, Socio-\neconomic implications for Ethiopia, Sudan and Egypt. c\u2013f, 30-year moving \naverage of the annual naturalized streamflow. The labels in parentheses are \nthe names of the climate models. The change in naturalized streamflow is the \npercentage change in the mean in 2021\u20132050 relative to 1981\u20132016. EOC change \nin precipitation is the percentage change in the mean in 2071\u20132100 relative to \n1981\u20132010. EOC change in temperature is the absolute change in the mean in \n2071\u20132100 relative to 1981\u20132010. The climate projections marked with thick \nblack outlines in a and with asterisks in c\u2013f were synthesized to address the \nEastern African Paradox (see Methods for the details).\n\nNature Climate Change | Volume 13 | January 2023 | 48\u201357\n52\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\n(3) simulating the economy-wide and river system infrastructure per-\nformance under the selected projections, and (4) designing an adap-\ntive plan for managing river system infrastructure. The four stages are \nnumbered in Fig. 2 from 1 to 4.\nIn the first stage, 29 transient climate change projections to \n2100 were constructed on the basis of bias-corrected CMIP6 Tier 1 \nsimulations41. Spanning the Shared Socio-economic Pathways (SSPs), \n20 general circulation model (GCM) simulations were selected to \nrepresent the joint distribution of end-of-century (EOC; 2071\u20132100) \nchanges in precipitation and temperature over the Nile Basin according \nto the CMIP6 ensemble. Additionally, nine projections with decreas-\ning precipitation trends were synthesized to counteract the uncertain \nEficient design\na\nb\nc\nd\ne\nf\ng\nh\ni\nDirection of better performance\nDirection of better performance\nChange in\nGERD\u2019s\nmean annual \nenergy\n(GWh)\nFavourable Sudanese design\nFavourable Ethiopian design\nFavourable Egyptian design\nChange in\nSudan\u2019s mean\nannual \nirrigation\n(bcm)\nChange in\nSudan\u2019s\nmean annual \nhydropower\n(GWh)\nChange in\nEgypt\u2019s mean \naccumulated \nreal GDP\n(US$ billion)\nChange in Ethiopia's real GDP\n(US$ billion)\nChange in Sudan\u2019s \nmean annual irrigation (bcm yr\u20131)\nChange in Sudan\u2019s \nmean annual hydropower (TWh yr\u20131)\nChange in Egypt\u2019s \nmean annual irrigation (bcm yr\u20131)\nChange in Egypt\u2019s \nmean annual hydropower (TWh yr\u20131)\nChange in\nEthiopia\u2019s mean\naccumulated \nreal GDP\n(US$ billion)\nChange in\nGERD\u2019s\nmean firm \npower\n(MW)\nChange in\nSudan\u2019s mean\naccumulated \nreal GDP\n(US$ billion)\nChange in\nEgypt\u2019s mean\nannual \nirrigation\n(bcm)\nChange in\nEgypt\u2019s mean\nannual \nhydropower\n(GWh)\nCompromise design\nZero change line (i.e., Washington draft proposal)\n2\n1\n0\n\u20131\n\u20132\n\u20133\n2\n1\n0\n\u20131\n\u20132\n\u20133\n2\n1\n0\n\u20131\n\u20132\n\u20133\n66.0\n38.8\n11.6\n\u201315.6\n\u201342.8\n\u201370.0\n490\n300\n110\n\u201380\n\u2013270\n\u2013460\n490\n300\n110\n\u201380\n\u2013270\n\u2013460\n490\n300\n110\n\u201380\n\u2013270\n\u2013460\n0.9\n0.6\n0.3\n\u20130.1\n\u20130.4\n\u20130.7\n0.9\n0.6\n0.3\n\u20130.1\n\u20130.4\n\u20130.7\n\u20135\n0\n5\n10\n15\n\u20134\nChange in Sudan\u2019s real GDP\n(US$ billion)\n\u20138\n\u20136\n\u20134\n\u20132\n0\n2\nChange in Egypt\u2019s real GDP\n(US$ billion)\n\u20132\n0\n2\n4\n6\n\u20131\n0\n1\n2\n3\n4\nChange in GERD\u2019s \nmean annual energy (TWh yr\u20131)\n\u20131.0\n\u20131.0\n\u20131.5\n\u20131.0\n\u20130.5\n0\n0.5\n1.0\n1.5\n\u20130.5\n0\n0.5\n\u20130.5\n0\n0.5\n1.0\n1.5\n\u20130.5\n0\n0.5\nFig. 4 | Trade-offs and synergies between Ethiopian, Sudanese and Egyptian \neconomy-wide and river system performance objectives. a, Parallel \ncoordinates plot of the performance under efficient designs of an adaptive  \nGERD management policy across 29 climate change projections for 2020\u20132045. \nb\u2013i, Box plots of some of the metrics shown in a for selected designs across the \n29 climate change projections. All change values are calculated from a baseline in \nwhich the GERD is operated on the basis of the Washington draft proposal.  \nThe box plots in b\u2013i correspond to the lines with similar colours in a. The ends of \nthe boxes in b\u2013i represent the upper and lower quartiles, the solid vertical lines \ninside the boxes mark the medians, the dashed vertical lines mark the means, \nthe circles show the data points and the whiskers extend to the maximum and \nminimum values, excluding the outliers. The firm power values are calculated on \nthe basis of a 90% reliability, and the GDP values are discounted at a 3% rate.\n\nNature Climate Change | Volume 13 | January 2023 | 48\u201357\n53\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nprecipitation increases often seen in GCMs over East Africa42,43 (Meth-\nods and Supplementary Table 1). The second stage of the framework \nis to calibrate and use a distributed hydrological model for the Nile \ndriven by historical climate time series and the 29 climate projections \nto generate naturalized historical and projected streamflow time series. \nFurthermore, time series for projected potential evapotranspiration \n(PET) from irrigation schemes and open water bodies were generated.\nBecause climate change scenarios have implications for river \nsystems and their economies, the third stage of the framework aims \nto capture these implications through integrated economy-wide and \nriver system simulators7. The river system simulator accounts for all \nmajor infrastructure in the Nile Basin in monthly time steps and uses \nthe naturalized streamflow and PET time series generated in the second \nstage. Furthermore, dynamic-recursive annual computable general \nequilibrium (CGE) models were developed for Ethiopia, Sudan and \nEgypt to simulate their economies.\nClimate projections may have different global socio-economic \ndevelopment pathways associated with them41. Accordingly, the CGE \nmodels were exogenously driven by national-level projections for \nlabour growth and population growth, total and sectoral productivi-\nties, and climate policies that differ with each SSP characterizing the \nclimate projections44\u201346. Population growth projections under the SSPs \nwere also used to change (that is, increase or decrease) municipal water \ndemands in the river system simulator.\nThe fourth stage of the framework seeks to design adaptive man-\nagement policies for the Nile water infrastructure. Climate adaptation \nplanning is achieved using an artificial-intelligence-based approach \n(multi-objective evolutionary algorithm) to design efficient adaptive \nmanagement plans. These plans involve adaptation mechanisms based \non new information gained about climate change impacts on the river \nsystem and riparian countries as the future unravels. The design frame-\nwork identifies efficient adaptive plans that maximize the riparian \ncountries\u2019 economy-wide and river system interests in different ways \nand reveal trade-offs.\nImplications of climate change for the Nile\nUnder the 29 examined projections, the results show varying impacts \nof climate change on the naturalized streamflow of the Nile and some \nsocio-economic characteristics of Ethiopia, Sudan and Egypt (Fig. 3). \nThe lower the GHG emissions and EOC forcing levels, the lower the \nchange in precipitation and streamflow (Fig. 3a). The 30-year moving \naverage naturalized streamflow data shown in Fig. 3c\u2013f indicate that the \nmean Nile streamflow could change by between \u221213% and +90% by 2050 \ncompared with 2020. The intra-annual variability of the naturalized \nstreamflow of the Nile and its main tributaries is projected to change \n(Extended Data Fig. 1), with the biggest changes occurring under SSP5 \nand high EOC temperature projections. The inter-annual streamflow \nvariability shows varying changes depending on different SSPs, forcing \nlevels, GCMs and time horizons, as depicted in Extended Data Fig. 2. \nThe increase in temperature imposed by climate change is projected to \nincrease PET (Extended Data Figs. 3 and 4), which would increase future \nirrigation water demands and evaporation from open water bodies.\nVarious SSPs that underpin climate projections have different \nimplications for baseline gross domestic product (GDP) and popu-\nlation growth and urbanization in Ethiopia, Sudan and Egypt over \n2021\u20132050 (Fig. 3b)44\u201346. The highest economic growth for each of the \nthree countries is projected under SSP5, whereas the lowest economic \ngrowth occurs under SSP3.\nCooperative adaptive management policy for the \nGERD\nWe use the Nile adaptive management framework to formulate and design \nan adaptive management policy for the GERD\u2019s initial filling and long-term \noperation for 2020\u20132045, involving short-term rules and interim and \nlong-term adaptation measures to cope with climate change uncertainties. \nThe formulation is based on cooperative behaviour whereby the riparian \ncountries consider each other\u2019s interests through adaptive measures. The \nformulation of the GERD adaptive filling and operation policy is described \nbriefly below and detailed in Extended Data Fig. 5.\nIn the adaptive formulation, water retention during the GERD\u2019s \ninitial filling phase is carried out in July and August and follows a \nstage-based approach (Supplementary Table 2) while maintaining a \nminimum outflow of 1.28 bcm per month, similar to the Washington \ndraft proposal19. From March to June, additional interim water releases \nduring the initial filling phase are made if the storage of the High Aswan \nDam (HAD) reservoir in Egypt falls below 60 bcm. This would enable \nquick initial filling of the GERD, as data show consistently high water \nstorage in the HAD reservoir in recent years (2020\u20132022)47.\nIn the long-term operation phase of the adaptive formulation, \nGERD water releases aim to generate a regular power target when res-\nervoir storage is above a level termed the power reduction threshold, \nbut this target is reduced as an interim precaution if GERD storage runs \nlower than a power reduction threshold to allow storage recovery during \nand following multi-year droughts. To help Egypt during droughts, an \ninterim minimum monthly drought mitigation water release from the \nGERD is activated if the storage of the HAD reservoir stays below 60 bcm \nover the past six months for an extended period; this period is termed \nthe drought trigger. This interim drought mitigation measure is acti-\nvated only if the GERD storage is above a threshold termed the drought \noutflow storage threshold. The operations of the seasonal storage dams \ndownstream of the GERD follow their historical rules during the GERD\u2019s \nfilling phase, whereas their reservoirs are kept as high as possible during \nthe long-term operation phase, similar to assumptions made by previ-\nous studies, assuming data sharing between Ethiopia and Sudan7,14\u201316.\nLong-term adaptation measures are applied to the GERD\u2019s regular \nand reduced power targets, the drought trigger of the HAD, and the \ninterim minimum monthly drought mitigation water release. The \nlong-term adaptation measures are triggered on a five-year interval, \ninvolving increasing or decreasing these four GERD management vari-\nables on the basis of the change in the mean annual inflow to the dam \nover the past five years in relation to the historical mean annual inflow \nover 1980\u20132019. For instance, if the mean annual inflow to the GERD \nincreased over the past five years compared with the historical mean, \nthe four above-mentioned GERD operation variables are increased \nproportionally, and vice versa.\nThe artificial-intelligence-based search component of the frame-\nwork was used to design the adaptive GERD management policy \ndescribed above. The search algorithm optimizes seven variables; \nthese variables and their upper and lower bounds are reported in \nSupplementary Table 3. The variables are designed to maximize nine \nobjectives of Ethiopia, Sudan and Egypt over 2020\u20132045 across the \n29 climate change projections: the accumulated GDP values of each \ncountry (three objectives); the GERD\u2019s 90% firm power (one objective); \nthe annual hydro-energy generation of the GERD, Sudan and Egypt \n(three objectives); and the annual irrigation water use in Sudan and \nEgypt (two objectives).\nEconomic and river system benefits of adaptive \nGERD management\nThe results reveal differences between the aggregated economy-wide \nand river system performance objectives over 2020\u20132045 under 1,032 \nefficient GERD adaptive management options and the performance \nunder the Washington draft proposal, as shown in the parallel coordi-\nnates plot48 depicted in Fig. 4a. The reader is referred to Supplemen-\ntary Section 2, Supplementary Table 2 and Extended Data Fig. 6 for \nthe implementation details of the Washington draft proposal. Four \nefficient adaptive GERD policy designs are highlighted in Fig. 4a (the \ngreen, blue, purple and cyan lines): a favourable design for each of \nEthiopia, Sudan and Egypt that results in the highest national accumu-\nlated GDP benefits, and an example compromise design that results in \n\nNature Climate Change | Volume 13 | January 2023 | 48\u201357\n54\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nFavourable Sudanese\ndesign\nFavourable Egyptian\ndesign\nCompromise\ndesign\nFavourable Ethiopian\ndesign\nSSP585 (ACCESS-CM2*)\nSSP585 (UKESM1-0-LL*)\nSSP585 (ACCESS-ESM1-5*)\nSSP585 (CESM2-WACCM*)\nSSP245 (ACCESS-ESM1-5*)\nSSP126 (UKESM1-0-LL*)\nSSP370 (CESM2*)\nSSP245 (CESM2*)\nSSP245 (MIROC6*)\nSSP126 (GFDL-ESM4)\nSSP245 (GFDL-ESM4)\nSSP126 (BCC-CSM2-MR)\nSSP245 (GISS-E2-1-G)\nSSP370 (GISS-E2-1-G)\nSSP245 (MPI-ESM1-2-HR)\nSSP245 (ACCESS-ESM1-5)\nSSP585 (MRI-ESM2-0)\nSSP370 (MRI-ESM2-0)\nSSP370 (GISS-E2-1-G)\nSSP126 (UKESM1-0-LL)\nSSP245 (MIROC6)\nSSP245 (CESM2)\nSSP370 (IPSL-CM6A-LR)\nSSP585 (GISS-E2-1-G)\nSSP370 (CESM2)\nSSP585 (ACCESS-ESM1-5)\nSSP585 (CESM2-WACCM)\nSSP585 (ACCESS-CM2)\nSSP585 (UKESM1-0-LL)\nSSP585 (ACCESS-CM2*)\nSSP585 (UKESM1-0-LL*)\nSSP585 (ACCESS-ESM1-5*)\nSSP585 (CESM2-WACCM*)\nSSP245 (ACCESS-ESM1-5*)\nSSP126 (UKESM1-0-LL*)\nSSP370 (CESM2*)\nSSP245 (CESM2*)\nSSP245 (MIROC6*)\nSSP126 (GFDL-ESM4)\nSSP245 (GFDL-ESM4)\nSSP126 (BCC-CSM2-MR)\nSSP245 (GISS-E2-1-G)\nSSP370 (GISS-E2-1-G)\nSSP245 (MPI-ESM1-2-HR)\nSSP245 (ACCESS-ESM1-5)\nSSP585 (MRI-ESM2-0)\nSSP370 (MRI-ESM2-0)\nSSP370 (GISS-E2-1-G)\nSSP126 (UKESM1-0-LL)\nSSP245 (MIROC6)\nSSP245 (CESM2)\nSSP370 (IPSL-CM6A-LR)\nSSP585 (GISS-E2-1-G)\nSSP370 (CESM2)\nSSP585 (ACCESS-ESM1-5)\nSSP585 (CESM2-WACCM)\nSSP585 (ACCESS-CM2)\nSSP585 (UKESM1-0-LL)\nClimate projections (higher Nile streamflow \u2192)\nClimate projections (higher Nile streamflow \u2192)\nClimate projections (higher Nile streamflow \u2192)\nSSP585 (ACCESS-CM2*)\nSSP585 (UKESM1-0-LL*)\nSSP585 (ACCESS-ESM1-5*)\nSSP585 (CESM2-WACCM*)\nSSP245 (ACCESS-ESM1-5*)\nSSP126 (UKESM1-0-LL*)\nSSP370 (CESM2*)\nSSP245 (CESM2*)\nSSP245 (MIROC6*)\nSSP126 (GFDL-ESM4)\nSSP245 (GFDL-ESM4)\nSSP126 (BCC-CSM2-MR)\nSSP245 (GISS-E2-1-G)\nSSP370 (GISS-E2-1-G)\nSSP245 (MPI-ESM1-2-HR)\nSSP245 (ACCESS-ESM1-5)\nSSP585 (MRI-ESM2-0)\nSSP370 (MRI-ESM2-0)\nSSP370 (GISS-E2-1-G)\nSSP126 (UKESM1-0-LL)\nSSP245 (MIROC6)\nSSP245 (CESM2)\nSSP370 (IPSL-CM6A-LR)\nSSP585 (GISS-E2-1-G)\nSSP370 (CESM2)\nSSP585 (ACCESS-ESM1-5)\nSSP585 (CESM2-WACCM)\nSSP585 (ACCESS-CM2)\nSSP585 (UKESM1-0-LL)\nYear\nYear\nYear\n\u20131.5\n\u20131.0\n\u20130.5\n0\n0.5\n1.0\n1.5\nChange in Ethiopia\u2019s real GDP (US$ billion)\n\u20130.6\n\u20130.4\n\u20130.2\n0\n0.2\n0.4\n0.6\nChange in Sudan\u2019s real GDP (US$ billion)\n\u20132\n\u20131\n0\n1\n2\nChange in Egypt\u2019s real GDP (US$ billion)\n2020\n2025\n2030\n2035\n2040\n2045\n2020\n2025\n2030\n2035\n2040\n2045\n2020\n2025\n2030\n2035\n2040\n2045\n2020\n2025\n2030\n2035\n2040\n2045\nYear\nYear\nYear\nYear\n2020\n2025\n2030\n2035\n2040\n2045\n2020\n2025\n2030\n2035\n2040\n2045\n2020\n2025\n2030\n2035\n2040\n2045\n2020\n2025\n2030\n2035\n2040\n2045\nYear\nYear\nYear\nYear\n2020\n2025\n2030\n2035\n2040\n2045\n2020\n2025\n2030\n2035\n2040\n2045\n2020\n2025\n2030\n2035\n2040\n2045\n2020\n2025\n2030\n2035\n2040\n2045\nYear\nFig. 5 | Heatmap matrix of changes in the Ethiopian, Sudanese and Egyptian \nreal GDP for four adaptive policy designs. The GDP changes are for the four \ndesigns of the GERD\u2019s adaptive policy highlighted in Fig. 4a compared with a \nbaseline in which the GERD is operated on the basis of the Washington draft \nproposal. Each row of the matrix shows GDP changes for one of the three \ncountries, whereas the matrix columns correspond to different designs of the \nadaptive management policy for the GERD. The labels in parentheses are the \nnames of the climate models. The changes in real GDP are discounted at a 3% rate. \nThe climate projections marked with asterisks were synthesized to address the \nEastern African Paradox.\n\nNature Climate Change | Volume 13 | January 2023 | 48\u201357\n55\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nat least 40% of the highest national accumulated GDP benefits for each \nof the three countries. The specifications of the four highlighted GERD \npolicy designs are in Supplementary Table 4 and Supplementary Fig. 2.  \nThe results show that the magnitudes of the economy-wide costs \nand benefits to Ethiopia, Sudan and Egypt vary due to the different \neconomies and the spatial and temporal characteristics of the river \nsystem in each country. There is a trade-off between the three countries \nin achieving the highest possible GDP performance. Compared with the  \nWashington draft proposal, the compromise design results in \neconomy-wide benefits to the three countries.\nEgypt\u2019s\nreal GDP\na\nb\nc\nd\ne\nf\ng\nh\ni\nSudan\u2019s\nreal GDP\nEthiopia\u2019s\nreal GDP\nInterim additional filling\nwater release\nPower reduction storage\nDrought trigger\nRegular power target\nInterim reduced power\ntarget\nDrought outflow storage\nthreshold\nInterim minimum drought\nmitigation water release\nInterim additional filling\nwater release\nPower reduction storage\nDrought trigger\nRegular power target\nInterim reduced power\ntarget\nDrought outflow storage\nthreshold\nInterim minimum drought\nmitigation water release\nInterim additional filling\nwater release\nPower reduction storage\nDrought trigger\nRegular power target\nInterim reduced power\ntarget\nDrought outflow storage\nthreshold\nInterim minimum drought\nmitigation water release\nInterim additional filling\nwater release\nPower reduction storage\nDrought trigger\nRegular power target\nInterim reduced power\ntarget\nDrought outflow storage\nthreshold\nInterim minimum drought\nmitigation water release\nInterim additional filling\nwater release\nPower reduction storage\nDrought trigger\nRegular power target\nInterim reduced power\ntarget\nDrought outflow storage\nthreshold\nInterim minimum drought\nmitigation water release\nInterim additional filling\nwater release\nPower reduction storage\nDrought trigger\nRegular power target\nInterim reduced power\ntarget\nDrought outflow storage\nthreshold\nInterim minimum drought\nmitigation water release\nInterim additional filling\nwater release\nPower reduction storage\nDrought trigger\nRegular power target\nInterim reduced power\ntarget\nDrought outflow storage\nthreshold\nInterim minimum drought\nmitigation water release\nInterim additional filling\nwater release\nPower reduction storage\nDrought trigger\nRegular power target\nInterim reduced power\ntarget\nDrought outflow storage\nthreshold\nInterim minimum drought\nmitigation water release\nInterim additional filling\nwater release\nPower reduction storage\nDrought trigger\nRegular power target\nInterim reduced power\ntarget\nDrought outflow storage\nthreshold\nInterim minimum drought\nmitigation water release\nEgypt\u2019s\nirrigation \nwater use\nEgypt\u2019s\nhydro-energy\ngeneration\nSudan\u2019s\nirrigation\nwater use\n0\n0.25\n0.50\n0.75\n1.0\nRelative influence\nGERD\u2019s\nfirm power\nSudan\u2019s\nhydro-energy\ngeneration\nGERD\u2019s\nhydro-energy\ngeneration\n0\n0.25\n0.50\n0.75\n1.0\nRelative influence\n0\n0.25\n0.50\n0.75\n1.0\nRelative influence\n0\n0.25\n0.50\n0.75\n1.0\nRelative influence\n0\n0.25\n0.50\n0.75\n1.0\nRelative influence\n0\n0.25\n0.50\n0.75\n1.0\nRelative influence\n0\n0.25\n0.50\n0.75\n1.0\nRelative influence\n0\n0.25\n0.50\n0.75\n1.0\nRelative influence\n0\n0.25\n0.50\n0.75\n1.0\nRelative influence\nFig. 6 | Rankings of the variables of the adaptive management policy for the \nGERD based on their relative influence on nine economic and river system \nperformance metrics. a\u2013i, The relative influence values are based on machine \nlearning and can range from zero to one, with zero indicating that the variable \ndoes not influence the performance metric, whereas one indicates that the \nvariable is the sole influencer of the performance metric. For each performance \nmetric, the sum of the relative influence values for all parameters is one.\n\nNature Climate Change | Volume 13 | January 2023 | 48\u201357\n56\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nThe results reveal varying costs and benefits across climate projec-\ntions (Fig. 4b\u2013i). For instance, in the most favourable design for each \ncountry taken individually, the discounted GDP over 2020\u20132045 could \nincrease by up to US$15.8, US$6.3 and US$3.0 billion for Ethiopia, Sudan \nand Egypt, respectively, compared with the Washington draft proposal. \nHowever, these country-centric designs lead to GDP losses for at least \none of the other two countries. In contrast, the mean GDP changes are \npositive for the three countries under the compromise design.\nThe temporal evolution of GDP changes under different climate \nprojections and GERD policy designs (Fig. 5) shows that adaptive GERD \nmanagement benefits Ethiopia and Sudan the most under climate pro-\njections with the highest streamflow. The favourable Egyptian design \nbenefits Egypt during multi-year droughts in climate projections with \nlow streamflow, but it reduces the overall Sudanese and Ethiopian \nbenefits (Fig. 5). The compromise system design results in balanced \nperformance across the three countries.\nThe seven optimized variables of the GERD adaptive policy show \nvarying influence on the Ethiopian, Sudanese and Egyptian perfor-\nmance objectives (Fig. 6). The Egyptian GDP is influenced the most by \nthe interim minimum drought mitigation water releases (Fig. 6a), as \nmost of Egypt\u2019s benefits and costs materialize during droughts through \nirrigation. In contrast, the most influential parameter for the Sudanese \nand Ethiopian GDPs is the GERD\u2019s reduced power target (Fig. 6b,c). \nFor Sudan, decreases in GERD water releases due to power reduction \ninfluence irrigation (Fig. 6f) and hydropower (Fig. 6g) because of the \nabsence of multi-year storage dams to buffer this variability.\nDiscussion\nBecause the Nile River is crucial to its population\u2019s economic develop-\nment and well-being, adaptation strategies are needed to cope with \nthe deep uncertainties associated with future climate change26,49. An \napproach that can identify efficient options for transboundary adapta-\ntion and demonstrate their economy-wide and river system benefits and \ntrade-offs could provide a platform for discussions on Nile adaptation \nstrategies. The adaptive planning framework introduced in this paper \ncan design adaptive policies for large infrastructures to cope with climate \nchange uncertainties. Using a meta-heuristic artificial-intelligence-based \nalgorithm for the search process provides twofold benefits. First, it ena-\nbles finding multi-dimensionally efficient (Pareto-optimal) solutions for \ncomplex and highly nonlinear interlinked river and economic systems. \nSecond, it optimizes on the basis of linked but independent simulation \nmodels developed by different disciplines. Although the proposed frame-\nwork can capture direct and induced impacts of climate change and \ninfrastructure management policies on river and economy systems, it \nshould be complemented with approaches to assess other impacts on \ngroundwater, river ecology and riparian populations7,50.\nThe analysis of the GERD\u2019s initial filling and long-term operation \nshows that adaptively managing the dam to maximize the national \nbenefits of any of the three countries would be costly for at least one \nof the other two countries. We show that a compromise adaptive man-\nagement approach could produce balanced benefits for the three \ncountries. These results demonstrate the opportunity cost of not \nimplementing collaborative adaptive solutions, especially under \nextreme climate change projections. It is high time to integrate climate \nchange adaptation into the decade-long negotiations between Ethio-\npia, Sudan and Egypt over the GERD and the broader Nile management \ndiscussion between the 11 riparian countries.\nAlthough the adaptive formulation examined in this study \nincludes some controlling features (for example, the 60-bcm HAD \ndrought threshold and the five-year adaptation period) that con-\nstrain basin-wide water management, it represents an incremental \nstep towards adaptive cooperation. Our goal is to show that any step \ntowards adaptive cooperation results in benefits. Such an incremental \napproach can help build trust between riparian countries and make an \neventual unconstrained basin-wide adaptive management easier to \nachieve, aiming to maximize basin-wide benefits, adopt benefit-sharing \nand build on the comparative advantages of each riparian country.\nPractical use of the proposed framework in Nile negotiations \nrequires riparian countries to negotiate an adaptive formulation and \ndefine the ranges within which decisions can be optimized. The philoso-\nphy behind the adaptive formulation should be guided by a long-term \nvision to counter future climate and socio-economic uncertainties. \nThe efficient cooperative adaptive designs emanating from the agreed \nformulation should then be negotiated to find a compromise solution \nthat balances national-level and basin-wide performance.\nOnline content\nAny methods, additional references, Nature Portfolio reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-022-01556-6.\nReferences\n1.\t\nIPCC Climate Change 2021: The Physical Science Basis (eds \nMasson-Delmotte, V. et al.) (Cambridge Univ. Press, 2021).\n2.\t\nWigley, T. M. L. & Raper, S. C. B. Natural variability of the climate \nsystem and detection of the greenhouse effect. Nature 344, \n324\u2013327 (1990).\n3.\t\nCrowley, J. T. Causes of climate change over the past 1000 years. \nScience 289, 270\u2013277 (2000).\n4.\t\nWang, W. C., Yung, Y. L., Lacis, A. A., Mo, T. A. & Hansen, J. E. \nGreenhouse effects due to man-made perturbations of trace \ngases. Science 194, 685\u2013690 (1976).\n5.\t\nParis Agreement (United Nations Framework Convention on \nClimate Change, 2015).\n6.\t\nRio+20 United Nations Conference on Sustainable Development \nThe Future We Want: Outcome Document of the United Nations \nConference on Sustainable Development (United Nations, 2012).\n7.\t\nBasheer, M. et al. Collaborative management of the Grand \nEthiopian Renaissance Dam increases economic benefits and \nresilience. Nat. Commun. 12, 5622 (2021).\n8.\t\nAgreement between the Republic of the Sudan and the United Arab \nRepublic for the Full Utilization of the Nile Waters (International \nWater Law Project, 1959); http://internationalwaterlaw.org/\ndocuments/regionaldocs/uar_sudan.html\n9.\t\nCasc\u00e3o, A. E. & Nicol, A. GERD: new norms of cooperation in the \nNile Basin? Water Int. 41, 550\u2013573 (2016).\n10.\t Salman, S. The Grand Ethiopian Renaissance Dam: the road to the \ndeclaration of principles and the Khartoum document. Water Int. \n41, 512\u2013527 (2016).\n11.\t\nTawfik, R. The Grand Ethiopian Renaissance Dam: a \nbenefit-sharing project in the Eastern Nile? Water Int. 41, 574\u2013592 \n(2016).\n12.\t Wheeler, K. G., Jeuland, M., Hall, J. W., Zagona, E. & Whittington, \nD. Understanding and managing new risks on the Nile with the \nGrand Ethiopian Renaissance Dam. Nat. Commun. https://doi.\norg/10.1038/s41467-020-19089-x (2020).\n13.\t Wheeler, K. et al. Exploring cooperative transboundary river \nmanagement strategies for the Eastern Nile Basin. Water Resour. \nRes. https://doi.org/10.1029/2017WR022149 (2018).\n14.\t Wheeler, K. G. et al. Cooperative filling approaches for the Grand \nEthiopian Renaissance Dam. Water Int. 41, 611\u2013634 (2016).\n15.\t Basheer, M. et al. Quantifying and evaluating the impacts of \ncooperation in transboundary river basins on the water\u2013energy\u2013\nfood nexus: the Blue Nile Basin. Sci. Total Environ. 630, 1309\u20131323 \n(2018).\n16.\t Basheer, M. Cooperative operation of the Grand Ethiopian \nRenaissance Dam reduces Nile riverine floods. River Res. Appl. 47, \n805\u2013814 (2021).\n\nNature Climate Change | Volume 13 | January 2023 | 48\u201357\n57\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\n17.\t Elagib, N. A. & Basheer, M. Would Africa\u2019s largest hydropower dam \nhave profound environmental impacts? Environ. Sci. Pollut. Res. \n28, 8936\u20138944 (2021).\n18.\t Joint Statement of Egypt, Ethiopia, Sudan, the United States and \nthe World Bank (United States Department of the Treasury, 2020); \nhttps://home.treasury.gov/news/press-releases/sm891\n19.\t Edrees, M. Letter dated 11 June 2021 from the Permanent \nRepresentative of Egypt to the United Nations addressed to the \nSecretary-Genera (United Nations, 2021); https://digitallibrary.\nun.org/record/3931750?ln=en\n20.\t Amde, T. A. Letter dated 14 May 2020 from the Permanent \nRepresentative of Ethiopia to the United Nations addressed to the \nPresident of the Security Council (United Nations, 2020); https://\ndigitallibrary.un.org/record/3862715?ln=en\n21.\t Taye, M. T., Willems, P. & Block, P. Implications of climate change \non hydrological extremes in the Blue Nile Basin: a review.  \nJ. Hydrol. Reg. Stud. 4, 280\u2013293 (2015).\n22.\t Di Baldassarre, G. et al. Future hydrology and climate in the River \nNile Basin: a review. Hydrol. Sci. J. 56, 199\u2013211 (2011).\n23.\t Bhattacharjee, P. S. & Zaitchik, B. F. Perspectives on CMIP5 model \nperformance in the Nile River headwaters regions. Int. J. Climatol. \n35, 4262\u20134275 (2015).\n24.\t Haasnoot, M., Kwakkel, J. H., Walker, W. E. & ter Maat, J. Dynamic \nadaptive policy pathways: a method for crafting robust decisions for \na deeply uncertain world. Glob. Environ. Change 23, 485\u2013498 (2013).\n25.\t Hui, R., Herman, J., Lund, J. & Madani, K. Adaptive water \ninfrastructure planning for nonstationary hydrology. Adv. Water \nResour. 118, 83\u201394 (2018).\n26.\t Marchau, V. A. W. J., Walker, W. E., Bloemen, P. J. T. M. & Popper, S. \nW. (eds) Decision Making under Deep Uncertainty: From Theory to \nPractice (Springer, 2019).\n27.\t Smith, M. et al. Adaptation\u2019s Thirst: Accelerating the Convergence \nof Water and Climate Action (Global Commission on Adaptation, \n2019).\n28.\t Hallegatte, S. Strategies to adapt to an uncertain climate change. \nGlob. Environ. Change 19, 240\u2013247 (2009).\n29.\t Reed, P. M. et al. Multisector dynamics: advancing the science \nof complex adaptive human\u2013Earth systems. Earth\u2019s Future 10, \ne2021EF002621 (2022).\n30.\t Walker, W. E., Haasnoot, M. & Kwakkel, J. H. Adapt or perish: \na review of planning approaches for adaptation under deep \nuncertainty. Sustainability 5, 955\u2013979 (2013).\n31.\t Kwadijk, J. C. J. et al. Using adaptation tipping points to prepare \nfor climate change and sea level rise: a case study in the \nNetherlands. WIREs Clim. Change 1, 729\u2013740 (2010).\n32.\t Kwakkel, J. H., Walker, W. E. & Marchau, V. Adaptive airport strategic \nplanning. Eur. J. Transp. Infrastruct. Res. 10, 249\u2013273 (2010).\n33.\t Kwakkel, J. H., Haasnoot, M. & Walker, W. E. Developing dynamic \nadaptive policy pathways: a computer-assisted approach for \ndeveloping adaptive strategies for a deeply uncertain world. \nClimatic Change 132, 373\u2013386 (2015).\n34.\t Zeff, H. B., Herman, J. D., Reed, P. M. & Characklis, G. W. \nCooperative drought adaptation: integrating infrastructure \ndevelopment, conservation, and water transfers into \nadaptive policy pathways. Water Resour. Res. https://doi.\norg/10.1002/2016WR018771 (2016).\n35.\t Fletcher, S., Lickley, M. & Strzepek, K. Learning about climate \nchange uncertainty enables flexible water infrastructure \nplanning. Nat. Commun. 10, 1782 (2019).\n36.\t Cohen, J. S. & Herman, J. D. Dynamic adaptation of water \nresources systems under uncertainty by learning policy structure \nand indicators. Water Resour. Res. 57, e2021WR030433 (2021).\n37.\t Ricalde, I. et al. Assessing tradeoffs in the design of climate \nchange adaptation strategies for water utilities in Chile. J. Environ. \nManage. 302, 114035 (2022).\n38.\t Pachos, K., Huskova, I., Matrosov, E., Erfani, T. & Harou, J. J. \nTrade-off informed adaptive and robust real options water \nresources planning. Adv. Water Resour. 161, 104117 (2022).\n39.\t Gold, D. F., Reed, P. M., Gorelick, D. E. & Characklis, G. W. Power \nand pathways: exploring robustness, cooperative stability, and \npower relationships in regional infrastructure investment and \nwater supply management portfolio pathways. Earth\u2019s Future 10, \ne2021EF002472 (2022).\n40.\t Beh, E. H. Y., Maier, H. & Dandy, G. C. Adaptive, multiobjective \noptimal sequencing approach for urban water supply \naugmentation under deep uncertainty. Water Resour. Res. https://\ndoi.org/10.1002/2014WR016254 (2015).\n41.\t O\u2019Neill, B. C. et al. The Scenario Model Intercomparison Project \n(ScenarioMIP) for CMIP6. Geosci. Model Dev. 9, 3461\u20133482  \n(2016).\n42.\t Wainwright, C. M. et al. \u2018Eastern African Paradox\u2019 rainfall decline \ndue to shorter not less intense Long Rains. NPJ Clim. Atmos. Sci. 2, \n34 (2019).\n43.\t Rowell, D. P., Booth, B. B. B., Nicholson, S. E. & Good, P. \nReconciling past and future rainfall trends over East Africa.  \nJ. Clim. 28, 9768\u20139788 (2015).\n44.\t Riahi, K. et al. The Shared Socioeconomic Pathways and their \nenergy, land use, and greenhouse gas emissions implications: an \noverview. Glob. Environ. Change 42, 153\u2013168 (2017).\n45.\t KC, S. & Lutz, W. The human core of the Shared Socioeconomic \nPathways: population scenarios by age, sex and level of \neducation for all countries to 2100. Glob. Environ. Change 42, \n181\u2013192 (2017).\n46.\t Crespo Cuaresma, J. Income projections for climate change \nresearch: a framework based on human capital dynamics. Glob. \nEnviron. Change 42, 226\u2013236 (2017).\n47.\t Water Level (Copernicus Global Land Service, 2022); https://land.\ncopernicus.eu/global/products/wl\n48.\t Inselberg, A. in Trends in Interactive Visualization: State-of-the-Art \nSurvey (eds Liere, R. et al.) 49\u201378 (Springer, 2009).\n49.\t Goulden, M., Conway, D. & Persechino, A. Adaptation to climate \nchange in international river basins in Africa: a review. Hydrol. Sci. \nJ. 54, 805\u2013828 (2009).\n50.\t Dasgupta, P. The Economics of Biodiversity: The Dasgupta Review \n(HM Treasury, 2021).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons license, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons license, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons license and your intended \nuse is not permitted by statutory regulation or exceeds the permitted \nuse, you will need to obtain permission directly from the copyright \nholder. To view a copy of this license, visit http://creativecommons.\norg/licenses/by/4.0/.\n\u00a9 The Author(s) 2023\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nMethods\nHere we introduce an adaptive planning framework for the Nile Basin \nthat combines climate change projections; integrated hydrological, \neconomy-wide and river system simulators; and multi-objective evo-\nlutionary and machine learning algorithms. Below is a description of \nthe components and data of the framework.\nClimate projections\nTwenty climate projections were selected from CMIP6 Tier 1 GCMs \non the basis of uniform sampling to cover the full range of available \nSSPs, radiative forcing and the joint distribution of EOC change in \nprecipitation and temperature over the Nile Basin. Processing, down-\nscaling and bias correction of the GCM simulations were driven by \nthe requirement for transient (2017\u20132100) three-hourly forcing data \nacross seven climate variables that govern the surface mass and energy \nbalances in the hydrological model (described in the next section). The \nseven variables are precipitation, temperature, incoming shortwave \nradiation, incoming longwave radiation, humidity, wind speed and \nsurface pressure. Given the large spatial domain, the low availability \nof subdaily model output for all of the relevant variables in CMIP6 and \nchallenges of multivariate bias correction51, we derived bias-corrected \ntransient projections by (1) resampling the 0.25\u00b0 historical baseline \nclimate dataset to match the (detrended) relative variability of the \nselected GCM projections and then (2) applying perturbation factors \nto reintroduce the change signal extracted from the GCMs following \na quantile delta mapping approach52.\nIn more detail, we first detrended the basin-average monthly \nfuture precipitation series for 2017\u20132100 in each GCM run on a \nmonth-wise basis using smoothed 37-year moving averages. Second, \nwe ranked the detrended future series month-wise and matched the \nranks to their equivalents in the historical baseline precipitation data-\nset of 1981\u20132016, allowing us to construct synthetic climate series for \n2017\u20132100 at the three-hour temporal resolution of the historical \nclimate datasets by inserting the relevant month of data from the his-\ntorical dataset at the right place in the future series. The synthesized \nseries thus have the same overall sequencing of relative intra-annual \n(seasonal), interannual and multi-annual variability as the original \nGCM projections in rank terms, while preserving the statistical prop-\nerties (including spatial, temporal and intervariable consistency53,54) \nand resolution of the historical baseline dataset (that is, without GCM \nbias). The historical baseline climate datasets used are the Multi-Source \nWeighted-Ensemble Precipitation (MSWEP)55 for precipitation and \nthe Princeton Global Forcing (PGF)56 for the other climate variables. \nThe 0.1\u00b0 MSWEP data were regridded to the 0.25\u00b0 PGF grid before the \nclimate series was synthesized.\nFor each of the 20 climate projections, we then reintroduced the \nclimate change signal to the synthesized 2017\u20132100 climate series, \nproducing bias-corrected and perturbed projections. The transient \nclimate change signal was reintroduced on a month-wise basis using a \n37-year moving window of quantile-based perturbations52. Smoothing \nwas applied to the perturbation factors via a moving average to avoid \nany jumps between successive windows. We note that this method does \nnot adjust the frequency of wet periods/days within a month, and it is \nnot designed to focus on projected changes in precipitation extremes, \ngiven the study\u2019s focus on water resources management on a monthly \nscale rather than flood frequency.\nSeveral previous studies have highlighted how the wetting trend \nin climate projections for eastern Africa stands in contrast to the \nobserved decline in precipitation from the 1980s to the late 2000s42. \nThis contradiction, known as the Eastern African Paradox, represents \na challenge for climate adaptation planning42,43. While the literature \nshows a growing understanding of how this issue is linked with regional \ncirculation dynamics and their representation in models57,58, there is \nnot yet a consensus on how the projections might be constrained. Yet, \nif the projections are taken at face value, there is a risk that adaptation \nmeasures are designed for a highly uncertain wetter climate that might \nnot materialize, with potentially severe socio-economic consequences.\nTo address this issue, we synthesized nine additional scenarios \nbased on 9 of the 20 initial projections by removing the overall wetting \ntendency in the CMIP6 ensemble for the East Africa region. The magni-\ntude of the tendency was estimated by quantifying the mean increase \nin EOC mean precipitation for a unit increase in mean temperature \naccording to the full GCM ensemble, given their strong relationship in \nthis context. The relationship was then used to adjust individual GCM \nprecipitation trends downwards, allowing us to select the nine projec-\ntions representing the most pessimistic scenarios in terms of precipita-\ntion decreases, which enables stress-testing adaptation policies across \na wide range of plausible futures (Fig. 3a). Supplementary Table 1 lists \nthe 29 projections and their main features. Daily PET was calculated \nfor the 29 projections using the FAO56 Penman\u2013Monteith method59.\nHydrological simulator\nTo generate time series of historical and projected naturalized stream-\nflow for the Nile Basin, we used and calibrated the Variable Infiltration \nCapacity (VIC version 5) land surface model60 and the Routing Appli-\ncation for Parallel Computation of Discharge (RAPID)61. The VIC and \nRAPID models were originally developed in a previous study to recon-\nstruct vector-based global naturalized streamflow62. The hydrological \nmodel has a spatial resolution of 0.25\u00b0 and a three-hourly temporal \nresolution. The historical streamflow time series was driven by MSWEP55 \nand temperature, incoming shortwave radiation, incoming longwave \nradiation, humidity, wind speed and surface pressure data from the \nPGF56. The 29 projected streamflow time series were driven by similar \nclimate variables obtained from the climate projections described \nin the previous section. The VIC and RAPID models were calibrated \nagainst a historical naturalized streamflow dataset of the Nile previ-\nously developed for the Eastern Nile Technical Regional Office63. The \nhydrological model was calibrated using baseflow parameters, depth \nof soil layers, variable infiltration curve and maximum soil moisture. \nSupplementary Table 5 shows the ranges of the values of the calibra-\ntion parameters for all grid cells in the hydrological model. The RAPID \nmodel routes streamflow using the Muskingum method64 and requires \na river network to determine Muskingum parameters (that is, the \ngradient coefficient and the weighting factor). The river network was \nobtained from the Hydrological data and maps on the basis of SHuttle \nElevation Derivatives at multiple Scales (HydroSHEDS)65. The hydrolog-\nical model was calibrated over the period 1982\u20131992 and validated over \n1993\u20132002 at four locations where naturalized flow data are available. \nSupplementary Table 6 shows the performance of the hydrological \nmodel in the calibration and validation periods, and Supplementary \nTable 7 summarizes the model performance ranking criteria adopted. \nAs elaborated in the next section, the combined hydrological and river \nsystem simulator was calibrated at additional locations.\nRiver system simulator\nTo simulate the performance of the Nile River system infrastructure \nunder climate change and various adaptive management options, we \ndeveloped a monthly river system model for the basin using Python \nWater Resources (Pywr)66. A river system model is a network representa-\ntion of the supplies, demands and infrastructures of water resources in \na river system67. Pywr is an open-source Python library that simulates \nresource networks. It enables representing water resources system \ninfrastructure (for example, dams, lakes, aquifer-based water sup-\nply and water abstraction locations) in a network structure driven \nby water supplies (for example, hydrological inflows) and demands \n(for example, irrigation, municipal and industrial water demands and \nhydropower) and system operating rules. Pywr uses a linear program-\nming approach to simulate water allocations at each time step (monthly \nin this study) subject to constraints imposed by system operation rules. \nSupplementary Fig. 3 shows a schematic of the Nile River system model \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nand its elements, and Supplementary Table 8 shows the main features of \nthe dams included in the model. The model uses the naturalized stream-\nflow time series of the Nile generated using the hydrological model \ndescribed in the previous section. Data for dams, lakes, reservoirs, and \nirrigation and municipal water demands were collected from different \nsources, including the Nile Basin Initiative and previous studies14,15,68\u201371.\nIn simulating future scenarios, the water demands of irrigation \nschemes and the net evaporation rates of lakes, reservoirs and swamps \nwere modified (increased or decreased) on the basis of the annual \nchange ratio of PET at their locations relative to the historical annual \nmean. Extended Data Figs. 3 and 4 show the projected annual change in \nPET relative to the historical mean at some selected dams and irrigation \nschemes in the Nile River system. Future municipal water demands were \ncalculated by applying the population growth rates projected under dif-\nferent climate change scenarios (that is, SSPs). The population growth \nprojections under different SSPs were obtained from the International \nInstitute for Applied Systems Analysis (IIASA) database44,45. Supple-\nmentary Section 3 describes the initial water demand assumptions for \nEthiopia, Sudan and Egypt.\nThe Nile River system simulator was calibrated and validated at \nten locations over 1995\u20132010 using historical river flow observations \nand reservoir and lake water levels. This calibration period was chosen \non the basis of the availability of common and continuous historical \nobserved data for the ten selected locations. Supplementary Fig. 4 \nand Supplementary Table 9 show the performance of the Nile River \nsystem model over the calibration and validation periods, Supple-\nmentary Table 10 reports the calibration parameters and their values, \nand Supplementary Table 7 reports the performance ranking criteria. \nTotal non-hydro (for example, thermal) generation for each country \nwas represented as a super generator used to fill the gap between \nhydropower generation and national electricity demand subject to \ngeneration capacity. The non-hydro generation capacity and national \nelectricity demands are updated annually on the basis of an iterative \nprocess between the river system and the economy-wide simulators, \nas described in a later section.\nEconomy-wide simulator\nThe standard open-source CGE model of the International Food Policy \nResearch Institute72 was modified and used to develop economy-wide \nmodels for Ethiopia, Sudan and Egypt. The production and consump-\ntion specifications in the CGE models are shown in Supplementary Figs. \n5 and 6. The CGE models were set up to run dynamically over multi-year \nperiods (that is, dynamic-recursive) following endogenous investment \nbehaviour, exogenous total and sectoral factor productivities, labour \ngrowth, and energy use efficiency trends.\nThe CGE models of Ethiopia, Sudan and Egypt include five agent \ntypes: households, the government, enterprises, industries (or eco-\nnomic activities) and the rest of the world. Households are disaggre-\ngated into rural and urban groups. Ethiopia\u2019s, Sudan\u2019s and Egypt\u2019s \nCGE models include 12, 15 and 14 economic activities, respectively. \nEconomic activities use the following factors of production: labour, \ngeneral capital, land, hydropower capital, non-hydro capital, renewa-\nbles capital, water supply capital, oil capital and gas capital. Labour \nand general capital are assumed mobile between sectors, whereas \nthe other factors are sector-specific. All capital types, except water \nsupply and hydropower capitals, grow or shrink over time on the basis \nof investment behaviour according to relative rates of return. Water \nsupply and hydropower capitals were not included in the year-to-year \ninvestment behaviour, as expansion in these infrastructures typically \nrequires abrupt investment in the Nile context.\nWe assume that commodity prices on the international market \nare exogenous, following the small open-economy assumption73. The \ngovernment is assumed to spend a fixed share of total absorption. Sav-\ning propensities are fixed, and the exchange rate is variable. The CGE \nmodels were calibrated to social accounting matrices for 2011 obtained \nfrom the International Food Policy Research Institute74\u201376. We then \nused the GTAP-Power 10 database for the year 201477 to disaggregate \nthe electricity sectors of the social accounting matrices. Five dynamic \nbaselines were calibrated for the CGE models of Ethiopia, Sudan and \nEgypt, with each baseline corresponding to an SSP. The baseline total \nfactor productivity values were set up such that the baseline models fol-\nlow projected economic growth (obtained from the IIASA database) for \neach SSP scenario46. After that, the total factor productivity values were \nfixed and applied exogenously. Labour growth (16\u201364 age group)45, \nurbanization78 and population growth45 projections for each of the \nthree countries were obtained from the IIASA database and applied \nexogenously to the CGE models. Also, projections for national-level \nsectoral productivity under each SSP were obtained from the Centre \nd\u2019\u00c9tudes Prospectives et d\u2019Informations Internationales79 and applied \nexogenously to the CGE models. Sectoral future energy use efficiency \nvalues were calibrated such that the baseline simulated country-level \ncarbon dioxide emissions in the baselines follow the growth pattern \nof the regional emission values of the Middle East and North Africa \nprojected by IIASA for each SSP80. Carbon dioxide emissions were \ncalculated in the CGE models by applying emission factors (tons of \ncarbon per terajoule) to the use of petroleum, gas and coal commodi-\nties in the economies.\nBecause the CGE models were calibrated on the basis of 2011 social \naccounting matrices, they were first run dynamically over 2011\u20132019 to \nbring the economies to the first year of the GERD\u2019s initial filling before \nscenarios for 2020\u20132045 were assessed. The temporal evolution of the \nvalues of some of the key driving parameters and baseline outputs of \nthe CGE models are shown in Supplementary Fig. 7.\nEconomy and river system coupling\nThe river system infrastructure and economy-wide simulators were \nconnected using a generic co-evolutionary framework developed in a \nprevious study7. The framework enables linking river system simula-\ntion models (with daily or monthly time steps) with dynamic-recursive \nannual CGE models. At each annual time step, the coupling framework \nperforms an iterative bidirectional communication between the river \nsystem and the CGE simulators to ensure coherence in the annual \nnational-scale irrigation water supply and demand, municipal water \nsupply and demand, hydropower generation, non-hydro generation \nand capacity, and electricity demand. In each iteration over an annual \ntime step, the river system model quantifies and spatially aggregates \nnational-scale irrigation and municipal water supplies and hydro \nand non-hydro generation on the basis of the river system\u2019s spatial \nand temporal constraints, infrastructure, and external drivers. This \ninformation is then passed to the CGE models as an external shock \non the basis of which changes to the economy\u2019s municipal and irriga-\ntion water demands, electricity demand, and non-hydro capacity are \ndetermined and passed back to the river system model for the next \niteration. Iterations over each annual time step can be terminated by \na maximum number of iterations and/or a convergence error. In this \nwork, we specified a maximum of three iterations between the river \nsystem and the economy-wide simulators for each annual time step and \na convergence error of US$5 million measured using the Ethiopian real \nGDP. The coupling framework is implemented using the open-source \nPython Network Simulation framework81. The reader is referred to \nBasheer et al.7 for further details about the co-evolutionary coupling \nframework used in this study.\nMulti-method artificial-intelligence-based design and \nlearning\nIn the Nile adaptive planning framework, an artificial-intelligence-based \nmulti-objective evolutionary algorithm (MOEA) provides the ability to \nidentify efficient adaptive management policies for river system infra-\nstructure. Supplementary Fig. 8 shows the interaction between the \nintegrated economy-wide and river system simulators and the MOEA.  \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nThe MOEA generates a design (that is, a set of decision variable values) \nfor the adaptive management policy (the yellow box in Supplemen-\ntary Fig. 8), which is then passed to the integrated economy-wide and \nriver system simulators. The integrated model performs a dynamic \nmulti-year multi-scenario simulation considering socio-economic \nand hydrological uncertainties related to climate change, resulting in \naggregated performance metrics (over time and projections) that are \nindicated as objectives. The adaptive management design and objec-\ntive values are then stored before proceeding to the next iteration, in \nwhich the MOEA suggests new adaptive management designs. The \niteration between the MOEA and the integrated simulators contin-\nues until a stopping criterion is met. In this study, we use a maximum \nnumber of iterations as a criterion to terminate iteration between \nthe integrated simulators and the MOEA. A maximum of 6,000 itera-\ntions (or function evaluations) was specified as a stopping criterion \nfor the search process of adaptive GERD management. With every \niteration, the MOEA learns from the previous generations of itera-\ntions and attempts to suggest an adaptive infrastructure management \ndesign that improves performance as measured by the objectives. \nOnce the stopping criterion is met, a non-dominated sorting process \nis performed to filter efficient adaptive management policies. We \nused the open-source Non-dominated Sorting Genetic Algorithm III \n(NSGA-III)82 as an MOEA for adaptive design. Platypus, an open-source \nPython-based framework for evolutionary computing83, was used, \nwhich supports NSGA-III.\nThe multi-objective search was performed for the adaptive man-\nagement formulation for the GERD with five random seeds. Each seed \nrepresents a unique starting condition for the search algorithm. The \nuse of multiple seeds allows checking that the search has converged \nto a global, approximately optimal set of solutions. A total of 30,000 \niterations were therefore performed during the search process (that is, \nfive seeds multiplied by 6,000 iterations). Optimization convergence \nwas tested by calculating the evolution of the hypervolume84 for each \nrandom seed (Supplementary Fig. 9).\nWe used machine learning as a post-processing step for under-\nstanding the relative influence of adaptive policy variables on \neconomy-wide and river system performance, on the basis of a simi-\nlar approach to a previous study on economy systems85. The values \nof the adaptive policy variables and the objectives generated during \nthe search iterations between the MOEA and the integrated simula-\ntors are used to train a machine learning model for each objective. \nAccordingly, the features of each machine learning model are the \nvariables of the adaptive policy, and the target is each of the objec-\ntives. The Random Forest Regression Machine Learning Algorithm86 \nwas used. After the machine learning models were trained, feature \nimportance was calculated, which, for each model, represents the \nrelative influence of features on an objective. The Random Forest \nRegression Algorithm was used through the open-source Scikit-learn \nPython library87.\nFor the GERD application, 80% of the data on the objectives and \ndecision variables were used to train 100 tree predictors for each \nmachine learning model, and 20% of the data were used to test per-\nformance. Maximum tree depth values from 1 to 30 were tested for \neach machine learning model. The lowest tree depth that provided a \ngood prediction ability while avoiding overfitting or underfitting the \ndata was selected for each machine learning model. Supplementary \nFig. 10 shows the performance of the machine learning models with \nthe training and testing data with different maximum tree depths and \nthe chosen maximum tree depth values.\nImplementation of the GERD adaptive management \nformulation\nThe GERD adaptive management formulation optimized in this study \nincludes nine objectives and seven decision variables. The objectives \nmaximize the following aggregate metrics over 2020\u20132045:\n\t1.\t\nThe mean (over projections) accumulated (discounted over \ntime) Ethiopian real GDP\n\t2.\t\nThe mean (over projections) accumulated (discounted over \ntime) Sudanese real GDP\n\t3.\t\nThe mean (over projections) accumulated (discounted over \ntime) Egyptian real GDP\n\t4.\t\nThe mean (over projections) 90% firm (over time) power gen-\neration of the GERD\n\t5.\t\nThe mean (over projections and time) annual energy generation \nof the GERD\n\t6.\t\nThe mean (over projections and time) annual energy from \nhydropower in Sudan\n\t7.\t\nThe mean (over projections and time) annual energy from \nhydropower in Egypt\n\t8.\t\nThe mean (over projections and time) annual irrigation water \nsupply in Sudan\n\t9.\t\nThe mean (over projections and time) annual irrigation water \nsupply in Egypt\nThe seven decision variables are (see Extended Data Fig. 5 for the \ndetails):\n\t1.\t\nInterim additional filling water release\n\t2.\t\nPower reduction storage\n\t3.\t\nRegular power target\n\t4.\t\nInterim reduced power target\n\t5.\t\nDrought trigger\n\t6.\t\nDrought outflow storage threshold\n\t7.\t\nInterim minimum drought mitigation water release\nThe upper and lower bounds of the seven decision variables are \nreported in Supplementary Table 3. Long-term adaptation measures \nare applied to the regular power target, the interim reduced power tar-\nget, the drought trigger and the interim minimum drought mitigation \nwater release. The adaptation measures involve increasing or decreas-\ning these four GERD operation variables dynamically on a five-year \ninterval on the basis of how the mean annual inflow to the GERD over \nthe past five years changed in relation to the historical mean annual \ninflow over 1980\u20132019.\nReporting summary\nFurther information on research design is available in the Nature Port-\nfolio Reporting Summary linked to this article.\nData availability\nThe economy-related input data supporting this study\u2019s findings are \navailable at Zenodo88: https://doi.org/10.5281/zenodo.5914757. The \nNile River system model and its data are not publicly available due to \nstate restrictions and contain information that could compromise \nresearch participant privacy/consent. The Nile model data can be \nmade available upon presentation of the necessary permissions \nfrom the relevant authorities that own the data. The CMIP6 climate \nprojections data can be accessed from https://esgf-node.llnl.gov/\nsearch/cmip6/. The baseline population, labour, urbanization and \neconomic growth data of Ethiopia, Sudan and Egypt associated with \nthe SSPs can be accessed from the IIASA database at https://tntcat.\niiasa.ac.at/SspDb/dsd?Action=htmlpage&page=10. The sectoral \nproductivity projections of Ethiopia, Sudan and Egypt were pro-\nduced on the basis of data from the Centre d\u2019\u00c9tudes Prospectives \net d\u2019Informations Internationales at http://www.cepii.fr/cepii/en/\nbdd_modele/bdd.asp. The MSWEP data are available from http://\nwww.gloh2o.org/mswep/. The PGF data are available from https://\nrda.ucar.edu/datasets/ds314.0/. The river network data used with \nRAPID are freely accessible from https://www.hydrosheds.org/. The \nWorld Bank data referred to in the manuscript are freely accessible \nfrom https://data.worldbank.org/.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nCode availability\nThe Pywr simulation model (version 1.9.1) is open-source and freely \navailable at https://github.com/pywr/pywr, and building Pywr models \ncan be facilitated by www.waterstrategy.org. The standard CGE model \nof the International Food Policy Research Institute is open-source and \nfreely accessible through the following link: https://www.ifpri.org/pub-\nlication/standard-computable-general-equilibrium-cge-model-gams\n-0. The VIC (version 5) model is freely accessible through the following \nlink: https://vic.readthedocs.io/en/master/. RAPID (version 1.8.0) is \nfreely accessible through the following link: http://rapid-hub.org/\nindex.html. The Python Network Simulation framework (version 0.1.5) \nis open-source and freely available in the following repository: https://\ngithub.com/UMWRG/pynsim. The multi-objective NSGA-III, the MOEA \nused in the multi-objective search, is open-source and freely available \nthrough Platypus (version 1.0.4) in the following repository: https://\ngithub.com/Project-Platypus/Platypus. The Random Forest Regres-\nsion Algorithm is open-source and freely available through Scikit-learn \n(version 0.24.2) at https://github.com/scikit-learn/scikit-learn.\nReferences\n51.\t Fran\u00e7ois, B., Vrac, M., Cannon, A. J., Robin, Y. & Allard, D. \nMultivariate bias corrections of climate simulations: which \nbenefits for which losses? Earth Syst. Dyn. 11, 537\u2013562  \n(2020).\n52.\t Cannon, A. J., Sobie, S. R. & Murdock, T. Q. Bias correction of \nGCM precipitation by quantile mapping: how well do methods \npreserve changes in quantiles and extremes? J. Clim. 28, 6938\u2013\n6959 (2015).\n53.\t Mehrotra, R. & Sharma, A. A resampling approach for correcting \nsystematic spatiotemporal biases for multiple variables in a \nchanging climate. Water Resour. Res. 55, 754\u2013770 (2019).\n54.\t Vrac, M. & Friederichs, P. Multivariate-intervariable, spatial, and \ntemporal-bias correction. J. Clim. 28, 218\u2013237 (2015).\n55.\t Beck, H. E. et al. MSWEP v2 global 3-hourly 0.1\u00b0 precipitation: \nmethodology and quantitative assessment. Bull. Am. Meteorol. \nSoc. 100, 473\u2013500 (2019).\n56.\t Sheffield, J., Goteti, G. & Wood, E. F. Development of a 50-year \nhigh-resolution global dataset of meteorological forcings for land \nsurface modeling. J. Clim. 19, 3088\u20133111 (2006).\n57.\t Walker, D. P., Marsham, J. H., Birch, C. E., Scaife, A. A. & Finney, \nD. L. Common mechanism for interannual and decadal \nvariability in the East African Long Rains. Geophys. Res. Lett. 47, \ne2020GL089182 (2020).\n58.\t King, J. A. & Washington, R. Future changes in the Indian Ocean \nWalker Circulation and links to Kenyan rainfall. J. Geophys. Res. \nAtmos. 126, e2021JD034585 (2021).\n59.\t Allen, R. G., Pereira, L. S., Raes, D. & Smith, M. FAO Irrigation and \nDrainage Paper: Crop Evapotranspiration (FAO, 1998).\n60.\t Liang, X., Lettenmaier, D. P., Wood, E. F. & Burges, S. J. A simple \nhydrologically based model of land surface water and energy \nfluxes for general circulation models. J. Geophys. Res. 99, \n14415\u201314428 (1994).\n61.\t David, C. H. et al. River network routing on the NHDPlus dataset.  \nJ. Hydrometeorol. 12, 913\u2013934 (2011).\n62.\t Lin, P. et al. Global reconstruction of naturalized river flows at 2.94 \nmillion reaches. Water Resour. Res. 55, 6499\u20136516 (2019).\n63.\t Development of the Eastern Nile Water Simulation Model (Deltares, \n2013).\n64.\t Gill, M. A. Flood routing by the Muskingum method. J. Hydrol. 36, \n353\u2013363 (1978).\n65.\t Lehner, B. & Grill, G. Global river hydrography and network \nrouting: baseline data and new approaches to study the world\u2019s \nlarge river systems. Hydrol. Process. 27, 2171\u20132186 (2013).\n66.\t Tomlinson, J. E., Arnott, J. H. & Harou, J. J. A water resource \nsimulator in Python. Environ. Model. Softw. 126, 104635 (2020).\n67.\t Wurbs, R. A. Generalized Models of River System Development and \nManagement (IntechOpen, 2011).\n68.\t Basheer, M., Sulieman, R. & Ribbe, L. Exploring management \napproaches for water and energy in the data-scarce \nTekeze-Atbara Basin under hydrologic uncertainty. Int. J. Water \nResour. Dev. 37, 182\u2013207 (2021).\n69.\t Basheer, M. & Elagib, N. A. Sensitivity of water\u2013energy nexus to \ndam operation: a water\u2013energy productivity concept. Sci. Total \nEnviron. 616\u2013617, 918\u2013926 (2018).\n70.\t Basheer, M. et al. Filling Africa\u2019s largest hydropower  \ndam should consider engineering realities. One Earth 3, 277\u2013281 \n(2020).\n71.\t Jeuland, M., Wu, X. & Whittington, D. Infrastructure development \nand the economics of cooperation in the Eastern Nile. Water Int. \nhttps://doi.org/10.1080/02508060.2017.1278577 (2017).\n72.\t Lofgren, H., Lee, R., Robinson, S., Thomas, M. & El-Said, M. A \nStandard Computable General Equilibrium (CGE) Model in GAMS \n(International Food Policy Research Institute, 2002).\n73.\t Armington, P. S. A theory of demand for products distinguished \nby place of production. Staff Pap. 16, 159\u2013178 (1969).\n74.\t Siddig, K., Elagra, S., Grethe, H. & Mubarak, A. A Post-separation \nSocial Accounting Matrix for the Sudan (International Food Policy \nResearch Institute, 2018); https://doi.org/10.2499/1024320695\n75.\t Al-Riffai, P. et al. A Disaggregated Social Accounting Matrix: \n2010/11 for Policy Analysis in Egypt (International Food Policy \nResearch Institute, 2016); http://ebrary.ifpri.org/cdm/ref/\ncollection/p15738coll2/id/130736\n76.\t Ahmed, H. A., Tebekew, T. & Thurlow, J. 2010/11 Social Accounting \nMatrix for Ethiopia: A Nexus Project SAM (International Food Policy \nResearch Institute, 2017); http://ebrary.ifpri.org/utils/getfile/\ncollection/p15738coll2/id/131505/filename/131720.pdf\n77.\t Chepeliev, M. Gtap-Power data base: version 10. J. Glob. Econ. \nAnal. 5, 110\u2013137 (2020).\n78.\t Jiang, L. & O\u2019Neill, B. C. Global urbanization projections for the \nShared Socioeconomic Pathways. Glob. Environ. Change 42, \n193\u2013199 (2017).\n79.\t Four\u00e9, J., B\u00e9nassy-Qu\u00e9r\u00e9, A. & Fontagn\u00e9, L. Modelling the world \neconomy at the 2050 horizon. Econ. Transit. Inst. Change 21, \n617\u2013654 (2013).\n80.\t Gidden, M. J. et al. Global emissions pathways under different \nsocioeconomic scenarios for use in CMIP6: a dataset of \nharmonized emissions trajectories through the end of the \ncentury. Geosci. Model Dev. 12, 1443\u20131475 (2019).\n81.\t Knox, S., Meier, P., Yoon, J. & Harou, J. J. A Python framework for \nmulti-agent simulation of networked resource systems. Environ. \nModel. Softw. 103, 16\u201328 (2018).\n82.\t Deb, K. & Jain, H. An evolutionary many-objective optimization \nalgorithm using reference-point-based nondominated sorting \napproach, part I: solving problems with box constraints. IEEE \nTrans. Evol. Comput. 18, 577\u2013601 (2014).\n83.\t Hadka, D. Platypus. GitHub https://github.com/Project-Platypus/\nPlatypus (2016).\n84.\t Zitzler, E., Thiele, L., Laumanns, M., Fonseca, C. M. & Da Fonseca, \nV. G. Performance assessment of multiobjective optimizers:  \nan analysis and review. IEEE Trans. Evol. Comput. 7, 117\u2013132 \n(2003).\n85.\t Basheer, M. et al. Balancing national economic policy  \noutcomes for sustainable development. Nat. Commun. 13, 5041 \n(2022).\n86.\t Breiman, L. Random Forests. Mach. Learn. 45, 5\u201332 (2001).\n87.\t Pedregosa, F. et al. Scikit-learn: machine learning in Python.  \nJ. Mach. Learn. Res. 12, 2825\u20132830 (2011).\n88.\t Basheer, M., Nechifor, V., Calzadilla, A., Harou, J. J., Data related to \na study on adaptive management of Nile infrastructure. Zenodo \nhttps://doi.org/10.5281/zenodo.5914757 (2022).\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nAcknowledgements\nM.B.\u2019s doctoral degree is funded by the Faculty of Science and \nEngineering of the University of Manchester. This work was supported \nby the UK Research and Innovation Economic and Social Research \nCouncil (grant no. ES/P011373/1) as part of the Global Challenges \nResearch Fund through the \u2018Future Design and Assessment of \nwater-energy-food-environment Mega Systems\u2019 (FutureDAMS) \nresearch project to J.J.H., V.N., A.C., S.G., D.P., N.F., J.S. and H.J.F. We \nthank GAMS Software GmbH for providing licences for mathematical \nsolvers compatible with parallel processing used for economy-wide \nsimulation on supercomputers. We acknowledge the use of the \nComputational Shared Facility and High-Performance Computing of \nthe University of Manchester. The views expressed in this paper are the \nsole responsibility of the authors and do not necessarily reflect those \nof their institutions.\nAuthor contributions\nM.B. wrote the original manuscript. M.B. performed the visualization \nof the results. All authors reviewed and edited the manuscript. \nM.B. developed the adaptive framework linking various data and \nmodel components. M.B. developed and calibrated the Nile River \nsystem model. V.N., M.B. and A.C developed and calibrated the \neconomy-wide models of Ethiopia, Sudan and Egypt. S.G., M.B. and \nJ.S. developed and calibrated the hydrological model of the Nile \nBasin. D.P., N.F. and H.J.F. processed and bias-corrected the climate \nprojections. M.B. performed the simulation and optimization work. \nM.B., J.J.H., V.N. and A.C. conceptualized the study. M.B. formulated \nthe adaptive policy of the Nile. J.J.H. acquired the funding. All authors \ncontributed to the validation and interpretation of the results.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at https://doi.org/10.1038/\ns41558-022-01556-6.\nSupplementary information The online version contains \nsupplementary material available at https://doi.org/10.1038/s41558-\n022-01556-6.\nCorrespondence and requests for materials should be addressed to \nJulien J. Harou.\nPeer review information Nature Climate Change thanks the \nanonymous reviewers for their contribution to the peer review of this \nwork.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nRange of projections\nClimate projection (2021-2050)\nHistorical (1981-2010)\n0\n20\n40\n60\n80\n1\n2\n3\n4\n5\n6\n7\n8\n9\n10 11 12\nPercentage (%)\nMonth\n(a) Nile Basin\n0\n20\n40\n60\n80\n1\n2\n3\n4\n5\n6\n7\n8\n9\n10 11 12\nPercentage (%)\nMonth\n(b) Blue Nile Basin\n0\n20\n40\n60\n80\n1\n2\n3\n4\n5\n6\n7\n8\n9\n10 11 12\nPercentage (%)\nMonth\n(c) White Nile Basin\n0\n20\n40\n60\n80\n1\n2\n3\n4\n5\n6\n7\n8\n9\n10 11 12\nPercentage (%)\nMonth\n(d) Tekeze-Atbara Basin\nExtended Data Fig. 1 | Mean monthly naturalized streamflow calculated as \na percentage of the mean annual naturalized streamflow. Mean monthly \nnaturalized streamflow of the Nile Basin (a) and its major sub-basins (b\u2013d). \nTekeze-Atbara is projected to witness the biggest change in the intra-annual \nnaturalized streamflow variability under the most extreme scenarios, followed \nby the Blue Nile.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nExtended Data Fig. 2 | Standardized streamflow index of the annual \nnaturalized streamflow of the Nile Basin relative to the mean and standard \ndeviation of 1981\u20132010 . Standardized streamflow index for different climate \nchange projections (a-ab). The index is calculated as follows: the annual \nnaturalized streamflow minus the mean annual naturalized streamflow in \n1981\u20132010; this difference is divided by the standard deviation of the annual \nnaturalized streamflow in 1981\u20132010.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nExtended Data Fig. 3 | Annual change in the mean potential evapotranspiration in large-scale irrigation schemes in the Nile Basin. The annual change values \n(a\u2013n) are calculated with respect to the mean potential evapotranspiration in 1981\u20132010, such that change values higher than one indicate an increase in annual \nevapotranspiration.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nExtended Data Fig. 4 | Annual change in the mean potential evaporation \nrates from open water bodies in the Nile Basin. The annual change values \n(a\u2013o) are calculated with respect to the mean potential evaporation rates in \n1981\u20132010, such that change values higher than one indicate an increase in rates. \nPotential evaporation rates from open water bodies were calculated following \nFAO56 recommendations using a multiplication factor applied to potential \nevapotranspiration. GERD stands for Grand Ethiopian Renaissance Dam, UASDC \nstands for Upper Atbara and Setit Dam Complex, KED stands for Khashm Elgirba \nDam, D/S stands for downstream, and HAD stands for High Aswan Dam.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nExtended Data Fig. 5 | Adaptive management policy of the Grand Ethiopian \nRenaissance Dam. The management policy shows the operation decisions for \nthe Grand Ethiopian Renaissance Dam (GERD) in the initial filling and long-term \noperation phases. HAD stands for High Aswan Dam. Labels numbered from 1 to 7 \nare the decision variables in the multiobjective optimization problem. The labels \nnumbered 3, 4, 5 and 7 are the decision variables adapted over time as climate \nchange unfolds.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01556-6\nExtended Data Fig. 6 | Annual water release rules of the Grand Ethiopian Renaissance Dam under the Washington draft proposal. a, Initial filling. b,Long-term \noperation. The Washington draft proposal contains additional rules to mitigate the impacts of droughts, as described in Supplementary Section 2.\n\n1\nnature research  |  reporting summary\nApril 2020\nCorresponding author(s):\nJulien Harou\nLast updated by author(s): Nov 1, 2022\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see our Editorial Policies and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nNo software was used\nData analysis\nThe Python water resources simulation model (Pywr version 1.9.1) is open-source and freely available at: https://github.com/pywr/pywr, and \nbuilding Pywr models can be facilitated by www.waterstrategy.org. The standard CGE model of the International Food Policy Research \nInstitute is open-source and freely accessible through the following link: https://www.ifpri.org/publication/standard-computable-general-\nequilibrium-cge-model-gams-0. The Variable Infiltration Capacity (VIC version 5) model is freely accessible through the following link: https://\nvic.readthedocs.io/en/master/. The Routing Application for Parallel computatIon of Discharge (RAPID version 1.8.0) is freely accessible \nthrough the following link: http://rapid-hub.org/index.html. The Python Network Simulation framework (Pynsim version 0.1.5) is open-source \nand freely available in the following repository: https://github.com/UMWRG/pynsim. The multiobjective Non-dominated Sorting Genetic \nAlgorithm (NSGA-III), the Multi-Objective Evolutionary Algorithm (MOEA) used in the multiobjective search, is open-source and freely \navailable through Platypus version 1.0.4 in the following repository: https://github.com/Project-Platypus/Platypus. The Random Forest \nRegression Algorithm is open-source and freely available through Scikit-learn version 0.24.2 at: https://github.com/scikit-learn/scikit-learn.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and \nreviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\n\n2\nnature research  |  reporting summary\nApril 2020\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nThe economy-related input data supporting this study's findings are available at Zenodo: http://doi.org/10.5281/zenodo.5914757. The Nile River system model and \nits data are not publicly available due to state restrictions and contain information that could compromise research participant privacy/consent. The Nile model data \ncan be made available upon presentation of necessary permissions from the relevant authorities that own the data. The CMIP6 climate projections data can be \naccessed from: https://esgf-node.llnl.gov/search/cmip6/. The baseline population, labor, urbanization, and economic growth data of Ethiopia, Sudan, and Egypt \nassociated with the SSPs can be accessed from the International Institute for Applied System Analysis (IIASA) database: https://tntcat.iiasa.ac.at/SspDb/dsd?\nAction=htmlpage&page=10. The sectoral productivity projections of Ethiopia, Sudan, and Egypt were produced based on data from the Centre d'\u00c9tudes \nProspectives et d'Informations Internationales (CEPII): http://www.cepii.fr/cepii/en/bdd_modele/bdd.asp. The Multi-Source Weighted-Ensemble Precipitation \n(MSWEP) data are available from: http://www.gloh2o.org/mswep/. The Princeton Global Forcing (PGF) data are available from: https://rda.ucar.edu/datasets/\nds314.0/. The river network data used with the Routing Application for Parallel computatIon of Discharge (RAPID) are freely accessible from: https://\nwww.hydrosheds.org/. The World Bank data referred to in the manuscript is freely accessible from: https://data.worldbank.org/.\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nEcological, evolutionary & environmental sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThis study introduces the first planning framework for adaptive management of river infrastructure systems to consider both the \nsocio-economic and hydrological uncertainties of climate change and policy. The framework uses climate and socio-economic data \nfrom the Coupled Model Intercomparison Project 6 (CMIP6) to drive integrated hydrological, economy-wide, and river system \nsimulators of the Nile Basin. Our adaptive planning framework uses artificial intelligence-based algorithms to design efficient adaptive \nplans for climate change. We use the framework to design a cooperative adaptive management policy for the Grand Ethiopian \nRenaissance Dam (GERD) that considers the socio-economic and river system interests of Ethiopia, Sudan, and Egypt.\nResearch sample\nThe targeted region of the study is the Nile Basin. The Nile River Basin, located in northeastern Africa, faces the threat of climate \nchange alongside physical and economic water scarcities, rapidly rising pressures on water resources due to population and \neconomic growth, and a politically complex transboundary water management system\nSampling strategy\nTwenty climate projections were selected from CMIP6 Tier-1 based on uniform sampling to cover the full range of available SSPs, \nradiative forcing, and the joint distribution of EOC change in precipitation and temperature over the Nile Basin. Processing and bias \ncorrection of the GCM simulations was driven by the requirement for transient (2017-2100) 3-hourly forcing data across seven \nclimate variables that govern the surface mass and energy balances in the hydrological model. We developed nine additional \nprojections based on 9 of the 20 initial projections by removing the overall wetting tendency in the CMIP6 ensemble for this region. \nData collection\nThe CMIP6 climate projections data were obtained from: https://esgf-node.llnl.gov/search/cmip6/. The baseline population, labor, \nurbanization, and economic growth data of Ethiopia, Sudan, and Egypt associated with the SSPs were obtained from the International \nInstitute for Applied System Analysis (IIASA) database: https://tntcat.iiasa.ac.at/SspDb/dsd?Action=htmlpage&page=10. The sectoral \nproductivity projections of Ethiopia, Sudan, and Egypt were produced based on data from the Centre d'\u00c9tudes Prospectives et \nd'Informations Internationales (CEPII): http://www.cepii.fr/cepii/en/bdd_modele/bdd.asp. The Multi-Source Weighted-Ensemble \nPrecipitation (MSWEP) data were obtained from: http://www.gloh2o.org/mswep/. The Princeton Global Forcing (PGF) data were \nobtained from from: http://hydrology.princeton.edu/data.pgf.php.\nTiming and spatial scale\nThe overall timeframe of the study is 1979-2100, including climate data downscaling and bias-correction (1979-2100), hydrological \nmodel development and simulation (1979-2100), economy simulation (2011-2045), and river system infrastructure simulation \n(1995-2045). The spatial domain of the study is the Nile Basin.\nData exclusions\nNo data were excluded from the analysis.\nReproducibility\nThe study experiments were designed based on numerical simulations. To ensure that the study experiments are reliable and \nreproducible, we calibrated and validated the simulation models at multiple sites in the Nile Basin over multi-year periods.\nRandomization\nThe Nile river system simulator was calibrated over the period 1995-2004  and validated over the period 2005-2010 at ten locations \nover using historical river flow observations and reservoir and lake water levels. This calibration period was chosen based on the \navailability of common and continuous historical observed data for the ten selected locations.\n\n3\nnature research  |  reporting summary\nApril 2020\nBlinding\nBinding is not relevant to the data because our data are acquired and processed systematically with established computational \npipelines.\nDid the study involve field work?\nYes\nNo\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology and archaeology\nAnimals and other organisms\nHuman research participants\nClinical data\nDual use research of concern\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\n\n\n    Scientific Research Findings:", "answer": "Based on 29 climate projections, we find that both the sign and magnitude of potential changes in naturalized streamflow of the Nile in 2021\u20132050 are highly uncertain. These uncertainties spark the need for an adaptive and cooperative approach. We show that cooperative adaptive management of the GERD yields compromise solutions with economy-wide benefits to Ethiopia, Sudan and Egypt compared with a proposal discussed in Washington, D.C. in 2020. Under an example compromise solution , the mean (based on 29 projections) discounted (at 3%) real gross domestic product (GDP) increases by US$0.77, 0.67 and 0.18 billion in 2020\u20132045 for Ethiopia, Sudan and Egypt, respectively, relative to the Washington draft proposal. These benefits are more pronounced under extreme climate scenarios, with rises in discounted real GDP of up to US$15.8, 6.3 and 3.0 billion over 2020\u20132045 for Ethiopia, Sudan and Egypt, respectively. Our results should be complemented by evaluating the impacts on ecology, groundwater and riparian populations.", "id": 51}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 12 | December 2022 | 1129\u20131135\n1129\nnature climate change\nhttps://doi.org/10.1038/s41558-022-01508-0\nArticle\nRatcheting of climate pledges needed to limit \npeak global warming\nGokul Iyer\u2009\n\u200a\u20091,4, Yang Ou\u2009\n\u200a\u20091,4, James Edmonds\u2009\n\u200a\u20091, Allen A. Fawcett\u2009\n\u200a\u20092, \nNathan Hultman\u2009\n\u200a\u20093, James McFarland2, Jay Fuhrman1, Stephanie Waldhoff\u2009\n\u200a\u20091  \n& Haewon McJeon\u2009\n\u200a\u20091\u2009\nThe new and updated emission reduction pledges submitted by countries \nahead of the Twenty-Sixth Conference of Parties represent a meaningful \nstrengthening of global ambition compared to the 2015 Paris pledges. Yet, \nlimiting global warming below 1.5\u2009\u00b0C this century will require countries to \nratchet ambition for 2030 and beyond. Here, we explore a suite of emissions \npathways to show that ratcheting near-term ambition through 2030 will be \ncrucial to limiting peak temperature changes. Delaying ratcheting ambition \nto beyond 2030 could still deliver end-of-century temperature change \nof less than 1.5\u2009\u00b0C but would result in higher temperature overshoot over \nmany decades with the potential for adverse consequences. Ratcheting \nnear-term ambition would also deliver benefits from enhanced non-CO2 \nmitigation and facilitate faster transitions to net-zero emissions systems in \nmajor economies.\nMany countries and other non-national actors announced new cli-\nmate ambition, actions and targets ahead of the Twenty-Sixth Confer-\nence of Parties (COP26), held in Glasgow in November 2021. COP26 \nprovided the first real demonstration of the 2015 Paris Agreement\u2019s \nmechanism to regularly revisit and enhance national climate strate-\ngies1. By the end of COP26, 151 countries submitted updated and new \nnationally determined contributions (NDCs) outlining plans to cut \nGHG emissions by 20302. Many countries also communicated official or \nunofficial long-term strategies (LTSs) that outline emission reduction \nstrategies through the mid-century3 and net-zero emissions targets4. \nAlthough the updated and new 2030 pledges suggest higher ambition \ncompared to the 2015 Paris pledges5,6, limiting global warming below \n1.5\u2009\u00b0C this century\u2014the aspirational goal of the Paris Agreement\u2014will \nrequire countries to further ratchet or increase ambition in 2030 and \nbeyond6\u201312. Importantly, recognizing the need for countries to ratchet \ntheir ambition beyond their current pledges, Article IV of the Glasgow \nClimate Pact accelerates the previously expected timeline for revising \nthese NDCs and calls for countries \u201cto revisit and strengthen the 2030 \ntargets in their nationally determined contributions \u2026 to align with the \nParis Agreement temperature goal by the end of 2022\u201d13. In addition, \nthe Pact calls for countries that have \u201cnot yet done so to communicate \nnew or updated nationally determined contributions and long-term low \ngreenhouse gas emission development strategies to net-zero emissions \nby or around mid-century\u201d. The Pact also \u201cemphasizes the urgent need \nfor Parties to increase their efforts to collectively reduce emissions \nthrough accelerated action\u201d.\nAs the international community responds to these calls for ratchet-\ning ambition, there is a strong need to understand both the long-term \ntemperature outcomes of ratcheting ambition in 2030 and beyond \nand what this ratcheting implies for sectoral and regional emissions. \nTo address this need, we explore a suite of high ambition emissions \npathways\u2014developed using the Global Change Analysis Model (GCAM; \nMethods)14\u2014in which countries are assumed to use various combina-\ntions of three strategies to ratchet ambition: (1) increasing near-term \nactions through 2030, (2) accelerating post-2030 emissions reductions \nand (3) moving forward the timing of dates that countries pledge to hit \nnet-zero emissions. We then use a reduced-form climate model (Hec-\ntor)15 to compute the end-of-century and peak temperature change \nimplications of the emissions pathways. Our study builds off and \nextends previous modelling studies that have explored high ambition \nemissions pathways6,8,16\u201320. In doing so, our study makes a timely con-\ntribution by exploring pathways that take the 2021 pledges made until \nReceived: 3 May 2022\nAccepted: 21 September 2022\nPublished online: 10 November 2022\n Check for updates\n1Joint Global Change Research Institute, Pacific Northwest National Laboratory and University of Maryland, College Park, MD, USA. 2US Environmental \nProtection Agency, Washington, DC, USA. 3Center for Global Sustainability, School of Public Policy, University of Maryland, College Park, MD, USA.  \n4These authors contributed equally: Gokul Iyer, Yang Ou. \n\u2009e-mail: haewon.mcjeon@pnnl.gov\n\nNature Climate Change | Volume 12 | December 2022 | 1129\u20131135\n1130\nArticle\nhttps://doi.org/10.1038/s41558-022-01508-0\nsection included in the Supplementary Information, we explore more \nambitious emission cuts in 2030 that lie within the literature range \n(Supplementary Fig. 1). We also examine the sensitivity of our results to \nthe methodology used to develop the NDC+ and NDC++ assumptions \n(Supplementary Section 2).\nBeyond 2030, countries are assumed to achieve the same level of \ndecarbonization rate\u2014defined as the annual rate of improvement in \nGHG emissions per unit of gross domestic product (GDP)\u2014as the rate \nbetween 2015 and 2030 or a minimum rate if their decarbonization \nrate is below this minimum rate. Our central assumption about the \npost-2030 minimum decarbonization rate is 2% and our sensitivity \nassumptions are 5% and 8%. These assumptions are consistent with \nprevious studies and are consistent with the average and high rates \nobserved historically (Methods).\nIn all our pathways, countries are assumed to achieve their official \nLTSs and net-zero pledges if any. Countries with LTSs are assumed to \nachieve their LTSs in the target year following a linear path (Methods). \nBeyond the target year, countries are assumed to follow a path defined \nby the minimum decarbonization rate. Our pathways vary in their \nassumptions about the timing of net-zero pledges if any. Our central \nassumption is that countries with net-zero pledges achieve net-zero \nemissions in the target year following a linear path and then continue \nto keep their emissions constant beyond that year. We consider two \nalternative sensitivity assumptions in which countries with net-zero \npledges advance the timing of achieving net-zero emissions by 5  \nor 10\u2009years.\nThe full combination of the above assumptions results in 27 \nemissions pathways. We note that our high ambition pathways are \nmeant to be illustrative and they do not imply feasibility, which would \nrequire accounting for a variety of considerations including ethical \nand political24\u201326.\nTemperature outcomes\nUsing a simple reduced-form climate model (Hector; Methods)15 we \nstudy the implications of the emissions pathways for end-of-century \nand peak global mean surface temperature changes (Fig. 2). Our results \nshow\u2014consistent with other studies\u2014that if countries achieve their \n2021 NDCs, official LTSs and net-zero pledges as stated, global surface \ntemperature change can be limited to <2\u2009\u00b0C warming this century5,6,9\u201312. \nIn addition, many of our high ambition pathways with the NDC+ and \nNDC++ emission levels in 2030 result in <1.5\u2009\u00b0C temperature change \nin 2100. Even if ratcheting of ambition is delayed to beyond 2030, \nend-of-century temperature change can be returned to <1.5\u2009\u00b0C. How-\never, that would require substantial ratcheting of post-2030 ambition \nbeyond historically observed decarbonization rates (Methods)\u2014as in \nthe pathways in which countries achieve the NDC emission level in 2030 \nfollowed by an 8% minimum decarbonization rate.\nRatcheting ambition in the near-term\u2014as in our NDC++ pathways\u2014\nhas marked implications for peak temperature changes. Ratcheting \nambition in the near-term results in lower levels of peak warming. For \ninstance, the peak temperature change in the pathway with NDC emis-\nsion level in 2030 followed by a 8% minimum decarbonization rate and \nnet-zero pledges in the specified target years is 1.77\u2009\u00b0C compared to \n1.82\u2009\u00b0C in the pathway with NDC emission level in 2030 followed by a \n2% minimum decarbonization rate. By contrast, the peak temperature \nchange in the pathway with NDC++ emissions in 2030 followed by a 2% \nminimum decarbonization rate and net-zero pledges in the specified \ntarget years is 1.68\u2009\u00b0C. Ratcheting ambition in the near- and long-term\u2014\nas in the pathways with NDC++ emission level in 2030 followed by an \n8% minimum decarbonization rate\u2014reduces peak temperature change \nfurther (peak temperature changes of 1.67\u2009\u00b0C if net-zero pledges are \nassumed to be achieved in the specified target years). This is an impor-\ntant finding since higher peak temperature changes and therefore \nhigher temperature overshoots\u2014that is, an exceedance of global mean \ntemperature change above the intended threshold before returning to \nthe end of COP26 as a starting point and providing important insights \non the long-term temperature change and sectoral and regional emis-\nsions implications of ratcheting ambition beyond those pledges (see \nSupplementary Section 1 for a detailed literature review).\nEmissions pathways\nOur emissions pathways (Fig. 1) explore a combination of three \nstrategies that countries might use to ratchet and achieve ambition:  \n(1) increasing ambition in the near-term through 2030, (2) increasing \npost-2030 decarbonization rates or (3) achieving net-zero pledges \nsooner.\nIn 2030, countries are assumed to achieve one of three ambition \nlevels, namely, NDC, NDC+ and NDC++. Our central assumption is that \ncountries achieve the updated or new pledges submitted until the end \nof COP26 (Methods). To construct the NDC+ and NDC++ sensitivities, \nwe begin with the ambition level implied in the updated or new pledges \nas assessed by Climate Action Tracker (CAT)21,22. CAT provides one \nof five rating categories for the country pledges (\u20181.5\u2009\u00b0C Paris Agree-\nment compatible\u2019, \u2018almost sufficient\u2019, \u2018insufficient\u2019, \u2018highly insufficient\u2019 \nand \u2018critically insufficient\u2019). In the NDC+ sensitivity, we assume that \ncountries rated \u2018critically insufficient\u2019 and \u2018highly insufficient\u2019 by CAT \nreduce their emissions by 30% below their current NDC. In the NDC++ \nsensitivity, we assume that countries rated \u2018critically insufficient\u2019, \n\u2018highly insufficient\u2019, \u2018insufficient\u2019 and \u2018not assessed\u2019 by CAT also reduce \ntheir emissions by 30% below their current NDC. Although arbitrary, \nthe 30% assumption enables us to explore the implications of greater \nambition and also helps ensure that global 2030 emissions are consist-\nent with existing high ambition scenarios. The GHG emissions in 2030 \nin our NDC++ pathways lie at the higher end of pathways explored in \nthe recent modelling literature6,23. In an additional sensitivity analysis \nGlobal GHG emissions (GtCO2e yr\u20131)\n\u201320\n\u201310\n0\n10\n20\n30\n40\n50\n60\n1990\n2000\n2010\n2020\n2030\n2040\n2050\nYear\n2060\n2070\n2080\n2090\n2100\nHistorical emissions\nUNEP \u201cBelow 1.5 \u00b0C\u201d\nrange in 2030\nCentral:  2% decarbonization\nNet-zero pledges achieved in target year\nMost ambitious: 8% decarbonization\nTiming of net-zero pledges advanced by 10 years\nNDC NDC+ NDC++\nFig. 1 | Global GHG emissions in the pathways modelled using the GCAM. \nThe emissions pathways vary across assumptions about ambition level in 2030, \npost-2030 minimum decarbonization rate and timing of net-zero for countries \nwith net-zero pledges. See text for detailed description of assumptions. The black \ncolour corresponds to the \u2018NDC\u2019 cases, orange to the \u2018NDC+\u2019 cases and blue to the \n\u2018NDC++\u2019 cases. Each colour group comprises nine pathways. The thick bold lines \nin each colour group correspond to the central assumptions about post-2030 \nminimum decarbonization (2%) and year of net-zero (target year as specified). \nThe thick dashed lines correspond to the most ambitious pathway within each \ncolour group. The lighter lines within each colour group correspond to different \nassumptions about the post-2030 minimum decarbonization rate and timing \nof net-zero pledges. The shaded green area represents 15\u201385 percentile range \nof 1.5\u2009\u00b0C pathways with no or limited overshoot from the IPCC SR1.5 report18. \nSee Supplementary Section 2 and Supplementary Table 1 for a mapping of our \npathways with the recently published IPCC Sixth Assessment Report20,52.\n\nNature Climate Change | Volume 12 | December 2022 | 1129\u20131135\n1131\nArticle\nhttps://doi.org/10.1038/s41558-022-01508-0\nbelow the intended level\u2014could expose natural and human systems \nto substantial risks potentially leading to irreversible and adverse \nconsequences such as the loss of some ecosystems27\u201329.\nFurthermore, ratcheting near-term ambition could markedly \nreduce the number of years of overshoot before returning to 1.5\u2009\u00b0C \nthis century (Supplementary Table 1 and Supplementary Fig. 2). For \nexample, the pathways with NDC emission level in 2030 followed by \na 2% or 5% minimum decarbonization rate and net-zero pledges in \nthe specified target years do not return to 1.5\u2009\u00b0C this century and the \nnumber of years of overshoot in the pathway with an 8% minimum \ndecarbonization rate is 67. By contrast, in the pathways with NDC++ \nemission level in 2030, the number of years of overshoot reduces to \n58, 56 and 47\u2009years respectively.\nAdvancing the timing of net-zero pledges can be an important \nratcheting strategy as it could provide the extra push required in the \nlong-term to return 2100 warming to <1.5\u2009\u00b0C and further reduce temper-\nature overshooting. For example, in the pathway with NDC+ emissions \nin 2030 followed by a 5% minimum decarbonization rate and net-zero \npledges in the specified target years, the 2100 and peak temperature \nchanges are, respectively, 1.51\u2009\u00b0C and 1.72\u2009\u00b0C. Advancing the timing of \nnet-zero pledges by 10\u2009years brings the 2100 and peak temperature \nchanges down to 1.46\u2009\u00b0C and 1.68\u2009\u00b0C.\nSectoral implications of ratcheting ambition\nRatcheting ambition would entail rapid reductions in CO2 emissions \nfrom all sectors of the energy system (Fig. 3 and Supplementary Fig. 4),  \nespecially through the mid-century. While some sectors (for exam-\nple, electricity, buildings and industry) decarbonize faster due to the \navailability of many low-carbon technology alternatives, others (for \nexample, transportation) decarbonize slower due to fewer options.\nRatcheting ambition in the near-term results in quicker transitions \nto net-zero emissions energy systems30\u201332. For example, in the pathway \nwith NDC emissions in 2030 followed by a 2% minimum decarboniza-\ntion rate, global CO2 emissions do not get to net-zero this century. \nRatcheting near-term ambition\u2014as in the pathway with NDC++ emis-\nsions in 2030 followed by 2% minimum decarbonization rate\u2014advances \nthe year of global net-zero CO2 emissions to 2053. Ratcheting both \nnear-term and long-term ambition\u2014as in the pathway with NDC++ \nemissions in 2030 followed by an 8% minimum decarbonization rate\u2014\nadvances the year of global net-zero CO2 emissions further to 2052 \n(Supplementary Table 1). Such accelerated declines in CO2 emissions \nare accompanied by rapid transformations throughout the global \nenergy system to phase out fossil fuel-based infrastructures and scale \nup low-carbon technologies such as renewables, nuclear and carbon \ncapture and storage (Supplementary Figs. 5\u20138).\nRatcheting near-term ambition also implies greater reductions \nin non-CO2 emissions, some of which have higher global warming \npotentials and shorter atmospheric lifetimes and therefore play an \nimportant role in both stabilizing long-term temperature change and \nlimiting peak near-term warming (Fig. 3). Non-CO2s respond to climate \npolicy in two ways33. First, fuel switching and associated phasing out of \ncarbon-intensive fuels due to climate policy reduce associated non-CO2 \nemissions (for example, fugitive methane emissions from resource pro-\nduction). Thus, higher CO2 ambition implies higher non-CO2 ambition. \nSecond, non-CO2 emissions that are largely unaffected by fuel switching \nsuch as hydrofluorocarbon (HFC) emissions from cooling energy use \nand industrial process emissions (perfluorocarbons (PFCs) and sulfur \nhexafluoride) respond to climate policy through the implementation \nof additional control measures. In our analysis, the NDC+ and NDC++ \npathways result, respectively, in 18% and 24% reduction in methane \nemissions from the energy system in 2030 relative to 2020. In terms \nof total methane emissions from energy and agricultural systems, the \nreduction is, respectively, 4% and 8%. In comparison, over a hundred \ncountries made a commitment under the Global Methane Pledge\u2014a \nkey outcome of COP26\u2014to collectively reduce methane emissions \nby at least 30% (ref. 34). Future work could explore higher ambition to \nreduce methane emissions\u2014especially from agriculture\u2014than sug-\ngested by our scenarios.\nNevertheless, some non-CO2 emissions such as methane emis-\nsions from cattle due to enteric fermentation are hard to abate35\u201337. \nHence, achieving net-zero GHG emissions beyond 2030 and continued \ndecarbonization over the longer term will require the deployment of \ncarbon dioxide removal (CDR) technologies38,39 to offset these emis-\nsions. Our pathways assume the availability of CDR technologies such \nas bioenergy in combination with carbon capture and storage (BECCS) \nand direct air capture (DAC) in addition to terrestrial sinks40 and assume \nthat the relative roles of CDR deployment versus mitigation in other \nsectors largely depend on economics. Although the scale of CDR in our \npathways is consistent with the extant literature (Supplementary Figs. 9 \nand 10), an important caveat that could affect the feasibility and scale of \n2% decarbonization\n5% decarbonization\n8% decarbonization\n2% decarbonization\n5% decarbonization\n8% decarbonization\n2% decarbonization\n5% decarbonization\n8% decarbonization\nGlobal surface temperature change (\u00b0C)\nNDC \nNDC+ \nNDC++ \n1.1\n1.2\n1.3\n1.4\n1.5\n1.6\n1.7\n1.8\n1.9\n2\nNet-zero pledges\nCentral: Net-zero pledges achieved in target year\nRatchet: Timing of net-zero pledges advanced by\n               5 and 10 years\nPost-2030 decarbonization rate\nCentral: 2% minimum decarbonization rate\nRatchet: 5% and 8% minimum decarbonization rate\n2030 ambition\nCentral: Countries achieve NDCs submitted until end\n               of COP26 (NDC)\nRatchet: Select countries increase 2030 ambition by\n               30% (NDC+)\nAlmost all countries increase 2030 ambition\nby 30% (NDC++) \n2100\nPeak  \nFig. 2 | Temperature change in 2100 (inner bars) and peak temperature \nchange (outer bars) outcomes of the emissions pathways explored in \nthis study. The grey colour corresponds to the \u2018NDC\u2019 cases, orange to the \n\u2018NDC+\u2019 cases and blue to the \u2018NDC++\u2019 cases. See text for description about \nthe construction of the NDC+ and NDC++ cases. The lighter shades within \neach colour group correspond to different assumptions about the post-2030 \nminimum decarbonization rate. While the first bar within each colour and \nshading group corresponds to the central case in which countries with net-zero \npledges are assumed to achieve their pledges in the target year as stated, the \nsecond and third bars assume that countries advance the accomplishment of \ntheir pledges by 5 and 10\u2009years respectively.\n\nNature Climate Change | Volume 12 | December 2022 | 1129\u20131135\n1132\nArticle\nhttps://doi.org/10.1038/s41558-022-01508-0\nCDR deployment that is not fully accounted for in our pathways is that \nthey could interact with societal priorities other than climate, creating \nvarying degrees of synergies and tradeoffs depending on the type and \nscale of CDR measures used41\u201344. In addition, since not all countries are \nequally endowed with CDR potential, achieving net-zero pledges and \nratcheting ambition might need to be supported by cooperative strat-\negies and/or trade38,45. A simple sensitivity analysis that explores the \nimplications of limited CDR availability suggests that the high ambition \npathways explored in this study are feasible under no availability of DAC \nbut that results in greater reductions in CO2 emissions from energy and \nindustrial sectors (Supplementary Figs. 11 and 12). Further research is \nrequired to better understand the role of CDR and emissions trading \nin high ambition emissions pathways16.\nRegional implications of ratcheting ambition\nThe implications of ratcheting ambition for regional emissions \ndepend on whether countries currently have net-zero pledges. For \ncountries with net-zero pledges (for example, China, India and the \nUSA), cumulative emissions grow and then remain flat beyond the \nyear of net-zero under the pathway with central assumptions (Fig. 4).  \nFor such countries, ratcheting ambition in the near-term\u2014as in the \nNDC++ pathways\u2014results in slower growth of emissions and a plateau-\ning of emissions at a lower peak level. In addition, ratcheting ambition \nin the near-term for such countries also facilitates an advancement of \nthe timing of net-zero CO2 emissions (Fig. 5). For example, ratcheting \nnear-term ambition from NDC to NDC++ emissions in 2030 in China, \nBrazil and the USA results in an advancement of the year of net-zero \nCO2 emissions from 2058, 2041 and 2046 to 2057, 2037 and 2044, \nrespectively. This advancement occurs despite the target years for \nthe official net-zero pledges\u2014which are modelled in terms of net-zero \nGHG emissions for the above countries\u2014remaining unchanged (see \nMethods and Supplementary Table 2 for details on how net-zero \npledges are modelled) because it facilitates a more rapid phase out \nof fossil-fuel-based infrastructure. Advancing the target year for net- \nzero pledges further advances the timing of net-zero CO2 emissions \n(Supplementary Table 3).\nBy contrast, cumulative emissions for emerging economies with-\nout net-zero pledges (for example, Middle East, Africa and Southeast \nAsia) grow throughout the century under the NDC pathways\u2014albeit at \na slower rate under the 8% minimum decarbonization rate assumption \n(Fig. 4). For such countries, ratcheting ambition both in the near-term \nand in the long-term\u2014as in the pathway with NDC++ emissions in 2030 \nfollowed by an 8% minimum decarbonization rate\u2014is critical to acceler-\nate the phase out of fossil-based infrastructure and consequently get \nto net-zero CO2 emissions sooner (Fig. 5).\nDiscussion\nThis study provides an ex-ante scientific underpinning to help design \nrevised and more ambitious pledges in response to the calls made in the \n2021 Glasgow Climate Pact and to understand their potential tempera-\nture implications during the century. Our results underscore the impor-\ntance that countries ratchet their ambition in the near-term\u2014through \n2030\u2014to reduce overshooting and thus maximize long-term climate \nbenefits. Our study also underscores the potential hazards of delaying \nthe timing of ratcheting ambition. Although limiting global warming to \n<1.5\u2009\u00b0C by the end of the century is possible even if ratcheting ambition \n2020\n2030\n2040\n2050\n2060\n2070\n2080\n2090\n2100\n2020\n2030\n2040\n2050\n2060\n2070\n2080\n2090\n2100\n\u221220\n0\n20\n40\n60\n\u221220\nYear\nYear\n0\n20\n40\n60\n\u221220\n0\n20\n40\n60\n\u201320\n0\n20\n40\n60\nGHG emissions (GtCO2e)\nF\u2212gases\nN2O Agriculture and land-use\nN2O energy\nCH4 Agriculture and land-use\nCH4 energy\nCO2 transportation\nCO2 other energy\nCO2 industry\nCO2 electricity\nCO2 buildings\nCO2 direct air capture\nCO2 Bioenergy\nCO2 land-use change\na\nb\nc\nd\nFig. 3 | GHG emissions by sector and species in a subset of the pathways \nexplored in the study. a, NDC/2% decarbonization rate/net-zero pledges in \ntarget years. b, NDC/8% decarbonization rate/net-zero pledges in target years. c, \nNDC++/2%decarbonization rate/net-zero pledges in target years. d, NDC++/8% \ndecarbonization rate/net-zero pledges in target years. See Supplementary Fig. \n3 for GHG emissions by sector and species in all of the pathways explored in this \nstudy and Supplementary Fig. 4 for a version of the figure with differences in \npanels b, c and d relative to a.\n\nNature Climate Change | Volume 12 | December 2022 | 1129\u20131135\n1133\nArticle\nhttps://doi.org/10.1038/s41558-022-01508-0\nis delayed, it would result in higher overshooting during the century \npotentially for a period spanning decades which could lead to irreversi-\nble and adverse consequences for human and natural systems. Delaying \nratcheting of ambition may also require accelerating post-2030 decar-\nbonization to rates that are substantially higher than historical rates. \nFor analysts, these results emphasize the need for future research to \n2020\n2030\n2040\n2050\n2060\n2070\n2080\n2090\n2100\n2020\n2030\n2040\n2050\n2060\n2070\n2080\n2090\n2100\n0\n500\n1,000\n1,500\n2,000\n2,500\n3,000\n3,500\n4,000\n0\n500\n1,000\n1,500\n2,000\n2,500\n3,000\n3,500\n4,000\nYear\nYear\n0\n500\n1,000\n1,500\n2,000\n2,500\n3,000\n3,500\n4,000\n0\n500\n1,000\n1,500\n2,000\n2,500\n3,000\n3,500\n4,000\nAustralia and New Zealand\nOther Europe\nRussia\nEU\nSouth America and Caribbean\nBrazil\nMexico\nCanada\nUSA\nAfrica\nMiddle East\nOther Asia\nIndia\nChina\nCumulative GHG emissions from 2020 (GtCO2e)\na\nb\nc\nd\nFig. 4 | Cumulative GHG emissions by region in a subset of the pathways \nexplored in the study. a, NDC/2% decarbonization rate/net-zero pledges in \ntarget years. b, NDC/8% decarbonization rate/net-zero pledges in target years.  \nc, NDC++/2%decarbonization rate/net-zero pledges in target years. d, NDC++/8% \ndecarbonization rate/net-zero pledges in target years. See Supplementary Fig. 13 \nfor GHG emissions by region in all of the pathways explored in this study.\n2030\n2040\n2050\n2060\n2070\n2080\n2090\n>2100\nYear of \nnet-zero CO2\n>2100\n2052\n2054\n2047\n2058\n2058\n2057\n2057\n2046\n2046\n2044\n2044\n>2100\n2087\n2087\n2051\n2070\n2070\n2070\n2070\n>2100\n2068\n>2100\n2062\n2041\n2041\n2037\n2037\n2045\n2045\n2042\n2042\n2046\n2046\n2044\n2044\n2044\n2044\n2042\n2042\n2050\n2050\n2050\n2050\na\nb\nc\nd\nFig. 5 | Year of net-zero CO2 emissions in a subset of the pathways explored in the \nstudy. a, NDC/2% decarbonization rate/net-zero pledges in target years. b, NDC/8% \ndecarbonization rate/net-zero pledges in target years. c, NDC++/2%decarbonization \nrate/net-zero pledges in target years. d, NDC++/8% decarbonization rate/net-zero \npledges in target years. See Supplementary Table 3 for year of net-zero CO2 emissions \nin all of the pathways explored in this study. Global maps in this figure are created \nusing an open-source R package53 and documented in ref. 54. \u201c>2100\u201d indicates the \nyear of net-zero CO2 emission occurs after 2100.\n\nNature Climate Change | Volume 12 | December 2022 | 1129\u20131135\n1134\nArticle\nhttps://doi.org/10.1038/s41558-022-01508-0\nexplore emissions pathways that focus not only end-of-century tem-\nperature targets but also alternative pathways that limit the degree \nof temperature overshoot during the century\u2014especially those with \nhigher near-term ambition than implied by the current set of NDCs. A \nfew studies have begun to explore such alternative pathways23,29 but \nmore community-wide studies\u2014including intermodel comparison \nefforts\u2014could help collect robust insights about the costs and benefits \nof ratcheting ambition in the near-term and the technological options \nthat could facilitate the implementation of higher near-term ambition.\nOur study suggests a strong potential for non-CO2 mitigation in \nfacilitating the higher ambition needed. Previous analyses have shown \nthat to limit temperature change to 1.5\u2009\u00b0C, mitigation strategies focused \nonly on CO2 reduction could require getting to net-zero two decades \nsooner than comprehensive strategies that include non-CO2s as well33. \nWhile the Global Methane Pledge34 is a step in the right direction to \nmotivate higher non-CO2 ambition, comprehensive strategies that \naccount for a wider suite of GHGs would ultimately be required to \nenable cost-effective emission reductions33,46.\nOur results also suggest that ratcheting near-term ambition could \nenable faster transitions required to accomplish net-zero pledges, \nespecially in major emitting economies. These transitions could very \nwell be accomplished with limited availability of nascent technologies \nsuch as CDR. However, the economic implications of these transitions \nwould hinge on the availability of CDR and other nascent technologies \nsuch as CCS and the ability of grid infrastructures to expand rapidly \nas technologies such as renewables scale up. The speed and scale at \nwhich these technologies can be deployed depend on a variety of \nfactors, such as costs, access to financial capital, supply-chain issues, \nland-use and geophysical constraints and other institutional, social and \nbehavioural factors26,47\u201350. Such factors could imply severe economic \nconsequences47,48. Future work should consider these factors and other \nreal-world political and ethical ramifications of ratcheting ambition \nto better understand the feasibility of the high ambition pathways \nexplored in this study51. Nonetheless, ratcheting near-term ambition \nand concurrently testing out and establishing policies and institutional \ninfrastructures that phase out fossil fuels and incentivize research \nand development and deployment of more nascent technologies in \nthe near-term will be crucial to facilitate deeper emissions reductions \nin the long-term needed to cost-effectively accomplish the long-term \ngoals of the Paris Agreement39.\nOnline content\nAny methods, additional references, Nature Research reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author contri-\nbutions and competing interests; and statements of data and code avail-\nability are available at https://doi.org/10.1038/s41558-022-01508-0.\nReferences\n1.\t\nThe Paris Agreement (UNFCCC, 2021); https://unfccc.int/\nprocess-and-meetings/the-paris-agreement/the-paris-agreement\n2.\t\nNationally Determined Contributions (NDCs) (UNFCCC, \n2021); https://unfccc.int/process-and-meetings/\nthe-paris-agreement/nationally-determined-contributions-ndcs/\nnationally-determined-contributions-ndcs#eq-1\n3.\t\nCommunication of Long-term Strategies (UNFCCC, 2021); https://\nunfccc.int/process/the-paris-agreement/long-term-strategies\n4.\t\nCAT Net Zero Target Evaluations (Climate Action \nTracker, 2021); https://climateactiontracker.org/global/\ncat-net-zero-target-evaluations/\n5.\t\nOu, Y. et al. Can updated climate pledges limit warming well \nbelow 2\u00b0C?. Science 374, 693\u2013695 (2021).\n6.\t\nEmissions Gap Report 2021: The Heat Is On\u2014A World of Climate \nPromises Not Yet Delivered (United Nations Environment \nProgramme, 2021).\n7.\t\nGlasgow\u2019s 2030 Credibility Gap: Net Zero\u2019s Lip Service to Climate \nAction (Climate Action Tracker, 2021).\n8.\t\nGrant, N. The Paris Agreement\u2019s ratcheting mechanism needs \nstrengthening 4-fold to keep 1.5\u2009\u00b0C alive. Joule https://doi.\norg/10.1016/j.joule.2022.02.017 (2022).\n9.\t\nMeinshausen, M. et al. Realization of Paris Agreement pledges \nmay limit warming just below 2\u2009\u00b0C. Nature 604, 304\u2013309 (2022).\n10.\t IPCC. Summary for Policymakers. In Climate Change 2022: Mitigation \nof Climate Change. Contribution of Working Group III to the Sixth \nAssessment Report of the Intergovernmental Panel on Climate \nChange, (eds Shukla, P.R. et al.) (Cambridge Univ. Press, 2022).\n11.\t\nLecocq, F. et al. in Climate Change 2022: Mitigation of Climate \nChange. Contribution of Working Group III to the Sixth Assessment \nReport of the Intergovernmental Panel on Climate Change (eds \nShukla, P. R. et al.) Ch. 4, (Cambridge Univ. Press, 2022).\n12.\t den Elzen, M. G. J. et al. Updated nationally determined \ncontributions collectively raise ambition levels but need \nstrengthening further to keep Paris goals within reach. Mitig. \nAdapt. Strateg. Glob. Change 27, 33 (2022).\n13.\t The Glasgow Climate Pact (UNFCCC, 2021); https://unfccc.int/\nsites/default/files/resource/cop26_auv_2f_cover_decision.pdf\n14.\t Calvin, K. et al. GCAM v5.1: representing the linkages between \nenergy, water, land, climate, and economic systems. Geosci. \nModel Dev. 12, 677\u2013698 (2019).\n15.\t Hartin, C. A., Patel, P., Schwarber, A., Link, R. P. & Bond-Lamberty, \nB. P. A simple object-oriented and open-source model for \nscientific and policy analyses of the global climate system\u2014\nHector v1.0. Geosci. Model Dev. 8, 939\u2013955 (2015).\n16.\t Holz, C., Siegel, L. S., Johnston, E., Jones, A. P. & Sterman, J. \nRatcheting ambition to limit warming to 1.5\u2009\u00b0C\u2013trade-offs between \nemission reductions and carbon dioxide removal. Environ. Res. \nLett. 13, 064028 (2018).\n17.\t Bharadwaj, B. & Brierley, C. Ratcheting up Ambition in Climate \nPolicy (Environmental Change Research Centre, 2017); https://\ndiscovery.ucl.ac.uk/id/eprint/1541174/1/ecrc_report_182_\nBharadwaj_Brierley.pdf\n18.\t IPCC Special Report on Global Warming of 1.5\u2009\u00b0C (eds \nMasson-Delmotte, V. et al.) (WMO, 2018).\n19.\t Iyer, G. C. et al. The contribution of Paris to limit global warming to \n2\u2009\u00b0C. Environ. Res. Lett. 10, 125002 (2015).\n20.\t Riahi, K. et al. in Climate Change 2022: Mitigation of Climate \nChange. Contribution of Working Group III to the Sixth Assessment \nReport of the Intergovernmental Panel on Climate Change (eds \nShukla, P. R. et al.) Ch. 3, (Cambridge Univ. Press, 2022).\n21.\t Countries (Climate Action Tracker, 2021); https://\nclimateactiontracker.org/countries/\n22.\t Rating System (Climate Action Tracker, 2021); https://\nclimateactiontracker.org/countries/rating-system/\n23.\t Riahi, K. et al. Cost and attainability of meeting stringent climate \ntargets without overshoot. Nat. Clim. Change 11, 1063\u20131069 (2021).\n24.\t van den Berg, N. J. et al. Implications of various effort-sharing \napproaches for national carbon budgets and emission pathways. \nClim. Change 162, 1805\u20131822 (2019).\n25.\t Rajamani, L. et al. National \u2018fair shares\u2019 in reducing greenhouse \ngas emissions within the principled framework of international \nenvironmental law. Clim. Policy 21, 983\u20131004 (2021).\n26.\t Peng, W. et al. Climate policy models need to get real about \npeople\u2014here\u2019s how. Nature 594, 174\u2013176 (2021).\n27.\t Hoegh-Guldberg, O. et al. in Global Warming of 1.5\u00b0C. An IPCC \nSpecial Report on the impacts of global warming of 1.5\u00b0C above \npre-industrial levels and related global greenhouse gas emission \npathways, in the context of strengthening the global response \nto the threat of climate change, sustainable development, and \nefforts to eradicate poverty (eds Masson-Delmotte, V. et al.) Ch. 3, \n(Cambridge Univ. Press, 2018).\n\nNature Climate Change | Volume 12 | December 2022 | 1129\u20131135\n1135\nArticle\nhttps://doi.org/10.1038/s41558-022-01508-0\n28.\t Geden, O. & L\u00f6schel, A. Define limits for temperature overshoot \ntargets. Nat. Geosci. 10, 881\u2013882 (2017).\n29.\t Drouet, L. et al. Net zero-emission pathways reduce the physical \nand economic risks of climate change. Nat. Clim. Change 11, \n1070\u20131076 (2021).\n30.\t Davis, S. J. et al. Net-zero emissions energy systems. Science \nhttps://doi.org/10.1126/science.aas9793 (2018).\n31.\t Azevedo, I., Bataille, C., Bistline, J., Clarke, L. & Davis, S. Net-zero \nemissions energy systems: what we know and do not know. \nEnergy Clim. Change https://doi.org/10.1016/j.egycc.2021.100049 \n(2021).\n32.\t DeAngelo, J. et al. Energy systems in scenarios at net-zero CO2 \nemissions. Nat. Commun. 12, 6096 (2021).\n33.\t Ou, Y. et al. Deep mitigation of CO2 and non-CO2 greenhouse \ngases toward 1.5\u2009\u00b0C and 2\u2009\u00b0C futures. Nat. Commun. 12, 6245 \n(2021).\n34.\t The Global Methane Pledge (Climate & Clean Air Coalition, 2021); \nhttps://www.globalmethanepledge.org/\n35.\t Herrero, M. et al. Greenhouse gas mitigation potentials in the \nlivestock sector. Nat. Clim. Change 6, 452\u2013461 (2016).\n36.\t Clark, M. A. et al. Global food system emissions could preclude \nachieving the 1.5\u00b0 and 2\u2009\u00b0C climate change targets. Science 370, \n705\u2013708 (2020).\n37.\t Harmsen, M. et al. The role of methane in future climate \nstrategies: mitigation potentials and climate impacts. Clim. \nChange https://doi.org/10.1007/s10584-019-02437-2  \n(2019).\n38.\t Iyer, G. et al. The role of carbon dioxide removal in net-zero \nemissions pledges. Energy Clim. Change https://doi.org/10.1016/j.\negycc.2021.100043 (2021).\n39.\t The Long-Term Strategy of the United States: Pathways to Net-Zero \nGreenhouse Gas Emissions by 2050 (United States Department \nof State, 2021); https://www.whitehouse.gov/wp-content/\nuploads/2021/10/US-Long-Term-Strategy.pdf\n40.\t Fuhrman, J. et al. The role of direct air capture and negative \nemissions technologies in the shared socioeconomic pathways \ntowards +1.5\u2009\u00b0C and +2\u2009\u00b0C futures. Environ. Res. Lett. https://doi.\norg/10.1088/1748-9326/ac2db0 (2021).\n41.\t Roy, J. et al. in Global Warming of 1.5\u00b0C. An IPCC Special Report \non the impacts of global warming of 1.5\u00b0C above pre-industrial \nlevels and related global greenhouse gas emission pathways, in \nthe context of strengthening the global response to the threat of \nclimate change, sustainable development, and efforts to eradicate \npoverty (eds Masson-Delmotte, V. et al.) Ch. 5, (Cambridge Univ. \nPress, 2018).\n42.\t Fuss, S. et al. Moving toward net-zero emissions requires  \nnew alliances for carbon dioxide removal. One Earth 3,  \n145\u2013149 (2020).\n43.\t Fuhrman, J. et al. Food\u2013energy\u2013water implications of negative \nemissions technologies in a +1.5\u2009\u00b0C future. Nat. Clim. Change 10, \n920\u2013927 (2020).\n44.\t Fuhrman, J., McJeon, H., Doney, S. C., Shobe, W. & Clarens, A. \nF. From zero to hero?: why integrated assessment modeling of \nnegative emissions technologies is hard and how we can do better. \nFront. Clim. https://doi.org/10.3389/fclim.2019.00011 (2019).\n45.\t Edmonds, J., Forrister, D., Clarke, L., de Clara, S. & Munnings, \nC. The Economic Potential of Article 6 of the Paris Agreement \nand Implementation Challenges (IETA, 2019); https://www.ieta.\norg/resources/International_WG/Article6/CLPC_A6%20report_\nno%20crops.pdf\n46.\t Purohit, P., Borgford-Parnell, N., Klimont, Z. & H\u00f6glund-Isaksson, \nL. Achieving Paris climate goals calls for increasing ambition of \nthe Kigali Amendment. Nat. Clim. Change 12, 339\u2013342 (2022).\n47.\t Iyer, G. et al. Diffusion of low-carbon technologies and the \nfeasibility of long-term climate targets. Technol. Forecast. Soc. \nChange 90, 103\u2013118 (2015).\n48.\t Iyer, G. et al. Improved representation of investment decisions in \nassessments of CO2 mitigation. Nat. Clim. Change 5, 436\u2013440 \n(2015).\n49.\t Tong, D. et al. Geophysical constraints on the reliability of solar \nand wind power worldwide. Nat. Commun. 12, 6146 (2021).\n50.\t Fuss, S. et al. Negative emissions\u2014Part 2: Costs, potentials and \nside effects. Environ. Res. Lett. https://doi.org/10.1088/1748-9326/\naabf9f (2018).\n51.\t Brutschin, E. et al. A multidimensional feasibility evaluation \nof low-carbon scenarios. Environ. Res. Lett. https://doi.\norg/10.1088/1748-9326/abf0ce (2021).\n52.\t IPCC Climate Change 2022: Mitigation of Climate Change (eds \nShukla, P. R. et al.) (Cambridge Univ. Press, 2022).\n53.\t Pebesma, E. Simple features for R: standardized support for \nspatial vector data. R. J. 10, 439\u2013446 (2018).\n54.\t Ou, Y. Source code for reproducing Fig5 (map) in Iyer & Ou, et al. \n2022 Ratcheting of climate pledges needed to limit peak global \nwarming. Zenodo https://doi.org/10.5281/zenodo.7082257 (2022).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\nThis is a U.S. Government work and not under copyright protection in \nthe US; foreign copyright protection may apply 2022\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01508-0\nMethods\nThe Global Change Analysis Model\nGCAM is an open-source model developed and maintained at the Pacific \nNorthwest National Laboratory\u2019s Joint Global Change Research Insti-\ntute. In this study, we use the version of the GCAM (v.5.3) used in ref. 5  \nand available in a public repository55. The full documentation of the \nmodel is available at the GCAM documentation page (http://jgcri.\ngithub.io/gcam-doc/) and the description here is a summary of the \nonline documentation and based on refs. 5,56\u201359.\nGCAM includes representations of five systems: economy, energy, \nagriculture and land-use, water and climate in 32 geopolitical regions \nacross the globe and the associated land allocation, water use and agri-\nculture production across 384 land subregions and 235 water basins. \nGCAM operates in 5-year time-steps from 2015 (calibration year) to \n2100 by solving for the equilibrium prices and quantities of various \nenergy, agricultural, water, land-use and GHG markets in each time \nperiod and in each region. GCAM is a dynamic recursive model. Hence, \nsolutions for each modelling period only depend on conditions in the \nlast modelling period. Outcomes of GCAM are driven by exogenous \nassumptions about population growth, labour participation rates and \nlabour productivity in the 32 geopolitical regions, along with represen-\ntations of resources, technologies and policy. GCAM tracks emissions \nof 24 gases, including GHGs, short-lived species and ozone precursors, \nendogenously based on the resulting energy, agriculture and land-use \nsystems as discussed in the following subsections.\nThe GCAM energy system contains representations of fossil \nresources (coal, oil and gas), uranium and renewable sources (wind, \nsolar, geothermal, hydro and biomass and traditional biomass) along \nwith processes that transform these resources to final energy carriers \n(electricity generation, refining, hydrogen production, gas processing \nand district heat), which are ultimately used to deliver goods and ser-\nvices demanded by end-use sectors (residential buildings, commercial \nbuildings, transportation and industry). Each of the sectors in GCAM \ninclude technological detail. For example, the electricity generation \nsector includes several different technology options to convert coal to \nelectricity such as pulverized coal with and without carbon capture and \nstorage (CCS) and coal integrated gasification combined cycle (IGCC) \nwith and without CCS. In every sector within GCAM, individual tech-\nnologies compete for market share on the basis of the levelized cost of \na technology. The cost of a technology in any period depends on (1) its \nexogenously specified non-energy cost, (2) its endogenously calculated \nfuel cost and (3) any cost of emissions, as determined by the climate \npolicy. The first term, non-energy cost, represents capital, fixed and \nvariable operation and maintenance costs incurred over the lifetime \nof the equipment (except for fuel or electricity costs), expressed per \nunit of output. For example, the non-energy cost of coal-fired power \nplant is calculated as the sum of overnight capital cost (amortized using \na capital recovery factor and converted to dollars per unit of energy \noutput by applying a capacity factor), fixed and variable operations and \nmaintenance costs. The second term, fuel or electricity cost, depends \non the specified efficiency of the technology, which determines the \namount of fuel or electricity required to produce each unit of output, \nas well as the cost of the fuel or electricity. The various data sources \nand assumptions are documented in the GCAM documentation page \n(http://jgcri.github.io/gcam-doc/).\nThe prices of fossil fuels and uranium are calculated endogenously. \nFossil fuel resource supply in GCAM is modelled using graded resource \nsupply curves that represent increasing cost of extraction as cumula-\ntive extraction increases. Wind and rooftop PV technologies include \nresource costs that are also calculated from exogenous supply curves \nthat represent marginal costs that increase with deployment, such \nas long-distance transmission line costs that would be required to \nproduce power from remote wind resources. Utility-scale solar pho-\ntovoltaic and concentrated solar power technologies are assumed to \nhave constant marginal resource costs regardless of deployment levels.\nIn GCAM, technology choice is determined by market competition. \nThe market share captured by a technology increases as its costs decline \nbut GCAM uses a logit model of market competition. This approach \nis designed to represent decision-making among competing options \nwhen only some characteristics of the options can be observed60,61 and \navoids a \u2018winner takes all\u2019 response.\nThe agriculture and land-use component of GCAM represents \ncompetition for land among alternative uses in 283 agro-economic \nzones within the 32 regions. Land is allocated between alternative \nuses such as food crops (including wheat, corn, rice, root and tuber \nand other grain), commercial biomass, forests, pasture, grassland \nand shrubs based on expected profitability according to a logit-share \nmechanism similar to the energy system. The profitability in turn \ndepends on the productivity of the land-based product (for example, \nmass of harvestable product per hectare), product price and non-land \ncosts of production (labour, fertilizer and so on). The productivity of \nland-based products is subject to change over time based on future \nestimates of crop productivity change. GCAM also tracks land from \ndesert, tundra and urban land. However, these are excluded from \neconomic competition and assumed to be fixed over time. Yields for all \ncrops are assumed to improve over time. These improvement rates vary \nby region, with higher improvement rates in developing regions. The \nenergy system and the agriculture and land-use systems are hard linked \n(coupled in code). Commercial biomass is demanded in the energy \nsystem while its supply is modelled in the agriculture and land-use \ncomponent. Fertilizer supply is represented in the energy system while \nfertilizer demand is modelled in the agriculture and land-use system. \nTraditional biomass is not modelled in the agriculture and land-use \nsystem but is instead represented through exogenous supply curves \nthat account for the opportunity cost associated with collecting tradi-\ntional biomass\u2014collecting traditional biomass requires labour which \nbecomes increasingly expensive as incomes rise.\nGCAM tracks emissions of a variety of GHG species: CO2, CH4, N2O, \nHFCs (HFC23, HFC32, HFC125, HFC134a, HFC143a, HFC152a, HFC227ea, \nHFC43, HFC236fa, HFC365mfc and HFC245fa), PFCs (CF4 and C2F6) \nand SF6. The CO2 emissions result from direct combustion of fossil \nfuels and conversion to other forms. Once a fossil fuel is extracted, \nthe carbon in the fuel is either emitted or sequestered. The total CO2 \nemissions in the base year of GCAM (currently 2015) is calibrated to the \nCarbon Dioxide Information Analysis Center database62 at the global \nlevel and fossil fuel consumption in the base year is calibrated to the \nInternational Energy Agency\u2019s Energy Balances Database63. Global aver-\nage emissions coefficients (for example, CO2 per GJ) are derived from \nthe ratio of the total emissions and the total fuel consumption for each \nfossil fuel (coal, oil and gas). In each model period, CO2 emissions from \na technology are calculated as the product of global average emission \ncoefficients obtained above and fuel consumption by that technology \nin that period. Agriculture and land-use change emissions depend on \nthe amount of land-use change, the equilibrium carbon density of the \necosystem and region-specific growth profiles64.\nGCAM also tracks non-CO2 emissions from the energy and agri-\ncultural and land-use systems. Historical emissions of CH4, N2O and \nF-gases are harmonized with the 2019 US Environmental Protection \nAgency (EPA) Global Non-CO2 Greenhouse Gas Emission Projections \nand Mitigation Potential report65. Historical emissions of short-lived \nforcing agents (BC and OC) and air pollutants (SO2, NOx and PM2.5) are \ncalibrated to the Community Emissions Data System66. These histori-\ncal emissions are then used to develop emission factors (emission per \nenergy input or service output of a specific technology). Emissions \nfactors are assumed to change over time if air pollution controls are \ntightened (local air pollutants only) or a carbon price is applied (GHGs \nonly; not all sectors). Future emissions are estimated as the product of \nthe projected economic activity, the corresponding emission factor \nfor a given technology and emissions reductions estimated through \nmarginal abatement cost (MAC) curves. MAC curves are based on ref. 65.\n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01508-0\nIn our pathways, non-CO2 emissions can be controlled by two mech-\nanisms. First, changes in activity (phasing out of carbon-intensive fuels \ndue to climate policy) will reduce non-CO2 emissions (for example, \nfugitive CH4 from natural gas production). Second, for emission sources \nwithout explicit representation of the underlying activity, emission \nreductions are calculated off of MAC curves that are parametrized to \nabatement technologies and abatement levels. MAC curves represent \nthe mitigation cost and corresponding emission reductions achievable \nfor each region, species and available source categories over time.\nThe version of GCAM used in this study includes important recent \ntechnological and socioeconomic trends. First, the effect of COVID-19  \non the global economy is reflected by incorporating the latest \ncountry-specific International Monetary Fund GDP growth projec-\ntions67. Second, electric power technology cost assumptions (capital \ncost, operation and maintenance cost and efficiency) follow recent \ntrends and projections and are based on the 2019 National Renew-\nable Energy Laboratory (NREL) Annual Technology Baseline68. These \nassumptions entail substantial capital and operation and mainte-\nnance cost reductions for most technologies, especially solar and wind \ntechnologies. Third, the version of GCAM used in this study includes \nelectrification options in the transportation sector including electric \nvehicles and electric trucks. Our transportation cost and energy inten-\nsity assumptions are based on the NREL Electrification Futures Study69.\nThe version of GCAM used in this study assumes the availability \nof three CDR options: afforestation, BECCS and DAC technologies. \nThe scale of each option is determined by economics. Our pathways \nincentivize afforestation by assuming a gradual transition\u2014by 2050\u2014to \na regime in which CO2 emissions from land-use changes are valued at \nthe same price as emissions from the energy system59,70. As described \nearlier, in GCAM, bioenergy competes for land with other land uses on \nthe basis of profitability. BECCS technologies are deployed in a variety \nof sectors within the GCAM energy system including refining, electric-\nity generation and hydrogen production. Our assumptions for DAC \ntechnologies are documented in Supplementary Table 4 and refs. 43,40.\nHector\nHector is the reduced-form carbon-cycle climate module that is avail-\nable for use in GCAM15,71 and is an open-source model. This study is \nbased on Hector v.2.5. Hector has a three-part carbon cycle: one-pool \natmosphere, three-pool land and four-pool ocean. The model\u2019s ter-\nrestrial carbon cycle includes primary production and respiration \nfluxes while also accommodating arbitrary geographic divisions, such \nas ecological biomes or political units. Hector\u2019s ocean component \nincludes a detailed representation of the inorganic carbon cycle, cal-\nculating air\u2013sea fluxes and ocean pH (ref. 71). Hector reproduces the \nglobal historical trends of atmospheric CO2, radiative forcing and \nsurface temperatures.\nGCAM interacts with Hector through emissions. At every time step, \nemissions from GCAM are passed to Hector. Hector then converts these \nemissions to concentrations when necessary and calculates the associ-\nated radiative forcing, as well as the response of the climate system (for \nexample, surface temperature and carbon fluxes).\nEmissions pathways\nThe representation of the NDCs in our central pathway is based on \nref. 5 and is explained in detail in the supplementary information to \nthat study. This study also includes 21 new and/or updated NDCs after \n30 September 2021, including those from China, Pakistan and many \nAfrican and Middle Eastern countries that were not included in ref. 5 \n(Supplementary Table 5). We assume that the NDCs are achieved as \nstipulated and focus on the climate outcomes of their successful imple-\nmentation. Examining the likelihood of individual regions achieving \ntheir submitted targets is beyond the scope of this study.\nOur representation includes only ratified and quantifiable uncon-\nditional NDC commitments, including absolute emissions limit, \npercentage emission reductions from a given reference level and emis-\nsion intensity targets. Parties whose commitments included: (1) only \nactions/policies, (2) non-GHG targets with no corresponding GHG \nemissions target or (3) only sector-specific GHG emissions reduction \ntargets without attempting to quantify the impact on their overall GHG \nfootprints are assumed to have target year emissions equal to the GCAM \nemissions in the default reference scenario without any climate policy \n(\u2018Reference\u2014No Policy\u2019 in ref. 5). Likewise, in cases where a country\u2019s \n2025 and 2030 emissions based on its NDC are lower than the default \nreference scenario in the same year, the NDC emissions are assumed to \nbe achieved as stipulated. In cases where a country\u2019s NDC emissions are \nhigher, emissions are assumed to be equal to the reference scenario. \nFor countries that included multiple types of commitments in their \nNDCs, such as economy-wide emissions reductions backed by sectoral \npolicies or targets, only the broadest commitment was considered. \nFor example, China\u2019s NDC representation in GCAM is based on its \ncommitment to reduce its carbon intensity of GDP by 65% relative to \n2005 and it does not explicitly model its targets for non-fossil energy \nconsumption or increased forest stock.\nSimilar to ref. 5, our pathways include LTSs and net-zero pledges. \nFor countries with LTSs that are different from a net-zero pledge (for \nexample, Mexico), emissions are assumed to meet their NDC commit-\nments in 2030 first. Beyond 2030, emissions linearly reduce to the LTS \nin the specified target year and then continue to follow a path defined \nby the decarbonization rate between 2015 and the LTS target year. For \ncountries with net-zero pledges, emissions are assumed to meet their \nNDC commitments in 2030 first. Beyond 2030, emissions linearly reduce \nto net-zero in the target year and then remain constant afterwards. In \nthe cases where countries have explicitly committed to net-zero CO2 \nemissions, such as South Korea, only CO2 emissions are constrained. This \nstudy also includes additional net-zero pledges that were announced \nafter the completion of the ref. 5 study. These include pledges from India, \nBrazil, Australia, New Zealand and Argentina (Supplementary Table 2). \nWhere the scope of net-zero targets is somewhat unclear (as in the case \nof Japan) or in cases where countries use terms such as \u2018carbon neutral\u2019 \nand \u2018net-zero GHG emissions\u2019 interchangeably, we follow the CAT assess-\nment and assume a net-zero GHG target. For example, China announced \na \u2019carbon neutrality\u2019 goal by 2060, which is assumed as a net-zero GHG \nemission target in our main analysis. This assumption is consistent with \nlatest official interpretations of China\u2019s net-zero pledge72.\nOur post-2030 decarbonization rate assumptions are consistent \nwith refs. 57,5. However, our definition of decarbonization rate is based \non all GHGs while the definitions used by ref. 57,5 are based only on fossil \nfuel and industrial CO2 emissions (note that the emissions scenarios \nmodelled in the studies of refs. 57,5 do include concurrent reductions \nin non-CO2s in response to CO2 reductions that are facilitated by the \ndecarbonization rate assumptions). Our central assumption about \nthe post-2030 minimum decarbonization rate is 2% and our sensitiv-\nity assumptions are 5% and 8%. While the 2% rate has been achieved \nroutinely in history and represents a moderate level of post-2030 \nmitigation, the 5% and 8% decarbonization rate assumptions can be \nconsidered as requiring more dedicated, stringent mitigation policies \n(Supplementary Fig. 14). For additional context, the 2% decarboni-\nzation rate falls under the higher end of the distribution of decar-\nbonization rates implied in the \u2018baseline\u2019 scenarios assessed by the \nIntergovernmental Panel on Climate Change (IPCC) Special Report on \n1.5\u2009\u00b0C (SR1.5) and the 5% and 8% assumptions lie at the peak of the distri-\nbution of decarbonization rates in scenarios limiting global warming \nto 1.5\u2009\u00b0C (Supplementary Figs. 15 and 16)18. We note that 2% minimum \ndecarbonization rate assumption is not binding for any region since \nthe implied 2015\u20132030 decarbonization rate in the NDCs for all regions \nis >2% (Supplementary Table 6).\nNotably, there is some interaction and overlap among the three \nstrategies explored in this study that countries might use to ratchet \nambition. With higher 2030 ambition, the post-2030 minimum \n\nNature Climate Change\nArticle\nhttps://doi.org/10.1038/s41558-022-01508-0\ndecarbonization rate assumption might no longer be binding in some \ncases. For example, in the case of India, the 2015\u20132030 decarbonization \nrates in the pathways with the NDC, NDC+ and NDC++ emission levels in \n2030 are, respectively, 2.1%, 4.4% and 4.4% (see Supplementary Table \n6 for 2015\u20132030 decarbonization rates under the NDC, NDC+ and \nNDC++ emission levels in 2030). Hence, the 2% post-2030 minimum \ndecarbonization rate assumption would be binding only in the NDC \ncases. In addition, advancing the timing of the net-zero pledges (for \ncountries with net-zero pledges) would result in higher post-2030 \ndecarbonization rates. However, it is important to note that our mini-\nmum decarbonization rate assumptions (2%, 5% and 8%) do not affect \nthe emission pathways of countries with net-zero pledges since these \ncountries are always assumed to achieve their pledges\u2014in the specified \ntarget years, 5\u2009years in advance or 10\u2009years in advance.\nData availability\nCountry ratings from the CAT are publicly available at https://climate-\nactiontracker.org/countries/. The latest (v.2020) Human Develop-\nment Index is publicly available at https://hdr.undp.org/data-center/\nhuman-development-index#/indicies/HDI. The datasets generated \nduring and analysed in the current study are available from a public \nrepository (https://doi.org/10.5281/zenodo.7069063). Source data \nare provided with this paper.\nCode availability\nGCAM is an open-source community model available at https://github.\ncom/JGCRI/gcam-core/releases. The version of GCAM and additional \ninput files associated with this study are available at https://doi.\norg/10.5281/zenodo.7069066.\nReferences\n55.\t Ou, Y. Source code and data for Ou et al. 2021 Updates to Paris \nclimate pledges improve chances of limiting global warming to \nwell below 2\u2009\u00b0C. Zenodo https://doi.org/10.5281/zenodo.5821125 \n(2021).\n56.\t Iyer, G. et al. Implications of sustainable development \nconsiderations for comparability across NDCs. Nat. Clim. \nChange 8, 124\u2013129 (2018).\n57.\t Fawcett, A. A. et al. Can Paris pledges avert severe climate \nchange? Science 350, 1168\u20131169 (2015).\n58.\t McJeon, H. et al. Limited impact on decadal-scale climate change \nfrom increased use of natural gas. Nature 514, 482\u2013485 (2014).\n59.\t Wise, M. et al. Implications of limiting CO2 concentrations for land \nuse and energy. Science 324, 1183\u20131186 (2009).\n60.\t Clarke, J. F. & Edmonds, J. Modelling energy technologies in a \ncompetitive market. Energy Econ. 15, 123\u2013129 (1993).\n61.\t McFadden, D. Econometric models for probabilistic choice \namong products. J. Bus. 53, S13\u2013S29 (1980).\n62.\t Boden, T. A., Marland, G. & Andres, R. J. Global, Regional, and \nNational Fossil-Fuel CO2 Emissions (Carbon Dioxide Information \nAnalysis Center, 2009).\n63.\t World Energy Balances 2019 (International Energy Agency, 2019).\n64.\t Wise, M., Calvin, K., Kyle, P., Luckow, P. & Edmonds, J. A. E. \nEconomic and physical modeling of land use in Gcam 3.0 and \nan application to agricultural productivity, land, and terrestrial \ncarbon. Clim. Change Econ. 05, 1450003 (2014).\n65.\t Global Non-CO2 Greenhouse Gas Emission Projections & Mitigation \n2015\u20132050 (United States Environmental Protection Agency, 2019).\n66.\t O\u2019Rourke, P. R. et al. CEDS v_2021_04_21 Release Emission Data \n(v_2021_02_05). Zenodo https://doi.org/10.5281/zenodo.4741285 \n(2021).\n67.\t Real GDP growth (International Monetary Fund, 2021); https://\nwww.imf.org/external/datamapper/NGDP_RPCH@WEO/OEMDC/\nADVEC/WEOWORLD\n68.\t Vimmerstedt et al. 2019 Annual Technology Baseline ATB Cost and \nPerformance Data for Electricity Generation Technologies (National \nRenewable Energy Lab, 2019).\n69.\t Jadun et al. Electrification Futures Study: End-use Electric \nTechnology Cost and Performance Projections through 2050 \n(National Renewable Energy Lab, 2017).\n70.\t Calvin, K. et al. Trade-offs of different land and bioenergy policies \non the path to achieving climate targets. Clim. Change 123, \n691\u2013704 (2014).\n71.\t Hartin, C. A., Bond-Lamberty, B., Patel, P. & Mundra, A. Ocean \nacidification over the next three centuries using a simple global \nclimate carbon-cycle model: projections and sensitivities. \nBiogeosciences 13, 4329\u20134342 (2016).\n72.\t XIE Zhenhua Explains 1+N Policy Framework for the Timeline \nand Roadmap of China\u2019s Carbon Peak and Neutrality Goals \n(National Center for Climate Change Strategy and International \nCooperation, accessed 4 March 2022); http://www.ncsc.org.cn/\nxwdt/gnxw/202107/t20210727_851433.shtml\nAcknowledgements\nThe research described in this paper was conducted with support \nfrom the US EPA IAA DW-089-92460001 (G.I., Y.O., J.E., J.F., S.W. and \nH.M.). The views and opinions expressed in this paper are those of \nthe authors alone and do not necessarily state or reflect those of \nthe Environmental Protection Agency or the US Government and no \nofficial endorsement should be inferred.\nAuthor contributions\nG.I. and Y.O. contributed equally to this study. G.I., Y.O., J.E.,  \nA.A.F., N.H. and H.M. designed the research. G.I. wrote the  \nfirst draft of the paper. Y.O. conducted the simulations. G.I., Y.O., \nJ.E., J.F., S.W. and H.M. contributed to the modelling tools. J.M. \ncontributed to the writing of the paper. All authors contributed  \nto writing the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains \nsupplementary material available at  \nhttps://doi.org/10.1038/s41558-022-01508-0.\nCorrespondence and requests for materials should be addressed to \nHaewon McJeon.\nPeer review information Nature Climate Change thanks Matthias \nWeitzel, Ioannis Dafnomilis and Kate Dooley for their contribution to \nthe peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\n\n    Scientific Research Findings:", "answer": "We find that ratcheting near-term ambition to 2030 will be crucial to limiting peak temperature changes this century. If ratcheting is delayed, it would result in higher temperature overshooting \u2014 that is, an exceedance of global mean temperature change above the intended threshold before returning to below the intended level \u2014 over many decades, with the potential for adverse and irreversible consequences for human and natural systems. Our results also suggest that ratcheting near-term ambition could facilitate faster transitions to net-zero emissions systems \u2014 especially in major economies \u2014 resulting in faster reductions in emissions from all sectors of the economy for both CO2 and non-CO2 emissions. Although these transitions can be accomplished with limited availability of nascent technologies such as CO2 removal, further research is required to better understand the role of such technologies in high-ambition emissions pathways.", "id": 52}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Climate Change | Volume 12 | November 2022 | 995\u2013998\n995\nnature climate change\nhttps://doi.org/10.1038/s41558-022-01501-7\nArticle\nHow the USA can benefit from risk-based \npremiums combined with flood protection\nLars T. de Ruig1,2\u2009\n, Toon Haer\u2009\n\u200a\u20091, Hans de Moel\u2009\n\u200a\u20091, Samuel D. Brody3, \nW. J. Wouter Botzen1,4,5, Jeffrey Czajkowski6 and Jeroen C. J. H. Aerts1,7\u2009\nFlood risk management in the USA is largely embedded in the National Flood \nInsurance Program (NFIP). Climate change and increasing exposure in flood \nplains pose a challenge to flood risk managers and make it vital to reduce \nrisk in the future. The proposed reforms are steering the NFIP to risk-based \npremiums, but it is uncertain if the reforms will result in unaffordability \nand incentivize risk-reduction investments or how the NFIP is affected by \nlarge-scale adaptation efforts. Using an agent-based model approach for \ncurrent and future scenarios, we demonstrate that risk-based premiums will \nyield a positive societal benefit (US$10\u2009billion) because they will incentivize \nhousehold risk-reduction investments. Moreover, our results show that \nproactive investment in large-scale adaptation measures complements a \ntransition to risk-based premiums to yield a higher overall societal benefit \n(US$26\u2009billion). We suggest that transitioning the NFIP to risk-based \npremiums can only be secured by additional investments in large-scale flood \nprotection infrastructure.\nFlooding is a devastating natural hazard, causing an average \n>US$100\u2009billion of damage every year1. Recent events of coastal and \nriver flooding in Europe and Asia have shown the huge impact of such \nevents on communities and policy-makers are struggling with how \nto anticipate future increase in flood risk due to climate change and \npopulation growth2. Without adaptation investments under the repre-\nsentative concentration pathway (RCP)\u20094.5 and shared socio-economic \npathway (SSP)\u20092 scenarios, fluvial flood risk for the USA is expected to \nincrease from about US$27\u2009billion to US$66\u2009billion per year (refs. 3,4), \nwhile coastal flood risk cost is expected to increase from US$1.8\u2009billion \nto US$189\u2009billion5. In the USA, the National Flood Insurance Program \n(NFIP) is the main program for managing flood risk. The NFIP pro-\nvides almost 5\u2009million policies to homeowners and businesses in the \nUSA, covering US$1.2\u2009trillion in assets and making it the largest flood \ninsurance market worldwide6. The program requires households in a \nparticipating community with a bank-backed mortgage living within a \n100-year flood zone to purchase mandatory flood insurance coverage. \nIt also requires that new developments in these zones meet certain \nbuilding codes. These low-lying flood zones are mapped by the Federal \nEmergency Management Agency (FEMA). However, the program has a \nUS$20.5\u2009billion debt due to, amongst other things, the setting of pre-\nmiums on the basis of national averages that do not reflect local risk, \nnew development in flood-impacted areas, and the lack of incentives \nfor homeowners to implement flood adaptation measures other than \nbuilding elevation7.\nSeveral reforms have been introduced to solve some of the issues. \nThese include the new Risk Rating 2.0 program, which more accurately \nsets premiums that reflect yearly risk for individual buildings8. While \nthe reforms are expected to increase mean insurance uptake and solve \nfinancial burden on the program, they are also likely to put pressure on \naffordability for low-income households living in high-risk flood zones. \nMoreover, there is uncertainty how the NFIP will perform under future \nclimate change conditions. However, existing studies on NFIP reforms \nonly focus on a specific region or on individual elements of the program \nReceived: 12 November 2021\nAccepted: 15 September 2022\nPublished online: 24 October 2022\n Check for updates\n1Institute for Environmental Studies (IVM), VU University Amsterdam, Amsterdam, the Netherlands. 2Royal HaskoningDHV, Amersfoort, the Netherlands. \n3Institute for a Disaster Resilient Texas (IDRT), Texas A&M University Galveston Campus, Galveston, TX, USA. 4Risk Management and Decision Processes \nCenter, The Wharton School, University of Pennsylvania, Philadelphia, PA, USA. 5Utrecht University School of Economics, Utrecht University, Utrecht, the \nNetherlands. 6Center for Insurance Policy and Research, National Association of Insurance Commissioners (NAIC), Kansas City, MO, USA. 7Deltares, Delft, \nthe Netherlands. \n\u2009e-mail: Lars.de.Ruig@vu.nl; Jeroen.Aerts@vu.nl\n\nNature Climate Change | Volume 12 | November 2022 | 995\u2013998\n996\nArticle\nhttps://doi.org/10.1038/s41558-022-01501-7\nhouseholds, governments and insurance markets is key in gaining a \ncomprehensive understanding of the NFIP and the effects of govern-\nment policies. Recent literature demonstrates the applicability of \nagent-based models (ABMs) for topics related to flood risk and the \neffects of individual decision-making, such as evacuation14, housing \nmarkets15, climate change migration16, community mitigation17,18 and \ninsurance markets19.\nThis study is based on a new flood risk and agent-based modelling \nframework (Methods; Supplementary Information), which simulates \ndynamic decisions by homeowners regarding whether to implement \nDRR (for example, elevation or flood-proofing of buildings) and \npurchase flood insurance. We account for heterogeneous consumer \nbehaviour and individual bounded rationality in risk perceptions20,21. \nFurthermore, governments can decide either proactively (using cost\u2013\nbenefit analysis based on yearly risk projections) or reactively (after a \nflood event) to invest in regional flood protection infrastructure. In \nturn, governmental proactive or reactive decision-making will influ-\nence the behaviour of homeowners. All the agents are provided with \nyearly risk projections from an underlying flood risk model, which \nis driven by climate change scenarios (RCP\u20094.5) and socio-economic \nprojections (SSP\u20092) until 2050 (Supplementary Information). We apply \nthis method to the conterminous US on a grid resolution of 30\u2009arcsec.\nTransition towards risk-based premiums\nThe effects of a transition from the current NFIP to risk-based premiums \nfor 2050 is illustrated in Fig. 1, in line with the patterns found for the \nshort-term effects in 2030 (Supplementary Fig. 2). Such transition leads \nto higher geographical heterogeneity of premium levels on a local scale. \nOn a national scale, the model predicts that premiums will become \nrelatively lower in coastal areas and higher in fluvial areas (although \nall premiums increase over time due to the effects of climate change \nand are limited to assessing the current conditions. Recent debates on \nthe restoration of aging infrastructure and studies on flood manage-\nment suggest that complementary government-based investments \nin large-scale flood protection infrastructure (for example, dikes) are \nrequired to anticipate future climate risks and the increasing exposure \nof assets in flood-prone areas2,3,9,10. There is a lack of a comprehensive \nUS-scale analysis of the NFIP that addresses how policy-holders will be \nimpacted in the future by reforms, investments in governmental flood \ninfrastructure and the ability of the NFIP to cope with climate change.\nWhile the new reforms have been assessed in various studies8, \nthis study aims to complement the upcoming reforms by showcasing \nhousehold behaviour patterns under different market structures and \nwith different governmental adaptation investment efforts under \nclimate change scenarios for fluvial and coastal flood risk. This is done \nby studying four indicators over time (2020\u20132050): insurance pen-\netration rates, unaffordability of premiums and investment costs of \nrisk-reduction measures, incentivization rate of building-scale disaster \nrisk reduction (DRR) and the program\u2019s debt. The sensitivity of these \nindicators under future changes provides an indication of whether the \nprogramme will be financially healthy in the future. It also indicates \nthe challenges that policy-holders may face. It should be noted that \nthis paper does not aim to replicate or fully assess the Risk Rating 2.0 \nprogram as the NFIP Risk Rating 2.0 Delay Act of 2021 has delayed \nroll-out to 30 September 2022. We also show how households are incen-\ntivized to invest in DRR at the local level and highlight the importance \nof proactive governmental large-scale flood protection measures to \ncomplement the performance of the NFIP.\nConventional flood risk assessment methods are often \nill-equipped to address flood management policy changes as they \naddress neither interactions between key stakeholders nor house-\nhold decision-making11\u201313. The spatial\u2013temporal interplay between \nPercentage change (%)\na\nb\nd\ne\nc\nPercentage change (%)\nPercentage change (%)\n\u2013100 \u201350\n\u201350 \u201325\n0\n25\n50\n0\n50\n100\n\u2013100 \u201350\n0\n50\n100\n25\n20\n15\n10\nPercentage (%)\n5\n0\n100\n80\n60\n40\n20\n0\n\u201320\nPercentage (%)\nCoastal\nFluvial\n2030\n2050\nScenario\nCurrent NFIP\nRCP 4.5, SSP 2\nRisk-based NFIP\nFig. 1 | Effects of NFIP reform 2050 (RCP\u20094.5\u2009+\u2009SSP\u20092). a\u2013c, Effects of \nNFIP premiums to risk-based premiums in terms of market penetration (a), \nunaffordability (b) and risk (c) are visualized as the mean percentage change on \na county level for 2050. d, Mean changes in penetration rates between coastal \nand fluvial risk areas. e, Mean insurer debt showing that debt might still increase \ndespite risk-based premiums (although at a lower rate than under current \nconditions). For d and e, error bars indicate the standard deviation (n\u2009=\u2009250 per \nscenario).\n\nNature Climate Change | Volume 12 | November 2022 | 995\u2013998\n997\nArticle\nhttps://doi.org/10.1038/s41558-022-01501-7\ncompared to present day). These effects result in a decrease of insur-\nance penetration rates in fluvial regions from 24.7% to 13.2% (partly \ndue to an increase in unaffordability). This implies that (on average) \nthe NFIP is currently underpricing these regions. Conversely, coastal \nregions display an increase from 10.2% to 17.1% in penetration rates \ndue to more attractive premiums under risk-based insurance pricing \nfor a share of households due to spatial variation in premiums. Still, a \nsubset of policy-holders will experience a steep increase of their pre-\nmium in these regions. Despite the lower penetration rates in fluvial \nregions, introducing risk-based premiums and offering premium dis-\ncounts based on the actual reduction of risk achieved by flood-proofing \nbuildings is expected to decrease the total average residential flood \nrisk (coastal and fluvial) across the USA by about US$1\u2009billion (\u22127.3%) \nby 2050. These results highlight the effectiveness of offering a pre-\nmium discount to implement a variety of DRR types (wet and dry \nflood-proofing), including through retrofitting.\nOur model demonstrates that unaffordability (Methods) is \nexpected to decrease from 4.5\u2009million households (23.8%) to 1\u2009million \nhouseholds (5.6%) for the 18.8\u2009million households at risk of floods \nnationwide in 2050 following a transition to risk-based premiums. \nHowever, the magnitude of the remaining unaffordability increases \nsubstantially to an average of US$2,000 per year per household. These \nfindings indicate that although risk-based premiums are lower than  \ncurrent NFIP premiums for many households, moving towards \nrisk-based premiums implies a sharp increase in premiums and  \nunaffordability for a subgroup of households living in high-risk \nareas. Unaffordability issues can potentially be overcome by offering  \ninsurance vouchers and providing inexpensive accessible loans for \nfinancing DRR measures by low-income households currently living \nin high flood risk zones7. Alternatively, further incentivizing DRR by \nhomeowners would make more people eligible for premium discounts. \nRelocation or managed retreat might also become necessary2. As seen \nin Fig. 1e, continuation of the program without anticipating climate \nchange or socio-economic development will further increase the  \ndebt of the NFIP by ~60% (from US$20.5\u2009billion to US$32.8\u2009billion). \nIntroducing risk-based premiums will significantly limit the rise in \nfuture debt but will not solve the problem entirely (debt will increase \nby 28% instead of 60%). An additional markup of premiums might be \nrequired to pay off current debt and make the program financially \nsound in the future.\nInsurance schemes must be interlinked with flood \nadaptation\nOur results show that additional large-scale flood adaptation investment \ncomplements a transition to risk-based premiums and household-level \nDRR measures22. Table 1 shows four policy scenarios: a baseline scenario \n(Sc1, current NFIP, reactive government) and three scenarios (Sc2, 3  \nand 4) with current and risk-based NFIP schemes and proactive or reac-\ntive government policies combinations. These are evaluated relative to \nthe baseline Sc1 (see Methods for scenario descriptions and Supplemen-\ntary Information for additional results). Table 1 displays the evaluation \nof these four scenarios for 2020\u20132050. It shows the present values of \ncategories that address the total societal costs of flood risk and flood \nmanagement. Total societal costs (F) are the expected cost of uncovered \nflood damage to properties and public assets (covered residential risk \nB subtracted from total flood risk A), insurance premium payments  \n(C; costs of covered risk) and the costs of flood risk-reduction measures \nincurred by governments (D) and households (E). The results show that \ntransitioning to risk-based premiums has a positive net present societal \nbenefit of about US$10\u2009billion for the period from 2020 to 2050, even \nwhen governments remain reactive towards investments in adaptation \ninfrastructure. If the government acts proactively alongside risk-based \npremiums, the net present societal benefit increases to US$26\u2009billion \nfor the same period. Large-scale flood protection measures also have \nrisk-reduction benefits over a longer lifespan than the 30-year period \nevaluated (2020\u20132050); hence, extending the analysis to the far future \n(for example, 2100) will favour proactive government policies even \nmore. Accordingly, the government can reduce a large share of (future) \nflood risk by increasing flood protection through the installation of \nlevees, the adoption of nature-based solutions or the implementation \nof other measures.\nThe community rating system (CRS) is a voluntary programme that \nincentivizes communities to actively engage in floodplain management \nactivities that exceed the minimum programme requirement. The CRS \ncould be used to further promote risk awareness and comprehensive \nfloodplain management by communities and local governments in \nexchange for NFIP premium discounts to policy-holders in the com-\nmunity22. However, the full potential of the current implementation \nof the CRS is not yet used because only a small share of participating \ncommunities are actively reducing risk and the system does not con-\nsistently reward measures that consider future climate change risks23. \nAs demonstrated by ref. 23, using the CRS to focus on regional flood \nprotection goes hand-in-hand with the aforementioned NFIP reforms, \nsince reducing flood risk through flood protection investments helps \nto keep insurance premiums affordable.\nRisk-based premiums and flood adaptation \ninfrastructure\nOur analysis suggests that either moving towards risk-based premi-\nums or proactively investing in large-scale flood protection yields \nTable 1 | Total societal costs related to flood risk\nA. Total flood risk  \n(of which total  \nresidential risk)\nB. Covered \nresidential risk\nC. Insurance \npremium \nexpenses\nD. Government \nflood protection \ninvestment costs\nE. Household \nflood-proofing \ninvestment costs\nF. Total \nsocietal costs \nA\u2009\u2212\u2009B\u2009+\u2009C\u2009+\u2009D\u2009+\u2009E\nPolicy scenarios\n(US$\u2009billion)\n(US$\u2009billion)\n(US$\u2009billion)\n(US$\u2009billion)\n(US$\u2009billion)\n(US$\u2009billion)\nSc1: Baseline, current NFIP and \nreactive government\nUS$497 [34] (US$198) [15]\nUS$44 [3]\nUS$60 [2]\nUS$124 [57]\nUS$5.2 [4]\nUS$643 [74]\nSc2: Current NFIP\u2009+\u2009proactive \ngovernment (relative to baseline)\n\u2212US$88 (\u2212US$31)\n\u2212US$3.2\n\u2212US$2.5\n+US$75\n\u2212US$1.2\n\u2212US$15\nSc3: Risk-based NFIP\u2009+\u2009reactive \ngovernment (relative to baseline)\n\u2212US$8.9 (\u2212US$9.0)\n\u2212US$19\n\u2212US$23\n+US$0.0\n+US$3.2\n\u2212US$10\nSc4: Risk-based NFIP\u2009+\u2009proactive \ngovernment (relative to baseline)\n\u2212US$94 (\u2212US$37)\n\u2212US$23\n\u2212US$28\n+US$74\n+US$1.0\n\u2212US$26\nResults are based on RCP\u20094.5 and SSP\u20092 for the period 2020\u20132050. Cost categories are: F, cumulative cost from 2020 to 2050, which is the sum of uncovered total risk (covered residential risk \n(B) subtracted from total risk (A)); C, insurance premium expenses; D, governmental flood protection investment costs; and E, household flood-proofing investment costs. Values are expressed \nas present values (US$\u2009billion, 2020 at 4% discount rate) relative to the baseline scenario Sc1. Negative values denote a societal improvement and positive values indicate a cost for society \ncompared with the baseline (note: numbers are rounded). Standard deviations are indicated in square brackets (n\u2009=\u2009250 per scenario).\n\nNature Climate Change | Volume 12 | November 2022 | 995\u2013998\n998\nArticle\nhttps://doi.org/10.1038/s41558-022-01501-7\nsignificantly higher societal benefits. The highest societal benefits \nare achieved when adopting both strategies. Despite high upfront \ninvestment costs for large-scale adaptation of ~US$75\u2009billion (compared \nto the baseline scenario), proactive investments are complementary \nto risk-based premiums. Together, they yield high average societal  \nbenefits ($26\u2009billion, net present value for 2020\u20132050, Table 1). While \nthe investment costs for large-scale adaptation seem high, the total \nsocietal benefits of US$26\u2009billion with adaptation investments sig-\nnificantly surpasses the expected societal benefit of US$10\u2009billion \nwhen only transitioning to risk-based premiums (Table 1). Further-\nmore, investing in flood protection infrastructure will reduce some \nof the equity issues (unaffordability) that arise when solely moving to \nrisk-based premiums (columns C and E in Table 1). It will also reduce the \nrisk to other governmental assets (for example, energy infrastructure \nand low-lying port areas). The remaining unaffordability issues for the \nsubset of households with a major increase in insurance premiums \ncould be addressed by offering insurance vouchers and providing \nlow-interest accessible loans to further incentivize the implementa-\ntion of DRR7 or relocation2.\nThe results of this study offer a timely contribution to both the \nRisk Rating 2.0 transition and the infrastructure bill. We show that \nthere is significant synergy to be achieved by moving towards both a \nrisk-based premium and a proactive strategy on large-scale adaptation. \nOur findings demonstrate that investing in large-scale adaptation \ndoes not reduce the effectiveness of risk-based premiums; in fact,  \nit contributes to the reduction of unaffordability.\nOnline content\nAny methods, additional references, Nature Research reporting sum-\nmaries, source data, extended data, supplementary information, \nacknowledgements, peer review information; details of author con-\ntributions and competing interests; and statements of data and code \navailability are available at https://doi.org/10.1038/s41558-022-01501-7.\nReferences\n1.\t\nMerz, B. et al. Causes, impacts and patterns of disastrous river \nfloods. Nat. Rev. Earth Environ. 2, 592\u2013609 (2021).\n2.\t\nMills, E. Insurance in a climate of change. Science 309,  \n1040\u20131044 (2005).\n3.\t\nWard, P. J. et al. A global framework for future costs and benefits \nof river-flood protection in urban areas. Nat. Clim. Change 7, \n642\u2013646 (2017).\n4.\t\nWinsemius, H. C. et al. Global drivers of future river flood risk. Nat. \nClim. Change 6, 381\u2013385 (2016).\n5.\t\nTiggeloven, T. et al. Global-scale benefit\u2013cost analysis of  \ncoastal flood adaptation to different flood risk drivers using \nstructural measures. Nat. Hazards Earth Syst. Sci. 20, 1025\u20131044 \n(2020).\n6.\t\nHorn, D. P. & Webel, B. W. Introduction to the National Food \nInsurance Program (NFIP, 2019).\n7.\t\nMichel-Kerjan, E. O. & Kunreuther, H. Redesigning flood \ninsurance. Science 333, 408\u2013409 (2011).\n8.\t\nRisk Rating 2.0 Overview (FEMA, 2019).\n9.\t\nMechler, R. & Schinko, T. Identifying the policy space for climate \nloss and damage. Science 354, 290\u2013292 (2016).\n10.\t Di Baldassarre, G., Kooy, M., Kemerink, J. S. & Brandimarte, L. \nTowards understanding the dynamic behaviour of floodplains \nas human-water systems. Hydrol. Earth Syst. Sci. 17, 3235\u20133244 \n(2013).\n11.\t\nDi Baldassarre, G. et al. Debates\u2014perspectives on \nsocio-hydrology: capturing feedbacks between physical and \nsocial processes. Water Resour. Res. 51, 4770\u20134781 (2015).\n12.\t Di Baldassarre, G. et al. Socio-hydrology: conceptualising human- \nflood interactions. Hydrol. Earth Syst. Sci. 17, 3295\u20133303 (2013).\n13.\t Grames, J., Prskawetz, A., Grass, D., Viglione, A. & Bl\u00f6schl, G. \nModeling the interaction between flooding events and economic \ngrowth. Ecol. Econ. 129, 193\u2013209 (2016).\n14.\t Dawson, R. J., Peppe, R. & Wang, M. An agent-based model for \nrisk-based flood incident management. Nat. Hazards 59, 167\u2013189 \n(2011).\n15.\t Filatova, T. Empirical agent-based land market: integrating \nadaptive economic behavior in urban land-use models. Comput. \nEnviron. Urban Syst. 54, 397\u2013413 (2015).\n16.\t Hassani-Mahmooei, B. & Parris, B. W. Climate change and internal \nmigration patterns in Bangladesh: an agent-based model. \nEnviron. Dev. Econ. 17, 763\u2013780 (2012).\n17.\t Tonn, G. L. & Guikema, S. D. An agent-based model of evolving \ncommunity flood risk. Risk Anal. 38, 1258\u20131278 (2018).\n18.\t Tonn, G. L., Guikema, S. & Zaitchik, B. Simulating behavioral \ninfluences on community flood risk under future climate \nscenarios. Risk Anal. 40, 884\u2013898 (2019).\n19.\t Han, Y. & Peng, Z. R. The integration of local government, \nresidents, and insurance in coastal adaptation: an agent-based \nmodeling approach. Comput. Environ. Urban Syst. 76, 69\u201379 \n(2019).\n20.\t Haer, T., Botzen, W. J. W. & Aerts, J. C. J. H. Advancing disaster \npolicies by integrating dynamic adaptive behaviour in risk \nassessments using an agent-based modelling approach. Environ. \nRes. Lett. 14, 044022 (2019).\n21.\t Aerts, J. C. J. H. et al. Integrating human behaviour dynamics into  \nflood disaster risk assessment. Nat. Clim. Change 8, 193\u2013199 \n(2018).\n22.\t Sadiq, A. A., Tyler, J. & Noonan, D. S. A review of community flood \nrisk management studies in the United States. Int. J. Disaster Risk \nReduct. 41, 101327 (2019).\n23.\t Blessing, R., Brody, S. D. & Highfield, W. E. Valuing floodplain \nprotection and avoidance in a coastal watershed. Disasters 43, \n906\u2013925 (2019).\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor holds exclusive rights to this \narticle under a publishing agreement with the author(s) or other \nrightsholder(s); author self-archiving of the accepted manuscript \nversion of this article is solely governed by the terms of such \npublishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited \n2022\n\nNature Climate Change \nArticle\nhttps://doi.org/10.1038/s41558-022-01501-7\nMethods\nModelling framework\nThe core of the modelling framework was developed for the contermi-\nnous US, including the entire USA coastline and all main river basins. It \nsimulated flood risk at a yearly time step with representative household \nadaptation at a resolution of 30\u02ba\u2009\u00d7\u200930\u02ba and government adaptation at \nthe county level. Homeowners could invest in DRR measures (eleva-\ntion or flood-proofing of buildings) or take out/cancel insurance and \ngovernments could invest in elevating dikes. Both these adaptations \nand the proposed NFIP reform policies were captured in four policy sce-\nnarios. The model was run 50 times for each of these policy scenarios \nwhile also assuming different climate scenarios and socio-economic \nscenarios. The framework builds upon earlier versions of DYNAMO20,24. \nNew components are the flood insurance market module, the address-\ning of coastal flood risk in addition to fluvial flood risk and the nation-\nwide application of the model. For the method description of the  \ncore model, please refer to ref. 20. See also Extended Data Fig. 1 and \nSupplementary Information.\nFlood risk model\nThe GLOFRIS flood risk model follows a commonly applied hazard\u2013\nexposure\u2013vulnerability model25,26. Coastal and fluvial inundation maps \nwere combined with land use to simulate the (future) exposure of assets \nand their values in flood zones. Depth\u2013damage curves were used to \ncombine hazard and exposure data to simulate flood risk (expected \nannual damage, EAD, in US$ per yr) for each individual grid cell and \ncounty. Floods can stochastically occur every year in each county on \nthe basis of their return period. See also Extended Data Fig. 1a and \nSupplementary Information.\nScenarios\nFluvial and coastal inundation maps are available for the current and \nfuture climate following the RCP\u20094.5 and RCP\u20098.5. The SSP scenarios27\u201332 \nwere used to represent the initial population numbers and to project \npopulation growth, income and economic growth for 2050. We applied \nthe SSP\u20092 and SSP\u20095 scenarios as they matched well with RCP\u20094.5 and \nRCP\u20098.5, respectively. SSP\u20092 was a middle-of-the-road scenario, while \nSSP\u20095 was an energy-intensive and resource-intensive scenario. The \nformer was used throughout the paper and the results of the SSP\u20095 \nscenario can be found in the Supplementary Information. See also \nExtended Data Fig. 1c.\nInput data\nFluvial flood hazard. The GLOFRIS fluvial inundation maps are based \non existing research3,33. In brief, daily time series of flood volumes \nwere constructed using hydrological and hydrodynamic model-\nling at a 0.5\u00b0\u2009\u00d7\u20090.5\u00b0 resolution. The GLOFRIS model was forced with \nEU-WATCH data for the period 1960\u20131999, representing historic con-\nditions. For future conditions, the GLOFRIS model was forced with \nfive global climate models (GCMs): HadGEM2-ES, IPSL-CM5A-LR, \nMIROC-ESM-CHEM, GFDL-ESM2M and NorESM1-M. Yearly maximum \nhydrological time series were extracted from the daily gridded flood \nvolumes and a Gumbel distribution was fitted accordingly. The result-\ning Gumbel parameters were used to estimate (future) return peri-\nods for each grid cell: 5, 10, 25, 50, 100, 250, 500 and 1,000\u2009yr. Finally,  \nGLOFRIS distributed flood volumes to a digital elevation model to \ncreate high-resolution (30\u02ba\u2009\u00d7\u200930\u02ba) inundation maps and their return \nperiods3,33. See also Extended Data Fig. 1b.\nCoastal flood hazard. For coastal inundation, as described in detail \nin ref. 5, the extreme sea levels are taken from the Global Tide and  \nSurge Reanalysis (GTSR) dataset34. GTSR is a global dataset of daily \nsea levels (tide and storm surge) for 1979\u20132014, which is based on \nthe Global Tide and Surge Model (GTSM). Within GTSM, tides are \nsimulated separately using the Finite Element Solution (FES2012) \nhydrodynamic model35. Surges are simulated using metrological data \nfrom the ERA-Interim global atmospheric reanalysis36. The GTSR data-\nset is enhanced with historical tropical cyclone tracks over the period \nof 1979\u20132004, using the International Best Track Archive for Climate \nStewardship archive, as GTSR is known for under-representing tropical  \ncyclones. The extremes are subsequently calculated using a Gumbel  \nfit of annual maxima using the maximum-likelihood method and  \nvalidated in refs. 34,37.\nTo calculate coastal inundation, overland inundation from \nnear-shore tide and surge levels were computed, after which the  \nnearest GTSR location are projected at the coastline. A resistance  \nfactor is used to simulate the reduction of flooding land inwards, \nas tides and storm surges have a limited time span and therefore  \ntheir flood peak and associated volume can only penetrate inland  \nto a certain degree5. Coastal flood maps were made on a resolution  \nof 30\u02ba\u2009\u00d7\u200930\u02ba for the same return periods as the fluvial flood maps:  \n5, 10, 25, 50, 100, 250, 500 and 1,000\u2009yr, respectively33. For future  \nconditions, mean sea level rise conditions were obtained from the \nRISES-AM project38 and simulated as a range of probabilistic outcomes. \nFor this paper, the 50th percentile was used. Future subsidence rates \nfrom the SUB_CR model and are included in the inundation model by \nadding the subsidence estimates to the MERIT digital terrain model39. \nIn the areas where fluvial and coastal flood cells overlapped, we applied \na simple method proposed by FEMA40 and selected the highest inunda-\ntion value.\nFor the continental scale and purpose of the study, the GLOFRIS \nmodel is sufficient. However, future research should focus on coupling \nABMs with higher resolution flood hazard models (for example, ref. 41) \nto increase the accuracy of risk simulations and to allow for improved \nanalyses on the local scale.\nExposure. The GlobalLand30 (ref. 42) database was used to estimate \nthe exposure of urban assets. Urban grid cells within the GlobalLand30 \ndataset were set to 75% residential, 15% commercial and 10% industrial \n(with an assumed building density of 20% for residential and 30% for \ncommercial or industrial)5,43. Future changes in residential surface \narea per cell were derived from the correlation between population \ngrowth and residential building surface growth20. The gross domestic \nproduct (GDP) growth from the SSPs was used to increase the value of \nproperties over time.\nVulnerability. The vulnerability was represented by depth\u2013dam-\nage curves (the relationship between inundation and the share of  \ndamage per land use type) and maximum damage values (the total \namount of damage per land use type). The HAZUS Multi-Hazard \nmodel44 of FEMA was used for the curves and the maximum damage  \nvalues were taken from ref. 43 following existing methods3,5. The \ndepth\u2013damage curves were altered to simulate the effect of household  \nDRR measures: dry flood-proofing (preventing flood water from  \nentering a building) or elevation of new buildings45. Dry flood-proofing \nhas a lifespan of 75\u2009yr (ref. 46), costs US$100\u2009m\u22122 (Supplementary  \nTable 5) and was implemented for water levels up to 1\u2009m (FEMA47) \nreducing damage by 85%48,49. However, overtopping due to higher  \nwater levels resulted in full damage. For elevation, the vulnerability  \ncurve was shifted so that damage would only occur above 1\u2009m of \ninundation.\nSocial vulnerability is represented by accounting for affordability \nin the adaptation choices of households. It is recommended for future \nresearch to also include other social vulnerability factors.\nCurrent flood protection. Not all low-lying flood zones have the poten-\ntial to be flooded due to the existing flood protection infrastructure \n(for example, dikes). The initial protection standards of levees in flood \nzones were based on the FLOPROS dataset50. Tiggeloven et al.5 and  \nothers used an approach where protection levels were based on GDP \n\nNature Climate Change \nArticle\nhttps://doi.org/10.1038/s41558-022-01501-7\nbut this resulted in very high protection levels for the USA (while obser-\nvations suggested otherwise)51,52. Therefore, the coastal protection \nstandards were set to 30-year exceedance levels51,52, which can be \nchanged by the governmental agents (equation (4)).\nPolicy scenarios\nA total of four policy scenarios (Sc1\u20134) were simulated. These consisted \nof combinations of the following NFIP insurance and governmental \npolicies:\n\u2022 \nSc1: current NFIP. This scenario simulated the current NFIP \nmarket structure as closely as possible. Observed premiums for \n100-year flood zones and low-risk flood zones were used.\n\u2022 \nSc2: NFIP with risk-based premiums. Risk-based premiums were \ncalculated by the flood risk model plus a loading factor of the \ncurrent NFIP premium setting (estimated at 49.7%).\n\u2022 \nSc3: reactive government. Re-evaluation of protection standards \nof levees only occurred after a flood event in a county. Measures \nwere implemented on the basis of a cost\u2013benefit analysis using \nrisk information from the model.\n\u2022 \nSc4: proactive government. Re-evaluation of protection  \nstandards occurred either every 6\u2009years or after a flood event  \nin the respective county. Measures were implemented on the  \nbasis of a cost\u2013benefit analysis using risk information from the \nmodel.\nEach policy scenario per RCP/SSP scenario was run for 50  \nrepetitions for each RCP/SSP scenario per five GCMs, for a total of  \n250 repetitions each (n\u2009=\u2009250).\nInsurance market (ABM)\nPremiums. Household decisions were based on yearly premiums \n(building and content coverage), which were simulated per grid cell of \n30\u02ba\u2009\u00d7\u200930\u02ba. See also Extended Data Fig. 1a. Start values of the premiums in \nSc1 and Sc2 (current NFIP) were 2014 county averages for 100-year flood \nzones and low-risk flood zones53. State averages for premiums were \nused if county-level information was lacking. These 2014 premiums \nwere adjusted to 2000 values at the start of the simulations. Next, these \npremiums were multiplied by the percentage change in yearly EAD \nsimulated by GLOFRIS. It was assumed that yearly premiums changed \nbefore the yearly government decision on protection standards and \nthe risk increase at that time step.\nPremium discounts. For the scenarios with risk-based premiums \n(Sc3 and Sc4), unloaded premiums were initially calculated by the \nflood risk model. Subsequently, average NFIP loading factors (49.7%) \nwere added54,55. Finally, a premium discount was applied using the CRS \n(Supplementary Information). Different discounts were applied for \nthe 100-year and the low-risk flood zones (based on the ref. 53 dataset). \nAlong with the CRS discounts, households that implemented DRR \n(elevation or flood-proofing) received a percentage premium discount \nthat reflected the risk reduction obtained from implementing the DRR \nmeasure for the risk-based scenarios.\nMandatory insurance. Part of the NFIP is a mandatory purchase \nrequirement for federally funded mortgages in a 100-year flood zone. \nEstimates show that, on average, 55% of the properties within a 100-year \nflood zone in participating communities are bound to the mandatory \npurchase requirement. However, research shows that an average of \nonly 78% of those households comply56. We applied the 55% mandatory \nshare for 100-year flood zones for participating communities and we \nbenchmarked the compliance rate on affordability and the expenditure \ncap (that is, we assumed that if a households could not afford the policy \npremium, then they would not comply with the mandatory purchase \nrequirement, resulting in an average compliance rate of 78%). After \nthe initial model setup, households in non-participating communities \ncan voluntarily adopt insurance over time but no mandatory require-\nment is enforced. It should be noted that this could lead to a higher \ninsurance demand for inland communities that are non-participatory \nin the present.\nExpenditure cap insurance. Following ref. 57 and ref. 58, we applied \nan expenditure cap definition for unaffordability. It was assumed  \nthat households could afford flood insurance if their annual pre-\nmium was within the expenditure cap of their annual income, which  \nwas benchmarked at 7.5%. Income was distributed per county  \nthrough a log-normal distribution based on mean and median \nincome from the US Census Bureau 201059. See also Supplementary \nInformation.\nHomeowner behaviour (ABM). Households that were not bound to \nthe mandatory requirement had a yearly decision to take or cancel \ninsurance. See also Extended Data Fig. 1a. First, (un)affordability was \ntested through the expenditure cap for insurance of 7.5%. Second, \ntwo strategies were compared following a subjective expected utility \n(EU)20,60,61 model:\nStrategy 1: take insurance, accepting the deductible\nStrategy 2: do not take or cancel insurance\nThe strategy that yielded the highest EU was chosen. The subjec-\ntive EU equation is as follows:\nEUs =\nPI\n\u222b\nPi\n\u03b2PiU (Wt \u2212\u03b3Di,t \u00d7 \u03b4s \u2212Cpremium,t \u2212dpremium,t) dP\n(1)\nEquation (1) calculates EUs for each strategy s. Each event i has a \nprobability Pi of occurring with a factor \u03b2 as perceived probabilities \n(see below). The total set of events I is the return periods of each flood \nevent (with return periods of 5, 10, 25, 50, 100, 250, 500, 1,000 and \n10,000\u2009years, respectively) and the probability of no flood event. The \nEUs is subsequently calculated as the approximation of the integral over \nI. Utility is calculated as a function of wealth W, uncovered damage D, \nfactor \u03b3 as perceived damage (see below), premium C and a premium \ndiscount d (if applicable to the scenario). Damage D per event i for \nyear t is calculated using the hazard\u2013exposure\u2013vulnerability model.\nRisk aversion. A general utility function following constant relative \nrisk aversion62\u201364 was assumed. In line with common findings62,63, house-\nholds were assumed to be slightly risk-averse in which case U(x)\u2009=\u2009ln(x).\nDeductibles. For strategy 1, homeowners had to pay a deductible  \n\u03b4 of 10% of the incurred damages, while strategy 2 had full damages  \n(no damage was covered, so \u03b42\u2009=\u20091, C\u2009=\u20090 and d\u2009=\u20090).\nPerception. Individuals act with bounded rationality in their decisions \non buying flood insurance. This was represented by the perceived prob-\nabilities \u03b2 and perceived damages \u03b3. Both factors were benchmarked \non the basis of empirical data by ref. 24 and simulated households over\u00ad\nestimating their risk after a flood event while underestimating their risk \nafter a period of no floods. Mathematically this is shown as:\n\u03b2 = 12.0639 \u00d7 \u03b13.71657\nt\n+ 0.08233\n(2a)\n\u03b3 = 0.442774 \u00d7 \u03b11.1671\nt\n+ 0.802826\n(2b)\nwhere \u03b1t\u2009=\u20091 if a flood occurs and \u03b1t\u2009=\u2009\u03b1t\u22121 /1.6 if no flood occurs. This \nexpression results in an increase during a flood event. In years with no \nstorm events (the grid cell experiences no inundation), \u03b2 and \u03b3 will sub-\nsequently decay to the inverse of the observed values in ~6\u2009years after \nthe storm event, in line with empirical evidence65\u201367. While homeowners \nwere aware of increasing risk over time, it is assumed that they were not \n\nNature Climate Change \nArticle\nhttps://doi.org/10.1038/s41558-022-01501-7\nfully informed on flood risk due to their bounded rationality. Therefore, \nat the start of the simulations, each agent is assigned a different risk \nincrease value picked from either a random-uniform distribution of the \nobjective risk increase (simulated by the risk model) or no increase at \nall. It should be noted that individual risk perception is the main driver \nfor changing behaviour but the decisions are also strongly influenced \nby the effectiveness of the adaptation measure and the income and \nwealth of the household, in line with coping appraisals68. For future \nwork, it is recommended to assess the effects of including other drivers \nsuch as influences by neighbourhood behaviour.\nHomeowner DRR and affordability. Affordability was tested before \nhouseholds could consider investing in individual adaptation meas-\nures (elevation for new properties and dry flood-proofing for exist-\ning properties). See also Supplementary Information. Elevation will \nprevent damages until the implementation height. Dry flood-proofing \nis preventing water from entering the property, which will lead to a \ndecrease in damages up to 85% for implementation height48,49. However, \ninundation higher than the implementation height will cause overtop-\nping and will result in full damages.\nAn expenditure cap for DRR was set to 2.5% (Supplementary \nInformation) to define affordability (similar to the cap for insurance). \nAdjustments were made as these are long-term investments that have \nbenefits over time but have high initial investment costs. Therefore, \nit was assumed that households could fund the investment through a \npersonal loan with 15% interest over 5\u2009years. The annual loan payment \nwas used to test the affordability24.\nIf the DRR options were affordable, each group (existing and new \nunprotected households) had the choice between two strategies:\nStrategy 1: implement disaster risk reducing measures\nStrategy 2: do nothing\nThis choice was determined on the basis of equation (3), which \ncalculates the subjective discounted expected utility (DEU) as follows:\nDEUs =\nPI\n\u222b\nPi\n\u03b2PiU (NPVs) dP =\nPI\n\u222b\nPi\n\u03b2PiU (\nT\n\u2211\nt=1\nWt\u2212\u03b3Di,t,s\n(1+r)t\n\u2212\nL\n\u2211\nt=0\nCannual,s\n(1+r)t ) dP\n=\nPI\n\u222b\nPi\n\u03b2PiU (\nT\n\u2211\nt=1\nWt\u2212\u03b3Di,t,s\n(1+r)t\n\u2212\nL\n\u2211\nt=0\nn\u2217C0,s\n1\u2212(1+n)L\n(1+r)t ) dP\n(3)\nThe DEU model is calculated for strategy s. The variables D, \u03b2, \u03b3, W, P  \nand i and the general utility function U(x) are similar to those in equation \n(1). The net present value (NPVs) is the sum of wealth Wt and the (reduced) \ndamages Di,t,s over the lifespan of either measure T, discounted to the pre-\nsent value using discount rate r. The discount rate is the pure time prefer-\nence of residents and is assumed to be 3%, following ref. 69. The investment \ncosts C0,s are US$100\u2009m\u22122 for dry flood-proofing and US$45\u2009m\u22122 for \nelevation (considering it only applied to new buildings). Following \nup the affordability metric, the investments are spread over 5\u2009years \nthrough a personal loan with 15% interest (Supplementary Table 5),  \nas an annual loan payment Cannual,s. For strategy 2 (without action), the \nNPVs contains full perceived damages and no investment costs.\nIt should be noted that DRR investments and the uptake/cancella-\ntion of flood insurance are not mutually exclusive, nor does one lead to \nanother. Each decision is simulated yearly and made decisions (such as \ninvesting in DRR) will impact future decisions during a model simulation.\nGovernment large-scale adaptation. Governments had the ability \nto adapt by raising protection standards (5-, 10-, 25-, 50-, 100-, 250-, \n500- and 1,000-year) each year per county. The initial fluvial protec-\ntion standards were taken from the FLOPROS database50. A 30-year \nprotection level was assumed for the coastal protection standards. \n(While some have aimed to determine coastal protection standards \non the basis of GDP correlations, this is often not realistic for the USA \nas found by ref. 51 and others52,70 (Supplementary Table 5).) Adaptation \nmeasures simulate increasing dike heights and all necessary additional \nadaptation measures (for example, beach nourishment and revet-\nments), although some areas might be overestimated whereas other \nwill be underestimated due to the scale of the model. During the initial \nmodelling setup, these protection levels were matched with water \nlevels to estimate an initial dike height. Protection standards could be \nincreased by increasing dike heights for rivers and coasts for a county. \nThe height increase was determined by using water levels associated \nwith different return periods to reach the necessary protection level. \nIf dikes were not upgraded over time, protection levels could decrease \ndue to sea level rise and climate change effects.\nThe decision to increase dike heights was based on a cost\u2013benefit \nanalysis approach, whereby the present value of investment and main-\ntenance costs was weighted against the present value of benefits from \nadaptation over time or mathematically:\nNPVPSi =\nN\n\u2211\nn=1\nL\n\u2211\nt=1\nBt,PSi,n\u2212Ct,PSi,n\n(1+r)t\n\u2212C0,PSi,n\n=\nN\n\u2211\nn=1\nL\n\u2211\nt=1\n(EADredt,PSi,n\u2212EADredt,PScurrent,n)\u2212(Ct,PSi,n\u2212Ct,PScurrent,n)\n(1+r)t\n\u2212C0,PSi,n\n(4)\nHere, NPVPSi is the net present value of investing in dike heights \nassociated with a protection standard PSi, calculated as the sum of each \ngrid cell n for a specific county and over the lifespan of a dike L (assumed \nat 100\u2009years)71. The benefits over time Bt are the difference between the \nreduction of EAD (EADred) between the new protection standard and \nthe current protection standard PScurrent. Similarly, the maintenance \ncosts Ct are the difference between the new and current protection \nstandard. Lastly, C0 are the investment costs, the discount rate r \n(assumed at 4%) and time t in years. The maintenance costs are assumed \nat US$0.1\u2009\u00d7\u2009106 per km and investment costs are assumed at US$8\u2009\u00d7\u2009106 \nper length (km) per height (m)48. The protection standard with the high-\nest NPV is chosen. If none of the protection standard has a positive NPV, \nnothing is done. For the reactive and proactive scenarios, we assume \nthat when the government decides on adaptation, it makes decisions \non the basis of perfect information on future developments of risk. As \nthe proactive government takes adaptation decisions frequently and \nthe reactive government very infrequently, we do capture an approxi-\nmate upper- and lower-bound of government decision-making.\nAgent interactions. Household and governmental agents interact \nthrough investing in either DRR measures or large-scale adaptation \nmeasures, respectively, reducing risk over time. Whenever a house-\nhold reduces their flood risk by implementing DRR measures, this will \ndirectly influence their decision on insurance, through changes in the \nperceived probability of floods or the associated damage and a lower \ninsurance premium for the risk-based insurance market scenarios. In \naddition, they reduce a (minor) share of regional flood risk, impacting \nthe decisions by the regional government.\nSimilarly, if governments reduce risk by implementing \nregional-scale adaptation measures, then this will be reflected in a \nlower perceived probability and damage by households depending \non the level of government protection. Via this mechanism, flood \nprotection by the government thus influences household\u2019s decisions \non adaptation investments and insurance uptake or cancellation.\nModelling robustness. We aimed to assess the sensitivity of proposed \nNFIP reforms under future conditions. To maximize the reliability of our \nmodel for this sensitivity analysis, we followed a three-step approach72 \n(Supplementary Information): (1) benchmarking, (2) validation and \n(3) sensitivity analysis.\nBenchmarking. Benchmarking of the GLOFRIS model has been exten-\nsively described by ref. 3,33. With a hit rate of 70%, the global model \nperforms well for the USA; however, it sometimes overestimates or \n\nNature Climate Change \nArticle\nhttps://doi.org/10.1038/s41558-022-01501-7\nunderestimates inundations (Supplementary Information). This \nuncertainty is apparent when comparing the damage simulations to \nobserved events (2000\u20132018): GLOFRIS underestimates three extreme \nhurricanes (Katrina, Sandy and Harvey) and overestimates other events \n(albeit to a lesser extent). We therefore applied a scaling factor (Supple-\nmentary Information) to better match the average observed damage. \nThe underestimation of extremes can be explained by the dominant \nflooding processes during these events, which are not captured by our \nmodelling setup (for example, levee breach; Supplementary Informa-\ntion) and also by the fact that tropical cyclones are not well-represented \nin the GTSR dataset34.\nFurthermore, the ABM model is upscaled from a benchmarked \nABM for New York City using the same modelling decision rules and \ntheory. However, the benchmarked parameters (risk aversion, protec-\ntion standard, investment costs of dry flood-proofing, expenditure cap \nof dry flood-proofing investments, loan interest rate and loan dura-\ntion) for New York City do not apply for the whole USA, which is why \nwe applied standardized values of the parameters in the behavioural \nrules, as mentioned above. In addition, we benchmarked the parameter \n\u2018expenditure cap for insurance\u2019, by testing different values (2.5%, 5% and \n7.5%) for this parameter and then ran the model 50 times for each of the \nparameter values. By comparing the number of policies per county as \nsimulated by the model with the nationwide database (Supplementary \nInformation), through a Spearman\u2019s Rho correlation test, the cap was \nset to 7.5% (\u03c1\u2009=\u20090.679; P\u2009=\u20090.000).\nValidation. Modelling outputs for the years 2000\u20132018 were compared \nto FEMA data (Supplementary Information) using two indicators (insur-\nance damage payouts and premium income). The model was run for 200 \nrepetitions to account for uncertainty. Supplementary Table 6 shows a \nslight underestimation of mean yearly annual insurance damage pay-\nouts (US$2.8\u2009billion versus US$3.0\u2009billion) and a slight overestimation of \npremium income (US$1.7\u2009billion versus US$1.3\u2009billion) when comparing \nthe model with the observed values, respectively. However, the standard \ndeviation of the observed values was much higher, especially for the \ndamages (0.9 versus 4.4 for the model and observed values). This last dif-\nference can be attributed to the underestimation of the damage by flood \nrisk model for the three extreme hurricanes in the period 2000\u20132018.\nSensitivity analysis. The results from the validation provide confidence \nthat, except for the three extreme events in the years 2000\u20132018, our \nmodelled values of both damage payouts and premium income are \nclose to their actual counterparts. However, to further test modelling \nuncertainty, a sensitivity analysis was conducted (Supplementary \nTables 7\u201310), varying four key decision variables in the ABM:\n\t(1)\t Reducing the expenditure cap for insurance of 7.5% decreases the \nshare of policy-holders up to 15% and 42% for an expenditure cap \nof 5.0% and 2.5%, respectively. This is largely caused by an increase \nin household affordability. However, since other variables are \nonly slightly influenced and the expenditure cap is benchmarked \non the observed data, the used benchmark of 7.5% seems valid.\n\t(2)\t Increasing the expenditure cap for DRR investments from 2.5% \nto 7.5% results in more households investing in DRR (up to \n54% for an expenditure cap of 7.5%) and consequently reduces \nresidential risk (up to 20%). However, the changes in results are \nrelatively uniform between scenarios and do not affect the main \nconclusions of the paper.\n\t(3)\t Increasing or decreasing the loan interest rates either decreases \nor increases the share of households that invest in DRR. How-\never, this has a relatively low impact on results (with residential \nrisk only changing by up to 4%).\n\t(4)\t Varying the governmental investment costs by \u00b120% has only \na minor impact on overall modelling results (for example, risk \nfluctuates by \u00b15% uniformly between scenarios).\nData availability\nLand-cover data were obtained from GlobeLand30,42. Inundation data \nwere obtained from the GLOFRIS cascade model33. Vulnerability curves \nwere obtained from HAZUS-MH model44. Maximum damage values are \navailable at ref. 43. NFIP insurance data are available at ref. 55. Income \ndata were obtained from the US Census Bureau59. Protection standard \ndatabase was obtained from FLOPROS50. NFIP Redacted Claims dataset is \navailable from FEMA73. FEMA and the Federal Government cannot vouch \nfor the data or analyses derived from these data after the data have been \nretrieved from the Agency\u2019s website(s) and/or Data.gov. Socio-economic \ndata were obtained from the International Institute for Applied Systems \nAnalysis74. The generated data that support the finding of this study \nare available in figshare with the identifiers: https://doi.org/10.6084/\nm9.figshare.17049416.v1. There are no restrictions on data availability.\nCode availability\nThe code for DYNAMO is available in Zenodo with the identifiers: \nhttps://doi.org/10.5281/zenodo.7025225. There are no restrictions \non code availability.\nReferences\n24.\t de Ruig, L. T. et al. An agent-based model for evaluating reforms \nof the National Flood Insurance Program: a benchmarked model \napplied to Jamaica Bay, NYC. Risk Anal. https://doi.org/10.1111/\nrisa.13905 (2022).\n25. Kron, W. Flood risk\u2009=\u2009hazard\u2009\u2022\u2009values\u2009\u2022\u2009vulnerability. Water Int. 30, \n58\u201368 (2005).\n26.\t de Moel, H., van Vliet, M. & Aerts, J. C. J. H. Evaluating the \neffect of flood damage-reducing measures: a case study of the \nunembanked area of Rotterdam, the Netherlands. Reg. Environ \nChange 14, 895\u2013908 (2014).\n27.\t Riahi, K. et al. The Shared Socioeconomic Pathways and their \nenergy, land use, and greenhouse gas emissions implications: an \noverview. Glob. Environ. Change 42, 153\u2013168 (2017).\n28.\t Samir, K. C. & Lutz, W. The human core of the shared \nsocioeconomic pathways: population scenarios by age, sex and \nlevel of education for all countries to 2100. Glob. Environ. Change \n42, 181\u2013192 (2017).\n29.\t Dellink, R., Chateau, J., Lanzi, E. & Magn\u00e9, B. Long-term economic \ngrowth projections in the Shared Socioeconomic Pathways.  \nGlob. Environ. Change 42, 200\u2013214 (2017).\n30.\t Crespo Cuaresma, J. Income projections for climate change \nresearch: a framework based on human capital dynamics.  \nGlob. Environ. Change 42, 226\u2013236 (2017).\n31.\t Leimbach, M., Kriegler, E., Roming, N. & Schwanitz, J. Future \ngrowth patterns of world regions\u2014a GDP scenario approach. \nGlob. Environ. Change 42, 215\u2013225 (2017).\n32.\t Jiang, L. & O\u2019Neill, B. C. Global urbanization projections for the \nShared Socioeconomic Pathways. Glob. Environ. Change 42, \n193\u2013199 (2017).\n33.\t Ward, P. J. et al. Aqueduct Floods Methodology (WRI, 2020);  \nwww.wri.org/publication/aqueduct-floods-methodology\n34.\t Muis, S., Verlaan, M., Winsemius, H. C., Aerts, J. C. J. H. & Ward, P. \nJ. A global reanalysis of storm surges and extreme sea levels.  \nNat. Commun. 7, 11969 (2016).\n35.\t Carr\u00e8re, L. & Lyard, F. Modeling the barotropic response of \nthe global ocean to atmospheric wind and pressure forcing\u2014\ncomparisons with observations. Geophys. Res. Lett. https://doi.\norg/10.1029/2002GL016473 (2003).\n36.\t Dee, D. P. et al. The ERA-Interim reanalysis: configuration and \nperformance of the data assimilation system. Q. J. R. Meteorol. \nSoc. 137, 553\u2013597 (2011).\n37.\t Muis, S. et al. A comparison of two global datasets of extreme \nsea levels and resulting flood exposure. Earths Future 5, 379\u2013392 \n(2017).\n\nNature Climate Change \nArticle\nhttps://doi.org/10.1038/s41558-022-01501-7\n38.\t Jevrejeva, S., Grinsted, A. & Moore, J. C. Upper limit for sea  \nlevel projections by 2100. Environ. Res. Lett. 9, 104008  \n(2014).\n39.\t Kooi, H., Bakr, M., de Lange, G., den Haan, E. & Erkens, G. User \nGuide to SUB-CR a MODFLOW Package for Land Subsidence and \nAquifer System Compaction that Includes Creep Deltares internal \nreport 11202275-008 (Deltares, 2018).\n40.\t Guidance for Flood Risk Analysis and Mapping Combined Coastal \nand Riverine (FEMA, 2015).\n41.\t Bates, P. D. et al. Combined modeling of US fluvial, pluvial, and \ncoastal flood hazard under current and future climates. Water \nResour. Res. 57, e2020WR028673 (2021).\n42.\t Jun, C., Ban, Y. & Li, S. Open access to Earth land-cover map. \nNature 514, 434 (2014).\n43.\t Huizinga, J., de Moel, H. & Szewczyk, W. Global Flood Depth\u2013\nDamage Functions: Methodology and the Database with  \nGuidelines (Publications Office of the European Union, 2017); \nhttps://doi.org/10.2760/16510\n44.\t Multi-Hazard Loss Estimation Methodology, Flood Model, HAZUS, \nTechnical Manual (FEMA, 2013).\n45.\t de Ruig, L. T., Haer, T., de Moel, H., Botzen, W. J. W. & Aerts, J. C. \nJ. H. A micro-scale cost\u2013benefit analysis of building-level flood \nrisk adaptation measures in Los Angeles. Water Resour. Econ. 32, \n100147 (2020).\n46.\t Aerts, J. C. J. H. & Botzen, W. J. W. Flood-resilient waterfront \ndevelopment in New York City: bridging flood insurance, \nbuilding codes, and flood zoning. Ann. N. Y. Acad. Sci. 1227, 1\u201382 \n(2011).\n47.\t Homeowner\u2019s Guide to Retrofitting\u2014Six Ways to Protect your Home \nfrom Flooding (FEMA, 2014).\n48.\t Aerts, J. C. J. H., Botzen, W. J. W., de Moel, H. & Bowman, M.  \nCost estimates for flood resilience and protection strategies in \nNew York City. Ann. N. Y. Acad. Sci. 1294, 1\u2013104 (2013).\n49.\t de Ruig, L. T. et al. An economic evaluation of adaptation \npathways in coastal mega cities: an illustration for Los Angeles. \nSci. Total Environ. 678, 647\u2013659 (2019).\n50.\t Scussolini, P. et al. FLOPROS: an evolving global database of \nflood protection standards. Nat. Hazards Earth Syst. Sci. 16, \n1049\u20131061 (2016).\n51.\t Hallegatte, S., Green, C., Nicholls, R. J. & Corfee-Morlot, J. Future \nflood losses in major coastal cities. Nat. Clim. Change 3, 802\u2013806 \n(2013).\n52.\t Nicholls, R. & Cazenave, A. Sea-level rise and its impact on coastal \nzones. Science 328, 1517\u20131520 (2010).\n53.\t Czajkowski, J., Villarini, G., Montgomery, M., Michel-Kerjan, E. O. & \nGoska, R. Assessing current and future freshwater flood risk from \nNorth Atlantic tropical cyclones via insurance claims. Sci. Rep. 7, \n41609 (2017).\n54.\t Technical Documentation of NFIP Actuarial Assumptions and \nMethods (FEMA, 2013).\n55.\t The National Flood Insurance Program: Financial Soundness and \nAffordability (Congressional Budget Office, 2017).\n56.\t Dixon, L. & Clancy, N. The National Flood Insurance Program\u2019s \nMarket Penetration Rate: Estimates and Policy Implications (RAND \nCorporation, 2006); https://www.rand.org/pubs/technical_\nreports/TR300.html\n57.\t Kousky, C. & Kunreuther, H. Addressing affordability in the \nNational Flood Insurance Program. J. Extreme Events 1, 1450001 \n(2014).\n58.\t Hudson, P. A comparison of definitions of affordability for flood \nrisk adaption measures: a case study of current and future \nrisk-based flood insurance premiums in Europe. Mitig. Adapt. \nStrateg. Glob. Change 23, 1019\u20131038 (2018).\n59.\t Income Data per County (US Census Bureau, 2010); https://www2.\ncensus.gov/geo/tiger/TIGER_DP/2010ACS/\n60.\t Von Neumann, J. & Morgenstern, O. Theory of Games and \nEconomic Behavior (Princeton Univ. Press, 1947).\n61.\t Hudson, P., Botzen, W. J. W. & Aerts, J. C. J. H. Flood insurance \narrangements in the European Union for future flood risk under \nclimate and socioeconomic change. Glob. Environ. Change 58, \n101966 (2019).\n62.\t Harrison, G. W., List, J. A. & Towe, C. Naturally occurring \npreferences and exogenous laboratory experiments:  \na case study of risk aversion. Econometrica 75, 433\u2013458  \n(2007).\n63.\t Bombardini, M. & Trebbi, F. Risk aversion and expected utility \ntheory: an experiment with large and small stakes. J. Eur. Econ. \nAssoc. 10, 1348\u20131399 (2012).\n64.\t Wakker, P. P. in Wakker, P. P. (ed) Prospect Theory for Risk and \nAmbiguity Ch. 3 (Cambridge Univ. Press, 2008).\n65.\t Kunreuther, H. Mitigating disaster losses through insurance.  \nJ. Risk Uncertain. 12, 171\u2013187 (1996).\n66.\t Bin, O. & Landry, C. E. Changes in implicit flood risk premiums: \nempirical evidence from the housing market. J. Environ. Econ. \nManag. 65, 361\u2013376 (2013).\n67.\t Kunreuther, H., Sanderson, W. & Vetschera, R. A behavioral model \nof the adoption of protective activities. J. Econ. Behav. Organ. 6, \n1\u201315 (1985).\n68.\t Bubeck, P., Botzen, W. J. W. & Aerts, J. C. J. H. A review of risk \nperceptions and other factors that influence flood mitigation \nbehavior. Risk Anal. 32, 1481\u20131495 (2012).\n69.\t Tol, R. S. J. The social cost of carbon. Annu. Rev. Resour. Econ. 3, \n419\u2013443 (2011).\n70.\t Nicholls, R. J. Coastal flooding and wetland loss in the  \n21st century: changes under the SRES climate and socio- \neconomic scenarios. Glob. Environ. Change 14, 69\u201386  \n(2004).\n71.\t Aerts, J. C. J. H. A review of cost estimates for flood adaptation. \nWater 10, 1646 (2018).\n72.\t Cirillo, P. & Gallegati, M. The empirical validation of an \nagent-based model. East Econ. J. 38, 525\u2013547 (2012).\n73.\t OpenFEMA Dataset: FIMA NFIP Redacted Claims (FEMA, accessed \n23 June 2020); https://www.fema.gov/openfema-data-page/\nfima-nfip-redacted-claims\n74.\t Keywan, R. et al. The Shared Socioeconomic Pathways and  \ntheir energy, land use, and greenhouse gas emissions \nimplications: an overview. Global Environmental Change 42, \n153\u2013168 (2017).\nAcknowledgements\nWe would like to thank M. Montgomery for sharing data and  \nM. Tesselaar for his feedback. This research received funding from \nthe Netherlands Organization for Scientific Research VIDI (45214005 \nto W.J.W.B.) and VICI (016140067 and 453-13-006 to J.A.) grant \nprogrammes and ERC advanced grant (884442 to J.A.).\nAuthor contributions\nThe work was conceptualized by L.T.de R., T.H., W.J.W.B. and J.A. The \nmethodology was developed by L.T.de R., T.H. and H.de M. Formal \nana\u00adlysis was undertaken by L.T.de R. and T.H. Visualization was by \nL.T.de R. and H.de M. Supervision and funding acquisition were by \nJ.A. The article was written by L.T.de R., H.de M., S.D.B., W.J.W.B., J.C. \nand J.C.J.H.A.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at  \nhttps://doi.org/10.1038/s41558-022-01501-7.\n\nNature Climate Change \nArticle\nhttps://doi.org/10.1038/s41558-022-01501-7\nSupplementary information The online version  \ncontains supplemen\u00adtary material available at  \nhttps://doi.org/10.1038/s41558-022-01501-7.\nCorrespondence and requests for materials should be addressed to \nLars T. de Ruig or Jeroen C. J. H. Aerts.\nPeer review information Nature Climate Change thanks Dylan Turner, \nYu Han, Brett Sanders and Brayton Noll for their contribution to the \npeer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\n\nNature Climate Change \nArticle\nhttps://doi.org/10.1038/s41558-022-01501-7\nExtended Data Fig. 1 | A schematic overview of the primary modelling \nsteps. A schematic overview of the primary modelling steps, showing the main \nbuildings blocks of the DYNamic climate impact Adaptation Model (DYNAMO): \n(a) modelling schematic: a flood risk model and an agent-based model, (b) input \ndata: flood maps, exposure data, flood protection data, income, (c) scenarios: \nsocio-economic and climate change scenarios until 2050, (d) policy scenarios: \ngovernmental adaptation policies and NFIP market structures, (e) outputs: flood \nrisk (EAD), insurance penetration rates, affordability, disaster risk reduction \n(DRR), and flood protection standards, (f) validation: damage and premium \nincome. The framework builds upon earlier applications and version of DYNAMO \nby Haer et al.21 for the EU and de Ruig et al.28 for New York City.\n\n\n    Scientific Research Findings:", "answer": "We show that risk-based premiums, which are spatially explicit on a household level, result in decreases for some households but substantial increases for others. Importantly though, we observe that offering premium discounts to incentivize building-level risk reduction measures decreases total residential flood risk across the United States by approximately US$1 billion per year (\u22127.3%) by 2050. Combined, introducing risk-based NFIP premiums that incentivize household risk reduction will yield a positive societal net benefit (US$10 billion in 30 years). Complementing this reform with pro-active government investments in large-scale flood protection yields an even higher overall societal net benefit (US$26 billion in 30 years). In conclusion, to fully minimize future flood risk, investments in large-scale flood protection are required in addition to NFIP rating reforms. Whereas our study specifically focusses on the United States, the concept of incentivizing risk reduction through risk-based pricing is highly relevant for adaptation planning globally.", "id": 53}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41558-022-01379-5\n1Department of Management, John Molson School of Business, Concordia University, Montr\u00e9al, Quebec, Canada. 2Department of Geography, Planning and \nEnvironment, Concordia University, Montr\u00e9al, Quebec, Canada. 3Loyola Sustainability Research Centre, Concordia University, Montr\u00e9al, Quebec, Canada. \n4Centre for Business, Climate Change and Sustainability, University of Edinburgh Business School, Edinburgh, UK. \u2709e-mail: anders.bjoern@concordia.ca\nT\nhere is increasing focus on the role of the private sector in \nmeeting global climate goals1\u20133. To this end, science-based \ntargets (SBTs) intend to align voluntary company-level emis-\nsion reduction targets with the global temperature goal of the Paris \nAgreement4\u20136. So far SBTs have been set by more than 1,000 com-\npanies, including many multinationals7,8. When reporting scope \n2 emissions, that is, emissions associated with the generation of \npurchased energy (primarily electricity, but also including steam, \nheat and cooling), the Science Based Targets initiative (SBTi) allows \ncompanies to use renewable energy certificates (RECs) to claim \nthe use of renewably generated electricity. Companies can then \nreport zero emissions for each unit of electricity consumption cov-\nered by purchased RECs9, regardless of the actual emissions pro-\nduced by the electricity grid at their location; however, RECs do \nnot reflect the physical electricity flow supplied to the companies \npurchasing them10, and there is evidence that RECs are unlikely to \nlead to additional renewable energy generation11\u201318. Consequently, \ncompany-level emission reductions reported through RECs are \nunlikely to reflect real reductions of global emissions, which has \nthe potential to compromise the alignment of SBTs with the Paris \ntemperature goal19,20. In recognition of this issue, several emis-\nsion accounting standards21,22 restrict or do not endorse the use of \nRECs, but RECs are nevertheless permitted by the Greenhouse Gas \nProtocol9 that forms the basis of SBTi\u2019s requirements.\nHere we combine information about SBTi-certified company \nemission reduction targets with data on REC purchases to assess \nthe effect of RECs on the alignment of companies\u2019 reported scope \n2 emission trajectories with the Paris temperature goal. We use the \nclimate change disclosures of 115 companies, which represents the \nsubset of companies with SBTs that have also disclosed data which \ncan be used to assess the contribution of RECs to their reported \n(2015\u20132019) and potential future emission reductions. Importantly, \nit was necessary that all companies in our sample reported their past \nemissions using both market- and location-based emission account-\ning methods (Box 1). We also distinguish here between RECs and \npower purchase agreements (PPAs), which represent a long-term \ncommitment by a company to purchase power from a particu-\nlar renewable energy project. Although empirical evidence is still \nneeded, we have adopted here the common assumption that PPAs \ndo lead to additional renewable energy production and real emis-\nsion reductions, as the long-term power price de-risks new proj-\nects and allows access to project finance14,15,17,18 (Box 1). By contrast, \nwe assume that RECs and similar market-based instruments are \nnon-additional, that is, not leading to additional renewable genera-\ntion capacity or real emissions reductions, and we use the term RECs \nto refer to all non-PPA instruments for that reason (see Methods \nfor details on these market-based instruments and terminology). \nAlthough existing literature suggest that RECs are non-additional \ndue to their low and uncertain prices11\u201318, some claim that RECs may \nstill contribute to the generation of more renewable energy in the \nlonger term by, in aggregation, signaling to the market that there \nis a demand for renewable energy9,23. Analyses so far do not find \nevidence to support the existence of such an indirect market effect, \nand we consequently do not consider this potential effect here. We \nacknowledge, however, the possibility that such longer-term indi-\nrect effects may become evident in future analyses.\nEffect of RECs on reported historical emission reductions\nThe sample of companies reported a combined 30.7% reduction \nin market-based scope 2 emissions between 2015 and 2019 (from \n68.9 to 47.8\u2009Mt CO2e per year, see solid black line in Fig. 1e). This \ncorresponds to a substantially higher reduction than the annual \n4.2% of base year emissions required by SBTi\u2019s linear 1.5\u2009\u00b0C global \nmitigation pathway (purple line). However, most of this reported \nemission reduction is caused by the companies\u2019 use of RECs (Fig. \n1d), which increased from covering 8% of their purchased energy \nin 2015 to 27% in 2019. Based on the existing empirical evidence, \nwe assume that this part of the reported reduction does not reflect \nactual reductions of emissions from the energy grid. Without the \nRECs contribution, market-based emissions would have reduced \nby only 9.9% between 2015 and 2019, with the resulting emission \ntrajectory closely aligning with the required annual 2.5% base year  \nRenewable energy certificates threaten the \nintegrity of corporate science-based targets\nAnders Bj\u00f8rn\u200a \u200a1,2,3\u2009\u2709, Shannon M. Lloyd\u200a \u200a1,3, Matthew Brander\u200a \u200a4 and H. Damon Matthews\u200a \u200a2,3\nCurrent greenhouse gas accounting standards allow companies to use renewable energy certificates (RECs) to report reduc-\ntions in emissions from purchased electricity (scope 2) as progress towards meeting their science-based targets. However, pre-\nvious analyses suggest that corporate REC purchases are unlikely to lead to additional renewable energy production. Here we \nshow that the widespread use of RECs by companies with science-based targets has led to an inflated estimate of the effective-\nness of mitigation efforts. When removing the emission reductions claimed through RECs, companies\u2019 combined 2015\u20132019 \nscope 2 emission trajectories are no longer aligned with the 1.5\u2009\u00b0C goal, and only barely with the well below 2\u2009\u00b0C goal of the Paris \nAgreement. If this trend continues, 42% of committed scope 2 emission reductions will not result in real-world mitigation. Our \nfindings suggest a need to revise accounting guidelines to require companies to report only real emission reductions as prog-\nress towards meeting their science-based targets.\nNature Climate Change | VOL 12 | June 2022 | 539\u2013546 | www.nature.com/natureclimatechange\n539\n\nArticles\nNature Climate Change\nBox 1 | Corporate scope 2 emissions accounting and the role of RECs and PPAs\nUnder the Greenhouse Gas Protocol9, companies are required \nto report their scope 2 emissions using both market- and \nlocation-based accounting. Market-based accounting allows \ncompanies to use market-based instruments, including RECs \nand PPAs, which give companies the right to claim the use of re-\nnewable electricity and report the emissions from its production \n(usually zero), rather than the emissions of the actual electricity \nmix they consume9. At a given site and year, a company calcu-\nlates market-based (MB) scope 2 (S2) emissions (E) by, first, mul-\ntiplying the part of their electricity consumption covered by the \nmarket-based instrument (CMBI) by the emission factor of the in-\nstrument (EFMBI) and, second, multiplying any uncovered electric-\nity consumption (C\u2009\u2013\u2009CMBI) by a residual grid mix emission factor \n(EFres), which represents the local grid without the electricity gen-\neration that has been claimed by RECs or PPAs: \nEMB,S2 = CMBI \u00d7 EFMBI + (C \u2212CMBI) \u00d7 EFres\nUnder location-based (LB) accounting, all companies on a grid \nmultiply their electricity consumption (C) by the average emission \nfactor (EFmix) for the grid mix, regardless of whether they have \npurchased RECs or entered into PPAs:\nELB,S2 = C \u00d7 EFmix\nFor companies with multiple sites and that consume other \nenergy products from a grid (heat, steam or cooling) in addition \nto electricity, emissions (whether market- or location-based) are \ncalculated for each energy product and site, followed by aggregation.\nWe use a simple example to illustrate the calculation of market- and \nlocation-based emissions for a company with one site that purchases \nelectricity from the local grid, which also provides electricity to other \nconsumers. The company uses no market-based instruments in Year \n1, enters into a multiyear PPA in Year 2, and purchases RECs from an \nexisting renewable energy generator in Year 3.\nIn Year 1, the company has the same market- and location-based \nemissions as it does not purchase any RECs or enter into any PPAs, \nand because the grid average and residual emission factors are the \nsame (as no grid customers purchase RECs from or enter into \nPPAs with the local windfarm).\nIn Year 2, a new windfarm becomes operational due to a \nmultiyear PPA that the company has entered into with the project \ndeveloper. The PPA achieves additionality as the project would \nnot have happened in the absence of the PPA, which allowed the \nproject developer to secure the necessary loans. Moreover, as total \nelectricity demand does not change, the coal power plant reduces \nits production. The PPA gives the company the right to apply an \nemission factor of zero to 100\u2009MWh of its electricity consumption, \nand the residual grid factor (which decreases slightly due to \nthe reduced generation from coal) is applied to its remaining \nelectricity consumption. The PPA allows the company to report \na substantial market-based emission reduction. The company\u2019s \nlocation-based emissions are also reduced due to the lower grid \naverage emission factor from the reduced generation from coal. \nHowever, the reported emission reduction is less than with \nthe market-based approach, as the emission benefit of the new \nwindfarm is shared equally by all consumers on the grid under \nlocation-based accounting.\nIn Year 3, the company, whose PPA is still active, purchases RECs \nfrom the pre-existing windfarm. The RECs are non-additional \nas they do not lead to the generation of additional renewable \nenergy, as the windfarm was operational in years 1 and 2 and \nwould have continued to operate in Year 3 in the absence of the \ncompany\u2019s RECs purchase. The RECs (covering 50\u2009MWh) allow \nthe company to report a reduction in market-based emissions \ncompared with Year 2, even though total grid emissions did not \nchange (889\u2009tCO2e). The reported reduction is therefore not real \nand the RECs effectively increase the market-based emissions of \nother energy consumers on the grid through an increase in the \nresidual emission factor. By contrast, the company\u2019s location-based \nemissions are the same as in Year 2 as the grid average emission \nfactor did not change.\n0 tCO2e\n889 tCO2e\nEFmix: 0.89 tCO2e per MWh\nEFres: 1.05 tCO2e per MWh\nCompany \n100 MWh\n200 MWh\nYear 3\n800 MWh\n800 MWh\nOther \nconsumers\n0 tCO2e\n100 MWh\nPPA\n100 MWh\nEFMBI: 0\nRECs\n50 MWh\nEFMBI: 0\n0 tCO2e\n889 tCO2e\nEFmix: 0.89 tCO2e per MWh\nEFres: 0.99 tCO2e per MWh\nCompany \n100 MWh\n200 MWh\nYear 2\n800 MWh\n800 MWh\nOther \nconsumers\n0 tCO2e\n100 MWh\nPPA\n100 MWh\nEFMBI: 0\n0 tCO2e\n1,000 tCO2e\nEFmix: 1.00 tCO2e per MWh\nEFres: 1.00 tCO2e per MWh\nCompany \n100 MWh\n200 MWh\nYear 1\n900 MWh\n800 MWh\nOther \nconsumers\nRECs (not real \nreduction)\n0\n50\n100\n150\n200\nEmissions (tCO2e per year) \nLocation-based\nPPA and \nchanges to EFres\nChanges to EFmix\n0\n50\n100\n150\n200\nEmissions (tCO2e per year) \nMarket-based\nElectricity grid\nElectricity grid\nElectricity grid\nYear 1\nYear 2\nYear 3\nYear 1\nYear 2\nYear 3\nNature Climate Change | VOL 12 | June 2022 | 539\u2013546 | www.nature.com/natureclimatechange\n540\n\nArticles\nNature Climate Change\nemission reduction of SBTi\u2019s linear well below 2\u2009\u00b0C pathway (com-\npare the dashed black and turquoise lines in Fig. 1e). This real \nemission reduction is driven by decarbonization of the residual \nelectricity grid (representing the grid mix without the energy \nproduction that has been claimed by RECs or PPAs; see Box 1)  \n(Fig. 1b) and to a lesser extent by a net-increase in company use \nof PPAs (Fig. 1c). By contrast, the increase in energy consumption \nhad a net-positive contribution to the change in emissions (Fig. 1a). \nHence, the appearance that the combined historical scope 2 emis-\nsion trajectory of companies with SBTs easily aligned with the 1.5\u2009\u00b0C \ngoal is strongly misleading and a consequence of heavy reliance on \nRECs, which are not associated with real emission reductions.\nReported location-based emissions (black line in Fig. 1h) \nreduced from 77.1 to 69.2\u2009MtCO2e per year (10.3%) over the \nperiod, far less than the reported market-based emissions. This is \nas RECs and PPAs are not considered in location-based accounting \n(Box 1), which is also why location-based emissions were higher \nthan market-based emissions in 2015. Instead, decarbonization of \nthe electricity grid was the main contributor to the reported emis-\nsion reduction (Fig. 1g), with the increase in energy consumption  \nagain having a net-positive contribution to the change in emissions \n(Fig. 1f). The location-based emissions trajectory overall barely \ncomplied with the well below 2\u2009\u00b0C goal (turquoise line) and is simi-\nlar to the market-based trajectory adjusted to exclude RECs (dotted \nblack line in Fig. 1e).\nCompanies headquartered in Europe and North America (88% \nof the sample) reported larger combined market-based emission \nreductions in 2015\u20132019 than companies based in Asia (11% of \nthe sample), but also relied more on RECs in their reporting. When \nremoving the contribution from RECs, North American compa-\nnies\u2019 combined market-based emission trajectory merely aligned \nwith the well below 2\u2009\u00b0C goal, whereas European companies in \naggregate did not align with any Paris goal (see Supplementary \nFig. 7). At the industry level, only companies in \u2018materials\u2019, \u2018hospi-\ntality and biotech\u2019, \u2018health care and pharma\u2019 (together accounting \nfor 21% of the sample companies) aligned with the 1.5\u2009\u00b0C goal in \naggregate after adjusting the emission trajectories for RECs (see  \nSupplementary Fig. 8).\nRenewable energy certificates also had a substantial influence on \nthe reported emission trajectories of individual companies; 89% of \nsample companies purchased RECs in the 2015\u20132019 period and \nthe sample companies\u2019 median market-based emission reduction \nchanged from 30.2% to 8.5% when removing the contribution from \nRECs. Likewise, the share of companies aligned with the 1.5\u2009\u00b0C goal, \n0\n2\n4\n6\n\u201312\n\u20138\n\u20134\n0\nEnergy consumption\nEnergy consumption\nGrid mix\nResidual grid mix\nPPAs\nRECs\n\u20132\n0\n2\n\u201315\n\u201310\n\u20135\n0\n2016\n2017\n2018\n2019\n45\n50\n55\n60\n65\n70\n2015\n2016\n2017\n2018\n2019\nMarket-based emissions (MtCO2e per year)\nReported emissions\nEmissions adjusted for RECs\nWell below 2 \u00b0C goal pathway\n1.5 \u00b0C goal pathway\nRECs \n(unlikely to \nbe real \nreductions)\nFactors \nother than \nRECs\na\nb\nc\nd\nf\ng\ne\n0\n2\n4\n6\n\u201312\n\u20138\n\u20134\n0\n2016\n2017\n2018\n2019\nChange in 2015 location-based emissions\n(MtCO2e per year)\nChange in 2015 market-based emissions (MtCO2e per year)\nh\n60\n65\n70\n75\n80\n2015\n2016\n2017\n2018\n2019\nLocation-based emissions\n(MtCO2e per year)\nFactors \nother than \nRECs\nFig. 1 | Combined scope 2 emissions of sample companies and the factors contributing to their reductions since 2015. a\u2013e, Market-based accounting.  \nf\u2013h, Location-based accounting. a\u2013h, The contributions of energy consumption (a,f), residual grid mix (b), grid mix (g), PPAs (c) and RECs (d) to the reported \nmarket- (e) and location-based (h) emission trajectories, with the contribution of RECs to reported market-based emission reductions highlighted in orange.\nNature Climate Change | VOL 12 | June 2022 | 539\u2013546 | www.nature.com/natureclimatechange\n541\n\nArticles\nNature Climate Change\naligned with the well below 2\u2009\u00b0C goal, and not aligned with either \ngoal changes from 68%, 8% and 24% to 36%, 12% and 52%, respec-\ntively, when excluding RECs (Fig. 2); 40% of the companies whose \nreported market-based emissions aligned with the 1.5\u2009\u00b0C goal did not \nalign with any Paris goal after adjusting for RECs (see the grey flows \nbetween first and second column of Fig. 2). Hence, market-based \nemission disclosures give the impression that three-quarters of the \ncompanies were in alignment with one of the Paris goals. However, \nwhen removing contributions from RECs, only half of companies \nstill demonstrated such alignment and, of these, alignment with \nthe less ambitious goal became more common. Much of the differ-\nence in the distribution of company alignment with the Paris goals \nbetween the market- and location-based accounting disappears after \nremoving the contribution of RECs (Fig. 2). However, for one-third \nof the companies the temperature goal alignments is different for \nadjusted market- and location-based accounting. For example, for \ncompanies that increased their use of PPAs over the period adjusted \nmarket-based emissions commonly aligned with a more ambitious \ntemperature goal than location-based emissions.\nScenario of future REC usage in pursuit of SBTs\nWe now turn from the sample companies\u2019 past reported emissions \nto their future commitments to reducing emissions through SBTs  \n(Fig. 3a). Few companies (6%) report scope 2 SBTs independent \nfrom other emission scopes. Most companies (82%) report scope 2 \nSBTs in combination with SBTs for scope 1 (covering direct emis-\nsions24) and the rest (12%) in combination with scope 1 and scope 3 \nSBTs (covering value chain emissions beyond scope 2; ref. 24). Most \ncompanies (89%) state that their SBT refers to market-based  \naccounting, whereas the SBTs of the remaining 11% refer to \nlocation-based accounting (in short, market- and location-based \nSBTs). This indicates that most companies aim to use RECs and \nPPAs in pursuing their targets.\nCompanies with market-based SBTs tend to commit to more \nambitious emission reductions (compare the orange and blue boxes \nin Fig. 3a). The trend is especially pronounced for SBTs covering \nscopes 1 and 2, where the median annual reduction is 1.7 percent-\nage points higher for market- over location-based SBTs (4.2% versus \n2.5% of base year emissions). This may be because market-based \naccounting offers a relatively low-cost means of appearing to reduce \nemissions14, and therefore companies using this approach are will-\ning to set more ambitious reduction targets; 58% of market-based \nSBTs align with the 1.5\u2009\u00b0C goal (below purple line in Fig. 3a) and \n28% with the well below 2\u2009\u00b0C goal (between the turquoise and \npurple line). The corresponding shares for location-based SBTs \nare 8% and 54%, respectively. Some SBTs do not align with either \ntemperature goal as SBTi, until 2019, approved targets aligning \nwith a less ambitious 2\u2009\u00b0C goal8 and one target-setting method \n(the sectoral decarbonization approach4,25) allows companies with \ncertain characteristics to reduce emissions at a lower rate than is  \nrequired globally.\nWe next estimated the SBTs specifically for scope 2 (Fig. 3b) from \nSBTs that cover scopes 1 and 2, or scopes 1, 2 and 3, assuming future \nscope 2 emission reduction will have the same relative contribution \nto total emission reductions as in the past (see Methods). These esti-\nmates illustrate the implications of a continuation of current trends, \nrather than an explicit prediction of future emissions pathways. The \nestimated market-based scope 2 SBTs are generally more ambitious \nthan the reported market-based SBTs for overarching emissions \nscopes and are closer to the reported SBTs specifically for scope 2 \n(Fig. 3a). Note, however, that 3% of these estimated market-based \nSBTs involve emission increases (above the 0% line in Fig. 3), as \nthe companies in question increased scope 2 emissions during the \npast reference period; 75% of the estimated market-based scope \n2 SBTs comply with the 1.5\u2009\u00b0C goal and 12% comply with the \nwell below 2\u2009\u00b0C goal. Across companies, the estimated combined \nmarket-based scope 2 SBTs involve an annual reduction of 7.2% of \nbase year emissions (weighted average by base year emissions; see \nFig. 3b) and would thereby seem to easily comply with the 1.5\u2009\u00b0C \ngoal (purple line). By contrast, the estimated location-based scope \n2 SBTs only barely comply with the well below 2\u2009\u00b0C goal collectively \n(the weighted average reduction rate is 2.7% of base year emissions). \nHowever, the potential use of RECs may overstate the apparent Paris \ngoal alignment of the market-based SBTs.\nTo investigate further, we estimated and removed the future con-\ntribution of RECs from the estimated market-based scope 2 SBTs \n(Fig. 3c) on the basis of the relative contributions of RECs to past \nscope 2 emissions reductions (see Methods). This exploratory sce-\nnario results in a combined real reduction rate of 3.6% of base year \nemissions across companies (see the weighted average in Fig. 3c), \nwhich is markedly lower than the 7.2% reduction rate for unad-\njusted market-based scope 2 SBTs (Fig. 3b) and merely complies \nwith the well below 2\u2009\u00b0C goal. Moreover, although nearly all (77 of \n102) companies pursuing market-based scope 2 SBTs seem to align \nwith the 1.5\u2009\u00b0C goal (Fig. 3b), far fewer (38 of 102) companies will \nin fact align with the 1.5\u2009\u00b0C goal if they continue their past pattern \nof REC usage (Fig. 3c). In that scenario, companies will most com-\nmonly not align with either temperature goal (45 of 102), whereas \na minority (19 of 102) will align with the less ambitious well below \n2\u2009\u00b0C goal. The 25\u201375th percentile range of estimated market-based \nemission reduction rates adjusted for RECs (Fig. 3c) is similar to the \ncorresponding range for estimated location-based SBTs (Fig. 3b),  \nwhich challenges the notion that market-based SBTs tend to \nbe more ambitious than location-based SBTs (Fig. 3a,b). Taken \ntogether, the implication of this future emission scenario is that \nAligned with 1.5 \u00b0C goal\nAligned with well below 2 \u00b0C goal\nNot aligned with either goal\n35%\n13%\n52%\nShare of companies\n51%\n9%\n40%\n68%\n8%\n24%\n36%\n12%\n52%\nReported\nmarket-based \nMarket-based\nadjusted for RECs\nLocation-based\nFig. 2 | Share of sample companies whose 2015\u20132019 scope 2 emission \ntrajectories aligned with temperature goals for three emission accounting \napproaches. The grey flows between the first and second column indicate \nthe shares of companies for which the temperature goal alignment changed \nwhen removing RECs from their market-based emission trajectories. See \nthe Supplementary Spreadsheet for individual company results.\nNature Climate Change | VOL 12 | June 2022 | 539\u2013546 | www.nature.com/natureclimatechange\n542\n\nArticles\nNature Climate Change\nan estimated 42% of the committed cumulative scope 2 emission \nreductions (101\u2009MtCO2e) from base year to target year will come \nfrom REC purchases and hence may not reflect actual reductions of \nglobal emissions (Fig. 4).\nAt the regional level, European and North American companies \ngenerally targeted higher future emission reductions than Asian \ncompanies. However, these regional differences substantially reduce \nafter adjusting the targets for estimated future RECs purchases (see \nSupplementary Fig. 9). Companies in half of the ten industries \n(together accounting for 45% of the sample companies) collec-\ntively align with the 1.5\u2009\u00b0C goal after adjusting the SBTs for RECs, \nwhereas companies in two industries align with the well below 2\u2009\u00b0C \ngoal and companies in three industries do not align with either (see \nSupplementary Fig. 10).\nImplications and outlook\nIn this study we assessed the use of RECs by companies with SBTs \nand the implications for their Paris alignment claims. The wide-\nspread use of RECs raises doubt on companies\u2019 apparent historic \nParis-aligned emission reductions, as it allows companies to report \nemission reductions that are not real. Moreover, a continuation of \nrecent trends would mean that nearly half of future scope 2 emis-\nsion reductions reported by companies with SBTs would not be real. \nOverall, our results confirm earlier suppositions19,20 and suggest that \ncorporate use of RECs in the pursuit of SBTs is the norm rather than \nthe exception. This is consistent with recent findings for a smaller \nsample of companies with net-zero targets18. SBTs must cover both \nscope 1 and scope 2 emissions6, and for many industries scope 2 \nemissions are the larger of the two (see Supplementary Figs. 11\u201313 \nfor a contribution analysis of our company sample). Consequently, \ncompanies\u2019 use of RECs threatens to undermine the integrity of \nSBTs as a whole. Together with recent findings of widespread \naccounting and reporting issues for scope 3 emissions26,27, our find-\nings should inform future work scrutinizing SBTs and companies\u2019 \nprogress against them5,8,28.\nAlthough our sample only covers 115 (14%) of the 813 compa-\nnies with SBTs at the time (due to data availability and the exclu-\nsion of energy generators, utilities and companies with intensity \ntargets), it is fairly representative in terms of regions and industries  \n(see Supplementary Figs. 5 and 6). We encountered several difficul-\nties in interpreting companies\u2019 disclosure of RECs and PPAs and our \nmain results (Figs. 1\u20134) are based on conservative interpretations \n(see Supplementary Section 1 for details). A set of parallel results \n(Supplementary Figs. 1\u20134) show that reported emission reductions \ncould be even more inflated by RECs than is the case for our main \nresults. Our evaluation of the Paris alignment of corporate scope 2 \nemission trajectories is based on the same global mitigation path-\nways used in SBTi\u2019s target progress assessment8 (4.2% and 2.5% base \nyear emission reductions annually for the 1.5\u2009\u00b0C and well below \n2\u2009\u00b0C goals, respectively). However, Paris-aligned mitigation sce-\nnarios commonly involve substantially higher reductions in direct \nemissions for the power sector than for the rest of the economy \n(for example, 7.2% and 3.7% reduction of 2020 emissions annually \nbetween 2020 and 2030 for the 1.5\u2009\u00b0C and well below 2\u2009\u00b0C goals, \nrespectively, in the mitigation scenario underlying SBTi\u2019s sectoral \ndecarbonization approach25,29). This would translate to a require-\nment for companies to reduce scope 2 emissions at a higher rate \nthan the global mitigation pathways, meaning that fewer companies \nwould be Paris-aligned than suggested by our study.\nOur findings have implications for SBTi\u2019s current approach of \nallowing companies to choose between market- and location-based \naccounting when setting SBTs and reporting target progress6. In \naddition to the problem that market-based accounting allows \nreporting emission reductions that are not real, there is a risk of \ndouble counting the emission benefits of renewable energy gen-\neration if one company claims the use of specific renewable energy \ngeneration using market-based accounting, whereas other compa-\nnies count that same renewable energy using the grid average emis-\nsion factor in their location-based accounting. There are at least two \nalternatives that would make it more likely that all reported scope \n2 emission reductions are real and renewable energy generation is \nonly counted once (see Table 1). First, SBTi could require all com-\npanies to use only location-based accounting. A potential drawback \nwith this option is that it would disincentivize companies from using \nPPAs or other market-based instruments that can lead to addi-\ntional renewable energy generation. In the second alternative, all \nReported\nEstimated\nAdjusted for RECs\nBox components:\nMedian\nWeighted average\n25\u201375th percentile range\nNon-outlier range\nAccounting approach:\nMarket-based\nLocation-based\nAdjusted market-based\nParis goal:\n1.5 \u00b0C\nWell below 2 \u00b0C\nScope 2\nScope 2\nN = 7\nN = 86\nN = 8\nN = 9\nN = 5\nN = 102\nN = 13\nN = 102\nScope 2\nScopes 1 and 2\nScopes 1, 2 and 3\n10\na\nc\nb\n5\n0\n\u20135\nAnnual percent emission change (% of base year)\n\u201310\n\u201315\nFig. 3 | Annual future emission reductions based on the SBTs of the sample companies. All SBTs have been annualized by dividing their targeted \npercentage emission reduction by the time span (target year minus base year). a, SBTs as reported by companies, covering different emissions scopes \nand emission accounting approaches. b, scope 2 SBTs estimated on the basis of companies\u2019 reported SBTs (from a) and past scope 2 to total emission \nreductions. c, Market-based scope 2 SBTs (from b) adjusted by removing the estimated contribution from RECs based on past REC-related to total scope \n2 emission reductions. The number of targets covered is indicated above each box. The non-outlier ranges (whiskers) are defined as 1.5-times the 25\u201375th \npercentile range. Outliers are not displayed. See the Supplementary Spreadsheet for individual company estimates.\nNature Climate Change | VOL 12 | June 2022 | 539\u2013546 | www.nature.com/natureclimatechange\n543\n\nArticles\nNature Climate Change\ncompanies could be required to use a restrictive version of market- \nbased accounting, involving mandatory demonstration of the  \nadditionality of market-based instruments (whether PPAs or RECs), \nthat is, evidence that the renewable energy generation would likely \nnot have occurred without the instrument30,31. The Net Zero Carbon \nBuildings Framework of the UK Green Buildings Council includes \nsuch a requirement22,32. However, experience from carbon-offsetting \nmarkets shows that demonstration of additionality is complicated \nand often contested33 and more research and method develop-\nment may therefore be needed for this option to be viable. Some \nMarket-based emission\nreductions: 234 MtCO2e\nMarket-based \nemission \nreductions from \nRECs (unlikely to \nbe real): \n101 MtCO2e\nMarket-based\nemission reductions\nfrom factors other\nthan RECs:\n133 MtCO2e\nLocation-based emission \nreductions from factors other \nthan RECs: \n4 MtCO2e\nFig. 4 | Estimated cumulative scope 2 emission reduction from base to target year on the basis of the sample companies\u2019 SBTs and the estimated role \nof RECs in achieving these reductions. The inner circle represents the estimated emission reduction commitments by companies whose SBTs refer to \nmarket- and location-based accounting. The outer ring represents the estimated contributions of RECs and other factors (see Fig. 1) to these reduction \ncommitments. The median base year and target year of the SBTs are 2017 and 2030, respectively.\nTable 1 | The current emission accounting requirement of SBTi and two (not exhaustive) alternatives that could potentially prevent \ndouble counting and the reporting of emission reductions that are not real\nAccounting requirements\nIssues with reporting \nemission reductions \nthat are not real?\nIssues with double counting?\nOther potential issues (not exhaustive)\nCompanies choose between \nlocation- and market-based \naccounting with no mandatory \ndemonstration of additionality \n(current requirement of SBTi)\nYes: companies may use \nnon-additional RECs.\nYes: a unit of generated \nrenewable energy can count \ntowards the market-based \nemission reduction of \none company and the \nlocation-based emission \nreductions of other companies.\nResources may be diverted from interventions linked to \nactual emission reductions (such as energy efficiency \nimprovements) to purchasing RECs.\nAll companies use \nlocation-based accounting\nNo.\nNo.\nMay disincentivize companies from contributing to \nnew renewable energy generation, and the associated \nemission reductions, through PPAs and other \nmarket-based instruments.\nAll companies use market-based \naccounting with mandatory \ndemonstration of additionality\nNo.\nNo.\nDemonstration of additionality is complicated and \nuncertain.\nUse of market-based instruments for claiming renewable \nenergy supply may be seen as misleading, since they \ndo not reflect the dependence of the user on the wider \nelectricity generation and transmission system.\nNature Climate Change | VOL 12 | June 2022 | 539\u2013546 | www.nature.com/natureclimatechange\n544\n\nArticles\nNature Climate Change\nstakeholders might also reject the use of market-based instruments \naltogether, whether additional or not, as this way of claiming to \nbe supplied by renewable energy does not reflect how renewable \nenergy supply relies on a broader system of grid-balancing, back-up \ncapacity and transmission services, which are often supported by \ntax payers or other energy consumers10.\nOur study highlights issues with the use of RECs in the context \nof SBTs. However, the Greenhouse Gas Protocol scope 2 guidance \nargues that the use of non-additional market-based instruments is \nnot a problem as the main goal is to allocate total grid emissions \nto individual consumers9. From that perspective, individual com-\npanies can legitimately use RECs to report emissions reductions \nthat do not reflect a global emission reduction, as the market-based \nemissions of all energy users on the grid sum to total grid emissions. \nBased on our analysis, we encourage the Greenhouse Gas Protocol \nto reconsider its stance on additionality for two reasons. First, as \nmany corporate energy users do not report emissions (nor do resi-\ndential households), a company\u2019s purchase of non-additional RECs \neffectively increases the market-based emissions of other actors  \nthat do not report emissions (through an increase in the residual  \nemission factor; see Box 1). This means that total reported \nmarket-based emissions will always overstate the actual grid-total \nemissions reduction due to incomplete reporting. Second, as \ncompanies increasingly disclose emissions in the context of tar-\ngets informed by the need to reduce global emissions (SBTs5 and \nnet-zero targets34,35), it is clearly misleading to stakeholders if \ncompanies can meet these targets without reducing global emis-\nsions. This would also be the case in a situation with complete \nreporting. We acknowledge that this second issue raises a related \nand broader limitation with the use of standard corporate emis-\nsions accounting (whether market- or location-based), in that \nchanges in emissions outside of the scopes 1, 2 and 3 accounting \nboundary are not shown. Apparent reductions within the bound-\nary may therefore not reflect total reductions in global emis-\nsions. For example, a company\u2019s decision to use bioenergy may \ncause emissions outside its value chain through indirect land use \nchange36. Accordingly, SBTi should consider options for comple-\nmenting standard scopes 1, 2 and 3 accounts with consequential \nemission accounting methods36 to ensure that actions taken to \nachieve STBs do not unintentionally increase emissions outside the  \naccounting boundary.\nConclusion\nOur study shows that the common voluntary corporate prac-\ntice of using RECs that are unlikely to drive additional renewable \nenergy production casts serious doubt on the veracity of reported \ncorporate emission trajectories and their apparent alignment \nwith the most ambitious Paris Agreement temperature goal. More \nbroadly, there is a need to critically consider the extent to which \nvoluntary corporate actions can be relied on for achieving a Paris- \naligned transition5.\nOnline content\nAny methods, additional references, Nature Research report-\ning summaries, source data, extended data, supplementary infor-\nmation, acknowledgements, peer review information; details of \nauthor contributions and competing interests; and statements of \ndata and code availability are available at https://doi.org/10.1038/\ns41558-022-01379-5.\nReceived: 28 November 2021; Accepted: 29 April 2022;  \nPublished online: 9 June 2022\nReferences\n\t1.\t Lui, S. et al. Correcting course: the emission reduction potential of \ninternational cooperative initiatives. Clim. Policy 0, 1\u201319 (2020).\n\t2.\t Kuramochi, T. et al. Beyond national climate action: the impact of region, \ncity, and business commitments on global greenhouse gas emissions. Clim. \nPolicy 20, 275\u2013291 (2020).\n\t3.\t Hsu, A. et al. A research roadmap for quantifying non-state and subnational \nclimate mitigation action. Nat. Clim. Change 9, 11\u201317 (2019).\n\t4.\t Bj\u00f8rn, A., Lloyd, S. & Matthews, D. From the Paris Agreement to corporate \nclimate commitments: evaluation of seven methods for setting \n\u201cscience-based\u201d emission targets. Environ. Res. Lett. 16, 054019 (2021).\n\t5.\t Bj\u00f8rn, A., Tilsted, J. P., Addas, A. & Lloyd, S. M. Current Climate Change \nReports. (2022); https://doi.org/10.1007/s40641-022-00182-w\n\t6.\t SBTi Corporate Manual TVT INF 002 v.2.0 (SBTi, 2021); https:// \nsciencebasedtargets.org/resources/files/SBTi-Corporate-Manual.pdf\n\t7.\t Companies Taking Action (SBTi, 2022); https://sciencebasedtargets.org/ \ncompanies-taking-action/\n\t8.\t From Ambition To Impact: How Companies Are Reducing Emissions At Scale \nWith Science-Based Targets (SBTi, 2021); https://sciencebasedtargets.org/ \nresources/files/SBTiProgressReport2020.pdf\n\t9.\t GHG Protocol Scope 2 Guidance: An Amendment to the GHG Protocol \nCorporate Standard (World Resources Institute, 2015); https://ghgprotocol. \norg/scope_2_guidance\n\t10.\tMonyei, C. G. & Jenkins, K. E. H. Electrons have no identity: setting right \nmisrepresentations in Google and Apple\u2019s clean energy purchasing. Energy \nRes. Soc. Sci. 46, 48\u201351 (2018).\n\t11.\tGillenwater, M., Lu, X. & Fischlein, M. Additionality of wind energy \ninvestments in the U.S. voluntary green power market. Renew. Energy 63, \n452\u2013457 (2014).\n\t12.\tGillenwater, M. Probabilistic decision model of wind power investment and \ninfluence of green power market. Energy Policy 63, 1111\u20131125 (2013).\n\t13.\tHamburger, \u00c1. & Harangoz\u00f3, G. Factors affecting the evolution of renewable \nelectricity generating capacities: a panel data analysis of European countries. \nJ. Energy Econ. Policy 8, 161\u2013172 (2018).\n\t14.\tBrander, M., Gillenwater, M. & Ascui, F. Creative accounting: a critical \nperspective on the market-based method for reporting purchased electricity \n(scope 2) emissions. Energy Policy 112, 29\u201333 (2018).\n\t15.\tNavigating the Nuances of Net-Zero Targets (New Climate Institute and Data \nDriven Envirolab, 2020).\n\t16.\tMulder, M. & Zomer, S. P. E. Contribution of green labels in electricity  \nretail markets to fostering renewable energy. Energy Policy 99,  \n100\u2013109 (2016).\n\t17.\tBriefing Document Corporate Procurement of Renewable Energy: Implications \nand Considerations (Climate Change Committee, 2020); https://www. \ntheccc.org.uk/publication/corporate-procurement-of-renewable-energy- \nimplications-and-considerations/\n\t18.\tCorporate Climate Responsibility Monitor 2022: Assessing the Transparency  \nand Integrity of Companies\u2019 Emission Reduction and Net-Zero Targets  \n(New Climate Institute, 2022); https://newclimate.org/2022/02/07/corporate- \nclimate-responsibility-monitor-2022/\n\t19.\tWalenta, J. Climate risk assessments and science-based targets: a review of \nemerging private sector climate action tools. Wiley Interdiscip. Rev. Clim. \nChange 11, e628 (2020).\n\t20.\tTrexler, M. & Schendler, A. Science-based carbon targets for the corporate \nworld: the ultimate sustainability commitment, or a costly distraction? J. Ind. \nEcol. 19, 931\u2013933 (2015).\n\t21.\tISO 14064-1. Greenhouse Gases\u2014Part 1: Specification with Guidance at the \nOrganization Level for Quantification and Reporting of Greenhouse Gas \nEmissions and Removals 2nd edn (International Organization for \nStandardization, 2018); https://www.iso.org/standard/66453.html\n\t22.\tRenewable Energy Procurement and Carbon Offsetting Guidance for Net Zero \nCarbon Buildings (UK Green Building Council, 2021); https://www.ukgbc. \norg/ukgbc-work/renewable-energy-procurement-carbon-offsetting- \nguidance-for-net-zero-carbon-buildings/\n\t23.\tHow Renewable Energy Certificates Make a Difference: The Impacts and \nBenefits of Buying Renewable Energy (Center for Resource Solutions, 2016); \nhttps://resource-solutions.org/wp-content/uploads/2016/03/How-RECs- \nMake-a-Difference.pdf\n\t24.\tThe Greenhouse Gas Protocol\u2014A Corporate Accounting and Reporting \nStandard (World Business Council For Sustainable Development and World \nResources Institute, 2004); https://ghgprotocol.org/sites/default/files/ \nstandards/ghg-protocol-revised.pdf\n\t25.\tKrabbe, O. et al. Aligning corporate greenhouse-gas emissions targets with \nclimate goals. Nat. Clim. Change 5, 1057\u20131060 (2015).\n\t26.\tKlaa\u00dfen, L. & Stoll, C. Harmonizing corporate carbon footprints. Nat. \nCommun. 12, 6149 (2021).\n\t27.\tBusch, T., Johnson, M. & Pioch, T. Corporate carbon performance data:  \nQuo vadis? J. Ind. Ecol. 26, 350\u2013363 (2020).\n\t28.\tGiesekam, J., Norman, J., Garvey, A. & Betts-Davies, S. Science-based  \ntargets: on target? Sustainability 13, 1657 (2021).\n\t29.\tSBTi Tool v.1.2.1 (SBTi, 2020); https://sciencebasedtargets.org/resources/ \nfiles/SBT-Tool-v1.2.1.xlsx\nNature Climate Change | VOL 12 | June 2022 | 539\u2013546 | www.nature.com/natureclimatechange\n545\n\nArticles\nNature Climate Change\n\t30.\tTrexler, M. C., Broekhoff, D. J. & Kosloff, L. H. A statistically-driven approach \nto offset-based GHG additionality determinations: what can we learn? \nSustain. Dev. Law Policy 6, 30\u201340 (2006).\n\t31.\tMethodological Tool: Tool for the Demonstration and Assessment of \nAdditionality v.07.0.0 (United Nations Framework Convention on Climate \nChange, 2012); https://cdm.unfccc.int/methodologies/PAmethodologies/ \ntools/am-tool-01-v7.0.0.pdf\n\t32.\tNet Zero Carbon Buildings: A Framework Definition (UK Green Building  \nCouncil, 2019); https://www.ukgbc.org/ukgbc-work/net-zero-carbon- \nbuildings-a-framework-definition/\n\t33.\tSchneider, L. Assessing the additionality of CDM projects: practical \nexperiences and lessons learned. Clim. Policy 9, 242\u2013254 (2009).\n\t34.\tHale, T. et al. Assessing the rapidly-emerging landscape of net zero targets. \nClim. Policy 22, 18\u201329 (2022).\n\t35.\tRogelj, J., Geden, O., Cowie, A. & Reisinger, A. Net-zero emissions targets are \nvague: three ways to fix. Nature 591, 365\u2013368 (2021).\n\t36.\tBrander, M. Comparative analysis of attributional corporate greenhouse gas \naccounting, consequential life cycle assessment, and project/policy level \naccounting: A bioenergy case study. J. Clean. Prod. 167, 1401\u20131414 (2017).\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2022\nNature Climate Change | VOL 12 | June 2022 | 539\u2013546 | www.nature.com/natureclimatechange\n546\n\nArticles\nNature Climate Change\nMethods\nTerminology for market-based instruments and their potential effects on \nrenewable energy generation. There is no formally agreed terminology for the \nmarket-based instruments that companies can use to obtain renewable energy \nattributes (that is, rights to claim usage of renewable energy) under the market-based \naccounting approach. A large number of instruments with similar characteristics \nexist and the same instrument may go by different names in different regions9. \nFor example, the Bloomberg database37 defines RECs broadly as the \u201c\u2026amount \nof bundled and unbundled energy attributes purchased in the form of tradable \ncertificates from an official registry, in thousands of megawatt hours. This field \nincludes voluntary certificates such as Guarantees of Origin sold within the \nRenewable Energy Certificate System and European Energy Certificate System as \nwell as Renewable Energy Certificates commonly used in the United States\u201d. SBTi \nidentifies RECs and PPAs as the two available market-based instruments for meeting \nrenewable electricity targets within its net-zero standard38, although other contractual \narrangements exist, such as supplier-specific emission factors, and green power \nproducts or tariffs. It should also be noted that PPAs themselves generally include the \nrenewable energy attributes associated with the power that is purchased (though it is \nalso possible to structure a PPA for the power only, without the associated attributes). \nFor simplicity in the terminology used we use the term REC for all contractual \narrangements for renewable energy attributes other than PPAs and we use the term \nPPA for PPAs that include renewable energy attributes, unless otherwise stated.\nWe use the term additionality in relation to the effect of RECs and other \nmarket-based instruments on the generation of renewable energy. The \nadditionality term originated in the carbon-offsetting literature31,33. In the context \nof using market-based instruments for renewable energy to report emission \nreductions, we here follow the definition of the term from the ISO 14064-2 \nstandard39: \u201cAdditionality, as a concept, describes the relationship between cause \nand effect. For any cause and effect, the effect can be described as being additional \nif it would not have occurred in the absence of the cause\u201d. In this study, the effect \ncan both refer to the generation of renewable energy and the development of new \ncapacity (for example, windmill or solar panels) for renewable energy generation.\nCompany data. Our initial sample was composed of the 813 companies with \napproved SBTs as per the online SBTi database7 on July 20th, 2021. The SBTs in \nthe SBTi database generally contain the following data of relevance to this study: \nsector, emission metric (absolute or intensity-based), emission scope (1, 2, 3 or \na combination), base year, target year and targeted percentage reduction in the \nemission metric. We first excluded 28 energy generators and utilities (3% of the \ninitial sample) due to our focus on companies that purchase energy. We excluded \nanother 101 companies (12% of the initial sample) that only have intensity-based \nSBTs covering scope 2 to avoid the additional uncertainty associated with \nconverting intensity targets to absolute emission targets. For companies with \nmultiple SBTs covering different emissions scopes and target years, we selected a \nsingle SBT, prioritizing targets specifically for scope 2, when available (otherwise, \nwe prioritized targets for scopes 1 and 2 combined over targets for scopes 1, 2 and \n3 combined), followed by prioritization of the shortest target time span (that is, \nthe difference between the base year and target year). In addition to the target data \nsourced from SBTi, we collected information on the scope 2 accounting approach \nthat each SBT refers to (market- or location-based) from company disclosures to \nCDP (formerly the Carbon Disclosure Project)40 (this information is not provided \nby SBTi; see Supplementary Section 2.4 for more details). Note, however, that only \naround half of companies with approved SBTs reported to CDP. We were therefore \nleft with 338 of the 813 companies after removing energy generators and utilities \n(28), intensity targets (101) and companies that did not report to CDP (346).\nFor these 338 remaining companies, we analyzed their past emissions for \n2015\u20132019. The start year (2015) aligns with the publication of the Greenhouse \nGas Protocol standard related to market-based accounting (see Box 1)9, the \nintroduction of a distinction between market- and location-based scope 2 \nemissions in the CDP questionnaire40 and the approval of the first SBTs by SBTi7. \nWe ended the period in 2019 to avoid abnormalities caused by COVID-19 and \nas 2019 is the most recent year for which complete data was available for many \ncompanies. We used companies\u2019 2015\u20132019 emissions disclosed to CDP and, in \ncases of missing or apparently erroneous data, complemented with emission  \ndata from the Bloomberg database37. We also used CDP data on companies\u2019 \npurchased energy (electricity, heat, steam and cooling) and purchased RECs  \nand PPAs (details in Supplementary Section 1). We adjusted ambiguous or \napparently erroneous datapoints (details in Supplementary Section 2). We excluded \n223 companies (27% of the initial sample and 66% of the restricted sample  \nof 338 companies) due to missing or poor-quality data, leaving 115 companies \n(14% of the initial sample) that are included in our analysis (details in \nSupplementary Section 2).\nThe resulting sample of 115 SBTs have been approved by SBTi between 2016 \nand 2021. We used SBTi\u2019s classification of headquarter regions and industries7 to \nassess the representativeness of the final company sample (Supplementary  \nSection 4) and estimate combined results at the region level (Supplementary  \nFigs. 7 and 9). For combined results at the industry level (Supplementary Figs. 8 \nand 10) and the industry-level assessment of scope 2 emissions relative to other \nemission scopes (Supplementary Figs. 11\u201313), we used the industry classifications \nof CDP40 as it involves a more manageable number of industries (10 versus the 49 \nof SBTi\u2019s classification system7).\nReference mitigation pathways for the Paris temperature goal. We used \nSBTi\u2019s linear 1.5\u2009\u00b0C and well below 2\u2009\u00b0C global mitigation pathways41, involving \nannual reductions of 4.2% and 2.5% of base year emissions, respectively. The \nSBTi developed these pathways from a subset of the pathways described in the \nSpecial Report on Global Warming of 1.5\u2009\u00b0C of the Intergovernmental Panel on \nClimate Change42. SBTi determined this subset by applying criteria related to \ntemperature limit probability, temporary overshoot of emission budget, year of \npeak emissions and near-term emission reduction rate, with the aim of isolating \npathways conforming with principles of plausibility, responsibility, objectivity, \nand consistency41. The SBTi notes that linearization of emission pathways over \nlong timespans can result in substantial deviations of the pathways\u2019 cumulative \nemissions and therefore recommends the use of the derived reduction rates (4.2% \nand 2.5%) for the shorter time span of 2020\u20132035. However, the SBTi also advises \ncompanies to apply these reductions rates to set SBTs for base years before 20206, \nand SBTi applied the requirement of the 1.5\u2009\u00b0C pathway (4.2% reduction in base \nyear emissions per year) as a benchmark for the combined emission trajectory of \ncompanies with SBTs in the 2015\u20132019 period in its latest target progress report8. \nFollowing SBTi, we here apply the annual emission reduction rates of the two \nSBTi pathways (4.2% and 2.5%, respectively) as references to evaluate the Paris \nalignment of past corporate emission trajectories (2015\u20132019) and future targeted \ntrajectories (median values 2017\u20132030) for the 115 companies.\nEstimating the contribution of scope 2 to total emission changes for 2015\u20132019. \nBased on the equation for calculating market-based scope 2 emissions (Box 1), \nwe used the following set of equations to calculate the contribution of changes \nin energy consumption, use of PPAs, use of RECs (collectively referred to as \nmarket-based instruments or MBI) and residual grid mix emission factor to \nreported changes in market-based scope 2 emissions for each year (that is, 2016, \n2017, 2018 and 2019, noted with \u2018t\u2019 below) relative to 2015 (Figs. 1 and 2).\nContribution to reported emission changes (\u0394E) from change in energy \nconsumption (C) relative to 2015 (in tCO2e per year):\n\u0394EMB,S2,\u0394C,t = EMB,S2,2015 \u00d7\nC2015 \u2212Ct\nC2015 \u2212CMBI, 2015\n(1)\nContribution to reported emission changes from change in energy \nconsumption that is covered by PPAs (CPPA) relative to 2015 (in tCO2e per year):\n\u0394EMB,S2,\u0394PPA,t = EMB,S2,2015 \u00d7 CPPA, t \u2212CPPA, 2015\nC2015 \u2212CMBI, 2015\n(2)\nContribution to reported emission changes from change in energy \nconsumption that is covered by RECs (CREC) relative to 2015 (in tCO2e per year):\n\u0394EMB,S2,\u0394REC,t = EMB,S2,2015 \u00d7 CREC, t \u2212CREC, 2015\nC2015 \u2212CMBI, 2015\n(3)\nContribution to reported emission changes from change in the residual grid \nmix emission factor (EFres) relative to 2015 (in tCO2e per year):\n\u0394EMB,S2,\u0394EF_res,t\n= (EMB,S2,2015 \u2212EMB,S2,t) \u2212(\u0394EMB,S2,\u0394C,t + \u0394EMB,S2,\u0394PPA,t + \u0394EMB,S2,\u0394REC)\n(4)\nNote that equation (4) estimates the contribution from a change in the residual \ngrid mix emission factors as the difference between the change in reported emissions \n(the first parenthesis) and the sum of the contributions to emission changes \nfrom changes in energy consumption, PPAs and RECs (the second parenthesis) \nas companies do not directly disclose the values of EFres behind their reported \nemissions to CDP. Note also that equations (1\u20134) are based on the approximation \nthat EFMBI is zero for all RECs and PPAs, which allows treating RECs and PPAs as if \nthey virtually eliminate the energy consumption they cover in equation (1\u20133). This \napproximation is needed as companies only reported EFMBI values to CDP for part \nof the study\u2019s time period (2016, 2017 and 2018). We consider the approximation \nthat EFMBI is zero in all cases sound, as biomass and biogas incineration is the only \ntype of renewable energy generation known to emit (small amounts of) GHGs \nduring operation (in addition to emission of biogenic CO2, which is considered \nclimate neutral in corporate accounting24), and because this energy source accounted \nfor a negligible share of the combined energy consumption claimed by RECs or \nPPAs (CMBI) across the company sample (less than 1% in 2019). We note that some \ncompanies reported non-zero values of EFMBI for other types of renewable energy \ngeneration such as hydropower, wind and solar, and we assume that this reflects \na misunderstanding of the CDP questionnaire (for example, reporting life-cycle \nemission factors instead of operational emission factors).\nWe next calculated adjusted market-based scope 2 emissions for each year \nfollowing 2015 (t) (aEMB,S2, in tCO2e per year) by removing the contribution from \nRECs to reported emission changes (see equation (3)):\naEMB,S2, t = EMB,S2,t + \u0394EMB,S2,\u0394REC,t\n(5)\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nFor location-based accounting, we used a similar set of equations, based on the \nequation for calculating location-based scope 2 emissions (Box 1), to calculate  \nthe contribution of changes in energy consumption and the average grid mix \nemission factor to reported changes in scope 2 emissions between 2015 and each \nfollowing year (t):\nContribution to reported emission changes from change in energy \nconsumption (C) relative to 2015 (in tCO2e per year):\n\u0394ELB,S2,\u0394C,t = ELB,S2,2015 \u00d7 C2015 \u2212Ct\nC2015\n(6)\nContribution to reported emission changes from change in the average grid \nmix emission factor (EFmix) relative to 2015 (in ton CO2e per year):\nELB,S2,\u0394EF_mix,t = (ELB,S2,2015 \u2212ELB,S2,t) \u2212\u0394ELB,S2,\u0394C,t\n(7)\nAnnualization of SBTs. As shown in Fig. 3a, to allow for comparison of \nthe targeted emission reductions across SBTs with varying timespans, we \nannualized the reported SBTs, formulated as a percentage reduction between the \ncompany-specific base year (B) and target year (Y), following Wang et al43:\nSBT\u2032 =\nSBT\nY \u2212B\n(8)\nThe resulting annualized SBT (SBT\u2032) indicates an annual percent reduction of \nbase year emissions and assumes a linear emission trajectory between base year \nand target year.\nEstimation of SBTs specific for scope 2. As shown in Fig. 3b, most companies \nin our sample (94%) do not have SBTs specifically for scope 2. Instead they have \ntargets that combine scope 1 and 2 (82%), or scope 1, 2 and 3 emissions (12%). \nTo derive SBTs specifically for scope 2 in these cases, we first calculated the past \ncontribution of scope 2 emission changes (S2Cpast, dimensionless) to the emission \nreductions for the emission scopes covered by a company\u2019s reported SBT:\nS2Cpast = ES2,b \u2212ES2,y\nESR,b \u2212ESR,y\n(9)\nHere, ES2 is the reported historic scope 2 emissions and ESR is the reported \nhistoric emissions for the scopes covered by the company\u2019s reported SBT, both \nreferring to market-based accounting for companies with market-based SBTs and \nlocation-based accounting for companies with location-based SBTs; b and y are \ncompany-specific start- and end-years within the 2015\u20132019 period, selected to \nrepresent the longest consecutive decrease in ESR. This approach (as opposed to a \ncommon start- and end-year) avoids past increases in ESR, which is desirable given \nthat S2Cpast forms the basis for projecting a company\u2019s future scope 2 emissions \nin the context of a targeted emission decrease across the emission scopes covered \nby its reported SBT. On average across our sample, the b to y period spans \n2.5 years. For one of the sample companies, ESR increased consistently in the \n2015\u20132019 period, meaning there was no company-specific basis for projecting \nscope 2 emissions. For that company we instead assigned S2Cpast the average value \ncalculated for the other sample companies with reported SBTs covering the same \nemission scopes (1\u20133, market-based), which was 13.4%.\nNext we extrapolated from the past reference period (b to y) by assuming that \nS2Cpast applies to the full period between the specified base year, B, and future \ntarget year, Y, of the reported company SBT and calculated the corresponding \ntargeted annual reductions in scope 2 emissions for that period (\u0394E\u2032\nS2) (again, \nmarket- or location-based, depending on the reported SBT):\n\u0394E\u2032\nS2 = ESR,B \u00d7 \u0394SBT\u2032 \u00d7 \u0394S2Cpast\n(10)\nFinally, we estimated the annualized SBT specifically for scope 2:\nSBT\u2032\nS2 = \u0394E\u2032\nS2\nES2,B\n(11)\nAdjustment of scope 2 SBTs for RECs. Given that many companies in our sample \nused RECs to report market-based emission reductions in the 2015\u20132019 period, \nit is likely that they will continue to use RECs as a contribution to meeting their \nscope 2-specific SBTs. To take this into account, we modified equation (9) to \nremove the contribution of RECs (Fig. 3c) to past changes in scope 2 emissions, \nby drawing on equations (3) and (5), which were modified to cover the b to y \nreference period (same as for equation (9)) instead of the 2015 to t period:\naS2Cpast =\nEMB,S2,b\u2212aEMB,S2,y\nEMB,SR,b\u2212EMB,SR,y\n=\nES2,mark,b\u2212(EMB,S2,y+\u0394EMB,S2,\u0394REC,y)\nEMB,SR,b\u2212EMB,SR,y\n=\nES2,mark,b\u2212\n(\nEMB,S2,y+EMB,S2,2015\u00d7\nCREC, y\u2212CREC, b\nCb\u2212CMBI, b\n)\nEMB,SR,b\u2212EMB,SR,y\n(12)\naS2Cpast can be interpreted as the contribution of a real change in scope 2 emissions \n(that is, without the contribution of RECs) to a reported change in emissions for \nthe scopes covered by the reported SBT. Following this, we calculated the adjusted \nscope 2-specific annualized SBTs (aSBT\u2032\nS2) by using equations (10) and (11) with \naS2Cpast as input instead of S2Cpast. For the sample company with a consistent \nincrease in ESR in the 2015\u20132019 period, we (similarly to above) assigned aS2Cpast \nthe average value calculated for the other sample companies that have SBTs that \ncover the same emission scopes (1\u20133, market-based), which was 11.9%.\nContribution of RECs to committed cumulative scope 2 emission reductions. \nBased on the above estimations of SBT\u2032\nS2 for market- and location-based targets, \nand aSBT\u2032\nS2 for market-based targets, we estimated the contribution of RECs to the \ncumulative emission reductions from the base year, B, to the target year, Y, for each \nSBT, again assuming a linear pathway (Fig. 4). First, we calculated the reductions \nin scope 2 emissions (market- or location-based, depending on the reported SBT) \nbetween Y and B:\n\u0394ES2 = ES2,B \u00d7 SBT\u2032\nS2 \u00d7 (Y \u2212B)\n(13)\nFor 33 companies, equation (13) resulted in a \u0394ES2 larger than ES2,B, which \nwould imply negative scope 2 emissions in Y. This is unlikely to occur (the median \nY is 2030) and we therefore assumed that these companies will stop reducing scope \n2 emissions when they reach a value of zero. We therefore corrected the value of \n\u0394ES2 to ES2,B in these cases and corrected Y to the time of zero scope 2 emissions:\nY = B +\n1\nSBT\u2032\nS2\n(14)\nFor each SBT, we then calculated the cumulative reduction of scope 2 emissions \nbetween B and Y:\ncum\u0394ES2 = \u0394ES2 \u00d7 (Y \u2212B) \u00d7 1\n2\n(15)\nFor market-based SBTs, we further calculated the annual scope 2 emission \nreduction between B and Y adjusted to remove the REC contribution (a\u0394ES2) \nby using aSBT\u2032\nS2 instead of SBT\u2032\nS2 as input in equation (13). In cases where \u0394ES2 \nis larger than ES2,B, we downscaled a\u0394ES2 on the basis of the correction of \u0394ES2 \n(described above):\na\u0394ES2 = ES2,B \u00d7 aSBT\u2032\nS2\nSBT\u2032\nS2\n(16)\nWe then calculated the adjusted cumulative emission reductions with the \ncontribution of RECs removed (a.cum\u0394ES2) by using a\u0394ES2 instead of \u0394ES2 as \ninput in equation (15). We then estimated the cumulative contribution from \nRECs to emission reductions (REC.cum\u0394ES2) as the difference between cum\u0394ES2 \nand a.cum\u0394ES2. Finally, we aggregated REC.cum\u0394ES2 for all market-based SBTs \nand compared it with the aggregated cum\u0394ES2 for all market-based SBTs and the \naggregated cum\u0394ES2 for all location-based SBTs.\nNote that five of the sample companies reduced their purchases of RECs \nbetween b and y (the start- and end-years in equation (9)) to such an extent that a \ncontinuation would lead to a projected negative use of RECs in Y or earlier, which \nwould not be possible. For the estimation of REC.cum\u0394ES2 of these five companies, \nwe therefore corrected aSBT\u2032\nS2, assuming that the use of RECs would decrease \nlinearly from B until reaching a value of zero in Y:\ncorrected.aSBT\u2032\nS2 = SBT\u2032\nS2 + (aSBT\u2032\nS2 \u2212SBT\u2032\nS2) \u00d7\nCREC,B\n(CREC, b \u2212CREC, y) \u00d7 Y\u2212B\ny\u2212b\n(17)\nNote that the SBT of one of the five companies has 2011 as B, which is before \nthe Greenhouse Gas Protocol began permitting the use of RECs through the \nmarket-based accounting approach9. To account for the reported use of RECs \nbetween b and y, we used b instead of B in equation (17) for this company. The \ncorrections of aSBT\u2032\nS2 for these five companies had a modest influence on results, \ncausing the aggregated REC.cum\u0394ES2 (illustrated in Fig. 4) to change from \n97\u2009MtCO2e to 101\u2009MtCO2e.\nData availability\nThis study is based on preexisting datasets, primarily references7,37,40. The data \nbehind all figures are available in the Supplementary Spreadsheet.\nCode availability\nNo custom code was developed for this study. All equations are given in Methods \nand we used Microsoft Excel for the analysis.\nReferences\n\t37.\tBloomberg Database (Bloomberg, 2021); https://www.bloomberg.com/\nprofessional/solution/bloomberg-terminal/\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\n\t38.\tSBTi Corporate Net-Zero Standard v.1.0 (SBTi, 2021); https://\nsciencebasedtargets.org/resources/files/Net-Zero-Standard.pdf\n\t39.\tISO 14064-2:2019. Greenhouse Gases\u2014Part 2: Specification with Guidance at \nthe Project Level for Quantification, Monitoring and Reporting of Greenhouse \nGas Emission Reductions or Removal Enhancements (International \nOrganization for Standardization, 2019); https://www.iso.org/standard/ \n66454.html\n\t40.\tCDP 2010\u20132020 Annual Questionnaire\u2014Investor and Supply Chain Version. \nAdditional Online Lookups in 2021 Annual Questionnaire (CDP, 2021);  \nhttps://www.cdp.net/en#a8888e63070314c2285625253a462815\n\t41.\tFoundations of Science-based Target Setting v.1.0 (SBTi, 2019); https://\nsciencebasedtargets.org/resources/files/foundations-of-SBT-setting.pdf\n\t42.\tHuppmann, D. et al. Scenario Analysis Notebooks for the IPCC Special Report \non Global Warming of 1.5\u2009\u00b0C (International Institute for Applied Systems \nAnalysis, 2018); https://doi.org/10.22022/SR15/08-2018.15428\n\t43.\tWang, D. D. & Sueyoshi, T. Climate change mitigation targets set by global \nfirms: overview and implications for renewable energy. Renew. Sustain. Energy \nRev. 94, 386\u2013398 (2018).\nAcknowledgements\nWe are very thankful for the valuable feedback of the three anonymous reviewers. This \nresearch was funded by the Natural Sciences and Engineering Research Council of \nCanada (NSERC) Discovery Grants Program (grant no. RGPIN/6956-2017 to S.L. and \nRGPIN-2017-04159 to H.D.M.), Concordia University Research Chair funding (H.D.M.) \nand the Concordia University Horizon Fellows Program (A.B.).\nAuthor contributions\nA.B. conceived the study idea. A.B. developed the study design with contributions \nfrom M.B., S.L. and H.D.M. M.B. synthesized the literature on market-based scope 2 \nemission accounting. A.B. performed the data analysis and produced the figures with \ncontributions from S.L. H.D.M. assisted with framing the manuscript. A.B. drafted the \nmanuscript with contributions from H.D.M., S.L. and M.B. All authors contributed to \nmanuscript editing and revisions.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary material \navailable at https://doi.org/10.1038/s41558-022-01379-5.\nCorrespondence and requests for materials should be addressed to Anders Bj\u00f8rn.\nPeer review information Nature Climate Change thanks Jing Meng and the other, \nanonymous, reviewer(s) for their contribution to the peer review of this work.\nReprints and permissions information is available at www.nature.com/reprints.\nNature Climate Change | www.nature.com/natureclimatechange\n\n\n    Scientific Research Findings:", "answer": "We analysed the emissions reporting of 115 companies with approved science-based targets. Eighty-nine per cent of the companies claimed emission reductions through the use of renewable energy certificates. In combination, the companies reported a 31% reduction in emissions related to purchased energy during the 2015\u20132019 period, which would align with the 1.5 \u00b0C goal of the Paris Agreement. However, two-thirds of this claimed emission reduction was due to the use of renewable energy certificates that were unlikely to have resulted in actual emission reductions. Hence, the real combined emission reduction was closer to 10%, which is only barely aligned with the well-below 2 \u00b0C goal. Our finding that ineffective renewable energy certificates have led to substantially overstated emission reductions probably also holds for companies that have made other ambitious emission reduction commitments, such as net-zero targets.", "id": 54}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41558-021-01268-3\n1Department of Political Science, University of California, Santa Barbara, CA, USA. 2Department of Political Science, Universit\u00e9 de Montr\u00e9al, Montr\u00e9al, \nQu\u00e9bec, Canada. 3Department of Political Science, University of British Columbia, Vancouver, British Columbia, Canada. 4Institute of Political Science and \nOeschger Center for Climate Change Research, University of Bern, Bern, Switzerland. \u2709e-mail: mildenberger@ucsb.edu\nI\nn recent years, over 3,500 US economists, including 27 Nobel prize \nwinners1, and 1,600 European economists2 have signed public \nstatements advocating carbon taxation. As a policy tool, carbon \ntaxes offer theoretical cost effectiveness, price stability and adminis-\ntrative simplicity. However, these potential economic advantages can \ncome with a steep political price. Carbon taxes have been rejected \nin referenda and elections3\u20137, been reversed after political back-\nlash8,9, been opposed by a substantial proportion of the public10\u201312  \nand generated political controversy whenever debated across \nadvanced democracies9,11,13\u201316. Scholars have identified diverse bar-\nriers to public acceptance of carbon taxes, including perceptions \nthat the policy will not reduce emissions, that it is too costly, that it is \nregressive and that it might undermine economic prosperity10,11,17\u201320.\nIn response to this opposition, interest has grown in deploying \ncarbon tax revenues to boost public support. Carbon tax revenues \ncan be directed towards environmental spending4,6,7,17,21\u201330 or can be \npaired with tax cuts, although this latter approach may be less effec-\ntive than green spending at reducing public opposition23,24,28,30,31 or \nonly effective for some voters22,32,33. However, earmarking through \neither spending or tax cuts may lack high visibility6,34, and voters  \nmay distrust that governments will deliver or maintain these \nbenefits23,33,35,36. Instead, it has been suggested that highly visible \nlump-sum rebates or \u2018dividends\u2019 could be more effective in winning \npublic support and reinforcing that support over time as beneficia-\nries become accustomed to regular dividends37. The hypothetical \npotential of climate rebates to increase public carbon pricing support \nhas been shown in the United States4,17,20,21,29, Canada38, Norway35, \nSwitzerland39, the United Kingdom4,30, Australia4, Germany17, \nTurkey27, France40 and India4. These studies offer strong reasons \nto expect that bundling carbon taxes with lump-sum rebates could \nincrease public acceptance.\nHowever, other studies offer more sobering assessment. Surveys \nmeasuring public support for hypothetical policies may overestimate \nvoters\u2019 support when confronted with real-world costs. For example,  \npolling overestimated voter support for failed carbon tax refe\u00ad\nrenda in Washington state5 and Switzerland39. Voters\u2019 perceptions  \nof rebate costs and benefits may also be shaped by competing par-\ntisan and interest-group narratives5,40,41. More broadly, voters are \noften unaware they receive even high-profile government benefits42.\nIn light of these competing perspectives, the political impact of \nclimate rebates requires testing in the context of real-world policies,  \nas they are implemented. Two such policies currently exist, in \nCanada and Switzerland. In Canada, the federal government \nimposed a carbon tax and rebate scheme for households as part \nof its national strategy for pricing carbon in 2019, which currently \napplies in four of the country\u2019s ten provinces (covering over half of \nthe Canadian population). The tax was initially set at 20 Canadian \ndollars (Can$20) per tonne, rising to Can$50 per tonne by 2022. \nIn December 2020, the government announced a revised schedule \nincreasing to Can$170 per tonne by 2030. The associated rebate, \ncalled the Climate Action Incentive Payment, is based on the number  \nof adults and children in each household, with a 10% increase \nfor rural households. It is delivered through an income tax credit  \nto one adult per household. All tax revenues are returned to the \nprovince of origin. Because provincial emissions per capita vary \nwidely, rebate sizes also vary by province. For example, the average  \ndividend in Saskatchewan is almost double that in Ontario. The \npolicy is highly progressive, with 80% of households receiving  \nmore in dividends than they pay in carbon taxes43. Supplementary \nSection 1 details Canadian federal and provincial carbon pricing.\nBy contrast, Switzerland established its climate rebate pro-\ngramme in 2008 as part of an escalating carbon tax that reached 96 \nSwiss francs (CHF96) per tonne by 2018. In Switzerland, roughly \ntwo-thirds of revenue is redistributed to businesses and the pub-\nlic. Public rebates are given on a per capita basis, with every person \n(including children) receiving an equal amount. Citizens receive \ntheir rebates as a discount on their health insurance premiums, \nwith annual notifications about this monthly benefit through health \ninsurance forms. In a June 2021 referendum, Swiss voters narrowly \nrejected a climate law that would have increased the carbon tax level \nand associated rebate amounts. Supplementary Section 2 details the \nSwiss policy context.\nLimited impacts of carbon tax rebate programmes \non public support for carbon pricing\nMatto Mildenberger\u200a \u200a1\u2009\u2709, Erick Lachapelle\u200a \u200a2, Kathryn Harrison3 and Isabelle Stadelmann-Steffen\u200a \u200a4\nRevenue recycling through lump-sum dividends may help mitigate public opposition to carbon taxes, yet evidence from real-world \npolicies is lacking. Here we use survey data from Canada and Switzerland, the only countries with climate rebate programmes, \nto show low public awareness and substantial underestimation of climate rebate amounts in both countries. Information was \nobtained using a five-wave panel survey that tracked public attitudes before, during and after implementation of Canada\u2019s 2019 \ncarbon tax and dividend policy and a large-scale survey of Swiss residents. Experimental provision of individualized information \nabout true rebate amounts had modest impacts on public support in Switzerland but potentially deleterious effects on support \nin Canada, especially among Conservative voters. In both countries, we find that perceptions of climate rebates are structured \nless by informed assessments of economic interest than by partisan identities. These results suggest limited effects of existing \nrebate programmes, to date, in reshaping the politics of carbon taxation.\nNature Climate Change | VOL 12 | February 2022 | 141\u2013147 | www.nature.com/natureclimatechange\n141\n\nArticles\nNaTuRE ClImaTE CHangE\nWe evaluate the effect of real-world dividends on public sup-\nport for carbon taxes in both countries. In Canada, we report an \noriginal five-wave panel survey of Canadians from February 2019 to \nMay 2020. Our sample included residents from five provinces, two \nsubject to the federal carbon tax (Saskatchewan and Ontario), one \nwith provincial emissions trading (Quebec) and two with provin-\ncial carbon taxes (British Columbia and Alberta). Alberta repealed \nits provincial tax midway through our study, prompting introduc-\ntion of a federal-tax and rebate scheme in 2020. We surveyed the \nsame respondents after the federal carbon tax was announced but \nbefore it was implemented (February 2019), soon after implementa-\ntion (April 2019), after residents of federal-tax provinces received \ntheir rebates (July 2019), after a federal election in which carbon \npricing was a prominent issue (November 2019) and one year after \npolicy implementation (May 2020). In wave 4, we embedded a sur-\nvey experiment exposing respondents to individualized informa-\ntion about their actual rebates. In Switzerland, we fielded a survey \nof 1,050 Swiss residents in December 2019, which also included an \nembedded survey experiment where respondents were exposed to \ninformation on their actual policy rebates. We provide full details \non Canadian and Swiss surveys in Methods.\nWe begin with data from our longitudinal Canadian survey,  \nvisualizing public support for carbon pricing by province over time \n(Fig. 1). Canadian carbon tax support remained relatively stable \nacross our panel. In the rebate province of Ontario, public support \nby wave 5 was within two percentage points of wave 1. However, \nbetween waves 1 and 4, carbon pricing support did increase in the \nrebate province of Saskatchewan, before declining through wave 5 \nfor a net gain of five percentage points to 32%. The final wave fol-\nlowed the onset of the COVID-19 pandemic, although COVID inci-\ndence in rebate provinces was still low; for example, in Saskatchewan, \nthere had been a total of only 176 cases and 2 COVID-related  \ndeaths in advance of wave 5 (ref. 44). Federal COVID-related financial  \nassistance was already available to respondents by this time. Trends \nin carbon pricing opposition are similar (Extended Data Fig. 1 and \nSupplementary Section 4). We also conduct exploratory statistical \nanalysis comparing trends in rebate versus non-rebate provinces \n(Supplementary Section 5).\nThese provincial averages mask strong partisan differences in \ncarbon pricing support (Fig. 2). Policy support was concentrated \namong Liberal Party of Canada supporters (the party that imple-\nmented the policy) versus Conservative Party of Canada support-\ners (the opposition party that strongly opposed it) and remained \nstable through time. Conservative opposition persisted in both \nfederal-tax (rebate) and provincial pricing (non-rebate) provinces. \nBy wave 5, 75% and 81% of Liberal Party supporters in Ontario and \nSaskatchewan, respectively, supported carbon pricing, compared \nwith 32% and 13% of Conservative Party supporters in these same \nprovinces. Partisan splits across rebate and non-rebate provinces \nshow similar trends (Extended Data Fig. 2 and Supplementary \nSection 6). These partisan differences persist even after condition-\ning on respondents\u2019 individual cost exposure (Extended Data Fig. 3 \nand Supplementary Section 7).\nFor rebate policies to offer political benefits to incumbent gov-\nernments, the public must perceive those benefits42,45. We test public \nknowledge about existing rebate programmes in both Canada and \nSwitzerland. We first test Canadian respondents\u2019 specific knowl-\nedge about their rebates. In wave 3, immediately after residents of \nOntario and Saskatchewan received their rebates, we asked respon-\ndents whether they had received a climate-related benefit as part of \ntheir federal income tax returns (Supplementary Section 8). Many \nCanadians did not know, including 17% in rebate provinces and \nbetween 33% and 36% in non-rebate provinces. In Ontario, only \n0\n25\n50\n75\nPublic support for carbon pricing (%)\nWave 1\n(February)\nWave 2\n(April)\nWave 3\n(July)\nWave\nWave 4\n(November)\nWave 5\n(May)\nBritish Columbia\nAlberta\nSaskatchewan\nOntario\nQu\u00e9bec\nFig. 1 | Support for carbon pricing by province across waves. The dotted \nline indicates when the federal carbon tax policy came into effect. The \nsolid line indicates the approximate period during which households \nreceived their climate rebates. The dashed line indicates the timing of a \nfederal election in which the carbon tax was highly salient. Respondents \nin Saskatchewan and Ontario received climate rebates. Data from \nrespondents who completed all five waves (n\u2009=\u2009899). Error bars depict \n95% confidence intervals. Supplementary Section 3 reproduces this figure \nfor the first four waves only (n\u2009=\u20091,190), finding identical trends across this \nexpanded sample.\n0\n25\nPublic support for carbon pricing (%)\n50\n75\n100\nWave 1\n(February)\nWave 2\n(April)\nWave 3\n(July)\nWave\nWave 4\n(November)\nWave 5\n(May)\nLiberal (Qu\u00e9bec and British Columbia)\nLiberal (Saskatchewan and Ontario)\nConservative (Qu\u00e9bec and British Columbia)\nConservative (Saskatchewan and Ontario)\nFig. 2 | Support for carbon pricing among Liberal Party and Conservative \nParty voters by rebate versus non-rebate province. The dotted line \nindicates when the policy came into effect. The solid line indicates the \napproximate period during which households received their climate \nrebates. The dashed line indicates the timing of a federal election in \nwhich the carbon tax was highly salient. Voters are classified according \nto their wave 1 (pre-policy implementation) party preferences. Albertan \nrespondents are excluded because Alberta switched from being a \nnon-rebate province to being a rebate province between waves 1 and 5. \nError bars depict 95% confidence intervals.\nNature Climate Change | VOL 12 | February 2022 | 141\u2013147 | www.nature.com/natureclimatechange\n142\n\nArticles\nNaTuRE ClImaTE CHangE\n55% of residents correctly believed they had received a rebate, while \nSaskatchewan residents were more aware (75%). By contrast, about \n11% and 13% of individuals in the non-rebate provinces of Alberta \nand British Columbia incorrectly reported rebate receipt.\nWe then asked respondents to estimate the size of any rebate  \nthey believed their household had received (Table 1). We compare \nperceived amounts to the true average rebate for our survey respon-\ndents (see Methods for details). Residents in non-rebate provinces \nnonetheless estimated a positive average rebate amount, a misper-\nception that continued after the fall 2019 election (Supplementary \nSection 9). In rebate provinces, our survey averages reflect a 40% \nunderestimation in Saskatchewan and 32% underestimation in \nOntario of true rebate amounts. Limiting our analysis to respon-\ndents who correctly believed they had received a rebate, the \nOntario average estimate was CDN$198 (standard error (s.e.) $13), \nonly a 9% underestimation, and the Saskatchewan average esti-\nmate was CDN$315 (s.e. $13), a 29% underestimation. Still, only \n24% of Ontario respondents and 19% of Saskatchewan respon-\ndents estimated a rebate amount falling within CDN$100 of their  \ntrue rebate (Extended Data Fig. 4 and Supplementary Section 10). \nThese misperceptions are associated with party preference. In both \nprovinces, respondents who consistently indicated they would vote \nfor the anti-carbon tax Conservative Party systematically estimated \nlower rebate amounts (Supplementary Section 10). We also find \npersistent confusion among respondents as to whether the pro-\nvincial or federal government is responsible for carbon pricing in \ntheir province, with some learning across the panel (Supplementary \nSection 11).\nWe conduct a similar analysis in Switzerland. Consistent with \nprevious surveys7,46, we find limited knowledge of the Swiss rebate \nTable 1 | Average estimated and true rebate sizes for sample, by \nprovince\nProvince\nAverage perceived \nrebate (CDN$)\nTrue average \nrebate (CDN$)\nReceived federal rebate\nSaskatchewan\n268 (13)\n444\nOntario\n149 (11)\n217\nDid not receive federal \nrebate\nBritish Columbia\n63 (9)\n0\nAlberta\n83 (9)\n0\nQu\u00e9bec\n54 (10)\n0\nStandard errors in parentheses. See Methods for details on calculating true average rebate.\n1.00\na\nb\nc\nd\n0.75\n0.50\nFraction of survey respondents\n0.25\n0\n1.00\n0.75\n0.50\n0.25\n0\nYes\nCantonal\ntax\nNone\nLess than\nCHF3\nCHF\n3\u201310\nCHF\n10\u201315\nDo not\nknow\nNational\ntax\nHealth\nbill\nPension\ncontribution\nElectricity\nbills\nDo not\nknow\nNo\nResponse\nResponse\nResponse\nResponse\nDo not know\nPublic\nrebates\nRenewable\nenergy\nGov.\nspending\nDo not know\n1.00\n0.75\n0.50\nFraction of survey respondents\n0.25\n0\n1.00\n0.75\n0.50\nFraction of survey respondents\nFraction of survey respondents\n0.25\n0\nFig. 3 | Knowledge about the Swiss scheme among December 2019 survey respondents. a\u2013d, Responses to survey questions: belief that Switzerland has \na carbon tax on fossil fuels (a), belief regarding what most carbon tax revenue is directed towards (b), belief regarding what tax revenue is redistributed to \nreduce (c), perceived monthly rebate size (d). Correct choices are highlighted in green.\nNature Climate Change | VOL 12 | February 2022 | 141\u2013147 | www.nature.com/natureclimatechange\n143\n\nArticles\nNaTuRE ClImaTE CHangE\npolicy. Although the policy has been in place for over ten years, only \n12% of respondents knew tax revenues were redistributed to the \npublic, and 85% did not know they received a health bill discount  \nassociated with the country\u2019s carbon tax (Fig. 3). Every Swiss  \nresident receives CHF5.35 per month (in 2019) as their rebate, but \nonly 13% of respondents knew (or correctly guessed) the monthly \nrebate was between CHF3 and CHF10.\nLow public awareness of rebates in Canada and Switzerland may \nstem from the indirect mechanisms through which governments \nin both countries redistribute their climate dividends. Using two \nsurvey experiments, we assess whether increasing rebate aware-\nness through individualized rebate information increases support \nfor existing and future carbon taxes. Here, low existing public \nknowledge allows us to randomize information about government \nbenefits that respondents already receive, providing a second-best \napproximation for experimental manipulation of rebate receipt \nitself. However, our experiments ultimately identify the effect of \ninformation about rebates on public support, not the direct effect of \nrebates themselves. These experiments also focus on testing infor-\nmation about policy benefits, rather than policy costs.\nIn Canada, half of wave 4 survey respondents from Ontario and \nSaskatchewan (n\u2009=\u2009605) were randomly assigned a custom mock-up \nof their own tax return with their true climate dividend promi-\nnently displayed (Supplementary Section 12 describes treatment; \nMethods describes calculation details; Supplementary Section 13 \nshows experimental balance.) Receiving treatment led respondents \nto increase perceptions of their household\u2019s rebate size, suggesting \nat least partial updating in the treatment group (Supplementary \nSection 14). However, treatment did not change carbon pricing sup-\nport (Fig. 4a: Difference-in-Means (DIM)\u2009=\u2009\u20130.0342, s.e.\u2009=\u20090.106, \nP\u2009=\u20090.747). Instead, information about their true benefit decreased \nrespondents\u2019 belief that the rebates were sufficient to cover their tax \nexposure (Fig. 4b: DIM\u2009=\u2009\u20130.136, s.e.\u2009=\u20090.0662, P\u2009=\u20090.0398). As such, \nCanadians who learned the true value of their rebates were signifi-\ncantly more likely to perceive themselves as net losers even though \nmost Canadians are net beneficiaries. This shift was concentrated \namong Conservative Party of Canada supporters (DIM\u2009=\u2009\u20130.213, \ns.e.\u2009=\u20090.102, P\u2009=\u20090.0391).\nIn our December 2019 Switzerland survey, half of respondents \nwere randomly assigned an encouragement treatment to leave  \ntheir computers mid-survey and retrieve their health insurance \nforms; respondents were then asked to report the size of their  \nbenefit. All treated respondents were then shown a sample health \nform with benefit size highlighted (Supplementary Section 15  \nprovides example), irrespective of whether they reported having  \nfound their personal form (Supplementary Section 16 shows  \nexperimental balance). Unlike in Canada, we find personal rebate  \ninformation increased support for the current scheme on a \nfour-point scale by around one-fifth of a standard deviation \n(DIM\u2009=\u20090.18885, s.e.\u2009=\u20090.06155, P\u2009<\u20090.01; Fig. 5a). These results hold \non both the right and left sides of the political spectrum but not \nfor centre-party supporters. However, treatment had no effect on \nsupport for either small (equivalent to CHF0.03 per litre increase \nin heating oil costs; DIM\u2009=\u20090.06213, s.e.\u2009=\u20090.09744, P\u2009=\u20090.524) or \nlarge (equivalent to CHF0.15 per litre increase in heating oil costs; \nDIM\u2009=\u20090.11182, s.e.\u2009=\u20090.09396, P\u2009=\u20090.235) increases in the Swiss  \ncarbon tax rate.\nBeyond low visibility, we also consider alternative reasons for \nthe weak effects of rebates on public opinion. In Canada, carbon \npricing preferences might have remained relatively stable despite \nrebates because the political benefits of revenue recycling came with \npolicy announcement (before our wave 1), not during implemen-\ntation (our panel period). Two pieces of evidence suggest this as \nunlikely. First, we find little baseline knowledge about the rebate \nin wave 1, which we would expect if anticipation of future rebates \nhad already increased support (Supplementary Section 11). Second, \nthe announcement of a federal rebate policy for Alberta occurred \nbetween waves 2 and 3, after a newly elected provincial government \nrepealed the provincial tax, which did not provide universal rebates. \nThis prompted the federal government to step in to announce it \nwould impose a tax and rebate policy over the objection of the pro-\nvincial government (as in Saskatchewan and Ontario.) However, \nwe find no announcement effect in Alberta, where carbon pricing \nsupport trends roughly in parallel with other provinces after policy \nannouncement (Fig. 1).\n4\na\nb\n3\n2\nGroup\nAll\nConservative\nLiberal\nNDP\nSupport for existing policy (4-point scale)\nBelief receive more than pay (scale from \u20131 to 1)\n1\n1.0\n0.5\n0\n\u20130.5\n\u20131.0\nControl\nTreatment\nTreatment status\nControl\nTreatment\nTreatment status\nGroup\nAll\nConservative\nLiberal\nNDP\nFig. 4 | The effect of rebate information on carbon pricing support in \nCanada. a,b, Exposure to individualized information about a respondent\u2019s \ntrue climate rebate amount in Canada did not shape carbon pricing support \n(a) but instead generated a backlash by making respondents believe they \npaid more in tax than they received as their rebate (b). Full sample is in \nblack, with subgroups defined by wave 4 party preference. Error bars depict \n95% confidence intervals. NDP, New Democratic Party.\nNature Climate Change | VOL 12 | February 2022 | 141\u2013147 | www.nature.com/natureclimatechange\n144\n\nArticles\nNaTuRE ClImaTE CHangE\nAnother possibility is that policy preferences remain condi\u00ad\ntioned primarily by partisanship. We find that Conservative \nParty supporters are more likely than Liberal Party supporters to \nacknowledge having seen negative ads about carbon pricing and \nto report that these ads made them less supportive of this policy \n(Supplementary Section 18). Similarly, respondents who report \nhaving voted for the Conservative Party in the Fall 2019 election  \nwere more likely to underestimate their rebates, even when exposed \nto information about their true rebate amount in our survey \nexperiment (Supplementary Section 19). More broadly, in the two \nfederal-tax provinces, supporters of the Liberal Party of Canada \nwere three to eight times more likely to support the carbon tax \nthan were Conservative Party supporters. Similarly, in Switzerland, \nleft-leaning voters were 48% more likely to support rebates relative \nto right-leaning voters. In short, partisanship does structure both \ncarbon tax preferences and patterns of rebate responsiveness.\nFinally, our Canadian results might be a function of survey design \neffects. However, we find no such effects using independent samples  \nof provincial respondents across the survey\u2019s first four waves \n(Supplementary Section 20). Accordingly, response consistency in \npanel surveys is unlikely to account for weak rebate effects47.\nOverall, our results speak to growing interest in recycling carbon \ntax revenues in the form of lump-sum rebates to mitigate persistent \npublic opposition to carbon taxes. We explore existing policies, as \nimplemented, in Canada and Switzerland using a new longitudinal  \nopinion panel as well as two survey experiments. We find only  \nlimited evidence that these existing policies have reshaped the  \npolitics of carbon pricing to date. Members of the public in both \ncountries remain ill-informed about the rebates they are already \nreceiving and systematically underestimate their size. These low \nlevels of awareness may stem from rebates delivered via a credit \nagainst a (tax or insurance) bill rather than a more-visible check in \nthe mail and, in the case of Canada, a highly politicized communi-\ncation environment. Still, experimental provision about individual \nrebate size only modestly increased support for the current policy in \nSwitzerland and did not increase support for even a small tax increase. \nIn Canada, information about rebate size did not increase policy  \nsupport, but instead led Conservative Party respondents to believe \nthe policy imposed net costs on their household.\nThese findings imply that one-time information does not sub-\nstantially affect policy support. While results from non-climate \ndomains may not extrapolate to carbon taxation19, previous studies  \nsuggest learning over time built public support for conges-\ntion taxes48\u201350 and solid-waste charges51. Yet public ignorance of  \ndividends has persisted for more than a decade in Switzerland, \nand our Canadian panel covered a period in which the carbon tax  \nwas highly salient by virtue of the policy\u2019s implementation, court \nchallenges, federal\u2013provincial conflict and partisan debate during \na federal election, a most likely case for public learning about the \nrebate scheme.\na\nb\nc\n4\n3\n2\n1\n3\n4\n2\nSupport for existing policy (4-point scale)\nSupport for small carbon tax increase\n(4-point scale)\n4\n3\n2\n1\nSupport for large carbon tax increase\n(4-point scale)\n1\nControl\nTreatment\nTreatment status\nControl\nTreatment\nTreatment status\nControl\nTreatment\nTreatment status\nGroup\nAll\nRight\nCentre\nLeft\nGroup\nAll\nRight\nCentre\nLeft\nGroup\nAll\nRight\nCentre\nLeft\nFig. 5 | The effect of rebate information on carbon pricing support in Switzerland. a\u2013c, Exposure to individualized information about a respondent\u2019s \ntrue climate rebate amount in Switzerland increased support for the existing policy (a) but not support for either small (b) or large (c) future carbon tax \nincreases. Full sample in black, with subgroups based on ideological position of preferred political party (Supplementary Section 17 provides classification \ndetails). Error bars depict 95% confidence intervals.\nNature Climate Change | VOL 12 | February 2022 | 141\u2013147 | www.nature.com/natureclimatechange\n145\n\nArticles\nNaTuRE ClImaTE CHangE\nAltogether, results from these studies paint a more complex \npicture of the benefits of lump-sum carbon tax rebates than did \nprevious surveys and laboratory experiments using hypothetical \npolicies. While the climate-dividend policies in Switzerland and \nCanada diverge from policy ideals, trading off public transparency \nfor administrative efficiency, we note that these are the only two \nextant examples of carbon tax and dividend globally. Both were \nimplemented in the context of partisan and interest-group debates, \nincluding widespread dissemination of selective or misleading \ninformation. As always, both policy design and attitudinal change \nmay still occur. The government of Canada has announced that \nfuture rebates, which will steadily increase in value, will be delivered  \nto households directly. However, in Switzerland, voters rejected \nan increase in the country\u2019s carbon tax rate, alongside increased \nrebates, in June 2021 when faced with intense politicization of \npolicy costs by opponents. The evolution and impact of new rebate \ndesigns, increasing tax rates and benefit sizes, and potential shifts in \npartisan positions remain for future research.\nOnline content\nAny methods, additional references, Nature Research report-\ning summaries, source data, extended data, supplementary infor-\nmation, acknowledgements, peer review information; details of \nauthor contributions and competing interests; and statements of \ndata and code availability are available at https://doi.org/10.1038/\ns41558-021-01268-3.\nReceived: 29 October 2020; Accepted: 15 December 2021;  \nPublished online: 24 January 2022\nReferences\n\t1.\t Economists\u2019 Statement on Carbon Dividends (Climate Leadership Council, \n2019).\n\t2.\t Economists\u2019 Statement on Carbon Dividends (European Association of \nEnvironment and Resource Economists, 2019).\n\t3.\t Harrison, K. A tale of two taxes: the fate of environmental tax reform in \nCanada. Rev. Policy Res. 29, 383\u2013407 (2012).\n\t4.\t Carattini, S., Kallbekken, S. & Orlov, A. How to win public support for a \nglobal carbon tax. Nature 565, 289\u2013291 (2019).\n\t5.\t Anderson, S., Marinescu, I. & Shor, B. Can Pigou at the polls stop us melting \nat the poles? National Bureau of Economic Research (NBER) Working Paper \nNo. 26146 (2019).\n\t6.\t Thalmann, P. The public acceptance of green taxes: 2 million voters express \ntheir opinion. Public Choice 119, 179\u2013217 (2004).\n\t7.\t Baranzini, A. & Carattini, S. Effectiveness, earmarking and labeling: testing \nthe acceptability of carbon taxes with survey data. Environ. Econ. Policy Stud. \n19, 197\u2013227 (2017).\n\t8.\t S\u00e9nit, C.-A. The Politics of Carbon Taxation in France: Preferences, Institutions, \nand Ideologies (Institut du D\u00e9veloppement Durable et des Relations \nInternationales, 2012).\n\t9.\t Mildenberger, M. Carbon Captured: How Labor and Business Control Climate \nPolitics (MIT Press, 2020).\n\t10.\tRhodes, E., Axsen, J. & Jaccard, M. Exploring citizen support for different \ntypes of climate policy. Ecol. Econ. 137, 56\u201369 (2017).\n\t11.\tStadelmann-Steffen, I. & Dermont, C. The unpopularity of incentive-based \ninstruments: what improves the cost\u2013benefit ratio? Public Choice 175, 37\u201362 \n(2018).\n\t12.\tLevi, S. Why hate carbon taxes? Machine learning evidence on the roles of \npersonal responsibility, trust, revenue recycling, and other factors across 23 \nEuropean countries. Energy Res. Soc. Sci. 73, 101883 (2021).\n\t13.\tRabe, B. G. Can We Price Carbon? (MIT Press, 2018).\n\t14.\tHarrison, K. The comparative politics of carbon taxation. Annu. Rev. Law  \nSoc. Sci. 6, 507\u2013529 (2010).\n\t15.\tDermont, C. & Stadelmann-Steffen, I. The role of policy and party \ninformation in direct-democratic campaigns. Int. J. Public Opin. Res. 32, \n442\u2013466 (2020).\n\t16.\tLachapelle, E. & Kiss, S. Opposition to carbon pricing and right-wing \npopulism: Ontario\u2019s 2018 general election. Environ. Politics 28, 970\u2013976 \n(2019).\n\t17.\tBeiser-McGrath, L. F. & Bernauer, T. Could revenue recycling make effective \ncarbon taxation politically feasible? Sci. Adv. 5, eaax3323 (2019).\n\t18.\tKirchg\u00e4ssner, G. & Schneider, F. On the political economy of environmental \npolicy. Public Choice 115, 369\u2013396 (2003).\n\t19.\tCarattini, S., Carvalho, M. & Fankhauser, S. Overcoming public resistance to \ncarbon taxes. Wiley Interdiscip. Rev. Clim. Change 9, e531 (2018).\n\t20.\tNowlin, M. C., Gupta, K. & Ripberger, J. T. Revenue use and public support \nfor a carbon tax. Environ. Res. Lett. 15, 084032 (2020).\n\t21.\tAmdur, D., Rabe, B. G. & Borick, C. P. Public views on a carbon tax depend \non the proposed use of revenue. Issues Energy Environ. Policy 13, 1\u20139 (2014).\n\t22.\tDol\u0161ak, N., Adolph, C. & Prakash, A. Policy design and public support for \ncarbon tax: evidence from a 2018 US national online survey experiment. \nPublic Adm. 98, 905\u2013921 (2020).\n\t23.\tBachus, K., Van Ootegem, L. & Verhofstadt, E. No taxation without \nhypothecation: towards an improved understanding of the acceptability of an \nenvironmental tax reform. J. Environ. Policy Plan. 21, 321\u2013332 (2019).\n\t24.\tHsu, S.-L., Walters, J. & Purgas, A. Pollution tax heuristics: an empirical  \nstudy of willingness to pay higher gasoline taxes. Energy Policy 36,  \n3612\u20133619 (2008).\n\t25.\tKallbekken, S. & Aasen, M. The demand for earmarking: results from a focus \ngroup study. Ecol. Econ. 69, 2183\u20132190 (2010).\n\t26.\tKallbekken, S. & S\u00e6len, H. Public acceptance for environmental taxes: \nself-interest, environmental and distributional concerns. Energy Policy 39, \n2966\u20132973 (2011).\n\t27.\tGevrek, Z. E. & Uyduranoglu, A. Public preferences for carbon tax attributes. \nEcol. Econ. 118, 186\u2013197 (2015).\n\t28.\tRotaris, L. & Danielis, R. The willingness to pay for a carbon tax in Italy. \nTransp. Res. D 67, 659\u2013673 (2019).\n\t29.\tKotchen, M., Turk, Z. & Leiserowitz, A. Public willingness to pay for a US \ncarbon tax and preferences for spending the revenue. Environ. Res. Lett. 12, \n094012 (2017).\n\t30.\tBristow, A. L., Wardman, M., Zanni, A. M. & Chintakayala, P. K. Public \nacceptability of personal carbon trading and carbon tax. Ecol. Econ. 69, \n1824\u20131837 (2010).\n\t31.\tLachapelle, E., Borick, C. P. & Rabe, B. Public attitudes toward climate \nscience and climate policy in federal systems: Canada and the United States \ncompared. Rev. Policy Res. 29, 334 (2012).\n\t32.\tJagers, S. C., Martinsson, J. & Matti, S. The impact of compensatory measures \non public support for carbon taxation: an experimental study in Sweden. \nClim. Policy 19, 147\u2013160 (2019).\n\t33.\tFairbrother, M. When will people pay to pollute? Environmental taxes, \npolitical trust and experimental evidence from Britain. Br. J. Polit. Sci. 49, \n661\u2013682 (2017).\n\t34.\tBeuermann, C. & Santarius, T. Ecological tax reform in Germany: handling \ntwo hot potatoes at the same time. Energy Policy 34, 917\u2013929 (2006).\n\t35.\tKallbekken, S., Kroll, S. & Cherry, T. L. Do you not like Pigou, or do you not \nunderstand him? Tax aversion and revenue recycling in the lab. J. Environ. \nEcon. Manage. 62, 53\u201364 (2011).\n\t36.\tCherry, T. L., Garc\u00eda, J. H., Kallbekken, S. & Torvanger, A. The development \nand deployment of low-carbon energy technologies: the role of economic \ninterests and cultural worldviews on public support. Energy Policy 68, \n562\u2013566 (2014).\n\t37.\tKlenert, D. et al. Making carbon pricing work for citizens. Nat. Clim. Change \n8, 669\u2013677 (2018).\n\t38.\tJagers, S., Lachapelle, E., Martinsson, J. & Matti, J. Bridging the ideological \ngap? How fairness perceptions mediate the effect of revenue recycling on \npublic support for carbon taxes in the United States, Canada and Germany. \nRev. Policy Res. https://doi.org/10.1111/ropr.12439 (2021).\n\t39.\tCarattini, S., Baranzini, A., Thalmann, P., Varone, F. & V\u00f6hringer, F. Green \ntaxes in a post-Paris world: are millions of nays inevitable? Environ. Resour. \nEcon. 68, 97\u2013128 (2017).\n\t40.\tDouenne, T. & Fabre, A. Yellow vests, pessimistic beliefs, and carbon tax \naversion. Am. Econ. J. Econ. Policy (in the press). https://www.aeaweb.org/\narticles?id=10.1257/pol.20200092&&from=f\n\t41.\tDavidovic, D., Harring, N. & Jagers, S. C. The contingent effects of \nenvironmental concern and ideology: institutional context and people\u2019s \nwillingness to pay environmental taxes. Environ. Politics 29, 674\u2013696 (2020).\n\t42.\tMettler, S. The Submerged State: How Invisible Government Policies Undermine \nAmerican Democracy (Univ. Chicago Press, 2011).\n\t43.\tFiscal and Distributional Analysis of the Federal Carbon Pricing System \n(Parliamentary Budget Office, 2019).\n\t44.\tQuenneville, G. & Hunter, A. Covid-19 in Saskatchewan: province marks \u2019sad \nmilestone\u2019 of first 2 deaths. CBC News (30 March 2020); https://www.cbc.ca/\nnews/canada/saskatoon/coronavirus-saskatchewan-1.5514563\n\t45.\tKumlin, S. & Stadelmann-Steffen, I. How Welfare States Shape the Democratic \nPublic: Policy Feedback, Participation, Voting, and Attitudes (Edward Elgar, \n2014).\n\t46.\tSchwegler, R., Gina, S., Schappi, B. & Iten, R. Klimaschutz und Gr\u00fcne \nWirtschaft - was meint die Bev\u00f6lkerung? Ergebnisse einer repr\u00e4sentativen \nBev\u00f6lkerungsbefragung Technical Report (INFRAS, 2015).\n\t47.\tBergmann, M. & Barth, A. What was I thinking? A theoretical framework for \nanalysing panel conditioning in attitudes and (response) behaviour. Int. J. Soc. \nRes. Methodol. 21, 333\u2013345 (2018).\nNature Climate Change | VOL 12 | February 2022 | 141\u2013147 | www.nature.com/natureclimatechange\n146\n\nArticles\nNaTuRE ClImaTE CHangE\n\t48.\tSchuitema, G., Steg, L. & Forward, S. Explaining differences in acceptability \nbefore and acceptance after the implementation of a congestion charge in \nStockholm. Transp. Res. A 44, 99\u2013109 (2010).\n\t49.\tHensher, D. A. & Li, Z. Referendum voting in road pricing reform: a review \nof the evidence. Transp. Policy 25, 186\u2013197 (2013).\n\t50.\tAndersson, D. & N\u00e4ss\u00e9n, J. The Gothenburg congestion charge scheme: a pre\u2013 \npost analysis of commuting behavior and travel satisfaction. J. Transp. Geogr. \n52, 82\u201389 (2016).\n\t51.\tCarattini, S., Baranzini, A. & Lalive, R. Is taxing waste a waste of time? \nEvidence from a Supreme Court decision. Ecol. Econ. 148, 131\u2013151  \n(2018).\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2022\nNature Climate Change | VOL 12 | February 2022 | 141\u2013147 | www.nature.com/natureclimatechange\n147\n\nArticles\nNaTuRE ClImaTE CHangE\nMethods\nOur paper draws from four new data sources. First, we report a new survey \ndataset, the Canadian Climate Opinion Panel (CCOP). Second, we report a survey \nexperiment embedded in the fourth wave of the CCOP. Third, we report a large-n \nsurvey of Swiss residents conducted in December 2019. Fourth, we report a survey \nexperiment embedded in this Swiss survey. We discuss each dataset and the \nmethods used to analyse it, in turn.\nCCOP. The CCOP was a custom five-wave public opinion panel survey \nadministered online to a sample drawn from the Leger 360 platform. This platform \nis a web-based pool of over 400,000 Canadians, 60% of which were recruited \nrandomly via random-digit dialling. From this pool, an initial sample of 3,313 \npanellists was generated for five Canadian provinces: Alberta (n\u2009=\u2009663), British \nColumbia (n\u2009=\u2009661), Ontario (n\u2009=\u2009660), Qu\u00e9bec (n\u2009=\u2009661) and Saskatchewan \n(n\u2009=\u2009668). These provinces were selected to ensure representation of provinces \nsubject to the federal carbon tax and dividend as well as provinces exempt from the \nfederal carbon tax because of provincial policies deemed equivalent to the federal \ncarbon price. Respondents were remunerated by Leger at a rate of CDN$1 to \nCDN$3 per wave depending on survey length.\nPanellists were invited to participate in the study and answered the first \nquestionnaire between 21 February and 5 March 2019. During this wave, we \nobtained 3,313 completes and a combined American Association of Public Opinion \nResearch (AAPOR) RR3 response rate of 18%. AAPOR RR3 rates incorporate \nan estimate of eligibility among respondents of unknown eligibility into the \nresponse-rate denominator, offering a conservative response-rate estimate52. \nPanellists were subsequently recontacted between 10 and 28 April 2019, after the \nfederal carbon tax policy came into effect on 1 April. During this second wave, \nresponses were received from 2,189 returning panellists from Alberta (n\u2009=\u2009437), \nBritish Columbia (n\u2009=\u2009434), Ontario (n\u2009=\u2009440), Qu\u00e9bec (n\u2009=\u2009439) and Saskatchewan \n(n\u2009=\u2009439). An additional 252 respondents (50 from each province except \nSaskatchewan, where 52 completed) were added to this wave as a check against \npanel experience, resulting in a total sample of 2,441. The combined AAPOR RR3 \nresponse rate for this portion of the fieldwork was 50%. A third invitation went \nout to panellists between 27 June and 19 July 2019, after the majority (over 96%) \nhad completed their income tax returns and thus would have received their rebate \nif eligible. In this wave, we secured completes from 1,509 panellists from Alberta \n(n\u2009=\u2009303), British Columbia (n\u2009=\u2009301), Ontario (n\u2009=\u2009301), Qu\u00e9bec (n\u2009=\u2009300) and \nSaskatchewan (n\u2009=\u2009304). Another 251 respondents (50 from each province except \nQuebec, where 51 completed) were added as a check against panel experience, for \na total of 1,760 completes. The AAPOR RR3 response rate for this third wave was \n49%. We then secured 1,440 completes in the fourth wave following the October \n2019 federal election, of which 1,190 were returning panellists. The remaining 250 \nover sample were equally distributed across the five provinces. The fieldwork for \nthis portion of the study was conducted between 22 November and 16 December \n2019. The AAPOR RR3 response rate for this portion of the fieldwork was 56%. \nFinally, a total of 899 panellists completed a fifth wave of our survey administered \nbetween 13 and 28 May 2020. This included 200 from British Columbia, 176 from \nAlberta, 161 from Saskatchewan, 193 from Ontario and 169 from Qu\u00e9bec. The \nAAPOR RR3 response rate for returning panellists in wave 5 was 76%. Our overall \nsample was broadly representative of population characteristics of each province, \nincluding age, gender, education and income (Supplementary Section 21). The \nsurvey received a human subjects review from the Universit\u00e9 de Montr\u00e9al\u2019s Comit\u00e9 \nd\u2019\u00e9thique de la recherche en arts et humanit\u00e9s (certificate CERAH-2019-016-D). \nAll survey respondents provided informed consent before beginning the survey. \nWe find no evidence of systematic attrition of respondents in our later waves on  \nthe basis of observed demographic characteristics (Supplementary Section 22).\nA concern in any public opinion panel is that respondents are repeatedly \nexposed to a topic, shaping their beliefs and opinions as a function of \npanel participation. Panel design effects of this sort can compromise the \nrepresentativeness of a panel over time. In the context of the present study, \npanel design effects are of even greater potential concern. Climate dividends, as \nimplemented by the Canadian federal government, are integrated into a complex \nfederal income tax system, which would tend to dampen respondents\u2019 awareness \nand understanding of the policy. Further, contentious political debates have created \na confusing messaging environment for Canadians about the structure, value and \npresence of carbon pricing policies in various provinces. If panel respondents, \nby virtue of their participation in our study, became more informed about and \nengaged with climate dividends, then preference shifts within the panel could be a \nmisleading indicator of the dividend\u2019s effects on the general public\u2019s preferences. To \nmeasure potential design effects, we collected a random sample of new respondents \nduring waves 2, 3 and 4 (n\u2009=\u2009252 in wave 2, n\u2009=\u2009251 in wave 3 and n\u2009=\u2009250 in wave \n4). These respondents were randomly sampled across the five survey provinces \nin equal proportion, equivalent to our sampling procedure in the broader panel \nsurvey (Supplementary Section 20).\nCanada survey experiment. Before deploying wave 4, we estimated the objective \nrebate received by each survey respondent from Ontario and Saskatchewan, using \ntheir province of residence, reported marital status (including common law), \nnumber of children residing with them as reported in wave 3 and whether their \nresidence is rural (for example, outside a census metropolitan area (CMA)) and \nthus eligible for an additional rebate. These factors completely determine dividend \nlevels within the current Canadian policy, which we calculated using Revenue \nCanada income tax worksheets. Note that dividend levels are not a function of \nincome in Canada. For CMA measurements, we determine the respondent\u2019s \nplace of residence using the Postal Code Conversion File provided by Statistics \nCanada, which gives us a range of geographic identifying variables (such as \nresidence in a CMA and electoral district) for each of the self-reported postal codes \ncollected in our survey. We summarize the rebate calculation process for 2019 in \nSupplementary Section 23. As part of an embedded survey experiment in wave 4,  \nwe randomly assigned half the respondents to receive a filled-out tax form that \nshowed them their own household rebate amount (Supplementary Section 12).\nDetails about question wording in our survey instrument are presented in \nSupplementary Section 24. All respondents were given the option of responding in \neither English (n\u2009=\u2009752) or French (n\u2009=\u2009147).\nSwiss public opinion survey. We fielded an online survey of 1,050 Swiss residents, \nquota sampled on age, gender and language, in December 2019. The survey was \nprovided in German and French but not Italian, which is the official language in \nthe canton of Ticino as well as some municipalities in Graub\u00fcnden. Nevertheless, \nthe survey covers respondents from all Swiss cantons. Overall, the sample quite \nclosely matches the Swiss population; however, as typical for such surveys, the \ngroups of the lower educated (secondary education I) as well as the oldest age \ngroups are somewhat underrepresented (Supplementary Section 26). A copy  \nof the Swiss survey instrument is also provided in Supplementary Section 26.  \nThe survey received a human subjects review from the University of California \nSanta Barbara\u2019s Office of Research Human Subjects Committee (protocol number \n21-19-0801). All survey respondents provided informed consent before beginning \nthe survey.\nSwiss survey experiment. As part of this December 2019 Swiss survey, half of \nrespondents were randomly assigned to an encouragement treatment, where we \nasked respondents to pause their survey and retrieve their most recent health \ninsurance form. Respondents were then asked to let us know what the size of \ntheir rebate benefit was. Because all Swiss residents receive the same amount, we \nthen displayed a sample document (in their language of survey response) to all \nrespondents in the treated group, whether they reported finding their bill or not \n(Supplementary Section 15 for example). We then measured respondent support \nfor the existing Swiss policy and their potential support for either a small (from \nCHF0.25 to CHF0.28 per l heating oil) or large (to CHF0.40 per l heating oil) tax \nincrease (respondents were randomly asked for their preferences on either of these \ntwo cost settings). A summary of all variables used as well as descriptive statistics \nfor the Swiss data can be found in Supplementary Section 27.\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nAll supporting data are available through a Harvard Dataverse replication archive \nat https://doi.org/10.7910/DVN/3WBCH9.\nCode availability\nAll supporting code is available through a Harvard Dataverse replication archive at \nhttps://doi.org/10.7910/DVN/3WBCH9.\nReferences\n\t52.\tStandard Definitions: Final Dispositions of Case Codes and Outcome Rates for \nSurveys (AAPOR, 2015).\nAcknowledgements\nThis research was supported by grants (initials of grant principal investigators in \nparentheses) from a Social Sciences and Humanities Research Council (SSHRC) \nPartnership Granthad (M.M., E.L. and K.H.), theSmart Prosperity Institute at the \nUniversity of Ottawa (M.M., E.L. and K.H.), the Economics and Environmental Policy \nResearch Network of Canada (M.M., E.L. and K.H.) and the Centre for International \nGovernance Innovation (#5597, K.H.), the Social Sciences and Humanities Research \nCouncil of Canada (#435-2017-1388, E.L.), the Institute for Social, Economic and \nBehavioral Research at UC Santa Barbara (M.M.) and the Hellman Fellows Fund (M.M.). \nWe thank participants at the UC Santa Barbara Environmental Politics Workshop, Swiss \nPolitical Science Association annual meeting, Environmental Politics and Governance \nconference, L. Fesenfeld, P. Bergquist, C. Fischer, P. Quirk, G. de Roche and C. Hazlett  \nfor comments on earlier drafts of this manuscript.\nAuthor contributions\nM.M., E.L. and K.H. designed, collected and analysed the Canadian data presented  \nhere. M.M. and I.S.-S. designed, collected and analysed the Swiss data presented here.  \nAll authors contributed to the writing of the paper.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTuRE ClImaTE CHangE\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at https://doi.org/10.1038/s41558-021-01268-3.\nSupplementary information The online version contains supplementary material \navailable at https://doi.org/10.1038/s41558-021-01268-3.\nCorrespondence and requests for materials should be addressed to Matto Mildenberger.\nPeer review information Nature Climate Change thanks Thomas Bernauer, Nicholas Rivers \nand the other, anonymous, reviewer(s) for their contribution to the peer review of this work.\nReprints and permissions information is available at www.nature.com/reprints.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTuRE ClImaTE CHangE\nExtended Data Fig. 1 | Opposition to carbon pricing by province across waves. Wave 1 was conducted in February 2019 and wave 5 in April 2020. The \ndotted line indicates when the federal carbon tax policy came into effect. The solid line indicates the approximate period during which households received \ntheir climate rebates. The dashed line indicates the timing of a federal election in which climate policy, including the carbon tax, was highly salient. \nRespondents in Saskatchewan and Ontario received a federal climate rebate associated with Canada\u2019s 2019 carbon tax. Other respondents were subject  \nto provincial carbon pricing policies that had few, if any rebate, components. Error bars give 95% confidence intervals.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTuRE ClImaTE CHangE\nExtended Data Fig. 2 | Support for carbon pricing among Liberal and Conservative voters, by rebate vs. non-rebate province. The dotted line indicates \nwhen the policy came into effect. The solid line indicates the approximate period during which households received their climate rebates. The dashed line \nindicates the timing of a federal election in which climate policy, including the carbon tax, was highly salient. Voters are classified according to their wave \n1 (pre-policy implementation) party preferences. Since Alberta only became subject to the federal tax, and thus eligible for the federal dividend, between \nwave 4 and 5 (and Albertan respondents had not received a rebate as of wave 5), we bundle Albertan data with the non-rebate provinces. Error bars give \n95% confidence intervals.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTuRE ClImaTE CHangE\nExtended Data Fig. 3 | Support for carbon pricing among Liberal and Conservative voters, by cost exposure. Cost exposure measured by whether \nrespondents report driving alone to work, or whether they report a different means of getting to work (transit, walk, cycle, carpool, work/study from home). \nIndividuals who chose the survey option \u2018This question doesn\u2019t apply to me\u2019 when asked how they get to work are excluded from the figure. The dotted line \nindicates when the policy came into effect. The solid line indicates the approximate period during which households received their climate rebates. The \ndashed line indicates the timing of a federal election in which climate policy, including the carbon tax, was highly salient. Voters are sorted according to \ntheir wave 1 (pre-policy implementation) party preferences and self-reported means of getting to work. Error bars give 95% confidence intervals.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTuRE ClImaTE CHangE\nExtended Data Fig. 4 | Distribution of perceived household rebate sizes for Canadian panel. Responses from respondents who remained in the panel as \nof wave 3 and resided in the rebate provinces of Ontario and Saskatchewan. The correct answer for each set of respondents is highlighted in green.\nNature Climate Change | www.nature.com/natureclimatechange\n\n\n    Scientific Research Findings:", "answer": "Neither Canada nor Switzerland has implemented rebates in a highly visible way. A decade after adoption, we find that 85% of respondents in Switzerland were unaware that they receive rebates. Awareness was higher soon after the implementation of Canada\u2019s carbon tax, but 25% and 45% of respondents in two provinces, respectively, that were subject to the federal tax did not know that they had received an income tax credit. Most citizens in both countries significantly underestimated the value of rebates received. When informed of the true value, Swiss respondents were slightly more supportive of the current tax but not of a tax increase. In Canada, information on rebates did not increase public support. Tracking policy support over time revealed no difference between Canadian provinces with and without rebates. Instead, we found that carbon tax support was conditioned more by respondents\u2019 partisan allegiances than by rebates.", "id": 55}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Analysis\nhttps://doi.org/10.1038/s41558-021-01217-0\n1Department of Environmental and Occupational Health and Justice, Rutgers School of Public Health, Center for Population-Level Bioethics, and \nDepartment of Philosophy, Rutgers University, New Brunswick, NJ, USA. 2Division of Social Sciences, Yale-NUS College, Singapore, Singapore. 3Energy \nand Resources Group, University of California, Berkeley, CA, USA. 4School of Public and International Affairs, Princeton University, Princeton, NJ, \nUSA. 5Mercator Research Institute on Global Commons and Climate Change (MCC), Berlin, Germany. 6Department of Economics of Climate Change, \nTechnische Universit\u00e4t Berlin, Berlin, Germany. 7Department of Global Health and Population, Harvard TH Chan School of Public Health, Harvard \nUniversity, Boston, MA, USA. 8Paris School of Economics, Paris, France. 9Joint Research Centre, European Commission, Seville, Spain. 10Potsdam \nInstitute for Climate Impact Research, Potsdam, Germany. 11Department of Economics, University of Oklahoma, Norman, OK, USA. 12CNRS, CIRED, \nNogent-sur-Marne, France. 13School of International Affairs & Department of Civil and Environmental Engineering, Pennsylvania State University, University \nPark, PA, USA. 14Gangarosa Department of Environmental Health, Rollins School of Public Health, Emory University, Atlanta, GA, USA. 15Department \nof Economics, University of Texas, Austin, TX, USA. 16International Institute for Applied Systems Analysis (IIASA), Laxenburg, Austria. 17Paris School of \nEconomics (CNRS), Centre d\u2019Economie de la Sorbonne, Paris, France. \u2709e-mail: mark.budolfson@rutgers.edu\nA \nfamiliar theme from research on climate policy and eco-\nnomic development is that there is an important trade-off \nbetween climate action and near-term poverty reduction; \nthis literature is based in part on results from existing cost\u2013benefit \nclimate policy models1\u20135, which assume that the burden of a nation\u2019s \nclimate mitigation must fall to some extent on the poor. If this \nassumption were correct, some trade-off between climate action \nand poverty alleviation would be inevitable. The key question would \nthen be to what extent benefitting the future poor through avoiding \nfuture climate damages can justify (from a development or equity \nperspective) reduced near-term development for the current poor6,7.\nHowever, these models ignore the possibility that the revenues \nfrom a carbon tax could be used in a progressive way that gener-\nates immediate net benefits for the current poor. A large literature \nhas now investigated the implications of these \u2018revenue recycling\u2019 \nopportunities and identified an equal per capita refund of the rev-\nenues as a salient option8\u201318. The evidence indicates that an equal \nper capita refund typically makes immediate net beneficiaries out \nof most citizens and is often more progressive and potentially more \nfeasible than other salient options for using revenues19\u201321.\nFindings from studies of revenue recycling have not been incor-\nporated into optimal policy analyses at the global level, including to \nmodel possible synergies with other development goals, for exam-\nple sustainable development goals (SDGs)22\u201324. This is an important \noversight, as many of the arguments that there are trade-offs between \nclimate action and poverty alleviation or other SDGs depend on the \npremise that climate action must harm the current poor25,26. Indeed, \nbecause the possibility of progressive revenue recycling is not taken \ninto account in existing optimal climate policy calculations, these \nmodels have a built-in bias against mitigation, since they imply that \nmitigation must entail costs for the poorest citizens within regions \nin the coming decades and, more generally, imply an intergenera-\ntional trade-off in well-being27,28.\nModelling progressive revenue recycling\nIn the climate economics literature, the \u2018initial burden\u2019 of a carbon \ntax\u2014the distribution of tax payments and mitigation costs before \nany possible redistribution of revenues\u2014is generally found to sub-\ntract from all income groups (and thus would increase poverty \nin the absence of redistribution) but in a way that is progressive \nin poorer countries and regressive in richer countries; in poorer \ncountries fossil fuels are disproportionately consumed (relative to \nincome) by richer citizens, whereas in rich countries fossil fuels are \ndisproportionately consumed by poorer citizens18,29,30. Therefore, as \npoorer countries get richer and consumption patterns change, this \nregionally differentiated driver of the initial burden of carbon taxes \nwill probably evolve.\nTo confirm this relationship and quantify these dynamics, we \nconducted a review of the literature on the initial burden of a car-\nbon or gasoline tax (Supplementary Section 1). We included studies \nfrom around the world to capture estimates for regions with differ-\nent levels of wealth. Figure 1 displays the results, reporting the rela-\ntionship between gross domestic product (GDP) per capita and the \ndistribution of the initial burden before redistribution of revenues \nClimate action with revenue recycling has benefits \nfor poverty, inequality and well-being\nMark Budolfson\u200a \u200a1\u2009\u2709, Francis Dennig\u200a \u200a2, Frank Errickson\u200a \u200a3,4, Simon Feindt\u200a \u200a5,6, Maddalena Ferranna7, \nMarc Fleurbaey8, David Klenert\u200a \u200a9, Ulrike Kornek5,10, Kevin Kuruc\u200a \u200a11, Aur\u00e9lie M\u00e9jean\u200a \u200a12, \nWei Peng\u200a \u200a13, Noah Scovronick\u200a \u200a14, Dean Spears\u200a \u200a15, Fabian Wagner16 and St\u00e9phane Zuber\u200a \u200a17\nExisting estimates of optimal climate policy ignore the possibility that carbon tax revenues could be used in a progressive \nway; model results therefore typically imply that near-term climate action comes at some cost to the poor. Using the Nested \nInequalities Climate Economy (NICE) model, we show that an equal per capita refund of carbon tax revenues implies that achiev-\ning a 2\u2009\u00b0C target can pay large and immediate dividends for improving well-being, reducing inequality and alleviating poverty. In \nan optimal policy calculation that weighs the benefits against the costs of mitigation, the recommended policy is characterized \nby aggressive near-term climate action followed by a slower climb towards full decarbonization; this pattern\u2014which is driven \nby a carbon revenue Laffer curve\u2014prevents runaway warming while also preserving tax revenues for redistribution. Accounting \nfor these dynamics corrects a long-standing bias against strong immediate climate action in the optimal policy literature.\nNature Climate Change | VOL 11 | December 2021 | 1111\u20131116 | www.nature.com/natureclimatechange\n1111\n\nAnalysis\nNature Climate Change\n(the consumption elasticity of the initial burden, where an elastic-\nity of \u03f5 means that if a person\u2019s consumption increases by 1%, that \nperson\u2019s initial burden increases by \u03f5%). An elasticity <1 means the \ninitial burden of the carbon tax falls disproportionately on the poor \n(the tax is regressive before redistribution of revenues), whereas a \nvalue >1 indicates that the tax burden falls disproportionately on \nthe rich (the tax is progressive before redistribution of revenues). \nWe use this relationship as an estimate of the distributional impli-\ncations of carbon taxation, assuming that the initial burden is dis-\ntributed within a region on the basis of the consumption elasticity \nestimated by the best-fit line in Fig. 1. As a region grows richer over \ntime, the elasticity used to estimate the distribution of its initial bur-\nden declines.\nTo investigate the impact of an equal per capita refund of tax rev-\nenues on well-being, poverty and inequality, we modify the Nested \nInequalities Climate Economy (NICE), a 12-region global climate \npolicy model that represents inequality within regions by grouping \nthe population into five equally populous quintiles, ranked from \npoorest to richest. We modify NICE to implement two distinct \npolicy scenarios. In the first scenario, the \u2018no recycling\u2019 scenario, \nmitigation costs affect consumption but tax payments do not. This \nis the standard assumption in this type of model and is implemented \nby returning tax revenues in proportion to the initial burden. In the \nsecond scenario, the \u2018recycling\u2019 scenario, the tax revenue in each \nregion is redistributed on an equal per capita basis. As a result, some \nquintiles are net beneficiaries in the recycling scenario if the refund \nis greater than the initial burden; this is in contrast to the no recy-\ncling scenario where all quintiles bear a net cost from the climate \npolicy. (See Methods and in particular equation (4), for a detailed \ndescription of the two scenarios.)\n2\u2009\u00b0C benefits for poverty, inequality and well-being\nAs a first demonstration of the potential impact of revenue recycling, \nwe model the difference in consumption of the poorest quintile in \nall NICE regions under a 2\u2009\u00b0C scenario relative to business-as-usual \n(BAU), both with equal per capita revenue recycling (the recycling \nscenario) and without it (the no recycling scenario) (Fig. 2a,b). \nThere is a similar pattern in all regions: without progressive revenue \nrecycling, climate action does indeed involve a substantial trade-off \nwhere the poorest lose from climate policy in the short-to-medium \nterm as they shoulder their share of mitigation costs without com-\npensation. In contrast, with the equal per capita dividend, climate \naction involves a synergy with poverty alleviation. Yet even in the \nrecycling scenario, consumption falls below BAU for several regions \nlater in the century. This occurs because it is after the point where \nthere are substantial revenues to be distributed (see section on the \ncarbon Laffer curve) but before the point where the benefits of cli-\nmate action are large. Nevertheless, consumption in the recycling \nscenario is always above the no recycling scenario in the early peri-\nods due to the benefits of redistribution. After the year 2100, both \ncases produce increasing benefits from avoided climate damage. \nNote that once carbon revenues disappear in the future, people will \nalso be much wealthier than their counterparts today.\nFocusing on inequality\u2014measured by the Gini index (Fig. \n2b,d)\u2014also demonstrates the benefits of progressive redistribu-\ntion. Equal per capita recycling generates a reduction in inequal-\nity in all regions while revenues are available for redistribution. \nOnce full decarbonization occurs and revenues disappear, miti-\ngation has a regressive impact compared with BAU due to the \nrelationship reported in Fig. 1 combined with the continued cost \nof decarbonization even after there are zero net emissions. The \nimpacts on inequality without recycling, which are determined by \nthe elasticity estimated in Fig. 1, are small overall and switch from \nprogressive to regressive once a region\u2019s GDP per capita surpasses \n~US$21,500 (Fig. 1).\nExamining the impact of the equal per capita refund on all quin-\ntiles in the United States, China and India\u2014chosen to represent \ncountries at different levels of wealth\u2014reveals that in all three coun-\ntries, more than half the population (namely, those in the lower part \nof the distribution) benefits in the near term, particularly those in \nthe bottom quintile (Fig. 3). In India, the poorest 40% never experi-\nence a loss relative to BAU over the full time horizon. This redistri-\nbution towards the lower quintiles has a positive effect on poverty \nalleviation by reducing the percentage of the population below the \npoverty line (Supplementary Tables 2\u20134).\nFurthermore, the progressive equal per capita dividend increases \naggregate well-being in every region relative to the BAU over the \nnext decades and in the far future (Supplementary Fig. 3). The inter-\ngenerational trade-off between costs of reducing emissions now \nand benefits in the future is weakened over the entire time horizon: \naggregate well-being over time is higher with the equal per capita \ndividend than without it in all regions and both are better overall \nthan BAU.\nAll results presented above assume that revenues raised in a \ngiven region are distributed only within that region. However, there \nare well-being- and justice-based arguments for redistributing total \nglobal revenues on an equal per capita basis globally21,31,32. Under \nthis redistribution framework, more dramatic improvements occur \nfor inequality and consumption in the poorest regions of the world \n(Supplementary Fig. 4).\nThe carbon Laffer curve\nThe stringent 2\u2009\u00b0C constraint means that the world will rapidly \ndecarbonize and so there will be less and less revenue from car-\nbon taxation to recycle. This highlights an important caveat to our \nstoryline: the positive effect of the carbon tax through progressive \nredistribution is initially strong but diminishes once the economy \ndecarbonizes enough for revenues to decline. In short, there is a \u2018car-\nbon Laffer curve\u2019. Conceptually, the Laffer curve is the widely recog-\nnized fact that tax revenue does not monotonically increase with the \ntax rate\u2014in the case of sufficiently large taxes, market transactions  \n2.0\n1.5\n1.0\nElasticity\n0.5\nUnited States\nEuropean Union\nMiddle East\nEurasia\nLatin America\nIndia\nChina\nJapan\nOther high income\nOther Asia\nAfrica\n0\n1,000\n5,000\n10,000\nGDP per capita (2005 US$)\n25,000\n50,000\nFig. 1 | Estimates from the literature on the distribution of the initial \nburden of a carbon or gasoline tax and the resulting relationship \nwith per capita GDP. This relationship (black line) is used to estimate \nthe consumption elasticity of the initial burden before any possible \nredistribution as a function of regional per capita GDP in each NICE model \nregion at each point in time. Section 1 of the Supplementary Information \ndescribes the methods of the literature review; Supplementary Table 1 \ncites all included studies, and Supplementary Figs. 1 and 2 detail multiple \nsensitivity tests.\nNature Climate Change | VOL 11 | December 2021 | 1111\u20131116 | www.nature.com/natureclimatechange\n1112\n\nAnalysis\nNature Climate Change\n(for example, fossil fuel use) reduce to the point where there is little \ntaxable activity to generate revenue33. As a quantitative illustra-\ntion of the carbon Laffer curve in NICE, Fig. 4 shows this nonlin-\near relationship between global near-term (2025) decarbonization  \nand tax revenue. Total revenue is highest in the 55\u201375% decarbon-\nization range and decreases thereafter until full decarbonization \nultimately implies that no revenue is generated (under full decar-\nbonization there are no industrial emission to be taxed).\n9\na\nb\nc\nd\n6\n3\nChange in consumption (% of BAU)\nChange in Gini index (relative to BAU)\nChange in consumption (% of BAU)\nChange in Gini index (relative to BAU)\n0\n\u20133\n0\nUnited States (40)\nEuropean Union (27)\nJapan (30)\nRussia (28)\nEurasia (40)\nChina (31)\nIndia (35)\nMiddle East (44)\nAfrica (52)\nLatin America (43)\nOther high income (31)\nOther Asia (43)\n\u20130.25\n\u20130.50\n\u20130.75\n\u20131.00\n\u20131.25\n2020\n2040\n2060\nGini index: no recycling\nGini index: recycling\nBottom quintile consumption: no recycling\nBottom quintile consumption: recycling\n2080\n2100\n2020\n2040\n2060\n2080\n2100\n0\n\u20130.25\n\u20130.50\n\u20130.75\n\u20131.00\n\u20131.25\n2020\n2040\n2060\n2080\n2100\n9\n6\n3\n0\n\u20133\n2020\n2040\n2060\nYear\nYear\nYear\nYear\n2080\n2100\nFig. 2 | Trade-offs between climate action, poverty alleviation and inequality turn into synergies with an equal per capita carbon dividend. a,b, For a \n2\u2009\u00b0C mitigation pathway, the change in per capita consumption of the bottom quintile in each region is shown, without (a) and with (b) equal per capita \nrecycling, compared with the BAU case with no climate policy. c,d, The change in the Gini index, without (c) and with (d) equal per capita recycling, where \na higher value indicates more inequality. The numbers in the legend are the initial Gini values. Diamond symbols identify the year of maximum carbon tax \nrevenue as a percentage of regional consumption. (Results assume that each region\u2019s aggregate climate damages are distributed to quintiles in proportion \nto consumption, an assumption that makes the welfare impact of damages essentially equivalent to what they would be in more aggregated cost\u2013benefit \nmodels including DICE, RICE, PAGE and FUND51,52. Further below and in Supplementary Fig. 14 we discuss results that modify this important assumption, \nshowing that our main findings hold with other damage specifications and distributions).\n6\nQ1 (poorest)\nQ2\nQ3\nQ4\nQ5 (richest)\n2020\n2040\n2060\nYear\nYear\nYear\nUnited States\na\nb\nc\nChina\nIndia\n2080\n2100\nYear of\nmaximum revenue\n4\n2\nChange in consumption (% of BAU)\nChange in consumption (% of BAU)\nChange in consumption (% of BAU)\n0\n\u20132\n6\n2020\n2040\n2060\n2080\n2100\n4\n2\n0\n\u20132\n6\n2020\n2040\n2060\n2080\n2100\n4\n2\n0\n\u20132\nFig. 3 | Change in consumption of all quintiles in the 2\u2009\u00b0C mitigation pathway with the equal per capita recycling compared with the BAU case with no \nclimate policy. a\u2013c, Change in consumption as a percentage of BAU over time for the United States (a), China (b) and India (c). The vertical dotted line in \neach panel identifies the year of maximum carbon tax revenue as a percentage of regional consumption. Q, income quintile of population.\nNature Climate Change | VOL 11 | December 2021 | 1111\u20131116 | www.nature.com/natureclimatechange\n1113\n\nAnalysis\nNature Climate Change\nThis relationship implies that an optimal climate policy with an \nequal per capita carbon dividend must balance the value to society \nof (1) lower CO2 emissions\u2014and thus reduced climate change\u2014that \nwill result from high carbon taxes and (2) some level of continuing \nemissions, which enables the progressive redistribution that tax rev-\nenues can fund. Note that unlike income tax where going beyond \nthe peak of the Laffer curve is inefficient, in the case of climate we \nought to go beyond that point to curb climate change.\nStrong action now and steady action later\nTo investigate the trade-off between the benefits of lowering emis-\nsions and the benefits of continued carbon tax revenue, we perform \nan optimal policy calculation; optimal policy refers to the policy \nthat maximizes (discounted) net benefits through time and does \nnot feature a temperature constraint as in the results above. With \nrevenue recycling, the model recommends high decarbonization \ninitially\u2014there are dual benefits of redistributable revenue and \nlower future temperatures\u2014but postpones full decarbonization \nfor many decades as redistribution continues (Fig. 5). Without the \nequal per capita revenue recycling, the model at first recommends \nmore moderate ambition, to protect the current poor from high \nmitigation costs, followed by a rapid increase in decarbonization \nto avoid extreme warming. Despite this different temporal pattern \nof mitigation, the maximum temperature rise is similar in both \nscenarios, although it peaks later with revenue recycling, a poten-\ntially valuable delay if it reduces the rate of temperature change and \nenables more time for adaptation34. The carbon tax and carbon divi-\ndend trajectories corresponding to the decarbonization paths are \nreported in Supplementary Figs. 7 and 8. (Unless otherwise stated, \nresults assume standard discounting parameters from the Regional \nIntegrated Climate Economy (RICE) model: pure time prefer-\nence\u2009=\u20091.5% per year; consumption elasticity of marginal utility\u2009=\u20091.5 \n(representing the diminishing marginal utility of consumption) and \ndistribution of climate damages proportional to consumption.)\nThe optimal decarbonization pathway is not exclusively driven \nby the motive to redistribute. To demonstrate this, the \u2018no dam-\nages\u2019 scenario depicts the optimal carbon tax with revenue recycling \nbut where climate damages are artificially set to zero regardless of \nwarming (Fig. 5, black line). In this case, the only benefit of a car-\nbon tax is the redistribution it allows. Global decarbonization that \nis optimal purely from this motive is substantial and ranges between \n~50 and 60%, as this ensures maximum redistribution to the poor. \nStill, this is much lower than the case where climate benefits exist \nalongside redistributional benefits, demonstrating that substantial \nincentive to decarbonize further remains even at such high levels of \ndecarbonization.\nAn equal per capita global redistribution leads to similar decar-\nbonization trajectories to those reported in Fig. 5 (which assume \nwithin-region redistribution only), a result driven largely by the \ncarbon Laffer curve (Supplementary Fig. 5). However, it would \nlead to far greater improvements in global well-being, particularly \nfor Africa, India and Other Asia (Supplementary Fig. 6 and associ-\nated text).\nDiscussion and sensitivity analyses\nWe have shown that an equal per capita refund of carbon tax rev-\nenues improves the well-being of individuals toward the bottom of \nthe income distribution and reduces poverty and inequality. The \nimplication is that adopting strong climate policy need not entail \na trade-off where the people of today (and the poor in particular) \nmust sacrifice for the benefit of future generations.\nThis finding contributes to the debate over whether there should \nbe a gradual ramp up to aggressive policy (for example, as advocated \nby Nordhaus27) or a large-scale push toward immediate maximum \nfeasible reductions (for example, as advocated by Stern28). Even with \nthe relatively high discounting parameters preferred by Nordhaus, \nprogressive revenue recycling leads to high levels of decarbonization \nimmediately\u2014comparable in the initial decades to strict climate \ntarget pathways (for example, 1.5 or 2\u2009\u02daC)\u2014followed by less decar-\nbonization in later periods (Fig. 5). With lower carbon emitted in \nthe atmosphere early on, and anticipating the carbon Laffer curve, \nthe initial period of high decarbonization is followed by a gradual \nlong-term increase toward full decarbonization to keep peak warm-\ning at a moderate level and preserve revenue for redistribution.\nThe temporal difference in optimal decarbonization pathways \nbetween scenarios with and without revenue recycling (the crossing \npattern seen in Fig. 5a) appears robust to several key uncertainties. \nWhile our main results assume background inequality remains con-\nstant in all regions, Supplementary Fig. 9 shows that the crossing \npattern persists in several scenarios involving narrowing or widen-\ning background inequality. The crossing is repeated in all scenarios \nbut is less extreme with more reductions in background inequal-\nity. When background inequality is lower, initial decarbonization is \nagain much higher with the progressive recycling but, unlike in the \nother scenarios, it then remains relatively high through time. This \noccurs because a greater decrease in background inequality reduces \nthe incentive to delay decarbonization to preserve tax revenues for \nredistribution, thus bringing forward the optimal date of full decar-\nbonization to avoid more climate harms.\nOur qualitative results are also robust to choices about key dis-\ncounting parameters, namely the rate of pure time preference and \nthe consumption elasticity of marginal utility. As explained in the \nMethods, normative and descriptive disagreements exist about \nthe appropriate value of these parameters27. Under a range of dis-\ncounting parameter combinations typically considered in the lit-\nerature, revenue recycling always induces stronger short-term \nemission reductions and a slower transition to full decarbonization \n(Supplementary Fig. 10).\nOur findings raise important questions about feasibility. One is \nwhether it is technically feasible to decarbonize as quickly in the \nearly periods as the model recommends. This question is beyond \nthe scope of this paper; however, we note that the initial decades \nof the optimal trajectory reported here are comparable to many \nIPCC 1.5\u20132\u2009\u02daC pathways35. Another question relates to negative \nemissions, both whether they are needed and how they would be \nfunded if all carbon dividends are redistributed. Consistent with \nsome IPCC scenarios, our trajectories do not require negative emis-\nsions (Supplementary Fig. 11). Nevertheless, even if a substantial \nfraction of revenues was diverted to subsidize negative emission \n1.5\n1.0\n0.5\nTotal revenue (trillions 2005 US$)\n0\n0\n25\n50\nGlobal decarbonization (%)\n75\n100\nFig. 4 | The carbon Laffer curve. The curve is illustrated by plotting \nnear-term decarbonization versus global revenue generated (here for \n2025). Global decarbonization is the percentage reduction in carbon \nemissions compared with a BAU scenario with no climate policy.\nNature Climate Change | VOL 11 | December 2021 | 1111\u20131116 | www.nature.com/natureclimatechange\n1114\n\nAnalysis\nNature Climate Change\ntechnologies, the benefits of redistributing the remaining dividends \nremains strong (Supplementary Fig. 12). However, we acknowledge \nthat negative emission technologies would have unprecedented and \ncurrently poorly understood implications.\nA second dimension of feasibility concerns public opinion and \npolitical will. An emerging literature indicates that communicat-\ning the co-benefits of climate action may increase policy support, \nin particular for co-benefits that lead to economic development \nand more compassionate communities36,37. Similarly, bundling \nclimate policy with social and economic programmes, a feature \nof widely discussed strategies across the political spectrum from \nthe Climate Leadership Council to the Green New Deal, may also \nincrease support for action38. Overall, the literature suggests that \nprogressive redistribution may have relatively broad appeal, at least \ngiven effective communication of the benefits, although this may \nbe tempered by evidence from Pigouvian taxation studies which \nindicates that people may be resistant to policies that start with \nhigh tax rates19,21,39\u201342.\nA third feasibility concern is whether governments would actu-\nally have the capacity to perform progressive transfers, even if there \nwas political will to do so. In Supplementary Fig. 13, we report  \noptimal policy results under imperfect recycling programmes, \nincluding that the bottom quintile does not receive any transfers or \nif a large proportion of the revenue was lost in policy implementa-\ntion cost; in both cases the pattern of high initial decarbonization \nfollowed by the gradual progression to full decarbonization remains \nintact, although is somewhat muted. Supplementary Fig. 13 also \nreports results for a scenario that is more progressive than an equal \nper capita redistribution.\nOne potential limitation of our study is that NICE does not \ninclude the full suite of policy levers available to alleviate distribu-\ntional concerns. Within countries, for example, one might consider \nchanges to income taxation. We model synergies between carbon \ntaxation and inequality reduction under the assumption that, apart \nfrom the distribution of mitigation costs and the distribution of \nthe tax revenue, inequality is not otherwise affected by economic \nincentive effects of the policy (Supplementary Information Section \n9 gives details about the relation with optimal taxation theory). \nComplementary work could use subregional agent-based or micro-\nsimulation models to estimate how such incentive effects and other \ninteraction effects may influence inequality levels and optimal pol-\nicy with revenue recycling.\nIn addition, some NICE regions consist of multiple countries. \nTherefore, our main results implicitly assume some level of inter-\nnational transfers between countries within these multicountry \nregions. This could be important for multicountry regions with \nheterogeneous levels of development and differing capacities. \nNevertheless, NICE represents several key countries as individual \nregions (United States, China, India, Russia and Japan) and avoids \ntransfers across regions in the main results.\nFurther studies could investigate the role of the distribution of \ncarbon tax revenues when regions apply different carbon taxes. In \nthe absence of international transfers such as those modelled in \nSupplementary Fig. 6, the assumption of a global carbon price is \ncertainly a constraint to the alleviation of distributional concerns, \nsince it requires a high policy burden from poor countries. In mod-\nels allowing for differential carbon prices by region, all high emitters \nare required to mitigate at least as much as under the global carbon \nprice assumption43\u201345. Our results here with a global carbon price \ncould thus be seen as a price floor for high emitters, as recently pro-\nposed by the International Monetary Fund45.\nWe also do not consider the question of horizontal inequality\u2014\nthat is the heterogeneous effects of a carbon tax on households with \nthe same income level but different consumption patterns\u2014which \nrecent evidence suggests may be important46\u201348. Including horizon-\ntal inequality would be a worthwhile extension of our work.\nRecent research also indicates that the damage functions used in \ncost\u2013benefit models, such as NICE, may underestimate future cli-\nmate impacts49. We test this possibility in two ways. First, we keep \nthe total damages the same but assume that they disproportionately \nharm the poor (thus having a greater well-being impact). Second, \nwe double the total size of the damages. Both cases display the \ncharacteristic crossing pattern of Fig. 5, although full decarboniza-\ntion occurs earlier (Supplementary Fig. 14). The crossing pattern \nis also evident when we replace the NICE climate module with the \nFAIR climate model (Supplementary Fig. 15), as recommended in a \nrecent report by the National Academies50.\nConclusions\nEstimates of optimal climate policy have ignored the possibility that \nrevenues from a carbon tax could be used in a progressive way that \ngenerates immediate net benefits for the current poor. As a con-\nsequence, they mistakenly imply that climate action must come at \nsome cost to overall well-being and especially to the poor. We have \nshown that this storyline of the climate, development and inequal-\nity nexus reverses when progressive revenue recycling is taken into \naccount. Our approach corrects a long-standing bias against strong \n100\na\nb\nGlobal decarbonization\nGlobal temperature\n80\n60\n40\nGlobal decarbonization (%)\nIncrease in temperature (\u00b0C)\n20\nRecycling\nNo recycling\nNo damages\n0\n4\n3\n2\n1\n2050\n2100\nYear\nYear\n2150\n2200\n2050\n2100\n2150\n2200\nFig. 5 | Optimal mitigation with and without equal per capita carbon \ndividend. a,b, Optimal decarbonization (a) and temperature (b) with and \nwithout revenue recycling, and a comparison case that assumes no climate \ndamages; the latter shows how much mitigation is driven by progressive \nredistribution alone, as opposed to being driven by avoided climate \ndamages. Global decarbonization in a is the percentage reduction in carbon \nemissions compared with a BAU scenario with no climate policy.\nNature Climate Change | VOL 11 | December 2021 | 1111\u20131116 | www.nature.com/natureclimatechange\n1115\n\nAnalysis\nNature Climate Change\nimmediate climate action. We find that with progressive revenue \nrecycling, aggressive climate action can pay large dividends for \nimproving well-being, reducing inequality and alleviating poverty. \nIn an optimal policy calculation, the recommended policy is charac-\nterized by aggressive near-term climate action followed by a slower \nclimb towards full decarbonization; this pattern prevents runaway \nwarming while also preserving tax revenues for redistribution. The \nbenefits from progressive use of carbon revenues are most pro-\nnounced in the early decades, when the revenues are largest and the \nneeds of the poor are most urgent.\nOnline content\nAny methods, additional references, Nature Research report-\ning summaries, source data, extended data, supplementary infor-\nmation, acknowledgements, peer review information; details of \nauthor contributions and competing interests; and statements of \ndata and code availability are available at https://doi.org/10.1038/\ns41558-021-01217-0.\nReceived: 23 August 2020; Accepted: 9 October 2021;  \nPublished online: 29 November 2021\nReferences\n\t1.\t Anthoff, R. & Tol, R. The Climate Framework for Uncertainty, Negotiation and \nDistribution (FUND) Technical Description, Version 3.9 (Univ. of Notre \nDame, 2014); https://www3.nd.edu/~nmark/Climate/Fund-3-\n9-Scientific-Documentation.pdf\n\t2.\t Stern, N. Stern Review: The Economics of Climate Change (Her Majesty\u2019s \nTreasury, 2006).\n\t3.\t Nordhaus, W. D. & Boyer, J. Warming the World: Economic Models of Global \nWarming (MIT Press, 2000).\n\t4.\t Nordhaus, W. D. Revisiting the social cost of carbon. Proc. Natl Acad. Sci. \nUSA 114, 1518\u20131523 (2017).\n\t5.\t Tol, R. The damage costs of climate change toward more comprehensive \ncalculations. Environ. Resour. Econ. 5, 353\u2013374 (1995).\n\t6.\t Anthoff, D. & Tol, R. S. in Climate Change and Common Sense: Essays in \nHonour of Tom Schelling (eds Hahn, R. W. & Ulph, A.) 260\u2013274 (Oxford \nUniv. Press, 2012).\n\t7.\t Hahn, R. W. & Ulph, A. in Climate Change and Common Sense: Essays in \nHonour of Tom Schelling (eds Hahn, R. W. & Ulph, A.) Ch. 1 (Oxford Univ. \nPress, 2012).\n\t8.\t Cullenward, D., Wilkerson, J. T., Wara, M. & Weyant, J. P. Dynamically \nestimating the distributional impacts of US climate policy with NEMS:  \na case study of the Climate Protection Act of 2013. Energy Econ. 55,  \n303\u2013318 (2016).\n\t9.\t Fawcett, A., McFarland, J., Morris, A. C. & Weyant, J. (eds) Special issue on \nEMF 32 study on US carbon tax scenarios Climate Change Econ. 9, https://\nwww.worldscientific.com/toc/cce/09/01 (2018).\n\t10.\tGoulder, L. H. & Hafstead, M. A. Tax reform and environmental policy: \noptions for recycling revenue from a tax on carbon dioxide. Resources for the \nFuture https://doi.org/10.2139/ssrn.2338210 (2013).\n\t11.\tGoulder, L. H., Hafstead, M. A., Kim, G. & Long, X. Impacts of a Carbon Tax \nacross US Household Income Groups: What Are the Equity-Efficiency \nTrade-Offs? (National Bureau of Economic Research, 2018).\n\t12.\tHorowitz, J., Cronin, J.-A., Hawkins, H., Konda, L. & Yuskavage, A. Methodology \nfor Analyzing a Carbon Tax (US Department of the Treasure, 2017).\n\t13.\tJorgenson, D. & Goettle, R. Double Dividend: Environmental Taxes and Fiscal \nReform in the United States (MIT Press, 2013).\n\t14.\tMathur, A. & Morris, A. Distributional effects of a carbon tax in broader US \nfiscal reform. Energy Policy 66, 326\u2013334 (2014).\n\t15.\tMetcalf, G. E. Paying for Pollution: Why a Carbon Tax is Good for America \n(Oxford Univ. Press, 2018).\n\t16.\tMurray, B. C., Bistline, J., Creason, J., Wright, E. & Kanudia, A. The EMF 32 \nStudy on Technology and Climate Policy Strategies for Greenhouse Gas \nReductions in the US Electric Power Sector: An Overview (Elsevier, 2018).\n\t17.\tRausch, S., Metcalf, G. E. & Reilly, J. M. Distributional impacts of carbon \npricing: a general equilibrium approach with micro-data for households. \nEnergy Econ. 33, S20\u2013S33 (2011).\n\t18.\tSterner, T. Fuel Taxes and the Poor: The Distributional Effects of Gasoline \nTaxation and their Implications for Climate Policy (RFF Press, 2012).\n\t19.\tKlenert, D. et al. Making carbon pricing work for citizens. Nat. Clim. Change \n8, 669\u2013677 (2018).\n\t20.\tKlenert, D., Schwerhoff, G., Edenhofer, O. & Mattauch, L. Carbon taxation, \ninequality and Engel\u2019s law: the double dividend of redistribution. Environ. \nResour. Econ. 71, 605\u2013624 (2018).\n\t21.\tCarattini, S., Kallbekken, S. & Orlov, A. How to win public support for a \nglobal carbon tax. Nature 565, 289\u2013291 (2019).\n\t22.\tMcCollum, D., Gomez Echeverri, L., Riahi, K. & Parkinson, S. SDG7: Ensure \nAccess to Affordable, Reliable, Sustainable and Modern Energy for All (IIASA, \n2017); http://pure.iiasa.ac.at/id/eprint/14621/1/SDGs-interactions-\n7-clean-energy.pdf\n\t23.\tNerini, F. F. et al. Mapping synergies and trade-offs between energy and the \nSustainable Development Goals. Nat. Energy 3, 10\u201315 (2018).\n\t24.\tSustainable Development Goals (United Nations Department of Economic  \nand Social Affairs, accessed 6 May 2020); https://sustainabledevelopment.un.\norg/sdgs\n\t25.\tDavies, J. B., Shi, X. & Whalley, J. The possibilities for global inequality and \npoverty reduction using revenues from global carbon pricing. J. Econ. \nInequal. 12, 363\u2013391 (2014).\n\t26.\tFranks, M., Lessmann, K., Jakob, M., Steckel, J. C. & Edenhofer, O. \nMobilizing domestic resources for the Agenda 2030 via carbon pricing. Nat. \nSustain. 1, 350\u2013357 (2018).\n\t27.\tNordhaus, W. D. A review of the \u2018Stern review on the economics of climate \nchange\u2019. J. Econ. Lit. 45, 686\u2013702 (2007).\n\t28.\tStern, N. The economics of climate change. Am. Econ. Rev. 98, 1\u201337 (2008).\n\t29.\tDorband, I. I., Jakob, M., Kalkuhl, M. & Steckel, J. C. Poverty and \ndistributional effects of carbon pricing in low- and middle-income \ncountries\u2014a global comparative analysis. World Dev. 115, 246\u2013257 (2019).\n\t30.\tSterner, T. Distributional effects of taxing transport fuel. Energy Policy 41, \n75\u201383 (2012).\n\t31.\tMarron, D. B. & Maag, E. How to Design Carbon Dividends (SSRN, 2018).\n\t32.\tGardiner, S., Caney, S., Jamieson, D. & Shue, H. Climate Ethics: Essential \nReadings (Oxford Univ. Press, 2010).\n\t33.\tLaffer, A. B. The Laffer curve: past, present, and future. Backgrounder 1765, \n1\u201316 (2004).\n\t34.\tBurke, M., Davis, W. M. & Diffenbaugh, N. S. Large potential reduction in \neconomic damages under UN mitigation targets. Nature 557, 549\u2013553 (2018).\n\t35.\tIPCC Special Report on Global Warming of 1.5\u2009\u00b0C (eds Masson-Delmotte, V. \net al.) (WMO, 2018).\n\t36.\tBain, P. G., Hornsey, M. J., Bongiorno, R. & Jeffries, C. Promoting \npro-environmental action in climate change deniers. Nat. Clim. Change 2, \n600\u2013603 (2012).\n\t37.\tBain, P. G. et al. Co-benefits of addressing climate change can motivate action \naround the world. Nat. Clim. Change 6, 154\u2013157 (2016).\n\t38.\tBergquist, P., Mildenberger, M. & Stokes, L. C. Combining climate, economic, \nand social policy builds public support for climate action in the US. Environ. \nRes. Lett. 15, 054019 (2020).\n\t39.\tCarattini, S., Carvalho, M. & Fankhauser, S. Overcoming public resistance to \ncarbon taxes. WIREs Clim. Change 9, e531 (2018).\n\t40.\tMarshall, G. et al. Guide to Communicating Carbon Pricing (World  \nBank, 2018).\n\t41.\tPahle, M. et al. Sequencing to ratchet up climate policy stringency. Nat. Clim. \nChange 8, 861\u2013867 (2018).\n\t42.\tCarattini, S., Baranzini, A., Thalmann, P., Varone, F. & V\u00f6hringer, F. Green \ntaxes in a post-Paris world: are millions of nays inevitable? Environ. Resour. \nEcon. 68, 97\u2013128 (2017).\n\t43.\tBudolfson, M. & Dennig, F. in Handbook on the Economics of Climate Change \n(eds Chichilnisky, G. et al.) 224\u2013238 (Edward Elgar Press, 2020).\n\t44.\tBudolfson, M. et al. Utilitarian benchmarks for emissions and pledges \npromote equity, climate, and development. Nat. Clim. Change 11,  \n827\u2013833 (2021).\n\t45.\tParry, I., Black, S. & Roaf, J. Proposal for an International Carbon Price Floor \nAmong Large Emitters (International Monetary Fund, 2021).\n\t46.\tCronin, J. A., Fullerton, D. & Sexton, S. Vertical and horizontal \nredistributions from a carbon tax and rebate. J. Assoc. Environ. Resour. Econ. \n6, S169\u2013S208 (2019).\n\t47.\tDouenne, T. The vertical and horizontal distributive effects of energy taxes: a \ncase study of a French policy. Energy J. 41, 231\u2013253 (2020).\n\t48.\tFischer, C. & Pizer, W. A. Horizontal equity effects in energy regulation. J. \nAssoc. Environ. Resour. Econ. 6, S209\u2013S237 (2019).\n\t49.\tCarleton, T. et al. Valuing the Global Mortality Consequences of Climate \nChange Accounting for Adaptation Costs and Benefits (Univ. of Chicago, 2018).\n\t50.\tValuing Climate Damages: Updating Estimation of the Social Cost of Carbon \nDioxide (National Academies Press, 2017).\n\t51.\tBudolfson, M., Dennig, F., Fleurbaey, M., Siebert, A. & Socolow, R. H. The \ncomparative importance for optimal climate policy of discounting, \ninequalities and catastrophes. Clim. Change 145, 481\u2013494 (2017).\n\t52.\tDennig, F., Budolfson, M. B., Fleurbaey, M., Siebert, A. & Socolow, R. H. \nInequality, climate impacts on the future poor, and carbon prices. Proc. Natl \nAcad. Sci. USA 112, 15827\u201315832 (2015).\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2021\nNature Climate Change | VOL 11 | December 2021 | 1111\u20131116 | www.nature.com/natureclimatechange\n1116\n\nAnalysis\nNature Climate Change\nMethods\nAll model code and data used to generate results for this article are archived53 \nand a running version is available at https://github.com/Environment-Research/\nrevenue_recycling.\nThe NICE model52 used here is a modification of the RICE model3,54, which \nwas developed by W. Nordhaus. RICE is the regional counterpart to the global \ndynamic integrated climate\u2013economy (DICE) model, which is one of three \nleading cost\u2013benefit models used by researchers and governments for regulatory \nanalysis, including to estimate the social cost of carbon55. RICE3,54 and NICE51,52 \nhave been described in great detail elsewhere. Since their basic architecture is \nthe same, we first describe this RICE architecture and then explain the model \ndevelopments that make RICE into NICE, noting from the outset that all models of \nthis class are reduced-form representations of reality with associated strengths and \nlimitations56,57.\nRICE is a regionally disaggregated optimization model that includes an \neconomic component and a geophysical (climate) component that are linked. \nRICE divides the world into 12 regions, some of which are single countries while \nothers are groups of countries. Each region has a distinct endowment of economic \ninputs including capital, labour and technology, which together produce that \nregion\u2019s gross output via a Cobb\u2013Douglas production function. Carbon emissions \nare a function of gross output and an exogenously determined, region-specific, \ncarbon intensity pathway. These carbon emissions can be abated (mitigating \nclimate change) at a cost to gross output via regional control policies that are \nselected so that in every period the marginal cost of abatement\u2014or carbon \nprice\u2014is the same for all regions. The climate module determines how unabated \ncarbon emissions affect global temperature and, ultimately, the future economy \nthrough climate-related damages. Region-specific damage functions capture this \nrelationship between increased temperature and economic damage, with poorer \nregions generally more vulnerable as a proportion of income.\nThe original RICE model is solved by choosing decarbonization and savings \nrates in all regions and periods to maximize an objective function which sums, over \nperiods and regions, a concave utility function of regional per capita consumption \nwith a discount factor applied to future values. To simplify the optimization, the \nsolution concept implemented in this study takes the savings rates as given\u2014rather \nthan solving for their optimal values\u2014and maximizes only over the control rates \n(decarbonization). In the default implementation of RICE, Negishi weights are \nadded to the objective function to ensure that the marginal cost of reducing \nemissions by a ton (the carbon price) is the same for all regions, period by period. \nNICE achieves equality of carbon prices without using Negishi weights52.\nThe NICE model extends RICE by disaggregating regional consumption \ninto five socioeconomic groups with consumption levels reflecting the current \ndistribution of consumption within the regions58. So as not to affect any of the \naggregate economic variables (investment, capital, output and so on), this is \ndone by splitting average regional consumption into five units (or quintiles) \nafter aggregate savings have been determined. The background consumption \ndistribution and the distributions of damage and mitigation cost are determined in \nthe way described below.\nWe denote regions by index i, quintiles by j and periods by t. Quantities \nwithout a j index are regional aggregates and are identical to the quantities in the \nmore aggregated RICE model. Net output Yit is given by\nYit = 1\u2212\u03bbit\n1+Dit Qit\n(1)\nwhere Qit denotes gross output, \u03bbit mitigation cost (opportunity costs of reducing \nCO2 emissions as a share of GDP) and Dit climate damages. The basic trade-off \nof the RICE model\u2014mitigation costs in the present for the reduction of climate \ndamages in the future\u2014is embodied in this equation. As mentioned above, in each \nperiod the regional mitigation costs are chosen so that they are consistent with a \nglobally uniform carbon price, which is implemented as a local tax, taxt, in each \nregion.\nDefining the aggregate savings rate, sit, and population, Lit, the average regional \nconsumption is\n\u00afcit = 1\u2212sit\nLit Yit\n(2)\nwhile the average gross consumption (predamage and premitigation cost) is\n\u00afcpre\nit\n= 1\u2212sit\nLit Qit = 1+Dit\n1\u2212\u03bbit \u00afcit\n(3)\nWe assume that gross consumption is distributed across population quintiles \naccording to a baseline distribution, yielding gross consumptions for each quintile. \nUnder the no recycling scenario, final consumption of each quintile is computed \nby subtracting climate damages and mitigation costs from gross consumption \naccording to distributions that reflect different exposures and vulnerabilities of \nconsumption groups to these impacts. Under the recycling scenario, carbon taxes \nare raised according to the same distribution as mitigation costs and redistributed \nas equal per capita payments within regions.\nThe baseline distribution is given by quintile weights, qijt, that denote the ratio \nbetween quintile consumption and average consumption. If for quintile j in region i \nand period t, qijt\u2009>\u20091, its consumption is greater than average regional consumption \nin that period; if qijt\u2009<\u20091, its consumption is less than the average. Since the five \nquintiles comprise equal proportions of the population, \u2211\nj qijt = 5 in all regions \nand periods. In the base implementation these quintile weights are fixed across \ntime and estimated to the current distribution of consumption in the region by \naggregating country level distributional data from the World Income Inequality \nDatabase58 to regional distributions. The aggregation is described in detail in \nSection 6 of the Supplementary Information.\nThe initial burden of the carbon tax is the sum of the mitigation costs and \ntax payments. Within a region, the initial burden is distributed across quintiles \naccording to the weights, \u03c4ijt. The substantive assumption of our analysis is that the \ntwo components of the initial burden\u2014the mitigation cost and the tax payment\u2014\nare distributed according to the same weights, \u03c4ijt, which are calculated on the basis \nof Fig. 1, as described in more detail below.\nWe denote by dijt the weights of the distribution of damage to consumption in \nregion i and period t, which we also describe in more detail below.\nWith this notation the consumption of quintile j in region i and period t is \ngiven by\ncijt =\n\u00afcpre\nit qijt\n\u001f \u001e\u001d \u001c\nGross consumption\n\u2212\u00afcitDitdijt\n\u001f \u001e\u001d \u001c\nDamage cost\n\u2212\nInitial burden\n\u001d\n\u001c\u001f\n\u001e\n\uf8eb\n\uf8ec\n\uf8ec\n\uf8ec\n\uf8ed\n\u00afcpre\nit \u03bbit\u03c4ijt\n\u001f\n\u001e\u001d\n\u001c\nMitigation cost\n+ Eit\nLit\ntaxt\u03c4ijt\n\u001f\n\u001e\u001d\n\u001c\nTax payments\n\uf8f6\n\uf8f7\n\uf8f7\n\uf8f7\n\uf8f8+ Eit\nLit\ntaxt\u03b4ijt\n\u001f\n\u001e\u001d\n\u001c\nRefund\n(4)\nThe value of the parameter \u03b4ijt in the expression for the refund distinguishes \nour two policy scenarios: no recycling and recycling. In the no recycling scenario, \ncarbon tax revenues are refunded within each region according to the distribution \nof the initial burden, so that \u03b4ijt\u2009=\u2009\u03c4ijt. From equation (4) we can see that this implies \nthat tax payments and the refund cancel each other out. Hence the carbon tax \ncomponents disappear, leaving the mitigation cost as the only impact of the climate \npolicy, as is standard in cost\u2013benefit Integrated Assessment Models (IAMs). That \nis the reason we call this the no recycling scenario. Under this implementation, all \nquintiles bear some cost from climate policy.\nIn the recycling scenario, carbon tax revenues are refunded equally per capita \nwithin each region, so that \u03b4ijt\u2009=\u20091. As a hypothetical example to illustrate the \ndistributional impact of this scenario, if \u03c4ijt\u2009=\u20091 for all quintiles, the tax would be \nraised equally per capita and cancelled out with the equal per capita dividend, \nresulting in the same situation as in the no recycling scenario. But in all of our \nmodel runs \u03c4ijt >1 for the top quintile and <1 for the bottom quintile, so that the \nrecycling scenario always yields a more equal distribution than the no recycling \nscenario when the same (positive) tax is applied.\nThe essential ingredients for the process of downscaling regional consumption \nto subregional consumption quintiles are the distributional weights qijt, dijt and \u03c4ijt . \nAs described in the Supplementary Information, the qijt for the first model period \nare estimated from current regional consumption distributions. Under our baseline \nassumption these remain constant over time and in Supplementary Fig. 9 we \nconsider alternative projections.\nFor the distributional weights of damage and of the initial burden (dijt and \u03c4ijt) \nwe assume a constant elasticity relationship to the consumption distribution:\ndijt = 5\nq\u03be\nijt\n\u2211\nk q\u03be\nikt\nand\n\u03c4ijt = 5\nq\u03c9it\nijt\n\u2211\nk q\u03c9it\nikt\nIn the main results of the paper we take the damage elasticity of consumption, \n\u03be, to be equal to 1 in all periods and in all regions. In Supplementary Fig. 14 we \nconsider alternative values of this parameter. Previous applications of the NICE \nmodel study the importance of this parameter to optimal carbon prices51,52.\nBecause the distributional weights, \u03c4ijt, of the initial burden are central to our \npolicy analysis and because there is substantial evidence that the consumption \nelasticity of the initial burden, \u03c9it, decreases with a region\u2019s per capita GDP, \nwe estimate a relationship between this elasticity and GDP per capita with a \nsimple ordinary least squares regression of the estimates from the literature on \nthe distributional impact of carbon and fuel taxes, summarized in Fig. 1. For \neach study, k, we estimated the elasticity, \u03c9k, as the slope of the regression of \nlog initial burden reported in the study with respect to log consumption level \nof the population quintile. In Fig. 1 (and Supplementary Fig. 1) these estimated \nelasticities are plotted against the (log) GDP per capita of the country-year on \nwhich the study is based, yk.\nThe result is an estimated relationship between the consumption elasticity of \nthe initial burden, \u03c9k, and the log of GDP per capita, log yk : \u03c9k = \u02c6\u03b1 + \u02c6\u03b2 log yk.\nThe analysis is described in more detail in Section 1 of the Supplementary \nInformation.\nNature Climate Change | www.nature.com/natureclimatechange\n\nAnalysis\nNature Climate Change\nTo project elasticities, \u03c9it, for each region and period in the model, we compute \nthe predicted elasticities \u02c6\u03c9it = \u02c6\u03b1 + \u02c6\u03b2yit according to the regression above for the \nmodel GDP per capita, yit, of region j in period t.\nData availability\nAll data used in our version of the model are archived53 and freely available at \nhttps://github.com/Environment-Research/revenue_recycling.\nCode availability\nAll model code used to generate results and create figures for this article is \narchived53 and freely available at https://github.com/Environment-Research/\nrevenue_recycling.\nReferences\n\t53.\tBudolfson, M. et al. Climate action with revenue recycling has benefits for \npoverty, inequality, and wellbeing. Zenodo https://doi.org/10.5281/\nzenodo.5552749 (2021).\n\t54.\tNordhaus, W. D. Economic aspects of global warming in a post-Copenhagen \nenvironment. Proc. Natl Acad. Sci. USA 107, 11721\u201311726 (2010).\n\t55.\tTechnical Update on the Social Cost of Carbon for Regulatory Impact \nAnalysis\u2014Under Executive Order 12866 (Interagency Working Group on \nSocial Cost of Carbon, United States Government, 2013).\n\t56.\tPindyck, R. S. Climate change policy: what do the models tell us? J. Econ. Lit. \n51, 860\u2013872 (2013).\n\t57.\tWeyant, J. Some contributions of integrated assessment models of global \nclimate change. Rev. Environ. Econ. Policy 11, 115\u2013137 (2017).\n\t58.\tWorld Income Inequality Database (UNU-WIDER, accessed 6 May 2020); \nhttps://www.wider.unu.edu/project/wiid-world-income-inequality-database\nAcknowledgements\nThis article has received funding from the NAVIGATE project of the European Union\u2019s \nHorizon 2020 research and innovation programme under grant no. 821124 (S.F., M. \nFleurbaey, U.K., A.M., F.W. and S.Z.) and the NIEHS-funded HERCULES Center \nP30ES019776 (N.S.). We thank C. Burnham and the Climate Futures Initiative at \nPrinceton University for support.\nAuthor contributions\nM.B., F.D., F.E., U.K., K.K. and N.S. are co-lead authors and contributed equally to \nthe study. M.B., F.D., D.K., F.E., M. Ferranna, U.K., K.K., A.M., N.S. and S.Z. designed \nthe research. M.B., F.D., S.F., M. Ferranna, D.K., U.K., K.K., A.M. and S.Z. conducted \nthe literature review on the distributional impact of a carbon tax. F.D., F.E. and K.K. \nconducted the modelling. F.D., K.K., S.F., M. Ferranna, M. Fleurbaey, D.K., U.K., A.M. \nand S.Z. led the social welfare and tax analysis. M.B. and N.S. wrote the first draft of the \nmanuscript with contributions from U.K. and K.K. All authors interpreted the results and \nedited the manuscript.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary material \navailable at https://doi.org/10.1038/s41558-021-01217-0.\nCorrespondence and requests for materials should be addressed to Mark Budolfson.\nPeer review information Nature Climate Change thanks Allen Fawcett and the other, \nanonymous, reviewer(s) for their contribution to the peer review of this work.\nReprints and permissions information is available at www.nature.com/reprints.\nNature Climate Change | www.nature.com/natureclimatechange\n\n\n    Scientific Research Findings:", "answer": "We find that a 2 \u00b0C target can be met while simultaneously increasing wellbeing, reducing inequality and alleviating poverty if each country or region imposes a substantial carbon tax and refunds the revenues to its citizens on an equal per capita basis. Furthermore, the benefits are often large, particularly for those at the lower end of the income distribution. The overall benefits to society are even greater if total carbon tax revenues are returned on an equal per capita basis globally, which directs more of the revenues towards the poorest populations in the world (rather than the poorest within each country or region). We also find that the optimal decarbonization trajectory is characterized by rapid reductions in emissions initially \u2014 which limits runaway climate change and allocates substantial revenues to the current poor \u2014 followed by a slower climb towards net zero emissions, which preserves some tax revenues for future generations.", "id": 56}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41558-020-0847-4\n1Department of Earth and Environmental Sciences, University of Leuven (KU Leuven), Heverlee, Belgium. 2Ecosystems Services and Management Program, \nInternational Institute for Applied System Analysis (IIASA), Laxenburg, Austria. 3RTI International, Durham, NC, USA. 4College of Science and Engineering, \nRitsumeikan University, Kusatsu, Japan. 5United States Environmental Protection Agency, Washington, DC, USA. 6Department of Economics and Social \nSciences, University of Natural Resources and Life Sciences, Vienna, Austria. \u2709e-mail: charlotte.janssens@kuleuven.be\nA\npproximately 11% of the world population in 2017, or 821\u2009mil-\nlion people, suffered from hunger1. Undernourishment has \nbeen increasing since 2014 due to conflict, climate vari-\nability and extremes, and is most prevalent in sub-Saharan Africa \n(23.2% of population), the Caribbean (16.5%) and Southern Asia \n(14.8%)1. Climate change is projected to raise agricultural prices2 \nand to expose an additional 77\u2009million people to hunger risks by \n2050 (ref. 3), thereby jeopardizing the UN Sustainable Development \nGoal to end global hunger4. Adaptation policies to safeguard food \nsecurity range from new crop varieties and climate-smart farming \nto reallocation of agricultural production2,5.\nInternational trade can be an important adaptation mechanism6,7. \nTrade links countries with a food deficit with countries with a food \nsurplus and raises consumption possibilities through specialization \naccording to comparative advantage. Climate change affects regions \nand crops differently8, possibly shifting regional comparative \nadvantages and altering trade patterns. Studies report that restrict-\ning trade exacerbates the impact of climate change on agricultural \nproduction, whereas liberalizing trade alleviates it9\u201314. However, the \ncurrent literature is incomplete in its scenario design and does not \ncomprehensively assess whether and\u2014if so, why\u2014the role of trade \nbecomes larger under climate change (see Methods; Supplementary \nText). The \u2018adaptation illusion hypothesis\u2019 argues that many farm \npractices are wrongly identified as adaptation because they have \nequal beneficial impacts with or without climate change15,16. Here \nwe investigated the case of adaptation through trade, and reveal \nwhether climate change alters the pattern of comparative advantage \nand increases the impact of trade integration on hunger. With the \nemerging integration between climate and trade policy agendas17, \na better understanding is needed to guide international policies to \nreduce hunger.\nPrevailing trade barriers may affect the adaptation potential of \ntrade. Border protection is widespread and has an important influ-\nence on agrifood trade18,19. Despite substantial liberalization efforts \nunder the ongoing Doha Round, tariffs remain high for agricultural \nproducts20. We investigated the impact of pre-Doha tariff levels as \nwell as further liberalization of agricultural tariffs. Other trade costs \nassociated with infrastructure, logistics and custom procedures are \nhigh, particularly in agricultural trade and in developing countries21. \nReducing such barriers could create larger trade gains than reduc-\ntions in border protection18. We compared the adaptation potential \nof trade liberalization, through the reduction in tariff barriers, and \ntrade facilitation, through the reduction in other trade costs.\nWe focused on global hunger projections to 2050 and analysed \nhow climate change and trade interact in their impact on hunger. \nOur economic (Global Biosphere Management Model (GLOBIOM)) \nand crop (Environment Policy Integrated Model (EPIC)) modelling \napproach (see Methods) is well established for investigating agricul-\ntural climate change impacts22\u201325. We advance on the current litera-\nture by analysing 60 integrated scenarios that capture variability in \ntrade barriers and in climate projections originating from general \ncirculation models (GCMs), emissions scenarios (Representative \nConcentration Pathways (RCPs)) and assumptions about CO2 fer-\ntilization. By statistically analysing the scenario sample, we assessed \nwhether, where and how climate change influences the effect of \ntrade on the risk of hunger.\nThe adaptive effect of international trade on global hunger\nBuilding on a previous study24, we used ten climate change and six \ntrade scenarios, and analysed hunger effects at the global and regional \nlevels. Four RCPs (2.6\u2009W\u2009m\u22122, 4.5\u2009W\u2009m\u22122, 6.0\u2009W\u2009m\u22122 and 8.5\u2009W\u2009m\u22122) \nare projected by HadGEM2\u2013ES. RCP\u20098.5 is also implemented with \nGlobal hunger and climate change adaptation \nthrough international trade\nCharlotte Janssens\u200a \u200a1,2\u2009\u2709, Petr Havl\u00edk2, Tam\u00e1s Krisztin2, Justin Baker\u200a \u200a3, Stefan Frank2, \nTomoko Hasegawa\u200a \u200a2,4, David Lecl\u00e8re\u200a \u200a2, Sara Ohrel\u200a \u200a5, Shaun Ragnauth\u200a \u200a5, Erwin Schmid\u200a \u200a6, \nHugo Valin\u200a \u200a2, Nicole Van Lipzig1 and Miet Maertens\u200a \u200a1\nInternational trade enables us to exploit regional differences in climate change impacts and is increasingly regarded as a \npotential adaptation mechanism. Here, we focus on hunger reduction through international trade under alternative trade sce-\nnarios for a wide range of climate futures. Under the current level of trade integration, climate change would lead to up to \n55\u2009million people who are undernourished in 2050. Without adaptation through trade, the impacts of global climate change \nwould increase to 73\u2009million people who are undernourished (+33%). Reduction in tariffs as well as institutional and infra-\nstructural barriers would decrease the negative impact to 20\u2009million (\u221264%) people. We assess the adaptation effect of trade \nand climate-induced specialization patterns. The adaptation effect is strongest for hunger-affected import-dependent regions. \nHowever, in hunger-affected export-oriented regions, partial trade integration can lead to increased exports at the expense of \ndomestic food availability. Although trade integration is a key component of adaptation, it needs sensitive implementation to \nbenefit all regions.\nNature Climate Change | VOL 10 | September 2020 | 829\u2013835 | www.nature.com/natureclimatechange\n829\n\nArticles\nNature Climate Change\nfour alternative climate models (GFDL\u2013ESM2M, NorESM1\u2013M, \nIPSL\u2013CM5A\u2013LR and MIROC\u2013ESM\u2013CHEM). RCP\u20092.6 represents \nclimate stabilization at 2\u2009\u00b0C, whereas RCP\u20098.5 represents a probable \ntemperature range of 2.6\u20134.8\u2009\u00b0C (ref. 26). We compared the stron-\ngest climate change impacts (RCP\u20098.5) with the intermediate climate \nscenarios (RCP\u20092.6 to RCP\u20096.0). EPIC projects yields for climatic \nconditions of each RCP\u2009\u00d7\u2009GCM combination including CO2 fertil-\nization that are compared to yields without climate change impacts \n(no climate change scenario). RCP\u20098.5\u2009\u00d7\u2009HadGEM2\u2013ES was also \nrun without CO2 fertilization effects, representing the worst pos-\nsible outcome. Our approach follows the ISI-MIP (www.isimip.org) \nFast Track Protocol, which considers scenarios with CO2 fertiliza-\ntion as the default, and prioritizes RCP\u20098.5\u2009\u00d7\u2009HadGEM2\u2013ES for CO2 \nsensitivity analyses. We provide a complete CO2 sensitivity analy-\nsis across RCPs in the Supplementary Text. In the baseline trade \nscenario, trade barriers were kept constant at 2010 level, but trade \npatterns vary endogenously across different climate impact scenar-\nios. The fixed imports scenario prevents agricultural imports from \nexceeding levels from the no climate change scenario. The pre-Doha \ntariffs scenario represents the trade environment before global trade \nliberalization launched by the Doha Round. In the facilitation sce-\nnario, additional costs from expanding trade volume beyond the \ncurrent level (for example infrastructure costs) were set close to \nzero. Under the tariff elimination scenario agricultural tariffs were \nprogressively phased out from \u221225% in 2020 to \u2212100% in 2050. The \nfacilitation\u2009+\u2009tariff elimination scenario combines the previous two \nscenarios. Socioeconomic developments were modelled with the \nsecond Shared Socioeconomic Pathway (SSP2)27. The scenarios are \ndiscussed further in the Methods.\nThrough adjustments in trade, supply and demand, the 2050 \nglobal population at risk of hunger under climate change and trade \nscenarios deviates substantially from the SSP2 baseline (the base-\nline trade\u2009+\u2009no climate change scenario; Fig. 1). Lower trade costs \nreduce importer prices, increase traded quantities and/or increase \nexporter prices, whereas lower climate-induced crop yields increase \nprices. On the supply side, this influences the optimal land alloca-\ntion within each pixel in terms of land cover, crop and management  \nsystem. On the demand side, regions determine the optimal level of \nconsumption and trade of each product in response to new price lev-\nels. Within-country distributional impacts of price changes through \nagricultural income effects were not considered (see Methods). \nIn the baseline trade scenario, price changes across RCP\u20098.5 sce-\nnarios lead to a reduction in global food availability of \u22120.2% to \n\u22123% compared with the baseline. The corresponding hunger \neffects are large\u2014an additional 7\u201355\u2009million people are projected \nto become undernourished (+6% to +45%). Across the RCP\u20098.5 \nscenarios, global cropland area changes by \u22122% to +3% and the \nshare of irrigated area increases from +1% to +7%. Total agricul-\ntural trade volume increases by +1% to +7% across RCP\u20098.5 sce-\nnarios through an expansion at the intensive and extensive margin \n(new flows representing 1\u20133% of total trade volume; Supplementary \nTable 1). Hunger impacts under intermediate climate change range \nfrom a decrease of 1\u2009million to an increase of 14\u2009million undernour-\nished people. In RCP\u20092.6, undernourishment is lower than in the \nno climate change scenario because crop yields in several regions \nincrease or remain unaffected partly due to the CO2 fertilization \neffect (Extended Data Fig. 1, Supplementary Fig. 12). When adap-\ntation through trade is constrained in the fixed imports scenario, \nhunger exacerbates across all of the RCP\u20098.5 scenarios, up to an \nadditional 73\u2009million undernourished people compared with the \nbaseline (+60%). By preventing endogenous market responses to \nclimate change, the fixed imports scenario results in lower global \ncrop production efficiency (\u22121% to \u22122.5%), lower global food avail-\nability (\u221210 to \u221237\u2009kcal per capita per day) and higher agricultural \nprices (+2% to +17%) across the RCP\u20098.5 scenarios compared with \nthe baseline trade scenario (Supplementary Table 2). The pre-Doha \ntariffs scenario leads to up to 81\u2009million additional undernourished \npeople compared with the baseline scenario (+67%), highlighting \nthe importance of trade integration that has already been achieved \nthrough the Doha Round in alleviating the potential long-term \nimpacts of climate change on hunger.\nThe facilitation and tariff elimination scenarios reduce the global \nrisk of hunger from climate change to a comparable extent, and the \nfacilitation\u2009+\u2009tariff elimination scenario can even compensate for \n0\n50\n100\n150\n200\nBaseline trade\nFixed imports\nPre-Doha tariffs\nFacilitation\nTariff elimination\nFacilitation +\ntariff elimination\nPopulation at risk of hunger (\u00d7106)\nGCM:\nNone\nGFDL\u2013ESM2M\nIPSL\u2013CM5A\u2013LR\nMIROC\u2013ESM\u2013CHEM\nNorESM1\u2013M\nHadGEM2\u2013ES\nClimate scenario:\nNo CC\nRCP 2.6\nRCP 4.5\nRCP 6.0\nRCP 8.5\nRCP 8.5 without the CO2 effect\nFig. 1 | Global population at risk of hunger in 2050 across climate change and trade scenarios. Climate change scenarios include the effect of CO2 \nfertilization on crop yields. RCP\u20098.5 is implemented with and without the CO2 effect. The black dotted horizontal line indicates the population at risk of \nhunger in the SSP2 baseline (122\u2009million).\nNature Climate Change | VOL 10 | September 2020 | 829\u2013835 | www.nature.com/natureclimatechange\n830\n\nArticles\nNature Climate Change\nthe impact of all but the most extreme climate change scenario. \nTrade liberalization and facilitation reduce hunger by enhancing \nclimate-induced trade adjustments\u2014across RCP 8.5 scenarios, total \nagricultural trade increases by 166% to 262% under the facilita-\ntion\u2009+\u2009tariff elimination scenario\u2014by reducing agricultural prices, \nand by increasing food availability and crop production efficiency \n(Supplementary Tables 1 and 2). The hunger effect under extreme \nclimate change (RCP\u20098.5 without the CO2 effect) is reduced by 31% \nunder the facilitation scenario, 11% under the tariff elimination sce-\nnario and 64% under the facilitation\u2009+\u2009tariff elimination scenario. \nThese effects are consistent with other studies that reported 44% \nlower hunger effects under market integration13 and 46% lower \nprice effects under trade liberalization10 (Supplementary Fig. 5).\nRegional perspective on climate change, hunger and trade\nThe hunger outcomes of the climate and trade scenarios differ \nsubstantially among the hunger-affected regions (Fig. 2). Climate \nchange has little impact on regions facing positive or small nega-\ntive crop yield impacts (Russia and West Asia (CSI), and the \nMiddle-East and North-Africa (MNA)) or maintaining a high crop \nyield (Latin American countries (LAC); Extended Data Fig. 1 (for \naverage crop yield), Supplementary Figs. 1\u20134 (for the four main \ncrops)). Regions with negative impacts on medium crop yields face \nlarger hunger impacts (East Asia (EAS) and Southeast Asia (SEA)). \nSouth Asia (SAS) and sub-Saharan Africa (SSA) face the most \nsevere hunger impacts from climate change. They experience nega-\ntive impacts on already low yields, also when including the impact \nof supply-side adaptation on yields (Extended Data Fig. 2). Across \nthe RCP\u20098.5 scenarios, projections for the baseline trade scenario \nrange from an increase of 13\u2013181% and 2\u201351% in the population at \nrisk of hunger for SAS and SSA, respectively. The effect of the trade \nscenarios on regional undernourishment is largest among baseline \nnet-importing regions (SSA, MNA, EAS and SAS) and regions in \nwhich climate change reduces net exports (SEA; Extended Data \nFigs. 3 and 4). The fixed imports scenario enlarges hunger impacts \nin the extreme climate change scenario in SSA, SAS and SEA by rais-\ning agricultural prices (Extended Data Figs. 5 and 6), increasing net \nexports in SEA, and reducing net imports in SSA and SAS. Adverse \neffects from trade restriction, such as the export bans observed \nduring the 2007\u20132008 world food crisis28,29 and those feared as a \nresult of the global COVID-19 pandemic30,31, may pose severe \nhunger risks under climate change. Under the pre-Doha tariffs \nscenario undernourishment in SSA, SAS and EAS is substantially \nhigher compared with the baseline trade scenario. Tariff liberaliza-\ntion between 2001 and 2010 reduced average import tariffs in SSA, \nSAS and EAS by around 30% (Supplementary Table 6). The lower  \n0\n10\nBaseline\nPopulation at\nrisk of hunger (\u00d710\n6)\nFixed \nimports\nPre-Doha\ntariffs\nFac. +\ntariff el.\nCAN\n0\n10\nCSI\n0\n10\n20\nEAS\n0\n10\nEUR\n0\n10\nLAC\n0\n10\n20\nMNA\n0\n10\nOCE\n0\n10\n20\n30\n40\n50\nSAS\n0\n10\n20\nSEA\n0\n10\n20\n30\n40\n50\n60\nSSA\n0\n10\nUSA\nRegion\nCAN\nCSI\nEAS\nEUR\nLAC\nMNA\nOCE\nSAS\nSEA\nSSA\nUSA\nClimate scenario\nNo CC\nRCP 2.6\nRCP 4.5\nRCP 6.0\nRCP 8.5\nRCP 8.5 without CO2 effect\nBaseline Fixed \nimports\nPre-Doha\ntariffs\nFac. +\ntariff el.\nBaseline Fixed \nimports\nPre-Doha\ntariffs\nFac. +\ntariff el.\nBaseline\nFixed \nimports\nPre-Doha\ntariffs\nFac. +\ntariff el.\nBaseline Fixed \nimports\nPre-Doha\ntariffs\nFac. +\ntariff el.\nBaseline Fixed \nimports\nPre-Doha\ntariffs\nFac. +\ntariff el.\nBaseline Fixed \nimports\nPre-Doha\ntariffs\nFac. +\ntariff el.\nBaseline Fixed \nimports\nPre-Doha\ntariffs\nFac. +\ntariff el.\nBaseline Fixed \nimports\nPre-Doha\ntariffs\nFac. +\ntariff el.\nBaseline Fixed \nimports\nPre-Doha\ntariffs\nFac. +\ntariff el.\nBaseline Fixed \nimports\nPre-Doha\ntariffs\nFac. +\ntariff el.\nPopulation at\nrisk of hunger (\u00d710\n6)\nPopulation at\nrisk of hunger (\u00d710\n6)\nPopulation at\nrisk of hunger (\u00d710\n6)\nPopulation at\nrisk of hunger (\u00d710\n6)\nPopulation at\nrisk of hunger (\u00d710\n6)\nPopulation at\nrisk of hunger (\u00d710\n6)\nPopulation at\nrisk of hunger (\u00d710\n6)\nPopulation at\nrisk of hunger (\u00d710\n6)\nPopulation at\nrisk of hunger (\u00d710\n6)\nPopulation at\nrisk of hunger (\u00d710\n6)\nFig. 2 | Population at risk of hunger in 2050 across climate change and trade scenarios in each region. The results from the GCM HadGEM2\u2013ES are \nshown; the full scenario set is provided in Extended Data Fig. 7. The following regions are included: USA, Russia and West Asia (CSI), East Asia (EAS), \nSoutheast Asia (SEA), South Asia (SAS), Middle East and North Africa (MNA), sub-Saharan Africa (SSA), Latin American Countries (LAC), Oceania \n(OCE), Canada (CAN) and Europe (EUR). The black dotted horizontal lines indicate the population at risk of hunger in the SSP2 baseline. Fac.\u2009+\u2009tariff el., \nfacilitation\u2009+\u2009tariff elimination.\nNature Climate Change | VOL 10 | September 2020 | 829\u2013835 | www.nature.com/natureclimatechange\n831\n\nArticles\nNature Climate Change\ntariffs reduce the overall level of trade costs by 2050 (Supplementary \nTable 7) and enable larger agricultural net imports in SSA, SAS \nand EAS across all climate scenarios (Extended Data Fig. 3). In \nthe MNA region, the average import tariff reduced marginally and \nin SEA it was already low (Supplementary Table 6). The facilica-\ntion\u2009+\u2009tariff elimination scenario reduces hunger in the SSA, MNA \nand EAS regions across all climate scenarios by decreasing average \ntrade costs (Supplementary Table 7), thereby reducing agricultural  \nprices and raising agricultural imports (Extended Data Figs. 3  \nand 5). In some cases, trade integration increases rather than \ndecreases the level of undernourishment in a region under cli-\nmate change. The largest adverse effects occur under the tariff \nelimination scenario in the SEA and SAS regions (Extended Data \nFig. 7). Whereas the facilitation scenario reduces hunger in the \nextreme climate change scenario by 16% and 8%, the tariff elimi-\nnation scenario increases hunger impacts by 4% and 16% in SEA \nand SAS, respectively. Both trade scenarios reduce average trade \ncosts (Supplementary Table 7), but the tariff elimination scenario \nincreases rice exports from SAS and SEA, thereby reducing domes-\ntic calorie availability. The facilication\u2009+\u2009tariff elimination scenario \ncompensates for calorie loss from rice exports through increased \nimports of other agricultural goods and decreases the hunger effect \nof extreme climate change by 26% and 11% in SEA and SAS, respec-\ntively. Our sensitivity analysis shows that the effects of trade on \nclimate-induced hunger are robust to CO2 fertilization assumptions \n(Supplementary Figs. 13 and 14).\nA larger role for trade under climate change\nTo reveal whether the effect of trade increases under climate change \nand, therefore, has a real adaptation role, we analysed hunger out-\ncomes from GLOBIOM on crop yield shifts projected by EPIC and \naverage trade costs in regional-level regression models (Table 1). \nWe interpret these results for a 5.4% reduction in crop yields and \na 23% reduction in average trade costs, which correspond to the \naverage impacts of climate change and the trade integration scenar-\nios, respectively. Regression results revealed that a 5.4% reduction \nin crop yields within a region leads to an average food availability \nreduction of 11\u2009kcal per capita per day (95% confidence interval \n(CI), 15\u20138\u2009kcal per capita per day) and an additional 0.52\u2009million \npeople at risk of hunger (CI\u2009=\u20090.25\u20130.79\u2009million). For a 23% decrease \nin trade costs, we project an increase in average food availability \nwithin a region of 13\u2009kcal per capita per day (CI\u2009=\u20099\u201316\u2009kcal per cap-\nita per day) and a decrease in undernourished people of 1.22\u2009million \n(CI\u2009=\u20091.52\u20130.93\u2009million). When excluding regions that experience \nnegative impacts in some trade scenarios (SAS and SEA), we found \na significant negative interaction effect between trade costs and crop \nyields (P\u2009=\u20090.014). For example, under extreme climate change (that \nis, a 20% crop yield reduction), the positive effect of a 23% reduc-\ntion in trade costs is 1.97\u2009million fewer people undernourished, \nconsisting of a direct (\u22121.50\u2009million) and a climate-induced trade \neffect (\u22120.47\u2009million). These results confirm the existence of posi-\ntive trade effects on food availability and hunger alleviation13,32 and \nreveal an additional climate-induced effect of lowering trade costs.\nWe ran the regressions presented in Table 1 with regional inter-\naction effects (Supplementary Table 3). In most of the regions, \nclimate-induced decreases in crop yields reduce food availability \nand increase hunger while reduced trade costs have opposite effects. \nThe food availability impacts of crop-yield changes are largest \nfor SAS, SSA and SEA, whereas the effect of trade costs is largest \nfor regions maintaining net imports under climate change (SSA, \nMNA and EAS). The corresponding impact on hunger is largest \nin low-income regions (SSA and SAS), followed by middle-income \nregions (EAS, MNA, and SEA). The interaction effect, which reveals \nwhether climate change alters the relationship between trade costs \nand hunger, is most pronounced in SSA, followed by EAS. Figure 3 \nplots the predicted hunger\u2013yield relationship in EAS and SSA for \ndifferent levels of trade cost, showing that hunger is less sensitive to \nclimate-induced yield changes under reduced trade costs.\nInter-regional specialization\nWe assessed the extent that climate change shifts the pattern of com-\nparative advantage of four important crops (corn, wheat, soya and \nrice; Fig. 4). Consistent with Ricardo\u2019s theory of comparative advan-\ntage, a region is regarded as having a comparative advantage when it \nspecializes in a certain crop, such that its share of world production \nTable 1 | Results from OLS estimation of the impact of crop \nyields, trade costs and their interaction on regional hunger and \nfood availability\nPopulation at risk of \nhunger (million)\nFood availability (kcal per \ncapita per day)\nSample\n(1) All \nregions\n(2) \nWithout \nSAS and \nSEA\n(1) All \nregions\n(2) Without \nSAS and \nSEA\nCrop yield \n(percentage change)\n\u22129.70***\n\u22121.80\n213.00***\n173.00***\n(2.60)\n(1.40)\n(29.00)\n(31.00)\nTrade cost \n(log[US$/106\u2009kcal])\n4.70***\n5.80***\n\u221249.00***\n\u221280.00***\n(0.58)\n(0.73)\n(7.40)\n(9.40)\nCrop yield\u2009\u00d7\u2009trade \ncost\n3.30\n\u22128.90**\n14.00\n191.00***\n(6.20)\n(3.60)\n(60.00)\n(74.00)\n*P\u2009<\u20090.1; **P\u2009<\u20090.05; ***P\u2009<\u20090.01. Regional fixed effects are included. The values in brackets show \nthe heteroskedastic robust s.e.; n\u2009=\u2009550 (1) and n\u2009=\u2009450 (2). Adjusted R2\u2009=\u20090.890 (1) and adjusted \nR2\u2009=\u20090.930 (2) for the hunger regressions; and adjusted R2\u2009=\u20090.950 (1) and adjusted R2\u2009=\u20090.920 (2) \nfor the food availability regressions. Observations are GLOBIOM output for the 11 world regions \nunder 5 different trade scenarios (baseline, pre-Doha tariffs, facilitation, tariff elimination and \nfacilitation\u2009+\u2009tariff elimination) and ten climate change scenarios in 2050. The regression models \nare described in the Methods.\nEAS\nSSA\n\u22120.4\n\u22120.2\n0\n0.2\n0.4 \u22120.4\n\u22120.2\n0\n0.2\n0.4\n\u221240\n0\n40\n80\n120\nCrop yield change due to climate change (decimal fraction)\nPopulation at risk of hunger (\u00d7106)\nTrade cost\nFirst decile\nMedian\nNinth decile\nFig. 3 | Fitted linear response of population at risk of hunger to \nclimate-induced crop yield change in EAS and SSA for different values \nof trade costs. The shaded areas indicate the 95% prediction intervals. \nPrediction on the basis of an ordinary least squares (OLS) estimation of \nthe regional level linear regression of the impact of crop yield change, trade \ncosts and their interaction on population at risk of hunger. The regression \nresults are shown in Supplementary Table 3 and the regression model \nis described in the Methods. The fitted response for all of the regions is \nshown in Extended Data Fig. 8.\nNature Climate Change | VOL 10 | September 2020 | 829\u2013835 | www.nature.com/natureclimatechange\n832\n\nArticles\nNature Climate Change\nincreases when trade costs decrease (see Methods; Supplementary \nText). Under no climate change, trade integration increases the \nglobal production share of the United States (USA) in corn; LAC in \nsoya; CSI, Europe (EUR) and LAC in wheat; and SAS and EAS in \nrice (Fig. 4a). Trade integration has similar impacts on specializa-\ntion under climate change (Fig. 4b). Figure 4c compares the spe-\ncialization of regions in response to trade-cost reduction; negative \nvalues indicate decreases and positive values indicate increases in \ncomparative advantage under climate change compared with no cli-\nmate change. For example, MNA still decreases its share of global \nwheat production in response to trade integration under climate \nchange, but to a lesser extent than under no climate change. The \nsmall and mainly insignificant values indicate that the pattern of \ncomparative advantage of the four crops remains similar under cli-\nmate change. Although climate change affects crop yields and cost \ncompetitiveness of regions, it does not substantially alter the relative \nposition between regions (Supplementary Figs. 8\u201310). This finding \nis corroborated by alternative indicators of comparative advantage, \nincluding crop shares in a region\u2019s total production, export shares in \na region\u2019s crop production and the revealed comparative advantage \n(RCA) index (Supplementary Figs. 6, 7 and 11).\nAdaptation to climate change occurs through changes in exist-\ning and new inter-regional trade flows (Supplementary Tables 8\u201311). \nAcross the RCP\u20098.5 scenarios, the largest export growth originates \nfrom major baseline producing regions (corn from USA and LAC, \nsoya from LAC and USA, rice from SAS and SEA, and wheat from \nEUR and Canada (CAN); Supplementary Fig. 9). The largest new \ntrade flows are new corn exports from USA to EAS, CAN, LAC and \nSEA, from EUR to MNA and from LAC to EAS; new soya exports \nfrom LAC to SAS and from USA to CAN and MNA; and new wheat \nexports from CSI to EUR, and from MNA to SSA. Climate change \ndoes not induce substantial new rice trade flows. There is uncertainty \nacross RCP\u20098.5 scenarios in bilateral trade patterns, but several exports \nto hunger-affected regions increase consistently (such as wheat from \nEUR to SSA, soya from LAC to SAS, or corn from LAC to MNA). \nHowever, hunger-affected regions are not only engaging in trade at \nthe importer side, but also increase certain exports (wheat in MNA, \ncorn in SSA, and rice in EAS and SAS; Extended Data Fig. 10).\nNo climate change\na\nb\nc\nCorn\nRice\nSoya\nWheat\n0\n20\n40\n60\n80\nShare of global production (%)\nTrade scenario\nBaseline\nFacilitation + tariff elimination\nClimate change\nCC \u2013 no CC\n\u22120.2\n0.0\n0.2\n\u22120.2\n0.0\n0.2\n0\n20\n40\n60\n80\n\u22120.2\n0.0\n0.2\n\u22120.2\n0.0\n0.2\n0\n20\n40\n60\n80\n\u22120.2\n0.0\n0.2\n\u22120.2\n0.0\n0.2\n0\n20\n40\n60\n80\n\u22120.2\n0.0\n0.2\n\u22120.2\n0.0\n0.2\nImpact of 1% trade-cost reduction\non share of global production (%)\nImpact of 1% trade-cost reduction\non share of global production (%)\nRegion\nUSA\nCAN\nEUR\nCSI\nEAS\nSEA\nSAS\nMNA\nSSA\nLAC\nOCE\nFig. 4 | Inter-regional specialization in corn, rice, soya and wheat in response to trade-cost reduction in 2050. a, The share of global production under \nno climate change in the baseline trade and facilitation\u2009+\u2009tariff elimination scenarios. b,c, The impact of 1% trade-cost reduction on the share of global \nproduction where the dependent variable is either the outcome under climate change (b) or the difference in outcome between climate change (CC) and \nno climate change (c). Each point shows the estimated impact of a 1% trade-cost reduction for each region on share of world production (%), with lines \ndenoting the corresponding 95% confidence interval (heteroskedastic robust s.e.). The regression models are described in the Methods.\nNature Climate Change | VOL 10 | September 2020 | 829\u2013835 | www.nature.com/natureclimatechange\n833\n\nArticles\nNature Climate Change\nDiscussion\nInternational trade contributes globally to climate change adapta-\ntion. The impact of the worst climate change scenarios on global \nrisk of hunger increases by 33\u201347% under restricted trade scenar-\nios, and decreases by 11\u201364% under open trade scenarios. The gain \nfrom reducing trade costs is largest for regions that remain import \ndependent under climate change. Climate change increases the role \nof trade in reducing the risk of hunger for some regions, although \nit does not substantially alter the pattern of comparative advantage \nof main staple crops. It is the ability to link food surplus with defi-\ncit regions that underpins trade\u2019s adaptation effect. These conclu-\nsions are robust across RCPs, and independent from the assumption \non CO2 fertilization effects. Finally, we found that the number of \nundernourished people increases with climate change, irrespective \nof trade scenarios. Thus, climate change mitigation remains crucial \nfor eradicating hunger.\nOur study is comprehensive in its scenario design and rigorous \nin its analysis of the processes driving adaptation through trade. \nNevertheless, it is important to emphasize that this study focuses \non the global scale in the long term. Trade policies and climate \nchange have important within-country distributional consequences \nthrough income and food-access effects33\u201335, which are theoretically \nambiguous and which our modelling approach does not consider. \nAcross households with different food access channels, from urban \nnet-consumers to rural subsistence farmers, impacts can differ even \nin their direction34. Also, current global studies, including ours, \nfocus on crop- and grass-yield impacts, and other direct and indi-\nrect climate change effects are not represented to date\u2014for example, \nheat stress on animals, pest and disease incidence, sea level rise or \nreduced pollination. Finally, we take a long-term equilibrium per-\nspective ignoring the negative effects of extreme weather events. All \nof these aspects require substantial new research.\nDespite the limitations described above, our study brings novel \npolicy implications. We found that liberalization that has already \nbeen achieved under the Doha Round substantially reduces \nclimate-induced hunger impacts. A careful approach to trade \nintegration covering different types of trade barriers can further \nlimit hunger risks. The full removal of agricultural tariffs leads \nto increases in food availability in SSA, MNA and EAS, but may \nincrease exports and lower regional food availability in SEA and \nSAS. Further trade facilitation can reduce undernourishment in \nall hunger-affected regions. However, the effective realization of \ntrade facilitation requires considerable investments in transport \ninfrastructure and technology. Especially in low-income regions, \nsuch as SSA, infrastructure is weak36. An estimated US$130\u2013170 \nbillion a year is needed to bridge the infrastructure gap in SSA \nby 2025 (ref.37). Infrastructure finance averaged US$75 billion in \nrecent years, with the largest contribution from budget-constrained \nnational governments37. Alternative financing through institutional \nand private investments, called for by the African Development \nBank Group and the World Bank Group36,37, could be not only cru-\ncial for economic growth, but also for climate change adaptation. In \nessence, our results demonstrate that trade instruments can mitigate \nan important part of the adverse hunger effects of long-term cli-\nmate change. Our results thereby confirm the importance of holistic \napproaches to international trade negotiations, and could prove also \nrelevant in the face of trade-policy reactions in more acute crisis \nsituations, such as the global COVID-19 pandemic.\nOnline content\nAny methods, additional references, Nature Research report-\ning summaries, source data, extended data, supplementary infor-\nmation, acknowledgements, peer review information; details of \nauthor contributions and competing interests; and statements of \ndata and code availability are available at https://doi.org/10.1038/\ns41558-020-0847-4.\nReceived: 19 July 2019; Accepted: 16 June 2020;  \nPublished online: 20 July 2020\nReferences\n\t1.\t FAO, IFAD, UNICEF, WFP & WHO The State of Food Security and Nutrition \nin the World 2018. Building Climate Resilience for Food Security and Nutrition \n(FAO, 2018).\n\t2.\t Nelson, G. C. et\u00a0al. Climate change effects on agriculture: economic responses \nto biophysical shocks. Proc. Natl Acad. Sci. USA 111, 3274\u20133279 (2014).\n\t3.\t 2019 Global Food Policy Report (IFPRI, 2019).\n\t4.\t Hoegh-Guldberg, O. et\u00a0al. in Special Report on Global Warming of 1.5\u2009\u00b0C  \n(eds Masson-Delmotte, V. et\u00a0al.) Ch. 3 (WMO, 2018).\n\t5.\t Hertel, T. W. Climate Change, Agricultural Trade and Global Food Security. \nThe State of Agricultural Commodity Markets (SOCO) 2018: Background Paper \n9 (FAO, 2018).\n\t6.\t Huang, H., von Lampe, M. & van Tongeren, F. Climate change and trade in \nagriculture. Food Policy 36, S9\u2013S13 (2011).\n\t7.\t Brown, M. E. et\u00a0al. Do markets and trade help or hurt the global food system \nadapt to climate change? Food Policy 68, 154\u2013159 (2017).\n\t8.\t Rosenzweig, C. et\u00a0al. Assessing agricultural risks of climate change in the 21st \ncentury in a global gridded crop model intercomparison. Proc. Natl Acad. Sci. \nUSA 111, 3268\u20133273 (2014).\n\t9.\t Stevanovi\u0107, M. et\u00a0al. The impact of high-end climate change on agricultural \nwelfare. Sci. Adv. 24, e1501452 (2016).\n\t10.\tWiebe, K. et\u00a0al. Climate Change impacts on agriculture in 2050 under a range \nof plausible socioeconomic and emissions scenarios. Environ. Res. Lett. 10, \n085010 (2015).\n\t11.\tGouel, C. & Laborde, D. The Crucial Role of International Trade in Adaptation \nto Climate Change Working Paper No. 25221 (National Bureau of Economic \nResearch, 2018).\n\t12.\tCostinot, A., Donaldson, D. & Smith, C. Evolving comparative advantage and \nthe impact of climate change in agricultural markets: evidence from 1.7 \nmillion fields around the world. J. Polit. Econ. 124, 205\u2013248 (2016).\n\t13.\tBaldos, U. L. C. & Hertel, T. W. The role of international trade in  \nmanaging food security risks from climate change. Food Secur. 7,  \n275\u2013290 (2015).\n\t14.\tCui, H. D., Kuiper, M., van Meijl, H. & Tabeau, A. Climate Change and \nGlobal Market integration: Implications for Global Economic Activities, \nAgricultural Commodities, and Food Security. The State of Agricultural \nCommodity Markets (SOCO) 2018: Background Paper (FAO, 2018).\n\t15.\tLobell, D. B. Climate change adaptation in crop production: beware of \nillusions. Glob. Food Sec. 3, 72\u201376 (2014).\n\t16.\tMoore, F. C., Baldos, U. L. C. & Hertel, T. Economic impacts of climate \nchange on agriculture: a comparison of process-based and statistical yield \nmodels. Environ. Res. Lett. 12, 065008 (2017).\n\t17.\tZimmermann, A., Benda, J., Webber, H. & Jafari, Y. Trade, Food Security and \nClimate Change: Conceptual Linkages and Policy Implications Background \nPaper for The State of Agricultural Commodity Markets (SOCO) 2018  \n(FAO, 2018).\n\t18.\tHoekman, B. & Nicita, A. Trade policy, trade costs, and developing country \ntrade. World Dev. 39, 2069\u20132079 (2011).\n\t19.\tDisdier, A. C. & van Tongeren, F. Non-tariff measures in agri-food trade: \nwhat do the data tell us? Evidence from a cluster analysis on OECD imports. \nAppl. Econ. Perspect. Policy 32, 436\u2013455 (2010).\n\t20.\tBureau, J., Guimbard, H. & Jean, S. Agricultural trade liberalisation in the \n21st century: has it done the business? J. Agric. Econ. 70, 3\u201325 (2018).\n\t21.\tArvis, J. F., Duval, Y., Shepherd, B., Utoktham, C. & Raj, A. Trade costs in the \ndeveloping world: 1996-2010. World Trade Rev. 15, 451\u2013474 (2016).\n\t22.\tMosnier, A. et\u00a0al. Global food markets, trade and the cost of climate change \nadaptation. Food Secur. 6, 29\u201344 (2014).\n\t23.\tLecl\u00e8re, D. et\u00a0al. Climate Change induced transformations of agricultural \nsystems: insights from a global model. Environ. Res. Lett. 9, 124018 (2014).\n\t24.\tBaker, J. et\u00a0al. Evaluating the effects of climate change on US agricultural \nsystems: sensitivity to regional impact and trade expansion scenarios. \nEnviron. Res. Lett. 13, 064019 (2018).\n\t25.\tHavl\u00edk, P. et\u00a0al. Climate Change Impacts and Mitigation in the Developing \nWorld, An Integrated Assessment of the Agriculture and Forestry Sectors. \nBackground Paper for the World Bank Report: \u201cShock Waves: Managing the \nImpacts of Climate Change on Poverty.\u201d Policy Research Working Paper \nWPS7477 (The World Bank, 2015).\n\t26.\tIPCC Climate Change 2014: Synthesis Report (eds Core Writing Team, \nPachauri, R. K. & Meyer L. A.) (IPCC, 2014).\n\t27.\tFricko, O. et\u00a0al. The marker quantification of the Shared Socioeconomic \nPathway 2: A middle-of-the-road scenario for the 21st century. Glob. Environ. \nChange 42, 251\u2013267 (2017).\n\t28.\tBou\u00ebt, A. & Laborde Debucquet, D. in Food Price Volatility and Its \nImplications for Food Security and Policy (eds Kalkuhl, M. et\u00a0al.) 167\u2013179 \n(Springer, 2016); https://doi.org/10.1007/978-3-319-28201-5_8\nNature Climate Change | VOL 10 | September 2020 | 829\u2013835 | www.nature.com/natureclimatechange\n834\n\nArticles\nNature Climate Change\n\t29.\tKornher, L. Maize Markets in Eastern and Southern Africa (ESA) in the \nContext of Climate Change. The State of Agricultural Commodity Markets \n(SOCO) 2018: Background Paper (FAO, 2018).\n\t30.\tGlauber, J., Laborde, D., Martin, W. & Vos, R. COVID-19: Trade Restrictions \nare Worst Possible Response to Safeguard Food Security (IFPRI, 2020); https://\nwww.ifpri.org/blog/covid-19-trade-restrictions-are-worst-possible-response- \nsafeguard-food-security\n\t31.\tSulser, T. & Dunston, S. COVID-19-Related Trade Restrictions on Rice and \nWheat Could Drive Up Prices and Increase Hunger (IFPRI, 2020); https://\nwww.ifpri.org/blog/covid-19-related-trade-restrictions-rice-and-wheat-could-\ndrive-prices-and-increase-hunger\n\t32.\tDithmer, J. & Abdulai, A. Does trade openness contribute to food security? A \ndynamic panel analysis. Food Policy 69, 218\u2013230 (2017).\n\t33.\tBureau, J. C. & Swinnen, J. EU policies and global food security. Glob. Food \nSec. 16, 106\u2013115 (2018).\n\t34.\tPorter, J. R. et\u00a0al. Food security and food production systems. in Climate \nChange 2014: Impacts, Adaptation, and Vulnerability (eds Field, C. B. et\u00a0al.) \n485\u2013533 (Cambridge Univ. Press, 2014).\n\t35.\tSwinnen, J. & Squicciarini, P. Mixed messages on prices and food security. \nScience 335, 405\u2013406 (2012).\n\t36.\tCalderon, C., Cantu, C. & Chuhan-Pole, P. Infrastructure Development in \nSub-Saharan Africa: A Scorecard Policy Research Working Paper WPS8425 \n(The World Bank, 2018).\n\t37.\tAfrican Economic Outlook 2018 (African Development Bank, 2018).\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2020\nNature Climate Change | VOL 10 | September 2020 | 829\u2013835 | www.nature.com/natureclimatechange\n835\n\nArticles\nNature Climate Change\nMethods\nModelling framework. We used the GLOBIOM, a recursive dynamic, spatially \nexplicit, economic partial equilibrium model of the agriculture, forestry and \nbioenergy sector with bilateral trade flows and costs that can model new trade \npatterns38. The model computes a market equilibrium in 10-year time steps \nfrom 2000 to 2050 by maximizing welfare (the sum of consumer and producer \nsurplus) subject to technological, resource and political constraints. In each time \nstep, market prices adjust endogenously to equalize supply and demand for each \nproduct and region. On the demand side, a representative consumer for each of 30 \neconomic regions optimizes consumption and trade in response to product prices \nand income. Food demand depends endogenously on product prices through an \nisoelastic demand function and exogenously on GDP and population projections39. \nWe mainly present model results aggregated to 11 regions (Supplementary Table 4):  \nUSA, CAN, EUR, OCE, SEA, SAS, SSA, MNA, EAS, CSI and LAC. GLOBIOM \nis a bottom-up model that builds on a high spatial grid-level resolution on the \nsupply side. Land is disaggregated into simulation units\u2014clusters of 5 arcmin \npixels that are created based on altitude, slope and soil class, 30 arcmin pixels, and \ncountry boundaries. GLOBIOM\u2019s crop production sector includes 18 major crops \n(barley, beans, cassava, chickpeas, corn, cotton, groundnut, millet, palm oil, potato, \nrapeseed, rice, soybean, sorghum, sugarcane, sunflower, sweet potato and wheat) \nunder 4 management systems (irrigated, high input; rainfed, high input; rainfed, \nlow input; and subsistence). The allocation of acreage by the crop and management \nsystem is determined by potential yields, production costs and expansion \nconstraints23. Crop production parameters are based on the detailed biophysical \ncrop model EPIC. Additional biophysical models were used to represent the \nlivestock (RUMINANT40) and forestry (G4M41) sectors. Further information on \nmodel structure and parameters was described previously42,43.\nAs a partial equilibrium model, GLOBIOM focuses only on specific sectors of \nthe economy and does not represent feedbacks on consumer income and GDP from \ntrade and climate change. However, the partial equilibrium model allows for more \ndetail in the represented sectors, and a more accurate assessment of biophysical \nimpacts. This is due to the high spatial and commodity resolution as well as the \nphysical rather than monetary representation of variables compared with the general \nequilibrium models that explicitly cover income feedbacks. Crop yields adjust \nendogenously through the management system or location of production, and \nexogenously according to long-term technological development and climate change \nimpacts23. The output from EPIC was used to compute, for each time step, yield \nshifters for each climate change scenario and each crop and management system at a \ndisaggregated spatial scale (simulation unit). EPIC simulates scenario-specific yields \non the basis of inputs from climate models (daily climatic conditions including solar \nradiation, minimum and maximum temperature, precipitation, wind speed, relative \nhumidity and CO2 concentration). Climate change impacts on livestock production \nare modelled through crop and grassland yield impacts on feed availability. EPIC \ncrop and grassland yield impacts, as well as their implementation in GLOBIOM, are \nfurther explained in Lecl\u00e8re et\u00a0al.23 and Baker et\u00a0al.24.\nInternational trade. International trade is represented in GLOBIOM through the \nEnke\u2013Samuelson\u2013Takayama\u2013Judge spatial equilibrium assuming homogenous \ngoods38,44. Bilateral trade flows between the 30 economic regions were determined \nby the initial trade pattern, relative production costs of regions and the \nminimization of trading costs38. The initial trade pattern was informed by the \nBACI database from CEPII averaging across 1998\u20132002 (ref. 45). Trade costs are \ncomposed of tariffs from the MAcMap-HS6 database46, transport costs47 and a \nnonlinear trade expansion cost. The MAcMap-HS6 2001 release from CEPII-ITC \nprovides ad valorem and specific tariffs, and shadow tariff rates of tariff rate quotas \nfor the model calibration in the base year 2000 (ref. 48). To incorporate trade \nliberalization developments under the Doha Round, the tariff data is updated in \nthe 2010 time step with the 2010 release of MAcMap-HS6 (ref. 49; Supplementary \nTable 6). We used the estimation by Hummels47 to compile input data on bilateral \ntransport costs on the basis of the distance between trade pairs and the weight\u2013\nvalue ratio of agricultural products. Transport costs were set to US$30 per ton \nminimum, on the basis of the fifth percentile of the OECD Maritime Transport \nCost database (2003\u20132007), and were kept constant at base year level over the \nsimulation period as the drivers of transport costs (for example, fuel prices and \ncontainerization50) are not represented in the partial equilibrium model. In the \nscenario simulations, the nonlinear expansion cost raises per-unit trade costs when \nthe traded quantity increases over time to model persistency in trade flows. A \nconstant elasticity function was used for trade flows observed in the base year, and \na quadratic function was used for new trade flows. The nonlinear element reflects \nthe cost of trade expansion in terms of infrastructure and capacity constraints in \nthe transport sector and was reset after each 10-year time step. Compared to other \nglobal economic models, GLOBIOM\u2019s trade representation is positioned between \nthe rigid Armington approach of general equilibrium models and the flexible world \npool market approach of many partial equilibrium models.\nRisk of hunger. We measured the population at risk of hunger, or the number \nof people whose food availability falls below the mean minimum dietary energy \nrequirement, on the basis of previous studies51\u201353. The following four parameters \nwere used: the mean minimum dietary energy requirement, the coefficient of \nvariation of the distribution of food within a country, the mean food availability \nin the country (kcal per capita per day) and total population. Minimum dietary \nenergy requirements are exogenously calculated on the basis of demographic \ncomposition (age, sex) of future population projections. Future changes in \nthe inequality of food distribution within a country are exogenous and follow \nprojected national income growth. This is based on an estimated relationship \nbetween income and the coefficient of variation of food distribution with \nobserved historical national-level data. Poor infrastructure, remoteness and a \nhigh prevalence of subsistence farming limit local markets in distributing food \nequally across households7. Income is lowest in SAS and SSA, regions in which \nthe share of land under subsistence farming is the largest (27% in SAS and 43% \nin SSA)54. Food availability in kcal per capita per day is endogenously determined \nby GLOBIOM at the regional level. One limitation of the approach is that it does \nnot include within-country distributional consequences of trade integration and/\nor climate change through income effects. Trade policies and climate change alter \nfood prices, which affects individual incomes, purchasing power and food access \ndepending on households being net consumers or net producers of food33. At the \naggregate regional level, the bias from not considering these distributional effects \nmay be upward or downward, depending on the share of net-consuming versus \nnet-producing households; degree of subsistence farming versus agricultural wage \nwork; and share of rural versus urban population in each country.\nClimate change adaptation. Climate change adaptation is defined by the IPCC \nas \u201cthe process of adjustment to actual or expected climate and its effects\u201d26. \nAdaptation of the agricultural sector to climate-induced changes in crop yields \nmay include adjustments in consumption, production and international trade2. \nDemand-side adaptation is captured in GLOBIOM by changes in regional \nconsumption levels in response to market prices. Supply-side adaptation includes \nthe reallocation of land for each crop by a grid-cell and management system, and \nthe expansion of cropland to other land covers23. Whereas Lecl\u00e8re et\u00a0al.23 assess \nsupply-side adaptation, here we focused on international market responses, in \nwhich our analysis approach is inspired by the \u2018adaptation illusion hypothesis\u2019 \npostulated by Lobell15 and confirmed by Moore et\u00a0al.55. They argue that farm-level \npractices identified as adaptation measures by many crop modelling studies \ncannot be referred to as climate adaptation as they have the same yield impact \nin current climate as under climate change. In a similar manner, we intended to \ninvestigate whether, where and, if so, why trade integration has a larger positive \nimpact on the risk of hunger under climate change. We defined the adaptation \neffect of trade as the sum of the effect of reducing trade costs on hunger under \ncurrent climate (direct trade effect), and any additional positive or negative \nimpact of trade integration under climate change (climate-induced trade effect). \nThe adaptation effect of trade can be understood through Ricardo\u2019s theory of \ncomparative advantage (Supplementary Text)11,12. Reducing trade costs promotes \ntrade according to comparative advantage56 and facilitates the role of trade as \na transmission belt in linking food-deficit and food-surplus regions57. Climate \nchange impacts differ across crops and regions8. Depending on the spatial \ndistribution of these impacts, the current pattern of comparative advantage may \nbe intensified, maintained or substantially altered. This may lead to increased \nfood deficits in certain regions. Trade is argued to have a larger role under climate \nchange as it facilitates adjustment to changes in comparative advantage11,12 and \nenables food surplus to be linked with food deficit regions6,7,57.\nScenario design. Our choice of climate change scenarios was determined by the \nISI-MIP Fast Track Protocol used by crop modellers to calculate crop and grass yield \nimpacts8,58. We used all four RCPs that reflect increasing levels of radiative forcing by \n2100 (the 2.6\u2009W\u2009m\u22122, 4.5\u2009W\u2009m\u22122, 6\u2009W\u2009m\u22122 and 8.5\u2009W\u2009m\u22122 scenarios)59 as projected by \nthe HadGEM2\u2013ES GCM60,61. RCP\u20098.5 was implemented with four additional GCMs \nto reflect uncertainty in climate models: GFDL\u2013ESM2M62, IPSL\u2013CM5A\u2013LR63, \nMIROC\u2013ESM\u2013CHEM64 and NorESM1\u2013M65. RCP\u20092.6 represents climate stabilization \nat 2\u2009\u00b0C and RCP\u20098.5 a temperature range of 2.6\u20134.8\u2009\u00b0C (ref. 26). Yield impacts are \nbased on simulations from the crop model EPIC23,24. Each RCP\u2009\u00d7\u2009GCM combination \nwas modelled including CO2 fertilization effects. RCP\u20098.5\u2009\u00d7\u2009HadGEM2\u2013ES was \nadditionally simulated without the CO2 effect, which reflects the most severe climate \nchange scenario. These scenarios represent the tier 1 set of ISI-MIP scenarios and \nclimate change impacts are simulated individually for all 18 GLOBIOM crops, except \nfor oil palm, and for grasslands. Scenarios without CO2 fertilization for RCPs other \nthan RCP\u20098.5 were considered to be of secondary importance in the ISI-MIP Fast \nTrack\u2014and in the latest simulation protocol for ISI-MIP 3b\u2014and were therefore \navailable only for the four main crops (corn, rice, soya and wheat). We carry out a \ncomprehensive sensitivity analysis with respect to the CO2 fertilization effect for \nall RCPs, however, as this requires extrapolating climate change impacts from the \nfour crops to the other crops, and thus would introduce inconsistency with the Tier \n1 scenarios, the analysis is presented separately in Supplementary Text. In the no \nclimate change scenario, exogenous yield change originates only from long-term \ntechnological development assumptions.\nWe implemented six trade scenarios to analyse the role of trade in climate \nchange adaptation. The first scenario\u2014fixed imports\u2014limits imports to the level \nobserved in the no climate change scenario or less. This represents restricting trade \nflow adjustments in response to climate change, or limiting trade as an adaptation \nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nmechanism. The second scenario, pre-Doha tariffs, excludes the tariff update \nin 2010, representing the trade environment before global trade liberalization \nlaunched by the Doha Round (a comparison of average tariff rates is provided in \nSupplementary Table 6). We also implemented three trade integration scenarios \nto assess promotion of the trade adaptation mechanism. In the first scenario, the \nfacilitation scenario, the nonlinear part of trade costs is set close to zero from \n2020 onwards on the basis of Baker et\u00a0al.24. This reflects the impact of reducing \ntransaction costs, infrastructure costs and other institutional barriers limiting the \nexpansion of trade. Trade facilitation is defined by the WTO as the \u201csimplification \nof trade procedures\u201d66. In economic literature it refers to the reduction in trade \ntransaction costs that are determined by the efficiency of customs procedures, \ninfrastructure services and domestic regulations 18,66. Other trade costs that are \nrelevant in agricultural trade, which were not included in this study, are non-tariff \nmeasures (NTMs). UNCTAD defines NTMs as \u201call policy-related trade costs \nincurred from production to final consumer, with the exclusion of tariffs\u201d67. Typical \nexamples of NTMs are technical measures, such as sanitary and phytosanitary \nmeasures (SPS), and price and quantity control measures, such as quotas and \nsubsidies. Some studies include also the above-mentioned transaction costs in the \ncategory of NTMs68,69, whereas others make the explicit distinction18,70. The per-unit \ntransport costs were kept constant at the base year level. In the second scenario, \ntariff elimination, all agricultural tariffs were progressively phased out between \n2020 and 2050, that is \u221225% in 2020, \u221250% in 2030, \u221275% in 2040 and \u2212100% in \n2050. This scenario leads to a 70% growth in total agricultural trade (Supplementary \nTable 1), comparable in magnitude to the agricultural import (+36%) and export \n(+60%) growth under tariff liberalization reported by Anderson and Martin71. The \nfinal scenario, facilitation\u2009+\u2009tariff elimination, is a combination of the previous two \nscenarios and presents the most extensive open trade scenario. In the baseline trade \nscenario, trade barriers are kept constant at 2010 levels, but trade patterns vary \nendogenously across the different climate impact scenarios. Supplementary Table 7 \nprovides a comparison of average trade costs across the different scenarios.\nSocioeconomic developments were modelled according to the SSP2, which \nreflects a middle-of-the-road scenario in which the population reaches 9.2\u2009billion \nby 2050 and income grows according to historical trends in each region27. The \ntechnological development assumed by SSP2 leads to an increase in global average \ncrop yields of 66% between 2000 and 2050 (Supplementary Table 12). The SSP \nscenarios are widely discussed and are often used as a basis for harmonizing key \nmacroeconomic assumptions for integrated assessment modelling of different \nclimate futures72. SSP2 projects a decrease in the global population at risk of hunger \nfrom 867\u2009million in 2000 to 122\u2009million by 2050. This because of an increase in \nfood consumption\u2014global food availability increases from 2,700 to 3,007\u2009kcal per \ncapita per day\u2014and an improved food distribution within regions, which are both \nrelated to the assumed income growth under SSP2 (ref. 53). Income projections lead \nto changes in food preferences. Under SSP2, the share of livestock products in diets \nincreases globally from 16% in 2000 to 17.3% in 2050, with the largest increases in  \nAsian regions73. Such changes affect the baseline trade pattern\u2014for example, \nincreased production and consumption of livestock products in SAS, EAS and SEA \nimply an increase in imports of feed crops such as corn and soya by 2050.\nHunger statistical analysis. We analysed the results of the scenario runs using a \nregional-level linear regression model to infer the underlying relationship between \ntrade costs, crop-yield changes and hunger as predicted by GLOBIOM. The \nfollowing models were estimated using OLS (Table 1):\nPopulation at risk of hungeritr \u00bc\n\u03b2 1\n\u00f0 \u00de\n1 Crop yieldir \u00fe \u03b2 1\n\u00f0 \u00de\n2 Trade costsitr\n\u00fe\u03b2 1\n\u00f0 \u00de\n3 Crop yieldir \u00b4 Trade costsitr \u00fe P\ni\n\u03b2 1\n\u00f0 \u00de\n4i Regioni \u00fe \u03b5 1\n\u00f0 \u00de\nitr\nFood availabilityitr \u00bc\n\u03b2 2\n\u00f0 \u00de\n1 Crop yieldir \u00fe \u03b2 2\n\u00f0 \u00de\n2 Trade costsitr\n\u00fe\u03b2 2\n\u00f0 \u00de\n3 Crop yieldir \u00b4 Trade costsitr \u00fe P\ni\n\u03b2 2\n\u00f0 \u00de\n4i Regioni \u00fe \u03b5 2\n\u00f0 \u00de\nitr\nWe estimated the models also with regional interaction terms (Fig. 3, \nSupplementary Table 3):\nPopulation at risk of hungeritr \u00bc\nP\ni\n\u03b2 3\n\u00f0 \u00de\n1i Crop yieldir \u00b4 Regioni \u00fe \u03b2 3\n\u00f0 \u00de\n2i Trade costsitr \u00b4 Regioni\n\ue010\n\u00fe\u03b2 3\n\u00f0 \u00de\n3i Crop yieldir \u00b4 Trade costsitr \u00b4 Regioni \u00fe \u03b2 3\n\u00f0 \u00de\n4i Regioni\n\ue011\n\u00fe \u03b5 3\n\u00f0 \u00de\nitr\nFood availabilityitr \u00bc\nP\ni\n\u03b2 4\n\u00f0 \u00de\n1i Crop yieldir \u00b4 Regioni \u00fe \u03b2 4\n\u00f0 \u00de\n2i Trade costsitr \u00b4 Regioni\n\ue010\n\u00fe\u03b2 4\n\u00f0 \u00de\n3i Crop yieldir \u00b4 Trade costsitr \u00b4 Regioni \u00fe \u03b2 4\n\u00f0 \u00de\n4i Regioni\n\ue011\n\u00fe \u03b5 4\n\u00f0 \u00de\nitr\nwhere Population at risk of hungeritr gives the number of people at risk of  \nhunger (million) and Food availabilityitr gives the food availability (kcal per capita \nper day) in 2050 in each region i, trade scenario t and climate change scenario r.  \nCrop yieldir gives the change in average crop yield (kcal\u2009ha\u22121) compared with \nthe average crop yield in the no climate change scenario in 2050. Trade costitr \ngives the log-transformed weighted average trade costs (US$ per 106\u2009kcal) \non all trade flows in 2050. To obtain a measure that reflects the implication \nof trade scenarios on overall trading costs, we calculated the trade-weighted \naverage of trade costs over all agricultural imports, exports and intraregional \ntrade flows for each region i, trade scenario t and climate change scenario r: \nAverage trade costitr \u00bc P\nk\nxiktr\nTotalxitr \u00b4 Trade costiktr\nI\n where xiktr are the trade flows of \ncrop k in, out and within region i in each scenario (t,r) and Totalxitr\nI\n is the sum of \nall trade flows in, out and within region i in each scenario (t,r). The variables Crop \nyieldir and Trade costitr are centred (demeaned) to solve structural multicollinearity. \nFor the regional fixed effects (Regioni) dummy variables were used.\n\u03b2 m\n\u00f0 \u00de\nki\nI\n are the slope coefficients to be estimated for variable k in regression \nmodel m (with k\u2009=\u20091,\u2009\u2026,\u20094 and m\u2009=\u20091,\u2009\u2026,\u20094). \u03b5 m\n\u00f0 \u00de\nitr\nI\n is an independently and identically \nnormally distributed error term with zero mean and \u03c32\nm\n\u00f0 \u00de\nI\n variance. Standard \nerrors were estimated robust to heteroscedasticity using the HC3 method as \nrecommended by Long and Ervin74. HC3 is a refined version of White\u2019s method \nfor estimating heteroskedastic s.e. (HC0). Long and Ervin74 demonstrated using \nMonte Carlo simulations that the HC3 method outperforms HC0 for small sample \nsizes (n\u2009<\u2009250). The calculation of s.e. of the regional interaction effects was \nperformed using the delta method. The F statistic of overall significance rejects \nthe null hypothesis at the 1% significance level for all of the models. The sample \nwas composed of GLOBIOM regional output under five different trade scenarios \n(baseline, pre-Doha tariffs, facilitation, tariff elimination and facilitation\u2009+\u2009tariff \nelimination) and ten climate change scenarios in 2050. The sample size was 550 \nfor models with regional fixed effects (11 regions\u2009\u00d7\u20095 trade\u2009\u00d7\u200910 climate change \nscenarios) and 450 for models with regional interaction terms (9 regions (EUR and \nCAN excluded)\u2009\u00d7\u20095 trade\u2009\u00d7\u200910 climate change scenarios). Summary statistics of all \nof the variables are shown in Supplementary Table 5.\nComparative advantage statistical analysis. To assess comparative advantage, we \nestimated linear regression models of the effect of trade-cost reduction on the share \nof production of a specific crop that region i represents in total world production \nof that crop in each trade scenario t and climate change scenario r (Share of world \nproductionitr); the share of a specific crop in a region\u2019s total crop production \n(Share of regional crop productionitr); and the share of a region\u2019s production of \na specific crop that is exported (Share of production exporteditr). The following \nmodels were estimated separately for wheat, corn, rice and soya using OLS (Fig. 4, \nSupplementary Figs. 16 and 17):\nShare of world productionitr \u00bc\nP\ni\n\u03b2 5\n\u00f0 \u00de\n1i Trade costsitr \u00b4 Regioni\n\u00fe\u03b2 5\n\u00f0 \u00de\n2i Regioni \u00fe \u03b5 5\n\u00f0 \u00de\nitr\nShare of regional crop productionitr \u00bc\nP\ni\n\u03b2 6\n\u00f0 \u00de\n1i Trade costsitr \u00b4 Regioni\n\u00fe\u03b2 6\n\u00f0 \u00de\n2i Regioni \u00fe \u03b5 6\n\u00f0 \u00de\nitr\nShare of production exporteditr \u00bc\nP\ni\n\u03b2 7\n\u00f0 \u00de\n1i Trade costsitr \u00b4 Regioni\n\u00fe\u03b2 7\n\u00f0 \u00de\n2i Regioni \u00fe \u03b5 7\n\u00f0 \u00de\nitr\nFor Fig. 4b, the dependent variable is the outcome under climate change, whereas, \nfor Fig. 4c, the dependent variable is the difference in outcome between climate \nchange and no climate change. Trade costsitr is the log-transformed trade-weighted \naverage of trade costs (US$ per ton) per region i, trade scenario t and climate \nchange scenario r (Supplementary Table 7). The variable Trade costitr was centred \n(demeaned) to solve structural multicollinearity. Dummy variables were used \nfor regional fixed effects (Regioni). Observations were taken from the nine \nRCP\u2009\u00d7\u2009GCM scenarios and four trade integration scenarios (baseline trade, \nfacilitation, tariff and facilitation\u2009+\u2009tariff); regions were excluded that have a deficit \nproduction in at least 90% of the trade and climate change scenarios; n\u2009=\u2009189 \n(corn), n\u2009=\u2009180 (rice), n\u2009=\u200998 (soya) and n\u2009=\u2009246 (wheat); s.e. was estimated robust \nto heteroscedasticity using the HC3 method and s.e. of regional interaction effects \nwas calculated using the delta method.\nData availability\nThe authors declare that the main data supporting the findings of this study are \navailable within the Article and the Supplementary Information. Additional data \nare available from the corresponding author on request. Source data are provided \nwith this paper.\nCode availability\nCode used for the statistical analysis of the scenario data is available from the \ncorresponding author on request.\nReferences\n\t38.\tMosnier, A. Tracking Indirect Effects of Climate Change Mitigation and \nAdaptation Strategies in Agriculture and Land Use Change With a Bottom-Up \nGlobal Partial Equilibrium Model (Univ. Natural Resources and Life Sciences \n(BOKU), 2014).\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\n\t39.\tValin, H. et\u00a0al. The future of food demand: understanding differences in \nglobal economic models. Agric. Econ. 45, 51\u201367 (2014).\n\t40.\tHerrero, M. et\u00a0al. Biomass use, production, feed efficiencies, and greenhouse \ngas emissions from global livestock systems. Proc. Natl Acad. Sci. USA 110, \n20888\u201320893 (2013).\n\t41.\tForsell, N. et\u00a0al. Assessing the INDCs\u2019 land use, land use change, and forest \nemission projections. Carbon Balance Manag. 11, 26 (2016).\n\t42.\tHavl\u00edk, P. et\u00a0al. Global land-use implications of first and second generation \nbiofuel targets. Energy Policy 39, 5690\u20135702 (2011).\n\t43.\tHavlik, P. et\u00a0al. Climate change mitigation through livestock system \ntransitions. Proc. Natl Acad. Sci. USA 111, 3709\u20133714 (2014).\n\t44.\tTakayama, T. & Judge, G. G. Spatial and Temporal Price and Allocation \nModels (North-Holland Publishing Company, 1971).\n\t45.\tGaulier, G. & Zignago, S. BACI: International Trade Database at the \nProduct-Level. The 1994-2007 Version Working Paper 1\u201335 (CEPII, 2010); \nhttps://doi.org/10.2139/ssrn.1994500\n\t46.\tBou\u00ebt, A., Decreux, Y., Fontagn\u00e9, L., Jean, S. & Laborde, D. Assessing applied \nprotection across the world. Rev. Int. Econ. 16, 850\u2013863 (2008).\n\t47.\tHummels, D. Toward a Geography of Trade Costs (Purdue University, 2001).\n\t48.\tBou\u00ebt, A., Decreux, Y., Fontagn\u00e9, L., Jean, S. & Laborde, D. A Consistent, \nAd-Valorem Equivalent Measure of Applied Protection Across the World: The \nMAcMap-HS6 Database Working Papers 2004-22 (CEPII Research Center, \n2004); http://cepii.fr/PDF_PUB/wp/2004/wp2004-22.pdf\n\t49.\tGuimbard, H., Jean, S., Mimouni, M. & Pichot, X. MAcMap-HS6 2007, an \nExhaustive and Consistent Measure of Applied Protection in 2007. J. Int. \nEcon. 130, 99\u2013122 (2012).\n\t50.\tHummels, D. Transportation costs and international trade in the second era \nof globalization. J. Econ. Perspect. 21, 131\u2013154 (2007).\n\t51.\tHasegawa, T. et\u00a0al. Consequence of climate mitigation on the risk of hunger. \nEnviron. Sci. Technol. 49, 7245\u20137253 (2015).\n\t52.\tHasegawa, T. et\u00a0al. Risk of increased food insecurity under stringent global \nclimate change mitigation policy. Nat. Clim. Change 8, 699\u2013703 (2018).\n\t53.\tHasegawa, T., Fujimori, S., Takahashi, K. & Masui, T. Scenarios for the risk of \nhunger in the twenty-first century using Shared Socioeconomic Pathways. \nEnviron. Res. Lett. 10, 014010 (2015).\n\t54.\tGlobal Spatially-Disaggregated Crop Production Statistics Data for 2000  \nVersion 3.0.7 (IFPRI, 2019); https://doi.org/10.7910/DVN/A50I2T\n\t55.\tMoore, F. C., Baldos, U., Hertel, T. & Diaz, D. New science of climate change \nimpacts on agriculture implies higher social cost of carbon. Nat. Commun. 8, \n1607 (2017).\n\t56.\tCostinot, A., Donaldson, D. & Komunjer, I. What goods do countries trade? A \nquantitative exploration of Ricardo\u2019s ideas. Rev. Econ. Stud. 79, 581\u2013608 (2012).\n\t57.\tClapp, J. Food Security and International Trade: Unpacking disputed narratives. \nThe State of Agricultural Commodity Markets 2015-2016 (FAO, 2015).\n\t58.\tWarszawski, L. et\u00a0al. The inter-sectoral impact model intercomparison project \n(ISI-MIP): project framework. Proc. Natl Acad. Sci. USA 111, 3228\u20133232 (2014).\n\t59.\tvan Vuuren, D. P. et\u00a0al. The representative concentration pathways: an \noverview. Climatic Change 109, 5\u201331 (2011).\n\t60.\tMartin, G. M. et\u00a0al. The HadGEM2 family of Met Office unified model \nclimate configurations. Geosci. Model Dev. 4, 723\u2013757 (2011).\n\t61.\tCollins, W. J. et\u00a0al. Development and evaluation of an Earth-system \nmodel\u2014HadGEM2. Geosci. Model Dev. 4, 1051\u20131075 (2011).\n\t62.\tDunne, J. P. et\u00a0al. GFDL\u2019s ESM2 global coupled climate-carbon Earth system \nmodels. part I: physical formulation and baseline simulation characteristics.  \nJ. Clim. 25, 6646\u20136665 (2012).\n\t63.\tDufresne, J. L. et\u00a0al. Climate change projections using the IPSL-CM5 Earth \nsystem model: from CMIP3 to CMIP5. Clim. Dyn. 40, 2123\u20132165 (2013).\n\t64.\tWatanabe, S. et\u00a0al. MIROC-ESM 2010: model description and basic results of \nCMIP5-20c3m experiments. Geosci. Model Dev. 4, 845\u2013872 (2011).\n\t65.\tBentsen, M. et\u00a0al. The Norwegian Earth System Model, NorESM1-M\u2014part 1: \ndescription and basic evaluation of the physical climate. Geosci. Model Dev. 6, \n687\u2013720 (2013).\n\t66.\tMo\u00efs\u00e9, E. & Sorescu, S. Trade Facilitation Indicators\u2014The Potential Impact of \nTrade Facilitation on Developing Countries\u2019 Trade OECD Trade Policy Papers \nNo. 144 (OECD Publishing, 2013).\n\t67.\tNon-Tariff Measures to Trade: Economic and Policy Issues for Developing \nCountries (UNCTAD, 2013).\n\t68.\tMinor, P. J. Time as a Barrier to Trade: A GTAP Database of Ad Valorem \nTrade Time Costs 2nd edn (ImpactEcon, 2013).\n\t69.\tPetri, P. A. & Plummer, M. G. The Economic Effects of the Trans-Pacific \nPartnership: New Estimates Working Paper 16-2 (Peterson Institute for \nInternational Economics, 2016).\n\t70.\tBalistreri, E. J., Maliszewska, M., Osorio-Rodarte, I., Tarr, D. G. &  \nYonezawa, H. Poverty, welfare and income distribution implications of \nreducing trade costs through deep integration in eastern and Southern  \nAfrica. J. Afr. Econ. 27, 172\u2013200 (2018).\n\t71.\tAnderson, K. & Martin, W. Agricultural Trade Reform and the Doha \nDevelopment Agenda (The World Bank and Palgrave Macmillan, 2006).\n\t72.\tRiahi, K. et\u00a0al. The Shared Socioeconomic Pathways and their energy, land \nuse, and greenhouse gas emissions implications: an overview. Glob. Environ. \nChange 42, 153\u2013168 (2017).\n\t73.\tValin, H. et\u00a0al. Agricultural productivity and greenhouse gas emissions: \ntrade-offs or synergies between mitigation and food security? Environ. Res. \nLett. 8, 035019 (2013).\n\t74.\tLong, J. S. & Ervin, L. H. Using heteroscedasticity consistent standard errors \nin the linear regression model. Am. Stat. 54, 217\u2013224 (2000).\nAcknowledgements\nWe thank H. Guimbard and staff at CEPII for their contribution in terms of trade policy \ndata and A. Mosnier for her support in the trade modelling work before this study. We \nacknowledge research funding from Research Foundation Flanders (FWO contract, \n180956/SW) and support from the US Environmental Protection Agency (EPA, contract \nBPA-12-H-0023; call order, EP-B15H-0143). The views and opinions expressed in this \npaper are those of the authors alone and do not necessarily state or reflect those of \nthe EPA, and no official endorsement should be inferred. This paper has also received \nfunding from the EU Horizon 2020 research and innovation programme under grant \nagreement no. 776479 for the project CO-designing the Assessment of Climate CHange \ncosts (https://www.coacch.eu/), and from the European Structural and Investments \nFunds for the project SustES, Adaptive Strategies for Sustainability of Ecosystems \nServices and Food Security in Harsh Natural Conditions (reg. no. CZ.02.1.01/0.0/0.0/16\n_019/0000797).\nAuthor contributions\nAll of the authors have contributed substantially to the manuscript. P.H., J.B., T.K. and \nC.J. developed the concept and designed scenarios. P.H., E.S., T.H., C.J. and D.L. provided \ncode and model simulations. C.J., T.K. and P.H. analysed the data. C.J., P.H., T.K., J.B. and \nM.M. interpreted the data and wrote the manuscript on which S.F., H.V., N.V.L., E.S., \nT.H., S.O. and S.R. commented.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at https://doi.org/10.1038/s41558-020-0847-4.\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41558-020-0847-4.\nCorrespondence and requests for materials should be addressed to C.J.\nPeer review information: Nature Climate Change thanks Maksym Chepeliev and the \nother, anonymous, reviewer(s) for their contribution to the peer review of this work.\nReprints and permissions information is available at www.nature.com/reprints.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nExtended Data Fig. 1 | Biophysical impact of climate change on average crop yield in each region by 2050 as projected by the EPIC crop model. Yields in \nton dry matter per ha. The x-axis indicates the crop yield under no climate change and y-axis the crop yield under climate change for different RCP x GCM \ncombinations without market feedback and adaptation measures. Under no climate change yields are determined by base year yield and assumptions on \ntechnological development over time, under climate change an additional climate impact shifter is applied. Points above the black line indicate an increase \nin crop yield, points below a decrease in crop yield.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nExtended Data Fig. 2 | Impact of climate change on average crop yield after supply-side adaptation in each region by 2050 as projected by GLOBIOM. \nYields in ton dry matter per ha. The x-axis indicates the crop yield under no climate change and y-axis the crop yield under climate change for different \nRCP x GCM combinations with GLOBIOM market feedback and supply-side adaptation (changes in management system and reallocation of production \nacross spatial units in response to price changes). Points above the black line indicate an increase in crop yield, points below a decrease in crop yield.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nExtended Data Fig. 3 | Net agricultural trade of baseline net importing regions in 2050 under trade and climate change scenarios. Net agricultural trade \nin ton dry matter. Fac. = Facilitation, Tariff elim. = Tariff elimination.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nExtended Data Fig. 4 | Net agricultural trade of baseline net exporting regions in 2050 under trade and climate change scenarios. Net agricultural trade \nin ton dry matter. Fac. = Facilitation, Tariff elim. = Tariff elimination.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nExtended Data Fig. 5 | Change in agricultural prices of baseline net importing regions in 2050 under trade and climate change scenarios compared to \nSSP2 baseline. Fac. = Facilitation, Tariff elim. = Tariff elimination.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nExtended Data Fig. 6 | Change in agricultural prices of baseline net exporting regions in 2050 under trade and climate change scenarios compared to \nSSP2 baseline. Fac. = Facilitation, Tariff elim. = Tariff elimination.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nExtended Data Fig. 7 | Change in population at risk of hunger in 2050 in hunger-affected regions under climate change and trade scenarios compared \nto SSP2 baseline. Fac. = Facilitation, Tariff elim. = Tariff elimination. The estimated risk of hunger in the other world regions is zero (CAN, EUR) or very low \n(OCE, USA).\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nExtended Data Fig. 8 | Plot of the fitted linear response of population at risk of hunger (million) to climate-induced crop yield change for different \nvalues of trade costs (1st decile, median, 9th decile). Shaded areas indicate prediction intervals. Prediction based on an OLS estimation of a regional level \nlinear regression of the impact of crop yield change, trade costs and their interaction on population at risk of hunger. Regression results are shown in \nSupplementary Table 3 and the regression model is described in Method.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nExtended Data Fig. 9 | Share of production volume that each region represents of total world production for corn, rice, soya and wheat in the SSP2 \nbaseline in 2050. The projected total world production by 2050 in the SSP2 baseline is 1213 Mt for corn, 884 Mt for rice, 309 Mt for soya and 794 Mt  \nfor wheat.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNature Climate Change\nExtended Data Fig. 10 | Net trade (1000 ton) in East Asia (EAS), Middle East and North Africa (MNA), South Asia (SAS) and Sub-Saharan Africa (SSA) \nfor corn, rice, soya and wheat under climate change and trade scenarios in 2050. Net agricultural trade in ton dry matter. Values above zero indicate net \nexports, negative values indicate net imports.\nNature Climate Change | www.nature.com/natureclimatechange\n\n\n    Scientific Research Findings:", "answer": "Under current trade barriers, a pessimistic scenario of high global warming (+4 \u00b0C by 2100) with no benefits from enhanced atmospheric CO2 on crops could cause up to an additional 55 million people to be undernourished by 2050, mostly in Sub-Saharan Africa and South Asia. If trade restrictions that prevent increased trading under climate change were imposed, the impact could increase to an additional 73 million people. Reduction in tariffs and improvements in trade infrastructure would limit the impact to an additional 20 million people. For export-oriented regions, however, partial trade integration could lead to lower domestic food availability. The findings show that trade policies clearly influence the sensitivity of hunger to climate change. The study focuses on the impacts of trade and climate change on food availability and does not account for income effects, which determine people\u2019s access to food. Furthermore, it does not investigate extreme weather events such as droughts and floods, which are likely to accentuate the importance of trade for adaptation.", "id": 57}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Analysis\nhttps://doi.org/10.1038/s41558-020-00956-w\n1Nelson Institute for Environmental Studies, Center for Sustainability and the Global Environment (SAGE), University of Wisconsin-Madison, Madison, \nWI, USA. 2Department of Agricultural and Applied Economics, University of Wisconsin-Madison, Madison, WI, USA. 3Department of Applied Economics, \nOregon State University, Corvallis, OR, USA. 4Forests Team, World Resources Institute, Washington, DC, USA. 5Department of Geographical Sciences, \nUniversity of Maryland, College Park, MD, USA. \u2709e-mail: fanny.moffette@wisc.edu\nC\nost-effective reduction of deforestation should be a corner-\nstone in the suite of climate change mitigation strategies. \nLand-use change accounts for 6\u201317% of global carbon emis-\nsions1. At any reasonable carbon price, avoided deforestation pro-\nvides 7.2\u20139.6 times the abatement potential of reforestation2. There \nare a range of strategies for accomplishing this\u2014protected areas3, \npayments for ecosystem services4 and supply-chain initiatives5, \namong others. However, all of these strategies require monitoring \nof deforestation activities. Given that deforestation rates are often \nhighest in low- and middle-income countries with limited resources \nto create effective monitoring systems6, the availability of low-cost \nmonitoring technologies may be an important support to policy \naimed at avoiding deforestation.\nThis Analysis estimates the impact of the Global Land Analysis \nand Discovery (GLAD) alerts of tree cover loss7 on deforestation \ntrends across 22 tropical countries between 2011 and 2018. The sys-\ntem that we examine is the first to offer loss alerts at high spatial \nresolution (\u223c30\u2009m) and frequency (up to every 8\u2009days, depending on \ncloud cover). The GLAD alerts use all available Earth observation \ndata from Landsat 7 and Landsat 8 satellites, which combined have \na revisit rate of 8\u2009days. Each day, all new images are first masked for \nclouds and shadow, and then a loss detection algorithm is applied to \nall unobscured land, evaluating the reflectance of the current obser-\nvation together with baseline Landsat metrics. Each 0.00025\u00b0 pixel \n(28\u2009m at the Equator) flagged as loss becomes an alert. The alerts \nmap all forms of tree cover loss including logging, clear-cuts and \nintensive fires in both natural forests and plantations. These alerts \nand the date of first detection are provided without charge through \na simple interface called Global Forest Watch (GFW).\nAvailable in Chinese, English, French, Indonesian, Portuguese \nand Spanish, GFW also enables users to select a specific area for \nwhich they can receive alerts via e-mail. The e-mails state the num-\nber of pixels with suspected deforestation within the subscribed \narea, list the ten most recent alerts with geographic coordinates \nand include a hyperlink to view the locations on a map, as well as \na link to download a file containing all alerts with coordinates and \ndetection dates (Supplementary Fig. B1). Access to and use of the \nalerts require relatively low technical capacity. They can be down-\nloaded on a cellphone during a temporary Internet connection and \na cellphone GPS without Internet can be used to navigate to the \nsite of probable deforestation. Before GLAD alerts, most ongo-\ning prevention of deforestation relied on voluntary reports and/\nor ranger patrols. Because rangers and volunteers are not omni-\npresent, this undermined policymakers\u2019 ability to locate emerging \ndeforestation hotspots.\nAssuming that free deforestation alerts reduce the cost to poli-\ncymakers of monitoring forests, thereby reducing the cost of imple-\nmenting deforestation policy, we examine two questions: whether \nmaking these alerts available affected deforestation rates, and \nwhether areas that were actively monitored by subscribers using the \nsystem saw a change in deforestation trends. Recent work8 in Brazil \nhas shown the substantial contributions of the government\u2019s own \nreal-time monitoring system to the reduction in national deforesta-\ntion trends. However, it is not known whether providing deforesta-\ntion monitoring information freely via a platform unaffiliated with \nany governmental institution and accessible from anywhere on the \nglobe could have an impact on land-use trends. Here we show that \nunder certain conditions, it can.\nWe use a random sample of 1\u2009km2 grid cells from all countries \nthat began receiving alerts before 2018, and estimate the impact \nof GLAD alerts on deforestation using two different strategies. We \nfirst examine whether having access to GLAD alerts for a particu-\nlar region decreases deforestation using the gradual rollout of the \nalert system that began in 2016 (see timeline in Fig. 1a). Second, \nwe study whether using GLAD alerts via subscription results in \na decrease in deforestation rates. In the case of both access and \nuse, we compare trends in areas covered earlier with those cov-\nered later, holding constant factors known to affect deforestation, \nsuch as location and biophysical characteristics. Figure 1b shows \nthe countries analysed for this study and the distribution of for-\nest within them. Our main outcome is the annual probability of  \nforest loss9 extracted at the grid-cell level for 2011 through 2018 \nThe impact of near-real-time deforestation alerts \nacross the tropics\nFanny Moffette\u200a \u200a1,2\u2009\u2709, Jennifer Alix-Garcia3, Katherine Shea4 and Amy H. Pickens5\nReducing deforestation to mitigate climate change necessitates monitoring of deforestation activity. However, while freely \navailable deforestation alerts on forest loss are available, the effect of these alerts and the presence of subscribers in a particu-\nlar area is unclear. Here, we show that subscriptions to alerts in 22 tropical countries decrease the probability of deforestation \nin Africa by 18% relative to the average 2011\u20132016 levels. There is no effect on other continents, and the availability of alerts \ndoes not significantly change deforestation outcomes. This decrease in Africa is higher in protected areas and concessions, \nsuggesting that alerts either increased capacity to enforce existing deforestation policy or induced the development of more \neffective anti-deforestation policies. Calculated using the social cost of carbon for avoided deforestation in Africa, we estimate \nthe alert system\u2019s value to be between US$149 million and US$696 million.\nNature Climate Change | VOL 11 | February 2021 | 172\u2013178 | www.nature.com/natureclimatechange\n172\n\nAnalysis\nNATURe ClimATe CHAnge\n(this outcome does not distinguish between natural and anthro-\npogenic sources of forest loss). This measure is equal to one if \nthere is a positive amount of deforestation during a given year, \nzero if there is none.\nTo evaluate the impact of alerts, we examine the data in two ways. \nFirst, we use the country as the unit of analysis and apply country \nfixed effects and year-by-continent fixed effects. Second, we use the \ngrid cell as the unit of analysis and parallel the country model. We \napply grid-cell and year-by-continent fixed effects, annual measures \nof average temperature, cumulative precipitation and the same flex-\nible time trends, with covariates measured at the grid-cell level. Our \napproach eliminates the effect of time-invariant characteristics that \ninfluence deforestation (for example soil suitability, slope and so \non), general macroeconomic events affecting all countries or spe-\ncific regions, impacts of weather and confounding variation result-\ning from changes in transport or production technologies over time. \nMore details can be found in the Methods.\nAlert availability does not affect deforestation rates\nTo estimate the impact of alert availability, we examine only coun-\ntries that had access to alerts by September 2018. In this case, the \ncounterfactual is countries that received alerts later. If it were the \ncase that forests with higher deforestation rates received alerts \nsooner than forests with lower rates, and if those rates continued \nto be the same over time, then even in the absence of an effect of \nalerts, the estimated impact (approximately the deforestation rate \nof early recipients minus that of late recipients) would show that \nthe availability of alerts had increased deforestation. In this case,  \nthe comparison we make would understate the impact of alert \navailability. The reverse is true if forests at low risk of deforestation \nreceive alerts first.\nTests of differences in deforestation rates between late- and \nearly-enrolled countries before 2016 suggest that the former provide \na plausible counterfactual for the latter, as there are no significant \ndifferences in deforestation in this pre-alert period (Supplementary \nTables A3\u2013A5). If we assume that time trends continue in a similar \nway across comparison groups in the absence of the alerts, then our \nresults are not likely to be biased.\nThere is no robust evidence of any impact of alert availability  \non average or in any particular region (Fig. 2 and Supplementary \nTable A2). Results using the alternative outcomes of per cent forest \nloss and winsorized per cent forest loss are qualitatively similar to \nthe binary outcome estimates (Supplementary Tables A6 and A7). \nAs it is plausible that countries may take time to adapt to the avail-\nability of alerts, we also examine a specification with a 1\u2009yr lag of \nalert availability (Supplementary Tables A8\u2013A10). Although results \nfrom this specification are not statistically different from zero, the \nsign of the average effect switches to become negative.\nLower deforestation in African forests with subscriptions\nTo assess the impact of using GLAD alerts for monitoring, we \nfirst identified users who have the ability to act on alert informa-\ntion. These users, through their subscriptions to GFW, can select \neither an existing jurisdiction, upload a shapefile or identify the \nboundaries of a region about which they wish to receive alerts. \nLeveraging answers to voluntary questions, we divide subscrip-\ntions into two categories: those with intent to control deforestation \nand those without. Those without intent include students, academ-\nics or staff from the World Resources Institute (WRI) or affiliates, \nwhereas those with intent to control include all other subscrip-\ntions to GLAD attached to areas smaller than 100\u2009Mha. The spatial \ndistribution and total number of subscriptions with and without \nFebruary 2014\nGFW (2.0) was\nlaunched but\ninitially provided\nonly annual data\nMarch 2016\nGLAD alerts \ufb01rst launched\nwith Peru, Republic of the\nCongo and Kalimantan\n(Indonesia)\nFebruary 2017\nBrunei, Malaysia, Indonesia, Papua New\nGuinea, Timor-Leste, Burundi, Cameroon,\nCentral African Republic, Equatorial\nGuinea, Gabon, Democratic Republic of\nthe Congo, Rwanda and Uganda\nAugust 2016\nBrazil\nNovember 2017\nColombia,\nEcuador, French\nGuyana, Guyana,\nSuriname and\nVenezuela\na\nb\nFig. 1 | Timeline of GLAD rollout and study region. a, The rollout of GLAD alerts on GFW. b, The forest cover within our study region in 2010 (green \nshading). Study countries are outlined in orange; we exclude the Amazon biome of Brazil in our estimations.\nNature Climate Change | VOL 11 | February 2021 | 172\u2013178 | www.nature.com/natureclimatechange\n173\n\nAnalysis\nNATURe ClimATe CHAnge\nintent are presented in Fig. 3. We provide summary statistics and \ndiscussion of these different classifications in the Methods and \nSupplementary Section B1.\nWe restrict the sample to the subgroup of forested grid cells from \nareas that have GLAD subscriptions with intent to control by 2018 \n(the results of those without intent are in Supplementary Tables  \nB8\u2013B11). In contrast to the previous estimation, the treatment vari-\nable in this analysis varies within countries. Because the sample \nis composed only of areas that ever had a GLAD subscription, we \neliminate one source of bias: the situation where forests at higher \nrisk of deforestation are more likely to be monitored than forests at \nlower risk of deforestation. However, the approach does not elimi-\nnate the possibility that forests at higher risk of deforestation may \nbe monitored sooner than forests at lower risk of deforestation. If \nthis is the case, using the same logic described above, our estimates \nprovide a lower bound of the effect of subscription to the GLAD \nalerts on deforestation.\nSummary statistics suggest that areas with early subscriptions \ntend to have had slightly higher deforestation before alert availabil-\nity (Table 1 and Supplementary Section B2). Pre-alert trends across \nthe different subscription groups (Supplementary Tables B3\u2013B5) \nwere not statistically significantly different using the binary defor-\nestation and winsorized outcomes, but show significant differences \nfor the per cent deforestation outcome. This supports the validity of \ncomparisons for the first two outcomes but not for per cent defores-\ntation. Additional robustness checks detailed in the Supplementary \nInformation confirm large, negative and statistically significant \neffects on forest in Africa.\nThe average effect of subscriptions on deforestation is negative \n(the opposite sign of the impact of alert availability), but statistically \ninsignificant and small compared with the average yearly 2011\u20132016 \ndeforestation probability (0.18). Results using winsorized per cent \ndeforestation and per cent deforestation as outcomes tell the same \nstory (Supplementary Tables B6 and B7). The effect of subscrip-\ntions on deforestation is probably a conservative estimate because \nsome areas are likely to receive more monitoring than others, which \nshould increase the standard errors and decrease the statistical sig-\nnificance of the estimated coefficients.\nAcross regions, we estimate significant avoided deforestation \nin the African subsample, but no robust effects in Asia and South \nAmerica. The coefficient for Africa implies an 18% decrease in \nthe probability of deforestation relative to the 2011\u20132016 lev-\nels (0.04/0.22) (Fig. 4). As the average deforestation in these cells \nis 1.86%, the estimated avoided deforestation per year is equal to \n495.27\u2009km2. If we apply a carbon density of 6,000\u2009MtCO2\u2009km\u22122 \n(the low end of the valuation for forests in this region10), and if \nwe value the amount of avoided carbon emissions resulting from \nalert subscriptions at the social cost of carbon (US$50\u2009t\u22121)11, we \ncalculate the alert system\u2019s value at US$149 million. With a car-\nbon density of 28,100\u2009MtCO2\u2009km\u22122 (the estimated value of the \ndense forests in our sample1), we estimate a value of US$696 mil-\nlion. Both numbers exceed the costs of developing and maintaining \nthe system. Furthermore, this number represents a lower bound, \nas impacts of using GLAD are larger with a 1\u2009yr lagged specifica-\ntion (Supplementary Tables B14\u2013B16), which suggests that when \nsubscribers have more experience with the alerts, their impact on \ndeforestation is greater. Although some cells had subscriptions \nfor multiple years, these total benefits are only calculated for the  \nlast year (when all subscriptions were in place) to avoid overstating \nthe impacts.\nFinally, by way of identifying a mechanism through which \ndeforestation might be controlled, we examine whether the effect \nof subscriptions varies with local policy efforts. For this analysis, \nwe include interactions between subscriptions and indicators for \nwhether or not a grid cell is more than 50% contained by a conces-\nsion or protected area. Information on concessions is not available \nfor all countries, so these estimates are for the subsample of coun-\ntries that have both concession and protected areas data (Methods). \nMarginal effects are presented in Fig. 5 and are extracted from \nSupplementary Table B2. The individual interactions are negative \nand statistically significant\u2014subscriptions have a stronger deterrent \neffect in both protected areas and in forest concessions.\n\u22120.05\n0\n0.05\n0.10\nMarginal effect of GLAD availability\nAll continents\nAfrica\nAsia\nSouth America\nGeographic coverage\nFig. 2 | Impact of GLAD availability on deforestation. The point estimates (shaded bars) and 95% confidence intervals are shown for the impact of GLAD \navailability on the probability of deforestation on average for all continents are shown as well as for Africa, Asia and South America. Each bar shows the \nimpact from a different estimation. The first is from column 6 in panel A of Supplementary Table A2 and the other three are from column 6 in Panel B of \nthe same table, showing the total effect (that is, the linear combination of the baseline and interaction effects).\nNature Climate Change | VOL 11 | February 2021 | 172\u2013178 | www.nature.com/natureclimatechange\n174\n\nAnalysis\nNATURe ClimATe CHAnge\nThese mechanisms are particularly relevant for the African sub-\nsample\u2014more of the subscribed forests in Africa are in protected \nareas and concessions (57% versus less than 33% in the other two \nregions). Separate estimates for these interactions by continent \nshow robust results only for African countries (Supplementary \nTables B19\u2013B21). There is also a consistently statistically significant \nSubscriptions with intent to control deforestation\nSubscriptions without intent to control deforestation\n0\n1\n2\n3\n4\n5\n6\n7\n8\n9\n10 11 12 13 14 15 16 17 18 19 20\nFig. 3 | Spatial distribution and number of subscriptions. Subscriptions with (top) and without (bottom) intent to control deforestation are shown.  \nDarker colours indicate greater numbers of subscriptions.\nTable 1 | Summary statistics\nEarly\nLate\nNever\nNormalized difference\nMean\ns.d.\nMean\ns.d.\nMean\ns.d.\n(early versus late)\nOutcomes\n Deforestation over 2011 to 2016 (0/1)\n0.23\n0.42\n0.18\n0.38\n0.23\n0.42\n0.09\n Per cent deforestation over 2011 to 2016\n0.78\n4.52\n0.42\n2.79\n0.55\n2.77\n0.07\n Winsorized per cent deforestation over 2011 \nto 2016\n0.30\n0.81\n0.21\n0.67\n0.30\n0.81\n0.08\nControls\n Average precipitation per day (mm)\n0.22\n0.20\n0.20\n0.08\n0.18\n0.08\n0.09\n Average temperature\n25.85\n2.71\n26.35\n3.05\n26.01\n2.57\n\u22120.12\n Protected areas (%)\n12.27\n32.45\n21.88\n41.04\n13.49\n33.80\n\u22120.18\n Distance to nearest port within country (km)\n289.64\n357.25\n371.06\n280.97\n616.58\n598.17\n\u22120.18\n Distance to nearest road (km)\n7.34\n11.31\n20.89\n25.49\n10.35\n15.80\n\u22120.49\n Distance to nearest urban centre with \n>100,000 inhabitants (km)\n141.52\n97.05\n234.90\n170.47\n139.54\n95.04\n\u22120.48\nNumber of grid cells\n205,457\n194,203\n262,844\n399,660\nEarly and late subscription groups are based on whether the grid cell had at least one subscription before or by April of 2017 (that is, when 50% of the sample was covered by a subscription; Supplementary \nFig. B2). Normalized differences are a scale-free measure of the difference in distributions between samples, which has advantage that it is directly interpretable in terms of how much average standard \ndeviation is the mean from one sample to the mean of the other sample17.\nNature Climate Change | VOL 11 | February 2021 | 172\u2013178 | www.nature.com/natureclimatechange\n175\n\nAnalysis\nNATURe ClimATe CHAnge\nnegative impact of concessions with subscriptions in the Asian sub-\nsample, although not in protected areas (Supplementary Fig. B3). \nThis effect is not visible in the average impact in Asia because con-\ncessions cover a small part of the estimation sample (14%). As our \nstrategy does not account for the endogenous placement of protected \nareas and concessions, we present these results as intriguing correla-\ntions. However, these correlations confirm the anecdotal evidence \non alerts: that they are used for enforcing protected area policy and \nfor controlling illegal deforestation within forest concessions.\nBenefits and limitations of use\nOur Analysis presents strong evidence that freely available \nforest-change detection data can support decreased deforestation. \nThe use of GLAD alerts through online subscriptions with intent \n\u22120.10\n\u22120.05\n0\n0.05\nMarginal effect of GLAD subscriptions\nAll continents\nAfrica\nAsia\nSouth America\nGeographic coverage\nFig. 4 | Impact of GLAD subscriptions on deforestation. The point estimates (shaded bars) and 95% confidence intervals are shown for the impact of \nGLAD subscriptions with intent to control deforestation on the probability of deforestation on average for all continents as well as for Africa, Asia and \nSouth America. Each bar shows the impact from a different estimation: the first is from column 6 in panel A of Supplementary Table B1 and the other  \nthree are from column 6 in panel B of the same table, showing the total effect (that is, the linear combination of the baseline and interaction effects).\n\u22120.10\n\u22120.05\n0\n0.05\nMarginal effect of subscriptions\nIn PA\nIn concession\nOutside PA/concession\nLocation\nFig. 5 | Estimates of GLAD subscriptions in concessions and protected areas on deforestation. The point estimates (shaded bars) and 95% confidence \nintervals are shown for the impact of GLAD subscriptions with intent to control deforestation on the probability of deforestation. Each bar shows a \nmarginal effect: the first is from subscriptions in areas without protected areas (PA) or concessions, the second is in protected areas and the third in \nconcessions. The estimates come from column 6 of Supplementary Table B2.\nNature Climate Change | VOL 11 | February 2021 | 172\u2013178 | www.nature.com/natureclimatechange\n176\n\nAnalysis\nNATURe ClimATe CHAnge\nto control deforestation led to an 18% decrease in the probability \nof forest loss in the African countries included in our sample. The \nestimated carbon benefits of the avoided deforestation from GLAD \nsubscriptions\u2014from US$149 million US$696 million\u2014is a low-end \nestimate. It is likely that real benefits are much higher, given that \nimpacts seem to be increasing over time, and that there are a num-\nber of substantial co-benefits alongside avoided carbon emissions, \nincluding biodiversity and watershed protection, that are not valued \nin the calculation.\nOur empirical approach supports a causal interpretation. Using \nthe timing of availability or subscription to the alert system helps \nalleviate an important source of bias, as all observed units are even-\ntually affected by the intervention. We control for confounding land \ncharacteristics, temporal shocks and regional trends. Although we \ncannot definitively lay out the causal chain that drives this result, \nthere are a number of reasons why the effects of the alert system \nwere large and statistically significant for only the African coun-\ntries in our sample. First, countries in Africa had minimal or no \nbroad-scale deforestation monitoring technology before GLAD. \nBecause of this, the introduction of freely available alerts con-\nstituted a new source of information to support policy interven-\ntions. Indeed, there is evidence that Cameroon\u2019s government relies \non GLAD for deforestation monitoring12, and that efforts to stem \ndeforestation in Africa have recently been boosted through regional \nagreements (https://afr100.org/content/home).\nIn contrast to Africa, in Asia there is some evidence that the \nprivate sector has already made large investments in monitoring \nof deforestation and fires13,14. In this situation the free alert system \nmay have provided additional information for only some areas of \nforest, as alerts were already available in part of the region. As our \nidentification of effects relies on a break in the deforestation trend \nwhen alerts became available or subscriptions activated, there \nwould be no detectable effect where monitoring was already taking \nplace via existing systems. We do observe a correlation between \nsubscriptions in concessions and decreases in deforestation in \nAsia, but these forests do not cover a large enough area to drive a \nregion-wide impact.\nIn Latin America, the effects may be muddied both by existing \nmonitoring systems and by political unrest. Regarding existing sys-\ntems, in 2017, the Peruvian government stopped using GLAD and \ncreated their own early warning system15. This means that GLAD \nsubscriptions within Peru are not necessarily representative of the \nareas where the government is taking more action. Colombia and \nVenezuela, which constitute a substantial portion of the South \nAmerican sample, were facing political challenges when subscrip-\ntions became available in November 2017. Finally, South American \ncountries in our sample, except Peru, have also had access to GLAD \nfor a shorter time period than Africa and Asia. It is likely to take \ntime for the implementation of deforestation policy to effectively \nuse the alert system, as evidenced by our analysis of lag effects.\nThe alerts that we analyse here are the highest resolution, most \naccessible and highest frequency system ever created. They are not, \nhowever, the first. During the early 2000s, the first near-real-time \ndeforestation monitoring system used coarse-resolution data \n(500\u2009\u00d7\u2009500\u2009m)16. These systems were accessible to only a few coun-\ntries and required specific technical skills to manipulate the data. \nUntil recently, data-driven forest monitoring depended, at best, on \nyearly forest cover maps6. Yearly wall-to-wall maps were relatively \nscarce until the release of the first worldwide annual deforestation \nmaps9, which greatly decreased the burden of data access. Still, \nannual maps are insufficient for stopping deforestation in action, \nand those available require GIS skills to use effectively. Although \nGFW has hosted a few lower-resolution and less-accessible alert \nsystems, including FORMA (Forest Monitoring For Action) and \nTerra-I, the GLAD system represents a substantial improvement in \nscale, coverage and accessibility (Supplement C).\nThe alert system has an important limitation: a loss event can only \nbe detected if there are no clouds above it when satellites pass over. \nIf an area of forest is cleared and there are clouds overhead when the \nsatellite images the area, it will not be detected. As images can only \nbe obtained every 8\u2009days, a month or more may pass before an image \nwithout clouds blocking the clearing can be obtained, particularly \nduring rainy seasons or in regions with persistent cloud cover. This \nmeans that there can be a time lag between the clearing event and the \nactual alert publication, so interventions to stop further deforestation \nmay also be delayed. This affects their utility, but it does not affect \nour impact estimates as we do not directly include the alerts in our \nestimation. The fact that we detect an impact suggests that even this \nimperfect measure provides additional important information to con-\ntrol deforestation. Furthermore, we do not think that variation in alert \nlags across continents drives the variation in our results\u2014forests in \nAfrica, Asia and South America have similar cloud-free image avail-\nability (an average of 76% months per year) during our study period. \nHowever, in the regions with subscriptions in Africa, cloud coverage \nwas 24.7% lower than in South America and 15.5% lower than in \nAsia, suggesting that alert lags may have been shorter in Africa.\nThis Analysis contributes to the ongoing discussion about how \nto address climate change through forest conservation policies in \ndeveloping countries and via zero-deforestation supply-chain ini-\ntiatives. It is particularly relevant because a substantial number of \ndeveloping countries do not have the resources to produce their \nown deforestation alert system. The globally consistent platform \nanalysed here can also facilitate worldwide investments in climate \nchange interventions.\nImportant unanswered questions remain about how the benefits \nof GLAD alerts will evolve in the long term, the most effective ways \nto encourage use and the technical limits of such large-scale moni-\ntoring. Nonetheless, we conclude that the GLAD system of freely \navailable near-real-time forest monitoring provides an immensely \nvaluable scaffold for the implementation of national forest policies, \nand subsequently to the world that benefits from the global public \ngoods provided by standing forests.\nOnline content\nAny methods, additional references, Nature Research report-\ning summaries, source data, extended data, supplementary infor-\nmation, acknowledgements, peer review information; details of \nauthor contributions and competing interests; and statements of \ndata and code availability are available at https://doi.org/10.1038/\ns41558-020-00956-w.\nReceived: 16 March 2020; Accepted: 26 October 2020;  \nPublished online: 4 January 2021\nReferences\n\t1.\t Baccini, A. et\u00a0al. Estimated carbon dioxide emissions from tropical \ndeforestation improved by carbon-density maps. Nat. Clim. Change 2, \n182\u2013185 (2012).\n\t2.\t Busch, J. et\u00a0al. Potential for low-cost carbon dioxide removal through tropical \nreforestation. Nat. Clim. Change 9, 463\u2013466 (2019).\n\t3.\t Sims, K. R. Conservation and development: evidence from Thai protected \nareas. J. Environ. Econ. Manag. 60, 94\u2013114 (2010).\n\t4.\t Alix-Garcia, J. M., Shapiro, E. N. & Sims, K. R. E. Forest conservation and \nslippage: evidence from Mexico\u2019s national payments for ecosystem services \nprogram. Land Econ. 88, 613\u2013638 (2012).\n\t5.\t Gibbs, H. K. et\u00a0al. Brazil\u2019s soy moratorium. Science 347, 377\u2013378 (2015).\n\t6.\t Romijn, E. et\u00a0al. Assessing change in national forest monitoring capacities of \n99 tropical countries. For. Ecol. Manag. 352, 109\u2013123 (2015).\n\t7.\t Hansen, M. C. et\u00a0al. Humid tropical forest disturbance alerts using Landsat \ndata. Environ. Res. Lett. 11, 34008 (2016).\n\t8.\t Assun\u00e7\u00e3o, J., Gandour, C. & Rocha, R. DETERring Deforestation in the \nAmazon: Environmental Monitoring and Law Enforcement (Climate Policy \nInitiative, 2017).\n\t9.\t Hansen, M. C. et\u00a0al. High-resolution global maps of 21st-century forest cover \nchange. Science 342, 850\u2013854 (2013).\nNature Climate Change | VOL 11 | February 2021 | 172\u2013178 | www.nature.com/natureclimatechange\n177\n\nAnalysis\nNATURe ClimATe CHAnge\n\t10.\tGibbs, H. K., Brown, S., Niles, J. O. & Foley, J. A. Monitoring and estimating \ntropical forest carbon stocks: making REDD a reality. Environ. Res. Lett. 2, \n045023 (2007).\n\t11.\tHoward, P. & Sylvan, D. Expert Consensus on the Economics of Climate \nChange (Institute for Policy Integrity, New York University School of Law, \n2015); https://www.edf.org/sites/default/files/expertconsensusreport.pdf\n\t12.\tBulletin d\u2019Alertes GLAD du 1er Trimestre de 2018 (Republique du Cameroun, \n2018); http://wri-sites.s3.amazonaws.com/forest-atlas.org/cmr.forest-atlas.org/\nresources/bulletins/Bulletin%20suivi%20du%20couvert%20forestier%20\n1%C3%A8me%20Trimestre.pdf\n\t13.\tPalm Oil Progress Report 2017 (PepsiCo, 2017); https://perma.cc/DXY2-XSKL\n\t14.\tPalm Oil Progress Update (Cargill, 2015).\n\t15.\tVargas, C., Montalban, J. & Leon, A. A. Early warning tropical  \nforest loss alerts in Peru using Landsat. Environ. Res. Commun. 1,  \n121002 (2019).\n\t16.\tFiner, M. et\u00a0al. Combating deforestation: from satellite to intervention. \nScience 360, 1303\u20131305 (2018).\n\t17.\tImbens, G. W. & Wooldridge, J. M. Recent developments in the econometrics \nof program evaluation. J. Econ. Lit. 47, 5\u201386 (2009).\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2021\nNature Climate Change | VOL 11 | February 2021 | 172\u2013178 | www.nature.com/natureclimatechange\n178\n\nAnalysis\nNATURe ClimATe CHAnge\nMethods\nSample. We use a stratified random sample of 1\u2009km2 grid cells drawn from across \ncountries with proportional allocation within states. To ensure a representative \nsample, countries are divided into two groups: those with more than 75,000 \nforested grid cells and those with fewer than 75,000 forested grid cells. For the first \ngroup, we randomly selected 10% of all grid cells. For the second group, we first \ndetermined the minimum random sample percentage according to Lohr18 and then \nproportionally sample this percentage of grid cells across states within countries. \nThe sample was limited to grid cells that had a minimum of 50% forest cover in 2010 \nand at least 25\u2009ha of forest with canopy cover \u226560%. Brazilian forests in the Amazon \nbiome are excluded from the analyses due to the existing monitoring system there. \nWe study all countries that had access to the alerts before 2018 (Fig. 1a).\nThe countries with fewer than 75,000 forested grid cells are: Burundi (70), \nRwanda (638), Timor-Leste (3,862), Brunei (4,830), Uganda (8,813), Equatorial \nGuinea (25,352) and French Guiana (74,855). Using the same formula to sample \nthe grid cells in the countries with more than 75,000 forested grid cells would lead \nto sampling less than 10%. To ensure representation within countries, sampling  \nfor these locations is stratified at the second administrative level. Supplementary \nTable A11 presents the population of forested grid cells, sample sizes and \ncorresponding sampling percentages. Sampling weights used in empirical \nestimations correspond to the inverse of the probability that the observation is \nincluded due to sampling design.\nTo control for the lack of precision from GLAD alerts in wetland areas due \nto changes in water levels, we exclude wetland areas. Specifically, we used the \nlandform map for Indonesia from the University of Maryland19, the published  \nmap of wetlands for the Congo Basin20 and the Pantanal biome from Brazil. For  \nthe countries without a regional map, we used the international CIFOR dataset21. \nThe final sample includes 662,504 grid cells.\nSubscriptions. Subscriptions are classified as with and without intent to control \ndeforestation. We exclude subscriptions for areas greater than 100\u2009Mha (this area \ncorresponds to a little more than the size of a large state (such as Mato Grosso, \nBrazil) or a small country (Portugal for example) as they are considered too large to \nbe used for monitoring and are most probably associated with users exploring the \nplatform. To create the with-intent group, we exclude all subscriptions associated \nwith the academic sector as well as those subscriptions created by staff from WRI \nor affiliates (together these constitute the without-intent subscriptions). The end \nresult is a total of 558 subscriptions with intent and 734 without. The number of \ngrid cells ever covered by a subscription with intent is 399,660, whereas those that \nhave ever been covered by a subscription without intent total 298,574.\nOf the 558 subscriptions with intent, 302 declared their primary job \nresponsibility using a drop down menu within the alerts system. Supplementary \nTable B22 details these responses. The majority of these were GIS specialists, with a \nrelatively large number of programme managers, technical staff, land-use planning \nspecialist, reporters, and forest/park managers. Of those that did not answer this \nquestion but did declare their location, the largest number was from Indonesia.\nThe subscriptions layer has time and spatial variation. Each polygon of the \nlayer corresponds to a subscribed area with a date of initiation of the subscription. \nAlthough it is the case that some subscribers terminate their subscriptions, to \ncircumvent the potential endogeneity problems with using this variation, we \nassume that once a subscription has been made, the user remained interested in  \nthe subscribed area until the end of our study period. Grid cells are marked as  \nwith subscription from the date of the first subscription.\nAnnual deforestation and forest cover. Our main dependent variable is annual \ndeforestation9 extracted at the grid-cell level for 2011 through 2018. We use this \nannual measure of deforestation (rather than the sum of alerts for that year) \nbecause the alerts are reliant on single observations (making them more susceptible \nto error) and are designed to be conservative, whereas the annual product maps \nloss are based on the entire year and can capture the area of loss more accurately9. \nWe prefer a binary measure of deforestation that is equal to one if there is a positive \namount of deforestation during a given year. The reason for this is that the two \ncontinuous outcomes (per cent deforested and a winsorized version of per cent \ndeforested) have substantial outliers, even in the case of the winsorized measure. \nFor example, the per cent deforestation outcome ranges from zero to 100. However, \nthe measure at the 75th percentile is zero, and at the 95th, 3.57. For the winsorized \nversion of this variable, the 75th percentile value is still zero, and the 95th is 3.29. \nThis type of skewness, particularly when combined with measurement error, can \nresult in considerable bias in regressions.\nOther covariates. We also extract a number of covariates to use as control \nvariables. As weather can influence both the choice to deforest and the \nmeasurement of forest cover, data on temperature are extracted from NASA Land \nsurface temperature22 and precipitation from Climate Hazards Group InfraRed \nPrecipitation with Station data (CHIRPS; ref. 23). To control for transportation costs \n(and because deforestation frontiers are known to be more intensely concentrated \nnear urban centres), we calculate three measures of distance. First, the distance \nto the nearest urban centre with more than 100,000 inhabitants, which can be \ndownloaded from the Joint Research Centre\u2019s Global Human Settlement Layer24. \nSecond, the distance to the nearest port according to the World Port Index from \nthe National Geospatial-Intelligence Agency25. Third, the distance to the nearest \nroad was calculated on the basis of the original road data of OpenStreetMap \n(available from Geofabrik26). Identifiers for the second administrative level are \nassigned to each grid cell using version 3.6 of GADM (ref. 27).\nWe determine whether a grid cell is contained within a protected area using \nlists of worldwide protected areas provided by the IUCN and UNEP-WCMC \n(ref. 28), and available on the GFW website. Data identifying the location of forest \nconcessions come from various country-specific sources and are available on \nthe GFW website. About half of the countries in our sample provided data on \nconcessions in at least one of the following categories: logging, oil palm and wood \nfibre. Countries are Brunei, Cameroon, Central African Republic, Democratic \nRepublic of the Congo, Equatorial Guinea, Gabon, Indonesia, Malaysia, Papua New \nGuinea and the Republic of Congo. Finally, for each grid cell, the per cent observed \nper year is the average per cent per month of the area with forest cover in 2010 with \nat least one clear observation (excludes areas obscured by cloud or shadow).\nEstimation strategy. The basic estimation strategy exploits variation in either \naccess to alerts (to estimate the impact of availability) or use of subscriptions.  \nThe estimation equation is:\ndi;c;k;y \u00bc \u03b2GLADi;c;y \u00fe \u03c8i\u00f0c\u00de \u00fe \u03b1k;y \u00fe \u03bc0Xi;c;k;y \u00fe \u03f5i;c;k;y\nwhere di,c,k,y is a binary variable indicating deforestation in grid cell i in country c \nof continent k in year y. GLADi,c,y is equal to the share of months that the country \nhad access to GLAD alerts in year y or, to assess the impact of subscriptions, an \nindicator equal to one once over 50% of a grid cell fell within a subscribed area for \nthe first time. In the first case, because the treatment variable only changed across \ncountries and time, we first collapse the data to the country-year level (eliminating \nthe subscript i in the equation). In the second, the treatment variable varies within \ncountries and across time, so we run the estimation at the grid-cell level. \u03b2 is the \nestimated treatment effect. The samples differ across estimates as well. To estimate \nthe impact of availability, we use the sample of cells that had access to GLAD alerts \nby 2018 and to assess the impact of subscriptions, those with a subscription at any \ntime before 2018.\nWe also include country (estimation of availability) or grid-cell (estimation \nof use) fixed effects, \u03c8i(c), which control for time-invariant characteristics of the \ncountry or grid-cell affecting deforestation (such as soil suitability and slope). \nIn both specifications, we also include continent-year fixed effects, \u03b1k,y, which \ncontrol for general macroeconomic events occurring for all countries within the \nsame continent and influencing overall deforestation trends (global food demand, \nworld agricultural prices and so on); and a vector of controls Xi,c,k,y that includes: \naverage temperature, cumulative precipitation, interactions of year indicators with \nthe per cent area in protected areas at the baseline and the distances to the nearest \nport, nearest road and nearest city. The distance/time interactions are included \nto help eliminate confounding variation resulting from changes in transport or \nproduction technologies that would affect the time trends of deforestation in more \nremote areas differentially. \u03bc is the vector of estimated parameters. Owing to the \nsmall number of clusters in the country-level estimation, we calculated confidence \nintervals using wild bootstrap clustering (a method that relies on random sampling \nwith replacement to calculate the confidence intervals). Residuals (\u03f5i,c,k,y) were \nclustered at the second administrative level for the subscription estimations. \nWe apply sampling weights in each regression so that point estimates can be \nconsidered to represent the population of forest.\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nDatasets analysed for this study are available from the corresponding author upon \nreasonable request.\nCode availability\nThe codes used to generate the figures and tables are available via Zenodo29.\nReferences\n\t18.\tLohr, S. Sampling: Design and Analysis 2nd edn (Cengage Learning, 2010).\n\t19.\tMargono, B. A., Potapov, P. V., Turubanova, S., Stolle, F. & Hansen, M. C. \nPrimary forest cover loss in Indonesia over 2000\u20132012. Nat. Clim. Change 4, \n730\u2013735 (2014).\n\t20.\tBwangoy, J.-R. B., Hansen, M. C., Roy, D. P., Grandi, G. D. & Justice, C. O. \nRemote sensing of environment wetland mapping in the Congo Basin using \noptical and radar remotely sensed data and derived topographical indices. \nRemote Sens. Environ. 114, 73\u201386 (2010).\n\t21.\tGlobal Wetlands Map (Center for International Forestry Research, 2017); \nhttps://doi.org/10.17528/cifor/006412\n\t22.\tLand Surface Temperature [DAY] (1 MONTH - TERRA/MODIS) (NEO, 2020); \nhttps://neo.sci.gsfc.nasa.gov/view.php?datasetId=MOD_LSTD_M\nNature Climate Change | www.nature.com/natureclimatechange\n\nAnalysis\nNATURe ClimATe CHAnge\n\t23.\tCHIRPS: Rainfall Estimates from Rain Gauge and Satellite Observations \n(CHIRPS, 2019); https://www.chc.ucsb.edu/data/chirps\n\t24.\tGHSL - Global Human Settlement Layer (European Commission, 2019); \nhttps://ghsl.jrc.ec.europa.eu\n\t25.\tNGA\u2019s World Port Index (NGA, 2017); https://services2.arcgis.com/\njUpNdisbWqRpMo35/arcgis/rest/services/WPI_Ports2017/FeatureServer\n\t26.\tGEOFABRIK (Geofabrik, 2019); https://www.geofabrik.de/data/download.html\n\t27.\tGADM data (GADM, 2018); https://gadm.org/data.html\n\t28.\tWorld Database on Protected Areas (WDPA) (IUCN and UNEP-WCMC, 2016); \nhttps://www.iucn.org/theme/protected-areas/our-work/quality-and-effectiveness/\nworld-database-protected-areas-wdpa\n\t29.\tMoffette, F., Alix-Garcia, J., Shea, K. & H. Pickens, A. The Impact of \nNear-real-time Deforestation Alerts Across the Tropics (Zenodo, 2020);  \nhttps://doi.org/10.5281/zenodo.4054099\nAcknowledgements\nWe thank S. Jamilla, E. Goldman and I. Collins for creating the database. We thank \nK. Chomitz, T. Coger, J. Engelmann, N. Harris, H. Nembhard, F. Stolle and N. Ullery \nand participants at the Environmental and Resources Seminar of the University of \nWisconsin-Madison and at the Applied Economics Seminar at Oregon State University \nfor comments. We acknowledge funding from the World Resources Institute; the \norganization had no input into the study design nor impact on the presentation of  \nthe results.\nAuthor contributions\nF.M. and J.A.-G designed research, performed econometric analyses and led the writing. \nK.S. managed the data compilation and contributed to the writing. A.H.P contributed to \nthe database.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41558-020-00956-w.\nCorrespondence and requests for materials should be addressed to F.M.\nPeer review information Nature Climate Change thanks Juliano Assun\u00e7\u00e3o,  \nJohannes Reiche and Juan Robalino for their contribution to the peer review  \nof this work.\nReprints and permissions information is available at www.nature.com/reprints.\nNature Climate Change | www.nature.com/natureclimatechange\n\n1\nnature research  |  reporting summary\nOctober 2018\nCorresponding author(s):\nFanny Moffette\nLast updated by author(s): Oct 2, 2020\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nAll data except the subscriptions and the percent observed (i.e. percent of the grid-cell that is not obscured by cloud or shadow) were \ndownloaded from publicly available websites. They were spatially compiled at the grid-cell level by ArcMap version 10.6 and the package \nArcPy. Different projections were used when creating the grid-cells: South America Albers Equal Area Conic, Africa Albers Equal Area \nConic, and Asia South Albers Equal Area Conic. For the first analysis (i.e. impact of GLAD availability), data were collapsed down to the \ncountry level with the standard \"collapse\" command in Stata-SE version 15.1. \nData analysis\nThe main analysis was conducted using the standard \"xtreg\" command in Stata-SE version 15.1\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. \nWe strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nThe codes used to generate the figures and tables are available in the Zenodo data repository (https://doi.org/10.5281/zenodo.4054099). All data necessary for \nreplication of the results in this paper are available from the corresponding author on reasonable request. Country and state GADM codes correspond are available \nfor download from GADM version 3.6 at https://gadm.org/data.html. The original Tree Cover 2010 and Deforestation data can be downloaded from https://\nearthenginepartners.appspot.com/science-2013-global-forest/download_v1.6.html. We excluded grid-cells in wetland areas using the wetland (landform) map for \nIndonesia from Margono et al. (2014) (https://glad.umd.edu/dataset/primary-forest-cover-loss-indonesia-2000-2012), the map for the Republic of the Congo and \nthe Democratic Republic of the Congo from Bwangoy et al. (2010) (http://dx.doi.org/10.1016/j.rse.2009.08.004), the map from NASA for the Amazon Basin (http://\n\n2\nnature research  |  reporting summary\nOctober 2018\ndx.doi.org/10.3334/ORNLDAAC/1284), and the Pantanal biome from Brazil. For the countries without a regional map, we used the international CIFOR dataset \n(https://www.cifor.org/globaladditional -wetlands/). The original precipitation data from CHIRPS can be downloaded from https://www.chc.ucsb.edu/data/chirps. \nThe original temperature data from NASA Land surface temperature can be downloaded from https://neo.sci.gsfc.nasa.gov/view.php?datasetId=MOD_LSTD_M. The \noriginal protected areas and concession areas (logging, oil palm, and forest) can be downloaded from www.globalforestwatch.org. The original major urban center, \ndefined as cities above 100,000 people, can be downloaded from JRC\u2019s Global Human Settlement (https://ghsl.jrc.ec.europa.eu ). The original port data can be \ndownloaded from World Port Index from the National Geospatial-Intelligence Agency (https://www.arcgis.com/home/item.html?\nid=b04b76b94059436e93757c301c10026c). The original road data (OpenStreetMap) can be downloaded from Geofabrik (https://www.geofabrik.de/data/\ndownload.html). The original subscription layer with information on subscribers is available from World Resources Institute but restrictions apply to the availability \nof these data. However, the grid-cell level subscription timing are included in the replication files available from the authors upon reasonable request. Original \nsubscription layer with information on subscriptions can be obtained upon reasonable request and with permission of World Resources Institute. The percent \nobserved (i.e. percent of the grid-cell that is not obscured by cloud or shadow) has been created by the University of Maryland and is available in the replication files \navailable from the authors upon reasonable request.\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nQuantitative methods were used to analyze the effect of the availability and the use of real-time forest monitoring on deforestation.\nResearch sample\nWe use a sample of 1 x 1 km grid-cells randomly drawn from forested grid-cells across all countries that began receiving alerts prior to \n2018. These countries are: Peru, Republic of the Congo, Brazil (outside of the Amazon biome), Brunei, Malaysia, Indonesia, Papua New \nGuinea, Timor Leste, Burundi, Cameroon, Central African Republic, Equatorial Guinea, Gabon, Democratic Republic of the Congo, \nRwanda, Uganda, Colombia, Ecuador, French Guyana, Guyana, Suriname, and Venezuela. \nSampling strategy\nFor the countries with less than 75,000 forested grid-cells, we follow Lohr (2000). For the countries with more than 75,000 forested grid-\ncells, we randomly draw 10% of all grid-cells. To ensure representation within country, the random sampling is stratified at the second-\nadministrative level. \nData collection\nNot applicable.\nTiming\nStudy covers the 2011-2018 period.\nData exclusions\nBrazilian forests in the Amazon biome are excluded from the analyses due to the existing monitoring system there since the mid-2000s. \nNon-participation\nNot applicable.\nRandomization\nNot applicable.\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology\nAnimals and other organisms\nHuman research participants\nClinical data\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\n\n\n    Scientific Research Findings:", "answer": "Use of the GLAD (global land analysis and discovery) alerts, through subscription to Global Forest Watch, decreased the probability of deforestation in Africa by 18% within the first two years relative to the average 2011\u20132016 levels. The simple availability of the alerts did not significantly impact deforestation, and we found no effect on other continents. Effects in Africa were driven by subscriptions within protected areas and logging concessions, which suggests that the alerts were used to fight illegal deforestation. Using the social cost of carbon, we estimate the value of the alert system to be in the range of US$149\u2013$696 million. However, the benefits are probably greater since co-benefits such as biodiversity are not included and the effectiveness of using the alerts may be increasing over time. Alert systems are likely to be effective in other regions as long as they provide earlier or more accessible reports of forest loss and policies designed to reduce deforestation are enforced.", "id": 58}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41558-020-00914-6\n1Department of Political Science, Washington University in St. Louis, St. Louis, MO, USA. 2Swiss Institute for International Economics and Applied \nEconomic Research, St. Gallen, Switzerland. 3Department of Political Science, Yale University, New Haven, CT, USA. 4Department of Political Science, \nStanford University, Stanford, CA, USA. \u2709e-mail: mbechtel.mail@gmail.com; kenneth.scheve@yale.edu\nT\nwo decades ago, over a dozen national academies of sciences \nurged policymakers to take prompt action to reduce emis-\nsions of greenhouse gases1. Although scholars have contin-\nued to assess the numerous ways in which climate change affects \nhumans2,3, animals4,5 and plants6,7 on Earth, governments have been \nhesitant to pursue policies able to successfully reduce emissions8,9 \nbecause of a costs\u2013participation dilemma: to be effective, climate \npolicy must raise the price of carbon and include most countries of \nthe world. Realizing both of these objectives is challenging because \nclimate action is voluntary and existing studies demonstrate that \npublics are averse to costs10\u201313. A potential response to this problem \nis the \u2018ramping-up principle\u2019, that is, the idea that the price of car-\nbon should be gradually increased to give publics time to adjust to \nstricter regulations8. Although intuitively appealing, it is not clear \nwhether publics do indeed prefer increasing cost paths over alterna-\ntive cost schedules or whether they simply favour climate policies \nwith average low costs. Previous work has explored the willingness \nof individuals to invest in energy efficiency improvements14,15 and \npublic approval of costly climate policy initiatives16\u201320. So far, there \nexists no systematic evidence on which intertemporal cost paths \nmaximize public support for climate action.\nMapping mass support for climate cost schedules\nTo characterize mass preferences over climate cost paths, we \ndevised a survey that we conducted among representative samples \nof adult populations (n\u2009=\u200910,075) in four major economies (France, \nGermany, the United Kingdom and the United States). The sociode-\nmographic margins for the raw and weighted samples along with the \npopulation margins for each country are shown in Supplementary \nTable 1. We first studied support for different climate cost schedules \nusing a direct question item to introduce the idea of an international \nagreement that would entail certain average costs per month and \nper household. The survey then presented respondents with four \ndifferent ways to distribute the costs of implementing an inter-\nnational climate agreement over time and asked them to indicate \nwhich cost schedule they would select in a referendum given a cer-\ntain cost level average (Fig. 1; see the Methods for the wording of \nthe questions). The cost level was held constant for respondents in \nthe United States, but we randomized whether the average monthly \nhousehold costs for implementing the agreement were expected to \nbe low (\u20ac28, \u20ac39 and \u00a315) or high (\u20ac113, \u20ac154 and \u00a360) in France, \nGermany and the United Kingdom, respectively. These values have \nbeen used in previous research10 and correspond to approximately \n0.5% and 2.0% of gross domestic product (GDP) expressed in aver-\nage monthly costs per household.\nThe share of respondents who preferred a constant allocation \nof climate costs over time as opposed to other options, including \nan increasing cost path, is shown in Figure 2. The results are virtu-\nally identical for the weighted data (see Extended Data Fig. 1). The \npooled responses show that 58% of all individuals selected a con-\nstant cost path whereas only 12% preferred an increasing cost path \n(Fig. 2a). When we inspected responses separately by country, we \nfound that the constant cost schedule received majority support in \neach of the four countries (Fig. 2b). We also explored whether pref-\nerences for climate cost paths reflected expectations about average \ncost levels. When breaking out the results by cost treatment, we still \nfound that a majority preferred a constant cost path over increasing, \ndecreasing and inverse U-shaped allocations in both the low-cost \nand the high-cost conditions. These results indicate that ramping \nup climate costs over time may provoke more public opposition \nthan a policy that keeps the costs of climate action stable, even if the \naverage cost level is relatively high (\u20ac113, \u20ac154 and \u00a360 per month \nand per household in France, Germany and the United Kingdom, \nrespectively).\nThe causal effects of cost schedules\nOur mapping of the distribution of support for different climate \ncost paths suggested a correlation between a constant distribution \nof costs and climate policy support. However, it remained unclear \nwhether a constant cost path does in fact cause a higher level of \nConstant carbon pricing increases support for \nclimate action compared to ramping up costs  \nover time\nMichael M. Bechtel\u200a \u200a1,2\u2009\u2709, Kenneth F. Scheve\u200a \u200a3\u2009\u2709 and Elisabeth van Lieshout\u200a \u200a4\nThe introduction of policies that increase the price of carbon is central to limiting the adverse effects of global warming. \nConventional wisdom holds that, of the possible cost paths, gradually raising costs relating to climate action will receive the \nmost public support. Here, we explore mass support for dynamic cost paths in four major economies (France, Germany, the \nUnited Kingdom and the United States). We find that, for a given level of average costs, increasing cost paths receive little sup-\nport whereas constant cost schedules are backed by majorities in all countries irrespective of whether those average costs are \nlow or high. Experimental evidence indicates that constant cost paths significantly reduce opposition to climate action relative \nto increasing cost paths. Preferences for climate cost paths are related to the time horizons of individuals and their desire to \nsmooth consumption over time.\nNature Climate Change | VOL 10 | November 2020 | 1004\u20131009 | www.nature.com/natureclimatechange\n1004\n\nArticles\nNaTUrE ClImaTE CHanGE\npublic approval than that caused by the widely discussed option \nof an increasing cost schedule, and how sizeable this effect is rela-\ntive to the level of costs. To answer these questions, we devised a \nrandomized climate cost conjoint experiment10,21. The experiment \npresented respondents with two climate policy scenarios and asked \nthem to indicate which of the two they prefer. Each respondent \ncompleted eight conjoint tasks, each of which specified both the \ntemporal allocation of costs that each scenario entailed (increasing, \nconstant or decreasing) and the associated average costs to house-\nholds (0.5%, 1.0%, 2.0% or 2.5% of GDP), along with other policy \nfeatures (see Methods). By independently randomizing both the \nlevel of costs and their temporal distribution, this approach allowed \nus to separate and directly compare the causal effects of these two \nforces (see Methods). The instructions to respondents, along with \nan example conjoint task, are shown in Supplementary Fig. 1. For \nour analysis, each policy profile was treated as one observation. We \nestimate the causal effects of climate cost paths on policy support \nby regressing whether a scenario was chosen on indicator variables \nfor each randomly assigned attribute value. The full results of the \nconjoint analysis are shown in Extended Data Fig. 2.\nOur estimates of how costs paths affect policy support, along \nwith 95% and 99% robust confidence intervals, are shown in Fig. 3a.  \nMoving from an increasing to a constant cost path caused a  \nsignificant increase in policy support by 7 percentage points com-\npared to the widely discussed option of ramping up costs over \ntime (the reference category), even when we explicitly specified \nand fixed the cost level of the policy proposal. At the same time, \ndecreasing cost schedules also raised climate policy support com-\npared to the increasing cost path. The causal effects estimated from \nthe conjoint analysis show how cost paths changed the average \nlevel of support, but may as such not be informative about the level \nof support for a particular policy package, which was 50% across \nall profiles by design in the forced-choice, paired-profile conjoint \nexperiment. We also note that the causal effect estimates are a func-\ntion of both the preference for a policy feature and the intensity of \nthat preference, which contrasts to the analysis above in which the \nestimates are simply a function of the number of individuals who \nprefer a given cost schedule (Fig. 2). Understanding the rationale \nfor the positive estimate of the decreasing cost path in the conjoint \nexperiment seems a productive inquiry for future research, but we \nfocus primarily on the comparison between constant costs and \nincreasing costs in the remainder of this paper because constant \ncosts are preferred to an increasing time path across both measure-\nment strategies.\nTo put the sensitivity to cost paths into perspective, we also \nvisualized the causal effects of cost levels with support for climate \naction (Fig. 3a). First, consistent with previous research, costs had \na significant and substantively noteworthy effect on public support. \nSecond, when using these estimates to benchmark the sensitivity to \nan increasing domestic cost path, the effect of switching from a con-\nstant to an increasing schedule was similar to doubling the average \nmonthly household costs from 0.5% of GDP (\u20ac28, \u20ac39, \u00a315 and $53) \nto 1.0% of GDP (\u20ac56, \u20ac77, \u00a330 and $107). These results are robust \nacross all four countries (Fig. 3b). Overall, both the pooled esti-\nmates and the results by country suggest that constant and decreas-\ning cost schedules increase the willingness to support climate action \ncompared to a policy with increasing costs over time.\nAlthough the results provide information about the causal effect \nof cost paths on the support of publics for climate action, opposi-\ntion to an increasing cost schedule could reflect concerns about ris-\ning cost levels. To address this, we explicitly informed respondents \nto the survey that, in the various scenarios, costs were distributed \ndifferently over time but average household costs across the plans \nremained unchanged. Furthermore, we re-estimated the cost path \neffects separately by cost level to explore whether the aversion to an \nincreasing cost schedule depended strongly on costs (Fig. 4).\nMonthly household costs\n2020 2022 2024 2026 2028 2030 2032 2034 2036 2038 2040\nYear\nMonthly household costs\n2020 2022 2024 2026 2028 2030 2032 2034 2036 2038 2040\nYear\nMonthly household costs\n2020 2022 2024 2026 2028 2030 2032 2034 2036 2038 2040\nYear\nMonthly household costs\n2020\n2022 2024 2026 2028 2030 2032 2034 2036 2038 2040\nYear\nFig. 1 | Cost paths presented to respondents. Respondents were asked to select their preferred time path from four options: constant, increasing, \ndecreasing and inverse U-shaped. The order in which these options were shown was randomized across respondents.\nNature Climate Change | VOL 10 | November 2020 | 1004\u20131009 | www.nature.com/natureclimatechange\n1005\n\nArticles\nNaTUrE ClImaTE CHanGE\nA constant cost schedule increased support by 8 percentage \npoints if costs were low (0.5% of GDP) and by 6 percentage points \nif costs were high (2.5% of GDP). This suggests that the aversion to \nincreasing cost paths is unlikely to be explained by concerns about \naverage cost levels. Comparison of decreasing cost paths to con-\nstant cost paths showed that decreasing schedules were only pre-\nferred over constant plans when costs were particularly high (2.0% \nof GDP or more). In addition, we considered the possibility that \nrespondents would have different expectations about when costs \nwould commence depending on the cost schedule. We devised an \nalternative version of the conjoint experiment that replaced the cost \npath attribute with information about the year in which household \ncontributions would start. These values were drawn from a continu-\nous set of integers ranging from 2020 to 2040. This version of the \nclimate action conjoint experiment was completed by 680 randomly \nchosen respondents from the United States.\nWe examined the causal effect of start year by regressing policy \nsupport on a full set of indicator values that binned the start year \nintegers into 2-year periods. If publics were mostly concerned about \ncost avoidance, a later starting period should significantly increase \npublic approval. The results (Fig. 5) indicate that changing the start \nyear had no significant effect on support for climate action. Overall, \nthis evidence is consistent with the idea that publics prefer cost sta-\nbility, a tendency that seems to be relatively independent of con-\ncerns related to cost levels.\nStability and predictability drive support for constant costs\nTo assess potential explanations for why publics prefer constant cost \npaths over the intuitively appealing ramping-up approach, our sur-\nvey included an open-ended question that asked respondents why \nthey selected a certain cost path. We then performed an exploratory \ntext analysis of the responses (see Methods) to determine which \n58%\n12%\n19%\n11%\nPooled\na\nb\nc\nd\n56%\n13%\n18%\n13%\n64%\n13%\n15%\n8%\n56%\n12%\n23%\n9.7%\n59%\n11%\n21%\n9.7%\nUnited States\nUnited Kingdom\nGermany\nFrance\nConstant\nIncreasing\nDecreasing\nInverse U\nBy country\n61%\n11%\n20%\n8.7%\nLow-cost treatment\n(France, Germany, United Kingdom)\n59%\n13%\n19%\n9.6%\nHigh-cost treatment\n(France, Germany, United Kingdom)\nFig. 2 | Preferences for distributing climate costs over time. a,b, The percentage of respondents who prefer constant, increasing, decreasing or inverse \nU-shaped intertemporal allocations of climate costs. Data are pooled (a) (n\u2009=\u200910,075) and by country (b) (France, n\u2009=\u20092,000; Germany, n\u2009=\u20092,000; United \nKingdom, n\u2009=\u20092,000; United States, n\u2009=\u20094,075). c, Low average household costs for implementing the agreement were set to \u20ac28, \u20ac39 and \u00a315 for France, \nGermany and the United Kingdom, respectively. d, High average household costs for implementing the agreement were set to \u20ac113, \u20ac154 and \u00a360 for \nFrance, Germany and the United Kingdom, respectively.\nNature Climate Change | VOL 10 | November 2020 | 1004\u20131009 | www.nature.com/natureclimatechange\n1006\n\nArticles\nNaTUrE ClImaTE CHanGE\nwords were predictive of an individual\u2019s time path choice22. The dis-\ntinctiveness and frequency of the 20 most distinguishing terms for \neach country and time path are shown in Fig. 6.\nWord stems such as \u2018budget\u2019, \u2018easier\u2019, and \u2018know\u2019 were systematic \npredictors of choosing the constant cost path, which indicates that \nthe popularity of this option was related to the desire of individuals \nto simplify budgeting and planning for the future. Additional word \nstems were related to stability concerns that could be expected if \nindividuals tried to smooth consumption over time. Consistent with \nthe rationale underlying the ramping-up approach8, individuals that \nsupported an increasing cost path in part justified this preference by \nhighlighting that this schedule allowed publics to gradually adjust \nto rising costs. However, respondents also chose this option in the \nhope that delayed costs would be felt less strongly, arguably because \nof wage increases and inflation, because they themselves would be \ntoo old or no longer alive, or because they generally discounted \nfuture income and consumption. Respondents who preferred \nhigher costs up front that decreased over time emphasized the need \nto make investments now, which they believed to be essential to \nconfronting climate change.\nThe text analysis suggests at least two theoretical explanations \nfor why people prefer a constant cost path over an increasing sched-\nule: consumption smoothing and time discounting. Consumption \nsmoothing refers to the extent that individual utility in intertem-\nporal choices is characterized by diminishing marginal utility so \nthat individuals prefer to consume similar amounts across time. A \nrecent study finds that experimental estimates of diminishing mar-\nginal utility predict preferences over large-stake payment plans (D. \nAycinena, S. Blazsek, L. Rentschler and C. Sprenger, personal com-\nmunication). Individuals with a strong preference for consumption \nsmoothing should prefer constant cost plans over increasing sched-\nules. Many studies have demonstrated the importance of time dis-\ncounting23,24 for understanding individual decisions on borrowing25 \nand saving26, the willingness to invest in clean energy14, as well as \nthe policy preferences of elected officials27. Individuals who exhibit \nrelatively high levels of patience (that is, low discounting) probably \nalso prefer constant to increasing plans.\nTo assess these two arguments, our survey instrument included \nan item to capture an individual\u2019s general preference for intertempo-\nral smoothing and time discounting (patience). The measurement \nof these preferences relies on the convex time budget approach28,29 in \nwhich respondents are asked to repeatedly choose a bundle of pay-\nments that will pay out some combination of a lower payout sooner \nor a higher payout later. Our respondents chose from a set of six \noptions that included both extreme cases in which the full payment \nis realized either sooner or later as well as mixed bundles. The dif-\nferences in those choices allowed separate identification of attitudes \ntoward consumption smoothing, present bias and time discounting. \nSee Methods for details about this measurement approach.\nWe investigated the importance of consumption smoothing \nand discounting by regressing cost path choice on a dichotomized \nIncreasing\nConstant\nDecreasing\n\u20ac28, \u20ac39, \u00a315, $53\n\u20ac56, \u20ac77, \u00a330, $107\n\u20ac113, \u20ac154, \u00a360, $213\n\u20ac141, \u20ac193, \u00a375, $267\nCost path\nCost path\na\nb\nCost level\nCost level\n\u201330\n\u201320\n\u201310\n0\n10\nChange in chosen policy profile (percentage points)\nPooled\nIncreasing\nConstant\nDecreasing\n\u20ac28, \u20ac39, \u00a315, $53\n\u20ac56, \u20ac77, \u00a330, $107\n\u20ac113, \u20ac154, \u00a360, $213\n\u20ac141, \u20ac193, \u00a375, $267\n\u201330\n\u201320\n\u201310\n0\n10\nChange in chosen policy profile (percentage points)\nUnited States\nUnited Kingdom\nGermany\nFrance\nBy country\nFig. 3 | Support for climate action as a function of cost paths and cost levels.  a,b, Results are pooled (a) (n\u2009=\u2009129,280 policy profiles) or by country  \n(b) (France, n\u2009=\u200932,000; Germany, n\u2009=\u200932,000; United Kingdom, n\u2009=\u200932,000; United States, n\u2009=\u200933,280 policy profiles). Dots with horizontal lines are  \npoint estimates from linear least squares regressions of climate policy chosen on randomly assigned cost path and cost level attributes. Error bars  \nindicate 95% and 99% confidence intervals computed from robust standard errors clustered by respondent. Results with weighting are shown in  \nExtended Data Figs. 3,4.\nIncreasing\nConstant\nDecreasing\nCost path\n0\n2\n4\n6\n8\n10\nChange in chosen policy profile (percentage points)\n0.5% of GDP: \u20ac28, \u20ac39, \u00a315, $53\n1% of GDP: \u20ac56, \u20ac77, \u00a330, $107\n2% of GDP: \u20ac113, \u20ac154, \u00a360, $213\n2.5% of GDP: \u20ac141, \u20ac193, \u00a375, $267\nFig. 4 | Support for climate action as a function of cost paths by cost level. \nCausal effects of climate cost paths on policy support estimated separately \nfor each randomly assigned cost level (0.5% of GDP, n\u2009=\u200932,305; 1% of \nGDP, n\u2009=\u200932,373; 2% of GDP, n\u2009=\u200932,367; 2.5% of GDP, n\u2009=\u200932,235 policy \nprofiles). Dots with horizontal lines are point estimates from linear least \nsquares regressions of climate policy chosen on randomly assigned cost \npath attributes. Error bars indicate 95% and 99% confidence intervals \ncomputed from robust standard errors clustered by respondent. Results \nwith weighting are shown in Extended Data Fig. 5.\nNature Climate Change | VOL 10 | November 2020 | 1004\u20131009 | www.nature.com/natureclimatechange\n1007\n\nArticles\nNaTUrE ClImaTE CHanGE\nsmoothing parameter variable, a measure of patience (both set to \none for respondents with above-median values and zero otherwise), \nand a full set of sociodemographic predictors in a multinomial pro-\nbit model. We find that individuals with a higher desire to smooth \nconsumption are significantly more likely to select the constant cost \npath over the increasing and decreasing climate cost paths. Similarly, \nrespondents who were more patient were significantly more likely \nto support a constant climate cost schedule over time-varying cost \npaths such as the ramping-up approach (Supplementary Table 2). \nOverall, the predictive patterns indicate that, consistent with the \nresults from our analysis of respondents\u2019 answers to our open-ended \nquestion, general attitudes towards time seem to play a systematic \nrole in understanding mass preferences for constant over dynamic \ncost paths.\nConclusion\nCredible climate policies will have to raise the price of carbon, and \nthe public are concerned about these costs even if they believe the \nscience of climate change and generally would like governments to \naddress the issue. One intuitively appealing approach to this prob-\nlem is to ramp up the costs of climate action over time. This lowers \ncosts in the near to medium term but also requires individuals to \nkeep adjusting to steadily increasing carbon prices. Our results indi-\ncate that such cost plans run the risk of reducing support for climate \npolicies because many individuals prefer to smooth their consump-\ntion over time. The ramping-up approach could remain politically \nfeasible if some voters focus on near-term costs or if policy experi-\nence causes mass preferences to become more favourable towards \ncostly climate action in general. But to the extent that policymakers \n2020\u20132021\n2022\u20132023\n2024\u20132025\n2026\u20132027\n2028\u20132029\n2030\u20132031\n2032\u20132033\n2034\u20132035\n2036\u20132037\n2038\u20132040\nStart year of cost paths\n\u201310\n\u20135\n0\n5\n10\nChange in chosen policy profile (percentage points)\nFig. 5 | Support for climate action as a function of cost paths by start year. \nResults from a conjoint experiment conducted in a separate section of \nthe United States survey that randomized the year in which contributions \nwould start (n\u2009=\u200910,880 policy profiles; see Methods for details). Dots with \nhorizontal lines are point estimates from linear least squares regressions \nof climate policy chosen on randomly assigned cost path attributes. Error \nbars indicate 95% and 99% confidence intervals computed from robust \nstandard errors clustered by respondent. Results with weighting are shown \nin Extended Data Fig. 6.\nCosts remain the same over the entire period\nCosts are lower initially and higher later\nCosts are higher initially and  lower later\nCosts are lower initially,\nhigher in the middle and lower later\nFrance\nGermany\nUnited Kingdom\nUnited States\n3\n4\n5\n6\n2\n3\n4\n5\n6\n2\n3\n4\n5\n6\n2\n3\n4\n5\n3\n4\n5\n6\n7\n4\n6\n8\n2.5\n5.0\n7.5\n10.0\n2.5\n5.0\n7.5\n10.0\n12.5\nLog(word count)\nLog odds ratio\nBudget\nChaqu\nConst\nDepens\nEquilibr\nEquit\nFacil\nGer\nMauvais\nMem\nN'y\nQuoi\nRest\nSait\nSavoir\nSimpl\nSomm\nStabilit\nStabl\nSurpris\nAusgab\nBelast\nBess\nBetrag\nBleib\nBleibt\nEinstell\nGenau\nGerecht\nGleich\nGleichbleib\nGleichma\nKalkuli\nKalkulierbar\nMonat\nPlan\nPlanbar\nPlanung\nRechn\nUberschaubar\nAmount\nBudget\nConsist\nConstant\nEasier\nExact\nFair\nFairer\nFix\nKeep\nKnow\nLike\nManag\nMonth\nMuch\nOutgo\nPlan\nRemain\nStay\nSteadi\nAmount\nBudget\nChang\nConsist\nConstant\nEasier\nEqual\nEven\nEveri\nFix\nKeep\nKnow\nMonth\nPlan\nPredict\nRate\nRemain\nSet\nStay\nSteadi\nAge\nAlor\nAugment\nAugmentent\nAuss\nCourb\nEsper\nHauss\nL'augment\nPetit\nPeu\nPeut\nPeuvent\nPrepar\nProgress\nProgressivit\nRas\nSalair\nToujour\nVont\nAlt\nAnder\nAnsteig\nEbenfall\nEgal\nEinkomm\nErhoh\nErhoht\nErleb\nGanz\nGewohn\nInflation\nJahr\nKostensteiger\nLangsam\nSteig\nSteiger\nSteigt\nTeur\nVielleicht\nAdapt\nAddit\nAllow\nDead\nEarn\nGive\nGradual\nHope\nHousehold\nIncreas\nInflat\nLive\nMoment\nOffset Old\nProbabl\nRise\nSlowli\nTime\nWage\nAdjust\nDead\nDollar\nEconomi\nEveryth\nGive\nGradual\nGreat\nHope\nIncreas\nInflat\nMoney\nNew\nOld\nPossibl\nPrepar\nSlowli\nTime\nWage\nYear\nAgir\nBaiss\nCher\nDebut\nDepart\nDiminu\nDon\nEffort\nEnsuit\nFil\nFur\nInvest\nMainten\nMesur\nMoin\nPlac\nPuis\nRapid\nSuit\nVit\nAnfang\nDass\nEntlast\nErfolg\nGeld\nGut\nInvesti\nInvestition\nLauf\nLieb\nMehr\nNahm\nPositiv\nSchnell\nSink\nSofort\nSollt\nSpat\nWenig\nZeit\nBenefit\nCheaper\nDecreas\nDrop\nEventu\nFirst\nFutur\nGet\nGoe\nImprov\nIniti\nInvest\nLater Less\nLong\nNow\nReduc\nTechnolog\nThing\nWorst\nBegin\nCheaper\nCost\nDecreas\nFirst\nFront\nFutur\nGet\nGoe\nIniti\nLater\nLess Lower\nNow\nOlder\nPay\nRather\nReduc\nSooner\nUpfront\n2040\nBaiss\nCommenc\nCourb\nDouc\nEnsuit\nFais\nJust\nMesur\nMilieu\nMis\nNecessair\nPar\nPermet\nPermettr\nPrepar\nProgress\nPuis\nRepartit\nTerm\nAllerding\nAnsteig\nAuswirk\nBest\nErscheint\nErst\nGewohn\nIrgendwann\nLangsam\nLogisch\nLosung\nMal\nMensch\nMitt\nStart\nWeg\nZeig\nZeitpunkt\nZiel\nZuerst\nAdjust\nBenefit\nEas\nEnd\nFund\nGive\nGradual\nImprov\nLater\nLook\nLow\nLower\nMight\nPrepar\nPromis\nReduc\nReward\nSave\nStart\nUse\nAdjust\nBack\nBuild\nChanc\nCurv\nEas\nEnd\nGet\nGive\nGradual\nHalf\nIniti\nLook\nLow\nLower\nMiddl\nPrepar\nRamp\nSlowli\nStart\nFig. 6 | Words associated with justifications for a given cost path. This plot shows the top 20 most distinguishing words in respondents\u2019 justifications for choos\u00ad\ning a given time path by cost schedule and country. The vertical axis plots the most distinguishing words on the basis of the log odds ratio. The horizontal axis \nshows the log frequency with which a term occurred. Words were reduced to their common stems (for example, surprise and surprising both became surpris).\nNature Climate Change | VOL 10 | November 2020 | 1004\u20131009 | www.nature.com/natureclimatechange\n1008\n\nArticles\nNaTUrE ClImaTE CHanGE\nseek to design policies that are transparent and meet meaningful \nemission reduction goals, constant cost plans promise more support \nfor climate action relative to ramping-up approaches. Moreover, \nowing to the delay in large-scale policy responses to climate change, \ncountries will probably have to pursue more progressive and costly \nclimate action to limit the adverse effects of global warming. The \ndrop in support because of higher costs associated with these more \nambitious policy efforts may be at least partially mitigated by select-\ning a set of attractive design features such as the constant distribu-\ntion of costs.\nOnline content\nAny methods, additional references, Nature Research report-\ning summaries, source data, extended data, supplementary infor-\nmation, acknowledgements, peer review information; details of \nauthor contributions and competing interests; and statements of \ndata and code availability are available at https://doi.org/10.1038/\ns41558-020-00914-6.\nReceived: 14 January 2020; Accepted: 20 August 2020;  \nPublished online: 21 September 2020\nReferences\n\t1.\t Australian Academy of Sciences et al. The science of climate change. Science \n292, 1261 (2001).\n\t2.\t Obradovich, N., Tingley, D. & Rahwan, I. Effects of environmental stressors \non daily governance. Proc. Natl Acad. Sci. USA 115, 8710\u20138715 (2018).\n\t3.\t Stevanovi\u0107, M. et al. The impact of high-end climate change on agricultural \nwelfare. Sci. Adv. 2, e1501452 (2016).\n\t4.\t Chen, I.-C., Hill, J. K., Ohlemu\u0308ller, R., Roy, D. B. & Thomas, C. D. Rapid \nrange shifts of species associated with high levels of climate warming. Science \n333, 1024\u20131026 (2011).\n\t5.\t Bonebrake, T. C. & Mastrandrea, M. D. Tolerance adaptation and \nprecipitation changes complicate latitudinal patterns of climate change \nimpacts. Proc. Natl Acad. Sci. USA 107, 12581\u201312586 (2010).\n\t6.\t Panetta, A. M., Stanton, M. L. & Harte, J. Climate warming drives local \nextinction: Evidence from observation and experimentation. Sci. Adv. 4, \neaaq1819 (2018).\n\t7.\t C\u00e1mara-Leret, R. et al. Climate change threatens New Guinea\u2019s biocultural \nheritage. Sci. Adv. 5, eaaz1455 (2019).\n\t8.\t Nordhaus, W. Climate change: the ultimate challenge for economics. Am. \nEcon. Rev. 109, 1991\u20132014 (2019).\n\t9.\t Keohane, R. O. & Victor, D. G. Cooperation and discord in global climate \npolicy. Nat. Clim. Change 6, 570\u2013575 (2016).\n\t10.\tBechtel, M. M. & Scheve, K. F. Mass support for global climate agreements \ndepends on institutional design. Proc. Natl Acad. Sci. USA 110,  \n13763\u201313768 (2013).\n\t11.\tDrews, S. & van den Bergh, J. C. J. M. What explains public support for \nclimate policies? A review of empirical and experimental studies. Clim. Policy \n16, 855\u2013876 (2015).\n\t12.\tAklin, M. & Urpelainen, J. Debating clean energy: frames, counter frames, \nand audiences. Glob. Environ. Change 23, 1225\u20131232 (2013).\n\t13.\tEgan, P. J. & Mullin, M. Climate change: US public opinion. Annu. Rev. Polit. \nSci. 20, 209\u2013227 (2017).\n\t14.\tNewell, R. G. & Siikam\u00e4ki, J. Individual time preferences and energy \nefficiency. Am. Econ. Rev. 105, 196\u2013200 (2015).\n\t15.\tFeldman, L. & Hart, P. S. Climate change as a polarizing cue: framing effects \non public support for low-carbon energy policies. Glob. Environ. Change 51, \n54\u201366 (2018).\n\t16.\tStoutenborough, W., Bromley-Trujillo, R. & Vedlitz, A. Public support for \nclimate change policy: consistency in the influence of values and attitudes \nover time and across specific policy alternatives. Rev. Policy Res. 31,  \n555\u2013583 (2014).\n\t17.\tHammar, H. & Jagers, S. C. Can trust in politicians explain individuals\u2019 support \nfor climate policy? The case of CO2 tax. Clim. Policy 5, 613\u2013625 (2006).\n\t18.\tTingley, D. & Tomz, M. Conditional cooperation and climate change. Comp. \nPolit. Stud. 47, 344\u2013368 (2014).\n\t19.\tBernauer, T. & Gampfer, R. How robust is public support for unilateral \nclimate policy? Environ. Sci. Policy 54, 316\u2013330 (2015).\n\t20.\tMildenberger, M. Support for climate unilateralism. Nat. Clim. Change 9, \n187\u2013188 (2019).\n\t21.\tHainmueller, J., Hopkins, D. J. & Yamamoto, T. Causal inference in conjoint \nanalysis: understanding multidimensional choices via stated preference \nexperiments. Polit. Anal. 22, 1\u201330 (2014).\n\t22.\tMonroe, B. L., Colaresi, M. P. & Quinn, K. M. Fightin\u2019words: lexical feature \nselection and evaluation for identifying the content of political conflict. Polit. \nAnal. 16, 372\u2013403 (2008).\n\t23.\tAndersen, S., Harrison, G. W., Lau, M. I. & Rutstr\u00f6m, E. E. Eliciting risk and \ntime preferences. Econometrica 76, 583\u2013618 (2008).\n\t24.\tFrederick, S., Loewenstein, G. & O\u2019Donoghue, T. Time discounting and time \npreference: a critical review. J. Econ. Lit. 40, 351\u2013401 (2002).\n\t25.\tMeier, S. & Sprenger, C. Present-biased preferences and credit card \nborrowing, Am. Econ. J. Appl. Econ. 2, 193\u2013210 (2010).\n\t26.\tSutter, M., Kocher, M. G., Gl\u00e4tzle-R\u00fctzler, D. & Trautmann, S. T. Impatience \nand uncertainty: experimental decisions predict adolescents\u2019 field behavior. \nAm. Econ. Rev. 103, 510\u2013531 (2013).\n\t27.\tSheffer, L., Loewen, P. J., Soroka, S., Walgrave, S. & Sheafe, T. \nNonrepresentative representatives: an experimental study of the decision \nmaking of elected politicians. Am. Polit. Sci. Rev. 112, 302\u2013321 (2018).\n\t28.\tAndreoni, J. & Sprenger, C. Estimating time preferences from convex budgets. \nAm. Econ. Rev. 102, 3333\u20133356 (2012).\n\t29.\tAndreoni, J., Kuhn, M. A. & Sprenger, C. Measuring time preferences: a \ncomparison of experimental methods. J. Econ. Behav. Organ. 116,  \n451\u2013464 (2015).\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2020\nNature Climate Change | VOL 10 | November 2020 | 1004\u20131009 | www.nature.com/natureclimatechange\n1009\n\nArticles\nNaTUrE ClImaTE CHanGE\nMethods\nSample. We carried out our survey in four major economies (France, Germany, \nthe United Kingdom and the United States). The survey was conducted online by \nYouGov using matched sampling. Propensity scores were used to match available \ninternet respondents (participants of the YouGov online panel) to a stratified \nrandom sample of the target adult population. The propensity score model \nincluded age, gender, years of education and region (European countries), or age, \ngender, race or ethnicity, years of education and region (the United States).\nUnited States. The survey period was 18 December 2018 to 3 January 2019. The \nsampling frame for the target population was constructed from the full 2016 \nAmerican Community Survey. All matched respondents were then assigned \nweights stratified on 2016 presidential vote, age, gender, race and education to \ncorrect for remaining imbalances. The final number of respondents was 4,075.\nFrance, Germany and the United Kingdom. The survey period was 31 March 2019 \nto 4 April 2019. The sampling frames for the target populations were constructed \nfrom the 2018 Eurobarometer survey for selection within strata by weighted \nsampling with replacements (using the person weights on the public use file). The \nfinal number of observations was 2,000 for France, 2,000 for Germany, and 2,000 \nfor the United Kingdom. The distributions of sociodemographic characteristics in \nthe population, the raw samples and the weighted samples by country are reported \nin Supplementary Table 1.\nMeasurements. We measured individual tendency to discount future events \nand the preference for consumption smoothing using the convex time budget \nmethod28,29. The allocation of payments xt and xt+k between period t and t\u2009+\u2009k was \nconsidered. The utility function of an individual was defined as:\nU xt; xt\u00fek\n\u00f0\n\u00de \u00bc\nx\u03b1\nt \u00fe \u03b2\u03b4kx\u03b1\nt\u00fek if t \u00bc 0\nx\u03b1\nt \u00fe \u03b4kx\u03b1\nt\u00fek if t>0\n\u001f\nwhere \u03b1 measures consumption smoothing as the utility function curvature, \u03b4 \ndenotes long-run time discounting, and \u03b2 captures present bias. We estimated \nthese parameters using the choices of respondents in a convex time budget task. \nIn this task, respondents were asked to select a bundle of payments that would be \nreceived at time t and t\u2009+\u2009k, where each choice comprised the cases in which the full \npayment occurred at time t and t\u2009+\u2009k and also included four convex combinations. \nFor example, a choice task might ask respondents to choose between a combination \nof US$19 today and $0 in 5 weeks, a combination of $0 today and $20 in 5 weeks, \nas well as the following four convex combinations of these two: $15.20 today and \n$4.00 in 5 weeks, $11.40 today and $8.00 in 5 weeks, $7.60 today and $12.00 in 5 \nweeks, and $3.80 today and $16.00 in 5 weeks.\nThis example compares payments today to payments in 5 weeks. Following \nexisting work28, we asked respondents to also assess budgets in which the earlier \npayment would occur at t\u2009>\u20090 and the later payment after that. Variation in \nindividual choices provided information to estimate the consumption smoothing \nparameter \u03b1 and the other parameters mentioned above using individual-level \nregressions.\nParameters were estimated by regressing the natural log of the ratio of the \nchosen earlier and later payments on the number of days to the first payment, \nthe number of days that the payment is delayed, and the natural log of the price \nratio of the later payments to the earlier payments. Our estimate of the smoothing \nparameter \u03b1 was the inverse of the coefficient on the price ratio. The time \ndiscounting (or patience) parameter was the exponent of the ratio of the coefficient \non the temporal delay to the coefficient on the natural log of the price ratio. To \ndeal with extreme values we trimmed the estimated parameters at the 5th and \n95th percentiles. We then transformed the estimated parameters into an indicator \nvariable that equals one if a respondent\u2019s consumption smoothing measure was \ngreater than the median and is zero otherwise (Supplementary Table 2). Similarly, \nwe constructed an indicator variable that took a value of one if a respondent\u2019s time \ndiscounting estimated parameter was above the median, thus corresponding to a \nrelatively high degree of patience.\nThe sociodemographic predictors that were used included gender, age, income, \neducation and children. Age categories were 18\u201334 years, 35\u201349 years, 50\u201364 years, \nand 65 years and above. Respondents were divided into three income categories \nsuch that the sample in each country was equally split between these groups. \nEducation was also split into three categories adjusted to the education system of \neach country, with groupings broadly corresponding to less than a high school \ndegree, a high school degree or more and a bachelor\u2019s degree or more. \u2018Children\u2019 \nwas measured as a binary variable that took a value of one if the respondent \nindicated having any children.\nCost schedule questions. The full survey instrument is available as part of the \nreplication archive for this study. Our main outcome variables were based on two \nitems, namely a climate agreement item and a climate cost path item.\nThe climate agreement item was as follows:\n\u201cAs you probably know, many experts say that countries should take action to \naddress global warming.\nSuppose [your country] is considering joining an international agreement to \nreduce greenhouse gas emissions. Implementing the agreement would mean that \neach household would have to pay on average [France, \u20ac28, \u20ac113; Germany, \u20ac39, \n\u20ac154; the United Kingdom, \u00a315, \u00a360; the United States, $107] more per month \nthrough, for example, higher energy prices.\nGenerally speaking, do you approve or disapprove of [France, Germany, the \nUnited Kingdom, the United States] joining such an agreement?\u201d\nRespondents were asked to answer on a scale from 1 (strongly approve) to 10 \n(strongly disapprove).\nThe climate cost path item was as follows:\n\u201cRegardless of your previous answer, suppose [France, Germany, the United \nKingdom, the United States] is going to implement that international agreement \nand the household costs would still be [same costs as above] per month on average. \nHowever, there are different ways of distributing these costs over time.  \nThe figures below indicate four alternatives. If you had to select one of the  \noptions in a referendum, which would you chose? Please carefully consider the \navailable options.\u201d\nThe options were presented to respondents as shown in Fig. 1. The preferences \nfor distributing climate costs over time, shown in Fig. 2, were calculated as the \nshare of respondents who selected a particular time path. We investigated the role \nof different individual-level covariates using a multinomial probit regression. The \nfour possible values of the outcome variable corresponded to the four time path \noptions available to respondents. Our independent variables were the measures of \nconsumption smoothing, patience and the sociodemographic covariates described \nabove (the results are reported in Supplementary Table 2).\nClimate cost conjoint experiment. The forced-choice conjoint experiment  \nstarted by explaining to respondents that they were about to see a pair of policies, \nfrom which they would be asked to choose their preferred option. The full \nintroductory text is available in Supplementary Fig. 1a. Respondents then  \nsaw two possible climate policies side by side, with information about eight \ndifferent aspects of each policy. They were then asked to choose which of the two \npolicy sets they would prefer to see enacted. An example of such a task is shown \nin Supplementary Fig. 1b. Each respondent was asked to make a choice regarding \neight different pairs of policies. Our analysis treats one policy scenario or proposal \nas one observation.\nThe randomized policy features comprised the policy\u2019s cost to the average \nhousehold, the distribution of those costs over time, and the percentage of revenues \nused for adaptation and mitigation efforts. Information was provided on each of \nthese features regarding the respondents\u2019 own country and other major economies. \nThe temporal distribution feature could take increasing, decreasing or constant \nvalues. The percentages of revenue spent varied between 0% and 100%, such that \nadaptation and mitigation together summed to 100%. The different household cost \nvalues for each country were equivalent to 0.5%, 1.0%, 1.5%, 2.0% and 2.5% of per \ncapita GDP (see Extended Data Fig. 2).\nEach of the policy features was randomized separately. One of the strengths \nof conjoint experiments is that this separate randomization allows for the direct \ncalculation of the treatment effect of different values of the features, that is, the \naverage marginal component effect21. This causal effect captures the change in \nprobability that a respondent will choose a particular profile with that value, \ncompared to a profile with the baseline value for that policy feature. We computed \nthis treatment effect using linear regression, where the unit of analysis was a \npolicy profile and the outcome of interest was whether the profile was selected \nby the respondent. Standard errors were clustered by respondent to account for \nwithin-individual correlation of the error term (see Extended Data Fig. 2 for the \nresulting treatment effects). To investigate differences between populations of \nrespondents, we estimated this regression separately for different subgroups.\nAll respondents in France, Germany and the United Kingdom completed \nthe climate cost path conjoint experiment. In the United States, this conjoint \nexperiment was administered to 2,080 randomly chosen respondents (about 50% \nof the total sample). Each respondent performed eight conjoint tasks in which \nthey assessed two profiles simultaneously. The results (Fig. 5) rely on an alternative \nversion of the climate conjoint experiment that replaced the cost schedule attribute \nwith the randomly assigned year in which household contributions would start. \nThis conjoint experiment was administered to 680 randomly chosen respondents \nin the United States. Each respondent again performed eight conjoint tasks \nin which they assessed two profiles simultaneously. Therefore, the number of \nobservations for this analysis is 10,880.\nQuantitative text analysis. After selecting one of the four time paths, respondents \nwere asked an open-ended question that read \u201cPlease let us know why you chose \nthis response\u201d. The text analysis that we performed on the responses to these \nquestions was not preregistered. We first applied standard text-as-data cleaning \nprocedures to these responses by removing all punctuation, capitalization and word \norderings to treat the responses as a \u2018bag of words\u2019. In each of the three languages, \nwe discarded common stop words (for example, \u2018the\u2019 and \u2018and\u2019) and we stemmed \neach word to combine related terms (for example, \u2018consistently\u2019 and \u2018consistency\u2019 \nboth became \u2018consist\u2019). We then discarded rare terms, namely those that occurred \nless than five times across answers from a particular country. After this cleaning \nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTUrE ClImaTE CHanGE\nprocess, we were left with 9,428 responses containing at least one useable term \n(United States, 3,844; United Kingdom, 1,887; France, 1,894; Germany, 1,803).\nTo identify the words that best predicted the time path option that a respondent \nhad selected, we applied the \u2018Fightin\u2019 Words\u2019 algorithm22. The basic intuition of \nthis measure is captured by a normal log odds ratio. For a given time path k and a \ngiven word j, we calculated the log odds ratio according to the probability of a text \nstring containing that word j being written by a respondent who chose option k or \nanother option:\nLog odds ratioj;k \u00bc log\nP kjj\n\u00f0\n\u00de\n1 \u001b P kjj\n\u00f0\n\u00de\n\u001f\n\u001e\n\u001b log\nP :kjj\n\u00f0\n\u00de\n1 \u001b P :kjj\n\u00f0\n\u00de\n\u001f\n\u001e\nThe approach developed in ref. 22 produces a similar measure, but with useful \nregularization. The model sets up the text corpus as a multinomial distribution, with \na Dirichlet prior on the probabilities of different topics based on different words:\n\u03c0 \ue018Dirichlet \u03b1\n\u00f0 \u00de\nyk \ue018Multinomial nk; \u03c0k\n\u00f0\n\u00de\nUsing an informative prior avoids overfitting to rare terms. We followed the \napproach in ref. 22 and used the average number of words per text such that \u03b1\u2009=\u20095 in \nour application. Furthermore, we emphasized words for which the estimates were \nmore certain by reweighting words using their variance. To this end, we employed \nthe z-score of the log odds ratio, which is the regularized log odds ratio divided by its \nstandard deviation. This z-score is the measure displayed on the vertical axis in Fig. 6.\nPreregistration. This study has been preregistered at the American Economic \nAssociation\u2019s registry for randomized controlled trials, under AEARCTR-0004090.\nEthics oversight. This study was approved by the internal review boards of \nWashington University in St. Louis (201803178) and Stanford University \n(eProtocol 46325).\nReporting Summary. Further information on research design is available in the \nNature Research Reporting Summary linked to this article.\nData availability\nData and replication materials are available at the Harvard Dataverse (https://doi.\norg/10.7910/DVN/VXJPN5).\nCode availability\nStatistical code are available as part of the replication materials at the Harvard \nDataverse (https://doi.org/10.7910/DVN/VXJPN5).\nAcknowledgements\nWe thank Clara Vandeweerdt for research assistance and audiences at Yale University and \nthe 2019 International Political Economy Society Conference for comments. M.M.B. and \nK.F.S. gratefully acknowledge financial support from the Swiss Network for International \nStudies and the Weidenbaum Center on the Economy, Government, and Public Policy at \nWashington University in St. Louis. K.F.S. thanks the Institute for Research in the Social \nSciences at Stanford University for a faculty fellowship.\nAuthor contributions\nM.M.B., K.F.S. and E.vL. contributed equally to the study design, data collection and \nanalysis, interpretation of the results and writing of the manuscript.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at https://doi.org/10.1038/s41558-020-00914-6.\nSupplementary information is available for this paper at https://doi.org/10.1038/\ns41558-020-00914-6.\nCorrespondence and requests for materials should be addressed to M.M.B. or K.F.S.\nPeer review information Nature Climate Change thanks Rebecca Bromley-Trujillo, \nChristopher Warshaw and the other, anonymous, reviewer(s) for their contribution to \nthe peer review of this work.\nReprints and permissions information is available at www.nature.com/reprints.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTUrE ClImaTE CHanGE\nExtended Data Fig. 1 | Preferences for distributing climate costs over time (weighted data). The percentage of respondents who prefer constant, \nincreasing, decreasing, or inverse U-shaped intertemporal allocations of climate costs (n\u2009=\u200910,075).\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTUrE ClImaTE CHanGE\nExtended Data Fig. 2 | The causal effects of cost path, cost level, and other policy attributes on public support. Dots with horizontal lines are point \nestimates from a linear least squares regression of climate policy chosen (n\u2009=\u2009129,280) on randomly assigned cost path, cost level, and revenue investment \nattributes. Error bars indicate 95% and 99% confidence intervals computed from robust standard errors clustered by respondent.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTUrE ClImaTE CHanGE\nExtended Data Fig. 3 | Support for climate action as a function of cost paths and cost levels (weighted data). Dots with horizontal lines are point \nestimates from linear least squares regressions of climate policy chosen on randomly assigned cost path and cost level attributes. Error bars indicate 95% \nand 99% confidence intervals computed from robust standard errors clustered by respondent, n(policy profiles)=129,280.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTUrE ClImaTE CHanGE\nExtended Data Fig. 4 | Support for climate action as a function of cost paths and cost levels by country (weighted data). Dots with horizontal lines are \npoint estimates from linear least squares regressions of climate policy chosen on randomly assigned cost path and cost level attributes. Error bars indicate \n95% and 99% confidence intervals computed from robust standard errors clustered by respondent, n(France, policy profiles)=32,000, n(Germany, policy \nprofiles)=32,000, n(United Kingdom, policy profiles)=32,000, n(United States, policy profiles)=33,280.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTUrE ClImaTE CHanGE\nExtended Data Fig. 5 | Support for climate action as a function of cost paths by cost level (weighted data). Causal effects of climate cost paths on policy \nsupport estimated separately for each randomly assigned cost level, n(0.5% of GDP, policy profiles)=32,305, n(1% of GDP, policy profiles)=32,373, \nn(2% of GDP, policy profiles)=32,367, n(2.5% of GDP, policy profiles)=32,235. Dots with horizontal lines are point estimates from linear least squares \nregressions of climate policy chosen on randomly assigned cost path attributes. Error bars indicate 95% and 99% confidence intervals computed from \nrobust standard errors clustered by respondent.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNaTUrE ClImaTE CHanGE\nExtended Data Fig. 6 | Support for climate action as a function of cost paths by cost level (weighted data). Results from a conjoint experiment conducted \nin a separate section of the United States survey that randomized the year in which contributions would start, n(policy profiles)=10,880, see Methods \nsection for details. Dots with horizontal lines are point estimates from linear least squares regressions of climate policy chosen on randomly assigned cost \npath attributes. Error bars indicate 95% and 99% confidence intervals computed from robust standard errors clustered by respondent.\nNature Climate Change | www.nature.com/natureclimatechange\n\n1\nnature research  |  reporting summary\nOctober 2018\nCorresponding author(s):\nMichael Bechtel\nLast updated by author(s): Jul 10, 2020\nReporting Summary\nNature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nStata version 15, R.\nData analysis\nStata version 15, R.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. \nWe strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A list of figures that have associated raw data \n- A description of any restrictions on data availability\nData and replication materials will be made available at the Harvard Dataverse. \nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\n\n2\nnature research  |  reporting summary\nOctober 2018\nBehavioural & social sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nSurvey\nResearch sample\nRepresentative sample of the adult populations in the US, UK, Germany, and France.\nSampling strategy\nWe fielded our survey in four major economies that belong to the group of Annex 1 countries (France, Germany, the United Kingdom, \nand the United States). The survey was conducted online by YouGov on representative samples of the adult populations. YouGov \nemploys matched sampling in which interviews are conducted from participants in YouGov\u2019s online panel. Matched sampling involves \ntaking a stratified random sample of the target population and then matching available internet respondents to the target sample using \npropensity scores. The propensity score model included age, gender, years of education, and region for the European countries and \ngender, age, race/ethnicity, region, and education for the United States.\nData collection\nComputer (online survey)\nTiming\nUnited States: The field period was December 18, 2018 to January 3, 2019.  \nFrance, Germany, United Kingdom: The field period was March 31, 2019 to April 04, 2019. \nData exclusions\nUnited States: YouGov interviewed 4081 respondents. 6 were excluded from the analyses because they did not meet the age \nrequirement. The final n was 4,075. \nUK: YouGov interviewed 2135 respondents who were then matched down to a sample of 2000 to produce the final dataset using \nmatched sampling (see above). \nGermany: YouGov interviewed 2093 respondents who were then matched down to a sample of 2000 to produce the final dataset using \nmatched sampling (see above). \nFrance: YouGov interviewed 2039 respondents who were then matched down to a \nsample of 2000 to produce the final dataset using matched sampling (see above). \nNon-participation\nThe universe of possible individual was YouGov's opt-in sample. Respondents were drawn from this sample using the matched sampling \napproach described above.\nRandomization\nRandom allocation.\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology\nAnimals and other organisms\nHuman research participants\nClinical data\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nAntibodies\nAntibodies used\nNA\nValidation\nNA\nEukaryotic cell lines\nPolicy information about cell lines\nCell line source(s)\nNA\nAuthentication\nNA\n\n3\nnature research  |  reporting summary\nOctober 2018\nMycoplasma contamination\nNA\nCommonly misidentified lines\n(See ICLAC register)\nNA\nPalaeontology\nSpecimen provenance\nNA\nSpecimen deposition\nNA\nDating methods\nNA\nTick this box to confirm that the raw and calibrated dates are available in the paper or in Supplementary Information.\nAnimals and other organisms\nPolicy information about studies involving animals; ARRIVE guidelines recommended for reporting animal research\nLaboratory animals\nNA\nWild animals\nNA\nField-collected samples\nNA\nEthics oversight\nNA\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nHuman research participants\nPolicy information about studies involving human research participants\nPopulation characteristics\nSee above.\nRecruitment\nThe universe of possible individual was YouGov's opt-in sample. Respondents were drawn from this sample using the matched \nsampling approach described above to obtain samples representative of adult populations in the countries studied.\nEthics oversight\nThe study was approved by the Internal Review Boards at Washington University in St. Louis (#201803178) and Stanford \nUniversity (eProtocol 46325).\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nClinical data\nPolicy information about clinical studies\nAll manuscripts should comply with the ICMJE guidelines for publication of clinical research and a completed CONSORT checklist must be included with all submissions.\nClinical trial registration\nNA\nStudy protocol\nNA\nData collection\nNA\nOutcomes\nNA\nChIP-seq\nData deposition\nConfirm that both raw and final processed data have been deposited in a public database such as GEO.\nConfirm that you have deposited or provided access to graph files (e.g. BED files) for the called peaks.\nData access links \nMay remain private before publication.\nNA\nFiles in database submission\nNA\n\n4\nnature research  |  reporting summary\nOctober 2018\nGenome browser session \n(e.g. UCSC)\nNA\nMethodology\nReplicates\nNA\nSequencing depth\nNA\nAntibodies\nNA\nPeak calling parameters\nNA\nData quality\nNA\nSoftware\nNA\nFlow Cytometry\nPlots\nConfirm that:\nThe axis labels state the marker and fluorochrome used (e.g. CD4-FITC).\nThe axis scales are clearly visible. Include numbers along axes only for bottom left plot of group (a 'group' is an analysis of identical markers).\nAll plots are contour plots with outliers or pseudocolor plots.\nA numerical value for number of cells or percentage (with statistics) is provided.\nMethodology\nSample preparation\nNA\nInstrument\nNA\nSoftware\nNA\nCell population abundance\nNA\nGating strategy\nNA\nTick this box to confirm that a figure exemplifying the gating strategy is provided in the Supplementary Information.\nMagnetic resonance imaging\nExperimental design\nDesign type\nNA\nDesign specifications\nNA\nBehavioral performance measures\nNA\nAcquisition\nImaging type(s)\nNA\nField strength\nNA\nSequence & imaging parameters\nNA\nArea of acquisition\nNA\nDiffusion MRI\nUsed\nNot used\nPreprocessing\nPreprocessing software\nNA\n\n5\nnature research  |  reporting summary\nOctober 2018\nNormalization\nNA\nNormalization template\nNA\nNoise and artifact removal\nNA\nVolume censoring\nNA\nStatistical modeling & inference\nModel type and settings\nNA\nEffect(s) tested\nNA\nSpecify type of analysis:\nWhole brain\nROI-based\nBoth\nStatistic type for inference\n(See Eklund et al. 2016)\nNA\nCorrection\nNA\nModels & analysis\nn/a Involved in the study\nFunctional and/or effective connectivity\nGraph analysis\nMultivariate modeling or predictive analysis\nFunctional and/or effective connectivity\nNA\nGraph analysis\nNA\nMultivariate modeling and predictive analysis\nNA\n\n\n    Scientific Research Findings:", "answer": "We find that the public has a clear preference for constant carbon pricing schedules. We also note that only 12% support an increasing cost path that would gradually ramp up costs over time. In addition, these preferences are surprisingly similar across countries. When we randomly provide half of the respondents with cost schedules that specify the average costs associated with a plan to be low and another half with a high cost version of our question, we still find that most people prefer constant cost schedules.\nAdditional results that rely on a climate policy conjoint experiment that randomizes both cost schedules and cost levels confirm this finding. Our study does not, however, rule out the possibility that low and stable cost plans may successfully introduce citizens to carbon pricing and build future support for higher carbon price plans in the long run.", "id": 59}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Articles\nhttps://doi.org/10.1038/s41558-021-01128-0\n1School of International Affairs and Department of Civil and Environmental Engineering, Pennsylvania State University, University Park, PA, USA. 2Joint \nGlobal Change Research Institute, Pacific Northwest National Laboratory, College Park, MD, USA. 3School of the Environment, Yale Program on Climate \nChange Communication, Yale University, New Haven, CT, USA. 4Center for Global Sustainability, School of Public Policy, University of Maryland, College \nPark, MD, USA. 5School of Global Policy and Strategy, University of California San Diego, La Jolla, CA, USA. 6Scripps Institution of Oceanography, \nUniversity of California San Diego, La Jolla, CA, USA. 7Mechanical and Aerospace Engineering, Jacobs School of Engineering, University of California San \nDiego, La Jolla, CA, USA. 8The Brookings Institution, Washington, DC, USA. \u2709e-mail: weipeng@psu.edu\nA\ns governments get serious about decarbonization, politi-\ncal leaders in large and politically diverse countries need to \ngrapple with huge variations in political and administrative \nfeasibility within their countries. That heterogeneity in interests and \ncapabilities has led many federal governments to encourage or tol-\nerate large internal variations in policy effort. Diverse studies have \npointed to the benefits of heterogeneous approaches for experimen-\ntation and learning1\u20134. Yet these realities in climate politics have \nnot been well represented in leading modelling frameworks, which \ntypically assume nationally uniform policy efforts5\u20138. This gap in \nmodelling work also reflects the widely held assumption by poli-\ncymakers that heterogeneous subnational policy efforts will be a lot \nmore costly than nationally uniform efforts9.\nTo assess the potential increase in cost from heterogeneous \nsubnational policy formation compared with theorized opti-\nmal uniform nationwide policy, we study the case of the United \nStates. The United States is the world\u2019s second largest emitter and \nalready displays one of the world\u2019s largest variations in subna-\ntional action. Bottom-up coalitions, such as the \u2018America is All In\u2019 \ninitiative, have gathered thousands of signatories from political \nleaders in cities, states, companies and universities that represent \na constituency of more than half the U.S. population10. More than \n30 states have completed state-level climate action plans, or are \nin the process of developing one11. Nearly 40 states have created \nrenewable portfolio standards or voluntary renewable energy \ngoals to facilitate low-carbon transition12. Meanwhile, some states \nare making limited effort and are under no sustained public pres-\nsure to strengthen it.\nHere we focus on heterogeneity in the stringency of policy efforts \nmade by each state. For ease, we will represent those policy efforts \nas state-varying prices on carbon, although in practice no enterprise \nuses only simple price instruments\u2014a variety of other mechanisms, \nsuch as renewable portfolio standards, low-carbon fuel standards \nand industrial policies are the norm13,14. Our approach, by design, \nallows us to assess the long-term dynamics that arise when states \nvary in their willingness to act and when the country, as a whole, \nvaries the national decarbonization targets it might pursue. We see \nthis long-term view of heterogeneity in action as a complement to \nstudies that have focused on much nearer-term policy heterogene-\nity and thus been able to look more granularly at specific policy \ninstruments15. (See Supplementary Note\u00a02 for a comparison with \nAmerica\u2019s Pledge study that examines nearer-term policies13.)\nMethodologically, we employ a process-based integrated assess-\nment model (IAM): the Global Change Assessment Model with \nstate-level detail in the United States (GCAM-USA; Methods)16. \nIt includes detailed representation of sectoral and technology \noptions, as well as the interactions between the economic, energy \nand land-use systems. These representations allow us to assess the \ndeployment trajectories of critical technologies in each state, the \ncompetition between different technologies in various sectors and \nthe impacts of varying cost and policy assumptions on infrastruc-\nture investments across space and time. Our analysis hence goes \nbeyond prior studies based on simpler economic models17\u201319 by \nidentifying critical sectors, technologies and processes that deter-\nmine the cost of heterogeneous policy efforts.\nScenario design\nWe examine 12 mitigation scenarios that vary across two dimensions:\n\ti.\t\nNational mitigation effort, measured by four targets of nation-\nal total greenhouse gas (GHG) emissions for 2050 (for exam-\nple, 20%, 40%, 60% and 80% below 2005 levels; Supplementary \nFig.\u00a01). We also include a 95% decarbonization target as a sen-\nsitivity analysis, which is close to the net-zero emissions target \nset by the Biden Administration (Supplementary Fig. 7).\nThe surprisingly inexpensive cost of state-driven \nemission control strategies\nWei Peng\u200a \u200a1\u2009\u2709, Gokul Iyer\u200a \u200a2, Matthew Binsted\u200a \u200a2, Jennifer Marlon\u200a \u200a3, Leon Clarke4, \nJames A. Edmonds\u200a \u200a2 and David G. Victor\u200a \u200a5,6,7,8\nTraditionally, analysis of the costs of cutting greenhouse gas emissions has assumed that governments would implement ideal-\nized, optimal policies such as uniform economy-wide carbon taxes. Yet actual policies in the real world, especially in large fed-\neral governments, are often highly heterogeneous and vary in political support and administrative capabilities within a country. \nWhile the benefits of heterogeneous action have been discussed widely for experimentation and leadership, little is known \nabout its costs. Focusing on the United States, we represent plausible variation (by more than a factor of 3) in the stringency of \nstate-led climate policy in a process-based integrated assessment model (GCAM-USA). For a wide array of national decarbon-\nization targets, we find that the nationwide cost from heterogeneous subnational policies is only one-tenth higher than nation-\nally uniform policies. Such results hinge on two critical technologies (advanced biofuels and electricity) for which inter-state \ntrade ameliorates the economic efficiencies that might arise with heterogeneous action.\nNature Climate Change | VOL 11 | September 2021 | 738\u2013745 | www.nature.com/natureclimatechange\n738\n\nArticles\nNATUrE CliMATE ChAnGE\n\tii.\t Subnational policy approach, modelled as three degrees of \nheterogeneity in the stringency of state-level climate policy. The \nUniform approach assumes a nationally uniform policy imple-\nmentation. It is represented as a uniform price on carbon, mod-\nelled by equalizing the marginal abatement cost (MAC) across \nstates in GCAM-USA. By contrast, the Hybrid and Heterogene-\nous approaches assume state heterogeneity in policy stringency, \nwhich is represented as state-varying carbon prices and MACs. \nThe MAC implicitly measures the highest willingness to pay \nby industries and households in each state to mitigate carbon \nemissions. A higher MAC implies a higher price on carbon and \nthus more stringent climate policy efforts.\nWe incorporate key political factors and processes in our sce-\nnario design to bring it closer to the realities. Because the United \nStates is a democracy, we use public opinions to proxy for the het-\nerogeneity in stringency of state policies. Indeed, there is a vast lit-\nerature on the important role for public attitudes in shaping climate \npolicy (see Supplementary Note\u00a01 as well as refs. 20\u201322). Under the \nHeterogeneous approach, we model the widest heterogeneity in \npolicy stringency (by more than a factor of 3) across all the states. \nUnder the Hybrid approach, we temper that heterogeneity by first \noverlaying a modest level of nationwide carbon policy that all states, \nat minimum, must adopt. When looking at state-level policies, \na majoritarian 50% define who gets elected, and thus the median \nvoter matters; we hence allow a non-linear, steep rise in policy strin-\ngency when support exceeds 50% (Figs.\u00a01 and 2 and Methods). At \nthe federal level, the rules for adoption are stricter and, since 1975, \nhave required a 60% cloture threshold before the body can pass leg-\nislation (rule 22 in ref. 23); we hence set the nationwide carbon policy \nto be the 40th percentile of the stringency levels across all the states. \nWe are mindful that state and federal policies arise through much \nmore complex and nuanced processes; our purpose here is simply \nto show a plausible method for quantifying degrees of heterogeneity \nand their implications for mitigation cost (see Sensitivity analysis \nsection for alternative formulations for policy heterogeneity).\nTo make computation easier, we set the relative ratios of state-level \ncarbon prices based on public opinions, then let the model compute \nthe whole set of carbon price values to achieve the national mitiga-\ntion targets (see Fig.\u00a02 for more details).\nHeterogeneity at state level\nWe categorize states into three groups, namely low-, medium-, and \nhigh-supporting states, depending on the level of current public \nsupport for climate policy (Fig.\u00a01). With the Uniform approach, \nthe carbon price is uniform across the three groups of states, but \nincreases dramatically as the national mitigation efforts become \nmore stringent (solid lines in Fig.\u00a0 3: US$74, US$210, US$670 \nand US$1,557 per ton CO2 for national targets of 20%, 40%, 60% \nand 80% decarbonization, respectively). The substantial increase \nin carbon price is a direct result of an increasing MAC, which \nsuggests that availability of low-cost mitigation options becomes \nincreasingly constrained to achieve more ambitious decarboniza-\ntion goals.\nCompared with the Uniform approach, the main effect of intro-\nducing heterogeneous subnational policies is to shift the burden of \nemissions reductions from low- to medium- and high-supporting \nstates, while the country overall still achieves the same national \nmitigation target. The effect of such a shift is driven by much lower \ncarbon prices in the low-supporting states and is greatest under the \nHeterogeneous approach, since it displays the widest heterogeneity \nin state-level effort.\nMitigation efforts by sector\nBy comparing scenarios having ambitious versus less ambitious \nnational mitigation effort (for example, the 80% and 40% decar-\nbonization scenarios with the Uniform policy approach in Fig.\u00a04), \nwe confirm the common finding in the literature that deep decar-\nbonization generally requires decarbonizing the electricity sector \nfirst (the sector with the lowest mitigation cost) then moves onto \n\u2018harder-to-abate\u2019 sectors such as industry, residential and transport \nsectors24\u201326. This sectoral pattern remains robust under substantial \nstate heterogeneity in policy stringency, indicating that mitiga-\ntion far beyond the electricity sector is necessary to achieve deep \ndecarbonization goals, irrespective of uniform or heterogeneous \napproaches.\nIntroducing state heterogeneity affects the sectoral alloca-\ntion of CO2 mitigation. At the national level, all three subnational \npolicy approaches achieve CO2 mitigation with similar sectoral \nRatio of the carbon price in each state to the carbon price in District of Columbia\n0.3\n0.4\n0.5\n0.6\n0.7\n0.8\n0.85\n0.9\n1\nLow support\nMedium support High support\nPercentage of adults who think their local officials should do more to address\nglobal warming (2018)\n45%\n50%\n55%\n60%\n65%\na\nb\nFig. 1 | State-level variations in public support for climate policy and its \nimpact on carbon pricing under the Heterogeneous approach. a, State \nheterogeneity in public support level for climate policy in 2018. We show the \npercent adult population in each state who think their local officials should \ndo more to address global warming (data from ref. 49). b, State heterogeneity \nin carbon prices under the Heterogeneous approach. Here we present the \nratio of the carbon price in each state (modelled as the marginal abatement \ncost of carbon policies) to the carbon price in the District of Columbia, where \nthe support rate is the highest. We group the 50 states and the District of \nColumbia into low-, medium- and high-supporting groups (indicated by red, \nyellow and blue, respectively; each group includes 17 states), based on their \ncurrent climate policy support levels as shown in a. We assume the state-level \ncarbon price varies non-linearly with public support level and decreases \ndramatically when the support level drops to below 50% according to the \nmedium voter theorem. This assumption leads to much lower carbon prices in \nstates with low support rate. More details are included in Fig.\u00a02, Methods and \nSupplementary Methods.\nNature Climate Change | VOL 11 | September 2021 | 738\u2013745 | www.nature.com/natureclimatechange\n739\n\nArticles\nNATUrE CliMATE ChAnGE\ncontributions. At the subnational level, moving from the Uniform to \nthe Hybrid and Heterogeneous approach, the low-supporting states \nreduce less CO2 from the electricity and refinery sectors, while \nthe medium- and high-supporting states reduce more from these  \ntwo sectors.\nWith a heterogeneous approach, mitigation in harder-to-abate \nsectors is especially important in states with higher support rate. \nFor instance, to achieve a national mitigation effort of 80% decar-\nbonization with a heterogeneous policy approach, the contribu-\ntion of non-electricity sectors to total CO2 mitigation from 2015 to \n2050 is only 48% in the low-supporting states, while it is 60% in \nthe medium- and high-supporting states (Fig.\u00a04). This suggests that \ndecarbonizing the electricity sector is a key strategy for all the states, \nand the major differences between states with low and high sup-\nport levels occur outside the electricity sector, where mitigation is  \nmore costly.\nImplications for nationwide cost\nThe nationwide mitigation cost increases dramatically from mod-\nerate to ambitious national mitigation efforts (Fig.\u00a0 5). However, \nmore surprising is that the difference in total cost does not vary \nmuch when the subnational policy approach is heterogeneous \nas compared with uniformity across the states. Under the Hybrid \nand Heterogeneous approaches, the economic cost of mitigation \ndrops in the low-supporting states by up to half; the medium- and \nhigh-supporting states take up most of that slack. And despite \nthe factor of 3 variation in carbon prices across states under the \nHeterogeneous approach, the national mitigation cost is only slightly \nhigher than with the Uniform approach. For a national mitigation \neffort of 20%, 40%, 60% and 80% decarbonization, the nationwide \ncost under the heterogeneous approach is only 14%, 9%, 4% and 5% \nhigher than the uniform approach, respectively.\nThis surprising result (that the overall costs do not rise much in spite \nof huge heterogeneity across states) is largely driven by the flexibility  \nacross states in terms of electricity trade and the location of \nenergy-intensive industries (for example, bio-refineries; see the \nelectricity and bioliquids production patterns in Supplementary \nFig.\u00a0 3). In other words, despite the subnational heterogeneity in \npolicy stringency, energy markets are tightly coupled across states, \nallowing much of the heterogeneity in policy to be arbitraged \nthrough trade activities in the energy markets. Comparing the \nHeterogeneous and Hybrid approach with the Uniform approach, \nmany states adjust the amount of electricity production, technology \nchoices and trading volume with neighbouring states. Nationally, \nto achieve an 80% cut in GHG emissions, we find ~10% more \ninter-grid electricity trade under the heterogeneous policy approach \nthan the uniform approach (see Extended Data Fig.\u00a02 for net elec-\ntricity trade across 15 grid regions; Supplementary Fig.\u00a0 5 shows \na map of the grid regions). There is also a shift in the geographic \npattern of where critical mitigation technologies, such as bioen-\nergy with carbon capture and storage (BECCS), are being deployed \n(Extended Data Fig.\u00a03 and Supplementary Fig.\u00a06). This is because, \nwith a heterogeneous approach, the mitigation efforts are shifted \nfrom low- to high-supporting states, pushing the green states to \nachieve deep cuts in their emissions by turning to negative emis-\nsions technologies such as BECCS.\nOur finding indicates that deep decarbonization is nonetheless \ncostly. The socioeconomic implications can be large and will vary \nacross states that have different levels of wealth, economic struc-\nture and carbon intensity (Extended Data Fig.\u00a04). There is growing \nattention on how various forms of justice and equity considerations \nintersect with deep decarbonization, which is revealed not least \nin current plans in the United States and Europe that frame deep \ndecarbonization as broad social changes.\n260\nHeterogeneous\nHybrid\nHybrid\nHeterogeneous\nUniform\n240\n220\n200\n180\nRatio relative to \nthe highest carbon price\nState-level carbon prices in 2050 (2015 US$ per ton CO2)\nPublic support level for climate policy\n(Percentage of adults who think their local offcials should do\nmore to address global warming)\n1.00\n0.75\n0.50\n0.25\n45%\n50%\n55%\n60%\n65%\n160\n140\n120\n100\n80\n60\nWY\nND\nMT\nWV\nIA\nSD\nVA\nKY\nNE\nOK\nAK\nIN\nKS\nTN\nAL\nMN\nAR\nID\nNH\nMO\nLA\nMS\nTX\nOH\nMI\nWA\nSC\nPA\nCO\nIL\nNM\nCT\nME\nAZ\nWI\nNV\nNC\nUT\nOR\nGA\nVT\nMA\nDE\nRI\nHI\nNJ\nMD\nCA\nFL\nNY\nDC\nFig. 2 | Model-computed state-level carbon prices in 2050 to achieve a national target of 40% decarbonization by 2050 relative to 2005 under three \nsubnational policy approaches. The carbon prices are modelled as the marginal abatement cost of carbon policies in GCAM-USA. The inset shows \ndifferent assumptions for how carbon price varies as a function of public support under the Hybrid and Heterogeneous approaches, respectively (more in \nMethods and Supplementary Methods). Based on these assumptions, the GCAM-USA model then computes the whole set of state-level carbon prices to \nachieve the national decarbonization target, as presented in the main figure.\nNature Climate Change | VOL 11 | September 2021 | 738\u2013745 | www.nature.com/natureclimatechange\n740\n\nArticles\nNATUrE CliMATE ChAnGE\nSensitivity analyses\nWe test the sensitivity of the results to six alternative formulations for \npolicy heterogeneity (Fig.\u00a06, Extended Data Fig.\u00a05 and Supplementary \nMethods), including: (1) assuming a linear relationship between pub-\nlic support and policy stringency (\u2018Linear\u2019), (2) increasing the state \nheterogeneity in policy stringency by varying the state-level carbon \nprices by a factor of 10 (\u2018+ range\u2019), (3) and (4) assuming no effort and \nhence zero carbon price in three or five lowest-supporting states (\u20183 \nzero\u2019 and \u20185 zero\u2019), (5) using the public opinion results from a different \nsurvey question on \u2018Do you think your Governor should do more to \naddress global warming?\u2019 (\u2018Gov\u2019) and (6) using existing commitments \nto climate action (instead of public opinion) to proxy for policy strin-\ngency (\u2018AP\u2019: America\u2019s Pledge). Most of these alternative formulations \ndo not alter the core result that the heterogeneous policy approach is \nonly marginally more expensive. However, the cost of deep decarbon-\nization can increase significantly if a few states with the lowest sup-\nport levels are not engaged in climate action at all (\u20183 zero\u2019 and \u20185 zero\u2019 \nscenarios). For instance, to achieve 80% decarbonization nationally, \na zero effort in those locales drives costs for the rest of the country \nup by 25\u201370%. At least some modest floor level of effort by all the \nstates is critical to avoid inter-state carbon leakage and a significant \ncost escalation.\nWe also consider three technology constraints based on what prior \nresearch has shown to be most pivotal in determining future mitiga-\ntion costs (Fig.\u00a06)27,28. For all three policy approaches, we find much \nhigher mitigation costs under: (a) limited electricity infrastructure \ninvestment and production, (b) no investment in carbon capture \nand storage (CCS) and (c) limited availability of biomass. The great-\nest impacts on mitigation costs come from constraining CCS and \nbiomass because this restricts the role of BECCS and forces greater \nreliance upon extremely expensive technologies outside the electric-\nity and refinery sectors (cf. Extended Data Fig.\u00a06 with Fig.\u00a04), such \nas hydrogen use in transport and industrial sectors. Relaxing the bio-\nmass constraint, by contrast, significantly lowers the mitigation costs \n(Supplementary Figs.\u00a0 11 and 12). Limiting electricity production \nhas a lesser impact on overall costs because, in the face of such con-\nstraints, reductions in energy demand through efficiency measures, \nincreases in natural gas use coupled with CCS, as well as more miti-\ngation in the refinery sector through BECCS, can provide additional \ndecarbonization. Detailed discussions are included in Supplementary \nNote\u00a03. Although these technology constraints raise the total national \ncost of mitigation, our main finding remains robust: a heteroge-\nneous policy approach, relative to a uniform approach, is only slightly  \nmore expensive.\n1,500\nNational mitigation efforts\nSubnational policy approaches\n20% decarbonization\nUniform\nHybrid\nHeterogeneous\nHigh-supporting states\nLow-supporting states\n40% decarbonization\n60% decarbonization\n80% decarbonization\n1,000\n500\nAverage carbon price in 2050 (2015 US$ per ton CO2)\n0\n500\n1,000\nReduction in energy-related CO2 emissions in 2050 relative to 2015\n(million ton CO2)\n1,500\nFig. 3 | Reduction in energy-related CO2 emissions and carbon prices in 2050 in low- and high-supporting states. The size of the dots represents four \nlevels of national mitigation effort: 20%, 40%, 60% and 80% decarbonization by 2050 relative to 2005. The solid, dotted and dashed lines represent \nthree subnational policy approaches: Uniform, Hybrid and Heterogeneous, respectively. Comparing the Hybrid and Heterogeneous approaches with the \nUniform approach, the lower carbon prices in the low-supporting states (modelled as lower marginal abatement cost) yield lesser CO2 reductions within \nthat category of states (that is, comparing dashed/dotted red lines with solid red lines), thus shifting more emissions reductions to the high-supporting \nstates (that is, comparing dashed/dotted blue lines with solid blue lines). The national average carbon price and national total CO2 emissions reduction are \npresented in Supplementary Fig.\u00a02.\nNature Climate Change | VOL 11 | September 2021 | 738\u2013745 | www.nature.com/natureclimatechange\n741\n\nArticles\nNATUrE CliMATE ChAnGE\nDiscussion\nU.S. policy analysts have famously celebrated subnational action \nby states as federalist laboratories of democracy29. Unshackled \nfrom central mandates, or where central governments are mired in \npolitical gridlock, the logic of federalism globally emphasizes the \nutility of local experiments and the practical reality that not every \n40% D\nUni\n64%\n61%\n60%\n60%\n44%\n45%\n52%\n68%\n67%\n65%\n43%\n43%\n40%\n61%\n60%\n59%\n39%\n40%\n40%\n63%\n62%\n42%\n43%\n43%\n0\na\nb\nc\nd\n\u20132,000\n\u20134,000\nMillion ton CO2\n0\n0\n0\n\u2013500\n\u20131,000\n\u20131,500\n\u2013500\n\u20131,000\n\u20131,000\n\u20132,000\n\u20131,500\nMillion ton CO2\nMillion ton CO2\nMillion ton CO2\nHyb\nHet\nUni\nHyb\nHet\nElectricity\nTransportation\nResidential/commercial\nIndustry\nRefinery\n80% D\n40% D\nUni Hyb\nHet\nUni\nHyb\nHet\n80% D\n40% D\nUni Hyb\nHet\nUni\nHyb\nHet\n80% D\n40% D\nUni Hyb\nHet\nUni\nHyb\nHet\n80% D\nFig. 4 | Reduction in energy-related CO2 emissions. a\u2013d, Reduction in energy-related CO2 emissions in 2050 relative to 2015 by sector at the national level \n(a) and for low- (b), medium- (c) and high-supporting (d) states. The labels \u201840% D\u2019 and \u201880% D\u2019 represent two levels of national mitigation efforts, that \nis, 40% and 80% decarbonization by 2050 relative to 2005. \u2018Uni\u2019, \u2018Hyb\u2019 and \u2018Het\u2019 represent three subnational policy approaches, that is, Uniform, Hybrid \nand Heterogeneous. The white percentages represent the share of the electricity sector in all-sector CO2 reduction. Results for other levels of national \nmitigation efforts are presented in Extended Data Fig.\u00a01. Detailed energy technology choices for electricity generation, liquids production and end-use \nsectors are reported in Supplementary Fig.\u00a04.\nUniform\n20% D\n20%\nLow\nHigh\nMedium\nLow\nLow\nLow\nHigh\nHigh\nHigh\nMedium\nMedium\nMedium\n40%\n60%\n80%\n100%\n0\n40% D\n60% D\nNational mitigation effort\n80% D\n0\n1%\n2%\n3%\nMitigation cost in 2050 as a fraction of projected GDP\n4%\n5%\n6%\nHybrid\nHeterogeneous\nContribution of three groups of states\nto national CO2 emissions in 2016\nUniform\nHybrid\nHeterogeneous\nUniform\nHybrid\nHeterogeneous\nUniform\nHybrid\nHeterogeneous\nFig. 5 | Carbon mitigation cost in 2050 as a fraction of projected 2050 gross domestic product (GDP), by national mitigation effort (20\u201380% \ndecarbonization, indicated as 20\u201380% D) and by subnational policy approach (Uniform, Hybrid and Heterogeneous). Colours reflect the state \ngroupings (Fig.\u00a01b), and the inset shows the contribution by low- (red), medium- (yellow) and high-supporting (blue) states to national total CO2 \nemissions in 2016 (ref. 50). See Extended Data Fig.\u00a04 for state-by-state results.\nNature Climate Change | VOL 11 | September 2021 | 738\u2013745 | www.nature.com/natureclimatechange\n742\n\nArticles\nNATUrE CliMATE ChAnGE\nnational subunit will have the same political preferences or admin-\nistrative capabilities.\nThe question of heterogeneity is hardly new to climate policy \nanalysis, although over history most of this debate has played out \nwith regard to global politics, where it has been taken as axiom-\natic that jurisdiction-wide policies would achieve the greatest lever-\nage on emissions and the lowest cost30,31. That assumption persists, \neven as the Paris Agreement and other policy frameworks have \nevolved to encourage high degrees of heterogeneity32\u201337. Where \nthese issues have been analysed closely, they appear to confirm the \ncosts of heterogeneity. Studies have found that heterogeneous action \nacross countries, driven by equity concerns38, delays in action39\u201341 or \nvariations in investment risks42, could increase the global mitigation \ncost by as much as 40%28. (See Supplementary Note\u00a04 for a fuller \nliterature review in the global context.)\nControlling pollution of warming gases may prove to be the most \nelaborate and costly environmental policy undertaking yet, and the \nanalysis we present here suggests that the extra nationwide economic \ncost of federalist variation in state-level policy can be very mod-\nest\u2014much smaller by an order of magnitude than variations in the \noverall level of nationwide effort. For policymakers, this insight sug-\ngests that federalist approaches could be more sustainable, because \nhigher costs are aligned with the states that are politically more will-\ning to bear them. Future research should aim to simulate or measure \n1.0%\na\nb\n0.8%\n10%\n8%\n6%\n4%\n2%\n0\n0.6%\n0.4%\nHigh\nMed\nLow\nHigh\nMed\nLow\nMitigation cost as percentage of GDP\nMitigation cost as percentage of GDP\n0.2%\n0\nUni\nHyb\nHet\nAlternative formulations\nfor polilcy heterogeneity\nLow electricity\ninfrastructure\n40% decarbonization\n80% decarbonization\nNo CCS\nLow biomass\nUni\nHyb\nHet\nUni\nHyb\nHet\nUni\nHyb\nHet\nHet\n(linear)\nHet\n(+range)\nHet\n(3 zero)\nHet\n(5 zero)\nHet\n(Gov)\nHet\n(AP)\nUni\nHyb\nHet\nAlternative formulations\nfor polilcy heterogeneity\nLow electricity\ninfrastructure\nNo CCS\nLow biomass\nUni\nHyb\nHet\nUni\nHyb\nHet\nUni\nHyb\nHet\nHet\n(linear)\nHet\n(+range)\nHet\n(3 zero)\nHet\n(5 zero)\nHet\n(Gov)\nHet\n(AP)\nFig. 6 | Sensitivity analysis of carbon mitigation cost for two levels of national mitigation effort. a,b, Carbon mitigation cost in 2050 as a fraction of \nprojected 2050 GDP for 40% (a) and 80% (b) decarbonization, with state groups coloured as in Fig.\u00a01. See Extended Data Fig.\u00a05 and Supplementary \nFigs.\u00a08\u201310 for other national decarbonization levels. The first three bars on the left represent our main scenarios shown in Fig.\u00a05. Alternative formulations of \npolicy heterogeneity include: \u2018Linear\u2019, assuming a linear relationship between public support and policy stringency; \u2018+ range\u2019, increasing state heterogeneity \nby varying the state-level carbon prices by a factor of 10; \u20183 zero\u2019 and \u20185 zero\u2019, assuming no effort in three or five lowest-supporting states (that is, zero \ncarbon price); \u2018Gov\u2019, using the public opinion results from a different survey question on \u2018Do you think your Governor should do more to address global \nwarming?\u2019; and \u2018AP\u2019 (America\u2019s Pledge), using existing commitments to climate action (instead of public opinion) to proxy for policy stringency. Technology \nconstraints include: \u2018Low Electricity Infrastructure\u2019, assuming reduced electricity infrastructure investment by constraining the national total electricity \nproduction to be 60% of the level in the main scenario that uses a uniform approach to achieve 80% decarbonization nationally; \u2018No CCS\u2019, assuming no \ninvestment in carbon capture and storage; and \u2018Low Biomass\u2019, assuming the national total biomass availability is half of the level in the main scenarios.\nNature Climate Change | VOL 11 | September 2021 | 738\u2013745 | www.nature.com/natureclimatechange\n743\n\nArticles\nNATUrE CliMATE ChAnGE\nthe benefits of heterogeneous action\u2014some agent-based modelling \nhas done this in other settings37\u2014and compare them with the costs \nof non-uniform action, such as using IAM-based cost assessments \nthat we demonstrate here.\nOur study shows that the magnitude of mitigation costs is \naffected by the availability of critical technologies (which in our case \ninclude low-carbon electricity and BECCS) and the ability to trade \nrelevant energy products across state borders. Whether these tech-\nnologies, especially untested ones such as BECCS, can be deployed \nat the speed and scale needed for deep decarbonization depends \non a variety of factors, such as economic cost, land use constraints \nand other sustainability considerations43,44. In addition, while our \nmodel allows for a certain degree of flexibility for inter-state trade \nof electricity and bioliquids, we often observe inflexibility of energy \ninvestments and trade in the real world, due to physical and regu-\nlatory constraints on electricity transmission, biomass resources, \nopposition to pipelines, etc.43\u201346. Future work should consider these \nreal-world constraints in modelling critical technologies at fine spa-\ntial scale and should disentangle more precisely whether it is the \ntechnology frontiers or the ability to trade across state borders that \ndrives the low cost of policy heterogeneity found in this study.\nA central policy puzzle in heterogeneous systems is how policy \nleaders\u2014who are willing to bear higher costs for policy action\u2014\ncan demonstrate effective mitigation strategies that make it easier \nfor other units to follow. For climate change, followership is par-\nticularly critical to effective policy, because the jurisdictions that are \nwilling to lead on climate policy account for just a tiny fraction of \nglobal emissions47. In our analysis, followership is especially needed \nin the medium-supporting states, given their large contribution to \nnational total emissions. The largest differences between leader and \nfollower units arise with so-called harder-to-abate sectors such as \nindustrial heating, buildings and transportation48. For policymak-\ners, this suggests that there is a very high premium on advances in \ndecarbonization of these sectors, because followership will be hard-\nest in these domains. If leaders can, through testing and deployment \nin their local markets, reduce abatement costs in these other sectors, \nthey could make followership much easier and increase total lever-\nage on emissions. For analysts, the role of these sectors in follower \njurisdictions suggests that we need to analyse the real potential for \nimproved decarbonization performance in these sectors in greater \ndepth. Where leadership can plausibly reduce costs, the model of \nheterogeneous policymaking can be particularly advantageous; \nwhere not, it could lead to policy cul-de-sacs, where followers never \nfollow and world emissions remain stuck at high levels.\nOnline content\nAny methods, additional references, Nature Research report-\ning summaries, source data, extended data, supplementary infor-\nmation, acknowledgements, peer review information; details of \nauthor contributions and competing interests; and statements of \ndata and code availability are available at https://doi.org/10.1038/\ns41558-021-01128-0.\nReceived: 9 December 2020; Accepted: 23 July 2021;  \nPublished online: 23 August 2021\nReferences\n\t1.\t Ostrom, E. Beyond markets and states: polycentric governance of complex \neconomic systems. Am. Econ. Rev. 100, 641\u2013672 (2010).\n\t2.\t De B\u00farca, G., Keohane, R. O. & Sabel, C. Global experimentalist governance. \nBr. J. Political Sci. 44, 477\u2013486 (2014).\n\t3.\t Rabe, B. G. Statehouse and Greenhouse (Brookings Institution Press, 2004).\n\t4.\t Sabel, C. & Victor, D. G. Governing global problems under uncertainty: \nmaking bottom-up climate policy work. Clim. Change 144, 15\u201327 (2017).\n\t5.\t McCollum, D. L. et al. Energy investment needs for fulfilling the Paris \nAgreement and achieving the Sustainable Development Goals. Nat. Energy 3, \n589\u2013599 (2018).\n\t6.\t Rogelj, J. et al. Paris Agreement climate proposals need a boost to keep \nwarming well below 2\u2009\u00b0C. Nature 534, 631 (2016).\n\t7.\t Rogelj, J. et al. Scenarios towards limiting global mean temperature increase \nbelow 1.5\u2009\u00b0C. Nat. Clim. Change 8, 325\u2013332 (2018).\n\t8.\t Fragkos, P. et al. Coupling national and global models to explore policy \nimpacts of NDCs. Energy Policy 118, 462\u2013473 (2018).\n\t9.\t Lowrey, A. Are states really more efficient than the federal government?  \nThe Atlantic (2 October 2017).\n\t10.\tAmerica is All In. https://www.americaisallin.com/\n\t11.\tU.S. State Climate Action Plans. Center for Climate and Energy Solutions \nhttps://www.c2es.org/document/climate-actionplans/ (2020).\n\t12.\tState Renewable Portfolio Standards and Goals. National Conference of State \nLegislatures https://www.ncsl.org/research/energy/\nrenewable-portfolio-standards.aspx (2020).\n\t13.\tFulfilling America\u2019s Pledge: How States, Cities, and Businesses are Leading the \nUnited States to a Low-Carbon Future (Bloomberg Philanthropies, Rocky \nMountain Institute and Center for Global Sustainability at the University of \nMaryland, 2018).\n\t14.\tLempert, R. et al. Pathways to 2050: Alternative Pathways for Decarbonizing \nthe U.S. Economy. C2ES Climate Innovation 2050 https://www.c2es.org/\ndocument/pathways-to-2050-scenarios-for-decarbonizing-the-u-s-economy \n(2019).\n\t15.\tHultman, N. E. et al. Fusing subnational with national climate action is \ncentral to decarbonization: the case of the United States. Nat. Commun. 11, \n5255 (2020).\n\t16.\tCalvin, K. et al. GCAM v5.1: representing the linkages between energy,  \nwater, land, climate, and economic systems. Geosci. Model Dev. 12,  \n677\u2013698 (2019).\n\t17.\tD\u2019Autume, A., Schubert, K. & Withagen, C. Should the carbon price be the \nsame in all countries? J. Public Econ. Theory 18, 709\u2013725 (2016).\n\t18.\tB\u00f6hringer, C. & Rutherford, T. F. Carbon taxes with exemptions in an open \neconomy: a general equilibrium analysis of the German tax initiative. J. \nEnviron. Econ. Manag. 32, 189\u2013203 (1997).\n\t19.\tBoeters, S. Optimally differentiated carbon prices for unilateral climate policy. \nEnergy Econ. 45, 304\u2013312 (2014).\n\t20.\tPage, B. I. & Shapiro, R. Y. Effects of public opinion on policy. Am. Political \nSci. Rev. 77, 175\u2013190 (1983).\n\t21.\tBurstein, P. The impact of public opinion on public policy: a review and an \nagenda. Political Res. Q. 56, 29\u201340 (2003).\n\t22.\tBromley-Trujillo, R. & Poe, J. The importance of salience: public opinion and \nstate policy action on climate change. J. Public Policy https://doi.org/10.1017/\nS0143814X18000375 (2018).\n\t23.\tRules of the Senate. U.S. Senate https://www.rules.senate.gov/\nrules-of-the-senate\n\t24.\tFawcett, A. A., Clarke, L. E. & Weyant, J. P. Introduction to EMF 24. Energy \nJ. 35, https://doi.org/10.5547/01956574.35.SI1.1 (2014).\n\t25.\tWilliams, J. H. et al. Pathways to deep decarbonization in the United States. \nThe U.S. Report of the Deep Decarbonization Pathways project of the \nSustainable Development Solutions Network and the Institute for Sustainable \nDevelopment and International Relations (2014).\n\t26.\tIyer, G. et al. Measuring progress from nationally determined contributions \nto mid-century strategies. Nat. Clim. Change 7, 871\u2013874 (2017).\n\t27.\tClarke, L. et al. International climate policy architectures: overview of the \nEMF 22 International Scenarios. Energy Econ. 31, S64\u2013S81 (2009).\n\t28.\tClarke, L. et al. in Climate Change 2014: Mitigation of Climate Change. IPCC \nWorking Group III Contribution to AR5 (Cambridge Univ. Press, 2014).\n\t29.\tBrandeis. New State Ice Co. v. Liebmann (1932).\n\t30.\tUnited Nations Convention on Climate Change. FCCC/INFORMAL/84 \nGE.05-62220 (E) 200705 (United Nations, 1992).\n\t31.\tWigley, T. M. L., Richels, R. & Edmonds, J. A. Economic and environmental \nchoices in the stabilization of atmospheric CO2 concentrations. Nature 379, \n240\u2013243 (1996).\n\t32.\tKeohane, R. O. & Victor, D. G. The regime complex for climate change. \nPerspect. Politics 9, 7\u201323 (2011).\n\t33.\tKeohane, R. O. & Victor, D. G. Cooperation and discord in global climate \npolicy. Nat. Clim. Change 6, 570 (2016).\n\t34.\tBiermann, F., Pattberg, P., van Asselt, H. & Zelli, F. The fragmentation of \nglobal governance architectures: a framework for analysis. Glob. Environ. \nPolitics 9, 14\u201340 (2009).\n\t35.\tOllivier-Mrejen, R., Michel, P. & Pham, M.-H. Chronicles of a Science \nDiplomacy Initiative on Climate Change. Science and Diplomacy (2018).\n\t36.\tHale, T. \u2018All hands on deck\u2019: the Paris Agreement and nonstate climate action. \nGlob. Environ. Politics 16, 12\u201322 (2016).\n\t37.\tHovi, J., Sprinz, D. F., S\u00e6len, H. & Underdal, A. The club approach: a gateway \nto effective climate co-operation? Br. J. Political Sci. 49, 1071\u20131096 (2019).\n\t38.\tBauer, N. et al. Quantification of an efficiency\u2013sovereignty trade-off in \nclimate policy. Nature 588, 261\u2013266 (2020).\n\t39.\tBosetti, V., Carraro, C., Sgobbi, A. & Tavoni, M. Delayed action and \nuncertain stabilisation targets. How much will the delay cost? Clim. Change \n96, 299\u2013312 (2009).\nNature Climate Change | VOL 11 | September 2021 | 738\u2013745 | www.nature.com/natureclimatechange\n744\n\nArticles\nNATUrE CliMATE ChAnGE\n\t40.\tJakob, M., Luderer, G., Steckel, J., Tavoni, M. & Monjon, S. Time to act now? \nAssessing the costs of delaying climate measures and benefits of early action. \nClim. Change 114, 79\u201399 (2012).\n\t41.\tSchaeffer, M. et al. Mid- and long-term climate projections for fragmented \nand delayed-action scenarios. Technol. Forecast. Soc. Change 90, 257\u2013268 \n(2015).\n\t42.\tIyer, G. C. et al. Improved representation of investment decisions in \nassessments of CO2 mitigation. Nat. Clim. Change 5, 436 (2015).\n\t43.\tFuss, S. et al. Negative emissions\u2014Part 2: costs, potentials and side effects. \nEnviron. Res. Lett. 13, 063002 (2018).\n\t44.\tClimate change and land: an IPCC special report on climate change, \ndesertification, land degradation, sustainable land management,  \nfood security, and greenhouse gas fluxes in terrestrial ecosystems  \n(IPCC, 2019).\n\t45.\tSanchez, D. L., Johnson, N., McCoy, S. T., Turner, P. A. & Mach, K. J. \nNear-term deployment of carbon capture and sequestration from  \nbiorefineries in the United States. Proc. Natl Acad. Sci. USA 115,  \n4875 (2018).\n\t46.\tCochran, J., Denholm, P., Speer, B. & Miller, M. Grid Integration and the \nCarrying Capacity of the U.S. Grid to Incorporate Variable Renewable Energy. \nhttps://www.osti.gov/servlets/purl/1215010 (2015).\n\t47.\tVictor, D. G. et al. Turning Paris into reality at the University of California. \nNat. Clim. Change 8, 183\u2013185 (2018).\n\t48.\tMission Possible: Reaching Net-Zero Carbon Emissions from \nHarder-to-Abate Sectors by Mid-Century. Energy Transitions Commission \nhttp://www.energy-transitions.org/sites/default/files/ETC_MissionPossible_\nFullReport.pdf (2018).\n\t49.\tHowe, P. D., Mildenberger, M., Marlon, J. R. & Leiserowitz, A. Geographic \nvariation in opinions on climate change at state and local scales in the USA. \nNat. Clim. Change 5, 596 (2015).\n\t50.\tU.S. Energy Information Administration. The State Energy Data System (SEDS).\nPublisher\u2019s note Springer Nature remains neutral with regard to jurisdictional claims in \npublished maps and institutional affiliations.\n\u00a9 The Author(s), under exclusive licence to Springer Nature Limited 2021\nNature Climate Change | VOL 11 | September 2021 | 738\u2013745 | www.nature.com/natureclimatechange\n745\n\nArticles\nNATUrE CliMATE ChAnGE\nMethods\nGCAM-USA model. The GCAM is an open-source global integrated assessment \nmodel (https://github.com/JGCRI/gcam-core/releases; for more information, see \nref. 16 and online documentation http://jgcri.github.io/gcam-doc/toc.html). GCAM \nrepresents key interactions across the economic, energy, land and climate systems \nin 32 geopolitical regions in the world. It is a market equilibrium model that solves \nfor the market prices and quantities of a large number of markets simultaneously. \nIt is dynamic recursive with myopic foresight (that is, the model solution in each \nmodel period depends on the conditions in that period or periods before it).\nIn this study, we use GCAM-USA v5.1, which is a version of GCAM \nwith state-level detail in the United States26. Like GCAM, GCAM-USA is an \nopen-source model. Detailed documentation for the GCAM-USA model is \navailable online (http://jgcri.github.io/gcam-doc/gcam-usa.html). Here we \nsummarize key model features that are relevant for this study. The model results for \nall main and supplementary scenarios are available from a public data repository51.\nGCAM-USA divides the energy and economic systems of the United States into \n50 states and Washington DC, with state-level representation of socioeconomics, \nenergy transformation (power generation and refining), carbon storage, renewable \nresources (wind and solar), electricity markets (with the representation of \nregional electricity grids) and consumer end-use energy demands (in buildings, \ntransportation and industrial sectors).\nEconomic growth, population changes and climate changes set the scale \nfor energy demands in buildings, transportation and industrial sectors. These \ndemands are supplied by a variety of fuels and technologies. GCAM-USA \nincludes technological detail in all sectors within the energy system. For example, \nthe model includes about 20 different power generation technologies, about 10 \ndifferent building service types and building technologies that vary across fuel \ninputs, and about 40 different transportation technologies that vary across modes, \nvehicle class, vehicle size and fuel input. The model also tracks details about \ncapital stock vintages in capital intensive sectors such as power and refining along \nwith simple algorithms to retire existing stock based on natural lifetimes and \neconomic conditions. Nested logit structures are used to share out new investments \nacross technologies to meet demands. The logit structures create competition \nbetween different fuels and technologies on the basis of relative costs52. It avoids \nwinner-take-all type of responses and helps capture heterogeneity in unmodelled \nfactors across various technology options.\nTo represent electricity trade, states are grouped into 15 electricity grid regions \nto reflect electricity market and planning areas (Supplementary Fig.\u00a05). Within \na grid region, we assume unconstrained trade and therefore common electricity \nprices across these states. Trade between grid regions is calibrated to historical levels \nto reflect existing economic conditions as well as implied physical transmission \ncapability. In future modelling periods, trade can change from calibrated levels as \nrelative regional electricity prices change. The model also includes flexibility in the \nlocation of energy-intensive industries such as refineries on the basis of the relative \ncosts of production across states based on a non-linear logit equation53. For oil \nrefining, our scenarios assume that expansion takes place in the states that currently \nengage in oil refining and that the relative competitiveness of those states remains \nconstant over time. For newer technologies such as biomass to liquids technologies, \nwe assume that the current distribution of production will remain largely static in the \ncoming years, which, in turn implies that biomass production will largely occur in \nthe same locations as today. Note that the model includes representation of negative \nemissions technologies (such as BECCS) in both electricity and refining sectors.\nIn addition to a detailed representation of the energy system within the United \nStates, the model includes global representations of the agriculture and land-use \nsystems at the basin scale. These systems are all hard-coupled in code. For example, \nbiomass production is modelled in the agriculture and land-use component of the \nmodel that creates a competition for land among various land uses such as biomass, \ncrop production, managed and unmanaged forests and livestock. Biomass produced \nin the agriculture and land-use component is demanded in the energy system. \nLikewise, fertilizers produced in the energy system are demanded by crops in the \nagriculture and land-use system. GCAM-USA maintains representations of 31 \ngeopolitical regions outside of the United States; hence, prices and changes occurring \nwithin the United States are consistent with international and global conditions.\nScenario design. To design mid-century scenarios, we set varying levels of \nnational mitigation effort, targeting the national total GHG emissions in 2050 \nto be 20%, 40%, 60%, 80% and 95% below 2005 levels, respectively. The 20\u201380% \ndecarbonization results are presented in the main text, and 95% decarbonization \nresults are presented in Supplementary Fig.\u00a07 as a sensitivity run to understand the \nrequired energy system changes in line with a 1.5\u00b0C global climate stabilization \ntarget and close to the Biden Administration\u2019s net-zero emissions target. We \nassume linear GHG mitigation pathways from 2015 to 2050 with 5-year interval \n(see trajectories for GHG targets in Supplementary Fig.\u00a01). Since GCAM-USA is \nembedded within the global GCAM model and allows for interactions between \nthe United States and the rest of the world through global markets, to avoid \ncross-country carbon leakage, we set decarbonization targets for other countries \nbased on ref. 53, which are consistent with the 2\u00b0C pathway.\nWe consider three subnational policy approaches to achieve the national \ntargets. Under the Uniform approach, the model solves a single MAC (and in turn \ncarbon price) nationally to meet the decarbonization target. The MAC is uniform \nacross states, which is determined by the marginal cost to mitigate the last unit of \nCO2 emissions nationally. Under the Hybrid and Heterogeneous approaches, we \nallow for heterogeneous MACs across states. We set the relative ratio of state-level \nMACs based on the present-day public support level for climate policy, then let the \nmodel solve the whole set of MACs for 51 states (Figs.\u00a01b and\u00a02 and Supplementary \nMethods). Note that the MACs capture the effects similar to a carbon price. A \nhigh MAC encourages the deployment of high-cost CO2 mitigation technologies \n(such as renewable electricity and BECCS) as well as a reduction in overall fossil \nenergy use. As long as the importing activities can reduce energy production and \nassociated emissions within the state boundary, our approach would not further \nrequire importing only low-carbon electricity or goods.\nUnder both the Hybrid and Heterogeneous approach, we assume a non-linear \nrelationship between the public support level and MAC (Figs.\u00a01b and Fig.\u00a02 and \nSupplementary Methods). Under the Heterogeneous approach, we explore the \nlargest heterogeneity in MACs across all the states. Under the Hybrid approach, \nwe further consider a modest level of nationwide carbon policy, and allow states \nto set their MAC only higher than this nationwide threshold. The threshold is set \nat the 40th percentile to represent the 60% voting threshold of passing legislation \nin the U.S. Senate. We also consider a variety of alternative formulations of policy \nheterogeneity, discussed in detail in\u00a0Supplementary Methods.\nCalculation of mitigation costs. To calculate the mitigation cost, we construct \nstate-level MAC curves using the CO2 emissions and MACs from 20% to 95% \ndecarbonization target runs. The model considers state-varying and time-varying \nMAC curves, driven by the variations in a wide range of factors such as existing \ninfrastructure, local availability of low-carbon resources and the evolving costs of \ntechnologies over time. We then integrate the area under the MAC curves in each \nstate to calculate state-level mitigation cost, and add up the cost to the national \nlevel. All monetary values in this paper are presented in 2015 US$.\nData availability\nThe datasets generated during and analysed in the current study are available from \na public repository (https://doi.org/10.5281/zenodo.5061357).\nCode availability\nThe GCAM and GCAM-USA model are available for download from https://\ngithub.com/JGCRI/gcam-core. Detailed model documentation is available online \nat http://jgcri.github.io/gcam-doc/gcam-usa.html.\nReferences\n\t51.\tPeng, W. et al. Datasets for \u2018The Surprisingly Inexpensive Cost of \nState-Driven Emission Control Strategies\u2019. Zenodo. https://doi.org/10.5281/\nzenodo.5061357 (2021).\n\t52.\tClarke, J. F. & Edmonds, J. A. Modelling energy technologies in a competitive \nmarket. Energy Econ. 15, 123\u2013129 (1993).\n\t53.\tFawcett, A. A. et al. Can Paris pledges avert severe climate change? Science \n350, 1168 (2015).\nAcknowledgements\nWe thank B. Keohane, D. Tingley, L. Stokes, J. Jenkins, K. Fisher-Vanden and participants at \nseminars at Penn State University (September 2019), Johns Hopkins University (September \n2019) and Princeton University (November 2019) on related themes. W.P. received a \nsummer research stipend from Penn State School of International Affairs. G.I., M.B. and \nJ.A.E. received support from the Global Technology Strategy Program. D.G.V. draws \nfunding, in part, from the Electric Power Research Institute, a nonprofit R&D organization \nfocused on the electric power sector. D.G.V. is also supported partly by donations to the \nScripps Institutional Oceanography for research on emergency responses to climate change.\nAuthor contributions\nW.P., G.I. and D.G.V conceived and designed the study. W.P., G.I. and M.B performed the \nmodel simulations with data input from J.M. W.P. analysed the data. W.P., G.I. and D.G.V \nwrote the manuscript with important input from all authors.\nCompeting interests\nD.G.V. is a consultant to the shareholder group Engine No. 1. The other authors declare \nno competing interests.\nAdditional information\nExtended data is available for this paper at https://doi.org/10.1038/s41558-021-01128-0.\nSupplementary information The online version contains supplementary material \navailable at https://doi.org/10.1038/s41558-021-01128-0.\nCorrespondence and requests for materials should be addressed to W.P.\nPeer review information Nature Climate Change thanks Aleh Cherp, Laurent Drouet and \nthe other, anonymous, reviewer(s) for their contribution to the peer review of this work.\nReprints and permissions information is available at www.nature.com/reprints.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNATUrE CliMATE ChAnGE\nExtended Data Fig. 1 | Reduction in energy-related CO2 emissions for four levels of national mitigation efforts under two subnational policy \napproaches. Here we show the reduction in CO2 emissions in 2050 relative to 2005. The first and second rows show the results under the uniform and \nheterogeneous approach, respectively. Different colors of the bars show the mitigation in different economic sectors. The white numbers represent the \npercent contribution of electricity sector to total CO2 mitigation. Low, Medium and High represents the low-, medium-, and high-supporting states, \nrespectively.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNATUrE CliMATE ChAnGE\nExtended Data Fig. 2 | Net electricity trade in 2050 for 15 grid regions. a, \u201880% Uniform\u2019: 80% national decarbonization with the uniform approach;  \nb, \u201880% Heterogeneous\u2019: 80% national decarbonization with the heterogeneous approach; c, Changes in \u201880% Heterogeneous\u2019 relative to \u201880% \nUniform\u2019. For a) and b), the background colors represent the net electricity export from a grid region (orange) or the net import into a grid region (blue). \nThe electricity transmission patterns remain largely the same under these two subnational policy approaches. For c), the background colors represent \nthe absolute differences in net electricity export (that is, generation minus demand) between these two approaches. The black numbers show the \npercent differences: the positive numbers indicate that a net exporting (importing) grid in \u201c80% Uniform\u201d further increases its export (import) in \u201c80% \nHeterogeneous\u201d, while the negative numbers indicate that a net exporting (importing) grid in \u201c80% Uniform\u201d reduces its export (import) in \u201c80% \nHeterogeneous\u201d. In other words, 13 out of the 15 grids increase their electricity trade with other grids under the Heterogeneous approach. The 15 electricity \ngrid regions are presented in Supplementary Fig.\u00a05.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNATUrE CliMATE ChAnGE\nExtended Data Fig. 3 | CO2 sequestration by bioenergy with carbon capture and storage (BECCS) technology in 2050. a, Total CO2 sequestration from \nBECCS; b, CO2 sequestration from BECCS for bioliquids production; c, CO2 sequestration from BECCS for electricity production. 20%-80% represent the \nlevels of national mitigation effort. \u2018U\u2019 and \u2018H\u2019 represent Uniform and Heterogeneous policy approaches, respectively. Low, Medium and High represent \nlow-, medium-, and high-supporting states.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNATUrE CliMATE ChAnGE\nExtended Data Fig. 4 | Variations in state-level carbon intensity and mitigation costs to achieve 80% decarbonization nationally. a, Carbon intensity \nin 2015; b, Reduction in carbon intensity in 2050 relative to 2015, to achieve 80% national decarbonization with uniform or heterogeneous approach; c, \nMitigation cost in 2050 as a fraction of 2050 GDP, to achieve 80% national decarbonization with uniform or heterogeneous approach. All economic values \nare presented in US$2015.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNATUrE CliMATE ChAnGE\nExtended Data Fig. 5 | Carbon mitigation costs for four levels of national mitigation efforts under alternative formations of policy heterogeneity. Here \nwe show the carbon mitigation cost in 2050 as a fraction of projected 2050 GDP, by three groups of states (low, medium and high-supporting states). \n\u2018Uni\u2019 stands for the uniform approach. Het, Het (Gov), Het (LN), Het (+range), Het (3 zero), Het (5 zero) and Het (AP) represent different heterogeneous \napproaches with detailed description in the Sensitivity Analysis section and Supplementary Methods. The inserted figure shows the contribution by low-, \nmedium-, and high-supporting states to national total CO2 emissions in 201650.\nNature Climate Change | www.nature.com/natureclimatechange\n\nArticles\nNATUrE CliMATE ChAnGE\nExtended Data Fig. 6 | Reduction in energy-related CO2 emissions when carbon capture and storage (CCS) technology is not available. Here we show \nthe reduction in energy-related CO2 emissions in 2050 relative to 2015 for: a, National total; b, Low-supporting states; c, Medium-supporting states;  \nd, High-supporting states. Different colors of the bars represent different economic sectors. The white numbers represent the percent contribution of \nthe electricity sector to total CO2 mitigation. 40% U \u2013 40% decarbonization with uniform approach; 40% H \u2013 40% decarbonization with heterogeneous \napproach; 80% U \u2013 80% decarbonization with uniform approach; 80% H - 80% decarbonization with heterogeneous approach.\nNature Climate Change | www.nature.com/natureclimatechange\n\n\n    Scientific Research Findings:", "answer": "As compared to an idealized nationally uniform policy, we find that varying the state-level policy stringency by a factor of three will increase the nationwide cost by only about 10%. Such results are robust under different national decarbonization targets, formulations for policy heterogeneity, and technology assumptions. The low cost hinges on two conditions. First is the availability of critical technologies (for example, low-carbon electricity and bioenergy with carbon capture and storage) and the ability to trade relevant energy products across state borders. Second is that there is at least some effort by every state. If a handful of states are not engaged at all, the leading states will need to adopt extremely expensive negative emissions technologies, which pushes up the nationwide cost. These two conditions could be difficult to meet in the real world given the technological, regulatory and political realities. Future work should consider potential constraints on technology adoption as well as trade and state politics.", "id": 60}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Cities | Volume 1 | October 2024 | 654\u2013664\n654\nnature cities\nArticle\nhttps://doi.org/10.1038/s44284-024-00116-7\nConnectivity in urbanscapes can cause \nunintended flood impacts from  \nstormwater systems\nVinh Ngoc Tran\u2009\n\u200a\u20091, Valeriy Y. Ivanov\u2009\n\u200a\u20091\u2009\n, Weichen Huang1, Kevin Murphy\u2009\n\u200a\u20091, \nFariborz Daneshvar\u2009\n\u200a\u20091, Jeff H. Bednar2, G. Aaron Alexander\u2009\n\u200a\u20093, Jongho Kim4 & \nDaniel B. Wright\u2009\n\u200a\u20093\nUrban flooding is intensifying worldwide, presenting growing challenges to \nurban communities. We posit that most of the flood management solutions \ncurrently employed are local in nature and fail to account for ways in which \nthe space\u2013time connectivity of floods is exacerbated by built infrastructure. \nWe examine the 2014 flood in Southeast Michigan to identify key factors \ncontributing to urban flooding and explore the implications of design \nchoices on inundation. Findings reveal that stormwater infrastructure \nthat neglects flood spatial connectivity can be ineffective in mitigating \nfloods, leading to inundation even in the absence of local rainfall. Different \nconfigurations of network connections\u2014including interfaces with natural \nchannels\u2014can significantly impact upstream surcharge, overflowing \nmanholes and inundation conditions. These results emphasize the need to \nconsider interconnectedness of flood processes in urban watershed systems \nto mitigate limitations inherent in the design of flood control and warning \nsystems, to enhance urban flood resilience.\nClimate change is increasing the likelihood of flooding1,2, with often \ndisastrous impacts, particularly in urban areas3,4. Flood-related global \neconomic losses reached a staggering $651 billion (US dollars) between \n2000 and 20195. The United States has been particularly impacted, with \nfloods causing 1,782 fatalities and damages exceeding $102 billion,  \naffecting 99% of US counties since 20006\u20138. The growing impacts of \nflood events across the United States and globally will continue to \nrise, with losses projected to soar by a factor of 20 by the end of the \ntwenty-first century9. These increasing impacts are driven not only \nby the increasing severity of extreme rainstorms, but also by rapid \nurbanization10, booming population densities and the dramatic expan-\nsion of highly connected transportation and infrastructure networks in \nflood-prone areas11,12. Rapid urbanization has substantially altered the \nnatural water cycle through the proliferation of impervious surfaces, \ninhibiting rainfall infiltration and resulting in elevated surface runoff13. \nThe disruption of natural drainage pathways has rendered urban areas \nincreasingly susceptible to inundation. Consequently, numerous major \ncities globally have witnessed a rise in the frequency and severity of \nflood events, as urbanization outpaces upgrades to flood mitigation \ninfrastructure (often referred to as stormwater systems)10.\nStormwater systems play an essential role in cities across the \nworld (Supplementary Fig. 1). Stemming from advances in science \nand technology over past decades, a new generation of flood mitiga-\ntion infrastructures, such as green-blue infrastructure (GBI) systems, \nare perceived as the future of urban flood management, designed to \nreplace traditional \u2018gray\u2019 approaches14. GBI refers to an integrated \napproach that combines \u2018green\u2019 infrastructure solutions, such as per-\nmeable pavements, rain gardens and green roofs, with \u2018blue\u2019 infra-\nstructure elements (or low-impact development), such as constructed \nwetlands, detention ponds and restored floodplains. Despite the \nReceived: 5 December 2023\nAccepted: 25 July 2024\nPublished online: 27 August 2024\n Check for updates\n1Department of Civil and Environmental Engineering, University of Michigan, Ann Arbor, MI, USA. 2Macomb County Public Works Commissioner  \nCandice S. Miller, Clinton Township, MI, USA. 3Department of Civil and Environmental Engineering, University of Wisconsin-Madison, Madison, WI, USA. \n4School of Civil and Environmental Engineering, University of Ulsan, Ulsan, South Korea. \n\u2009e-mail: ivanov@umich.edu\n\nNature Cities | Volume 1 | October 2024 | 654\u2013664\n655\nArticle\nhttps://doi.org/10.1038/s44284-024-00116-7\nreflect the ways in which drainage system interconnectedness  \ncan create vulnerabilities through the aggregate \u2018network effects\u2019 \nof localized engineering solutions. Additionally, modeling tools \nemployed in engineering design practices are outdated and lack the \ncapability to describe the coupled dynamics between flows over the \nland surface, through natural and artificial surface channels, and \nthrough sewer systems. Similar deficiencies in design practices can \nbe found worldwide28\u201333, and exist despite the availability of more \nmodern modeling tools that are capable of performing sophisticated \nsimulations. These practices have impacted scientific research focus-\ning on the optimization of drainage systems, as most studies have relied \nsolely on simplified one-dimensional (1D) models that simulate water \nconveyance within sewer systems, while neglecting interactions with \nsurface water (see Group 3 in Supplementary Table 1, which contains a \nrelevant literature review). Such model limitations may result in poorly \ninformed design of drainage systems and may compromise urban flood \nresilience and prevention capacities4,34,35.\nState-of-the-science urban flood modeling can accurately simulate \nthe motion of flood waves over the surface and within stormwater infra-\nstructure, while accounting for the complexities of surface conditions \n(vegetation, roads, bridges, buildings and so on)36\u201338. Augmented by the \ngrowing volumes of high-resolution geospatial data for urbanscapes, \nit is now possible to explore flooding behaviors in urban environments \nwith far fewer simplifying assumptions. As pointed out already, previous  \nstudies have primarily focused on model developments of specific \nphysical processes37 and have thus lacked comprehensive insights \nregarding the connectivity of floods in urban domains as shaped by \nnatural drainage elements and gray infrastructure. Further research  \nis required to advance such understanding. By investigating flood \npropagation through a complex urban environment with multiple \nprocess interactions, our study provides novel insights into emer-\ngent flood behaviors shaped by both natural and man-made drainage \nsystems. The knowledge gained and capabilities demonstrated in this \nresearch could help to provide sustainable and resilient approaches  \nto addressing the intensification of global urban flood hazards.\nSpecifically, we investigate the following. First, what is the  \nnature of the complexity of flood connectivity in urban landscapes? \nSecond, can sewer systems exacerbate urban floods? We hypothesize \nthat (1) in certain conditions, connectivity via stormwater infrastruc-\nture can lead to flooding in a particular location, even in the absence \napparent merits of GBI (for example, eco-sustainability and climate-\nchange resilience)15,16, they are costly, require a longer implementation \nperiod, and necessitate continuous maintenance to ensure efficient \ndrainage and retention capabilities, with a continued reliance on gray \ninfrastructure17,18. Therefore, gray infrastructure continues to be the \nprimary approach for mitigating urban flooding17,19,20.\nThe repeated damages caused by urban flooding have raised ques-\ntions regarding the efficacy of costly drainage systems, whether gray \nor GBI. For example, the United States spends $7\u201310 billion dollars \nannually for the construction and renovation of stormwater systems, \nwith larger expenditures needed over the long term21. With the gene\u00ad\nrally poor state of aging infrastructure in the United States22, even \nmoderate rainfall events (a return period of ~10 years)23,24 can exceed \nthe capacity of drainage systems8. Although current inadequacies stem \npartly from outdated stormwater design standards25,26, we posit that it is \nalso due to an insufficient understanding of \u2018flood connectivity\u2019 within \nurban landscapes, defined as a complex interplay of flooding mecha-\nnisms (Fig. 1) and drainage pathways within urban settings, arising  \nfrom the spatiotemporal distribution of hydraulic head (the term used \nto describe the summed potential, kinetic and pressure energies of \nstormwater flows). This flood connectivity encompasses the interac-\ntions and interdependencies among runoff, surface and subterranean \nsewer and drainage systems, rivers and other natural and built water \nbodies, and infrastructure components (for example, buildings and \nroadways) that can, by design or in an unintended way, convey storm-\nwater during floods. Although numerous studies have been conducted \nin recent decades focusing on understanding urban flooding (Supple-\nmentary Table 1 provides a detailed literature review), most previous \nstudies have focused on a few specific, isolated physical phenomena, \nsuch as assessing the impact of infrastructure and rainfall on flooding, \nor case-study-specific aspects of water exchange between drainage \nsystems and surface flows. Therefore, a comprehensive exploration \nof flood connectivity in complex urban areas and a holistic synthesis \nof process interactions remain to be explored.\nInsufficient appreciation of flood connectivity can be demon-\nstrated by outdated and oversimplified stormwater system design \nguidelines22,27. For example, the American Society of Civil Engineers \n(ASCE) stormwater design guidelines were last revised in 2006 \n(GUIDE2006 hereafter)28. While offering an extensive set of drainage \nnetwork design criteria, the GUIDE2006 guidelines do not adequately \nPipe\nRiver\nUrban watershed of interest\nNearby watershed\nPast climate median water level\nFuture/current median water level\nVadose zone\nFlooding\nManhole\nFlow direction\nOutfall\nGroundwater\n[FC3]\n[FC1]\n[FC2]\nFig. 1 | Illustration of key urban flooding concepts. A cross-section of an \nurban watershed, depicting residential areas (buildings), green spaces (trees), \nan open channel (river cross-section) and an underground stormwater system \n(manholes, pipes and outfalls). Outfall discharging into the river connects the \nurban stormwater network with a larger drainage system. The arrows indicate \npotential flow directions of water. Black arrows indicate surface flow directions, \nblue arrows show flows from the ground surface into the sewer system and out \nof the outfall locations (for example, open channels), and red arrows depict \nbackwater occurrences when open channel flows reverse into the drainage \nsystem through outfalls, potentially causing surcharging at manholes. The blue \nwavy line represents the normal water level (historical conditions), and the red \nwavy line the current or future river water level due to the influence of climate \nchange and urbanization. FC1 represents flooding due to high water levels \nin the river, causing overflow and inundation of the surrounding floodplain \n(river-induced or \u2018fluvial\u2019 flooding). FC2 represents flooding caused by intense \nrainfall and insufficient natural or engineered drainage capacity of the area \n(rainfall-induced or \u2018pluvial\u2019 flooding). FC3 represents flooding caused by the \nhydraulic connection of stormwater infrastructure with stream channels and \nother drainage basins. FC3 flooding can be caused by water accumulation due to \neither fluvial or pluvial flooding or backup flow from the river through drainage \npipes, leading to surcharge at the manholes (infrastructure-induced flooding). \nFC, flooding concept.\n\nNature Cities | Volume 1 | October 2024 | 654\u2013664\n656\nArticle\nhttps://doi.org/10.1038/s44284-024-00116-7\nof \u2018local\u2019 precipitation, and (2) improper positioning of stormwater \ndrainage outlets and curb inlets (called manholes hereafter) can  \nexacerbate flooding.\nResults\nFlood of 2014 in Southeast Michigan\nWe present a case study that demonstrates how engineering design \ninfrastructure interacts with runoff drainage processes at different \nspatial scales of origin, including local scales, typical of engineering \ndesign, and the larger scales that connect upstream and downstream \ndrainage conditions. The study was carried out in relation to a highly \nurbanized watershed in Southeast Michigan that has multiple con-\nventional elements of gray infrastructure designed to deliver storm \nrunoff to nearby channels. We designed the study by focusing on a \nsingle extreme storm on 11 August 2014 (referred to as STORM2014 \nhereafter) that struck Southeast Michigan and caused catastrophic \nflooding. The storm was termed a 1-in-100-year event, with rainfall \ntotaling between four and six inches in four hours39. The economic \ncosts of the storm were substantial, with a total of over $1.1 billion dam-\nages to over 118,000 homeowners and businesses40, translating to the \nlargest flood-related disaster that the Federal Emergency Management \nAgency (FEMA) identified during 2014. The area was declared a federal \ndisaster zone in that September40. The study focuses on the Warren \narea, which is situated within the confines of the inner-ring Detroit \nsuburb of Madison Heights and has an area of ~8.8\u2009km2 (Fig. 2a,b). This \nregion is considered to be a part of the Red Run watershed, dissected \nby a network of diverse drainage elements such as open channels (for \nexample, Bear Creek; Fig. 2b and Supplementary Fig. 2), culverts, under-\nground stormwater conveyance systems, and their outlets (outfalls; \nFig. 2b,c). In contrast to this drainage complexity, the National Flood \nHazard Layer from FEMA (Fig. 2b) shows 1% or 0.2% annual probability \nfloodplains mainly confined to the vicinity of Bear Creek41. Nonetheless, \nthroughout the STORM2014 event, the majority of the Warren area, \ncomprising residential neighborhoods and major roadways, experi-\nenced profound inundation, as depicted in the simulation results in \nFig. 2d. Specifically, 58% of the domain area was affected by severe \nflooding, with water levels exceeding 0.5\u2009m reported by eyewitnesses \nin the area of the General Motors production facility (Supplementary \nFig. 2b). Because the area is typical of urban watersheds connected to \na larger drainage system, this domain was selected for an analysis of \n42\u00b0 30' N\n42\u00b0 35' N\n83\u00b0 10' W\n83\u00b0 05' W\n83\u00b0 W\na\nEsri, HERE, Garmin, NGA, USGS, NPS\n 2 mi \n 5 km \nMI\nb\nc\nOutfall 1\nOutfall 2\nRed Run watershed\nWarren area\nBuilding footprint\n1% Annual chance flood\n0.2% Annual chance flood\nRiver/open channel\nCulvert\nPipe\nInlet\nOutlet\nManhole\nOutfall 1\nOutfall 2\nd    HR-Integrated\n0\n0.4\n0.8\n1.2\n1.6\n2.0\ne    NR-Integrated\n0\n0.4\n0.8\n1.2\n1.6\n2.0\nf    HR-Controlled\n0\n0.4\n0.8\n1.2\n1.6\n2.0\ng    \u2206h (d\u2013f)\n\u20130.2\n\u20130.1\n0\n0.1\n0.2\nFig. 2 | Flood maps of Warren, Michigan for the STORM2014 (11 August 2014) \nevent. a,b, The Warren area identified by the dashed line in b is situated in the \nRed Run watershed (a) in southeastern Michigan. The orange and green dots in \na indicate the inlet and outlet locations of the watershed, respectively, with the \ninlet being the discharge point of the George W. Kuhn Retention Treatment Basin. \nThe Warren area shown in b includes buildings (gray polygons), open channels \n(for example, Bear Creek), culverts (dark blue lines), underground pipes \n(green lines), manholes (brown dots) and two outfalls (green and red squares). \nc, Photographs of the outfalls. The blue and orange areas in b represent the \nregions at risk of flooding with a 1% and 0.2% annual chance flood, respectively, \nprovided by FEMA. d, Simulation results for the maximum inundation depth \nof the flood event using STORM2014, where the land-cover conditions and \nculvert systems closely reproduce actual drainage conditions, allowing for both \ndrainage and backwater phenomena at the outfalls (referred to as the \u2018integrated\u2019 \noutfall condition). The rainfall is uniformly distributed across the entire Red \nRun watershed in this scenario (referred to as homogeneous rain, HR). e, The \nmaximum inundation depth results for a simulation configuration that mimics \nthat of d, but with rain occurring only outside the Warren area (referred to as \nno-rain, NR). f, The maximum inundation depth results using HR, but assuming \nthat all water collected by storm sewers discharges into the open channel; that \nis, the backwater effect in drainage pipes is not enabled (referred to as the HR \n\u2018controlled\u2019 outfall). g, The difference of the flood depths in d and f. The locations \nof surcharging manholes in d\u2013f are shown with brown dots. The values on the \ncolor bars in d\u2013g are expressed in meters.\n\nNature Cities | Volume 1 | October 2024 | 654\u2013664\n657\nArticle\nhttps://doi.org/10.1038/s44284-024-00116-7\nthe interactions of urban flooding mechanisms. To replicate historical \nflood events, we gathered an extensive array of data of high accuracy \nand spatial resolution that influence the flood response in this area, \nincluding topography, land use/land cover, road networks, buildings, \nstormwater systems and open channels. Notably, we also conducted \nfield surveys to validate and refine the data, particularly regarding the \nlocations of drainage outfalls, where strong interconnections exist \nbetween the drainage system, open channels and ground surfaces. \nFurther details on the specific data used and the model configurations \nare provided in the Methods.\nNo local rain, but flooding\nThe design of the stormwater system in GUIDE2006 is based on rainfall \nscenarios that represent different levels of precipitation that can cause \nflooding if the stormwater infrastructure is unable to accommodate the \ninflux of runoff quickly enough. In other words, GUIDE2006 assumes \nthat rainfall occurrence is a prerequisite for flooding. However, we show \nthat even in the absence of rainfall in local runoff source areas, runoff \nfrom heavy rain outside their boundaries can still cause inundation \ndue to the spatial hydraulic connectivity facilitated by the stormwater \ninfrastructure. The results in Fig. 2e highlight that a substantial portion \nof the studied area is susceptible to flooding: water exceeds 0.1\u2009m in \ndepth over ~75% of the area, with 27% having a depth of over 0.5\u2009m. Such \nan outcome indicates how the spatial connectivity of flood processes \ncan alter local perception of flood vulnerability conditions. Specifically, \nscenarios with intense precipitation only in neighboring areas must \nbe considered when assessing flood risks and planning mitigation for \na region of interest, even in the absence of rainfall occurring locally \nwithin that area.\nOne source of flooding in this \u2018no local rain\u2019 situation can be \nattributed to the increased water levels in open channels causing the \nwell-known phenomenon of fluvial flooding (Fig. 1, FC1). Under this \nmechanism, the water levels exceed the confines of the riverbanks, \nand streamflow spills into floodplain areas. Additionally, simulated \nresults show that runoff from the surrounding land areas (where rain-\nfall occurs) can flow along roadways into the Warren area, leading to a \nlocalized pluvial flooding (denoted by FC2 in Fig. 1). A third likely cause \nidentified is related to backwater surcharge through the stormwater \npipe network and the catch basins in source areas. Specifically, extreme \nwater levels within open channels induce backwater flow at stormwater \noutfalls, causing a substantial surge of reverse flow into topographically \nupstream source areas of the drainage system, leading to surcharge at \nmanholes and flooding in areas that did not experience rainfall. This \nflooding mechanism reflects concept FC3\u2014infrastructure-induced \nflooding that can impact areas distant from open channels.\nIndeed, the results of explicit modeling of flood connectivity pre-\nsent evidence that stormwater networks design with a \u2018local mindset\u2019 \nmay fail severely. Not only can they become inadequate to accomplish \ntheir designated task to drain stormwater, but they can even reverse \nstormwater systems\u2019 work and aggravate flooding due to sewer sur-\ncharge in residential areas that have not experienced rainfall (Fig. 2e).  \nThe spatial connectivity of flooded urbanscape may thus lead to the \nhydraulic head distribution expansion of inundation into regions  \nthat are distant from a river or do not experience extreme rainfall.\nNo backwater flow into outfalls, but heavier local flooding\nThe location, invert elevation and size of outfalls, as well as their associ-\nated sewer pipes, play a major role in determining the drainage capacity \nof a stormwater collection system. As stipulated in GUIDE2006, the \nentire volume of generated runoff within the system is assumed to \nflow out freely from outfalls. Although this permits an assessment of \nthe maximum drainage capacity, such an assumption may underesti-\nmate the influence of flood connectivity at outfalls, for example, the \ndistribution of hydraulic head in open channels or storage tanks. If the \nhydraulic head at the outfall section generated by the passing flood \nwave in the channel exceeds the hydraulic head in the stormwater sys-\ntem, then backwater flow commences and channel water will flow into \nthe sewer system, impeding stormwater drainage and possibly even \ncausing surcharge at upstream locations. The lack of consideration \nof flood connectivity in complex urbanscapes may lead to incorrect \nassessments of system drainage capacity and therefore an inaccurate \nrepresentation of possible inundation levels at sites requiring flood \nprotection.\nWe examined how accounting for flood connectivity or lack \nthereof may impact flood levels. Specifically, the former case (referred \nto as \u2018Integrated\u2019 outfalls) accounted for the hydraulic coupling of water \nlevels in open channels at the outfalls, mimicking the real-world flood \ncondition to the best of our abilities. In the latter case (\u2018Controlled\u2019), \nall stormwater at the outfalls is assumed to discharge freely into the \nreceiving channel of Bear Creek; that is, the outflow is controlled to \nhave the rate of a pipe with a free overall downstream condition, not \nimpeded by the presence of the receiving water body. Comprehensive \nhydrodynamic modeling of these two distinct scenarios of outfall \nfunctioning reveals peculiar results. Notably, with the outfalls hydrauli-\ncally disconnected (the \u2018Controlled\u2019 case) from the rest of the flooded \nwatershed, the resultant flooding is more extreme than that of the \nintegrated outfall case, which facilitated backwater effects (Fig. 2d,f). \nSpecifically, in areas near the outfalls at the Bear Creek channel, inun-\ndation levels for the \u2018Controlled\u2019 case exceeded the \u2018Integrated\u2019 case \nby over 0.1\u2009m (Figs. 2g and 3a\u2013c). This result can be explained by the \nconsiderable drainage from the \u2018Controlled\u2019 sewer system discharging \ninto the open channel: water levels at the outfall sites thus tend to be \nhigher and cause more severe flooding in nearby areas. In contrast, \nin the more realistic \u2018Integrated\u2019 case, if the water level in the channel \nexceeds the hydraulic head of the stormwater system at the outfall \nlocation, backwater flow will occur and water from the open channel \nwill flow into the sewer system. This results in lower water levels in near \noutfall areas experiencing backwater, leading to reduced flooding in \nthe \u2018Integrated\u2019 case. Conversely, with large discharges entering the \nsewer system, upstream areas (such as in the southern region) can \nbe impacted when the sewer\u2019s rainwater intake capacity is reduced, \npotentially increasing flooding due to slower drainage.\nNeglecting the connectivity between a sewer system and adjacent \nflooded areas and consequent backwater effects can overestimate \nflood levels. This can potentially result in poor design of the number, \npositioning and dimensions of manholes, pipes and outfalls. Larger \nsewers and outfall dimensions can facilitate rapid drainage, yet they \ncan also allow rapid backflow. As shown in Fig. 3d,e, a large outfall (with \na diameter of 3.8\u2009m) can discharge high flows (up to 40.9\u2009m3\u2009s\u22121) in the \n\u2018Controlled\u2019 case, whereas the backflow into the sewer can be up to \n63.5\u2009m3\u2009s\u22121 in the \u2018Integrated\u2019 case. The smaller outfall (with the diameter \nof 0.3\u2009m) had lower discharge and backflow (up to 9.7 and 16.7\u2009m3\u2009s\u22121 for \nthe controlled and integrated cases, respectively).\nIt is still far from being standard engineering practice to realisti-\ncally simulate flooding with coupled overland- and pipe-flow models, \nsuch as the one used in this study (results shown in Fig. 2). Instead, \nengineers mainly rely on simplified stormwater models (for example, \n1D channel flow models coupled with conceptual/lumped rainfall-\nrunoff modules28,30\u201333,42,43. In such formulations, the total discharge \nfrom outfalls is assumed to entirely \u2018disappear\u2019 (or become accommo-\ndated by a stormwater storage such as pond or tank) rather than flow \ninto open channels. Water levels in open channels (at outfall locations \nor elsewhere) will thus remain unaffected and no backwater effect \ncan occur. The assumption that stormwater discharge vanishes after \ndischarging through outfall can clearly have a strong impact on flood \nlevel assessments, and can misjudge drainage system performance. \nSpecifically, inundation levels are broadly underestimated across the \nstudy area (especially near open channels) by up to 20\u2009cm, as compared \nto the more real-world \u2018HR-Integrated\u2019 case (Supplementary Fig. 3a,b). \nThese estimates indicate that even assumptions for outfalls can lead to \n\nNature Cities | Volume 1 | October 2024 | 654\u2013664\n658\nArticle\nhttps://doi.org/10.1038/s44284-024-00116-7\nsignificantly different assessments of stormwater system performance. \nThis highlights that the lack of appreciation of flood connectivity may \nresult in poorly informed stormwater system designs.\nMore overflowed manholes, but lower surcharge amounts\nDifferent outfall configurations should produce distinct differences \nin the surcharge in catch basins. In principle, when the sewer system \nsimultaneously receives a substantial influx of water from both the \nground surface through manholes in source areas and from the open \nchannel through outfalls, this can induce higher surcharge and a larger \nnumber of surcharging manholes in low-lying areas (that is, where low \nhydraulic head is expected). The case study results only partially align \nwith this conjecture, showing that accounting for backwater effects (the \n\u2018Integrated\u2019 case) results in surcharging in a slightly larger number of \nmanholes (27 versus 25), as depicted in Fig. 2d,f. Figure 3f,g indicates that \nthe surcharge rates in manholes near the outfalls are comparatively lower \nas compared to the \u2018Controlled\u2019 case (for example, manholes 1 and 2). The \ndifferences in the surcharge rates were negligible for manholes further \nfrom the outfalls or at higher elevations (manholes 3, 4 and 5; Fig. 3h\u2013j).\nThe higher surcharge rates at manholes (for example, manhole 1,  \nFig. 3f near outfall 2) in the \u2018Controlled\u2019 case are attributed to the  \nhigher water volumes \u2018injected\u2019 into the flooded area from a nearby \ndraining outfall. Such an \u2018injection\u2019 leads to larger volumes of water \ninflow into the adjacent manholes, for example, manhole 1 and nearby \nmanholes (Fig. 2f) and, if paired with a limited drainage capacity of the \noutfall, (for example, outfall 2, which has a small 0.8-m diameter), can \nexacerbate the surcharge (Fig. 3f,g). These results demonstrate how \noutfall configuration/drainage capacity can impact an assessment of \nlocal flooding conditions. This underscores the complexity and vital \nimportance of flood connectivity for understanding the performance \nof stormwater infrastructure. Ignoring flood connectivity between land \nsurfaces, infrastructure and open channels in urban areas can lead to \nan incomplete understanding of how water moves through a complex \nurban flooding environment. Failing to account for this interconnect-\nedness misinforms engineering design and flood mitigation strategies.\nDiscussion\nExisting evaluations of urban flooding have so far neglected to account \nfor the interactions of complex mechanisms associated with flooding \ncaused by human-engineered stormwater networks22,23,37,38,42,44,45. Our \nresults demonstrate that the interaction of flooding types\u2014facilitated \nby the connectivity created by human-engineered infrastructure\u2014is an \nimportant aspect of urban flood dynamics. Ignoring such a connectiv-\nity may lead to intensified urban floods, especially in regions where \nriver-induced floods meet stormwater infrastructure (Fig. 2d,f), even \nwhen rainfall does not occur over the area served by the infrastruc-\nture. To expand the applicability of the study findings to a broader \nrange of flooding cases, we conducted simulations for the 14 largest \nevents (including STORM2014; Supplementary Fig. 4) over the past \n15 years in the Warren study area (rainfall return periods \u22651\u2009year) with \noutfalls that have hydraulic connection with the receiving channel, \nthereby mimicking real-world conditions. The simulation maximum \nflooded areas are shown in Fig. 4a. Even for the annual flood (that is, a \nreturn period of 1\u2009year), the total area flooded to a depth of over 0.5\u2009m \ncomprises up to 7% of the total area of the Warren area (excluding the \narea of open channels). With the increase in rainfall intensity (shown in  \na\nOutfall 1\nOutfall 2\nManhole 1\nManhole 2\nManhole 3\nManhole 4\nManhole 5\nb    Outfall 1\n5\n10\n15\n20\n180\n182\n184\n186\n188\nElevation (m)\nc    Outfall 2\n5\n10\n15\n20\n180\n182\n184\n186\n188\nOutfall elevation\n0.2% annual chance flood\n1% annual chance flood\n2% annual chance flood\n10% annual chance flood\nHR-Integrated\nHR-Controlled\nd    Outfall 1\n5\n10\n15\n20\n\u2013100\n\u201350\n0\n50\nDischarge (m3 s\u22121)\ne    Outfall 2\n5\n10\n15\n20\n\u2013100\n\u201350\n0\n50\nf    Manhole 1\n5\n10\n15\n20\n\u20135\n0\n5\n10\n15\nDischarge (m3 s\u22121)\nDischarge (m3 s\u22121)\ng    Manhole 2\n5\n10\n15\n20\n\u20135\n0\n5\n10\n15\nh    Manhole 3\n5\n10\n15\n20\n\u20135\n0\n5\n10\n15\ni    Manhole 4\n5\n10\n15\n20\n\u20135\n0\n5\n10\n15\nj    Manhole 5\n5\n10\n15\n20\n\u20135\n0\n5\n10\n15\nHour\nHour\nFig. 3 | Simulated water levels and flow rates at the outfalls and surcharged \nmanholes for STORM2014. a, Map depicting the locations of outfalls and select \nmanholes for the results shown in b\u2013j. b,c, Simulated water surface elevations  \nat two outfall locations (see legend). The gray region represents the invert  \n(pipe bottom-to-top) elevation range of the outfalls. Colored lines represent  \nthe floodwater levels at probabilities ranging from 10% to 0.2% annual chance \nflood as provided by FEMA. d,e, Outflow (positive values) or inflow (negative \nvalues) rates at the outfalls for the HR-Integrated and HR-Controlled scenarios. \nf\u2013j, Overflow (positive) and drainage (negative) rates for five manholes, indicated \nby circles in a.\n\nNature Cities | Volume 1 | October 2024 | 654\u2013664\n659\nArticle\nhttps://doi.org/10.1038/s44284-024-00116-7\nFig. 4b for STORM2014), the area of severe flooding and the flood level \nalso increase (Fig. 4c). It should be noted that the drainage system  \nin this area is designed to handle runoff resulting from a 10-year  \nstorm, with the assumption that all rainfall will be converted to  \nrunoff and flow into the drainage system46. However, in reality, not all \nrunoff efficiently concentrates in the drains due to the retention of \nsome water in low-lying areas, causing localized flooding.\nWe also carried out simulations for all events to evaluate the \nimpact of the stormwater system on flooding under varying rainfall \nconditions and outfall configurations, in the same fashion as was  \ndone in the simulation configurations for STORM2014 (Supplemen-\ntary Figs. 5\u201317). Generally, the rainfall magnitude certainly influences  \nthe performance of the stormwater system, with fewer instances \nof manhole overflow occurring under lower rainfall intensities. We \nstill observe the same flooding phenomenon implied by the flood \nconnectivity. For instance, the number of overflowing manholes in  \nthe HR-Integrated case is higher than in the HR-Controlled case. The \ncase of rainfall in the watershed outside of the Warren area can still \ncause flooding and overflow at several manholes in that area. Overall,  \nsimulation results for 14 rainfall events of different magnitudes  \ncorroborate the persistence of the analyzed flood dynamics in the \ncase study urban area caused by flood spatiotemporal connectivity.\nAnother observable outcome is the rapid onset of flooding after \nrainfall. As depicted in Fig. 4c, impactful flooding with inundation \ndepths exceeding 0.5\u2009m occurs within a narrow time window following \nthe most intense precipitation (initiating at hour 15:00), within 0.5\u20131\u2009h. \nThis signifies that to mitigate flooding, the drainage system must not \nonly store substantial rainwater volumes but also drain rapidly to \nminimize local water accumulation. However, the influx of water from \nthe surrounding areas or backwater effects at outfalls can exceed pipe \nand manhole conveyance capacities and thus cause drainage \u2018bottle-\nnecks\u2019. Failure to adequately consider flood connectivity can lead to the \ndesign of infrastructure (for example, massive underground storages) \nthat may fail to counter flooding (see results in Supplementary Text 1). \nOptimization-based approaches are probably required to alleviate the \nissue of such drainage \u2018bottlenecks\u2019.\nThe recognition of the potential hazard imposed by flooding due \nto connectivity, even in the absence of rainfall in the design area, is an \nimportant but overlooked reality of urban stormwater systems. It is \nnoted that most design guidelines worldwide28,30\u201333 rely on scenarios \nwith only \u2018direct rainfall\u2019 occurring in the area of interest (often uni-\nform), without considering influx/outflux from/to nearby intercon-\nnected areas. With flooding waters contributed by sources other \nthan direct rainfall, assessments of drainage capacity can be highly \nuncertain. Similar to above, stormwater system design may require \noptimization of its configuration.\nIn the United States, FEMA maps require further updates in \nflood-prone areas, as already highlighted in previous research47\u201349. \nThe current perception is that FEMA\u2019s hazard maps, which have not \nbeen updated for over a decade, are outdated due to climate change \nand land-use alterations41. Critically, we show here that FEMA maps \nare further outdated because the methods (modeling and flood con-\ncepts) used to generate them do not consider the complexity of flood \nconnectivity, especially in urbanscapes. Hence, flood risk assessments \ncan become particularly unreliable for urban zones where the different  \nflood processes interact (Fig. 1), compromising their utility. For  \nthe specific case study considered here, the simulated water levels  \nduring STORM2014 at the two discussed outfall locations were higher \nthan what FEMA maps show for the 0.2% annual flood probability \n(a 500-year return period). This was also confirmed by eyewitness  \naccounts. This demonstrates the necessity for more accurate  \nmodels to mimic the spatial interactions of flood mechanisms across \nrivers, surfaces, drainage systems and other infrastructure types.\nThe execution of this study is not without limitations. First, the \nhydrodynamic model employed, while advanced and sophisticated, is \nsubject to inherent uncertainties arising from physical process simpli-\nfications, parameterizations and input data quality. Second, the use of \ncrowdsourced data, such as images from the internet, to estimate flood \ndepths at a few locations introduced subjectivity and uncertainties into \nthe model evaluation process. This arises from the potential misalign-\nment between the timing and locations of the crowdsourced data with \nthe simulation configuration. Third, the study focuses on a single event, \nand therefore carries the signature of the specifics of this case study. \nThe findings have been generalized to highlight study broad-level impli-\ncations, but other urban areas with differing topographic, hydrological \nand infrastructure characteristics will require similarly detailed analy-\nsis to understand the impacts of flood space\u2013time connectivity. These \nlimitations underscore the necessity for further research with diverse \ncase studies and the requirement for observations of inundation depth \ndata and drainage system operational data. Notwithstanding these \na\n0\n2\n4\n6\n8\nFlooded area (km2)\n0\n0.5\n1.0\n1.5\n2.0\nInundation depth (m)\nP = 1.0-year\nP = 1.8-year\nP = 1.9-year\nP = 3.2-year\nP = 4.1-year\nP = 4.1-year\nP = 6.1-year\nP = 6.6-year\nP = 7.3-year\nP = 7.8-year\nP = 8-year\nP = 21.3-year\nP = 31.9-year\nP = 110.3-year\nc\n11 Aug, 6:00\n11 Aug, 12:00\n11 Aug, 18:00\n12 Aug, 00:00\n2014   \n0\n2\n4\n6\n8\nFlooded area (km2)\n[0\u20130.1] m\n(0.1\u20130.5] m\n(0.5\u20131] m\n(1\u20131.5] m\n(1.5\u20132] m\n(2\u20135] m\nb\n0\n10\n20\nRainfall\n(mm per 15 min)\nFig. 4 | Relationship between flood inundation and flooded area.  \na, Relationship between the flooded Warren area and the maximum flood level \ncalculated for the 14 largest rainfall events for the Warren area between 2009 and \n2021 (Supplementary Fig. 4). The legend indicates the rainfall frequency (return \nperiod P) for each event, which is calculated based on the maximum rainfall \nwithin 3\u2009h and uses the intensity\u2013duration\u2013frequency (IDF) curve provided by \nNOAA. b,c, The rainfall time series (b) and flooded area for six flood levels (c) \nduring the STORM2014 event (the return period is 110.3\u2009years). In the legend of \nc, parentheses indicate exclusion of the endpoint and square brackets indicate \ninclusion of the endpoint.\n\nNature Cities | Volume 1 | October 2024 | 654\u2013664\n660\nArticle\nhttps://doi.org/10.1038/s44284-024-00116-7\nlimitations, this Article presents a novel contribution by accentuat-\ning the importance of considering flood connectivity in urban flood \nmitigation strategies and the potential shortcomings of traditional \nstormwater management approaches that neglect the interconnected-\nness of flood processes.\nIn summary, urban flooding is a result of complex interactions \nbetween precipitation, terrain and land use, surface hydrology and \nhuman infrastructure systems. Disentangling the complex interac-\ntions that flood connectivity produces in urbanscapes is imperative for \nimproved urban flood resilience and mitigation. By explicitly analyzing \nsuch interactions for a case study with documented flooding issues, \nthis work emphasizes that it is vital to distinguish the distinct mecha-\nnisms of fluvial, pluvial and infrastructure-induced flooding, so that \nbetter engineering solutions can be developed. Holistic perspectives \nencompassing the complete connectivity between all facets of the \nurban flooding system, rather than compartmentalized viewpoints, \nneed to be prioritized as cities strive to enhance their flood resilience \nand sustainability.\nMethods\nCase study and dataset\nThe main study area, Warren, encompasses the urban domain of the Red \nRun watershed located in southeast Michigan (Fig. 2a,b). This is a com-\nplex urban area with a substantial proportion of impervious surfaces. \nThe land-use cover obtained from the National Land Cover Database \n2019 (NLCD, https://www.mrlc.gov) indicates a predominant presence \nof developed areas with buildings, driveways, pavements and parking, \nwith an impervious surface ratio reaching 94.48% (Supplementary  \nFig. 18). The terrain in this region is relatively flat, with a northward \nslope adjacent Bear Creek. According to FEMA reports, some Warren  \nareas fall within the 100- and 500-year flood zones. To ensure a  \nrealistic simulation, a substantial amount of digital and field obser\u00ad\nvational data were collected to set up the model accurately.\nThe overall simulation domain encompasses the entire Red Run \nwatershed in which the Warren region is nested (Fig. 2a). This was done \nto ensure the modeling of spatial flood processes, such as surface \ninflows into the Warren area from other sub-basins and the simulation \nof flood waves in the Red Run River and its tributaries and their impact \non the Warren area. High-resolution elevation data (0.6-m resolution)  \nwere used to set up the overland flow model (Lidar Point Cloud \ndata kindly provided by the USGS, https://apps.nationalmap.gov/ \ndownloader/) (Supplementary Fig. 19a). Land-use and land-cover data \nwith a 30-m resolution obtained from the NLCD were used to estimate \nroughness parameters for the model. Similarly, impervious surface \ndata with a 30-m resolution, also derived from the NLCD, were used \nto estimate infiltration parameters. It is worth noting that other areas \nwithin the Red Run watershed also exhibited high impervious surface \nratios, ranging mostly between 47% and 95%.\nInfrastructures (buildings and stormwater systems) play a crucial  \nrole in surface flow routing and water drainage of the urbanized  \nareas. The most recent building footprint data were obtained from the  \nSoutheast Michigan Council of Governments (https://maps.semcog.\norg/BuildingFootprints), encompassing a total of 2,128 buildings \nwithin the Warren area (Fig. 2b and Supplementary Fig. 20). Data on  \nthe stormwater system, including manholes, outfalls and pipes, were \ncollected and compiled in collaboration with Macomb and Oakland \ncounties. A total of 21 culverts, 3,417 pipes, 3,393 manholes and  \n75 outfalls were identified and incorporated into the simulation for \nthe entire Red Run watershed (Fig. 2a,b and Supplementary Fig. 2).\nThe rainfall data were extracted from the Detroit City Airport \nweather station, which is part of the Automated Surface Observing \nSystems network. In addition to the STORM2014 used primarily in \nour analysis, data for 13 other major rainfall events from 2009 to 2021 \nwere also collected to simulate and assess the impact of variability in \nrainfall conditions on flooding. All rainfall events were collected over \na duration of 24\u2009h, with rainfall measurements taken at 1-min intervals \nand processed into 15-min intervals (Supplementary Fig. 4). Flow data \nat the inlet location (Fig. 2a) of the Red Run watershed were provided \nby the Clinton River Watershed Council upon request.\nField support for correcting outfall characteristics\nOutfalls play a crucial role in the stormwater system, as they are strate-\ngically designed to receive and discharge the water volume collected \nby the manholes, minimizing the risk of flooding. To ensure the closest \npossible simulation to reality, field observations were conducted on 21 \nJanuary 2023 to refine the design, location and functionality of the two \nimportant outfalls in the Warren area (Fig. 2b,c). Both outfalls are situ-\nated in the northern part of Warren, with their discharge gates flowing \ninto Bear Creek. These are open-type outfalls that allow water from Bear \nCreek to flow inside the stormwater pipes if the hydraulic head in the \nchannel is higher than the hydraulic head at the outfall locations. The \nfield measurements included the width of the outfalls, invert and crown \nelevations, and the bankfull width of Bear Creek around the outfall area.\nUrban flood modeling\nA coupled model integrating complex hydrologic, hydraulic and  \nmorphologic processes, which has been verified in previous case  \nstudies, was used in the research as the high-fidelity urban flood \nmodel50,51. The hydrology module, the TIN-based real time integrated \nbasin simulator (tRIBS), can simulate various hydrological processes \nsuch as canopy interception, evapotranspiration from the bare soil and \ncanopy, vertical and lateral moisture fluxes in the subsurface, and diverse \nrunoff generation mechanisms (for example, saturation excess, infiltra-\ntion excess, perched subsurface stormflow and groundwater exfiltra-\ntion) with appropriate inputs of meteorological data, topography,  \nland-use and soil-type data. Taking into account these hydrologic  \nprocesses enabled the model to simulate the hydrodynamics of \noverland flow (overland flow model, OFM52), relying on physically \nmodeled wave velocities within a domain of arbitrary geometric con-\nfiguration. The OFM model solves the full form of the 2D Saint-Venant \nequations (that is, the shallow water equations), which are derived  \nby depth-integrating the Navier\u2013Stokes equations. The governing  \nequations consist of a continuity equation and two momentum  \nequations for two perpendicular horizontal directions:\n\u2202U\n\u2202t + \u2202E\n\u2202x + \u2202G\n\u2202y = S\nwhere U is vector of flow variables, E and G are the flux terms in the  \nx and y directions, respectively, and S is the source vector:\nU =\n\u23a1\u23a2\u23a2\u23a2\n\u23a3\nh\nuh\nvh\n\u23a4\u23a5\u23a5\u23a5\n\u23a6\n, E =\n\u23a1\n\u23a2\n\u23a2\n\u23a2\n\u23a3\nuh\nu2h +\n1\n2 gh2\nuvh\n\u23a4\n\u23a5\n\u23a5\n\u23a5\n\u23a6\n, G =\n\u23a1\n\u23a2\n\u23a2\n\u23a2\n\u23a3\nvh\nuvh\nv2h +\n1\n2 gh2\n\u23a4\n\u23a5\n\u23a5\n\u23a5\n\u23a6\n,\nS =\n\u23a1\n\u23a2\n\u23a2\n\u23a2\n\u23a2\n\u23a3\ni\n\u2212gh\n\u2202zb\n\u2202x \u2212CDu\u221au2 + v2\n\u2212gh\n\u2202zb\n\u2202y \u2212CDv\u221au2 + v2\n\u23a4\n\u23a5\n\u23a5\n\u23a5\n\u23a5\n\u23a6\nwhere h represents flow depth, u, v are the flow velocities in the  \nx- and y-axis directions in the Cartesian system of coordinates, g is \nthe gravitational acceleration constant, i is the net runoff production \nrate, zb is the bed elevation and CD\u2009=\u2009gn2h\u22121/3 is the surface drag  \ncoefficient, where n is the Manning roughness coefficient. Detailed \ndescriptions of the numerical solution with the finite-volume method \nare available in ref. 52.\nThe solution appears to be feasible for many scenarios involving \noverland flow; however, due to numerical considerations, it is necessary \n\nNature Cities | Volume 1 | October 2024 | 654\u2013664\n661\nArticle\nhttps://doi.org/10.1038/s44284-024-00116-7\nto constrain the time step of the time-explicit finite-volume method \nby the mesh\u2019s smallest cell area, thus presenting a distinct compu\u00ad\ntational problem. For instance, in high-resolution applications, the \ntypical time step is of the order of \u223c\ud835\udcaa\ud835\udcaa (10\u22121)\u2009s, with a cell size of \u223c\ud835\udcaa\ud835\udcaa \n(101)\u2009m. Alternately, by using a time-implicit numerical scheme for \nsolving the shallow water equations, limitations on time-stepping are \nfewer and the overall number of solution steps is reduced, generating \na potential computational benefit. However, this is limited to no greater \nthan one order of magnitude due to the increased need of iterations \nwithin each time step to handle transient flow situations52.\nThe EXTRAN module of the Storm Water Management Model53 is \nfully coupled with tRIBS-OFM for simulating a stormwater drainage  \nsystem. EXTRAN is a 1D dynamic sewer network model based on 1D \nSaint-Venant equations. Water transport in conduits and nodes is \ncalculated based on the continuity and momentum equations, written \nrespectively as\n\u2202A\n\u2202t + \u2202q\n\u2202l = 0,\n\u2202q\n\u2202t +\n\u2202(q2/A)\n\u2202l\n+ gC \u2202H\n\u2202l + gCSl + gAel = 0\nwhere l is the distance along the conduit, C is the cross-sectional  \narea of the conduit, A denotes the cross-sectional area of the conduit, \nq and H are the flow rate and hydraulic head, respectively, of water in \nthe conduit at time t, Sl denotes the friction slope of the conduit and  \nel is the local energy loss per unit length of conduit.\nSurface water calculated from the OFM can move into the sewer \nnetwork through manholes; conversely, if the flow of the pipe network \nexceeds its drainage volume, a reverse (surcharge) flow occurs and \nbecomes the source term for surface water. Drainage refers to the flow \nof surface water to the manhole, and surcharge refers to the reverse \nflow of the flow from the manhole to the surface. These bidirectional \ndischarges, drainage and surcharge, can be calculated from the weir \nor orifice equations by comparing the hydraulic head at the manhole, \nthe ground elevation, and the water depth of the surface water. Specifi-\ncally, if the hydraulic head at the manhole (H) is lower than the ground \nelevation (zb), that is, H\u2009<\u2009zb, drainage, qd can be computed with a weir \nequation:\nqd = cwwwh\u221a2gh\nwhere cw is the weir discharge coefficient for the manholes, and ww \nis the width of the weir crest. Conversely, if the hydraulic head at \nthe manhole rises and is greater than the topographic elevation of  \nthe ground surface (but the hydraulic head at the manhole is still less \nthan the surface water level), that is, H\u2009>\u2009zb, drainage can be calculated \nby the orifice formula:\nqd = coAm\u221a2g (h + zb \u2212H)\nwhere co is the orifice discharge coefficient for the manholes, and Am \nis the area of the manhole mouth. Lastly, if the hydraulic head at the  \nmanhole exceeds the surface water level, that is, H\u2009>\u2009(h\u2009+\u2009zb), the  \nsurcharge, qsur, can be calculated according to the orifice formula:\nqsur = coAm\u221a2g (H \u2212h \u2212zb)\nMesh generation\nThis research necessitates sophisticated mesh generation to ensure \nthe most accurate and robust simulation outcomes. The intricate pro-\ncess of mesh generation that considers the complex geometry of the \nurban environment is summarized visually in Supplementary Fig. 21.  \nPrimarily, the Hydrology Analysis Tools of the ESRI ArcGIS Package  \nare used to delineate the watershed area (183.6\u2009km2) starting with the \nhigh resolution of 0.6\u2009m. Subsequently, the ArcGIS tool \u2018Raster2TIN\u2019 \nis used to construct a triangulated irregular network (TIN) from the \ndesignated watershed boundary. The TIN is capable of representing \nthe terrain\u2019s geometrical structure using substantially fewer nodes \nthan the original raster-based landscape representation.\nIn this study we focused on the Warren area (Supplementary Fig. 19b),  \nand thus divided it into cells of a considerably smaller size than  \nthose outside the region. We established minimum and maximum \ngrid resolution sizes of 1 and 100\u2009m2, respectively, inside the area  \nof Warren, and in the outer parts, the maximum cell size could  \nnot exceed 5,000\u2009m2. This refinement of the resolution helped  \nreduce the number of nodes and generated cells, thus alleviating the \ncomputational burden. Note that the Warren area does not correspond \nto a headwater catchment\u2014that is, surface flows and subsurface pipe \nflows can enter this rectangle-shaped domain from all of its sides.\nWe incorporated building footprints into the TIN to account \nfor the impact of buildings on flood wave propagation through the \nurbanscape. Initially, the building footprints were simplified using the \n\u2018Simplify Polygon\u2019 tool in ArcGIS, ensuring that the minimum length of \neach footprint polygon edge was 5\u2009m. These simplified footprints were \nthen merged with the TIN to identify the triangle cells representing \nbuildings, and the edges coinciding with the footprint polygons were \ndesignated as walls. Supplementary Fig. 21 provides an illustration \nof the process for the Warren region to demonstrate this procedure. \nThe simplification of building footprints was necessary for two main \nreasons: (1) to prevent the generation of very small triangles (this \nconstraint was imposed to maintain an appropriate time step in the \nfinite-volume method implemented in OFM for simulating overland \nflow) and (2) to reduce the number of cells in the overall mesh repre-\nsenting the watershed, thereby reducing the computational burden \nof the numerical simulation.\nOverall, the TIN resulted in 116,215 mesh nodes and 232,292 triangle  \ncells (Supplementary Fig. 19b). The smallest cell area is 1.3\u2009m2, and the \naverage cell size generated for the Warren area is ~47\u2009m2. It is worth \nnoting that, even with this modestly sized TIN, the simulation time for a \n24-h rainfall event requires up to 240\u2009h of wall-clock time to complete. \nIt should be noted that in our study, technologies to speed up model \nruntime, such as parallelization or the use of surrogate modeling54, \nwere not employed. However, when such computationally expensive \nmodels need to be used in applications such as design optimization \nand control/operation of stormwater systems, as well as real-time \nforecasting, such technologies are particularly necessary to provide \nsimulation/forecast results in a timely manner.\nAll of the land cover is developed area, with the impervious surface \nranging from 47% to 95%. In the Warren area, the impervious surface \naccounts for up to 94.48%. The land use and impervious surface infor-\nmation was downloaded from NLCD 2019 (https://www.mrlc.gov/).\nExperimental design\nThe research comprised two modeling experiments to address two \nresearch questions. First, we evaluated two scenarios of the spatial \ndistribution of rainfall in causing flooding within the Warren area. For \nthis purpose, STORM2014 precipitation data were configured as a grid \nwith a resolution of 500\u2009m\u2009\u00d7\u2009500\u2009m using rainfall data obtained from \nthe Detroit City Airport rain gauge. This gridded product was used as \nmodel input in the first scenario, \u2018HR-Integrated\u2019 (results are presented \nin Fig. 2d). The drainage system was designed to closely resemble real-\nworld conditions, allowing for water exchange between the inside and \noutside of the Warren drainage system based on the dynamically chang-\ning distribution of hydraulic head over the surface and below ground \nin storm sewers. In the second scenario, the gridded rainfall data for \nthe Warren area were set to 0\u2009mm for the duration of the STORM2014 \nevent (that is, 24\u2009h). This scenario was named \u2018NR-Integrated\u2019 (Fig. 2e).\nIn the second experiment, we explored the functioning of two \nstormwater sewer outfalls (Fig. 2b,c) using three different model \nconfigurations. The first set-up represents the most realistic design \n\nNature Cities | Volume 1 | October 2024 | 654\u2013664\n662\nArticle\nhttps://doi.org/10.1038/s44284-024-00116-7\nby allowing backwater to occur (water from the open channel flows \nback into the sewer system through the outfall) (termed \u2018Integrated\u2019 \noutfalls). The second set-up assumes the current sewer system is  \nunaffected by water levels at the outfall locations, maximizing drainage  \ncapacity by allowing unrestricted discharge from the stormwater  \nsystem into the open channel without backflow from Bear Creek (termed \n\u2018Controlled\u2019 outfalls). The third set-up precludes any water exchange \nat the outfalls between the stormwater system and open channel, with \nthe total outflow from the system stored in an underground storage  \nsized to contain all runoff from the Red Run watershed under  \nthe STORM2014 event (a volume of ~24.77\u2009million\u2009m3) rather than \nentering the open channel. We term this case \u2018Water loss\u2019, the results \nof which are shown in Supplementary Fig. 3.\nThe Manning\u2019s roughness coefficient is the most influential para\u00ad\nmeter for overland and sewer flows55. For surface flow simulated with \nthe OFM model, we assume a spatially uniform value of 0.015, which  \nis in the middle of the interval for concrete surfaces (0.012\u20130.018)29. \nManning\u2019s coefficients for all the conduits were set to 0.013  \n(with concrete as the pipe material)56.\nData mining and model evaluation\nIn terms of its suitability to mimic the important details of hydro\u00addynamic \nprocesses such as overland flow, backwater, hydraulic jump, as well as \nthe influence of vegetation and buildings on surface flow, the tRIBS-OFM  \nmodel has been extensively tested and validated in numerous  \nprevious studies52,54,57. To analyze the performance of the tRIBS-OFM \nmodel with regard to its capability to depict the interaction between the \nstormwater system and surface flow, we conducted a benchmark experi-\nment to evaluate simulated surcharge in manholes, as demonstrated \nin Supplementary Fig. 22. A detailed description of this case study  \nis provided by the United Kingdom Environment Agency58. This \nevidence supports that tRIBS-OFM produces reliable results when  \nsimulating complex hydraulic dynamics and interactions between \nthe stormwater system and the surface, comparable with that of  \nother models that used the same dataset58. The competent performance \nby tRIBS-OFM makes it a great choice for studies of urban flooding.\nDue to the limitations of flood data availability in the study area, \nconventional model validation is not feasible. Therefore, an alternate \napproach was adopted to validate the model results by collecting \nflood evidence from various internet sources. The manual process of \ndata collection and analysis involved several steps to ensure reliability \nand accuracy:\n\u2022 \nKeywords and search operators: an ad hoc combination of  \nkeywords was used for searching, including \u20182014 flood\u2019, \n\u2018Macomb County\u2019, \u2018Oakland County\u2019, \u2018Warren MI\u2019, \u2018Bear Creek \nflood\u2019 and \u2018Michigan flood disaster 2014\u2019. These keywords were \nentered into the Google search engine. From the obtained \nsearch results, images published in reports, newspapers and \nblogs were selected and evaluated for inclusion in the study.\n\u2022 \nQuality assessment and metadata extraction: in a subsequent \nfiltering step, the obtained images were visually analyzed to \ndetermine flooded locations and water levels relative to visible \nobjects such as people, vehicles or traffic signs. Details about \nstreet names or intersections were used to estimate and extract \ncoordinates. The extracted metadata include information such \nas coordinates, flood severity, flooded location (street name) \nand the website link from which the image was sourced.\n\u2022 \nUncertainty assessment: estimating water levels from images \ninvolves subjective interpretation and conjectures necessitating \nan evaluation of relevant confidence intervals. Typically, there  \nare two approaches to estimating this uncertainty: (1) assuming  \nthat the uncertainty is proportional to the estimated water \ndepth or (2) considering that the uncertainty is associated \nwith each reference class59. In this study, we opted for the first \napproach, assuming that the standard deviation was equal to  \n20% of the estimated water depth. The final results for the esti\u00ad\nmation of water levels at six locations are described in Supple-\nmentary Table 2.\nThe flood depth data were subsequently used to validate the model \nresults (Supplementary Fig. 23). One limitation of these flood depth \nestimations is the lack of certain information regarding the time at \nwhich the images were acquired. Therefore, the accuracy of such a \ncomparison of the model results is constrained. We thus considered \nthe obtained flood depths as a threshold for \u2018being flooded\u2019 and used \nthem to compare with the model results to determine whether the \nmodel exceeds this threshold at any simulation time. In essence, the \nmodel results confirm that at the locations marked as \u2018flooded\u2019, the \noccurrence of flooding is consistent with the simulation results, with \nrelatively minor discrepancies in flood depths. For example, at loca-\ntion 1, the model yields a maximum flood depth of ~0.85\u2009m, with an \nestimated data range from 0.8 to 1.2\u2009m. Similar results can be observed \nfor other locations.\nFundamentally, an accurate description of what occurs in  \nreality requires substantial efforts, not only in terms of the model\u2019s \ncapabilities but also regarding the accuracy of the data used. This \nstudy does not aim to assert that the model used in the study can pre-\ncisely depict all relevant flood dynamics during STORM2014 across the \nentire study area (Red Run watershed and Warren area). Rather, we aim \nto show that, in conditions with complex permutation of the storm\u00ad\nwater infrastructure, terrain, land use and buildings, the sophisticated, \nstate-of-the-science model can provide insights into potential flood \nphenomena that have not been examined previously.\nReporting summary\nFurther information on research design is available in the Nature \nPortfolio Reporting Summary linked to this Article.\nData availability\nThe land use cover was obtained from the National Land Cover Database \n2019 (https://www.mrlc.gov). High-resolution elevation data were \nprovided by the USGS (https://apps.nationalmap.gov/downloader/). \nThe building footprint data were obtained from the Southeast Michigan \nCouncil of Governments (https://maps.semcog.org/BuildingFoot-\nprints). Data on the stormwater system, including manholes, outfalls \nand pipes, were collected and compiled in collaboration with Macomb \nand Oakland Counties. The rainfall data were from the Automated \nSurface Observing Systems network (https://www.weather.gov/asos/). \nThe simulation dataset (~600\u2009gigabytes) archived on the Globus Cloud  \nis available upon written request to the authors. Source data are  \nprovided with this paper.\nCode availability\nData processing and analysis were performed using MATLAB 2022b \n(standard version). The source code is publicly accessible via GitHub \nat https://github.com/vinhngoctran/RedRun_processing.\nReferences\n1.\t\nDonat, M. G., Lowry, A. L., Alexander, L. V., O\u2019Gorman, P. A. & \nMaher, N. More extreme precipitation in the world\u2019s dry and wet \nregions. Nat. Clim. Change 6, 508\u2013513 (2016).\n2.\t\nPrein, A. F. et al. The future intensification of hourly precipitation \nextremes. Nat. Clim. Change 7, 48\u201352 (2016).\n3.\t\nScott, D. T., Gomez-Velez, J. D., Jones, C. N. & Harvey, J. W. \nFloodplain inundation spectrum across the United States.  \nNat. Commun. 10, 5194 (2019).\n4.\t\nPaprotny, D., Sebastian, A., Morales-Napoles, O. & Jonkman, S. N.  \nTrends in flood losses in Europe over the past 150 years.  \nNat. Commun. 9, 1985 (2018).\n\nNature Cities | Volume 1 | October 2024 | 654\u2013664\n663\nArticle\nhttps://doi.org/10.1038/s44284-024-00116-7\n5.\t\nUNDRR Annual Report 2021 (United Nations Office for Disaster Risk \nReduction, 2021).\n6.\t\nNWS Preliminary US Flood Fatality Statistics (National Weather \nService, 2023).\n7.\t\nStorm Events Database (NOAA National Centers for Environmental \nInformation, 2023).\n8.\t\nUS Billion-Dollar Weather and Climate Disasters (NOAA National \nCenters for Environmental Information, 2023); https://doi.org/ \n10.25921/stkw-7w73\n9.\t\nHirabayashi, Y. et al. Global flood risk under climate change.  \nNat. Clim. Change 3, 816\u2013821 (2013).\n10.\t Rentschler, J. et al. Global evidence of rapid urban growth in flood \nzones since 1985. Nature 622, 87\u201392 (2023).\n11.\t\nRevi, A. et al. Urban areas. In Climate Change 2014: Impacts, \nAdaptation and Vulnerability. Part A: Global and Sectoral Aspects. \nContribution of Working Group II to the Fifth Assessment Report of \nthe Intergovernmental Panel on Climate Change Ch. 8, 535\u2013612 \n(Cambridge Univ. Press, 2014).\n12.\t Nowak, D. J. & Walton, J. T. Projected urban growth (2000\u20132050) \nand its estimated impact on the U.S. forest resource. J. Forestry \n103, 383\u2013389 (2005).\n13.\t Andreadis, K. M. et al. Urbanizing the floodplain: global changes \nof imperviousness in flood-prone areas. Environ. Res. Lett. 17, \n104024 (2022).\n14.\t Golden, H. & Hoghooghi, N. Green infrastructure and its \ncatchment-scale effects: an emerging science. WIREs Water 1, \ne1254 (2018).\n15.\t Radhakrishnan, M., Pathirana, A., Ashley, R. M., Gersonius, B. & \nZevenbergen, C. Flexible adaptation planning for water sensitive \ncities. Cities 78, 87\u201395 (2018).\n16.\t Wong, T. H. & Brown, R. R. The water sensitive city: principles for \npractice. Water Sci. Technol. 60, 673\u2013682 (2009).\n17.\t From Gray to Green\u2014Helping Communities Adopt Green \nInfrastructure (EPA, 2023).\n18.\t Alves, A., Vojinovic, Z., Kapelan, Z., Sanchez, A. & Gersonius, B. \nExploring trade-offs among the multiple benefits of green-blue-\ngrey infrastructure for urban flood mitigation. Sci. Total Environ. \n703, 134980 (2020).\n19.\t Wilbanks, T. et al. Climate Change and Infrastructure, Urban \nSystems, and Vulnerabilities. Technical Report for the U.S. \nDepartment of Energy in Support of the National Climate \nAssessment (Springer, 2013).\n20.\t Collentine, D. & Futter, M. N. Realising the potential of natural \nwater retention measures in catchment flood management: \ntrade-offs and matching interests. J. Flood Risk Manag. 11, 76\u201384 \n(2018).\n21.\t EFAB Report: Evaluating Stormwater Infrastructure Funding and \nFinancing (EPA, 2023).\n22.\t Qiao, X.-J., Kristoffersson, A. & Randrup, T. B. Challenges to \nimplementing urban sustainable stormwater management from \na governance perspective: a literature review. J. Clean. Prod. 196, \n943\u2013952 (2018).\n23.\t Rosenzweig, B. R. et al. Pluvial flood risk and opportunities for \nresilience. WIREs Water 5, e1302 (2018).\n24.\t ten Veldhuis, J. A. E. How the choice of flood damage metrics \ninfluences urban flood risk assessment. J. Flood Risk Manag. 4, \n281\u2013287 (2011).\n25.\t Backhaus, A., Dam, T. & Jensen, M. B. Stormwater management \nchallenges as revealed through a design experiment with \nprofessional landscape architects. Urban Water J. 9, 29\u201343 (2012).\n26.\t Wright, D. B., Bosma, C. D. & Lopez-Cantu, T. U.S. hydrologic \ndesign standards insufficient due to large increases in frequency \nof rainfall extremes. Geophys. Res. Lett. 46, 8144\u20138153 (2019).\n27.\t Mays, L. W. Stormwater Collection Systems Design Handbook \n(McGraw-Hill, 2001).\n28.\t Standard Guidelines for the Design of Urban Stormwater Systems \n(ASCE, 2006).\n29.\t Arcement, G. J. & Schneider, V. R. Guide for Selecting Manning\u2019s \nRoughness Coefficients for Natural Channels and Flood Plains  \n(US GPO, 1989).\n30.\t Woods-Ballard, B. et al. The SuDS Manual (CIRIA, 2007).\n31.\t Drainage Services Department. Stormwater Drainage Manual: \nPlanning, Design and Management (Government of the Hong \nKong Special Administrative Region, 2018).\n32.\t Allen, M. D. et al. Water Sensitive Urban Design: Basic Procedures \nfor \u2018Source Control\u2019 of Stormwater. A Handbook for Australian \nPractice (ed. Argue, J.) (Univ. South Australia, 2004).\n33.\t Bradford, A. & Gharabaghi, B. Evolution of Ontario\u2019s stormwater \nmanagement planning and design guidance. Water Quality Res. J. \n39, 343\u2013355 (2004).\n34.\t Wing, O. E. J. et al. Inequitable patterns of US flood risk in the \nAnthropocene. Nat. Clim. Change 12, 156\u2013162 (2022).\n35.\t Framing the Challenge of Urban Flooding in the United States \n(National Academies Press, 2019).\n36.\t Rosenzweig, B. R. et al. The value of urban flood modeling.  \nEarth\u2019s Future 9 (2021); https://doi.org/10.1029/2020ef001739\n37.\t Guo, K., Guan, M. & Yu, D. Urban surface water flood modelling\u2014a \ncomprehensive review of current models and future challenges. \nHydrol. Earth Syst. Sci. 25, 2843\u20132860 (2021).\n38.\t Forero-Ortiz, E., Mart\u00ednez-Gomariz, E. & Ca\u00f1as Porcuna, M.  \nA review of flood impact assessment approaches for \nunderground infrastructures in urban areas: a focus on transport \nsystems. Hydrol. Sci. J. 65, 1943\u20131955 (2020).\n39.\t Carmichael, C., Danks, C. & Vatovec, C. Assigning blame:  \nhow local narratives shape community responses to extreme \nflooding events in Detroit, Michigan and Waterbury, Vermont. \nEnviron. Commun. 14, 300\u2013315 (2020).\n40.\t Shepardson, D. White House Approves Michigan Disaster \nDeclaration (Detroit News Washington Bureau, 2014).\n41.\t National Flood Hazard Layer (FEMA, 2023).\n42.\t Yazdanfar, Z. & Sharma, A. Urban drainage system planning and \ndesign\u2014challenges with climate change and urbanization:  \na review. Water Sci. Technol. 72, 165\u2013179 (2015).\n43.\t Chen, S. S. et al. Designing sustainable drainage systems in \nsubtropical cities: challenges and opportunities. J. Clean. Prod. \n280, 124418 (2021).\n44.\t Schmitt, T. G. & Scheid, C. Evaluation and communication  \nof pluvial flood risks in urban areas. WIREs Water 7, e1401  \n(2020).\n45.\t Prosdocimi, I., Kjeldsen, T. R. & Miller, J. D. Detection and \nattribution of urbanization effect on flood extremes using \nnonstationary flood-frequency models. Water Resour. Res. 51, \n4244\u20134262 (2015).\n46.\t Storm Water Management Plan (City of Warren, 2012); https://\nwww.cityofwarren.org/departments/engineering-division/\n47.\t Mapping the Zone: Improving Flood Map Accuracy (National \nAcademies Press, 2009).\n48.\t Kousky, C. Financing flood losses: a discussion of the national \nflood insurance program. Risk Manag. Insurance Rev. 21, 11\u201332 \n(2018).\n49.\t Pralle, S. Drawing lines: FEMA and the politics of mapping flood \nzones. Clim. Change 152, 227\u2013237 (2019).\n50.\t Ivanov, V. Y., Vivoni, E. R., Bras, R. L. & Entekhabi, D. Catchment \nhydrologic response with a fully distributed triangulated irregular  \nnetwork model. Water Resour. Res. https://doi.org/10.1029/ \n2004wr003218 (2004).\n51.\t Ivanov, V. Y., Vivoni, E. R., Bras, R. L. & Entekhabi, D. Preserving \nhigh-resolution surface and rainfall data in operational-scale \nbasin hydrology: a fully-distributed physically-based approach.  \nJ. Hydrol. 298, 80\u2013111 (2004).\n\nNature Cities | Volume 1 | October 2024 | 654\u2013664\n664\nArticle\nhttps://doi.org/10.1038/s44284-024-00116-7\n52.\t Kim, J., Warnock, A., Ivanov, V. Y. & Katopodes, N. D. Coupled \nmodeling of hydrologic and hydrodynamic processes including \noverland and channel flow. Adv. Water Res. 37, 104\u2013126 (2012).\n53.\t Rossman, L. A. Storm Water Management Model User\u2019s Manual, \nVersion 5.0 (US Environmental Protection Agency, 2010).\n54.\t Ivanov, V. Y. et al. Breaking down the computational barriers \nto real\u2010time urban flood forecasting. Geophys. Res. Lett. 48, \ne2021GL093585 (2021).\n55.\t Ozdemir, H., Sampson, C. C., de Almeida, G. A. M. & Bates, P. D. \nEvaluating scale and roughness effects in urban flood modelling \nusing terrestrial LIDAR data. Hydrol. Earth Syst. Sci. 17, 4015\u20134030 \n(2013).\n56.\t Chow, V. Open Channel Hydraulics (McGraw Hill Education,  \n1959).\n57.\t Kim, J., Ivanov, V. Y. & Katopodes, N. D. Hydraulic resistance to \noverland flow on surfaces with partially submerged vegetation. \nWater Resour. Res. https://doi.org/10.1029/2012wr012047 (2012).\n58.\t N\u00e9elz, S. & Pender, G. Benchmarking the Latest Generation of 2D \nHydraulic Modelling Packages (UK Environment Agency, 2013).\n59.\t Songchon, C., Wright, G. & Beevers, L. The use of crowdsourced \nsocial media data to improve flood forecasting. J. Hydrol. 622, \n129703 (2023).\nAcknowledgements\nV.N.T. and V.Y.I. acknowledge the support of the US National Science \nFoundation CMMI program award no. 2053429 and the Department \nof Defense, Department of the Navy, the Office of Naval Research \naward no. N00014-23-1-2735, Environmental Protection Agency \ngrant #2020\u20102509. J. Kim was supported by the National Research \nFoundation of Korea (NRF) grant funded by the Korea government \n(MSIT)(NRF-2022R1A2C2008584. D.B. Wright and G.A. Alexander \nacknowledge the support of the U.S. National Science Foundation \nCMMI program award #2053358. We acknowledge informative \ndiscussions with P. Seelbach, B. Kerkez, J. Bednar, G. O\u2019Neil, C. Brown, \nC. Purdy, A. Asher, M. Thomas and S. Bergt that assisted this study.\nAuthor contributions\nV.N.T. and V.Y.I. designed the study. V.N.T., W.H., K.M. and F.D. collected \nthe data. V.N.T. conducted the experiments. V.N.T. and V.Y.I. analyzed \nthe data and wrote and revised the paper with inputs from all  \nco-authors. All authors contributed to the interpretation of results, \nwriting and revision of the paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s44284-024-00116-7.\nCorrespondence and requests for materials should be addressed to \nValeriy Y. Ivanov.\nPeer review information Nature Cities thanks Matt Bartos and the \nother, anonymous, reviewer(s) for their contribution to the peer review \nof this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature America, \nInc. 2024\n\n1\nnature portfolio  |  reporting summary\nApril 2023\nCorresponding author(s):\nValeriy Y. Ivanov\nLast updated by author(s): Jul 10, 2024\nReporting Summary\nNature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nAll datasets were manually collected.\nData analysis\nData processing and analysis were performed using MATLAB 2022b (standard version). The source code is publicly available in a GitHub \nrepository (https://github.com/vinhngoctran/RedRun_processing).\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and \nreviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A description of any restrictions on data availability \n- For clinical datasets or third party data, please ensure that the statement adheres to our policy \n \nThe land use cover was obtained from the National Land Cover Database 2019 (https://www.mrlc.gov). High-resolution elevation data was provided by the USGS, \nhttps://apps.nationalmap.gov/downloader/). The building footprint data was obtained from the Southeast Michigan Council of Governments (https://\nmaps.semcog.org/BuildingFootprints). Data on the stormwater system, including manholes, outfalls, and pipes, were collected and compiled in collaboration with \n\n2\nnature portfolio  |  reporting summary\nApril 2023\nthe Macomb and Oakland Counties. The rainfall data were the Automated Surface Observing Systems network (https://www.weather.gov/asos/). The simulation \ndataset (approximately 600 Gigabytes) archived on the Globus Cloud is available upon written request to the authors.\nResearch involving human participants, their data, or biological material\nPolicy information about studies with human participants or human data. See also policy information about sex, gender (identity/presentation), \nand sexual orientation and race, ethnicity and racism.\nReporting on sex and gender\nN/A\nReporting on race, ethnicity, or \nother socially relevant \ngroupings\nN/A\nPopulation characteristics\nN/A\nRecruitment\nN/A\nEthics oversight\nN/A\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nEcological, evolutionary & environmental sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nThe main purpose of the study is to examine how built infrastructure exacerbates the space-time connectivity of urban floods, using \nthe 2014 flood in Southeast Michigan as a case study. The findings highlight the importance of considering the interconnectedness of \nflood processes in urban watershed systems when designing flood control and warning systems to enhance urban flood resilience.\nResearch sample\nThe research case study examines the 2014 flood event in Southeast Michigan (Warren City) to identify key factors contributing to \nurban flooding.\nSampling strategy\nN/A\nData collection\nAll data was collected manually through the websites provided in the manuscript.\nTiming and spatial scale\nThe data was collected during the period of 2021-2023 from various sources (including sources published on websites as well as data \ngathered from governments through processes of exchange and collaboration).\nData exclusions\nN/A\nReproducibility\nAll model simulations are reproducible.\nRandomization\nN/A\nBlinding\nN/A\nDid the study involve field work?\nYes\nNo\nField work, collection and transport\nField conditions\nField work was conducted to collect and verify information about the stormwater system outfalls in the Warren area.\nLocation\nWarren, Michigan, U.S (latitude = 42.52319, longitude = -83.02367)\nAccess & import/export\nSince the survey locations were in public areas, the measurement activities did not encounter any difficulties or require any permits.\n\n3\nnature portfolio  |  reporting summary\nApril 2023\nDisturbance\nN/A\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology and archaeology\nAnimals and other organisms\nClinical data\nDual use research of concern\nPlants\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nNovel plant genotypes\nN/A\nSeed stocks\nN/A\nAuthentication\nN/A\nPlants\n\n\n    Scientific Research Findings:", "answer": "Starting with analysis of unexpected flooding in southeast Michigan, USA, in 2014, we find that human-engineered infrastructure can introduce additional stormwater flow connectivity in urban environments that may increase flood risk. Specifically, flooding may worsen where river channels are connected to underground infrastructure. This has immediate implications for stormwater design, as differing assumptions about flood connectivity between river flooding and infrastructure can yield markedly different estimates of potential inundation. The ubiquitous practice of ignoring this connectivity \u2014 assessing the performance of infrastructure elements in isolation and with overly simplistic rainfall scenarios \u2014 can lead to substantial errors, which result in reduced infrastructure effectiveness. Paradoxically, design errors can reverse stormwater network functionality in the real world and lead, for instance, to flooding even in the absence of local rainfall. Although this study illuminates broader challenges in stormwater network design and flood risk management, solutions will need to be tailored to the needs and unique characteristics of individual communities.", "id": 61}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Cities | Volume 1 | October 2024 | 665\u2013676\n665\nnature cities\nhttps://doi.org/10.1038/s44284-024-00121-w\nArticle\nOperationalizing climate justice in the \nimplementation of Boston\u2019s Building \nPerformance Standard\nClaudia V. Diezmart\u00ednez\u2009\n\u200a\u2009\u2009\n, Benjamin K. Sovacool & Anne G. Short Gianotti\u2009\n\u200a\u2009\nCities are moving toward the implementation of more just urban climate \nactions, but the politics and processes of operationalizing climate justice \nin practice remain understudied. Here we examine the implementation \nof climate justice through Boston\u2019s Building Emissions Reduction and \nDisclosure Ordinance (BERDO), a landmark Building Performance Standard \nthat reflects a transformative shift towards carbon neutrality in cities. We \nutilize a rich mixed-methods research design that is rooted in 5 months \nof participant observation within the City of Boston\u2019s Environment \nDepartment, 20 expert interviews and a systematic content analysis of \nhundreds of policy documents. We find that implementing BERDO entails \npolitical contestation over differing conceptions of resistance and power \nrelations around climate justice. Justice becomes subject to scope and scale \ndiscrepancies, the processes of bureaucratization and even weaponization. \nIn documenting these tensions, we provide insights into the complex \nchallenges that cities may face as they begin to operationalize climate \njustice on the ground.\nCities are now taking the lead in implementing just urban transitions\u2014\nfair and equitable transitions toward low-carbon and resilient urban \nsocieties1,2. Given the global push towards urbanization into metro\u00ad\npolitan areas3, and the devolution of some authority in climate deci-\nsion-making to subnational actors, cities have been leaders in climate  \naction for over two decades4\u20136. More recently, city governments have \nbeen increasingly recognizing the connections between climate change \nand social justice, and making headway with the integration of justice \nand equity concerns into climate plans and policy implementation \ntools7\u20139. However, these efforts have been met with questions about \nwhether and how cities will take just climate action from planning  \ninto practice.\nScholars have repeatedly criticized the gaps between the rhetoric \nand reality of urban climate action10\u201313, which we define here as poli-\ncies and programs to mitigate and adapt to climate change. However, \nexisting literature has focused primarily on either analysing the devel-\nopment of plans, policies and targets or evaluating the post-facto \noutcomes of cities\u2019 programs2,11,12,14\u201318. This collective scholarship has \nrevealed multiple barriers to urban climate action, from funding con-\nstraints and limited capacities to lack of political will and issues of \nauthority16,19,20. These barriers result in, at best, insufficient reductions \nin greenhouse gas emissions21 and, at worst, the exacerbation of climate \nvulnerabilities and injustice in cities22\u201325.\nAlthough cities purportedly continue to underdeliver on climate \naction and justice, the politics and dynamics of policy implementa-\ntion\u2014the process through which city governments translate goals \nand plans into operational and enforceable programs\u2014remain under\u00ad\nstudied14,15,26. Rather than recognizing the complexities of operation-\nalizing climate action, existing theory suggests that either unique \n\u201cconfigurations\u201d of enabling factors16 or the mere removal of barriers \nshould enable cities to successfully implement climate policies19,27. \nThis thinking obscures the \u201cpolitics and contested nature of low car-\nbon urbanism\u201d28 and hinders the analysis of policy implementation \nthrough political and justice lenses15,27, with scholars ultimately failing \nto address how the \u201crecognition of socially vulnerable groups either \ncarries through or drops out of the policy implementation process\u201d18.\nReceived: 24 March 2024\nAccepted: 6 August 2024\nPublished online: 13 September 2024\n Check for updates\nDepartment of Earth and Environment, Boston University, Boston, MA, USA. \n\u2009e-mail: cvdiezm@bu.edu\n\nNature Cities | Volume 1 | October 2024 | 665\u2013676\n666\nArticle\nhttps://doi.org/10.1038/s44284-024-00121-w\nThe history and implementation process of BERDO\nThe implementation of BERDO began shortly after the adoption of \nthe ordinance with the launch of rule-making (Fig. 1a). City officials \ndescribed the goal of rule-making as \u201ccreating clarity\u201d for building own-\ners on how to implement different compliance mechanisms to achieve \nemissions reductions (Table 1) and to define how justice mandates that \nwere embedded in the ordinance (Table 2) work \u201cnot just in concept but \nin action\u201d. Regulations were adopted in three phases. Each phase was \ndedicated to specific regulatory topics and included the engagement \nof multiple actors (Fig. 1b).\nTo operationalize climate justice, city staff simultaneously sought \n(1) to elevate community voices in regulatory decisions through the \nformation of a Community Advisory Group (CAG) that consists of com-\nmunity advocates and leaders working on environmental justice, social \njustice, affordable housing and climate action and (2) to maintain the \nbuy-in of the regulated parties that are subject to BERDO, including real \nestate, hospitals, universities and other building owners (Fig. 1b). The \nconvening of a justice-centered CAG for rule-making was also unique  \nto BERDO compared with other existing BPS programs. However,  \nunlike the Review Board, which has decision-making authorities in \nmultiple aspects of implementation, the CAG served in an advisory \nrole specifically in the rule-making process.\nCAG members met regularly with city officials to advise on how \nto align regulatory and implementation decisions with justice ideals \nand community priorities. For city staff, the CAG provided a space to \n\u201cexplicitly talk about equity\u201d and enable \u201cground truthing\u201d with com-\nmunity leaders that have \u201ctrue expertise and what is needed or wanted \nin a community\u201d. The CAG also provided a buttress against opposition \nfor regulatory decisions that were designed for justice, as stated by a \ncity official: \u201cwe had the backup to say \u2018we are talking to community \nmembers, and this is a concern that\u2019s being brought up\u2019\u201d. CAG members \ndescribed their role as representing the needs of their neighborhoods, \nto ensure that implementation \u201cwas aligned with equity\u201d and does not \ndisproportionately burden vulnerable communities, while making \nsure that BERDO \u201chad teeth\u201d and guardrails against \u201cthe tricks\u201d that \nbuilding owners may use to circumvent obligations. Perceptions from  \nCAG members about this process are summarized in Fig. 2.\nWhereas CAG members provided direction for incorporating  \njustice in rule-making, city staff saw engaging with regulated parties \nas an important component of fulfilling justice goals in the long term. \nAs one city official stated: \u201cwe need to bring all building owners along  \nin this process to actually see building decarbonization happen\u201d.  \nRegulated parties engaged in rule-making through public comment \nletters, public hearings and ad hoc meetings with city officials. In \nthis regard, city officials explained that \u201chaving a clear process for \nhearing and responding to feedback was really important\u201d. \u201cPeople \nnotice that, and it matters. Even if they\u2019re not all going to be happy \nwith exactly where you ended up, they understand how you got there \nand felt like they weren\u2019t left out of the process\u201d and it \u201cbuilds some \ngoodwill with stakeholders, even when we weren\u2019t taking all of their \nrecommendations\u201d.\nBeyond rule-making, implementing the early stages of BERDO also \ninvolved operationalizing the procedural and recognitional aspects of \njustice in everyday decisions and practices (see definitions in Methods). \nThis included managing the reporting and third-party verification pro-\ncesses, a help desk, outreach and education for building owners and the \nReview Board\u2019s nomination and seating process. Through these pro-\ncesses, city officials sought to implement justice by (1) creating resources \nto help with \u201cgetting everyone subject to BERDO across the finish line\u201d \nfor reporting and emissions compliance and by (2) giving \u201ctime, atten-\ntion, and resources towards under-resourced residents and owners\u201d.\nJustice contestations in policy implementation\nRule-making not only set the groundwork for BERDO but also  \nprovided a forum for CAG members and regulated parties to contest \nHere we examine the politics and processes through which city \ngovernments operationalize climate action and climate justice (see \ndefinition in Methods) via a case study of Boston\u2019s Building Perfor-\nmance Standard (BPS), one of the latest policy approaches for build-\ning decarbonization in the United States. Buildings are often one of \nthe largest sources of greenhouse gas emissions in urban areas29,30, \naccounting for about 37% of global energy-related carbon emissions \nand more than 34% of direct energy consumption31. Energy efficiency \nis one of the most common climate interventions across cities world-\nwide6,32, and the mitigation sector for which US cities articulate justice \nconcerns most frequently7. Existing policies, however, have produced \nonly marginal improvements in building energy intensity33. Energy \ndemand and emissions from buildings have continued to increase \nglobally, and few local efforts are aligned with achieving net-zero \noperational emissions from buildings33.\nAs one the first BPS programs adopted in the United States, \nBoston\u2019s Building Emissions Reduction and Disclosure Ordinance \n(BERDO) represents a radical shift in building decarbonization. \nBERDO is an ordinance (local law) adopted in 2021 that mandates \nlarge residential buildings (with 15 or more units) and commercial \nbuildings (\u226520,000\u2009ft2/1,858 m2) in Boston to progressively reduce their \ngreenhouse gas emissions to reach net-zero by 2050. BERDO requires  \nthe building owners (1) to report their annual energy and water use, \n(2) to verify reported data through a third party on a regular basis and \n(3) to reduce their annual emissions below an emissions standard that \ncorresponds to their building-use type(s). Reporting and third-party \nverification have been required for all buildings since 2022. Emis-\nsions compliance begins in 2025 for larger buildings (\u226535 units or \n\u226535,000\u2009ft2/3,251 m2) and in 2030 for smaller buildings (15\u201334 units \nor 20,000\u201334,999\u2009ft2/1,858\u20133,252 m2). BERDO is lauded for being \none of the few existing BPS programs that explicitly incorporates \njustice mandates within its ordinance. This includes the creation of a \ncommunity-driven Review Board that has substantial decision-making \nauthority over the implementation of the program, and the establish-\nment of the Equitable Emissions Investment Fund to support building \ndecarbonization projects that benefit environmental justice communi-\nties. Whereas other BPS programs include committees with advisory  \nfunctions (for example, New York City has an Advisory Board for  \nLocal Law 97), BERDO is unique in its delegation of implementation \npowers to a resident-driven Review Board that provides permanent \ncommunity oversight over the program.\nWe investigate the first two years of implementation of the BERDO \nprogram with a particular focus on rule-making, the process through \nwhich the City of Boston developed rules and regulations for imple-\nmenting and enforcing BERDO. We use a mixed-methods approach \nthat combines 5 months of participant observation within the City of \nBoston\u2019s Environment Department, 20 interviews with city staff and \ncommunity leaders involved in the implementation of BERDO and a \nsystematic content analysis of over 200 policy documents related to \nrule-making. We find that policy implementation served as an impor-\ntant site of political contestation and resistance around climate justice. \nContestations about justice, equity and fairness were mobilized by \ndifferent actors to advance their interests in rule-making, ultimately \nshaping the implementation of BERDO itself. We provide insights into \nthe complex challenges that cities may face as they begin to operation-\nalize climate justice on the ground, and argue for a shift in scholarship \nto politicize policy implementation and reframe climate justice as a \npractice that continues beyond policy planning.\nResults\nDrawing from our mixed-methods research design (Methods), we \norganize our analytical insights according to three core themes: (1) the \nhistory and implementation process of BERDO, (2) contested climate \njustice claims that arise during implementation and (3) challenges in \ntranslating climate justice from theory into practice.\n\nNature Cities | Volume 1 | October 2024 | 665\u2013676\n667\nArticle\nhttps://doi.org/10.1038/s44284-024-00121-w\nand reinterpret the justice goals and mandates that are embedded in \nthe ordinance. These justice contestations concentrated on (1) distribu-\ntive justice, (2) procedural justice and (3) justice as recognition (Fig. 3) \n(see definitions in Methods).\nWhereas the BERDO program may produce impacts at multiple \nscales, contestations about the distribution of benefits and burdens \ncentered primarily on building owners, tenants and environmental \njustice communities in Boston. CAG members generally advocated for \nremoving perceived regulatory loopholes that would enable building \nowners to avoid or delay compliance, rely on compliance mechanisms \nthat do not provide localized benefits or pass burdens onto tenants \nthrough increases in rent charges or energy bills. As explained in a \nBERDO \nadopted\nNov\nMar\nDec\nFeb\nJan\nApr\nMay\nJun\nFeb\nOct\nJan\nNov\nDec\nSep\nAug\nJul\nMar\nJul\nApr\nJun\nMay\nAug\nSep\nOct\nNov\n2022\n2023\nPublic technical working sessions\nMeetings with CAG\nMeetings with Boston residents\nPublic hearings of the BERDO Review Board\nFirst reporting and \nthird-party verification deadline\nSecond reporting\ndeadline\nOct \n2021\nPhase 1\nPhase 2\nPhase 3\nRule-making\nfinalizes\nRule-making\nbegins\nPublic listening sessions\nPublic hearings of the Boston Air Pollution Control Comission\nDec\n2023\nInformal public comment period\nfor regulations proposals\nFormal public comment period\nfor draft regulations\n*  Ad hoc meetings with regulated \n   parties were conducted throughout \n   the rule-making process\nReview Board\ncreated\na    BERDO rule-making process\nb    Key actors in the implementation of BERDO\nReview Board\nSix members nominated by \ncommunity-based organizations\nTwo members nominated by \nany individual or organization\nOne city councilor\n\u2022 Oversees implementation of BERDO\n\u2022 Issues penalties and fines\n\u2022 Makes funding decisions for the Equitable Emissions Investment Fund\n\u2022 Approves flexibility measures and sets conditions of approval\n\u2022 Recommends updates to regulations and policies \n\u2022 Comply with BERDO\n\u2022 Engage and provide public comments on rule-making\n\u2022 Nominate Review Board members\n\u2022 Provide direct advice on rule-making to city oficials\n\u2022 Engage and provide public comments on rule-making\n\u2022 Engage Boston residents in rule-making process\n\u2022 Nominate Review Board members as \n  community-based organizations\n\u2022 Nominate at least two-thirds \n  of the Review Board members\nCAG\nRegulated\nparties\nNon-profit\nafordable\nhousing\nSocial justice\nadvocacy\nEnvironmental justice\nadvocacy\nClimate advocacy\nUniversities\nReal estate\nHospitals\nCultural\ninstitutions\nOther building\nowners\nBoston\ncommunity-based\norganizations\nNominate\nNominate\nNominate\nFig. 1 | BERDO rule-making process and key actors. a, Rule-making consisted of \nthree phases. Phase 1 set rules for reporting and third-party verification. Phase 2 \nset rules related to the Review Board, emissions factors, compliance mechanisms \nassociated with renewable energy and other administrative regulations. Phase \n3 set rules related to flexibility measures, the Equitable Emissions Investment \nFund and penalties and fines. Symbols are used to represent public comment \nperiods and different engagement components that were part of the rule-\nmaking process. b, Key actors in the implementation of BERDO include the CAG, \nregulated parties and the Review Board. All CAG members are community-based \norganizations and some (non-profit affordable housing) are also regulated \nparties. Community-based organizations nominate at least two-thirds of the \nReview Board.\n\nNature Cities | Volume 1 | October 2024 | 665\u2013676\n668\nArticle\nhttps://doi.org/10.1038/s44284-024-00121-w\ncomment letter from CAG members: \u201cwe are concerned that the ben-\nefits of BERDO compliance (for example, building improvements, \nimproved air quality or jobs) may not adequately reach those in envi-\nronmental justice populations, and that the burdens of compliance \n(for example, financial costs or displacement) may be inequitably \nshouldered by those same populations\u201d.\nRegulated parties rarely advocated for the creation of additional \nbenefits or the distribution of existing benefits. Most of their requests \nwere related to providing flexibility for building owners and mini-\nmizing costs and regulatory burdens, often arguing that regulations \nwould \u201cunfairly\u201d, \u201cunduly\u201d or \u201coverly\u201d burden them. In some instances, \nregulated parties were reluctant to accept regulations that mandated \nbuilding owners to distribute benefits in the community. In public feed-\nback related to conditions of approval for flexibility measures, some \nregulated parties argued that distributing benefits such as housing \nand energy affordability, health and climate resilience \u201care outside the \nscope of emissions reductions in covered buildings\u201d and \u201cgo beyond \nthe intent of the BERDO ordinance\u201d. In other cases, regulated parties \nmobilized the idea of environmental justice benefits to advance regula-\ntory decisions that favor them. For example, in public comment letters \nregarding the Equitable Emissions Investment Fund, some regulated \nparties pushed for the prioritization of projects that produce larger \nemissions reductions as this \u201cis the greatest benefit to environmental \njustice and all populations\u201d. Such a decision is likely to favor applica-\ntions from larger carbon-intensive institutions as opposed to more \nholistic projects that benefit the community in other ways.\nContestations about procedural justice largely revolved around \nthe Review Board, which has substantial authority in the implemen-\ntation of BERDO (Table 2). Regulated parties sought to strengthen \ntheir representation in the Review Board by urging the city to select \nmembers with technocratic expertise and expand the definition of \n\u201ccommunity-based organizations\u201d (which can nominate two-thirds \nof the Review Board) to include the business community. Unsuccess-\nful in these efforts, regulated parties then advocated for regulations \nthat would limit the Review Board\u2019s discretion, give them access to \nthe Review Board through working groups or enable owners to easily \nappeal Review Board decisions.\nBy contrast, CAG members supported giving power to the Review \nBoard, although it proved challenging to balance maintaining the \nReview Board\u2019s discretion versus outlining decision-making processes \nwith prescriptive justice measures. CAG members also advocated for \ntenants and residents to have a voice in Review Board processes and \nregulatory decision-making, which led to additional meetings with \nBoston residents during rule-making. Whereas both regulated par-\nties and CAG members supported some monitoring and disclosure \nrequirements, only the CAG focused on tracking environmental justice \nmetrics and implementing outreach and education efforts beyond \nbuilding owners.\nContestations about distributive and procedural justice revealed \nunderlying debates about who is recognized as vulnerable and deserv-\ning of justice under BERDO. Although BERDO regulates all types of \nbuilding owners, CAG members largely focused on restricting the \nbehavior of landlords and large institutions, sometimes overlooking \nthe burdens placed on other building owners. Some CAG members \npushed for more nuanced discussions and highlighted how certain \nowners are also vulnerable. As stated in a CAG meeting: \u201cif we want \njustice, justice is not just displacing burdens on private owners\u201d. This \nstatement frames building owners as subjects of justice and challenges \nthe assumption that protecting tenants at the expense of owners is \nalways just. Whereas CAG members generally agreed on prioritizing \nbenefits to environmental justice populations, communities of color, \ntenants and low- and moderate-income residents, specifics about how \nto operationalize such priorities were often contested. For instance, \nwhen discussing the Equitable Emissions Investment Fund, some sug-\ngested that money should only be directed to environmental justice \nneighborhoods and were wary of including \u201cgeographic equity\u201d as \na part of the Review Board\u2019s funding criteria, fearing that this would \nchannel resources to wealthier neighborhoods. Others argued that \nregulations should recognize the differences between environmental \njustice neighborhoods, and that projects in \u201cwealthy\u201d communities can \nalso serve vulnerable residents. Reflecting on issues of recognition, a \ncity official commented: \u201cI have a specific responsibility to even the \nscales for folks at the margins because they\u2019ve historically been uncared \nfor, but that doesn\u2019t need to express itself as callousness or disregard to \nthose in the middle. [\u2026] In this role, I don\u2019t think it is right or ethically \njust for me to just not care about people. That is a practice of dehu-\nmanization and is a big problem within our movements too. I would \nlike us to get to the point where we lack the ability to demonize anyone\u201d.\nAlthough justice contestations featured prominently in regula-\ntions related to BERDO\u2019s justice components (Table 2), concerns about \njustice were also mobilized in regulatory topics that are not commonly \nassociated with justice and/or without an explicit justice mandate that \nemanates from the ordinance. For instance, concerns about asthma \nin children and air pollution in communities of color were mobilized \nby CAG members to support the assignment of an emissions factor to \nfossil fuel-derived district steam. This decision was opposed by some \nregulated parties, who argued that such rule places an \u201cunfair financial \nburden\u201d and \u201cdisproportionately burdens district steam customers\u201d. \nIn advocating to restrict the use of PPAs, CAG members argued that \ndespite existing challenges in renewable energy markets, \u201cwe can-\nnot allow the current reality to further entrench existing inequity \nat the expense of environmental justice populations\u201d. Conversely, \nregulated parties argued that relaxing third-party verification require-\nments to enable in-house data verification would be \u201ca key workforce \nTable 1 | BERDO compliance mechanisms\nCompliance \nmechanism\nDescription\nDirect emissions \nreductions in \nbuildings\nOwners may directly reduce emissions from \nelectricity and fossil fuel consumption through \nbuilding retrofitting and fuel switching. This includes \nimprovements such as upgrading to high-efficiency \nelectric appliances and lighting, the electrification of \nheating, cooling and cooking systems, insulation and \nbuilding-envelope improvements\nEnrollment into \nBoston\u2019s municipal \naggregation \nprogram\nOwners and tenants may reduce emissions from \nelectricity consumption by enrolling into Boston\u2019s \nCommunity Choice Electricity program, a municipal \naggregation program through which the City of Boston \npurchases Massachusetts (MA) Class I Renewable \nEnergy Certificates (RECs) on behalf of customers\nLocal renewable \nenergy generation\nOwners may reduce emissions from electricity \nconsumption through renewable energy generation \nlocated on-site (for example, rooftop solar) or off-\nsite (for example, community solar, Power Purchase \nAgreements (PPAs) in the ISO New England grid)\nEligible RECs\nOwners may mitigate emissions from electricity \nconsumptions by either purchasing and retiring \nunbundled MA Class I RECs or purchasing and retiring \nbundled MA Class I RECs as part of a PPA in the ISO \nNew England grid\nEligible PPAs\nOwners may mitigate emissions from electricity \nconsumption by entering into a long-term energy \ncontract with a generator of non-emitting renewable \nenergy that is located outside the ISO New England \ngrid. The PPA must meet an \u201cadditionality\u201d requirement, \nand associated RECs must be retired as part of the PPA\nAlternative \nCompliance \nPayments (ACPs)\nOwners may mitigate emissions from electricity and \nfossil fuel consumption through ACPs. ACPs are priced \nat US$234 for every metric ton of CO2-equivalent \nemitted above the emissions standards. ACPs are not \nfines, but rather a compliance pathway\n\nNature Cities | Volume 1 | October 2024 | 665\u2013676\n669\nArticle\nhttps://doi.org/10.1038/s44284-024-00121-w\ndevelopment opportunity\u201d that \u201cwould create a powerful incentive for \nlarge existing buildings to hire energy efficiency experts long term\u201d. \nThese examples show how justice contestations appear in seemingly \nopaque or technical decisions and beyond spaces that are explicitly \nlabeled for justice. As one city official explained, even where the ordi-\nnance does not specifically require it, \u201cregulations can still accommo-\ndate additional thoughts on what it means to advance equity goals\u201d.\nEven though CAG members and regulated parties were not always \nsuccessful in influencing implementation decisions via mobilizing \nideas of justice, justice contestations came to define how the City of \nBoston operationalized climate justice and shaped the implementation \nof BERDO itself. This could be seen in the evolution of the rule-making  \nprocess with the addition of meetings with Boston residents and  \nadditional public comment periods as a result of requests from the  \nCAG and regulated parties. Contestations were also reflected in the  \nfinal rule-making language, which was drafted and revised to resonate \nwith multiple, and sometimes conflicting, justice ideals that were \nmobilized throughout the process (Fig. 4).\nChallenges in translating climate justice into practice\nInterviewees identified several common barriers to policy implementa-\ntion in BERDO, including capacity and budget constraints, technical \ncomplexity, uncertainty and data quality/availability. However, we \nfound that additional distinctive challenges stemmed from the pursuit \nof implementing climate justice itself. We categorized these challenges \ninto (1) scope and scale discrepancies, (2) the bureaucratization of \njustice and (3) the weaponizing of justice.\nThe first challenge in implementing climate justice is that it \ninherently requires action at multiple scales and beyond the scope \nof a single program. As stated by an interviewee, \u201cthe ordinance was \nreflective of some community priorities and goals that would never \nreally be able to be addressed through just BERDO alone\u201d. In addi-\ntion to the legal constraints attached to any given program, scope \nand scale discrepancies partially originate from the multiplicity of \nmeanings of justice. For some, operationalizing justice in BERDO \nmeant implementing the program in a way that avoids harm and dis-\ntributes benefits inherently produced by building decarbonization. \nFor others, justice also meant \u201cincreasing the pot of benefits that are \non the table and then distributing those\u201d, potentially crossing the \nlegal scope of BERDO.\nThis issue was best exemplified in discussions about gentrifi-\ncation and displacement, with advocates persistently pushing for \nBERDO to include tenant protections. For context, the majority of \nBoston\u2019s households (65%) are renters34. Across Greater Boston, \n65% of Black residents and 70% of Latinx residents are renters (com-\npared with only 33% of White residents), and more than half of the \narea\u2019s renters are cost-burdened35. Symptoms of green gentrification \nhave already been reported in Boston and found to be associated \nwith urban greening, climate initiatives and redevelopment strate-\ngies36,37. With over 80% of Boston\u2019s census blocks considered to be \n\u201cenvironmental justice populations\u201d38, a great portion of BERDO \nbuildings are located in, and impact the lives of, multiple vulnerable \ncommunities (Supplementary Appendix A). In this landscape, the \nimplementation of BERDO may be perceived, rightly or wrongly, \nas another installment in patterns of historical marginalization. \nAlthough BERDO includes a goal that is related to housing justice \n(Table 2), municipalities in Massachusetts lack the authority to insti-\ntute rent control. Therefore, BERDO offered limited avenues for inte-\ngrating blanket rent stabilization measures and tenant protections \nduring rule-making. This example showcases how, even if a policy is \nexplicitly designed with a justice lens and where city staff and advo-\ncates agree on principles such as \u201cclimate justice is housing justice\u201d, \nindividual climate programs may not always have the scale or scope \nfor operationalizing those goals.\nScope and scale discrepancies were omnipresent throughout \nrule-making, with BERDO implementation serving as a battleground \nfor multiple issues that could not be directly or fully addressed through \nthis program, from evictions and slow permitting processes to the \nmarketing of renewable natural gas in the Northeastern US. This often \nled to tensions, and some CAG members felt that they were unable to \n\u201cset the agenda\u201d during rule-making (Fig. 2). Some members explained \nthat \u201cmaking the regulations was not always in line with what the group \nwanted to talk about\u201d and \u201cthe response that \u2018we can\u2019t do anything \nabout rent control because that\u2019s a state issue\u2019 is really unsatisfying \nfor people and can be really disempowering\u201d. Such frustrations were \nillustrated in a statement during a public meeting: \u201cwe keep talking \nabout buildings and not people\u201d.\nThe second challenge in translating climate justice into practice is \nthat it necessitates the bureaucratization or standardization of justice \ngoals into concrete processes, measures or criteria. This creates the \nrisk of reducing justice to box-checking or scoring exercises that do \nnot fully reflect justice ideals. As one city official explained \u201c[Before] \nI understood environmental justice as a way to dismantle power and \ndistribute power, [but] to write policy, you have to write environmental \nTable 2 | Key justice components embedded in the BERDO \nprogram\nComponent\nDescription\nGoal\nTo \u201creduce the emissions of air pollutants, including \ngreenhouse gases, from building energy production and \nconsumption, and thereby to encourage efficient use of \nenergy and water, develop further investment in building a \ngreen economy, including by encouraging the hiring and \ntraining of green jobs, protect public health, and promote \nequitable access to housing\u201d\nReview Board\nA nine-member independent board that provides \ncommunity oversight over the implementation of BERDO. \nTwo-thirds of the Review Board (six members) must be \nnominated by community-based organizations. One \nseat is reserved by the Chair of Boston City Council\u2019s \nEnvironmental Justice, Resiliency, and Parks Committee. \nTwo seats may be nominated by anyone. The Review \nBoard has the authority to make funding decisions for the \nEquitable Emissions Investment Fund, approve and set \nconditions for flexibility measures requested by owners, \nenforce the ordinance, issue penalties and fines, propose \nupdates to emissions standards and the price of ACPs, \nand recommend revisions to regulations and compliance \nmechanisms\nEquitable \nEmissions \nInvestment \nFund\nA special purpose fund that collects all ACPs (Table 1) \nand fines made pursuant to BERDO. The Review Board \nmakes funding decisions, provided that the fund must \nbe used to support local building carbon abatement \nprojects in Boston and must prioritize projects that benefit \nenvironmental justice communities and populations who \nare disproportionately affected by air pollution\nFlexibility \nmeasures with \nconditions of \napproval\nOwners may apply to obtain flexibility in complying \nwith emissions standards. Flexibility measures must be \napproved by the Review Board, and the Review Board may \nset conditions of approval, including conditions related to \nenvironmental justice. Flexibility measures include:\n\u2022 \u0007Building Portfolios, which enable owners to comply with \na single emissions standard across a group (\u2018portfolio\u2019) \nof buildings that share the same owner or Institutional \nMaster Plan (a development plan that is approved by the \nBoston Planning & Development Agency)\n\u2022 \u0007Individual Compliance Schedules, which enable owners \nto request an alternative emissions reduction based on \na baseline year. Individual Compliance Schedules must \nestablish absolute emissions limits that decline every \nfive years on a linear basis or better. On the basis of their \nselected baseline year, owners must achieve a 50% \nreduction in absolute emissions by 2030 and a 100% \nreduction by 2050\n\u2022 \u0007Hardship Compliance Plans, which enable owners to \nrequest alternative emissions standards and/or emissions \nreduction schedules if facing an eligible hardship in \ncomplying with the default emissions standards\n\nNature Cities | Volume 1 | October 2024 | 665\u2013676\n670\nArticle\nhttps://doi.org/10.1038/s44284-024-00121-w\njustice in a measurable, actionable way, [\u2026] which it\u2019s not the way the \npeople really talk about it\u201d. This can be seen in environmental jus-\ntice assessments that rank initiatives based on a limited set of justice  \nmetrics or in definitions and indexes of environmental justice com-\nmunities, which, despite compiling multiple socio-demographic vari-\nables, may not always accurately reflect who the local residents actually  \nperceive as vulnerable within their city. Justice is in many ways impervi-\nable to codification. As stated by a city official: \u201cI don\u2019t think you can \n100% ever codify justice. There are some things like love, which I believe \nhas to be a critical component to how we live in the world, and how \nam I supposed to write that? How do we do this with love and grace?\u201d.\nThe bureaucratization of justice is not only driven by the state\u2019s \nneed to standardize rules and processes. Rather, it can be purposefully \nor inadvertendly promoted by myriad actors. For instance, bureauc-\nratization can come from regulated parties who are seeking to limit \njustice and demanding to constrain justice-oriented decision-making \n\u201cI think that the frustration of the group was that [the process] was on the timeline of having to \npromulgate the regulations, instead of stepping back and saying to the group \u2018what are the \nmajor issues that you want to talk about?\u2019\u201d\n\u201cI got the impression that some people on the CAG really wanted [the process] to be more \nbottom-up than it was. But it was much more bottom-up than other people even tried to do, and \nI think it became more bottom-up as people gave feedback about how the meetings were going\u201d\nTop-down\n\u201cIt was really an experiment on how to have policy-making and a regulatory process be more collaborative\u201d\n\u201cThe fact that we were able to bring in the perspectives of people who otherwise would not have had \ntheir feedback heard, whether or not it was incorporated, is important and meaningful\u201d\n\u201cThe fact that the city engaged a group in this advisory board was unique\u201d\nNew and unique\n\u201cI think the city was very open and willing to learn\u201d\n\u201cI found [the process] to be very thoughtful and intentional\u201d\n\u201cI love that the city made a huge efort to include people in every community, and it's really open and \nwilling to talk to people.\u201d\n\u201cI think [the city] really make an efort to make sure that stakeholders in EJ communities fully participate \nin it, and I can say that we try our best to fully participate\u201d\n\u201cI found the CAG process to be very organized. The fact that there was a repository of documents [...], \nthe determination of the city oficials to be present at the meetings [...], they had ofice hours. \nIf you wanted to talk to someone, you could talk to someone\u201d\nOpen and intentional\n\u201cThis whole process to get community input was very structured and clear\u201d\n\u201cThis is a great model and other cities should do that\u201d\n\u201cEngagement is, I think, the most important thing that has happened\u201d\n\u201cI can see my participation in the regulations\u201d\nImpactful\n\u201cWe can see in the regulations, where the community voice has been integrated. Did we get to community \nownership? I don't think so, but I think this is a much better place than before\u201d\n\u201cAs an advisory group member, it was challenging. It's just so much information. I think we really \nwould have benefited from some in-person longer session meetings\u201d\n\u201cThe city heard things that we didn't hear from other people, and they also had to make decisions \nthat are based on things that we may not have been aware. So that felt frustrating\u201d\n\u201cOne of the biggest challenges that I felt throughout the process was being able to break down \nand explain what a lot of these highly technical policies mean and what they would actually \nlike result in\u201d\nComplex\n\u201cI know that it's easier said than done when it comes to public processes, but even more time \nwould have been better. You got to go slow to go fast and build trust\u201d\n\u201cI think that the city did the best it could in the situation [...] you've got to be moving things along \nat a pace, whereas the group didn't always feel informed or understanding what was going on\u201d\n\u201cWe were under a tight timeline to come up with the content of the regulations\u201d\n\u201cIt was a rushed process because of its timeline\u201d\nQuick\n\u201cIt was a very collaborative process\u201d\n\u201cI didn't feel like I was working against the city in any way. I really felt like we were working together and \nthat we wanted the same things\u201d\nCollaborative\n\u201cI never felt there was anybody trying to hide anything. It was pretty transparent\u201d\nFig. 2 | Perceptions of CAG process. Summary of perceptions from members about the process and approach of the CAG. Data come from interviews with CAG \nmembers. EJ, environmental justice.\n\nNature Cities | Volume 1 | October 2024 | 665\u2013676\n671\nArticle\nhttps://doi.org/10.1038/s44284-024-00121-w\nHealth and\nquality of life\nBuilding owners\nBuilding tenants\nEnvironmental justice communities\nBoston residents\nCommunities outside Boston\nand near power plants and\ngas infrastructure\nEnergy justice\n(reduced energy\nburdens, access to\nheating and cooling)\nNew green jobs and\nworkforce development\nClimate change\nmitigation\nGlobal community\nDirect compliance costs\nand regulatory burdens\nEnergy injustice\n(increased energy\nburdens)\nGentrification \nand displacement\nWho should benefit and how? \nWho should be protected from burdens and how?\nWhat benefits \nshould be distributed?\nWhat burdens should be\nminimized or avoided?\nDefinition of community-based organization\nOpportunity to convene a\nReview Board working group\nOpportunity to provide public comment\nfor Review Board decisions\nOpportunity to appeal\nReview Board decisions\nProcedural justice: contestations about power, representation and accountability\nWho can nominate Review Board members?\nWho has access to the Review Board and\ntheir decision-making processes and how?\nHow much discretion should the Review Board have?\nRequirements or guidelines to make decisions\nrelated to flexibility measures\nReview Board\nWho is represented in the Review Board?\nNominated by community-based organizations\nNominated by any individual or institution\nCity councilor, Chair of the Environmental Justice,\nResiliency, and Parks Committee\nRequirements related to expertise \nand/or sector representation\nRequirements or guidelines to make decisions\nrelated to the Equitable Emissions Investment Fund\nDistributive justice: contestations about the distribution of benefits and burdens\nWho is the Review Board and the\nCity of Boston accountable to and how?\nMonitoring and disclosure requirements\nEducation and outreach\nJustice as recognition: contestations about who is recognized as vulnerable and deserving of justice\nBuilding owners\nEnvironmental\njustice communities\nBuilding tenants\nLanlords\nLarge institutions\nLow-income owners\nModerate-income owners\nCondominium owners\nSmall businesses\nCultural institutions\nAny owner\nLow-income tenants\nModerate-income tenants\nSmall-business tenants\nAny tenant\nOficially recognized environmental justice \nneighborhoods\nVulnerable residents in any neighborhood\nFig. 3 | Justice contestations in the rule-making process of BERDO. Different actors framed and mobilized multiple contestations around ideals of justice, equity \nand fairness to advance their interests and shape the implementation decisions. Contestations focused on issues of distributive justice, procedural justice and justice \nas recognition.\n\nNature Cities | Volume 1 | October 2024 | 665\u2013676\n672\nArticle\nhttps://doi.org/10.1038/s44284-024-00121-w\nto \u2018objective\u2019 or \u2018data-based scoring systems\u2019. It can also come well-\nintentioned advocates who are seeking to further justice but pushing \nfor blanket justice requirements that may serve a purpose in some \ncontexts but are ill-fitting in others. Both approaches bureaucratize \njustice, either by forcing its objectification and quantification or by \nreducing it to for-the-sake-of-it mandates.\nAdditional challenges emerge when justice is weaponized to stall, \nimpede or co-opt implementation. Throughout rule-making, some CAG \nmembers and regulated parties pushed for implementation delays by \nmobilizing ideas of justice and fairness. Some interviewees questioned \nthe value of such hold-ups, which can expose policies to political tur-\nmoil, put them at risk of being legally challenged and ignore that \u201cwe \nalso have a deadline from a climate perspective\u201d. On the one hand, just \nimplementation does require time to enable meaningful engagement \nof community voices, particularly those who have been historically left \nout. As stated by a CAG member: \u201csome of us have not historically been \nasked to be involved in these types of groups, so even giving people \ntime to just mature into those roles requires more time\u201d. On the other \nhand, claims of procedural justice can also be strategically misused \nby some actors to impede implementation. A city official explained: \n\u201cyou have some people that are accusing [Boston] of going too slow \nbecause of equity, and other people who, from the equity perspective, \nare accusing [Boston] of going too fast\u201d.\nBeyond implementation delays, city staff and CAG members \nexpressed concerns that flexibility measures and funding opportuni-\nties created for under-resourced building owners could be exploited \nby large institutions with enough resources to navigate the system and \nmake their case to the Review Board. A CAG member warned: \u201csome of \nthe exceptions could swallow the good intentions of the ordinance\u201d. \nRegulations also needed to balance the ideals of procedural justice \nwith the risk of public participation being perversely used to challenge \nReview Board decisions, either to advance NIMBY (\u2018not in my back yard\u2019) \nclaims or to redress past issues with building owners. A city official \nexplained: \u201c[BERDO] isn\u2019t your opportunity to right wrongs that you \nperceive as being done. Everybody has to be treated in the same way for \nconsistency [\u2026] so that no single decision can be picked apart under \nthe eyes of the law and overturned\u201d.\nDiscussion\nOur analysis shows that contestations about justice have formed an \ninextricable component of the implementation of BERDO. Ideals of \njustice, equity and fairness were mobilized by all actors to advance \ntheir interests and influence implementation decisions. These contesta-\ntions impacted how BERDO was operationalized on the ground, both in \nterms of the process (for example, the engagement of residents in rule-\nmaking) and the resulting implementation decisions (for example, the \nadopted regulatory language). This was the case for many regulatory \ndecisions, including seemingly technocratic topics without explicit \njustice mandates. At the same time, policy implementation served as \na site for contesting and reinterpreting the ordinance\u2019s justice goals \nand mandates, ultimately redefining what justice means and who is \ndeserving of justice in the context of BERDO.\nThe implementation of BERDO illustrates that the politics of  \nclimate justice can transcend the stage of policy planning and be used \nas a tool by which to ensure, transform or impede implementation. This \nsupports previous theories suggesting that climate policy is shaped by \ncontestations over justice39 and that the ways in which justice is under-\nstood in a particular place are critical determinants of how climate \nprograms come to be developed and implemented2,39,40. Rather than \nsettling on a single definition of justice, the multiplicity of meanings \nof justice under BERDO resulted in regulations that sought to resonate \nwith multiple, and sometimes conflicting, justice ideals that were \nmobilized by different actors. This reveals climate justice not only as \na principle of climate planning or as a policy outcome, as often treated \nin the literature14,15, but also as a highly disputed political process in \nwhich competing ideals of justice are contested and translated into \nimplementation decisions.\nOur research highlights important caveats for the implementation \nof climate justice. First, whereas visions of climate justice in the city, \nsuch as the \u201cGreen New Deal\u201d and \u201cjust urban transitions\u201d, necessitate \nmulti-sectoral and multi-level governance approaches14, implementing \nclimate justice through a single climate program is inevitably subject to \nconstraints of scope and scale. In that regard, community-driven imple-\nmentation approaches such as the CAG and Review Board must begin \nwith a shared understanding of the scope and scale of the program at \nJustice \ncontestations\nDistribution\nProcedure\nRecognition\nCAG\nRegulated parties\nCity of Boston\nThe right for building tenants to access the \nReview Board and have power in \ndecision-making processes\nEnsuring that flexibility measures do not\nprevent environmental justice communities\nfrom receiving the benefits of BERDO\nExamples of justice contestations\nRecognizing vulnerable building owners\nand building tenants\nAdding requirements and conditions related to environmental justice:\nCertain building portfolios will require \u201cplans to prioritize distribution of\nbenefits associated with BERDO compliance in buildings in the building\nportfolio that are located in environmental justice populations and\nafordable housing\u201d\nTo approve hardship compliance plans, \u201con a case-by-case basis, the\nReview Board may include special conditions relevant to the distribution\nof benefits to environmental justice populations and advancing the\npurpose of the ordinance\u201d\nCreating avenues for tenants to communicate with the Review Board and\nparticipate in decision-making processes:\n\u201cThe Review Board shall hold one meeting per year dedicated to hear\nconcerns raised by tenants of residential buildings covered by the\nordinance and provide information directed towards tenants of said\nbuildings\u201d\n\u201cThe Review Board shall initiate a proceeding to evaluate a petition from\nthe greater of twenty percent of tenants or five tenants of a building\nincluded in a building portfolio to terminate the building portfolio\u201d\nCreating avenues for vulnerable building owners to access relief\nthrough flexibility mechanisms and for building tenants to\nbenefit from funding from the Equitable Emissions Investment Fund:\n\u201cLow-income owners of building(s) that provide afordable housing \nto low-income tenants\u201d are explicitly called out as one of circumstances \nand characteristics for accessing hardship compliance plans\n\u201cBenefits to afordable housing and tenant protections\u201d are one of the\nevaluation criteria that the Review Board shall use to make funding\ndecisions for the Equitable Emissions Investment Fund\nExamples of regulatory language reflecting justice contestations\nFig. 4 | Examples of the impact of justice contestations in rule-making. Examples of how justice contestations are reflected in regulatory language.\n\nNature Cities | Volume 1 | October 2024 | 665\u2013676\n673\nArticle\nhttps://doi.org/10.1038/s44284-024-00121-w\nhand. This is not to say that policies should be implemented in a vacuum \nor avoid challenging existing power relations and siloing practices \nwithin a city. However, a mismatch between the goals and the reality \nof the legal boundaries of implementating climate action can frustrate \nthe operation of programs and weaken trust between city officials, \nadvocates and those who they aim to protect.\nSecond, the implementation of climate justice will inherently \ninvolve a degree of \u201cadministrative ordering of nature and society\u201d41. \nImplementation requires rendering justice legible through bureau-\ncratic tools that, for instance, standardize environmental justice com-\nmunities into a legal definition or simplify justice into a checklist  \nof regulatory requirements or a series of benefits to be distributed. \nOn the one hand, this bureaucratization of justice can result in what \nwe call \u201cordinary innovations\u201d\u2014small but meaningful changes that \nreimagine policy tools and governing practices to steer cities toward \nthe social, ethical and political decisions that are needed for just urban \ntransitions.\nOn the other hand, however sophisticated, bureaucratic tools are \n\u201cprojects of legibility\u201d41 that are never fully realized. Intangibles such as \ncare, compassion and understanding for others cannot be fully codified \nin artifacts such as regulations and will need to evolve into everyday \npractices and \u201cquiet acts\u201d42 of justice through the mundane decisions \nthat implementors such as city officials and Review Board members \nwill take moving forward. As stated by a city official: \u201cas much as it is \ncodified in the ordinance and supported in regulations, environmental \njustice is a practice as well\u201d. This emphasizes how transformative action \ntakes place not only through grand interventions but also through \nmundane and \u201cindividually smaller actions\u201d that can collectively shift \nsystems over time43,44. These everyday practices often occur in \u201cthe \nmiddle space\u201d between institutional and community action and, as \nwith BERDO rule-making, are neither purely top-down nor bottom-up \nefforts43.\nFinally, the prospective weaponizing of justice highlights the \nneed for a critical lens in the implementation of justice. Policymakers \nshould interrogate who is mobilizing justice claims, to what extent and \nwhy, and question when, how and who may misuse or co-opt justice \nmeasures. Well-intentioned environmental laws have been perversely \nused to block projects that climate policies seek to promote. That has \nbeen seen with lawsuits under the California Environmental Quality \nAct, which have been weaponized to advance economic agendas from \nthe private sector and put forward NIMBY claims against housing \nand greening projects that will serve people of color and diversify \ncommunities45. Under the National Environmental Policy Act, Indig-\nenous justice concerns have similarly been co-opted by elite groups \nto oppose offshore wind development, thereby perpetuating colonial \nrelations46. Even community-led initiatives can be \u201clater discovered to \ncreate more problems, more injustices\u201d, with policies often focused \non finding the \u201cpath of least resistance\u201d in the near term rather than \nenvironmentally sound and just solutions in the long term47. New \nclimate justice measures run the same risk of exacerbating environ-\nmental and social vulnerabilities if co-opted by actors to advance \ntheir own interests or used in bad faith or in ways that undermine \nparticipatory processes.\nFurthermore, our research highlights what is lost when policy \nimplementation is omitted from studies of climate governance or \nwhen it is over-simplified into combinations of barriers and enabling \nfactors, without recognition of the inherent political complexities of \nthe policy process. As cities increasingly integrate justice into climate \naction7\u20139,48, scholars, advocates and practitioners should look at policy \nimplementation as a key site of political contestation and resistance, \nwhere opposing conceptions of justice are fought out and translated  \ninto action. Future scholarship should examine climate justice  \nnot only as a goal or outcome of policies but also as a set of mundane \npractices and \u201cquiet\u201d42 acts of care and resistance that continuously \nunfold throughout the implementation of climate action.\nMethods\nWe use a mixed-methods approach that combines participant obser-\nvation, semistructured interviews and content analysis to examine  \nthe politics and process of implementing urban climate justice  \nthrough BERDO.\nDefining urban climate justice\nDrawing from existing scholarship, we define urban climate justice \nas a concept and social movement that (1) recognizes the inequita-\nble impacts of climate change in cities49,50, (2) acknowledges that \nclimate change is driven by the historical and structural processes \nof environmental racism, settler colonialism, heteropatriarchy and \nracial capitalism, all of which structure human\u2013environment inter-\nactions51\u201354, (3) highlights the inequitable impacts of urban climate \naction39,55,56 and (4) advances the pursuit of justice through climate \naction in cities26,39.\nThe study of urban climate justice has evolved from and alongside \nenvironmental justice scholarship1,49,57\u201360. This literature has docu-\nmented how communities of color, Indigenous communities and \nhistorically marginalized populations have been disproportionately \nexposed to environmental hazards, denied environmental benefits \nand excluded from decision-making processes. Environmental justice \nscholars have also exposed the connections between environmental \nvulnerabilities and structural issues such as environmental racism, \nwhite supremacy, settler colonialism and heteropatriarchy47,49,50,61\u201364. \nClimate justice expands on this scholarship by analysing how these \ninequities and structures are also manifested through climate change \nand climate policy across geographies and at multiple scales1,49,57.\nAlthough justice has been conceptualized in several distinct ways, \nwe understand justice to be composed of three dimensions or tenets: \n(1) distributive justice, (2) procedural justice and (3) justice as recog-\nnition. Distributive justice refers to the fair allocation of the benefits \nand burdens of climate change and climate policy39,65. Procedural jus-\ntice refers to inclusive participation, engagement, transparency and \naccountability in decision-making processes39,63,65,66. Finally, justice as \nrecognition refers to the respect and valuing of all people in climate \ngovernance and requires the acknowledgement of historic and ongoing \ninequities and the pursuit of efforts to reconcile these inequities39,65,67. \nSome scholars also add \u201crestorative justice\u201d to highlight the need for \nthe healing, reconciliation and rebuilding of relationships, communi-\nties and the environment63,68.\nOne additional conceptual clarification underlies our understand-\ning of urban climate justice. There is no single agreed definition of \n\u201curban\u201d or \u201cthe city\u201d in the literature. Different scholars delineate cities \nbased on population size, population density, political boundaries, \nboundaries of mass transit systems, percentages of vegetation and \nimpervious surface area and the residents\u2019 own perception and expe-\nrience of place3,69. Here, we consider urban to be \u201can area with legally \ndefined boundaries with recognized urban status and their own local \ngovernment\u201d3.\nDefining policy implementation\nWe understand implementation as the process of translating public \npolicies into operational and enforceable programs70. We primarily  \ninvestigate implementation through the process of rule-making. \nThrough rule-making, city governments develop and issue specific \nregulations that establish rules and parameters by which to implement \nand enforce a policy71. In this way, rule-making is one of the first and \nmost critical steps of the policy implementation process. Rule-making \nis also a critical site to analyse the operationalization of climate justice \non the ground. Rule-making is where issues of politics and power most \nclearly intersect with policy implementation by providing an arena for \ncity governments and other actors to contest and reinterpret the justice \ngoals and mandates that were already embedded in climate policies \nduring the policy planning process72.\n\nNature Cities | Volume 1 | October 2024 | 665\u2013676\n674\nArticle\nhttps://doi.org/10.1038/s44284-024-00121-w\nParticipant observation\nC.V.D. conducted participant observation by working as a Policy  \nFellow at the City of Boston\u2019s Environment Department. In this role, \nC.V.D. actively participated in the implementation of BERDO, with a \nparticular emphasis on the rule-making process, which involved work-\ning directly with city staff (1) to engage with community advocates \n(the CAG), Boston residents and regulated parties throughout the \nregulations process; (2) to prepare materials for public meetings; (3) to \nreview and analyse public feedback; (4) to draft and revise the regula-\ntions language; and (5) to support other implementation activities as \nneeded. More than 580\u2009h of participant observation were carried out \nover a period of 5 months, from June to December 2023. This period \ncovered most of Phase 3 of the rule-making process (Fig. 1). During \nparticipant observation, detailed notes were taken that focused on \nthe justice themes that emerged during rule-making, whether and how \ndifferent actors articulate, contest and reinterpret justice concerns, \nand how the said concerns were translated (or not) into a specific \nregulatory or implementation decision. All notes were anonymized \nand transcribed into a digital format for analysis. Participant observa-\ntion provided us with an in-depth insight into the social, cultural and \npolitical context in which BERDO unfolds, and enabled us to directly \ntrack and experience the process through which the justice goals and \nmandates embedded in BERDO were contested, reinterpreted and \nultimately translated into specific regulatory and implementation \ndecisions. Participant observation also enabled us to add nuance \nto the data collected through interviews and content analysis and \nfacilitated the identification of \u201csubtleties of meaning\u201d73 among city \nstaff and community advocates.\nInterviews\nWe conducted 20 semistructured interviews with city staff, commu-\nnity advocates and leaders involved in the implementation of BERDO. \nInterviews are well suited for tracing the chronology of events and \nmovement of policy ideas74 and for building an understanding of how \n\u201ccertain events, practices, or knowledges are constructed and enacted \nwithin particular contexts\u201d75, making them effective tools with which to \nanalyse both the history and politics of policy implementation and situ-\nated views on climate action and justice. We used the semistructured \ninterview approach to ensure that all interviews covered key topics, \nwhile allowing a conversational approach that enables respondents to \ntell their own stories and for new topics to emerge. Whereas the exact \nwording and order of questions were tailored to each respondent, all \ninterviews consisted of a series of open-ended questions that solicited \ninformation about the respondent\u2019s role, influence and perceptions on \nthe BERDO implementation process. This included the respondents\u2019 \nperceptions about the justice implications of BERDO, the successes \nand challenges of the implementation process, justice concerns that \nhave been addressed during implementation, justice concerns that \nremain unaddressed, and opportunities and challenges for the future \nimplementation of BERDO. Sample interview scripts can be found in \nSupplementary Appendix B.\nRespondents were identified through participant observation. \nOur final interview sample included ten city staff involved in the rule-\nmaking process and/or other implementation activities for BERDO \nand ten community advocates who were part of the CAG during rule-\nmaking. Some respondents had been involved with BERDO since the \npolicy planning process, whereas others became involved at different \nphases of the rule-making process and implementation of BERDO. \nInterviews ranged between 30 and 65\u2009min, but most lasted 50\u2009min. \nInterviews were conducted in person or over Zoom. All interviews were \naudio recorded and transcribed for analysis.\nQuotes from interview respondents are identified in the main text \nas coming from a \u201ccity official\u201d, \u201ccity staff\u201d or \u201cCAG member\u201d. The term \n\u201cinterviewee\u201d is also used sporadically to grant additional anonymity \nto the respondents.\nContent analysis\nWe conducted a content analysis of relevant policy documents, records \nfrom public meetings and public comments related to the rule-making  \nprocess. We used these data to complement and triangulate our  \nanalysis from participant observation and interviews, and to confirm \nthe elements of the regulations that were explicitly linked to equity and \njustice at different stages of the rule-making process. We gave special \nattention to justice controversies related to the implementation of \nBERDO and the role and positions of regulated parties, community advo-\ncates and other actors in proposing any specific regulations language \nor implementation decisions and strategies. This analysis included the \nfinal regulatory language adopted through the rule-making process, \nthe minutes and materials from 15 public hearings of the Boston Air \nPollution Control Commission, minutes and materials from 12 public \nhearings of the Review Board, minutes and materials from 11 CAG meet-\nings, minutes and materials from 12 public meetings held by the City of \nBoston\u2019s Environment Department, 134 public comment letters received \nas part of the rule-making process, and nine documents that included \ncity staff responses to public comment letters. A list of public meetings \nincluded in this analysis can be found in Supplementary Appendix C.\nData analysis\nWe coded all participant observation, interview and content analysis \ndata using an iterative qualitative process of inductive coding. We first \ncoded the data according to emergent themes revealed by each source \nindependently (that is, participant observation notes, interview tran-\nscripts and policy documents). We then conducted multiple iterative \nrounds of focused coding to homogenize our analysis across all data \nsources. The final coding protocol included the following themes:  \n(1) justice contestations; (2) the implementation process; (3) implemen-\ntation decisions and outcomes; and (4) implementation challenges. \nThe final coding protocol can be found in Supplementary Appendix D.  \nWe used NVivo 20 software for all coding.\nEthics and confidentiality\nThis research was approved by the Boston University Institutional \nReview Board (Exempt Research number 6907X) and complies with all \nrelevant ethical regulations. City staff and community advocates that \nwere part of the CAG received a letter explaining the goals and scope of \nthe research project before the beginning of the participant observa-\ntion process. All interview respondents received a letter of informed \nconsent before participating in an interview. Participants were offered \nno compensation. We took all reasonable measures to protect the \nconfidentiality of participants, which included reporting the findings \nfrom participant observation and interview responses anonymously.\nReporting summary\nFurther information on research design is available in the Nature \nPortfolio Reporting Summary linked to this article.\nData availability\nParticipant observation, interview transcripts and analysed data are \nnot publicly available because they contain information that would \ncompromise the research participants\u2019 confidentiality and undermine \nthe process of informed consent.\nCode availability\nNo custom algorithms or code were used in the collection or analysis \nof the data. All data were analysed using NVivo 20 software.\nReferences\n1.\t\nHughes, S. & Hoffmann, M. Just urban transitions: toward a research \nagenda. Wiley Interdiscip. Rev. Clim. Change 11, e640 (2020).\n2.\t\nBulkeley, H. Climate changed urban futures: environmental politics \nin the Anthropocene city. Environ. Polit. 30, 266\u2013284 (2021).\n\nNature Cities | Volume 1 | October 2024 | 665\u2013676\n675\nArticle\nhttps://doi.org/10.1038/s44284-024-00121-w\n3.\t\nSovacool, B. K. & Brown, M. A. Twelve metropolitan carbon \nfootprints: a preliminary comparative global assessment. Energy \nPolicy 38, 4856\u20134869 (2010).\n4.\t\nStehle, T., Hickmann, M. L. & H\u00f6hne, C. Urban climate politics \nin emerging economies: a multi-level governance perspective. \nUrbanisation 7, 9\u201325 (2022).\n5.\t\nHsu, A., H\u00f6hne, N., Kuramochi, T., Vilari\u00f1o, V. & Sovacool, B. K.  \nBeyond states: harnessing sub-national actors for the deep \ndecarbonisation of cities, regions, and businesses. Energy Res. \nSoc. Sci. 70, 101738 (2020).\n6.\t\nBulkeley, H. Cities and the governing of climate change.  \nAnnu. Rev. Environ. Resour. 35, 229\u2013253 (2010).\n7.\t\nDiezmart\u00ednez, C. V. & Short Gianotti, A. G. US cities increasingly \nintegrate justice into climate planning and create policy tools for \nclimate justice. Nat. Commun. 13, 5763 (2022).\n8.\t\nCannon, C., Chu, E., Natekal, A. & Waaland, G. Translating and \nembedding equity-thinking into climate adaptation: an analysis  \nof US cities. Reg. Environ. Change 23, 30 (2023).\n9.\t\nCaggiano, H., Kocaku\u015fak, D., Kumar, P. & Tier, M. O. U.S. cities\u2019 \nintegration and evaluation of equity considerations into climate \naction plans. npj Urban Sustain. 3, 50 (2023).\n10.\t Bulkeley, H. & Betsill, M. Rethinking sustainable cities: multilevel \ngovernance and the \u2018urban\u2019 politics of climate change. Environ. \nPolit. 14, 42\u201363 (2005).\n11.\t\nvan der Heijden, J. When opportunity backfires: exploring the \nimplementation of urban climate governance alternatives in three \nmajor US cities. Policy Soc. 40, 116\u2013135 (2021).\n12.\t Keskitalo, E. H. C., Juhola, S., Baron, N., Fyhn, H. & Klein, J. \nImplementing local climate change adaptation and mitigation \nactions: the role of various policy instruments in a multi-level \ngovernance context. Climate 4, 7 (2016).\n13.\t Woodruff, S. C. & Stults, M. Numerous strategies but limited \nimplementation guidance in US local adaptation plans. Nat. Clim. \nChange 6, 796\u2013802 (2016).\n14.\t Westman, L. & Cast\u00e1n Broto, V. Transcending existing paradigms: \nthe quest for justice in urban climate change planning. Local \nEnviron. 26, 536\u2013541 (2021).\n15.\t Hughes, S. The politics of urban climate change policy: toward a \nresearch agenda. Urban Aff. Rev. 53, 362\u2013380 (2017).\n16.\t Ryan, D. From commitment to action: a literature review on \nclimate policy implementation at city level. Clim. Change 131, \n519\u2013529 (2015).\n17.\t Laurian, L. & Crawford, J. Organizational factors of environmental \nsustainability implementation: an empirical analysis of US cities \nand counties. J. Environ. Policy Plan. 18, 482\u2013506 (2016).\n18.\t Malloy, J. T. & Ashcraft, C. M. A framework for implementing \nsocially just climate adaptation. Clim. Change 160, 1\u201314 (2020).\n19.\t Eisenack, K. et al. Explaining and overcoming barriers to climate \nchange adaptation. Nat. Clim. Change 4, 867\u2013872 (2014).\n20.\t F\u00fcnfgeld, H., Fila, D. & Dahlmann, H. Upscaling climate change \nadaptation in small- and medium-sized municipalities: current \nbarriers and future potentials. Curr. Opin. Environ. Sustain. 61, \n101263 (2023).\n21.\t Burley Farr, K., Song, K., Yeo, Z. Y., Johnson, E. & Hsu, A. Cities \nand regions tackle climate change mitigation but often focus  \non less effective solutions. Commun. Earth Environ. 4, 439 \n(2023).\n22.\t Sovacool, B. K. Who are the victims of low-carbon transitions? \nTowards a political ecology of climate change mitigation. Energy \nRes. Soc. Sci. 73, 101916 (2021).\n23.\t Carley, S. & Konisky, D. M. The justice and equity implications of \nthe clean energy transition. Nat. Energy 5, 569\u2013577 (2020).\n24.\t Anguelovski, I. et al. Equity impacts of urban land use planning \nfor climate adaptation: critical perspectives from the global north \nand south. J. Plan. Educ. Res. 36, 333\u2013348 (2016).\n25.\t Hardy, R. D., Milligan, R. A. & Heynen, N. Racial coastal formation: \nthe environmental injustice of colorblind adaptation planning for \nsea-level rise. Geoforum 87, 62\u201372 (2017).\n26.\t Bulkeley, H., Carmin, J. A., Cast\u00e1n Broto, V., Edwards, G. A. S. & \nFuller, S. Climate justice and global cities: mapping the emerging \ndiscourses. Global Environ. Change 23, 914\u2013925 (2013).\n27.\t Biesbroek, R. et al. Opening up the black box of adaptation \ndecision-making. Nat. Clim. Change 5, 493\u2013494 (2015).\n28.\t Luque-Ayala, A., Bulkeley, H. & Marvin, S. in Rethinking Urban  \nTransitions: Politics in the Low Carbon City (eds Luque-Ayala, A. et al.) \n13\u201336 (Routledge, 2018); https://doi.org/10.4324/9781315164779-2\n29.\t Salimifard, P. et al. Climate policy impacts on building energy use, \nemissions, and health: New York City local law 97. Energy 238, \n121879 (2022).\n30.\t Walter, T. & Mathew, P. GHG policy impacts for Seattle\u2019s buildings: \ntargets, timing, and scope. Build. Cities 2, 283\u2013301 (2021).\n31.\t 2022 Global Status Report for Buildings and Construction: Towards \na Zero\u2011emission, Efficient and Resilient Buildings and Construction \nSector (United Nations Environment Program, 2022).\n32.\t Sethi, M., Lamb, W., Minx, J. & Creutzig, F. Climate change \nmitigation in cities: a systematic scoping of case studies. Environ. \nRes. Lett. 15, 093008 (2020).\n33.\t Global Status Report for Buildings and Construction \u2013 Beyond \nFoundations: Mainstreaming Sustainable Solutions to Cut \nEmissions From the Buildings Sector (United Nations Environment \nProgramme, 2024); https://doi.org/10.59117/20.500.11822/45095\n34.\t City of Boston Office of Housing Boston Housing Conditions & \nReal Estate Trends Report (City of Boston Planning Department, \n2022); https://www.bostonplans.org/getattachment/ \n066b23c5-cab9-4731-a338-f6e57e3ef55f\n35.\t Kennedy, A. et al. The Greater Boston Housing Report Card 2023: \nWith a Special Analysis of Community Land Trusts (The Boston \nFoundation, 2023); https://www.tbf.org/-/media/tbf/reports-and-\ncovers/2023/gbhrc2023-full-report.pdf\n36.\t Anguelovski, I. et al. Green gentrification in European and  \nNorth American cities. Nat. Commun. 13, 3816 (2022).\n37.\t Planas-Carbonell, A., Anguelovski, I., Oscilowicz, E., P\u00e9rez-del-\nPulgar, C. & Shokry, G. From greening the climate-adaptive city \nto green climate gentrification? Civic perceptions of short-lived \nbenefits and exclusionary protection in Boston, Philadelphia, \nAmsterdam and Barcelona. Urban Clim. 48, 101295 (2023).\n38.\t 2020 Environmental Justice Populations (Commonwealth of \nMassachusetts, 2022); https://s3.us-east-1.amazonaws.com/\ndownload.massgis.digital.mass.gov/shapefiles/census2020/ \nEJ%202020%20updated%20municipal%20statistics% \n20Nov%202022.pdf\n39.\t Bulkeley, H., Edwards, G. A. S. & Fuller, S. Contesting climate \njustice in the city: examining politics and practice in urban climate \nchange experiments. Global Environ. Change 25, 31\u201340 (2014).\n40.\t Cox, S. in Urban Climate Justice: Theory, Praxis, Resistance  \n(eds Rice, J. L. et al.) 205\u2013219 (Univ. Georgia Press, 2023).\n41.\t Scott, J. C. Seeing Like a State: How Certain Schemes to Improve \nthe Human Condition Have Failed (Yale Univ. Press, 1998).\n42.\t Ranganathan, M. & Bratman, E. From urban resilience to abolitionist \nclimate justice in Washington, DC. Antipode 53, 115\u2013137 (2021).\n43.\t See, J., Fuentes, A. S., Porio, E. & Wilmsen, B. Conceptualising and \nenacting pathways to transformative climate justice: examples \nfrom the Philippines. Local Environ. https://doi.org/10.1080/13549\n839.2024.2353040 (2024).\n44.\t Scoones, I. et al. Transformations to sustainability: combining \nstructural, systemic and enabling approaches. Curr. Opin. \nEnviron. Sustain. 42, 65\u201375 (2020).\n45.\t Hernandez, J. California environmental quality act lawsuits and \nCalifornia\u2019s housing crisis. Hastings Environ. Law J. 24, 21\u201371 \n(2018).\n\nNature Cities | Volume 1 | October 2024 | 665\u2013676\n676\nArticle\nhttps://doi.org/10.1038/s44284-024-00121-w\n46.\t Bacchiocchi, E., Sant, I. & Bates, A. Energy justice and the \nco-opting of indigenous narratives in U.S. offshore wind \ndevelopment. Renew. Energy Focus 41, 133\u2013142 (2022).\n47.\t Pellow, D. N. Garbage Wars: The Struggle for Environmental Justice \nin Chicago (MIT Press, 2002).\n48.\t Diezmart\u00ednez, C. V. & Short Gianotti, A. G. Municipal finance \nshapes urban climate action and justice. Nat. Clim. Change 14, \n247\u2013252 (2024).\n49.\t Schlosberg, D. & Collins, L. B. From environmental to climate \njustice: climate change and the discourse of environmental \njustice. Wiley Interdiscip. Rev. Clim. Change 5, 359\u2013374 (2014).\n50.\t Bullard, R. D. & Wright, B. (eds) Race, Place, and Environmental \nJustice After Hurricane Katrina: Struggles to Reclaim, Rebuild, and \nRevitalize New Orleans and the Gulf Coast (Taylor & Francis, 2009).\n51.\t Rice, J. L., Levenda, A. & Long, J. in Urban Climate Justice: Theory, \nPraxis, Resistance (eds Rice, J. L. et al.) 1\u201326 (Univ. Georgia Press, \n2023).\n52.\t Whyte, K. Too late for indigenous climate justice: ecological and \nrelational tipping points. Wiley Interdiscip. Rev. Clim. Change 11, \ne603 (2020).\n53.\t Rice, J. L., Long, J. & Levenda, A. Against climate apartheid: \nconfronting the persistent legacies of expendability for climate \njustice. Environ. Plan. E https://doi.org/10.1177/2514848621999286 \n(2021).\n54.\t Sultana, F. Critical climate justice. Geogr. J. https://doi.org/10.1111/\ngeoj.12417 (2021).\n55.\t Rice, J. L., Cohen, D. A., Long, J. & Jurjevich, J. R. Contradictions of \nthe climate-friendly city: new perspectives on eco-gentrification \nand housing justice. Int. J. Urban Reg. Res. 44, 145\u2013165 (2019).\n56.\t Shi, L. et al. Roadmap towards justice in urban climate adaptation \nresearch. Nat. Clim. Change 6, 131\u2013137 (2016).\n57.\t Agyeman, J., Schlosberg, D., Craven, L. & Matthews, C. Trends and \ndirections in environmental justice: from inequity to everyday life, \ncommunity, and just sustainabilities. Annu. Rev. Environ. Resour. \n41, 321\u2013340 (2016).\n58.\t Toxic Wastes and Race in the United States: A National Report on \nthe Racial and Socio-Economic Characteristics of Communities \nwith Hazardous Waste Sites (United Church of Christ, Commission \nfor Racial Justice, 1987).\n59.\t Bullard, R. Dumping in Dixie: Race, Class, and Environmental \nQuality (Westview Press, 1990).\n60.\t Taylor, D. E. Race, Class, Gender and American Environmentalism \n(Diane Publishing Co., 2002).\n61.\t Pulido, L. Rethinking environmental racism: white privilege and \nurban development in Southern California. Ann. Assoc. Am. Geogr. \n90, 12\u201340 (2000).\n62.\t Pulido, L. A critical review of the methodology of environmental \nracism research. Antipode 28, 142\u2013159 (1996).\n63.\t Maung, R. & Pellow, D. N. in Handbook of Environmental Sociology \n(eds Schaefer Caniglia, B. et al.) 35\u201352 (Springer, 2021);  \nhttps://doi.org/10.1007/978-3-030-77712-8_3\n64.\t Mohai, P., Pellow, D. & Roberts, J. T. Environmental justice.  \nAnnu. Rev. Environ. Resour. 34, 405\u2013430 (2009).\n65.\t Walker, G. Beyond distribution and proximity: exploring the \nmultiple spatialities of environmental justice. Antipode 41, \n614\u2013636 (2009).\n66.\t Hughes, S., Giest, S. & Tozer, L. Accountability and data-driven \nurban climate governance. Nat. Clim. Change 10, 1085\u20131090 \n(2020).\n67.\t Hughes, S. Principles, drivers, and policy tools for just climate \nchange adaptation in legacy cities. Environ. Sci. Policy 111, 35\u201341 \n(2020).\n68.\t McCauley, D. & Heffron, R. Just transition: integrating climate, \nenergy and environmental justice. Energy Policy 119, 1\u20137 (2018).\n69.\t Short Gianotti, A. G., Getson, J. M., Hutyra, L. R. & Kittredge, D. B. \nDefining urban, suburban, and rural: a method to link perceptual \ndefinitions with geospatial measures of urbanization in central \nand eastern Massachusetts. Urban Ecosyst. 19, 823\u2013833 (2016).\n70.\t Rosenbaum, W. A. Environmental Politics and Policy (Sage \nPublications, 2020).\n71.\t Rinfret, S. R. Frames of influence: U.S. environmental rulemaking \ncase studies: frames of influence. Rev. Policy Res. 28, 231\u2013246 (2011).\n72.\t Kerwin, C. M. & Furlong, S. R. Rulemaking: How Government \nAgencies Write Law and Make Policy (Sage Publications, 2019).\n73.\t Baxter, J. & Eyles, J. Evaluating qualitative research in social \ngeography: establishing \u2018rigour\u2019 in interview analysis. Trans. Inst. \nBr. Geogr. 22, 505\u2013525 (1997).\n74.\t Wood, A. Tracing policy movements: methods for studying \nlearning and policy circulation. Environ. Plan. A 48, 391\u2013406 (2016).\n75.\t Secor, A. J. in Research Methods in Geography: A Critical \nIntroduction (eds Jones, J. P. III & Gomez, B.) 194\u2013205  \n(Wiley, 2010).\nAcknowledgements\nWe are grateful to all of the city officials and community leaders \nwho generously shared their time, knowledge and experiences for \nthis research. This work was supported by a Boston Area Research \nInitiative Seed Grant (C.V.D.), a Fellowship by the Switzer Foundation \n(C.V.D.) and the National Science Foundation NSF-2314889 (A.G.S.G.) \nand NSF-1735087 (C.V.D.).\nAuthor contributions\nC.V.D. and A.G.S.G. conceptualized the work. C.V.D., B.K.S. and A.G.S.G. \ncarried out the methodology and interpretation of the results. C.V.D. \ncollected and analysed the data in addition to visualization. The \noriginal draft was written by C.V.D., and B.K.S. and A.G.S.G. reviewed \nand edited the paper. Funding acquisition was by C.V.D. and A.G.S.G. \nProject supervision was by A.G.S.G. (lead) and B.K.S. (supporting).\nCompeting interests\nC.V.D. continues to be an employee of the City of Boston\u2019s Environment \nDepartment. The co-authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains supplementary \nmaterial available at https://doi.org/10.1038/s44284-024-00121-w.\nCorrespondence and requests for materials should be addressed to \nClaudia V. Diezmart\u00ednez.\nPeer review information Nature Cities thanks Juan Palacios,  \nTanesha Thomas and the other, anonymous, reviewer(s) for their \ncontribution to the peer review of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n\u00a9 The Author(s), under exclusive licence to Springer Nature America, \nInc. 2024\n\n\n    Scientific Research Findings:", "answer": "Our research demonstrates that climate justice has been a key component of implementing BERDO. Boston city officials sought to enact climate justice by elevating community voices (advocacy groups, community leaders and residents) in rulemaking processes, while maintaining the buy-in of parties that are subject to BERDO (real estate developers, hospitals, universities and other building owners). In this process, multiple actors could contest and reinterpret the justice mandates embedded in BERDO (for example, what prioritizing benefits for environmental justice populations means, or who the review board should represent). Different actors mobilized arguments about justice to advance their interests (for example, protecting tenants versus minimizing regulatory burdens) and ultimately shaped how BERDO has been operationalized on the ground, both in the planning process and the resulting implementation decisions. This unique case study of the implementation of an equity-oriented policy revealed distinctive challenges in operationalizing climate justice, including scope and scale discrepancies, and the bureaucratization and weaponization of justice.", "id": 62}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465241255946\nJournal of Health and Social Behavior\n2025, Vol. 66(1) 2\u00ad\u201317\n\u00a9 The Author(s) 2024\nDOI: 10.1177/00221465241255946\njournals.sagepub.com/home/hsb\nOriginal Article\nAs U.S. health disparities widen and the intergenera-\ntional transmission of advantage in families and \ncommunities strengthens, increasing attention is \nbeing paid to the processes underlying these trends. \nOf particular interest is early life, when children are \ninfluenced by previous generations in ways that will \nbe consequential for decades to come. Sociological \nresearch has pinpointed the shaping of children\u2019s \neveryday lives by parents in ways that foster chil-\ndren\u2019s long-term health and socioeconomic well-\nbeing as one important mechanism of the \nintergenerational transmission of advantage. Notions \nof children as agents in their own lives, rather than \npassive, innocent receptacles of socialization (see \nPugh 2014), and research on the importance of local \nnorms and institutions (Brown-Saracino 2015) con-\ntextualize families within communities as a site for \nreproducing inequalities.\nAs scholars are increasingly emphasizing, the \nconcept of \u201chealth lifestyles\u201d is useful for under-\nstanding how social and health inequalities play out \nin individuals\u2019 everyday lives (Cockerham 2005; \nKrueger, Bhaloo, and Rosenau 2009; Weber [1922] \n1978). Cockerham (2005:55) defined health life-\nstyles as \u201ccollective patterns of health-related \nbehavior based on choices from options available to \npeople according to their life chances.\u201d Understand\u00ad\nings of health lifestyles in early life and as part of \nthe intergenerational reproduction of inequalities \nare still nascent. Furthermore, although a long theo-\nretical tradition suggests that multiple lifestyle \noptions are available to people with similar social \n1255946 HSBXXX10.1177/00221465241255946Journal of Health and Social BehaviorMollborn et al.\nresearch-article2024\n1Stockholm University, Sweden\n2University of Colorado Boulder, Boulder, CO, USA\n3U.S. Census Bureau, Washington, DC, USA\n4Good Nutrition Ideas, Eugene, OR, USA\nCorresponding Author:\nStefanie Mollborn, Department of Sociology, Stockholm \nUniversity, Stockholm, SE-106 91, Sweden. \nEmail: mollborn@sociology.su.se\nChildren\u2019s Health Lifestyles  \nand the Perpetuation of \nInequalities\nStefanie Mollborn1,2\n, Jennifer A. Pace3, and Bethany Rigles4\nAbstract\nHealth lifestyles are a well-theorized mechanism perpetuating health and social inequalities, but empirical \nresearch has not yet documented crucial aspects: (1) health lifestyles\u2019 collective nature or content beyond \nbehaviors and (2) how people choose among available lifestyles in their social contexts. We conducted \ninterviews, observations, and focus groups with families in two middle- to upper-middle-class communities. \nContemporary class-privileged parenting involves constructing an individualized health lifestyle reliant on \nan expansive understanding of health and composed of parents\u2019 identities and narratives, children\u2019s health \nbehaviors and identity expressions, and community norms. Children\u2019s predominant health lifestyles in \nour sample vary by focus on parent versus child identity expression and on future achievements versus \npresent well-being. Parents expect health lifestyles to influence future socioeconomic attainment and \nhealth inequalities. Understanding how health lifestyles encompass more than behaviors and are locally \ncontextualized and how people choose them within structural constraints can inform research and policy.\nKeywords\nchildhood, family, health behavior, health lifestyle, inequality, parenting\n\nMollborn et al.\t\n3\nclass, most research has treated class-based parent-\ning and its implications for children as relatively \nmonolithic within class (e.g., Lareau 2011). With a \nhealth lifestyles approach, scholars can articulate \nvariation within social categories and multilevel \nconceptualizations of contexts to understand how \nchildren\u2019s everyday lives and health are shaped by \nparents, children, and communities in ways that \nmay reinforce future inequalities\u2014a phenomenon \nabout which not enough is yet known in childhood \n(Pugh 2014).\nEmpirical work on health lifestyles lags behind \ntheory in crucial ways. Despite a growing consen-\nsus that lifestyles likely extend beyond behaviors, \nextant (mostly quantitative) research has measured \nonly health behaviors. Furthermore, although it has \nlong been understood that health lifestyles are col-\nlective (Cockerham, R\u00fctten, and Abel 1997; \nFrohlich and Potvin 1999), empirical research has \nexamined them in individuals. Importantly, health \nlifestyles are theorized to be contextually specific \n(Cockerham et al. 2004), but almost all research has \nused national or geographically dispersed samples. \nFinally, Cockerham (2005:61) theorized, but empir-\nical work has not yet shown, that people \u201calign their \ngoals, needs, and desires with their probabilities for \nrealizing them and choose a lifestyle according to \ntheir assessments of the reality of their resources \nand class circumstances.\u201d This \u201cinterplay between \nlife choices and life chances\u201d (Cockerham \n2005:60), in which people choose lifestyles from \namong available options, is fundamental to health \nlifestyles theory but has not been empirically docu-\nmented. We seek to address these gaps through a \nstudy that situates families within community col-\nlectivities and creates space for new understandings \nof health, health behaviors, and health lifestyles to \narise from the data.\nThis study\u2019s goals are to (1) articulate an induc-\ntive model of the components of children\u2019s health \nlifestyles in two middle- to upper-middle-class \ncommunities and (2) identify predominant health \nlifestyles for community children, how they are \nchosen, and potential longer-term implications. Our \nqualitative interview, observational, and focus \ngroup data from families with elementary-aged \nchildren show that through health lifestyles, parents \nattempt to shape children\u2019s everyday lives in multi-\nfaceted ways that they imagine will affect long-\nterm well-being. Children\u2019s health lifestyles are \ndistinct from adult lifestyles because they blend \nparent and child identity expression and because of \nthe salience of later life for understanding how and \nwhy they form.\nWe articulate an inductively derived theoretical \nmodel of the components of children\u2019s health life-\nstyles and describe their predominant types in study \ncommunities. As Cockerham (2023) and others \n(e.g., Krueger et al. 2009; Mollborn and Modile \n2022) have foreshadowed, this model integrates \nwith and expands the standard definition of health \nlifestyles beyond health-related behaviors to \ninclude understandings of health and health-related \nnorms, narratives, and identities. Although we rely \nprimarily on interviews, multimethod data incorpo-\nrating observed behavior, private narratives, and \npublic talk document parental identity work, com-\nmunity structures and normative processes, and \nchild identity expression. Through health lifestyles, \nparents, schools, and communities together influ-\nence children\u2019s behaviors, identities, and futures. \nHealth lifestyles are thus a pathway through which \nsocial and health advantages persist across lives \nand generations.\nBackground\nThe Transmission of Advantage in Early Life\nU.S. socioeconomic inequalities have been increas-\ning, and intergenerational class mobility lags behind \nthat of peer countries (Saez 2008). Life expectancy \nis falling (Muennig et al. 2018), and links between \nsocioeconomic status and health are tightening \n(Masters, Hummer, and Powers 2012). Two impli-\ncations of these trends motivate health lifestyles \nresearch. First, strong processes transmit social \nadvantage and disadvantage across lives and gener-\nations. Lifestyle factors are important for under-\nstanding socioeconomic disparities in later health \n(Puka et al. 2022). Second, socially advantaged par-\nents are experiencing increasing pressures to pre-\nserve \nchildren\u2019s \nsocioeconomic \nand \nhealth \nadvantages to avoid worsening circumstances \namong the disadvantaged (Link and Phelan 1995; \nNelson 2010). Alongside structural phenomena, \nsuch as the intergenerational transmission of wealth \nand persistent residential segregation, research has \nfocused on family processes\u2014including health-\nrelated parenting (Augustine, Prickett, and Kimbro \n2017; Christensen 2004)\u2014to understand intergen-\nerational replications of social advantage. Indeed, \nboth healthy behaviors and socioeconomic advan-\ntage are increasingly being consolidated in families \n(Maralani and Portier 2021). The life course per-\nspective, including theories on accumulation and \ntransmission of advantage, often grounds this work \n(Elder 1994).\n\n4\t\nJournal of Health and Social Behavior 66(1) \nHealth Lifestyles\nContemporary intensive parenting and child iden-\ntity development are manifested in health lifestyle \nformation. Although children are infrequently \naddressed, a long-standing literature has examined \nindividuals\u2019 lifestyles as manifestations of social \nclass and other conditions. A lifestyle is an individu-\nal\u2019s collection of everyday behaviors, expressing \ngroup-based identity, regulated by group norms, and \nconstrained by social structures (Cockerham 2013). \nSocial class results in a specific set of lifestyle \noptions being available, and individuals choose \namong them as an expression of group membership \n(Weber [1922] 1978). Cockerham (2005) and others \n(e.g., Frohlich and Potvin 1999; Mollborn et al. \n2014) have focused specifically on health lifestyles, \ndefined previously, as a core domain of lifestyles. \nPrevious research on health lifestyles focused on \nadults can be extended to children. Children have \nrelatively less control over their health lifestyles but \nexercise agency and express identities in conjunc-\ntion with parental control (Pugh 2014).\nNational quantitative studies measuring behav-\niors abound, but little research examines health life-\nstyles in specific contexts or collectivities \n(Cockerham 2023). One quantitative study found \nthat health lifestyle behaviors cluster in localized \nways (Lee et al. 2015), and another mapped health \nlifestyles in two schools (adams et al. 2021). A \nqualitative study found that Scottish adults under-\nstood health lifestyles as reflecting identity and \nlived experience (McGarrol 2020), and another \ncontrasted Norwegian youths\u2019 health lifestyle for-\nmation by class standing (Eriksen et al. 2024). \nResearch has not empirically documented aspects \nof health lifestyles beyond behaviors or how people \nchoose them based on locally available options.\nHealth lifestyles are not just a resource for \nfuture health but also a cultural symbol that itself \ngenerates inequalities (Mollborn, Lawrence, and \nSaint Onge 2021). Korp (2008) emphasized the \nsymbolic power of \u201chealthy\u201d lifestyles as both \nmanifestations of inequality and phenomena that \ncreate inequality. Shared notions of \u201chealthy\u201d or \n\u201cunhealthy\u201d lifestyles legitimize some behaviors \nand delegitimize others. Lifestyles are an effective \nform of class distinction (Abel 2008; Bourdieu \n1986a), especially because of the increasing moral \nvalue being placed on health (Luna 2019). Indeed, \nlaypeople consider \u201chealthy\u201d lifestyle behaviors \ncrucial to well-functioning families (Williamson  \net al. 2018). Because parenting, like health, has strong \nmoral dimensions (Shirani, Henwood, and Coltart \n2012), children\u2019s health lifestyles create class dis-\ntinctions (Mollborn, Rigles, and Pace 2021). Class-\nadvantaged parents use food to teach classed values \nabout discipline and enact distinction, linking nar-\nratives and behaviors (Elliott and Bowen 2018; \nFielding-Singh 2019; Wills et al. 2011). These pro-\ncesses have longer-term implications, as Eriksen  \net al. (2024:149) wrote: \u201cThe dense integration of \nhealth lifestyle and family life instils a rigorous \nhealth orientation in the upper-class child\u2019s habitus, \na bodily disposition for health practices equipping \nthem to live\u2014and to wish for\u2014a healthy lifestyle \nnow and as they grow older.\u201d\nConceptualizing lifestyles as a blend of struc-\nture and agency (Cockerham 2005), we find that \nchildren\u2019s health lifestyles are rooted in parents\u2019 \nexpansive understandings of health, which blur \nboundaries across physical and psychological well-\nbeing, social integration, and academic achieve-\nment (Pace, Mollborn, and Rigles 2022; Warner \n2010). They are broader than researchers\u2019 implicit \nunderstandings of health when operationalizing \nhealth lifestyles. Health lifestyles encompass chil-\ndren\u2019s behaviors (including traditional health \nbehaviors and others linked to wider conceptualiza-\ntions of health, such as socializing and doing home-\nwork). Although content likely varies, norms in the \ncommunities that we studied prescribe that chil-\ndren\u2019s health lifestyles must reflect substantial \nparental identity investments and coherent parent-\ning narratives articulating the lifestyle\u2019s benefits for \nhealth, well-being, and future success. Children\u2019s \nidentity expression is also dictated by community \nnorms. Health lifestyles create advantages for chil-\ndren while simultaneously reaffirming parents\u2019 \nadvantaged class standing. Health lifestyles likely \nhave implications for socioeconomic attainment, \nhealth, and future lifestyles and identities, fueling \nthe intergenerational perpetuation of inequalities.\nParenting and Inequalities\nThe literatures on health lifestyles and parenting \nconverse infrequently, and this study exemplifies \nhow they can inform each other. Parenting practices \nreproduce advantages, instilling cultural, social, and \nhuman capital and setting children up for success or \nfailure in interactions with social institutions \n(Bourdieu 1986b; Lareau 2011) in what parents \nview as an uncertain world (Nelson 2010). Structural \nopportunities heavily constrain these practices. \nClass-advantaged parents leverage resources to \nselect structures considered beneficial for children, \nsuch as residential areas, child care, and schools \n\nMollborn et al.\t\n5\n(Augustine, Cavanagh, and Crosnoe 2009; Lareau, \nEvans, and Yee 2016; Mirowsky and Ross 2015). \nBeyond \u201cstatus safeguarding\u201d (Milkie and Warner \n2014) via institutions, resource-advantaged parents \ntypically raise children in ways that intensify their \nlikelihood of future success. Schneider, Hastings, \nand LaBriola (2018) identified larger class dispari-\nties in parental financial investments in children in \nstates with higher income inequality, driven by \nincreases among highest income parents.\nThis exemplifies \u201cintensive parenting,\u201d a strat-\negy in which advantaged parents, especially moth-\ners, leverage abundant resources to instill cultural \ncapital, health, and educational benefits in children \n(Bourdieu 1986b; Hays 1996; Shirani et al. 2012). \nAnalyses of mothers\u2019 time investments with chil-\ndren suggests that they matter for academic and \nbehavioral outcomes (Fomby and Musick 2018). \nCalarco (2014) found that parents\u2019 strategies repre-\nsented deliberate mobilization of class privilege. A \nbackdrop of growing economic insecurity height-\nens parents\u2019 sense of risk (Cooper 2014).\nGroup-based identities (Stets and Burke 2000) \nshape parenting. Collett, Vercel, and Boykin (2015) \narticulated parenting identity processes for fathers. \nA parent has a socially influenced identity standard \nfor a \u201cgood parent\u201d and seeks to align their parent-\ning behaviors with the standard to minimize discrep-\nancies and verify their identity, reducing negative \nemotions and social judgments and increasing posi-\ntive ones. The more flexible the identity standard is, \nthe more leeway there is to avoid negative repercus-\nsions. Tsushima and Burke (1999) emphasized the \nimportance for parenting identity of aligning parent-\ning tasks such as managing children\u2019s time with \nabstract goals like fostering autonomy. Resources \nfacilitate this alignment. Children\u2019s identity expres-\nsion can also intervene to complicate parents\u2019 efforts \n(Chin and Phillips 2004). Pugh (2009) found that \nchildren pursued goals such as engaging with popu-\nlar culture and being independent from adults\u2019 con-\ntrol, which often conflicted with adults\u2019 goals. \nExpanding research on agency to include often \noverlooked groups like children is important for life \ncourse theory (Landes and Settersten 2019).\nHealth is not directly addressed in much parent-\ning literature, but some studies incorporate it. \nIntensive parenting among socially advantaged par-\nents can be \u201cindividualist parenting\u201d (Reich 2016), \nreflecting a parent\u2019s vision for the child\u2019s future in \nways that are guided by experts but do not always \nfollow standardized recommendations. This can \nsometimes result in advantaged parents making \nchild health decisions not driven by medical \ninterpretations of the child\u2019s best interests (King, \nJennings, and Fletcher 2014; Reich 2016).\nData and Methods\nData and Procedures\nTo understand the empirical scope of health life-\nstyles, this study\u2019s design was different from typical \nhealth lifestyle approaches that focus on a specific \nset of physical health-related behaviors and their \nfrequencies. We sought to collect qualitative data on \nchildren\u2019s health lifestyles, avoiding imposing pre-\nexisting notions around what constitutes health and \nhealth behaviors for parents, what children\u2019s health \nlifestyles consist of, and whether parents talk about \nhealth lifestyles. Recruitment materials said the \nstudy was about \u201cparents, kids, and well-being.\u201d \nData collection strategies, procedures, and instru-\nments were refined through pilot research. The data \ncombined in-home family observations, parent \ninterviews and focus groups, and key informant \ninterviews in two neighboring middle-class com-\nmunities in the U.S. West\u2014\u201cGreenville\u201d and \n\u201cSpringfield\u201d\u2014from September 2015 to May 2016. \n(All names and some potentially identifying details \nhave been altered.) Our primary data source was 55 \nparent interviews: 35 with parents who also partici-\npated in a home observation (N = 30 families; in 5 \nfamilies, both parents requested to be interviewed \ntogether) and 20 with parents who only participated \nin an interview. We conducted six focus groups \n(three for each community), including 21 parents \n(some of whom had also done interviews and/or \nobservations). The nine key informants interacted \nwith families in the local area (e.g., sports coaches, \npediatricians, teachers). The 30 observation families \n(typically observed from the end of school to the \nstart of the bedtime routine on one school night) \nincluded a fourth or fifth grader age 9 to 11. Parents \nin other interviews and focus groups had at least one \nelementary-age child. We chose these ages because \nfamily influences are still substantial but have been \njoined by peers, school, and child agency.\nWe abductively revised our study design in \nresponse to emergent evidence and methodological \nconsiderations (Timmermans and Tavory 2012), \nchanging our sampling strategy to better balance \ncommunity data collection and create community-\nspecific parent focus groups. Our broad-based recruit-\nment strategies were designed to diversify the sample \n(Lofland and Lofland 2006). Rather than relying pri-\nmarily on snowball sampling, we recruited partici-\npants online and through local parenting email \n\n6\t\nJournal of Health and Social Behavior 66(1) \nlistservs, personal contacts, referrals, and public fly-\ners. A few participants recommended the study to oth-\ners. When we reached 10 parent interviews from a \nschool, we stopped collecting data from its families. \nThe resulting nonrepresentative sample was sociode-\nmographically varied, included many neighborhoods \nand social networks, and incorporated families from \n23 elementary schools and homeschoolers.\nOur data collection team included one faculty \nmember and two graduate students (all White \nwomen) and three undergraduates (an African \nAmerican woman, an Asian American man, and a \nWhite man). Families received $200 for a home \nobservation with interview, and other participants \nreceived $50. The study received institutional \nreview board approval. The faculty member or grad-\nuate student conducted interviews and focus groups, \nwhich were audio recorded and transcribed. The \nsemistructured interviews started with questions that \ndid not prompt about health, regarding the child\u2019s \ndaily routine, how parents navigate the child\u2019s pref-\nerences, how their child\u2019s life compares to their own \nchildhood, what parenting messages they try to con-\nvey, and what parenting is like in their community. \nParents\u2019 strong focus on health emerged unprompted \nin these sections. Later, we prompted about health, \nincluding how parents define \u201chealth\u201d and \u201cwell-\nbeing,\u201d what shapes children\u2019s health, and so on. \nObservations were conducted by either the faculty \nmember or a graduate student and an undergraduate. \nObservers followed different family members in dif-\nferent spaces, yielding two sets of field notes for \nmany interactions and one for others.\nParent participants\u2019 average age was 43, and \n80% were mothers. Seventy-seven percent were \nmarried, 17% were divorced/separated, 4% were \nsingle, and one parent was widowed. Eighty-six \npercent of parents identified as White, 8% identi-\nfied as Asian American, and 6% identified as \nLatino; a substantial minority were foreign-born. \nChildren were 2 to 15 years old, with most in fourth \nor fifth grade. Based on reported parent and partner \neducation and occupation and housing quality in \nthe family observations, we coded 59% of families \nas upper-middle class, 29% as middle-class or \nmixed (e.g., higher education but lower income), \nand 12% as working class or poor. Most parents had \ngrown up in a similar social class as measured by \nparental occupations, but some were class-mobile.\nField Sites\nThe study\u2019s communities, both midsized cities in \nthe same large metropolitan area (U.S. Census \nBureau 2017), were middle- to upper-middle-class. \nThey were demographically quite similar, with \nmedian household incomes close to the state aver-\nage and high proportions of residents identifying as \nWhite. Middle-class Springfield was more socio-\neconomically and ethnically diverse than upper-\nmiddle-class Greenville; its median housing value \nwas half as high, and half as many residents had a \nbachelor\u2019s degree (U.S. Census Bureau 2017). Both \ncommunities had unusually high rates of positive \nhealth behaviors and low obesity rates and were \nlocated in a geographic region that was considered a \npolitically liberal health mecca attracting highly \neducated new residents.\nAnalysis\nElectronic copies of transcriptions and field notes \nwere manually coded using NVivo qualitative soft-\nware and summary spreadsheets. Analysis of themes \nreported here was inductive: We had no a priori \nexpectation about what children\u2019s health lifestyles \nconsisted of, how health or health behaviors were \ndefined, or how they would vary. We analyzed \nobservational field notes together with parent inter-\nviews to compare personal accounts to observed \nbehaviors, allowing themes to arise organically and \ncoding for some predetermined themes. People\u2019s \npublic and private accounts and behaviors are often \ninconsistent in sociologically meaningful ways \n(Swidler 2001). Because the communities\u2019 health \nlifestyles were similar even though distribution and \ndegree varied, we combined the communities here.\nWe viewed the interviews and focus groups as \nopportunities for participants to actively construct \nnarratives (Holstein and Gubrium 1995). Through \nnarratives situated in specific social contexts, people \nconstruct identities, justify actions, and manage oth-\ners\u2019 impressions (Swidler 2001). Narratives, which \nshed light on norms, individual and group identities, \nand inequalities, turned out to be an important aspect \nof children\u2019s health lifestyles. Our goal was not to \nadjudicate whether parenting, specific health life-\nstyles, or their consequences are good or bad.\nResults\nMany parents were familiar with the idea of a \nhealthy lifestyle, and there was some unprompted \nuse of the term. But mostly parents broadly articu-\nlated \u201cthe way we raise them\u201d (in Laura\u2019s words), \noften with eloquent narratives, identity statements, \nand nuanced understandings of health and commu-\nnity norms underlying the behavioral routines they \n\nMollborn et al.\t\n7\ncarefully fostered and repeatedly linked to health \nand well-being. Grounded in these data, we induc-\ntively modeled the components of children\u2019s health \nlifestyles as articulated by parents and described \nprevalent types in the study communities.\nComponents of Children\u2019s Health Lifestyles\nWe found that children\u2019s health lifestyles combine \nbehavioral and nonbehavioral aspects. Dan, a \nWhite, upper-middle-class Springfield father, illus-\ntrates this complexity when describing 11-year-old \nBrittany\u2019s typical weekday routine:\nShe gets to sleep in. She gets to get a nice \nhomemade breakfast. We walk to school. I \npersonally\u2014Mary [Brittany\u2019s stepmother] and I \nboth feel\u2014it\u2019s a much better life for a child to \nget a good, full night\u2019s sleep and start the day \nwith a good breakfast. Have some family time . . . \norganic, cage free, all that good stuff.\nIn representing Brittany\u2019s everyday lifestyle, Dan \nmakes health salient and blends behavioral routines \nwith narratives, representing his parent identity, that \nunderlie and justify them. Dan continues, \u201cMary \nand I are both\u2014I will call it old-fashioned\u2014but we \neat dinner together. I don\u2019t answer the phone; nei-\nther of us will answer the phone at dinnertime. So \nwe talk during dinner, and we spend a lot of quality \ntime together, not just being under the same roof.\u201d \nMealtime behaviors express his and Mary\u2019s \u201cold-\nfashioned\u201d identities. Dan implicitly distinguishes \nhis family\u2019s lifestyle from his idea of a typical mod-\nern family that he believes eats separately.\nDan continues with a parenting narrative on \nboundary setting, which he links to well-being \nthroughout his interview.\nAnd ever since Brittany was really little, I\u2019ve \nalways been\u2014what\u2019s the word? Fanatical? I \nthink kids do better with consistency, and you \ncan also add into that, boundaries. So she\u2019s \nalways had a set bedtime, which has obviously \ngotten a little later as she\u2019s gotten older. . . . And \nI think all human beings operate better, function \nbetter with a routine.\nDan does repeated narrative work to distinguish his \nparenting favorably from that of other parents, \nwhom he views as overly hands-off. He also \nacknowledges the importance of community for \nfacilitating his desired health lifestyle: \u201cOne of the \nreasons I bought this house is, it\u2019s close to the ele-\nmentary school, it\u2019s close to middle school, it\u2019s \nclose to the high school. So since she\u2019s an only \nchild, I wanted her to grow up with a lot of friends in \nher neighborhood that she would get to know all \nthrough school.\u201d Mary and Dan organize their work \nschedules and exercise time to maximize interaction \nwith Brittany. We observed rooms dedicated to her \nhobbies. Manifesting Dan and Mary\u2019s parenting \nidentities and representing a broad understanding of \nhealth from sleep and diet to family interactions and \npeer contact, Brittany\u2019s everyday routine is care-\nfully curated to create a health lifestyle that they \nbelieve sets her up for future success.\nReflecting this example, Table 1 describes our \ninductively derived model of the components of \nchildren\u2019s health lifestyles. As contemporary health \nlifestyles theory posits (Cockerham 2023), they \ncombine health behaviors and multiple nonbehav-\nioral aspects. Qualitative research on young adults \nhas suggested that \u201cnot just health behaviors, but \nidentities, narratives, norms, and understandings of \nhealth [are] core aspects of health lifestyles\u201d \n(Mollborn and Modile 2022). But previous scholar-\nship has not systematically demonstrated how these \naspects integrate to create health lifestyles or form in \nthe interplay among parents, children, and commu-\nnity collectivities. We address each aspect in turn.\nUnderstandings of health (parent).\u2002 The first com-\nponent of children\u2019s health lifestyles, understand-\nings of health, undergirds the others. As described, \nour previous research (see Pace et al. 2022) has \narticulated these expansive understandings of chil-\ndren\u2019s health that parents draw on to craft children\u2019s \nhealth lifestyles, including behaviors, identities, \nnarratives, and norms. Aspects include physical \nhealth status and health behaviors; psychological \nhealth; achievement in the academic, athletic, and \nextracurricular realms; and social connectedness \n(Pace et al. 2022). Sofia articulated such a multifac-\neted understanding: \u201cI think everything has to be in \nTable 1.\u2002 Inductively Derived Components of \nChildren\u2019s Health Lifestyles.\nHealth Lifestyle Components\n1. Understandings of health (parent)\n2. Health behaviors (child)\n3. Parenting narratives (parent)\n4. Identities (parent and child)\n5. Norms (community collectivities)\n\n8\t\nJournal of Health and Social Behavior 66(1) \nbalance to be healthy . . . I think it\u2019s important to \nhave friends, it\u2019s important to exercise, it\u2019s impor-\ntant to eat well, it\u2019s important to . . . .\u201d Her husband \njumped in: \u201cto know other cultures.\u201d Similarly, \nEmma said, \u201cMental and emotional [health] is really \nimportant, so we kind of focus on stress, health, and \nhealthy lifestyle.\u201d\nHealth behaviors (child).\u2002 The previously empiri-\ncally documented aspect of children\u2019s health life-\nstyles is their health behaviors. Reflecting parents\u2019 \nexpansive understandings of health, we found that \nparents related a wide variety of child behaviors to \nhealth, including those typically related to physical \nhealth, like diet and exercise, but also those related \nto psychological well-being, academic achieve-\nment, and social connection. Brittany\u2019s lifestyle \ndescribed previously is one example, as are Linda \nand her 9- and 11-year-olds. Self-describing as \n\u201cprobably an average family, average parent,\u201d Linda \ndetails her efforts to oversee a wide variety of \nhealth-related behaviors in her children, focusing on \nnutrition, sports participation, unstructured play, \nsleep, curbing behaviors that she fears can lead to \nfood and technology addictions, and managing \nbehaviors that she relates to psychological resil-\nience, self-esteem, and the capacity to learn. We \ncannot rigorously investigate the frequencies of \nchildren\u2019s health behaviors with our qualitative \ndata, a task for future research.\nParenting narratives (parent).\u2002 As in Linda\u2019s and \nDan\u2019s interviews, most parents link children\u2019s \nhealth to thoughtful, well-articulated, health-ori-\nented narratives around parenting. In our communi-\nties, intensive parenting\u2014which falls mostly on \nmothers, although some fathers, like Dan, are heav-\nily involved\u2014entails constructing a narrative that \njustifies their child\u2019s individually tailored health \nlifestyle. These narratives are similar to but more \ncomplex and multifaceted than the health lifestyle \nnarratives class-privileged U.S. young adults (but \nnot those less privileged) said they had learned from \nparents growing up (Mollborn and Modile 2022). \nReflecting other trends toward the neoliberal indi-\nvidualization of lives (Reich 2016), parents do not \nsimply adopt available cultural templates for chil-\ndren\u2019s lifestyles. Rather, customized parenting \nchoices and narrative justifications are expected, \nreflecting parents\u2019 unique identities. Resource con-\nstraints on health lifestyle construction are mostly \ninvisible in parents\u2019 narratives.\nAndrea says her parenting is informed by \u201ca lot \nof reading\u201d because \u201cyou just want them to be \nhealthy.\u201d Andrea\u2019s narrative highlights the \u201ccon-\nstant struggle\u201d of balancing children\u2019s \u201cmental,\u201d \n\u201cphysical,\u201d and \u201cspiritual\u201d health. She focuses on \n\u201cthe routines and their structure and how you talk to \nthem and how you get through those hard days and \nnights where everything is just a nightmare. It\u2019s so \nkey. And it really does pay off.\u201d She responds when \nasked if her 8- and 11-year-olds are healthy: \u201cYeah. \nThey have friends, they go to birthday parties, \nthey\u2019re active. Thank God they don\u2019t have ongoing \nissues with dyslexia or learning disorders. I\u2019m \nblessed. I think they\u2019re both a little overweight, but \nI\u2019m sure that will even out.\u201d She went on to discuss \nher efforts to \u201copen up a conversation\u201d with her \ndaughter (but not her son) about balancing her \ndesire to be sedentary and create art with the need \nfor cardiovascular activity. Andrea\u2019s narrative pres-\nents her as a health-focused mother who applies \nhard work and discipline to parent successfully. \nParents\u2019 narratives complement behaviors and sub-\njective understandings of health as key aspects of \nchildren\u2019s health lifestyles in these class-privileged \ncommunities.\nIdentities (parent and child).\u2002 A child\u2019s health life-\nstyle reflects parents\u2019 identities, represented through \nbehaviors and narratives in a public way that is \nappraised by other community members. But poten-\ntial judgment is not the only reason parents\u2019 identi-\nties are highly invested in constructing lifestyles: \nParents view health lifestyles as crucial for shaping \nchildren\u2019s well-being and success. Parent and child \nidentity expression through children\u2019s health life-\nstyles is normatively prescribed in these communi-\nties: Parents should determine the best lifestyle to \nreflect their identities, and it should also be por-\ntrayed as aligned with the child\u2019s identity.\nParent identities are palpable throughout their \nnarratives, such as Andrea\u2019s portrayal of her hard \nwork to parent successfully around health and \nDan\u2019s self-promoting comparisons of his boundary-\nsetting efforts to make Brittany\u2019s lifestyle healthier. \nSimilarly, Dawn discusses at length how her iden-\ntity influences her parenting of her 9- and 5-year-\nolds, from fostering resiliency and \u201ca deep sense of \nself-worth and self-love\u201d to instilling respect and an \nunderstanding of the importance of eating nutri-\ntiously. Her narrative repeatedly cites outside evi-\ndence that her children\u2019s identities are developing \nalong these lines, from them happily drinking green \nsmoothies to a neighbor commenting on how \u201ccom-\nfortable,\u201d \u201cgracious,\u201d and \u201cwell adjusted\u201d they are.\nDawn\u2019s attention to her children\u2019s identity \nexpressions is typical, but parent and child identity \n\nMollborn et al.\t\n9\nexpressions often conflict. Children, by expressing \ntheir own preferences, disrupt the smoothness of \nparents\u2019 attempted socialization into lifestyles. In \nobservations, common pushback or conflict came \nfrom children wanting to use technology more, eat \nunhealthier foods, or move less than parents pre-\nferred. Parents and children struggling for control \nof children\u2019s behaviors caused tensions. Brittany \n(described previously) typifies study children when \npushing back against parent control. She resists \nlimits on dessert, having her food cut up, sugges-\ntions about physical activities she should do, and \nrules around technology use. Brittany works to \nexpress her identity, although in a defeated way that \nanticipates an ultimate lack of control (which was \nstarker than in many families). These realities of \nchild identity expressions complicate the neater \nnarratives parents often present, such as Dan\u2019s \naccounts of Brittany\u2019s compliance.\nFurther complicating these dynamics, many par-\nents in our sample articulate a norm that a health \nlifestyle should reflect the child\u2019s identity. Parents \noften work hard to make children\u2019s behaviors both \nfit their parent identities and credibly appear to \nreflect the child\u2019s identity. Christine describes the \nintensive parenting efforts involved in accommo-\ndating 10-year-old Noah\u2019s preferences:\nIt took us a long time to find something Noah \nwanted to do. When he was younger we had him \nin gymnastics, tae kwon do, swimming. We tried \neverything until we hit soccer, and then he just \nloved it. So he really didn\u2019t want to do anything \nelse until he hit basketball. And so now his entire \nfocus is basketball and soccer. And that\u2019s what \nhe wants. . . . So all that is his choice.\nChristine describes Noah\u2019s sports involvement as \n\u201chis choice,\u201d but Noah\u2019s agency was actively con-\nstructed and constrained by his parents repeatedly \nenrolling him in sports until he found two he enjoys. \nNoah is being pressured to \u201cprefer\u201d a health lifestyle \nthat involves \u201cloving\u201d playing multiple sports. He \nlikely understood that his parents\u2019 efforts to put him \non sports teams would not cease until he chose two \nsports, which constrains his identity expression \neven as Christine portrays him as choosing.\nIt would have been far easier for Christine sim-\nply to choose two sports, so she is expending con-\nsiderable effort to encourage Noah\u2019s identity \nexpression. Thus, even though children\u2019s identity \nexpression often causes problems for children\u2019s \nenactment of their parents\u2019 preferred behaviors, \nparents encourage it as an important facet of \nmiddle-class health lifestyles. Participants differ, \nhowever, in how highly they prioritize children\u2019s \nidentity expression (see the following).\nNorms (community collectivities).\u2002 Community, or \ncollectivity, norms are the final component of chil-\ndren\u2019s health lifestyles that we identified. They reg-\nulate what is acceptable in other aspects and are \ntools used by parents to reinforce lifestyles. Many \nparents explicitly rely on community norms to rein-\nforce their chosen lifestyle. Nick describes health-\nrelated norms in Greenville as a draw for moving \nthere, such as valuing \u201cexercise and organic and \ngrowing food in your backyard . . . I feel like my \nfifth grader, after being in a Greenville school, has \nbeen fairly indoctrinated about the food, which I \nnever was growing up. You know, the food groups, \nand creating a healthy meal, and eating different \ncolors when we\u2019re eating, and the importance of \nthat.\u201d Nick feels that strong norms in his son\u2019s \nschool reinforce his health lifestyle.\nYet community norms can threaten chosen health \nlifestyles, such as when an \u201cachievement of indepen-\ndence\u201d lifestyle is condemned by other parents (see \nthe following). In this class-advantaged sample \nwhere geographic mobility is normative, many par-\nents, including Nick, acknowledge these dynamics, \nsaying they chose where to raise their children based \non the community norms they would experience in \nenacting their preferred health lifestyles. This exem-\nplifies the formation of \u201coverrider enclaves\u201d resist-\ning the \u201cdefault American lifestyle\u201d posited by \nMirowsky and Ross (2015). Thus, parent preferences \ncan shape the community norms in the child\u2019s life-\nstyle. Hector says of Greenville\u2019s norms that encour-\nage healthy eating and physical activity, \u201cWe brought \nthem [the children] to the environment where the \nvalues we valued are there. And they are getting \nthem, not just from us, which in hindsight has been \namazing. But it\u2019s sort of what brought us here, \nright?\u201d Although she can feel judged for her lifestyle \napproach (see the following), his wife, Sofia, echoes, \n\u201cI feel we have a lot in sync with other parents and \nthe way they teach their children here.\u201d Many par-\nents note synergies or tensions between community \nand family in fostering health lifestyles.\nParents sometimes acknowledge that commu-\nnity resources facilitate norms. A Greenville focus \ngroup parent described the community as \u201cfamily \noriented.\u201d Another parent tentatively linked that to \nresources: \u201cI wonder if it has something to do with that \nit\u2019s a more affluent community, and so, you know, \nthere\u2019s more free time maybe, or more ability for \nparents to be involved because they\u2019re not trying to \n\n10\t\nJournal of Health and Social Behavior 66(1) \ndrag all of their jobs.\u201d The first parent concurred, \nand a third noted that \u201call that land\u201d\u2014public open \nspace\u2014contributed to community cohesion and \nnorms. Parents\u2019 understanding of community \naspects of children\u2019s lifestyles is often nuanced, \narticulating selection and causation and relating \nresources to norms.\nLocalized Health Lifestyle Options\nUnderstandings of health, behaviors, narratives, \nidentities, and norms together comprise children\u2019s \nhealth lifestyles, but the lifestyles\u2019 actual content \nvaries within our sample. As health lifestyle theory \nsuggests, community norms prescribe that parents \nshould intensively construct a deliberate lifestyle \nfor their children from among acceptable options. \nBefore articulating differences, we emphasize that \nthese middle-class communities\u2019 predominant life-\nstyles have important similarities in health-related \nnorms: (1) a strong focus on nutrition, exercise, and \nacademic achievement; (2) intensive parenting \nefforts required to sustain the lifestyle; and (3) the \nhigh salience of the child\u2019s future for their health \nlifestyle.\nWe summarize the four prevalent health life-\nstyles in our communities, primarily illustrated \nusing four families. Table 2 articulates the lifestyles \nand the two dimensions that differentiate them: the \nrelative importance of parent versus child identity \nexpression and the relative focus on the future ver-\nsus present. Although most parents blend elements \nof multiple health lifestyles in constructing a life-\nstyle, most children\u2019s lifestyles primarily fall into \none type. We argue that the interweaving of differ-\nent health lifestyle components (understandings of \nhealth, behaviors, narratives, identities, and norms) \nstrengthens the power of these lifestyles, likely \nmaking them more effective for transmitting class \nadvantage into adulthood than a lifestyle solely \ncomposed of behaviors would be.\nWe expect that different communities likely con-\ntain different prevalent children\u2019s health lifestyles \nbecause societal notions around parenting intersect \nwith community demographics and localized cul-\ntures (see Brown-Saracino 2015). Our communities\u2019 \ncultural focus on physical health, athleticism, and \noutdoor activities shapes health lifestyles, poten-\ntially making them more salient for parenting than \nelsewhere and attracting new families who seek \nTable 2.\u2002 Typology of Childhood Health Lifestyles and Selected Associated Behaviors and Parent \nNarratives Documented in Our Middle- to Upper-Middle-Class Communities.\nWhose Identity Expression Is Prioritized?\nPrevalent Behaviors and \nAttitudes\n\u2002\nParent\nChild\nTemporal focus?\n\u2002 \u2003 Long-term \nElite achievement \nlifestyle\nAchievement of \nindependence  \nlifestyle\nBehaviors: more structured \nexercise, less unstructured \nplay, more (implicit) body \nregulation\nNarratives: competencies/\nachievement focus\n\u2002 \u2003 Short-term\nFamily connection \nlifestyle\nLet kids be kids  \nlifestyle\nBehaviors: more \nunstructured exercise and \nplay, less body regulation\nNarratives: well-being/stress \nfocus\nPrevalent behaviors \nand attitudes\nBehaviors: more \nrestricted technology \nuse\nNarratives: focus on \nparents developing \nchild identity, parents \nfeeling less alone and \nmore successful\nBehaviors: less \nrestricted technology \nuse\nNarratives: focus on \nchild developing own \nidentity, parents \nfeeling more alone \nand worried about \nbeing successful\n\u2002\n\nMollborn et al.\t\n11\nsuch lifestyles. Family and community resources are \nalso important for enabling these lifestyles. The pre-\ndominant lifestyles that we identify likely appear in \nsome other communities but are presumably joined \nby additional lifestyles yet to be identified.\nElite achievement.\u2002 The most prevalent, norma-\ntively dominant, and typically most resource-intensive \nhealth lifestyle for community children, the \u201celite \nachievement\u201d lifestyle, combines a future-oriented \nfocus on skill building with high levels of parental \nidentity expression in everyday interactions (see \nTable 2). Achievement is part of parents\u2019 under-\nstandings of child health (Pace et al. 2022). Parents \nfrom upper-middle-class childhood backgrounds \nwere disproportionately represented in this health \nlifestyle. The elite achievement lifestyle has much \nin common with Lareau\u2019s (2011) \u201cconcerted culti-\nvation\u201d parenting style, which involves heavily \nstructured time spent on extracurricular activities \nand school achievement. Parents seeking to enact an \nelite achievement lifestyle for their children differ in \nemphasizing sports, academics, or both. Unstruc-\ntured playtime and technology use are limited. They \nview their child\u2019s current success, including in \nachieving a fit body, as linked to future well-being \nand sometimes compromise children\u2019s current \nhealth and stress levels to achieve these goals (Moll-\nborn et al. 2021). However, like parents in other life-\nstyles, they tend to view it as expressing their \nchildren\u2019s preferences and making them happy. \nElite achievement parents feel that they are crafting \na \u201cnormal\u201d health lifestyle for their child and tend to \nview their parenting as relatively successful.\nFor example, Laura and 10-year-old Jacob are \nfrom a White, upper-middle-class Greenville family. \nLaura says, \u201cI feel like so much of our peer group is \npretty homogeneous\u201d in their parenting and that she \nfollows the dominant norm. Jacob is \u201cvery active,\u201d \nplaying on multiple sports teams many seasons and \ndoing supplemental academic work. Laura empha-\nsizes Jacob\u2019s identity expression in his lifestyle: \n\u201cMy ultimate goal for my son is really just to have \nfun with his activities.\u201d She also links Jacob\u2019s extra-\ncurricular involvement to future success: \u201cIf you \ndon\u2019t get in on the ground floor in some of these \nsports, you\u2019re never going to have a shot at playing \nin college.\u201d Laura worries about the academic impli-\ncations of Jacob\u2019s problematically \u201ctoo easy\u201d work-\nload, although he enjoys having little homework. \nShe feels she needs to provide supplemental aca-\ndemic preparation for his future. Finally, like parents \nin other health lifestyles, Laura connects Jacob\u2019s \nhealth lifestyle to parent and child identities: \u201cWe\u2019ve \njust always gone really hard. Like, I just go really \nhard. Their dad goes really hard. And that\u2019s just kind \nof the way we raise them. Yeah, a rolling stone gath-\ners no moss.\u201d Similarly, Kaya enrolled a sometimes \nreluctant 9-year-old Madeline in multiple sports to \ncombat overweight and her physical tendency to \n\u201cdo the minimal amount of work necessary\u201d while \nalso being \u201cextremely involved\u201d in fostering \nMadeline\u2019s academic achievement and interest in \nnutritious cooking.\nAchievement of independence.\u2002 Like elite achieve-\nment, the \u201cachievement of independence\u201d health \nlifestyle focuses on the future and achievement, but \nit emphasizes child identity expression in everyday \nactivities, working toward a distinct goal: the child\u2019s \nindependent acquisition of skills and maturity to \nfacilitate successful adulthood. This health lifestyle \ninvolves considerable parent management despite \nfocusing on child independence. Many parents \nespousing an achievement of independence lifestyle \ndefine future success differently from elite achieve-\nment parents, emphasizing the importance of raising \ntheir child into an autonomous and well-functioning \nadult regardless of socioeconomic success. They \nsometimes let children make unwise choices to \nserve the longer-term goal of fostering independent \ndecision-making and understanding consequences. \nAchievement of independence children tend to have \nstructured activities but more freedom in technol-\nogy use. Parents typically feel more alone in their \nparenting choices.\nJasmine, Aaron, and 9-year-old Evie belong to a \nmultiracial working- to middle-class Springfield \nfamily. Aaron said, \u201cWe are super intentional with \nwhat we do with the kids.\u201d Evie does more chores \nthan most study children, often toward a stated goal \nof fostering adult skills. When Evie asks Aaron \nwhether she is \u201cdone now\u201d with toothbrushing, \nAaron responds, \u201cYou\u2019re getting to the age where \nyou can decide if you\u2019re done or not. Like, you \ncan\u2019t keep on coming to me as an adult and be like, \n\u2018Is two minutes long enough to brush?\u2019 . . . I mean, \nyou have to answer these questions for yourself.\u201d \nWhether Evie brushes her teeth long enough today \nis secondary to the future-oriented goal of indepen-\ndently enacting appropriate health behaviors. \nJasmine is clear about the lifestyle they foster: \u201cI \nfeel like we are \u2018free range,\u2019 if you want to put \nlabels on all these things. We really give our kids \nthe opportunity to learn by themselves. Not neces-\nsarily alone, but I\u2019m going to watch you try that and \nsee if you can do it.\u201d Aaron articulates their goal of \nfuture independence and competence: \u201cI want her \n\n12\t\nJournal of Health and Social Behavior 66(1) \nto be confident. I want her to be able to choose col-\nlege if she chooses. I want her to choose to travel if \nshe wants to.\u201d They feel on track with meeting \nthese goals because \u201cI feel like she\u2019s healthy in her \nhead, where she knows pretty much who she is.\u201d\nYet Jasmine and Aaron sometimes feel the sting \nof others\u2019 judgement, suggesting that achievement \nof independence is less normative than elite \nachievement. Describing an outdoor interaction \nwith a hovering parent, Aaron says, \u201cIf I see you on \ntop of your kid, it\u2019s not a big deal to me. But if they \nsee us letting the kids go, they definitely have to say \nsomething [negative to me]. I just don\u2019t understand \nthat.\u201d Similarly, Sofia schedules many activities for \nher children and seeks to instill hard work, but \nwhen they are with other children, if \u201cthey\u2019re not \ngetting into really dangerous situations, I let them \nbe.\u201d She thinks friends judge her for \u201cgiving our \nchildren too much freedom\u201d and says \u201cyou get that \nlittle smirk\u201d from other parents.\nFamily connection.\u2002 The third prevalent health \nlifestyle combines the focus on parent identity \nexpression that is similar to elite achievement with a \nnew parenting dimension: emphasizing the child\u2019s \npresent over the future. Sometimes in explicit or \nimplicit resistance to elite achievement, \u201cfamily \nconnection\u201d parents prioritize current well-being \nand lower stress over future considerations, \nachieved by fostering social relationships, particu-\nlarly with family. This health lifestyle demands con-\nsiderable time and interaction, with its focus turned \ninward toward family more than outward toward \nstructured activities. Family connection children \noften restrict technology use and share unstructured \nexercise and play with family. Parents frequently \narticulate narratives of opting out of elite achieve-\nment and feeling secure in their parenting.\nSharon and 9-year-old Finn are in a White, mid-\ndle- to upper-middle-class Springfield family. \nSharon\u2019s family spends lots of time together engag-\ning in healthy behaviors: \u201cThe more we can be out-\ndoors together, the better. . . . We like to ski together \nas a family and do little trips together.\u201d Finn plays \non a traveling sports team. Although this activity \ncould signal elite achievement, Sharon situates his \ninvolvement within their family connection life-\nstyle, saying, \u201cI love [traveling overnight to games \nfor] soccer because it has brought us together as a \nfamily.\u201d Sharon restricts Finn\u2019s extracurricular \nreading\u2014supplemental academic preparation that \ncould foster future socioeconomic success\u2014when \nit threatens their \u201cfamily time.\u201d Like other family \nconnection parents, Sharon directly links her \nparenting identity to family members\u2019 healthy \nlifestyles:\nI think being healthy is having enough fresh air \nin your day and having good food to eat, having \nexercise, having family time, having friend time. \nI feel that\u2019s what I do in my life. For myself and \nmy family, I try to puzzle together\u2014like balance \nit out, so that we\u2019re all getting nourished on all \nthose levels. . . . And sometimes it\u2019s unbalanced. \nSo it\u2019s like, okay, we need to have some family \ntime . . . I kind of orchestrate it, I think, in our \nfamily.\nSharon carefully \u201corchestrates\u201d balance in her fam-\nily to improve current well-being and health, talking \nabout \u201cfamily time\u201d as an antidote to stressors and \nimbalances. Similarly, Mark and Rachel curtail tech-\nnology use and encourage joint physical activities to \ncreate more family time with their 7- and 9-year-\nolds. Rachel says, \u201cThe message is that you\u2019re a part \nof this family, and you need to contribute. And that\u2019s \npart of how we are great together. There\u2019s this mes-\nsage that we work together, and it\u2019s great, and we all \nneed to help each other.\u201d Although many family con-\nnection parents discuss opting out of some high-\npressure goals for their children\u2019s future to reduce \nstress, be healthier, and increase family interaction \ntime, they view their health lifestyle as equally inten-\nsively constructed and challenging to maintain.\nLet kids be kids.\u2002 The fourth health lifestyle simi-\nlarly focuses on present mental health and stress \nlevels, but the child\u2019s identity expression, more than \nthe parents\u2019, is considered crucial for well-being. \nDespite some similarities to Lareau\u2019s (2011) working-\nclass \u201caccomplishment of natural growth\u201d parenting \nstyle, \u201clet kids be kids\u201d parents intensively manage \nchildren\u2019s lives, constructing the lifestyle to accom-\nmodate child preferences. Let kids be kids children \nare enrolled in structured activities but have rela-\ntively more unstructured exercise and playtime and \nmore freedom when using technology. Their parents \nmake these decisions deliberately, but some still \nquestion their long-term implications.\nAnna and 10-year-old Chloe are in a White, \nupper-middle-class Greenville family. Anna articu-\nlates her parenting goal as inculcating \u201cthat feeling \nof, you\u2019re okay just the way you are. . . . And every-\nbody has a unique gift to offer, and you just have to \nfigure out what it is.\u201d These statements foreshadow \nthe future but foreground current well-being and \nchild identity expression. Anna says that like her, \nChloe is \u201cintroverted\u201d and \u201cusually wants to just \n\nMollborn et al.\t\n13\ncome home after school\u201d instead of going to sports \nand playdates. Anna \u201cputs [her children\u2019s] prefer-\nences first,\u201d but because Chloe has trouble making \nfriends, Anna has decided to schedule one playdate \na week with a friend of Chloe\u2019s choosing. Chloe \nwants to join a traveling sports team that Anna \nthinks is \u201cvery intense, and I\u2019m a little concerned . . \n. but she loves this team. So we\u2019re just going to go \nfor it.\u201d Reflecting this lifestyle, Anna helps Chloe \nexpress her identity even when privately critiquing \nher choices. Yet Anna sounds conflicted when com-\nparing herself to elite achievement parents\u2014an \nindication that this lifestyle may be less normative \nin Greenville\u2014worrying that her friends\u2019 daughter \n\u201cis going to, like, Princeton or something, and \nmaybe I should be pushing more.\u201d\nLike Anna, Julie goes to great lengths to facili-\ntate 9-year-old Callie\u2019s identity expression. Callie \nparticipates in multiple sports, and their family \nfocuses on nutrition, but self-expression is core to \nJulie\u2019s narrative. She says, \u201cIt\u2019s very, very impor-\ntant for Callie and I that she expresses herself. . . . \nWe believe in self-control and self-regulating as a \nway of teaching her self-care. . . . It\u2019s very impor-\ntant that she knows when she\u2019s in a situation that is \nnot okay, that she can speak up.\u201d Reflecting this \nemphasis, Julie says Callie is allowed several \u201cper-\nsonal days\u201d when she can choose to miss school, \nand there are no parental controls on family techno-\nlogical devices.\nDiscussion\nImplications for Inequalities\nThis qualitative study innovates by documenting the \ncomponents and prevalent types of child health life-\nstyles in two middle-class communities. Well-\ndeveloped theory has posited that health lifestyles \nextend beyond behaviors and that people choose \nlifestyles from among options available by social \nstatus, but empirical scholarship has not docu-\nmented whether or how this happens. Our study \nstarts to bridge this gap, empirically unpacking the \nlast box in Cockerham\u2019s (2023) theoretical model, \nwhich contains the construct of health lifestyles and \nlinks it to social reproduction. We found that parents \nconstruct expansive understandings of \u201chealth\u201d for \ntheir children (Pace et al. 2022), which undergird \nchildren\u2019s health lifestyles that are further com-\nprised of health-related behaviors, narratives, iden-\ntities, and norms. Parents craft health lifestyles from \namong locally available options and in concert with \nchild identity expression. Parents navigate potential \njudgments of their health lifestyle choices and artic-\nulate awareness of community norms that inform \ndecisions. These lifestyles are likely important for \nreproducing families\u2019 and communities\u2019 social, \nsocioeconomic, and health advantages.\nAlthough our cross-sectional data cannot docu-\nment long-term effects, parents imagine that their \nchildren\u2019s health lifestyles matter for identities and \nfuture lifestyles, health, and socioeconomic attain-\nment. We agree that this is likely. By prioritizing \nchild versus parent identity expression and present \nwell-being versus future competencies, parents feel \nthat their lifestyle choices inform their child\u2019s \nfutures. Anna\u2019s narrative of \u201ctrying to develop a \nperson who is intrinsically motivated to do this \nstuff, not for me\u201d drives her let kids be kids lifestyle \napproach. She is attempting both to shape Chloe\u2019s \nbehaviors and instill an identity that will presum-\nably shape her future lifestyle. Similarly, Laura \nteaching her son to \u201cgo really hard\u201d is about both \nelite achievement and a narrative around discipline \nand hard work, which can be used to justify advan-\ntaged class position (Luna 2019; Mollborn and \nModile 2022). Such mastery of behaviors and self-\npresentations that match cultural logics of classed \ninstitutions (Gage-Bouchard 2017) is linked to \nsocioeconomic attainment (Calarco 2018; Lareau \n2011). Despite some differences, all health lifestyle \ntypes in our study communities are class-privileged \nand feature intensive parenting, socialization into \nentitlement, and development of thin bodies \nthrough diet and physical activity, which facilitate \nclass distinctions and privilege in adulthood \n(Calarco 2014; Mollborn et al. 2021). Thus, con-\nstructing health lifestyles through behaviors, narra-\ntives, identities, norms, and understandings of \nhealth is a mechanism through which class and \nhealth distinctions can perpetuate intergeneration-\nally and across life. Qualitative research comparing \nclass-advantaged and -disadvantaged young adults \nhas indeed found this mechanism to be salient \n(Mollborn and Modile 2022). Collecting more data \ndirectly from children and following them longitu-\ndinally could help articulate these processes.\nImplications for Health Lifestyles Research\nThis study\u2019s effort to harmonize health lifestyles \ntheory with extant empirical research has research \nimplications. Most empirical health lifestyles work \nis quantitative, using nationally representative data \nand latent class analyses with specific health behav-\niors as inputs. This method identifies predominant \nconfigurations of inputs, which researchers label as \n\n14\t\nJournal of Health and Social Behavior 66(1) \nprevalent lifestyles in a national population but can-\nnot adjudicate whether each behavior should be part \nof the lifestyle or if behaviors or other components \nhave been missed. This research has spurred many \nadvances, such as identifying frequent combinations \nof healthy and unhealthy behaviors within individu-\nals and articulating life course and network dynam-\nics of health lifestyles (for a review, see Mollborn et al. \n2021). More progress is needed in several areas.\nFirst, parents\u2019 expansive understandings of chil-\ndren\u2019s health (Pace et al. 2022) justify incorporating \na wider variety of health behaviors. For example, \nCockerham\u2019s (2023) expanded theoretical model \narticulates technology use and COVID-related \nsocial distancing as newly important behaviors for \nhealth lifestyles. Behaviors related to psychological \nwell-being, social connection, and academic \nachievement could also be included, according to \nour parent participants. Latent class analysis frame-\nworks can accommodate broad operationalizations.\nSecond, both theory and our analysis suggest \nthat health lifestyles research should operationalize \nlifestyles more broadly beyond behaviors. We found \nthat health lifestyles combine understandings of \nhealth and health-related behaviors, identities, nar-\nratives, and norms (Cockerham 2023; Krueger et al. \n2009; Mollborn and Modile 2022). Without this \nexpansive conceptualization, efforts to understand \nhealth behaviors or parenting practices may lack \nimportant context and be difficult to change through \npolicies. Qualitative research combining public and \nprivate narratives with observations of behaviors \nand interactions, as we have done, is promising for \narticulating these processes. But quantitative \nresearch can also make strides. On a related topic, \nLankes (2022) conducted latent class analyses iden-\ntifying \nprevalent \nU.S. \nintensive \nparenting \napproaches that incorporated a variety of behavioral \nand attitudinal measures from survey data. \nBehaviors and attitudes combined in interesting \nways that furthered theoretical understandings. We \nargue that appropriate operationalizations of health \nlifestyles depend on research context but should ide-\nally be grounded in target individuals\u2019, rather than \nresearchers\u2019, conceptualizations of health. These \nunderstandings can set boundaries around a study\u2019s \noperationalization of health lifestyle by guiding the \nselection of behaviors for inclusion and possible \nmeasures of norms or attitudes, identities, and narra-\ntives depending on the method.\nThird, in moving beyond behaviors, research \ncan broaden beyond individuals to examine fami-\nlies and communities or collectivities. Situating \nindividuals\u2019 agency within linked lives is important \nyet overlooked (Landes and Settersten 2019). And \nhealth lifestyles are by definition collective \n(Cockerham 2005; Frohlich and Potvin 1999), an \nempirically neglected aspect. We found that par-\nents, children, and communities were all fundamen-\ntal to components of children\u2019s health lifestyles. \nAlthough a few studies have examined collective \ninfluences on individuals\u2019 health lifestyles, future \nresearch still needs to move beyond the individual \nlevel to show how lifestyles themselves and their \nstructure\u2013agency interplay are collective.\nFourth, this study joins a few qualitative and \nquantitative investigations (e.g., adams et al. 2021; \nLee et al. 2015; McGarrol 2020) in identifying \nlocalized health lifestyle options. We innovate in \nexploring how people negotiate choosing among \noptions, but more research is needed. Almost noth-\ning is known about how local health lifestyle \noptions vary systematically or intersect with indi-\nviduals\u2019 characteristics when forming a person\u2019s \nhealth lifestyle. Most extant research is conducted \nnationally, but theories expect health lifestyles to \nplay out locally. This means that many aspects of \nour findings likely do not apply in socially disad-\nvantaged settings. For example, Eriksen et al. \n(2024) and Mollborn and Modile (2022) found that \nnarratives about health lifestyles linked to strongly \nhealth-focused identities were prevalent among \nyoung people from class-advantaged but not class-\ndisadvantaged \nbackgrounds. \nHealth \nlifestyle \noptions should be examined using quantitative data \nthat include respondents clustered within widely \nvarying communities, and qualitative research \nshould document a diversity of local communities. \nEriksen et al\u2019s (2024) longitudinal qualitative study \nin Norway is one such step.\nWe argue that moving health lifestyles research \nin these directions, using a variety of methods, can \nmore fully articulate the power of health lifestyles \nfor understanding how advantage is perpetuated \nacross lives and generations and how health life-\nstyles themselves are a tool that creates and props \nup inequalities (Korp 2008). Class-advantaged chil-\ndren\u2019s health lifestyles in our communities involve \nlearning specific behaviors that community norms \ndeem \u201chealthy\u201d and worthy of reward, which \nbecome embodied through thinness and fitness. \nThey also entail developing identities organized \naround the fundamental importance of expansive \nunderstandings of health and health behaviors, \nwhich are expressed in narratives that link health to \ndiscipline and moral worth. It is this combination of \ncharacteristics, beyond specific health behaviors, \nthat propels children toward advantaged futures.\n\nMollborn et al.\t\n15\nUnderstanding health lifestyles in this way \ncould spur more effective and appropriate policy \nefforts. Changing people\u2019s health behaviors is noto-\nriously difficult\u2014in part because a target behavior \nlikely combines with other behaviors within a \nhealth lifestyle that is also comprised of identities, \nnarratives, norms, and understandings of health. \nReturning to Brittany, her behaviors have been \ncarefully planned and justified by Dan and Mary, \nwho relate their chosen lifestyle for Brittany to their \nown traditional, family-oriented identities and their \nmoral worth as parents. Policy efforts from \nBrittany\u2019s community or school to change one of \nher health behaviors are unlikely to succeed unless \nthey engage with her broader health lifestyle, its \nlinks to both her present and future, her parents\u2019 \ndeep identity investment, her own identity expres-\nsion, and the lifestyle\u2019s embeddedness within local-\nized community norms. A successful effort to \nchange Brittany\u2019s behavior must therefore move \nwell beyond the individual child and her family\u2014\nwhich is harder to implement but potentially much \nmore effective. This empirical reality could trans-\nlate into starkly different policy approaches to inter-\nvening in social inequalities and children\u2019s health.\nAcknowledgments\nWe thank Joshua Goode, Fred Pampel, Richard Jessor, \nElizabeth Lawrence, Olowudara Oloyede, Kevin Le, \nAndrew Bennett, and Amber Bunner.\nFunding\nThe authors disclosed receipt of the following financial \nsupport for the research, authorship, and/or publication of \nthis article: This study was supported by National Science \nFoundation (NSF) Grants SES 1423524 and 1729463. We \nalso thank the Eunice Kennedy Shriver National Institute \nof Child Health and Human Development (NICHD)-\nfunded University of Colorado Population Center (P2C \nHD066613) and the Lund University Centre for Economic \nDemography for development, administrative, and/or \ncomputing support. The content is solely the responsibility \nof the authors and does not necessarily represent the offi-\ncial views of the NSF, NICHD, or the National Institutes.\nORCID iD\nStefanie Mollborn \n https://orcid.org/0000-0002-6683-9146\nReferences\nAbel, Thomas. 2008. \u201cCultural Capital and Social \nInequality in Health.\u201d Journal of Epidemiology \nand Community Health 62(7):e13. doi:10.1136/\njech.2007.066159.\nadams, jimi, Elizabeth M. Lawrence, Joshua A. Goode, \nDavid R. Schaefer, and Stefanie Mollborn. 2021. \n\u201cPeer Network Processes in Adolescents\u2019 Health \nLifestyles.\u201d Journal of Health and Social Behavior \n63(1):125\u201341.\nAugustine, Jennifer March, Shannon E. Cavanagh, and \nRobert Crosnoe. 2009. \u201cMaternal Education, Early \nChild Care, and the Reproduction of Advantage.\u201d \nSocial Forces 88(1):1\u201329.\nAugustine, Jennifer March, Kate C. Prickett, and Rachel \nTolbert Kimbro. 2017. \u201cHealth-Related Parenting \namong U.S. Families and Young Children\u2019s Physical \nHealth.\u201d Journal of Marriage and Family 79(3):816\u201332.\nBourdieu, Pierre. 1986a. Distinction: A Social Critique \nof the Judgement of Taste. London, UK: Routledge \n& Kegan Paul.\nBourdieu, Pierre. 1986b. \u201cThe Forms of Capital.\u201d Pp. \n241\u221258 in Handbook of Theory and Research for the \nSociology of Education, edited by J. G. Richardson. \nNew York, NY: Greenwood Press.\nBrown-Saracino, Japonica. 2015. \u201cHow Places Shape \nIdentity: The Origins of Distinctive LBQ Identities \nin Four Small U.S. Cities.\u201d American Journal of \nSociology 121(1):1\u201363.\nCalarco, Jessica McCrory. 2014. \u201cCoached for the \nClassroom: \nParents\u2019 \nCultural \nTransmission \nand \nChildren\u2019s \nReproduction \nof \nEducational \nInequalities.\u201d \nAmerican \nSociological \nReview \n79(5):1015\u201337.\nCalarco, \nJessica \nMcCrory. \n2018. \nNegotiating \nOpportunities: How the Middle Class Secures \nAdvantages in School. New York, NY: Oxford \nUniversity Press.\nChin, Tiffani, and Meredith Phillips. 2004. \u201cSocial \nReproduction and Child-Rearing Practices: Social \nClass, Children\u2019s Agency, and the Summer Activity \nGap.\u201d Sociology of Education 77(3):185\u2013210.\nChristensen, Pia. 2004. \u201cThe Health-Promoting Family: \nA Conceptual Framework for Future Research.\u201d \nSocial Science & Medicine 59(2):377\u201387.\nCockerham, William C. 2005. \u201cHealth Lifestyle Theory \nand the Convergence of Agency and Structure.\u201d \nJournal of Health and Social Behavior 46(1):51\u201367.\nCockerham, William C. 2013. \u201cBourdieu and an Update \nof Health Lifestyle Theory.\u201d Pp. 127\u221254 in Medical \nSociology on the Move: New Directions in Theory, edited \nby W. C. Cockerham. Dordrecht, Netherlands: Springer.\nCockerham, William C. 2023. \u201cHealth Lifestyle Theory \nin a Changing Society: The Rise of Infectious \nDiseases and Digitalization.\u201d Journal of Health and \nSocial Behavior 64(3):437\u201351.\nCockerham, William C., Brian P. Hinote, Pamela \nAbbott, and Christian Haerpfer. 2004. \u201cHealth \nLifestyles in Central Asia: The Case of Kazakhstan \nand Kyrgyzstan.\u201d Social Science & Medicine \n59(7):1409\u201321.\nCockerham, William C., Alfred R\u00fctten, and Thomas \nAbel. 1997. \u201cConceptualizing Contemporary Health \nLifestyles: Moving beyond Weber.\u201d Sociological \nQuarterly 38(2):321\u201342.\n\n16\t\nJournal of Health and Social Behavior 66(1) \nCollett, Jessica L., Kelcie Vercel, and Olevia Boykin. \n2015. \u201cUsing Identity Processes to Understand \nPersistent Inequality in Parenting.\u201d Social Psychology \nQuarterly 78(4):345\u201364.\nCooper, Marianne. 2014. Cut Adrift: Families in Insecure \nTimes. Berkeley: University of California Press.\nElder, Glen H., Jr. 1994. \u201cTime, Human Agency, and \nSocial Change: Perspectives on the Life Course.\u201d \nSocial Psychology Quarterly 57(1):4\u201315.\nElliott, Sinikka, and Sarah Bowen. 2018. \u201cDefending \nMotherhood: Morality, Responsibility, and Double \nBinds in Feeding Children.\u201d Journal of Marriage \nand Family 80(2):499\u2013520.\nEriksen, Ingunn Marie, Kari Stefansen, Tonje Fjogstad \nLangnes, and Kristin Walseth. 2024. \u201cThe Formation \nof Classed Health Lifestyles during Youth: A Two-\nGenerational, Longitudinal Approach.\u201d Sociology of \nHealth & Illness 46(1):137\u201352.\nFielding-Singh, Priya. 2019. \u201cYou\u2019re Worth What You \nEat: Adolescent Beliefs about Healthy Eating, \nMorality, and Socioeconomic Status.\u201d Social Science \n& Medicine 220:41\u201348.\nFomby, Paula, and Kelly Musick. 2018. \u201cMothers\u2019 Time, \nthe Parenting Package, and Links to Healthy Child \nDevelopment.\u201d Journal of Marriage and Family \n80(1):166\u201381.\nFrohlich, Katherine L., and Louise Potvin. 1999. \n\u201cCollective Lifestyles as the Target for Health \nPromotion.\u201d Canadian Journal of Public Health \n90(Suppl. 1):S11\u201314.\nGage-Bouchard, Elizabeth A. 2017. \u201cCulture, Styles \nof Institutional Interactions, and Inequalities in \nHealthcare Experiences.\u201d Journal of Health and \nSocial Behavior 58(2):147\u201365.\nHays, Sharon. 1996. The Cultural Contradictions of \nMotherhood. New Haven, CT: Yale University Press.\nHolstein, James A., and Jaber F. Gubrium. 1995. The \nActive Interview. Vol. 37. Thousand Oaks, CA: Sage.\nKing, Marissa D., Jennifer Jennings, and Jason \nM. \nFletcher. \n2014. \n\u201cMedical \nAdaptation \nto \nAcademic Pressure: Schooling, Stimulant Use, and \nSocioeconomic Status.\u201d American Sociological \nReview 79(6):1039\u201366.\nKorp, Peter. 2008. \u201cThe Symbolic Power of \u2018Healthy \nLifestyles .\u2019\u201dHealth Sociology Review 17(1):18\u201326.\nKrueger, Patrick M., Tajudaullah Bhaloo, and Pauline \nVaillancourt Rosenau. 2009. \u201cHealth Lifestyles in \nthe United States and Canada: Are We Really So \nDifferent?\u201d Social Science Quarterly 90(5):1380\u2013402.\nLandes, Scott D., and Richard A. Settersten, Jr. 2019. \n\u201cThe Inseparability of Human Agency and Linked \nLives.\u201d Advances in Life Course Research 42:1\u201310.\nLankes, Jane. 2022. \u201cNegotiating \u2018Impossible\u2019 Ideals: \nLatent Classes of Intensive Mothering in the United \nStates.\u201d Gender & Society 36(5):677\u2013703.\nLareau, Annette. 2011. Unequal Childhoods: Class, \nRace, and Family Life. Berkeley: University of \nCalifornia Press.\nLareau, Annette, Shani Adia Evans, and April Yee. \n2016. \u201cThe Rules of the Game and the Uncertain \nTransmission of Advantage: Middle-Class Parents\u2019 \nSearch for an Urban Kindergarten.\u201d Sociology of \nEducation 89(4):279\u201399.\nLee, Chung Gun, Dong-Chul Seo, Susan E. Middlestadt, \nand Hsien-Chang Lin. 2015. \u201cDoes the Relationship \nbetween Cigarette Smoking and Other Key Health \nBehaviors Vary by Geographic Area among U.S. \nYoung Adults? A Multilevel Analysis.\u201d International \nJournal of Behavioral Medicine 22(4):481\u201388.\nLink, Bruce G., and Jo Phelan. 1995. \u201cSocial Conditions \nas Fundamental Causes of Disease.\u201d Journal of \nHealth and Social Behavior 35:80\u201394.\nLofland, John, and Lyn H. Lofland. 2006. Analyzing \nSocial Settings. Long Grove, IL: Waveland Press.\nLuna, Jessie K. 2019. \u201cThe Ease of Hard Work: \nEmbodied Neoliberalism among Rocky Mountain \nFun Runners.\u201d Qualitative Sociology 42(2):251\u201371.\nMaralani, Vida, and Camille Portier. 2021. \u201cThe \nConsolidation of Education and Health in Families.\u201d \nAmerican Sociological Review 86(4):670\u201399.\nMasters, Ryan K., Robert A. Hummer, and Daniel A. \nPowers. 2012. \u201cEducational Differences in U.S. \nAdult Mortality: A Cohort Perspective.\u201d American \nSociological Review 77(4):548\u201372.\nMcGarrol, Sarah. 2020. \u201cContextualizing Lifestyles: \nHow Socially Contrasting Places in Fife, Scotland \nInfluence Lay Understandings of Lifestyle and \nHealth Behaviours in Relation to Coronary Heart \nDisease.\u201d Health & Place 66:102432. doi:10.1016/j.\nhealthplace.2020.102432.\nMilkie, Melissa A., and Catharine H. Warner. 2014. \n\u201cStatus Safeguarding: Mothers\u2019 Work to Secure \nChildren\u2019s Place in the Social Hierarchy.\u201d Pp. 66\u201385 \nin Intensive Mothering: The Cultural Contradictions \nof Modern Motherhood, edited by L. R. Ennis. \nToronto, Canada: Demeter Press.\nMirowsky, John, and Catherine E. Ross. 2015. \u201cEducation, \nHealth, and the Default American Lifestyle.\u201d Journal \nof Health and Social Behavior 56(3):297\u2013306.\nMollborn, Stefanie, Laurie James-Hawkins, Elizabeth \nLawrence, and Paula Fomby. 2014. \u201cHealth \nLifestyles in Early Childhood.\u201d Journal of Health \nand Social Behavior 55(4):386\u2013402.\nMollborn, Stefanie, Elizabeth M. Lawrence, and Jarron \nM. Saint Onge. 2021. \u201cContributions and Challenges \nin Health Lifestyles Research.\u201d Journal of Health \nand Social Behavior 62(3):388\u2013403.\nMollborn, \nStefanie, \nand \nAdenife \nModile. \n2022. \n\u201c\u2018Dedicated to Being Healthy\u2019: Young Adults\u2019 \nDeployments \nof \nHealth-Focused \nCultural \nCapital.\u201d Social Science & Medicine 293:114648. \ndoi:10.1016/j.socscimed.2021.114648.\nMollborn, Stefanie, Bethany Rigles, and Jennifer A. \nPace. 2021. \u201c\u2018Healthier Than Just Healthy\u2019: Families \nTransmitting Health as Cultural Capital.\u201d Social \nProblems 68(3):574\u201390.\n\nMollborn et al.\t\n17\nMuennig, Peter A., Megan Reynolds, David S. Fink, \nZafar Zafari, and Arline T. Geronimus. 2018. \n\u201cAmerica\u2019s Declining Well-Being, Health, and Life \nExpectancy: Not Just a White Problem.\u201d American \nJournal of Public Health 108(12):1626\u201331.\nNelson, Margaret K. 2010. Parenting out of Control: \nAnxious Parents in Uncertain Times. New York, NY: \nNYU Press.\nPace, Jennifer A., Stefanie Mollborn, and Bethany Rigles. \n2022. \u201c\u2018I Just Want Her to Be Happy and Healthy\u2019: \nDeploying Broad Understandings of Health in Class-\nPrivileged Parenting.\u201d SSM\u2013Qualitative Research in \nHealth 2:100183. doi:10.1016/j.ssmqr.2022.100183.\nPugh, Allison J. 2009. Longing and Belonging: Parents, \nChildren, \nand \nConsumer \nCulture. \nBerkeley: \nUniversity of California Press.\nPugh, Allison J. 2014. \u201cThe Theoretical Costs of Ignoring \nChildhood: Rethinking Independence, Insecurity, \nand Inequality.\u201d Theory and Society 43(1):71\u201389.\nPuka, Klajdi, Charlotte Buckley, Nina Mulia, Aur\u00e9lie M. \nLasserre, J\u00fcrgen Rehm, and Charlotte Probst. 2022. \n\u201cEducational Attainment and Lifestyle Risk Factors \nAssociated with All-Cause Mortality in the U.S.\u201d \nJAMA Health Forum 3(4):e220401. doi:10.1001/\njamahealthforum.2022.0401.\nReich, Jennifer A. 2016. Calling the Shots: Why Parents \nReject Vaccines. New York, NY: New York \nUniversity Press.\nSaez, Emmanuel. 2008. \u201cStriking It Richer: The \nEvolution of Top Incomes in the United States.\u201d \nPathways Magazine Winter:6\u20137.\nSchneider, Daniel, Orestes P. Hastings, and Joe LaBriola. \n2018. \u201cIncome Inequality and Class Divides in \nParental \nInvestments.\u201d \nAmerican \nSociological \nReview 83(3):475\u2013507.\nShirani, Fiona, Karen Henwood, and Carrie Coltart. 2012. \n\u201cMeeting the Challenges of Intensive Parenting \nCulture: Gender, Risk Management, and the Moral \nParent.\u201d Sociology 46(1):25\u201340.\nStets, Jan E., and Peter J. Burke. 2000. \u201cIdentity Theory \nand Social Identity Theory.\u201d Social Psychology \nQuarterly 63(3):224\u201337.\nSwidler, Ann. 2001. Talk of Love: How Culture Matters. \nChicago, IL: University of Chicago Press.\nTimmermans, \nStefan, \nand \nIddo \nTavory. \n2012. \n\u201cTheory Construction in Qualitative Research: \nFrom Grounded Theory to Abductive Analysis.\u201d \nSociological Theory 30(3):167\u201386.\nTsushima, Teresa, and Peter J. Burke. 1999. \u201cLevels, \nAgency, and Control in the Parent Identity.\u201d Social \nPsychology Quarterly 62(2):173\u201389.\nU.S. \nCensus \nBureau. \n2017. \n\u201cData \nProfiles.\u201d \nhttps://data.census.gov/profile?q=United%20\nStates&g=010XX00US.\nWarner, Catharine H. 2010. \u201cEmotional Safeguarding: \nExploring the Nature of Middle-Class Parents\u2019 \nSchool Involvement.\u201d Sociological Forum 25(4): \n703\u201324.\nWeber, Max. [1922]1978. Economy and Society: \nAn Outline of Interpretive Sociology. Berkeley: \nUniversity of California Press.\nWilliamson, Deanna L., Margo Charchuk, Kaysi \nE. Kushner, Berna J. Skrypnek, and Nicole \nY. Pitre. 2018. \u201cFamilies That Do Well: Lay \nConceptualizations of Well-Functioning, Healthy, \nStrong, and Good Families.\u201d Family Relations \n67(4):467\u201382.\nWills, Wendy, Kathryn Backett-Milburn, Mei-Li \nRoberts, and Julia Lawton. 2011. \u201cThe Framing \nof Social Class Distinctions through Family Food \nand Eating Practices.\u201d The Sociological Review \n59(4):725\u201340.\nAuthor Biographies\nStefanie Mollborn, PhD, is a professor of sociology at \nStockholm University, Sweden and an affiliate of the \nInstitute of Behavioral Science at the University of \nColorado Boulder. She researches inequalities, well-\nbeing, and health among young people and families in \nEurope and the United States using quantitative and quali-\ntative methods. Her current research projects focus on \nhealth lifestyles and early life inequalities, sexual minority \nhealth disparities and family structure, the life course \nimpacts of contraceptive access, and intensive parenting \nattitudes.\nJennifer A. Pace, PhD, is a researcher at the Center for \nBehavioral Science Methods at the U.S. Census Bureau. \nHer research areas include family and childhood health, \nidentity construction, and inequalities. Her current proj-\nects focus on children\u2019s health lifestyles, vaccination in \npregnancy, and women\u2019s experiences with family care-\nwork during health crises.\nBethany Rigles, PhD, is the director of research and eval-\nuation and co-owner of Good Nutrition Ideas. She uses \nmixed research methodologies to explore health dispari-\nties among diverse people with disabilities and to create \nhealth interventions for vulnerable populations.\n\n\n    Scientific Research Findings:", "answer": "\u2022 Through health lifestyles, parents attempted to shape children\u2019s everyday lives in multifaceted ways that they imagined would affect their long-term well-being and socioeconomic success.\n\u2022 Class-privileged parenting involved constructing an individualized health lifestyle reliant on expansive understandings of health and composed of parents\u2019 identities and narratives, children\u2019s health behaviors and identity expressions, and community norms.\n\u2022 Children\u2019s predominant health lifestyles varied by focus on parent versus child identity expression and children\u2019s future achievements versus present well-being. An \u201celite achievement\u201d lifestyle, with structured exercise, limited screen time, and increased body size regulation, focused on parent identity expression and future child achievements, predominated in upper-middle-class settings, and had some downsides for children.", "id": 63}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465241230839\nJournal of Health and Social Behavior\n2024, Vol. 65(3) 324\u00ad\u2013339\n\u00a9 American Sociological Association 2024\nDOI: 10.1177/00221465241230839\njournals.sagepub.com/home/hsb\nOriginal Article\nIn the past two decades, states have emerged as \nimportant players in the immigration policy arena, \npassing immigration legislation at a scale not seen in \nover a century (Chacon 2019; Gulasekaram and \nRamakrishnan 2015). State legislatures have passed \nmore than 4,000 immigration-related laws and reso-\nlutions since 2005, spanning domains including pub-\nlic benefits, immigration enforcement, driver\u2019s \nlicenses and identification systems, education, and \nemployment (Johnston and Morse 2013; Morse \n2019). While some of these policies aim to drive out \nimmigrants by enacting new immigration enforce-\nment schemes and restricting immigrants\u2019 access to \nservices and benefits, others seek to support immi-\ngrant integration by expanding protections from \ndeportation and facilitating access to health care, \nsocial services, higher education, and other resources.\nThese policies have important health implica-\ntions for the well-being of immigrants and their \nchildren. For one, the expansion of public health \ninsurance to immigrants is associated with greater \nhealth insurance enrollment and preventive care \nutilization and in some cases, improved health out-\ncomes (Drewry et al. 2015; Rosenberg et al. 2022; \nSchut and Boen 2022; Swartz et al. 2017, 2019; \nWherry et al. 2017). Furthermore, a growing body \nof research shows that non-health-related state \nimmigration laws can also impact the health and \npsychological well-being of immigrants and their \nfamilies (Crookes et al. 2021; Perreira and Pedroza \n2019; Philbin et al. 2018).\nDespite growing attention to the links between \nstate immigration policies and health, we have a \n1230839 HSBXXX10.1177/00221465241230839Journal of Health and Social BehaviorMoinester and Stanhope\nresearch-article2024\n1Washington University in St. Louis, St. Louis, MO, USA\n2Emory University, Atlanta, GA, USA\nCorresponding Author:\nMargot Moinester, Washington University in St. Louis, \nBox 1112, St. Louis, MO 63130, USA. \nEmail: mmoinester@wustl.edu\nExtending Driver\u2019s Licenses to \nUndocumented Immigrants: \nComparing Perinatal Outcomes \nFollowing This Policy Shift\nMargot Moinester1\n and Kaitlyn K. Stanhope2\nAbstract\nResearch shows that restrictive immigration policies and practices are associated with poor health, but \nfar less is known about the relationship between inclusive immigration policies and health. Using data \nfrom the United States natality files, we estimate associations between state laws granting undocumented \nimmigrants access to driver\u2019s licenses and perinatal outcomes among 4,047,067 singleton births to \nMexican and Central American immigrant birthing people (2008\u20132021). Fitting multivariable log binomial \nand linear models, we find that the implementation of a license law is associated with improvements in \nlow birthweight and mean birthweight. Replicating these analyses among U.S.-born non-Hispanic White \nbirthing people, we find no association between the implementation of a license law and birthweight. \nThese findings support the hypothesis that states\u2019 extension of legal rights to immigrants improves the \nhealth of the next generation.\nKeywords\ndriver\u2019s licenses, Hispanic or Latino, immigration policy, perinatal health, United States\n\nMoinester and Stanhope\t\n325\nlimited understanding of whether and to what \nextent the enactment of inclusive immigration poli-\ncies at the state level can improve health. Extant \nscholarship largely focuses on the deleterious \nhealth effects of restrictive state-level immigration \npolicies (Hardy et al. 2012; Torche and Sirois 2019; \nWang and Kaushal 2019; Young, Crookes, and \nTorres 2022) or on the health effects of states\u2019 \nimmigration policy climate, a measure that captures \na range of immigration-related policies enacted in a \ngiven \nstate \n(Dondero \nand \nAltman \n2020; \nHatzenbuehler et al. 2017; Schut and Boen 2022; \nStanhope et al. 2019; Sudhinaraset et al. 2021; \nYoung et al. 2019). Among the studies that examine \nhow specific state-level inclusive immigration poli-\ncies impact health, the focus has primarily been on \npolicies that extend access to public health insur-\nance to immigrants (for notable exceptions, see \nKoball, Kirby, and Hartig 2022; Potochnick, May, \nand Flores 2019; Schut and Boen 2022). How non-\nhealth-related inclusive immigration policies shape \nthe health and well-being of immigrants and their \nchildren remains far less clear. Additionally, to what \nextent individual inclusive policies can positively \naffect health independent of states\u2019 broader policy \nclimate is also not well understood.\nThe present study expands understanding of the \nlinks between state immigration policies and health \nby leveraging 14 years of data from the U.S. natal-\nity files to assess how state laws that authorize \nundocumented immigrants to obtain driver\u2019s \nlicenses are associated over time with key indica-\ntors of perinatal health\u2014early entry into prenatal \ncare, gestational age at delivery, and birthweight\u2014\namong Mexican and Central American immigrants. \nIdentification (ID) laws represent a prominent \ndomain of immigration policymaking with impor-\ntant health implications. Government-issued IDs \nare required to access a range of economic and \nmaterial resources critical for health, including \nbank accounts, utilities, prescription medications, \nhousing, and safety net programs (de Graauw 2014; \nLeBr\u00f3n et al. 2018). A driver\u2019s license is a form of \ngovernment-issued ID that has particularly strong \nhealth-promoting potential for immigrants. For one, \ngiven the spatially decentralized character of much \nof the United States and limited public transit sys-\ntems (Jones 2008), a driver\u2019s license provides an \nundocumented immigrant with not only the identi-\nfication needed to obtain the aforementioned \nresources but also the mobilty often required to \naccess these resources and occupational opportuni-\nties. Furthermore, given the recent devolution of \nimmigration enforcement to local law enforcement, \ntraffic stops now constitute a key pathway into the \nU.S. immigration enforcement system (Moinester \n2019; Waslin 2013). Access to a driver\u2019s license \nsubstantially reduces the chances that a routine traf-\nfic stop leads to arrest and deportation, thereby \naddressing a considerable source of stress for \nundocumented immigrants and their families (Kline \n2017; Rhodes et al. 2015).\nAt the population level, Mexican and Central \nAmerican immigrants are especially likely to expe-\nrience the health benefits of this policy change. \nThrough its immigration policies and enforcement \nactions over the past several decades, the U.S. fed-\neral government \u201cirregularized\u201d long-established \npatterns of migration from Latin America (Menj\u00edvar \nand Kanstroom 2013) and, in turn, actively con-\nstructed \u201cillegality\u201d as a racialized social condition \nthat is essentially synonymous with \u201cMexicanness\u201d \nand Latino-ness (Armenta and Vega 2017; De \nGenova 2002).1 As a result of these policy changes, \nLatine immigrants now comprise over 70% of the \nU.S. undocumented population, with 67% of this \npopulation hailing from Mexico and Central \nAmerica (Migration Policy Institute 2019).2 \nAdditionally, Latine immigrants are disproportion-\nately stopped by the police and turned over to immi-\ngration authorities when a driver\u2019s license cannot be \nfurnished (Armenta and Vega 2017; Gardner and \nKohli 2009; Moinester 2019). Risk of deportation \nfollowing driving without a valid license is thus \nespecially pronounced for this population.\nFindings from this study highlight the potential \nof an individual state policy to positively shape the \nlives of Mexican and Central American immigrants \nand their children amid a highly conflictual federal \nand state immigration policy climate. Over the past \ntwo decades, deportations from the United States \nhave reached record levels, and anti-immigrant \nrhetoric has intensified (Callaghan et al. 2019; \nGolash-Boza 2015). Many states have simultane-\nously passed laws that intensify surveillance and \nimmigration enforcement (Young and Wallace \n2019). Within this context, we find that states\u2019 \nmove to extend access to a driver\u2019s license to \nundocumented immigrants is associated with \nimprovements in low birthweight and mean birth-\nweight for Mexican and Central American immi-\ngrants. We further find no association between \nthese laws and birthweight among U.S.-born non-\nHispanic White birthing people\u2014a group for whom \nno effect from the law was expected. Given that \nbirthweight is a critical measures of early develop-\nment (Torche and Conley 2015), this study\u2019s find-\nings underscore how states\u2019 extension of legal \nrights to immigrants can improve the health of the \nnext generation.\n\n326\t\nJournal of Health and Social Behavior 65(3) \nBackground\nTheorizing Driver\u2019s License Laws \nRelationship to Perinatal Health\nAs Young and Wallace\u2019s (2019) health and policy \ncontext framework elucidates, state-level immigra-\ntion policies can affect health by heightening the \ncriminalization of immigrants or by increasing \nimmigrants\u2019 integration into U.S. society. Criminal\u00ad\nization policies can deleteriously affect health by \nincreasing stress and experiences of discrimina-\ntion. Alternately, integration policies can improve \nhealth by facilitating access to health-promoting \nresources.\nAccess to driver\u2019s licenses for undocumented \nimmigrants is a dynamic area of immigration poli-\ncymaking that shapes both the criminalization of \nimmigrants and their integration into U.S. society. \nSince 2013, 16 states and the District of Columbia \nhave enacted legislation that expands access to driv-\ner\u2019s licenses to undocumented immigrants. In total, \n19 states and the District of Columbia currently \nextend driving privileges to undocumented immi-\ngrants (National Conference of State Legislatures \n2023). As illustrated in our conceptual model in \nFigure 1, prior scholarship and Young and Wallace\u2019s \n(2019) framework suggest two key mechanisms \nthrough which these license laws may improve peri-\nnatal health: (1) reduced stress and (2) increased \naccess to financial and health care resources. We \ndetail each of these mechanisms in the following.\nDriver\u2019s Licenses and Stress\nThe expansion of the legal right to a driver\u2019s license \nmay improve Mexican and Central American immi-\ngrants\u2019 perinatal health by lessening deportation \nfears and subsequent stress. Exposure to stress prior \nto and during pregnancy can increase risk of adverse \nbirth outcomes, including low birthweight and pre-\nterm birth. Chronic stress leads to wear and tear on \nphysiologic systems, altering the body\u2019s neuroendo-\ncrine and immune systems, and possibly resulting in \nsubcellular weathering or accelerated aging (Hobel, \nGoldstein, and Barrett 2008; Kramer and Hogue \n2009; Traylor et al. 2020). Research links restrictive \nimmigration policies and enforcement actions to \nelevated risk of preterm birth and lower birthweight \namong Latine birthing people, with stress operating \nas the presumed mechanism (Novak, Geronimus, \nand Martinez-Cardoso 2016; Ro, Bruckner, and \nDuquette-Rury 2020; Stanhope et al. 2019; Torche \nand Sirois 2019).\nDriving without a license is a source of stress for \nmany undocumented immigrants and their families. \nIn large swaths of the country, a car is needed to \naccess jobs, schools, health care services, places of \nworship, and basic necessities such as grocery \nstores (Sanchez, Stolz, and Ma 2003). Yet for immi-\ngrants unable to obtain a driver\u2019s license, driving \nproduces risk of police discovery, arrest, and even-\ntual deportation and, in turn, chronic worry and \nstress (Armenta 2017; Kline 2017; Stuesse and \nColeman 2014).\nFigure 1.\u2002 Conceptual Model of the Relationship between Driver\u2019s License Laws and Perinatal Health.\n\nMoinester and Stanhope\t\n327\nRisk of arrest and deportation following a traffic \nstop has intensified for immigrants over the past 15 \nyears\u2014a period marked by enhanced immigration \nenforcement in the U.S. interior and increased coop-\neration between local law enforcement and federal \nimmigration authorities (Coleman 2012; Moinester \n2019). Because of these dual dynamics, routine traffic \nstops now function as a key funnel into the immigra-\ntion enforcement system (Armenta 2017). Racialized \npolicing practices and the common elision of race and \nimmigration status have led Latine drivers, specifi-\ncally, to disproportionately be stopped by the police \nand arrested when found driving without a U.S. \nlicense (Armenta and Vega 2017; Moinester 2019).\nAltogether, evidence suggests that living in a \nstate without legal access to driving may act as a \nchronic stressor for undocumented Mexican and \nCentral American immigrants and their families, \nwith implications for perinatal health. Pregnant per-\nsons may experience stress prior to pregnancy, fear-\ning for themselves or family members that a traffic \nstop will result in arrest and deportation. By reduc-\ning the criminalization of immigrants, driver\u2019s \nlicense laws may lessen deportation fears and subse-\nquent stress, potentially improving birth outcomes.\nDriver\u2019s Licenses and Access to \nResources\nExtending driver\u2019s licenses to undocumented immi-\ngrants may also improve immigrants\u2019 perinatal \nhealth by facilitating access to a diverse set of health-\npromoting resources, including financial resources \nand health care. Economic disadvantage is associ-\nated with adverse birth outcomes (Auger et al. 2012; \nde Graaf, Steegers, and Bonsel 2013; Forde et al. \n2019), and research shows that acute improvements \nin financial well-being may improve birth outcomes \n(Glassman et al. 2013; Hoynes, Miller, and Simon \n2015). A government-issued ID, such as a driver\u2019s \nlicense, provides immigrants with the required docu-\nmentation needed to access various financial \nresources, including setting up bank accounts, cash-\ning checks, and applying for loans (de Graauw 2014; \nLeBr\u00f3n et al. 2018). A driver\u2019s license also facilitates \nthe mobility often required to obtain these resources.\nFurthermore, by making it easier for undocu-\nmented immigrants to commute to work, driver\u2019s \nlicense laws may also lead to improved access to \njobs and potentially, increased earnings. The one \nstudy to date that assesses this possibility finds that \nthe availability of driver\u2019s licenses leads men who \nare likely to be undocumented to increase their \nweekly work hours and therefore their take-home \npay. This study further finds that availability of driv-\ner\u2019s licenses broadened the set of job opportunities \nfor undocumented immigrants (Amuedo-Dorantes, \nArenas-Arroyo, and Sevilla 2020). Given the strong \nassociation between financial hardship and stress \n(O\u2019Neill et al. 2005), the improved financial well-\nbeing resulting from access to a driver\u2019s license may \nalso impact perinatal health by reducing stress.\nAccess to driver\u2019s licenses may also improve \nimmigrant\u2019s perinatal health by increasing access to \ntimely prenatal care. Prenatal care functions as a \nmechanism for identifying and managing risk fac-\ntors that may contribute to poor pregnancy out-\ncomes (Conway and Kutinova 2006). Early entry \ninto prenatal care is particularly important for birth-\ning people without regular access to preventive care \noutside of pregnancy to identify and manage under-\ndiagnosed chronic conditions.\nDisparities by nativity and ethnicity in prenatal \ncare utilization exist (Fuentes-Afflick, Hessol, and \nP\u00e9rez-Stable 1998; Goldfarb et al. 2017; Korinek \nand Smith 2011). Although limited access to insur-\nance is likely a primary driver of these disparities \n(Fuentes-Afflick et al. 2006; Korinek and Smith \n2011; Reed et al. 2005), limited access to reliable \ntransportation options, including restrictive driver\u2019s \nlicense policies, is an important contributor, too \n(Koball et al. 2022; Korinek and Smith 2011; \nRhodes et al. 2015). Altogether, this body of work \nsuggests that extending driving privileges to undoc-\numented individuals may faciliate immigrants\u2019 \nincorporation into society, increasing financial \nresources and timely entry into prenatal care, with \npositive implications for Mexican and Central \nAmerican immigrants\u2019 birth outcomes.\nStudy Overview and Hypotheses\nWe examine the effect of enacting a state law that \nextends driver\u2019s license access to undocumented \nimmigrants on key indicators of perinatal health\u2014\npreterm birth, birthweight, and timing of entry into \nprenatal care. Based on our conceptual model \n(Figure 1), we hypothesize that the implementation \nof this law would improve perinatal outcomes for \nforeign-born Mexican and Central American birth-\ning people exposed to this legislation prior to gesta-\ntion both by reducing deportation worry and \nsubsequent stress and increasing access to financial \nand health care resources. We further hypothesize \nthat the effects may be strongest for those exposed \nto the law for longer periods prior to conception, \nallowing potential decreases in chronic stress and \nincreases in resources to take full effect.\n\n328\t\nJournal of Health and Social Behavior 65(3) \nData And Methods\nData and Sample\nWe used data from the U.S. restricted use natality files, \n2008 to 2021 (National Vital Statistics System 2021) \nand linked these data to county- and state-level charac-\nteristics using the resident county and year. We started \nthe analysis in 2008 because this is the year when \nSecure Communities was launched and funding for \n287(g) dramatically expanded\u2014two programs that \ninvolve local law enforcement in immigration enforce-\nment and have been critical to the rise in deportations \nfollowing a traffic stop (Moinester 2019). These pro-\ngrams contribute to immigrants\u2019 fears of driving with-\nout a license and, in turn, may be consequential for \nperinatal health (Rhodes et al. 2015).\nThe target population for this study was \nMexican and Central American birthing people \nwho were likely to be impacted by the passage of a \nlicense law. We believe that undocumented birthing \npeople would be most strongly impacted because \nthey would personally receive an additional privi-\nlege and reduction in deportation risk because of \nthe implementation of this law. The impact of these \nlaws may also spill over to birthing people who are \nnot themselves undocumented but whose partner or \nother loved ones are. These persons may experience \na reduction in stress because of the passage of the \nlaws or increase in financial resources due to \nchanges for others in their household. Legal status \nis not included in birth certificate data. However, to \nmake the study population more likely to be undoc-\numented or have undocumented partners or loved \nones, we restricted the analysis to birthing people \nborn in Central America or Mexico. These regions \naccount for 67% of the U.S. undocumented popula-\ntion (Migration Policy Institute 2019). As a sensi-\ntivity analysis, we stratified by maternal education \nto determine whether associations are stronger \namong individuals without a high school education, \nwho are more likely to be undocumented (Passel \nand Cohn 2019).\nNatality data from 2008 to 2021 included infor-\nmation from both the standard and revised birth \ncertificates. The revised birth certificate was intro-\nduced in 2003 but not adopted by all states until \n2016. The standard birth certificate contains limited \ninformation, and for some variables, such as entry \ninto prenatal care, the information is of lower qual-\nity (Gregory et al. 2019). For this reason, we used \ninformation from the standard and revised birth cer-\ntificates for birthweight and preterm birth analyses \nbut only information from the revised birth certifi-\ncate for entry into prenatal care analyses.\nOf all singleton births conceived between 2008 \nand 2020 (49,262,494), we excluded births with \nmissing or implausible (<20 or >43 weeks) gesta-\ntional age (excluded 2,404,419), missing or implau-\nsible (<500 or >5000 grams) birthweight (excluded \n119,191), missing or invalid resident state or county \n(excluded 49,939), or missing nativity (excluded \n125,925). We then restricted to foreign-born birth-\ning people born in Mexico or Central America \n(excluded 35,903,130). We further excluded other-\nwise eligible births in Washington, Utah, and New \nMexico, which had a law in place prior to our study \nperiod (excluded 170,648), and in Massachusetts, \nRhode Island, Virginia, and Minnesota, which \nimplemented a law following our study period \n(excluded 138,519) for an analytic data set of \n4,047,067 births. For entry into prenatal care analy-\nses, we further excluded births in states prior to the \nadoption of the revised birth certificate or with \nmissing data on entry into prenatal care (excluded \n251,001). Overall, no included variable was miss-\ning for greater than 7% of observations; the highest \nproportion of observations had missing or implau-\nsible data on gestational age (4.9%). The study \nreceived institutional review board approval.\nIndependent Variable\nWe defined a license law as a law allowing individuals \nto apply for a state driver\u2019s license regardless of immi-\ngration status. We used data from the National \nConference of State Legislatures (2023) to identify \nwhich states have passed a license law, the enactment \ndate, and the date the law went into effect (i.e., imple-\nmentation date). To date, 19 states and the District of \nColumbia have enacted a license law. Thirteen of \nthese laws were implemented during our study period. \nWe excluded Washington, New Mexico, Utah, \nMassachusetts, Rhode Island, Virginia, and Minnesota \nfrom the analysis because license laws in these states \nwere implemented either before or after our study \nperiod and there was no within-state variation in \nexposure to the law. For a full list of these laws, see \nTable S1 in the online version of the article.\nWe classified the exposure status of each birth \nbased on the resident state at delivery and the esti-\nmated date of conception. To calculate estimated \ndate of conception, we used gestational age at deliv-\nery, month, and year of delivery. We assumed 15 as \nthe day of the month for all births because birthday \nis not available in natality data. We considered births \nfor which the law was fully implemented at concep-\ntion as exposed because these individuals would \nhave experienced all benefits of the law (the \n\nMoinester and Stanhope\t\n329\nprivileges of driving and accessing resources that \nrequire a government-issued ID and the reduction in \ndeportation-related stress) for their entire gestation.\nTo assess how the relationship between the \nimplementation of a license law and perinatal \nhealth changes over time, we created a set of binary \nindicators based on the number of months prior to \nconception that the law had been implemented in \nthe resident state for a secondary, lagged analysis. \nFor example, an individual who resided in \nCalifornia (law effective January 1, 2015) at deliv-\nery and conceived in July 2015 in California would \nhave been assigned as exposed at conception and \nexposed for six months prior to conception. \nHowever, an individual who resided in California at \ndelivery and conceived in January 2015 would have \nbeen assigned the primary exposure of exposed at \nconception but not the six-month lag.\nDependent Variables\nWe examined four dependent variables: (1) preterm \nbirth (binary, <37 completed weeks gestation), (2) \nlow birthweight (binary, <2500 grams), (3) birth-\nweight (continuous, grams), and (4) entry into pre-\nnatal care in the first trimester (binary, entered in \nfirst three months). We measured gestational age in \nweeks using the combined last menstrual period and \nphysical estimated gestational age variable (Dietz  \net al. 2014).\nControls\nWe considered several sets of potential confounders \nof the relationship between implementation of a \nlicense law and perinatal outcomes based on a priori \nknowledge. First, we considered whether the com-\nposition of birthing people may vary across places \nand bias the observed association. Second, we con-\nsidered whether economic and demographic charac-\nteristics of place may drive the implementation of a \nlicense law and impact perinatal outcomes. Third, \nwe considered the overall state immigration climate \nas a potential confounder.\nIndividual.\u2002 We included information on parity \n(primiparous or multiparous), age (continuous), and \nspecific country of origin as individual covariates.\nState demographic and economic climate.\u2002 To \ncharacterize differences in demographic and eco-\nnomic characteristics that may act as confounders, \nwe included information on county and state \ncharacteristics. For counties, we included rurality \n(urban or rural, binary), defined using classifica-\ntions from the National Center for Health Statistics \n(2017). For states, we included the following mea-\nsures: the percentage of adults (\u226525 years of age) \nwith a high school education, the percentage of \nadults (\u226516 years of age) who were unemployed, the \npercentage of Latine individuals in the state who \nwere noncitizens, and the percentage of foreign-\nborn Mexican and Central American individuals \nwho were noncitizens. These time-varying variables \nwere created using data from the American Com-\nmunity Survey, three-year estimates for 2008 to \n2009 and five-year estimates for 2010 to 2020.\nImmigration policy and enforcement climate.\u2002 We \ncharacterized differences in counties\u2019 and states\u2019 \npolicy and immigration enforcement climates using \nseveral indicators. At the county level, we assessed \nthe presence of a 287(g) agreement for each year. \nFor example, a 287(g) agreement that went into \neffect in 2010 and ended in 2012 was recorded in \nour data as being present in 2010, 2011, and 2012 \nfor individuals residing in that county at delivery. At \nthe state level, we assessed the annual immigration \ndetainer rate per 100,000 noncitizens. We further \ncapture states\u2019 overall immigration policy climate \nannually using the Immigration Climate Index (ICI; \nPham and Van 2014, 2019). To avoid overlap with \nthe exposure, which is included in the ICI following \nits enactment, we created a modified ICI. For births \nconceived prior to the enactment of the license law, \nthis took the value of the cumulative ICI (2005 to \nconception year), representing all relevant enacted \npolicies in the state. For births conceived following \nthe enactment of the license law, we subtracted 3 \npoints\u2014the value representing the license law\u2014\nfrom the cumulative ICI. As indicators of overall \nsupport for perinatal health in the state, we exam-\nined whether the Children\u2019s Health Insurance Pro-\ngram (CHIP) unborn child option was active in the \nstate, which provides Medicaid coverage to other-\nwise ineligible pregnant immigrants during their \npregnancy, and whether the state had expanded \nMedicaid under the Affordable Care Act.3 Both of \nthese variables were year specific. Data on detainers \nwere provided by the Transactional Records Access \nClearinghouse (2024), and data on the noncitizen \npopulation came from the American Community \nSurvey. Data on state adoption of the CHIP unborn \nchild option and Medicaid expansion came from the \nKaiser Family Foundation (Brooks et al. 2022; Kai-\nser Family Foundation 2023).\n\n330\t\nJournal of Health and Social Behavior 65(3) \nComparison Group\nWe considered two potential comparison groups to \nrepresent the counterfactual comparison. First, indi-\nviduals residing in states that never enacted license \nlaws were a potential comparison group and offer \nthe advantage of controlling for secular trends. \nHowever, after descriptive analyses (Table 2), we \nobserved differences in enacting and nonenacting \nstates at baseline that widened over time, including \nfor the ICI, state detainer rate, and Medicaid expan-\nsion status. Thus, we considered this group an inap-\npropriate comparison group. Second, we considered \nindividuals residing in states that would eventually \nenact a license law but had not yet enacted the law \nto be a potential comparison group. This group \noffers the advantage of a better counterfactual com-\nparison for state economic and immigration climate \n(Table 2). Furthermore, due to the large number of \nstates and wide variation in time of enactment of \nlicense laws, it allows for a robust comparison \ngroup. We selected this second group (not yet enact-\ners) as the comparison group.\nAnalytic Approach\nFor our primary analysis, we compared individuals \nresiding in a state with a license law in place (imple-\nmented) at the estimated date of conception to those \nresiding in a state where a license law was not imple-\nmented before the pregnancy but would be imple-\nmented \nfollowing \nthe \npregnancy, \nexcluding \nindividuals partially exposed (e.g., individuals for \nwhom the law was implemented during their preg-\nnancy). We fit multivariable log binomial models for \npreterm birth, low birthweight birth, and entry into \nprenatal care and linear models for birthweight to cal-\nculate risk ratios (or beta estimates) and 95% confi-\ndence intervals (CIs). We fit models including state \nfixed effects to account for unmeasured characteris-\ntics of the state that may have influenced the enact-\nment of a license law. We included covariates \nsequentially and fit four models for each outcome: \n(1) unadjusted (Model 0), (2) individually adjusted \n(Model 1, controlling for age [linear and quadratic], \nparity [binary: primiparous or multiparous], specific \ncountry of origin [Mexico, Honduras, El Salvador, \nGuatemala, or other], state fixed effects, year and \nmonth of conception), and (3) adjusting for state \ndemographic and economic climate (Model 2, add-\ning state percentage of adults [\u226525] with a high \nschool education, percentage of adults [\u226516] who \nwere unemployed, percentage of Latine individuals \nin the state who were noncitizens, percentage of for-\neign-born Mexicans and Central Americans who \nwere noncitizens, and county rurality), and (4) adjust-\ning for the cumulative ICI minus license laws (Model \n3 + adding the adjusted ICI value). Finally, we fit a set \nof models with lagged exposures (6, 12, 18, and 24 \nmonths) to assess how the relationship between \nlicense laws and perinatal health changes over time. \nFor each lagged model, we excluded individuals par-\ntially exposed to that lag (e.g., individuals for whom \nthe law was implemented between 6, 12, 18, or 24 \nmonths prior to conception and the birth date).\nSensitivity Analyses\nWe conducted several sensitivity analyses to inter-\nrogate the impact of analytic choices on our \nobserved results. First, we repeated models among \nnon-Hispanic, U.S.-born White birthing people as a \nplacebo analysis. This group should not have been \nconsiderably impacted by the policy decision to \nextend access to driver\u2019s license to undocumented \nimmigrants. Repeating the analysis for this group \nwould tell us if there were other time-varying pro-\ncesses that may have occurred at the same time and \nimpacted perinatal health generally, although it does \nnot account for processes specific to immigrant \ncommunities (e.g., stress due to changes in national \nimmigration climate during the Trump presidency). \nSecond, we fit models stratifying by individual edu-\ncational attainment (less than high school or high \nschool graduate or higher). Education levels of \nundocumented immigrants are well below those of \ndocumented immigrants (Passel and Cohn 2019). A \nstronger association among individuals with lower \neducation would be consistent with our hypothesis \nthat these laws would have the greatest impact on \nundocumented persons. Finally, to determine the \nimpact that the COVID-19 pandemic may have had \non results, we repeated the analysis restricting to \nbirths occurring in 2008 to 2019.\nResults\nDescriptive Analysis\nTable 1 provides descriptive statistics of Mexican- \nand Central American-born individuals delivering a \nlive-born infant in states that either never enacted a \nlicense law or enacted and implemented one during \nour study period. Statistics are provided for 2008 \n(the first year of data in this sample) and 2020 (the \nlast full year of data in this sample). Overall, pre-\nterm birth rates were lower in states that enacted \nlicense laws compared to those that did not, although \nrates in both groups declined over time. Similarly, \nMexican- and Central American-born birthing \n\nMoinester and Stanhope\t\n331\npeople delivering in states that enacted license laws \nwere more likely to enter prenatal care in the first \ntrimester. Birthweight was similar across groups \nand over time. Other birthing person characteristics \nwere similar across groups (i.e., education, parity, \nmarital status). In both 2008 and 2020, a greater pro-\nportion of birthing people lived in counties with \n287(g) agreements in place in states that never \nTable 1.\u2002 Characteristics of 368,305 Mexican- and Central American-Born Birthing People Delivering \nLive Born, Singleton Infants in States That Have and Have Not Enacted Laws Providing Licenses to \nUndocumented Immigrants, 2008 and 2020, 43 States and the District of Columbia.a\nNever Enacted\nEnacted\n\u2002\n2008\n2020\n2008\n2020\n\u2002\nN = 52,752\nN = 143,416\nN = 55,443\nN = 116,694\nBirthing person characteristicsb\u2002\n% (n)\n% (n)\n% (n)\n% (n)\nPreterm birth\n14.8 (7,779)\n12.5 (17,919)\n13.5 (7,491)\n11.1 (12,946)\nVery preterm birth\n2.3 (1,231)\n1.5 (2,178)\n2.1 (1,144)\n1.3 (1,525)\nLow birthweight\n6.4 (3,372)\n5.7 (8,098)\n6.2 (3,448)\n5.7 (6,597)\nVery low birthweight\n1.3 (683)\n.8 (1,157)\n1.2 (681)\n.8 (939)\nMean birthweight, grams (SD)\n3,283.8 (557.5)\n3,284.6 (521.9)\n3,293.3 (551.2)\n3,298.3 (526.8)\nE\u0007ntered prenatal care in first \ntrimester\n54.8 (28,922)\n56.9 (81,659)\n75.1 (41,647)\n73.7 (85,983)\nMean age (SD)\n27 (6.1)\n29.2 (6.2)\n27.8 (6.1)\n30.2 (6.2)\nPrimiparous\n25.6 (13,509)\n21.9 (31,471)\n26.3 (14,584)\n21.4 (24,968)\nMarriedc\n50.6 (26,715)\n52.2 (74,767)\n49.6 (27,506)\n46.1 (23,207)\nBirthing person educationd\n\u2003 Eighth grade or less\n31.2 (16,294)\n22.3 (31,643)\n29.9 (16,145)\n21.3 (23,707)\n\u2003 Some high school\n30.2 (15,802)\n22.4 (31,791)\n30.2 (16,304)\n21.5 (23,938)\n\u2003 High school graduate\n26.3 (13,757)\n32.7 (46,427)\n26.2 (14,173)\n32.3 (35,892)\n\u2003 Some college/associates\n8.1 (4,251)\n13.6 (19,269)\n9.6 (5,177)\n16.2 (17,981)\n\u2003 College or more\n4.2 (2,177)\n9.2 (13,039)\n4.1 (2,221)\n8.7 (9,699)\nCountry of origin\n\u2003 Belize\n.1 (56)\n.1 (165)\n.2 (122)\n.3 (306)\n\u2003 Costa Rica\n.3 (137)\n.4 (555)\n.1 (65)\n.4 (449)\n\u2003 El Salvador\n4 (2,114)\n7.2 (10,330)\n6 (3,343)\n12.9 (15,010)\n\u2003 Guatemala\n4.2 (2,237)\n13.6 (19,430)\n4.2 (2,311)\n11.6 (13,565)\n\u2003 Honduras\n3.5 (18,69)\n13 (18,576)\n1.7 (955)\n6.5 (7,626)\n\u2003 Mexico\n86.2 (4,5464)\n64.1 (91,956)\n87 (48,226)\n67.2 (78,423)\n\u2003 Nicaragua\n1.2 (628)\n1.2 (1,779)\n.6 (313)\n.9 (1,017)\n\u2003 Panama\n.5 (247)\n.4 (625)\n.2 (108)\n.3 (298)\nCounty-level characteristics\n13.1 (6,920)\n12.1 (17,302)\n2.7 (1,472)\n2.3 (2,727)\n\u2003 Lived in a rural countye\n24.5 (1,2943)\n19 (27,184)\n6.8 (3,771)\n2.6 (3,053)\n\u2003 287(g) contract in placef\n14.8 (7,779)\n12.5 (17,919)\n13.5 (7,491)\n11.1 (12,946)\naExcluding seven states that implemented license laws outside of our study period.\nbMeasured using data from the National Vital Statistics System (2021).\ncMissing 344,884 observations.\ndMissing 266,570 observations.\neNational Center for Health Statistics (2017).\nfDefined as any police department within the county (including municipal) having an active 287(g) agreement with \nImmigration Customs Enforcement any time during the year. These agreements were obtained using the Internet \nArchive.\n\n332\t\nJournal of Health and Social Behavior 65(3) \nenacted license laws compared to those that did. The \nproportion of birthing people living in counties with \n287(g) agreements declined over time in both \ngroups. Finally, individuals living in states that \nnever enacted a license law were more likely to live \nin a rural county than individuals residing in states \nthat passed a license law.\nState demographic and economic characteristics \nwere similar across groups at each time point (Table \n2). However, indicators of immigration enforcement \nclimate show a more positive climate for immigrants \nin states that enacted license laws with increasing \ninclusivity in 2020 compared to 2008. The median \nICI was \u22127 (interquartile range [IQR] = 11) in states \nthat never enacted a license law in 2008 and \u221215 \n(IQR = 47) in 2020. In states that enacted a license \nlaw, the ICI was 2.0 (IQR = 4.0) in 2008 and 39  \n(IQR = 27) in 2020. Similarly, the median detainer \nrate was lower in both 2008 and 2020 in states that \nenacted license laws than in states that never enacted \na license law. States that enacted license laws were \nalso more likely to adopt the unborn child option \nunder CHIP and expand Medicaid than states that \nnever enacted a license law.\nMultivariable Log Binomial and Linear \nModels\nAfter adjusting for year and month; individual, \ncounty, and state characteristics; and state ICI, \nCentral American- and Mexican-born individuals \nresiding in states at delivery with a license law in \nplace since the estimated date of conception had \nlower risks of preterm birth (fully adjusted relative \nrisk [RR] = .90; 95% CI = [.89, .92]) and low birth-\nweight birth (adjusted RR = .93; 95% CI = [.91, \nTable 2.\u2002 Characteristics of State Demographics, Economy, and Immigration Climate in States That \nHave and Have Not Enacted Laws Providing Licenses to Undocumented Immigrants, 2008 and 2020, 43 \nStates and the District of Columbia.a\nNever Enacted, \n2008\nNever Enacted, \n2020\nEnacted, 2008\nEnacted, 2020\nN (States)\n31 States\n31 States\n12 States + DC\n12 States + DC\nState-Level Characteristics\nMedian (IQR)\nMedian (IQR)\nMedian (IQR)\nMedian (IQR)\nMedian Immigration Climate \nIndexb\n\u20137 (11)\n\u201315 (47)\n2 (4)\n39 (27)\nMedian state detainer rate (of \n100,000 noncitizens)c\n712.9 (1,115.4)\n709.4 (403.6)\n422.5 (715.5)\n263.3 (265.6)\nPercentage with a high school \n(or more) educationd\n85.7 (7.2)\n90.4 (5.1)\n86.9 (3)\n90.2 (1.7)\nPercentage of Latines who are \nnoncitizensd\n27.2 (17.4)\n20 (13.2)\n28.9 (10.2)\n21.9 (7)\nPercentage of foreign-born \nMexican and Central \nAmerican-born who are \nnoncitizensd\n79 (7.4)\n73.6 (10.7)\n79.5 (9.6)\n72 (8.5)\nPercentage of adults \nunemployedd\n8.6 (1.5)\n7.2 (1.3)\n8.3 (1)\n7.3 (1)\nMedicaid expansion under \nAffordable Care Act, % (n)e\n25.8 (8)\n54.8 (17)\n100 (13)\n100 (13)\nState CHIP unborn child \noption in place, % (n)f\n22.6 (7)\n29 (9)\n46.2 (6)\n46.2 (6)\nNote: IQR = interquartile range; CHIP = Children\u2019s Health Insurance Program.\naExcluding seven states that implemented license laws outside of our study period.\nbPham and Van (2014, 2019).\ncTransactional Records Access Clearinghouse ( 2024).\ndFrom American Community Survey, three-year estimates for 2008 to 2009 and five-year estimates for 2010 to 2020.\neKaiser Family Foundation (2023).\nfBrooks et al. (2022).\n\nMoinester and Stanhope\t\n333\n.96]), a slight increase in mean birthweight (adjusted \n\u03b2 = 5.2; 95% CI = [1.6, 8.7]), and lower probability \nof first-trimester prenatal care (adjusted RR = .98; \n95% CI = [.97, .98]) compared to individuals resid-\ning in a state without a license law implemented \nprior to conception (Table 3). Additionally, varying \nthe lag in the exposure (i.e., classifying individuals \nas exposed only if the law was in place 6, 12, 18, or \n24 months prior to conception) did not change the \nresults for first-trimester prenatal care (Table 4). For \npreterm birth, low birthweight, and mean birth-\nweight, the estimated associations were slightly \nstronger the greater the lag time, with the greatest \nchange occurring for birthweight.\nSensitivity Analyses\nWe conducted a series of additional analyses to \nassess assumptions underlying our analysis and \ninterpretation. First, a causal interpretation of our \nresults assumes there were no other events occur-\nring at or around the passing of these laws in each \nstate that affected perinatal outcomes for Mexican/\nCentral American-born birthing people living in the \nstate. To test this possibility, we used a placebo \napproach and examined the effect of passage of a \ndriver\u2019s license law on U.S.-born, non-Hispanic \nWhite birthing people (see Table S2 in the online \nversion of the article), a group for whom no effect \nfrom the law was expected. In the fully adjusted \nmodel, there was no observed association for low \nbirthweight, mean birthweight, or entry into prena-\ntal care and a slight inverse association for preterm \nbirth (RR = .97; 95% CI = [.96, .98]). Second, we fit \nmodels stratifying by birthing person\u2019s educational \nattainment (Table S3 in the online version of the \narticle). In fully adjusted models, we observed a \nslightly stronger protective association among indi-\nviduals who did not have a high school education \nfor preterm birth, low birthweight, and mean birth-\nweight. However, the observed association with \nfirst-trimester entry into prenatal care remained in \nthe opposite direction of our hypothesis (lower \nprobability of entry for individuals exposed to the \nlaw since conception). Lastly, to determine whether \nthe COVID-19 pandemic impacted our results, we \nrepeated our analysis excluding births in 2020 to \n2021 (Table S4 in the online version of the article). \nTable 3.\u2002 Models (Log Binomial and Linear) Estimating Associations between Exposure to a Law \nSince Conception Allowing Driver\u2019s Licenses for Undocumented Immigrants and Perinatal Outcomes, \nMexican- and Central American-Born Birthing People in the 12 States and the District of Columbia That \nImplemented a License Law between 2008 and 2020, N = 1,903,939.a\nPreterm Birth\nLow Birthweight\nBirthweight\nFirst-Trimester \nEntry into Prenatal \nCareb\n\u2002\nRR [95% CI]\nRR [95% CI]\n\u03b2 [95% CI]\nRR [95% CI]\nUnadjusted\n1.02 [.96, 1.07]\n1.09 [1.06, 1.12]\n\u201315.9 [\u201325.1, \u20136.7]\n1.01 [.97, 1.05]\nIndividually adjustedc\n.94 [.92, .96]\n.94 [.92, .97]\n3.7 [.7, 6.7]\n.99 [.98, .99]\nA\u0007djusted for state \neconomic and \ndemographic covariatesd\n.91 [.89, .93]\n.93 [.91, .96]\n5.7 [2.2, 9.2]\n.97 [.97, .98]\nAdditionally adjusted for \nImmigrant Climate Index\n.90 [.89, .92]\n.93 [.91, .96]\n5.2 [1.6, 8.7]\n0.98 [.97, .98]\nNote: RR = relative risk; CI = confidence interval.\naData obtained from the National Vital Statistics System (2021); American Community Survey, three-year estimates \nfor 2008 to 2009 and five-year estimates for 2010 to 2020; National Center for Health Statistics (2017); and Pham \nand Van (2014, 2019).\nbIncluding births in states following the implementation of the revised birth certificate (1,832,026).\ncAdjusted for fixed effects for state, year of conception (continuous and quadratic), parity (binary), birthing person\u2019s \nage (continuous and quadratic), month of conception, and country of origin.\ndAdjusted for fixed effects for state, year of conception (continuous and quadratic), month of conception, parity \n(binary), birthing person\u2019s age (continuous and quadratic), country of origin, rurality (county-level, binary), \nunemployment rate (state), percentage of Latine individuals who are noncitizens (state), percentage of foreign-born \nMexican- and Central American-born individuals who are noncitizens, and percentage of adults without a high school \ndiploma.\n\n334\t\nJournal of Health and Social Behavior 65(3) \nThe direction and magnitude of the associations did \nnot change, but the association for birthweight was \nno longer statistically significant.4\nDiscussion\nOver the past several years, states have increasingly \nenacted policies that extend access to driver\u2019s licenses \nto undocumented immigrants. We exploit state and \ntemporal variation in the implementation of these \npolicies to estimate their association with perinatal \nhealth outcomes for birthing people from Mexico and \nCentral America. We proposed that access to a driv-\ner\u2019s license law may improve perinatal health by \nreducing chronic stress and/or expanding access to \nfinancial and health care resources and that the \nstrength of the association would increase over time. \nOverall, our results provide support for these hypoth-\neses. Adjusted models showed reduced risk of pre-\nterm birth and low birthweight and higher mean \nbirthweight comparing individuals in a state with a \nlicense law in place during gestation to those residing \nin a state who would eventually implement a license \nlaw but had not yet during that pregnancy. For pre-\nterm birth, low birthweight, and birthweight, we fur-\nther observed a lagged association\u2014the estimated \nassociations were slightly stronger the greater the lag \ntime. Our sensitivity analysis with U.S.-born, \nnon-Hispanic White birthing people suggests that the \nlaws do not appear to impact birthweight among indi-\nviduals who were already eligible to receive driver\u2019s \nlicenses prior to the passage of the law and who are \nunlikely to be perceived as undocumented. However, \nwe observed a slight protective association for pre-\nterm birth among U.S.-born, non-Hispanic White \nbirthing people, suggesting that other secular pro-\ncesses may have driven part of the observed associa-\ntion for Mexican- and Central American-born \nbirthing people as well. Our sensitivity analysis strat-\nifying results by education suggests that the imple-\nmentation of a license law is associated with greater \nimprovements in preterm birth, low birthweight, and \nbirthweight among Mexican and Central American \nbirthing people with less than a high school education \ncompared to those with at least a high school degree. \nImmigrants with lower levels of education are more \nlikely to be undocumented and thus directly impacted \nby these laws (Passel and Cohn 2019).\nOverall, our results and sensitivity analyses pro-\nvide the strongest support for a protective effect of \nlicense laws on low birthweight and birthweight. \nAlthough we find an association between license \nlaws and improvements in preterm birth, our sensi-\ntivity analysis with non-Hispanic White birthing \npeople suggests that the protective effect for pre-\nterm birth may have partially resulted from secular \nTable 4.\u2002 Models (Log Binomial and Linear) Estimating Associations between Implementation of a Law \nprior to Conception Allowing Driver\u2019s Licenses for Undocumented Immigrants and Perinatal Outcomes, \nMexican- and Central American-Born Birthing People in the 12 States and the District of Columbia That \nImplemented a License Law between 2008 and 2020, N = 1,903,939.a\nNumber of \nMonths in \nPlace prior to \nConceptionb\nPreterm Birth\nLow Birthweight\nBirthweight\nFirst-Trimester \nEntry into \nPrenatal carec\nN\nRR [95% CI]d\nRR [95% CI]d\nbeta [95% CI]d\nRR [95% CI]d\n1 or more\n1,903,939\n.90 [.89, .92]\n.93 [.91, .96]\n5.2 [1.6, 8.7]\n1.01 [.97, 1.05]\n6 or more\n1,830,680\n.89 [.87, .91]\n.93 [.9, .96]\n7.5 [3.4, 11.5]\n.98 [.97, .98]\n12 or more\n1,760,122\n.88 [.86, .9]\n.92 [.89, .96]\n8.6 [3.8, 13.4]\n.97 [.96, .98]\n18 or more\n1,698,533\n.88 [.86, .91]\n.91 [.87, .95]\n10.8 [5.6, 16.1]\n.97 [.96, .98]\n24 or more\n1,638,869\n.87 [.84, .9]\n.88 [.84, .93]\n16.3 [10.4, 22.2]\n.97 [.97, .98]\nNote: RR = relative risk; CI = confidence interval.\naData obtained from the National Vital Statistics System (2021); American Community Survey, three-year estimates \nfor 2008 to 2009 and five-year estimates for 2010 to 2020; National Center for Health Statistics (2017); and Pham \nand Van (2014, 2019).\nbFor each lag, partially exposed individuals were excluded (e.g., individuals for whom the law was implemented in \nbetween the specific lag cutoff and the birth date).\ncIncluding births in states following the implementation of the revised birth certificate (1,832,026).\ndAdjusted for fixed effects for state, year (continuous and quadratic), parity (binary), birthing person\u2019s age (continuous \nand quadratic), country of origin, rurality (county-level, binary), unemployment rate (state), percentage of Latine \nindividuals who are noncitizens (state), percentage of foreign-born Mexicans and Central Americans who are \nnoncitizens, percentage of adults without a high school diploma, Immigration Climate Index.\n\nMoinester and Stanhope\t\n335\nprocesses that impacted both groups and are not \naccounted for by our set of controls. These results \nare consistent with prior research on the impact of \ninclusive immigration policies at the federal level \non perinatal outcomes, which found that the \nDeferred Action for Childhood Arrivals program \nwas associated with improvements in low birth-\nweight and birthweight but not preterm birth \n(Hamilton, Langer, and Patler 2021). Our findings \nfurther align with work demonstrating associations \nbetween state-level restrictive immigration policies \nand birthweight (Torche and Sirois 2019). The \nobserved associations for low birthweight and \nmean birthweight may have resulted from a reduc-\ntion in stress and/or increased access to economic \nresources following the implementation of the pol-\nicy. These two mechanisms are both plausible \nimpacts of a driver\u2019s license law as shown through \nprior research (Amuedo-Dorantes et al. 2020; \nRhodes et al. 2015).\nIn contrast to our hypothesis, we did not observe \nan increase in first-trimester entry into prenatal care \nfor individuals exposed to the law, and in adjusted \nmodels, we observed a slightly lower probability of \nfirst-trimester entry into care for exposed individu-\nals. This finding suggests that the observed associa-\ntion between license laws and perinatal outcomes is \nnot driven by access to prenatal care. There are \nmultiple potential explanations for this result. First, \nin enacting states, most of the study population \nentered prenatal care in the first trimester prior to \nthe implementation of the law (73.8%; Table 1), \nand there may have been little room for improve-\nment. Second, although lack of access to driver\u2019s \nlicenses has been described as a barrier to entry into \nprenatal care for Latine birthing people in the past \n(Rhodes et al. 2015), many significant barriers \n(e.g., insurance, cost, distance to care, language \nbarriers) remain. Access to a driver\u2019s license would \nnot have addressed these barriers, and other secular \nprocesses may have exacerbated them.\nAltogether, the findings from this study under-\nscore the potential for an individual state policy to \npositively shape the lives of immigrants and their \nfamilies amid a shifting and largely exclusive fed-\neral immigration climate. The laws included in this \nstudy were implemented between 2013 and \n2020\u2014a period characterized by increasingly nega-\ntive discourse around immigration at the national \nlevel, especially since 2015, and an increasingly \nrestrictive federal immigration enforcement climate \nunder President Trump (Callaghan et al. 2019; \nCapps et al. 2018). Within this context, we find that \nthe extension of the legal right to drive to undocu-\nmented immigrants is associated with health \nbenefits for the children of Mexican and Central \nAmerican immigrants. Our findings, in turn, sug-\ngest that state policies that reduce the criminaliza-\ntion of immigrants and facilitate immigrants\u2019 \nintegration into U.S. society may act as a buffer \nagainst a restrictive federal immigration climate.\nThe results of this study should be interpreted in \nlight of several limitations. First, we relied on natal-\nity data to define place of birth, residence, and out-\ncomes, which may be subject to misclassification, \nparticularly for entry into prenatal care. To address \nthis problem, we only used data from the revised \nbirth certificate for entry into prenatal care, which \nhas shown adequate quality (57% to 70% agree-\nment between birth certificates and medical \nrecords; Gregory et al. 2019). Second, we used data \nfrom both the standard and revised versions of the \nbirth certificate, which were implemented in differ-\nent years in different states and may impact data \nquality. Third, we imperfectly identified the popula-\ntion possibly impacted by the law by restricting to \nbirthing people born in Mexico or Central America. \nThis identification strategy may exclude key popu-\nlations (e.g., other immigrant groups less likely to \nbe undocumented) or include nonimpacted popula-\ntions (e.g., lawful permanent residents and natural-\nized citizens without close network ties who are \nundocumented). We speculate that if we were able \nto identify only undocumented individuals, the pol-\nicy effects would be larger. This is supported by the \nfact that estimated effects were stronger among \nindividuals without a high school education, who \nare more likely to be undocumented (see Table S3 \nin the online version of the article). Fourth, we \nassign the exposure based on resident state at deliv-\nery and timing of conception. Some individuals \nmay have moved states during the pregnancy and \nthus have the exposure misclassified. If individuals \nwere more likely to move into states following the \nimplementation of the law, this bias could be \ntoward the null. We speculate that this likely \nimpacts a minor proportion of individuals.\nThe findings from this study suggest several \ndirections for future research. More research is \nneeded on the mechanisms linking driver\u2019s license \nlaws to improved perinatal health. Although our \nfindings suggest that the observed associations are \nnot driven by access to prenatal care, we are unable \nto assess the role of reduced stress and increased \naccess to financial resources. To fully disentangle \nthe mechanisms linking license laws and perinatal \nhealth requires other kinds of data and analyses. \nAdditional research is also needed on how driver\u2019s \nlicense laws intersect with other state-level policies \nto shape perinatal health. In controlling for the ICI, \n\n336\t\nJournal of Health and Social Behavior 65(3) \nthis study shows that driver\u2019s license laws are asso-\nciated with improved birthweight independent of \nstates\u2019 broader policy climate. Yet given that many \nstates possess a combination of integrative and \ncriminalization policies (Young and Wallace 2019), \nit may be that driver\u2019s license laws lead to greater \nimprovements in birthweight in certain policy con-\ntexts. Future research should explore this possibil-\nity and variation in the impact of license laws on \ndifferent racial-ethnic groups. In the context of a \npolarized and oscillating federal immigration cli-\nmate, additional evidence on if, how, and for whom \nstate supportive policies can improve immigrant \nwell-being is urgently needed.\nAcknowledgments\nWe thank Susan B. Long at the Transactional Records \nAccess Clearinghouse at Syracuse University, Pham \nHoang Van, and Huyen Pham for their help accessing the \ndata we use in this study. We thank Michael R. Kramer \nand Michael Esposito for their feedback on the work in \nprogress. Any errors are our own.\nFunding\nThe authors disclosed receipt of the following financial \nsupport for the research, authorship, and/or publication of \nthis article: This work was conducted with support from the \nNational Science Foundation (Award No. SES-1655497); \nthe National Heart, Lung, and Blood Institute (Award No. \nK99HL161355); and the American Bar Foundation.\nORCID iD\nMargot Moinester \n https://orcid.org/0000-0002-2734-1863\nSUPPLEMENTAL MATERIAL\nTables S1 to S6 are available in the online version of the \narticle.\nNotes\n\u2002 1.\t For example, the end of the Bracero Program and \npassage of the Hart-Cellar Act curtailed long- \nestablished legal migration pathways from Mexico \nand Central America while not addressing the eco-\nnomic demand for labor within the United States \nor the root causes of migration across the southern \nborder (Calavita 2010; Massey and Pren 2012). \nFurthermore, policies such as the Immigration \nReform and Control Act substantially increased the \ncost and risk of crossing the United States-Mexico \nborder, leading undocumented migrants to reduce \ntheir return trips to their home country after reach-\ning the United States (Massey and Pren 2012).\n\u2002 2.\t The term Latine is used to capture multiple genders \nand nonbinary individuals. Throughout this article, \nwe use the term Latine rather than Latina to acknowl-\nedge that not all birthing people identify as female.\n\u2002 3.\t We examined Medicaid expansion and CHIP in our \ndescriptive analyses but not in our regression analy-\nses because of the lack of variation in both variables \namong states that implemented a driver\u2019s license \nlaw (see Table 2).\n\u2002 4.\t This change in significance is likely due to the \nsmaller sample size. In addition to excluding two \nyears of data, we excluded New Jersey, New York, \nand Oregon because no births prior to 2020 were \nconceived following the implementation of the \nlicense law in these states.\nReferences\nAmuedo-Dorantes, Catalina, Esther Arenas-Arroyo, and \nAlmudena Sevilla. 2020. \u201cLabor Market Impacts of \nStates Issuing of Driver\u2019s Licenses to Undocumented \nImmigrants.\u201d \nLabour \nEconomics \n63:101805. \ndoi:10.1016/j.labeco.2020.101805.\nArmenta, Amada. 2017. Protect, Serve, and Deport: \nThe Rise of Policing as Immigration Enforcement. \nOakland: University of California Press.\nArmenta, Amada, and Irene Vega. 2017. \u201cLatinos and \nthe Crimmigration System.\u201d Pp. 221\u201336 in Race, \nEthnicity and Law. Vol. 22, edited by M. Deflem. \nBingley, UK: Emerald Publishing Limited.\nAuger, Nathalie, Alison L. Park, Philippe Gamache, \nRobert \nPampalon, \nand \nMark \nDaniel. \n2012. \n\u201cWeighing the Contributions of Material and Social \nArea Deprivation to Preterm Birth.\u201d Social Science \n& Medicine 75(6):1032\u201337.\nBrooks, Tricia, Allexa Gardner, Aubrianna Osorio, \nJennifer \nTolbert, \nBradley \nCorallo, \nMeghana \nAmmula, and Sophia Moreno. 2022. Medicaid and \nCHIP Eligibility and Enrollment Policies as of \nJanuary 2022: Findings from a 50-State Survey. San \nFrancisco, CA: Kaiser Family Foundation.\nCalavita, Kitty. 2010. Inside the State: The Bracero \nProgram, Immigration, and the I.N.S. New Orleans, \nLA: Quid Pro Books.\nCallaghan, Timothy, David J. Washburn, Katharine \nNimmons, Delia Duchicela, Anoop Gurram, and \nJames Burdine. 2019. \u201cImmigrant Health Access \nin Texas: Policy, Rhetoric, and Fear in the Trump \nEra.\u201d BMC Health Services Research 19(1):342. \ndoi:10.1186/s12913-019-4167-1.\nCapps, Randy, Muzaffar Chishti, Julia Gelatt, Jessica \nBolter, and Ariel G. Ruiz Soto. 2018. Revving up the \nDeportation Machinery: Enforcement and Pushback \nunder Trump. Washington, DC: Migration Policy \nInstitute.\nChacon, Jennifer M. 2019. \u201cImmigration Federalism in \nthe Weeds.\u201d UCLA Law Review 66:1330\u201393.\nColeman, Mathew. 2012. \u201cThe \u2018Local\u2019 Migration \nState: The Site-Specific Devolution of Immigration \n\nMoinester and Stanhope\t\n337\nEnforcement in the U.S. South.\u201d Law & Policy \n34(2):159\u201390.\nConway, Karen Smith, and Andrea Kutinova. 2006. \n\u201cMaternal Health: Does Prenatal Care Make a \nDifference?\u201d Health Economics 15(5):461\u201388.\nCrookes, Danielle M., Kaitlyn K. Stanhope, Ye Ji Kim, \nElizabeth Lummus, and Shakira F. Suglia. 2021. \n\u201cFederal, State, and Local Immigrant-Related \nPolicies and Child Health Outcomes: A Systematic \nReview.\u201d Journal of Racial and Ethnic Health \nDisparities 9:478\u201388.\nDe Genova, Nicholas P. 2002. \u201cMigrant \u2018Illegality\u2019 and \nDeportability in Everyday Life.\u201d Annual Review of \nAnthropology 31:419\u201347.\nde Graaf, Johanna P., Eric A. P. Steegers, and Gouke \nJ. Bonsel. 2013. \u201cInequalities in Perinatal and \nMaternal Health.\u201d Current Opinion in Obstetrics & \nGynecology 25(2):98\u2013108.\nde Graauw, Els. 2014. \u201cMunicipal ID Cards for \nUndocumented Immigrants Local Bureaucratic \nMembership in a Federal System.\u201d Politics & Society \n42(3):309\u201330.\nDietz, Patricia M., Jennifer M. Bombard, Yalonda L. \nHutchings, John P. Gauthier, Melissa A. Gambatese, \nJean Y. Ko, Joyce A. Martin, and William M. Callaghan. \n2014. \u201cValidation of Obstetric Estimate of Gestational \nAge on U.S. Birth Certificates.\u201d American Journal of \nObstetrics and Gynecology 210(4):335.e1\u2013e5.\nDondero, Molly, and Claire E. Altman. 2020. \u201cImmigrant \nPolicies as Health Policies: State Immigrant Policy \nClimates and Health Provider Visits among U.S. \nImmigrants.\u201d SSM - Population Health 10:100559. \ndoi:10.1016/j.ssmph.2020.1005.\nDrewry, Jonathan, Bisakha Sen, Martha Wingate, Janet \nBronstein, E. Michael Foster, and Milton Kotelchuck. \n2015. \u201cThe Impact of the State Children\u2019s Health \nInsurance Program\u2019s Unborn Child Ruling Expan\u00ad\nsions on Foreign-Born Latina Prenatal Care and Birth \nOutcomes, 2000\u20132007.\u201d Maternal and Child Health \nJournal 19(7):1464\u201371.\nForde, Allana T., Danielle M. Crookes, Shakira F. Suglia, \nand Ryan T. Demmer. 2019. \u201cThe Weathering \nHypothesis as an Explanation for Racial Disparities \nin Health: A Systematic Review.\u201d Annals of \nEpidemiology 33:1\u201318.e3.\nFuentes-Afflick, Elena, Nancy A. Hessol, Tamar Bauer, \nMary J. O\u2019Sullivan, Veronica Gomez-Lobo, Susan \nHolman, Tracey E. Wilson, and Howard Minkoff. \n2006. \u201cUse of Prenatal Care by Hispanic Women \nafter Welfare Reform.\u201d Obstetrics & Gynecology \n107(1):151\u201360.\nFuentes-Afflick, Elena, Nancy A. Hessol, and Eliseo J. \nP\u00e9rez-Stable. 1998. \u201cMaternal Birthplace, Ethnicity, \nand Low Birth Weight in California.\u201d Archives of \nPediatrics & Adolescent Medicine 152(11):1105\u201312.\nGardner, Trevor George, II, and Aarti Kohli. 2009. \u201cThe \nC.A.P. Effect: Racial Profiling in the ICE Criminal \nAlien Program.\u201d SSRN Scholarly Paper ID 3168885. \nSocial Science Research Network, Rochester, NY.\nGlassman, Amanda, Denizhan Duran, Lisa Fleisher, \nDaniel Singer, Rachel Sturke, Gustavo Angeles, \nJodi Charles, et al. 2013. \u201cImpact of Conditional \nCash Transfers on Maternal and Newborn Health.\u201d \nJournal of Health, Population, and Nutrition 31(4 \nSuppl. 2):S48\u201366.\nGolash-Boza, Tanya. 2015. Deported: Immigrant \nPolicing, Disposable Labor, and Global Capitalism. \nNew York, NY: NYU Press.\nGoldfarb, Samantha S., Whitney Smith, Anne E. \nEpstein, Stevie Burrows, and Martha Wingate. 2017. \n\u201cDisparities in Prenatal Care Utilization among \nU.S. versus Foreign-Born Women with Chronic \nConditions.\u201d Journal of Immigrant and Minority \nHealth 19(6):1263\u201370.\nGregory, Elizabeth C. W., Joyce A. Martin, Erica Lee \nArgov, and M. J. Osterman. 2019. \u201cAssessing the \nQuality of Medical and Health Data from the 2003 \nBirth Certificate Revision: Results from New York \nCity.\u201d National Vital Statistics Reports 68(8):1\u201320.\nGulasekaram, Pratheepan, and S. Karthick Ramakrishnan. \n2015. The New Immigration Federalism. New York, \nNY: Cambridge University Press.\nHamilton, Erin R., Paola D. Langer, and Caitlin Patler. \n2021. \u201cDACA\u2019s Association with Birth Outcomes \namong Mexican-Origin Mothers in the United \nStates.\u201d Demography 58(3):975\u201385.\nHardy, Lisa J., Christina M. Getrich, Julio C. Quezada, \nAmanda Guay, Raymond J. Michalowski, and Eric \nHenley. 2012. \u201cA Call for Further Research on the Impact \nof State-Level Immigration Policies on Public Health.\u201d \nAmerican Journal of Public Health 102(7):1250\u201353.\nHatzenbuehler, Mark L., Seth J. Prins, Morgan Flake, \nMorgan Philbin, M. Somjen Frazer, Daniel Hagen, \nand Jennifer Hirsch. 2017. \u201cImmigration Policies \nand Mental Health Morbidity among Latinos: A \nState-Level Analysis.\u201d Social Science & Medicine \n174:169\u201378.\nHobel, Calvin J., Amy Goldstein, and Emily S. Barrett. \n2008. \u201cPsychosocial Stress and Pregnancy Outcome.\u201d \nClinical Obstetrics and Gynecology 51(2):333\u201348.\nHoynes, Hilary, Doug Miller, and David Simon. 2015. \n\u201cIncome, the Earned Income Tax Credit, and Infant \nHealth.\u201d American Economic Journal: Economic \nPolicy 7(1):172\u2013211.\nJohnston, Gillian, and Ann Morse. 2013. 2012 \nImmigration-Related Laws and Resolutions in the \nStates. Denver, CO: National Conference of State \nLegislatures.\nJones, David W. 2008. Mass Motorization + Mass \nTransit: An American History and Policy Analysis. \nBloomington: Indiana University Press.\nKaiser Family Foundation. 2023. \u201cStatus of State \nMedicaid Expansion Decisions: Interactive Map.\u201d \nhttps://www.kff.org/medicaid/issue-brief/status-of-\nstate-medicaid-expansion-decisions-interactive-map/.\nKline, Nolan. 2017. \u201cPathogenic Policy: Immigrant \nPolicing, Fear, and Parallel Medical Systems in the \nU.S. South.\u201d Medical Anthropology 36(4):396\u2013410.\n\n338\t\nJournal of Health and Social Behavior 65(3) \nKoball, Heather, James Kirby, and Seth Hartig. 2022. \n\u201cThe Relationship between States\u2019 Immigrant-\nRelated Policies and Access to Health Care among \nChildren of Immigrants.\u201d Journal of Immigrant and \nMinority Health 24(4):834\u201341.\nKorinek, Kim, and Ken R. Smith. 2011. \u201cPrenatal Care \namong Immigrant and Racial-Ethnic Minority \nWomen in a New Immigrant Destination: Exploring \nthe Impact of Immigrant Legal Status.\u201d Social \nScience & Medicine 72(10):1695\u2013703.\nKramer, Michael R., and Carol R. Hogue. 2009. \u201cWhat \nCauses Racial Disparities in Very Preterm Birth? \nA Biosocial Perspective.\u201d Epidemiologic Reviews \n31:84\u201398.\nLeBr\u00f3n, Alana M. W., William D. Lopez, Keta Cowan, \nNicole L. Novak, Olivia Temrowski, Maria Ibarra-\nFrayre, and Jorge Delva. 2018. \u201cRestrictive ID \nPolicies: Implications for Health Equity.\u201d Journal of \nImmigrant and Minority Health 20(2):255\u201360.\nMassey, Douglas S., and Karen A. Pren. 2012. \u201cUnintended \nConsequences of U.S. Immigration Policy: Explaining \nthe Post-1965 Surge from Latin America.\u201d Population \nand Development Review 38(1):1\u201329.\nMenj\u00edvar, Cecilia, and Daniel Kanstroom. 2013. \n\u201cIntroduction\u2014Immigrant \u2018Illegality\u2019: Constructions \nand Critiques.\u201d Pp. 1\u201336 in Constructing Immigrant \n\u201cIllegality\u201d: Critiques, Experiences, and Responses, \nedited by C. Menj\u00edvar and D. Kanstroom. New York, \nNY: Cambridge University Press.\nMigration Policy Institute. 2019. \u201cProfile of the \nUnauthorized Population: United States.\u201d https://\nwww.migrationpolicy.org/data/unauthorized-immi \ngrant-population/state/US.\nMoinester, Margot. 2019. \u201cA Look to the Interior: \nTrends in U.S. Immigration Removals by Criminal \nConviction Type, Gender, and Region of Origin, \nFiscal Years 2003\u20132015.\u201d American Behavioral \nScientist 63(9):1276\u201398.\nMorse, Ann. 2019. \u201cReport on State Immigration \nLaws 2018.\u201d Immigrant Policy Project, National \nConference of State Legislatures. http://www.ncsl.\norg/research/immigration/report-on-state-immigra \ntion-laws.aspx.\nNational Center for Health Statistics. 2017. \u201cData \nAccess\u2014Urban Rural Classification Scheme for \nCounties.\u201d \nhttps://www.cdc.gov/nchs/data_access/\nurban_rural.htm.\nNational Conference of State Legislatures. 2023. \u201cStates \nOffering Driver\u2019s Licenses to Immigrants.\u201d https://\nwww.ncsl.org/immigration/states-offering-drivers-\nlicenses-to-immigrants.\nNational Vital Statistics System. 2021. \u201c2008-2021 \nNatality \nData.\u201d \nhttps://www.cdc.gov/nchs/nvss/\nbirths.htm.\nNovak, Nicole L., Arline T. Geronimus, and Aresha M. \nMartinez-Cardoso. 2016. \u201cChange in Birth Outcomes \namong Infants Born to Latina Mothers after a \nMajor Immigration Raid.\u201d International Journal of \nEpidemiology 46(3):839\u201349.\nO\u2019Neill, Barbara, Benoit Sorhaindo, Jing Jian Xiao, and \nE. Thomas Garman. 2005. \u201cFinancially Distressed \nConsumers: Their Financial Practices, Financial \nWell-Being, and Health.\u201d Journal of Financial \nCounseling and Planning 16(1):73\u201387.\nPassel, Jeffrey S., and D\u2019vera Cohn. 2019. \u201cU.S. \nUnauthorized Immigrants Are More Proficient \nin English, More Educated than a Decade Ago.\u201d \nPew Research Center. https://www.pewresearch.\norg/fact-tank/2019/05/23/u-s-undocumented-immi \ngrants-are-more-proficient-in-english-more-edu \ncated-than-a-decade-ago/.\nPerreira, Krista M., and Juan M. Pedroza. 2019. \u201cPolicies \nof Exclusion: Implications for the Health of \nImmigrants and Their Children.\u201d Annual Review of \nPublic Health 40:147\u201366.\nPham, Huyen, and Pham Hoang Van. 2014. \u201cMeasuring \nthe Climate for Immigrants: A State-by-State \nAnalysis.\u201d Pp. 21\u201340 in Strange Neighbors: The \nRole of States in Immigration Policy, edited by  \nG. J. Chin and C. Hessick. New York, NY: New \nYork University Press.\nPham, Huyen, and Pham Hoang Van. 2019. \u201cSubfederal \nImmigration Regulation and the Trump Effect.\u201d New \nYork University Law Review 94(1):125\u201370.\nPhilbin, \nMorgan \nM., \nMorgan \nFlake, \nMark \nL. \nHatzenbuehler, and Jennifer S. Hirsch. 2018. \u201cState-\nLevel Immigration and Immigrant-Focused Policies \nas Drivers of Latino Health Disparities in the United \nStates.\u201d Social Science & Medicine 199:29\u201338.\nPotochnick, Stephanie, Sarah F. May, and Lisa Y. \nFlores. 2019. \u201cIn-State Resident Tuition Policies \nand the Self-Rated Health of High-School-Aged \nand College-Aged Mexican Noncitizen Immigrants, \nTheir Families, and the Latina/o Community.\u201d \nHarvard Educational Review 89(1):1\u201329.\nReed, Mary M., John M. Westfall, Caroline Bublitz, \nCatherine Battaglia, and Alexandra Fickenscher. \n2005. \u201cBirth Outcomes in Colorado\u2019s Undocumented \nImmigrant Population.\u201d BMC Public Health 5:100. \ndoi:10.1186/1471-2458-5-100.\nRhodes, Scott D., Lilli Mann, Florence M. Sim\u00e1n, \nEunyoung Song, Jorge Alonzo, Mario Downs, \nEmma Lawlor, et al. 2015. \u201cThe Impact of Local \nImmigration Enforcement Policies on the Health of \nImmigrant Hispanics/Latinos in the United States.\u201d \nAmerican Journal of Public Health 105(2):329\u201337.\nRo, Annie, Tim A. Bruckner, and Lauren Duquette-\nRury. 2020. \u201cImmigrant Apprehensions and Birth \nOutcomes: Evidence from California Birth Records \n2008\u20132015.\u201d Social Science & Medicine 249:112849. \ndoi:10.1016/j.socscimed.2020.112849.\nRosenberg, Julia, Veronika Shabanova, Sarah McCollum, \nand Mona Sharifi. 2022. \u201cInsurance and Health \nCare Outcomes in Regions Where Undocumented \nChildren Are Medicaid-Eligible.\u201d Pediatrics 150(3): \ne2022057034. doi:10.1542/peds.2022-057034.\nSanchez, Thomas W., Rich Stolz, and Jacinta S. Ma. \n2003. Moving to Equity: Addressing Inequitable \n\nMoinester and Stanhope\t\n339\nEffects of Transportation Policies on Minorities. \nCambridge, MA: The Civil Rights Project at Harvard \nUniversity.\nSchut, Rebecca Anna, and Courtney Boen. 2022. \n\u201cState Immigration Policy Contexts and Racialized \nLegal Status Disparities in Health Care Utilization \namong U.S. Agricultural Workers.\u201d Demography \n59(6):2079\u2013107.\nStanhope, Kaitlyn K., Carol R. Hogue, Shakira F. \nSuglia, Juan S. Leon, and Michael R. Kramer. \n2019. \u201cRestrictive Sub-federal Immigration Policy \nClimates and Very Preterm Birth Risk among \nU.S.-Born and Foreign-Born Hispanic Mothers \nin the United States, 2005\u20132016.\u201d Health & Place \n60:102209. doi:10.1016/j.healthplace.2019.102209.\nStuesse, \nAngela, \nand \nMathew \nColeman. \n2014. \n\u201cAutomobility, Immobility, Altermobility: Surviving \nand Resisting the Intensification of Immigrant \nPolicing.\u201d City & Society 26(1):51\u201372.\nSudhinaraset, May, Rebecca Woofter, Maria-Elena \nDe Trinidad Young, Amanda Landrian, Dovile \nVilda, and Steven P. Wallace. 2021. \u201cAnalysis \nof State-Level Immigrant Policies and Preterm \nBirths by Race-Ethnicity among Women Born in \nthe U.S. and Women Born outside the U.S.\u201d JAMA \nNetwork Open 4(4):e214482. doi:10.1001/jamanet-\nworkopen.2021.4482.\nSwartz, Jonas J., Jens Hainmueller, Duncan Lawrence, \nand Maria I. Rodriguez. 2017. \u201cExpanding Prenatal \nCare to Unauthorized Immigrant Women and the \nEffects on Infant Health.\u201d Obstetrics and Gynecology \n130(5):938\u201345.\nSwartz, Jonas J., Jens Hainmueller, Duncan Lawrence, \nand Maria I. Rodriguez. 2019. \u201cOregon\u2019s Expansion \nof Prenatal Care Improved Utilization among \nImmigrant Women.\u201d Maternal and Child Health \nJournal 23(2):173\u201382.\nTorche, Florencia, and Dalton Conley. 2015. \u201cA Pound \nof Flesh: The Use of Birthweight as a Measure of \nHuman Capital Endowment in Economics Research.\u201d \nPp. 632\u201349 in The Oxford Handbook of Economics \nand Human Biology, edited by J. Komlos and I. R. \nKelly. New York, NY: Oxford University Press.\nTorche, \nFlorencia, \nand \nCatherine \nSirois. \n2019. \n\u201cRestrictive Immigration Law and Birth Outcomes \nof Immigrant Women.\u201d American Journal of \nEpidemiology 188(1):24\u201333.\nTransactional Records Access Clearinghouse. 2024. \n\u201cICE Detainer Records.\u201d https://trac.syr.edu/php \ntools/immigration/detain/about_data.html.\nTraylor, Claire S., Jasmine D. Johnson, Mary C. \nKimmel, and Tracy A. Manuck. 2020. \u201cEffects \nof Psychological Stress on Adverse Pregnancy \nOutcomes and Nonpharmacologic Approaches for \nReduction: An Expert Review.\u201d American Journal \nof Obstetrics & Gynecology MFM 2(4):100229. \ndoi:10.1016/j.ajogmf.2020.100229.\nWang, Julia Shu-Huah, and Neeraj Kaushal. 2019. \n\u201cHealth and Mental Health Effects of Local \nImmigration Enforcement.\u201d International Migration \nReview 53(4):970\u20131001.\nWaslin, Michele L. 2013. \u201cDriving while Immigrant: \nDriver\u2019s \nLicense \nPolicy \nand \nImmigration \nEnforcement.\u201d Pp. 3\u201322 in Outside Justice: \nImmigration and the Criminalizing Impact of \nChanging Policy and Practice, edited by D. C. \nBrotherton, D. L. Stageman, and S. P. Leyro. New \nYork, NY: Springer.\nWherry, Laura R., Rachel Fabi, Adam Schickedanz, and \nBrendan Saloner. 2017. \u201cState and Federal Coverage \nfor Pregnant Immigrants: Prenatal Care Increased, \nNo Change Detected For Infant Health.\u201d Health \nAffairs 36(4):607\u201315.\nYoung, Maria-Elena De Trinidad, Danielle M. Crookes, \nand Jacqueline M. Torres. 2022. \u201cSelf-Rated Health \nof Both U.S. Citizens and Noncitizens Is Associated \nwith State-Level Immigrant Criminalization Policies.\u201d \nSSM - Population Health 19:101199. doi:10.1016/j.\nssmph.2022.101199.\nYoung, Maria-Elena De Trinidad, Gabriela Leon-Perez, \nChristine R. Wells, and Steven P. Wallace. 2019. \n\u201cInclusive State Immigrant Policies and Health \nInsurance among Latino, Asian/Pacific Islander, \nBlack, and White Noncitizens in the United States.\u201d \nEthnicity & Health 24(8):960\u201372.\nYoung, Maria-Elena De Trinidad, and Steven P. Wallace. \n2019. \u201cIncluded, but Deportable: A New Public \nHealth Approach to Policies That Criminalize and \nIntegrate Immigrants.\u201d American Journal of Public \nHealth 109(9):1171\u201376.\nAuthor Biographies\nMargot Moinester is an assistant professor in the \nDepartment of Sociology at Washington University in St. \nLouis. Her research focuses on immigration, policing, \nhealth, and inequality in the United States. Her research \nhas been supported by the National Science Foundation \nand the Russell Sage Foundation.\nKaitlyn Stanhope is an assistant professor in the \nDepartment of Epidemiology in the Emory Rollins School \nof Public Health. Her research focuses on understanding \nhow structural and interpersonal stressors impact health \noutcomes for birthing people and their infants and identi-\nfying policy, community, and clinical interventions that \nmay mitigate those negative impacts. Her research has \nbeen supported by the National Heart, Lung, and Blood \nInstitute and Health Resources and Service Administration.\n\n\n    Scientific Research Findings:", "answer": "\u2022 For Mexican and Central American immigrants, implementation of a license law is associated with fewer low birth weight births and higher mean birth weight\u2014critical measures of early development.\n\u2022 The association between license laws and birth weight was stronger the longer the law was in effect.\n\u2022 License law implementation was not associated with birth weight for U.S.-born non-Hispanic White pregnant people, a group for whom no effect from the law was expected.", "id": 64}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465231205266\nJournal of Health and Social Behavior\n2024, Vol. 65(2) 162\u00ad\u2013181\n\u00a9 American Sociological Association 2023\nDOI: 10.1177/00221465231205266\njournals.sagepub.com/home/hsb\nOriginal Article\nOne of the most consequential changes in the \nAmerican labor market in the past half-century is \nthe demise of the labor movement (Farber et\u00a0 al. \n2018; Rosenfeld 2014). Whereas one third of pri-\nvate sector workers were union members in the \n1950s, today, about 6% of the private sector is \nunionized. Although most research on the conse-\nquences of union decline focuses on its economic \nimpact, a growing literature has addressed the con-\nsequences of union membership for health and well-\nbeing (Leigh and Chakalov 2021; Reynolds and \nBrady 2012). Through the improvement of work-\nplace and occupational conditions (Donado 2015; \nHagedorn et\u00a0al. 2016; Keune 2013), higher earnings \n(VanHeuvelen 2018; Western and Rosenfeld 2011), \nexpanded access to fringe benefits (Rosenfeld \n2014), and broad sociopsychological advantages \n(Blanchflower and Bryson 2020; Flavin and \nShufeldt 2016), union membership has been shown \nto influence multiple dimensions of physical and \nmental health and well-being.\nThe current study uses multilevel growth curve \nmodels applied to 39 waves of data from the Panel \nStudy of Income Dynamics between years 1970 \nand 2019 to assess how the accumulated experience \nof union membership across the entirety of one\u2019s \ncareer, or what we term cumulative unionization, \naffects disparities of general health, functional limi-\ntations, and chronic conditions among adults ages \n60 and older. The measure of cumulative unioniza-\ntion, rather than a static measure of union status \ncollected at a particular point in time, captures \ncareer-long duration of union attachment, a source \n1205266 HSBXXX10.1177/00221465231205266Journal of Health and Social BehaviorHan et al.\nresearch-article2023\n1University of Minnesota, Twin Cities, Minneapolis, MN, \nUSA\n2Bocconi University, Milano, Italy\nCorresponding Author:\nTom VanHeuvelen, Department of Sociology, University \nof Minnesota, Twin Cities, Social Sciences Building, 267 \n19th Ave S, Minneapolis, MN 55455, USA. \nEmail: tvanheuv@umn.edu\nCumulative Unionization and \nPhysical Health Disparities \namong Older Adults\nXiaowen Han1\n, Tom VanHeuvelen1\n,  \nJeylan T. Mortimer1\n, and Zachary Parolin2\nAbstract\nWhereas previous research shows that union membership is associated with improved health, static \nmeasurements have been used to test dynamic theories linking the two. We construct a novel measure of \ncumulative unionization, tracking individuals across their entire careers, to examine health consequences \nin older adulthood. We use data from the Panel Study of Income Dynamics (1970\u20132019) and predict \nself-rated health, functional limitations, and chronic health conditions in ages 60 to 79 using cumulative \nunionization measured during respondents\u2019 careers. Results from growth models show that unionized \ncareers are associated with .25 SD to .30 SD improvements in health among older adults across all \nmeasures. Analyses of life course mechanisms reveal heterogeneous effects across unionization timing, age \nin older adulthood, and birth cohort. Moreover, subgroup analyses reveal unionization to partially, but not \nfully, ameliorate disparities based on privileged social positions. Our findings reveal a substantial and novel \nmechanism driving older adulthood health disparities.\nKeywords\ncumulative unionization, health disparities, heterogeneous effects, life course, older adulthood\n\nHan et al.\t\n163\nof advantage that accumulates unequally across \nsocial groups (Dannefer 2021; O\u2019Rand 1996). The \nunequal accumulation of union experience further \ntranslates into health disparities in later life.\nAfter documenting overall positive associations \nbetween a larger proportion of one\u2019s career spent in \nunions and positive physical health outcomes at \nolder ages, we move on to examine key sources of \nheterogeneity. First, we consider heterogeneity in \nunion\u2013health associations across three temporal \ndimensions of life course processes: variation in \nage timing of career unionization, ages across older \nadulthood, and birth cohorts. We find a significant \nlag, where unionization at younger ages is more \npredictive of older adulthood health, while the asso-\nciations of cumulative unionization unsurprisingly \ndecay as individuals move into older ages. And \nassociations are concentrated among earlier birth \ncohorts or those who spent more years in a labor \nmarket with strong and federally supported labor \nunions (Jacobs and Myers 2014).\nSecond, we examine whether cumulative union \nassociations are heterogeneous across social cate-\ngories based on structural positions of advantage: \neducational attainment, sex, and race. Associations \nare primarily concentrated among individuals with-\nout a college degree and men, although the benefi-\ncial associations are slightly more pronounced \namong non-White individuals. These findings \nreflect the critique of labor unions as narrowly \nfocused on the well-being of the midtwentieth cen-\ntury\u2019s White working class (Cowie 2010) and \nlabor\u2019s tumultuous relationship with civil rights \n(Frymer 2011). But they also reveal cumulative \nunionization to be a distinct input into health out-\ncomes that cannot be fully reduced to other forms \nof advantage based on social position.\nThe current study has significant implications \nfor health scholars and policymakers. It reveals that \nunionization, measured at the career level, has a \nmeaningful input into health disparities that extend \ninto older adulthood. Neither cross-sectional stud-\nies nor longitudinal studies that begin in the late \ncareer can detect these disparities. We argue that the \nbenefits of unionization for older adults is an under-\nappreciated dimension of the positive health out-\ncomes recently examined by health scholars \n(Hagedorn et\u00a0al. 2016; Leigh and Chakalov 2021). \nResults from this research build on the growing lit-\nerature that extends work and occupation mecha-\nnisms to the career level (Donnelly 2022; Parolin \nand VanHeuvelen 2023). Finally, our findings con-\ntribute to literature focusing on the policy and insti-\ntutional sources of population health and well-being \n(Bakhtiari, Olafsdottir, and Beckfield 2018; Montez \net\u00a0al. 2020).\nBackground\nUnions represent a critical labor market institution \npromoting economic attainment, stability, and ben-\nefits for ordinary and less powerful workers \n(Ahlquist 2017; Farber et\u00a0al. 2018; Rosenfeld 2014). \nScholars have identified the decline of private sector \nunions, covering one third of workers in the 1950s \nto 6% today (Eidlin 2018), as a transformational \nprocess for the American labor market (Card, \nLemieux, and Riddell 2004; Rosenfeld 2019). \nFortin, Lemieux, and Lloyd (2021), for example, \nfind that the combination of minimum wage stagna-\ntion and union decline accounted for two thirds of \ninequality growth since the 1980s, whereas Farber \net\u00a0al. (2018) show consistent union earnings premi-\nums over three quarters of a century amid union \nmembership compositional changes.\nYet union benefits extend beyond economic out-\ncomes. Health scholars have begun to focus \nsquarely on how labor union membership contrib-\nutes to health and well-being (Leigh and Chakalov \n2021; Reynolds 2021). Recent studies suggest that \nunion members enjoy better health outcomes than \nnonunion members. In particular, union members \ntend to have lower rates of workplace injuries \n(Economou and Theodossiou 2015), work in firms \nwith higher rates of OSHA inspections (Weil 1991), \nand have lower overall exposure to hazardous envi-\nronments (Ehrenberg, Smith, and Hallock 2021). In \nthe wake of union decline, contingent, precarious, \nand unpredictable work has risen, which is shown \nto have negative effects on worker health \n(Schneider and Harknett 2019).\nWhy would unions produce positive health out-\ncomes for workers? Scholars point to multiple \ndirect and indirect mechanisms. Unions increase \nwages (VanHeuvelen 2018), particularly middle \nand lower incomes (Firpo, Fortin, and Lemieux \n2009), and keep households out of poverty (Brady, \nBaker, and Finnigan 2013; VanHeuvelen and \nBrady 2022). More economic resources should \ntranslate straightforwardly to better opportunities \nto avoid and manage health problems. Unions also \ncompress earnings distributions (Jacobs and \nDirlam 2016; Western and Rosenfeld 2011). \nInsofar as income matters not only absolutely but \nrelatively for individual well-being (Alderson and \nKatz-Gerro 2016; Frank 2016), a compressed dis-\ntribution may have wide rippling effects (Curran \nand Mahutga 2018).\n\n164\t\nJournal of Health and Social Behavior 65(2) \nMoreover, union members have higher rates of \nvarious fringe benefits, such as retirement and pen-\nsion plans (Rosenfeld 2014), and occupational \nsecurity (Brown 2006; Malinowski, Minkler, and \nStock 2015), which increase predictability and \nopportunity to plan for the future. As a result, \nunionized workers can more easily devote contem-\nporaneous spending to health needs and have access \nto health care to address health issues. Union mem-\nbers on average have higher rates of paid sick leave, \nvacation leave, and paid family leave, all of which \nshould increase access to health care and recovery \ntime away from work in the case of a negative \nhealth event (Leigh and Chakalov 2021; Reynolds \n2021; Rosenfeld 2014). Union members are also \nmore likely to have health insurance and higher \nquality health insurance (Hagedorn et\u00a0 al. 2016; \nReynolds and Brady 2012), which may enable more \nimmediate treatment of health problems and enable \naccess to preventive care, which would reduce  \nlonger-term health problems.\nUnion members typically have lower rates of \nexcessive overtime work and less exposure to pre-\ncarious and contingent work, which may reduce \nworkplace \nstress \n(Hirsch, \nMacpherson, \nand \nDuMond 1997; Morantz 2017; Trejo 1993). Unions \nprovide an institutionalized platform for workers to \norganize collectively against unsafe conditions; the \nclassic work by Wallace (1987) shows that higher \nunion membership rate is associated with signifi-\ncant declines in fatal and nonfatal injury rates \namong miners. Unions also enhance social stability \nand support, promoting stable marriage (Schneider \nand Reich 2014), greater sociality and support at \nwork (Hagedorn et\u00a0 al. 2016), and more frequent \ncommunity and charity engagement (Zullo 2013). \nIt is unlikely that any of these mechanisms would \nprovide positive health outcomes that can be caus-\nally determined to originate exclusively from union \nmembership. However, the combination of pre-\nsumed indirect and peripheral benefits that are \nthemselves shown to be associated with improved \nhealth and well-being suggests that unionization \nmakes a nontrivial contribution to better health \namong workers.\nCumulative Unionization as a \nMechanism of Health\nIn the current study, we extend the analytic focus to \nthe influence of cumulative unionization that occurs \nover the course of an individual\u2019s career on health \noutcomes that occur in older adulthood, which we \ndefine as ages 60 and older. The vast majority of \nstudies connecting unionization to health outcomes \nfocus on cross-sectional associations or longitudinal \nassociations bounded in short time periods. \nAlthough valuable, we argue that such a focus likely \nmisses key dynamics by which unionization con-\ntributes to health and well-being. Above and beyond \nthe point-in-time measure of union status, cumula-\ntive unionization contributes holistic life course \ninformation on divergent career experiences, imply-\ning potential mechanisms underlying the accumula-\ntion of union coverage.\nAccording to mechanisms of cumulative advan-\ntage (Dannefer 2003), a person with consistent \nunion attachment over the work career might have \nentered a unionized job in early adulthood, enabling \nthe attainment of future benefits, greater stability, \nand higher overall career earnings. These accumu-\nlated advantages may bring health benefits that are \nalso accumulated and extended until later in life \nthrough the aforementioned mechanisms. In addi-\ntion, our argument draws on recent stratification \nresearch that demonstrates the significance and the \nnonlinearities of outcomes such as earnings mea-\nsured at the career level (Kim, Tamborini, and \nSakamoto 2015; Parolin and VanHeuvelen 2023; \nSakamoto, Tamborini, and Kim 2018). Our focus \nalso builds on similar recent studies on cumulative \ncareer experience and health outcomes.\nFor example, using Health and Retirement \nStudy data, Donnelly (2022) finds that precarious \nwork during ages 50 to 65 predicted health and \nmortality at age 65 onward, suggesting that repeated \nexposure to decrements in job quality produces a \nunique contribution to health outcomes in subse-\nquent life stages. She draws on a stress and life \ncourse perspective (Pearlin et\u00a0al. 2005), which pre-\ndicts that repeated exposure to stressful life events \nwill accumulate across the life course to negatively \nimpact health. Similarly, Wahrendorf et\u00a0al. (2021) \nfind an association between retrospectively col-\nlected information on negative labor market experi-\nences, such as unemployment, involuntary job loss, \nand downward mobility, and health after the age of \n50. Together, these lines of research motivate the \nneed to extend employment mechanisms beyond \nnarrow point-in-time associations.\nWe conceptualize repeated exposure to union \nmembership as a protective buffer against potential \nnegative health events. The benefits of unionization \ninvolve immediate support to address negative \nhealth events, reduction of workplace and job \nsearch stress, and an infrastructure for individuals \nto accumulate resources to successfully transition \nout of the labor market at traditional retirement \n\nHan et al.\t\n165\nages. Insofar as union membership produces higher \nearnings, greater job security, and more predictable \naccess to fringe benefits, it is reasonable to antici-\npate that the accumulation of these benefits will be \nobservable in health outcomes in older age. These \nmechanisms should allow for (1) more contempora-\nneously available resources to address negative \nhealth events and (2) longer-term acquisition of \nresources to better address living conditions and \nhealth concerns at older ages. Moreover, insofar as \nunion membership provides job security and more \npredictable pathways of advancement, union mem-\nbership may act as a buffer against the accumula-\ntion of stress across the life course, which has been \nshown to negatively impact health (Donnelly 2022; \nPearlin et\u00a0al. 2005).\nLife Course Dynamics of  \nCumulative Unionization\nAlthough we first consider cumulative unionization, \nor the total life course duration of union member-\nship, the life course timing of union exposure \nreveals another temporal dimension that could help \nexplain the union\u2013health associations. According to \nthe life course principle of timing, the developmen-\ntal consequences of life experiences depend on \nwhen they occur in one\u2019s life (Elder, Johnson, and \nCrosnoe 2003). The very meaning of the same life \nevent can change significantly at different develop-\nmental stages. Life course scholars have given con-\nsiderable attention to the formative effects of life \nexperiences and environments at early stages on a \nnumber of life outcomes, including psychological \norientations, socioeconomic attainment, and health. \nTherefore, they often refer to adolescence and  \nearly adulthood as \u201cimpressionable years\u201d in human \ndevelopment (Alwin and McCammon 2003). \nSimilarly, across occupational career stages, the \ntiming of a career experience like unionization \ncould have strong implications for health outcomes \nin later life.\nWe assess unionization at different career stages \nto determine whether older adult health is differen-\ntially affected by the timing of union membership. \nTwo hypotheses are plausible. On the one hand, \nhealth outcomes in older adulthood might be more \naffected by contemporaneous or recent experiences \nof unionization. Union membership may thus have \nmore immediate influence on health outcomes later \nin the career, insofar as union membership provides \naccess to health insurance and higher pay and sig-\nnals greater probability of employment in a higher \nquality job. Alternatively, union membership may \nbe most consequential in the early years. Budd \n(2010) shows that the large majority of workers \nwho have ever been unionized are union members \nprior to the age of 30. According to cumulative \nadvantage mechanisms, this initial advantage of \nearly unionization might either be translated into \ncontinued union membership or provide the oppor-\ntunity to develop relevant work experience and pur-\nsue more advantageous job opportunities, unionized \nor not, that facilitate accumulation of health bene-\nfits over the following career stages. Moreover, the \nbenefits of unionization, including higher pay, a \nmore stable career, and access to fringe benefits, \nmay be most acutely felt at early ages, when indi-\nviduals are still working to establish their careers. \nAt later ages, as individuals enter more stable career \nstages, they may be able to collect more diverse \nhealth benefits from sources other than unions or \nwork in general.\nNext, we consider the dynamic nature of physi-\ncal health conditions in later life. The principle of \nlife-span development suggests that aging is a life-\nlong process, dependent on circumstances and \nevents throughout the life course (Elder et\u00a0al. 2003). \nWe therefore focus on growth trajectories of health \noutcomes and follow workers beyond the end of \ntheir careers. To the extent that career mechanisms \ninfluence older adulthood health, we anticipate that \nthese will decay the further a respondent moves \nfrom their working ages because any effects of \ncareer mechanisms will mix with more immediate \nlife course and health issues that arise in older \nadulthood. Moreover, the life course principle of \ntime and place argues that the life course outcomes \nof individuals are embedded in and shaped by the \nhistorical times and places they experience over \ntheir lifetime (Elder et\u00a0al. 2003). The diverse his-\ntorical circumstances of individuals from different \nbirth cohorts may matter significantly for their \ncareer decisions, opportunities, and experiences. \nSpecifically, we are able to capture respondents \nwho entered the labor market during the golden age \nof unionization (Cowie 2010) and the era of rapid \nunion decline, led by the Carter and Reagan admin-\nistrations, which diminished the support of federal \ninstitutions for labor (Jacobs and Myers 2014; \nRosenfeld 2014). It is reasonable to suppose that \nthe latter group would not be as protected by a \nweakened labor movement. Guided by the two \nadditional life course principles of life-span devel-\nopment and of time and space, we hypothesize that \nthe effects of cumulative unionization on health \nvary across age in later life and across birth cohorts \nliving through shifting labor market contexts.\n\n166\t\nJournal of Health and Social Behavior 65(2) \nSubgroup Advantage\nWe anticipate that union effects will operate differ-\nently across demographic subgroups. We consider \nthree key sources of subgroup heterogeneity: educa-\ntional attainment, gender, and race. Unions have \nhistorically been most consequential for raising \nwages in the middle and lower portions of the earn-\nings distribution (Firpo et\u00a0al. 2009, 2018). Union \nprotection has thus been a significant source of eco-\nnomic protection and stability for a working class \nthat often lacks alternative mechanisms of social \nclosure to produce economic protection, most nota-\nbly a college degree (Parolin and VanHeuvelen \n2023). Particularly for the cohorts we study, which \nhave union membership more firmly rooted in blue \ncollar work (Cowie 2010), we anticipate larger \neffects among those without a college degree.\nWe develop contradictory expectations for gen-\nder and race. On the one hand, union protection \nmight be weaker and less consistent among margin-\nalized groups. Unions have a fraught history regard-\ning discrimination and segregation along race and \ngender lines (Frymer 2011; Lichtenstein 2012). \nRacial integration often occurred in spite of union \nwishes and under threat of litigation. Although \nsome unions moved to organize racial minorities, \nthis was not a uniform outcome (Frymer 2011). \nSimilarly, women often faced exclusion, discrimi-\nnation, and harassment from unions in early wom-\nen\u2019s organizing efforts (Cowie 2010). Although \nmany dimensions of union benefits, such as the \nwage premium, have similar magnitudes across \ngroups (Rosenfeld and Kleykamp 2012; Western \nand Rosenfeld 2011), mistreatment or opportunity \nhoarding among more privileged groups might \nhave consequences that counteract or undermine \nany economic benefit.\nAlthough we consider this argument more \nlikely, one could reasonably draw alternative expec-\ntations. Labor unions were not the sole source of \ndiscrimination and unfair discretion faced by \nwomen and racial minorities among the cohorts we \nstudy. Indeed, compared to other systems of dis-\ncrimination and mistreatment, unions may have \nprovided a more robust system of equalization \ncompared to other options faced by women and \npeople of color (Lichtenstein 2012). For example, \nWilson, Roscigno, and Huffman (2015) found that \nstates that enabled the public sector to operate with \nmore market-driven discretion resulted in greater \nracial inequalities. Similarly, Biasi and Sarsons \n(2022) found that changes to public sector bargain-\ning in Wisconsin, which weakened formal union \nstructures and increased opportunity for discretion, \nresulted in substantially higher gender inequalities \namong public sector teachers. Therefore, despite \nthe real fraught histories of discrimination and \nexclusion, the equalizing effects of unions may \nhave overall positive effects for women and racial \nminorities.\nThe Present Study\nWe anticipate that cumulative unionization will \naffect physical health outcomes for older adults, the \ngroup that would reflect the consequences of career-\nlong processes, with low rates of unionization pro-\nducing health disadvantages among older adults that \nhave hitherto gone unobserved. We assess the life \ncourse dynamics of unionization in terms of (1) the \nrespondent\u2019s union membership at different career \nstages, (2) the extent to which cumulative unioniza-\ntion effects extend into older ages, and (3) cohort \nvariation. We also consider demographic heteroge-\nneity in terms of more and less advantaged sub-\ngroups based on educational attainment, race, and \nsex. Our argument leads to the following hypothe-\nses, which we examine next.\nHypothesis 1: Respondents who spent more of \ntheir careers as union members have better phys-\nical health in older adulthood.\nHypothesis 2: Cumulative unionization is more \nconsequential for physical health in older adult-\nhood than contemporaneous union membership.\nHypothesis 3a: Hypothesis 1 associations are \nprimarily the result of early career unionization.\nHypothesis 3b: Hypothesis 1 associations are \nprimarily the result of late career unionization.\nHypothesis 4: Hypothesis 1 associations decline \nat older ages.\nHypothesis 5: Hypothesis 1 associations are \nmore concentrated among earlier birth cohorts.\nHypothesis 6a: Hypothesis 1 associations are \nconcentrated among noncollege workers.\nHypothesis 6b: Hypothesis 1 associations are \nconcentrated among men.1\nHypothesis 6c: Hypothesis 1 associations are \nconcentrated among White individuals.\nData And Methods\nData\nWe used the Panel Study of Income Dynamics \n(PSID), the longest running U.S. longitudinal sur-\nvey, between 1970 and 2019. The PSID began with \n\nHan et al.\t\n167\na nationally representative sample of about 5,000 \nfamilies and 18,000 individuals. Information from \nthese individuals and their descendants was col-\nlected annually from 1968 to 1995 and biennially \nthereafter. The PSID was ideal for our purposes. The \nlong-running administration of the PSID allows for \nthe prospective tracking of individual employment \ncharacteristics, avoiding inaccuracies produced by \nretrospective accounting of one\u2019s career. Moreover, \nthe PSID began collecting extensive health informa-\ntion of individuals in the 1990s. Main data came \nfrom the WZB-PSID file, which incorporates infor-\nmation from the Cross-National Equivalent File \n(Brady and Kohler 2022).\nWe constructed a sample based on the following \nconditions: respondents have been observed (1) a \nminimum of 15\u2009times between ages 18 and 60,2 (2) \na minimum of once prior to age 30, (3) a minimum \nof once at or beyond age 60, and (4) employed at \nleast once between ages 18 and 60. The combina-\ntion of these characteristics ensured that we cov-\nered respondents over a wide swath of their careers \nto collect a sufficient representation of their level of \nunionization.\nOur sample of older adults was made up by \n2,735 respondents in the PSID . We constructed a \nsample of all available person-years of these 2,735 \nrespondents when they were ages 60 or older, \nresulting in a final sample of 11,354 person-years. \nThe PSID\u2019s low attrition, between 92% and 98%, \nsuggested that unequal attrition plays a minimal \nrole in results (Schoeni et\u00a0al. 2013).\nDependent Variables\nWe focused on three key dependent variables mea-\nsured when respondents were ages 60 and older. \nFirst, self-rated health was a five-category indicator \nranging from poor (1) to excellent (5). Second was \nthe number of reported current or former chronic \nconditions. Following the logic used by Brady et\u00a0al. \n(2022), we focused on life-threatening conditions: \nhigh blood pressure, cancer, diabetes, heart attack/\ndisease, lung disease, and stroke.3 Third, we con-\nstructed a scale of seven activities of daily living \n(ADLs) that represent functional limitations: \nwhether the respondent has difficulty bathing, \ndressing, eating, getting in and out of a bed/chair, \ngoing outdoors, using the toilet, and walking.\nIndependent Variables\nOur key variable was cumulative unionization, fol-\nlowing Parolin and VanHeuvelen (2023). It was \nbased on prospective measurements of a respon-\ndent\u2019s employment and union membership statuses \nbetween ages 18 and 60. We used the ratio of waves \nthat a respondent is a union member over waves \nemployed. For our respondents, cumulative union-\nization was measured as a time invariant.\nCumulative unionization has two noteworthy val-\nues: 0% and 100%. However, the health benefits of \ncumulative unionization from 0% to 100% might not \nbe linearly additive. Some respondents who were \nnever unionized may have had a stable trajectory in a \nnonunionized and prestigious occupation, such as \nmanagement, or may have unobserved advantages, \nsuch as network support. Thus, they might be more \nadvantaged with respect to socioeconomic and health \nbenefits from work than someone who spent a small \nproportion of their career in a union, a proportion \nthat might reflect employment instability. We there-\nfore included a squared term for cumulative union-\nization and pay special attention to career churn that \nmay explain any health differences for those between \n0% and 100% unionized. Appendix Tables 2a to 2c \nin the online version of the article show a variety of \nalternative specifications of our cumulative union-\nization measure. Although main results were sup-\nported with only a linear term, the quadratic fit \nintroduces some complexity that requires important \ncaveats against a simpler argument. We preferred to \ninclude this complexity because it maps more closely \nto the descriptive nature of our data (supplemental \nanalyses showed that nonlinear functional forms \nimprove model fit) and the complexity introduced \nshould motivate future research on the nature of \ncumulative unionization.\nWe took two additional steps in the computation \nof our cumulative unionization measure. First, fol-\nlowing Parolin and VanHeuvelen (2023), we \nimputed biennially skipped years using the mean of \ntwo adjacent nonmissing waves for employment \nand union membership.\nSecond, Parolin and VanHeuvelen (2023) restrict \ntheir sample to men because of missing union infor-\nmation for PSID spouses in 1970 to 1975 and 1977 to \n1978 (almost entirely female in these years). We \nimputed spouse membership using probabilities com-\nputed from 1976 and 1979 waves that include spouse \nunion membership. We first estimated a logit model \npredicting union membership among employed \nspouses in these two waves using two-digit industry \nand occupations, education, race, and state. We then \ncomputed predicted probabilities of union member-\nship among employed spouses in 1970 to 1975 and \n1977 to 1978. Nonemployed spouses received zero \nvalues for both employment status and union \n\n168\t\nJournal of Health and Social Behavior 65(2) \nmembership. Main results without imputation are \nincluded in the Appendix in the online version of the \narticle.\nIn addition to overall cumulative unionization \nbetween ages 18 and 60, we constructed measures \nfor specific age intervals: 18 to 30, 31 to 40, 41 to \n50, and 51 to 60. Each measure represented the per-\ncentage of employed years in the age interval in \nwhich a respondent was covered by or a member of \na union.\nTo ensure that cumulative measures did not \nproxy for contemporary union membership, we \nincluded a time-varying binary indicator for current \nunion membership for each observation at ages 60 \nand older. In the Appendix in the online version of \nthe article, we show replicated results controlling \nfor year-specific union membership information \nbetween ages 25 and 54.\nControls\nWe adjusted for controls, both time varying and \ninvariant, that may confound the association \nbetween unionization and health.\nDemographic controls included age, year of \nbirth, sex, race (White, Black, other), and four-cate-\ngory census region of residence (Northeast, South, \nMidwest, West). We included two measures of mar-\nital status: total years married and whether the \nrespondent is married in a particular wave.\nWe included an indicator of whether a respon-\ndent had a college degree by the age of 40.4\nWe included three health-related measures: the \nrespondent\u2019s retrospective self-assessment of gen-\neral health at ages 0 to 16,5 smoking status at ages \n60 to 79, and the proportion of observed waves that \na respondent had health insurance from 1999 \nonward, the first year health insurance information \nis consistently available.\nWe included eight measures of contemporary \nand career employment. First, we included mea-\nsures of a respondent\u2019s employment status at ages \n60+: working, retired, unemployed, or not in the \nlabor force. Second, we included the respondent\u2019s \nages 60+ posttax and transfer total household \nincome, divided by the square root of the number of \nhousehold members. Third, we constructed 19 \nindustry \ncategories \nfollowing \nWestern \nand \nRosenfeld (2011) and assigned respondents the \nmodal category we observe at ages 18 to 60. \nAppendix materials in the online version of the arti-\ncle include replication with percentage waves in \nindustry categories. Fourth, we constructed 12 large \noccupation categories and assigned the modal \nobserved category from ages 18 to 60. Fifth, we \nassigned the respondent\u2019s modal region of resi-\ndence from ages 18 to 60. Sixth, we included a \nmeasure of total years employed between ages 18 \nand 60 (percentage employed yielded the same \nresults). Seventh, we included the percentage of \nyears that a respondent was self-employed. Eighth, \nwe combined information from occupation and \nindustry, harmonized to 1990 Census Bureau clas-\nsification schemes. We made all possible combina-\ntions of the 200 industry categories and 275 \noccupation categories observed in the PSID and \ncounted the number of unique combinations each \nrespondent held between ages 18 and 60. We used \nthis measure to proxy labor market position \nchanges. Appendix materials in the online version \nof the article include results with and without these \ncareer controls.\nTable 1 includes descriptive statistics for our \nmain analytic sample.\nWe observed that the average time in a union \nduring one\u2019s career was around 16%. However, this \nmixed those who were ever and never union mem-\nbers. About 43% of the sample was never in a union \nduring the period observed, whereas 6% of respon-\ndents age 60 or older were currently unionized. \nNearly 30% of the observed career was unionized \namong the 57% who were ever union members. \nRegarding the timing of unionization, we observed \nslightly higher rates of unionization at ages 41 to 50 \nand the lowest at ages 18 to 30, but the magnitude \nof differences was not great, with means varying \nfrom 14% to 17%.6 We saw that the restrictions to \ncollect both contemporaneous career and older \nadulthood health resulted in a sample of individuals \nages 60 to 79 born between years 1940 and 1959, \nwith health outcomes collected in the 2001 to 2019 \nsurvey waves.\nMethods\nThe primary goal of the current article is to assess \nthe contribution of time-invariant cumulative union-\nization between ages 18 and 60 to time-varying \nolder adult health outcomes from ages 60 and \nonward. Therefore, we used a series of growth curve \nmodels to predict older adult health using cumula-\ntive career information. We used a general setup as \nfollows:\nhealth\nage\ncunion\ncunion\nit\nage\nit\ncunion\ni\ncunion\ni\n=\n+\n+\n+\n+\n+\n\u03b2\n\u03b2\n\u03b2\n\u03b2\n0\n2\n2\nx\u03b2\u03b2\n\uf0f2it\ni\nage\ni\n\u03b2\n\u03b3\n\u00b5\n\u03b2\n\u03b3\n\u00b5\n0\n00\n0\n10\n1\n=\n+\n=\n+\n.\n\b\n\nHan et al.\t\n169\nTable 1.\u2002 Descriptive Statistics of Sample for Main Analysis.\nVariable\nMean\nStandard Deviation\nMinimum\nMaximum\nDependent variables\n\u2003 Self-rated health\n3.245\n1.054\n1\n5\n\u2003 Functional limitations scale\n.401\n1.081\n0\n7\n\u2003 Total number of life-threatening  \nchronic health conditions\n1.151\n1.077\n0\n6\nUnionization\n\u2003 Cumulative unionization 18\u201360a\n.163\n.260\n0\n1\n\u2003 \u2003 Was never in a union\n.431\n0\n1\n\u2003 \u2003 Cumulative unionization for \nrespondents ever in a unionb\n.286\n.289\n.005\n1\n\u2003 Current union memberc\n.059\n0\n1\n\u2003 Cumulative unionization 18\u201330\n.137\n.243\n0\n1\n\u2003 Cumulative unionization 31\u201340\n.165\n.306\n0\n1\n\u2003 Cumulative unionization 41\u201350\n.174\n.335\n0\n1\n\u2003 Cumulative unionization 51\u201360\n.167\n.337\n0\n1\nSelection of key time-varying  \nindependent variables\n\u2003 Age\n65.124\n4.290\n60\n79\n\u2003 Year of birth\n1948.433\n4.966\n1940\n1959\n\u2003 Survey year\n2013.558\n4.627\n2001\n2019\n\u2003 Employment status\n\u2003 \u2003 Currently working\n.453\n0\n1\n\u2003 \u2003 Unemployed\n.019\n0\n1\n\u2003 \u2003 Retired\n.415\n0\n1\n\u2003 \u2003 Other\n.113\n0\n1\n\u2003 Total posttax and transfer-equivalized \nincome, 2021 dollars\n50,416.64\n55,353.41\n0\n1,010,548\n\u2003 Years employed ages 18\u201360\n30.460\n8.605\n1\n43\n\u2003 Percentage years self-employed ages \n18\u201360\n.090\n.179\n0\n1\nSelection of key time-invariant  \nindependent variables\n\u2003 Female\n.547\n0\n1\n\u2003 Has a college degree by age 40\n.411\n0\n1\n\u2003 Respondent race\n\u2003 \u2003 White\n.669\n0\n1\n\u2003 \u2003 Black\n.286\n0\n1\n\u2003 \u2003 Other\n.045\n0\n1\nData Source: 1970 to 2019 Panel Study of Income Dynamics.\nNote: Sample criteria for the 2,729 individuals with 11,095 person-years includes (1) 15 observations between ages 18 \nand 60, (2) at least one observation ages 60+, (3) minimum one observation age 30 or younger, and (4) minimum one \nyear observed employed.\naCumulative unionization is measured as the percentage of observed employed waves between ages 18 and 60 that a \nrespondent is a union member. Cumulative unionization measures are time-invariant, and so summary statistics are \ncollected from persons, not person-years.\nbVery low values originate from imputation measures of spouses prior to 1979.\ncRefers to whether respondent age 60 and older is a union member. Respondents not currently employed are coded \nas 0.\n\n170\t\nJournal of Health and Social Behavior 65(2) \nWe included a random coefficient for respon-\ndent age. When we estimated models to life course \ntiming, we replaced cumulative unionization with \nfour measurements of cumulative unionization, \nmeasured in specific age intervals:\n\t\n\u03b2\n\u03b2\n\u03b2\n\u03b2\ncu\ni\ncu\ni\ncu\ni\ncu\ncu\ncu\ncu\ncu\n1830\n18302\n2\n3140\n31402\n1830\n1830\n3140\n3\n+\n+\n+\n140\n4150\n4150\n5160\n2\n4150\n41502\n2\n5160\n5\ni\ncu\ni\ncu\ni\ncu\ni\ncu\ncu\ncu\ncu\n+\n+\n+\n+\n\u03b2\n\u03b2\n\u03b2\n\u03b2\n1602\n2\n5160\ncu\ni .\n\b\nWe first estimated models with age-specific \ncumulative unionization measure between ages 50 \nand 60 and then added measures of earlier ages sub-\nsequentially to determine the effect of a particular \nera. To examine age and cohort effects, we inter-\nacted our main and squared cumulative unioniza-\ntion terms first with respondent age, 60 to 79, and \nthen with respondent birth cohort, 1940 to 1959. \nWhen we examined subgroup heterogeneity, we \nestimated growth curve models separately among \ngroups with and without a college degree, men and \nwomen, and White and non-White individuals.\nWe used a growth curve model for three main \nreasons. The first was to align with methodologies \nused in recent studies of cumulative employment \non health trajectories in older adulthood (Donnelly \n2022). The second was to streamline across main \nand age trajectory results. Third, our growth curve \nmodel provided several advantages for our data \nstructure: It handled repeated person observations, \nadjusted for unobserved person-level heterogeneity \nin aging, and provided a general modeling frame-\nwork to include both time-varying and time-invari-\nant measures. An extended discussion of the growth \ncurve models is included in the Appendix in the \nonline version of the article.\nResults\nFigure 1 shows local polynomial smoothed plots of \nthe simple bivariate association between cumulative \nunionization and our three health outcomes\u2014self-\nrated health, functional limitations, and chronic \nconditions\u2014for the total sample and separately by \ncollege degree attainment.\nSeveral descriptive trends are notable. First, for \nthe overall sample, respondents with the highest \nrates of cumulative unionization have the most posi-\ntive health outcomes. Respondents who spent 100% \nof their careers in unions have mean self-rated health \nat around 3.5 compared to around 3.1 for those \nspending 40% of their careers in unions and around \n3.3 for those who were never unionized. Similarly, \nalways unionized respondents have functional \nFigure 1.\u2002 Bivariate Associations between Older Adult Health Outcomes and Cumulative Unionization\nData Source: 1970 to 2019 Panel Study of Income Dynamics.\nNote: Graphs depict local polynomial smoothed fits between health outcomes\u2014self-rated health, functional limitation \nscale, and the number of chronic health conditions\u2014against the percentage of total observed employed waves \nunionized between ages 18 and 60.\n\nHan et al.\t\n171\nlimitation rates of around .25 compared to between \n.40 and .55 for those between 0% and 50% of their \ncareer unionized. Always unionized respondents \nsimilarly have around .1 to .2 fewer life-threatening \nchronic conditions compared to those at 50% or less \nof their career unionized.\nWe see significant variation of trends by educa-\ntional attainment. Cumulative unionization has a \nmuch more modest association for those with a col-\nlege degree. There is little difference in self-rated \nhealth and chronic health conditions for the most \nand least unionized college graduates. Cumulative \nunionization is much more consequential among \nthose without a college degree. Those who spent \ntheir entire career in a union have better outcomes \nacross each health measure, with rates roughly con-\nverging to those among those with a college educa-\ntion. These descriptive findings help motivate the \nexamination of unionization as a career process but \nalso suggest the presence of significant heterogene-\nity across key subgroups in the data.\nGrowth Curve Models Predicting Older \nAge Health Outcomes\nWe next turn to multilevel growth curve regression \nmodels predicting our three health outcomes using \ncumulative unionization measures and controls. We \nfirst estimate simple associations between union \nmembership at ages 60+ and cumulative unioniza-\ntion with our health outcomes. We then add controls \nto our regression models. Results are shown in \nTable 2.\nAcross each of our three outcomes, we observe \na similar pattern of results. Among older adults, \nunion membership measured at the current wave \ndoes not predict health outcomes when controls are \nadded to regression models. However, cumulative \nunionization operates in a similar way across each \noutcome. We observe a curvilinear pattern with \nmore positive associations for more unionized \ncareers. These patterns of association for cumula-\ntive unionization remain with the inclusion of both \ncontrols and contemporaneous union membership. \nFigure 2 visualizes the predicted values by cumula-\ntive unionization from Models 5, 10, and 15.\nFigure 2 shows a similar curvilinear pattern \nacross our three health outcomes. Older adults who \nspent the entirety of their careers in unions have bet-\nter self-rated health, fewer functional limitations, \nand fewer chronic conditions. The worst health out-\ncomes are found among those who spent approxi-\nmately 30% of their observed waves in a union. Yet \nrespondents with fully unionized careers also have \nbetter health outcomes than those who were never in \na union, with .206 higher self-rated health (p\u2009<\u2009.05), \n.17 fewer functional limitations (p\u2009<\u2009.10), and .14 \nfewer life-threatening chronic conditions (p\u2009=\u2009.21). \nNotably, these results adjust for industry, occupa-\ntion, and total number of industry\u2013occupation cate-\ngories held during a respondent\u2019s career (for \ncomparisons excluding career measures, see the \nAppendix in the online version of the article). Effect \nsizes are relatively large. Maximum differences in \npredicted values range from .22 SD to .31 SD of the \nspecific health outcome. The overall finding is one \nin which a fully unionized career corresponds with \nbeneficial older adulthood health.\nLife Course Timing of Unionization\nWe next examine life course timing of cumulative \nunionization. To what extent is overall cumulative \nunionization\u2019s influence dependent on the particular \ncareer timing of union membership? To what extent \nare cumulative unionization effects concentrated \namong cohorts who spent a longer period of their \ncareers in a labor market with a strong labor pres-\nence? And to what extent does the association of \ncumulative unionization and health outcomes decay \nas individuals age into older adulthood?\nTable 3 replicates results incorporating four age-\nspecific cumulative unionization measures: the per-\ncentage of observed waves where the respondent is \nunionized between ages 18 and 30, 31 and 40, 41 \nand 50, and 51 and 60, in place of the career total \nmeasure. We first include the 51 to 60 measure and \nthen sequentially add younger age measures. This \nmodeling setup assesses the extent to which con-\ntemporary older adulthood health reflects recent \nunion membership or whether earlier exposure to \nunion membership contributes to older adulthood \nhealth outcomes. Descriptive information about \ninterage correlations are included in the Appendix \nin the online version of the article.\nAcross each health outcome, we find a similar \ngeneral pattern: Unionization at earlier ages is more \npredictive of older adulthood health than unioniza-\ntion at later ages. Unionization at ages 18 to 30 is \nsignificantly associated with more positive self-\nrated health and fewer chronic conditions, and \nunionization in ages 31 to 40 is significantly associ-\nated with fewer functional limitations. We do find \nsignificant associations for ages 41 to 50 for self-\nrated health, but the most consistent pattern across \nhealth outcomes is found in earlier career stages; \nany associations for older ages (e.g., Models 3, 7) \nare removed when younger aged union membership \n\n172\t\nJournal of Health and Social Behavior 65(2) \nis included. As found by Budd (2010), unionization \nrates by age 30 are highly predictive of unionization \nprospects for one\u2019s entire career. And Parolin and \nVanHeuvelen (2023) show that career unionization \nproduces disproportionately high lifetime earnings \npartially due to growth in union wage premiums. It \nis plausible that developing higher earnings and \nhigher employment stability in early career and \nmidcareer should produce significant downstream \nconsequences for later career earnings and employ-\nment and, subsequently, improved health outcomes. \nNotably, our results would not be found if we used \ndata that measured the effect of only later career \nunionization on older adulthood health outcomes.\nAge Timing of Health Outcomes\nTo what extent do cumulative unionization associa-\ntions hold constant across the age range in our sam-\nple of older adults, ages 60 to 79? We address this \nquestion by interacting both cumulative unioniza-\ntion and its squared term with age. Predicted values \nfrom these models are shown in Figure 3.\nRegarding self-rated health, we see a clear \ndecline of the protective effects of cumulative \nunionization as respondents move from ages 60 to \n79. For example, among respondents who were \nunion members in every observed wave of their \ncareer, self-rated health decreases from about 3.7 at \nage 60 to about 3.25 at age 70 and about 2.8 at age \nTable 2.\u2002 Growth Curve Models Predicting Older Age Health.\nSelf-Rated Health\n\u2002\n(1)\n(2)\n(3)\n(4)\n(5)\nCurrent union member\n.1183**\n(.038)\n\u2013.0237\n(.039)\n\u2013.0179\n(.040)\nCumulative unionization\n\u2013.7896**\n(.244)\n\u2013.6929**\n(.224)\n\u2013.6810**\n(.226)\nCumulative unionization squared\n1.1131***\n(.310)\n.8929**\n(.278)\n.8874**\n(.278)\nN\n11,405\n11,354\n11,010\n11,005\n11,005\n\u2002\nFunctional Limitations\n\u2002\n(6)\n(7)\n(8)\n(9)\n(10)\nCurrent union member\n\u2013.2186***\n(.042)\n\u2013.0230\n(.044)\n\u2013.0331\n(.045)\nCumulative unionization\n.7518**\n(.243)\n.6257**\n(.234)\n.6497**\n(.236)\nCumulative unionization squared\n\u20131.1036***\n(.308)\n\u2013.8052**\n(.290)\n\u2013.8158**\n(.290)\nN\n11,213\n11,162\n10,826\n10,821\n10,821\n\u2002\nLife-Threatening Chronic Conditions\n\u2002\n(11)\n(12)\n(13)\n(14)\n(15)\nCurrent union member\n\u2013.1189***\n(.034)\n.0304\n(.036)\n.0263\n(.037)\nCumulative unionization\n.6355*\n(.251)\n.5320*\n(.252)\n.5137*\n(.254)\nCumulative unionization squared\n\u2013.8591**\n(.318)\n\u2013.6600*\n(.313)\n\u2013.6520*\n(.313)\nN\n11,369\n11,318\n10,971\n10,966\n10,966\nControls?\nX\nX\nX\nData Source: 1970 to 2019 Panel Study of Income Dynamics.\nNote: Standard errors are in parentheses. Controls are listed in \u201cData and Methods\u201d section. Total N\u2009=\u200911,405.\n*p\u2009<\u2009.05. **p\u2009<\u2009.01. ***p\u2009<\u2009.001 (two-tailed tests).\n\nHan et al.\t\n173\n79. By age 79, we do not observe any significant \nassociation.\nRegarding functional limitations and chronic \nconditions, we highlight an overall conclusion and \na caveat. The main effects of cumulative unioniza-\ntion remain significant for both functional limita-\ntions and chronic conditions, whereas interaction \nterms are nonsignificant. These findings suggest no \nchange in cumulative unionization\u2019s effect over \ntime as a respondent ages. At the same time, the lin-\near combination of the main and interaction terms, \nfor cumulative unionization\u2019s main and squared \ncoefficients, show insignificant associations with \nfunctional limitations and chronic conditions by the \ntime respondents reach age 79. Although these do \nnot show up in the significance of interaction coef-\nficients, they suggest that the effect of cumulative \nunionization becomes, at minimum, surrounded by \nconsiderable uncertainty at older ages.\nIn total, we can conclude that although the more \nlasting effects of cumulative unionization on physi-\ncal health are uncertain, there is at minimum a \ndecade of older adulthood in which the protective \neffects of career unionization extend.\nBirth Cohort Timing\nWe next consider whether cumulative unionization \nassociations remain consistent across the 19 birth \ncohorts in our sample. To do so, we interact year of \nbirth with our cumulative unionization measure. We \nfind overall significance of the interactions for self-\nrated health and functional limitations: joint signifi-\ncance Wald test, \u03c72(2) = 5.41 and 18.82, p\u2009<\u2009.10 and \np\u2009<\u2009.001, respectively. Figure 4 plots the predicted \neffects by birth cohort. We observe that the inverted \nU shape from the main results is most pronounced \namong the earliest birth cohorts in our sample, or \nthose born in 1940. The association flattens out \nacross birth cohorts, until negligible associations are \nfound among those born in 1959.\nSubgroup Heterogeneity\nTo what extent are results disproportionately con-\ncentrated among education, sex, and racial groups? \nWe reestimate growth curve regression models sep-\narately by subgroup and present results in Table 4.\nWe observe health effects of cumulative union-\nization to be primarily concentrated among indi-\nviduals without college degrees and among men. \nFor education, the curvilinear fit for cumulative \nunionization is found across all three health out-\ncomes for noncollege workers. Unionization has no \nsignificant association with either physical health \nor functional limitations for college-educated work-\ners. For sex, although the curvilinear fit is jointly \nsignificant for all three outcomes for men and over-\nall insignificant for women, the discrete change \nbetween men and women is insignificant across all \nFigure 2.\u2002 Predicted Health Outcomes by Cumulative Unionization.\nData Source: 1970 to 2019 Panel Study of Income Dynamics.\nNote: Predicted values calculated from Models 5, 10, and 15 of Table 2. Shaded areas represent 95% confidence \nintervals. SRH\u2009=\u2009self-rated health.\n\n174\nTable 3.\u2002 Age-Specific Cumulative Unionization Measures.\nSelf-Rated Health\nFunctional Limitations\nLife-Threatening Chronic Conditions\n\u2002\n(1)\n(2)\n(3)\n(4)\n(5)\n(6)\n(7)\n(8)\n(9)\n(10)\n(11)\n(12)\nCumulative unionization\n\u2003 Age 51\u201360\n\u2013.1909\n(.254)\n\u2013.2091\n(.271)\n\u2013.2394\n(.272)\n\u2013.2635\n(.276)\n\u2013.0918\n(.242)\n\u2013.0033\n(.251)\n.0321\n(.254)\n.1002\n(.258)\n.2847\n(.283)\n.2684\n(.302)\n.3122\n(.305)\n03240\n(.310)\n\u2003 Age 51\u201360 squared\n.2126\n(.267)\n.1985\n(.271)\n.2309\n(.273)\n.2253\n(.276)\n.0635\n(.254)\n.0314\n(.252)\n.0003\n(.255)\n\u2013.0374\n(.258)\n\u2013.3521\n(.297)\n\u2013.3869\n(.302)\n\u2013.4446\n(.306)\n\u2013.4475\n(.311)\n\u2003 Age 41\u201350\n\u2013.6017*\n(.256)\n\u2013.5469*\n(.261)\n\u2013.4712\n(.265)\n.4644\n(.238)\n.3715\n(.243)\n.3001\n(.247)\n.3248\n(.284)\n.3110\n(.291)\n.3209\n(.297)\n\u2003 Age 41\u201350 squared\n.7148**\n(.257)\n.6649*\n(.261)\n.5917*\n(.265)\n\u2013.5844*\n(.238)\n\u2013.5048*\n(.243)\n\u2013.4491\n(.247)\n\u2013.2767\n(.285)\n\u2013.2824\n(.292)\n\u2013.2785\n(.297)\n\u2003 Age 31\u201340\n\u2013.3474\n(.236)\n\u2013.1811\n(.245)\n.5438*\n(.220)\n.4999*\n(.228)\n.1027\n(.264)\n.0148\n(.275)\n\u2003 Age 31\u201340 squared\n.4005\n(.253)\n.2802\n(.257)\n\u2013.6180**\n(.236)\n\u2013.5701*\n(.239)\n\u2013.0889\n(.283)\n\u2013.0896\n(.288)\n\u2003 Age 18\u201330\n\u2013.5492*\n(.233)\n.1782\n(.216)\n.6597*\n(.261)\n\u2003 Age 18\u201330 squared\n.5867*\n(.269)\n\u2013.2811\n(.250)\n\u2013.6594*\n(.302)\nData Source: 1970 to 2019 Panel Study of Income Dynamics.\nNote: Standard errors are in parentheses. Models include all independent variables discussed in the \u201cData and Methods\u201d section. Sample size in full models. Self-rated health n = \n9,562; functional limitations n = 9,391; chronic health conditions n = 9,533.\n+p\u2009<\u2009.10. *p\u2009<\u2009.05. **p\u2009<\u2009.01.\n\nHan et al.\t\n175\nlevels of cumulative unionization, meaning no \nstraightforward conclusion for gender can be \ndrawn. For race, we observe cumulative unioniza-\ntion to operate similarly across White and non-\nWhite groups. Cumulative unionization is jointly \nsignificant in all cases except chronic conditions for \nWhite respondents.\nTo better illustrate the subgroup heterogeneity \nof cumulative unionization, Figure 5 presents pre-\ndicted values of self-rated health from the six \nFigure 3.\u2002 Predicted Health Outcomes, by Cumulative Unionization by Age.\nData Source: 1970 to 2019 Panel Study of Income Dynamics.\nNote: Results computed from full analytic sample using growth curve models including all control variables discussed in \nthe \u201cData and Methods\u201d section. Cumulative unionization and its squared term both interacted with age.\nFigure 4.\u2002 Predicted Health Outcomes by Cumulative Unionization across Birth Cohorts.\nData Source: 1970 to 2019 Panel Study of Income Dynamics.\nNote: Results are computed from the full analytic sample using growth curve models including all control variables \ndiscussed in the \u201cData and Methods\u201d section. Cumulative unionization and its squared term both interacted with \nbirth cohort, which ranges from 1940 to 1959. Joint significance of interactions (main and squared term) are \nsignificant for self-rated health, \u03c72(2)\u2009=\u200911.96, p\u2009<\u2009.01, and functional limitations, \u03c72(2)\u2009=\u200919.73, p\u2009<\u2009.001, but not chronic \nconditions, \u03c72(2)\u2009=\u2009.78 p\u2009=\u2009.68.\n\n176\t\nJournal of Health and Social Behavior 65(2) \nTable 4.\u2002 Subgroup Analysis of Older Adulthood Health by Cumulative Unionization.\nEducation\nGender\nRace\n\u2002\nNoncollege\nCollege\nMen\nWomen\nWhite\nNon-White\nSelf-Rated Health\n(1)\n(2)\n(3)\n(4)\n(5)\n(6)\nCumulative unionization\n\u2013.7871**\n(.286)\n\u2013.4022\n(.366)\n\u2013.6584*\n(.317)\n\u2013.5211\n(.324)\n\u2013.7603**\n(.276)\n\u2013.7644*\n(.377)\nCumulative unionization \nsquared\n1.0653**\n(.347)\n.4702\n(.463)\n.9340*\n(.375)\n.5636\n(.427)\n.9598**\n(.341)\n1.0852*\n(.467)\nCurrent union member\n.0132\n(.054)\n\u2013.0559\n(.058)\n\u2013.0374\n(.057)\n.0026\n(.056)\n.0038\n(.048)\n\u2013.0528\n(.070)\nFunctional Limitations\n(7)\n(8)\n(9)\n(10)\n(11)\n(12)\nCumulative unionization\n.8497**\n(.298)\n.3053\n(.384)\n.5051\n(.330)\n.6626+\n(.338)\n.8097**\n(.290)\n.2352\n(.393)\nCumulative unionization \nsquared\n\u2013.9961**\n(.361)\n\u2013.4913\n(.484)\n\u2013.7570+\n(.389)\n\u2013.7261\n(.444)\n\u2013.9228**\n(.358)\n\u2013.5260\n(.486)\nCurrent union member\n\u2013.0573\n(.062)\n\u2013.0134\n(.067)\n.0403\n(.064)\n\u2013.1083+\n(.063)\n\u2013.0231\n(.055)\n\u2013.0291\n(.079)\nChronic Conditions\n(13)\n(14)\n(15)\n(16)\n(17)\n(18)\nCumulative unionization\n.7854*\n(.320)\n.0260\n(.410)\n.5975+\n(.356)\n.4282\n(.363)\n.3754\n(.311)\n.6955+\n(.418)\nCumulative unionization \nsquared\n\u2013.8906*\n(.389)\n\u2013.1653\n(.519)\n\u2013.6124\n(.420)\n\u2013.7469\n(.480)\n\u2013.4578\n(.385)\n\u2013.9338+\n(.520)\nCurrent union member\n\u2013.0419\n(.050)\n.0955+\n(.053)\n\u2013.0009\n(.052)\n.0644\n(.051)\n.0537\n(.044)\n\u2013.0034\n(.064)\nData Source: 1970 to 2019 Panel Study of Income Dynamics.\nNote: Standard are errors in parentheses. All regression models include controls discussed in the \u201cData and Methods\u201d section. Sample \nsizes: Model 1 n = 6,394; Model 2 n = 4,642; Model 3 n = 5,044; Model 4 n = 5,992; Model 5 n = 8,056; Model 6 n = 2,980; Model 7 n = \n6,319; Model 8 n = 4,530; Model 9 n = 4,949; Model 10 n = 5,900; Model 11 n = 7,890; Model 12 n = 2,959; Model 13 n = 6,366; Model \n14 n = 4,627; Model 15 n = 5,025; Model 16 n = 5,968; Model 17 n = 8,029; Model 18 n = 2,964.\n+p\u2009<\u2009.10. *p\u2009<\u2009.05. **p\u2009<\u2009.01 (two-tailed test).\nsubgroup regression models (top row) and group \ndifferences at cumulative unionization levels (bot-\ntom row). Beyond the conclusions drawn from \nTable 4, we see that health outcomes for more vul-\nnerable social categories, noncollege and non-\nWhite individuals, converge to the more positive \nhealth levels of the more advantaged groups, col-\nlege and White individuals, at the highest levels of \ncumulative unionization. We can also observe that \nalthough the positive effect of cumulative unioniza-\ntion is present for both White and non-White \ngroups, the magnitude of the effect of high cumula-\ntive unionization is much greater for non-White \nindividuals in the sample.\nNo single and simple conclusion can be drawn \nregarding unionization\u2019s operation across privi-\nleged social categories, but a few general observa-\ntions can be made. We see that a unionized career \nallows for certain less privileged groups, such as \nworkers without a college degree and non-White \nworkers, to attain older adulthood health outcomes \nmore typically found in more advantaged popula-\ntions. Little can be confidently concluded from our \nanalyses of men and women, aside from the possi-\nbility that the health consequences may be some-\nwhat more concentrated among men than women. \nConsidered as a whole, cumulative unionization \nrepresents a distinct input into the production of \nolder age health disparities that neither fully reflect \nnor fully ameliorate disparities rooted in heteroge-\nneity of social categories defined by relative advan-\ntage and disadvantage.\nDiscussion\nLabor union decline has been widely consequential \nfor middle and lower paid workers, resulting in \nworsened job security, reduced fringe benefits, \n\nHan et al.\t\n177\nlower earnings, social disintegration, and height-\nened inequality (Cowie 2010; Fortin et\u00a0 al. 2021; \nRosenfeld 2014; Western and Rosenfeld 2011), all \nconsequential for health outcomes (Hagedorn et\u00a0al. \n2016; Leigh and Chakalov 2021; Reynolds and \nBrady 2012). The present study examines how \nunion benefits that accrue over an entire career pre-\ndict physical health among older adults. Using 39 \nwaves of PSID data to measure both cumulative \nunion membership during a respondent\u2019s career \nbetween ages 18 and 60 and physical health at ages \n60 to 79, we draw three main sets of conclusions.\nFirst, we find that the accumulated history of \nunionization contributes to older adulthood health. \nAmong older adults, those who spent the entirety of \ntheir career in unions have better self-rated health, \nfewer functional limitations, and fewer life- \nthreatening chronic conditions than those who were \npartially or never union members.\nWhy would the health consequences of union \nmembership operate at the career level for older \nadults? Insofar as union membership provides mate-\nrial benefits to workers who may otherwise have \nlower economic protection, unionized workers \nshould have improved capacity to pursue better \nhealth outcomes. Moreover, insofar as union mem-\nbership improves local working conditions, job secu-\nrity, and fringe benefits, union membership should \nprovide a buffer to stressful employment events that \nmight accumulate to produce negative chronic health \nconditions. These mechanisms imply unionization \nbenefits that rest on predictability and security over \nlong stretches of time, which contemporaneous mea-\nsurements of health and unionization may not detect. \nThese mechanisms admittedly remain speculative, \nneeding future empirical testing.\nSecond, life course dynamics produce heteroge-\nneous effects. Union membership in the early career \nis more predictive of older age health than member-\nship at the later stage of one\u2019s career, potentially \ndue to the critical role of early career stage in real-\nizing union benefits. As workers enter more stable \ncareer stages, health benefits from sources other \nthan a union might emerge. Moreover, early union \nexperience might facilitate positive selection into \nunionized jobs and related benefits in later career \nstages. Yet our cumulative measures within age \nintervals might not capture variations in point-in-\ntime health benefits of union status at each age; \nsuch variability should be tested in future research. \nFigure 5.\u2002 Predicted Self-Rated Health by Cumulative Unionization, across Education, Gender, and \nRacial Subgroups.\nData Source: 1970 to 2019 Panel Study of Income Dynamics.\nNote: Results are computed from the full analytic sample using growth curve models including all control variables \ndiscussed in the \u201cData and Methods\u201d section. Predicted values are collected from six separate growth curve models \nestimated within each subgroup. Bottom panel shows group marginal effects with 95% confidence intervals.\n\n178\t\nJournal of Health and Social Behavior 65(2) \nAcross age in later life, we find that cumulative \nunionization effects decline in older ages of older \nadulthood. As physical health continues to change \namong adults at the older ages, it is plausible to \nexpect that career mechanisms would decline in \ntheir influence.\nAcross birth cohorts, we find the associations \nmore clearly present in the earliest cohorts of our \nsample, suggesting that cumulative unionization \ninteracts with the social context in which our \nrespondents came of age. It is sensible to observe a \ndecline in the influence of cumulative unionization \namong more recent cohorts. Our earlier cohorts, \nwho entered the labor market during the peak of \nunion power, experienced the strongest effects of \nunionization, whereas those entering the labor mar-\nket during the sharp decline of union membership, \npower, and federal support enjoyed much weaker \nlonger-term benefits (Cowie 2010; Rosenfeld \n2014). Previous studies show a decline in labor\u2019s \nability to control wage setting in the private sector \nin the wake of the Reagan administration\u2019s policy \nchanges (Jacobs and Myers 2014) and a decline of \nstrike efficacy (Rosenfeld 2006). Results here sug-\ngest a weakening of the health consequences as \nwell. This variation suggests, unsurprisingly, that \nunion effects vary depending on labor market insti-\ntutional context.\nThird, we find significant heterogeneity across \nsubgroups. Our findings are most pronounced \namong those without a college degree and generally \nlarger for non-White individuals. Insofar as associ-\nations vary by gender, they are more pronounced \namong men, although the strength of this contrast is \nweak. Union power was most firmly rooted in \nworking-class men among the cohorts examined in \nthis study (Cowie 2010; Lichtenstein 2012), and so \nit is unsurprising to find these groups to be the pri-\nmary beneficiaries. As with wages and employment \nsecurity, unionization provides a source of health \nprotection for those with lower levels of education. \nAlthough unions were resistant to racial integration \n(Frymer 2011), Black workers quickly unionized to \ngain protection against widespread risks of labor \nmarket discrimination (Rosenfeld and Kleykamp \n2012). Cumulative unionization thus provides a dis-\ntinct mechanism for older age health disparities \nrather than simply reinforcing or ameliorating those \nof other group differences rooted in systems of \nadvantage and disadvantage.\nOur study is the first to identify unionization as a \ncareer-level mechanism producing health disparities \nin older adulthood. Until recently, no high-quality \nlongitudinal data existed that allowed for the track-\ning of individuals over the entirety of their careers \nand into older adulthood ages. Our study clearly \nillustrates the longer-term consequences of union \nmembership for many individuals who would other-\nwise face high risk of low pay and job insecurity. The \nlong-term dynamics of career attainment and oppor-\ntunity will be of greater social consequence as a \nlarger share of the American population enters older \nadulthood.\nAlthough the mechanisms that we infer to \nunderstand the results are established in stratifica-\ntion and labor literatures, they remain speculative in \ntheir precise connection to older adulthood health \ndisparities. Future research should establish the \nresponsible mechanisms for this study\u2019s findings. \nWe assumed that union membership itself was the \ndriver of older adulthood health outcomes. Yet \nunions have well-established influence on egalitar-\nian state-level policy outcomes. To what extent \ndoes cumulative unionization reflect longer resi-\ndence in more egalitarian policy environments? To \nwhat extent does unionization prevent early mortal-\nity? If unions help prevent death during one\u2019s career \nand improve health outcomes, our results might be \nconservative when considered alongside uneven \nmortality along unionization. More research on this \ntopic is needed. The group with partially unionized \ncareers consistently and surprisingly had the worst \nhealth outcomes. Although we control for many \ncareer characteristics, this group may experience \nsignificant unmeasured volatility, either life course \nor economic. In Appendix Table 1 in the online ver-\nsion of the article, we include descriptive informa-\ntion for this group compared to never and highly \nunionized groups. Although beyond the scope of \nthe current study, this group deserves greater atten-\ntion. Although the later life health disadvantages of \npartial union exposure imply longer-term conse-\nquences of career instability and changes, our find-\nings do not rule out the possibility that union status \nmight still provide contemporaneous health bene-\nfits for these workers.\nORCID iDs\nXiaowen Han \n https://orcid.org/0000-0002-9029- \n5302\nTom VanHeuvelen \n https://orcid.org/0000-0002- \n7504-8186\nJeylan T. Mortimer \n https://orcid.org/0000-0001- \n9192-1873\nZachary Parolin \n https://orcid.org/0000-0003-4065- \n3711\n\nHan et al.\t\n179\nSupplemental Material\nAppendix Tables 1 through 5, Appendix Figures 1 through \n6, and the Supplementary Material are available in the \nonline version of the article.\nNotes\n1.\t\nWe argued that union effects could potentially move \nin either direction for racial minorities and women. \nWe structure our hypotheses pointing in one direc-\ntion for simplicity and because we more strongly \nsuspect these outcomes.\n2.\t\nNo clear threshold exists for the number of measure-\nments during the occupational career. Many studies \nof permanent income use 20\u2009years of career infor-\nmation (Kim et\u00a0al. 2015; Parolin and VanHeuvelen \n2023). We assessed variation in this decision by \nchoosing thresholds between 10 and 30 observa-\ntions, but results did not substantively change.\n3.\t\nUnlike Brady et\u00a0al. (2022), we kept our result as a \ncount. Sensitivity analyses showed that respondents \nwith multiple chronic conditions have significantly \nworse health.\n4.\t\nThis included an associate\u2019s degree, bachelor\u2019s \ndegree, and advanced degree.\n5.\t\nThis question was asked in multiple waves. We took \nthe average of all available answers provided by \nrespondents for a single time-invariant measure.\n6.\t\nAdjacent age periods tended to have high correlations, \nranging between .63 and .77. The earliest, 18 to 30, and \noldest, 51 to 60, age ranges had a correlation of .36.\nReferences\nAhlquist, John S. 2017. \u201cLabor Unions, Political \nRepresentation, and Economic Inequality.\u201d Annual \nReview of Political Science 20:409\u201332. doi:10.1146/\nannurev-polisci-051215-023225.\nAlderson, Arthur S., and Tally Katz-Gerro. 2016. \n\u201cCompared to Whom? Inequality, Social Compa\u00ad\nrison, and Happiness in the United States.\u201d Social \nForces 95(1):25\u201354.\nAlwin, Duane F., and Ryan J. McCammon. 2003. \n\u201cGenerations, Cohorts, and Social Change.\u201d Pp. \n23\u201349 in Handbook of the Life Course, edited by  \nJ. T. Mortimer and M. J. Shanahan. New York, NY: \nKluwer.\nBakhtiari, Elyas, Sigrun Olafsdottir, and Jason Beckfield. \n2018. \u201cInstitutions, Incorporation, and Inequality: \nThe Case of Minority Health Inequalities in Europe.\u201d \nJournal of Health and Social Behavior 59(2):248\u201367.\nBiasi, Barbara, and Heather Sarsons. 2022. \u201cFlexible \nWages, Bargaining, and the Gender Gap.\u201d The \nQuarterly Journal of Economics 137(1):215\u201366.\nBlanchflower, David G., and Alex Bryson. 2020. \u201cNow \nUnions Increase Job Satisfaction and Well-Being.\u201d \nWorking Paper Series, Working Paper No. 27720. \nNational Bureau of Economic Research, Cambridge, \nMA.\nBrady, David, Regina S. Baker, and Ryan Finnigan. \n2013. \u201cWhen Unionization Disappears: State-Level \nUnionization and Working Poverty in the United \nStates.\u201d American Sociological Review 78(5):872\u2013\n96. doi:10.1177/0003122413501859.\nBrady, David, Christian Guerra, Ulrich Kohler, and \nBruce Link. 2022. \u201cThe Long Arm of Prospective \nChildhood Income for Mature Adult Health in \nthe United States.\u201d Journal of Health and Social \nBehavior 63(4):543\u201359.\nBrady, David, and Ulrich Kohler. 2022. \u201cThe WZB-PSID \nFile\u201d [Data set]. Berlin, Germany: WZB-Berlin \nSocial Science Center.\nBrown, Marianne P. 2006. \u201cLabor\u2019s Critical Role in \nWorkplace Health and Safety in California and \nbeyond\u2014As Labor Shifts Priorities, Where Will \nHealth and Safety Sit?\u201d NEW SOLUTIONS: A \nJournal of Environmental and Occupational Health \nPolicy 16(3):249\u201365.\nBudd, John W. 2010. \u201cWhen Do U.S. Workers First \nExperience Unionization? Implications for Revi\u00ad\ntalizing the Labor Movement.\u201d Industrial Relations: \nA Journal of Economy and Society 49(2):209\u201325.\nCard, David, Thomas Lemieux, and W. Craig Riddell. \n2004. \u201cUnions and Wage Inequality.\u201d Journal of \nLabor Research 25(4):519\u201359. doi:10.1007/s12122-\n004-1011-z.\nCowie, Jefferson. 2010. Stayin\u2019 Alive: The 1970s and the \nLast Days of the Working Class. New York, NY: The \nNew Press.\nCurran, Michaela, and Matthew C. Mahutga. 2018. \n\u201cIncome Inequality and Population Health: A Global \nGradient?\u201d Journal of Health and Social Behavior \n59(4):536\u201353.\nDannefer, Dale. 2003. \u201cCumulative Advantage/Disadv\u00ad\nantage and the Life Course: Cross-Fertilizing \nAge and Social Science Theory.\u201d The Journals of \nGerontology, Series B: Psychological Sciences and \nSocial Sciences 58(6):S327\u201337.\nDannefer, Dale. 2021. Age and the Reach of Sociological \nImagination: Power, Ideology and the Life Course. \nNew York, NY: Routledge.\nDonado, Alejandro. 2015. \u201cWhy Do Unionized Workers \nHave More Nonfatal Occupational Injuries?\u201d ILR \nReview 68(1):153\u201383.\nDonnelly, Rachel. 2022. \u201cPrecarious Work in Midlife: \nLong-Term Implications for the Health and Mortality \nof Women and Men.\u201d Journal of Health and Social \nBehavior 63(1):142\u201358.\nEconomou, Athina, and Ioannis Theodossiou. 2015. \u201cJoin \nthe Union and Be Safe: The Effects of Unionization \non Occupational Safety and Health in the European \nUnion.\u201d Labour 29(2):127\u201340.\nEhrenberg, Ronald G., Robert S. Smith, and Kevin F. \nHallock. 2021. Modern Labor Economics: Theory \nand Public Policy. New York, NY: Routledge.\n\n180\t\nJournal of Health and Social Behavior 65(2) \nEidlin, Barry. 2018. Labor and the Class Idea in \nthe United States and Canada. New York, NY: \nCambridge University Press.\nElder, Glen H., Monica Kirkpatrick Johnson, and Robert \nCrosnoe. 2003. \u201cThe Emergence and Development \nof Life Course Theory.\u201d Pp. 3\u201319 in Handbook of \nthe Life Course, edited by J. T. Mortimer and M. J. \nShanahan. New York, NY: Kluwer.\nFarber, Henry S., Daniel Herbst, Ilyana Kuziemko, and \nSuresh Naidu. 2018. \u201cUnions and Inequality over \nthe Twentieth Century: New Evidence from Survey \nData.\u201d Working Paper Series, Working Paper No. \n24587. National Bureau of Economic Research, \nCambridge, MA.\nFirpo, Sergio, Nicole M. Fortin, and Thomas Lemieux. \n2009. \u201cUnconditional Quantile Regressions.\u201d Econo\u00ad\nmetrica 77(3):953\u201373. doi:10.3982/ECTA6822.\nFirpo, Sergio P., Nicole M. Fortin, and Thomas Lemieux. 2018. \n\u201cDecomposing Wage Distributions Using Recentered \nInfluence Function Regressions.\u201d Econometrics 6(2): \n28. doi:10.3390/econometrics6020028.\nFlavin, Patrick, and Gregory Shufeldt. 2016. \u201cLabor \nUnion Membership and Life Satisfaction in the \nUnited States.\u201d Labor Studies Journal 41(2):171\u201384. \ndoi:10.1177/0160449X16643321.\nFortin, Nicole M., Thomas Lemieux, and Neil Lloyd. \n2021. \u201cLabor Market Institutions and the Distribution \nof Wages: The Role of Spillover Effects.\u201d \nJournal of Labor Economics 39(S2):S369\u2013412. \ndoi:10.1086/712923.\nFrank, Robert H. 2016. Success and Luck. Princeton, NJ: \nPrinceton University Press.\nFrymer, \nPaul. \n2011. \nBlack \nand \nBlue: \nAfrican \nAmericans, the Labor Movement, and the Decline \nof the Democratic Party. Princeton, NJ: Princeton \nUniversity Press.\nHagedorn, Jenn, Claudia Alexandra Paras, Howard \nGreenwich, and Amy Hagopian. 2016. \u201cThe Role \nof Labor Unions in Creating Working Conditions \nThat Promote Public Health.\u201d American Journal \nof Public Health 106(6):989\u201395. doi:10.2105/\nAJPH.2016.303138.\nHirsch, Barry T., David A. Macpherson, and J. Michael \nDuMond. 1997. \u201cWorkers# X0027; Compensation \nRecipiency in Union and Nonunion Workplaces.\u201d \nILR Review 50(2):213\u201336.\nJacobs, David, and Jonathan C. Dirlam. 2016. \u201cPolitics \nand Economic Stratification: Power Resources and \nIncome Inequality in the United States.\u201d American \nJournal of Sociology 122(2):469\u2013500. doi:10.1086/ \n687744.\nJacobs, David, and Lindsey Myers. 2014. \u201cUnion \nStrength, Neoliberalism, and Inequality: Contingent \nPolitical Analyses of U.S. Income Differences since \n1950.\u201d American Sociological Review 79(4):752\u201374. \ndoi:10.1177/0003122414536392.\nKeune, Maarten. 2013. \u201cTrade Union Responses to \nPrecarious Work in Seven European Countries.\u201d \nInternational Journal of Labor Research 5(1):59\u201378.\nKim, ChangHwan, Christopher R. Tamborini, and \nArthur Sakamoto. 2015. \u201cField of Study in \nCollege and Lifetime Earnings in the United \nStates.\u201d Sociology of Education 88(4):320\u201339. \ndoi:10.1177/0038040715602132.\nLeigh, J. Paul, and Bozhidar Chakalov. 2021. \u201cLabor \nUnions and Health: A Literature Review of Pathways \nand Outcomes in the Workplace.\u201d Preventive \nMedicine Reports 24:101502. doi:10.1016/j.pmedr \n.2021.101502.\nLichtenstein, Nelson. 2012. State of the Union: A \nCentury of American Labor. Princeton, NJ: Princeton \nUniversity Press.\nMalinowski, Beth, Meredith Minkler, and Laura Stock. \n2015. \u201cLabor Unions: A Public Health Institution.\u201d \nAmerican Journal of Public Health 105(2):261\u201371.\nMontez, Jennifer Karas, Jason Beckfield, Julene Kemp \nCooney, Jacob M. Grumbach, Mark D. Hayward, \nHuseyin Zeyd Koytak, Steven H. Woolf, and Anna \nZajacova. 2020. \u201cU.S. State Policies, Politics, \nand Life Expectancy.\u201d The Milbank Quarterly \n98(3):668\u201399.\nMorantz, Alison D. 2017. \u201cWhat Unions Do for \nRegulation.\u201d Annual Review of Law and Social \nScience 13:515\u201334.\nO\u2019Rand, Angela M. 1996. \u201cThe Precious and the \nPrecocious: Understanding Cumulative Disadvantage \nand Cumulative Advantage over the Life Course.\u201d \nThe Gerontologist 36(2):230\u201338.\nParolin, Zachary, and Tom VanHeuvelen. 2023. \u201cThe \nCumulative Advantage of a Unionized Career for \nLifetime Earnings.\u201d ILR Review 76(2):434\u201360.\nPearlin, Leonard I., Scott Schieman, Elena M. Fazio, and \nStephen C. Meersman. 2005. \u201cStress, Health, and \nthe Life Course: Some Conceptual Perspectives.\u201d \nJournal of Health and Social Behavior 46(2):205\u201319.\nReynolds, Megan M. 2021. \u201cHealth Power Resources \nTheory: A Relational Approach to the Study of \nHealth Inequalities.\u201d Journal of Health and Social \nBehavior 62(4):493\u2013511.\nReynolds, Megan M., and David Brady. 2012. \u201cBringing \nYou More than the Weekend: Union Membership \nand Self-Rated Health in the United States.\u201d Social \nForces 90(3):1023\u201349. doi:10.1093/sf/sor023.\nRosenfeld, Jake. 2006. \u201cDesperate Measures: Strikes \nand Wages in Post-accord America.\u201d Social Forces \n85(1):235\u201365. doi:10.1353/sof.2006.0140.\nRosenfeld, Jake. 2014. What Unions No Longer Do. \nCambridge, MA: Harvard University Press.\nRosenfeld, Jake. 2019. \u201cU.S. Labor Studies in the \nTwenty-First Century: Understanding Laborism \nwithout Labor.\u201d Annual Review of Sociology 45:449\u2013\n65. doi:10.1146/annurev-soc-073018-022559.\nRosenfeld, Jake, and Meredith Kleykamp. 2012. \n\u201cOrganized Labor and Racial Wage Inequality in \nthe United States.\u201d American Journal of Sociology \n117(5):1460\u2013502. doi:10.1086/663673.\nSakamoto, Arthur, Christopher R. Tamborini, and \nChangHwan Kim. 2018. \u201cLong-Term Earnings \n\nHan et al.\t\n181\nDifferentials between African American and White \nMen by Educational Level.\u201d Population Research \nand Policy Review 37(1):91\u2013116. doi:10.1007/\ns11113-017-9453-1.\nSchneider, Daniel, and Kristen Harknett. 2019. \u201cConse\u00ad\nquences of Routine Work-Schedule Instability \nfor Worker Health and Well-Being.\u201d American \nSociological Review 84(1):82\u2013114. doi:10.1177/ \n0003122418823184.\nSchneider, Daniel, and Adam Reich. 2014. \u201cMarrying \nAin\u2019t Hard When You Got a Union Card? \nLabor Union Membership and First Marriage.\u201d \nSocial Problems 61(4):625\u201343. doi:10.1525/sp \n.2014.12316.\nSchoeni, Robert F., Frank Stafford, Katherine A. \nMcgonagle, and Patricia Andreski. 2013. \u201cResponse \nRates in National Panel Surveys.\u201d The ANNALS of the \nAmerican Academy of Political and Social Science \n645(1):60\u201387. doi:10.1177/0002716212456363.\nTrejo, Stephen J. 1993. \u201cOvertime Pay, Overtime Hours, \nand Labor Unions.\u201d Journal of Labor Economics \n11(2):253\u201378.\nVanHeuvelen, Tom. 2018. \u201cMoral Economies or Hidden \nTalents? A Longitudinal Analysis of Union Decline \nand Wage Inequality, 1973\u20132015.\u201d Social Forces \n97(2):495\u2013530. doi:10.1093/sf/soy045.\nVanHeuvelen, Tom, and David Brady. 2022. \u201cLabor \nUnions and American Poverty.\u201d ILR Review \n75(4):891\u2013917. doi:10.1177/00197939211014855.\nWahrendorf, Morten, Hanno Hoven, Christian Deindl, \nThorsten Lunau, and Paola Zaninotto. 2021. \n\u201cAdverse Employment Histories, Later Health \nFunctioning and National Labor Market Policies: \nEuropean Findings Based on Life-History Data from \nShare and ELSA.\u201d The Journals of Gerontology, \nSeries B: Psychological Sciences and Social Sciences \n76(Suppl. 1):S27\u201340.\nWallace, Michael. 1987. \u201cDying for Coal: The Struggle \nfor Health and Safety Conditions in American Coal \nMining, 1930\u201382.\u201d Social Forces 66(2):336\u201364. \ndoi:10.2307/2578744.\nWeil, David. 1991. \u201cEnforcing OSHA: The Role of Labor \nUnions.\u201d Industrial Relations: A Journal of Economy \nand Society 30(1):20\u201336.\nWestern, Bruce, and Jake Rosenfeld. 2011. \u201cUnions, \nNorms, and the Rise in U.S. Wage Inequality.\u201d \nAmerican \nSociological \nReview \n76(4):513\u201337. \ndoi:10.1177/0003122411414817.\nWilson, George, Vincent J. Roscigno, and Matt Huffman. \n2015. \u201cRacial Income Inequality and Public Sector \nPrivatization.\u201d \nSocial \nProblems \n62(2):163\u201385. \ndoi:10.1093/socpro/spv001.\nZullo, Roland. 2013. \u201cOrganized Labor\u2019s Civic \nNiche.\u201d Nonprofit and Voluntary Sector Quarterly \n42(4):781\u2013802. doi:10.1177/0899764012450534.\nAuthor Biographies\nXiaowen Han is a PhD candidate in sociology at the \nUniversity of Minnesota and affiliated with the Minnesota \nPopulation Center and Life Course Center. Her research \ninterests focus on the social inequalities in multidimen-\nsional job quality and its consequences on health dispari-\nties over the life course.\nTom VanHeuvelen is an associate professor in the \nDepartment of Sociology at the University of Minnesota. \nHis research broadly centers around social inequality, \nstratification, quantitative methodology, and historical/\ncomparative sociology. His work has appeared in the \nAmerican Sociological Review, American Journal of \nSociology, Social Forces, and other outlets.\nJeylan T. Mortimer is a professor emeritus in the \nDepartment of Sociology at the University of Minnesota. \nHer research focuses on the life course social psychology, \ninequality, and health, particularly at the transition to \nadulthood.\nZachary Parolin is an assistant professor of social policy \nat Bocconi University. His research focuses on the mea-\nsurement, causes, and consequences of poverty and social \ninequality across the United States and European Union.\n\n\n    Scientific Research Findings:", "answer": " \u2022 Individuals who spent a longer portion of their career as union members reported substantially better physical health across all indicators in older adulthood compared to similar individuals who spent less time in unions.\n\u2022 Associations between union membership and health are strongest among men, those with less than a college degree, and Black individuals.\n\u2022 The greatest health benefits were found among those who joined a union early in their careers, before the age of 30.", "id": 65}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465231194043\nJournal of Health and Social Behavior\n2024, Vol. 65(1) 2\u00ad\u201319\n\u00a9 American Sociological Association 2023\nDOI: 10.1177/00221465231194043\njournals.sagepub.com/home/hsb\nOriginal Article\nPreventive health care use can reduce the risk of dis-\nease, disability, and death (U.S. Department of \nHealth and Human Services [USDHHS] 2022). For \nexample, sigmoidoscopy and colonoscopy screen-\nings are important tools in reducing deaths from \ncolorectal cancer (Brenner, Stock, and Hoffmeister \n2014), flu shots decrease the risk of serious illness if \ninfected (Ferdinands et al. 2021), and COVID-19 \nvaccinations prevented an estimated 14.4 to 19.8 \nmillion deaths worldwide during the first year of use \n(Watson et al. 2022). \u201cPreventive care\u201d generally \nincludes spending on programs for public health \ninformation, education, and counseling; immuniza-\ntion, early disease detection, and healthy condition \nmonitoring programs (e.g., for monitoring preg-\nnancy, child growth and development, general \nhealth checkups); epidemiological surveillance; and \nemergency preparedness and disaster response \n(Kamal and Hudman 2020). Despite the clear health \nbenefits preventive care offers, the majority of U.S. \nadults do not receive all recommended clinical \npreventive services (Borsky et al. 2018). Thus, \nunderstanding and addressing the determinants of \npreventive care use is critical to reducing prevent-\nable deaths, which are higher in the United States \nthan all other OECD countries, and vital for improv-\ning population health (Kamal and Hudman 2020).\nA growing literature points to structural sexism as \nan important upstream driver of population health in \nthe United States (Homan 2019, 2021a, 2021b; \nKavanagh and Graham 2019; Krieger 2020). \nStructural sexism refers to systematic gender \ninequality in power and resources in a given social \ncontext (Homan 2019). Structural sexism directs \nattention beyond individual sexist beliefs and \n1194043 HSBXXX10.1177/00221465231194043Journal of Health and Social BehaviorDore et al.\nresearch-article2023\n1Emory University, Atlanta, GA, USA\n2Florida State University, Tallahassee, FL, USA\nCorresponding Author:\nEmily C. Dore, Department of Sociology, Emory \nUniversity, 1555 Dickey Drive, Atlanta, GA 30322, USA. \nEmail: emily.catherine.dore@emory.edu\nStructural Sexism and \nPreventive Health Care Use  \nin the United States\nEmily C. Dore1\n, Surbhi Shrivastava1,  \nand Patricia Homan2\nAbstract\nPreventive health care use can reduce the risk of disease, disability, and death. Thus, it is critical to \nunderstand factors that shape preventive care use. A growing body of research identifies structural sexism \nas a driver of population health, but it remains unknown if structural sexism is linked to preventive care \nuse and, if so, whether the relationship differs for women and men. Gender performance and gendered \npower and resource allocation perspectives lead to competing hypotheses regarding these questions. \nThis study explores the relationship between structural sexism and preventive care in gender-stratified, \nmultilevel models that combine data from the Behavioral Risk Factor Surveillance System with state-level \ndata (N = 425,454). We find that in states with more structural sexism, both men and women were less \nlikely to seek preventive care. These findings support the gender performance hypothesis for men and the \ngendered power and resource allocation hypothesis for men and women.\nKeywords\ngender inequality, health disparities, preventive health care, state context, structural sexism\n\nDore et al.\t\n3\nbehaviors to highlight the institutional arrangements \nin a society that privilege men and subordinate \nwomen (Homan 2019; Ridgeway and Correll 2004). \nStudies have shown that structural sexism in U.S. \nstate-level institutions is linked to a variety of nega-\ntive health outcomes among women, men, and chil-\ndren, including worse self-rated health, more chronic \nconditions, worse physical functioning, and higher \ninfant and adult mortality rates (Homan 2019; \nKavanagh, Shelley, and Stevenson 2017; Kawachi  \net al. 1999; Koenen, Lincoln, and Appleton 2006).\nOne pathway through which structural sexism \nmay harm health is via its impact on health care. \nThere is very little research on the topic, but one \nstudy by Rapp et al. (2022) found that higher state-\nlevel structural sexism is associated with decreased \nhealth care access and more barriers to affordability \namong women. No studies to date have examined \nthe relationship between structural sexism and pre-\nventive care use and how it might differ among \nwomen and men. This represents a major gap in \nknowledge that can hinder efforts to improve health \nand health equity in the United States.\nTo address this gap, the present study uses the \nBehavioral Risk Factor Surveillance System \n(BRFSS) combined with state-level data to explore \nthe relationship between structural sexism and nine \ndifferent types of preventive health care use in  \ngender-stratified, multilevel models. Gender perfor-\nmance theories and gendered power and resource \nallocation perspectives generate the study\u2019s two \nhypotheses about the relationship between structural \nsexism and preventive care. In high-sexism environ-\nments, gender performance theory suggests that \nmen would be less likely and women more likely to \naccess needed preventive care, and gendered power \nand resource allocation perspective posits that both \nmen and women would be less likely to access care. \nFindings suggest that both theoretical perspectives \ncontribute to our understanding of health care use \namong men, whereas only a gendered power and \nresource allocation perspective is relevant for \nwomen. Results demonstrate a pattern of universal \nharm, suggesting that reducing structural sexism is a \npromising approach to increasing preventive care \nuse for both women and men in the United States. \nThis study provides novel empirical evidence and \nimportant theoretical insights with policy implica-\ntions for preventive health care use.\nBackground\nThe overall picture of preventive health care use in \nthe United States is concerning. Preventive care \nspending in the United States has declined as a share \nof total health expenditure from 3.7% in 2000 to \n2.9% in 2018 (Kamal and Hudman 2020). This \ndeclining investment in preventing and controlling \nrisk exposure is problematic because more than a \nquarter of all personal health care spending in the \nUnited States in 2016 was due to modifiable risk \nfactors and preventable illnesses (Bolnick et al. \n2020). Qualitative research finds that while provid-\ners know that preventive services can reduce the \nburden of disease, most providers, including hospi-\ntals and physicians, are paid to treat rather than to \nprevent disease (Levine et al. 2019). Studies find \nthat most Americans do not access the benefits of \npreventive health care. For example, only 8% of \nU.S. adults age 35 or older received all recom-\nmended, high-priority, appropriate clinical preven-\ntive services in 2015, and nearly 5% received none \nof the routine preventive screenings (Borsky et al. \n2018). The present study focuses on preventive \nhealth care use because of concerns around the pat-\nterns of preventable deaths and illnesses in the \nUnited States, low rates of preventive health care \nuse, and the promise of preventive health care for \nimproving population health.\nThere are a number of factors that can affect the \nuse of preventive health care services. Preventive \nhealth care use among U.S. adults is consistently \nand significantly associated with the level of health \ninsurance coverage and having a usual place and \nprovider (Blewett et al. 2008; DeVoe et al. 2003; \nFaulkner and Schauffler 1997). Other studies have \nfound that a higher purpose in life (Kim, Strecher, \nand Ryff 2014) and higher life satisfaction (Kim, \nKubzansky, and Smith 2015) are also associated \nwith a greater likelihood of preventive health care \nuse. Among adolescents, a greater proportion of \nthose aged 16 to 17 years compared to younger age \ngroups did not have a usual place for preventive \ncare and did not receive a well-child checkup or \nvisit a dentist in the past 12 months (Black, Nugent, \nand Vahratian 2016). There are racial and ethnic \ndifferences in preventive health care use as well, \nsuch as low rates of breast cancer screening among \nAfrican American women and cancer screening \namong Asian Americans (Hou, Sealy, and Kabiru \n2011; Jones, Katapodi, and Lockhart 2015).\nFurthermore, preventive care use is also highly \ngendered. Research on U.S. adults shows that \nwomen score higher than men on a composite mea-\nsure of preventive care use (Asch et al. 2006), and \nmen are less likely to have their blood pressure and \ncholesterol checked, less likely to visit the dentist, \nand less likely to get flu shots compared to women \n\n4\t\nJournal of Health and Social Behavior 65(1) \n(Vaidya, Partha, and Karmakar 2012). Yet at the \nsame time, research shows that when accessing \nhealth care, women are less likely than men to \nreceive the most advanced or effective treatments \nand more likely to experience physician bias and \ndiscrimination that can undermine their health \n(Borkhoff et al. 2008; Chapman, Kaatz, and Carnes \n2013; Greenwood, Carnahan, and Huang 2018). \nExamining these patterns of gender differences is \ncrucial, but this approach alone provides an incom-\nplete picture because it does not capture the impact \nof structural factors on these outcomes. More spe-\ncifically, this approach does not consider how gen-\nder inequality varies across social contexts in ways \nthat may influence the uptake and receipt of health \ncare services and the effect on health of both men \nand women (Homan 2019; Schofield 2015). A \nstructural sexism approach shifts the focus from \ncomparisons between men and women to within-\ngender comparisons across social contexts charac-\nterized by varying levels of systematic gender \ninequality (i.e., structural sexism). This approach \nilluminates how the unequal gendered distribution \nof valued resources and opportunities in a society \nmay shape the health of all its members.\nStructural Sexism, Health, and Health Care\nA nascent body of research has linked structural \nsexism in U.S. state-level institutions to a variety of \nnegative health outcomes, including worse self-\nrated health, more chronic conditions, and worse \nphysical functioning among both men and women; \nhigher mortality risk among men; more depressive \nsymptoms, higher risk of disordered eating, and \nmore unnecessary c-sections among women; and \nhigher infant mortality rates (Beccia et al. 2022; \nChen et al. 2005; Homan 2019; Kavanagh et al. \n2017, 2018; Kawachi et al. 1999; Nagle and Samari \n2021). While this growing evidence sheds light on \nthe associations between structural sexism and pop-\nulation health outcomes, the role of health care \nremains poorly understood. Only one study has \nexamined U.S. state-level structural sexism and \nhealth care access among men and women, and it \nfound that higher structural sexism is associated \nwith more barriers to health care access for women \nbut not men (Rapp et al. 2022). To our knowledge, \nno study has focused on understanding structural \nsexism as a determinant of preventive health care \nuse in the United States.\nNevertheless, Rapp et al.\u2019s (2022) study and \nprominent theories of gender and health point to a \nnumber of ways for structural sexism to impact \npreventive health care use, perhaps differentially for \nwomen and men. Particularly relevant are gender \nperformance and gendered power and resource allo-\ncation perspectives (Courtenay 2000; Kavanagh and \nGraham 2019), which generate hypotheses regard-\ning the patterns of association between structural \nsexism and preventive health care use among \nwomen and men. A gender performance perspective \nsuggests that structural sexism fosters hegemonic \ngender norms that men and women enact through \nthe performance of health-related behaviors, includ-\ning preventive health care use. A gendered power \nand resource allocation perspective argues that \nstructural sexism disempowers women, which sets \nin motion social, political, and economic processes \nthat limit the health-promoting resources available \nto everyone (e.g., health care and social program \nspending). Figure 1 illustrates the patterns of associ-\nation predicted by the two theoretical perspectives, \nwhich we discuss in detail in the following.\nStructural sexism and gender performance.\u2002 A gen-\nder performance perspective links structural sexism \nto preventive health care use through its connection \nto gender relations and norms that shape men\u2019s and \nwomen\u2019s behavior. From this perspective, gender is \nan interactive accomplishment, something people do \nin everyday life rather than who or what people are \n(West and Zimmerman 1987). Individuals \u201cdo gen-\nder\u201d when they orient their behavior toward widely \naccepted norms of masculinity and femininity (West \nand Zimmerman 1987). Moreover, gender relations \nmanifest at macro levels such as in the economy and \nthe state (Connell 2012). For instance, institutions \nsuch as transnational corporations can maintain gen-\ndered division of labor in the workplace. Therefore, \na sexist environment at macro levels can shape pro-\ncesses, even if indirectly, that reinforce doing gender \nat meso and micro levels. Additionally, research and \ntheory suggest that the subordination of women in \npatriarchal societies is linked to stronger dominance \nhierarchies among men, resulting in increased com-\npetition and amplifying the importance of conform-\ning to hegemonic masculine norms (Connell 1987, \n2012; Wilkinson 2005).\nMasculinities and health theory argues that men \ndemonstrate their conformity to hegemonic mascu-\nline ideals of strength, bravery, stoicism, self- \nreliance, control, dominance, and sexual prowess/\nvirility through risk-taking and unhealthy behaviors \nin order to preserve their status and patriarchal priv-\nilege (Cheng 1999; Connell 1987, 2005, 2012; \nConnell and Messerschmidt 2005; Gray et al. 2002; \nKavanagh and Graham 2019; Mahalik, Burns, and \n\nDore et al.\t\n5\nSyzdek 2007). Indeed, studies have shown that \nadherence to these masculine norms is linked to \ngreater substance use, violence, sexual risk behav-\niors, health care avoidance, and a variety of other \nnegative health-related beliefs and behaviors \n(Courtenay 2000; Fleming and Agnew-Brune 2015; \nMahalik et al. 2007; Seidler et al. 2016). In terms of \npreventive care, there is evidence that masculinity \nnorms (particularly avoidance of femininity, risk-\ntaking, and self-reliance) are inversely associated \nwith colorectal cancer screening (Christy, Mosher, \nand Rawl 2014).\nNorms of femininity are largely framed in oppo-\nsition to masculinity. Emphasized femininity is the \ncomplement to hegemonic masculinity and involves \norienting one\u2019s behavior toward the interests and \ndesires of men (Connell 1987). Such femininity typ-\nically entails gentleness, nurturing, passivity, beauty, \nyouth, fragility, and a domestic/family orientation. \nWhile there is much less research on femininity and \nhealth, positive health beliefs and behaviors, includ-\ning health care use, are typically considered  \nfeminine-typed behavior (Courtenay 2000; Fleming \nand Agnew-Brune 2015). Additionally, the feminine \nideals of youth, beauty, and nurturing may lead to \nincreased contact with health care providers to \nobtain services for women themselves and their \nchildren (Daly and Groes 2017). In sum, to the \nextent that structural sexism increases pressure to \nconform to the hegemonic ideals of masculinity and \nemphasized femininity (Connell 1987, 2012), we \nwould expect greater structural sexism to be associ-\nated with higher levels of preventive health care use \namong women and lower levels among men (see \nFigure 1a).\nStructural sexism and gendered power and resource \nallocation.\u2002 A gendered power perspective links struc-\ntural sexism to preventive health care use through the \nimpact of women\u2019s (dis)empowerment on social, \npolitical, and economic processes that allocate \nresources relevant for population health. At the state \nlevel, more liberal policies expand economic regula-\ntions, protect marginalized groups, and are associated \nwith longer life expectancies (Montez et al. 2020). \nSuch policies are more likely to support women and \nbe supported by women in power (Kavanagh and \nGraham 2019). Evidence from around the world also \nshow that when women are empowered socially and \npolitically, there are greater investments in education, \nhealth care, public health, and other social programs \nthat tend to improve health for the entire population \n(Boehmer and Williamson 1996; Bolzendahl and \nBrooks 2007; Little, Dunn, and Deen 2001; Miller \n2008; Young 2001). Differences in state-level poli-\ncies can also affect health care use through barriers to \naccessing services, such as lack of insurance options \nlike expansion of Medicaid, unavailability of flexible \nappointments, inadequate transportation, and poor \nsocial support for childcare, among other factors \n(National Academies of Sciences, Engineering, and \nMedicine 2018). This leads to the hypothesis that \nstates with higher levels of structural sexism may \noffer less generous safety-net policies and allocate \nfewer resources to health care, leading to lower levels \nof preventive health care use among both men and \nwomen, as illustrated in Figure 1b. This pattern \nwould be consistent with the findings of Homan \n(2019) that state-level structural sexism exposure \nwas universally harmful for health, negatively \nimpacting outcomes for both men and women.\nFigure 1.\u2002 Potential Relationships between Structural Sexism and Preventive Care Use Based on \nGender Theories.\n\n6\t\nJournal of Health and Social Behavior 65(1) \nBecause structural sexism involves material and \nsocial advantages for men, it is also theoretically \npossible that men could leverage their greater per-\nsonal resources to achieve good health even in the \nabsence of health-promoting collective conditions, \nand we might thereby observe a positive associa-\ntion between greater sexism and preventive care \namong men. However, this is unlikely based on \nmasculinity and health research and previous find-\nings on structural sexism and men\u2019s health that \nhave found harmful effects of state-level structural \nsexism on men\u2019s health (Homan 2019; Kavanagh  \net al. 2018). Thus far, no empirical studies have \nidentified a health benefit of state-level sexism \nexposure for men. For this reason, we do not picture \nor further explicate this additional possible but \nunlikely scenario.\nThe Present Study\nIn this study, we examine the association between \nstate-level structural sexism and preventive health \ncare use among men and women in the United States. \nWe construct a measure of structural sexism based on \nHoman (2019) and estimate gender-stratified multi-\nlevel models that capture state-level context and \nindividual-level demographics. We ask: (1) Is struc-\ntural sexism associated with preventive health care \nuse among women and men? and (2) If so, are the \npatterns more consistent with theories of gender \nnorms and health behaviors or gendered power and \nresource allocation? It is important to note that both \ntheoretical perspectives predict a negative relation-\nship between structural sexism and preventive care \nuse for men. Therefore, we can only confirm/deny \nthat one or both of these theories applies to this case. \nBut the theoretical perspectives generate conflicting \nhypotheses regarding the direction of the relationship \namong women, and we can therefore determine \nwhich is most supported by the evidence.\nData And Methods\nWe build on previous work on structural sexism \n(Homan 2019; Homan and Burdette 2021) and \nexamine the relationship between individual-level \nhealth data with a measure of structural sexism at \nthe state level and other relevant state-level and \nindividual-level covariates. For individual-level \ndata, we used the BRFSS from 2018 (Centers for \nDisease Control and Prevention, 2018). For state-\nlevel environments, we included controls from the \nU.S. Census Bureau and a measure of structural sex-\nism constructed from various administrative \nsources, including the U.S. Census Bureau, the \nBureau of Labor Statistics, the Center for American \nWomen and Politics, and Guttmacher Institute. We \nlinked these state-level measures and individual-\nlevel data from BRFSS to examine the association \nbetween structural sexism and gendered use of pre-\nventive health care.\nSample\nThe sample was composed of men and women from \nthe BRFSS national survey in 2018. BRFSS is the \nlargest continuously conducted health survey in the \nworld and collects annual, cross-sectional data from \nrespondents in all 50 states about health behaviors \nand conditions and demographics. Some questions \nvary each year, thus we used 2018 data because it \nwas the most recent year that had the most applica-\nble and complete data on preventive health care use. \nIt was also important to avoid 2020 data due to the \nimpact of COVID-19 on health care use. Our ana-\nlytic sample consisted of 425,454 individuals, of \nwhich 192,854 were men and 232,600 were women. \nThe sample sizes varied for each preventive health \ncare service outcome, depending on how many indi-\nviduals responded to each of the questions.\nDependent Variables\nOur analysis involved several dependent variables \nthat measured the use of preventive health care from \nBRFSS. Several of these BRFSS variables have dif-\nferent versions that factor age of respondent and fre-\nquency of use of the service. For these measures, we \nanalyzed the versions that included the largest sample \nsizes to provide the most complete picture of gen-\ndered patterns of health care use. For example, there \nwere two versions of the question about mammo-\ngrams: (1) women who have ever had a mammogram \nand (2) women age 40+ who have had a mammo-\ngram in the past 2 years. We used the former measure \nbecause the 2018 mammogram recommendations \nvary for women 40 years and older, and therefore, the \nlatter measure may miss important variation by limit-\ning the time period. Most health care services apply \nto both men and women; however, there were three \nthat were asked only of women and one that was \nasked only of men (described in the following).\nAll respondents were asked about a variety of \nhealth care services. These included dichotomous \nmeasures of whether respondents had a person they \nthought of as their personal doctor, if they had ever \nhad a colonoscopy/sigmoidoscopy or been tested \nfor human immunodeficiency virus (HIV), and \n\nDore et al.\t\n7\nwhether within the past year they had visited a doc-\ntor for a routine checkup, gotten a flu shot, or vis-\nited a dentist, dental hygienist, or dental clinic. For \nwomen, they were also asked if they had ever had a \nmammogram, pap test, or human papillomavirus \n(HPV) test. For men, they asked if they had ever \nhad a prostate-specific antigen (PSA) test. All vari-\nables were coded 0/1 so that 1 indicated the partici-\npant had used the health care service within the \nspecified timeframe.\nIndependent Variables\nWe constructed a state-level measure of structural \nsexism based on Homan (2019) from a variety of data \nsources. The sexism measure is a composite score of \npolitical, economic, cultural, and physical/reproduc-\ntive factors. We used data that were as chronically \nclose to 2018 to be relevant to the individual BRFSS \ndata, depending on data availability for each source. \nFor the political measure, we calculated percentage \nof state legislature seats occupied by men in 2018 \nusing data from the Center for American Women and \nPolitics (2023). For the cultural measure, we included \nthe percentage of state population composed of reli-\ngious conservatives (evangelical Protestant or Latter-\nDay Saints) from Pew Research in 2014 (Pew \nResearch Center 2014). This is an important indicator \nof structural sexism because conservative religious \ninstitutions endorse gender essentialist beliefs and \nrestrict women to subordinate roles in the church, the \nfamily, and society at large (Barr 2021; Chaves and \nEagle 2015; Council on Biblical Manhood and \nWomanhood [CBMW] 2023; Homan 2019). The \nprevalence of religious conservatives in a state is \nassociated with traditional/patriarchal gender norms \nnet of individual attitudes (Moore and Vanneman \n2003). For economic measures, we calculated a ratio \nof men\u2019s to women\u2019s labor force participation rates, \nages 16+, a ratio of men\u2019s to women\u2019s poverty rate \nfrom IPUMS USA (Ruggles et al. 2022), and a ratio \nof men\u2019s to women\u2019s median usual weekly earnings \nof full-time wage and salary workers in 2018 (U.S. \nBureau of Labor Statistics 2019). We then standard-\nized and summed all variables to create a structural \nsexism index (\u03b1 = .79) and standardized the final \nstructural sexism index for an easier interpretation of \nresults. A higher value on the structural sexism index \nindicates more structural sexism.\nAdditional Covariates\nAt the individual level, we controlled for a range of \ncharacteristics known to be associated with health \ncare use. These included continuous measures of \nage and income, parental status (has at least one \nchild under 18 in the house or no children under 18 \nin the house), marital status (married or a member of \nan unmarried couple or not married/in a couple), \nand categorical measures of education (less than \nhigh school, high school graduate, some college, \nand bachelor degree), and race (White, Black, \nAmerican Indian or Alaskan Native, Asian, Native \nHawaiian or Pacific Islander, other race, multiracial, \nand Hispanic). We also controlled for a categorical \nvariable of self-rated health (poor, fair, good, very \ngood, and excellent) given that individuals in worse \nhealth may be more likely to seek health care ser-\nvices. Finally, we controlled for insurance status to \nisolate the effect of structural sexism on preventive \nhealth care use given that Rapp et al. (2022) showed \nthat higher state-level sexism is associated with \nbeing uninsured for both men and women.\nBecause we are examining the associations \nbetween individual health care behaviors with \nstate-level sexism contexts, we also controlled for \nother relevant state-level measures. These included \nracial composition (% non-White) and poverty rate \nin 2018 from IPUMS USA (Ruggles et al. 2022), \nthe state Gini coefficient (Frank 2021), and if the \nstate is in the South based on U.S. census defini-\ntions (U.S. Census Bureau 2017).\nAnalytic Strategy\nFirst, we calculated descriptive statistics for the state-\nlevel measures and individual-level measures. For the \nindividual-level descriptive statistics, we stratified by \ngender and estimated t tests for continuous variables, \ntest of proportion for dichotomous variables, and chi-\nsquare tests for categorical variables to understand \nstatistical differences between men and women. Then, \nwe stratified the sample by gender and ran several \nmultilevel logit models, with individuals nested \nwithin states, for each health care service. These mod-\nels predicted the use of each preventive health care \nservice as a function of state-level structural sexism \nexposure. Each model included all previously men-\ntioned state-level and individual-level covariates. We \nretained the largest sample size for each outcome, \nwhich means that the sample size varies depending on \nhow many individuals responded to each question. \nThe sample sizes range from 115,012 to 182,582 for \nwomen and 95,317 to 159,469 for men.\nSupplemental analyses.\u2002 We ran several supple-\nmental analyses to check the robustness of our find-\nings to different specifications. First, we replicated \n\n8\t\nJournal of Health and Social Behavior 65(1) \nthe main analyses but included a control for if the \nstate had expanded Medicaid by 2018 (Kaiser Fam-\nily Foundation 2023). Medicaid expansion increased \naccess to health insurance, and health insurance \nfacilitates access to care, thus it is possible this state-\nlevel policy would affect use of preventive care. \nHowever, a gendered power and resource allocation \nperspective suggests that Medicaid expansion deci-\nsions are a consequence of structural sexism (i.e., \nlower structural sexism leads to greater investment \nin social and health policy that improves population \nhealth), and indeed, some studies have used Medic-\naid expansion as an indicator of structural sexism \n(Rapp et al. 2022). Therefore, controlling for Med-\nicaid expansion risks underestimating the impact of \nstructural sexism on preventive care to the extent \nthat it functions as a mechanism. Thus, we chose to \ninclude this measure in supplemental models (see \nAppendix Tables 1 and 2 in the online version of the \narticle) rather than the main analyses. For the sec-\nond supplemental analysis (see Appendix Tables 3 \nand 4 in the online version of the article), we \nremoved self-rated health as a covariate from the \nmain model. We used self-rated health as a measure \nof general health, which is likely to be associated \nwith health care use in general because individuals \nwho are sick are more likely to receive care than \nindividuals who are not sick. We chose to include \nthis control in our main analyses to remove the \nimpact of health status on preventive care use, but \nwe did not include it for a supplemental analysis to \nunderstand if it affects the relationship between \nstructural sexism and preventive health care use.\nThe third supplemental analysis was a modifica-\ntion to the measure of structural sexism. One recent \nstudy suggested the removal of religious conserva-\ntism as an indicator of structural sexism (McKetta  \net al. 2022); therefore, we estimated models with this \nitem left out of the sexism index (see Appendix \nTables 5 and 6 in the online version of the article). \nResults were substantively similar, and we chose to \nretain the religious conservatism item in our main \nanalysis because theory and previous research high-\nlight the important role of conservative religious insti-\ntutions as a foundational element of sexist oppression \nin the United States (Barr 2021; CBMW 2023; \nChaves and Eagle 2015; Homan 2019; Homan and \nBurdette 2021; Moore and Vanneman 2003). Fourth, \nwe replicated our analyses with samples limited to \nonly ages recommended by the U.S. Preventive \nServices Task Force for the applicable preventive \ncare services (see Appendix Tables 7 and 8 in the \nonline version of the article). Because the screening \nand testing services, such as mammography and HPV \ntests, are generally recommended only for specific \nages, individuals within these age ranges may be \nmore likely to get the services regardless of structural \nsexism. However, we included all ages in the main \nanalyses because recommendations may vary by \ncharacteristics other than age, such as family history. \nFinally, we tested an interaction between gender and \nstructural sexism for the full sample to understand \nwhether the effect of structural sexism differs by gen-\nder (see Appendix Table 9 in the online version of the \narticle). Results remained largely consistent across all \nthese supplemental models, indicating that our key \nfindings and conclusions are robust to a variety of \nalternative specifications. See the Appendix in the \nonline version of the article for further details.\nResults\nDescriptive statistics are shown in Table 1 for state-\nlevel measures and Table 2 for individual-level \nmeasures. Earnings ratio, labor force ratio, and pov-\nerty ratio all have means that are above 1, indicating \ngender inequality that favors men. Proportion of \nmen in state legislature is 74%, signaling that men \nconsistently outnumber women in local govern-\nment, and proportion of women without abortion \naccess is almost 50%, again showing women in a \ndisadvantaged position, this time regarding repro-\nductive health care access.\nIndividual-level descriptive statistics show that \nwomen are more likely to receive preventive health \ncare services compared to men and that the differ-\nences are statistically significant. For example, \n83.8% of women had seen a doctor in the past year \ncompared to 76.3% of men. Similarly, 87.6% of \nwomen have a person they consider a personal doc-\ntor compared to 78.0% of men. These were the two \nmost highly endorsed measures, while the health \ncare services least likely to be used were getting a \nflu shot in the past year (36.7% of men and 41.8% \nof women) and ever having tested for HIV (33.0% \nof men and 33.5% of women). Women were \nslightly older on average (56.2 years compared to \n53.5 years), were less likely to be married (52.0% \ncompared to 59.0%), were more likely to be parents \n(36.9% compared to 25.3%), and had lower \nincomes ($50,382 on average compared to \n$56,439). The remainder of individual descriptive \nstatistics were statistically different although per-\nhaps not meaningfully different, including the \nracial breakdown, self-rated health, and education.\nThe results for the gender-stratified multilevel \nmodels that predict use of preventive health care \nservices conditional on exposure to structural \n\nDore et al.\t\n9\nsexism are in Table 3 for women and Table 4 for \nmen. Figure 2 visualizes the results for both men \nand women in a forest plot with a darker color rep-\nresenting significant results. Overall, we found that \nboth women and men were less likely to use pre-\nventive services in states with more structural sex-\nism. Women were less likely to have had a \ncolonoscopy or sigmoidoscopy (OR = .94, p = \n.018), to have tested for HIV (OR = .82, p < .001), \nand to have had a mammogram (OR = .95, p = \n.009), a pap test (OR = .94, p = .011), and an HPV \ntest (OR = .87, p < .001) in states with more struc-\ntural sexism compared to women in states with less \nstructural sexism. Men were similar in that they \nwere less likely to have a personal doctor (OR = \n.91, p = .031), to have had a colonoscopy or sig-\nmoidoscopy (OR = .93, p = .006), and to have \ntested for HIV (OR = .85, p < .001) in states with \nmore structural sexism compared to men in states \nwith less structural sexism. However, men also \nwere more likely to have had a PSA test in states \nwith more structural sexism (OR = 1.06, p = .003), \nwhich was the only service positively associated \nwith structural sexism.\nThe coefficients can be interpreted as the ORs \nfor using a particular service associated with a 1 \nstandard deviation increase in the structural sexism \nindex. For example, a 1 standard deviation increase \nin the structural sexism index is associated with a \n.87 OR of women getting tested for HPV, which \ntranslates to 13% lower odds. To further aid in the \ninterpretation of these results, we used our model to \ncalculate the predicted probabilities of HPV testing \namong women across the range of sexism \nexposures observed in our sample: A woman \nexposed to the lowest observed level of sexism has \na .51 predicted probability of testing for HPV, while \na woman exposed to the highest level of sexism has \nonly a .38 predicted probability of testing for HPV.\nDiscussion\nThis study measures the association between state-\nlevel exposure to structural sexism on gendered use \nof preventive health care. We build on the growing \nliterature examining the relationship between struc-\ntural sexism and health (Homan 2019; McKetta  \net al. 2022; Rapp et al. 2022; Rapp, Volpe, and \nNeukrug 2021). Preventive health care use is an \nimportant outcome to study in this context due to its \nrelevance to population health outcomes and the \ngendered patterns of use (USDHHS 2022). Most \nstudies focus on individual-level factors in deter-\nmining health care use despite the growing empha-\nsis on upstream structural determinants in the health \ndisparities literature (Braveman and Gottlieb 2014; \nHoman and Brown 2022; Montez 2017; Montez, \nHayward, and Zajacova 2021). Our research ques-\ntions were: (1) Is structural sexism associated with \npreventive health care use among women and men? \nand (2) If so, are the patterns more consistent with \ntheories of gender norms and health behaviors or \ngendered power and resource allocation? We had \ntwo hypotheses based on two different perspectives. \nFirst, based on the gender performance perspective, \nwe hypothesized that in states with more structural \nsexism, men would be less likely to use preventive \nhealth care and women would be more likely to use \nTable 1.\u2002 State-Level Data and Descriptive Statistics (N = 50).\nMeasure\nData Source\nMean (SD)\nRange\nStructural Sexism Index\n.00 (1)\n[\u22121.71, 2.84]\n\u2003 Earnings ratio (M:W)\nBureau of Labor Statistics\n1.25 (.07)\n[1.13, 1.47]\n\u2003 Labor force ratio (M:W)\nIPUMS American Community Survey\n1.15 (.03)\n[1.08, 1.26]\n\u2003 Poverty ratio (W:M)\nIPUMS American Community Survey\n1.05 (.06)\n[.89, 1.22]\n\u2003 Proportion men in state  \nlegislature\nCenter for American Women in \nPolitics\n.74 (.08)\n[.60, .89]\n\u2003 Proportion women without  \nabortion access\nGuttmacher Institute\n.47 (.26)\n[.03, .96]\nPercentage religious conservatives\nPEW Research Center\n.29 (.12)\n[.10, .62]\nRacial composition\nU.S. Census Bureau\n.21 (.12)\n[.05, .74]\nPoverty rate\nU.S. Census Bureau\n.12 (.03)\n[.07, .20]\nGini coefficient\nFrank (2021)\n.61 (.03)\n[.55, .70]\nSouthern region\nU.S. Census Bureau\n16 states\nNote: W:M\u2009=\u2009women-to-men; M:W\u2009=\u2009men-to-women.\n\n10\t\nJournal of Health and Social Behavior 65(1) \npreventive health care compared to their same- \ngender counterparts in states with less sexism. \nSecond, based on the gendered power and resource \nallocation perspective, we argued that in states with \nmore structural sexism, there would be less funding, \neducation, and infrastructure for health care such \nthat both men and women would be less likely to \nuse preventive health care compared to their coun-\nterparts in states with less sexism.\nOverall, our results indicate a strong negative \nrelationship between exposure to structural sexism \nat the state level and the use of preventive health \nTable 2.\u2002 Descriptive Statistics, Behavioral Risk Factor Surveillance System, 2018, N = 425,454.\nMen\nWomen\nGender Difference\n\u2002\n(n = 192,854)\n(n = 232,600)\np Value\nPreventive care use\n\u2003 Visited a doctor in past year\n76.3%\n83.8%\n.001\n\u2003 Have a personal doctor\n78.0%\n87.6%\n.001\n\u2003 Flu shot within past year\n36.7%\n41.8%\n.001\n\u2003 S\u0007igmoidoscopy or colonoscopy within \npast year\n73.0%\n75.6%\n.001\n\u2003 Ever tested for HIV\n33.0%\n33.5%\n.006\n\u2003 Visited dentist in past year\n65.5%\n70.6%\n.001\n\u2003 Ever had a mammogram\n\u2014\n79.2%\n\u2014\n\u2003 Ever had a pap test\n\u2014\n94.1%\n\u2014\n\u2003 Women who ever had HPV test\n\u2014\n43.2%\n\u2014\n\u2003 Ever had a PSA test\n51.7%\n\u2014\n\u2014\nIndividual-level controls\n\u2003 Age\n53.5 (17.6)\n56.2 (17.2)\n.001\n\u2003 Race\n.001\n\u2003 \u2003 White\n77.1%\n77.3%\n\u2002\n\u2003 \u2003 Black\n7.1%\n9.0%\n\u2002\n\u2003 \u2003 American Indian or Alaskan Native\n1.8%\n1.8%\n\u2002\n\u2003 \u2003 Asian\n2.6%\n1.9%\n\u2002\n\u2003 \u2003 Native Hawaiian or Pacific Islander\n.4%\n.3%\n\u2002\n\u2003 \u2003 Other race\n.9%\n.6%\n\u2002\n\u2003 \u2003 Multiracial\n2.1%\n1.9%\n\u2002\n\u2003 \u2003 Hispanic\n8.0%\n7.2%\n\u2002\n\u2003 Married/in a couple\n59.0%\n52.0%\n.001\n\u2003 Income\n56,439 (28,658)\n50,382 (29,123)\n.001\n\u2003 Insurance\n90.3%\n93.1%\n.001\n\u2003 Parent\n25.3%\n36.9%\n.001\n\u2003 Education\n.001\n\u2003 \u2003 Less than high school\n7.8%\n7.0%\n\u2002\n\u2003 \u2003 High school degree\n28.2%\n26.8%\n\u2002\n\u2003 \u2003 Some college\n26.1%\n29.0%\n\u2002\n\u2003 \u2003 Bachelor degree\n37.9%\n37.2%\n\u2002\n\u2003 Self-rated health\n.001\n\u2003 \u2003 Excellent\n16.8%\n16.1%\n\u2002\n\u2003 \u2003 Very good\n32.5%\n33.1%\n\u2002\n\u2003 \u2003 Good\n32.2%\n31.3%\n\u2002\n\u2003 \u2003 Fair\n13.4%\n14.0%\n\u2002\n\u2003 \u2003 Poor\n5.1%\n5.5%\n\u2002\nNote: Descriptive statistics use data from the Behavioral Risk Factor Surveillance System and other sources listed in \nTable 1. HIV = human immunodeficiency virus; HPV = human papillomavirus; PSA = prostate-specific antigen.\n\n11\nTable 3.\u2002 Associations between Structural Sexism and Preventive Health Care Use among Women, Odds Ratios.\nVisited a \nDoctor in Past \nYear\nHas a  \nPersonal \nDoctor\nFlu Shot  \nin Past Year\nEver Had a \nSigmoidoscopy \nor \nColonoscopy\nEver  \nTested  \nfor HIV\nVisited  \nDentist in  \nPast Year\nEver Had a \nMammogram\nEver Had  \na Pap Test\nEver Had  \nHPV Test\nStructural sexism\n.99\n.92\n1.03\n.94*\n.82***\n1.02\n.95**\n.94*\n.87***\nAge\n1.03***\n1.04***\n1.03***\n1.05***\n.96***\n1.01***\n1.14***\n1.05***\n.96***\nRace (reference: White)\n\u2003 Black\n2.06***\n1.15***\n.81***\n1.27***\n2.49***\n.97\n1.52***\n.96\n1.38***\n\u2003 AIAN\n1.31***\n.54***\n.99\n.76***\n1.75***\n1.08\n1.22**\n.72***\n1.45***\n\u2003 Asian\n1.14**\n.78***\n1.37***\n.66***\n.43***\n.98\n.79***\n.17***\n.38***\n\u2003 NHPI\n1.17\n.76*\n.94\n.60***\n.73**\n.80*\n1.30*\n.47***\n.78*\n\u2003 Other race\n.93\n.77**\n.85*\n.85\n1.46***\n.97\n1.07\n.51***\n1.19*\n\u2003 Multiracial\n1.05\n.89*\n.89**\n.97\n1.65***\n.79***\n1.04\n1.01\n1.30***\n\u2003 Hispanic\n1.40***\n.80***\n1.13***\n.93\n1.07**\n1.32***\n1.33***\n.73***\n1.00\nEducation (reference: < HS)\n\u2003 HS degree\n1.03\n1.28***\n.87***\n1.27***\n.92**\n1.37***\n1.18***\n1.28***\n1.03\n\u2003 Some college\n1.01\n1.38***\n.96\n1.56***\n1.37***\n1.63***\n1.25***\n1.94***\n1.35***\n\u2003 College degree\n.99\n1.36***\n1.28***\n1.84***\n1.57***\n2.34***\n1.23***\n3.47***\n1.62***\nMarried/in a couple\n.98\n1.11***\n.95***\n1.20***\n.81***\n1.02\n1.10***\n2.43***\n1.06***\nIncome\n1.00***\n1.00***\n1.00***\n1.00***\n1.00\n1.00***\n1.00***\n1.00***\n1.00***\nInsurance\n3.67***\n4.39***\n2.21***\n2.42***\n.98\n1.94***\n1.36***\n1.11**\n1.16***\nParent\n.90***\n1.12***\n1.01\n.67***\n1.73***\n.93***\n.85***\n2.42***\n1.69***\nSelf-rated health\n.84***\n.85***\n.95***\n.89***\n.82***\n1.28***\n.92***\n.94***\n.97***\nIntercept\n.67\n1.02\n.24**\n.10***\n1.08\n.16***\n.004***\n1.07\n4.83***\nLevel 2 variance (SE)\n.04***\n(.01)\n.08***\n(.02)\n.03***\n(.01)\n.02***\n(.01)\n.04***\n(.01)\n.01***\n(.002)\n.01***\n(.003)\n.02***\n(.01)\n.02***\n(.01)\nn\n181,713\n182,582\n178,178\n115,012\n168,197\n182,082\n176,741\n176,040\n131,406\nNote: All models control for state-level measures of poverty rate, Gini coefficient, racial composition, and southern region. Models use data from the Behavioral Risk Factor \nSurveillance System and other sources listed in Table 1. The sample sizes range from 115,012 to 182,582 observations depending on the outcome. HIV = human immunodeficiency \nvirus; HPV = human papillomavirus; AIAN\u2009=\u2009American Indians and Alaska Natives; NHPI\u2009=\u2009Native Hawaiian and Pacific Islander; HS\u2009=\u2009high school.\n*p\u2009<\u2009.05. **p\u2009<\u2009.01. ***p\u2009<\u2009.001.\n\n12\nTable 4.\u2002 Associations between Structural Sexism and Preventive Health Care Use among Men, Odds Ratios.\nVisited a Doctor  \nin Past Year\nHas a Personal \nDoctor\nFlu Shot in Past \nYear\nEver had a \nSigmoidoscopy  \nor Colonoscopy\nEver Tested for \nHIV\nVisited Dentist in \nPast Year\nEver Had a PSA \nTest\nStructural sexism\n.98\n.91*\n1.02\n.93**\n.85***\n1.00\n1.06**\nAge\n1.04***\n1.05***\n1.03***\n1.06***\n.98***\n1.01***\n1.09***\nRace (reference: White)\n\u2003 Black\n1.79***\n1.07*\n.90***\n1.15***\n3.02***\n.94**\n1.45***\n\u2003 AIAN\n1.26***\n.79***\n1.00\n.71***\n1.75***\n.99\n.76***\n\u2003 Asian\n1.32***\n1.05\n1.40***\n.58***\n.51***\n.82***\n.51***\n\u2003 NHPI\n1.33**\n1.21\n1.08\n.71*\n.84\n.80*\n.59***\n\u2003 Other race\n.93\n.72***\n.80**\n.76**\n1.67***\n.89\n.82*\n\u2003 Multiracial\n1.06\n.93\n1.02\n.94\n1.58***\n.79***\n.83**\n\u2003 Hispanic\n1.25***\n.79***\n1.12***\n.86***\n1.23***\n1.28***\n.99\nEducation (reference: < HS)\n\u2003 HS degree\n1.15***\n1.29***\n.99\n1.36***\n1.07*\n1.39***\n1.46***\n\u2003 Some college\n1.22***\n1.46***\n1.16***\n1.78***\n1.38***\n1.67***\n1.92***\n\u2003 College degree\n1.20***\n1.58***\n1.69***\n2.22***\n1.49***\n2.43***\n2.43***\nMarried/in a couple\n1.07***\n1.20***\n1.12***\n1.40***\n.75***\n1.19***\n1.27***\nIncome\n1.00***\n1.00***\n1.00***\n1.00***\n1.00\n1.00***\n1.00***\nInsurance\n3.66***\n4.27***\n2.58***\n2.75***\n1.07**\n2.03***\n1.82***\nParent\n.87***\n.97*\n.94***\n.67***\n1.29***\n.90***\n.61***\nSelf-rated health\n.83***\n.88***\n.92***\n.91***\n.89***\n1.21***\n1.01\nIntercept\n.18***\n.16*\n.09***\n.02***\n.44\n.13***\n.0005***\nLevel 2 variance (SE)\n.03***\n(.01)\n.08***\n(.02)\n.03***\n(.01)\n.03***\n(.01)\n.03***\n(.01)\n.01***\n(.003)\n.01***\n(.003)\nn\n158,644\n159,469\n155,392\n95,317\n146,983\n159,048\n110,839\nNote: All models control for state-level measures of poverty rate, Gini coefficient, racial composition, and southern region. Models use data from the Behavioral Risk Factor \nSurveillance System and other sources listed in Table 1. The sample sizes range from 95,317 to 159,469 observations depending on the outcome. HIV = human immunodeficiency \nvirus; PSA = prostate-specific antigen; AIAN\u2009=\u2009American Indians and Alaska Natives; NHPI\u2009=\u2009Native Hawaiian and Pacific Islander; HS\u2009=\u2009high school.\n*p\u2009<\u2009.05. **p\u2009<\u2009.01. ***p\u2009<\u2009.001.\n\nDore et al.\t\n13\ncare services among both men and women. These \nresults partially supported our hypothesis based on \na gender performance perspective because men \nwere less likely to use preventive health care in \nstates with more structural sexism, consistent with \nthe idea that performing masculinity entails health \ncare avoidance and negative health beliefs and \nbehaviors (Courtenay 2000). However, a gender \nperformance perspective was not generally sup-\nported for women, who were also less likely to use \npreventive health care in states with more structural \nsexism but who were hypothesized to increase \nhealth care utilization in conjunction with perform-\ning femininity. Our results fully supported a gen-\ndered power and resource allocation perspective \nbecause both men and women were overall less \nlikely to use preventive health care services in \nstates with more sexism. This pattern would be \nexpected if the disempowerment of women in sex-\nist environments leads to the contraction of health-\npromoting resources for everyone to access.\nThese findings align with much of the literature \non structural sexism and health but contribute a key \npiece by examining the gendered use of preventive \nhealth care. Most relevant to the current study, \nRapp et al. (2022) found that women in states with \nmore sexism face more barriers to accessing and \naffording health care, while men\u2019s access is unaf-\nfected by structural sexism. Although they did not \nexamine preventive health care explicitly, it is pos-\nsible these findings apply to preventive health care \nand can help to explain our findings that women in \nstates with more sexism were less likely to access \npreventive health care. However, they do not help \nto explain our finding that men were generally also \nless likely to access preventive health care in states \nwith more sexism. Nagle and Samari (2021) found \nthat birthing people living in states with more sex-\nism were more likely to have had a cesarean sec-\ntion. These findings align with the results of this \nstudy because women in states with more sexism \nwere not receiving care in the recommended way or \nfrequency. Finally, Homan (2019) found that struc-\ntural sexism was associated with more chronic con-\nditions, worse self-rated health, and worse physical \nfunctioning for women and worse physical func-\ntioning for men. Our study in conjunction with \nRapp et al. (2022) and Nagle and Samari (2021) \nprovide evidence on negative interactions with the \nhealth care system in sexist states, suggesting pos-\nsible explanations for Homan\u2019s (2019) findings that \nwomen and men in sexist states have worse health.\nThere was one particularly interesting finding in \nour study that did not fit the same pattern as the \nother results. In states with more sexism, men were \nmore likely to get PSA tests compared to men in \nstates with less sexism, which was the opposite \nrelationship for all other forms of preventive care \namong men in this study. While initially somewhat \nsurprising, this can also be understood from a modi-\nfied gender performance perspective. From a hege-\nmonic masculinity viewpoint, prostate cancer \n(which can be detected by PSA tests, allowing for \ntimely treatment) presents a unique threat to mascu-\nlinity because it often causes incontinence and erec-\ntile dysfunction, which can be perceived as \nweakness, lack of control over bodily function, and \nfailure to fulfill masculine ideals of sexual prowess \nFigure 2.\u2002 Associations between Structural Sexism and Odds of Using Preventive Health Care Services \namong Women and Men.\nNote: Models use data from the Behavioral Risk Factor Surveillance System and other sources listed in Table 1. The \nsample sizes range from 115,012 to 182,582 for women and 95,317 to 159,469 for men.\n\n14\t\nJournal of Health and Social Behavior 65(1) \n(Gray et al. 2002). Thus, while it may be generally \nconsidered unmasculine to seek out health care \n(especially preventive care that is not an urgent \nnecessity), men may make exceptions for care that \nsustains sexual function. Indeed, a qualitative study \nof men\u2019s constructions of masculinity and their \nhelp-seeking behavior in Scotland found that while \nthere was widespread agreement with the hege-\nmonic view that real men should not need help or \nconsult with physicians for \u201cminor\u201d symptoms or \npain, men were much more in favor of help-seeking \nwhen it was used as a means to preserve other \nenactments of masculinity that they valued, espe-\ncially their physical strength and sexual perfor-\nmance (O\u2019Brien, Hunt, and Hart 2005). Thus, to the \nextent that exposure to high levels of structural sex-\nism exerts increased pressure to conform to hege-\nmonic norms of masculinity, men in these \nenvironments must navigate competing demands of \nmasculine invincibility and self-reliance that \nencourage health care avoidance versus the mascu-\nline imperative to preserve the functioning of the \npenis\u2014itself the very symbol of manhood, power, \nand sexuality (Cheng 1999; Gray et al. 2002; Potts \n2000). In sum, our findings suggest that while \nhigher structural sexism exposure decreases men\u2019s \nuse of preventive care services in general (either \nthrough performing masculinity, decreased health \ncare resources and accessibility, or both), it may \nincrease the demand for preventive health care \ndirectly affecting men\u2019s sexual performance, which \nfunctions as a central element of their identity as \nmen. Future research should investigate the associ-\nations between structural sexism and other sexual \nand urological services and outcomes to further \nexplore this issue.\nLimitations and Future Research\nThis research makes an important contribution to \nthe literature on structural factors that shape health \ncare use and, therefore, health outcomes. However, \nthere are limitations to this research worth noting. \nFirst, we use cross-sectional data from a single year \nand can therefore not claim a causal relationship. \nSecond, we do not explicitly test mechanisms, so \nthe pathways through which structural sexism may \nshape preventive care remain unclear. Rapp et al. \n(2022) provides evidence that states with more sex-\nism bar women from accessing health care gener-\nally, but determining the specific factors and barriers \naffecting women\u2019s and men\u2019s preventive health care \nuse in particular requires further research. We sug-\ngest that future researchers further investigate gen-\nder performance perspectives by combining \nstructural sexism measures with individual-level \ndata on gender norms/beliefs and health care use \nand examining the role of cultural sexism (i.e.,  \ncontextual-level gender norms/ideology; see Price \net al. 2021). Additionally, researchers can further \ninvestigate gendered power and resources alloca-\ntion perspective by examining the relationships \nbetween structural sexism and specific social poli-\ncies (e.g., welfare spending, unemployment insur-\nance, Medicaid expansion, education spending, \nhealth care spending, parental leave, etc.) that redis-\ntribute resources in ways that improve population \nhealth and may therefore act as mechanisms \nbetween structural sexism and preventive health \ncare use. Third, our measures of preventive health \ncare are self-reported and may be subject to recall \nbias, and we also only include the available mea-\nsures of preventive health care in BRFSS. The mea-\nsures are also dichotomous, which may miss more \nnuanced responses. Thus, investigating the associa-\ntions between structural sexism and a variety of \nother measures of health care service use and health \nbehaviors is an important area for future research.\nFourth, we captured several important aspects \nof structural sexism and other state-level and  \nindividual-level covariates that may impact preven-\ntive health care use, but our measure of structural \nsexism is not exhaustive. Future research should \ndevelop additional measures of structural sexism \nand examine the strengths and limitations of differ-\nent measurement approaches. Fifth, there are a vari-\nety of other state-level social, economic, and \npolitical characteristics not included in our models \nthat may be relevant for preventive care and may or \nmay not be related to structural sexism. While these \nfactors could potentially confound the relationship \nbetween sexism and health or health care, it is also \npossible that they serve as important mechanisms. \nThus, careful examination of the complex relation-\nships between structural sexism and various fea-\ntures of state contexts is an important area for future \nresearch.\nSixth, we have focused specifically on the role \nof structural sexism in shaping preventive care \ngiven recent evidence of its associations with \nhealth, but it does not operate in isolation. Homan, \nBrown, and King (2021) provide a theoretical \nframework to examine health and health care out-\ncomes through a structural intersectional lens. \nBecause gendered expectations can be both classed \nand racialized, future research should investigate \nuse of preventive health care at the intersection of \nmultiple oppressive systems. This will involve \nmeasuring multiple forms of structural oppression \n(e.g., structural sexism and structural racism) and \n\nDore et al.\t\n15\nexamining their individual and joint effects on \nhealth care use among individuals with varied inter-\nsectional identity categories (e.g., Black women \nwith a college degree vs. White men without a high \nschool diploma). Research on structural sexism has \nthus far used data with predominantly White sam-\nples, suggesting an intersectional approach is cru-\ncial for providing further theoretical insight.\nIt is also important to consider the role of sexu-\nality in affecting men\u2019s health-seeking behavior. \nStudies find that gay men are likelier to get PSA \ntesting than their heterosexual counterparts (Ma  \net al. 2021; Wilcox Vanden Berg et al. 2022). \nHowever, whether exposure to higher structural \nsexism may or may not affect greater PSA testing \namong gay men warrants further study. Finally, our \nsample is limited by data constraints in the BRFSS \nsuch that we can only study the associations \nbetween structural sexism and preventive health \ncare among cisgender men and women. It is essen-\ntial that future research investigate the impact of \nstructural sexism among transgender and nonbinary \npeople and develop new measures of structural sex-\ngender-sexuality-based oppression such as cishet-\neropartiarchy and gender binarism (Everett et al. \n2022; Krieger 2020).\nConclusion\nStructural sexism is an important determinant of \npreventive health care use for both men and women. \nThese findings suggest a few possible explanations \nrelated to the gender performance and gendered \npower and resource allocation perspectives. Specific \nto men, structural sexism may exacerbate gender \nnorms that disincentivize them from using care to \navoid appearing vulnerable. For both men and \nwomen, a lack of resources and state-level supports \nmay prevent them from receiving preventive health \ncare services. This work contributes to a growing \nbody of research on the universal harm of structural \nsexism and the urgency of dismantling oppressive \ngender systems to improve population health.\nFunding\nThe authors disclosed receipt of the following financial \nsupport for the research, authorship, and/or publication of \nthis article: Research reported in this publication was sup-\nported by the National Institute on Minority Health and \nHealth Disparities of the National Institutes of Health \nunder Award No. F31MD017935, and the Network on \nLife Course Health Dynamics and Disparities in 21st \nCentury America (Award No. 2 R24 AG 045061-06) from \nthe National Institutes on Aging. The content is solely the \nresponsibility of the authors and does not necessarily \nrepresent the official views of the National Institutes of \nHealth.\nORCID iDs\nEmily C. Dore \n https://orcid.org/0000-0002-0899-0200\nPatricia Homan \n https://orcid.org/0000-0003-3609-8188\nSUPPLEMENTAL MATERIAL\nAppendices Table 1 through 9 and Figures 1 through 5 are \navailable in the online version of the article.\nReferences\nAsch, Steven M., Eve A. Kerr, Joan Keesey, John L. \nAdams, Claude M. Setodji, Shaista Malik, and \nElizabeth A. McGlynn. 2006. \u201cWho Is at Greatest \nRisk for Receiving Poor-Quality Health Care?\u201d New \nEngland Journal of Medicine 354(11):1147\u201356. \ndoi:10.1056/nejmsa044464.\nBarr, Beth Allison. 2021. The Making of Biblical \nWomanhood: How the Subjugation of Women Became \nGospel Truth. Grand Rapids, MI: Brazos Press.\nBeccia, Ariel L., S. Bryn Austin, Jonggyu Baek, \nMadina Ag\u00e9nor, Sarah Forrester, Eric Y. Ding, \nWilliam M. Jesdale, and Kate L. Lapane. 2022. \n\u201cCumulative Exposure to State-Level Structural \nSexism and Risk of Disordered Eating: Results \nfrom a 20-Year Prospective Cohort Study.\u201d Social \nScience & Medicine 301:114956. doi:10.1016/j.socs-\ncimed.2022.114956.\nBlack, Lindsey I., Colleen N. Nugent, and Anjel \nVahratian. 2016. \u201cAccess and Utilization of Selected \nPreventive Health Services among Adolescents Aged \n10\u201317.\u201d NCHS Data Brief No. 246. Hyattsville, MD: \nNational Center for Health Statistics.\nBlewett, Lynn A., Pamela Jo Johnson, Brian Lee, and Peter \nB. Scal. 2008. \u201cWhen a Usual Source of Care and \nUsual Provider Matter: Adult Prevention and Screening \nServices.\u201d Journal of General Internal Medicine \n23(9):1354\u201360. doi:10.1007/S11606-008-0659-0.\nBoehmer, Ulrike, and John B. Williamson. 1996. \u201cThe \nImpact of Women\u2019s Status on Infant Mortality \nRate: A Cross-National Analysis.\u201d Social Indicators \nResearch 37(3):333\u201360.\nBolnick, Howard J., Anthony L. Bui, Anne Bulchis, \nCarina Chen, Abigail Chapin, Liya Lomsadze, Ali H. \nMokdad, Francois Millard, and Joseph L. Dieleman. \n2020. \n\u201cHealth-Care \nSpending \nAttributable \nto \nModifiable Risk Factors in the U.S.A.: An Economic \nAttribution Analysis.\u201d The Lancet Public Health \n5(10):e525\u201335. doi:10.1016/S2468-2667(20)30203-6.\nBolzendahl, Catherine, and Clem Brooks. 2007. \n\u201cWomen\u2019s Political Representation and Welfare \nState Spending in 12 Capitalist Democracies.\u201d Social \nForces 85(4):1509\u201334. doi:10.1353/sof.2007.0061.\nBorkhoff, Cornelia M., Gillian A. Hawker, Hans J. \nKreder, Richard H. Glazier, Nizar N. Mahomed, \nand James G. Wright. 2008. \u201cThe Effect of Patients\u2019 \n\n16\t\nJournal of Health and Social Behavior 65(1) \nSex on Physicians\u2019 Recommendations for Total \nKnee Arthroplasty.\u201d CMAJ: Canadian Medical \nAssociation Journal 178(6):681\u201387. doi:10.1503/\ncmaj.071168.\nBorsky, Amanda, Chunliu Zhan, Therese Miller, Quyen \nNgo-Metzger, Arlene S. Bierman, and David Meyers. \n2018. \u201cFew Americans Receive All High-Priority, \nAppropriate Clinical Preventive Services.\u201d Health \nAffairs \n(Milwood) \n37(6):925\u201328. \ndoi:10.1377/\nHLTHAFF.2017.1248.\nBraveman, Paula, and Laura Gottlieb. 2014. \u201cThe Social \nDeterminants of Health: It\u2019s Time to Consider \nthe Causes of the Causes.\u201d Public Health Reports \n129(Suppl. 2):19\u201331.\nBrenner, Hermann, Christian Stock, and Michael \nHoffmeister. \n2014. \n\u201cEffect \nof \nScreening \nSigmoidoscopy and Screening Colonoscopy on \nColorectal Cancer Incidence and Mortality: Systematic \nReview and Meta-analysis of Randomised Controlled \nTrials and Observational Studies.\u201d BMJ (Clinical \nResearch Ed.) 348:g2467. doi:10.1136/bmj.g2467.\nCenter for American Women and Politics. 2023. \u201cWomen \nin State Legislatures 2018.\u201d https://cawp.rutgers.edu/\nfacts/levels-office/state-legislature/women-state-\nlegislatures-2018.\nCenters for Disease Control and Prevention. 2018. \nBehavioral Risk Factor Surveillance System Survey \nData. Atlanta, GA: U.S. Department of Health and \nHuman Services, Centers for Disease Control and \nPrevention.\nChapman, Elizabeth N., Anna Kaatz, and Molly \nCarnes. 2013. \u201cPhysicians and Implicit Bias: How \nDoctors May Unwittingly Perpetuate Health Care \nDisparities.\u201d Journal of General Internal Medicine \n28(11):1504\u201310. doi:10.1007/s11606-013-2441-1.\nChaves, Mark, and Allison Eagle. 2015. \u201cReligious \nCongregations in 21st Century America: A Report \nfrom the National Congregations Study.\u201d Department \nof Sociology, Duke University. https://sites.duke.\nedu/ncsweb/files/2019/02/NCSIII_report_final.pdf\nChen, Ying-Yeh, S. V. Subramanian, Doloros Acevedo-\nGarcia, and Ichiro Kawachi. 2005. \u201cWomen\u2019s \nStatus and Depressive Symptoms: A Multilevel \nAnalysis.\u201d Social Science & Medicine 60(1):49\u201360. \ndoi:10.1016/j.socscimed.2004.04.030.\nCheng, Cliff. 1999. \u201cMarginalized Masculinities and \nHegemonic Masculinity: An Introduction.\u201d The \nJournal of Men\u2019s Studies 7(3):295\u2013315. doi:10.3149/\njms.0703.295.\nChristy, Shannon M., Catherine E. Mosher, and Susan \nM. Rawl. 2014. \u201cIntegrating Men\u2019s Health and \nMasculinity Theories to Explain Colorectal Cancer \nScreening Behavior.\u201d American Journal of Men\u2019s \nHealth 8(1):54\u201365. doi:10.1177/1557988313492171.\nConnell, R. W. 1987. Gender and Power: Society, the \nPerson, and Sexual Politics. Stanford, CA: Stanford \nUniversity Press.\nConnell, Raewyn W. 2005. Masculinities. Rev. ed. \nCambridge, UK: Polity Press.\nConnell, \nRaewyn. \n2012. \n\u201cGender, \nHealth, \nand \nTheory: Conceptualizing the Issue, in Local and \nWorld Perspective.\u201d Social Science & Medicine \n74(11):1675\u201383. \ndoi:10.1016/j.socscimed.2011.06 \n.006.\nConnell, R. W., and James W. Messerschmidt. 2005. \n\u201cHegemonic Masculinity: Rethinking the Concept.\u201d \nGender & Society 19(6):829\u201359. doi:10.1177/ \n0891243205278639.\nCouncil on Biblical Manhood and Womanhood. \n2023. \u201cThe Council on Biblical Manhood and \nWomanhood.\u201d http://cbmw.org/.\nCourtenay, Will H. 2000. \u201cConstructions of Masculinity \nand Their Influence on Men\u2019s Well-Being: A \nTheory of Gender and Health.\u201d Social Science and \nMedicine \n50(10):1385\u2013401. \ndoi:10.1016/S0277-\n9536(99)00390-1.\nDaly, Moira, and Fane Groes. 2017. \u201cWho Takes the \nChild to the Doctor? Mom, Pretty Much All of the \nTime.\u201d Applied Economics Letters 24(17):1267\u201376. \ndoi:10.1080/13504851.2016.1270410.\nDeVoe, Jennifer E., George E. Fryer, Robert Phillips, \nand Larry Green. 2003. \u201cReceipt of Preventive \nCare among Adults: Insurance Status and Usual \nSource of Care.\u201d American Journal of Public Health \n93(5):786\u201391. doi:10.2105/AJPH.93.5.786.\nEverett, \nBethany \nG., \nAubrey \nLimburg, \nPatricia \nHoman, and Morgan M. Philbin. 2022. \u201cStructural \nHeteropatriarchy and Birth Outcomes in the \nUnited \nStates.\u201d \nDemography \n59(1):89\u2013110. \ndoi:10.1215/00703370-9606030.\nFaulkner, Lisa A., and Helen Haplin Schauffler. 1997. \n\u201cThe Effect of Health Insurance Coverage on \nthe Appropriate Use of Recommended Clinical \nPreventive \nServices.\u201d \nAmerican \nJournal \nof \nPreventive Medicine 13(6):453\u201358.\nFerdinands, Jill M., Manjusha Gaglani, Emily T. Martin, \nArnold S. Monto, Donald Middleton, Fernanda \nSilveira, H.Keipp Talbot, Richard Zimmerman, and \nManish Patel. 2021. \u201cWaning Vaccine Effectiveness \nagainst Influenza-Associated Hospitalizations among \nAdults, 2015\u20132016 to 2018\u20132019, United States \nHospitalized Adult Influenza Vaccine Effectiveness \nNetwork.\u201d Clinical Infectious Diseases 73(4):726\u201329.\nFleming, Paul J., and Christine Agnew-Brune. 2015. \n\u201cCurrent Trends in the Study of Gender Norms and \nHealth Behaviors.\u201d Current Opinion in Psychology \n5:72\u201377. doi:10.1016/j.copsyc.2015.05.001.\nFrank, Mark W. 2021. \u201cThe U.S. State-Level Income \nInequality Data.\u201d https://www.shsu.edu/eco_mwf/\ninequality.html.\nGray, Ross E., Margaret I. Fitch, Karen D. Fergus, \nEric Mykhalovskiy, and Kathryn Church. 2002. \n\u201cHegemonic Masculinity and the Experience of \nProstate Cancer: A Narrative Approach.\u201d Journal of \nAging and Identity 7(1):43\u201362.\nGreenwood, Brad N., Seth Carnahan, and Laura Huang. \n2018. \u201cPatient\u2013Physician Gender Concordance and \nIncreased Mortality among Female Heart Attack \n\nDore et al.\t\n17\nPatients.\u201d Proceedings of the National Academy \nof \nSciences \n115(34):8569\u201374. \ndoi:10.1073/\npnas.1800097115.\nHoman, \nPatricia. \n2019. \n\u201cStructural \nSexism \nand \nHealth in the United States: A New Perspective \non Health Inequality and the Gender System.\u201d \nAmerican \nSociological \nReview \n84(3):486\u2013516. \ndoi:10.1177/0003122419848723.\nHoman, Patricia. 2021a. \u201cSexism and Health: Advancing \nKnowledge through Structural and Intersectional \nApproaches.\u201d American Journal of Public Health \n111(10):1725\u201327. doi:10.2105/AJPH.2021.306480.\nHoman, Patricia A. 2021b. \u201cStructural Sexism and Life-\nCourse Health: Implications for Public Policy.\u201d  \nPp. 31\u201340 in Life Course Implications of US Public \nPolicy, edited by J. M. Wilmoth and A. S. London: \nRoutledge.\nHoman, Patricia A., and Tyson H. Brown. 2022. \u201cSick \nand Tired of Being Excluded: Structural Racism in \nDisenfranchisement as a Threat to Population Health \nEquity.\u201d Health Affairs 41(2):219\u201327. doi:10.1377/\nhlthaff.2021.01414.\nHoman, Patricia, Tyson H. Brown, and Brittany King. \n2021. \u201cStructural Intersectionality as a New \nDirection for Health Disparities Research.\u201d Journal \nof Health and Social Behavior 62(3):350\u201370. \ndoi:10.1177/00221465211032947.\nHoman, Patricia, and Amy Burdette. 2021. \u201cWhen \nReligion Hurts: Structural Sexism and Health in \nReligious Congregations.\u201d American Sociological \nReview 86(2):234\u201355.\nHou, Su-I., Diadrey-Anne Sealy, and Caroline W. Kabiru. \n2011. \u201cClosing the Disparity Gap: Cancer Screening \nInterventions \namong \nAsians\u2014A \nSystematic \nLiterature Review.\u201d Asian Pacific Journal of Cancer \nPrevention 12(11):3133\u201339.\nJones, Tarsha P., Maria C. Katapodi, and Joan S. \nLockhart. \n2015. \n\u201cFactors \nInfluencing \nBreast \nCancer Screening and Risk Assessment among \nYoung African American Women: An Integrative \nReview of the Literature.\u201d Journal of the American \nAssociation of Nurse Practitioners 27(9):521\u201329. \ndoi:10.1002/2327-6924.12223.\nKaiser Family Foundation. 2023. \u201cStatus of State \nMedicaid Expansion Decisions: Interactive Map.\u201d \nhttps://www.kff.org/medicaid/issue-brief/status-\nof-state-medicaid-expansion-decisions-interactive-\nmap/.\nKamal, Rabah, and Julie Hudman. 2020. \u201cWhat Do \nWe Know about Spending Related to Public \nHealth in the U.S. and Comparable Countries?\u201d \nPeterson-KFF Health System Tracker. https://www \n.healthsystemtracker.org/chart-collection/what-do-\nwe-know-about-spending-related-to-public-health-\nin-the-u-s-and-comparable-countries/.\nKavanagh, Shane A., and Melissa Graham. 2019. \u201cHow \nGender Inequity Impacts on Men\u2019s Health. An \nExploration of Theoretical Pathways.\u201d International \nJournal of Men\u2019s Social and Community Health \n2(1):e11\u201321. doi:10.22374/ijmsch.v2i1.5.\nKavanagh, Shane A., Julia M. Shelley, and Christopher \nStevenson. 2017. \u201cDoes Gender Inequity Increase \nMen\u2019s Mortality Risk in the United States? A \nMultilevel Analysis of Data from the National \nLongitudinal Mortality Study.\u201d SSM - Population \nHealth 3:358\u201365. doi:10.1016/j.ssmph.2017.03.003.\nKavanagh, Shane A., Julia M. Shelley, and Christopher \nStevenson. 2018. \u201cIs Gender Inequity a Risk Factor \nfor Men Reporting Poorer Self-Rated Health in \nthe United States?\u201d PLoS One 13(7):e0200332. \ndoi:10.1371/journal.pone.0200332.\nKawachi, Ichiro, Bruce P. Kennedy, Vanita Gupta, and \nDeborah Prothrow-Stith. 1999. \u201cWomen\u2019s Status \nand the Health of Women and Men: A View from \nthe States.\u201d Social Science & Medicine 48(1):21\u201332. \ndoi:10.1016/S0277-9536(98)00286-X.\nKim, Eric S., Laura D. Kubzansky, and Jacqui Smith. \n2015. \u201cLife Satisfaction and Use of Preventive Health \nCare Services.\u201d Health Psychology 34(7):779\u201382. \ndoi:10.1037/HEA0000174.\nKim, Eric S., Victor J. Strecher, and Carol D. Ryff. 2014. \n\u201cPurpose in Life and Use of Preventive Health Care \nServices.\u201d Proceedings of the National Academy \nof \nSciences \n111(46):16331\u201336. \ndoi:10.1073/\nPNAS.1414826111.\nKoenen, Karestan C., Alisa Lincoln, and Allison \nAppleton. 2006. \u201cWomen\u2019s Status and Child Well-\nBeing: A State-Level Analysis.\u201d Social Science & \nMedicine 63(12):2999\u20133012. doi:10.1016/j.socs-\ncimed.2006.07.013.\nKrieger, Nancy. 2020. \u201cMeasures of Racism, Sexism, \nHeterosexism, and Gender Binarism for Health \nEquity Research: From Structural Injustice to \nEmbodied Harm\u2014An Ecosocial Analysis.\u201d Annual \nReview of Public Health 41:37\u201362.\nLevine, Susan, Erin Malone, Akaki Lekiachvili, and Peter \nBriss. 2019. \u201cHealth Care Industry Insights: Why the Use \nof Preventive Services Is Still Low.\u201d Preventing Chronic \nDisease 16:180625. doi:10.5888/PCD16.180625.\nLittle, Thomas H., Dana Dunn, and Rebecca E. Deen. \n2001. \u201cA View from the Top: Gender Differences \nin Legislative Priorities among State Legislative \nLeaders.\u201d \nWomen \n& \nPolitics \n22(4):29\u201350. \ndoi:10.1300/J014v22n04_02.\nMa, Sung Jun, Oluwadamilola T. Oladeru, Katy Wang, \nKristopher Attwood, Anurag K. Singh, Daphne A. \nHaas-Kogan, and Paula M. Neira. 2021. \u201cProstate \nCancer Screening Patterns among Sexual and \nGender Minority Individuals.\u201d European Urology \n79(5):588\u201392. doi:10.1016/j.eururo.2020.11.009.\nMahalik, James R., Shaun M. Burns, and Matthew Syzdek. \n2007. \u201cMasculinity and Perceived Normative Health \nBehaviors as Predictors of Men\u2019s Health Behaviors.\u201d \nSocial Science & Medicine 64(11):2201\u2013209. \ndoi:10.1016/j.socscimed.2007.02.035.\nMcKetta, Sarah, Seth J. Prins, Deborah Hasin, Megan E. \nPatrick, and Katherine M. Keyes. 2022. \u201cStructural \nSexism and Women\u2019s Alcohol Use in the United \nStates, 1988\u20132016.\u201d Social Science & Medicine \n301:114976. doi:10.1016/j.socscimed.2022.114976.\n\n18\t\nJournal of Health and Social Behavior 65(1) \nMiller, Grant. 2008. \u201cWomen\u2019s Suffrage, Political \nResponsiveness, and Child Survival in American \nHistory.\u201d The Quarterly Journal of Economics \n123(3):1287\u2013327.\nMontez, \nJennifer \nKaras. \n2017. \n\u201cDeregulation, \nDevolution, and State Pre-emption Laws\u2019 Impact \non U.S. Mortality Trends.\u201d American Journal of \nPublic \nHealth \n107(11):1749\u201350. \ndoi:10.2105/\nAJPH.2017.304080.\nMontez, Jennifer Karas, Jason Beckfield, Julene Kemp \nCooney, Jacob M. Grumbach, Mark D. Hayward, \nHuseyin Zeyd Koytak, Steven H. Woolf, and Anna \nZajacova. 2020. \u201cU.S. State Policies, Politics, and \nLife Expectancy.\u201d Milbank Quarterly 98(3):1\u201332. \ndoi:10.1111/1468-0009.12469.\nMontez, Jennifer Karas, Mark D. Hayward, and Anna \nZajacova. 2021. \u201cTrends in U.S. Population Health: \nThe Central Role of Policies, Politics, and Profits.\u201d \nJournal of Health and Social Behavior 62(3):286\u2013\n301. doi:10.1177/00221465211015411.\nMoore, Laura M., and Reeve Vanneman. 2003. \u201cContext \nMatters: Effects of the Proportion of Fundamentalists \non Gender Attitudes.\u201d Social Forces 82(1):115\u201339. \ndoi:10.1353/sof.2003.0099.\nNagle, Amanda, and Goleen Samari. 2021. \u201cState-\nLevel Structural Sexism and Cesarean Sections \nin the United States.\u201d Social Science & Medicine \n289:114406. doi:10.1016/j.socscimed.2021.114406.\nNational Academies of Sciences, Engineering, and \nMedicine. 2018. Health-Care Utilization as a Proxy \nin Disability Determination. Washington, DC: The \nNational Academies Press.\nO\u2019Brien, Rosaleen, Kate Hunt, and Graham Hart. 2005. \n\u201c\u2018It\u2019s Caveman Stuff, but That Is to a Certain \nExtent How Guys Still Operate\u2019: Men\u2019s Accounts \nof Masculinity and Help Seeking.\u201d Social Science \n& Medicine 61(3):503\u201316. doi:10.1016/j.socscimed \n.2004.12.008.\nPew Research Center. 2014. \u201cEvangelical Protestants.\u201d \nhttps://www.pewresearch.org/religion/religious-\nlandscape-study/religious-tradition/evangelical-prot \nestant/.\nPotts, Annie. 2000. \u201c\u2018The Essence of the Hard \non\u2019: Hegemonic Masculinity and the Cultural \nConstruction of \u2018Erectile Dysfunction.\u2019\u201d Men and \nMasculinities \n3(1):85\u2013103. \ndoi:10.1177/10971\n84X00003001004.\nPrice, Maggi A., Sarah McKetta, John R. Weisz, Jessie \nV. Ford, Micah R. Lattanner, Hilary Skov, Elizabeth \nWolock, and Mark L. Hatzenbuehler. 2021. \u201cCultural \nSexism Moderates Efficacy of Psychotherapy: \nResults from a Spatial Meta-analysis.\u201d Clinical \nPsychology: Science and Practice 28(3):299\u2013312. \ndoi:10.1037/cps0000031.\nRapp, Kristen Schorpp, Vanessa V. Volpe, Tabitha L. \nHale, and Dominique F. Quartararo. 2022. \u201cState-\nLevel Sexism and Gender Disparities in Health \nCare Access and Quality in the United States.\u201d \nJournal of Health and Social Behavior 63(1):2\u201318. \ndoi:10.1177/00221465211058153.\nRapp, Kristen Schorpp, Vanessa V. Volpe, and Hannah \nNeukrug. 2021. \u201cState-Level Sexism and Women\u2019s \nHealth Care Access in the United States: Differences \nby Race-Ethnicity, 2014\u20132019.\u201d American Journal \nof Public Health 111(10):1796\u2013805. doi:10.2105/\nAJPH.2021.306455.\nRidgeway, Cecilia L., and Shelley J. Correll. 2004. \n\u201cUnpacking the Gender System: A Theoretical \nPerspective \non \nGender \nBeliefs \nand \nSocial \nRelations.\u201d Gender and Society 18(4):510\u201331. \ndoi:10.1177/0891243204265269.\nRuggles, Steven, Sarah Flood, Ronald Goeken, Megan \nSchouweiler, and Matthew Sobek. 2022. IPUMS \nUSA: Version 12.0 [Dataset]. Minneapolis, MN: \nIPUMS.\nSchofield, Toni. 2015. A Sociological Approach to \nHealth Determinants. Port Melbourne, Australia: \nCambridge University Press.\nSeidler, Zac E., Alexei J. Dawes, Simon M. Rice, John L. \nOliffe, and Haryana M. Dhillon. 2016. \u201cThe Role of \nMasculinity in Men\u2019s Help-Seeking for Depression: \nA Systematic Review.\u201d Clinical Psychology Review \n49:106\u201318. doi:10.1016/j.cpr.2016.09.002.\nU.S. Bureau of Labor Statistics. 2019. \u201cHighlights of \nWomen\u2019s Earnings in 2018.\u201d https://www.bls.gov/\nopub/reports/womens-earnings/2018/home.htm.\nU.S. Census Bureau. 2017. \u201cCensus Bureau Region and \nDivision Codes and Federal Information Processing \nSystem (FIPS) Codes for States.\u201d https://www2.cen \nsus.gov/programs-surveys/popest/geographies/2016/\nstate-geocodes-v2016.xls.\nU.S. Department of Health and Human Services. 2022. \n\u201cPreventive Care.\u201d Healthy People 2030. https://\nhealth.gov/healthypeople/objectives-and-data/\nbrowse-objectives/preventive-care.\nVaidya, Varun, Gautam Partha, and Monita Karmakar. \n2012. \u201cGender Differences in Utilization of \nPreventive Care Services in the United States.\u201d \nJournal of Women\u2019s Health 21(2):140\u201345.\nWatson, Oliver J., Gregory Barnsley, Jaspreet Toor, \nAlexandra B. Hogan, Peter Winskill, and Azra C. \nGhani. 2022. \u201cGlobal Impact of the First Year of \nCOVID-19 Vaccination: A Mathematical Modelling \nStudy.\u201d The Lancet Infectious Diseases 22(9):1293\u2013\n302. doi:10.1016/S1473-3099(22)00320-6.\nWest, Candace, and Don H. Zimmerman. 1987. \u201cDoing \nGender.\u201d Gender & Society 1(2):125\u201351. doi:10.117\n7/0891243287001002002.\nWilcox Vanden Berg, Rand N., Spyridon P. Basourakos, \nJonathan Shoag, Douglas Scherr, and Bashir Al \nHussein Al Awamlh. 2022. \u201cProstate Cancer \nScreening for Gay Men in the United States.\u201d Urology \n163:119\u201325. doi:10.1016/j.urology.2021.07.027.\nWilkinson, Richard G. 2005. The Impact of Inequality: \nHow to Make Sick Societies Healthier. London: \nRoutledge.\n\nDore et al.\t\n19\nYoung, Frank W. 2001. \u201cStructural Pluralism and Life \nExpectancy in Less-Developed Countries: The Role \nof Women\u2019s Status.\u201d Social Indicators Research \n55(2):223\u201340. doi:10.1023/A:1010982822560.\nAuthor Biographies\nEmily C. Dore is a PhD student in the Department of \nSociology at Emory University. She previously earned her \nmaster\u2019s of public health and master\u2019s of social work \ndegrees at Boston University. She studies the impact of \nstate contexts and social policies on health disparities \nthroughout the life course. Ms. Dore\u2019s recent work has \nbeen published in Health Affairs, BMC Research Notes, \nand Psychology of Violence. Her research is currently sup-\nported by the National Institute of Minority Health and \nHealth Disparities.\nSurbhi Shrivastava is a PhD student in the Department of \nSociology at Emory University. She has a master\u2019s of pub-\nlic health from Tata Institute of Social Sciences and a \nbachelor\u2019s of dental surgery from Manipal University, \nIndia. Her current research focuses on obstetric violence, \nmaternal health, and social inequality in India and Brazil. \nHer previous work has been published in Journal of \nBiosocial Science, Indian Journal of Medical Ethics, and \nBMJ Tobacco Control.\nPatricia (Trish) Homan is an associate professor of soci-\nology and the associate director of the Public Health \nProgram at Florida State University. She is also an associ-\nate of Florida State University\u2019s Pepper Institute on Aging \nand Public Policy and the Center for Demography and \nPopulation Health. Her research focuses on developing \ntheory and measurement for structural sexism, structural \nracism, and other forms of structural oppression and exam-\nining how these forces shape health. Her work has been \npublished in American Sociological Review, Demography, \nAmerican Journal of Public Health, The Milbank \nQuarterly, Health Affairs, Social Forces, Social Science & \nMedicine, Journal of Health and Social Behavior, The \nGerontologist, and The Journals of Gerontology: Series B, \namong other outlets. Her research has won multiple \nnational awards, including the 2022 NIH Matilda White \nRiley Early Stage Investigator Award, the 2022 Early \nCareer Gender Scholar Award from SWS South, the 2021 \nASA Sex & Gender Section Distinguished Article Award, \nand the 2019 Roberta G. Simmons Outstanding Dissertation \nAward from the ASA Medical Sociology Section. Dr. \nHoman earned her PhD in sociology from Duke University \nbefore joining Florida State University in 2018.\n\n\n    Scientific Research Findings:", "answer": "\u2022 We find that overall, both men and women use less preventive health care in states with higher levels of structural sexism.\n\u2022 In general, these findings support the gender performance theory for men, because men are less likely to seek care in a context encouraging traditional masculine traits, and the gendered power and resource allocation theory for both men and women, because states in which women are less empowered are less likely to have health-promoting policies for all.\n\u2022 One finding does not follow this pattern: Men are more likely to test for prostate cancer in states with more structural sexism.", "id": 66}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465221109202\nJournal of Health and Social Behavior\n2023, Vol. 64(1) 2\u00ad\u201320\n\u00a9 American Sociological Association 2022\nDOI: 10.1177/00221465221109202\njournals.sagepub.com/home/hsb\nOriginal Article\nThe United States continues to have unenviable \npopulation health. U.S. life expectancy stagnated \nfor five years (2010 to 2014) and subsequently \ndeclined for another three (2015 to 2017)\u2014an \nunheralded decline among high-income nations \n(Crimmins and Zhang 2019; Woolf and Schoomaker \n2019). Underneath these troubling trends are sub-\nstantial and increasing disparities across U.S. states \n(Montez et\u00a0 al. 2020; Wilmoth, Boe, and Barbieri \n2011; Woolf and Schoomaker 2019). For instance, \nin 2018, the difference in life expectancy between \nWest Virginia (74.7) and Hawaii (81.9) was a \nmarked 7.2 years\u2014up from 4.3 in 2000. To under-\nstand why the United States is fairing so poorly, \nscholars are increasingly pointing toward macro-\nlevel structural explanations (e.g., Beckfield and \nBambra 2016; Gutin and Hummer 2021; Torche and \nRauf 2021). Income inequality and public policy \nrepresent two such structural factors because both \nmirror national trends in life expectancy, are \nstrongly tied to population health, and exhibit wide \ndifferences across states (e.g., Kaufman et\u00a0al. 2020; \nMontez et\u00a0 al. 2020; Pickett and Wilkinson 2015; \nVenkataramani, O\u2019Brien, and Tsai 2021).\nIncome inequality and public policy are two \nmajor axes of the United States\u2019s sociopolitical \nlandscape that have experienced fundamental shifts \nin recent years. Income inequality in the United \nStates has been rising for over half a century \n1109202 HSBXXX10.1177/00221465221109202Journal of Health and Social BehaviorMcFarland et al.\nresearch-article2022\n1Florida State University, Tallahassee, FL, USA\n2University of Texas at San Antonio, San Antonio, TX, \nUSA\n3Syracuse University, Syracuse, NY, USA\nCorresponding Author:\nMichael J. McFarland, PhD, Department of Sociology \nand Center for Demography and Population Health, \nFlorida State University, 113 Collegiate Loop, P.O. Box \n3062270, Tallahassee, FL 32306-2270, USA. \nEmail: mmcfarland@fsu.edu\nIncome Inequality and \nPopulation Health: Examining \nthe Role of Social Policy\nMichael J. McFarland1\n, Terrence D. Hill2\n,  \nand Jennifer Karas Montez3\nAbstract\nStudies of the relationship between income inequality and life expectancy often speculate about the role \nof policy, but direct empirical research is limited. Drawing on the neo-materialist perspective, we examine \nwhether the longitudinal association between income inequality and life expectancy is mediated and \nmoderated by policy liberalism in U.S. states (2000\u20132014). More liberal policy contexts are characterized \nby greater efforts to regulate the economy, redistribute income, and protect vulnerable groups and lesser \nefforts to penalize deviant social behavior. We find that state-level income inequality is inversely associated \nwith policy liberalism and life expectancy. The association between income inequality and life expectancy \nwas not mediated by policy liberalism but was moderated by it. The association is attenuated in states \nwith more liberal policy contexts, supporting the neo-materialist perspective. This finding illustrates how \nstates like New York and California (with liberal policy contexts) can exhibit high income inequality and \nhigh life expectancy.\nKeywords\nincome inequality, life expectancy, public policy\n\nMcFarland et al.\t\n3\n(McCall and Percheski 2010; Schaeffer 2020) and \nnow exceeds that in peer countries (Organisation for \nEconomic Cooperation and Development 2021). It \nhas grown to such an extent that the top 20% of U.S. \nhouseholds acquired more income in 2018 than did \nthe bottom 80% (Schaeffer 2020). At the same time, \nthe United States has undergone a rapid period of \nhyper-political divergence in the types of policies \nenacted across U.S. states. Policy environments \namong U.S. states are increasingly moving further \nright or left on the political spectrum (Montez et\u00a0al. \n2020). Although the U.S. government distributes \npolicy-making authority across levels of govern-\nment, states increasingly define the policy context in \nwhich Americans live (Grumbach 2018).\nResearchers, however, know little about the \nimplications of their intersection vis-\u00e0-vis popula-\ntion health. This article draws from the extensive \nliterature on income inequality and population \nhealth. Two issues from this literature are particu-\nlarly relevant. First, why is income inequality asso-\nciated with decreased life expectancy? Second, \nunder which conditions is the association more or \nless pronounced? Some work has attempted to \nanswer the first question by pointing to potential \npathways related to human capital investments, \nsocial welfare programs, public services, health \ncare infrastructure, social capital, and health behav-\nior (e.g., Clarkwest 2008; Dunn, Burgess, and Ross \n2005; Elgar and Aitken 2011; Elgar, Stefaniak, and \nWohl 2020). Other studies have considered the sec-\nond question by demonstrating that the association \nbetween income inequality and life expectancy or \nmortality appears stronger in places that have less \neconomic development, greater financial insecurity, \nand fewer social welfare programs (Curran and \nMahutga 2018; Rambotti 2015; Ross et\u00a0 al. 2000, \n2005; Wilkinson and Pickett 2008).\nThe present study builds on this robust literature \nby simultaneously considering mortality conse-\nquences of state policy environments (i.e., principles, \ndecisions, and methods of action proposed and \nselected by governmental bodies). Although \u201cpol-\nicy\u201d is regularly mentioned as an implication of \nresearch linking income inequality and population \nhealth (e.g., implicating policies on taxation and \nredistribution schemes), studies rarely theorize about \nthe functions of policies or use direct measures of \nthem. When policy is measured, assessments are \ngenerally restricted to a single policy or a narrow \nrange of policies. Consequently, a more comprehen-\nsive measure of the policy environment might help \nelucidate the association between income inequality \nand mortality. We build on previous research by \nexamining whether the longitudinal association \nbetween income inequality and life expectancy is \nmediated and moderated by a comprehensive assess-\nment of policy environments in U.S. states from \n2000 through 2014.\nIncome inequality and \npopulation health\nPlaces with more inequitable income distributions \nhave lower life expectancies (e.g., Curran and \nMahutga 2018; Elgar et\u00a0al. 2020; Hill et\u00a0al. 2019; \nHill and Jorgenson 2018; Jorgenson et\u00a0 al. 2020, \n2021; Pickett and Wilkinson 2015; Thombs et\u00a0al. \n2020; Wilkinson and Pickett 2009; Zheng 2012). \nThese associations persist across indicators of \nincome inequality (e.g., Gini coefficient, income \nshare measures, and the Theil entropy index) and \nlongevity (e.g., life expectancy, infant mortality).\nIn response to early and credible concerns over \ncross-sectional designs, modeling decisions, and \nstatistical rigor (Beckfield 2004; Gravelle 1998; \nMellor and Milyo 2001, 2003), more recent national \nand international studies have employed longitudi-\nnal designs and a range of advanced statistical tech-\nniques (Curran and Mahutga 2018; Hill et\u00a0al. 2019; \nHill and Jorgenson 2018; Jorgenson et\u00a0 al. 2020, \n2021; Liao and De Maio 2021; Thombs et\u00a0al. 2020; \nZheng 2012). Evaluating the consistency in the lit-\nerature, Kondo and colleagues (2009:7) concluded \nthat their \u201cmeta-analysis of cohort studies including \naround 60 million participants found that people \nliving in regions with high income inequality have \nan excess risk for premature mortality.\u201d Less is \nknown, however, about why income inequality is \nlinked to worse population health.\nThere are three general theoretical perspectives \nto explain why income inequality might undermine \npopulation health: the psychosocial perspective, the \nsocial capital perspective, and the neo-materialist \nperspective. Although these theoretical perspectives \nhave been contrasted in the literature (Clarkwest \n2008; Elgar and Aitken 2011; Lynch et\u00a0 al. 2000; \nMcLeod et\u00a0al. 2003), they \u201care not mutually exclu-\nsive\u201d and are \u201clikely to reinforce each other\u201d \n(Lochner et\u00a0al. 2001:390). The distinction between \nthese perspectives lies primarily in the types of causal \nmechanisms purported to link income inequality to \npopulation health.\nThe psychosocial perspective points to deleterious \neffects of relative deprivation. Such deprivation may \ncontribute to widespread negative self-\u00adappraisals \n(e.g., a sense of worthlessness from social compari-\nsons), alienation (e.g., a sense of powerlessness \n\n4\t\nJournal of Health and Social Behavior 64(1) \nfrom the subjective experience of stratification), \nemotional distress (e.g., anxiety from a sense of \ninsecurity and anger from a sense of injustice), risky \ncoping behaviors (e.g., heavy alcohol consumption \nand smoking through self-medication), and, over \ntime, the breakdown of the physiological systems of \nthe human body (Hill et\u00a0 al. 2019; Wilkinson and \nPickett 2008). Lynch and colleagues (2000:1201) \nexplain \u201cthat income inequality affects health \nthrough perceptions of place in the social hierarchy \nbased on relative position according to income\u201d and \nthat \u201csuch perceptions produce negative emotions \nsuch as shame and distrust that are translated \u2018inside\u2019 \nthe body into poorer health via psycho-neuro-endo-\ncrine mechanisms and stress induced behaviors such \nas smoking.\u201d\nThe social capital perspective asserts that income \ninequality may undermine interpersonal trust, social \ncohesion, cooperation, and collective political \nefforts to support vulnerable populations of society \nthrough processes related to widespread status com-\npetition and class resentment (Elgar et\u00a0al. 2020; Hill \net\u00a0al. 2019). In support of this perspective, income \ninequality has been associated with lower levels of \ngroup membership, social trust, and civic engage-\nment in the United States and other countries (Elgar \net\u00a0al. 2020; Wilkinson and Pickett 2009).\nThe neo-materialist perspective posits that income \ninequality weakens broader commitments to the \ngeneral interests of society by concentrating wealth \nand power among elites through political capture \nand the subsequent demoralization of voters from \nlower social classes. This, in turn, undermines resis-\ntance to neoliberal policies like tax cuts, deregula-\ntion, and divestment in community resources and \nsocial services that promote public health and safety, \nsuch as education and environmental regulations \n(Clarkwest 2008; Coburn 2004; Curran and \nMahutga 2018; Elgar and Aitken 2011; Hill et\u00a0al. \n2019; Jorgenson et\u00a0al. 2020; Jorgenson, Schor, and \nHuang 2017; Lynch et\u00a0al. 2000; McLeod et\u00a0al. 2003; \nNeumayer and Pl\u00fcmper 2016). According to this \nperspective, political units that have a high amount \nof income inequality are less likely to invest in their \npopulations in terms of human, physical, cultural, \ncivic, and health resources.\nIntersection of income \ninequality and policy \ncontexts\nWe view the links between income inequality and \npopulation health as working through or being \nmodulated by the public policy environment or \nboth. In discussing these issues, we conceptualize \nthe policy environment as occurring on a liberal to \nconservative political continuum, consistent with \nrecent formulations in political science (Grumbach \n2018). Liberal-leaning policies tend to expand eco-\nnomic regulations and protect marginalized groups, \nwhile conservative-leaning policies tend to endorse \ntraditional values and disapprove of state interven-\ntion in markets (Grumbach 2018).\nPolicy Contexts as a Mediator of \nIncome Inequality and Health\nThe idea that income inequality might undermine \npopulation health through policy mechanisms com-\nports with the neo-materialist and social capital per-\nspectives (Clarkwest 2008; Curran and Mahutga \n2008; Elgar and Aitken 2011; Lynch et\u00a0 al. 2000; \nMcLeod et\u00a0 al. 2003). Lynch and colleagues \n(2000:1202) explain that income inequality and \nrelated political processes \u201cinfluence the private \nresources available to individuals and shape the \nnature of public infrastructure\u2014education, health ser-\nvices, transportation, environmental controls, avail-\nability of food, quality of housing, occupational \nhealth regulations\u2014that form the \u2018neo-material\u2019 \nmatrix of contemporary life.\u201d Although income \ninequality has been shown to undermine public \nspending on education, library books, income redis-\ntribution, income assistance, food assistance (food \nstamps), health insurance, and medical care \n(Clarkwest 2008; Coburn 2004; Curran and \nMahutga 2018; Elgar and Aitken 2011; Neumayer \nand Pl\u00fcmper 2016; Ross et\u00a0al. 2000, 2005), there is \nlittle empirical evidence to suggest that policies \nrelated to medical care and public social expendi-\ntures might mediate the association between income \ninequality and life expectancy (Clarkwest 2008; \nElgar and Aitken 2011; Le Grand 1987; Wilkinson \nand Pickett 2009).\nAlthough there is little evidence of an indirect \neffect of income inequality on life expectancy \nthrough policy, this mediation model is theoretically \nviable because policy is known to shape the social, \neconomic, and behavioral determinants of popula-\ntion health (Bambra, Smith, and Pearce 2019; \nDawes 2020; Montez et\u00a0al. 2017; Woolhandler et\u00a0al. \n2021). For example, U.S. state policy regulates \naccess to human capital by funding schools, income by \nsetting a minimum wage, prenatal care via Medicaid, \nchildhood nutrition through the Supplemental \nNutrition Assistance Program, tobacco by way of \nexcise taxes, and much more. For these reasons, \n\nMcFarland et al.\t\n5\nBambra and colleagues (2019) aptly frame policies \nand the political choices behind them as the \u201ccauses \nof the causes of the causes\u201d of geographic inequali-\nties in population health.\nWhile many studies have isolated the health \nbenefits of specific policies, such as minimum \nwage (e.g., Kaufman et\u00a0al. 2020), earned income \ntax credits (e.g., Strully, Rehkopf, and Xuan 2010), \nand Medicaid (e.g., Bhatt and Beck-Sagu\u00e9 2018), \nothers have focused on the health consequences of \noverarching policy contexts (Beckfield and Bambra \n2016; Bradley et\u00a0al. 2016; Montez et\u00a0al. 2020). For \ninstance, lower life expectancies in the United \nStates relative to other high-income countries are \nlargely explained by the miserly social policies of \nthe United States (Beckfield and Bambra 2016). \nFurthermore, U.S. states with a more liberal policy \ncontext have exhibited longer life expectancies \nover a 45-year period than states with more conser-\nvative policy contexts (Montez et\u00a0al. 2020).\nLiberal policy contexts are often linked to good \npopulation health because they shape the material \nand social determinants of health in salubrious ways \n(Brennenstuhl, Quesnel-Vall\u00e9e, and McDonough \n2012; McCartney et\u00a0 al. 2019). They are generally \naimed at solving enduring social problems that com-\npromise the health and well-being of populations, \nespecially vulnerable populations. For instance, \nenvironmental policies protect populations by limit-\ning exposure to toxins. Health and welfare policies \nsupport the sick and poor through financial assis-\ntance and the provision of needed services. Civil \nrights policies empower socially marginalized \ngroups by legislating against widespread institutional \ndiscrimination. Educational spending invests in the \nhuman capital of workers and the future of state \neconomies. In addition, a body of research using \ncross-national data shows that social democratic wel-\nfare state regimes offer more social protection, \nincluding health care access, and tend to have better \npopulation health outcomes than their less liberal \ncounterparts (e.g., Beckfield and Bambra 2016; \nBrennenstuhl et\u00a0al. 2012). The next section discusses \npotential synergies between state policy contexts and \nincome inequality in relation to population health.\nPolicy Contexts as a Moderator of \nIncome Inequality and Health\nAlthough previous studies of income inequality and \npopulation health have emphasized the potential \nmediating influence of policy, the neo-materialist \nperspective may be extended by also framing policy \nas an effect modifier of inequality (Curran and \nMahutga 2018; McLeod et\u00a0al. 2003). In addition to \nasking how policy might help to mediate or explain \nthe association between income inequality and life \nexpectancy, we might also ask whether the associa-\ntion between income inequality and life expectancy \nvaries across policy environments. This idea is gen-\nerally supported by the fact that countries with more \ngenerous policies (e.g., pensions, unemployment \ninsurance, health care, and social services) tend to \nlimit the degree to which populations are dominated \nby the market forces of neoliberalism and broader \nsystems of social stratification (Bambra 2007).\nFollowing these lines of research, we expect that \nthe association between state-level income inequal-\nity and life expectancy will be moderated or attenu-\nated in U.S. states with more liberal policy \nenvironments. The idea is that state policy contexts \nthat invest in the general welfare may help offset the \nprocesses through which income inequality erodes \npopulation health. While income inequality tends to \nundermine social equality, social capital, and mate-\nrial resources, liberal policy environments tend to \nsupport the equitable distribution of goods and ser-\nvices (Arts and Gelissen 2002; Bambra 2007).\nOur expectation derives from studies finding that \nthe magnitude of the association between income \ninequality and population health differs across pop-\nulation subgroups and country contexts. In general, \nthe effect of income inequality on health appears to \nbe strongest for disadvantaged individuals and \ncountries with weak economic safety nets. For \ninstance, the association between U.S. state-level \nincome inequality and individual-level mortality is \nmore pronounced among near-poor whites than their \nnondisadvantaged counterparts (Lochner et\u00a0 al. \n2001). Studies also demonstrate that the association \nbetween country-level income inequality and coun-\ntry-level life expectancy is stronger at higher levels \nof relative poverty (e.g., Rambotti 2015).\nThe second indicator is that income inequality is \noften unrelated to life expectancy and other indica-\ntors of population health in countries with more \ncomprehensive social programs (Coburn 2004; \nLaporte and Ferguson 2003; McLeod et\u00a0al. 2003; \nRoss et\u00a0 al. 2000, 2005). Ross and colleagues \n(2000:898) speculated that \u201cthe lack of a significant \nassociation between income inequality and mortal-\nity in Canada may indicate that the effects of \nincome inequality on health are not automatic and \nmay be blunted by the different ways in which \nsocial and economic resources are distributed in \nCanada and in the United States.\u201d Coburn (2004:49) \nechoed this sentiment, arguing that \u201csocial support \nprograms, from Medicare to education and social \n\n6\t\nJournal of Health and Social Behavior 64(1) \ninfrastructure\u201d are likely to \u201creduce the effect of \nincome inequality on health in Canada more so than \nin the U.S.\u201d Ross and colleagues (2005:108) later \nbroadened the comparisons by noting that \u201cthe \nidentification of a relationship between inequality \nand mortality within the United States and Great \nBritain, and the absence for Australia, Canada, and \nSweden, provides some evidence for the idea that \nthere is no necessary association between income \ninequality and population health.\u201d\nThe third indicator is that the association \nbetween income inequality and life expectancy can \nbe weaker in countries with greater economic devel-\nopment. A study by Curran and Mahutga (2018) \nfound that the effect of income inequality was more \npronounced in the poorest countries. They explained \nthat \u201cpoorer countries with high inequality may \ninvest less in public goods; however, unlike high-\nincome countries with high inequality, they lack the \neconomic and administrative resources to maintain \ninfrastructure for all\u201d (p. 539).\n2000 to 2014.\u2002 The 2000 to 2014 time period is of \nparticular interest because it followed major pieces of \nnational legislation that made states more important \nin shaping income inequality and increased levels of \ninequality nationally. The passage of welfare reform \nin 1996, for instance, resulted in state policies having \nmore power to influence income inequality. Indeed, \nliberal-leaning states saw slower rates of growth in \nincome inequality after this time than their conserva-\ntive-leaning counterparts (Kelly and Witko 2012). \nOur time period also comes on the heels of major \npieces of tax legislation (e.g., the 1997 Taxpayer \nRelief Act, the 2001 and 2003 Bush Tax Cuts) that \ndisproportionately benefited high-income taxpayers \nand expanded income inequality (Center on Budget \nand Policy Priorities 2017). Furthermore, state-level \npolicy contexts polarized considerably during this \ntime, especially after 2000 (Grumbach 2018:Figure \n1; Montez et\u00a0al. 2020:Figure 3). By 2014, the left\u2013\nright divergence in state policy contexts was at its \nhighest point since 1970.\nHypotheses\nBased on previous theory and research, we formu-\nlated three hypotheses to guide subsequent analyses.\nHypothesis 1: Income inequality will be \ninversely associated with life expectancy.\nHypothesis 2: The inverse association between \nincome inequality and life expectancy will be \nmediated by policy liberalism.\nHypothesis 3: The inverse association between \nincome inequality and life expectancy will be \nattenuated in states with higher levels of policy \nliberalism.\nData and methods\nThe data for this study included state-year observa-\ntions for 2000 to 2014 for all 50 states. Our analytic \nsample included 750 observations (15 years \u00d7 50 \nstates). This study combined state-level data from a \nvariety of sources described below.\nMeasures\nAnnual data on life expectancy by state were \nobtained from the United States Mortality Database \n(https://usa.mortality.org/). Life expectancy is an \nideal indicator of overall population health because \nit reflects age-specific mortality rates spanning all \nages for a particular year.\nOur key indicator of income inequality was the \nincome share of the top 10% of earners based off \npretax gross income reported to the Internal \nRevenue Service. These data were obtained from \nthe U.S. State-Level Income Inequality Database, \ndeveloped by Mark Frank (2015) and constructed \nfrom individual tax filing data available from the \nInternal Revenue Service. From the same source, \nwe also obtained alternative measure of income \ninequality to use in supplemental analyses. These \nmeasures included the share of the top .1%, 1%, \nand 5% of income earners and the Gini index.\nThe main source of data on state policies came \nfrom Grumbach (2018). These measures, shown in \nTable 1, contained 135 policies spanning 16 domains: \nabortion, campaign finance, civil rights and liberties, \ncriminal justice, education, environment, gun control, \nhealth and welfare, housing and transportation, immi-\ngration, private sector labor, public sector labor, \nLGBT rights, marijuana, taxes, and voting. Evidence \non how these policy domains are plausibly linked to \nlife expectancy are described elsewhere (Kemp, \nGrumbach, and Montez 2022). For each state, the \ndata contained a score for each domain annually from \n2000 through 2014.\nThe scoring mechanisms for each policy domain \nwere established by Grumbach (2018). First, 135 \nindividual policies were classified as liberal or con-\nservative. Liberal was defined as the use of state \npower to regulate the economy, redistribute income \nand wealth, protect vulnerable groups, or limit the \ngovernment\u2019s ability to penalize irregular social \nbehavior. Conservative was defined as the reverse. \n\nMcFarland et al.\t\n7\nFor example, gay marriage bans were classified as a \nconservative policy. Second, individual policies \nwere scored on a 0 to 1 scale during the 2000 to \n2014 period, with values ranging from conservative \n(low) to liberal (high). For instance, a state\u2019s score \nfor medical marijuana policy is 0 in years when it \nwas illegal and 1 in years when it was legal. \nNondichotomous policies (e.g., K\u201312 spending) \nwere normalized across all state-year observa-\ntions\u2014giving a 0 to 1 range. Next, annual scores \nTable 1.\u2002 U.S. State Policy Domains and Their Constituent Policies.\nDomain\nPolicies\nAbortion\nAbortion insurance restriction, abortion legal, consent post-Casey, consent pre-Casey, \nemergency contraception, gestation limit, Medicaid covers abortion, parental notice, \npartial birth abortion ban, physician required, waiting required\nCampaign \nfinance\nCorporate contribution ban, limit on individual contributions (yes, no), dollar limit on \nindividual contributions ($), limit on political action committee (PAC) contributions \n(yes, no), dollar limit on PAC contributions ($), public funding elections\nCivil rights and \nliberties\nBible allowed in public schools, corporal punishment ban, discrimination ban public \naccommodations, Equal Rights Amendment (ERA) ratification, fair employment \ncommission, gender discrimination ban, gender equal pay law, moment of silence \nin public school, no-fault divorce, physician-assisted suicide, public breastfeeding, \nReligious Freedom Rights Amendment, reporters right to source confidentiality, \nstate Americans with Disabilities Act (ADA), state ERA\nCriminal justice\nDeath penalty repeal, determinate sentencing, DNA motions, three strikes,  \ntruth-in-sentencing\nEducation\nCharter school law; Higher education spending; K-12 spending; School choice\nEnvironment\nBottle bill, California car emissions standards, endangered species, e-waste, \nGreenhouse Gas (GHG) cap, renewables fund, solar tax credit, state National \nEnvironmental Policy Act (NEPA)\nGun control\nAssault weapon ban, background checks (dealers), background checks (private), Brady \nlaw, dealer licenses required, gun registration, open carry, Saturday night special ban, \nstand your ground\nHealth and \nwelfare\nAffordable Care Act (ACA) exchange, Aid to Families with Dependent Children \n(AFDC) payment level, AFDC Up, Children\u2019s Health Insurance Program (CHIP) \neligibility (children), CHIP eligibility (infants), CHIP eligibility (pregnant women), \nexpanded dependent coverage, Medicaid adoption, Medicaid expansion, Pre-Balanced \nBudget Act CHIP eligibility, senior prescription drugs, Temporary Assistance for \nNeedy Families (TANF) eligibility, TANF payment level, welfare drug test, welfare \ntime limit\nHousing and \ntransportation\nGrowth management, Lemon law, rent control ban, tort limit\nImmigration\nDrivers\u2019 licenses for undocumented, English official language, e-verify, e-verify ban, in-\nstate tuition for undocumented, state cash benefits for recent immigrants, state food \nbenefits for recent immigrants, state health benefits for recent immigrants\nLabor (private \nsector)\nDisability insurance; local minimum wage ban, local sick leave law ban, minimum wage, \npaid family leave, paid sick leave, prevailing wage, right to work, unemployment \ncompensation\nLabor (public \nsector)\nBan on agency fees (state), collective bargaining (firefighters), collective bargaining \n(local), collective bargaining (police), collective bargaining (state), collective \nbargaining (teachers)\nLGBT rights\nCivil unions and marriage, gay marriage ban, hate crime law, LGB discrimination ban \npublic accommodations, LGB employment discrimination ban, sodomy ban\nMarijuana\nMarijuana decriminalization, medical marijuana\nTaxes\nCorporate tax rate, earned income tax credit, estate tax, income tax, sales tax, tax \nburden, top capital gains rate, top income rate\nVoting\nAbsentee voting, early voting, motor voter, voter identification\n\n8\t\nJournal of Health and Social Behavior 64(1) \nfor each policy domain were calculated as the sum \nof the liberal policy scores minus the sum of the \nconservative policy ones. The summed scores for \neach domain were then standardized across states \nand years to give a 0 to 1 scale.\nThis study included various state-level time-\nvarying control variables, including percentage \nBlack, percentage Hispanic, median household \nincome (in thousands of dollars), percentage of the \npopulation between 18 and 64 years old, population \nsize, percentage of population with a college \ndegree, and percentage of the population that was \nforeign born. Controls were chosen based on prior \ninequality and health research (e.g., Hill et\u00a0al. 2019; \nHill and Jorgenson 2018; Rambotti 2015). Median \nhousehold income, in particular, was chosen to iso-\nlate the association in the income inequality and life \nexpectancy from overall levels of income in the \npopulation. Similarly, we needed to assure this \nassociation did not exist because more educated \npeople live in places with lower levels of inequality. \nControl variables were obtained from the following \nsources: the U.S. Bureau of Labor Statistics Local \nArea Statistics Project, U.S. Census Bureau Small \nArea Income and Poverty Estimates, U.S. Census \nBureau Population and Housing Estimates, the \n2000 Decennial Census, and the 1% American \nCommunity Survey PUMS files.\nAnalytical Strategy\nOur analytical strategy utilized regression models \nwith fixed effects for states and years. Fixed effects \ncontrol for omitted variables that are time-invariant \nby examining variability within states rather than \nbetween states. We used the xtreg command in Stata \nalong with the fe option to estimate regression mod-\nels with fixed effects for years and states, with robust \nstandard errors clustered by state. The time fixed \neffects were accounted for by the inclusion of year-\nspecific intercepts. Linear models with fixed effects \nassume linear additive effects. All models accounted \nfor serial correlation using the vce (robust) command \nin Stata. In the tables and text, we provide coeffi-\ncients, their uncertainty (standard errors or confi-\ndence intervals), and p value thresholds up to .10 to \nimprove transparency and reproducibility of our \nresults (Wasserstein and Lazar 2016).\nThe influence of income inequality on policy \nand life expectancy likely has a meaningful tempo-\nral lag (Clarkwest 2008; Mellor and Milyo 2003; \nZheng 2012). While the fixed effects approach has \nseveral advantages, particularly the ability to rule \nout time-invariant sources of potential spuriousness, \nits interpretation vis-\u00e0-vis temporal ordering is \nunclear. Two-way fixed effects coefficients can be \ninterpreted as the average difference in intraunit \nchanges in the dependent variable at time point t for \neach one-unit intraunit increase in the explanatory \nvariable at time point t, averaged across time points. \nBecause the model addresses change averaged \nacross time points, they tell us little about the tem-\nporal lag needed for an independent variable to \ninfluence an outcome. To partially address this \nissue, we show both conventional fixed effects \nresults and results after incorporating a five-year lag \nvariable for income inequality and policy \u00adliberalism. \nWe chose a five-year lag because the only study \n(that we are aware of) to systematically examine \nhow the association between inequality and mortal-\nity changes by the lag time found that the associa-\ntion becomes pronounced at five years (Zheng \n2012). To model life expectancy from 2000 to 2014 \nwith five-year lags on income inequality and policy \nliberalism, we also incorporated inequality and pol-\nicy indicators from 1995 to 1999.\nOur specific modeling strategy was as follows. \nFirst, we examined how changes in income inequal-\nity predict changes in policy liberalism. Next, we \nexamined the relationships among income inequal-\nity, policy liberalism, and life expectancy with six \nmodels. Using these models, we were able to assess \nthe extent to which policy liberalism mediates or \nmoderates the association between income inequal-\nity and life expectancy.\nUsing several models, we tested whether the \nassociation between income inequality and life \nexpectancy is mediated by policy liberalism. \nPrevious studies using fixed effects (Anthony, \nDiPerna, and Amato 2014; Huang, Oshima, and \nKim 2010) used the Sobel (1982) test of indirect \neffects. Studies show that the Monte Carlo method \nfor assessing mediation (MCMAM) has greater sta-\ntistical power than the standard Sobel test (Bauer, \nPreacher, and Gil 2006; MacKinnon, Lockwood, \nand Williams 2004). Unlike the Sobel test, which \nassumes that the sampling distribution of the indi-\nrect effect is normally distributed, the MCMAM \nuses model estimates and asymptotic variances and \ncovariances to simulate the sampling distribution \nfor the indirect effect. For this reason, we employed \nthe MCMAM to formally test the indirect effect of \nincome inequality on life expectancy through pol-\nicy liberalism.\nThe formula for the indirect effect is \nE a b\nab\nj\nj\na b\nj\nj\n(\n) =\n+\u03c3\n (Bauer et\u00a0al. 2006). The aver-\nage indirect effect is a function of the average effect \nof X on M (a), the average effect of M on Y (b), and \n\nMcFarland et al.\t\n9\nthe covariance between the two effects (\u03c3a b\nj\nj). To \nsimulate the sampling distribution of the average \nindirect effect, a multinormal distribution is defined \nwith means equal to a b and\na b\nj\nj\n\u001f\n\u001e\n\u001e\n, ,\n\u03c3\n and covariance \nmatrix equal to the estimated covariance matrix of \nthese estimates. Random values from this multinor-\nmal distribution are entered into the aforementioned \nformula to estimate the average indirect effect. \nResampling provides a simulated sampling distri-\nbution for the average indirect effect.\nThe MCMAM uses parameter estimates and \ntheir asymptotic variances to simulate random \ndraws from the joint distribution of a (the regression \nof policy liberalism on income inequality) and b (the \nregression of life expectancy on policy liberalism) \nand to compute the product of these values. This \nprocedure was repeated 20,000 times, and the result-\ning distribution of a \u00d7 b was used to estimate a 95% \nconfidence interval (CI) around the observed value \nof a \u00d7 b (Selig and Preacher 2008). The procedure \nrequired (1) the coefficient for the regression of pol-\nicy liberalism on income inequality, or a (Table 3, \nModel 2); (2) the coefficient for the regression of \nlife expectancy on policy liberalism, or b (Table 4, \nModel 5); (3) the asymptotic sampling variance of \na; (4) the asymptotic sampling variance of b; and (5) \nthe covariance of a and b. The MCMAM output \nprovided a 95% CI for the indirect effect.\nLastly, we estimated several models to test for \nmoderation by including an interaction term between \nincome inequality and policy liberalism. This model \nis of the following form:\ny\nx\nz\nx z\nit\nit\nit\nit\ni\nt\nit\n*\n*\n*\n* *\n,\n=\n+\n+\n+\n+\n+\n\u03b2\n\u03b2\n\u03b2\n1\n2\n3\nw\n\u03b5\n\b\n(1)\nwhere subscript i represents each unit of analysis \n(i.e., state-year), subscript t represents year, y is life \nexpectancy for each state at each year, x represents \nthe income share of the top 10% income earners, z \nrepresents policy liberalism, \u221di is the state-specific \ndisturbance term, wt is the year-specific disturbance \nterm that is constant across all states, and \u03b5 is the \ndisturbance term unique to each state at each year. \nThe \u201c*\u201d represent deviations from state-specific \nmeans where \ny\ny\ny\nit\nit\ni\n* =\n\u2212\n, x\nx\nx\nit\nit\ni\n* =\n\u2212\n, and \nz\nz\nz\nit\nit\ni\n* =\n\u2212\n. \u03b23 is of particular interest because it \ntells whether the association between changes in \nincome inequality and life expectancy varies by \npolicy liberalism.\nTo better understand how the association \nbetween income inequality and life expectancy var-\nies by policy liberalism, we plotted the results from \nEquation 1 and calculated marginal effects of \nincome inequality at various levels of policy liber-\nalism. Marginal effects are calculated by taking the \npartial derivative of Equation 1. Equation 2 shows \nthat the income inequality\u2013life expectancy associa-\ntion will vary by policy liberalism when \u03b23  \u2260 0.\n\t\n\u2202\n\u2202\n=\n+\ny\nx\nzit\n*\n*\n*\n\u03b2\n\u03b2\n1\n3\n\t\n(2)\nMarginal effects allowed us to quantify the asso-\nciation between income inequality and life expec-\ntancy at various levels of policy liberalism. Marginal \neffects were calculated using the margins dydx com-\nmand in Stata.\nResults\nTable 2 presents the descriptive statistics for our \nstudy variables, including key variables at various \npoints in time. It shows notable variation in our key \nstudy variables. For instance, in the most equitable \nstate (Alaska), the top 10% of wage earners con-\ntrolled 33% of income, but in the least equitable \nstate (New York), they controlled 60% in 2014. The \nreported means also increased over time for all three \nof our key variables.\nTable 3, Model 2 shows that increases in income \ninequality from five years prior predict decreases in \npolicy liberalism, net of state fixed effects and the \ntime-varying covariates. While the coefficient for \nincome inequality is negative in both models, it is \nnot statistically different from zero in Model 1, \nwhich underscores the importance of considering \npast income inequality when predicting subsequent \nstate policy liberalism.\nTable 4 presents the results examining life expec-\ntancy and reveals three important points. First, \nModel 1 shows that increases in income inequality \nare associated with decreases in life expectancy (b = \n\u2212.031; p < .001) when nonlagged, while Model 4 \nshows this association for lagged income inequality \nis not statistically significant (b = \u2013.007; p = .347). \nThe former association corresponded with a standard \ndeviation increase in income inequality being associ-\nated with a .15-year decrease in life expectancy. This \nestimate was found by standardizing our income \ninequality measure and rerunning Model 1 (not \nshown). Second, both the nonlagged policy liberal-\nism score (Model 2, b = 2.164, p = .07) and the \nlagged score (Model 5, b = 2.143, p < .05) are associ-\nated with life expectancy, net of income inequality. \nThird, the association between income inequality \nand life expectancy did not show any meaningful \n\u00b5\n\n10\t\nJournal of Health and Social Behavior 64(1) \nchange with the inclusion of policy liberalism in \neither the lagged or nonlagged models. This latter \npoint hints that the association between income \ninequality and life expectancy is not mediated by \npolicy liberalism. We next performed a formal medi-\nation analysis to assess the veracity of this claim.\nThe mediation analysis used information pro-\nvided in Tables 3 and 4. The direct effect of income \ninequality on life expectancy using the lagged \nregression models is \u2013.007 (Table 4, Model 4), while \nthe indirect effect is the product of \u2013.002 (Table 3, \nModel 2) and 2.143 (Table 4, Model 5), or \u2013.004. \nThe total effect is the sum of the direct effect (\u2013.007) \nand indirect effects (\u2013.004), or \u2013.011. Figure 1 pres-\nents the uncertainty of these mediation estimates \nusing the MCMAM procedure. Given that the 95% \nCI (\u2013.007, .002) contains zero, our analysis supports \nthe null hypothesis of no mediation and suggests \nthat the indirect effect of income inequality on life \nexpectancy through policy liberalism is not likely \ndifferent from zero, 90% CI (\u2013.006, .001). We also \ntested our proposed mediation model using the tra-\nditional Sobel test, and the results of this test (z = \n\u20131.50, SE = .003, p = .13) are also consistent with \nthe null hypothesis of no indirect effect. Evidence \nfor the null hypothesis is also found when estimates \nfrom the nonlagged models are used (not shown).\nNext, we examine whether policy liberalism \nmoderates the association between income inequal-\nity and life expectancy. Model 6 in Table 4 shows \nthat the association between lagged income inequal-\nity and life expectancy varies by policy liberalism \nsuch that the association becomes weaker with \nincreases in policy liberalism (b = .142, p < .01). \nTable 2.\u2002 Descriptive Statistics (N = 750).\nMinimum\nMaximum\nMean\nStandard Deviation\nLife expectancy\u2013overall\n73.39\n81.70\n77.92\n1.70\nLife expectancy\u20132000\n73.39\n79.51\n76.87\n1.45\nLife expectancy\u20132007\n74.38\n80.60\n78.01\n1.59\nLife expectancy\u20132014\n74.93\n81.70\n78.73\n1.70\nIncome share of top 10%\u2013overall\n32.97\n62.17\n44.13\n5.12\nIncome share of top 10%\u20132000\n33.27\n58.93\n43.58\n5.82\nIncome share of top 10%\u20132007\n36.08\n62.17\n46.02\n5.32\nIncome share of top 10%\u20132014\n32.97\n60.42\n46.10\n5.31\nPolicy liberalism\u2013overall\n.18\n.77\n.45\n.15\nPolicy liberalism\u20132000\n.19\n.63\n.38\n.13\nPolicy liberalism\u20132007\n.20\n.71\n.42\n.15\nPolicy liberalism\u20132014\n.15\n.74\n.43\n.18\nPercentage Black\n.31\n37.51\n10.51\n9.53\nPercentage Hispanic\n.68\n47.67\n9.77\n9.64\nPercentage foreign born\n27.00\n1.00\n8.21\n5.98\nMedian household income (in thousands  \nof dollars)\n30.19\n73.85\n47.87\n8.63\nPercentage population age 18\u201364 years\n58.83\n66.04\n62.50\n1.30\nTotal population (in thousands)\n493.79\n38,600.00\n6,000.17\n6,612.77\nPercentage population college degree\n10.71\n30.56\n18.81\n3.99\nNote: All variables are from state-level data from 2000 to 2014.\nFigure 1.\u2002 Monte Carlo Distribution of Indirect \nEffects of Income Inequality on Life Expectancy.\n\nMcFarland et al.\t\n11\nThis pattern looks similar for the nonlagged model \n(b = .096, p = .056). Furthermore, when alternative \nmeasures of inequality are utilized, including the \nincome share of the top .1%, 1%, and 5%, the inter-\naction term between income inequality and policy \nliberalism remains significant at p < .05. To better \nunderstand the finding from Model 6 in Table 4, we \nexamine these relationships graphically, in Figure 2, \nand provide marginal effects.\nFigure 2 shows how the association between \nincome inequality and life expectancy varies by pol-\nicy liberalism based on the lagged model (Model 6 of \nTable 4). Income inequality is more strongly \ninversely associated with life expectancy among \nstates with relatively low levels of policy liberalism. \nAmong states with average to high levels of policy \nliberalism, the association between income inequal-\nity and life expectancy is reduced to a statistical zero \nthroughout most of this half of the policy liberalism \ndistribution. Also intriguing, the combination of \nhigh-income inequality and low policy liberalism \nresults in particularly low levels of life expectancy. \nLife expectancies are all similar, and relatively high, \nwhen inequality is low, as seen in the left side of \nFigure 2. At higher levels of income inequality, this \npattern diverges, however. States with either low-\nincome inequality or high policy liberalism have rel-\natively high life expectancy, but states with both did \nnot have higher life expectancy than those with only \none. States with neither low inequality nor high pol-\nicy liberalism had particularly low life expectancy.\nTable 5 shows more precisely how the associa-\ntion between income inequality and life expectancy \nvaries by policy liberalism. Income inequality is \nmost strongly related to life expectancy in states \nwith the lowest levels of policy liberalism. For \ninstance, in the lagged model, the coefficient for \nincome inequality when policy liberalism is at  \nthe bottom 10th percentile is \u2013.033 but is .005 at  \nthe 75th percentile\u2014more than a 100% increase. \nIndeed, income inequality is not statistically associ-\nated with life expectancy by the 50th percentile  \nof policy liberalism. Overall, Table 5 presents evi-\ndence that any harmful influence of income \nTable 3.\u2002 Two-Way Fixed Effects Coefficients for the Regression of Policy Liberalism on Income \nInequality (N = 750).\nNonlagged\nLagged 5 Yearsa\n\u2002\nModel 1\nModel 2\nIncome inequality\n\u2003 Income share top 10%\n\u2212.001\n(.001)\n\u2212.002**\n(.001)\nTime-varying controls\n\u2003 Percentage Black\n.005\n(.007)\n.005\n(.007)\n\u2003 Percentage Hispanic (natural log)\n\u2212.143**\n(.045)\n\u2212.150**\n(.044)\n\u2003 Median household income\n.001\n(.001)\n.002\n(.001)\n\u2003 Percentage working age\n.006\n(.007)\n.006\n(.007)\n\u2003 Total population (square root)\n.000***\n(.000)\n.000***\n(.000)\n\u2003 Percentage college degree\n.004**\n(.001)\n.004**\n(.001)\n\u2003 Percentage foreign born\n.009*\n(.003)\n.008*\n(.003)\nConstant\n.438\n(.519)\n.536\n(.510)\nR2\n.452\n.462\nNote: Robust standard errors in parentheses.\naThese models also utilize measures of income inequality and policy liberalism from 1995 to 1999.\n*p < .05, **p < .01, ***p < .001.\n\n12\t\nJournal of Health and Social Behavior 64(1) \ninequality is smaller or nonexistent in states with \nmore liberal policy environments. In the following, \nwe probe this finding by disaggregating our mea-\nsure of policy liberalism into specific policy \ndomains because it is plausible that a limited num-\nber of policy domains are driving these patterns \nrather than the broader policy environment.\nTable 6 shows the marginal associations between \nincome inequality and life expectancy for different \nlevels of each policy domain used to create the over-\nall policy liberalism score using lagged models. Each \npolicy domain was included in the equivalent of \nModel 6 in Table 4. At least two important points can \nbe gleaned from Table 6. First, it provides evidence \nthat our decision to measure policy liberalism as one \nvariable was a judicious one. Among the 16 policy \ndomains, 13 are consistent with the prior finding that \n76.5\n77\n77.5\n78\n78.5\n79\nLife Expectancy\n35\n40\n45\n50\n55\n60\nIncome Inequality (Income Share of Top 10% Lagged 5 Years)\n1st percentile\n10th percentile\n25th percentile\n50th percentile\n75th percentile\n95th percentile\nPolicy Liberalism (Lagged 5 Years)\nFigure 2.\u2002 Life Expectancy by Income Inequality \nand Policy Liberalism Based on Model 6 from \nTable 4.\nTable 4.\u2002 Fixed Effects Coefficients Showing Life Exepctancy Regressed on Income Inequality and Policy \nLiberalism (N = 750).\nNonlagged\nLagged by 5 Years\n\u2002\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nModel 6\nIndependent variables\n\u2003 Income share top 10%\n\u2212.031***\n(.008)\n\u2212.030***\n(.008)\n\u2212.069**\n(.024)\n\u2212.007\n(.008)\n\u2212.005\n(.008)\n\u2212.064**\n(.024)\n\u2003 Policy liberalism\n2.164\u2020\n(1.171)\n\u22122.341\n(3.138)\n2.143*\n(.982)\n\u22124.354\n(2.873)\n\u2003 Income share \u00d7 Policy liberalism\n.096\u2020\n(.049)\n.142**\n(.052)\nTime-varying control variables\n\u2003 Percentage Black\n.020\n(.096)\n.009\n(.090)\n.001\n(.092)\n.039\n(.099)\n.027\n(.098)\n.020\n(.098)\n\u2003 Percentage Hispanic (natural log)\n\u2212.827\u2020\n(.453)\n\u2212.518\n(.443)\n\u2212.338\n(.435)\n\u2212.921\u2020\n(.481)\n\u2212.663\n(.465)\n\u2212.402\n(.446)\n\u2003 Median household income\n.003\n(.013)\n\u2212.001\n(.012)\n.001\n(.011)\n.011\n(.013)\n.008\n(.013)\n.008\n(.012)\n\u2003 Percentage working age\n.139*\n(.064)\n.125*\n(.059)\n.102\n(.057)\n.121\u2020\n(.069)\n.112\u2020\n(.063)\n.091\n(.060)\n\u2003 Total population (square root)\n.002**\n(.001)\n.003***\n(.001)\n.003***\n(.001)\n.002*\n(.001)\n.002**\n(.001)\n.002***\n(.001)\n\u2003 Percentage college degree\n.037**\n(.013)\n.029*\n(.012)\n.027*\n(.012)\n.040**\n(.013)\n.033*\n(.012)\n.034**\n(.012)\n\u2003 Percentage foreign born\n.087*\n(.038)\n.069*\n(.034)\n.060\u2020\n(.034)\n.091*\n(.039)\n.078*\n(.038)\n.069\u2020\n(.038)\nConstant\n65.039***\n(5.067)\n64.09***\n(4.467)\n66.945***\n(4.536)\n65.195***\n(5.414)\n64.425***\n(4.860)\n67.661***\n(4.760)\nR2\n.903\n.907\n.909\n.899\n.902\n.905\nNote: Robust standard errors in parentheses. Year fixed effects not shown.\naThese models also utilize measures of income inequality and policy liberalism from 1995 to 1999.\n\u2020p < .10, *p < .05, **p < .01, ***p < .001.\n\nMcFarland et al.\t\n13\nthe association between income inequality and life \nexpectancy is weaker with increases in policy liberal-\nism, although only 5 interactions are statistically sig-\nnificant from zero at conventional levels. The \nnegative association of income inequality is largest in \nmagnitude within states having the lowest levels of \npolicy liberalism, and it decreases as the policy liber-\nalism percentile increases. Civil rights, taxation, and \neducation were the only policy domains to not meet \nthis general pattern. Second, certain policies were \nmore important than others in understanding the \nintersection between income inequality and life \nexpectancy. In particular, the interactions between \nincome inequality and several policy domains, \nincluding marijuana legalization, environmental pro-\ntection, gun control, LGBT rights, and health and \nwelfare, were particularly pronounced. For instance, \nthe association between income inequality and life \nexpectancy was \u2013.025 for states at the 10th percentile \nfor environmental protection but .001 for those at the \n75th percentile\u2014more than a 100% change. In short, \nthe evidence in Table 6 supports our argument that \noverall state-level policy liberalism is important for \nunderstanding the association between income \ninequality and life expectancy while also pointing to \nthe differential importance of particular policy \ndomains in these relationships. The marginal effects \nwhen income inequality and policy domains were \nnonlagged five years were consistent with the overall \npatterns reported here and shown in Table A1 in the \nAppendix.\nRobustness and Sensitivity Analyses\nWe performed several ancillary analyses that \nstrengthened evidence for this finding. First, to \nensure that states with extremely low or high \nincome inequality were not driving our key results, \nwe ran several additional models that excluded \neach state from the analysis. Table A2 in the \nAppendix shows tests for moderation based on \nModel 6 from Table 4 after excluding the eight \nstates with the highest and lowest levels of income \ninequality in 2014. These states include New York, \nFlorida, Connecticut, Massachusetts, Alaska, \nHawaii, Iowa, and Nebraska. It shows that the same \npattern of findings reported previously. We also ran \nthese models after excluding the other 42 states (not \nshown). These results were also consistent with our \nreported findings. Second, we tested whether our \nresults were driven by our measure of income \ninequality by reestimating our analyses using the \nshare of the top .1%, 1%, and 5% of income earn-\ners. We found that choice of these income inequal-\nity measures did not alter our substantive findings. \nThese results are available on request.\nWe also included the Gini coefficient as an \nadditional measure of income inequality. However, \nour findings for this measure differed substantively \nfrom our more reliable findings for the other mea-\nsures of income inequality. Income inequality, as \nmeasured by the Gini coefficient, shared a stronger \npositive association with life expectancy in states \nwith higher levels of policy liberalism. This pattern \nis not unprecedented in the literature. For example, \nNeumayer and Pl\u00fcmper (2016) found that the Gini \ncoefficient was positively associated with life \nexpectancy inequality among developed countries. \nHill and Jorgenson (2018) showed that the Gini \nindex shares a different relationship with life \nexpectancy than other inequality indicators. These \nauthors explained that most income inequality \nindicators capture the extent to which there is a \nconcentration of income among the wealthy (i.e., \nTable 5.\u2002 Partial Slope Coefficients for the Association between Income Inequality and Average Life \nExpectancy as a Function of Policy Liberalism (N = 750).\nPercentiles for Policy \nLiberalism\nIncome Share and Policy Liberalism \nNonlaggeda\nIncome Share and Policy Liberalism \nLagged 5 Yearsb\n1\n\u2212.049 (.015)**\n\u2212.037 (.015)*\n10\n\u2212.047 (.014)**\n\u2212.033 (.014)*\n25\n\u2212.042 (.012)***\n\u2212.024 (.011)*\n50\n\u2212.032 (.008)***\n\u2212.008 (.008)\u2002\n75\n\u2212.022 (.007)**\n.005 (.008)\u2002\n95\n\u2212.009 (.011)\n.024 (.012)\u2020\nNote: Huber-White robust standard errors in parentheses.\naPartial slope coefficients were calcuated based on Model 3 in Table 4.\nbPartial slope coefficients were calcuated based on Model 6 in Table 4.\n\u2020p < .10, *p < .05, **p < .01, ***p < .001.\n\n14\t\nJournal of Health and Social Behavior 64(1) \nthe upper end of the income distribution) relative \nto the rest of the population, but the Gini coeffi-\ncient, by contrast, deals with how much the actual \nincome distribution deviates from the Lorenz curve \nand does not directly capture the specific location \nin the distribution where inequality is occurring. \nFor this reason, the Gini coefficient is more sensi-\ntive to changes in the middle of the distribution \nthan to changes at the highs and lows of the distri-\nbution (Burns 2015). Because our framework \nimplicates the higher end of the income distribu-\ntion as having a disproportionate influence on pop-\nulation health, the use of the Gini index is not \nappropriate. For these reasons, we highlight the \ntheoretical and empirical consistency of our results \nacross several income share measures.\nThird, we replicated our analysis using sex-\nspecific life expectancy measures. This ancillary \nanalysis was conducted because (1) others have \nfound gender differences in the relationship \nbetween income inequality and state-level mortal-\nity (e.g., Hill and Jorgenson 2018) and (2) gender-\nstratified structural exposures are known to shape \npopulation health (Homan 2019). Our finding that \nthe association between income inequality and \nlife expectancy varies by level of policy \u00adliberalism \ndid not differ by sex (these results are available \non request).\nFinally, it is plausible that the key associations \nreported previously were due to changes in other \ntime-varying characteristics, particularly economic \ngrowth, poverty rates, and mass liberal ideology. \nWe included a measure of gross domestic product \nper capita in all of our models. The inclusion of this \nvariable did not have any meaningful impact on \nregression estimates and did not change our sub-\nstantive results. Similarly, we included an indicator \nmeasuring the poverty rate after excluding our mea-\nsure of median income, and the inclusion of this \nvariable did not change any of our substantive \nresults. Furthermore, we controlled for the interac-\ntion between income inequality and the poverty \nrate. This did not affect our substantive results \neither. Finally, we utilized a measure of mass liberal \nideology developed by Tausanovitch and Warshaw \n(2013) derived on the policy preferences of 275,000 \nAmericans. The inclusion of this measure did not \nchange our substantive results, which is consistent \nwith the idea that our key results are driven by \nsocial policy rather than the underlying political \nand cultural beliefs of the broader population. All of \nthese results are available on request.\nTable 6.\u2002 Marginal Effects Showing the Association between Income Inequality and Life Expectancy as a \nFunction of Specific Policy Domains (N = 750).\nPolicy Domains\nInteraction \nTerm \nCoefficient\nPolicy \nLiberalism \nat 1st \nPercentile\nPolicy \nLiberalism \nat 10th \nPercentile\nPolicy \nLiberalism \nat 25th \nPercentile\nPolicy \nLiberalism \nat 50th \nPercentile\nPolicy \nLiberalism \nat 75th \nPercentile\nPolicy \nLiberalism \nat 95th \nPercentile\n1\nAbortion\n.047\u2020\n\u2212.030\u2020\n\u2212.023\n\u2212.018\n\u2212.006\n.006\n.011\n2\nCampaign finance\n.027\n\u2212.019\n\u2212.013\n\u2212.012\n\u2212.006\n\u2212.004\n.005\n3\nCivil rights\n\u2212.005\n\u2212.004\n\u2212.006\n\u2212.007\n\u2212.007\n\u2212.008\n\u2212.009\n4\nCriminal justice\n.034\n\u2212.022\u2020\n\u2212.015\n\u2212.015\n\u2212.008\n\u2212.002\n.005\n5\nMarijuana\n.025*\n\u2212.013\n\u2212.013\n\u2212.013\n\u2212.013\n.000\n.022\n6\nEnvironment\n.069**\n\u2212.029**\n\u2212.025**\n\u2212.020*\n\u2212.012\n.001\n.026*\n7\nGuns\n.054*\n\u2212.025**\n\u2212.020*\n\u2212.020*\n\u2212.014\u2020\n\u2212.004\n.023\n8\nHousing and \ntransportation\n.003\n\u2212.007\n\u2212.007\n\u2212.007\n\u2212.007\n\u2212.006\n.001\n9\nImmigration\n.025\n\u2212.015\n\u2212.013\n\u2212.013\n\u2212.010\n\u2212.006\n.001\n10\nLGBT\n.086***\n\u2212.039***\n\u2212.039***\n\u2212.025**\n\u2212.011\n.004\n.025*\n11\nVoting\n\u2212.016\n.000\n\u2212.002\n\u2212.004\n\u2212.005\n\u2212.009\n\u2212.013\n12\nTaxes\n.029\n\u2212.019*\n\u2212.017**\n\u2212.008\n\u2212.004\n.001\n.007\n13\nEducation\n\u2212.045\n.010\n.003\n\u2212.002\n\u2212.006\n\u2212.011\n\u2212.017\n14\nLabor\n.050\n\u2212.018\n\u2212.017\n\u2212.015\n\u2212.002\n.001\n.008\n15\nPublic labor\n.014\n\u2212.017\n\u2212.017\n\u2212.014\n\u2212.014\n\u2212.004\n\u2212.002\n16\nHealth and welfare\n.117**\n\u2212.053***\n\u2212.039**\n\u2212.025**\n\u2212.015\u2020\n\u2212.003\n.013\nNote: Estimates from 16 fixed effects models with individual policy domains included in the interaction between lagged \npolicy liberalism and lagged income inequality. Models included all study covariates.\n\u2020p < .10, *p < .05, **p < .01, ***p < .001.\n\nMcFarland et al.\t\n15\nDiscussion\nAlthough numerous studies have shown that income \ninequality is associated with lower life expectancy, \nlittle is known about the role of policy environ-\nments. In an effort to build on previous work, we \ntested whether the longitudinal association between \nincome inequality and life expectancy was mediated \nand moderated by a comprehensive assessment of \nstate-level policy environments in the United States.\nOur first hypothesis proposed that income \ninequality would be inversely associated with life \nexpectancy. This hypothesis received support in our \nnonlagged models. These findings corroborate the \nconclusions of other recent studies of income \ninequality and life expectancy (Curran and Mahutga \n2018; Hill et\u00a0 al. 2019; Hill and Jorgenson 2018; \nJorgenson et\u00a0al. 2020, 2021).\nOur second hypothesis posited that the inverse \nassociation between income inequality and life \nexpectancy would be partially mediated by policy \nliberalism. We found that income inequality was \ninversely associated with policy liberalism. We also \nobserved that policy liberalism was positively asso-\nciated with life expectancy. Because the product of \ntwo these paths (i.e., the indirect effect) was small \nand not statistically significant from zero, our Monte \nCarlo analysis supported a null interpretation of no \nmediation. There are several reasons why this find-\ning is important. It represents the first longitudinal \nmediation test of the neo-materialist perspective and \ndoes not support the widely cited idea that income \ninequality undermines population health through \npolicy mechanisms (Clarkwest 2008; Curran and \nMahutga 2008; Elgar and Aitken 2011; Lynch et\u00a0al. \n2000; McLeod et\u00a0al. 2003). With this in mind, future \nmediation studies should focus more on mecha-\nnisms that flow from the psychosocial and social \ncapital perspectives on income inequality (Kawachi \net\u00a0al. 1997; Wilkinson and Pickett 2009).\nOur final hypothesis stated that the inverse asso-\nciation between income inequality and life expec-\ntancy would be moderated or attenuated in states \nwith higher levels of policy liberalism. Our modera-\ntion analyses support this hypothesis, showing that \nincome inequality exhibited a weaker or nonexistent \nrelationship with life expectancy among states with \naverage to high policy liberalism. This finding helps \nus to understand how states like Massachusetts, \nNew York, and California can rank high on both \nincome inequality and life expectancy. When con-\nsidered along with our null mediation model, these \nresults suggest that the neo-materialist perspective \nmight benefit from focusing more on framing policy \nenvironments as a modifier of income inequality \nthan as a mechanism of income inequality (Curran \nand Mahutga 2018; McLeod et\u00a0al. 2003).\nOur finding that policy liberalism moderates the \nassociation between income inequality and life \nexpectancy may help reconcile cross-national \nincome inequality research. While the bulk of work \nexamining income inequality and life expectancy \n(or mortality) across countries finds a negative asso-\nciation (see Pickett and Wilkinson 2015), some do \nnot. In particular, the association sometimes disap-\npears after accounting for time-invariant sources of \nspuriousness (e.g., Bl\u00e1zquez-Fern\u00e1ndez, Cantarero-\nPrieto, and Pascual-Saez 2018; Hu, van Lenthe, and \nMackenbach 2015; Leigh and Jencks 2007). \nBecause cross-national research frequently differs in \nstudy design characteristics that are strongly linked \nto policy environments, such as countries and years \nof investigation, the inclusion of country-level pol-\nicy environment indicators as moderators may rec-\noncile these disparate findings. Our results, if they \nare applicable across countries, suggest that associa-\ntion between income inequality and life expectancy \n(1) will be weaker in more policy liberal countries \nand (2) will weaken over time to the extent that \ncountries embrace a more liberal policy orientation. \nThese two phenomena may, therefore, be obfuscat-\ning the expected association. Future work should \naddress this issue.\nThe results of our final hypothesis also implore \nfuture work to continue probing who is most and \nleast likely to be affected by inequality. One impor-\ntant aspect of policy liberalism is concern for socially \ndisadvantaged or marginalized populations. To the \nextent that vulnerable members of society dispropor-\ntionately benefited from policy liberalism, we would \nexpect subpopulation differences in the link between \nincome inequality and life expectancy\u2014particularly \nthose along socioeconomic and racial lines.\nOur analyses should be considered in the con-\ntext of three limitations. Although models with \nfixed effects provide rigorous statistical tests, these \nmodels have several potential limitations (Hill et\u00a0al. \n2020; Kropko and Kubinec 2020). First and fore-\nmost, fixed effects approaches do not reflect causal \nassociations. While our modeling strategy allowed \nus to account for time-invariant and several time-\nvarying sources of heterogeneity, including mass \npolitical ideology in ancillary models, we cannot \nexclude the possibility of unobserved time-varying \nfactors. All said, our key results were robust to a \nvariety of alternate modeling choices, and the most \nlikely sources of spuriousness were included in our \nmodels.\n\n16\t\nJournal of Health and Social Behavior 64(1) \nWe also note that our null mediation test could \nreflect the conservative nature of models with fixed \neffects. Given that our mediation analysis tested \nwhether the product of two paths (the association \nbetween changes in income inequality and changes in \npolicy liberalism and the association between changes \nin policy liberalism and changes in life expectancy) \nwas different from zero, we may have lacked the sta-\ntistical power to find evidence for mediation. Although \nwe are encouraged by several statistically significant \ntests in our focal regression models, the power of our \nmediation analysis could still be limited by modest \nwithin-state changes in income inequality, policy lib-\neralism, and life expectancy.\nFinally, we note that changes in states\u2019 popula-\ntion composition during the study period could \nimpact our results. For example, education levels (a \nstrong predictor of mortality) may have risen more \nin some states than others, some states may have \nexperienced in-migration of healthier individuals, \nand other states may have experienced out-migra-\ntion of healthy individuals. Although these pro-\ncesses should not be ignored when interpreting our \nfindings, prior studies indicate that they do not \nmaterially affect our findings (Couillard et\u00a0al. 2021; \nFenelon 2013; Montez et\u00a0al. 2019). A recent study \nexamined state-level trends in adult mortality dur-\ning the 1992 to 2016 period and concluded that the \ndiverging trends are primarily due to changes in \nstates\u2019 policy contexts, not to changes in the educa-\ntion or income levels of their populations (Couillard \net\u00a0 al. 2021). Another study examined state-level \ntrends in the association between education level \nand mortality during the 1985 to 2011 period and \nfound that changes in the education\u2013mortality \nassociation within states were not due to trends in \ninterstate migration or immigration (Montez et\u00a0al. \n2019).\nConclusion\nThe United States experienced momentous shifts \nin income inequality and state-level policy envi-\nronments in recent years. This study examined \nhow the confluence of these major dimensions of \nthe political economy shaped population health \nand thereby filled a lacuna in income inequality \nand health research. For nearly four decades, stud-\nies of income inequality and population health \nhave been speculating about the role of policy. \nAlthough the original neo-materialist perspective \nframed policy as a mediator, more recent scholar-\nship has framed policy as a moderator. Our analy-\nses clearly favor the more recent framing, showing \nthat more liberal state policies have the capacity to \ncompletely offset the effects of income inequality \non life expectancy.\nOne policy recommendation has been to reduce \nincome inequality. Our findings suggest that invest-\nments in liberal social policies that support the gen-\neral welfare of the population may also limit the \ndegree to which populations are dominated by the \nmarket forces of neoliberalism and broader systems \nof social stratification. For instance, states with \nboth high life expectancy and income inequality \n(e.g., Connecticut, New Jersey, New York) tend to \ninvest in health and welfare programs that pay pop-\nulation health dividends. Our work suggests other \nhigh-inequality states might benefit from similar \ninvestments.\n\nMcFarland et al.\t\n17\nAppendix\nTable A1.\u2002 Marginal Effects Showing the Association between Nonlagged Income Inequality and Life \nExpectancy as a Function of Specific Policy Domains (N = 750).\nPolicy Domains\nInteraction \nTerm \nCoefficient\nPolicy \nLiberalism \nat 1st \nPercentile\nPolicy \nLiberalism \nat 10th \nPercentile\nPolicy \nLiberalism \nat 25th \nPercentile\nPolicy \nLiberalism \nat 50th \nPercentile\nPolicy \nLiberalism \nat 75th \nPercentile\nPolicy \nLiberalism \nat 95th \nPercentile\n1\nAbortion\n.007\n\u2212.033*\n\u2212.032**\n\u2212.031**\n\u2212.030***\n\u2212.028**\n\u2212.026*\n2\nCampaign finance\n.014\n\u2212.037**\n\u2212.033***\n\u2212.033***\n\u2212.030***\n\u2212.029**\n\u2212.024*\n3\nCivil rights\n.006\n\u2212.034\n\u2212.033*\n\u2212.032**\n\u2212.031***\n\u2212.030***\n\u2212.029**\n4\nCriminal justice\n.017\n\u2212.038**\n\u2212.035***\n\u2212.035***\n\u2212.032***\n\u2212.028**\n\u2212.025*\n5\nMarijuana\n.025*\n\u2212.037***\n\u2212.037***\n\u2212.037***\n\u2212.037***\n\u2212.019\u2020\n\u2212.013\n6\nEnvironment\n.087***\n\u2212.061***\n\u2212.055***\n\u2212.045***\n\u2212.029***\n\u2212.018*\n.015\n7\nGuns\n.057**\n\u2212.044***\n\u2212.044***\n\u2212.038***\n\u2212.035***\n\u2212.024**\n.013\n8\nHousing and \ntransportation\n.007\n\u2212.032**\n\u2212.032**\n\u2212.032**\n\u2212.030***\n\u2212.028**\n\u2212.026*\n9\nImmigration\n.001\n\u2212.031**\n\u2212.031**\n\u2212.031**\n\u2212.031***\n\u2212.031**\n\u2212.031*\n10\nLGBT\n.070***\n\u2212.059***\n\u2212.047***\n\u2212.047***\n\u2212.036**\n\u2212.012\n\u2212.001\n11\nVoting\n\u2212.042\n\u2212.012\n\u2212.017\n\u2212.021*\n\u2212.026**\n\u2212.035***\n\u2212.045**\n12\nTaxes\n.012\n\u2212.033***\n\u2212.033***\n\u2212.029**\n\u2212.028**\n\u2212.025**\n\u2212.023*\n13\nEducation\n\u2212.050\n\u2212.010\n\u2212.012\n\u2212.025**\n\u2212.027**\n\u2212.040**\n\u2212.043**\n14\nLabor\n.035\n\u2212.044**\n\u2212.038**\n\u2212.037**\n\u2212.030***\n\u2212.025**\n\u2212.020\u2020\n15\nPublic labor\n.018\n\u2212.039**\n\u2212.039**\n\u2212.036**\n\u2212.024**\n\u2212.021**\n\u2212.021**\n16\nHealth and welfare\n.092*\n\u2212.068***\n\u2212.045***\n\u2212.040***\n\u2212.031***\n\u2212.022**\n\u2212.007\nNote: Estimates from 16 fixed effects models with individual policy domains included in the interaction between policy \nliberalism and income inequality. Models included all study covariates.\n\u2020p < .10, *p < .05, **p < .01, ***p < .001.\nTable A2.\u2002 Tests for Moderation After Excluding States with the Four Highest and Lowest Levels of \nIncome Inequality in 2014.\nIncome Inequality and Policy Liberalism are Lagged 5 Years\nNew York\nFlorida\nConnecticut Massachusetts\nAlaska\nHawaii\nIowa\nNebraska\nState Excluded\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nModel 6\nModel 7\nModel 8\nIncome share top \n10%\n\u2212.055*\n(.023)\n\u2212.069*\n(.025)\n\u2212.063*\n(.025)\n\u2212.063*\n(.024)\n\u2212.065*\n(.025)\n\u2212.065*\n(.023)\n\u2212.061*\n(.024)\n\u2212.064*\n(.024)\nPolicy liberalism\n\u22122.699\n(2.265)\n\u22124.626\n(2.944)\n\u22124.163\n(3.004)\n\u22124.174\n(2.969)\n\u22124.484\n(2.996)\n\u22124.753\u2020\n(2.763)\n\u22123.888\n(2.934)\n\u22124.196\n(2.925)\nIncome share \u00d7 \nPolicy liberalism\n.111*\n(.047)\n.150**\n(.054)\n.138*\n(.055)\n.140*\n(.054)\n.142*\n(.054)\n.146**\n(.051)\n.134*\n(.052)\n0.140*\n(.053)\nNote: Robust standard errors in parentheses.\nEstimates based on fixed effects analyses with all study covariates (not shown).\n\u2020p < .10, *p < .05, **p < .01.\nORCID iDs\nMichael J. McFarland \n https://orcid.org/0000-0001-8793- \n3169\nTerrence Hill \n https://orcid.org/0000-0003-3798-7753\nJennifer Karas Montez \n https://orcid.org/0000-0001- \n9472-615X\nReferences\nAnthony, Christopher J., James C. DiPerna, and Paul \nAmato. 2014. \u201cDivorce, Approaches to Learning, and \nChildren\u2019s Academic Achievement: A Longitudinal \nAnalysis of Mediated and Moderated Effects.\u201d \nJournal of School Psychology 52(3):249\u201361.\n\n18\t\nJournal of Health and Social Behavior 64(1) \nArts, Wil, and John Gelissen. 2002. \u201cThree Worlds \nof Welfare Capitalism or More? A State-of-the-\nArt Report.\u201d Journal of European Social Policy \n12(2):137\u201358.\nBambra, Clare. 2007. \u201cGoing beyond the Three Worlds \nof Welfare Capitalism: Regime Theory and Public \nHealth Research.\u201d Journal of Epidemiology & \nCommunity Health 61(12):1098\u2013102.\nBambra, Clare, Katherine E. Smith, and Jamie Pearce. \n2019. \u201cScaling Up: The Politics of Health and Place.\u201d \nSocial Science & Medicine 232:36\u201342.\nBauer, Daniel, Kristopher Preacher, and Karen Gil. 2006. \n\u201cConceptualizing and Testing Random Indirect \nEffects and Moderated Mediation in Multilevel \nModels: New Procedures and Recommendations.\u201d \nPsychological Methods 11(2):142\u201363.\nBeckfield, Jason. 2004. \u201cDoes Income Inequality Harm \nHealth? New Cross-National Evidence.\u201d Journal of \nHealth and Social Behavior 45(3):231\u201348.\nBeckfield, Jason, and Clare Bambra. 2016. \u201cShorter Lives \nin Stingier States: Social Policy Shortcomings Help \nExplain the U.S. Mortality Disadvantage.\u201d Social \nScience & Medicine 171:30\u201338.\nBhatt, Chintan B., and Consuelo M. Beck-Sagu\u00e9. 2018. \n\u201cMedicaid Expansion and Infant Mortality in the \nUnited States.\u201d American Journal of Public Health \n108(4):565\u201367.\nBl\u00e1zquez-Fern\u00e1ndez, Carla, David Cantarero-Prieto, and \nMarta Pascual-Saez. 2014. \u201cDoes Rising Income \nInequality Reduce Life Expectancy? New Evidence \nfor 26 European Countries (1995\u20132014).\u201d Global \nEconomic Review 47(4):464\u201379.\nBradley, Elizabeth H., Maureen Canavan, Erika Rogan, \nKristina Talbert-Slagle, Chima Ndumele, Lauren \nTaylor, and Leslie A. Curry. 2016. \u201cVariation in \nHealth Outcomes: The Role of Spending on Social \nServices, Public Health, and Health Care, 2000\u201309.\u201d \nHealth Affairs 35(5):760\u201368.\nBrennenstuhl, Sarah, Am\u00e9lie Quesnel-Vall\u00e9e, and Peggy \nMcDonough. 2012. \u201cWelfare Regimes, Population \nHealth and Health Inequalities: A Research \nSynthesis.\u201d Journal of Epidemiology & Community \nHealth 66(5):397\u2013409.\nBurns, J. K. 2015. \u201cPoverty, Inequality and a Political \nEconomy of Mental Health.\u201d Epidemiology and \nPsychiatric Sciences 24(2):107\u201313.\nCenter on Budget and Policy Priorities. 2017. \u201cThe Legacy \nof the 2001 and 2003 \u2018Bush\u2019 Tax Cuts.\u201d https://www.\ncbpp.org/research/federal-tax/the-legacy-of-the-2001-\nand-2003-bush-tax-cuts#_ftn13.%20https://www.\ncbpp.org/research/chart-book-the-bush-tax-cuts.\nClarkwest, Andrew. 2008. \u201cNeo-materialist Theory \nand the Temporal Relationship between Income \nInequality and Longevity Change.\u201d Social Science & \nMedicine 66(9):1871\u201381.\nCoburn, David. 2004. \u201cBeyond the Income Inequality \nHypothesis: Class, Neo-liberalism, and Health \nInequalities.\u201d Social Science & Medicine 58(1):41\u201356.\nCouillard, Benjamin K., Christopher L.L. Foote, Kavish \nGandhi, Ellen Meara, and Jonathan Skinner. 2021. \n\u201cRising Geographic Disparities in U.S. Mortality.\u201d \nJournal of Economic Perspectives 35(4):1\u201327.\nCrimmins, Eileen M., and Yuan S. Zhang. 2019. \u201cAging \nPopulations, Mortality, and Life Expectancy.\u201d \nAnnual Review of Sociology 45:69\u201389.\nCurran, Michaela, and Matthew C. Mahutga. 2018. \n\u201cIncome Inequality and Population Health: A Global \nGradient?\u201d Journal of Health and Social Behavior \n59(4):536\u201353.\nDawes, Daniel E. 2020. The Political Determinants of \nHealth. Baltimore, MD: Johns Hopkins University \nPress.\nDunn, James R., Bill Burgess, and Nancy A. Ross. 2005. \n\u201cIncome Distribution, Public Services Expenditures, \nand All-Cause Mortality in U.S. States.\u201d Journal of \nEpidemiology & Community Health 59(9):768\u201374.\nElgar, Frank, and Nicole Aitken. 2011. \u201cIncome \nInequality, Trust and Homicide in 33 Countries.\u201d \nEuropean Journal of Public Health 21(2):241\u201346.\nElgar, Frank, Anna Stefaniak, and Michael Wohl. 2020. \n\u201cThe Trouble with Trust: Time-Series Analysis of \nSocial Capital, Income Inequality, and COVID-19 \nDeaths in 84 Countries.\u201d Social Science & Medicine \n263:113365. doi:10.1016/j.socscimed.2020.113365.\nFenelon, Andrew. 2013. \u201cGeographic Divergence in \nMortality in the United States.\u201d Population and \nDevelopment Review 39(4):611\u201334.\nFrank, Mark W. 2015. \u201cThe U.S. Income Inequality Page \nof Mark W. Frank.\u201d http://www.shsu.edu/eco_mwf/\ninequality.html.\nGravelle, Hugh. 1998. \u201cHow Much of the Relation \nbetween \nPopulation \nMortality \nand \nUnequal \nDistribution of Income Is a Statistical Artefact?\u201d \nBritish Medical Journal 316(7128):382\u201385.\nGrumbach, Jacob M. 2018. \u201cFrom Backwaters to Major \nPolicymakers: Policy Polarization in the States, \n1970\u20132014.\u201d Perspectives on Politics 16(2):416\u201335.\nGutin, Iliya, and Robert A. Hummer. 2021. \u201cSocial \nInequality and the Future of U.S. Life Expectancy.\u201d \nAnnual Review of Sociology 47:13\u201343.\nHill, Terrence, Andrew Davis, J. Micah Roos, and \nMichael French. 2020. \u201cLimitations of Fixed-Effects \nModels for Panel Data.\u201d Sociological Perspectives \n63(3):357\u201369.\nHill, Terrence, and Andrew Jorgenson. 2018. \u201cBring \nout Your Dead!: A Study of Income Inequality and \nLife Expectancy in the United States, 2000\u20132010.\u201d \nHealth & Place 49:1\u20136.\nHill, Terrence, Andrew Jorgenson, Peter Ore, Kelly \nBalistreri, and Brett Clark. 2019. \u201cAir Quality and \nLife Expectancy in the United States: An Analysis \nof the Moderating Effect of Income Inequality.\u201d \nSSM - Population Health 7:100346. doi:10.1016/j.\nssmph.2018.100346.\nHoman, Patricia. 2019. \u201cStructural Sexism and Health \nin the United States: A New Perspective on Health \nInequality and the Gender System.\u201d American \nSociological Review 84(3):486\u2013516.\nHu, Yannan, Frank J. van Lenthe, and Johan P. Mackenbach. \n2015. \u201cIncome Inequality, Life Expectancy and \n\nMcFarland et al.\t\n19\nCause-Specific Mortality in 43 European Countries, \n1987\u20132008: A Fixed Effects Study.\u201d European \nJournal of Epidemiology 30(8):615\u201325.\nHuang, Jin, Karen Oshima, and Youngmi Kim. 2010. \n\u201cDoes Food Insecurity Affect Parental Characteristics \nand Child Behavior? Testing Mediation Effects.\u201d \nSocial Service Review 84(3):381\u2013401.\nJorgenson, Andrew, Terrence Hill, Brett Clark, Ryan \nThombs, Peter Ore, Kelly Balistreri, and Jennifer Givens. \n2020. \u201cPower, Proximity, and Physiology: Does Income \nInequality and Racial Composition Amplify the Impacts \nof Air Pollution on Life Expectancy in the United \nStates?\u201d Environmental Research Letters 15(2):024013. \nhttps://iopscience.iop.org/article/10.1088/1748-9326/\nab6789#:~:text=The%20results%20indicate%20\nthat%20income,inequality%20and%20larger%20\nblack%20populations.\nJorgenson, Andrew, Juliet Schor, and Xiaorui Huang. \n2017. \u201cIncome Inequality and Carbon Emissions in \nthe United States: A State-Level Analysis, 1997\u2013\n2012.\u201d Ecological Economics 134:40\u201348.\nJorgenson, Andrew, Ryan Thombs, Brett Clark, Jennifer \nGivens, Terrence Hill, Xiaorui Huang, Orla Kelly, \nand Jared Fitzgerald. 2021. \u201cInequality Amplifies the \nNegative Association between Life Expectancy and \nAir Pollution: A Cross-national Longitudinal Study.\u201d \nScience of the Total Environment 758:143705. \ndoi:10.1016/j.scitotenv.2020.143705.\nKaufman, John A., Leslie K. Salas-Hern\u00e1ndez, Kelli A. \nKomro, and Melvin D. Livingston. 2020. \u201cEffects of \nIncreased Minimum Wages by Unemployment Rate \non Suicide in the U.S.A.\u201d Journal of Epidemiology & \nCommunity Health 74(3):219\u201324.\nKawachi, Ichiro, Bruce P. Kennedy, Kimberly Lochner, \nand Deborah Prothrow-Stith. 1997. \u201cSocial Capital, \nIncome Inequality, and Mortality.\u201d American \nJournal of Public Health 87(9):1491\u201398.\nKelly, Nathan J., and Christopher Witko. 2012. \n\u201cFederalism and American Inequality.\u201d The Journal \nof Politics 74(2):414\u201326.\nKemp, Blakelee, Jacob M. Grumbach, and Jennifer Karas \nMontez. 2022. \u201cU.S. State Policy Contexts and \nPhysical Health among Midlife Adults.\u201d Socius 6. \ndoi:10.1177/23780231221091324.\nKondo, Naoki, Grace Sembajwe, Ichiro Kawachi, Rob \nVan Dam, S. V. Subramanian, and Zentaro Yamagata. \n2009. \u201cIncome Inequality, Mortality, and Self Rated \nHealth: Meta-analysis of Multilevel Studies.\u201d British \nMedical Journal 339(7731):1178\u201381.\nKropko, \nJonathan, \nand \nRobert \nKubinec. \n2020. \n\u201cInterpretation and Identification of Within-Unit and \nCross-sectional Variation in Panel Data Models.\u201d \nPLOS One 15(4):e0231349. doi:10.1371/journal.\npone.0231349.\nLaporte, Audrey, and Brian Ferguson. 2003. \u201cIncome \nInequality and Mortality: Time Series Evidence from \nCanada.\u201d Health Policy 66(1):107\u201317.\nLe Grand, Julian. 1987. \u201cInequalities in Health: Some \nInternational Comparisons.\u201d European Economic \nReview 31:182\u201391.\nLeigh, \nAndrew, \nand \nChristopher \nJencks. \n2007. \n\u201cInequality and Mortality: Long-Run Evidence from \na Panel of Countries.\u201d Journal of Health Economics \n26(1):1\u201324.\nLiao, Tim, and Fernando De Maio. 2021. \u201cAssociation \nof Social and Economic Inequality with Coronavirus \nDisease 2019 Incidence and Mortality across U.S. \nCounties.\u201d JAMA Network Open 4(1):e2034578. \ndoi:10.1001/jamanetworkopen.2020.34578.\nLochner, Kim, Elsie Pamuk, Diane Makuc, Bruce \nKennedy, and Ichiro Kawachi. 2001. \u201cState-Level \nIncome Inequality and Individual Mortality Risk: A \nProspective, Multilevel Study.\u201d American Journal of \nPublic Health 91(3):385\u201391.\nLynch, John, George Davey Smith, George Kaplan, \nand James House. 2000. \u201cIncome Inequality and \nMortality: Importance to Health of Individual Income, \nPsychosocial Environment, or Material Conditions.\u201d \nBritish Medical Journal 320(7243):1200\u20131204.\nMacKinnon, David, Chondra Lockwood, and Jason \nWilliams. 2004. \u201cConfidence Limits for the Indirect \nEffect: Distribution of the Product and Resampling \nMethods.\u201d Multivariate Behavioral Research 39(1): \n99\u2013128.\nMcCall, Leslie, and Christine Percheski. 2010. \u201cIncome \nInequality: New Trends and Research Directions.\u201d \nAnnual Review of Sociology 36:329\u201347.\nMcCartney, Gerry, Wendy Hearty, Julie Arnot, Frank \nPopham, Andrew Cumbers, and Robert McMaster. \n2019. \u201cImpact of Political Economy on Population \nHealth: A Systematic Review of Reviews.\u201d American \nJournal of Public Health 109(6):e1\u201312.\nMcLeod, Christopher, John Lavis, Cameron Mustard, and \nGreg Stoddart. 2003. \u201cIncome Inequality, Household \nIncome, and Health Status in Canada: A Prospective \nCohort Study.\u201d American Journal of Public Health \n93(8):1287\u201393.\nMellor, Jennifer, and Jeffrey Milyo. 2001. \u201cReexamining \nthe Evidence of an Ecological Association between \nIncome Inequality and Health.\u201d Journal of Health \nPolitics, Policy and Law 26(3):487\u2013522.\nMellor, Jennifer, and Jeffrey Milyo. 2003. \u201cIs Exposure to \nIncome Inequality a Public Health Concern? Lagged \nEffects of Income Inequality on Individual and Population \nHealth.\u201d Health Services Research 38(1):137\u201351.\nMontez, Jennifer Karas, Jason Beckfield, Julene Kemp \nCooney, Mark D. Hayward, Jacob M. Grumbach, \nHuseyin Zeyd Koytak, Steven H. Woolf, and Anna \nZajacova. 2020. \u201cU.S. State Policies, Politics, and Life \nExpectancy.\u201d The Milbank Quarterly 98(3):668\u201399.\nMontez, Jennifer Karas, Mark D. Hayward, and Douglas \nA. Wolf. 2017. \u201cDo U.S. States\u2019 Socioeconomic and \nPolicy Contexts Shape Adult Disability?\u201d Social \nScience & Medicine 178:115\u201326.\nMontez, Jennifer Karas, Anna Zajacova, Mark D. \nHayward, Steven H. Woolf, D. Chapman, and Jason \nBeckfield. 2019. \u201cEducational Disparities in Adult \nMortality across U.S. States: How Do They Differ, \nand Have They Changed Since the Mid-1980s?\u201d \nDemography 56(2):621\u201344.\n\n20\t\nJournal of Health and Social Behavior 64(1) \nNeumayer, Eric, and Thomas Pl\u00fcmper. 2016. \u201cInequalities \nof Income and Inequalities of Longevity: A Cross-\ncountry Study.\u201d American Journal of Public Health \n106(1):160\u201365.\nOrganisation for Economic Co-operation and Development. \n2021. \u201cIncome Inequality (Indicator).\u201d https://data.\noecd.org/inequality/income-inequality.htm.\nPickett, Kate, and Richard Wilkinson. 2015. \u201cIncome \nInequality and Health: A Causal Review.\u201d Social \nScience & Medicine 128:316\u201326.\nRambotti, Simone. 2015. \u201cRecalibrating the Spirit Level: \nAn Analysis of the Interaction of Income Inequality \nand Poverty and Its Effect on Health.\u201d Social Science \n& Medicine 139:123\u201331.\nRoss, Nancy, Danny Dorling, James Dunn, G\u00f6ran \nHenriksson, John Glover, John Lynch, and Gunilla \nRingb\u00e4ck Weitoft. 2005. \u201cMetropolitan Income \nInequality and Working-Age Mortality: A Cross-\nsectional Analysis Using Comparable Data from Five \nCountries.\u201d Journal of Urban Health 82(1):101\u201310.\nRoss, Nancy, Michael C. Wolfson, James Dunn, Jean-\nMarie Berthelot, George Kaplan, and John Lynch. \n2000. \u201cRelation between Income Inequality and \nMortality in Canada and in the United States: Cross-\nsectional Assessment Using Census Data and Vital \nStatistics.\u201d British Medical Journal 320:898\u2013902.\nSchaeffer, Katherine. 2020. \u201c6 Facts about Economic \nInequality in the U.S.\u201d https://www.pewresearch.\norg/fact-tank/2020/02/07/6-facts-about-economic-\ninequality-in-the-u-s/.\nSelig, James, and Kristopher Preacher. 2008. \u201cMonte \nCarlo Method for Assessing Mediation: An \nInteractive Tool for Creating Confidence Intervals \nfor Indirect Effects\u201d [Computer software]. Available \nfrom http://quantpsy.org/.\nSobel, Michael. 1982. \u201cAsymptotic Intervals for Indirect \nEffects in Structural Equations Models.\u201d Sociological \nMethodology 13:290\u2013312.\nStrully, Kate W., David H. Rehkopf, and Ziming Xuan. \n2010. \u201cThe Effects of Prenatal Poverty on Infant \nHealth: State Earned Income Tax Credits and Birth \nWeight.\u201d American Sociological Review 75(4):534\u201362.\nTausanovitch, Chris, and Christopher Warshaw. 2013. \n\u201cMeasuring Constituent Policy Preferences in \nCongress, State Legislatures, and Cities.\u201d The Journal \nof Politics 75(2):330\u201342.\nThombs, Ryan, Dennis Thombs, Andrew Jorgenson, \nand Taylor Harris Braswell. 2020. \u201cWhat Is Driving \nthe Drug Overdose Epidemic in the United States?\u201d \nJournal of Health and Social Behavior 61(3):275\u201389.\nTorche, Florencia, and Tamkinat Rauf. 2021. \u201cThe Political \nContext and Infant Health in the United States.\u201d \nAmerican Sociological Review 86(3):377\u2013405.\nVenkataramani, Atheendar S., Rourke O\u2019Brien, and \nAlexander C. Tsai. 2021. \u201cDeclining Life Expectancy \nin the United States: The Need for Social Policy as \nHealth Policy.\u201d JAMA 325(7):621\u201322.\nWasserstein, Ronald L., and Nicole A. Lazar. 2016. \u201cThe \nASA Statement on p-Values: Context, Process, and \nPurpose.\u201d The American Statistician 70(2):129\u201333.\nWilkinson, Richard, and Kate Pickett. 2008. \u201cIncome \nInequality \nand \nSocioeconomic \nGradients \nin \nMortality.\u201d American Journal of Public Health \n98(4):699\u2013704.\nWilkinson, Richard, and Kate Pickett. 2009. The Spirit \nLevel: Why Greater Equality Makes Societies \nStronger. New York, NY: Bloomsbury Press.\nWilmoth, John R., Carol Boe, and Magali Barbieri. \n2011. \u201cGeographic Differences in Life Expectancy \nat Age 50 in the United States Compared with \nOther High-Income Countries.\u201d Pp. 333\u201366 in \nInternational Differences in Mortality at Older Ages: \nDimensions and Sources, edited by E. M. Crimmins, \nS. H. Preston, and B. Cohen. Washington, DC: The \nNational Academies Press.\nWoolf, Steven H., and Heidi Schoomaker. 2019. \u201cLife \nExpectancy and Mortality Rates in the United States, \n1959\u20132017.\u201d JAMA 322(20):1996\u20132016.\nWoolhandler, Steffie, David U. Himmelstein, Sameer \nAhmed, Zinzi Bailey, Mary T. Bassett, Michael Bird, \nJacob Bor, et\u00a0al. 2021. \u201cPublic Policy and Health in \nthe Trump Era.\u201d The Lancet 397(10275):705\u201353.\nZheng, Hui. 2012. \u201cDo People Die from Income \nInequality of a Decade Ago?\u201d Social Science & \nMedicine 75(1):36\u201345.\nAuthor Biographies\nMichael McFarland is an associate professor in sociol-\nogy and a faculty associate of the Center for Demography \nand Population Health at Florida State University. His \nrecent research focuses on the adverse effects of policing \nfor health, early life lead exposure and mental health, and \nrecently, social policy and health. Recent publications can \nbe found in the Proceedings of the National Academy of \nSciences (PNAS), Social Science & Medicine, and Energy \nEconomics.\nTerrence Hill is a professor of sociology at the University \nof Texas at San Antonio. His research examines social \ninequalities in health and human suffering. To date, he \nhas published 138 peer-reviewed manuscripts. His publi-\ncations appear in a range of journals, such as the Journal \nof Health and Social Behavior, Social Science & \nMedicine, Society & Mental Health, The Journals of \nGerontology, The Gerontologist, American Journal of \nPublic Health, Labour Economics, and Social Work. \nAccording to GoogleScholar, his work has been cited \n6,891 times.\nJennifer Karas Montez is a professor of sociology, \nGerald B. Cramer faculty scholar in aging studies, director \nof the Center for Aging and Policy Studies, and codirector \nof the Policy, Place, and Population Health Lab at Syracuse \nUniversity. Her research investigates trends and disparities \nin population health since the 1980s and the growing influ-\nence of U.S. state policies and politics on those outcomes. \nA major focus of her work is understanding why health \ntrends are particularly worrisome for women, for people \nwithout a college degree, and for those living in states in \nthe South and Midwest.\n\n\n    Scientific Research Findings:", "answer": "\u2022 We found states with higher income inequality tend to have lower life expectancy.\n\u2022 However, the inverse association between income inequality and life expectancy existed only among states with low levels of policy liberalism. More liberal state policies regarding LGBT rights, the environment, health and welfare, and gun safety were important in mitigating the harmful effects of income inequality on life expectancy.\n\u2022 The findings illustrate how states like California and New York with liberal policy contexts can exhibit both high income inequality and high life expectancy.", "id": 67}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Differences in Determinants: Racialized Obstetric Care and \nIncreases in U.S. State Labor Induction Rates\nRyan K. Masters1,2, Andrea M. Tilstra2,3, Daniel H. Simon1,2, Kate Coleman-Minahan2,4\n1University of Colorado Boulder, Boulder, CO, USA\n2University of Colorado Population Center, Boulder, CO, USA\n3Oxford University, Oxford, UK\n4University of Colorado Anschutz Medical Campus, Aurora, CO, USA\nAbstract\nInduction of labor (IOL) rates in the United States have nearly tripled since 1990. We examine \nofficial U.S. birth records to document increases in states\u2019 IOL rates among pregnancies to Black, \nLatina, and White women. We test if the increases are associated with changes in demographic \ncharacteristics and risk factors among states\u2019 racial-ethnic childbearing populations. Among \npregnancies to White women, increases in state IOL rates are strongly associated with changes \nin risk factors among White childbearing populations. However, the rising IOL rates among \npregnancies to Black and Latina women are not due to changing factors in their own populations \nbut are instead driven by changing factors among states\u2019 White childbearing populations. The \nresults suggest systemic racism may be shaping U.S. obstetric care whereby care is not \u201ccentered \nat the margins\u201d but is instead responsive to characteristics in states\u2019 White populations.\nKeywords\nhealth care inequities; induction of labor; natality data; obstetric practices; structural racism\nRacial-ethnic inequities in health care have been widely reported in the United States \nwhereby the care and treatment of White people is often prioritized more than that of \nmarginalized populations (Institute of Medicine [IOM] 2003). Evidence for discrimination \nand unequal care in the United States has been documented in numerous settings (e.g., Daw \n2015; Geiger 2003; Lewey and Choudhry 2014; Morris et al. 2010). Unequal treatment \nin health care operates through multiple mechanisms, including policy creation and \nenforcement (Krieger 2001, 2012), the organization of U.S. health care systems (Popescu et \nal. 2010; Williams and Jackson 2005), medical training and culture (Cogburn 2019; Good \net al. 2003), and interpersonal interactions between patients and providers (Hoffman et al. \n2016). Combined, evidence suggests that systemic racism creates and maintains health care \nCorresponding Author: Ryan K. Masters, Department of Sociology, University of Colorado Boulder, Ketchum Hall 264, UCB 327, \nBoulder, CO 80309-0327, USA., ryan.masters@colorado.edu. \nSUPPLEMENTAL MATERIAL\nAppendices A through L are available in the online version of the article.\nHHS Public Access\nAuthor manuscript\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nPublished in final edited form as:\nJ Health Soc Behav. 2023 June ; 64(2): 174\u2013191. doi:10.1177/00221465231165284.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nsystems that underserve and harm communities of color in the United States. That is, the \nU.S. medical system has a history of centering care on the needs of dominant or majority \npopulations (i.e., White patients) rather than centering care \u201cat the margins\u201d or considering \nthe care needs of marginalized populations (Hardeman, Medina, and Kozhimannil 2016).\nFailure to center U.S. obstetric care at the margins has likely produced unequal care and \ndiscriminatory services in obstetric settings (Davis 2019, 2020; Liese et al. 2021; Logan et \nal. 2022; Vedam et al. 2019). Indeed, \u201cobstetric racism\u201d (Davis 2019, 2020) is likely partly \nresponsible for high rates of poor maternal and neonatal health outcomes among U.S. Black \nand Latina populations, such as elevated risks of maternal mortality and infant mortality \n(Mathews, MacDorman, and Thoma 2015; Petersen et al. 2019).1 In her studies of obstetric \nracism, Davis (2019, 2020) documents callous medical treatment of Black women during \npregnancy and highlights multiple instances in which they are disrespected and their birthing \npreferences are discounted and ignored. The racialized experiences of people during prenatal \nand obstetric care are documented further in an emerging body of literature that implicates \nsystemic and interpersonal racism as drivers of inadequate care for obstetric patients of color \n(Chantarat, Van Riper, and Hardeman 2022; Janevic et al. 2020; Logan et al. 2022).\nIn the current study, we are primarily concerned with the rising use of obstetric interventions \namong pregnancies to U.S. women and how systemic racism has likely shaped these trends. \nSpecifically, we consider racial-ethnic differences in the mechanisms driving the large \nincreases in rates of induction of labor (IOL) among U.S. pregnancies. From 1990 to 2017, \nthe average state IOL rate among singleton pregnancies to Black, Latina, and White women \nincreased from 12.5% to 34.4% (National Center for Health Statistics [NCHS] 2020). We \ncontend that the rising use of IOL in the United States provides a good case to illustrate \nhow obstetric care is not being centered at the margins. Although racial-ethnic differences in \nIOL rates are small in the United States and the rising use of IOL has occurred among all \nrace-ethnic populations (Martin et al. 2017; Tilstra and Masters 2020), we examine how the \nfactors associated with increasing IOL rates differ for race-ethnic groups.\nIn the United States, birth attendants are afforded great discretion in decision-making, \nguidelines for IOL are not well defined, and risk assessments of pregnancy and labor often \nuse highly subjective indications (American College of Obstetricians and Gynecologists \n[ACOG] 2007, 2019a, 2019b; Marconi 2019). Combined, U.S. obstetric environments likely \nallow implicit biases and obstetric racism to influence IOL decisions (Davis 2019; Liese \net al. 2021; Vedam et al. 2019). More broadly, systemic and cultural racism shape social \nconditions and practices that generate racial inequities in health care access and heath \npolicies (Cogburn 2019). These racialized processes likely shape obstetric practices in the \nUnited States. As a result, it is possible that rising IOL rates among pregnancies to White \nwomen are partly responding to changes in the health and risk factors of this childbearing \npopulation. By contrast, rising IOL rates among U.S. Black and Latina women may not \nhave occurred because of the changes or needs in these childbearing populations but, rather, \nbecause of the standardization of U.S. obstetric care practices based on the changes and \n1.For parsimony and consistency, we use \u201cmaternal\u201d and \u201cwomen\u201d when referring to birthing people. However, the identities of \nbirthing people include all gender identities.\nMasters et al.\nPage 2\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nneeds among White women. In short, obstetric racism and the failure to center U.S. obstetric \ncare at the margins likely has produced more \u201cinterventions without explanation\u201d among \nBlack and Latina women than among White women (Davis 2019:569).\nIn this article, we first document increases in state IOL rates among pregnancies to \nU.S. states\u2019 Black, Latina, and White childbearing populations between 1990 and 2017. \nThe trends show similar monotonic increases in U.S. states\u2019 IOL rates among all three \npopulations, although trends among White women exhibit some nonlinearity during the \n2000s and 2010s. We then estimate how states\u2019 IOL trends are affected by changes in \nrisk factors for \u201chigh-risk pregnancy\u201d among states\u2019 childbearing populations.2 Evidence \nsuggests that increases in state IOL rates among pregnancies to White women were likely \nresponding to changes in the demographic composition and changes in risk factors among \nstates\u2019 White childbearing populations (e.g., increases in births to women with obesity and \nincreases in the prevalence of maternal hypertension and maternal diabetes). By contrast, \nthe increases in state IOL rates among pregnancies to Black and Latina women are not \nassociated with changes in demographics or changes in risk factors among states\u2019 Black \nand Latina childbearing populations. Instead, we find evidence to suggest that increases in \nU.S. states\u2019 IOL rates among Black and Latina women were strongly shaped by changes \nin the demographics and risk factors of the states\u2019 White childbearing populations. Taken \ntogether, our findings suggest a clear example in which U.S. obstetric care is not being \ncentered at the margins (Hardeman et al. 2016) given that the rising IOL rates among all \nthree racial-ethnic groups appear to be responding only to changes in risk factors among \nstates\u2019 White childbearing populations.\nBACKGROUND\nIn the United States, White people, on average, have greater access to and receive higher \nquality health care than people of color (IOM 2003). Inequities in care are extensively \ndocumented and span many clinical settings and health conditions, including cardiovascular \ncare (Lewey and Choudhry 2014), diabetes treatment (Peek, Cargill, and Huang 2007), \nkidney transplantations (Daw 2015; Malek et al. 2011), mental health services (Neighbors \net al. 2007), addiction treatment (Hansen, Parker, and Netherland 2020; Hansen and Skinner \n2012), cancer screenings (Lansdorp-Vogelaar et al. 2012; Morris et al. 2010; Tehranifar et \nal. 2009), and HIV/AIDS treatment (Bogart et al. 2010). The experience of unequal care \nmanifests across multiple settings both directly and indirectly related to the health care \nsystem. The United States has a long and appalling history of racism in policy creation and \nenforcement. Policies are often enacted to maintain existing power structures and prioritize \nthe dominant power group (Krieger 2001, 2012). Because the dominant power group in \nthe United States has been historically White, policies and policy enforcement disadvantage \nand exclude people of color, often resulting in deleterious health consequences. Examples \ninclude the War on Drugs, which affected how pain was recognized and treated for Black \npatients; banks\u2019 redlining policies; highway construction; mass incarceration; and other \n2.\u201cHigh-risk\u201d pregnancy has no definition, but pregnancy risks are higher for women older than age 35, for women with preexisting \nhealth conditions (e.g., high blood pressure, obesity, diabetes), for people who engage in certain health behaviors while pregnant (e.g., \nsmoking cigarettes, using alcohol), and for multiple pregnancies (e.g., carrying twins or higher order multitudes; NICHD 2018).\nMasters et al.\nPage 3\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nsegregationist measures that affect the neighborhoods and broader environments in which \npeople of color live and work (e.g., Bailey et al. 2017; Roberts 1999; Rothstein 2017).\nUnequal care can also be a result of the organization of U.S. health care systems. Residential \nsegregation, unequal education and opportunities, and discriminatory hiring practices \nhave produced worse care and hindered adequate services in marginalized communities \n(Guagliardo 2004; Hayanga et al. 2009; Odom Walker et al. 2010; White, Haas, and \nWilliams 2012). This, in turn, limits Black Americans\u2019 access to high-quality medical \ncare (Popescu et al. 2010; Williams and Jackson 2005), including Black infants receiving \nneonatal intensive care that is, on average, of poorer quality than care for White infants \n(Horbar et al. 2019).\nHealth care education and culture is built on a history of racism that continues to inform \nthe training and guidance received in health care professional education. The history of \nmedicine is deeply rooted with horrific examples of racial exploitation and neglect (Feagin \nand Bennefield 2014; Washington 2006), from fabricating biological differences by race to \nexperimentation and performing procedures on people of color without consent. This history \nis compounded by the lack of training on implicit bias in medical education (Green et al. \n2021; Holmes 2012; Nieblas-Bedolla et al. 2020). In addition to affecting the behaviors \nand beliefs of medical professionals, this history has also seeped into the algorithms used \nin health care systems to guide care decisions, which have been shown to exhibit racial \nbias (Obermeyer et al. 2019). Cultural racism also more broadly reflects the ideology and \nintent of health policy and practice wherein whiteness is often embedded and centered in \nevaluations, metrics, and expectations of care (Cogburn 2019).\nFinally, interpersonal racism, or racism experienced via interactions between individuals, is \na common form of medical racism that directly biases individual-level care. Two-thirds of \nstudies analyzed in a meta-analysis found evidence of interpersonal racism in the medical \nsetting (Paradies, Truong, and Priest 2014). This can manifest as implicit bias against Black \nand Latina patients (Blair et al. 2013; Hoffman et al. 2016), and indeed, Black women \nreport more mistreatment and disrespect during childbirth than White women (Altman et al. \n2019; Logan et al. 2022; McLemore et al. 2018; Slaughter-Acey et al. 2016; Vedam et al. \n2019). Across all levels, the U.S. health care system directly and implicitly centers itself on \nthe needs of the White population, often resulting in the poor access and mistreatment of \npatients from marginalized racial-ethnic groups. In response, scholars have called for health \ncare providers and researchers to \u201ccenter at the margins\u2014that is, to shift our viewpoint from \na majority group\u2019s perspective to that of the marginalized group or groups\u201d (Hardeman et \nal. 2016:2114). Hardeman et al. (2016) encourage the health care community to redefine \n\u201cnormal\u201d and center the perspectives and needs of marginalized groups at the forefront of \ncare.\nIn U.S. obstetric care, examples of health care prioritizing the welfare and needs of \nthe White population include inequitable access to assisted reproductive technology for \nBlack women compared to White women (Ethics Committee of the American Society for \nReproductive Medicine 2015) and White infants receiving higher-quality care in neonatal \nintensive care units than non-White infants (Profit et al. 2017; Sigurdson et al. 2019). \nMasters et al.\nPage 4\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nThe failure to center obstetric care at the margins manifests as obstetric racism more \nbroadly, which Davis (2019:562) notes \u201cincludes, but is not limited to, critical lapses in \ndiagnosis; being neglectful, dismissive, or disrespectful; causing pain; and engaging in \nmedical abuse through coercion to perform procedures or performing procedures without \nconsent.\u201d Although obstetric racism often manifests through interpersonal racism, existing \nwork also highlights how systemic and structural biases are deeply rooted in U.S. health \ncare systems, including policies, organization, and education. Indeed, Davis (2019:561) sees \n\u201cobstetric racism [as] an extension of racial stratification\u201d and the result of \u201cthe historically \nconstituted stigmatization of Black women.\u201d The coercive nature of obstetric racism is \nparticularly important to consider in the context of U.S. childbirth. Childbirth in the United \nStates presents a unique health care interaction that necessitates the balance of (1) the health \nand risks for the pregnant person and the fetus with (2) the pregnant person\u2019s preferences \nand (3) the providers\u2019 preferences and decisions. In many cases, these risks and preferences \nmay compete during pregnancy care and labor. This strain can be exacerbated by the \ngreat amount of power and flexibility in obstetric decision-making by health care providers \n(ACOG 2007) and the highly subjective risk criteria across many birth procedures, including \nlabor inductions (Marconi 2019).\nThe use of induction of labor, or \u201cthe initiation of uterine contractions before the \nspontaneous onset of labor by medical and/or surgical means for the purpose of delivery,\u201d \nhas steadily increased among U.S. pregnancies since the 1990s (Ventura et al. 1999:92). \nIn 2015, nearly one-quarter of all U.S. births were induced (Martin et al. 2017), up from \njust 10% in 1990 (Osterman and Martin 2014). Labor induction is an important obstetric \nintervention for minimizing risks to maternal and fetal health, and increases in IOL have \nalso come on the heels of efforts to reduce cesarean deliveries (Nicholson et al. 2004, 2009a, \n2009b; Nicholson, Yeager, and Macones 2007). Indeed, scholars and practitioners use the \nterm \u201cpreventative labor induction\u201d when considering how elective IOL might be used to \nreduce risks associated with pregnancy and childbirth (Caughey 2009; Caughey et al. 2009; \nGrobman et al. 2018; Nicholson et al. 2008). However, IOL is often overused in the United \nStates, as evidenced by research suggesting that two-thirds of the increase in IOL during \nthe 1990s was a result of \u201cnonmedically indicated\u201d inductions (Ramsey, Ramin, and Ramin \n2000), and gestational distributions of U.S. births have been dramatically changed by the \nincreasing use of IOL at select gestations (Tilstra and Masters 2020).\nIn the United States, IOL rates do not substantially differ across race-ethnic childbearing \npopulations, and the increasing trends in IOL have occurred in all race-ethnic groups in \nsimilar ways (Martin et al. 2017). Although IOL rates among pregnancies to U.S. Black, \nLatina, and White childbearing populations are similar, we suspect that the reasons for \nthe high and rising use of IOL are different for White, Black, and Latina women in the \nUnited States. For instance, inductions among pregnancies to Latina and Black women \nmay be more likely to occur due to nonmedically indicated reasons than inductions \namong pregnancies to White women. Multilevel racism in the U.S. health care system \nhas possibly contributed to the perception that U.S. pregnancies among Black and Latina \nwomen are more likely to be \u201chigh risk\u201d than pregnancies among White women. The \ndeeply ingrained racist perceptions of risk in the U.S. health care system may affect IOL \ndecisions, where providers intervene in pregnancies to Black and Latina women to reduce \nMasters et al.\nPage 5\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nharm from these perceived risks. Also, higher rates of interventions without consent during \nBlack and Latina pregnancies reflect the callous and egregious care that patients of color \nreceive during childbirth (Logan et al. 2022). Indeed, for women of color in the United \nStates, \u201cneglect, lack of information, dismissiveness, disrespect, and interventions without \nexplanation [emphasis added], permeate maternal care and coalesce into obstetric racism\u201d \n(Davis 2019:569). These forms of obstetric racism have likely shaped the rising use of IOL \nin the United States in significantly racialized ways.\nDATA AND METHODS\nWe examined trends in U.S. states\u2019 IOL rates using the National Vital Statistics Systems \n(NVSS) restricted birth data for years 1990 through 2017 (NCHS 2020).3 To reduce the \nconfounding effects of multiparous women and multiple pregnancies on risk for obstetric \ninterventions (Denona et al. 2020; Donahue et al. 2010), we restricted the analytic samples \nto include only singleton first births among non-Hispanic White, non-Hispanic Black, \nand Hispanic women (henceforth, White, Black, and Latina). The data are composed of \n41,126,037 singleton first births: 26,446,616 to White women, 6,252,741 to Black women, \nand 8,426,680 to Latina women. We aggregated the data at the state level by mother\u2019s \nrace-ethnicity to create separate analytic samples for births among states\u2019 White, Black, \nand Latina childbearing populations (for the creation of the analytic samples, see Figure \nS1 in the appendix in the online version of the article). Due to small counts of births, \nwe omitted Idaho, Maine, Montana, North Dakota, South Dakota, Vermont, and Wyoming \nfrom the analytic sample for births to Black women, and we omitted Maine, Vermont, and \nWest Virginia from the analytic sample for births to Latina women. The analytic sample \nfor births to White women is composed of all 50 states plus the District of Columbia (DC) \nacross 28 years (1,428 state-years), and the analytic samples for births to Black and Latina \nwomen were limited to 43 states plus DC (1,232 state-years) and 47 states plus DC (1,344 \nstate-years), respectively.4\nMeasures\nWe calculated state-level time-varying measures of obstetric interventions, maternal \ndemographics, and risk factors for \u201chigh-risk pregnancy\u201d among states\u2019 Black, Latina, \nand White childbearing populations. Our outcome measure was the proportion of births \nin state i in year j in which labor was induced, where i = \u00a0Alabama, \u2026, Wyoming and \nj = 1990, \u2026, 2017.5 Induction of labor is a characteristic of delivery that is comparable across \nthe 1989 U.S. Standard Certificate of Live Birth and the 2003 U.S. Standard Certificate of \nLive Birth (Martin et al. 2007), and IOL coding in the NVSS is also comparable across \nU.S. states.6 As a control variable, we also created a measure of the cesarean delivery rate \n3.Data analyzed are Restricted-Use Vital Statistics Birth Data, which can be accessed via application: https://www.cdc.gov/nchs/nvss/\nnvss-restricted-data.htm.\n4.Results from models fitted to pregnancies among White and Latina women in reduced analytic samples composed of the same 44 \nstates are nondifferent from results in the article and are available in Appendix J in the online version of the article.\n5.For details about imputed values, see Appendix E in the online version of the article.\n6.Records of IOL in Wisconsin during the 1990s and early 2000s were prone to error. We adjusted IOL rates in Wisconsin for all years \n1990 through 2002. See Appendix D in the online version of the article.\nMasters et al.\nPage 6\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\namong first-birth singletons born to states\u2019 Black, Latina, and White women, calculated as \nthe proportion of singleton first births in state i delivered by cesarean in year j.\nSeveral factors determine whether a pregnancy is considered high risk (Holness 2018; \nNational Institute of Child Health and Human Development [NICHD] 2018). Risk of \npregnancy complications is higher for women older than age 35, and pregnancies can also be \naffected by several health behaviors such as smoking cigarettes, drinking alcohol, and using \nother substances. Maternal health problems such as high blood pressure, obesity, diabetes, \nthyroid disease, infections, and heart or blood disorders can also influence pregnancy risks \n(NICHD 2018). Here, we included several state-level measures of maternal demographics \nas the proportion of births in state i in year j to women younger than 20 years (i.e., \nproportion of births to teenagers), age 35 years and older (i.e., proportion of births to \nwomen of advanced maternal age [AMA]),7 married at time of birth, born in the United \nStates, with an educational level less than high school degree, and with an educational level \nat or above a bachelor\u2019s degree. Direct measures of risk factors for high-risk pregnancy \nincluded the proportion of births in state i in year j to women who used tobacco products \nat any time during pregnancy, experienced high gestational weight gain (i.e., 40 pounds or \nmore), had diabetes (gestational or pre-pregnancy), and had hypertension (gestational or pre-\npregnancy). Additional state-level proxies for high-risk pregnancies included the proportion \nof births in state i in year j delivered preterm (<37 weeks of gestation) and delivered late \nterm or postterm (\u226541 weeks of gestation).8\nState-Level Growth Curve Models\nWe modeled U.S. states\u2019 IOL rates between 1990 and 2017 separately for Black, Latina, \nand White women. We fit generalized linear mixed models to estimate the year-specific \nstate-level rates as outcomes of a fixed effect linear slope, a random intercept, a random \nslope, and state-specific residual variance:\nY ij = \u03c00i + \u03c01i( Y earij \u22121990) /5 + \u03b5ij\n\u03c00i = \u03b300 + \u03bc0i\n\u03c01i = \u03b310 + \u03bc1i .\n(1)\nAssuming \u03bc0i\n\u03bc1i~\nN([ 0\n0] ,[ \u03c30\n2 \u03c301\n\u03c310 \u03c31\n2])  and \u03b5ij \u223cN( 0, \u03c3\u03b5\n2)  where Y ij is the IOL rate for state i in year \nj, where i = \u00a0Alabama, \u2026, Wyoming and j = 1990, \u2026, 2017; \u03b300 is the average IOL rate among \nU.S. states in year 1990; \u00b50i is the estimated deviation from \u03b300 for state I in 1990; \u03b310\nis the average five-year change in IOL rate between 1990 and 2017; \u00b51i is the estimated \ndeviation from \u03b310 for state i; and \u03b5ij is the Level 1 residual variance for state i in year j. We \ndivided the slope by 5 for interpretation reasons (i.e., the estimated coefficient indicates the \n7.Advanced maternal age itself designates a pregnancy as \u201chigh risk,\u201d but we include it as a measure of states\u2019 childbearing \n\u201cdemographics\u201d alongside maternal age less than 20 years.\n8.We included measures of states\u2019 economic indicators and inequality such as the Gini index, child poverty rate, unemployment rate, \nhousing price index (in 1990 dollars), and rates of Supplemental Nutrition Assistance Program (SNAP) benefits. Results from models \nthat these controls are in Appendix I in the online version of the article.\nMasters et al.\nPage 7\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nexpected change in states\u2019 IOL rates over a five-year time span), and increasing the slope \nsize improves estimates of the variance component of \u00b51i, \u03c31\n2 (Singer and Willett 2003).\nWe contrasted estimates from the random slope model in Equation 1 with estimates from \nan unconditional means model to illustrate (a) the amount of variation in IOL rates that \nexists within states and between states; (b) the amount of within-states variation, \u03c3\u03b5\n2, that is \naccounted for by a linear approximation of a state-specific slope, \u03c01i; and (c) the amount \nof between-states variation in the IOL rate in 1990, \u03c30\n2, and the amount of between-states \nvariation in the change of IOL rate between 1990 and 2017, \u03c31\n2. Results from these models \nindicate the extent to which increases in IOL rates are similar or different across U.S. states \nand the extent to which the increases are similar or different for racial-ethnic populations.\nState-Level Fixed Effect Panel Regression\nWe then fitted trends in IOL rates among pregnancies to U.S. states\u2019 Black, Latina, and \nWhite women using state-level fixed effects panel regressions:\nY ij = \u03b20 + \u03b32State2 + \u2026 + \u03b3nStaten\n+\u03b419911991 + \u2026 + \u03b420172017 + \u03bcij,\n(2)\nwhere Y ij is the IOL rate among births to Black, White, or Latina women in state i in \ntime j; \u03b20 is the IOL rate in the referent state in 1990; \u03b3n are the coefficients associated \nwith the binary state regressors, Staten; \u03b4j are the coefficients associated with the binary \ntime regressors, 1991, \u2026 , 2017; and \u00b5ij is the error term. These \u201cwithin-estimator\u201d models \ncontrol for all time-invariant characteristics of U.S. states while also accounting for yearly \ntrends in IOL rates shared across states (Halaby 2004). The models are well suited for \nidentifying average trends in U.S. states\u2019 IOL rates while controlling for time-invariant state \ncharacteristics and for estimating how states\u2019 time-varying characteristics are associated \nwith changes in states\u2019 IOL rates.\nWe refitted the models to include time-varying indicators of the demographic profiles \nand risk factors for high-risk pregnancy among states\u2019 racial-ethnic-specific childbearing \npopulations:\nY ij = \u03b20 + \u03b32State2 + \u2026 + \u03c1nStaten + \u03b419911991\n+\u2026 + \u03b420172017 + \u03b2teen%teenij + \u03b2AMA%AMAij\n+\u03b2married%marriedij + \u03b2USborn%USbornij\n+\u03b2 < HS% < HSij + \u03b2BA%BAij + \u03b2tobacco%tobaccoij\n+\u03b2diabetes%diabetesij + \u03b2\u210eypertension%\u210eypertensionij\n+\u03b2weig\u210et%weig\u210etij + \u03b2cesarean%cesareanij\n+\u03b2preterm%pretermij + \u03b2posttern%posttermij + \u03bcij,\n(3)\nwhere \u03b2teen is the coefficient associated with the percent of births in state i to Black, \nLatina, or White women occurring among teenagers in year j, %teenij; \u03b2AMA is the coefficient \nMasters et al.\nPage 8\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nassociated with the percent of births among women age 35 years or older, %AMAij; \u03b2married\nis the coefficient associated with the percent of births among married women, %marriedij; \n\u03b2USborn is the coefficient associated with the percent of births among U.S.-born women, \n%USbornij; \u03b2 < HS is the coefficient associated with the percent of births among women with \nless than high school education, % < HSij; \u03b2BA is the coefficient associated with the percent \nof births among women with a college degree, %BAij; \u03b2tobacco is the coefficient associated \nwith the percent of births among women who used tobacco while pregnant, %tobaccoij; \n\u03b2diabetes is the coefficient associated with the percent of births among women with diabetes, \n%diabetesij; \u03b2\u210eypertension is the coefficient associated with the percent of births among women \nwith hypertension, %\u210eypertensionij; \u03b2weig\u210et is the coefficient associated with the percent of \nbirths among women with gestational weight gain greater than 40 pounds, %weig\u210etij; \u03b2cesarean\nis the coefficient associated with the percent of births delivered cesarean, %cesareanij; \u03b2preterm\nis the coefficient associated with the percent of births delivered at gestational week <37, \n%pretermij; and \u03b2postterm is the coefficient associated with the percent of births delivered after \ngestational week 41, %posttermij.\nWe then refitted the models to U.S. states\u2019 IOL rates among births to Black, Latina, and \nWhite women to include time-varying indicators of the demographic profiles and risk factors \nof the states\u2019 other race-ethnic childbearing populations. For example, Equation 4 regresses \nU.S. states\u2019 IOL rates among births to Black women on changes in the demographics and \nmaternal risk factors of the states\u2019 White childbearing populations:\nY ij = \u03b20 + \u03b32State2 + \u2026 + \u03b3nStaten + \u03b419911991\n+\u2026 + \u03b420172017 + \u03b2W teen%W teenij\n+\u03b2W AMA%W AMAij + \u03b2W married%W marriedij\n+\u03b2W USborn%W USbornij + \u03b2W < HS%W < HSij\n+\u03b2W BA%W BAij + \u03b2W tobacco%W tobaccoij\n+\u03b2W diabetes%W diabetesij\n+\u03b2W \u210eypertension%W \u210eypertensionij\n+\u03b2W weig\u210et%W weig\u210etij + \u03b2W cesarean%W cesareanij\n+\u03b2W preterm%W pretermij\n+\u03b2W posttern%W posttermij + \u03bcij .\n(4)\nAfter estimating Equations 2 through 4, we plotted the model-based expected IOL rates \nacross each year j holding all other covariates at their 1990 mean levels using the margins \nmodule in Stata 17. We contrasted the average trends in U.S. states\u2019 IOL rates among \nBlack, Latina, and White women estimated from Equation 2 (i.e., the observed rates) with \nthe adjusted trends estimated from Equation 3 to examine the extent to which changes in \nthe demographics and risk factors among states\u2019 Black, Latina, and White childbearing \npopulations are associated with the observed trends in states\u2019 IOL rates among these \npopulations. Then, we contrasted the average trends in U.S. states\u2019 IOL rates estimated \nfrom Equation 2 with the adjusted trends estimated from Equation 4 to examine the extent \nto which changes in the demographics and risk factors among states\u2019 other race-ethnic \nMasters et al.\nPage 9\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nchildbearing populations are associated with the changes in IOL rates among births to states\u2019 \nBlack, Latina, and White childbearing populations.9\nRESULTS\nTable 1 contains state-level descriptive statistics of the Black, Latina, and White analytic \nsamples in 1990, 2004, and 2017. We present the mean state IOL rates among singleton \nfirst births across these years and the demographic profiles and risk factors for high-risk \npregnancy among U.S. states\u2019 Black, Latina, and White childbearing populations.\nAverage state IOL rates among births to U.S. Black women increased from about 11% in \n1990 to 23% in 2004 and to 33% in 2017. Similar increases in IOL are observed among \nsingleton first births born to U.S. Latina (10%, 21%, 31%) and White women (14%, 28%, \n36%). Across this time, we also see substantial changes in the demographic profiles of \nstates\u2019 childbearing populations. Most noteworthy is the increasingly older age distributions \nof the states\u2019 childbearing populations, with large reductions in the proportion of births \nto teens (e.g., 43%\u201317% among Black women) and concomitant increases among women \nwith AMA (e.g., 5%\u201310% among White women). We also see sizable decreases in the \nproportion of births to women with education levels less than high school (e.g., 37%\u201323% \namong Latina women), which are offset by large proportionate increases in births to women \nwith college degrees (e.g., 9%\u201320% among Black women and 24%\u201344% among White \nwomen). The proportion of births to women who are married decreased substantially in the \nLatina (60%\u201340%) and White (76%\u201366%) childbearing populations. Finally, the proportion \nof births among immigrants increased in the Black childbearing population (7%\u201321%), and \nthe proportion of births among immigrants in the Latina childbearing population increased \nbetween 1990 and 2004 (47%\u201362%) and then decreased from 2004 to 2017 (62%\u201340%).\nChanges in risk factors for high-risk pregnancy are also observed in all three childbearing \npopulations. The proportion of births to women who used tobacco while pregnant dropped \n(e.g., 10%\u20134% among Black women, 18%\u20138% among White women), and the proportion of \nsingleton first births born in late-term or postterm gestations also declined (e.g., 23%\u201315% \namong Black women, 28%\u201318% among Latina women). Yet across this same time, in all \nthree populations, we see increasing rates of gestational diabetes (2% in 1990 to 5%\u20137% in \n2017), hypertension (4%\u20135% in 1990 to 8%\u201313% in 2017), and high gestational weight gain \n(e.g., 24%\u201328% among Black women). Among the White childbearing population, we also \nsee a small increase in the proportion of births born at premature gestations (8% to 10% to \n9%).10\nTo see the full year-over-year changes in U.S. states\u2019 IOL rates among births to Black, \nLatina, and White women, we plot IOL rates in each state in Figure 1 (gray lines) and \nindicate the yearly mean rate among all states (black lines).\n9.For results from several sensitivity analyses, such as combining Equation 3 and Equation 4 and using three-year lagged effects of \npredictors, see Appendix K in the online version of the article.\n10.Correlations between characteristics and risk factors among states\u2019 White, Latina, and Black populations in 1990, 2004, and 2017 \nare in Appendix L in the online version of the article.\nMasters et al.\nPage 10\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nIn 1990, the mean IOL rate among states\u2019 singleton first births born to Black, Latina, and \nWhite women were 10.8%, 9.6%, and 13.5%, respectively. The rates varied considerably \nacross states. For example, among pregnancies to Black women, IOL rates ranged from \n4.9% in Mississippi to 21.9% in Kentucky, and among White women, rates ranged from \n8.0% in California to 21.5% in Oregon (for each state\u2019s IOL rate among Black, Latina, and \nWhite women in 1990, 2004, and 2017, Appendix B in the online version of the article). \nYet IOL rates increased among all states\u2019 childbearing populations between 1990 and 2017. \nThe trends exhibit similar monotonic increases in all states\u2019 IOL rates among the three \npopulations, although trends among White women exhibit some nonlinearity during the \n2000s and 2010s. In 2017, the average IOL rate among U.S. states\u2019 Black, Latina, and White \nwomen were 33.4%, 31.0%, and 35.9%, respectively. As was the case in 1990, state-based \nvariation in these 2017 rates is high, ranging from 19.3% among Black and Latina women in \nCalifornia to 54.0% among White women in West Virginia.\nTable 2 presents results from growth curve models fitted to U.S. states\u2019 IOL rates among \nsingleton first births born to Black, Latina, and White women for all years 1990 through \n2017. Results from the unconditional means model (Panel A) suggest that only 14% \n(Latina), 17% (Black), and 23% (White) of variation in U.S. IOL rates between 1990 \nand 2017 occurred between states ( i . e ., \u03c30\n2(\n\u03c3\u03b5\n2 + \u03c30\n2)) . Thus, 77% (White) to 86% (Latina) \nof variation in U.S. IOL rates occurred within states over time ( i . e ., \u03c3\u03b5\n2(\n\u03c3\u03b5\n2 + \u03c30\n2)) , reflecting \nthe large increases in IOL rates observed in Figure 1. Results from the random slope \nmodel (Panel B) indicate that 77% (Black), 78% (Latina), and 79% (White) of the within-\nstates variation is accounted for by a linear approximation of changes in states\u2019 IOL rates (\ni . e .,( \u03c3\u03b5RSM\n2\n\u2212\u03c3\u03b5UMM\n2)\n\u03c3\u03b5UMM\n2)\n. Indeed, when linear ordinary least squares models are fitted \nseparately to each state\u2019s IOL rates in these childbearing populations, the median R2 is about \n.85 (Appendix C in the online version of the article). Results also indicate that estimates \nof the slope do not substantively differ for Black (.039), Latina (.034), and White women \n(.036), suggesting that states\u2019 IOL rates increased among these childbearing populations in \nsimilar ways (i.e., about 3.4%\u20133.9% every five years). Taken together, the findings from \nthe generalized mixed models suggest near-linear increases in U.S. states\u2019 IOL rates among \nWhite, Black, and Latina childbearing populations that occurred across U.S. states in similar \nways. The similarity in these trends is evident by the very small slope variances in Table \n1 (\u03c31\n2) and is also observed in Figure 2, which plots predicted Bayes estimates of IOL rates \namong states\u2019 Black, Latina, and White childbearing populations.\nIn Figure 3, we plot the mean state IOL rates estimated from Equations 2 through 4 to \ncontrast the observed mean IOL rates among states\u2019 Black, Latina, and White women (i.e., \n\u201cObserved\u201d) with the estimated mean IOL rates among states\u2019 Black, Latina, and White \nwomen while controlling for changes in these populations\u2019 demographics and risk factors \n(i.e., \u201cControl Own Characteristics\u201d) and the estimated mean IOL rates among states\u2019 Black, \nLatina, and White women while controlling for the demographics and risk factors of the \nstates\u2019 other race-ethnic childbearing populations. For the latter, we plot mean IOL rates \nfor Black and Latina women while controlling for the demographics and risk factors of \nstates\u2019 White childbearing populations (i.e., \u201cControl White Characteristics\u201d), and we plot \nMasters et al.\nPage 11\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nmean IOL rates for White women while controlling for the demographics and risk factors \nof states\u2019 Black childbearing populations (for detailed results, see Appendices F\u2013H in the \nonline version of the article).\nResults from Equation 3 indicate that U.S. states\u2019 IOL rates among pregnancies to Black \nwomen are associated with both the demographic profiles and risk factors of states\u2019 Black \nchildbearing populations (Appendix F in the online version of the article). Yet, as shown \nin Figure 3 (Panel A), changes in these demographics and risk factors between 1990 and \n2017 do not account for the upward trend in states\u2019 IOL rates among pregnancies to Black \nwomen. We see that the mean IOL rate indicated by the \u201cControl Own Characteristics\u201d \nline (dashed black) is nondifferent from the mean IOL rate indicated by \u201cObserved\u201d line \n(solid black). Thus, the increases in states\u2019 IOL rates among pregnancies to Black women \nare estimated to occur even while controlling for changes in demographic characteristics \nand risk factors among states\u2019 Black childbearing populations. Results from \u201cEquation \n(4)-White\u201d (Appendix F in the online version of the article) indicate that changes in the \ndemographics and risk factors of states\u2019 White childbearing populations are statistically \nand substantively associated with states\u2019 IOL rates among pregnancies to Black women. \nFurthermore, changes in risk factors among states\u2019 White childbearing populations account \nfor much of the rising IOL rates among states\u2019 Black childbearing populations. As shown \nin Figure 3 (Panel A), the rise in the mean IOL rate indicated by the \u201cControl White \nCharacteristics\u201d line (solid gray) is much lower than the upward trend of the \u201cObserved\u201d \nline. This suggests that increases in IOL among states\u2019 Black women would have been much \nsmaller if the demographic and risk factors of states\u2019 White childbearing populations had not \nchanged between 1990 and 2017.\nWe find similar results from models fitted to states\u2019 IOL rates among Latina women. \nIn Figure 3 (Panel B), there are no significant differences between the \u201cObserved\u201d line \n(solid black) and \u201cControl Own Characteristics\u201d line (dashed black) during the time period \n1990 to 2005, but the \u201cControl Own Characteristics\u201d line is significantly lower than the \n\u201cObserved\u201d rates for the time period 2005 to 2017. Although the differences are not \nsubstantively large, these findings suggest that a small fraction of the rising IOL rates \namong states\u2019 Latina women are associated with changes in the demographics and risk \nfactors of the states\u2019 Latina childbearing populations. Yet states\u2019 IOL rates among Latina \nwomen are also statistically and substantively associated with changes in demographic \ncharacteristics and risk factors among states\u2019 White childbearing populations (\u201cEquation \n(4)-White,\u201d Appendix G in the online version of the article). As shown in Figure 3 (Panel \nB), the \u201cControl White Characteristics\u201d line (solid gray) indicates that the average IOL \nrate among pregnancies to Latina women would not have substantively increased between \n1990 and 2017 if states\u2019 White childbearing populations had not experienced changes in \ntheir demographic composition or maternal risk factors. Thus, the rising trends in states\u2019 \nIOL rates among pregnancies to Latina women are largely explained by changes in the \ndemographic composition and risk factors of states\u2019 White childbearing populations.\nResults from Equation 2 (Appendix H in the online version of the article) suggest \nthat U.S. states\u2019 IOL rates among pregnancies to White women are associated with \nboth the demographic profiles and risk factors in states\u2019 White childbearing populations. \nMasters et al.\nPage 12\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nFurthermore, we find evidence suggesting that changes in these demographic characteristics \nand risk factors of states\u2019 White childbearing populations are strongly associated with the \nrising IOL rates among pregnancies to states\u2019 White women between 1990 and 2017. \nThis is clearly shown in Figure 3 (Panel C), which suggests that states\u2019 IOL rates among \npregnancies to White women (\u201cObserved\u201d line, solid black) would not have increased if \nthe demographic profiles and risk factors of states\u2019 White childbearing populations had \nremained at 1990 levels (\u201cControl Own Characteristics\u201d line, dashed black). Also shown \nin Figure 3 is the lack of associations between states\u2019 IOL trends among pregnancies to \nWhite women and changes in demographics and risk factors of states\u2019 Black childbearing \npopulations (\u201cControl Black Characteristics\u201d line, solid gray). We see that the average IOL \nrate among states\u2019 White childbearing populations is estimated to have increased between \n1990 and 2017 even if the demographic characteristics and maternal risk factors of states\u2019 \nBlack childbearing populations had not changed during this time.\nTogether, results presented in Figure 3 indicate that U.S. states\u2019 rising IOL rates among \npregnancies to Black, Latina, and White women are strongly associated with changes in the \ndemographic profiles and risk factors in states\u2019 White childbearing populations. By contrast, \nstates\u2019 rising IOL rates among Black and Latina women are not strongly associated with \nchanges in the demographic characteristics or risk factors in these populations. Nor are \nchanges in the demographic characteristics or risk factors in Black and Latina childbearing \npopulations associated with the rising state IOL rates among pregnancies to White women. \nThus, the increasing use of IOL in the United States appears to be a national-level \nphenomenon occurring in all states and among pregnancies to Black, Latina, and White \nwomen in similar ways and strongly responding to changes in the demographic profiles and \nrisk factors in White childbearing populations.\nDISCUSSION\nRecent changes in U.S. obstetric practices provide a good case for identifying processes \nthat might generate racial-ethnic inequities in U.S. health care more broadly. In many \nrespects, obstetric care in the United States reflects the same policies, hospital systems, \nmedical trainings and cultures, and interpersonal racism that often shape medical care of \nmarginalized populations (Cogburn 2019; IOM 2003). In other respects, challenges unique \nto obstetric practice might amplify the racialized care for women of color in the United \nStates. The guidelines for risk management of pregnancies and risk assessment for labor \ncomplications can be unclear, and in these environments, \u201cneglect, lack of information, \ndismissiveness, disrespect, and interventions without explanation\u201d can shape obstetric care \nfor U.S. women of color (Davis 2019:569). Indeed, at the individual level, Davis (2019) \nand others have documented many examples and forms of \u201cobstetric racism\u201d in maternal, \nprenatal, and labor care, including increased likelihood of receiving care without consent \n(Janevic et al. 2020; Logan et al. 2022). In this study, we sought to examine how obstetric \nracism and systemic racism more broadly may have shaped use of IOL at the population \nlevel in the United States.\nWe first documented trends in U.S. states\u2019 IOL rates among singleton first-birth pregnancies \nto Black, Latina, and White women. Results show that mean rates of IOL among states\u2019 \nMasters et al.\nPage 13\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nBlack, Latina, and White women (combined) nearly tripled between 1990 and 2017 (i.e., \nincreased from 12.5% to 34.4%) and that the increases in IOL among these populations \noccurred across all states in similar ways. At first glance, the racial-ethnic similarities in \n(1) high IOL rates (i.e., 31%\u201336% in 2017), (2) the rising use of IOL (i.e., increases of \n3.4%\u20133.9% per five years), and (3) the ubiquity of these trends throughout the United States \nappear to provide evidence against obstetric racism shaping the increased use of IOL in the \nUnited States. How can one implicate racism in a health outcome if that outcome does not \nsubstantively differ across racial-ethnic groups? Thus, we hypothesized that the mechanisms \nunderlying the rising trends in IOL likely differ for U.S. racial-ethnic groups and devised \na simple test. If IOL has been increasingly used among U.S. pregnancies to reduce risk of \nadverse birth and maternal health outcomes (Nicholson et al. 2009a), then the rising rates of \nIOL among U.S. pregnancies should be associated with changes in demographic profiles and \nrisk factors for high-risk pregnancy.\nThe evidence here suggests that this has, indeed, been the case for pregnancies among \nstates\u2019 White childbearing populations. Among pregnancies to White women, low maternal \neducation, gestational tobacco use, maternal diabetes, maternal hypertension, and high \ngestational weight gain were all estimated to be positively associated with risk of IOL. \nChanges in these and other risk factors among states\u2019 White childbearing populations \naccounted for all the increase in states\u2019 IOL rates among pregnancies to White women. \nBy contrast, we did not find this to be the case for IOL use among pregnancies to Black or \nLatina women. The increases in states\u2019 IOL rates among pregnancies to Black and Latina \nwomen were not explained by changes in these populations\u2019 demographic profiles or risk \nfactors. Rather, the rising use of IOL among states\u2019 Black and Latina women is strongly \nassociated with the changes in demographics and risk factors of states\u2019 White childbearing \npopulations. Thus, it appears that changes in U.S. states\u2019 IOL use are associated only with \nchanges in the characteristics of states\u2019 White childbearing populations.\nTaken together, evidence here suggests that the increasing use of IOL in the United States \nhas not been centered on the needs of the Black and Latina childbearing populations \nbut, instead, is likely responding to the changing needs and/or preferences of the White \nchildbearing population. Although we did not identify the underlying reasons for these \nassociations, we offer two potential explanations for future research to consider. One \npossible explanation is the way in which standards of obstetric care are created for the \nmajority or \u201cnormal\u201d pregnant person (i.e., White women) and then applied to all patients. \nThe processes underlying this explanation reflect health care providers\u2019 desires and efforts \nto reduce perceived risks associated with pregnancy and childbirth. That is, obstetric \ninterventions such as IOL can be used to minimize the risk of adverse birth outcomes and \nadverse maternal outcomes from high-risk pregnancies (Nicholson et al. 2004, 2007, 2009a, \n2009b). In their efforts to minimize risk among the majority White populations, hospitals, \nclinics, and care providers in the United States may be shifting their practices to identify \n\u201chigh-risk\u201d pregnancies and minimize poor outcomes of these pregnancies via IOL. Indeed, \nsince 1990, the gestational age distribution of births in the United States has shifted from \nIOLs primarily occurring at later gestational ages (40 weeks or more) to occurring more \nfrequently at earlier gestational ages (weeks 37\u201339; Tilstra and Masters 2020).\nMasters et al.\nPage 14\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nEvidence presented here suggests that this new normative practice of using obstetric \ninterventions at earlier gestations likely arose in response to the (perceived) needs of White \nwomen but that the normative practice may have come to shape care for all U.S. populations \nin similar ways. Thus, standards of pregnancy care are created and adjusted to meet the \nneeds of the majority or \u201cnormal patient\u201d (i.e., White women alone) and then applied to all \nracial-ethnic groups, regardless of patients\u2019 needs and preferences (Hardeman et al. 2016). \nProviders may also care more about minimizing harm and preventing adverse outcomes \namong White patients because the lives and children of White women are implicitly more \nvalued in the United States (Harris and Wolfe 2014). In either case, obstetric racism has \nresulted in prioritizing care for White women.\nA second possible explanation is that racist \u201crisk perception\u201d in U.S. clinical and obstetric \npractices have differentially affected the use of IOL for managing pregnancy and childbirth \namong U.S. racial-ethnic populations. Here, pregnancies among Black and Latina women \nmay be perceived and diagnosed as being riskier and less healthy than pregnancies to \nWhite women. Public attention to maternal and infant morbidity and mortality among U.S. \ncommunities of color, health care training and education, and implicit biases might influence \nproviders\u2019 perceptions of the health and needs of patients from these communities (Feagin \nand Bennefield 2014; Horbar et al. 2019; Washington 2006). Thus, providers may perceive, \nassess, and treat patients based on their skin color and/or on racist perceptions of their \nfamilial background and communities instead of their individual risk factors for carrying a \nhealthy pregnancy to full term or engaging in person-centered care and listening to their \nneeds and preferences, as providers are more likely to do with White patients (Altman et al. \n2019; Logan et al. 2022; McLemore et al. 2018; Slaughter-Acey et al. 2016; Vedam et al. \n2019).\nAdditionally, misperceptions about pain sensitivity might influence providers\u2019 reception to \nBlack and Latina women\u2019s desires and preferences in obstetric care settings. Foundational \npractices in obstetrics and gynecology were strongly influenced by the racist care of J. \nMarion Sims, who developed surgical procedures by operating without consent and without \nanesthesia on enslaved Black women (Washington 2006). Sims\u2019s experiments contributed \nto the persistent false belief that Black and White patients have biological differences in \npain perception (Hoffman et al. 2016). Indeed, the expansive and pervasive history of \nunequal health care in the United States has fostered an environment ripe for obstetric \nracism. Thus, our findings might implicate racism in the assessment of patients\u2019 needs and \nin clinical decision-making, which are consistent with a large body of evidence identifying \nracial-ethnic inequities in delivery of medical care in the United States (IOM 2003).\nFindings here provide complementary, structural-level evidence that align with individual-\nlevel studies of obstetric racism (Davis 2019, 2020; Janevic et al. 2020; Logan et al. 2022). \nWe demonstrate that the pattern of adjusting obstetric care to meet the changing composition \nand risk factors of the White population exists at the state level, thus showing that the \npatterns of obstetric racism persist at a population level in the United States. This contributes \nto an ever-growing body of literature implicating the pervasiveness of racism across U.S. \ninstitutions. Examples of institutions outside health care not centering at the margins include \nhow systemic racism infiltrates the education system through segregation and standardizing \nMasters et al.\nPage 15\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nWhite children as the norm in testing (Knoester and Au 2017; McGee 2020; Vaught and \nCastagno 2008) and how practices such as redlining and the behaviors of housing market \nprofessionals discriminate against communities of color and contribute to racial segregation \nby viewing White homeowners and tenants living arrangements and practices as \u201cnormal\u201d \n(Korver-Glenn 2021; Rothstein 2017). These and other institutions continue to engage in \nand teach racist behaviors that perpetuate racial differences in several economic, social, and \nhealth outcomes. Here, we provide further evidence that the field of obstetrics also likely \nengages in behavior and practices that prioritizes the (perceived) needs and desires of the \nWhite childbearing population.\nThis study has several limitations. First, data are composed at the state level and might \nfail to measure how changes in risk factors and IOL move together at other spatial levels \nwithin states (e.g., county or hospital). For example, hospitals that disproportionately serve \nBlack and Latina populations might be changing obstetric practices for reasons not observed \nin these data but that are correlated with changes in demographics and risk factors among \nstates\u2019 White childbearing populations. Second, the NVSS data contain a limited set of \nmeasures for \u201chigh-risk\u201d pregnancy. The rising use of IOL among states\u2019 Black and Latina \nwomen may be responding to changes in these childbearing populations that we are not \nobserving in these data. Yet our findings were robust to sensitivity analyses that controlled \nfor a number of possible economic-related confounders (see Notes 8 and 9). Third, we are \nnot measuring or directly observing \u201cobstetric racism\u201d or forms of systemic racism. Rather, \nwe are implicating differences in determinants of IOL trends as evidence for racism in the \nuse of IOL. This limitation, however, can be perceived as a strength of the study. One need \nnot document differences in outcomes by race-ethnicity or directly measure exposures to \nracism to suggest that the processes generating those outcomes might be shaped by racial-\nethnic differences in the mechanisms. Indeed, in this case, if researchers had documented \nonly the racial-ethnic similarities in U.S. IOL rates and trends, they would have concluded \nthat there are no racial-ethnic inequities in IOL practices.\nAlthough results from this study must be interpreted with these limitations in mind, our \nfindings are consistent with an extensive literature documenting health care inequity in the \nUnited States and provide strong evidence that U.S. obstetric care has not been centered \non the needs of Black and Latina childbearing populations. Results here indicate that \nrising use of IOL among pregnancies to Latina and Black women are likely responding \nto characteristics in White childbearing populations, not characteristics in these populations \nthemselves. These findings provide population-level evidence to support claims of obstetric \nracism documented in individual-level data and therefore provide an empirical foundation on \nwhich researchers can study the underlying mechanisms for racial-ethnic inequities in U.S. \nobstetric care.\nSupplementary Material\nRefer to Web version on PubMed Central for supplementary material.\nACKNOWLEDGMENTS\nThe authors are grateful to the anonymous referees and the JHSB editors for their helpful comments.\nMasters et al.\nPage 16\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nFUNDING\nThe authors disclosed receipt of the following financial support for the research, authorship, and/or publication of \nthis article: This research was funded by a grant from the Eunice Kennedy Shriver Institute of Child Health Human \nand Human Development (RHD099359A) and benefited from support provided to the University of Colorado \nPopulation Center (Project 2P2CHD066613-06) from the Eunice Kennedy Shriver Institute of Child Health Human \nand Human Development. The content is solely the responsibility of the authors and does not necessarily represent \nthe official views of the National Institute of Child Health and Human Development or the National Institutes.\nBiographies\nKate Coleman-Minahan, PhD, RN, FNP-BC, is an assistant professor at the University \nof Colorado College of Nursing and a research affiliate of the University of Colorado \nPopulation Center. She is a nurse practitioner and social scientist who studies reproductive \nhealth equity, including access to contraception and abortion among young people and \nimmigrants.\nRyan K. Masters is an associate professor of sociology at the University of Colorado \nBoulder, a faculty associate of the University of Colorado Population Center, and a fellow of \nthe Institute of Behavioral Science.\nDaniel H. Simon is a PhD candidate in the Department of Sociology at the University of \nColorado Boulder and a University of Colorado Population Center graduate research affiliate \nin the Institute of Behavioral Science. He is a social demographer who studies the structural \ndeterminants of suicide mortality and population health inequality.\nAndrea M. Tilstra, PhD, is a postdoctoral researcher in the Leverhulme Centre for \nDemographic Science, the Department of Sociology, and Nuffield College at the University \nof Oxford. She is a social demographer who studies the population health consequences \nof policy changes, institutional practices, and public health crises, revealing how these \nstructural factors may exacerbate or equalize existing health inequalities.\nREFERENCES\nAltman Molly R., Oseguera Talita, McLemore Monica R., Kantrowitz-Gordon Ira, Franck Linda S., \nand Lyndon Audrey. 2019. \u201cInformation and Power: Women of Color\u2019s Experiences Interacting \nwith Health Care Providers in Pregnancy and Birth.\u201d Social Science & Medicine 238. doi:10.1016/\nj.socscimed.2019.112491.\nAmerican College of Obstetricians and Gynecologists. 2007. \u201cCommittee \nOpinion Number 390: Ethical Decision-Making in Obstetrics and \nGynecology\u201d https://www.acog.org/clinical/clinical-guidance/committee-opinion/articles/2007/12/\nethical-decision-making-in-obstetrics-and-gynecology.\nAmerican College of Obstetricians and Gynecologists. 2019a. \u201cCommittee Opinion Number 765: \nAvoidance of Nonmedically Indicated Early Term Deliveries and Associated Neonatal Morbidities.\u201d \nObstetrics & Gynecology 133(2):e156\u201363. [PubMed: 30681546] \nAmerican College of Obstetricians and Gynecologists. 2019b. \u201cCommittee Opinion Number \n764: Medically Indicated Late-Preterm and Early Term Deliveries.\u201d Obstetrics & Gynecology \n133(2):e151\u201355. [PubMed: 30681545] \nBailey Zinzi D., Krieger Nancy, Ag\u00e9nor Madina, Graves Jasmine, Linos Natalia, and Bassett Mary \nT.. 2017. \u201cStructural Racism and Health Inequities in the USA: Evidence and Interventions.\u201d The \nLancet 389(10077):1453\u201363.\nMasters et al.\nPage 17\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nBlair Irene V., Havranek Edward P., Price David W., Hanratty Rebecca, Fairclough Diane L., Farley \nTillman, Hirsh Holen K., and Steiner John F.. 2013. \u201cAssessment of Biases against Latinos and \nAfrican Americans among Primary Care Providers and Community Members.\u201d American Journal of \nPublic Health 103(1):92\u201398. [PubMed: 23153155] \nBogart Laura M., Wagner Glenn J., Galvan Frank H., and Klein David J.. 2010. \u201cLongitudinal \nRelationships between Antiretroviral Treatment Adherence and Discrimination due to HIV-\nSerostatus, Race, and Sexual Orientation among African American Men with HIV.\u201d Annals of \nBehavioral Medicine 40(2):184\u201390. [PubMed: 20552416] \nCaughey Aaron B. 2009. \u201cPreventive Induction of Labor: Can Its Use Lower the Cesarean Delivery \nRate?\u201d Journal of Women\u2019s Health 18(11):1743\u201345. doi:10.1089/jwh.2009.1559.\nCaughey Aaron B., Sundaram Vandana, Kaimal Anjali J., Gienger Allison, Cheng Yvonne W., \nMcDonald Kathryn M., Shaffer Brian L., and Bravata Douglas M.. 2009. \u201cSystematic Review: \nElective Induction of Labor versus Expectant Management of Pregnancy.\u201d Annals of Internal \nMedicine 151(4):252\u201363. [PubMed: 19687492] \nChantarat Tongtan, Van Riper David C., and Hardeman Rachel R.. 2022. \u201cMultidimensional Structural \nRacism Predicts Birth Outcomes for Black and White Minnesotans.\u201d Health Services Research \n57(3):448\u201357. [PubMed: 35468220] \nCogburn Courtney D. 2019. \u201cCulture, Race, and Health: Implications for Racial Inequities and \nPopulation Health.\u201d The Milbank Quarterly 97(3):736\u201361. [PubMed: 31512293] \nDavis D\u00e1na-Ain. 2019. \u201cObstetric Racism: The Racial Politics of Pregnancy, Labor, and Birthing.\u201d \nMedical Anthropology 38(7):560\u201373. [PubMed: 30521376] \nDavis D\u00e1na-Ain. 2020. \u201cReproducing while Black: The Crisis of Black Maternal Health, Obstetric \nRacism, and Assisted Reproductive Technology.\u201d Reproductive Biomedicine & Society Online \n11:56\u201364. [PubMed: 33305027] \nDaw Jonathan. 2015. \u201cExplaining the Persistence of Health Disparities: Social Stratification and \nthe Efficiency\u2013Equity Trade-Off in the Kidney Transplantation System.\u201d American Journal of \nSociology 120(6):1595\u2013640.\nDenona Branko, Foley Michael, Mahony Rhona, and Robson Michael. 2020. \u201cDiscrimination by \nParity Is a Prerequisite for Assessing Induction of Labour Outcome\u2014Cross-Sectional Study.\u201d \nBMC Pregnancy Childbirth 20(1):709. doi:10.1186/s12884-020-03334-8. [PubMed: 33225906] \nDonahue Sara M.A., Kleinman Ken P., Gillman Matthew W., and Oken Emily. 2010. \u201cTrends in Birth \nWeight and Gestational Length among Singleton Term Births in the United States: 1990\u20132005.\u201d \nObstetrics & Gynecology 115(2 Part 1):357\u201364. [PubMed: 20093911] \nEthics Committee of the American Society for Reproductive Medicine. 2015. \u201cDisparities in Access \nto Effective Treatment for Infertility in the United States: An Ethics Committee Opinion.\u201d Fertility \nand Sterility 104(5):1104\u201310. [PubMed: 26364838] \nFeagin Joe, and Bennefield Zinobia. 2014. \u201cSystemic Racism and U.S. Health Care.\u201d Social Science & \nMedicine 103:7\u201314. [PubMed: 24507906] \nGeiger H Jack. 2003. \u201cRacial and Ethnic Disparities in Diagnosis and Treatment: A Review of \nthe Evidence and a Consideration of Causes.\u201d Pp. 417\u201354 in Unequal Treatment: Confronting \nRacial and Ethnic Disparities in Health Care, edited by Smedley BD, Stith AY, and Nelson AR. \nWashington, DC: The National Academies Press.\nGood Mary-Jo D., James Cara, Good Byron J., and Becker Anne E.. 2003. \u201cThe Culture of Medicine \nand Racial, Ethnic, and Class Disparities in Healthcare.\u201d Pp. 594\u2013625 in Unequal Treatment: \nConfronting Racial and Ethnic Disparities in Health Care, edited by Smedley BD, Stith AY, and \nNelson AR. Washington, DC: The National Academies Press.\nGreen Tiffany L., Zapata Jasmine Y., Brown Heidi W., and Hagiwara Nao. 2021. \u201cRethinking Bias \nto Achieve Maternal Health Equity: Changing Organizations, Not Just Individuals.\u201d Obstetrics & \nGynecology 137(5):935\u201340. [PubMed: 33831936] \nGrobman William A., Rice Madeline M., Reddy Uma M., Tita Alan T.N., Silver Robert M., \nMallett Gail, Hill Kim, et al. 2018. \u201cLabor Induction versus Expectant Management for \nLow-Risk Nulliparous Women.\u201d New England Journal of Medicine 379(6):513\u201323. doi:10.1056/\nNEJMoa1800566. [PubMed: 30089070] \nMasters et al.\nPage 18\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nGuagliardo Mark F. 2004. \u201cSpatial Accessibility of Primary Care: Concepts, Methods, and \nChallenges.\u201d International Journal of Health Geographics 3(1):1\u201313. [PubMed: 14748927] \nHalaby Charles N. 2004. \u201cPanel Models in Sociological Research: Theory into Practice.\u201d Annual \nReview of Sociology 30:507\u201344.\nHansen Helena, Parker Caroline, and Netherland Jules. 2020. \u201cRace as a Ghost Variable in (White) \nOpioid Research.\u201d Science, Technology, & Human Values 45(5):848\u201376.\nHansen Helena, and Skinner Mary E.. 2012. \u201cFrom White Bullets to Black Markets and Greened \nMedicine: The Neuroeconomics and Neuroracial Politics of Opioid Pharmaceuticals.\u201d Annals of \nAnthropological Practice 36(1):167\u201382.\nHardeman Rachel R., Medina Eduardo M., and Kozhimannil Katy B.. 2016. \u201cStructural Racism and \nSupporting Black Lives\u2014The Role of Health Professionals.\u201d New England Journal of Medicine \n375(22):2113\u201315. [PubMed: 27732126] \nHarris Lisa H., and Wolfe Taida. 2014. \u201cStratified Reproduction, Family Planning Care, and the \nDouble Edge of History.\u201d Current Opinion in Obstetrics and Gynecology 26(6):539\u201344. [PubMed: \n25379770] \nHayanga Awori J., Waljee Akbar K., Kaiser Heather E., Chang David C., and Morris Arden M.. \n2009. \u201cRacial Clustering and Access to Colorectal Surgeons, Gastroenterologists, and Radiation \nOncologists by African Americans and Asian Americans in the United States: A County-Level \nData Analysis.\u201d Archives of Surgery 144(6):532\u201335. [PubMed: 19528386] \nHoffman Kelly M., Trawalter Sophie, Axt Jordan R., and Oliver M. Norman. 2016. \u201cRacial Bias in \nPain Assessment and Treatment Recommendations, and False Beliefs about Biological Differences \nbetween Blacks and Whites.\u201d Proceedings of the National Academy of Sciences 113(16):4296\u2013\n301.\nHolmes Seth M. 2012. \u201cThe Clinical Gaze in the Practice of Migrant Health: Mexican Migrants in the \nUnited States.\u201d Social Science & Medicine 74(6):873\u201381. [PubMed: 21992736] \nHolness Nola. 2018. \u201cHigh-Risk Pregnancy.\u201d Nursing Clinics of North America 53(2):241\u201351. \ndoi:10.1016/j.cnur.2018.01.010. [PubMed: 29779516] \nHorbar Jeffrey D., Edwards Erika M., Greenberg Lucy T., Profit Jochen, Draper David, Helkey Daniel, \nLorch Scott A., et al. 2019. \u201cRacial Segregation and Inequality in the Neonatal Intensive Care \nUnit for Very Low-Birth-Weight and Very Preterm Infants.\u201d JAMA Pediatrics 173(5):455\u201361. \n[PubMed: 30907924] \nInstitute of Medicine. 2003. Unequal Treatment: Confronting Racial and Ethnic Disparities in Health \nCare Washington, DC: The National Academies Press. doi:10.17226/12875.\nJanevic Teresa, Piverger Naissa, Afzal Omara, and Howell Elizabeth A.. 2020. \u201c\u2018Just because You \nHave Ears Doesn\u2019t Mean You Can Hear\u2019\u2014Perception of Racial-Ethnic Discrimination during \nChildbirth.\u201d Ethnicity & Disease 30(4):533\u201342. [PubMed: 32989353] \nKnoester Matthew, and Au Wayne. 2017. \u201cStandardized Testing and School Segregation: Like Tinder \nfor Fire?\u201d Race Ethnicity and Education 20(1):1\u201314.\nKorver-Glenn Elizabeth. 2021. Race Brokers: Housing Markets and Segregation in 21st Century Urban \nAmerica New York, NY: Oxford University Press.\nKrieger Nancy. 2001. \u201cTheories for Social Epidemiology in the 21st Century: An Ecosocial \nPerspective.\u201d International Journal of Epidemiology 30(4):668\u201377. [PubMed: 11511581] \nKrieger Nancy. 2012. \u201cMethods for the Scientific Study of Discrimination and Health: An Ecosocial \nApproach.\u201d American Journal of Public Health 102(5):936\u201344. [PubMed: 22420803] \nLansdorp-Vogelaar Iris, Kuntz Karen M., Knudsen Amy B., Van Ballegooijen Marjolein, Zauber \nAnne G., and Jemal Ahmedin. 2012. \u201cContribution of Screening and Survival Differences to \nRacial Disparities in Colorectal Cancer Rates.\u201d Cancer Epidemiology and Prevention Biomarkers \n21(5):728\u201336.\nLewey Jennifer, and Choudhry Niteesh K.. 2014. \u201cThe Current State of Ethnic and Racial Disparities \nin Cardiovascular Care: Lessons from the Past and Opportunities for the Future.\u201d Current \nCardiology Reports 16(10):530. doi:10.1007/s11886-014-0530-3. [PubMed: 25135343] \nLiese Kylea L., Davis-Floyd Robbie, Stewart Karie, and Cheyney Melissa. 2021. \u201cObstetric \nIatrogenesis in the United States: The Spectrum of Unintentional Harm, Disrespect, Violence, \nand Abuse.\u201d Anthropology & Medicine 28(2):188\u2013204. [PubMed: 34196238] \nMasters et al.\nPage 19\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nLogan Rachel G., McLemore Monica R., Julian Zo\u00eb, Stoll Kathrin, Malhotra Nisha, GVtM Steering \nCouncil, and Saraswathi Vedam. 2022. \u201cCoercion and Non-consent during Birth and Newborn \nCare in the United States.\u201d Birth 49(4):749\u201362. [PubMed: 35737547] \nMalek Sayeed K., Keys Brandon J., Kumar Sanjaya, Milford Edgar, and Tullius Stefan G.. 2011. \n\u201cRacial and Ethnic Disparities in Kidney Transplantation.\u201d Transplant International 24(5):419\u201324. \n[PubMed: 21166727] \nMarconi Anna M. 2019. \u201cRecent Advances in the Induction of Labor.\u201d F1000Research 8:F1000 \nFaculty Rev-1829. doi:10.12688/f1000research.17587.1.\nMartin Joyce A., Hamilton Brady E., Osterman Michelle J., Driscoll Anne K., and Mathews TJ. \n2017. \u201cBirths: Final Data for 2015.\u201d NVSS 66(1). https://www.cdc.gov/nchs/data/nvsr/nvsr66/\nnvsr66_01.pdf.\nMartin Joyce A., Hamilton Brady E., Sutton Paul D., Ventura Stephanie J., Menacker Fay, Kirmeyer \nSharon, and Munson Martha L.. 2007. \u201cBirths: Final Data for 2005.\u201d NVSS 56(6). https://\nwww.cdc.gov/nchs/data/nvsr/nvsr56/nvsr56_06.pdf.\nMathews TJ, MacDorman Marian F., and Thoma Marie E.. 2015. \u201cInfant Mortality Statistics from \nthe 2013 Period Linked Birth/Infant Death Dataset.\u201d NVSS 64(9). https://stacks.cdc.gov/view/cdc/\n32752.\nMcGee Ebony Omotola. 2020. \u201cInterrogating Structural Racism in STEM Higher Education.\u201d \nEducational Researcher 49(9):633\u201344.\nMcLemore Monica R., Altman Molly R., Cooper Norlissa, Williams Shanell, Rand Larry, and Franck \nLinda. 2018. \u201cHealth Care Experiences of Pregnant, Birthing, and Postnatal Women of Color at \nRisk for Preterm Birth.\u201d Social Science & Medicine 201:127\u201335. [PubMed: 29494846] \nMorris Arden M., Rhoads Kim F., Stain Stain C., and Birkmeyer John D.. 2010. \u201cUnderstanding \nRacial Disparities in Cancer Treatment and Outcomes.\u201d Journal of the American College of \nSurgeons 211(1):105\u201313. [PubMed: 20610256] \nNational Center for Health Statistics. 2020. [Natality Birth Files, 1990\u20132017]\nNational Institute of Child Health and Human Development. 2018. \u201cHigh-Risk Pregnancy\u201d https://\nwww.nichd.nih.gov/health/topics/high-risk.\nNeighbors Harold W., Caldwell Cleopatra, Williams David R., Nesse Randolph, Taylor Robert J., \nBullard Kai M., Torres Myriam, and Jackson James S.. 2007. \u201cRace, Ethnicity, and the Use of \nServices for Mental Disorders: Results from the National Survey of American Life.\u201d Archives of \nGeneral Psychiatry 64(4):485\u201394. [PubMed: 17404125] \nNicholson James M., Caughey Aaron B., Stenson Morghan H., Cronholm Peter F., Kellar Lisa, Bennett \nIan, Margo Katie, and Straton Joseph B.. 2009a. \u201cThe Active Management of Risk in Multiparous \nPregnancy at Term: Association between a Higher Preventive Labor Induction Rate and Improved \nBirth Outcomes.\u201d American Journal of Obstetrics and Gynecology 200(3):250.e1\u201313.\nNicholson James M., Cronholm Peter, Kellar Lisa C., Stenson Morghan H., and Macones George \nA.. 2009b. \u201cThe Association between Increased Use of Labor Induction and Reduced Rate of \nCesarean Delivery.\u201d Journal of Women\u2019s Health 18(11): 1747\u201358.\nNicholson James M., Kellar Lisa C., Cronholm Peter F., and Macones George A.. 2004. \n\u201cActive Management of Risk in Pregnancy at Term in an Urban Population: An Association \nbetween a Higher Induction of Labor Rate and a Lower Cesarean Delivery Rate.\u201d American \nJournal of Obstetrics and Gynecology 191(5):1516\u201328. doi:10.1016/j.ajog.2004.07.002. [PubMed: \n15547519] \nNicholson James M., Parry Samuel, Caughey Aaron B., Rosen Sarah, Keen Allison, and Macones \nGeorge A.. 2008. \u201cThe Impact of the Active Management of Risk in Pregnancy at Term on \nBirth Outcomes: A Randomized Clinical Trial.\u201d American Journal of Obstetrics and Gynecology \n198(5):511.e1\u2013e15.\nNicholson James M., Yeager David L., and Macones George. 2007. \u201cA Preventive Approach to \nObstetric Care in a Rural Hospital: Association between Higher Rates of Preventive Labor \nInduction and Lower Rates of Cesarean Delivery.\u201d Annals of Family Medicine 5(4):310\u20139. \ndoi:10.1370/afm.706. [PubMed: 17664497] \nMasters et al.\nPage 20\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nNieblas-Bedolla Edwin, Christophers Briana, Nkinsi Naomi T., Schumann Paul D., and Stein Elizabeth \n2020. \u201cChanging How Race Is Portrayed in Medical Education: Recommendations from Medical \nStudents.\u201d Academic Medicine 95(12):1802\u2013806. [PubMed: 32379145] \nObermeyer Ziad, Powers Brian, Vogeli Christine, and Mullainathan Sendhil. 2019. \u201cDissecting Racial \nBias in an Algorithm Used to Manage the Health of Populations.\u201d Science 366(6464):447\u201353. \n[PubMed: 31649194] \nOdom Walker, Gery Ryan Kara, Ramey Robin, Nunez Felix L., Beltran Robert, Splawn Robert G., and \nBrown Arleen F.. 2010. \u201cRecruiting and Retaining Primary Care Physicians in Urban Underserved \nCommunities: The Importance of Having a Mission to Serve.\u201d American Journal of Public Health \n100(11):2168\u201375. [PubMed: 20935263] \nOsterman Michelle J., and Martin Joyce A.. 2014. \u201cRecent Declines in Induction of Labor by \nGestational Age\u201d Centers for Disease Control and Prevention, National Center for Health Statistics \nNational Vital Statistics Report. https://stacks.cdc.gov/view/cdc/23533.\nParadies Yin, Truong Mandy, and Priest Naomi. 2014. \u201cA Systematic Review of the Extent and \nMeasurement of Healthcare Provider Racism.\u201d Journal of General Internal Medicine 29(2):364\u2013\n87. [PubMed: 24002624] \nPeek Monica E., Cargill Algernon, and Huang Elbert S.. 2007. \u201cDiabetes Health Disparities.\u201d Medical \nCare Research and Review 64(5 Suppl.):101S\u201356S. [PubMed: 17881626] \nPetersen Emily E., Davis Nicole L., Goodman David, Cox Shanna, Syverson Carla, Seed Kristi, \nShapiro-Mendoza Carrie, Callaghan William M., and Barfield W. 2019. \u201cRacial-Ethnic Disparities \nin Pregnancy-Related Deaths\u2014United States, 2007\u20132016.\u201d Morbidity and Mortality Weekly \nReport 68(35):762\u201365. [PubMed: 31487273] \nPopescu Iona, Nallamothu Brahmajee K., Vaughan-Sarrazin Mary S., and Cram Peter. 2010. \u201cRacial \nDifferences in Admissions to High-Quality Hospitals for Coronary Heart Disease.\u201d Archives of \nInternal Medicine 170(14):1209\u201315. [PubMed: 20660838] \nProfit Jochen, Gould Jeffrey B., Bennett Mihoko, Goldstein Benjamin A., Draper David, Phibbs \nCiaran S., and Lee Henry C.. 2017. \u201cRacial-Ethnic Disparity in NICU Quality of Care Delivery.\u201d \nPediatrics 140(3):e20170918. doi:10.1542/peds.2017-0918. [PubMed: 28847984] \nRamsey Patrick S., Ramin Kirk D., and Ramin Susan M.. 2000. \u201cLabor Induction.\u201d Current Opinion in \nObstetrics and Gynecology 12(6):463\u201373. [PubMed: 11128407] \nRoberts Dorothy E. 1999. Killing the Black Body: Race, Reproduction, and the Meaning of Liberty \nNew York, NY: Vintage.\nRothstein Richard. 2017. The Color of Law: A Forgotten History of How Our Government Segregated \nAmerica New York, NY: Liveright Publishing Corporation.\nSigurdson Krista, Mitchell Briana, Liu Jessica, Morton Christine, Gould Jeffrey B., Lee Henry \nC., Nicole Capdarest-Arest, and Jochen Profit. 2019. \u201cRacial-Ethnic Disparities in Neonatal \nIntensive Care: A Systematic Review.\u201d Pediatrics 144(2):e20183114. doi:10.1542/peds.2018-3114. \n[PubMed: 31358664] \nSinger Judith D., and Willett John B.. 2003. Applied Longitudinal Data Analysis: Modeling Change \nand Event Occurrence New York, NY: Oxford University Press.\nSlaughter-Acey Jaime C., Sealy-Jefferson Shawnita, Helmkamp Laura, Caldwell Cleopatra H., Osypuk \nTheresa L., Platt Robert W., Straughen Jennifer K., Dailey-Okezie Rhonda K., Abeysekara Purni, \nand Misra Dawn P.. 2016. \u201cRacism in the Form of Microaggressions and the Risk of Preterm Birth \namong Black Women.\u201d Annals of Epidemiology 26(1):7\u201313. [PubMed: 26549132] \nTehranifar Parisa, Neugut Alfred I., Phelan Jo C., Link Bruce G., Liao Yuyan, Desai Manisha, and \nBeth Terry Mary. 2009. \u201cMedical Advances and Racial-Ethnic Disparities in Cancer Survival.\u201d \nCancer Epidemiology and Prevention Biomarkers 18(10):2701\u2013708.\nTilstra Andrea M., and Masters Ryan K.. 2020. \u201cWorth the Weight? Recent Trends in Obstetric \nPractices, Gestational Age, and Birth Weight in the United States.\u201d Demography 57(1):99\u2013121. \n[PubMed: 31997231] \nVaught Sabrina E., and Castagno Angelina E.. 2008. \u201c\u2018I Don\u2019t Think I\u2019m a Racist\u2019: Critical Race \nTheory, Teacher Attitudes, and Structural Racism.\u201d Race Ethnicity and Education 11(2):95\u2013113.\nVedam Saraswathi, Stoll Kathrin, Taiwo Tanya K., Rubashkin Nicholas, Cheyney Melissa, Strauss \nNan, McLemore Monica, et al. 2019. \u201cThe Giving Voice to Mothers Study: Inequity and \nMasters et al.\nPage 21\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nMistreatment during Pregnancy and Childbirth in the United States.\u201d Reproductive Health \n16(1):77. doi:10.1186/s12978-019-0729-2. [PubMed: 31182118] \nVentura Stephanie J., Martin Joyce A., Curtin Sally M., and Mathews TJ. 1999. \u201cBirths: Final Data for \n1997.\u201d NVSS 47(18). https://stacks.cdc.gov/view/cdc/84459.\nWashington Harriet A. 2006. Medical Apartheid: The Dark History of Medical Experimentation on \nBlack Americans from Colonial Times to the Present New York, NY: Doubleday Books.\nWhite Kellee, Haas Jennifer S., and Williams David R.. 2012. \u201cElucidating the Role of Place in \nHealth Care Disparities: The Example of Racial-Ethnic Residential Segregation.\u201d Health Services \nResearch 47(3 Pt. 2):1278\u201399. [PubMed: 22515933] \nWilliams David R., and Braboy Jackson Pamela. 2005. \u201cSocial Sources of Racial Disparities in \nHealth.\u201d Health Affairs 24(2):325\u201334. [PubMed: 15757915] \nMasters et al.\nPage 22\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nFigure 1. \nU.S. States\u2019 Labor Induction Rates among Pregnancies to Black Women, Latina Women, \nand White Women, 1990 to 2017.\nSource: National Vital Statistics System restricted natality data, 1990 to 2017.\nMasters et al.\nPage 23\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nFigure 2. \nPredicted Bayes Estimates of U.S. States\u2019 Labor Induction Rates among Pregnancies to \nBlack Women, Latina Women, and White Women, 1990 to 2017.\nSource: National Vital Statistics System restricted natality data, 1990 to 2017.\nMasters et al.\nPage 24\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nFigure 3. \nEstimates of Mean Labor Induction Rates among U.S. States\u2019 Black, Latina, and White \nChildbearing Populations, 1990 to 2017.\nSource: National Vital Statistics System restricted natality data, 1990 to 2017.\nNote: \u201cObserved\u201d lines are mean induction of labor (IOL) rates estimated from Equation \n2, the \u201cControl Own Characteristics\u201d lines are mean IOL rates estimated from Equation 3 \nwhile holding constant states\u2019 1990 levels of demographic characteristics and risk factors, \n\u201cControl White Characteristics\u201d lines are mean IOL rates estimated from Equation 4 while \nholding constant states\u2019 1990 levels of demographic characteristics and risk factors among \nWhite childbearing populations, and \u201cControl Black Characteristics\u201d lines are mean IOL \nrates estimated from Equation 4 while holding constant states\u2019 1990 levels of demographic \ncharacteristics and risk factors among Black childbearing populations.\nMasters et al.\nPage 25\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\n\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nMasters et al.\nPage 26\nTable 1.\nDescriptive Statistics of Analytic Samples of Singleton First Births among U.S. States\u2019 Black, Latina, and White Childbearing Populations in 1990, 2004, \nand 2017.\nBlack\nLatina\nWhite\n1990\n2004\n2017\n1990\n2004\n2017\n1990\n2004\n2017\nCharacteristics of labor/delivery\n Labor induction\n.10\n.23\n.33\n.10\n.21\n.31\n.14\n.28\n.36\n Cesarean delivery\n.23\n.30\n.32\n.24\n.25\n.25\n.24\n.28\n.28\nMaternal demographics\n Teen maternal age\n.43\n.35\n.17\n.32\n.31\n.21\n.20\n.16\n.08\n Advanced maternal age\n.02\n.05\n.07\n.02\n.04\n.06\n.05\n.09\n.10\n Education <high school\n.31\n.25\n.13\n.37\n.44\n.23\n.15\n.12\n.06\n Education BA+\n.09\n.14\n.20\n.11\n.11\n.17\n.24\n.37\n.44\n Married\n.28\n.25\n.29\n.60\n.42\n.40\n.76\n.67\n.66\n U.S.-born\n.93\n.85\n.79\n.53\n.38\n.60\n.96\n.95\n.94\nPregnancy risk factors\n Gestational tobacco use\n.10\n.07\n.04\n.09\n.04\n.03\n.18\n.13\n.08\n Diabetes\n.02\n.03\n.05\n.02\n.03\n.07\n.02\n.03\n.06\n Hypertension\n.05\n.07\n.13\n.04\n.05\n.08\n.05\n.07\n.11\n Weight gain 40+ lbs.\n.24\n.29\n.28\n.24\n.26\n.25\n.27\n.34\n.30\n Preterm gestation\n.16\n.15\n.14\n.11\n.11\n.10\n.08\n.10\n.09\n Postterm gestation\n.23\n.17\n.15\n.28\n.20\n.18\n.31\n.20\n.19\nStates\n44\n44\n44\n48\n48\n48\n51\n51\n51\nSource: National Vital Statistics System restricted natality data, 1990 to 2017.\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\n\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nAuthor Manuscript\nMasters et al.\nPage 27\nTable 2.\nResults from Unconditional Growth Curve Models Fitted to States\u2019 Labor Induction Rates among Pregnancies to Black, Latina, and White Childbearing \nPopulations, 1990\u20132017.\nBlack\nLatina\nWhite\nb\nSE\nb\nSE\nb\nSE\nPanel A: Unconditional means model\n Intercept\n.227\n.006\n.208\n.004\n.265\n.006\n Variance components\n  \u03c30\n2\n.0012\n.0003\n.0007\n.0002\n.0015\n.0003\n  \u03c3\u03b5\n2\n.0060\n.0002\n.0044\n.0002\n.0050\n.0002\nPanel B: Random slope model, unstructured error-covariance matrix\n 5-year time\n.039\n.002\n.034\n.001\n.036\n.002\n Intercept\n.121\n.007\n.116\n.005\n.168\n.005\n Variance components\n  \u03c31\n2\n.0001\n.0000\n.0001\n.0000\n.0002\n.0000\n  \u03c30\n2\n.0021\n.0005\n.0011\n.0003\n.0013\n.0003\n  \u03c301\n\u2212.0003\n.0001\n\u2212.0002\n.0001\n\u2212.0002\n.0001\n  \u03c3\u03b5\n2\n.0014\n.0001\n.0010\n.0000\n.0010\n.0000\nN states\n44\n48\n51\nState-years\n1,232\n1,344\n1,428\nSource: National Vital Statistics System restricted natality data, 1990 to 2017.\nNote: b indicates estimated coefficient, and SE indicates standard error.\nJ Health Soc Behav. Author manuscript; available in PMC 2024 February 07.\n\n\n    Scientific Research Findings:", "answer": "\u2022 Between 1990 and 2017, increases in IOL among pregnancies to White women are explained by changes in risk factors among White childbearing populations.\n\u2022 Between 1990 and 2017, increases in IOL among pregnancies to Black women are not explained by changes in risk factors among their own populations; rather, these increases are largely explained by changes in risk factors among White childbearing populations.\n\u2022 Together, changes in the demographic characteristics and risk factors of White childbearing populations are driving the rising use of IOL in the United States for all racial-ethnic populations.", "id": 68}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465231166334\nJournal of Health and Social Behavior\n2023, Vol. 64(3) 336\u00ad\u2013353\n\u00a9 The Author(s) 2023\nDOI: 10.1177/00221465231166334\njournals.sagepub.com/home/hsb\nOriginal Article\nThe inverse relationship between socioeconomic sta-\ntus (SES) and mental health is a cornerstone of the \ngeneral argument linking social inequality to well-\nbeing (Adler et al. 1994; Aneshensel 2009; Pearlin \n1989; Thoits 2010). The SES\u2013mental health associa-\ntion is also one of the most consistent in social sci-\nence (Thoits 2010), spanning over a century of \nresearch (Dohrenwend and Dohrenwend 1969; Faris \nand Dunham 1939; Hollingshead and Redlich 1958; \nJarvis 1855; Srole et al. 1962) and continuing to the \npresent day (Miech and Shanahan 2000; Warren \n2009). This consistency within countries, predomi-\nnantly estimated in the North American context, sup-\nports a social causation interpretation, net of social \nselection (Warren 2009; Wheaton 1978), indicating \nSES as a primary social determinant of mental health. \nHowever, cross-national comparisons across coun-\ntries demonstrate that the strength of the association \nbetween SES indicators, such as education, occupa-\ntion, and income, and mental health outcomes differs \n(Rai et al. 2013), implying contextual social determi-\nnants may be driving differences in the effect of SES \nbetween countries while also demonstrating the gen-\nerality of the existence of this association across very \ndifferent national contexts (Pr\u00e4g, Mills, and Wittek \n2016).\nAt the same time, attention has also focused on \nthe role of the institutional arrangements of welfare \nstate policies as generalized social determinants of \n1166334 HSBXXX10.1177/00221465231166334Journal of Health and Social BehaviorParbst and Wheaton\nresearch-article2023\n1University of Toronto, Toronto, ON, Canada\nCorresponding Author:\nMatthew Parbst, Department of Sociology, University of \nToronto, 725 Spadina Avenue, Toronto, ON M5S 2J4, \nCanada. \nEmail: matt.parbst@mail.utoronto.ca\nThe Effect of Welfare State \nPolicy Spending on the \nEqualization of Socioeconomic \nStatus Disparities in Mental \nHealth\nMatthew Parbst1\n and Blair Wheaton1\nAbstract\nThis article examines whether and how the relationship between socioeconomic status (SES) and \ndepression is modified by welfare state spending using the 2006, 2012, and 2014 survey rounds of the \nEuropean Social Survey (ESS) merged with macroeconomic data from the World Bank, Eurostat, and \nSOCX database (N = 87,466). Welfare state spending effort divided between social investment and social \nprotection spending modifies the classic inverse relationship between SES and depression. Distinguishing \npolicy areas in both social investment and social protection spending demonstrates that policy programs \ndevoted to education, early childhood education and care, active labor market policies, old age care, \nand incapacity account for differences in the effect of SES across countries. Our analysis finds that social \ninvestment policies better explain cross-national differences in the effect of SES on depression, implying \npolicies focused earlier in the life course matter more for understanding social disparities in the mental \nhealth of populations.\nKeywords\ndepression, social investment, social protection, socioeconomic status, welfare state\n\nParbst and Wheaton\t\n337\nhealth, often due to their effects on SES (Beckfield \net al. 2015; \u00d3lafsd\u00f3ttir and Beckfield 2020; \nReynolds 2021). To assess the effects of policy, we \nutilize spending effort as an operational indicator of \npolicy differences, recognizing that policies may \nexist with varying levels of funding and that the \nhypothetical presence of a policy without support-\nive spending is likely to have little to no effect \n(Castles 2008). Recent reviews demonstrate that the \nspending effort in general areas of social policy is \nassociated with SES differences in mental health, \nwhere increases in spending generally reduce SES \ndifferences both between (McAllister et al. 2018) \nand within countries (Simpson et al. 2021). \nBetween-country differences highlight the impor-\ntance of contextual differences in policy environ-\nments for SES inequalities in mental health, \naveraged over time, whereas within-country analy-\nsis demonstrate changes in either the direction or \nmagnitude of specific policy effects over a given \nperiod. Between-country studies find greater levels \nof active labor market (Niedzwiedz et al. 2016) and \nsickness benefit spending (van der Wel et al. 2015) \nreduce educational differences in mental health. \nWithin-country analyses demonstrate that SES \ninequalities are impacted by the direction of spend-\ning effort so that increases reduce and reductions \nincrease \ninequalities \nacross \nfamily \npolicies \n(Lebihan and Mao Takongmo 2018; Milligan and \nStabile 2011), social assistance (Herbst 2013), and \nold age security (Carrino, Glaser, and Avendano \n2020).\nOur approach incorporates both between- and \nwithin-country effects of policy by using cross-\nlevel interactions to understand if cumulative his-\ntorical realities or short-term changes through time \nin policy investment have distinct and/or differen-\ntial impacts on the association between SES and \nmental health at the individual level (Bell and Jones \n2015). Each taps a distinct realm of the role of state \npolicy. Between-country differences, partitioned \nfrom within-country differences over time, reflect \noverall averaged differences across countries that \npersist through time. Between-country differences \nallow for the identification of cumulative impacts \nmore straightforwardly. For example, the cumula-\ntive effects of relatively stable levels of investment \nin childcare could affect both the consequences of \nhaving children and the expectation of support if \nyou do not but foresee this possibility. This scenario \nsuggests that policy investments can have indirect \neffects on decision-making and experience at the \nindividual level, possibly encouraging entry into \nthe labor force. Providing access to labor force \nentry can in turn reduce the worst consequences of \nlow SES for mental health. Within-country differ-\nences reflect the effects of changes in policy within \ncountries over time. In this case, recent changes in \naccess options may be as important as the level of \nsupport, following the hypothesis that changes in \nresources can affect well-being as much as or even \nmore than the level of resources (Davis 1984). In \nour approach, both level of and changes in spending \nare separated and captured as distinct possibilities. \nAn important issue this allows for is the case where \npolicy patterns between and within countries trend \nin opposite directions, resulting in conflated and \ncounteracting forces that could mask the effects of \neach. This could happen, for example, when coun-\ntries with historically high levels of spending effort \nin a particular social policy\u2014resulting in a weaker \nassociation between SES and mental health\u2014begin \nspending less over time, thus reactivating the link \nbetween SES and mental health.\nWe argue that distinguishing cumulative histori-\ncal levels of spending from within-country trends \nprovides a more holistic and thus complete theoreti-\ncal understanding of how policy processes may \nimpact the relationship between individual social \nstatus (or position) and mental health. Our approach \nalso implies that the impact of policy on mental \nhealth is not ahistorical. It is important to resolve \nwhether the current state of linkages between social \nstatuses and mental health depend on current fund-\ning or the historical carry-forward of the norms set \nin place by average levels of funding through time.\nAs such, we seek to add to the institutional per-\nspective on the role of social policy in the associa-\ntion between SES and mental health by asking: \nWhat policies are capable of weakening that  \nassociation\u2014and thus of reducing the mental health \nrisk for lower SES groups\u2014both historically \nbetween countries and via trends in spending over \ntime within countries? In general, we specify these \neffects as different cross-level interactions in a \nthree-level multilevel model using an approach sug-\ngested by Bell and Jones (2015) and Giesselmann \nand Schmidt-Catran (2019).\nIn answering this question, we focus on both the \nsocial protection and social investment policy para-\ndigms and their impacts on SES inequalities. First, \nsocial protection policies seek to mitigate the \nunforeseen and adverse life circumstances accom-\npanying disadvantaged statuses by promoting a \nminimum standard of living (Titmuss 1974; \nWilensky 1975). The main mechanism of social \nprotection policies is the transfer of income to indi-\nviduals or households through either the traditional \n\n338\t\nJournal of Health and Social Behavior 64(3)\nfocal policies of the paradigm, such as pensions, \nsickness, and unemployment benefits (Esping-\nAndersen 1990), or the expanded policies of trans-\nfers to families and welfare services, such as health \ncare and housing (Castles 2008; Esping-Andersen \nand Myles 2009). Second, the social investment \nparadigm focuses on policies that promote human \ncapital formation and increased household labor \nmarket participation (Esping-Andersen et al. 2002; \nHemerijck 2017). Focal policies include early edu-\ncation and care programs (ECEC), general educa-\ntion, postsecondary education and vocational \ntraining, active labor market policies (ALMPs), and \nold age care (Kvist 2015). Both perspectives ulti-\nmately focus on methods by which welfare states \nreduce inequality and poverty (Brady 2009; Plavgo \nand Hemerijck 2021) but highlight different path-\nways to achieve these outcomes. We argue that the \ndifferences in the mechanisms involved in these \npathways may have differential consequences for \nthe link between SES and mental health. Social \ninvestment policies focus on individual-level pre-\nvention of pathways leading to inequality and pov-\nerty, such as poor education and labor market \nattachment. Social protection policies reflect reac-\ntive policies and are more responsive to immediate \nneeds, for example, by transferring income or the \nprovision of services when individuals cannot meet \ntheir needs from the labor market.\nWe focus on depression as the mental health \noutcome. Even though we necessarily focus on a \nsingle measure, variations in depression in a popu-\nlation remain a central marker of differences in the \nmental health of populations overall given its prev-\nalence, its seriousness and chronicity as a problem, \nand its overlap with associated mental health indi-\ncators, such as anxiety (Kessler et al. 1995, 2005; \nMirowsky and Ross 2003). By studying depression, \nwe also focus on the state\u2019s role in the alleviation of \nsuffering rather than the maximization of well-\nbeing, which may be more volatile in general popu-\nlations (Diener and Diener 1996) and also depends \nmore heavily on interpersonal rather than macro-\nsocial inputs (Diener et al. 1999; Veenhoven 2000).\nWe assess each component of SES on mental \nhealth both across general policy processes of the \nwelfare state and by disaggregating each of the pro-\ncesses to determine first, which components of SES \nare affected most and second, which kinds of spe-\ncific policies have the most significant impact. As \nsuch, we argue our study makes three contributions \nto the research tradition linking SES to mental \nhealth. First, it assesses whether the relationship \nbetween SES and depression varies significantly \nacross and within 21 modern welfare states. \nSecond, this study specifies and tests some of the \ninstitutional policy mechanisms (indicated by \nsocial investment vs. social protection spending \neffort) utilizing cross-level interactions that allow \nfor the detection of differential policy effects on the \nrelationship between SES and depression and that \nare beyond what can be considered in a main effects \nanalysis of SES differences (Bell and Jones 2015). \nThird, we disaggregate specific policies in these \npolicy paradigms to specify better which areas of \nthe welfare state matter most in reducing SES-\nbased disparities in mental health. Overall, our \nresults suggest that greater social policy spending \neffort provides \u201ccumulative\u201d mental health advan-\ntages (Thoits 2010) at the individual level.\nBackground\nSES and Mental Health at the \nIndividual Level\nEducation is argued to be the primary causal force \nlinking SES to mental health across the life course \n(Mirowsky and Ross 2007). A central role of educa-\ntion in the production of mental health is the social-\nization of beliefs and skills that together result in \nhigher levels of important personal coping \nresources, such as cognitive flexibility or a sense of \ncontrol over life outcomes (Ross and Mirowsky \n2010; Wheaton 1983). Those with high education \ntend to have more complex jobs, resulting in more \nintellectual flexibility (Kohn and Schooler 1982). \nEducation also has indirect effects through its effect \non economic resources, such as workplace authority \nand income, that help maintain better mental health \n(Kessler 1982; Turner, Wheaton, and Lloyd 1995). \nHowever, the mental health benefits to education \nare not limitless. There are potentially diminishing \nreturns to education on mental health among the \nhighly educated, who receive low economic returns \nrelative to their educational attainment (Bracke, Van \nDe Straat, and Missinne 2014).\nWhereas increasing education predominantly \nplaces individuals in advantageous life circum-\nstances beneficial for mental health (Miech and \nShanahan 2000; Ross and Mirowsky 2010), occupa-\ntion (Kohn et al. 1990) and income (Lynch, Kaplan, \nand Shema 1997) also may have independent effects \non mental health. Occupational status, for example, \nreflects structural job conditions that have both \ndirect and indirect effects on mental health (Kohn \nand Schooler 1973; Link, Lennon, and Dohrenwend \n1993), including the mental health advantages of \n\nParbst and Wheaton\t\n339\nmore complex, more creative, and more meaningful \nwork also involving greater control over the terms \nof work relative to lower status jobs that are routine \nand manual in nature (Kohn et al. 1990). Higher \nincomes decrease depression by reducing levels of \nfinancial strain and economic hardship (Mirowsky \nand Ross 2003) but more broadly, by creating choice \nand thus fostering a sense of mastery. Finally, SES \nin general structures the availability of social \nresources individuals may draw on (Thoits 1995), \nparticularly through the effect of income on access \nto social capital (Song 2011).\nCross-National Welfare State Impacts \non SES and Generalized Health\nStudies on the impact of welfare states on SES have \nexamined self-reported health (Bambra et al. 2009; \nBambra, Netuveli, and Eikemo 2010; Dahl and van \nder Wel 2013; Eikemo et al. 2008), the World Health \nOrganization mental well-being index (van der Wel \net al. 2015), and depression (Hansen, Slagsvold, and \nVeenstra 2017; Niedzwiedz et al. 2016). A majority \nof the research in this area utilizes Esping-\nAndersen\u2019s (1990) three regime typology to test dif-\nferences in the welfare state\u2019s impact of SES on \nhealth (Bergqvist, Yngwe, and Lundberg 2013), \nalthough studies focusing on social protection \nspending find greater spending reduces educational \ndifferences in self-rated health (Dahl and van der \nWel 2013) and mental well-being (van der Wel et al. \n2015).\nTwo cross-national studies focus on the relation-\nship between education and depression. Niedzwiedz \net al. (2016) find a moderating effect of active labor \nmarket spending on the association between levels \nof education and depression. However, evidence for \nthis conclusion can be questioned because they do \nnot consider the random (unmeasured) effects of \ncountries or time due to the pooling of waves in \ntheir sample, thus potentially conflating between- \nand within-country processes. Finally, Hansen et al. \n(2017) utilized separate ordinary least sqaures \nregressions to calculate the odds of the risk of \ndepression for seniors (age 60\u201380) for 10 countries. \nAlthough not testing any policies directly, they \nobserve smaller differences in the effect of educa-\ntion on depression in socially democratic welfare \nstates and the largest within post-Soviet states.\nResearch Questions\nWe ask two main research questions and consider \nthe relative importance of social investment and \nsocial protection in each of them. Our first research \nquestion addresses considerations of the average \ndifferences between countries and their effects on \nmental health: Are enduring differences in the \nspending levels between countries associated with a \nweakening of the association between SES and \nmental health?\nThis question emphasizes the importance of \nconsidering average policy climates across time. \nPolicy levels are relatively stable over short periods \ndue to \u201cpath dependency\u201d logic where removal or \nreduction of benefits is contested (Pierson 2000). \nWe expect to find enduring differences between \ncountries that continuously spend higher versus \nlower amounts on social investment and social pro-\ntection spending. We argue that this expectation \nreflects a \u201ccumulative advantage\u201d for mental health, \nespecially among individuals in societies where \ncontinuous investment at high levels promotes \noverall relative structural advantages that individu-\nals may benefit from, especially for those with low \nstatus. Whereas changes over time in a country test \nthe impact of the direction and magnitude of social \npolicy effects, the average context between coun-\ntries assesses differences due to generalized differ-\nences in the levels of social policy that persist \nthrough time (Thoits 2010).\nOur second question focuses on the impact of the \ndirection and magnitude of changes in social policy \nover time: Do changes in social investment and \nsocial protection policies help account for variation \nin the inverse association between SES and mental \nhealth in countries over time? In assessing both \nquestions, we assess which of our two types of poli-\ncies are most relevant to the weakening of the link \nbetween SES and mental health and the specific \npolicies that may be doing so. Our data cover a rela-\ntively short historical period (2006\u20132014). As a \nresult, this may hamper our ability to detect the full \nimpact of changing policies, especially because path \ndependency logic suggests that change is incremen-\ntal and takes time. Nevertheless, we contend that \nexplicitly separating between- and within-country \ntrends is important both conceptually and empiri-\ncally (Bell, Fairbrother, and Jones 2019). This ques-\ntion parallels the classic status versus change issue \nin the literature on general well-being: Is it level of \nincome or changes in income that are most relevant \nto well-being in general (Davis 1984)?\nBoth questions we address imply that some \ncombination of social investment and social protec-\ntion spending will modify the SES mental health \nassociation. We assess this by examining the mod-\nerating impact of social investment and social \n\n340\t\nJournal of Health and Social Behavior 64(3)\nprotection spending and their component policies \nfor each component of SES on depression.\nData and Methods\nIndividual data were taken from the 2006, 2012, and \n2014 survey rounds of the European Social Survey \n(ESS) of representative samples of adults (restricted \nto 18 and above). Rounds had to be specifically cho-\nsen based on the availability of depression mea-\nsures. Rounds varied in sample size from 35,210 to \n42,351, and typical within-country samples per year \nvaried from 949 to 2,925 and were combined for a \ntotal pooled sample of 113,078. Included in this \nstudy were 21 countries: Austria, Belgium, Czech \nRepublic, Denmark, Estonia, Finland, France, \nGermany, Ireland, Israel, Lithuania, Netherlands, \nNorway, Poland, Portugal, Slovakia, Slovenia, \nSpain, Sweden, Switzerland, and the United \nKingdom. Countries excluded from this analysis \nwere Bulgaria, Cyprus, Hungary, Iceland, Ukraine, \nand the Russian Federation, either due to the \nunavailability or quality of welfare state data or \ninclusion in the ongoing ESS at only one point in \ntime. Using listwise deletion, we obtained a total \nanalytic sample of 83,091, with attrition occurring \nmainly from occupation (6,588) and income mea-\nsures (16,496).\nDependent Variable\nDepression was an averaged index of eight survey \nitems taken from the widely used Center for \nEpidemiologic Studies Depression Scale (CES-D; \nVan de Velde et al. 2010). The eight CES-D scale \nitems used here load consistently on a single latent \nconstruct and behave similarly across countries and \ndemographic groups (Van de Velde et al. 2010). We \nadjusted averaged item scores to range from 0, or no \nweekly symptoms of depression, to 24, which repre-\nsents reporting the most severe response on every \nsurvey item.1 Each CES-D item question asked \nrespondents to report on the last week only. Question \nitems included were: felt depressed, felt everything \nyou did was an effort, your sleep was restless, you \nwere happy, you felt lonely, you enjoyed life, you \nfelt sad, and you could not get going. Depression \nhad a Cronbach\u2019s alpha of .84.\nSES\nEducation was measured in years ranging from 0 to \n25.2 We included a squared education term to account \nfor diminishing returns to education (Bracke et al. \n2014). Household occupational status translated \nISCO-88 occupational codes in the 2006 survey \nwave and ISCO-08 occupational codes in 2012  \nand 2014 surveys into the International Socio-\nEconomic Index of Occupational Status (ISEI) scale \n(Ganzeboom and Treiman 1996) utilizing the updated \nISEI-08 coding (Ganzeboom 2010). The ISEI index \nwas applied to the respondent\u2019s reported occupation \nand, if indicated, their partner\u2019s occupation returning \nthe highest score ranging from 1.10 to 8.90. Finally, \nhousehold income was calculated in deciles for both \nsurvey years and ranges from 1 to 10.3\nState-Level Moderators\nFollowing Kuitto (2016), social investment was a \nmeasure combining four areas of welfare state \nspending reported as a percentage of total gross \ndomestic product (GDP): education, ECEC, ALMP, \nand old age care. Education spending was extracted \nfrom the World Bank (2021).4 ECEC, ALMP, and \nold age care were extracted from the SOCX \nOrganization for Economic Cooperation and \nDevelopment (OECD) spending database (OECD \n2021).5 Also following Kuitto (2016), we measured \nsocial protection spending as a percentage of GDP \nof total cash transfers spending on incapacity, old \nage, survivors, unemployment, and other social \nexclusion spending, taken from the SOCX database \n(OECD 2021).6 However, we also included service \nspending on health care and housing in this cate-\ngory, also taken from the SOCX database, due to its \neffects on poverty and inequality and thus effects on \na minimum standard of living (Esping-Andersen \nand Myles 2009). We present the disaggregation of \neach policy paradigm in Table 1.\nControls\nControls in this analysis included confounders not \non the causal pathway between individual SES and \ndepression, such as age and gender (Mirowsky and \nRoss 2003) and father\u2019s educational status, to incor-\nporate the intergenerational transmission of SES. At \nthe country level, we controlled for economic devel-\nopment as measured by GDP per capita based on \npurchasing power parity (divided by 1,000) in con-\nstant 2011 international dollars taken from the \nWorld Bank (2021) and income inequality, as mea-\nsured by the GINI coefficient for household market \nincome, ranging from 0 to 1 where 1 equals com-\nplete inequality (Solt 2019).7 Both country mea-\nsures impact psychological well-being (Wilkinson \nand Pickett 2010). Finally, we estimated the effects \n\nParbst and Wheaton\t\n341\nof each category of spending, controlling for the \nother category in the same model. Thus, for exam-\nple, if we estimated the effect of social investment in \ngeneral or a component of social investment (educa-\ntion spending), this was in the presence of the asso-\nciated effects of social protection.\nFinally, we considered each ESS round as a con-\ntrol allowing for the observation of secular changes \nin our dependent variable across time.\nAnalysis\nWe employed three-level hierarchical linear models \n(HLMs) to answer our research questions, incorpo-\nrating both between-country averaged differences \nover time and within-country trends. In these mod-\nels, the individual was Level 1, the country-year \nwas Level 2, and the country was Level 3 \n(Giesselmann and Schmidt-Catran 2019). Our ana-\nlytic strategy addressed core issues in using multiple \ncross-sectional, cross-national studies to estimate \ndifferences between countries and over time within \ncountries using pooled time-series cross-sectional \ndata structures and small number of countries \n(Schmidt-Catran, Fairbrother, and Andre\u00df 2019). \nWe followed the approach of Giesselmann and \nSchmidt-Catran (2019) to estimate our cross-level \ninteractions between spending and SES but \nextended their model to include a random coeffi-\ncient for SES at both the country (Level 3) and \ncountry-year levels (Level 2) utilizing a within\u2013\nbetween coding structure (Bell and Jones 2015). We \ndid so because first, it is consequential for estimates \nof Level 3 variables in the model (Heisig, Schaeffer, \nand Giesecke 2017) and second, to account for \nunobserved differences in the effect of SES on \ndepression both between and within countries (Bell \net al. 2019). The use of the three-level model along \nwith the between and within coding in HLM allowed \nfor the simultaneous estimation of both effects.\nOur HLM strategy also accounted for small \ncluster sizes (Level 3 N) by utilizing a Satterthwaite \napproximation following Elff et al. (2021) given \nour number of countries was 21. All models used \nESS poststratification weight. We describe a full \naccounting for the model structure in Appendix A in \nthe online version of the article.\nIn our models, we distinguished \u201cbetween\u201d \nfrom \u201cwithin\u201d effects. This distinction had a some-\nwhat different meaning applied to state-level policy \ndifferences versus individual measures of SES. At \nthe state level, the between measure reflected the \naverage level of spending over time at the country \nlevel, while the within measure reflected changes in \nlevels of spending in each country over time. At the \nindividual level, the between measure of SES in \neffect controlled for the average level of SES over \ntime in each country, while the within measure \nreflected the relative position of individuals on SES \nrelative to that mean. Including the between effect \nof SES controlled for the compositional mean of \nSES across countries. Individual differences in SES \nwere captured by the within measure\u2014analogous \nto a group-mean centered approach at the individ-\nual level, but with the group mean included in the \nmodel (Hox 2010).\nRobustness Checks\nFollowing Schmidt-Catran et al. (2019) and Van der \nMeer, Te Grotenhuis, and Pelzer (2010), we con-\nducted an outlier analysis and report the Cook\u2019s D in \nAppendix B in the online version of the article. All \ncountries with a Cook\u2019s D value higher than 4/N \nwere removed from analysis to examine their influ-\nence on model results, and any model modifications \nare noted in tables. Outlier model results are avail-\nable by request.\nResults\nDescriptive Overview\nDescriptive information in Table 2 highlights the \ngeneral trends of the combined sample. Beginning \nwith individual-level variables, depressive symp-\ntoms remain constant through Round 3 and Round 6 \nat 5.61 symptoms per week and decrease to 5.25 in \nRound 7 with relatively consistent variation across \nwaves. Turning to controls, over half of the sample \nis female in all three waves. From the 2006 to 2014 \nTable 1.\u2002 Social Investment and Social Protection \nPolicy Paradigm Overview.\nPolicy Paradigm\nDisaggregated Policies\nSocial \ninvestment\nEarly childhood education and \ncare (ECEC)\nEducation\nActive labor market policies\nOld age care\nSocial \nprotection\nFamily transfers\nUnemployment transfers\nOld age transfers\nSurvivors transfers\nIncapacity transfers\nHealth\nHousing\n\n342\nTable 2.\u2002 Descriptive Overview of European Social Survey Rounds 3, 6, and 7.\nRound 3 (N = 24,702)\nRound 6 (N = 29,950)\nRound 7 (N = 28,439)\n\u2002\nMean or % (SD)\nRange\nMean or % (SD)\nRange\nMean or % (SD)\nRange\nDependent variable\nDepression\n5.61 (4.01)\n0\u201324\n5.60 (4.12)\n0\u201324\n5.24 (3.96)\n0\u201324\nIndividual-level variables\nControls\nFemale\n52%\n52%\n52%\n\u2002\nFather\u2019s level of education\nAbove secondary\n15%\n19%\n19%\n\u2002\nAge\n48.82 (16.90)\n18\u201397\n5.44 (17.09)\n18\u2013102\n5.83 (17.23)\n18\u2013114\nFocal independent variables\nEducation (years)\n12.67 (4.12)\n0\u201325\n12.99 (3.98)\n0\u201325\n13.18 (3.93)\n0\u201325\nHousehold occupation\n4.68 (2.18)\n1.10\u20138.90\n4.63 (2.20)\n1.10\u20138.90\n4.74 (2.22)\n1.10\u20138.90\nHousehold income (deciles)\n5.61 (2.45)\n1\u201310\n5.31 (2.77)\n1\u201310\n5.35 (2.75)\n1\u201310\nCountry-level variables\nControls\nGross domestic product (thousands)\n39.85 (10.77)\n18.27\u201363.85\n37.57 (10.23)\n23.22\u201363.00\n40 (9.43)\n25.31\u201363.42\nGini\n.47 (.04)\n.39\u2013.52\n.48 (.04)\n.37\u2013.52\n.47 (.04)\n.37\u2013.52\nFocal moderators\nSocial protection\n18.47 (3.92)\n11.91\u201325.45\n19.73 (4.43)\n13.30\u201328.92\n2.06 (4.41)\n13.23\u201328.97\nFamily cash\n1.03 (.48)\n.32\u20132.08\n.98 (.61)\n.20\u20132.43\n.98 (.56)\n.18\u20132.10\nUnemployment\n1.01 (.76)\n0\u20133.16\n1.09 (.98)\n0\u20133.24\n1.05 (.88)\n0\u20133.24\nOld age cash\n6.67 (2.13)\n2.40\u201310.33\n7.34 (2.09)\n4.20\u201312.20\n7.78 (2.37)\n4.12\u201312.25\nHealth\n5.48 (1.44)\n2.45\u20137.92\n5.96 (1.54)\n2.95\u20138.72\n6.13 (1.60)\n2.91\u20138.81\nSurvivors cash\n1.04 (.76)\n0\u20132.16\n.91 (.76)\n.01\u20132.34\n.96 (.75)\n.01\u20132.34\nIncapacity cash\n2.16 (1.04)\n1.22\u20133.42\n2.01 (.49)\n1.53\u20133.31\n2.00 (.54)\n1.28\u20133.29\nHousing\n.43 (.44)\n0\u20131.78\n.54 (.73)\n0\u20133.32\n.41 (.39)\n0\u20131.54\nSocial investment\n7.29 (1.95)\n5.03\u201312.31\n7.57 (2.31)\n4.89\u201312.90\n7.67 (2.31)\n5.01\u201312.93\nEarly education and care programs\n.64 (.35)\n.25\u20131.33\n.75 (.37)\n.32\u20131.58\n.78 (.36)\n.38\u20131.64\nEducation\n5.37 (.90)\n3.72\u20137.73\n5.62 (1.11)\n3.91\u20137.66\n5.63 (1.12)\n3.99\u20137.67\nActive labor market policies\n.69 (.33)\n.07\u20131.46\n.65 (.42)\n.14\u20131.93\n.69 (.43)\n.13\u20132.03\nOld age care\n.56 (.65)\n0\u20132.13\n.54 (.65)\n0\u20132.16\n.57 (.66)\n0\u20132.18\n\nParbst and Wheaton\t\n343\nwaves, there is an increasing percentage of individ-\nuals whose father\u2019s education is above secondary, \nranging from 15% in Round 3 to 19% in Round 7. \nThe average age increased slightly, from nearly 49 \nyears to around 51 years. Finally, SES trends show \nboth education and occupation slightly increasing \nover time while income levels marginally decreased \non average.\nCountry-level controls are mostly stable across \nwaves except for GDP per capita. GDP per capita \ndeclined in Round 6 and then increased beyond \nRound 3 levels in Round 7. Gini remained rela-\ntively constant across waves.\nSocial protection and investment spending \nincreased across this time, and a nontrivial amount \nof spending differences exists across areas. \nComparatively, states spend more on social protec-\ntion spending on average. Social protection spend-\ning also generally has the larger average increase \nacross waves, with a proportionate increase of \nabout 2% of GDP. In social protection spending, the \ntwo largest areas of spending are old age cash ben-\nefits and health. These two areas have driven the \noverall increase in social protection because all \nother programs remain relatively constant. Social \ninvestment only marginally increased, with less \nthan a 0.5% increase over this period due to mar-\nginal increases in education and ECEC. All within-\ncountry trends are in Appendix C in the online \nversion of the article.\nEstimating the Role of State-Level Policies in \nthe SES\u2013Depression Relationship\nTable 3 examines the relationship between the three \nindicators of SES and depression. Random coeffi-\ncient models are presented to reflect the possibility \nthat the effect of SES varies across countries, based \non unspecified country-level differences. The test of \nthe random variation of coefficients for SES is a \nlogical first step that signals the possibility, or not, \nof specific cross-level interactions. We also include \na country-year random slope that acknowledges \nwithin-country trends.\nModels for the fixed effects of SES, including \nthe random effects noted previously, are tested in \nModels 1 through 3\u2014one each for each component \nof SES. In Models 1, 2, and 3, we build a model for \nthe fixed effects of SES by considering the standard \ncausal sequence for education, occupation, and \nincome (Sewell, Hauser, and Featherman 1976). \nThus, in the model testing the fixed and random \neffects of occupation, we control for education and \ninclude a random effect for occupation (Model 2), \nand in the model from income, we control for both \neducation and occupation and include a random \neffect for income (Model 3). The fixed effects in \nthese models include controls at the individual level \nfor the survey round,8 gender, age, and father\u2019s \neducation.\nOur focus in Model 1 is on the within and \nbetween effects of education and the squared term \nfor the possibility of diminishing returns of educa-\ntion. The focus of our research questions is on the \nwithin measures at the individual level, while the \nbetween measures capture aggregate differences in \naverage SES only but must be controlled. Table 3 \nshows that the within differences in education are \nnegatively related to depression, while the between \ndifferences are not. This is consistent with expecta-\ntion. State-level differences in SES are likely to be \nsmaller than individual differences within coun-\ntries. The squared education term at the individual \nlevel is also significant in the expected positive \ndirection, indicating a diminishing return to high \nlevels of education. This model also shows a sig-\nnificant random slope component of the effect of \neducation between countries and within countries, \nsignaling that this association does vary across \ncountries, and this may signal important variation \nin the effect of SES due to state-level differences. \nFinally, the random effect for the squared education \nterm is only significant at the between-country \nlevel.\nModels 2 and 3 add occupational status and \nincome progressively to the education model. The \neffect of occupational status before adding income is \nnotably larger than in the final model\u2014reflecting \npartial mediation by the effect of income\u2014and \nagain, the random coefficients in this model are sig-\nnificant, in this case, both between and within coun-\ntries. Model 3 also indicates that there is a between \neffect of occupation. Substantively, this indicates that \ncountries that generally have a higher level of house-\nhold occupational status decrease depression at the \nindividual level. Finally, adding income in Model 3 \npresents the full model with SES fixed and random \neffects and demonstrates that income also has signif-\nicant fixed and random between and within effects. \nOnly individual-level differences in income are \nrelated to depression: Between-country differences \nin average income over time have no effect.\nAll SES components in this final model have \ndistinct negative associations with depression\u2014no \none component absorbs the effects of other compo-\nnents. This finding is important because analyses \nonly using one or two components will underesti-\nmate the total effect of SES. Overall, Table 3 \n\n344\t\nJournal of Health and Social Behavior 64(3)\nTable 3.\u2002 Random Coefficient Models\u2019 Fixed and Random Effects for Socioeconomic Status Predicting \nDepression (N = 83,091).\n(1) Education\n(2) Household \nOccupation\n(3) Household Income\nFixed components\nIntercept\n8.248\n(5.614)\n9.116\n(5.593)\n8.149\n(5.806)\nControls\n\u2003 Round 6\n\u2212.255**\n(.079)\n\u2212.297***\n(.070)\n\u2212.242*\n(.102)\n\u2003 Round 7\n\u2212.424***\n(.081)\n\u2212.488***\n(.071)\n\u2212.414**\n(.104)\n\u2003 FemaleW\n.837***\n(.026)\n.837***\n(.026)\n.713***\n(.026)\n\u2003 FemaleB\n4.562\n(5.985)\n8.241\n(6.005)\n10.653\n(6.825)\nFather\u2019s education\n\u2003 Above secondaryW\n.027\n(.037)\n.143***\n(.037)\n.186**\n(.037)\n\u2003 Above secondaryB\n1.729\n(2.514)\n5.264\n(2.739)\n5.396\n(3.234)\n\u2003 AgeW\n.008***\n(.001)\n.013***\n(.001)\n.008***\n(.001)\n\u2003 AgeB\n.013\n(.098)\n\u2212.048\n(.096)\n\u2212.081\n(.117)\nFocal associations\n\u2003 EducationW\n\u2212.436***\n(.062)\n\u2212.289***\n(.016)\n\u2212.247***\n(.015)\n\u2003 EducationB\n\u2212.716\n(.510)\n.485\n(.498)\n.368\n(.526)\n\u2003 EducationW\n2\n.011***\n(.002)\n.009***\n(.001)\n.008***\n(.001)\n\u2003 EducationB\n2\n.019\n(.025)\n\u2212.024\n(.024)\n\u2212.018\n(.025)\n\u2003 OccupationW\n\u2212.243***\n(.019)\n\u2212.123**\n(.008)\n\u2003 OccupationB\n\u22121.765**\n(.538)\n\u22121.481*\n(.565)\n\u2003 IncomeW\n\u2212.288***\n(.020)\n\u2003 IncomeB\n.067\n(.349)\nRandom components\nBetween country\n\u2003 Intercept\n.877*\n(.430)\n.311*\n(.134)\n.386*\n(.202)\n\u2003 Education\n.071**\n(.028)\n\u2002\n\u2003 Education2\n.000*\n(.000)\n\u2002\n\u2003 Occupation\n.006**\n(.002)\n\u2002\n(continued)\n\nParbst and Wheaton\t\n345\n(1) Education\n(2) Household \nOccupation\n(3) Household Income\n\u2003 Income\n.006**\n(.003)\nWithin country\n\u2003 Intercept\n.051***\n(.015)\n.054***\n(.016)\n.087***\n(.025)\n\u2003 Education\n.001*\n(.000)\n\u2002\n\u2003 Occupation\n.002*\n(.001)\n\u2002\n\u2003 Income\n.004**\n(.001)\nNote: All within-country effects are denoted by W, and all between-country effects are denoted by B. Data are from \nRounds 3, 6, and 7 of the European Social Survey.\n*p\u2009<\u2009.05. **p\u2009<\u2009.01. ***p\u2009<\u2009.001 (two-tailed).\nTable 3.\u2002 (continued)\ndemonstrates a consistency with SES findings in \nthe literature and the generality of its association \nwith mental health within countries. At the same \ntime, the significance of the accompanying random \neffects for all components of SES also suggests the \npossibility of cross-level interactions with state-\nlevel policy differences.\nTable 4 presents models that include both indi-\nvidual-level and contextual effects and thus assess \nboth of our research questions. Results are pre-\nsented separately for within versus between compo-\nnents of both SES and spending, although again, \nour focus is on the within effects. Each component \nof SES includes two models: the first testing the \neffect of social investment (SI) and the second test-\ning the effect of social protection (SP) while also \ncontrolling for the other form of spending.\nResults include controls for the contextual \neffects of GDP and Gini. Except for the household \nincome models for SI and SP spending and the edu-\ncation model for SI, the within effect of GDP per \ncapita is negative and significant, indicating that \nincreasing economic development levels across \ntime reduce depression. There are also negative \nbetween effects for GDP per capita across all mod-\nels. This indicates that countries with historically \nhigher GDP per capita also tend to have lower lev-\nels of depression. However, counter to expecta-\ntions, the effect of income inequality\u2014as measured \nby Gini\u2014is not significant at either the between or \nwithin level of analysis.\nThe interaction effects estimated for each SES \ncomponent and each spending measure directly \naddress our first and second research questions. We \nconsider first the interaction between differences in \nSES and spending within countries, primarily to \naddress \npotential \nunobserved \nconfounding \n(Giesselmann and Schmidt-Catran 2019). None of \nthese interactions are significant, as expected. In \nanswering our first research question, we focus \nattention on the interaction effect between levels of \nspending and within-country variation in SES. All \nsignificant interactions in these results involve SI \nspending. The positive interaction term signifies \nthat higher spending counteracts the negative slope \nfor the effect of that component of SES, signifying \na moderated negative slope. For the lone significant \nsquared term for the moderating effect of SI on edu-\ncation, the negative term signifies that SI spending \nreduces the diminishing returns to high education. \nThe most consistent result here is the interaction \nbetween individual differences in SES and overall \nsocial investment differences between countries \nover time. This finding suggests that a general cli-\nmate of social investment does reduce depression \ndifferences between high and low SES at the indi-\nvidual level. In results not displayed, we note that \nall SI models explain away the negative depression \ntrend across waves, suggesting it to be an important \nmechanism for societal well-being.\nRegarding our second question, we fail to detect \nany within-country trends in policy that mitigate \nthe effects of SES across the observed period in \nthese data. We noted earlier the possible problems \nof detecting these sources of influence over delim-\nited time frames.\n\n346\t\nJournal of Health and Social Behavior 64(3)\nTable 4.\u2002 Between and Within Moderating Fixed Effects of Social Investment and Social Protection \nSpending on Socioeconomic Status and Depression (N = 83,091).\nSES Component\nEducation\nHousehold Occupation\nHousehold Income\nWelfare State Policy\nSI\nSP\nSI\nSP\nSI\nSP\nIntercept\n9.133\n(5.365)\n8.990\n(5.335)\n12.242*\n(4.971)\n12.291*\n(4.905)\n12.244\n(5.469)\n12.067\n(5.46)\nIndividual level\nFocal associations\n\u2003 SES Componentw\n\u22121.035***\n(.159)\n\u2212.641\n(.300)\n\u2212.411***\n(.053)\n\u2212.307**\n(.089)\n\u2212.497\n(.052)\n\u2212.311**\n(.099)\n\u2003 SES componentb\n\u2212.303\n(.480)\n\u2212.298\n(.477)\n\u22121.206\n(.576)\n\u22121.201\n(.568)\n.428\n(.350)\n.407\n(.349)\n\u2003 Education2\nw\n.026***\n(.005)\n.015\n(.009)\n.009***\n(.001)\n.009***\n(.001)\n.008***\n(.001)\n.008***\n(.001)\n\u2003 Education2\nb\n.007\n(.024)\n.007\n(.024)\n\u2212.000\n(.022)\n.001\n(.021)\n.007\n(.025)\n.006\n(.025)\nContextual level\nControls\n\u2003 GDP per capitaw\n\u2212.087\n(.043)\n\u2212.097*\n(.045)\n\u2212.091*\n(.042)\n\u2212.090*\n(.041)\n.020\n(.065)\n.006\n(.065)\n\u2003 GDP per capitab\n\u2212.067**\n(.016)\n\u2212.068**\n(.016)\n\u2212.053**\n(.015)\n\u2212.053**\n(.015)\n\u2212.051*\n(.017)\n\u2212.051*\n(.017)\n\u2003 GINIw\n5.092\n(3.121)\n4.868\n(3.140)\n3.716\n(3.025)\n3.332\n(2.996)\n3.934\n(4.634)\n3.623\n(4.733)\n\u2003 GINIb\n1.561\n(3.330)\n1.551\n(3.311)\n\u22121.155\n(3.334)\n\u22121.135\n(3.290)\n.259\n(3.789)\n.118\n(3.783)\nFocal moderators\n\u2003 Social investmentw\n\u22121.676\n(2.182)\n\u2212.038\n(.111)\n\u2212.304\n(1.559)\n.004\n(.098)\n\u22121.388\n(1.027)\n\u2212.120\n(.153)\n\u2003 Social investmentb\n\u2212.107\n(.072)\n\u2212.118\n(.070)\n\u2212.084\n(.067)\n\u2212.104\n(.065)\n\u2212.138\n(.076)\n\u2212.119\n(.076)\n\u2003 Social protectionw\n\u2212.029\n(.040)\n\u2212.248\n(.294)\n\u2212.040\n(.044)\n.129\n(.274)\n.015\n(.068)\n\u2212.041\n(.257)\n\u2003 Social protectionb\n.022\n(.037)\n.028\n(.037)\n.004\n(.033)\n.007\n(.033)\n.012\n(.037)\n.011\n(.037)\nInteractions\n\u2003 SESb \u00d7 Policyw\n.128\n(.166)\n.017\n(.022)\n.068\n(.323)\n\u2212.036\n(.061)\n.220\n(.180)\n.002\n(.050)\n\u2003 SESw \u00d7 Policyb\n.082**\n(.021)\n.011\n(.015)\n.023**\n(.007)\n.003\n(.005)\n.029***\n(.007)\n.001\n(.005)\n\u2003 Ed2\nw \u00d7 Policyb\n\u2212.002*\n(.001)\n.000\n(.000)\n\u2002\n\u2003 SESw \u00d7 Policyw\n\u2212.007\n(.012)\n\u2212.002\n(.004)\n\u2212.023\n(.019)\n\u2212.007\n(.006)\n.025\n(.023)\n.001\n(.007)\nNote: All within-country effects are denoted by w, and all between-country effects are denoted by b. Data are from \nRounds 3, 6, and 7 of the European Social Survey. SI\u2009=\u2009social investment; SP\u2009=\u2009social protection; SES\u2009=\u2009socioeconomic \nstatus; GDP\u2009=\u2009gross domestic product; Ed\u2009=\u2009education.\n*p\u2009<\u2009.05. **p\u2009<\u2009.01. ***p\u2009<\u2009.001 (two-tailed).\n\nParbst and Wheaton\t\n347\nWe display the moderating between-country \neffect of social investment spending on the impact \nof SES in Figure 1. This figure displays predicted \nvalues of depression by levels of education, occu-\npation, and household income, respectively, at three \nlevels of SI spending\u2014low, average, and high, \nwhere \u201clow\u201d and \u201chigh\u201d represent 1 SD below and \nabove the mean of all countries and average is the \nmean for all countries. Each component of SES is \ncountry-mean centered with standard deviation \nunits above and below the mean on the x-axis. \nExamining differences between the three levels of \nspending demonstrates a clear pattern that general-\nizes across all components of SES: The slope for \nthe effect of SES is weaker as the level of social \ninvestment increases. It is important to note that \nthis figure especially makes clear the fact that indi-\nviduals face different levels of depression at the \nsame level of low status depending on the country \nin which they live. Differences at high levels of \nSES are not affected, and specifically, depression \nlevels are not worse due to state investments at high \nlevels of status. Put another way, the impact of \nsocial policy here is where it is intended to be and \ndoes not degrade mental health levels of the advan-\ntaged. Generally, this finding suggests that the con-\nsequences of fundamental individual differences in \nsocial status depend in part on social context.\nFinally, we turn to Table 5 to assess our two \nresearch questions across disaggregated social \ninvestment and protection policies. We present a \ntable indicating whether between- and within- \npolicy differences interactions are significant for \nthe effect of SES.9 Each significant interaction is \nnoted by direction (\u00b1) and level of significance. In \nSI policies, we find a consistent pattern of overall \nbetween-country SI spending moderating the \neffects of education, occupation, and income. There \nare two exceptions: ALMPs do not reduce the posi-\ntive slope for education squared and between ECEC \ndoes not moderate the effect of household occupa-\ntion. Social protection has much more limited influ-\nence on the SES\u2013depression association: Only \nincapacity spending moderates the effect of house-\nhold income on depression.\nOut of 24 possible models involving social pro-\ntection spending, only one has significant interac-\ntion effects in comparison to nearly all models of \nsocial investment. These findings conform to our \nexpectations that living in countries with higher \nlevels of certain types of spending bestows a mental \nhealth advantage for those who live there, specifi-\ncally through the reduction of status differences in \nmental health, and that SI spending specifically is \nmore consequential for this association across the \nSES spectrum.\nDiscussion and \nConclusion\nOur study contributes to the foundational history of \nresearch on the association between SES and mental \nhealth by incorporating an institutional perspective \nof social investment and social protection spending \nand applying it between and within countries cross-\nnationally across time. This builds on earlier cross-\nnational studies (Niedzwiedz et al. 2016; Pr\u00e4g et al. \n2016; Rai et al. 2013) using an analytic strategy that \nexplicitly accounts for between- and within-country \ntrends across time. In doing so, we find a significant \nFigure 1.\u2002 Between-Country Moderating Effect of Social Investment Spending on Education, \nOccupation, and Income on Depression.\nNote: Data are from Rounds 3, 6, and 7 of the European Social Survey.\n\n348\t\nJournal of Health and Social Behavior 64(3)\nwithin effect of each component of SES on depres-\nsion at the individual level, adding to evidence con-\nsistent with a social causation interpretation of this \nassociation (Gangl 2010) outside of the North \nAmerican context, while also finding significant \npolicy effects that alter this association.\nIn answering our first research question, we find \nevidence to support the notion that SI spending is \nmore consequential for the SES\u2013mental health \nassociation. Both generally and in terms of specific \ndisaggregated policies, SI reduces mental health \ninequities between low and high status across all \ncomponents of SES. Of particular importance is the \neffect of SI on the education and depression link. \nWe find\u2014in addition to a general mitigation of the \nimplications of status differences\u2014an impact of SI \nthat also linearizes the effect of education and thus \npotentially extends its impact to both low and high \nstatus. This finding contributes to other findings on \nthe nonlinear effects of education in Europe \n(Bracke et al. 2014). Our findings also are consis-\ntent with Niedzwiedz et al.\u2019s (2016) finding of \nALMPs reducing educational inequalities in \ndepression, but we generalize this finding to other \nforms of SI and across components of SES. Social \nprotection has a much more limited impact. The \nlone significant SP policy that we find is the moder-\nating effect of incapacity cash transfers on the asso-\nciation between household income and depression. \nNevertheless, we demonstrate that more generous \nwelfare states matter for reducing mental health \ninequalities and argue that this article is one of the \nfirst to find direct evidence for Thoits\u2019s (2010) con-\nception of cumulative advantage at the macro level \nwhere individuals who live in high spending coun-\ntries enjoy mental health advantages. Our evidence \nalso points to the importance of the timing of policy \ninterventions in the life course because SI policies \nare typically more focused on formative stages of \nthe life course and less on later life.\nOur second research question focusing on \nwithin-country change in policies yielded little evi-\ndence for the cross-level effects of policy. These \nfindings are in fact consistent with other within-\ncountry studies on mental health that observe rela-\ntively small changes, both positive and negative, in \nthe level of spending on mental health outcomes in \nthe population (Lee and Wolf 2014; Pak 2020). For \nexample, over the years we observe in this study, \nspending remains relatively stable across most \nspending areas in both SI and SP. However, one \nbenefit of our model specification is that we were \nTable 5.\u2002 Disaggregation of Focal Between (B) and Within (W) Interaction Fixed Effects on Within \nEffects of Socioeconomic Status (N = 83,091).\nSES Component\nEducationw\nEducationw\n2\nHousehold \nOccupationw\nHousehold Incomew\nBetween and within\nB\nW\nB\nW\nB\nW\nB\nW\nSocial investment\n\u2003 ECEC\nX+**\nX\u2013**\nX+***\n\u2002\n\u2003 Education\nX+**\nX\u2013**\nX+**\nX+**\n\u2002\n\u2003 ALMPs\nX+**\nX+**\nX+*\n\u2002\n\u2003 Old age care\nX+**\nX\u2013*\nX+*\nX+**\n\u2002\nSocial protection\n\u2003 Family cash\n\u2002\n\u2003 Unemployment\n\u2002\n\u2003 Old age cash\n\u2002\n\u2003 Health care\n\u2002\n\u2003 Survivors cash\n\u2002\n\u2003 Incapacity cash\nX+*\n\u2002\n\u2003 Housing\n\u2002\nNote: X marks significant interactions and direction. All within-country effects are denoted by W, and all between-\ncountry effects are denoted by B. Models adjusted for influence include: household income and ECEC (Denmark, \nFinland, Poland), all components of SES and ALMPs (Denmark for education and household income, Poland and \nNorway for household occupation), education and old age care (Poland), and household income and incapacity \n(Denmark and Norway). Data are from Rounds 3, 6, and 7 of the European Social Survey. SES = socioeconomic \nstatus; ECEC = early education and care programs; ALMPs = active labor market policies.\n*p\u2009<\u2009.05. **p\u2009<\u2009.01. ***p\u2009<\u2009.001 (two-tailed).\n\nParbst and Wheaton\t\n349\nable to observe cross-national differences in the \neffect of continuously high versus low levels of \nspending between countries, which demonstrated \nthat SI policies are important for cross-national dif-\nferences. Generally, our null findings suggest \nfocusing on longer periods of time with greater \nspending variation or on specific policies in coun-\ntries that may be more historically volatile.\nOur article indirectly speaks to a core theme in \nthe long literature linking SES to mental health. \nGoing back to at least Faris and Dunham (1939) and \nincluding the classic community studies of the mid-\ntwentieth century (Hollingshead and Redlich 1958) \nand a history of investigations on the causal interpre-\ntation of this relationship (Warren 2009; Wheaton \n1978), the emphasis throughout has been on consis-\ntency of this relationship across time, measures, and \nsamples. This consistency of findings aligns with the \ncentral premises of fundamental cause theory (Link \nand Phelan 1995), with the exception of interest in \nSES variation by gender and race (Jackson and \nWilliams 2006). However, most of this research is \nbased on North American findings.\nOur study suggests that there is variability in the \nstrength of the link between SES and mental health \nnot only at the individual level (Kessler and \nMacRae 1981) but also at the macro level of state \npolicy while not questioning the consistency of this \nassociation overall. Our article adds another layer \nto the discussion of SES differences in mental \nhealth by considering the social context within \nwhich SES differences occur. The link between \nSES and mental health is not exactly immutable or \nunchanging across social contexts: Clearly, it \ndepends on features of social context not usually \nspecified in the traditional literatures in sociology \naddressing this issue.\nOur results also have broader implications for \nsocial policy research in general. By distinguishing \nthe two theoretical pathways of social protection \nand social investment, we demonstrated that social \ninvestment spending more dramatically reduces the \ninverse link between SES and depression. At the \nsame time, we find no systematic evidence suggest-\ning that changes in the direction or magnitude of \nspending (i.e., within-country trends) impact the \nSES\u2013depression association, either due to the gen-\neral or specific domains of either social investment \nor social protection policy. In effect, it is the consis-\ntent historical presence of spending that matters.\nWhile our study answers important questions \nabout cross-national differences in the link between \nSES and mental health, it also raises other questions \nthat need to be answered in further research. A \ncentral question concerns the mechanisms through \nwhich social policies reduce the linkage between \nSES and depression. Given evidence that more gen-\nerous welfare states reduce levels of economic \nstrain (Levecque et al. 2011), a plausible next step \nis to investigate how the role of configurations of \nstress histories and coping resources at the individ-\nual level that play out over the life course\u2014follow-\ning the tradition of the stress process (Pearlin 1989; \nPearlin et al. 1981)\u2014are impacted by social invest-\nment and social protection. There are, in fact, some \nclues in the results in Table 4. Policies that target \nECEC and education have broad effects on reduc-\ning the mental health disadvantages of low educa-\ntion, occupation, and income, suggesting that the \nimpacts of these policies may operate fundamen-\ntally through increased access to educational oppor-\ntunities, which is consequential for both labor force \nplacement and mobility and indirectly for income. \nIn addition, low status individuals may dispropor-\ntionately benefit from ECEC and old age care in \npart because of increased labor force attachment \namong women (Boeckmann, Misra, and Budig \n2015; Costa-Font and Vilaplana-Prieto 2022), in \nturn indirectly increasing household income while \nalso providing women with mental health benefits \n(Kessler and McRae 1981). ALMPs may further \nsupport this earlier advantage. These policy pro-\ncesses may in turn alter the distribution of chronic \nstress exposures over lives (Turner et al. 1995) and \nthe development of a sense of control or mastery \n(Mirowsky and Ross 2007) within countries.\nOur study is not without limitations. Due to \nrestrictions caused by the specific waves of the ESS \nthat included the measurement of depression, we \nhave limited waves of data. Ideally, we would want \nmore than the three waves available for analysis \nand/or more widely spaced waves. For example, it \nis recommended to have at least four or more waves \nto detect the complex within-country effects under \ninvestigation in this article (Giesselmann and \nSchmidt-Catran 2019). In addition, our measures \ndo not deal with variation in policy delivery, \nwhether targeted or universal (Brady and Burroway \n2012), that may more directly affect depression. \nOur measures of policy also do not include mea-\nsures of duration of benefits or their generosity for \neach type of policy (Scruggs and Allan 2008), \nwhich is demonstrated to affect mental well-being \n(Glass, Simon, and Andersson 2016). Finally, varia-\ntion in the effect of spending effort during this \nperiod itself may have been due to other factors \nsuch as the 2008 world financial crisis, an issue that \ncould partially confound observed associations.\n\n350\t\nJournal of Health and Social Behavior 64(3)\nDespite these limitations, our analysis contrib-\nutes theoretically and empirically to the growing \nbody of evidence that suggests social inequality in \nmental health depends on the policy context one \nlives within. Our findings also strongly suggest that \nsocial policy may be another \u201cfundamental cause\u201d \nof health inequalities (Beckfield et al. 2015) in \naddition to previous focuses on individual-level \ndeterminants (Link and Phelan 1995). Our findings \nsuggest that the reduction of human suffering is \nindeed within the control of policymakers, and as \nsuch, considerations of policy outcomes should \nincorporate the understanding that changing policy \nmay also mean changing the balance of suffering \nacross an entire society.\nORCID iDs\nMatthew Parbst \n https://orcid.org/0000-0002-5949-6122\nBlair Wheaton \n https://orcid.org/0000-0002-6705-4594\nSUPPLEMENTAL MATERIAL\nAppendices A through D are available in the online version of \nthe article.\nNotes\n\u2002 1.\t Individuals who had more than four items of the \nindex missing were treated as missing at random \nand deleted.\n\u2002 2.\t Top coded at the 99.5% of the sample distribution \ndue to extreme values.\n\u2002 3.\t Because of different codings of household income \nbetween the 2006 and 2012 and 2014 waves, 2006 \nwas coded into deciles through collapsing income cat-\negories into the appropriate distributional properties. \nThe same procedure is used in Glass et al. (2016).\n\u2002 4.\t Belgium 2012 was extracted from the Eurostat data-\nbase. Britain 2012 is derived by averaging across \n2011 and 2013 data points.\n\u2002 5.\t Data extracted in 2017.\n\u2002 6.\t All data available by request.\n\u2002 7.\t Original scale was divided by 100.\n\u2002 8.\t Survey wave is appropriate in this context com-\npared to each survey year because there is no sig-\nnificant differences between years within each wave \nrespectively.\n\u2002 9.\t Full models are in Appendix D in the online version \nof the article.\nReferences\nAdler, Nancy E., Thomas Boyce, Margaret A. Chesney, \nSheldon Cohen, Susan Folkman, Robert L. Kahn, and \nS. Leonard Syme. 1994. \u201cSocioeconomic Status and \nHealth: The Challenge of the Gradient.\u201d American \nPsychologist 49(1):15\u201324.\nAneshensel, Carol S. 2009. \u201cToward Explaining Mental \nHealth Disparities.\u201d Journal of Health and Social \nBehavior 50(4):377\u201394.\nBambra, Clare, Gopalakrishnan Netuveli, and Terje A. \nEikemo. 2010. \u201cWelfare State Regime Life Courses: \nThe Development of Western European Welfare State \nRegimes and Age-Related Patterns of Educational \nInequalities in Self-Reported Health.\u201d International \nJournal of Health Services 40(3):399\u2013420.\nBambra, Clare, Daniel Pope, Viren Swami, Debbie \nStanistreet, Albert-Jan Roskam, Anton Kunst, \nand Alex Scott-Samuel. 2009. \u201cGender, Health \nInequalities, and Welfare State Regimes: A Cross-\nNational Study of 13 European Countries.\u201d Journal \nof Epidemiology & Community Health 63(1):38\u201344.\nBeckfield, Jason, Clare Bambra, Terje A. Eikemo, Tim \nHuijts, Courtney McNamara, and Claus Wendt. \n2015. \u201cAn Institutional Theory of Welfare State \nEffects on the Distribution of Population Health.\u201d \nSocial Theory & Health 13(3\u20134):227\u201344.\nBell, Andrew, Malcolm Fairbrother, and Kelvyn \nJones. 2019. \u201cFixed and Random Effects Models: \nMaking an Informed Choice.\u201d Quality & Quantity \n53(2):1051\u201374.\nBell, Andrew, and Kelvyn Jones. 2015. \u201cExplaining \nFixed Effects: Random Effects Modeling of Time-\nSeries Cross-Sectional and Panel Data.\u201d Political \nScience Research and Methods 3(1):133\u201353.\nBergqvist, Kersti, Monica \u00c5berg Yngwe, and Olle \nLundberg. 2013. \u201cUnderstanding the Role of Welfare \nState Characteristics for Health and Inequalities\u2014\nAn Analytical Review.\u201d BMC Public Health \n13(1):1234\u201354.\nBoeckmann, Irene, Joya Misra, and Michelle J. Budig. \n2015. \u201cCultural and Institutional Factors Shaping \nMothers\u2019 Employment and Working Hours in \nPostindustrial Countries.\u201d Social Forces 93(4):1301\u2013\n33.\nBracke, Piet, Vera Van De Straat, and Sarah Missinne. \n2014. \u201cEducation, Mental Health, and Education\u2013\nLabor Market Misfit.\u201d Journal of Health and Social \nBehavior 55(4):442\u201359.\nBrady, David. 2009. Rich Democracies, Poor People: \nHow Politics Explain Poverty. Oxford, UK: Oxford \nUniversity Press.\nBrady, \nDavid, \nand \nRebekah \nBurroway. \n2012. \n\u201cTargeting, \nUniversalism, \nand \nSingle-Mother \nPoverty: A Multilevel Analysis across 18 Affluent \nDemocracies.\u201d Demography 49(2):719\u201346.\nCarrino, Ludovico, Karen Glaser, and Mauricio Avendano. \n2020. \u201cLater Retirement, Job Strain, and Health: \nEvidence from the New State Pension Age in the \nUnited Kingdom.\u201d Health Economics 29(8):891\u2013912.\nCastles, Francis G. 2008. \u201cWhat Welfare States Do: A \nDisaggregated Expenditure Approach.\u201d Journal of \nSocial Policy 38(1):45\u201362.\nCosta-Font, Joan, and Cristina Vilaplana-Prieto. 2022. \n\u201cInvesting in Care for Old Age? An Examination \nof Long-Term Care Expenditure Dynamics and Its \nSpillovers.\u201d Empirical Economics 64:1\u201330.\n\nParbst and Wheaton\t\n351\nDahl, Espen, and Kjetil A. van der Wel. 2013. \n\u201cEducational Inequalities in Health in European \nWelfare States: A Social Expenditure Approach.\u201d \nSocial Science & Medicine 81:60\u201369.\nDavis, James A. 1984. \u201cNew Money, an Old Man/Lady, \nand \u2018Two\u2019s Company\u2019: Subjective Welfare in the \nNORC General Social Surveys 1972\u20131982.\u201d Social \nIndicators Research 15(4):319\u2013350.\nDiener, Ed, and Carol Diener. 1996. \u201cMost People Are \nHappy.\u201d Psychological Science 7(3):181\u201385.\nDiener, Ed, Eunkook M. Suh, Richard E. Lucas, and \nHeidi L. Smith. 1999. \u201cSubjective Well-Being: \nThree Decades of Progress.\u201d Psychological Bulletin \n125(2):276\u2013302.\nDohrenwend, Bruce P., and Barbara S. Dohrenwend. \n1969. Social Status and Psychological Disorder: A \nCausal Inquiry. New York, NY: Wiley-Interscience.\nEikemo, Terje A., Clare Bambra, Kerry Joyce, and Espen \nDahl. 2008. \u201cWelfare State Regimes and Income-\nRelated Health Inequalities: A Comparison of 23 \nEuropean Countries.\u201d The European Journal of \nPublic Health 18(6):593\u201399.\nElff, Martin, Jan Paul Heisig, Merlin Schaeffer, and \nSusumu Shikano. 2021. \u201cMultilevel Analysis with \nFew Clusters: Improving Likelihood-Based Methods \nto Provide Unbiased Estimates and Accurate \nInference.\u201d British Journal of Political Science \n51:412\u201326.\nEsping-Andersen, G\u00f8sta. 1990. The Three Worlds of \nWelfare Capitalism. Cambridge, UK: Polity Press.\nEsping-Andersen, \nG\u00f8sta, \nDuncan \nGallie, \nAnton \nHemerijck, and John Myles, eds. 2002. Why We \nNeed a New Welfare State. New York, NY: Oxford \nUniversity Press.\nEsping-Andersen, G\u00f8sta, and John Myles. 2009. \n\u201cEconomic Inequality and the Welfare State.\u201d Pp. \n639\u201344 in The Oxford Handbook of Economic \nInequality, edited by W. Salverda, B. Nolan, and T. \nM. Smeeding. Oxford, UK: Oxford University Press.\nFaris, Robert E. Lee, and Henry Warren Dunham. 1939. \nMental Disorders in Urban Areas: An Ecological \nStudy of Schizophrenia and Other Psychoses. \nChicago, IL: University of Chicago Press.\nGangl, Markus. 2010. \u201cCausal Inference in Sociological \nResearch.\u201d Annual Review of Sociology 36:21\u201347.\nGanzeboom, Harry B. G. 2010. \u201cA New International \nSocio-Economic Index [ISEI] of Occupational \nStatus for the International Standard Classification \nof Occupation 2008 [ISCO-08] Constructed with \nData from the ISSP 2002-2007.\u201d Paper presented at \nAnnual Conference of International Social Survey \nProgramme, Lisbon, Portugal.\nGanzeboom, Harry B. G., and Donald J. Treiman. \n1996. \u201cInternationally Comparable Measures of \nOccupational Status for the 1988 International \nStandard Classification of Occupations.\u201d Social \nScience Research 25(3):201\u201339.\nGiesselmann, Marco, and Alexander W. Schmidt-Catran. \n2019. \u201cGetting the within Estimator of Cross-Level \nInteractions in Multilevel Models with Pooled Cross-\nSections: Why Country Dummies (Sometimes) \nDo Not Do the Job.\u201d Sociological Methodology \n49(1):190\u2013219.\nGlass, Jennifer, Robin W. Simon, and Matthew A. \nAndersson. 2016. \u201cParenthood and Happiness: \nEffects of Work\u2013Family Reconciliation Policies \nin 22 OECD Countries.\u201d American Journal of \nSociology 122(3):886\u2013929.\nHansen, \nThomas, \nBritt \nSlagsvold, \nand \nMarijke \nVeenstra. 2017. \u201cEducational Inequalities in Late-\nLife Depression across Europe: Results from the \nGenerations and Gender Survey.\u201d European Journal \nof Ageing 14(4):407\u201318.\nHeisig, Jan Paul, Merlin Schaeffer, and Johannes \nGiesecke. 2017. \u201cThe Costs of Simplicity: Why \nMultilevel Models May Benefit from Accounting for \nCross-Cluster Differences in the Effects of Controls.\u201d \nAmerican Sociological Review 82(4):796\u2013827.\nHemerijck, Anton, ed. 2017. The Uses of Social \nInvestment. Oxford, UK: Oxford University Press.\nHerbst, Chris M. 2013. \u201cWelfare Reform and the \nSubjective Well-Being of Single Mothers.\u201d Journal \nof Population Economics 26(1):203\u201338.\nHollingshead, August B., and Fredrick C. Redlich. 1958. \nSocial Class and Mental Illness: A Community Study. \nNew York, NY: Wiley.\nHox, Joop J. 2010. Multilevel Analysis: Techniques and \nApplications. 2nd ed. New York, NY: Routledge, \nTaylor, and Francis.\nJackson, Pamela Braboy, and David R. Williams. 2006. \n\u201cThe Intersection of Race, Gender, and SES: Health \nParadoxes.\u201d Pp. 131\u201362 in Gender, Race, Class, & Health: \nIntersectional Approaches, edited by A. J. Schultz and \nL. Mullings. San Francisco, CA: Jossey-Bass.\nJarvis, Edward. 1855. Report on Insanity and Idiocy in \nMassachusetts, by the Commission on Lunacy, under \nResolve of the Legislature of 1854. Boston, MA: \nWilliam White.\nKessler, Ronald C. 1982. \u201cA Disaggregation of the \nRelationship between Socioeconomic Status and \nPsychological Distress.\u201d American Sociological \nReview 47(6):752\u201364.\nKessler, Ronald C., Patricia Berglund, Olga Demler, \nRobert Jin, Kathleen R. Merikangas, and Ellen E. \nWalters. 2005. \u201cLifetime Prevalence and Age-of-\nOnset Distributions of DSM-IV Disorders in the \nNational Comorbidity Survey Replication.\u201d Archives \nof General Psychiatry 62(6):593\u2013602.\nKessler, Ronald C., and James A. McRae. 1981. \u201cTrends \nin the Relationship between Sex and Psychological \nDistress: \n1957\u20131976.\u201d \nAmerican \nSociological \nReview 46(4):443\u201352.\nKessler, Ronald C., William Saunders, Cindy Foster, \nand Paul Stang. 1995. \u201cSocial Consequences of \nPsychiatric Disorders I: Educational Attainment.\u201d \nAmerican Journal of Psychiatry 7(152):1026\u201332.\nKohn, Melvin L., Atsushi Naoi, Carrie Schoenbach, Carmi \nSchooler, and Kazimierz M. Slomczynski. 1990. \n\n352\t\nJournal of Health and Social Behavior 64(3)\n\u201cPosition in the Class Structure and Psychological \nFunctioning in the United States, Japan, and Poland.\u201d \nAmerican Journal of Sociology 95(4):964\u20131008.\nKohn, \nMelvin \nL., \nand \nCarmi \nSchooler. \n1973. \n\u201cOccupational \nExperience \nand \nPsychological \nFunctioning: An Assessment of Reciprocal Effects.\u201d \nAmerican Sociological Review 38(1):97\u2013118.\nKohn, Melvin L., and Carmi Schooler. 1982. \u201cJob \nConditions \nand \nPersonality: \nA \nLongitudinal \nAssessment of Their Reciprocal Effects.\u201d American \nJournal of Sociology 87(6):1257\u201386.\nKuitto, Kati. 2016. \u201cFrom Social Security to Social \nInvestment? Compensating and Social Investment \nWelfare Policies in a Life-Course Perspective.\u201d \nJournal of European Social Policy 26(5):442\u201359.\nKvist, Jon. 2015. \u201cA Framework for Social Investment \nStrategies: Integrating Generational, Life Course, and \nGender Perspectives in the EU Social Investment \nStrategy.\u201d Comparative European Politics 13(1):131\u201349.\nLebihan, Laetitia, and Charles-Olivier Mao Takongmo. \n2018. \u201cThe Impact of Universal Child Benefits \non Family Health and Behaviours.\u201d Research in \nEconomics 72(4):415\u201327.\nLee, Hee Seung, and Douglas A. Wolf. 2014. \u201cAn \nEvaluation of Recent Old-Age Policy Innovations in \nSouth Korea.\u201d Research on Aging 36(6):707\u201330.\nLevecque, Katia, Ronan Van Rossem, K. De Boyser, \nSarah Van de Velde, and Piet Bracke. 2011. \n\u201cEconomic Hardship and Depression across the Life \nCourse: The Impact of Welfare State Regimes.\u201d \nJournal of Health and Social Behavior 52(2):262\u201376.\nLink, Bruce G., Mary Clare Lennon, and Bruce P. \nDohrenwend. 1993. \u201cSocioeconomic Status and \nDepression: The Role of Occupations Involving \nDirection, Control, and Planning.\u201d American Journal \nof Sociology 98(6):1351\u201387.\nLink, Bruce G., and Jo Phelan. 1995. \u201cSocial Conditions \nas Fundamental Causes of Disease.\u201d Journal of \nHealth and Social Behavior 35(Extra Issue):80\u201394.\nLynch, John W., George A. Kaplan, and Sarah J. Shema. \n1997. \u201cCumulative Impact of Sustained Economic \nHardship on Physical, Cognitive, Psychological, \nand Social Functioning.\u201d New England Journal of \nMedicine 337(26):1889\u201395.\nMcAllister, Ashley, Sara Fritzell, Melody Almroth, Lisa \nHarber-Aschan, Sofia Larsson, and Bo Burstr\u00f6m. \n2018. \u201cHow Do Macro-level Structural Determinants \nAffect Inequalities in Mental Health? A Systematic \nReview of the Literature.\u201d International Journal for \nEquity in Health 17:180\u201394\nMiech, Richard A., and Michael J. Shanahan. 2000. \n\u201cSocioeconomic Status and Depression over the \nLife Course.\u201d Journal of Health and Social Behavior \n41(2):162\u201376.\nMilligan, Kevin, and Mark Stabile. 2011. \u201cDo Child \nTax Benefits Affect the Well-Being of Children? \nEvidence from Canadian Child Benefit Expansions.\u201d \nAmerican Economic Journal: Economic Policy \n3(3):175\u2013205.\nMirowsky, John, and Catherine E. Ross. 2003. Social \nCauses of Psychological Distress. 2nd ed. New \nBrunswick, NJ: Transaction Publishers.\nMirowsky, John, and Catherine E. Ross. 2007. \u201cLife \nCourse Trajectories of Perceived Control and Their \nRelationship to Education.\u201d American Journal of \nSociology 112(5):1339\u201382.\nNiedzwiedz, Claire L., Richard J. Mitchell, Niamh K. \nShortt, and Jamie R. Pearce. 2016. \u201cSocial Protection \nSpending and Inequalities in Depressive Symptoms \nacross Europe.\u201d Social Psychiatry and Psychiatric \nEpidemiology 51(7):1005\u201314.\n\u00d3lafsd\u00f3ttir, Sigrun, and Jason Beckfield. 2020. \u201cHealth \nInequalities: What Do We Know and What Do \nWe Need to Know?\u201d Social Science & Medicine \n267:113575. doi:10.1016/j.socscimed.2020.113575.\nOrganization \nfor \nEconomic \nCooperation \nand \nDevelopment. \n2021. \n\u201cSocial \nExpenditure \n- \nAggregated \nData.\u201d \nhttps://stats.oecd.org/Index.\naspx?datasetcode=SOCX_AGG#.\nPak, Tae-Young. 2020. \u201cSocial Protection for Happiness? \nThe Impact of Social Pension Reform on Subjective \nWell-Being of the Korean Elderly.\u201d Journal of Policy \nModeling 42(2):349\u201366.\nPearlin, Leonard I. 1989. \u201cThe Sociological Study of \nStress.\u201d Journal of Health and Social Behavior \n30(3):241\u201356.\nPearlin, Leonard I., Elizabeth G. Menaghan, Morton A. \nLieberman, and Joseph T. Mullan. 1981. \u201cThe Stress \nProcess.\u201d Journal of Health and Social Behavior \n22(4):337\u201356.\nPierson, \nPaul. \n2000. \n\u201cIncreasing \nReturns, \nPath \nDependence, and the Study of Politics.\u201d American \nPolitical Science Review 94(2):251\u201367.\nPlavgo, Ilze, and Anton Hemerijck. 2021. \u201cThe Social \nInvestment \nLitmus \nTest: \nFamily \nFormation, \nEmployment, and Poverty.\u201d Journal of European \nSocial Policy 31(3):282\u201396.\nPr\u00e4g, Patrick, Melinda C. Mills, and Rafael Wittek. \n2016. \u201cSubjective Socioeconomic Status and Health \nin Cross-National Comparison.\u201d Social Science & \nMedicine 149:84\u201392.\nRai, Dheeraj, Pedro Zitko, Kelvyn Jones, John Lynch, \nand Ricardo Araya. 2013. \u201cCountry- and Individual-\nLevel Socioeconomic Determinants of Depression: \nMultilevel Cross-National Comparison.\u201d British \nJournal of Psychiatry 202(3):195\u2013203.\nReynolds, Megan M. 2021. \u201cHealth Power Resources \nTheory: A Relational Approach to the Study of \nHealth Inequalities.\u201d Journal of Health and Social \nBehavior 62(4):493\u2013511.\nRoss, Catherine E., and John Mirowsky. 2010. \u201cWhy \nEducation Is the Key to Socioeconomic Differentials \nin Health.\u201d Pp. 33\u201351 in Handbook of Medical \nSociology. 6th ed., edited by C. E. Bird, A. M. \nFremont, and S. Timmermans. Nashville, TN: \nVanderbilt University Press.\nSchmidt-Catran, Alexander W., Malcolm Fairbrother, and \nHans-J\u00fcrgen Andre\u00df. 2019. \u201cMultilevel Models for \n\nParbst and Wheaton\t\n353\nthe Analysis of Comparative Survey Data: Common \nProblems and Some Solutions.\u201d KZfSS K\u00f6lner \nZeitschrift F\u00fcr Soziologie und Sozialpsychologie \n71(S1):99\u2013128.\nScruggs, Lyle A., and James Allan. 2008. \u201cSocial \nStratification and Welfare Regimes for the Twenty-\nFirst Century: Revisiting the Three Worlds of \nWelfare Capitalism.\u201d World Politics 60(4):642\u201364.\nSewell, William H., Robert M. Hauser, and David L. \nFeatherman, eds. 1976. Schooling and Achievement \nin American Society. New York, NY: Academic \nPress.\nSimpson, Julija, Viviana Albani, Zoe Bell, Clare Bambra, \nand Heather Brown. 2021. \u201cEffects of Social Security \nPolicy Reforms on Mental Health and Inequalities: \nA Systematic Review of Observational Studies in \nHigh-Income Countries.\u201d Social Science & Medicine \n272:113717. doi:10.1016/j.socscimed.2021.113717.\nSolt, Frederick. 2019. The Standardized World Income \nInequality Database: Version 8.0 [dataset]. Harvard \nDataverse. doi:10.7910/DVN/LM4OWF.\nSong, Lijun. 2011. \u201cSocial Capital and Psychological \nDistress.\u201d Journal of Health and Social Behavior \n52(4):478\u201392.\nSrole, Leo, Thomas S. Langner, Stanley T. Michael, \nMarvin K. Opler, and Thomas A. C. Rennie. 1962. \nMental Health in the Metropolis: The Midtown \nManhattan Study. Vol. 1. New York, NY: McGraw-\nHill Book Company, Inc.\nThoits, Peggy A. 1995. \u201cStress, Coping, and Social \nSupport Processes: Where Are We? What Next?\u201d \nJournal of Health and Social Behavior 35(Extra \nIssue):53\u201379.\nThoits, Peggy A. 2010. \u201cStress and Health: Major \nFindings and Policy Implications.\u201d Journal of Health \nand Social Behavior 51(1 Suppl.):S41\u2013S53.\nTitmuss, Richard M. 1974. Social Policy: An Introduction, \nedited by B. Abel-Smith and K. Titmuss. London: \nGeorge Allen & Unwin.\nTurner, R. Jay, Blair Wheaton, and Donald A. Lloyd. \n1995. \u201cThe Epidemiology of Social Stress.\u201d \nAmerican Sociological Review 60(1):104\u201325.\nVan de Velde, Sarah, Piet Bracke, Katia Levecque, and B. \nMeuleman. 2010. \u201cGender Differences in Depression \nin 25 European Countries after Eliminating \nMeasurement Bias in the CES-D 8.\u201d Social Science \nResearch 39(3):396\u2013404.\nVan der Meer, Tom, Manfred Te Grotenhuis, and Ben \nPelzer. 2010. \u201cInfluential Cases in Multilevel \nModeling: A Methodological Comment.\u201d American \nSociological Review 75(1):173\u201378.\nvan der Wel, Kjetil A., Clare Bambra, Nico Dragano, Terje \nA. Eikemo, and Thorsten Lunau. 2015. \u201cRisk and \nResilience: Health Inequalities, Working Conditions, \nand Sickness Benefit Arrangements: An Analysis of \nthe 2010 European Working Conditions Survey.\u201d \nSociology of Health & Illness 37(8):1157\u201372.\nVeenhoven, Ruut. 2000. \u201cWell-Being in the Welfare \nState: Level Not Higher, Distribution Not More \nEquitable.\u201d Journal of Comparative Policy Analysis: \nResearch and Practice 2(1):91\u2013125.\nWarren, John. 2009. \u201cSocioeconomic Status and \nHealth across the Life Course: A Test of the Social \nCausation and Health Selection Hypotheses.\u201d Social \nForces 87(4):2125\u201353.\nWheaton, Blair. 1978. \u201cThe Sociogenesis of Psychological \nDisorder: Reexamining the Causal Issues with \nLongitudinal Data.\u201d American Sociological Review \n43(3):383\u2013403.\nWheaton, Blair. 1983. \u201cStress, Personal Coping \nResources, \nand \nPsychiatric \nSymptoms: \nAn \nInvestigation of Interactive Models.\u201d Journal of \nHealth and Social Behavior 24(3):208\u201329.\nWilensky, Harold L. 1975. The Welfare State and \nEquality: Structural and Ideological Roots of Public \nExpenditures. Berkeley: University of California \nPress.\nWilkinson, Richard G., and Kate Pickett. 2010. The Spirit \nLevel: Why Equality Is Better for Everyone. New \nYork, NY: Penguin Books.\nWorld Bank. 2021. \u201cGovernment Expenditure on \nEducation, Total (% of GDP).\u201d https://data.world-\nbank.org/indicator/SE.XPD.TOTL.GD.ZS?view= \nchart.\nAuthor Biographies\nMatthew Parbst is a PhD candidate in the Department of \nSociology at the University of Toronto. Mr. Parbst\u2019s \nresearch interests include social policy, social determi-\nnants of health, and comparative research. His current \nresearch focuses on examining how social policy modifies \nindividual-level social determinants of health relationships \nwith mental health in the cross-national perspective.\nBlair Wheaton is currently the S.D. Clark Professor of \nSociology at the University of Toronto. He received his \nPhD from the University of Wisconsin in 1976 and taught \nat Yale University and McGill University before moving \nto the University of Toronto in 1989. He was the first \nrecipient of the Leonard I. Pearlin Award for Distinguished \nContributions to the Sociology of' Mental Health in 2000 \nand was elected to the Sociological Research Association \nin 2010. His research focuses on both the life course and \nsocial contextual approaches to understanding mental \nhealth over multiple life stages. Currently, he is following \nup the children from the Toronto Family Study 30 years \nlater to investigate the long-term consequences of growing \nup in gender-egalitarian households on work, family, and \nhealth outcomes.\n\n\n    Scientific Research Findings:", "answer": "\u2022 Our main findings generalize the SES and mental health link across the effects of education, occupation, and income across 21 countries while demonstrating significant policy effects that alter this association.\n\u2022 Our main policy findings suggest that Social Investment, rather than Social Protection, both generally and via specific disaggregated policies, reduces mental health inequalities between low and high individual status across countries.\n\u2022 countries with higher levels of Social Investment report more equal levels of mental health across education, occupation, and income. This is reflected by the flattening of the differences between low and high status as investment increases.", "id": 69}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465231172177\nJournal of Health and Social Behavior\n2023, Vol. 64(4) 470\u00ad\u2013485\n\u00a9 American Sociological Association 2023\nDOI: 10.1177/00221465231172177\njournals.sagepub.com/home/hsb\nOriginal Article\nMedical uncertainty has interested sociologists for \ndecades (Fox 1957, 1980; Gerrity, DeVellis, and \nLight 1992; Mackintosh and Armstrong 2020). \nUncertainty refers to the unknown in medical prac-\ntice and is significant due to its role in medical \nknowledge, diagnosis, and decision-making (Fox \n2000; \nJenkins, \nJessen, \nand \nSteffen \n2005). \nUncertainty has various sources, such as the inher-\nent limits of what is knowable in medicine (Atkinson \n1984). Clinical experience, standardization, guide-\nlines, and technology can alleviate some aspects of \nuncertainty but introduces others (e.g., Rapp 2004; \nRothman 1993). Sociological research has primar-\nily focused on the sources of medical uncertainty \nand how professionals and patients experience and \nmanage it (Dumit 2006; Jutel 2009; Littlejohn and \nKimport 2017).\nWe build on prior work by exploring an unex-\namined source of medical uncertainty: the political \ncontext of care. We use the case of abortion in Ohio, \na state with one of the most restrictive legislative \ncontexts in the United States prior to the \n1172177 HSBXXX10.1177/00221465231172177Journal of Health and Social BehaviorCzarnecki et al.\nresearch-article2023\n1Case Western Reserve University, Cleveland, OH, \nUSA\n2University of Cincinnati, Cincinnati, OH, USA\n3Ohio State University, Columbus, OH, USA\n4Mayo Clinic, Rochester, MN, USA\nCorresponding Author:\nDanielle Czarnecki, Department of Sociology, Case \nWestern Reserve University, 226 Mather Memorial \nBuilding, 10900 Euclid Avenue, Cleveland, OH 44106-\n7124, USA. \nEmail: dxc735@case.edu\nState of Confusion: Ohio\u2019s \nRestrictive Abortion  \nLandscape and the Production \nof Uncertainty in Reproductive \nHealth Care\nDanielle Czarnecki1\n, Danielle Bessett2, Hillary J. Gyuras2, \nAlison H. Norris3, and Michelle L. McGowan2,4\nAbstract\nThis study examines an underexplored source of medical uncertainty: the political context of care. Since \n2011, Ohio has passed over 16 abortion-restrictive laws. We know little about how this legislation affects \nreproductive health care outside of abortion clinics. Drawing on focus groups and interviews with genetic \ncounselors and obstetrician-gynecologists, we examine how abortion legislation impacts their work. We \nfind that interpretation and implementation of legislation is not straightforward and varies by institution \nand region of the state. An ever-changing legislative landscape combined with uneven implementation \nof restrictions into policy produces uncertainty in reproductive health care. We also found uncertainty \nabout the legal consequences of abortion in restrictive contexts, with obstetrician-gynecologists reporting \ngreater concerns given their proximity to care provision. We argue that uncertainty can result in stricter \ninterpretations of regulations than necessitated by the law, thereby amplifying the impacts of an already \nrestrictive context for abortion care.\nKeywords\nabortion, genetic counseling, ob-gyns, reproduction, uncertainty\n\nCzarnecki et al.\t\n471\noverturning of Roe (Guttmacher Institute 2018), to \nshow how the political context of care produces \nuncertainty. The political context of care includes \n(1) cultural politicization, such as stigma; (2) the \nlegal, regulatory, and policy contexts; and (3) the \neconomic organization of health care. We focus on \nhow an influx of laws and regulations that are often \nunclear and incongruent with evidence-based medi-\ncine necessitate and complicate the interpretive \nwork of institutions and providers. Frequent policy \nchanges disrupt standards and practices, economic \nand legal penalties stigmatize abortion care, and \nvariation in policies across institutions and states \nlimits information seeking to address uncertainty. \nThis politicized climate as a source of uncertainty is \nwhat we seek to understand.\nThis form of uncertainty matters because \npatients\u2019 lives and providers\u2019 professional integrity \nand livelihoods are at stake. Health care providers \nface the threat of criminalization, experience dis-\ntress when unable to provide necessary care, and \nare burdened with increased labor when navigating \ncomplicated restrictions and institutional con-\nstraints (Field et al. 2022; Heuerman et al. 2022). \nMedical exceptions to abortion restrictions, such as \nlife-threatening circumstances, leave clinicians \nuncertain about whether their clinical interpretation \nwill be viewed as legally defensible (Arey et al. \n2022). These laws ultimately challenge the grounds \non which care decisions should be made, from med-\nical evidence to other metrics, thereby eroding \npatient and provider autonomy in decision-making. \nAs the politicization of care expands to other areas \nof health care, medical uncertainty may increase.\nWe interviewed genetic counselors and obstetrician-\ngynecologists (ob-gyns) working in hospital set-\ntings to explore how abortion legislation affects \ntheir work. We find that the interpretation and \nimplementation of legislation is not straightforward \nand varies by institution. An ever-changing legisla-\ntive landscape combined with different institutional \ninterpretations of legal restrictions produces uncer-\ntainty for those working in reproductive health care \nsettings. This uncertainty occurs within a climate of \nfear and stigma surrounding abortion care provision \nin restrictive contexts, impacting health care profes-\nsions differently depending on their professional \nstatus and proximity to abortion care. We argue that \na political context that produces uncertainty can \nresult in more limiting interpretations and imple-\nmentation of legal restrictions than necessitated by \nthe law, thereby amplifying the impacts of an \nalready restrictive context. This chilling effect cap-\ntures how the political context of care can constrain \ncare beyond what law and policy dictate because \nclinicians and institutions err on the side of cau-\ntion when uncertain whether providing clinically \nappropriate care is legal. With Roe overturned, the \nlegislative context remains in flux such that repro-\nductive rights depend on one\u2019s state of residence \n(McCann 2022). Practicing in a state of uncer-\ntainty is more relevant than ever for many health \ncare providers.\nOur contribution to the medical sociology litera-\nture is threefold. First, we expand the sources of \nmedical uncertainty to include the political context \nof care. Second, we demonstrate how health care \nproviders experience uncertainty and how it limits \naccess to care to a greater degree than the law. \nThird, we discuss the impact and implications of \nuncertainty for abortion care and beyond.\nBackground\nSources of Medical Uncertainty\nMedical sociologists have explored the sources of \nmedical uncertainty and how clinicians manage \nthem (Atkinson 1984; Gerrity et al. 1992; Light \n1979; Mackintosh and Armstrong 2020). Fox (1957, \n1980, 2000) identified three sources of uncertainty \nthat arise during medical training: an insufficient \ngrasp of medical knowledge, the limits of medical \nknowledge, and an inability to distinguish between \nthe two. Clinicians routinely grapple with decision-\nmaking amid uncertainty, such as identifying a med-\nical problem (Barker 2008) and determining how \nbest to treat it (Light 1979). Although uncertainty is \npart of medicine, it can also lead to physician burn-\nout and decreased engagement, with consequences \nfor patient care (Begin et al. 2022).\nCultural and technological change can also gen-\nerate new sources of uncertainty (Alderson 2009; \nJenkins et al. 2005; Mesman 2008). In genetic \ncounseling, uncertainty about screening results is \ncommon \n(Brookes-Howell \n2006). \nDiagnostic \nuncertainty following newborn screening can leave \nnewborns and parents in a liminal state as \u201cpatients-\nin-waiting\u201d (Timmermans and Buchbinder 2010). \nCultural shifts may also contribute to uncertainty. \nFor instance, childhood becoming a domain of sci-\nentific study factored in the diagnostic category of \nattention-deficit/hyperactivity disorder (ADHD; \nRafalovich 2013). Some clinicians experience \nuncertainty and ambivalence about prescribing \nstimulants for ADHD that potentially cause patient \nharm. While cultural and technological change con-\ntributing to uncertainty has been documented, less \n\n472\t\nJournal of Health and Social Behavior 64(4) \nexplored are changes from the politicization of \nmedical care that may also result in uncertainty.\nPhysicians manage uncertainty in different \nways, such as by acquiring more knowledge and \nskills, adopting a treatment model, or by \u201cmuting\u201d \nthe voice of uncertainty (Fox 2000; Light 1979). In \na study of pediatric residents\u2019 use of evidence-based \nmedicine (EBM), Timmermans and Angell (2001) \nfound that EBM provided guidelines to navigate \nuncertainty but also produced new \u201cresearch-\nbased\u201d uncertainties. In the context of transgender \nmedicine, some clinicians navigate uncertainty and \ntheir lack of knowledge about transgender identity \nby relying on EBM even though affirming gender \nidentity is not an illness (shuster 2016). Uncertainty \nabout transgender care can also result in stigmatiza-\ntion and discrimination (Poteat, German, and \nKerrigan 2013). Another study found that internal \nmedicine residents experiencing uncertainty faced \nbarriers to obtaining attending physicians\u2019 input, \nsuch as professional hierarchy and admitting \nknowledge gaps, which compromised patient care \n(Farnan et al. 2008).\nUncertainty characterizes various aspects of \nreproduction. Prenatal screening may entail diag-\nnostic uncertainty regarding the implications for the \npregnant person and fetus (Markens, Browner, and \nPreloran 2003; Rapp 2004; Rothman 1993). For \nthose who want to avoid pregnancy, there is consid-\neration of contraceptive method, risk, and side \neffects (Geampana 2019). A study of contraceptive \ncounseling found that how clinicians communicate \nuncertain side effects impacts patient decision-\nmaking and autonomy and has health consequences \n(Littlejohn and Kimport 2017). Changes in the \norganization of medicine also contribute to uncer-\ntainty. Following the corporatization of hospitals, \nMorris (2014) found that greater economic pres-\nsures combined with fear of obstetric malpractice \nsuits fueled the rise in caesarian sections in the \nUnited States, limiting organizational and profes-\nsional risk at the expense of patients\u2019 well-being.\nWhile uncertainty characterizes all of medicine, \nabortion is medically relatively straightforward. \nUncertainty about abortion care often originates \nfrom outside rather than within the medical context. \nFor example, those seeking care in restrictive con-\ntexts may be uncertain about whether care is legally \navailable and contend with social myths about \nabortion that influence decision-making (Heymann \net al. 2022; Kimport 2021). Therefore, it appears \nthat a critical unexplored source of uncertainty is \nthe political context of care.\nThe Politicization of Abortion Care in \nthe United States\nAbortion in the United States has a history of politi-\ncization that shapes contemporary care provision \n(e.g., Cohen and Joffe 2021; Ginsburg 1998; Luker \n1985). Despite abortion being a safe and common \nprocedure, it has been marginalized such that the \nmajority of abortions are performed in dedicated \nabortion clinics (Freedman 2010). Anti-abortion \nviolence and legislative battles further constrain \nabortion access and increase stigma (Harris et al. \n2013). Since 2010, a surge in state-level abortion \nrestrictions has further politicized abortion care, \nwith close to half of all restrictions being enacted \nduring this period (Guttmacher Institute 2019). In \n2022, the Supreme Court of the United States over-\nturned Roe v. Wade, which eliminated the federal \nright to abortion and resulted in immediate legal \nrestrictions\u2014including complete bans\u2014throughout \nparts of the country (Guttmacher Institute 2022). \nThe legislative context continues to rapidly change \nin many states (McCann 2022).\nPrior studies on abortion restrictions focus on \ntheir impact on the provision of care and decreased \naccess to abortion. Restrictions result in increased \ncosts and travel distances, delays in obtaining care, \nand medical practices that raise rather than reduce \nmedical risk (Fuentes et al. 2016; Grossman and \nWhite 2018). Others show that restrictions increase \nclinic closures and service fluctuations (McGowan, \nNorris, and Bessett 2020). Recent research exam-\nines abortion seekers\u2019 and the public\u2019s knowledge \nof abortion legality and barriers to accurate infor-\nmation (Bessett et al. 2015; White et al. 2016). For \nexample, uncertainty about the legality of abortion \nin Ohio and believing that it was illegal increased \nduring periods of legislative activity to propose a \nsix-week abortion ban (Gallo et al. 2021).\nA growing body of literature explores providers\u2019 \nexperiences of recent abortion regulations. For exam-\nple, abortion providers in North Carolina challenged \nregulations by creatively scripting state-mandated \nabortion counseling (Buchbinder 2016). Restrictions \nand separating abortion care from mainstream medi-\ncine also contribute to challenges in training abortion \nproviders, limiting the number of providers (Gyuras \net al. 2023). Only 64% of U.S. obstetrics and gynecol-\nogy residency programs offer routine abortion train-\ning (Steinauer et al. 2018), which will likely decrease \nin a post-Dobbs context. Even trained providers find \nthemselves unable to provide care for various rea-\nsons, such as work policies (Freedman 2010).\n\nCzarnecki et al.\t\n473\nAlso important, yet underexplored, is how \nreproductive health care providers working outside \nof abortion clinics experience abortion restrictions \nand whether restrictions contribute to uncertainty. \nThese providers also interact with patients who \nneed and desire abortion care, which is offered\u2014\nalthough less frequently\u2014in some hospitals \n(Czarnecki et al. 2019; McLemore, Kools, and Levi \n2015). A recent study of genetic counselors work-\ning in Ohio showed that legislation negatively \nimpacted the provider\u2013patient relationship by com-\npressing the timeline for decision-making, inhibit-\ning capacity for patient advocacy, and increasing \nemotional burnout (Heuerman et al. 2022). Prior \nresearch on regulations in religious hospitals offers \ninsights into constraints on reproductive health care \n(Stulberg, Jackson, and Freedman 2016). Across \nthe United States, Catholic hospital policies restrict \ncare\u2014including for miscarriage management \n(Freedman and Stulberg 2013). Abortion clinic clo-\nsures further decrease referral options (Zurek and \nO\u2019Donnell 2019). A mixed-methods study assess-\ning abortion policies in U.S. teaching hospitals \nfound that over half limited abortion beyond what \nis mandated by state law, particularly in the \nMidwest and South (Zeldovich et al. 2020). This \nresearch shows that the political context of care \nshapes medical practice and training.\nThe Political Context of Abortion  \nCare in Ohio\nSince 2011, Ohio has passed numerous abortion-\nrestrictive laws. Prior to the recent Dobbs decision, \nOhio was one of the most abortion-restrictive states \n(Norris et al. 2020). The day that Roe was over-\nturned, a federal judge granted the state of Ohio\u2019s \nrequest to allow a bill that had been suspended by a \nlegal challenge under Roe to take effect immedi-\nately (SB23). Doing so made providing abortion \nillegal in almost all circumstances upon detection of \nembryonic cardiac activity, typically around six \nweeks\u2019 gestation. The embryonic cardiac activity \nban remained in effect until mid-September 2022, \nwhen a state court judge blocked the bill for the \nduration of a new legal challenge to the ban. As of \nthis writing, the embryonic cardiac activity ban \nremains blocked. The dynamic status of Ohio\u2019s \nabortion restrictions demonstrates the evolving leg-\nislative climate in the post-Dobbs era.\nNumerous Ohio restrictions remain in effect, \nincluding mandatory waiting periods, counseling \nrequirements, restrictions constraining minors seek-\ning abortion care, prohibitions on public insurance \ncoverage for abortion and on facilities performing \nthem, viability and gestational restrictions, transfer \nagreement requirements, a Down syndrome abor-\ntion ban, and abortion method bans (Heuerman et al. \n2022). Providers may face considerable civil and \ncriminal penalties, such as felony charges, prison \nsentences, fines amounting to several thousands of \ndollars, and loss of medical licensure, should they \nbe found to be in violation of certain laws, including \nrestrictions on dilation and evacuation1 (D&E) pro-\ncedures, a ban on abortions sought due to fetal \nDown syndrome, and a 22-week gestational limit.\nOur study examines how the distinct character \nof recent legislative restrictions\u2014their number, \nscope, and frequency\u2014contributes to uncertainty \nthat may impact patient care. Uncertainty is likely \neven more common and challenging in a post-\nDobbs context where the legal context in many \nstates is in greater flux (McCann 2022).\nData and Methods\nWe conducted focus groups with prenatal genetic \ncounselors and ob-gyns employed in a variety of \nOhio hospital settings (e.g., private, public, aca-\ndemic, religious). Focus groups are a useful method \nfor exploratory research that enables respondents \nwith similar characteristics, such as profession, to \nengage in conversation. Participants who were \nunable to attend focus groups shared their experi-\nences in individual, semistructured interviews. The \nUniversity of Cincinnati\u2019s Institutional Review \nBoard approved this study.\nParticipants were recruited through professional \nsociety memberships and listservs, advocacy \ngroups, university-affiliated hospitals, and snow-\nball sampling. Participants were eligible if they \nwere ob-gyns or genetic counselors who worked in \nOhio for at least six months between 2010 and \n2020. We excluded providers working in dedicated \nabortion clinics because the impact of abortion reg-\nulations on those who work in abortion clinics has \nbeen well documented. Less understood are the \nimplications of these laws for general obstetric-\ngynecological practice and for abortion-adjacent \nhealth care providers\u2014such as genetic counselors\u2014\nwho work in the same state regulatory context. \nGenetic counselors counsel patients on their repro-\nductive options, including abortion, prior to prena-\ntal testing and following prenatal diagnosis. \nOb-gyns interact with patients who need a range of \nreproductive health care, such as abortion, contra-\nception, and management of miscarriage and ecto-\npic pregnancy.\n\n474\t\nJournal of Health and Social Behavior 64(4) \nWe conducted four ob-gyn focus groups, ranging \nfrom 2 to 19 participants per group, and five individ-\nual interviews with ob-gyns (n = 35). We conducted \nfive genetic counselor focus groups, ranging from 2 \nto 6 participants per group, and two individual inter-\nviews with genetic counselors (n = 19). The total \nnumber of participants in the study was 54. \nParticipants provided verbal consent and received an \nOhio legislative timeline of abortion restrictions \nbetween 2010 and 2019 (Heuerman et al. 2022). \nParticipants were asked to review the timeline and \ndiscuss how regulations impacted their work. A  \nsemistructured interview guide ensured comprehen-\nsive coverage of the topic. Although some partici-\npants reported the type of hospital at which they \nwere employed (e.g., private, public), some \nexpressed uncertainty about their institution type, \nthus we were unable to definitively collect that infor-\nmation from all respondents. When a respondent\u2019s \nhospital type was provided, we note it in the results.\nInterviews and focus groups were audio-\nrecorded. Focus groups were 90 minutes; individual \ninterviews lasted 45 to 60 minutes. The recordings \nwere transcribed and deidentified, including provi-\nsion of pseudonyms for participants. At least two \nmembers of the research team thematically coded \neach transcript using ATLAS.ti, a qualitative analy-\nsis software. Coders met regularly to discuss \nthemes and resolve discrepancies. Analysis con-\nsisted of an iterative process of thematic categoriza-\ntion, \nconsultation \nwith \nrelevant \nliterature, \ndiscussion of themes, and revisiting the data and \nrefining codes (Emerson, Fretz, and Shaw 2011). \nMajor themes emerged: uncertainty and confusion, \nnegative impacts of restrictions on patients and pro-\nviders, ethical dilemmas, varied interpretations, \nregional variation, and inequalities in access.\nResults\nWe found that Ohio\u2019s rapidly shifting legislative \nlandscape of abortion regulation produced uncer-\ntainty for ob-gyns and genetic counselors working \nin hospitals. Hospitals\u2019 interpretation and imple-\nmentation of abortion legislation\u2014some of which \nwas not in effect at the time of the study\u2014deepened \nuncertainty about what is permissible. We show that \nuncertainty can result in more stringent restrictions \nthan necessitated by the law, further limiting access \nto care. We also found uncertainty about the legal, \nprofessional, and personal consequences of provid-\ning or facilitating access to abortion care in a restric-\ntive context, with ob-gyns\u2019 closer proximity to \nabortion care increasing their perception of threats \nto their professional and personal integrity and live-\nlihood. Importantly, all respondents understood \nabortion as health care and were frustrated when \nthey were unable to provide appropriate patient care.\n\u201cIt Was Always Hard to Keep Up with\u201d: \nUncertainty about Legal Restrictions\nAlthough respondents voiced familiarity with abor-\ntion legislation, there was widespread uncertainty and \nfrustration regarding the legal status and content of \nseveral bills, such as abortion method restrictions, rea-\nson bans, and prohibitions on the use of public funds. \nThis uncertainty was partly due to the challenges of \nkeeping up with a changing legislative context that \nincludes numerous bills with varied legal statuses \n(proposed, passed, enacted, enjoined,2 partially \nenjoined). Anna, an ob-gyn, explained, \u201cIt\u2019s always so \nconfusing what went up for votes, what went through, \nand what got blocked by a judge and what was enacted \n. . . it was always hard to keep up with.\u201d\nThe swift pace of legislative changes left pro-\nviders in a chronic state of uncertainty in the work-\nplace, with patients\u2019 options being subject to change \nfrom one day to the next. For instance, genetic \ncounselors described how routine referrals became \na process marked by uncertainty, which is exacer-\nbated by abortion clinic closures in the state, limit-\ning referral options. Genetic counselors managed \nuncertainty by regularly calling abortion clinics for \nthe most current legal information about patient \noptions. Bridget, a genetic counselor working in an \nacademic setting, explained:\nEvery week I had to call [the clinic] that we \ntypically refer to and say, \u201cWhat is the status this \nweek? . . . What is my time limit now?\u201d I found \nit frustrating that I would have to make that call \nevery single week to know how to guide \n[patients].\nTo navigate uncertainty, some respondents adapted \ntheir care approaches. For example, prenatal genetic \ncounselors typically see patients as they are nearing the \nstate\u2019s 21 weeks and 6 days gestational limit on abor-\ntion because imaging technology used to detect fetal \nabnormalities is effective at 18-weeks\u2019 gestation and \nlater. Aware of the compressed timeline for decision-\nmaking, some genetic counselors counseled patients \nconsidering termination to start fulfilling the numerous \nlegal requirements for abortion care because \u201cevery \nday counts.\u201d Other providers encouraged earlier pre-\nnatal testing, but technological limits and uneven utili-\nzation of prenatal care limit this strategy.\n\nCzarnecki et al.\t\n475\nOb-gyns also described how uncertainty about \nchanging laws affected their practices. In 2018, the \nOhio legislature restricted D&E procedures, the \nmost common second trimester abortion method \nand the only method available after 15 weeks in \nOhio. Ob-gyns described how the dynamic political \ncontext affects patient care:\nSandy:\u00a0After the D&E ban was officially passed . . . \n[the courts] were figuring out injunctions and \nstuff. We had a patient come in for a miscar-\nriage, which regardless of the method would \nhave been okay at that point. But it was around \n14 weeks, we were going to do a D&C and \nended up doing kind of like a modified D&E, \nand right before we started doing that part, my \nchief resident turns to me and goes, \u201cWait, we\u2019re \nallowed to do this right?\u201d Because literally just \nin terms of that information, it\u2019s just not widely \navailable or widespread.\nThe focus group conversation continued at length \nabout whether a D&E would be legal in a variety of \npatient circumstances that ob-gyns encounter in \ntheir work, showing the extent of uncertainty on \ntheir practice.\nProfessional organizations, advocacy groups, \nand their employers were rarely cited as useful \nsources of information about abortion legislation. \nRespondents described that hospitals typically \nrefrained from formally discussing legislation. \nAnna, aforementioned ob-gyn, could not recall her \nemployers ever addressing these laws with clini-\ncians, \u201cThe institutions that I worked at . . . never \ndirectly addressed these laws or their regulations on \nabortion to the med students or the fellows or any-\nbody.\u201d Respondents often learned what was permis-\nsible through social media, the news, and on an ad \nhoc basis when a patient presented for care.\nLegal stipulations that are not medically legible \nwere another source of uncertainty. Tamara, an ob-\ngyn resident at a public hospital, discussed her frus-\ntration with a proposed bill that would, if in effect, \nrequire that clinicians reimplant an ectopic preg-\nnancy into the uterus or face murder charges:\nIt doesn\u2019t even make sense. It\u2019s not a thing we \nare capable of doing. So, it\u2019s crazy to me that \nthere\u2019s a lot of made-up medicine . . . that\u2019s not \nsafe to patients, ectopics are dangerous. People \ncan bleed out very quickly.\nTamara was puzzled by a bill that would require doc-\ntors to perform something that was not medically \nappropriate or technically possible, thereby under-\nmining EBM and professional autonomy.\nAn ob-gyn resident, Tracy, who worked in a \nreligious hospital, summed up what we often heard \nfrom respondents: an awareness of the laws without \na full understanding of their status or what it meant \nfor their work:\nIt\u2019s hard to know [the laws], you have attendings \nthat kind of know, or the hospital policies, I kind \nof have a sense of what the trajectory has been, \nbut I don\u2019t know what the current thing is, I \nwould have to look it up or ask someone. That\u2019s \nhard to keep up with.\nTracy\u2019s comment points to multiple actors, such as \nsenior clinicians and hospital administrators, \ninvolved in translating laws into hospital policy. \nImplicit in her comment is that while policies \ninformed by legislation exist in her workplace, they \nmay not fully reflect or reference the law.\n\u201cLegal versus What\u2019s Available Is \nNot the Same Thing\u201d: Institutional \nInterpretations of Abortion Legislation \nand Regional Variation\nAnother factor contributing to uncertainty was that \ninterpretations of legislation could vary by institu-\ntion and region of the state. Although the most obvi-\nous reason for institutional variation is due to \nrestrictions based on funding or institution type, we \nfound that this does not fully explain the variation. \nRespondents also perceived that institutional cul-\ntures, personal beliefs, fear and stigma, or a combi-\nnation thereof accounted for some variation. We \nheard from respondents that Ohio\u2019s hospitals\u2014both \nreligious and nonreligious\u2014tend to have abortion \nrestrictions that are more conservative than the law, \nimpacting patients\u2019 access to abortion care more \nthan legislation alone would suggest. In fact, laws \nnot in effect during the study period still informed \ninstitutional policies and practices.\nIn 2011, Ohio imposed a public facilities ban for \n\u201cnon-therapeutic abortions,\u201d which remains in \neffect. We found that some respondents\u2014including \nthose working at the same institution\u2014were uncer-\ntain about their institution type, such as whether it \nwas public, private, or nonprofit. This uncertainty is \nunsurprising given that hospital mergers and \ndiverse hospital funding sources can make \u201cinstitu-\ntion type\u201d unclear in practice, making the applica-\nbility of an institution-based law potentially unclear \n\n476\t\nJournal of Health and Social Behavior 64(4) \n(Creswell and Abelson 2013). Institution-based \nrestrictions and uncertainty or inaccuracies about \nthem matter because they decrease access and \nbecause health care providers communicate \npatients\u2019 care options. Relaying inaccurate informa-\ntion to patients may further decrease access to  \nservices that remain legal. Moreover, hospital net-\nworks may include different kinds of institutions \nwith distinct regulations. Susan, a genetic counselor \nworking in a public hospital, noted:\nIn our perinatology practice, one of the off sites, \n[perinatologists] go to visit a Catholic hospital \nwhere they cannot even address [abortion] at all. \nAnd so, if we want to talk more about [patient] \noptions, let\u2019s have you come to the main hospital.\nRequiring patients to visit a different building in the \nsame hospital network to discuss options thoroughly \ncomplicates a routine aspect of patient care and \nundermines provider autonomy and authority in \nrelaying options. Respondents explained that these \ncumbersome processes burdened patients with \nabortion stigma. The added layer of uncertainty \nabout institution type makes it challenging to know \nnot only what the restrictions are but also the reason for \nthem (laws, institutional type, Catholic directives).\nOb-gyns also had different understandings of \nwhat legislation was in effect and what it meant for \ntheir work, even at the same institution. For exam-\nple, a 2017 bill prohibits abortion for reasons \nrelated to fetal Down syndrome. The bill was \nenjoined and not in effect at the time of the study. \nCarla and Pamela had different understandings of \nthe bill\u2019s status:\nCarla:\u00a0I don\u2019t think it\u2019s [the Down syndrome ban] in \neffect that I know of.\nPamela:\u00a0I thought it was in effect . . . so we say that \n[patients] may have to cross state lines. For \npatients that don\u2019t bring it up, I feel I still need \nto bring it up and explain that this may be a rea-\nson they may or may not be able to have a termi-\nnation. It\u2019s part of my standard genetic \ncounseling for OB patients.\nAlthough at the time a doctor in Ohio could legally \nperform an abortion for a Down syndrome diagno-\nsis, Pamela thought the bill was in effect and coun-\nseled all of her patients that they would need to \ntravel out of state if they opted to terminate for that \nreason. While this is not uncertainty, it points to \nconfusion and misunderstanding stemming from the \npolitical context, where abortion legislation is \nfrequently proposed, passed, partially implemented, \nor not in effect. Moreover, Ohio\u2019s attorney general \nhas communicated inaccurate information about \nlegislation to the media (Bischoff 2022).\nLegal and institutional definitions of terms such \nas \u201clethal\u201d and \u201clife-threatening\u201d can be ambiguous \nin practice and may not correspond to clinical real-\nties. We often heard that institutions and providers \nerred toward more conservative interpretations. \nKaren, an ob-gyn at a private hospital, described \nhow when a patient presented hemorrhaging, the \nhospital and providers disagreed on when the case \nbecame \u201cemergent.\u201d Although the medical evi-\ndence left little doubt as to the outcome, the legal \nuncertainty required providers to stand by and let \nthe patient\u2019s condition worsen before intervening:\nIt was a very heated situation. We had to wait \u2019til \nher blood count was low enough to call it \nemergent. So, we waited \u2019til her hemoglobin was \nbelow seven . . . even at that point, there was a lot \nof back and forth, but that was legally through \nthe hospital. We had to wait until that point. So, \nthere\u2019s definitely a lot of frustration . . . because \n[providers] knew what they wanted to do, but \nthey couldn\u2019t legally do that.\nKaren described this as a case of medical but not \nlegal certainty. Although clinicians had clarity about \nthe likely course of the condition, the hospital \nrequired more extreme evidence of the life-threatening \ncircumstances to show they had met the law\u2019s \nrequirements.\nSimilarly, Carol, a genetic counselor, told us that \nafter legislature passed the Down syndrome bill, \nher hospital changed their policies to ban not only \nterminations for Down syndrome\u2014even though the \nregulation was not actually in effect\u2014but for any \n\u201cnon-lethal\u201d diagnoses. She described the sudden \nshift in her hospital\u2019s policy and the subsequent \nday-to-day challenges of determining what consti-\ntutes a \u201clethal\u201d diagnosis:\nIt was really unfortunate because we literally \nhad the surgery [to terminate a pregnancy with a \nDown syndrome diagnosis] planned for this \nparticular patient, and we had to call her the day \nbefore her surgery and say, \u201cOh sorry, we can\u2019t \ndo this for you anymore.\u201d . . . It makes it more \ndifficult for me to figure out, is this considered\u2014\nwell, for any genetic condition\u2014is this \nconsidered lethal enough? . . . Is my institution \nlikely to consider this lethal or not?\n\nCzarnecki et al.\t\n477\nCarol\u2019s hospital, which she characterized as \u201cpretty \nliberal,\u201d shifted their policy to be more conservative \nthan the law. When asked why she thought the pol-\nicy shifted to such an extreme, Carol explained:\nWell, the publicity generated from the enactment \nof that law prompted our department head to \nquestion why we were doing this, and then that \nled to legal getting involved, and I think with the \npolitical environment that\u2019s when things \nchanged and they [hospital leadership] started \ncracking down on us.\nCarol cited the political climate as shifting her hos-\npital\u2019s policy toward greater restriction than the law \nwould require, demonstrating the chilling effect of \nthe political context.\nSimilarly, institutions had different definitions \nof clinical and legal terms. Judy, an ob-gyn, had \nexperience working in two different public hospi-\ntals in two regions of the state. She described how \nthe 2011 public facilities ban for \u201cnon-therapeutic \nabortions\u201d applied to both institutions but \u201clegal \nhigher ups\u201d at each had different interpretations of \nwhat \u201cnon-therapeutic\u201d meant:\nIt\u2019s whether or not we can provide terminations \nin a government-supported hospital, basically. \nSome have chosen to interpret the law that if \nit\u2019s not immediately potentially lethal to the \nmom, we can\u2019t do a termination in the hospital \nwhereas other hospitals say, \u201cIf we can say that \nthis anomaly will not result in a live born baby \nthan any risk\u2014any pregnancy carries a risk\u2014\nthere\u2019s a risk to the mom then with no benefit to \nthe fetus, we can justify performing the \nprocedures.\u201d\nThe hospital legal departments\u2019 different interpreta-\ntions of \u201cnon-therapeutic\u201d resulted in differences in \nabortion provision, with one being more restrictive \nthan the other.\nHowever, we also heard from respondents who \ndrew on traditional notions of medical uncertainty \nto challenge their institutions\u2019 interpretations. Erin \ndescribed how clinicians challenged her hospital\u2019s \nlegal team\u2019s interpretation of \u201clife-threatening,\u201d \nwhich hinged on a degree of certainty of predicting \ndeath that was not possible:\nWe\u2019ve been back in front of our legal team and \nsaying these circumstances, you can never \nguarantee if anyone dies. . . I can\u2019t guarantee she \nis going to die. And so, we\u2019ve had a little bit \nstronger push the other way, with a legal team \nand they seem to then go that way.\nErin\u2019s experience of her hospital\u2019s shift points to \ninternal pressures from clinicians, hospital adminis-\ntrators, and ethics and legal teams that can shape \ninstitutional policy.\nReferral practices were also marked by uncer-\ntainty. In 2017, a gestational limit bill went into \neffect that bans abortion at 20-weeks postfertiliza-\ntion, which is 22 weeks since the person\u2019s last men-\nstrual period. The bill, often referred to as the \n\u201c20-week ban,\u201d itself suggests a stricter interpreta-\ntion of two fewer weeks than the law actually per-\nmits. Although the law is confusing, other factors \nmay impact an institution\u2019s gestational limit, such \nas fear of legal liability, lack of providers, staff, and \ninstitutional pressures. Genetic counselors debated \nthe gestational limit up to which their local abortion \nclinic provided care:\nMartha:\u00a0I think it\u2019s 17 and 6 and there\u2019s a consistent \nthree-week waiting period and they require one \nweek in between the first and second appoint-\nment . . .\nSarah:\u00a0[Patients] have to know by like 13 and 6 to \nmeet all of those requirements. . .\nI asked [the clinic] why they changed that from \n19 and 6 to less, and the person I spoke to on \nthe phone honestly couldn't give me an \nanswer. . . I haven\u2019t been able to get an actual \nanswer. I assumed it was because of legisla-\ntion, but it\u2019s the same physicians who do it at \n[hospital] and there's a different time limit \nthere.\nThis exchange shows uncertainty about the clinic\u2019s \ngestational limit, confusion about the legally \nrequired wait time between abortion counseling and \nprovision appointments (24 hours in Ohio), and \nwhether the clinic\u2019s gestational limit changed due to \nlegislation or institutional factors. There is not only \nuncertainty about one\u2019s own institution but also \nwhere patients are referred for care.\nRespondents noted that local political climate \nshaped hospital policy across the state. Mary, an ob-\ngyn working in southwest Ohio, explained that her \nhospital\u2019s interpretation of the law is more conser-\nvative than other regions of the state:\nTalking to other providers at other locations, \n[they don\u2019t seem] to be as restrictive. . . I think \nthe law is restrictive, but our institution\u2019s \n\n478\t\nJournal of Health and Social Behavior 64(4) \ninterpretation is so severely conservative \n[compared to] other parts of the state.\nPatients present at hospitals with complicated \npregnancies. Ob-gyns discussed variation in the \ntreatment of one such complication, premature \nmembrane rupture, which can lead to sepsis. Andrea, \nan ob-gyn, grimly stated, \u201cI\u2019ve had one recently that \nwas 16 weeks, no fluid on ultrasound, and I had to \nlook at the patient and say, \u2018I\u2019m afraid that you\u2019re \ngoing to get septic and die,\u2019 but I can\u2019t do anything \nabout it.\u201d Others shared varied institutional interpre-\ntations for such cases that spanned from permitting \ncare, waiting until a patient becomes septic, waiting \nuntil patients are near death (although still uncertain \nabout whether they could provide care), sending \npatients out of state when a \u201cheartbeat\u201d is detected, \nand requiring thorough documentation.\nEven hospitals considered to be less conserva-\ntive imposed hurdles to providing care. A group of \nob-gyn residents, who described their institution as \nless conservative, elaborated:\nWe don\u2019t do elective abortion, so there needs to \nbe an agreement between a couple clinicians that \nit\u2019s indicated, then the ethics department has to \nbe called, and it is frequently more of a formality. \n. . . It\u2019s not a barrier, but it\u2019s a check in the system. \n. . . They\u2019ll ask for discussion between usually \ntwo different providers, oftentimes two different \nmaternal fetal medicine [MFM] physicians, or \nan ob-gyn and a MFM, and then the Department \nChair, and an ethicist . . . we have to do a lot of \ndocumentation. . . If it\u2019s in that gray zone, then \ngetting other physicians to also approve their \ndecision . . . sometimes it\u2019s consulting other \nproviders from other specialties like cardiology \nor rheumatology to weigh in on how this \npregnancy would impact this patient.\nThe hospital sought to mitigate uncertainty through \na convoluted process of approvals and documenta-\ntion about how pregnancy would affect a patient \nphysically. In doing so, they undermined profes-\nsional autonomy and judgment and contributed to \nabortion stigma.\nAs we have shown, a source of uncertainty in \nclinical settings is that institutions across the state \nhave variable policies, with many being more limit-\ning than the law. Ad hoc institutional guidance that \nconflicts with medical expertise constrained care \nand endangered patients.\n\u201cI Don\u2019t Want to Go to Jail\u201d: Uncertain \nConsequences for Providing Care\nUncertainty in reproductive health care settings out-\nside of abortion clinics includes the potential legal, \npersonal, and professional consequences for \ninvolvement in abortion care. Uncertainty about \nlegal restrictions occurs within a climate of fear and \nstigma surrounding abortion care in restrictive con-\ntexts that impact ob-gyns and genetic counselors \ndifferently. Genetic counselors and ob-gyns are dif-\nferently positioned in their proximity to abortion \ncare. Ob-gyns may find themselves directly \ninvolved with the provision of abortion care. If they \nviolate the law, they are subject to civil and criminal \nprosecution, along with other risks such as loss of \nlivelihood and exposing themselves and their fami-\nlies to risk of violence. At the time of this research, \nthis played out differently for genetic counselors, \nwho voiced their profession\u2019s concerns with profes-\nsional legitimacy and public distancing from abor-\ntion. All respondents tried to adhere to the law \nwithout exception.\nUncertainty about consequences surfaced when \nclinicians were legally or institutionally obligated \nto provide or withhold forms of care that might \nresult in increasing medical risk. For instance, ob-\ngyns expressed concern that the D&E ban would \n\u201crequire us to perform procedures [to cause fetal \ndemise] that have no medical benefit to the woman \nand potentially have a risk.\u201d The law requires that \nphysicians induce fetal demise prior to a D&E \nunless maternal life or health are at risk. This \nrequirement introduces invasive procedures to \ncause fetal demise despite no evidence of a benefit \nfor this approach as standard practice (Guttmacher \nInstitute 2017). At the time of the study, providers \nfound to be in violation of the law could face a \nfourth-degree felony charge. Charlotte, an ob-gyn, \nshared her fear:\nWhat worries me about . . . the felony [charge] is \nthat we might allow a woman to hemorrhage to \ndeath, right? Because we\u2019re so afraid . . . I\u2019m \nscared to do a D&E on someone coming in \nhemorrhaging with her second trimester loss. \nAnd I am going to want to be self-serving, and \nsay, \u201cOh well, I am scared . . . I don\u2019t want to go \nto jail. What are my kids going to do when I\u2019m in \njail?\u201d and that is horrible. I have been trained to \nserve society, and I should do that. To save this \nwoman\u2019s life because I have the skills to.\n\nCzarnecki et al.\t\n479\nFulfilling a medical obligation to a patient could \nplace a clinician\u2019s professional and personal well-\nbeing in jeopardy. Conversely, not providing essen-\ntial care could put a patient\u2019s life in danger.\nSimilarly, another ob-gyn, Emily, explained the \nrole of fear and uncertainty about the consequences \nof abortion provision in an ever-changing policy \ncontext:\nIf you know that overnight that there could be a \nnew law in place that you could be inadvertently \nbreaking, I think that is a deterrent for lots of \nproviders who . . . \u201cI\u2019m not aware of the laws, I \ndon\u2019t feel comfortable, and because of that I\u2019m \nnot going to do this\u201d . . . we all take an oath that \nwe are going to take care of people and we all do, \nand I think none of us would sit there and \nnecessarily let someone hemorrhage to death, \nbut at the same time, if your livelihood, your \n[medical] license, your family, and potentially \nyour freedom is all on the line, you\u2019re not going \nto provide for a patient, unless they\u2019re actively \ndying in front of you, a procedure that you\u2019re not \nsure is legal . . . because we do see these laws \npop up out of nowhere.\nThe chilling effect of the political context of care \nincludes the cumulative toll of restrictions, uncer-\ntainty, and constraints on the ability to provide care \nthat leads some to avoid abortion provision alto-\ngether, even for those who support abortion. For \nexample, Anna shifted her earlier conservative \nviews on abortion during her medical training and \nconsidered herself to be one of the most well-trained \nabortion providers in Ohio. She had worked at a \nhospital that was particularly restrictive. Despite her \nsupport of abortion, she explained that it was easier \nto work there because she did not have to closely \nfollow the changing legislation:\nIn honesty, I\u2019m probably (pause) not proud to \nsay that there\u2019s a sense of, just acceptance of this \nkind of environment . . . [restrictions are] \nchanging all the time, it\u2019s so hard to know what\u2019s \ngoing on with the laws, it can be really hard to \nknow even what battles to fight if you wanted to. \nSo, I think . . . it\u2019s easier to just work at a place \nthat doesn't allow any abortions and kind of stay \nout of the fray.\nAnna and others were candid about feeling worn \ndown and fearful from the onslaught of restrictions \nand wanting to avoid uncertainty. For some, more \nrestrictive institutional interpretations were viewed \nas providing \u201cblack and white\u201d rules that could \nlessen the moral and ethical ambiguities uncertainty \ncaused and the potential legal and professional \nrepercussions. But the cost of this was reinforcing \ninstitutional restrictions on patients\u2019 access to abor-\ntion and institutional buck-passing (Freedman \n2010). A few ob-gyns challenged their institutions \nand were able to shift policy. But there is consider-\nable risk that leads others to \u201cstay out of the fray.\u201d\nGenetic counselors are not directly involved \nwith abortion provision, which they perceived as \nproviding a degree of protection from prosecution. \nYet they still experienced uncertainty about being \nimplicated. Susan explained:\nI feel there is some protection from me being a \ngenetic counselor who currently doesn't bill \nunder my name because . . . I\u2019m not doing the \nprocedure. So, I can be an advocate for my \npatient and tell them this is what we know, this is \nwhat we don\u2019t know . . . I do feel some protection \nat the end of the day that should there ever be \nsome sort of scenario in which a law was broken \n. . . I would hope that I would not be liable for \nanything, but I don\u2019t know. . . My name is not on \nthe letter, so I hope . . . I don\u2019t know; this is scary. \nIt\u2019s a scary, scary thing to be in.\nSusan illustrates how uncertainty and a climate of \nfear also pervade professions adjacent to abortion \ncare. Several laws are written so that who can be \nprosecuted remains vague, furthering uncertainty, \nfear, and the reach of restrictive legislation. Genetic \ncounselors could not distance themselves com-\npletely from abortion care and the impact of restric-\ntive legislation and uncertainty.\nWe have shown that politically produced uncer-\ntainty is burdensome for health care providers. \nThose hoping to evade such uncertainty have few \noptions that would not further limit abortion access. \nUltimately, we have shown that the political context \nproduces uncertainty and fear that are far-reaching \nand together create a chilling effect that restricts \nabortion beyond the law.\nDiscussion\nAbortion is one of the most common and politicized \nmedical procedures. Although prior scholars have \nidentified sources of uncertainty that arise from \nwithin medical contexts (Atkinson 1984; Fox 2000; \nGerrity et al. 1992; Light 1979), abortion is \n\n480\t\nJournal of Health and Social Behavior 64(4) \nillustrative of how the politicization of care can also \ncreate uncertainty. Our study shows how an influx \nof abortion legislation produced uncertainty for ob-\ngyns and genetic counselors in hospital settings. \nAbortion\u2014a relatively uncomplicated procedure or \nmedication protocol\u2014was not the source of uncer-\ntainty. Rather, respondents reported uncertainty \nabout legality, institutional policy, and the conse-\nquences of providing care in an abortion-restrictive \nstate. Research on abortion restrictions has illumi-\nnated the detrimental impact on access, costs, and \nstigma (Fuentes et al. 2016; Grossman and White \n2018; White et al. 2019). We show that uncertainty \ncan also function to restrict abortion beyond what \nthe law dictates. Our findings describe an additional \nsource of medical uncertainty and contribute to \nunderstanding the politics of abortion by showing \nthat the political context of care is a source of medi-\ncal uncertainty.\nPrior research establishes that cultural and tech-\nnological change may contribute to medical uncer-\ntainty; we show that the politicization of care may \nalso contribute to uncertainty. We identify at least \nthree ways that the politicization of abortion con-\ntributed to uncertainty. First, a stark increase in \nabortion-restrictive legislation and litigation over \nthe past decade left our respondents unsure about \nthe status of bills and laws. We also found that \npending and enjoined legislation nevertheless \nshaped hospital policies. Second, despite profes-\nsional enthusiasm for standardized, evidence-based \nmedicine, hospitals across the state have different \ninterpretations of legislation. Some clinicians prac-\ntice at multiple institutions, and policies can differ \nbetween institutions and regionally. Among respon-\ndents who reported their institution type, we found \nthat policy differences were not reducible to institu-\ntion type. Because funding sources and the compo-\nsition of hospital organizations are complex, some \nrespondents were unsure of their institution type or \nthe rationale for some institutional policies. Such \nlack of understanding is important because certain \nlaws apply to specific kinds of institutions. \nVariation in policy across institutions and regions \ncan contribute to uncertainty in referral practices as \nwell. A recent study found that half of physicians \nacross different specialties did not know how to \nrefer patients for abortion (Anderson et al. 2022). \nTurning to others to mitigate uncertainty is limited \nin this context because policies vary and colleagues \nmay have inaccurate understandings of them. \nAttempts to resolve uncertainty may also bring \nunwanted attention that constrains practice. Third, \nthere is uncertainty about the legal, personal, and \nprofessional consequences for abortion care provi-\nsion, where criminal and civil penalties may apply. \nThis parallels Morris\u2019s (2014) study on caesarian \nsections showing that practice changes resulted \nfrom hospitals\u2019 concern about liability. The political \ncontext of care may derail the broader trend toward \nevidence-based medicine when laws and policies \nare not grounded in medicine (Timmermans and \nBerg 2010).\nUncertainty from the political context of care is \nsignificant because it profoundly impacts patients \nand health care providers. Although uncertainty is \ninherent to medicine, clinicians do not typically \nwonder whether providing evidence-based medi-\ncine will result in a felony charge or loss of medical \nlicensure. Politically produced uncertainty prevents \nhealth care professionals from providing medically \nappropriate care, compromising their professional \nobligations (Field et al. 2022). Relatively straight-\nforward protocols, such as managing ectopic preg-\nnancies and miscarriages, become fraught with \nuncertainty. Medical exceptions to abortion laws \nconfuse more than clarify when criteria are not \ndefined in medical terms and are unlikely to capture \nthe full range of medical circumstances (Arey et al. \n2022). Uncertainty from politicizing abortion care \nimpinges on both patient and provider autonomy \nbecause law and politics dictate medical practice \nrather than medical expertise. Strict abortion bans \nmay also contribute to uncertainty about what con-\nstitutes an abortion and clinical hesitation to inter-\nvene (Zernike 2022).\nUncertainty has worsened in a post-Dobbs con-\ntext, as the legality of abortion in many states has \nbeen in flux (McCann 2022). On June 24, 2022, the \nday that Roe was overturned, Ohio banned abortion \nprovision after detection of embryonic cardiac activ-\nity, with a very narrow and poorly understood excep-\ntion for patients with serious medical conditions. \nShortly after, a 10-year-old rape survivor in Ohio \nwas forced to travel to Indiana for abortion care. \nOhio\u2019s attorney general then incorrectly stated that it \nwould have been legal to have the abortion in Ohio at \nthe time (Bischoff 2022). The embryonic cardiac \nactivity ban does not include a rape exception (Ohio \nRevised Code 2019). The state\u2019s leading legal \nauthority providing misinformation about state law \nunderscores the challenges that politicizing health \ncare presents for those providing and seeking care.\nPrior studies show that uncertainty can reinforce \nphysician authority (e.g., Littlejohn and Kimport \n2017). We found that uncertainty can also under-\nmine authority by shifting the litmus for decision-\nmaking \nfrom \nmedical \nevidence \nto \nlegal \n\nCzarnecki et al.\t\n481\ninterpretations. The political context of care can \nerode physicians\u2019 autonomy as they face criminal, \nprofessional, and personal threats to their lives \nwhen regulatory policies conflict with medical \nexpertise (Field et al. 2022). Fear about the poten-\ntial consequences of providing care in an ever-\nchanging policy context could result in ob-gyn \nreluctance to provide abortion care at all, demon-\nstrating the chilling effect of politicized health care. \nRespondents were frustrated that laws not grounded \nin medicine interfere with their obligation to patients. \nIn contrast to studies showing that vague abortion \npolicies may allow for greater physician discretion \n(Zeldovich et al. 2020), we found that vagueness \noften intensified uncertainty in a context where the \nthreat of criminalization looms. Overinterpretation \nof the law was the norm in Ohio, further demon-\nstrating the myriad ways in which a state\u2019s political \nclimate can impede reproductive health promotion \n(Blake and McGowan 2020).\nAlthough genetic counselors voiced less fear \nabout repercussions because they are not perform-\ning abortions, they nevertheless expressed uncer-\ntainty about liability for facilitating care. With the \npassage of abortion bans for reasons related to fetal \ncharacteristics, genetic counselors will be further \ninvolved in navigating legislation (Heuerman et al. \n2022). Moreover, anti-abortion bills inspired by \nTexas\u2019s SB8 empower ordinary citizens to sue any-\none facilitating abortion, potentially implicating an \narray of health care professionals and exacerbating \nuncertainty. Future research should consider the \nconditions under which uncertainty erodes or \nexpands medical authority.\nThe political context of care may produce uncer-\ntainty \nwherever \nmedical \ncare \nis \ncontested. \nTransgender health care, an area already marked by \nuncertain expertise, has parallels with the politiciza-\ntion of abortion (shuster 2016). A Texas hospital \nrecently halted gender-affirming therapies out of fear \nof potential ramifications after a politician equated \ngender-affirming care to \u201cchild abuse\u201d (Goodman \n2022). Targeting clinicians providing gender- \naffirming care echoes the anti-abortion movement\u2019s \ntargeting of abortion providers, fueling a climate of \nfear. Infertility clinicians have also expressed con-\ncern that post-Roe personhood bills may extend to \nIVF embryos, generating uncertainty in fertility \nmedicine (Hoffman 2022). Although our study \nfocuses on a shifting legislative and policy context, \npolitical context encompasses more than legislation \nand policy. For example, the politicization of the \nCOVID-19 pandemic contributed to uncertainty \nabout prevention, treatments, and vaccines (Abbas \n2022; Kreps and Kriner 2020). Uncertainty arising \nfrom the political context may have negative impli-\ncations for patient trust in medicine as expertise and \nevidence is undermined or criminalized.\nConclusion\nBecause the landscape of abortion care in the United \nStates is expected to continue to change, future \nresearch should examine how restrictive policies \ncontribute to uncertainty in various clinical settings. \nPeople unable to access timely abortion care may \npresent at hospital settings later in pregnancy with \nmore complicated care needs (Arey et al. 2022). \nBecause the use of self-managed abortion will likely \nincrease and may be further criminalized, emer-\ngency medicine is poised to become a setting \nmarked by uncertainty related to abortion care and \nlegality (Harris and Grossman 2020). The relation-\nship between medicine and the law in different care \ncontexts will be important to examine. A recent \nstudy on drug testing during pregnancy identified \nhow medical practice became a form of criminaliza-\ntion and mode of governance (McCabe 2022). In \nour study, criminalization of health care providers \nwas a greater concern given current state legislation. \nHow reproductive health care providers navigate \nthe shifting legislative context and how patients will \nbe impacted will be important to examine.\nOur findings have implications for advocacy \nand policymaking. Lack of clear information about \nlegislation and how it applies to patient care was a \nprominent theme from our focus groups. Professional \nsocieties and institutions should ensure that health \ncare providers are equipped with information about \nhow legislation impacts care provision, including \nreferral practices (Heuerman et al. 2022). Health \ncare providers may also require knowledge about \nthe policies and nuances of out-of-state referrals \n(Goyal et al. 2022). Providers will be better able to \ncare for their patients if they have guidance about \nhow to navigate legislation that may not be medi-\ncally legible in clinical practice. Professional soci-\neties \nand \nadvocacy \norganizations \npreparing \nrecommendations for providers should also con-\nsider that institutional policies can vary and contrib-\nute to uncertainty. The factors discussed here are \nnot inherent complications of abortion care\u2014they \nare politically produced complexities that impact \nhealth care providers and patients. By identifying \nthe political context of care as a source of uncer-\ntainty in medical practice, we expand our under-\nstanding of both medical uncertainty and barriers to \nabortion access.\n\n482\t\nJournal of Health and Social Behavior 64(4) \nAcknowledgments\nThe authors thank Jessie Hill, JD, for providing guidance \non understanding and discussing Ohio legal regulations. \nWe also thank Dr. Meredith Pensak for her insights on \nabortion provision in Ohio; and Cynthia Beavin, Dyanna \nBurnham, and Colleen Kavanagh for their research assis-\ntance. We are grateful to the editors and three reviewers \nfor their insightful feedback.\nFunding\nThe authors disclosed receipt of the following financial \nsupport for the research, authorship, and/or publication of \nthis article: Data collection was supported by a grant from \nan anonymous foundation.\nORCID iD\nDanielle Czarnecki \n https://orcid.org/0000-0002-7299-9074\nNotes\n\u2002 1.\t Dilation and evacuation (D&E) involves dilating \nthe cervix and removing pregnancy tissue from the \nuterus. It is used for terminations and miscarriage \nmanagement after 15 weeks\u2019 gestation.\n\u2002 2.\t \u201cEnjoined\u201d refers to laws that were passed but are \nnot currently in effect due to a court order. \u201cPartially \nenjoined\u201d means some aspects of the law are in effect.\nReferences\nAbbas, Ali Haif. 2022. \u201cPoliticizing COVID-19 Vaccines in \nthe Press: A Critical Discourse Analysis.\u201d International \nJournal for the Semiotics of Law - Revue Internationale \nde S\u00e9miotique Juridique 35(3):1167\u201385.\nAlderson, Priscilla. 2009. \u201cUncertainty in Medical \nInnovation: Experienced Pioneers in Neonatal Care.\u201d \nSociology of Health & Illness 31(7):1111\u201312.\nAnderson, Elizabeth M., Sarah K. Cowan, Jenny A. \nHiggins, Nicholas B. Schmuhl, and Cynthie K. \nWautlet. 2022. \u201cWilling but Unable: Physicians\u2019 \nReferral Knowledge as Barriers to Abortion Care.\u201d \nSSM - Population Health 17:101002. doi:10.1016/j.\nssmph.2021.101002.\nArey, Whitney, Klaira Lerma, Anitra Beasley, Lorie \nHarper, Ghazaleh Moayedi, and Kari White. 2022. \n\u201cA Preview of the Dangerous Future of Abortion \nBans\u2014Texas Senate Bill 8.\u201d New England Journal \nof Medicine 387(5):388\u201390.\nAtkinson, Paul. 1984. \u201cTraining for Certainty.\u201d Social \nScience & Medicine 19(9):949\u201356. doi:10.1016/ \n0277-9536(84)90324-1.\nBarker, Kristin K. 2008. \u201cElectronic Support Groups, \nPatient-Consumers, and Medicalization: The Case \nof Contested Illness.\u201d Journal of Health and Social \nBehavior 49(1):20\u201336.\nBegin, Arabella Simpkin, Michael Hidrue, Sara Lehrhoff, \nMarcela G. del Carmen, Katrina Armstrong, \nand Jason H. Wasfy. 2022. \u201cFactors Associated \nwith Physician Tolerance of Uncertainty: An \nObservational Study.\u201d Journal of General Internal \nMedicine 37(6):1415\u201321.\nBessett, Danielle, Caitlin Gerdts, Lisa L. Littman, Megan \nL. Kavanaugh, and Alison Norris. 2015. \u201cDoes State-\nLevel Context Matter for Individuals\u2019 Knowledge \nabout Abortion, Legality and Health? Challenging \nthe \u2018Red States v. Blue States\u2019 Hypothesis.\u201d Culture, \nHealth & Sexuality 17(6):733\u201346.\nBischoff, Laura A. 2022. \u201cOhio A.G. Dave Yost Cast \nDoubt on 10-Year-Old Rape Victim Case, Now \n\u2018Rejoices\u2019 at Arrest.\u201d The Columbus Dispatch, \nAugust \n21. \nhttps://www.dispatch.com/story/\nnews/2022/07/13/ohio-attorney-general-rejoices-\narrest-child-rape-suspect/10048250002/.\nBlake, Valarie K., and Michelle L. McGowan. 2020. \n\u201cFilling a Federal Void: Promises and Perils of State \nLaw in Addressing Women\u2019s Health Disparities.\u201d \nThe Journal of Law, Medicine & Ethics 48(3):485\u2013\n90.\nBrookes-Howell, Lucy Claire. 2006. \u201cLiving without \nLabels: The Interactional Management of Diagnostic \nUncertainty in the Genetic Counselling Clinic.\u201d \nSocial Science & Medicine 63(12):3080\u201391.\nBuchbinder, Mara. 2016. \u201cScripting Dissent: U.S. \nAbortion Laws, State Power, and the Politics \nof Scripted Speech.\u201d American Anthropologist \n118(4):772\u201383.\nCohen, David S., and Carole Joffe. 2021. Obstacle \nCourse: The Everyday Struggle to Get an Abortion \nin America. Berkeley: University of California Press.\nCreswell, Julie, and Reed Abelson. 2013. \u201cNew Laws \nand Rising Costs Create a Surge of Supersizing \nHospitals.\u201d The New York Times, August 12.\nCzarnecki, Danielle, Renee R. Anspach, Raymond G. De \nVries, Mercedez D. Dunn, Katrina Hauschildt, and \nLisa H. Harris. 2019. \u201cConscience Reconsidered: \nThe Moral Work of Navigating Participation in \nAbortion Care on Labor and Delivery.\u201d Social \nScience & Medicine 232:181\u201389.\nDumit, Joseph. 2006. \u201cIllnesses You Have to Fight to Get: \nFacts as Forces in Uncertain, Emergent Illnesses.\u201d \nSocial Science & Medicine 62(3):577\u201390.\nEmerson, Robert M., Rachel I. Fretz, and Linda L. Shaw. \n2011. Writing Ethnographic Fieldnotes. 2nd ed. \nChicago, IL: University of Chicago Press.\nFarnan, Jeanne M., J. K. Johnson, D. O. Meltzer, H. \nJ. Humphrey, and V. M. Arora. 2008. \u201cResident \nUncertainty in Clinical Decision-Making and Impact \non Patient Care: A Qualitative Study.\u201d BMJ Quality \n& Safety 17(2):122\u201326.\nField, Meredith P., Hillary Gyuras, Danielle Bessett, \nMeredith J. Pensak, Alison H. Norris, and Michelle \nL. McGowan. 2022. \u201cOhio Abortion Regulations and \nEthical Dilemmas for Obstetrician-Gynecologists.\u201d \nObstetrics & Gynecology 140(2):253\u201361.\n\nCzarnecki et al.\t\n483\nFox, Renee C. 1957. \u201cTraining for Uncertainty.\u201d Pp. 207\u2013\n42 in The Student-Physician: Introductory Studies in \nthe Sociology of Medical Education, edited by R. K. \nMerton, G. Reader, and P. L. Kendall. Cambridge, \nMA: Harvard University Press.\nFox, Renee C. 1980. \u201cThe Evolution of Medical \nUncertainty.\u201d The Millbank Memorial Fund Quarterly \n58(1):1\u201349.\nFox, Renee C. 2000. \u201cMedical Uncertainty Revisited.\u201d \nPp. 409\u201325 in The Handbook of Social Studies in \nHealth and Medicine, edited by G. L. Albrecht,  \nR. Fitzpatrick, and S. C. Scrimshaw. London: SAGE.\nFreedman, Lori. 2010. Willing and Unable: Doctors\u2019 \nConstraints in Abortion Care. Nashville, TN: \nVanderbilt University Press.\nFreedman, Lori R., and Debra B. Stulberg. 2013. \n\u201cConflicts in Care for Obstetric Complications \nin Catholic Hospitals.\u201d AJOB Primary Research \n4(4):1\u201310.\nFuentes, Liza, Sharon Lebenkoff, Kari White, Caitlin \nGerdts, Kristine Hopkins, Joseph E. Potter, and \nDaniel Grossman. 2016. \u201cWomen\u2019s Experiences \nSeeking Abortion Care Shortly after the Closure \nof Clinics Due to a Restrictive Law in Texas.\u201d \nContraception 93(4):292\u201397.\nGallo, Maria F., John B. Casterline, Payal Chakraborty, \nAlison Norris, Danielle Bessett, and Abigail Norris \nTurner. 2021. \u201cPassage of Abortion Ban and \nWomen\u2019s Accurate Understanding of Abortion \nLegality.\u201d American Journal of Obstetrics and \nGynecology 225(1):63.e1\u2013e8\nGeampana, Alina. 2019. \u201c\u2018One Blood Clot Is One Too \nMany\u2019: Affected Vocal Users\u2019 Negative Perspectives \non Controversial Oral Contraceptives.\u201d Qualitative \nHealth Research 29(10):1519\u201330.\nGerrity, Martha S., Robert F. DeVellis, and Donald W. \nLight. 1992. \u201cUncertainty and Professional Work: \nPerceptions of Physicians in Clinical Practice.\u201d \nAmerican Journal of Sociology 97(4):1022\u201351.\nGinsburg, Faye D. 1998. Contested Lives: The Abortion \nDebate in an American Community. Updated ed. \nBerkeley: University of California Press.\nGoodman, J. David. 2022. \u201cHow Medical Care for \nTransgender Youth Became \u2018Child Abuse\u2019 in \nTexas.\u201d The New York Times, March 11.\nGoyal, Vinita, Rosann Mariappuram, Anna Rupani, and \nKari White. 2022. \u201cTexas Senate Bill 8: Implications \non the Ground and Lessons Learned.\u201d Panel \nPresentation at the National Abortion Federation\u2019s \n45th Annual Meeting, Orlando, FL, May 2.\nGrossman, Daniel., and Kari White. 2018. \u201cAbortion \n\u2018Reversal\u2019\u2014Legislating without Evidence.\u201d The \nNew England Journal of Medicine 379(16):1491\u201393.\nGuttmacher Institute. 2017. \u201cD&E Abortion Bans: The \nImplications of Banning the Most Common Second-\nTrimester Procedure.\u201d https://www.guttmacher.org/\ngpr/2017/02/de-abortion-bans-implications-banning-\nmost-common-second-trimester-procedure.\nGuttmacher Institute. 2018. \u201cState Abortion Policy \nLandscape: From Hostile to Supportive.\u201d https://\nwww.guttmacher.org/article/2019/08/state-abortion-\npolicy-landscape-hostile-supportive.\nGuttmacher Institute. 2019. \u201cState Facts about Abortion: \nOhio.\u201d https://www.guttmacher.org/fact-sheet/state-\nfacts-about-abortion-ohio.\nGuttmacher Institute. 2022. \u201c100 Days Post-Roe: At \nLeast 66 Clinics across 15 U.S. States Have Stopped \nOffering Abortion Care.\u201d https://www.guttmacher.\norg/2022/10/100-days-post-roe-least-66-clinics-\nacross-15-us-states-have-stopped-offering-abortion-\ncare.\nGyuras, Hillary J., Meredith P. Field, Olivia Thornton, \nDanielle Bessett, and Michelle L. McGowan. 2023. \n\u201cThe Double-Edged Sword of Abortion Regulations: \nDecreasing Training Opportunities while Increasing \nKnowledge Requirements.\u201d Medical Education \nOnline 28(1):2145104. doi:10.1080/10872981.2022\n.2145104.\nHarris, \nLisa \nH., \nand \nDaniel \nGrossman. \n2020. \n\u201cComplications of Unsafe and Self-Managed \nAbortion.\u201d New England Journal of Medicine \n382(11):1029\u201340.\nHarris, Lisa H., Lisa Martin, Michelle Debbink, and Jane \nHassinger. 2013. \u201cPhysicians, Abortion Provision, \nand the Legitimacy Paradox.\u201d Contraception \n87(1):11\u201316.\nHeuerman, Anne C., Danielle Bessett, Armand H. \nMatheny Antommaria, Leandra K. Tolusso, Nicki \nSmith, Alison H. Norris, and Michelle L. McGowan. \n2022. \n\u201cExperiences \nof \nReproductive \nGenetic \nCounselors with Abortion Regulations in Ohio.\u201d \nJournal of Genetic Counseling 31(3):641\u201352.\nHeymann, Orlaith, Tamika Odum, Alison H. Norris, \nand Danielle Bessett. 2022. \u201cSelecting an Abortion \nClinic: The Role of Social Myths and Risk Perception \nin Seeking Abortion Care.\u201d Journal of Health and \nSocial Behavior 63(1):90\u2013104.\nHoffman, Jan. 2022. \u201cInfertility Patients and Doctors \nFear Abortion Bans Could Restrict I.V.F.\u201d The New \nYork Times, July 5.\nJenkins, Richard, Hanne Jessen, and Vibeke Steffen. \n2005. Managing Uncertainty: Ethnographic Studies \nof Illness, Risk, and the Struggle for Control. \nCopenhagen, Denmark: Museum Tusculanum Press.\nJutel, Annemarie. 2009. \u201cSociology of Diagnosis: A \nPreliminary Review.\u201d Sociology of Health & Illness \n31(2):278\u201399.\nKimport, Katrina. 2021. No Real Choice: How Culture \nand Politics Matter for Reproductive Autonomy. \nNew Brunswick, NJ: Rutgers University Press.\nKreps, Sarah E., and Douglas L. Kriner. 2020. \u201cModel \nUncertainty, Political Contestation, and Public \nTrust in Science: Evidence from the COVID-19 \nPandemic.\u201d Science Advances 6(43):eabd4563. \ndoi:10.1126/sciadv.abd4563.\nLight, Donald. 1979. \u201cUncertainty and Control in \nProfessional Training.\u201d Journal of Health and Social \nBehavior 20(4):310\u201322.\nLittlejohn, Krystale E., and Katrina Kimport. 2017. \n\u201cContesting \nand \nDifferentially \nConstructing \n\n484\t\nJournal of Health and Social Behavior 64(4) \nUncertainty: Negotiations of Contraceptive Use \nin the Clinical Encounter.\u201d Journal of Health and \nSocial Behavior 58(4):442\u201354.\nLuker, Kristin. 1985. Abortion and the Politics of \nMotherhood. Berkeley: University of California \nPress.\nMackintosh, Nicola, and Natalie Armstrong. 2020. \n\u201cUnderstanding and Managing Uncertainty in \nHealth Care: Revisiting and Advancing Sociological \nContributions.\u201d Sociology of Health & Illness \n42(S1):1\u201320.\nMarkens, Susan, C. H. Browner, and H. Mabel Preloran. \n2003. \u201c\u2018I\u2019m Not the One They\u2019re Sticking the \nNeedle into\u2019: Latino Couples, Fetal Diagnosis, and \nthe Discourse of Reproductive Rights.\u201d Gender & \nSociety 17(3):462\u201381.\nMcCabe, Katharine. 2022. \u201cCriminalization of Care: \nDrug Testing Pregnant Patients.\u201d Journal of Health \nand Social Behavior 63(2):162\u201376.\nMcCann, Allison. 2022. \u201c\u2018Chaos and Confusion\u2019 in \nStates Where Abortion Is on Again, off Again.\u201d The \nNew York Times, August 11.\nMcGowan, Michelle L., Alison H. Norris, and Danielle \nBessett. 2020. \u201cCare Churn\u2014Why Keeping Clinic \nDoors Open Isn\u2019t Enough to Ensure Access to \nAbortion.\u201d The New England Journal of Medicine \n383(6):508\u201310.\nMcLemore, Monica R., Susan Kools, and Amy J. Levi. \n2015. \u201cCalculus Formation: Nurses\u2019 Decision-\nMaking in Abortion-Related Care.\u201d Research in \nNursing & Health 38(3):222\u201331.\nMesman, Jessica. 2008. Uncertainty in Medical \nInnovation: Experienced Pioneers in Neonatal Care. \nNew York, NY: Springer.\nMorris, Theresa. 2014. \u201cC-Section Epidemic.\u201d Contexts \n13(1):70\u201372.\nNorris, Alison H., Payal Chakraborty, Kaiting Lang, \nRobert B. Hood, Sarah R. Hayford, Lisa Keder, \nDanielle Bessett, et al. 2020. \u201cAbortion Access in \nOhio\u2019s Changing Legislative Context, 2010\u20132018.\u201d \nAmerican Journal of Public Health 110(8):1228\u201334.\nOhio Revised Code. 2019. \u201cOhio Revised Code. Ch. 26, \n\u00a7 2919.192. July 2019.\u201d https://codes.ohio.gov/ohio-\nrevised-code/section-2919.192.\nPoteat, Tonia, Danielle German, and Deanna Kerrigan. \n2013. \u201cManaging Uncertainty: A Grounded Theory \nof Stigma in Transgender Health Care Encounters.\u201d \nSocial Science & Medicine 84:22\u201329.\nRafalovich, \nAdam. \n2013. \n\u201cAttention \nDeficit-\nHyperactivity Disorder as the Medicalization of \nChildhood: Challenges from and for Sociology.\u201d \nSociology Compass 7(5):343\u201354.\nRapp, Rayna. 2004. Testing Women, Testing the Fetus: \nThe Social Impact of Amniocentesis in America. New \nYork, NY: Routledge.\nRothman, Barbara Katz. 1993. The Tentative Pregnancy: How \nAmniocentesis Changes the Experience of Motherhood. \nNew York, NY: W. W. Norton & Company.\nshuster, stef m. 2016. \u201cUncertain Expertise and the \nLimitations of Clinical Guidelines in Transgender \nHealthcare.\u201d Journal of Health and Social Behavior \n57(3):319\u201332.\nSteinauer, Jody E., Jema K. Turk, Tali Pomerantz, \nKristin Simonson, Lee A. Learman, and Uta Landy. \n2018. \u201cAbortion Training in U.S. Obstetrics and \nGynecology \nResidency \nPrograms.\u201d \nAmerican \nJournal of Obstetrics & Gynecology 219(1):86.\ne1\u2013e6.\nStulberg, Debra B., Rebecca A. Jackson, and Lori R. \nFreedman. 2016. \u201cReferrals for Services Prohibited \nin Catholic Health Care Facilities.\u201d Perspectives on \nSexual and Reproductive Health 48(3):111\u201317.\nTimmermans, Stefan, and Alison Angell. 2001. \n\u201cEvidence-Based Medicine, Clinical Uncertainty, \nand Learning to Doctor.\u201d Journal of Health and \nSocial Behavior 42(4):342\u201359.\nTimmermans, Stefan, and Marc Berg. 2010. The Gold \nStandard: \nThe \nChallenge \nof \nEvidence-Based \nMedicine. Philadelphia, PA: Temple University \nPress.\nTimmermans, Stefan, and Mara Buchbinder. 2010. \n\u201cPatients-in-Waiting: Living between Sickness and \nHealth in the Genomics Era.\u201d Journal of Health and \nSocial Behavior 51(4):408\u201323.\nWhite, Kari, Sarah E. Baum, Kristine Hopkins, Joseph \nE. Potter, and Daniel Grossman. 2019. \u201cChange in \nSecond-Trimester Abortion after Implementation of \na Restrictive State Law.\u201d Obstetrics and Gynecology \n133(4):771\u201379.\nWhite, Kari, Victoria deMartelly, Daniel Grossman, \nand Janet M. Turan. 2016. \u201cExperiences Accessing \nAbortion Care in Alabama among Women Traveling \nfor Services.\u201d Women\u2019s Health Issues 26(3):298\u2013\n304.\nZeldovich, Varvara B., Corinne H. Rocca, Callie Langton, \nUta Landy, Elizabeth S. Ly, and Lori R. Freedman. \n2020. \u201cAbortion Policies in U.S. Teaching Hospitals: \nFormal and Informal Parameters Beyond the Law.\u201d \nObstetrics and Gynecology 135(6):1296\u2013305.\nZernike, Kate. 2022. \u201cWhat Does \u2018Abortion\u2019 Mean? Even \nthe Word Itself Is up for Debate.\u201d The New York \nTimes, October 18.\nZurek, Melanie, and Jenny O\u2019Donnell. 2019. \u201cAbortion \nReferral-Making in the United States: Findings and \nRecommendations from the Abortion Referrals \nLearning Community.\u201d Contraception 100(5):360\u2013\n66.\nAuthor Biographies\nDanielle Czarnecki is a visiting assistant professor in the \nDepartment of Sociology at Case Western Reserve \nUniversity. Her research explores contested medical tech-\nnologies and practices, with a focus on reproduction. Her \nnew research examines women\u2019s experiences with multi-\nfetal pregnancy reduction.\n\nCzarnecki et al.\t\n485\nDanielle Bessett (she/her) is a professor of sociology and \nfaculty affiliate of both Women\u2019s, Gender, and Sexuality \nStudies and the Medical Scientist Training Program at the \nUniversity of Cincinnati, Ohio. Her research centers on the \nsociology of reproduction. Bessett directs UC\u2019s Kunz \nCenter for Social Research and co-leads OPEN, which \nconducts rigorous, interdisciplinary research to assess the \nreproductive health and well-being of Ohioans in the con-\ntext of federal and state policies.\nHillary J. Gyuras is a senior research associate at the \nUniversity of Cincinnati with the Ohio Policy Evaluation \nNetwork. Her research focuses on the politicization of \nsexual and reproductive health care and seeks to center the \nvoices of those impacted by abortion restrictions.\nAlison H. Norris is an associate professor in Ohio State \nUniversity\u2019s College of Public Health and College of \nMedicine. Dr. Norris is the co-principal investigator of \nOPEN, an interdisciplinary research program that studies \nthe impacts of reproductive-health-related policies and \nlaws on the health and well-being of Ohioans. Dr. Norris \nalso leads a longitudinal epidemiologic cohort study inves-\ntigating sexual and reproductive health decision-making \namong rural Malawians, access to contraception and abor-\ntion, prevention of unwanted pregnancy and sexually \ntransmitted diseases, and knowledge and stigma about \nsexual and reproductive health.\nMichelle L. McGowan is a senior associate consultant II \nin the Biomedical Ethics Research Program and the \nDepartment of Quantitative Health Sciences at Mayo \nClinic and a visiting scholar in Women\u2019s, Gender, and \nSexuality Studies at University of Cincinnati. She is an \nempirical bioethicist whose research focuses on the ethical \nand social implications of reproductive health policies. \nShe leads a study on the impact of clinic closures and \nchanges in service delivery on abortion clinics, health care \nproviders, and abortion seekers.\n\n\n    Scientific Research Findings:", "answer": "\u2022 The political context of care was a source of medical uncertainty. The political context of care included cultural politicization, such as stigma; the legal, regulatory, and policy contexts; and the economic organization of health care.\n\u2022 Interpretation and implementation of policy\u2014including policies based on laws that are not in effect\u2014differed across institutions and regions of the state. Frequent policy changes disrupted standards of care, increasing confusion and uncertainty about laws and policies.\n\u2022 Uncertainty around abortion policies led to institutions and reproductive health care professionals having stricter interpretations of what is allowed under the law, amplifying the impact of an already restrictive context for abortion care.", "id": 70}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465211058152\nJournal of Health and Social Behavior\n2022, Vol. 63(2) 162\u00ad\u2013176\n\u00a9 American Sociological Association 2021\nDOI: 10.1177/00221465211058152\njhsb.sagepub.com\nArticle\nWomen have long been targeted as a special popula-\ntion in law and drug policy and are often subject to \ndistinct forms of coercive disciplining for their sub-\nstance use (Campbell 2002). Pregnant substance \nusers are uniquely exposed to competing medical \nand legal responses while receiving pregnancy-\nrelated health care. Despite recent, louder prescrip-\ntions to treat addiction as a disease (Pew Research \nCenter 2014), pregnant and postpartum substance \nusers continue to experience extraclinical interven-\ntions such as criminal prosecution and child welfare \ninterventions (Paltrow and Flavin 2013). State man-\ndatory reporting laws and child abuse statutes that \ncodify substance use during pregnancy as child mal-\ntreatment create a statutory environment that \nempowers practitioners to investigate their patients \nfor suspected illegal activity. Recent estimates indi-\ncate that tens of thousands of women a year are sub-\nject to child welfare investigations after medical staff \nidentify and report them to the state for prenatal sub-\nstance use (U.S. Department of Health and Human \nServices 2021).1 This article examines the blending \nof medical and legal approaches in perinatal care \nthrough an analysis of medical professionals\u2019 \njustifications and rationales for drug testing pregnant \npatients. Expanding current understandings of the \nrelationship between medicalization and criminaliza-\ntion processes, this study reveals the mechanisms that \nenable medical providers to engage in investigatory \nwork in the clinic. I also discuss the consequences of \nmedico\u2013legal hybridity, whereby criminal\u2013legal \n\u00adlogics get taken up in medical practice in such a way \nthat medical routines, practices, and interactions \nbecome a form of criminalization.\nDespite media coverage and numerous legal com-\nmentaries on the topic of the criminalization of preg-\nnant substance users, there are few empirical accounts \nof this process (Paltrow and Flavin 2013; Stone \n2015). This study elucidates a process about which \nwe currently know little: how criminal\u2013legal logics \nguide medical actors\u2019 decision-making processes in \n1058152 HSBXXX10.1177/00221465211058152Journal of Health and Social BehaviorMcCabe\nresearch-article2021\n1University of Notre Dame, Notre Dame, IN, USA\nCorresponding Author:\nKatharine McCabe, Reilly Center for Science, \nTechnology, and Values, University of Notre Dame, 436 \nGeddes Hall, Notre Dame, IN 46556, USA. \nEmail: kmccabe6@nd.edu\nCriminalization of Care: Drug \nTesting Pregnant Patients\nKatharine McCabe1\nAbstract\nThis article reveals how law and legal interests transform medicine. Drawing on qualitative interviews \nwith medical professionals, this study shows how providers mobilize law and engage in investigatory \nwork as they deliver care. Using the case of drug testing pregnant patients, I examine three mechanisms \nby which medico-legal hybridity occurs in clinical settings. The first mechanism, clinicalization, describes \nhow forensic tools and methods are cast in clinical terminology, effectively cloaking their forensic intent. \nIn the second, medical professionals informally rank the riskiness of illicit substances using both medical \nand criminal-legal assessments. The third mechanism describes how gender, race, and class inform forensic \ndecision-making and criminal suspicion in maternal health. The findings show that by straddling both \nmedical and legal domains, medicine conforms to the standards and norms of neither institution while also \nsuspending meaningful rights for patients seeking care.\nKeywords\ncriminalization, drug testing, intersectionality, maternal substance use, provider bias\n\nMcCabe\t\n163\nhypermedicalized perinatal care settings. Drawing on \ninsights from medical sociology and socio\u2013legal \nstudies, this research examines an important site of \ntransformation where the lines between medical care \nand legal intervention significantly blur, resulting in a \nloss of rights and protections for patients and fore-\nclosing a site of possible institutional support for \nwomen who use substances.\nBackground\nDrug Testing Pregnant Patients\nDrug testing is a familiar phenomenon. Most adults \nin the United States will be drug tested in their life-\ntimes\u2014as a job applicant; while serving in the mili-\ntary, attending school, or playing competitive sports; \nor perhaps through a routinized system of surveil-\nlance, such as a substance treatment program or as a \nrequirement for probation. Drug testing technolo-\ngies have been available for over 60 years (Reynolds \n2005). In the 1980s, the passage of federal guide-\nlines for drug testing employees resulted in the cre-\nation of numerous safeguards for test subjects, \nincluding forensic testing standards, specimen- \nhandling procedures, confirmatory testing require-\nments, and confidentiality guidelines (Substance \nAbuse and Mental Health Services Administration \n[SAMHSA] 2012).\nDrug testing in clinical settings has emerged \noutside of the jurisdiction of federal regulations that \nguide employment testing. Clinical drug testing is \nnotably less sophisticated and less transparent than \nworkplace testing (SAMHSA 2012; Warner, \nWalker, and Friedmann 2003). In clinical settings, \nforensic standards like \u201cchain of custody\u201d protocol \n(i.e., protocol for handling biological specimens to \nmeet legal evidentiary standards) are not routinely \nfollowed, patients are often not informed that their \nbiological material will be tested, and confirmatory \ntesting is rarely carried out to rule out testing errors \n(Jarvis et\u00a0al. 2017; Warner et\u00a0al. 2003). Unlike fed-\neral workplace drug testing standards, there are no \nwidely applied standards governing clinical drug \ntesting that ensure that clinical testing is nondis-\ncriminatory, well regulated, consensual, or confi-\ndential (Jarvis et\u00a0al. 2017; SAMHSA 2012; Warner \net\u00a0al. 2003).\nDrug testing pregnant patients exists in this \nunstandardized and unregulated context. Drug testing \npatients and their newborns is a routine practice in \nperinatal care (American College of Obstetricians \nand Gynecologists [ACOG] 2017; Terplan and \nMinkoff 2017). However, drug tests are not clinical \ndiagnostic tools (ACOG 2017; Moeller, Lee, and \nKissack 2008; SAMHSA 2012; Warner et\u00a0al. 2003), \nnor do they predict future maltreatment of offspring \n(Wood 2006). For pregnant patients, the risks associ-\nated with being drug tested in the hospital are incom-\nparable to other populations and other testing \nscenarios. For instance, job applicants drug tested by \nemployers do not risk having their children removed \nfrom their care if they test positive, whereas simply \nentering a prenatal clinic to receive care exposes \npatients to the threat of intervention by outside \nauthorities. This criminal\u2013legal exposure is explicitly \nsexed (i.e., cis-men are not exposed to the same risk \nof being investigated for child abuse for a positive \ndrug test while receiving care in clinical settings). The \nAmerican College of Obstetricians and Gynecologists \n(ACOG) has authored statements discouraging medi-\ncal staff from conducting urine drug tests on pregnant \npatients (ACOG 2017) and has warned that such test-\ning may result in clinicians acting as \u201cagents of law \nenforcement\u201d (ACOG 2020) and in women avoiding \nprenatal care for fear of losing custody of their chil-\ndren or being criminally prosecuted (ACOG 2011).\nThe history of how pregnant substance users \nbecame targets of this facet of the \u201cwar on drugs\u201d \nhas its antecedents in the crack cocaine crisis of the \n1980s and 1990s. At the height of the social panic \naround crack cocaine, media portrayals and scien-\ntific research promoted the notion that crack \ncocaine was exceptionally destructive (Reinarman \nand Levine 2004) and that pregnant crack users \nwere inflicting irreparable damage on their off-\nspring (Roberts 1999). These campaigns relied on \nracist stereotypes and images that cast black repro-\nduction as causally linked to social disorganization \nand community disadvantage (Collins 1989; \nRoberts 1999). As a result, black mothers experi-\nenced disproportionate drug testing in hospitals and \nreporting to child welfare services (Chasnoff, \nLandress, and Barrett 1990; Roberts 2002). Early \nforecasts about the exceptional effects of crack \ncocaine exposure have since proven to be scientifi-\ncally inaccurate and hyperbolic (Hurt et\u00a0al. 2009), \nyet the initial reactions created enduring path \ndependencies between punitive, extraclinical sys-\ntems and reproductive health care (Bridges 2019).\nMedical systems and child protective systems \nare linked through federal policy and state laws. The \nChild Abuse Prevention and Treatment Act \n(CAPTA), a federal spending provision for state \nchild protection services, requires that health care \nproviders notify state child protective agencies \nwhen infants are born substance affected. However, \nCAPTA does not direct medical staff to actively \n\n164\t\nJournal of Health and Social Behavior 63(2)\nsearch for evidence that their patients use illicit sub-\nstances, nor does it encourage states to classify illicit \nsubstance use as child maltreatment (National \nAdvocates for Pregnant Women 2018). Twenty-\nthree states and the District of Columbia have gone \nbeyond federal guidelines by amending their civil \nstate statutes to include substance use during preg-\nnancy as a form of child maltreatment (Guttmacher \nInstitute 2021). The state of Illinois, where this \nstudy was conducted, classifies evidence of a con-\ntrolled substance in the blood, urine, or meconium \n(first stool) of a newborn infant as child neglect \n(Illinois Abused and Neglected Child Act 1975). \nFederal and state laws establish a baseline for state \nintervention at birth, but drug testing and reporting \npractices vary drastically not only by state but \nwithin states, municipalities, and hospitals (Drescher- \nBurke and Price 2005; Miller et\u00a0al. 2014).\nAlthough the policy and legal landscape encour-\nage providers to assume an active role in detecting \nand reporting patients on behalf of the state, the \nuptake of legal tasks in medicine cannot solely be \nunderstood as driven by \u201claw from on high.\u201d The \npenetration of law into medicine is uneven and con-\ntingent on institutional arrangements and profes-\nsional prerogatives (Heimer 1999). There are \nmyriad reasons that clinicians may be reluctant to \nreport suspected child abuse, including a lack of \nknowledge of state laws, concerns about patient \nretention, and previous poor interactions with child \nprotective services (Gunn, Hickson, and Cooper \n2005; Vulliamy and Sullivan 2000). On the other \nhand, maternal and newborn care providers may be \nparticularly receptive to prenatal drug use laws and \nperceive the state as a useful partner in managing \n\u201cdifficult\u201d patients (Abel and Kruger 2002; Heimer \n1999). Law empowers and informs medical prac-\ntices, but it is not wholly determinative. In most set-\ntings (barring settings with universal testing \nprotocol), provider discretion determines whether \npatients will be tested and reported to the state.\nMedico\u2013Legal Hybridity in Health Care\nOne neglected area this study highlights is the ways \nin which actors within hospital settings are engaged \nin processes of managing patients through punitive \nstate systems. The limited evidence available sug-\ngests that law enforcement and child welfare systems \nare more frequently called on to intervene in health \ncare interactions in public facilities that serve racially \ndiverse and low-income populations (Bridges 2011; \nLara-Mill\u00e1n 2014). Medical systems are key not just \nin the referral process\u2014the process of identifying \nand reporting women to child welfare services\u2014but \nmedicine is transformed through interactions with the \nstate, becoming a site of criminalization in and of \nitself. Although much of the literature to date exam-\nines the intersections of medicalization and criminal-\nization processes within sites that are often explicitly \ndisciplinary, e.g., semicarceral settings (Haney 2010), \ndrug courts (Tiger 2012), and mandated rehab \n(McKim 2014), this research shows how criminal\u2013\nlegal logics and tasks also get taken up within \nstaunchly medicalized settings.\nPrior work on medico\u2013legal hybridity focuses \non jurisdictional skirmishes between medical and \nlegal domains. Timmermans and Gabe (2002) theo-\nrize that different \u201cborderland\u201d scenarios emerge in \nresponse to medico\u2013legal conflict. In some scenar-\nios, medical and legal domains jockey for power \nand predominance over one another, in others, a \nform of semipermanent hybridity or a \u201cthird way\u201d \nemerges wherein something new is produced either \nthrough a blending or a rejection of methods from \nboth sides (Timmermans and Gabe 2002). Elizabeth \nChiarello (2015) shows that in hybrid environ-\nments, where frontline workers engage in medical \nand legal work, workers both embrace and eschew \ntheir discretionary power to enforce the law because \nit may strain their relationships with their patients \nor disturb their sense of identity as healers. One \nexample of embracing and eschewing discretion is \nthrough a process referred to as \u201cchanneling,\u201d in \nwhich workers use their discretion to select which \npatients will be moved through the \u201cpipeline\u201d from \nmedical to legal domains (Chiarello and Morrill \n2020).\nMaternal and infant care providers do not neces-\nsarily see themselves as working for the state, nor \nare they trained in forensic methods. Therefore, dis-\ncovering how and why these professionals become \nengaged in investigatory work is essential to under-\nstanding how legal interests get picked up and cir-\nculate within medicine. The question arises: What \nmotivates providers to drug test their patients and \nexpose them to significant state involvement? One \nway that providers may contend with the contradic-\ntions of medico\u2013legal hybridity is through the pro-\ncess of channeling patients who are deemed \nirremediable by medical means to more intensive \nand coercive fields of management such as child \nwelfare systems. Kelley Fong (2020) observes that \nreporting professionals like health care workers fre-\nquently call on child welfare not in response to spe-\ncific child maltreatment but, rather, in a paternalistic \nattempt to rehabilitate indigent families and engage \nthe state in their oversight.\n\nMcCabe\t\n165\nStrategies used to manage substance-using \nwomen are rooted in anxieties about gender and \nrace that frame understandings of responsible \nreproduction and self-governance. Women\u2019s sub-\nstance use threatens gendered expectations regard-\ning social reproduction, self-responsibility, and \nresponsibility to dependent others (Campbell 2002). \nHowever, gender, race, and class operate together \nto shape representations of \u201cideal motherhood,\u201d and \nthe reproduction of women of color is more likely \nto be policed through the state\u2019s criminal\u2013legal  \nand child welfare systems (Roberts 1999, 2002). \nMedical actors\u2019 routine engagement with child \n\u00adwelfare systems may heighten raced, classed, and \ngendered evaluations of potential parenting quality \nas providers grapple with the power to investigate \nand report patients based on criminal\u2013legal \nsuspicion.\nThis article considers how law is mobilized by \nhealth care workers and reveals that medicine \naccommodates law but not without significantly dis-\ntorting medical routines, rationality, and norms. \nBuilding off the concept of channeling (Chiarello \nand Morrill 2020), I show that providers do not sim-\nply refer patients to outside authorities who will \npolice them; rather, the very existence of a pipeline \ntransforms the provision of care in such a way that \nclinicians engage in policing and enforcement them-\nselves. The following analysis uses interview data to \ndetail three mechanisms that enable medical profes-\nsionals to engage in forensic investigatory work: (1) \nclinicalization\u2014applying clinical terms, definitions, \nand routines to forensic approaches; (2) negotiating \ncompeting risk frameworks\u2014the tendency to swap \nmedical risk calculations for legal risk assessments \nin the face of uncertainty; and (3) discretionary \n\u00adsuspicion\u2014drawing on racial, classed, and \u00adgendered \nnotions of \u201crespectable motherhood\u201d (Sweet 2019) \nto target patients for or protect patients from investi-\ngation. These mechanisms reveal that criminaliza-\ntion processes are integrated into care in subtle and \nsometimes covert ways. This less than obvious inte-\ngration enables forensic work to take place in clini-\ncal settings in a routine and unremarkable manner.\nThis study illuminates the transformation of \ncontemporary medical authority. Prior theoretical \nwork suggested that medicalization\u2014or the trans-\nformation of social problems into medical \u00adconditions\u2014 \nwas the primary way that medicine expanded its \njurisdiction (Conrad 2007). However, these find-\nings show that medico\u2013legal hybridity broadens \nphysicians\u2019 scope of practice and authority over \npatients. The consequences of hybridity in perinatal \ncare are troubling. By straddling care and legal \nenforcement, clinically initiated investigations cre-\nate a new form of \u201cstructural vulnerability\u201d \n(Bourgois et\u00a0 al. 2017) for patients who, through \ninteractions with providers, are denied normative \nmedical care and are exposed to legal interventions \nwithout basic legal protections.\nData and Methods\nThese data were derived from a larger qualitative \nstudy examining multisystem collaboration in \nresponse to perinatal substance use in Chicago, \nIllinois, and included in-depth interviews with \nfrontline workers and stakeholders (N = 71) and \ninterviews with local women who had experienced \nchild welfare interventions at the birth of their chil-\ndren (N = 28). Although these data informed this \nanalysis, the findings presented here are a subset of \nin-depth qualitative interviews conducted with 30 \nmedical providers in the Chicago area. To recruit \nmedical professionals who had knowledge of insti-\ntutional responses to perinatal substance use, I \nemployed a purposive sampling and snowball \napproach. I selected five hospitals in the Chicago \narea as recruitment sites. These sites were selected \nbecause they served distinct patient populations (in \nterms of race-ethnicity and class status) and included \none safety net hospital, one teaching hospital, two \nprivate hospitals, and one small community hospi-\ntal. Each hospital had publicly available contact \ninformation for their obstetric and neonatal physi-\ncians. I sent emails to providers whose contact \ninformation was accessible (N = 116) to invite them \nto participate in an interview about their \u201cobserva-\ntions and experiences working with\u201d substance-\nusing patients. The email recruitment resulted in  \n21 participants agreeing to meet for an interview.  \nAn additional 9 participants were subsequently \nrecruited via snowball sampling, for a total of 30 \ninterviews (see Table 1 for sample characteristics).\nThe resulting sample of providers was heteroge-\nneous. Through snowball sampling, I was able to \nhome in on actors who were known within their \ninstitutions to have knowledge and experience \nworking with substance-using patients (e.g., clini-\ncians with a special interest in perinatal substance \nuse or perinatal social workers who were often \nresponsible for making hotline calls to child wel-\nfare and communicating with patients after they \nhave been tested). None of the hospitals in this sam-\nple had instituted formal universal screening proto-\ncol, and all of the providers in this sample discussed \ndrug testing as a discretionary process. These find-\nings are not intended to generalize to any one group \n\n166\t\nJournal of Health and Social Behavior 63(2)\nof professionals or setting. The benefit of a diverse \nsample is that it provides various professional van-\ntage points at different points of care to reveal the \neffect of hybridity on decision-making processes \nthat may be transferable to the study of other pro-\nfessional groups, institutions, or social problems \nspanning medical and legal institutions.\nI conducted semistructured interviews in person \nand over the phone (eight individuals were unable \nto meet in person) that lasted on average 45 minutes \nto an hour, with the longest interview spanning over \n3 hours. The interview schedule centered around \nthe provision of care for substance-using patients, \nincluding questions regarding how substance use \nwas determined, drug screening and testing prac-\ntices, hospital procedures and treatment protocol, \nprofessionals\u2019 attitudes toward substance-using \npatients, medical interventions for maternal sub-\nstance use, and child welfare involvement.\nInterviews were recorded and transcribed verba-\ntim. Data were coded and analyzed using Atlas.ti. \nParticipants were given pseudonyms to protect their \nidentity. Data were analyzed inductively following a \ngrounded approach (Charmaz 2014). Interviews first \nunderwent open coding. During the initial coding \nphase, I simultaneously created detailed memos \nreflecting on my impressions of the emerging codes \nand patterns. These memos and initial codes guided \nthe next stage of focused coding, during which new \ncodes were created to capture salient themes within \nand across interviews and initial codes were con-\ndensed (Charmaz 2014). For the purposes of this \nanalysis, I focused on data from medical providers \nand themes that emerged from codes relevant to drug \ntesting, clinical decision-making, medical surveil-\nlance, risk, and collaborations between medicine and \nchild welfare. Ethical approval was obtained through \nthe appropriate institutional review board.\nResults\nClinicalization\nClinicalization is a process whereby providers use \nclinical terminology, definitions, and frameworks to \ndescribe forensic methods and tools. Unlike medi-\ncalization, clinicalization does not describe a pro-\ncess of transformation (i.e., a behavior or condition \nis not being transformed into a medical one). Rather, \nthe process of clinicalization superimposes clinical \ntechniques and language over essentially forensic \napproaches. Clinicalization was evident in the ways \nthat providers discussed drug testing as routine and \ncouched testing in the language of patient care. One \nobstetrician (OB) explained that \u201cIf a patient tells \nme that she\u2019s using marijuana, I usually don\u2019t do a \nurinary tox screen. But if they admit to anything \nmore than that, I\u2019ll tell them that we just need to do \na tox screen just to\u2009.\u2009.\u2009.\u2009just to check. And then most of \nthem are like okay, that\u2019s\u2009.\u2009.\u2009.\u2009that\u2019s fine.\u201d Framing \ndrug testing as \u201cjust checking\u201d presents it as neutral \nand routine when in fact, drug test results have the \npotential to drastically affect patients\u2019 lives, expos-\ning them to child welfare interventions, coerced \ndrug treatment, and long-term state surveillance. In \nthe following account, a physician (OB, teaching \nhospital) shares his experience with drug testing and \nattaining consent:\nDoctor:\u00a0We don\u2019t necessarily go through and do \nverbal consent for every test in the prenatal bat-\ntery that you get when you first show up. We\u2019re \ngonna do a bunch of tests to check for all kinds \nof things, including syphilis. We say something \nlike \u201cin this situation one of the things that we \nwanna check to make sure that we\u2019re taking the \nbest care of you is a urine drug screen. Is that \nokay with you?\u201d\nTable 1.\u2002 Medical Professionals Characteristics.\nn\n%\nEmployed by institution type\n\u2003 Private hospital\n12\n40\n\u2003 Teaching hospital\n11\n37\n\u2003 Public hospital\n6\n20\n\u2003 Community hospital\n1\n3\nProvider type\n\u2003 Obstetrician\n12\n40\n\u2003 Hospital social worker\n7\n23\n\u2003 Neonatologist\n3\n10\n\u2003 Hospital administrator\n3\n10\n\u2003 Maternal fetal medicine specialist\n2\n6\n\u2003 Neonatal nurse\n1\n3\n\u2003 Pediatrician\n1\n3\n\u2003 Psychiatrist\n1\n3\nRace\n\u2003 White\n20\n67\n\u2003 Black\n7\n23\n\u2003 South Asian\n1\n3\n\u2003 East Asian\n1\n3\n\u2003 Arab\n1\n3\nGender\n\u2003 Women\n24\n80\n\u2003 Men\n6\n20\nTotal\n30\n100\n\nMcCabe\t\n167\nInterviewer:\u00a0And normally, women are just like \n\u201cthat\u2019s okay?\u201d\nDoctor:\u00a0Yes. They say yes. It\u2019s a little bit like plead-\ning the fifth. If you say no, then we start to \nwonder.\nDrug testing was presented as just one test in a \nbattery of tests that patients may expect to be sub-\njected to while receiving prenatal care. Refusal of a \ndrug test carries different consequences relative to \nrefusing other sorts of tests (e.g., screening for sex-\nually transmitted infections). The language of \nforensic investigation blends into routinized screen-\ning as a part of clinical assessment and providing \nthe \u201cbest care\u201d to patients. Refusing drug testing is \ncompared to \u201cpleading the fifth,\u201d a legal designa-\ntion to signal that a witness is exercising their con-\nstitutional right not to incriminate themselves in a \ncriminal investigation. However, unlike criminal \nlegal proceedings, the right not to self-incriminate \nis unavailable to patients in perinatal care settings \nas they navigate care. Often patients are uninformed \nthat testing is taking place or what the potential \nconsequences of a positive test are (Warner et\u00a0al. \n2003). In cases where patients are made aware, to \nrefuse such testing would signal an admission of \nwrongdoing. Drug testing presents a double bind \nfor patients where to refuse medical care becomes a \nform of legal incrimination. Given the unique and \nhigh-stakes risks posed to patients and low clinical \nutility of drug testing, framing drug testing as neu-\ntral in clinical communications obscures the social \nand legal risks posed by drug testing and reveals \ngaps in the informed consent process.\nAnother type of clinicalization arose when med-\nical professionals moved seamlessly between clini-\ncal considerations and forensic interests when \ndiscussing drug testing decisions. Placing distinct \nand heterogeneous rationales on the same plane of \nconsideration, providers blurred the lines between \nclinical screening and forensic testing, which \nobfuscated the significant legal and social conse-\nquences of testing for forensic purposes. An OB \n(community hospital) explained how she decided \nwhether to drug test her pregnant patients:\nWe always ask. Any patient is screened \nverbally. Patients will sometimes admit to \nmarijuana usage and sometimes they won\u2019t. We \nuse the circumstances of the admission to drive \nwhether we are going to do urine screening for \ndrug use. If a patient comes in with a clinical \npicture that suggests substance abuse, uh \nplacental abruption.\u2009.\u2009.\u2009.\u2009Then we\u2019ll go ahead. A \nlack of prenatal care in and of itself\u2014during \nthe whole course of the pregnancy\u2014kind of \nraises our index of suspicion for substance \nabuse. So, we use different pieces of our \nmedical history, and sometimes just our eyes \nand ears to determine whether we\u2019re going to \nreally focus on having the patient admit drug \nuse. And if they don\u2019t, in the interest of the \nbaby, then we\u2019ll just go ahead and do a urine \ndrug screen.\nThe account illustrates several ways that clinical \nand forensic considerations meld into one another. \nThe obstetrician uses the language of forensic \ninvestigation throughout; the emphasis on attaining \n\u201cadmissions\u201d from the patient, an \u201cindex of suspi-\ncion,\u201d and using \u201cour eyes and ears\u201d all point to an \nethos of surveillance guiding testing decisions. \n\u201cIndex of suspicion\u201d is terminology that is regu-\nlarly used in medicine to describe the extent to \nwhich a disease or condition is being weighed as a \nlikely diagnosis. However, in this context, \u201cindex of \nsuspicion\u201d takes on nonclinical connotations \nbecause it is unclear if substance use is being sin-\ngled out as a disease that the clinician would like to \ntreat or as a risk that must be detected and reported \n\u201cin the interest of the baby.\u201d\nIn the aforementioned account, medical lingo and \nthe language of clinical discovery overlap with ter-\nminology one might expect in forensic investigations \nor court proceedings. For instance, an interest in \nacquiring \u201cadmissions\u201d from patients presents a kind \nof medico\u2013legal hybridity. In law, an \u201cadmission\u201d is \na piece of evidence that can be entered in the adjudi-\ncation of a legal case. In health care encounters, an \nadmission is a part of standard medical history- \ntaking and is often conducted under the auspices of a \nprivate and confidential interview for diagnostic pur-\nposes. The distinction has practical implications \nbecause patients are not \u201cput on notice\u201d or otherwise \ninformed that the information that they share with \nmedical staff can be used as evidence against them \nby state actors. Therefore, investigatory methods are \nintroduced into clinical practices without incorporat-\ning the underlying legal safeguards that would be a \npart of a standard legal investigation.\nIt was common for providers to frame drug test-\ning decisions by listing several incongruous risk \nfactors side by side. In the following account, an \nOB (teaching hospital) explains how decisions \naround drug testing were made at his institution:\n\n168\t\nJournal of Health and Social Behavior 63(2)\nRisk factors. She\u2019s problematic. Short of testing \neverybody, which we do not do, routinely. It\u2019s \nbased on, you know, suspicion.\u2009.\u2009.\u2009.\u2009Class, poverty, \nnot attending prenatal care, or certainly any \ndirect evidence, right? Track marks. Appearing \ninebriated. But we should be testing in those \ncircumstances. That\u2019s direct evidence. I guess \nthe other indicator to test\u2014now I\u2019m thinking \nabout the other clinical indicators, somebody \ncomes in [with] severely high blood pressures, \nor abrupting. Those are things that could be \ncaused by cocaine. And there\u2019s medical reason \nalso to know because certain drugs we would \ngive to treat high blood pressure could be \nextremely dangerous if there\u2019s cocaine in your \nsystem.\nThere were only a few medically indicated rea-\nsons providers cited for conducting a drug test to \nclarify the clinical picture. For example, abnor-\nmally high blood pressure and placental abruption \ncould be related to cocaine use. Additionally, a phy-\nsician may have concerns about how certain illicit \nsubstances might interact with other medications. \nHowever, these clinical concerns were placed next \nto other sorts of information or social \u201crisk factors,\u201d \nsuch as utilization of prenatal care, signs of poverty \nor class status, verbal admissions of substance use, \nor other \u201cdirect evidence,\u201d all of which are not \ndirectly linked to the medical management of the \npatient (I discuss biases in care in more detail in the \nfollowing findings sections).\nClinicalization was also evident in medical pro-\nfessionals\u2019 use of the language of clinical \u201cscreen-\ning\u201d in place of the straightforward language of \n\u201cdrug testing.\u201d Drug testing does not align with the \nguidelines for clinical screening and is not a diag-\nnostic tool (Wilson and Jungner 1968). In any form \n(blood, urine, meconium, or umbilical cord sam-\nples), drug testing does not confirm that someone \nhas a substance use disorder. To do so requires eval-\nuation by an appropriate clinician using diagnostic \ncriteria. By referring to drug testing as screening, \nproviders place forensic methods into a clinical tax-\nonomy that, in effect, masks the investigatory intent \nof testing from patients and spares providers from \ndirectly naming and potentially wrestling with their \nown policing and enforcement activities.\nAlthough drug testing was posed in clinical \nterms, it was rare for providers to discuss maternal \nsubstance use with health-based frames. Instead, it \nwas often described as a behavior that may cause or \nexacerbate other kinds of medical conditions, such \nas placental abruption, elevated blood pressure, or \nwithdrawal symptoms in infants. Clinicians did not \nmention medically treating the kinds of conditions \nthat co-occur with problem substance use, such as \noverdose, infection of the heart valves related to \ninjection substance use, inadequate nutrition, car-\ndiovascular problems, risks associated with with-\ndrawal symptoms, risk of HIV, or hepatitis C \nexposure and transmission. These findings support \nthe argument that substance use remains a moral-\nized problem despite numerous efforts to medical-\nize the condition (Conrad 2007).\nWhen criminal suspicion enters the clinic, nor-\nmative standards of care and informed consent go \nout. Breaches in processes that protect patients, \nsuch as informed consent and confidentiality, the \nsubversion of medical categories (e.g., using the \nterminology of \u201cscreening\u201d to describe a process of \nforensic evidence collection), and scaling back \ndiagnostic norms (i.e., for determining a substance \nuse disorder), are a few ways that drug testing \nrevokes patient-hood and norms of care.\nNegotiating Competing Risk \nFrameworks\nMedical staff negotiated medical and legal risk \nframeworks while assigning varying degrees of risk \nto illicit and licit substances. Some substances were \ndeemed benign even if they posed teratogenic or \nother health risks, whereas others were associated \nwith criminality. There was a notable amount of \ninconsistency in providers\u2019 risk assessments as they \nvacillated between health-based and criminal\u2013legal \nrisk, revealing that presumptions of harm are unsta-\nble. Ranking substances allowed medical staff to \nexercise their discretion by establishing gradients of \ncriminality and making finer distinctions between \npatients. Hybridity in the form of competing risk \nframes shapes decision-making in the clinic as pro-\nviders grapple with their own, often contradictory, \nassessments of risk.\nOver the past decade, the United States has seen \na dramatic shift in legal, policy-based, and cultural \nresponses to certain illicit substances. Successful \nefforts to legalize marijuana use and shifting atti-\ntudes toward criminal justice responses to substance \nuse (Pew Research Center 2014) have paved the \nway for medical and lay populations to reevaluate \ntheir beliefs about harm posed by substance use ver-\nsus its legality. Marijuana use has had a confounding \neffect on providers\u2019 assessments of risk and harm. In \nthe following account, an OB (public hospital) \nexplains why she believes perinatal marijuana use \nshould be exempt from child welfare interventions:\n\nMcCabe\t\n169\nWhat are the effects of marijuana? The reason \nwhy I don\u2019t call [child welfare] is because I feel \nlike marijuana\u2014as far as I can tell\u2014seems to be \nfairly benign. I don\u2019t want someone to be in a \nsystem for something that\u2019s fairly\u2009.\u2009.\u2009.\u2009which \ntobacco is worse, and it\u2019s legal. So, I don\u2019t call \nanybody on that person. I\u2019m not trying to have \nsomeone be at risk for losing their children over \nsomething that\u2019s not as noxious as tobacco use.\nThe doctor ranks substances by presumed \ndevelopmental or teratogenic effects, situating \nmarijuana use as \u201cmore benign\u201d and less \u201cnoxious\u201d \nthan smoking tobacco. In the same instance, she \ncontends with the legality of tobacco use, which \nhas been repeatedly shown to adversely affect \ndevelopment (DiFranza, Aligne, and Weitzman \n2004). Due to the unresolved and incongruous \nrelationship between legal and harmful, the pro-\nvider establishes her own threshold for interven-\ntion, drawing the line at involving child welfare for \nmarijuana use. The OB (teaching hospital) quoted \nbelow makes a similar distinction between mari-\njuana and cigarette use:\nSome [doctors] have come to the realization that \n[marijuana use] is so common that there\u2019s not \nmuch that we can do about it, other than \ncounseling women that this is not safe. What \nI\u2019ve heard in terms of counseling is it\u2019s, at the \nvery least, as harmful as cigarettes, which \nunfortunately doesn\u2019t help much because \ncigarettes are legal and incredibly harmful for \nyour health. So, it almost equates it to something \nthat\u2019s not illegal.\nNoting the ubiquity of marijuana use and its sta-\ntus relative to harmful but legal substances (e.g., \ncigarettes), the physician resigns her powers of \nintervention to educating and counseling patients \nrather than reporting them to outside authorities. In \nthese cases, medical knowledge of the harmful tera-\ntogenic effects of legal substances, such as alcohol \nand tobacco, is at odds with existing legal designa-\ntions. Providers seemed conflicted when weighing \nthese distinctions. Counseling parents about the \npotential effects of marijuana is a health-based \nresponse that promotes patient care, a sign that mar-\nijuana usage is becoming medicalized in maternal \nhealth settings. Yet it is also presented as a limited \nand less desirable response. Patient counseling is \nequated with being unable \u201cto do much about\u201d mar-\nijuana use. Providers may feel ill-equipped to coun-\nsel patients about substance use or may view child \nwelfare involvement as a more effective, \u201cstrong-\narm\u201d approach for getting patients to stop engaging \nin substance use.\nDowngrading the risk associated with marijuana \nuse was a recurring pattern across interviews. \nHowever, shifting attitudes toward marijuana did \nnot encourage a critical reexamination of the pre-\nsumed risks associated with other illicit substances, \nspecifically, cocaine and heroin, which were more \nlikely to be assessed using criminological criteria. A \nperinatal social worker (private hospital) shared her \nworries about suspected heroin-using patients visit-\ning their infants in neonatal care:\nWe\u2019ve had times where\u2014especially parents who \nwe are concerned are using heroin\u2014that it\u2009.\u2009.\u2009.\u2009it\u2019s \njust\u2009.\u2009.\u2009.\u2009like the nurse feels really unsafe. But I \ncan kind of understand why she feels unsafe. [E]\nspecially when it\u2019s overnight and there\u2019s not as \nmany people around. I think it\u2019s\u2009.\u2009.\u2009.\u2009most of the \ntime it\u2019s coming from a good place. It\u2019s \nnot\u2009.\u2009.\u2009.\u2009trying to be judgmental. It\u2019s more I\u2019m \nscared for me as a person, but, also, I\u2019m scared \nbecause of this baby.\nSubstance-using patients disrupt the normative \noperation of care in the neonatal setting. Heroin use \nis singled out as introducing an element of threaten-\ning uncertainty and a looming sense of potential \ncriminality. Concerns about infant safety are men-\ntioned alongside staff\u2019s personal feelings of \nunsafety, revealing that risk assessments become \namorphous the more risk is cast as anxieties around \ncriminality.\nUnlike responses to marijuana, providers did \nnot discuss sharing health-related information or \nproviding counseling to mothers who used cocaine \nor heroin. As the OB (teaching hospital) quoted \nbelow explained, the degree of uncertainty about a \nsubstance shapes attitudes toward child welfare \ninvolvement at birth:\nIf you have a patient who is using recreational \nmarijuana, I don\u2019t think that that\u2019s an appropriate \nuse of resources or the [child neglect statute]. I \nthink if you have somebody who is an opiate-\ndependent-on-maintenance-medication mother, \nthat\u2014again\u2014may not be an appropriate use of \nlaw or resources, because that\u2019s a known entity. \nThat\u2019s a known issue. Then obviously for  \nthose people who are using more heroin\u2009.\u2009.\u2009.\u2009or \ncocaine\u2009.\u2009.\u2009.\u2009or those other substances, I think it is \nan appropriate use. I as a medical professional, I \nhave no way of knowing how to evaluate if that \n\n170\t\nJournal of Health and Social Behavior 63(2)\ninfant is safe to leave the hospital with that \nmother. And I don\u2019t want to have to assess that. I \nwant no part of it. I have no idea how to go about \nfiguring that out. And I would want somebody \nwith a lot more skill and experience to be able to \ncome and talk to the mother to kind of figure that \nout with her.\nAttitudes and decisions toward involving child \nwelfare were not black and white but were made \nmeaningful in a context marked by medico\u2013legal \nhybridity. Medical staff drew from available medi-\ncal and criminal\u2013legal risk assessments to rank sub-\nstances and establish thresholds for forensic action. \nBy positioning certain substances as representing \n\u201cknown\u201d or \u201cunknown\u201d risks, the previously \nquoted physician gestures to the limits of medical \nsurveillance. Methadone treatment is categorized as \na \u201cknown\u201d risk factor because it falls within the \nrealm of medical management. Substances such as \ncocaine, heroin, and other opiates are \u201cunknown\u201d \nrisk factors that allude to a sprawling web of contin-\ngencies embodied in the desires of patients and \ntheir environments, social networks, and intimate \nrelationships. \u201cUnknown\u201d risk factors are described \nas beyond the scope of medicine, unmanaged and \nunadministered by any system of expertise.\nSome substances, like marijuana, are more \nlikely to be ignored or treated in a therapeutic man-\nner (like tobacco and alcohol use). Substances, \nsuch as cocaine and heroin, remain static and are \nconsistently associated with criminality and \nunknown social risks that require management \nfrom coercive and invasive systems. Moreover, \nuncertainty regarding the developmental effects \nassociated with illicit substance use makes legal \ncategories more salient and overrides medical\u2013sci-\nentific rationality. The scientific consensus on the \nlong-term effects of illicit substance exposure in \nutero is largely unsettled, particularly regarding the \nrelative risks posed by licit exposures, including \ntobacco and alcohol exposure (Behnke et\u00a0al. 2013; \nHurt et\u00a0al. 2009; Irner 2012). Providers may adopt \nlegal risk frames to fill in gaps in scientific knowl-\nedge. A physician who practiced at a teaching hos-\npital explained that legality supplants medical \nrationales and possibilities for care when providers \ndrug tests patients:\nWhy aren\u2019t you testing for SSRIs? Why aren\u2019t \nyou testing for protein, for tobacco? Oh, we need \nto look for evidence of intrauterine growth \nrestriction [IUGR]. They say: \u201cwell, cocaine \ncould cause this.\u201d It could. Maybe it does. I don\u2019t \nknow. The jury\u2019s still out on that. But you know \nwhat definitely does cause IUGR? Cigarette \nsmoking. Why aren\u2019t we testing for that?\u2009.\u2009.\u2009.\u2009The \nnumber one cause of preventable cognitive \nimpairment in this country is in-utero alcohol \nexposure. Why aren\u2019t you looking for this? And \nthey\u2019re like: \u201coh, because that\u2019s legal.\u201d That\u2019s \nproof that what you\u2019re doing has a primary \nprogrammatic intent of law enforcement. When \nyou say, \u201cbecause it\u2019s legal,\u201d you\u2019re showing me \nthat their primary programmatic intent is law \nenforcement or administrative law, or something \nlike that. But it\u2019s not medical.\nAlthough most of the medical staff I interviewed \nwere not as openly critical of drug testing as the \naforementioned provider, these findings show med-\nical professionals grappling with and expressing \nuncertainty and in some cases, discomfort about \nassessing risk using child welfare metrics. Providers \ncontort their own medical and scientific sensibili-\nties to make risk assessments about substance use \nduring pregnancy. Providers translate clinical con-\ncerns and uncertainty about developmental health \ninto legal risk. Where medical science is unclear, \nthe statutory environment suggests a definitive \ncourse of action based on concrete legal categories. \nThe process of swapping clinical for legal risk \nassessments enables providers to channel patients \nwho embody clinical and social \u201cunknowns\u201d to sur-\nveillant systems that will oversee and manage their \nprecarity.\nDiscretionary Suspicion\nRace, class, gender, and other patient traits like \nhealth status and age were salient in providers\u2019 \naccounts of how patients were selected for testing or \nshielded from suspicion. Counter to evidence that \nracial bias in medical interactions is implicit or \nunconscious (Feagin and Bennefield 2014), I found \nthat providers often mentioned race when discuss-\ning decisions to drug test. For example, a pediatri-\ncian explained that drug testing was overwhelmingly \ncarried out on black patients relative to other racial \ngroups where he practiced. He explained: \u201cThe big \nnumbers [disparities] are coming from oversurveil-\nlance of blacks. If you get 50 blacks and 50 whites \nthat are testing positive, maybe you had to test 2,000 \nblacks and 100 whites to get that 50/50.\u201d Over\u00ad\nsurveillance of black patients aligns with research \non drug testing and racial selection in perinatal care \nsettings (Chasnoff et\u00a0al. 1990; Roberts et\u00a0al. 2015). \nStudies find that black women and their infants are \n\nMcCabe\t\n171\nmore likely to be tested and reported to child protec-\ntive services than white women and their infants \n(Ellsworth, Stevens, and D\u2019Angio 2010; Kunins \net\u00a0al. 2007; Roberts et\u00a0al. 2015). Race was also made \nmeaningful in the context of specific institutional \nprocesses and operated in an intersectional manner. \nThe same physician (pediatrician, teaching hospital) \nremarked:\nThis is more of a personal anecdotal kind of \nthing\u2014when there\u2019s a white woman that looks to \nbe middle or upper class, and we\u2019re like where did \nyou get prenatal care, and she\u2019s pushin\u2019 a [baby] \nout, she\u2019ll be like \u201cI don\u2019t know\u2009.\u2009.\u2009.\u2009some\u2009.\u2009.\u2009.\u2009some \ngroup, I don\u2019t know.\u201d And we\u2019re like \u201cokay, \nokay\u2009.\u2009.\u2009.\u2009we\u2019ll get your records later.\u201d And if it\u2019s \nthe\u2009.\u2009.\u2009.\u2009if it\u2019s a black teenager, and she\u2019s like I don\u2019t \nknow. Then it\u2019s like \u201chmm\u2009.\u2009.\u2009.\u2009get the test.\u201d\nThe candid admission reveals that the race, \nclass, and age of patients may determine if clinical \ninformation (e.g., lack of prenatal care) is inter-\npreted as a legal red flag. Prenatal care utilization \nwas mentioned across several interviews as a risk \nfactor and criteria for testing. Providers interpret \nprenatal care utilization through their presumptions \nabout the social positionality of their patients, \ndetermining whether their suspicion will be roused \nor shrugged off. History of prenatal care expedi-\ntiously signals information about patients\u2019 resources \nand riskiness and is imbued with social weight that \nis not often questioned in the day-to-day provision \nof care (Bridges 2011). For example, prenatal care \nutilization is race and class stratified; black, \nHispanic, and Native American women experience \nhigher levels of unmet need (Osterman and Martin \n2018). Prenatal care utilization can also indicate \nnumerous aspects of disadvantage, including lim-\nited transportation capability, lack of insurance, and \ninability to take off time from work or caretaking \nresponsibilities (Roberts and Pies 2011). Using pre-\nnatal care as a trigger to drug test is complicated by \nthe fact that prenatal care use is determined by these \nstructural factors.\nA neonatologist who worked at a private hospi-\ntal in the city explained the circumstances under \nwhich providers tested patients. She stated that \namong their middle- and upper-class privately \ninsured patient population, race and class often \nclouded their ability to determine who to test:\nI mean the [patients] that strike out are [the ones \nwith] multiple substances. In a population like \nthis, you wouldn\u2019t suspect it. Oh my god, this \nmom was doing everything under the sun, but \n[we] just never thought about it. And they tell the \nteaching fellows: \u201cyou have to suspect it.\u201d It was \na yoga instructor and they\u2019re like: \u201cyeah, but \nshe\u2019s not doing drugs. No, she couldn\u2019t be doing \ndrugs.\u201d But she was an alcoholic, just doing \neverything. And her baby\u2019s urine tested positive \nfor everything.\nThat a \u201cyoga instructor\u2009.\u2009.\u2009.\u2009couldn\u2019t be doing \ndrugs\u201d suggests that patients who present certain \nlifestyle markers or affirm feminized ideals of self-\ncare and well-being are beyond criminal\u2013legal sus-\npicion. The same physician later reflected on her \nown history of testing patients and explained that \nsuspicion rests on racial categories: \u201cIt\u2019s based on \nthe clinician\u2019s discretion.\u2009.\u2009.\u2009.\u2009You know, if I hear, \nyou know, African American teenage mom is kinda \nloopy, I\u2019ll say send it [the drug test]. It\u2009.\u2009.\u2009.\u2009it\u2019s \nterrible.\u201d\n\u201cAfrican American teenage mom\u201d is an imag-\nined patient, and unlike a \u201cyoga instructor,\u201d the way \nthat she is imagined exposes her to institutional \nharm. Racial stereotypes work alongside institu-\ntionally situated biases to shape discretionary suspi-\ncion. Political and cultural discourses about the \ndeservedness and full citizenship of single black \nmothers that structure debates about resource allo-\ncation and welfare reform often center reproduction \nand child rearing. These discourses reverberate \nwithin prenatal and perinatal care settings where \nthe politics of reproduction meet the material reali-\nties of pregnancy and birth care (Bridges 2011; \nManzer and Bell 2021). The practice of drug testing \nallows providers to \u201cdo something\u201d with their sus-\npicion by mirroring forms of exclusion and disqual-\nification that exist in the broader social political \nrealm.\nBlack practitioners expressed frustration with \nwhat they perceived as racially differential responses \nto substance-using patients. Black medical staff \nwere critical of racial inequality, although they \nwere limited in their ability to change such prac-\ntices. As one provider (OB, private hospital) \nexplained, she was the only African American in \nthe room trying to explain racially biased drug test-\ning to her colleagues:\nHow we identify these patients matters. I\u2019m \nAfrican American, so I\u2019m the only African \nAmerican person in the room and I brought it up. \nI\u2019m like, \u201cwell, you know, statistically speaking, \nI\u2019m the person who would have been judged [if I \nwere a patient], where this young lady who may \n\n172\t\nJournal of Health and Social Behavior 63(2)\nbe Caucasian or Asian, she\u2019s in college and she \nmay have a problem also.\u201d\nRace, gender, class, and age operate intersec-\ntionally and heuristically in clinical settings. These \nheuristics activated providers\u2019 discretionary suspi-\ncion. However, I also found that providers resisted \ndetecting, labeling, or testing privileged patients. \nAn OB (public hospital) explained how providers \nstruggled to \u201cbelieve\u201d that the opioid crisis had \narrived at the local level, hinting at presumed \nunderlying social differences between prescription \nopioid and heroin users:\nWe keep talking about this \u201copioid crisis,\u201d \nbut\u2009.\u2009.\u2009.\u2009I just don\u2019t think people really believed \nit.\u2009.\u2009.\u2009.\u2009[I]n the city, people have this thought of: \n\u201cOh,\u2009.\u2009.\u2009.\u2009it\u2019s gonna be these people on heroin.\u201d \nBut I mean it\u2019s a lot of patients who are taking \nprescription drugs. I think we\u2019re kind of \nunderestimating that for those patients. It\u2019s really \neasy to say: \u201cOh, this patient\u2019s on heroin, this is \na problem.\u201d But I think we\u2019re normalizing \npeople being on prescription medication.\nThe opioid crisis has been made visible or invis-\nible in racialized ways (Netherland and Hansen \n2017). This was evident in the coded language that \nmedical staff used to refer to prescription opioid \nusers. For example, relative to heroin, prescription \nopioids were described as less visible and less prob-\nlematic. The \u201cpharmaceuticalization\u201d of the opioid \ncrisis situates the crisis as medical and race-neutral, \nobscuring how institutional responses are racially \nsegregated in ways that protect white users from \npunitive responses (Netherland and Hansen 2017). \nMedical staff who worked at hospitals that served \nprivately insured, majority white patient popula-\ntions described the opioid crisis as invisible, largely \nbecause drug testing was not being carried out on \npatients and opioid users were not being detected. A \nneonatologist who worked at a private hospital \nexplained the staff were unsure if there had been an \nincrease in substance-exposed infants due to the \nlack of regular testing: \u201cWe haven\u2019t seen a change. \nBut I suspect that\u2019s because we\u2019re not looking. We \ndo not look. Pretty much routinely.\u201d A perinatal \nsocial worker (private hospital) who worked at the \nsame hospital noted:\nWe\u2019re not really great about screening and some \nof our OBs are under the assumption that their \npatients wouldn\u2019t be using prescription drugs or \nany kinds of drugs. And unfortunately, I think \nsome get missed.\u2009.\u2009.\u2009.\u2009OBs feel like their patients \ndon\u2019t do that and we all obviously know that\u2019s \nnot the case. It\u2019s affected all economic statuses.\nThe desire to shield privately insured and class-\nprivileged patients from suspicion resulted in pro-\nviders avoiding testing certain patient populations, \nwhich kept them out of the reach of punitive sur-\nveillance systems. It is widely accepted that implicit \nbiases shape health care interactions (Feagin and \nBennefield 2014). These findings show that implicit \nand explicit biases restructure the provision of care \nwhen providers are empowered to investigate their \npatients and highlight provider discretion as a key \nprocess of legal assimilation.\nDiscussion\nDrug testing in perinatal care is anything but neu-\ntral, objective, and therapeutic. Drug testing mobi-\nlizes action in a way that expands physician \ndiscretion and legal enforcement capacity and trans-\nforms clinical interactions and routines into a form \nof criminalization. Through processes of clinical-\nization, negotiating competing risk frameworks, \nand discretionary suspicion, medical actors assimi-\nlate legal frameworks and investigatory work into \ntheir day-to-day routines and interactions. In this \nanalysis, I have examined these processes sepa-\nrately, but they often work in concert. For instance, \nproviders\u2019 ranking of illicit and licit substances also \nreflects discretionary suspicion because illicit drugs \nare imbued with social and cultural meanings \ninformed by racial, class, and gender politics.\nClinicalization normalizes and obfuscates how \nextraclinical aims guide the provision of care while \ndiscretionary suspicion limits the scope of provid-\ners\u2019 enforcement activities. Both processes, because \nthey are masked and applied in a limited manner, \nallow physicians to distance themselves from their \nown policing activities. For researchers who are \ninterested in understanding how hybridity is \nachieved, particularly in a field that is invested in \nprofessional sovereignty, these results show that \nhybridity thrives when the legal intent of certain \ntasks is ambiguous. Institutional risk culture and \nprofessional attitudes toward uncertainty also deter-\nmine the degree to which criminal-legal logics \nshape decision-making in care. Perhaps it is no sur-\nprise drug testing technologies would become \ninfused with legal and extraclinical implications \nwithin perinatal care, known for being a risk-averse \nand risk-saturated area of care (Studdert et\u00a0al. 2005). \nMedical risk thinking in perinatal care settings often \n\nMcCabe\t\n173\nplaces maternal interests and fetal/infant interests on \nopposite sides of the spectrum (Harris 2000). Thus, \nmedical risk frameworks may already complement \nthe child welfare model of assessing risk, streamlin-\ning the process of channeling patients from the hos-\npital to the child welfare system.\nOn its face, handing patients off to the state may \nappear to reinforce the boundaries between medi-\ncine and law. However, channeling counterintui-\ntively reflects physician discretion and deference to \noutside actors and systems. Medical discretionary \npower is often thought of as delimited by standard-\nized practices, shared decision-making, and the law \n(Timmermans and Angell 2001). Physicians are not \nsolely empowered through their clinical expertise, \nbut rather, their powers are enlarged through their \ncapitulation to legal and state definitions of harm \nand risk. Moreover, channeling is a structural pro-\ncess that bridges medicine and child welfare sys-\ntems: The more the bridge is utilized, the more \nsolidified the partnership between the two institu-\ntions becomes, and the less ability there is to forge \npathways to alternative systems (e.g., substance \ntreatment) or develop other internal processes to \nrespond to perinatal substance use.\nThis article documents biases in drug testing \ndecisions. Sometimes these biases were implicit \n(i.e., referencing prenatal care as a proxy for other \ntypes of risks or constructing opioid users as invisi-\nble). Surprisingly, biases were also often explicit. \nProviders acknowledged that factors such as race, \nclass, and age influenced their willingness to inves-\ntigate patients. Discretionary suspicion operates at \nthe intersection of race, class, gender, age, and \nhealth status through moral constructs of mothering \nand drug use. The construction of women\u2019s sub-\nstance use as more harmful and dangerous relative to \nmen\u2019s is grounded in an essentialist view of \u00ad\u201cmaternal \ninstinct\u201d and \u201cideal motherhood\u201d (Campbell 1999). \nSanctions aimed at pregnant substance users are \npartly aimed at their \u201cfailure to absorb the cost of \nsocial reproduction\u201d (Campbell 1999:909). In addi-\ntion, respectable mothering is interpreted through \ngendered race and class norms (Campbell 2002; \nCollins 1989; McKim 2014). In maternal care and \nchild welfare systems, suspicion is rooted in a dis-\ntrust of marginalized women to parent their chil-\ndren (Luna and Luker 2013; Roberts 2002). \nEnabling providers to police patients based on the \nsuspicion of maternal competence weaponizes \n\u00adphysicians\u2019 preexisting biases and stratifies health \ncare along another dimension (i.e., increased risk  \nof becoming system-involved through provider \ninteractions).\nThere are stark consequences of medico\u2013legal \nhybridity in perinatal care. Using health care deliv-\nery systems to capture patients with substance use \nissues creates unintended negative health outcomes, \nresulting in more women seeking prenatal care late \ninto their pregnancies or avoiding it altogether, \nbirthing unassisted at home, or attempting to self-\ndetox without the supervision of health profession-\nals (ACOG 2011; Stone 2015). Recent research \nsuggests that the postpartum period is a particularly \nvulnerable time for women with opioid use disor-\nders when rates of fatal overdose significantly \nincrease (Schiff et\u00a0al. 2018). By rerouting patients \ntoward punitive systems and away from health care \nresources, providers retract medical care in a \nmoment when substance-using women need it most.\nBy straddling clinical and forensic forms of \nsocial control, drug testing practices conform to \nneither. Drug testing practices violate the norms of \nclinical practice and standards of legal investiga-\ntion. Whether it concerns the clinical utility of drug \ntesting and its viability as a diagnostic tool or \npatients\u2019 rights and the ethics of informed consent, \ndrug testing rarely rose to the level of meeting the \ncriteria for normative medical intervention. On the \nother hand, drug testing was not a purely forensic \nform of intervention because it did not conform to \nseveral legal standards (i.e., laws that protect \n\u00adindividuals from incriminating themselves and \nforensic standards that protect witnesses from \n\u00adevidence tampering, forensic mishandling of \n\u00adbiological specimens, and testing errors). This \nresearch demonstrates that when institutions come \ntogether and create new hybrid formations, they do \nnot simultaneously usher in protective structures, \n\u00adleaving individuals in the borderland exposed to \nmultiinstitutional violence.\nLimitations and Generalizability\nThis study was conducted in Chicago, Illinois. \nAlthough these data were collected before recre-\national marijuana use was legalized in Illinois, it is \npossible that attitudes toward marijuana use are \ngenerally more permissive in Chicago relative to \nother parts of the country. Future research should \nexamine regional variations in responses to sub-\nstance use in perinatal care settings. In addition, this \nanalysis is premised on the uptake of legal tasks by \nindividual providers in settings where testing is \nbased on provider discretion. Studies examining \nhow law shapes provider\u2013patient interactions in \nsystems where universal drug testing is carried out \nmay reveal different patterns and outcomes. Another \n\n174\t\nJournal of Health and Social Behavior 63(2)\nlimitation of this study is that the data reflect provid-\ners\u2019 interpretations and perceptions of hospital test-\ning protocol. In the future, researchers should \nexamine official, documented hospital protocol to \nunderstand how formal organizational policy shapes \nprovider practices. Finally, given sample size limi-\ntations and the wide range of state and organiza-\ntional responses to substance use during pregnancy, \nthese findings are not intended to capture practices \nin perinatal care by and large. Rather, these findings \nreflect dilemmas workers may contend with in set-\ntings where they are given latitude over the degree \nto which they adopt legal tasks or not.\nConclusion\nUnderstanding how therapeutic and punitive control \nblend together is essential for scholars studying \nhealth governance in the twenty-first century. There \nare several possible areas of collision between puni-\ntive state and medical systems. Future global health \ncrises are expected to become more frequent and \ndestructive and will likely necessitate strong part-\nnerships between public health, medicine, and the \nstate. Furthermore, the increasing politicization of \nhealth care access and rights for marginalized \ngroups, including efforts to limit health care access \nto immigrants and transgender people, and ongoing \nefforts to restrict access to abortion services will \nlikely create myriad opportunities for health care \nworkers to enforce laws, actively challenge them, or \nignore them. Finally, researchers and practitioners \nshould examine the day-to-day criminalization of \nhealth care, including practices such as forensic \ntesting, mandated reporting, sharing private medical \nrecords for prosecutorial purposes, and court- \nmandated or coerced treatment. How medical actors \nmobilize law and exercise discretion reflects larger \npopulation control methods and reveals new vulner-\nable populations and social fault lines at the inter-\nstices of medicine and punitive state systems.\nAcknowledgments\nI am grateful to Anna Kirkland and Sandy Levitsky for \ntheir commitment to this research and for their generative \nfeedback along the way. Thanks to Claire Decoteau and \nPaige Sweet for reading and providing feedback on previ-\nous versions of the article and to the three conscientious \nreviewers for their insights.\nFunding\nThe author disclosed receipt of the following financial \nsupport for the research, authorship, and/or publication of \nthis article: This research was funded by the National \nScience Foundation Award No. DDRI 1802773.\nORCID iD\nKatharine McCabe \n https://orcid.org/0000-0002-2526- \n5561\nNote\n1.\t\nThe Comprehensive Addiction and Recovery Act, \npassed into law in 2016, requires that states report \nthe number of infants referred to child protective \nservices for prenatal substance exposure to the \nChildren\u2019s Bureau. In 2019, available data indicated \nthat a total of 38,625 infants were reported to child \nprotective agencies for prenatal substance use by \nhealth professionals. Of these reports, 83% were \ninvestigated. These figures likely greatly underes-\ntimate the total number of infants reported; do not \ninclude data for Florida, Iowa, New York, North \nDakota, or Pennsylvania; and underreport totals for \nseveral states, including Tennessee and Illinois.\nReferences\nAbel, Ernest L., and Michael Kruger. 2002. \u201cPhysician \nAttitudes Concerning Legal Coercion of Pregnant \nAlcohol and Drug Abusers.\u201d American Journal of \nObstetrics and Gynecology 186(4):768\u201372.\nAmerican College of Obstetricians and Gynecologists. 2011. \n\u201cSubstance Abuse Reporting and Pregnancy: The Role \nof the Obstetrician\u2013Gynecologist: Committee Opinion \nNo. 473.\u201d Obstetrics & Gynecology 117(1):200\u2013201.\nAmerican College of Obstetricians and Gynecologists. \n2017. \u201cOpioid Use and Opioid Use Disorder in \nPregnancy: Committee Opinion No. 711.\u201d Obstetrics \n& Gynecology 130(2):e81\u201394.\nAmerican College of Obstetrics and Gynecologists. 2020. \n\u201cOpposition to Criminalization of Individuals dur-\ning Pregnancy and the Postpartum Period.\u201d https://\nwww.acog.org/clinical-information/policy-and-\nposition-statements/statements-of-policy/2020/\nopposition-criminalization-of-individuals-pregnancy-\nand-postpartum-period.\nBehnke, Marylou, Vincent C. Smith, Committee on \nSubstance Abuse, and Committee on Fetus and \nNewborn. 2013. \u201cPrenatal Substance Abuse: Short- \nand Long-Term Effects on the Exposed Fetus.\u201d \nPediatrics 131(3):e1009\u201324. doi:10.1542/peds.2012-\n3931.\nBourgois, Philippe, Seth M. Holmes, Kim Sue, and \nJames Quesada. 2017. \u201cStructural Vulnerability: \nOperationalizing the Concept to Address Health \nDisparities in Clinical Care.\u201d Academic Medicine \n92(3):299\u2013307.\nBridges, Khiara. 2011. Reproducing Race: An Ethnography \nof Pregnancy as a Site of Racialization. Berkeley: \nUniversity of California Press.\n\nMcCabe\t\n175\nBridges, Khiara. 2019. \u201cRace, Pregnancy, and the Opioid \nEpidemic: White Privilege and the Criminalization \nof Opioid Use during Pregnancy.\u201d Harvard Law \nReview 133(3):770\u2013851.\nCampbell, Nancy D. 1999. \u201cRegulating Maternal Instinct: \nGoverning Mentalities of Late Twentieth-Century \nUS Illicit Drug Policy.\u201d Signs: Journal of Women in \nCulture and Society 24(4):895\u2013923.\nCampbell, Nancy. 2002. Using Women: Gender, Drug \nPolicy, and Social Justice. New York, NY: Routledge.\nCharmaz, Kathy. 2014. Constructing Grounded Theory. \nLos Angeles, CA: SAGE.\nChasnoff, Ira J., Harvey J. Landress, and Mark E. Barrett. \n1990. \u201cThe Prevalence of Illicit-Drug or Alcohol Use \nduring Pregnancy and Discrepancies in Mandatory \nReporting in Pinellas County, Florida.\u201d New England \nJournal of Medicine 322(17):1202\u2013206.\nChiarello, Elizabeth. 2015. \u201cThe War on Drugs Comes \nto the Pharmacy Counter: Frontline Work in the \nShadow of Discrepant Institutional Logics.\u201d Law & \nSocial Inquiry 40(1):86\u2013122.\nChiarello, Elizabeth, and Calvin Morrill. 2020. \u201cA Multi-\nfield Logics Approach to Theorizing Relationships \nbetween Healthcare and Criminal Justice.\u201d Pp. 153\u2013\n68 in Research Handbook on Socio\u2013Legal Studies \nof Medicine and Health, edited by M. A. Jacob and \nA. Kirkland. Cheltenham, UK; Northampton, MA: \nEdward Elgar Publishing.\nCollins, Patricia Hill. 1989. \u201cA Comparison of Two \nWorks on Black Family Life.\u201d Signs: Journal of \nWomen in Culture and Society 14(4):875\u201384.\nConrad, Peter. 2007. The Medicalization of Society: \nOn the Transformation of Human Conditions into \nTreatable Disorders. Baltimore, MA: Johns Hopkins \nUniversity Press.\nDiFranza, Joseph R., C. Andrew Aligne, and Michael \nWeitzman. 2004. \u201cPrenatal and Postnatal Environ\u00ad\nmental Tobacco Smoke Exposure and Children\u2019s \nHealth.\u201d Pediatrics 113(3):1007\u201315.\nDrescher-Burke, Krista, and Amy Price. 2005. Responding \nto Substance Exposed Newborns: An Exploratory Study \nof Policies and Practices. Berkeley, CA: The National \nAbandoned Infants Assistance Resource Center.\nEllsworth, Marc A., Timothy P. Stevens, and Carl T. \nD\u2019Angio. 2010. \u201cInfant Race Affects Application \nof Clinical Guidelines When Screening for Drugs of \nAbuse in Newborns.\u201d Pediatrics 125(6):e1379\u201385.\nFeagin, Joe, and Zinobia Bennefield. 2014. \u201cSystemic \nRacism and US Health Care.\u201d Social Science & \nMedicine 103(2104):7\u201314.\nFong, Kelley. 2020. \u201cGetting Eyes in the Home: \nChild Protective Services Investigations and State \nSurveillance of Family Life.\u201d American Sociological \nReview 85(4):610\u201338.\nGunn, Veronica L., Gerald B. Hickson, and William \nO. Cooper. 2005. \u201cFactors Affecting Pediatricians\u2019 \nReporting of Suspected Child Maltreatment.\u201d \nAmbulatory Pediatrics 5(2):96\u2013101.\nGuttmacher Institute. 2021. \u201cSubstance Use during \nPregnancy.\u201d https://www.guttmacher.org/state-pol-\nicy/explore/substance-use-during-pregnancy.\nHaney, Lynne. 2010. Offending Women. Berkeley: \nUniversity of California Press.\nHarris, Lisa H. 2000. \u201cRethinking Maternal\u2013Fetal \nConflict: Gender and Equality in Perinatal Ethics.\u201d \nObstetrics & Gynecology 96(5):786\u201391.\nHeimer, Carol A. 1999. \u201cCompeting Institutions: Law, \nMedicine, and Family in Neonatal Intensive Care.\u201d \nLaw and Society Review 30(1):17\u201366.\nHurt, Hallam, Laura M. Betancourt, Elsa K. Malmud, \nDavid M. Shera, Joan M. Giannetta, Nancy L. \nBrodsky, and Martha J. Farah. 2009. \u201cChildren \nwith and without Gestational Cocaine Exposure: A \nNeurocognitive Systems Analysis.\u201d Neurotoxicology \nand Teratology 31(6):334\u201341.\nIllinois Abused and Neglected Child Reporting Act of \n1975, 325 ILCS 5/4.4 (1975).\nIrner, Tina Birk. 2012. \u201cSubstance Exposure in Utero \nand Developmental Consequences in Adolescence: \nA Systematic Review.\u201d Child Neuropsychology \n18(6):521\u201349.\nJarvis, Margaret, Jessica Williams, Matthew Hurford, \nDawn Lindsay, Piper Lincoln, Leila Giles, Peter \nLuongo, and Taleen Safarian. 2017. \u201cAppropriate \nUse of Drug Testing in Clinical Addiction Medicine.\u201d \nJournal of Addiction Medicine 11(3):163\u201373.\nKunins, Hillary Veda, Eran Bellin, Cynthia Chazotte, \nEvelyn Du, and Julia Hope Arnsten. 2007. \u201cThe \nEffect of Race on Provider Decisions to Test for \nIllicit Drug Use in the Peripartum Setting.\u201d Journal \nof Women\u2019s Health 16(2):245\u201355.\nManzer, Jamie L., and Ann V. Bell. 2021. \u201c\u2018We\u2019re a \nLittle Biased\u2019: Medicine and the Management of \nBias through the Case of Contraception.\u201d Journal of \nHealth and Social Behavior 62(2):120\u201335.\nLara-Mill\u00e1n, Armando. 2014. \u201cPublic Emergency Room \nOvercrowding in the Era of Mass Imprisonment.\u201d \nAmerican Sociological Review 79(5):866\u201387.\nLuna, Zakiya, and Kristin Luker. 2013. \u201cReproductive \nJustice.\u201d Annual Review of Law and Social Science \n9(1):327\u201352.\nMcKim, Allison. 2014. \u201cRoxanne\u2019s Dress: Governing \nGender and Marginality through Addiction Treatment.\u201d \nSigns: Journal of Women in Culture and Society \n39(2):433\u201358.\nMiller, Catherine, Amy Lanham, Christopher Welsh, \nShaalini Ramanadhan, and Mishka Terplan. 2014. \n\u201cScreening, Testing, and Reporting for Drug and \nAlcohol Use on Labor and Delivery: A Survey of \nMaryland Birthing Hospitals.\u201d Social Work in Health \nCare 53(7):659\u201369.\nMoeller, Karen E., Kelly C. Lee, and Julie C. Kissack. \n2008. \u201cUrine Drug Screening: Practical Guide for \nClinicians.\u201d Mayo Clinic Proceedings 83(1):66\u201376.\nNational Advocates for Pregnant Women. 2018. \n\u201cUnderstanding CAPTA and State Obligations.\u201d \n\n176\t\nJournal of Health and Social Behavior 63(2)\nhttps://www.nationaladvocatesforpregnantwomen \n.org/understanding-capta-and-state-obligations/.\nNetherland, Julie, and Helena Hansen. 2017. \u201cWhite \nOpioids: Pharmaceutical Race and the War on Drugs \nThat Wasn\u2019t.\u201d BioSocieties 12(2):217\u201338.\nOsterman, Michelle, and Joyce Martin. 2018. \u201cTiming \nand Adequacy of Prenatal Care in the United States, \n2016.\u201d National Vital Statistics Reports 67(3):1\u201314.\nPaltrow, Lynn M., and Jeanne Flavin. 2013. \u201cArrests \nof and Forced Interventions on Pregnant Women \nin the United States, 1973\u20132005: Implications for \nWomen\u2019s Legal Status and Public Health.\u201d Journal \nof Health Politics, Policy and Law 38(2):299\u2013343.\nPew Research Center. 2014. \u201cAmerica\u2019s Changing \nDrug Policy Landscape.\u201d https://www.pewresearch.\norg/\u00adpolitics/2014/04/02/americas-new-drug-policy- \nlandscape/.\nReinarman, Craig, and Harry G. Levine. 2004. \u201cCrack \nin the Rearview Mirror: Deconstructing Drug War \nMythology.\u201d Social Justice 31(1/2):182\u201399.\nReynolds, Lawrence A. 2005. \u201cHistorical Aspects of \nDrugs-of-Abuse Testing in the United States.\u201d Pp. 1\u2013\n10 in Drugs of Abuse: Body Fluid Testing, edited by \nR. C. Wong, and Y. T. Harley. Totowa, NJ: Humana \nPress.\nRoberts, Dorothy E. 1999. Killing the Black Body: Race, \nReproduction, and the Meaning of Liberty. New \nYork, NY: Vintage Books.\nRoberts, Dorothy E. 2002. Shattered Bonds: The Color of \nChild Welfare. New York, NY: Basic Civitas Books.\nRoberts, Sarah C. M., and Cheri Pies. 2011. \u201cComplex \nCalculations: How Drug Use during Pregnancy \nBecomes a Barrier to Prenatal Care.\u201d Maternal and \nChild Health Journal 15(3):333\u201341.\nRoberts, Sarah C. M., E. Zahnd, Carolyn Sufrin, and M. A. \nArmstrong. 2015. \u201cDoes Adopting a Prenatal Substance \nUse Protocol Reduce Racial Disparities in CPS \nReporting Related to Maternal Drug Use? A California \nCase Study.\u201d Journal of Perinatology 35(2):146\u201350.\nSchiff, Davida M., Timothy Nielsen, Mishka Terplan, \nMalena Hood, Dana Bernson, Hafsatou Diop, Monica \nBharel, et\u00a0al. 2018. \u201cFatal and Nonfatal Overdose among \nPregnant and Postpartum Women in Massachusetts.\u201d \nObstetrics and Gynecology 132(2):466\u201374.\nStone, Rebecca. 2015. \u201cPregnant Women and Substance \nUse: Fear, Stigma, and Barriers to Care.\u201d Health & \nJustice 3(1):1\u201315.\nStuddert, David M., Michelle M. Mello, William M. \nSage, Catherine M. DesRoches, Jordon Peugh, Kinga \nZapert, and Troyen A. Brennan. 2005. \u201cDefensive \nMedicine among High-Risk Specialist Physicians \nin a Volatile Malpractice Environment.\u201d JAMA \n293(21):2609\u201317.\nSubstance \nAbuse \nand \nMental \nHealth \nServices \nAdministration. 2012. Clinical Drug Testing in \nPrimary Care: Technical Assistance Publication \n(TAP) 32. HHS Publication No. (SMA) 12-4668. \nRockville, MD: Substance Abuse and Mental Health \nServices Administration.\nSweet, Paige L. 2019. \u201cThe Paradox of Legibility: \nDomestic Violence and Institutional Survivorhood\u201d. \nSocial Problems 66(3):411\u201327.\nTerplan, Mishka, and Howard Minkoff. 2017. \u201cNeonatal \nAbstinence Syndrome and Ethical Approaches to the \nIdentification of Pregnant Women Who Use Drugs.\u201d \nObstetrics & Gynecology 129(1):164\u201367.\nTiger, Rebecca. 2012. Judging Addicts: Drug Courts and \nCoercion in the Justice System. New York: NYU \nPress.\nTimmermans, Stefan, and Alison Angell. 2001. \u201cEvidence-\nBased Medicine, Clinical Uncertainty, and Learning \nto Doctor.\u201d Journal of Health and Social Behavior \n42(4):342\u201359.\nTimmermans, Stefan, and Jonathan Gabe. 2002. \n\u201cIntroduction: Connecting Criminology and Sociology \nof Health and Illness.\u201d Sociology of Health & Illness \n24(5):501\u201316.\nU.S. Department of Health and Human Services. 2021. \n\u201cChild Maltreatment 2019.\u201d https://www.acf.hhs.\ngov/cb/research-data-technology/ statistics-research/\nchild-maltreatment.\nVulliamy, Anne P., and Richard Sullivan. 2000. \n\u201cReporting Child Abuse: Pediatricians\u2019 Experiences \nwith the Child Protection System.\u201d Child Abuse & \nNeglect 24(11):1461\u201370.\nWarner, Elizabeth A., Robert M. Walker, and Peter D. \nFriedmann. 2003. \u201cShould Informed Consent Be \nRequired for Laboratory Testing for Drugs of Abuse \nin Medical Settings?\u201d The American Journal of \nMedicine 115(1):54\u201358.\nWilson, James Maxwell Glover, and Gunnar Jungner \n1968. Principles and Practice of Screening for \nDisease. Geneva: World Health Organization.\nWood, Elissa. 2006. The Costs and Utility of Parental \nDrug-Testing in Child Protection: A Review of the \nAvailable Literature and Commentary. Sydney: \nNational Drug and Alcohol Research Centre.\nAuthor Biography\nKatharine McCabe is a postdoctoral researcher with the \nJohn J. Reilly Center for Science, Technology, and Values \nat the University of Notre Dame. She studies the intersec-\ntion of medicine and law, inequalities in maternal health \ncare, and social and policy responses to perinatal sub-\nstance use.\n\n\n    Scientific Research Findings:", "answer": "\u2022 Providers framed criminal suspicion in medical terms. By blending clinical definitions and routines with forensic interests, the legal intent of drug testing was obscured from patients.\n\u2022 Providers grappled with and expressed uncertainty about the risks associated with different types of chemical exposures. In the face of uncertainty regarding the effects of certain substances, providers swapped clinical for legal risk assessments.\n\u2022 Providers made decisions about drug testing based on patient traits such as perceived race, class, and age. Providers also shielded some patient populations from forensic testing based on their perceived class status.", "id": 71}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465211067209\nJournal of Health and Social Behavior\n2022, Vol. 63(3) 337\u00ad\u2013356\n\u00a9 American Sociological Association 2022\nDOI: 10.1177/00221465211067209\njhsb.sagepub.com\nOriginal Article\nBackground\nFor more than a century, governments have designed \npolicies intended to address drug use, dependence, \nand mortality (Musto 1999). Through what Foucault \n(1991) termed governmentality, policies may induce \ninstitutions to foster individual-level behavioral \nchoices that cohere with the aims of the government. \nThroughout the twentieth century, governmental pol-\nicy regarding drug supply increasingly focused on the \ncriminal justice system (MacCoun and Reuter 2001). \nBy relying on policies that define certain drugs as \nillicit, there is a natural reliance by the state on the \ncriminal justice system to curb supply. According to \nFoucault (1982), however, use of policing to influence \nindividual-level behavior is inefficient and often \ncounterproductive. Surveillance systems that induce \nself-governing behavior rather than direct coercion \nare more efficient because they reduce the resources \nrequired to align behavior with governmental aims \n(Foucault 1995). Policies may employ influence on \nbehavior across a continuum from the tacit forms of \nbehavioral persuasion akin to Foucault\u2019s panopticon \n1067209 HSBXXX10.1177/00221465211067209Journal of Health and Social BehaviorVuolo et al.\nresearch-article2022\n1The Ohio State University, Columbus, OH, USA\n2Purdue University, West Lafayette, IN, USA\nCorresponding Author:\nBrian C. Kelly, Department of Sociology, Purdue \nUniversity, 700 W State St, West Lafayette, IN 47907, \nUSA. \nEmail: bckelly@purdue.edu\nSurveillance, Self-Governance, \nand Mortality: The Impact of \nPrescription Drug Monitoring \nPrograms on U.S. Overdose \nMortality, 2000\u20132016\nMike Vuolo1\n, Laura C. Frizzell1, and Brian C. Kelly2\nAbstract\nPolicy mechanisms shaping population health take numerous forms, from behavioral prohibitions to \nmandates for action to surveillance. Rising drug overdoses undermined the state\u2019s ability to promote \npopulation-level health. Using the case of prescription drug monitoring programs (PDMPs), we contend \nthat PDMP implementation highlights state biopower operating via mechanisms of surveillance, whereby \nprescribers, pharmacists, and patients perceive agency despite choices being constrained. We consider \nwhether such surveillance mechanisms are sufficient or if prescriber/dispenser access or requirements \nfor use are necessary for population health impact. We test whether PDMPs reduced overdose mortality \nwhile considering that surveillance may require time to reach effectiveness. PDMPs reduced opioid \noverdose mortality 2 years postimplementation and sustained effects, with similar effects for prescription \nopioids, benzodiazepines, and psychostimulants. Access or mandates for action do not reduce mortality \nbeyond surveillance. Overall, PDMP effects on overdose mortality are likely due to self-regulation under \nsurveillance rather than mandated action.\nKeywords\ndrugs, mortality, overdose, policy, surveillance\n\n338\t\nJournal of Health and Social Behavior 63(3) \nthrough direct mandates for action. In this regard, \ncoercion may occur not only through explicit man-\ndates but also in implicit forms that give the appear-\nance of choice as well. Surveillance systems within \nthe scope of the health care system may appear, at face \nvalue, to be less coercive.\nBecause the health care system does not monitor \nillicit drug supply and largely is reactive to patterns of \nsubstance use, these inefficiencies were further \nbrought to light by the overdose crisis. By fundamen-\ntally altering the source of psychoactive drugs, the \nrise of prescription drug misuse in the twenty-first \ncentury disrupted the state\u2019s control of drug use, \ndependence, and mortality through criminalization \nsuch that \u201claw enforcement is unlikely to reach PPD \n[psychoactive prescription drugs] diversion effec-\ntively\u201d (Fischer, Bibby, and Bouchard 2010:2062). In \nterms of supply for nonmedical use, most sources are \nlegally written prescriptions for the users themselves \nor for friends or family members as opposed to acqui-\nsition from drug dealers (Hulme, Bright, and Nielsen \n2018). Furthermore, surges in overdoses during the \ntwenty-first century outpaced prior drug trends and \nundermined the state\u2019s ability to control population-\nlevel health via what Foucault (1990) referred to as \nbiopower\u2014a \u201ctechnology of power\u201d for disciplining \nand controlling populations. With the health care sys-\ntem now a key source of drug supply, combined with \na lapse in biopower, other policies were needed to \naddress the state\u2019s interests. Questions remain about \nhow policies may best achieve these ends, particu-\nlarly whether state surveillance of medical processes \nalone may reduce adverse outcomes or mandates for \naction are necessary.\nWithin this article, we examine the role of pol-\nicy implementation within U.S. states, specifically \nprescription drug monitoring programs (PDMPs), \non overdose mortality. As defined by the Centers \nfor Disease Control and Prevention (CDC; 2017), \n\u201cA prescription drug monitoring program (PDMP) \nis an electronic database that tracks controlled sub-\nstance prescriptions in a state.\u201d Drugs are deemed \ncontrolled substances by federal government sched-\nuling, typically defined as psychoactive drugs with \nabuse potential, and include both legally prescribed \nmedications and drugs with complete prohibitions. \nAs described in the following, independent state-\nlevel systems create considerable variability in cer-\ntain details, such as who can access PDMPs and \nwho is required to check the system. Typically, \nhowever, PDMPs not only track all prescribed con-\ntrolled substances but also produce reports at both \nthe patient and population level to identify overpre-\nscribing (Perrone and Nelson 2012).\nWe consider the role of PDMPs as a means to \nreduce overdose mortality by disrupting the circula-\ntion of prescription drugs available for consump-\ntion by the population. We argue that the \nimplementation of PDMPs is an example of state \nbiopower operating via a mechanism of surveil-\nlance (Foucault 1990, 1995). This approach allows \nmultiple actors\u2014physicians, pharmacists, and \npatients\u2014to perceive a level of action and agency \ndespite their choices being constrained (Bird and \nRieker 2008; Foucault 1982) and thus should be \neffective regardless of whether providers have \naccess to or are required to check these systems. \nSuch surveillance facilitates self-monitoring under \nconditions whereby individuals may be checked \nand scrutinized, regardless of whether this actually \noccurs. That is, unverifiable surveillance is a more \nefficient form of biopower than direct behavioral \ncoercion. Ultimately, a use of biopower in the form \nof PDMPs constrains the choices of multiple actors \nand in doing so impedes the circulation of con-\ntrolled substance medications, which in the aggre-\ngate impacts overdose mortality. In the analyses \nthat follow, we test this mechanism by first analyz-\ning the efficacy of PDMPs\u2019 general system of \n\u00adsurveillance prior to analyses of policies that spe-\ncifically grant access to PDMP databases to provid-\ners/dispensers or require providers/dispensers to \ncheck this system prior to writing or filling a pre-\nscription for a controlled substance. Thus, we con-\nsider that unverifiable surveillance of institutional \npractices and patient behavior orchestrated by pub-\nlic policy may shape patterns of drug-related mor-\ntality. We also consider that the effect of such \nsystems of surveillance may have an institutional \nlag given that self-monitoring under conditions of \nsurveillance may take time to scale up.\nThe Case of Overdose in the Twenty-\nFirst Century and Policy Responses\nIncreases in drug-related morbidity, mortality, and \ndependence over the past two decades affected U.S. \npopulation health in adverse ways. The rise in pre-\nscription drug misuse since the 1990s was consider-\nable (Atluri, Sudarshan, and Manchikanti 2014; \nBirnbaum et\u00a0 al. 2011; Dart et\u00a0 al. 2015; Maxwell \n2011), followed by more recent increases in heroin \nuse and problems related to synthetic opioids such \nas fentanyl (Carlson et\u00a0al. 2016). In this manner, the \ncrisis has experienced multiple waves and continues \nto expand because drugs such as cocaine and meth-\namphetamine are increasingly adulterated with fen-\ntanyl (Ciccarone 2021). These multiple waves have \n\nVuolo et al.\t\n339\ncontributed to growth in drug consumption, drug \ndependence, and overdose mortality over time. In \naddition to opioids, consumption and adverse out-\ncomes from other prescribed controlled substances \nalso escalated (Bachhuber et\u00a0al. 2016; Scholl et\u00a0al. \n2019). Increasing dependence on opioids and other \ndrugs led to growth in the unmet treatment burden \n(McCabe, Cranford, and West 2008). Beyond drug \ndependence, the United States experienced massive \ngrowth in overdoses over the past 20 years, which \ncontributed to lower life expectancies (Case and \nDeaton 2015; Dowell et\u00a0 al. 2017; Imtiaz, Probst, \nand Rehm 2018). Surging overdose mortality con-\ntributed greatly to years of life lost across socioeco-\nnomic groups (Ho 2017). Between 1999 and 2019, \nnearly 850,000 people died in the United States \nfrom a drug overdose, and early evidence indicates \nfurther escalation of overdose deaths during 2020 \n(CDC 2020). Considering the adverse effects on \npopulation health, the drug overdose crisis poten-\ntially undermined state biopower.\nIn response to problems emerging from these \npatterns of substance use, state-level governments \nimplemented various policy measures to combat \nescalating trends in abuse, dependence, and over-\ndose. Within the United States, drug policies operate \nin a multilevel manner. Whereas federal-level poli-\ncies intervene on the international drug trade, \nincluding heroin and illicitly manufactured fentanyl, \nPDMPs represent actions taken at the U.S. state \nlevel to curb pharmaceutical drugs acquired within \nthe context of the medical system. Although modest \nwithin the larger umbrella of U.S. drug control, \nPDMPs nonetheless directly target medicalized sub-\nstances recognized as fueling the overdose crisis. \nEvaluation of such policies and the distinctions in \nhow they are implemented remains imperative to \nidentify efficacious solutions to reduce overdose \nmortality and other substance abuse problems. We \nfocus on the implementation of PDMPs\u2014given \ntheir role in surveillance, long history, wide cover-\nage, and application to several classes of substances \n(Gugelmann and Perrone 2011)\u2014in shifting over-\ndose mortality while accounting for the broader pol-\nicy context by considering the passage of other \nrelevant policies over the period of observation.\nOn a basic level, all PDMPs create a database of \nprescriptions for controlled substances that allow for \nsurveillance over potential exploitations by either \nproviders or patients, although these systems of sur-\nveillance primarily affect the institutional duties of \nprescribers/dispensers. Although PDMPs have \ndrawn attention during the overdose crisis, some \nPDMPs preceded the rise in overdose mortality \nduring the twenty-first century; the earliest were \nimplemented in New York approximately a century \nago, and even electronic systems emerged approxi-\nmately 30 years ago. Thus, many currently practic-\ning physicians and pharmacists have been familiar \nwith these systems in principle even if not directly \nsubject to their regulation until more recently. With \nrespect to recent implementation, prescribers and \npharmacists often received information about these \nsystems from state medical and pharmacy boards \nand through continuing education. Information on \nhow prescribers/dispensers can access the state \nPDMP is typically maintained online via state web-\nsites. Use of PDMPs can be variable, especially \ngiven that practitioners indicate it can be time con-\nsuming; nonetheless, as of 2014, 87% of physicians \nin states with a PDMP indicated they had the capac-\nity to access PDMP information at the site of their \nclinical practice and had used a PDMP to check \npatient prescriptions (Rutkow et\u00a0al. 2015).\nAs of 2000, PDMPs were active in only 12 U.S. \nstates. In other U.S. states, essentially no surveil-\nlance system existed to assist in identifying poten-\ntial misuse. As the crisis became more visible, \nPDMP implementation was rapid, with PDMPs in \n22 U.S. states by 2005, 39 by 2010, and 49 by 2015, \nwith Missouri the lone holdout; St. Louis County \nestablished a PDMP during 2017 in response to the \nstate-level absence. Prior to that, all PDMP policies \nwere passed at the state level (and Washington, \nD.C.) rather than locally. Although PDMPs have a \nlengthy history, there is variability across and \nwithin U.S. states over time (Pauly et\u00a0 al. 2018), \nsummarized in Figure 1. For example, although \nprescribers/dispensers have not always had access \nto PDMP records, all active PDMPs now grant such \naccess. Furthermore, as of 2012, none required \nphysicians/dispensers to check the PDMP database \nwhen filling a prescription. By 2016, 16 U.S states \nhad passed such laws for prescribers, and 5 states \nhad passed such a law for dispensers. Given this \nsmall proportion and timeframe within which \nrequirements to check the PDMP registry were \nactive, we argue and empirically demonstrate that \nany effect of PDMPs observed over the last two \ndecades is due to the invisible system of surveil-\nlance they create rather than newer, explicit require-\nments for professionals to check them.\nPolicy and Surveillance as State \nBiopower\nPolicy interventions on behavior can take a range of \nforms, from the provision of information for \n\n340\t\nJournal of Health and Social Behavior 63(3) \ndecision-making to the elimination of choice through \nprohibition, and these imply varying impacts on per-\nsonal autonomy (Griffiths and West 2015). A key \ncomponent of the implementation of PDMPs is the \ncreation of a condition of surveillance. The aware-\nness that one may be monitored can create a self-\ngoverning system of social control even when it is \nunverifiable that one is actively being monitored \n(Foucault 1995). When prescribers and pharmacists \nbecome immersed in a system of state-level over-\nsight on the prescription and distribution of con-\ntrolled substances, they may subtly self-check and \nalter their behavior. This self-check on institution-\nally based practices can take multiple forms. For \nexample, when aware of PDMP surveillance, pre-\nscribers may consciously or subconsciously ques-\ntion whether to write a prescription, resulting in \nfewer prescriptions overall. Similarly, they may pre-\nscribe fewer doses per prescription, particularly for \nnew patient cases, resulting in fewer pills in circula-\ntion overall. In rarer instances, PDMPs may \n\u00adfacilitate reporting behaviors by which medical pro-\nfessionals under conditions of unverifiable surveil-\nlance alert authorities to suspected misuses by \npatients or even other prescribers/dispensers. In a \nPDMP system where prescribers/dispensers do not \nhave direct access to the database, this surveillance \nprocess would be the only possibility, such that any \neffect would be strictly through a mechanism of \nself-regulation under surveillance. In other words, \nbehavior can change even in the absence of inten-\ntional efforts to access the system. In a PDMP sys-\ntem where prescribers/dispensers have database \naccess, they can also self-govern by actively check-\ning the system either as habit or if something seems \namiss with a patient or request, with use of PDMPs \ncommon in U.S. states providing access (Rutkow \net\u00a0al. 2015). Importantly, these mechanisms may be \nin effect even if providers are not required to explic-\nitly check a monitoring system when writing or fill-\ning a prescription. As such, a system of surveillance \nlooming over prescribing, dispensing, and care-\nseeking behaviors may instill self-governance even \nif individuals are not mandated to act under these \npolicies.\nFor Foucault (1995), this internalized self- \ngovernance is the perfect exercise of state power \nbecause control is exercised over the greatest \n\u00adnumber of individuals while reducing costs to the \nsystem from third-party oversight, a form of power \nbest exemplified by the panopticon. Along with \ninstitutions such as schools and criminal justice, \nhealth was one of the key examples within which \nFoucault (1995) described panopticon-like power. \nFoucault conceptualized biopower in referring to \nthe power states possess to influence population-\nlevel health, life, death, reproduction, and longev-\nity. In this manner, biopower is a technology of \npower for disciplining and controlling populations \nthrough techniques such as surveillance, which \nstimulate the population to govern its behavior in \nways that conform to governmental aims (Foucault \n1990). Although Foucault was critical of power in \nmany instances, he also recognized the potential for \nbiopower to be a positive mechanism contributing \nto population-level well-being (Foucault 1990). \nInfectious disease surveillance remains a key exam-\nple of the opportunity to leverage surveillance in \nthe interest of population health (Fairchild 2003). In \nthe case of overdose mortality, states have capital-\nized on biopower to implement PDMPs that \n\u00adconstrain the autonomy of health professionals at \nthe individual level for an overall population-level \nbenefit via increased surveillance.\nState power, however, is not simply disciplining \nindividuals into conformity through coercion. \nSummarizing Foucault (1982), Garland (1997:175) \nstates that governmental power \u201cconstructs individ-\nuals who are capable of choice and action, shapes \nthem as active subjects, and seeks to align their \nchoices with the objectives of governing authori-\nties.\u201d Within medical sociology, constrained \nchoices theory contends directly with this dynamic \n(Bird and Rieker 2008). Although originally used to \nexplain gender differences in health, it provides a \nconceptual framework for multilevel processes that \nlead to differential shaping of health behaviors \n(Vuolo, Kadowaki, and Kelly 2016). That is, indi-\nviduals are agentic actors in terms of their \nFigure 1.\u2002 Prescription Drug Monitoring \nPrograms (PDMPs) across State and Year.\nSource: Data from Prescription Drug Abuse Policy \nSystem.\n\nVuolo et al.\t\n341\nhealth-related decision-making but constrained \nwithin the policy, community, and family environ-\nments (among others) within which they are situ-\nated. A Foucauldian interpretation adds that the \nstate has a vested interest in guiding these con-\nstraints. Furthermore, most work has focused on \nconstraints for individual health decisions rather \nthan also incorporating how constraints affect pro-\nviders and institutions.\nNotably, surveillance through PDMPs \u00adconstrains \nthe decision-making of multiple actors (Bird and \nRieker 2008)\u2014patients, physicians, and pharmacists\u2014 \nand in doing so provides the state with opportuni-\nties to guide the constraints of multiple types of \nindividuals central to the circulation of prescribed \ncontrolled substances and ultimately, overdose \nmortality. With choices effectively constrained but \nwith enough perceived agency, surveillance affects \nthe decision-making processes that shape the diver-\nsion of prescription drugs and drug dependence that \nfacilitates illicit drug use. For health professionals, \nthe suspension of professional licensing by the state \ncreates a potential sanction should surveillance \ndetect irregularities on the part of these prescribers \nand pharmacists. The PDMP system may also con-\nstrain and reduce the practice of \u201cdoctor shopping\u201d \nby patients. This constraint may occur after patients \nbecome aware that the state may be monitoring \ntheir receipt of prescriptions and that this catalog of \nprescriptions filled by multiple doctors may be used \nas evidence against them, particularly those who \nmay face distribution charges. The opportunity to \nintervene on multiple actors through self-regulation \nunder surveillance enhances policy impact, which \nindicates the full scope of biopower.\nWithin the context of constrained choices, the \nsurveillance effects of PDMPs apply across popu-\nlations when implemented in a U.S. state. In this \nmanner, the comprehensive nature of surveillance \ngrounded within PDMPs allows their effects to be \nnoticeable precisely because they create a system of \nsurveillance that applies universally across the pop-\nulation; this universality is a key characteristic of \nbiopower. Although even in a strong system of sur-\nveillance, some individuals find ways to evade \ndetection (Goffman 2009). Thus, a small pool of \nunscrupulous prescribers/dispensers could still seek \nways to divert prescriptions, whereas users may \ndetermine ways to work within the current system \nof surveillance. Yet the overall system of surveil-\nlance creates constraints for a large proportion of \nprescribers, pharmacists, and patients that inhibit \nthe circulation and potential diversion of medica-\ntions. The one-on-one doctor\u2013patient behavior, \nalthough surely a component, does not alone indi-\ncate how changes in aggregate have an effect. Most \nindividuals who misuse prescription medications \nobtain these drugs from friends and family, with \nprimarily heavy users turning to the black market or \ndoctor shopping (Hulme et\u00a0al. 2018). As such, the \nproblem of mortality extends beyond any single \nindividual momentary constrained behavior for \nhow a provider writes a prescription; the aggregate \neffect of individual constrained choices on local \navailability matters for misuse and mortality.\nOf policies that have been passed, PDMPs most \ndirectly address the state\u2019s interest in controlling \ndrug supply because they create a system of surveil-\nlance advantageous to the state in its universal \napplication across all controlled substance prescrip-\ntions. Other policies enacted, such as enhanced nal-\noxone access and Good Samaritan laws, require \nindividual-level action, which may depend on vary-\ning individual and community resources, and are \nnot subject to the same level of governmental over-\nsight in their use. In some ways, these policies \nrelease a certain level of governmental control. By \nabsolving users of responsibility should they call \nstate actors such as first responders, Good \nSamaritan laws prevent scrutiny by the criminal \njustice system (Haegerich et\u00a0 al. 2014)\u2014albeit in \nways that may yield better mortality outcomes. By \nreducing the need for hospital admission after use \nof an overdose reversal drug, naloxone access laws \n(Davis and Carr 2017) curtail surveillance by the \nhealth system. Notably, these policies create a \nmechanism whereby the state can intervene only on \nadverse outcomes rather than patterns of drug con-\nsumption more generally.\nImportantly, surveillance-driven constraints \nmay require time to shift self-governance in a man-\nner that produces measurable results. Clinical \ninteractions have cultural components rather than \nsimply being crude measurements of physiology \n(Dubbin, Chang, and Shim 2013; Light and Levine \n1988). As such, there may be a lag period as cul-\nture within institutions changes. Cultural shifts in \ncognition often occur in a tacit and incremental \nmanner rooted in practice rather than a paradig-\nmatic shift (Luhrmann 1991). Within the context of \nthe institution of medicine, shifts requiring behav-\nioral change in clinical interactions via self-moni-\ntoring may require time to scale up. In this regard, \nthere is a lag between the recognition of surveil-\nlance and the routinization of self-governance in \nclinical practice. As awareness of surveillance and \nresulting self-consciousness gradually become \ningrained in clinical practice, institutional and \n\n342\t\nJournal of Health and Social Behavior 63(3) \npersonal behavioral shifts emerge that lead to \nobservable effects on the circulating pool of con-\ntrolled medication and its subsequent misuse. \nThese forms of self-governance may also initially \noccur unevenly, depending on familiarity and \nexperience \nwith \npatients \n(Chiarello \n2015). \nSpecifically, the surveillance effect of PDMPs may \nbecome more readily situated within clinical \nencounters with new patients, whereas such regu-\nlation within the context of long-term patient rela-\ntionships may require more time to build as \nconsciousness of surveillance more fully takes \nhold.\nAlthough Foucault\u2019s arguments about the power \ninherent in surveillance is persuasive, we must also \nconsider the possibility that surveillance and self-\ngovernance alone may not be sufficient to affect \noverdose mortality. As noted previously, U.S. states \nmore recently implemented policies with specific \nrequirements for action on the part of prescribers/\ndispensers (Pauly et\u00a0al. 2018). The requirement that \nindividuals engage in specific types of action can be \na key policy lever for shifting health behaviors by \ncreating a system of constraints that influences \ndecision-making within a broader context (Bird and \nRieker 2008). Additionally, direct mandates for pre-\nscriber/dispenser action may create a more immedi-\nate effect by eliminating any lag in behavioral \nscale-up. As such, we also assess whether require-\nments for action drive any effect on overdose mor-\ntality, which would shift evidence away from the \nsurveillance and self-governance effect described \npreviously. However, we note that for Foucault \n(1982), such requirements that apply power through \nbehavioral coercion are unnecessary, less efficient, \nand even counterproductive. He stressed that power \nis exercised over free subjects and that governing is \ncultivated through the ability of individuals to make \nchoices for themselves while being steered toward \nchoices aligned with governmental aims.\nThe Empirical Case for Studying PDMPs\nIn addition to considering an important theoretical \nmechanism of surveillance on a critical population \nhealth outcome, there is an empirical case to be \nmade for our approach. While U.S. states increas-\ningly enacted policies in response to growing drug \ndependence and overdose problems, research \ntoward comprehensively examining the influence of \nPDMPs on trends in fatal overdose over a period of \nrapid policy implementation continues to build. \nRecent reviews have categorized these results as \n\u201cmixed\u201d (Finley et\u00a0al. 2017), \u201cinsufficient\u201d (Fink \net\u00a0 al. 2018), \u201climited\u201d (Rhodes et\u00a0 al. 2019), and \n\u201cconflicting\u201d (Moride et\u00a0al. 2019). Limitations of \nexisting studies on these policies include: (a) exami-\nnations of policies in single locations, (b) consider-\nations of single policies without attention to the \nwider policy context, (c) truncated periods of obser-\nvation, and (d) inability to account for unobserved \nheterogeneity. Most studies also do not account for \ntremendous heterogeneity within states.\nUtilizing sociological theories and integrating \nmultiple data sets, we build on this prior research to \nmove beyond each of these limitations through the \ninclusion of U.S. counties across the nation, con-\ntrolling for multiple policies to account for broader \npolicy contexts, an extended time period of policy \nimplementation, and use of lead and lag structures \nwithin fixed effects models of drug-related mortal-\nity. We examine overdose rates at the county level \nusing restricted access mortality data. Although \noverdose-related policies are enacted at the state \nlevel, analysis of mortality at the county level per-\nmits a more nuanced assessment given considerable \nwithin-state heterogeneity, an important distinction \ngiven noted differences between rural and urban \nareas in terms of prescription drug misuse and mor-\ntality (Keyes et\u00a0al. 2014; Peters et\u00a0al. 2020). For \ninstance, rural counties in central Pennsylvania \npresent a very different context than Philadelphia \nCounty despite being within the same state (Monnat \n2018). We first examine the possible impact of \nPDMPs, net of other policies, on overall opioid \noverdose mortality in the United States. Importantly, \nwe incorporate the time since policy passage \nbecause we anticipate that the effects of surveil-\nlance may not be immediate. We follow by looking \nat these effects on overdose mortality specific to \nprescription opioids, prescription benzodiazepines, \nand stimulants to show the potential effects of \nPDMPs beyond the general opioid overdose crisis. \nAdditionally, we consider whether these effects \nhold for subsequent PDMP policies that grant pre-\nscriber/dispenser access or require prescribers/dis-\npensers to check the system before writing or filling \na prescription. In this manner, our analyses permit \nan assessment that contrasts the effects of surveil-\nlance-stimulated self-governance and mandated \naction. Furthermore, most existing studies are \ndescriptive and lack a theoretical foundation, creat-\ning an urgent need for theoretically driven studies \nof prescription drug misuse (Watkins 2016). By \nincorporating sociological theories, we are better \nable to highlight the mechanisms by which PDMPs \nmight mitigate mortality and extend considerations \nof policy mechanisms to other health domains.\n\nVuolo et al.\t\n343\nData and Methods\nOverdose Mortality Data\nWe utilized CDC restricted access mortality files \nfrom 2000 to 2016. The database contains every \ndeath in the United States in each year. Cause of \ndeath codes allowed identification of drug over-\ndoses, which we then aggregated to the county \nlevel. We calculated age-adjusted overdose rates per \n100,000 for each county. Unlike public use data, \nrestricted data contain no suppressed information, \nwhich permits more precise analysis of counties \nwith fewer than 10 deaths per year.\nOur first outcome is age-adjusted death rate due \nto any opioid overdose within counties each year. \nWe utilized the CDC\u2019s operational definition for \nopioid overdose (CDC 2014). Such deaths have \nICD-10 underlying cause codes of \u201cdrug poison-\ning,\u201d X40-X44, X60-X64, X85, and Y10-Y14, \ntogether with ICD-10 multiple causes codes \nT40.0-T40.4 and T40.6. The latter multiple cause \ncodes specify opioids as a cause of death, which \ninclude prescription opioids, such as hydrocodone \n(Vicodin) and oxycodone (OxyContin), and mor-\nphine, heroin, and synthetic opioids, such as fen-\ntanyl. In addition to total opioid overdoses, we \nexamined overdoses from prescription opioids \n(T40.2) separately as an outcome and benzodiaze-\npines (T42.4) and psychostimulants with abuse poten-\ntial (T43.6) in separate analyses. Benzodiazepines, \nalso known as tranquilizers, are prescribed for condi-\ntions such as anxiety and insomnia; well-known \nexamples include alprazolam (Xanax), clonazepam \n(Klonopin), and diazepam (Valium). Psychostimulants \ninclude prescriptions, such as methylphenidate (Ritalin) \nand amphetamines (Adderall), and methamphet-\namine but not cocaine, which is classified sepa-\nrately (T40.5). We also note that deaths attributed to \noverdose can include multiple coded substances. \nBecause polydrug use heightens overdose risk \nbeyond that of a single substance, examining the \nrole of PDMPs on reductions in deaths involving \nprescription drugs in conjunction with another sub-\nstance remains important. We considered polydrug \nuse by additionally examining each of the separate \nclasses in combination with an opioid.\nPolicy Data\nWe utilized the Prescription Drug Abuse Policy \nSystem for a comprehensive listing of policy pas-\nsage in each U.S. state between 2000 and 2016. We \nthen created a state-year data set of the presence of \npolicies across the observation period. Our central \nvariable of interest is presence of a state-level \nPDMP (coding described in the following). We also \nconsidered alternative specifications via variables \nfor whether prescribers/dispensers have access to \nthe PDMP and whether prescribers/dispensers are \nrequired to check the PDMP when writing/filling a \nprescription. To account for other policies that may \naffect overdose rates, we included covariates for \nnaloxone access expanded beyond medical settings, \nGood Samaritan laws absolving criminal/civil lia-\nbility in reporting an overdose, pain clinic restric-\ntions, and medical marijuana laws (not passed \nbecause of the overdose crisis but an important pol-\nicy covariate given their pertinence to managing \npain).\nCounty- and State-Level Covariates\nFrom the U.S. Census Bureau\u2019s American Community \nSurvey (ACS) and decennial censuses, we included \ncounty-level time-varying covariates, specifically \npopulation density, median household income, and \npercentages foreign-born, female-headed house-\nholds, non-Hispanic Black, Hispanic, and over 25 \nwith a bachelor\u2019s degree. We included an interaction \nbetween percentage foreign-born and Hispanic \ngiven their high correlation and a significant inter-\naction effect in several model specifications. To \ninclude all counties, we required 5-year ACS esti-\nmates because only larger counties are available \nover shorter estimates. Because these begin in 2009, \nwe linearly interpolated years between the 2000 \ncensus and 2009 (with the exception of population \ndensity because estimates are available annually). \nWe also included county-level unemployment rate, \navailable annually from the Bureau of Labor Statistics. \nFrom the Annual Survey of State Government \nFinances, we included state-level per capita spending \non education, public welfare, hospitals, and health.\nAnalysis\nWe used fixed effects models to determine the effect \nof PDMPs on overdose mortality (Allison 2009; \nHalaby 2004; Winship and Morgan 1999). The \nmethod\u2019s strength is the elimination of unobserved \nheterogeneity by differencing all predictors and the \noutcome from its county-specific average. Fixed \neffects estimators are robust to observed or unob-\nserved time-invariant variables, which removes any \nconstant county-level effects. To account for the \nwell-known increase in overdose rates and provide \nestimates of the effect of PDMPs independent of \nthat increase, all models included fixed effects for \n\n344\t\nJournal of Health and Social Behavior 63(3) \nyear. Finally, we included a standard error cluster \ncorrection to account for dependencies between \ncounties within states. There are three additional \nassumptions of fixed effects models: exogenous \ntreatment selection, the parallel trends assumption, \nand causal direction (Vaisey and Miles 2017). \nAppropriate tests demonstrate no evidence of \nendogenous selection (Morgan and Winship 2007; \nVaisey and Miles 2017). The coding of our indepen-\ndent variable addressed the latter two assumptions.\nBecause policies may take time to reach effec-\ntiveness for outcomes related to overdose (Bao \net\u00a0al. 2016), we took advantage of the fact that all \nU.S. states but one (Missouri) implemented a \nPDMP prior to 2016 to recode the measure of the \npolicy as centered around the year of implementa-\ntion. That is, the point of passage was coded as 0; \nyears prior were coded \u22121, \u22122, and so on; and years \nafter were coded 1, 2, and so on. This approach \nincorporated lead and lag effects directly into the \nmodel without sacrificing data. As the number of \nU.S. states at the extremes becomes small and thus \nthe estimates become less precise, we capped the \nnumber of years in either direction at \u00b15.1 In mod-\nels that consider PDMP prescriber/dispenser access \nand requirements, we used traditional leads and \nlags because the lack of consistent implementation \nof these forms of PDMPs prohibit the prior \napproach, and as noted previously, requirements for \naction may yield more immediate effects. The tests \nfor leads for both approaches allowed us to test for \nreverse causality and parallel trends. For the for-\nmer, significant leads would indicate that an \nincrease or decrease in overdoses preceded policy \nimplementation and that the relationship actually is \ntemporally reversed.2 For the latter, plotting the \nleads allows determination of whether the period \nprior to adoption was different across counties prior \nto the point of adoption. All other policies were \ncoded as their status on the first of any given year. \nWe lagged all other covariates by 1 year to ensure \nappropriate temporal order.\nThe United States contains 3,142 counties, \namounting to 53,414 possible county-year observa-\ntions from 2000 to 2016. However, 13 counties \nrestructured boundaries such that we started those \nseries at later years when their boundaries cohered \nwith the present because prior years were not com-\nparable within the CDC data. In our main PDMP \nmodels (not including Missouri), there were 3,025 \ncounties across 17 years, which (accounting for \nrestructuring) included 48,239 observations. In our \nmodels for PDMP access and restrictions, the use of \ntraditional leads and lags eliminated data on both \nsides of the time series (2 years at the beginning for \nlags and 1 year at the end for leads). In these mod-\nels, there were 3,140 counties and 46,939 \nobservations.\nResults\nDescriptive Statistics\nTable 1 provides descriptive statistics, pooled over \ncounties and years. Although our models use differ-\nent subsets of data, we show the descriptive statis-\ntics from 2000 to 2016 (accounting for restructuring) \nbecause this covers the years used in our main mod-\neling approach (but includes Missouri). On average \nacross the observation period, the county-level age-\nadjusted overdose rate for all opioids was 5.54 per \n100,000. Overdoses involving prescription opioids \n(2.79), benzodiazepines (1.58), and stimulants (.66) \nwere notable. Figure 2 shows trends in these out-\ncomes over time averaged at the county level, con-\nfirming well-known increases in all opioid \noverdoses and flattening of those due specifically to \nprescription opioids.\nBenzodiazepines stabilized about the same time \nas prescription opioids, whereas psychostimulants \nhave seen increases in recent years. Given more \nrecent implementation in most U.S. states, \nexpanded naloxone laws (13.6%), Good Samaritan \nlaws (9.0%), and pain clinic restriction laws \n(10.5%) have lower coverage than PDMPs across \nthe entire period. County-years with medical mari-\njuana laws comprise 15.0% over the period. Finally, \nTable 1 displays averages for demographic and \nexpenditure covariates.\nAlthough Figure 1 showed the growth in PDMPs \nacross U.S. states, we also note the total percentage \nof county-years because this is our level of analysis \nfor each PDMP variable. With relatively wide cover-\nage, 66.1% of county-years had a PDMP during the \nperiod of observation. Prescribers/dispensers were \ngranted access in 58.5% and 56.0% of county-years, \nrespectively. As a recent change, only 3.9% and \n1.8% of county-years have PDMP checking require-\nments for prescribers and dispensers, respectively.\nPresence of a PDMP\nNext, we move to the results of fixed effects models \nfor the presence of any PDMP system, which \naccounts for the creation of a system of surveillance. \nFor ease of presentation, we show results only for \nPDMPs and use figures to interpret the results. The \nfull models, inclusive of all controls and fixed effects \n\nVuolo et al.\t\n345\nfor county and year, can be found in Appendix A. \nFigure 3 shows the relationship between the timing of \nPDMP adoption and the overall opioid overdose rate. \nThe vertical line in the center of each figure repre-\nsents the year of adoption, which is the baseline com-\nparison year. Each point estimate then represents the \ncoefficient for that year (equal to that in Appendix A) \nand provides the difference between the year of adop-\ntion and a given number of years in either direction, \ncontrolling for other covariates. There are no signifi-\ncant differences in the overall opioid overdose rate in \nany year prior to adoption compared to the year of \nadoption. Statistically, this finding means there are no \n\u201clead\u201d effects, which would indicate that change pre-\nceded policy implementation. Such significant leads \ncould imply state-level policy was passed in response \nto overdose rates in a county and that the relationship \nis temporally reversed. This pattern also assuages \nconcerns that the period prior to adoption was differ-\nent across counties prior to adoption (i.e., the parallel \ntrends assumption). The years after adoption then \nprovide differences occurring due to the presence of a \nTable 1.\u2002 Descriptive Statistics (Pooled over Counties and Year 2000\u20132016, N = 53,277).\nAverage (SD)/%\nCounty-level age-adjusted overdose rates\n\u2002\n\u2003 Overall\n5.54 (14.82)\n\u2003 Prescription opioid\n2.79 (13.11)\n\u2003 Prescription opioid with another opioid\n.63 (11.87)\n\u2003 Benzodiazepines\n1.58 (4.93)\n\u2003 Benzodiazepines with an opioid\n1.28 (4.31)\n\u2003 Psychostimulants\n.66 (2.68)\n\u2003 Psychostimulants with an opioid\n.23 (1.36)\n\u2003 Heroin\n.71 (2.57)\nState-level policies\n\u2002\n\u2003 PDMP system\n66.09%\n\u2003 PDMP with prescriber access\n58.51%\n\u2003 PDMP with dispenser access\n56.04%\n\u2003 PDMP prescriber requirement\n3.89%\n\u2003 PDMP dispenser requirement\n1.80%\n\u2003 Good Samaritan law\n9.00%\n\u2003 Naloxone access law\n13.55%\n\u2003 Medical marijuana\n14.98%\n\u2003 Pain management clinic law\n10.50%\nState-level per capita expenditures ($)\n\u2002\n\u2003 Education\n1.73 (.43)\n\u2003 Public welfare\n1.28 (.44)\n\u2003 Hospitals\n.18 (.13)\n\u2003 Health\n.16 (.08)\nCounty-level demographics\n\u2002\n\u2003 Unemployment rate\n6.32 (2.74)\n\u2003 Population density (per mi2)\n213.21 (1235.20)\n\u2003 Median household income (thousands)\n42.13 (11.63)\n\u2003 % living in poverty\n11.43 (5.66)\n\u2003 % with bachelor\u2019s degree\n13.38 (6.58)\n\u2003 % foreign-born\n4.12 (5.35)\n\u2003 % female-headed households\n6.38 (2.45)\n\u2003 % Black\n9.26 (14.35)\n\u2003 % Hispanic\n7.90 (12.92)\nNote: Data from Centers for Disease Control and Prevention, Prescription Drug Abuse Policy System, American \nCommunity Survey, and decennial censuses. PDMP = prescription drug monitoring program.\n\n346\t\nJournal of Health and Social Behavior 63(3) \nPDMP. In the first year, there is no significant differ-\nence compared to the year of adoption. However, by \nYear 2, significant differences emerge, and this \nreduction is sustained through 4 years after adoption. \nSpecifically, county-level overdoses per 100,000 are \npredicted to be .88, 1.17, and 1.59 lower in years 2, 3, \nand 4, respectively, from adoption.\nWe would presume that PDMPs will be more \ndirectly effective for prescription opioids than illicit \nopioids, although we recognize there is potential for \na cascading effect as opioid-dependent individuals \ntransition to heroin use. Thus, in Figure 4, we dis-\nplay two versions of the prescription opioid over-\ndose rate. On the left, we show the model for all \nprescription opioid overdoses. Although the general \npattern is the same, the effect is statistically signifi-\ncant only 4 years after implementation. When look-\ning at overdoses that included another opioid, \nhowever, we see a sustained negative effect on the \noverdose rate starting 2 years after implementation. \nThus, PDMPs may have been particularly effective \nat curbing polydrug use involving prescription \nopioids.\nGiven that PDMPs have a general effect on pre-\nscribed controlled substances, there is no reason to \nconfine the analysis strictly to opioids. In Figure 5, \nwe provide evidence that PDMPs are also effective \nat lowering the overdose mortality rate for benzodi-\nazepines and psychostimulants. Similar to the over-\nall opioid rate, there is no immediate effect in the \nfirst year after implementation for benzodiazepines. \nHowever, the rate is significantly lower in Years 2, \n3, and 4 by .51, .82, and .87, respectively. For stim-\nulants, the results are significant in Year 1 (\u2013.26) \nand sustained through Year 4 (\u2013.45). Although not \nshown, if we limit our analysis to benzodiazepine \nor stimulant overdoses including an opioid, the pat-\ntern is virtually identical, again indicating a benefit \nfor polydrug-related mortality rates.\nFor fixed effects models, a common validity \nassessment to a significant finding is to consider an \noutcome where there should be no effect. As a \nrobustness check, we ran our models for cocaine, \nwhich has no prescription analogue outside a hospi-\ntal setting and should not be affected by PDMPs, \nand indeed found no significant effects of PDMPs \non cocaine overdose mortality (results not shown).\nPDMPs with Prescriber/Dispenser \nAccess or Requirements to Check\nTo clarify whether surveillance is the mechanism \nby which PDMPs operate, we also isolate PDMP \nsystems that either provide access to prescribers/\ndispensers or mandate prescribers/dispensers to \ncheck the system. Unlike our prior model specifica-\ntion that centered PDMP on year of implementa-\ntion, we include traditional lags and leads. Although \nthis results in data loss on both sides of the time \nseries, this approach is necessary because it is not \npossible to center on year of implementation \nbecause these forms of PDMPs are more recent and \nFigure 3.\u2002 PDMP Effect on County-Level Opioid \nOverdose Rate Using Fixed Effects Regression.\nNote: Data from Centers for Disease Control and \nPrevention, Prescription Drug Abuse Policy System, \nAmerican Community Survey (ACS), and decennial \ncensuses. Model includes all policy and contextual \ncontrols described in text and year and county \nfixed effects. Differences from the horizontal 0 line \nrepresent effect sizes, with confidence intervals that \ndo not overlap indicating statistical significance at p < \n.05. See Appendix A for table of full model. PDMP = \nprescription drug monitoring program; OD = overdose; \ndiff. = difference.\nFigure 2.\u2002 Average County-Level Overdose \nRates over Time.\nNote: Data from Centers from Disease Control and \nPrevention.\n\nVuolo et al.\t\n347\nless common, particularly requirements to check. \nWe include a 2-year lag because 2 years is where \nwe start to see effects in the models for any PDMP \nsystem. We include a single lead because any addi-\ntional leads would remove data from the most \nrecent years when PDMP checking requirements \nwere being implemented.\nTable 2 summarizes results for these four PDMP \nvariables by the same five outcomes previously \ndescribed and heroin. Simply put, we find no evi-\ndence that overdose mortality was associated with \npolicies that specifically grant prescriber/dispenser \naccess to the PDMP system. Although almost all the \nlag coefficients are in the hypothesized negative \nFigure 4.\u2002 PDMP Effect on County-Level Prescription Opioid Overdose Rate Using Fixed Effects \nRegression.\nNote: Data from Centers for Disease Control and Prevention, Prescription Drug Abuse Policy System, American \nCommunity Survey, and decennial censuses. Models include all policy and contextual controls described in text and \nyear and county fixed effects. Differences from the horizontal 0 line represent effect sizes, with confidence intervals \nthat do not overlap indicating statistical significance at p < .05. See Appendix A for table of full models. PDMP = \nprescription drug monitoring program; OD = overdose; diff. = difference.\nFigure 5.\u2002 PDMP Effect on County-Level Benzodiazepine and Stimulant Overdose Rate Using Fixed \nEffects Regression.\nNote: Data from Centers for Disease Control and Prevention, Prescription Drug Abuse Policy System, American \nCommunity Survey, and decennial censuses. Models include all policy and contextual controls described in text and \nyear and county fixed effects. Differences from the horizontal 0 line represent effect sizes, with confidence intervals \nthat do not overlap indicating statistical significance at p < .05. See Appendix A for table of full models. PDMP = \nprescription drug monitoring program; OD = overdose; diff. = difference; implem. = implementation.\n\n348\t\nJournal of Health and Social Behavior 63(3) \ndirection for access, the only evidence of a signifi-\ncant reduction is for the effect of prescriber access \non psychostimulants, but only in the first lag, and a \ncoterminous effect for heroin. Given the significant \nlead, there is even some evidence that fatal psycho-\nstimulant overdoses were already on the decline \nwhen states implemented dispenser access.\nFor policies with requirements to check, we find \nno additional benefits beyond general implementa-\ntion of PDMPs. There is little evidence that requir-\ning prescribers/dispensers to check the PDMP \nreduced overdose mortality. We find a significant \nlead for all fatal opioid overdoses, implying that \nthese requirements may have been implemented in \nresponse to high overdose rates. Considering when \nthese particular policies were passed (Figure 1), \nthey were implemented as prescription opioid over-\ndose mortality was stabilizing and heroin overdoses \nwere increasing. In fact, in models for heroin over-\ndoses, we find a significant lead (b = 1.42, p < .01, \nfor prescribers; b = .89, p < .001, for dispensers). \nThus, as all opioid overdose mortality increased, \nU.S. states may have tried to strengthen PDMP \n\u00adsystems via these requirements, but this overdose \nrise was due to substances (heroin, fentanyl) less \nlikely to be affected directly by such changes. We \nalso find a significant lead for benzodiazepines. \nSeparating benzodiazepine overdoses that did and \ndid not include an opioid (not shown), we find only \na significant lead when the overdose included an \nopioid such that it was unlikely that overdoses due \nsolely to benzodiazepines drove this particular \u00adpolicy \nadoption.\nDiscussion\nFor over a century, the state has demonstrated a \nvested interest in controlling drug use, dependence, \nand overdose. Until recently, criminal justice has \nbeen the main institution used to curb supply and \nwider patterns of use (MacCoun and Reuter 2001). \nThe overdose crisis and rates of drug dependence \nescalating during the twenty-first century funda-\nmentally shifted the focal institution for the \nTable 2.\u2002 Fixed Effects Panel Models of Overdose Rates by PDMP Type.\nAll Opioid \nOverdoses\nAll Prescription \nOpioid \nOverdoses\nPrescription Opioid \nOverdoses with \nAnother Opioid\nAll \nBenzodiazepines \nOverdoses\nAll \nPsychostimulant \nOverdoses\nAll Heroin \nOverdoses\nAccess granted\n\u2002\nPrescriber\n\u2002\n\u2003 Lead 1\n\u2013.622\n\u2013.592\n\u2013.077\n\u2013.522\n\u2013.113\n\u2013.058\n\u2003 Coterminous\n.225\n.124\n.056\n.016\n.129\n\u2013.112*\n\u2003 Lag 1\n.169\n.061\n\u20130.038\n\u2013.156\n\u2013.204**\n\u2013.135\n\u2003 Lag 2\n\u2013.473\n\u2013.220\n\u2013.122\n\u2013.279\n\u2013.020\n\u2013.099\nDispenser\n\u2002\n\u2003 Lead 1\n\u2013.675\n\u2013.602\n\u2013.063\n\u2013.538\n\u2013.160*\n\u2013.022\n\u2003 Coterminous\n.277\n.117\n.005\n\u2013.001\n.125\n\u2013.076\n\u2003 Lag 1\n.004\n\u2013.039\n\u2013.064\n\u2013.139\n\u2013.186*\n\u2013.167\n\u2003 Lag 2\n\u2013.346\n\u2013.191\n\u2013.079\n\u2013.362\n\u2013.029\n\u2013.060\nRequired to check\n\u2002\nPrescriber\n\u2002\n\u2003 Lead 1\n2.886**\n.912\n.177\n1.872*\n.024\n1.416**\n\u2003 Coterminous\n.112\n.698\n.187\n.061\n.119\n\u2013.066\n\u2003 Lag 1\n.993\n.618\n.235\n1.591**\n.241\n.387\n\u2003 Lag 2\n2.084\n.796\n.659\n.640\n1.091***\n.907\nDispenser\n\u2002\n\u2003 Lead 1\n3.310**\n2.109\n.260\n4.138**\n.186\n.885***\n\u2003 Coterminous\n\u2013.204\n.588\n.948*\n\u20131.670*\n.608\n.634\n\u2003 Lag 1\n.884\n\u2013.745\n\u2013.841\n1.386\n.130\n.778\n\u2003 Lag 2\n1.237\n.887\n.888\n\u2013.086\n.664\n\u2013.024\nNote: Data from Centers for Disease Control and Prevention, Prescription Drug Abuse Policy System, American \nCommunity Survey, and decennial censuses. Models include all controls, county and year fixed effects, and state-level \ncluster-corrected standard errors. PDMP = prescription drug monitoring program.\n*p < .05, **p < .01, ***p < .001.\n\nVuolo et al.\t\n349\ndeployment of state control (Foucault 1982) given \nthat these trends were fueled by a rise in prescription \ndrug misuse where criminal justice approaches \nbecame ineffective (Fischer et\u00a0al. 2010). The over-\ndose crisis also necessitated an expansion of the use \nof biopower (Foucault 1990) as population health \nbecame affected in adverse ways beyond the scope \nof existing state intervention (Case and Deaton \n2015; Dowell et\u00a0al. 2017; Ho 2017; Imtiaz et\u00a0al. \n2018). Thus, U.S. states responded to these trends \nby implementing policies that create and enhance \nhealth care systems of surveillance and, in doing so, \nintroduced a more tacit form of governmentality. \nWith continued concern regarding overdose mortal-\nity, assessments of policy efficacy remain vital for \nunderstanding the social dynamics of drug trends \nand identifying equitable policymaking tools for \nintervention. We contend that biopower in the form \nof PDMP implementation constrains the choices of \nprescribers, dispensers, and patients and ultimately \nhow people may misuse prescription drugs, as \ninferred from reduced overdose mortality.\nThe results of these analyses indicate that PDMP \nsystems have been effective in reducing overdose \nmortality. As hypothesized, net of other overdose-\nrelated policies, PDMPs reduce the overall rate of \nopioid overdose mortality within 2 years of imple-\nmentation through the creation of a system of sur-\nveillance and self-regulation that functions without \ndirect requirements for action. In this regard, these \npolicies highlight the role of unverifiable surveil-\nlance in structuring behavioral routines (Foucault \n1995). Beyond a general effect on opioid overdose \nmortality, PDMPs reduce overdose rates specific to \nprescription opioids, benzodiazepines, and psycho-\nstimulants and also account for polydrug use. \nImportantly, this analysis accounts for within-state \nheterogeneity by assessing county-level mortality \nand improves on prior limitations by accounting for \nmultiple policies across the nation over an extended \nperiod of time.\nThrough the combination of lead and lag struc-\ntures and fixed effects, the results also highlight that \nthe effects of PDMPs require time to reach efficacy, \nbut once they do, they have a sustained effect. Prior \nresearch suggested that policy implementation may \nrequire time to impact opioid prescribing (Bao et\u00a0al. \n2016). We contend that a system of surveillance and \nself-governance builds via cultural shifts in institu-\ntional practices, which may occur in a more incre-\nmental manner (Luhrmann 1991). Because policies \nfocused on shifting routines within clinical encoun-\nters through a mechanism of surveillance and self-\nscrutiny, these effects are not immediate. Considering \nFoucault\u2019s conception of the panopticon, it is per-\nhaps reasonable to expect that shifts in self-gover-\nnance require time to become habituated under \npossible but unverifiable surveillance, whereas the \nproverbial immediacy of a guard standing in front of \na prisoner effects more immediate change (Foucault \n1995). Because overdose mortality has grown over \ndecades, the problem likely will not be resolved \nquickly by a single policy, but these results indicate \nthat such policies may lessen adverse mortality \ntrends. Generally, the research presented here indi-\ncates that policy implementation expanding states\u2019 \nbiopower (Foucault 1990) can be a key means to \nreduce mortality at the population level even if these \neffects do not emerge immediately. Beyond the case \nof fatal overdose, our results provide general evi-\ndence that policies rooted in establishing a system of \nsurveillance within institutions can affect population-\nlevel health outcomes.\nThe role of perceived agency within a system \nof structured constraints is also highlighted \nthrough these results, carrying forward Foucault\u2019s \n(1991) conception of governmentality. The imple-\nmentation of surveillance itself is sufficient to pro-\nduce changes in these outcomes even if they do \nnot directly coerce behavior. When we separated \nPDMP systems that did or did not grant prescrib-\ners/dispensers access or require them to check the \nsystem, we did not find significant effects. Thus, \nfollowing Foucault (1990, 1995), we argue that \nthe effect we observe is likely due to the system of \nsurveillance and self-regulation that initial imple-\nmentation of PDMP systems created. The systems \nsuccessfully leveraged state biopower to influence \npopulation-level health by disciplining relevant \nactors to self-govern their behavior. As noted pre-\nviously, the knowledge that one is potentially \nbeing observed, but without certainty, and the pos-\nsibility of sanctions from the state may alter \nbehavior among multiple actors such as prescrib-\ners, pharmacists, and patients, thus potentially \nenhancing the effect of these policies. That is, \nPDMPs create a form of surveillance whereby \nmultiple actors\u2019 behavior is constrained (Bird and \nRieker 2008), which in the aggregate helps shift \ndrug-related mortality curves. This distinction is \nimportant because governmental power is more \nsustainable if actors perceive themselves as capa-\nble of choice rather than acting out of coercion \n(Foucault 1982). Such constraints may be particu-\nlarly effective for actors operating within the con-\ntext of institutions.\nAlthough we did not observe significant effects \nfor the requirement of prescribers/dispensers to \n\n350\t\nJournal of Health and Social Behavior 63(3) \ncheck, we are not advocating that these policies are \nnot worthwhile. As of yet, they are not having an \nadditional effect on overdose mortality. One possi-\nbility is that they have not been in effect long \nenough to observe effects, as was important in our \nmain analysis. However, we observed lead effects \nfor the overall overdose rate and heroin overdose \nrate, implying that such requirements may have \nbeen passed in places where overdose rates were \nhigh. If heroin (and possibly fentanyl) were the \nreason that U.S. states attempted to strengthen their \nPDMP laws with requirements to check, we would \nnot anticipate finding an effect on illicit opioids \nbecause these substances are less directly affected \nby PDMPs. Furthermore, we found that presence \nof a PDMP system had a significant effect across \nmultiple drug classes such that the systems were \nworking on their intended target of controlled pre-\nscription drugs. As such, we must consider that \nself-policing in the wake of surveillance had \nalready occurred by the time these requirements to \ncheck were in place, and thus additional mandates \nfor action did not further enhance the effect of \nthese policies. Mandates to check may nonetheless \nhelp retain the impact of these policies after the \nnews cycles have shifted away from concerns \nabout opioids and overdose.\nThe effects of PDMP systems on reduced benzo-\ndiazepine and stimulant overdose mortality highlight \nthe broad scope of such systems of institutional sur-\nveillance; the policy is efficacious beyond a single \nclass of controlled substances. The findings for these \ndrug classes further support our central hypothesis \nregarding the constraining power of surveillance sys-\ntems because the expectation is that this oversight \nwould affect behavior generally and not necessarily \nfor opioids alone. Furthermore, as noted, these find-\nings are similar if we consider the case of polydrug \nuse between each class of substances and opioids. \nSubstantively, these analyses highlight the policy\u2019s \nimpact on classes of substances that have received \nless national attention than opioids but that have also \nseen significant growth in dependence and overdose \nduring the past two decades (Rigg and Ford 2014; \nScholl et\u00a0al. 2019). Although the findings on stimu-\nlants are complicated by unclear distinctions in cause \nof death codes between prescription stimulants and \nillegal stimulants such as methamphetamine, the \neffect is promising. Even regarding methamphet-\namine, PDMPs may reduce the ability to obtain pre-\ncursors (e.g., pseudoephedrine) for methamphetamine \nproduction, which is captured by surveillance sys-\ntems in some U.S. states. Furthermore, within an \nAmerican political context in which calls for reduced \ngovernmental involvement are common, our results \naugment prior research indicating that the surveil-\nlance effect of PDMPs has a measurable impact on \npremature mortality and should be maintained even if \nthe United States reduces overdose mortality in future \nyears.\nAlthough this study has numerous strengths, \nwe must consider some limitations. First, schol-\nars have identified that mortality data quality for \ndrug overdose varies across U.S. states and may \naffect estimates of overdose mortality (Buchanich \net\u00a0 al. 2018). Nonetheless, we follow standard \nCDC protocols for definitions of overdoses by \nsubstance (CDC 2014). Second, counties are \nimperfect geographic units. Although the nar-\nrower geographic unit is an improvement over \nstate-level measures, the size and number of \ncounties vary across U.S. states. Third, with \ncounty-level assessments, we cannot include \nindividual-level characteristics as covariates. \nFourth, although we included a large battery of \ncounty- and state-level controls and county and \nyear fixed effects, we recognize that other time-\nvarying factors may affect the overdose rate. \nHowever, our methodology is robust to observ-\nable and unobservable static factors; only factors \nthat differentially change over the period of \nobservation can bias the results. Finally, although \nwe used a robust statistical methodology, we note \nthat we cannot make direct observational conclu-\nsions about individual actors\u2019 behaviors. We \nhighly encourage additional qualitative work \namong physicians, dispensers, and patients \nregarding the surveillance effect of PDMPs and \nother relevant policies (e.g., Chiarello 2015; \nRigg, March, and Inciardi 2010). Relatedly, \nalthough we identify a pathway for the applica-\ntion of the theory of constrained choices in the \naggregate (i.e., by inferring the impact of deci-\nsions made by prescribers, dispensers, and \npatients on mortality), we do not have data on \neach of these decisions at the individual level. \nWe note that this is not a specific test of individ-\nual decision-making as completed elsewhere \n(Vuolo et\u00a0al. 2016).\nConclusion\nPDMPs are effective policies to reduce fatal opioid \noverdoses and have contributed to reductions in mor-\ntality related to benzodiazepines and psychostimu-\nlants as well. Notably, as shown with other research \non prescribing (Pauly et\u00a0al. 2018), the effect is not \nimmediate because institutional and interpersonal \n\nVuolo et al.\t\n351\nshifts require time for the \u00adsurveillance-driven effect \nto take hold. Thus, \u00adefficacy of this policy requires \ntime as institutional and interpersonal shifts coalesce, \nbut they have a sustained effect. Additionally, our \nanalyses show that the system of state surveillance \ncreated by PDMPs was enough to affect overdose \nmortality rates across multiple drug classes without \nexplicitly mandating action by relevant actors. In \nterms of impact, if we consider the effect of PDMPs \nin place for 4 years and eliminate the reduction in opi-\noid overdose mortality in these states, we would \nanticipate about 54,814 more deaths, an increase of \n11.1%.3\nOur findings have relevance beyond PDMPs for \npublic health, wherein surveillance is the main form \nof state intervention. The area of occupational \nhealth and safety is a clear example of surveillance \nsystems having an effect on self-governing of both \nindividuals and institutional actors leading to posi-\ntive changes in outcomes. For example, firms \nincreased compliance with safety and health regula-\ntions following a novel OSHA practice of issuing \npress releases about facilities with violations \n(Johnson 2020). Improved safety and health \n\u00adconditions occurred once the threat of sanction due \nto \u00adsurveillance emerged\u2014achieving the same \nimprovement in compliance from the press releases \nthat would require over 200 additional on-site \ninspections. Importantly for Foucault (1995), these \nsystems of surveillance affect our everyday lives \nbeyond these institutions, whereby practitioners \nand individuals make health-related choices that \nalign with the state\u2019s interests due to the perception \nof potential state intervention. In this manner, \nchoices constrained by policy implementation have \na measurable impact on health.\nAlthough the promise of policy implementation \ngives reason for hope in curbing prescription drug \noverdoses, the problem of adulteration and syn-\nthetic opioids, such as fentanyl, raises concern \nbecause these substances may be less impacted by \npolicies such as PDMPs. Put simply, policy \napproaches aimed at controlling pathways to drug \nuse through the medical system are less likely to \nmeet the challenges of evolving drug markets. \nFoucault (1995) warns that state surveillance may \nshift to the criminal justice system where other \ninstitutions fail because criminal justice fills any \ngap in the perfection of self-governing behavior. \nThus, the rise of fentanyl adulteration could mark a \nresurgence in prior emphases on criminalization \napproaches. However, even for illicit drugs, dispro-\nportionate reliance on criminalization is not a fore-\ngone conclusion. Expanding health systems for \nearly intervention with people who use drugs and \npromoting access for drug-dependent persons to \nreceive needed treatment services while still creat-\ning surveillance are likely preferable to an expan-\nsion of the coercive power of the criminal justice \nsystem. We may only reduce morbidity and mortal-\nity from substance use through sustained and \n\u00adcomprehensive efforts of prevention, health promo-\ntion, clinical care, and treatment alongside policy \nimplementation.\n\n352\t\nJournal of Health and Social Behavior 63(3) \nAppendix A.\u2002 Fixed Effects Regressions of County-Level Overdose Rates.\nAll Opioid \nOverdoses\nAll Prescription \nOpioid Overdoses\nPrescription \nOpioid Overdoses \nwith Another \nOpioid\nAll \nBenzodiazepine \nOverdoses\nAll \nPsychostimulant \nOverdoses\nYear (vs. 2001)\n\u2002\n2002\n.761**\n.416*\n.232\n.083\n.070\n\u2002\n(.238)\n(.181)\n(.141)\n(.088)\n(.060)\n2003\n1.305***\n.822**\n.432\n.276\n.090\n\u2002\n(.351)\n(.288)\n(.244)\n(.216)\n(.079)\n2004\n1.841***\n1.156**\n.585\n.554\n.233*\n\u2002\n(.470)\n(.369)\n(.356)\n(.275)\n(.101)\n2005\n2.612***\n1.719**\n.779\n.775*\n.270*\n\u2002\n(.589)\n(.496)\n(.449)\n(.308)\n(.124)\n2006\n3.787***\n2.497***\n1.055\n1.290**\n.184\n\u2002\n(.758)\n(.618)\n(.564)\n(.369)\n(.143)\n2007\n4.776***\n3.123***\n1.247\n1.703***\n.216\n\u2002\n(.854)\n(.720)\n(.679)\n(.436)\n(.155)\n2008\n4.956***\n3.684***\n1.400\n2.055***\n.255\n\u2002\n(.981)\n(.846)\n(.796)\n(.538)\n(.171)\n2009\n5.600***\n4.193***\n1.725\n2.269***\n.390\n\u2002\n(1.129)\n(.996)\n(.951)\n(.612)\n(.207)\n2010\n5.485***\n4.572***\n1.727\n2.585**\n.480\n\u2002\n(1.225)\n(1.132)\n(.978)\n(.772)\n(.257)\n2011\n6.457***\n5.533***\n2.059\n3.038**\n.575\n\u2002\n(1.310)\n(1.269)\n(1.153)\n(.958)\n(.290)\n2012\n6.510***\n5.549***\n2.165\n2.998**\n.629\n\u2002\n(1.446)\n(1.366)\n(1.300)\n(.952)\n(.344)\n2013\n6.651***\n5.784***\n2.323\n3.220**\n1.035**\n\u2002\n(1.470)\n(1.421)\n(1.377)\n(.938)\n(.373)\n2014\n7.993***\n6.843**\n3.310\n3.266**\n1.208**\n\u2002\n(2.095)\n(2.067)\n(2.165)\n(.934)\n(.392)\n2015\n8.170***\n6.347***\n2.645\n3.669**\n1.712***\n\u2002\n(1.694)\n(1.573)\n(1.450)\n(1.060)\n(.422)\n2016\n9.215***\n6.792***\n3.040\n3.885***\n2.298***\n\u2002\n(1.778)\n(1.639)\n(1.620)\n(1.103)\n(.466)\nPDMP (vs. year 0)\n\u2002\n\u20135 or less\n1.275\n1.038\n.164\n1.021\n.081\n\u2002\n(.674)\n(.583)\n(.167)\n(.567)\n(.131)\n\u20134\n.395\n.387\n.054\n.548\n\u2013.057\n\u2002\n(.575)\n(.444)\n(.136)\n(.391)\n(.106)\n\u20133\n.701\n.532\n\u2013.008\n.453\n\u2013.085\n\u2002\n(.443)\n(.342)\n(.110)\n(.333)\n(.082)\n\u20132\n.082\n.140\n\u2013.015\n.144\n\u2013.061\n\u2002\n(.324)\n(.237)\n(.086)\n(.230)\n(.074)\n\u20131\n\u2013.329\n\u2013.062\n\u2013.044\n.136\n\u2013.128\n\u2002\n(.226)\n(.173)\n(.065)\n(.181)\n(.068)\n1\n.208\n.134\n\u2013.064\n\u2013.129\n\u2013.260***\n\u2002\n(.283)\n(.180)\n(.061)\n(.142)\n(.065)\n2\n\u2013.881*\n\u2013.344\n\u2013.340*\n\u2013.507*\n\u2013.279**\n\u2002\n(.359)\n(.246)\n(.169)\n(.252)\n(.092)\n3\n\u20131.173*\n\u2013.547\n\u2013.302*\n\u2013.820*\n\u2013.330**\n\u2002\n(.482)\n(.353)\n(.126)\n(.359)\n(.103)\n(continued)\n\nVuolo et al.\t\n353\nAll Opioid \nOverdoses\nAll Prescription \nOpioid Overdoses\nPrescription \nOpioid Overdoses \nwith Another \nOpioid\nAll \nBenzodiazepine \nOverdoses\nAll \nPsychostimulant \nOverdoses\n4\n\u20131.594**\n\u2013.799*\n\u2013.304**\n\u2013.869*\n\u2013.452**\n\u2002\n(.546)\n(.359)\n(.112)\n(.385)\n(.138)\n5 or more\n\u20131.175\n\u2013.710\n\u2013.393*\n\u2013.733\n\u2013.380*\n\u2002\n(.596)\n(.402)\n(.161)\n(.508)\n(.170)\nNaloxone access\n.429\n\u2013.167\n\u2013.228\n\u2013.104\n\u2013.177\n\u2002\n(.604)\n(.381)\n(.278)\n(.315)\n(.170)\nGood Samaritan law\n.155\n.446\n.395\n.286\n\u2013.079\n\u2002\n(.929)\n(.429)\n(.342)\n(.275)\n(.232)\nPain clinic law\n.237\n\u2013.247\n.078\n.132\n\u2013.038\n\u2002\n(1.021)\n(.579)\n(.242)\n(.495)\n(.158)\nMedical marijuana\n2.208*\n\u2013.129\n.120\n\u2013.319\n\u2013.217\n\u2002\n(.928)\n(.427)\n(.230)\n(.352)\n(.236)\nUnemployment rate\n.099\n.088\n\u2013.006\n.100\n.001\n\u2002\n(.069)\n(.048)\n(.019)\n(.077)\n(.015)\nPopulation density\n\u2013.000\n.000\n.001\n.000\n\u2013.000\n\u2002\n(.002)\n(.001)\n(.001)\n(.000)\n(.000)\nMedian household \nincome\n\u2013.049\n\u2013.048\n\u2013.012\n\u2013.053\n\u2013.021\n\u2002\n(.041)\n(.027)\n(.011)\n(.029)\n(.012)\nPoverty rate\n.044\n\u2013.067\n\u2013.085\n\u2013.014\n.022\n\u2002\n(.092)\n(.082)\n(.087)\n(.043)\n(.022)\n% bachelor\u2019s degree\n\u2013.048\n\u2013.050*\n\u2013.026\n\u2013.028\n\u2013.000\n\u2002\n(.029)\n(.025)\n(.024)\n(.021)\n(.010)\n% foreign born\n.079\n\u2013.014\n\u2013.108\n.077\n\u2013.037\n\u2002\n(.193)\n(.148)\n(.132)\n(.060)\n(.033)\n% Hispanic\n\u2013.602\n\u2013.524\n\u2013.364\n\u2013.193*\n.017\n\u2002\n(.301)\n(.277)\n(.259)\n(.085)\n(.021)\n% Hispanic *\n.002\n.003\n.005\n\u2013.001\n\u2013.001\n% foreign-born\n(.005)\n(.004)\n(.004)\n(.001)\n(.001)\n% female-headed \nhouseholds\n.748\n.680\n.612\n.070\n.009\n(.487)\n(.508)\n(.545)\n(.043)\n(.015)\n% Black\n\u2013.009\n\u2013.105*\n\u2013.070\n\u2013.039\n\u2013.020\n\u2002\n(.066)\n(.052)\n(.058)\n(.034)\n(.022)\nState education \nexpenditures\n\u20131.845\n\u2013.901\n\u2013.742\n.260\n\u2013.462\n(1.574)\n(1.267)\n(.851)\n(.852)\n(.318)\nState public welfare \nexpenditures\n1.570\n\u2013.085\n\u2013.065\n.347\n.860*\n(1.126)\n(.643)\n(.368)\n(.633)\n(.332)\nState hospital \nexpenditures\n2.004\n.499\n.831\n.373\n.769\n(2.564)\n(2.246)\n(.899)\n(2.174)\n(.712)\nState health \nexpenditures\n9.873***\n4.661*\n1.696*\n1.035\n1.250\n(2.682)\n(1.934)\n(.687)\n(1.672)\n(1.025)\nConstant\n1.738\n3.787*\n1.653\n2.219\n.690\n\u2002\n(2.494)\n(1.853)\n(1.311)\n(1.267)\n(.586)\nObservations\n48,239\n48,239\n48,239\n48,239\n48,239\nNote: Data from Centers for Disease Control and Prevention, Prescription Drug Abuse Policy System, American \nCommunity Survey, and decennial censuses. Models include state-level cluster corrected standard errors. PDMP = \nPrescription drug monitoring program.\n*p < .05, **p < .01, ***p < .001.\nAppendix A.. (continued)\n\n354\t\nJournal of Health and Social Behavior 63(3) \nAcknowledgments\nWe thank Rich Carpiano, Dave Melamed, Trent Mize, \nSarah Mustillo, Eric Schoon, and Dan Winchester for \nvaluable feedback on this article.\nFunding\nThe authors disclosed receipt of the following financial \nsupport for the research, authorship, and/or publication of \nthis article: This work was funded by the National Institute \non Drug Abuse (Grant No. R21DA046447). The funding \nagency had no role in the research; the views expressed in \nthis article do not represent those of the funding agency.\nORCID iDs\nMike Vuolo \n https://orcid.org/0000-0002-2916-0046\nBrian C. Kelly \n https://orcid.org/0000-0003-2039-1347\nNotes\n1.\t\nGiven that the extremes of 5 years in each direction \ncontain data for all years prior or after that point, \nour figures display through 4 years in each direc-\ntion. Alternative specifications of models with dif-\nferent top/bottom coding (e.g., \u00b17 years) revealed \nno substantive differences in the results. We also \nnote that we found no differences by whether a \nstate was an early or late adopter, with late adoption \ndefined as 2007 or later.\n2.\t\nAlthough reverse causality is possible within the \nmodels, as would be detected through significant \nleads, reciprocation is not possible because pre-\nscription drug monitoring programs (PDMPs) \nremain in effect for the remainder of observations \nonce implemented. That is, the overdose rate cannot \naffect PDMP adoption once in effect.\n3.\t\nThis amount was calculated by taking all state-\nyears at the point where a PDMP was in effect for \n4 years, increasing their crude opioid overdose \nrate by the coefficient at the 4-year mark (shown \nin Figure 3), and taking the difference between the \nobserved number of deaths and that expected from \nthis increased overdose rate. This estimate is con-\nservative because it ignores the reduction in Years 2 \nand 3 and is only for opioids, thus not including any \nadditional lives lost to benzodiazepines or psycho-\nstimulants when not in the presence of an opioid.\nReferences\nAllison, Paul D. 2009. Fixed Effects Regression Models. \nLos Angeles, CA: SAGE Publications.\nAtluri, Sairam, Gururau Sudarshan, and Laxmaiah \nManchikanti. 2014. \u201cAssessment of the Trends in \nMedical Use and Misuse of Opioid Analgesics from \n2004 to 2011.\u201d Pain Physician 17:E119\u201328.\nBachhuber, Marcus A., Sean Hennessy, Chinazo O. \nCunningham, and Joanna L. Starrels. 2016. \u201cIncreasing \nBenzodiazepine Prescriptions and Overdose Mortality \nin the United States, 1996\u20132013.\u201d American Journal \nof Public Health 106(4):686\u201388.\nBao, Yuhua, Yijun Pan, Aryn Taylor, Sharmini \nRadakrishnan, Feijun Luo, Harold Alan Pincus, \nand Bruce R. Schackman. 2016. \u201cPrescription Drug \nMonitoring Programs Are Associated with Sustained \nReductions in Opioid Prescribing by Physicians.\u201d \nHealth Affairs 35(6):1045\u201351.\nBird, Chloe E., and Patricia P. Rieker. 2008. Gender and \nHealth: The Effects of Constrained Choices and Social \nPolicies. New York: Cambridge University Press.\nBirnbaum, Howard G., Alan G. White, Matt Schiller, \nTracy Waldman, Jody M. Cleveland, and Carl L. \nRoland. 2011. \u201cSocietal Costs of Prescription Opioid \nAbuse, Dependence, and Misuse in the United \nStates.\u201d Pain Medicine 12(4):657\u201367.\nBuchanich, Jeanine M., Lauren C. Balmert, Karl E. \nWilliams, and Donald S. Burke. 2018. \u201cThe Effect \nof Incomplete Death Certificates on Estimates of \nUnintentional Opioid-Related Overdose Deaths \nin the United States, 1999\u20132015.\u201d Public Health \nReports 133(4):423\u201331.\nCarlson, Robert G., Ramzi W. Nahhas, Silvia S. Martins, \nand Raminta Daniulaityte. 2016. \u201cPredictors of \nTransition to Heroin Use among Initially Non-\nopioid Dependent Illicit Pharmaceutical Opioid \nUsers: A Natural History Study.\u201d Drug and Alcohol \nDependence 160:127\u201334.\nCase, Anne, and Angus Deaton. 2015. \u201cRising Morbidity \nand Mortality in Midlife among White Non-Hispanic \nAmericans in the 21st Century.\u201d Proceedings of the \nNational Academy of Sciences 112(49):15078\u201383.\nCenters for Disease Control and Prevention. 2014. \nSpecific Drugs Involved in Drug Poisoning Deaths, \n2008\u20132013. Atlanta, GA: Centers for Disease \nControl and Prevention.\nCenters for Disease Control and Prevention. 2017. What \nStates Need to Know about PDMPs. Atlanta, FA: \nCenters for Disease Control and Prevention. https://\nwww.cdc.gov/drugoverdose/pdmp/states.html.\nCenters for Disease Control and Prevention. 2020. The \nThree Waves of Opioid Overdose Deaths. Atlanta, \nGA: Centers for Disease Control and Prevention.\nChiarello, Elizabeth. 2015. \u201cThe War on Drugs Comes \nto the Pharmacy Counter: Frontline Work in the \nShadow of Discrepant Institutional Logics.\u201d Law & \nSocial Inquiry 40(1):86\u2013122.\nCiccarone, Daniel. 2021. \u201cThe Rise of Illicit Fentanyls, \nStimulants and the Fourth Wave of the Opioid \nOverdose Crisis.\u201d Current Opinion in Psychiatry \n34(4):344\u201350.\nDart, Richard C., Hilary L. Surratt, Theodore J. Cicero, \nMark W. Parrino, S. Geoff Severtson, Becki Bucher-\nBartelson, and Jody L. Green. 2015. \u201cTrends in \nOpioid Analgesic Abuse and Mortality in the \nUnited States.\u201d New England Journal of Medicine \n372(3):241\u201348.\nDavis, Corey S., and Derek Carr. 2017. \u201cState Legal \nInnovations to Encourage Naloxone Dispensing.\u201d \n\nVuolo et al.\t\n355\nJournal of the American Pharmacists Association \n57(2 Supplement):S180\u201384.\nDowell, Deborah, Elizabeth Arias, Kenneth Kochanek, \nRobert Anderson, Gery P. Guy, Jr., Jan L. Losby, \nand Grant Baldwin. 2017. \u201cContribution of \nOpioid-Involved Poisoning to the Change in Life \nExpectancy in the United States, 2000\u20132015.\u201d JAMA \n318(11):1065\u201367.\nDubbin, Leslie A., Jamie Suki Chang, and Janet K. Shim. \n2013. \u201cCultural Health Capital and the Interactional \nDynamics of Patient-Centered Care.\u201d Social Science \n& Medicine 93:113\u201320.\nFairchild, Amy L. 2003. \u201cDealing with Humpty Dumpty: \nResearch, Practice, and the Ethics of Public Health \nSurveillance.\u201d The Journal of Law, Medicine & \nEthics 31(4):615\u201323.\nFink, David S., Julia P. Schleimer, Aaron Sarvet, Kiran \nK. Grover, Chris Delcher, Alvaro Castillo-Carniglia, \nJune H. Kim, Ariadne E. Rivera-Aguirre, Stephen \nG. Henry, Silvia S. Martins, and Magdalena Cerd\u00e1. \n2018. \u201cAssociation between Prescription Drug \nMonitoring Programs and Nonfatal and Fatal Drug \nOverdoses: A Systematic Review.\u201d Annals of \nInternal Medicine 168(11):783\u201390.\nFinley, Erin P., Ashley Garcia, Kristen Rosen, Don \nMcGeary, Mary Jo Pugh, and Jennifer Sharpe Potter. \n2017. \u201cEvaluating the Impact of Prescription Drug \nMonitoring Program Implementation: A Scoping \nReview.\u201d BMC Health Services Research 17(1):420.\nFischer, Benedikt, Meagan Bibby, and Martin Bouchard. \n2010. \n\u201cNon-medical \nUse \nand \nDiversion \nof \nPsychotropic Prescription Drugs in North America: A \nReview of Sourcing Routes and Control Measures.\u201d \nAddiction 105(12):2062\u201370.\nFoucault, Michel. 1982. \u201cThe Subject and Power.\u201d Pp. \n208\u201328 in Michel Foucault: Beyond Structuralism \nand Hermeneutics. 2nd ed., edited by H. L. Dreyfus \nand P. Rabinow. Chicago, IL: University of Chicago \nPress.\nFoucault, Michel. 1990. The History of Sexuality: An \nIntroduction, Volume I. New York, NY: Vintage \nBooks.\nFoucault, Michel. 1991. \u201cGovernmentality.\u201d Pp. 87\u2013104 \nin The Foucault Effect: Studies in Governmentality, \nedited by G. Burchell, C. Gordon, and P. Miller. \nChicago, IL: University of Chicago Press.\nFoucault, Michel. 1995. Discipline & Punish: The Birth \nof the Prison. New York, NY: Vintage Books.\nGarland, David. 1997. \u201c\u2018Governmentality\u2019 and the Problem \nof Crime: Foucault, Criminology, Sociology.\u201d \nTheoretical Criminology 1(2):173\u2013214.\nGoffman, Alice. 2009. \u201cOn the Run: Wanted Men in a \nPhiladelphia Ghetto.\u201d American Sociological Review \n74(3):339\u201357.\nGriffiths, P. E., and C. West. 2015. \u201cA Balanced \nIntervention Ladder: Promoting Autonomy through \nPublic Health Action.\u201d Public Health 129(8):1092\u201398.\nGugelmann, Hallam M., and Jeanmarie Perrone. 2011. \n\u201cCan Prescription Drug Monitoring Programs Help \nLimit Opioid Abuse?\u201d JAMA 306(20):2258\u201359.\nHaegerich, Tamara M., Leonard J. Paulozzi, Brian J. \nManns, and Christopher M. Jones. 2014. \u201cWhat \nWe Know, and Don\u2019t Know, about the Impact of \nState Policy and Systems-Level Interventions on \nPrescription Drug Overdose.\u201d Drug and Alcohol \nDependence 145:34\u201347.\nHalaby, Charles N. 2004. \u201cPanel Models in Sociological \nResearch: Theory into Practice.\u201d Annual Review of \nSociology 30:507\u201344.\nHo, Jessica Y. 2017. \u201cThe Contribution of Drug Overdose \nto Educational Gradients in Life Expectancy in the \nUnited States, 1992\u20132011.\u201d Demography 54(3): \n1175\u2013202.\nHulme, Shann, David Bright, and Suzanne Nielsen. \n2018. \u201cThe Source and Diversion of Pharmaceutical \nDrugs for Non-medical Use: A Systematic Review \nand Meta-analysis.\u201d Drug and Alcohol Dependence \n186:242\u201356.\nImtiaz, Sameer, Charlotte Probst, and J\u00fcrgen Rehm. 2018. \n\u201cSubstance Use and Population Life Expectancy in \nthe USA: Interactions with Health Inequalities and \nImplications for Policy.\u201d Drug and Alcohol Review \n37(S1):S263\u201367.\nJohnson, Matthew S. 2020. \u201cRegulation by Shaming: \nDeterrence Effects of Publicizing Violations of \nWorkplace Safety and Health Laws.\u201d American \nEconomic Review 110(6):1866\u2013904.\nKeyes, Katherine M., Magdalena Cerd\u00e1, Joanne E. \nBrady, Jennifer R. Havens, and Sandro Galea. 2014. \n\u201cUnderstanding the Rural\u2013Urban Differences in \nNonmedical Prescription Opioid Use and Abuse \nin the United States.\u201d American Journal of Public \nHealth 104(2):e52\u201359.\nLight, Donald, and Sol Levine. 1988. \u201cThe Changing \nCharacter of the Medical Profession: A Theoretical \nOverview.\u201d The Milbank Quarterly 66(Suppl. 2): \n10\u201332.\nLuhrmann, Tanya M. 1991. Persuasions of the Witch\u2019s \nCraft: Ritual Magic in Contemporary England. \nCambridge, MA: Harvard University Press.\nMacCoun, Robert J., and Peter Reuter. 2001. Drug War \nHeresies: Learning from Other Vices, Times, and \nPlaces. New York, NY: Cambridge University Press.\nMaxwell, Jane C. 2011. \u201cThe Prescription Drug Epidemic \nin the United States: A Perfect Storm.\u201d Drug and \nAlcohol Review 30(3):264\u201370.\nMcCabe, Sean E., James A. Cranford, and Brady T. \nWest. 2008. \u201cTrends in Prescription Drug Abuse and \nDependence, Co-occurrence with Other Substance \nUse Disorders, and Treatment Utilization: Results \nfrom Two National Surveys.\u201d Addictive Behaviors \n33(10):1297\u2013305.\nMonnat, Shannon M. 2018. \u201cFactors Associated with \nCounty-Level Differences in U.S. Drug-Related \nMortality Rates.\u201d American Journal of Preventive \nMedicine 54(5):611\u201319.\nMorgan, Stephen L., and Christopher Winship. 2007. \nCounterfactuals and Causal Inference: Methods \nand Principles for Social Research. New York, NY: \nCambridge University Press.\n\n356\t\nJournal of Health and Social Behavior 63(3) \nMoride, \nYola, \nDanae \nLemieux-Uresandi, \nGenaro \nCastillon, Cristiano Soares de Moura, Louise Pilote, \nMareva Faure, and Sasha Bernartsky. 2019. \u201cA \nSystematic Review of Interventions and Programs \nTargeting Appropriate Prescribing of Opioids.\u201d Pain \nPhysician 22(3):229\u201340.\nMusto, David F. 1999. American Disease: Origins of \nNarcotic Control. New York, NY: Oxford University \nPress.\nPauly, Nathan James, Svetla Slavova, Chris Delcher, \nPatricia Freeman, and Jeffery Talbert. 2018. \n\u201cFeatures of Prescription Drug Monitoring Programs \nAssociated with Reduced Rates of Prescription \nOpioid-Related Poisonings.\u201d Drug and Alcohol \nDependence 184:26\u201332.\nPerrone, Jeanmarie, and Lewis S. Nelson. 2012. \u201cMedication \nReconciliation for Controlled Substances\u2014An \u2018Ideal\u2019  \nPrescription-Drug Monitoring Program.\u201d New \nEngland Journal of Medicine 366(25):2341\u201343.\nPeters, David J., Shannon M. Monnat, Andrew L. \nHochstetler, and Mark T. Berg. 2020. \u201cThe \nOpioid Hydra: Understanding Overdose Mortality \nEpidemics and Syndemics across the Rural\u2013Urban \nContinuum.\u201d Rural Sociology 85(3):589\u2013622.\nRhodes, Emily, Maria Wilson, Alysia Robinson, \nJill A. Hayden, and Mark Asbridge. 2019. \u201cThe \nEffectiveness of Prescription Drug Monitoring \nPrograms at Reducing Opioid-Related Harms and \nConsequences: A Systematic Review.\u201d BMC Health \nServices Research 19(1):784.\nRigg, Khary K., and Jason A. Ford. 2014. \u201cThe Misuse of \nBenzodiazepines among Adolescents: Psychosocial \nRisk Factors in a National Sample.\u201d Drug and Alcohol \nDependence 137:137\u201342.\nRigg, Khary K., Samantha J. March, and James A. Inciardi. \n2010. \u201cPrescription Drug Abuse & Diversion: Role \nof the Pain Clinic.\u201d Journal of Drug Issues 40(3): \n681\u2013701.\nRutkow, Lainie, Lydia Turner, Eleanor Lucas, Catherine \nHwang, and G. Caleb Alexander. 2015. \u201cMost \nPrimary Care Physicians Are Aware of Prescription \nDrug Monitoring Programs, but Many Find the Data \nDifficult to Access.\u201d Health Affairs 34(3):484\u201392.\nScholl, Lawrence, Puja Seth, Mbabazi Kariisa, Nana \nWilson, and Grant Baldwin. 2019. \u201cDrug and \nOpioid-Involved Overdose Deaths\u2014United States, \n2013\u20132017.\u201d Morbidity and Mortality Weekly Report \n67(5152):1419\u201327.\nVaisey, Stephen, and Andrew Miles. 2017. \u201cWhat You \nCan\u2014and Can\u2019t\u2014Do with Three-Wave Panel Data.\u201d \nSociological Methods & Research 46(1):44\u201367.\nVuolo, Mike, Joy Kadowaki, and Brian C. Kelly. 2016. \u201cA \nMultilevel Test of Constrained Choices Theory: The \nCase of Tobacco Clean Air Restrictions.\u201d Journal of \nHealth and Social Behavior 57(3):351\u201372.\nWatkins, William C. 2016. \u201cA Social Learning Approach \nto Prescription Drug Misuse among College Students.\u201d \nDeviant Behavior 37(6):601\u201314.\nWinship, Christopher, and Stephen L. Morgan. 1999. \u201cThe \nEstimation of Causal Effects from Observational \nData.\u201d Annual Review of Sociology 25:659\u2013706.\nAuthor Biographies\nMike Vuolo is an associate professor in the Department of \nSociology at The Ohio State University and current editor \nof Sociological Methodology. His research interests \ninclude crime, law, and deviance; health; work and educa-\ntion; substance use; the life course; and statistics and \nmethodology.\nLaura C. Frizzell is a PhD candidate in the Department of \nSociology at The Ohio State University. Her research centers \non the experiences of racial-ethnic, gender, and sexual minor-\nities within the criminal legal system, from involvement in \ncrime and substance use to reentry after incarceration.\nBrian C. Kelly is a professor in the Department of \nSociology at Purdue University. His research focuses on \ncontextual influences of health, mainly substance use, \nsexual health, and mental health.\n\n\n    Scientific Research Findings:", "answer": "\u2022 PDMPs reduced overall opioid overdose mortality starting two years after implementation with sustained effects, with similar effects for prescription opioids, benzodiazepines, and psychostimulants.\n\u2022 Neither prescriber/dispenser access nor mandates to utilize the PDMP system reduce mortality beyond the effect of surveillance.\n.\u2022 Overall, PDMP effects on overdose mortality are likely due to self-regulation occurring within the context of surveillance rather than actions mandated by policy adjustments.", "id": 72}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465221074915\nJournal of Health and Social Behavior\n2022, Vol. 63(4) 472\u00ad\u2013490\n\u00a9 American Sociological Association 2022\nDOI: 10.1177/00221465221074915\njournals.sagepub.com/home/hsb\nOriginal Article\nThe COVID-19 pandemic has been one of the most \ndisruptive world events in modern times in terms of \nhealth and fundamentally reshaping the economy \nand people\u2019s social lives. Moreover, this impact has \nnot been equally shared. Some people, because of \nexisting social and economic resources, have been \nbetter able to shield themselves from transmission of \nthe disease and the various negative consequences \nreverberating from the pandemic. In \u00adparticular, \nBlack, Latino, and Native American \u00adpopulations in \nthe United States have experienced higher rates of \ninfection and, when infected, have been more likely \nto experience severe disease and death (Millett et\u00a0al. \n2020; Peek et\u00a0al. 2021; Ramos and Zamudio 2020). \nOther work has shown how the effects of the pan-\ndemic are also spaced and placed, with poor and \nminority neighborhoods suffering disproportion-\nately from the disease and having limited access to \ncare sites and testing locations (McMinn et\u00a0al. 2020; \nYang, Choi, and Sun 2021). In this analysis, we \nfocus on racial-ethnic disparities in access to the vac-\ncine in the United States specifically as it relates to \nracial-ethnic residential segregation.\nOn December 11, 2020, the U.S. Food and Drug \nAdministration issued the first emergency use \nauthorization for the Pfizer-BioNTech vaccine for \nCOVID-19 in individuals 16 years of age and older \nin the United States. Eight days later, a second vac-\ncine, Moderna, was approved for use in individuals \nages 18 and older. However, distribution of the vac-\ncine has been a logistical hurdle necessitating \nextremely low temperature cold storage and person-\nnel needs, all while having to maintain COVID-19 \n1074915 HSBXXX10.1177/00221465221074915Journal of Health and Social BehaviorAnderson and Ray-Warren\nresearch-article2022\n1University of Houston, Houston, TX, USA\nCorresponding Author:\nKathryn Freeman Anderson, Department of Sociology, \nUniversity of Houston, 3551 Cullen Blvd, PGH Building, \nRoom 450, Houston, TX 77204-3012, USA. \nEmail: kateanderson@uh.edu\nRacial-Ethnic Residential \nClustering and Early COVID-19 \nVaccine Allocations in Five \nUrban Texas Counties\nKathryn Freeman Anderson1\n and Darra Ray-Warren1\nAbstract\nPrevious research has indicated that racial-ethnic minority communities lack a wide variety of health-\nrelated organizations. We examine how this relates to the early COVID-19 vaccine rollout. In a series of \nspatial error and linear growth models, we analyze how racial-ethnic residential segregation is associated \nwith the distribution of vaccine sites and vaccine doses across ZIP codes in the five largest urban counties \nin Texas. We find that Black and Latino clustered ZIP codes are less likely to have vaccine distribution sites \nand that this disparity is partially explained by the lack of hospitals and physicians\u2019 offices in these areas. \nMoreover, Black clustering is also negatively related to the number of allocated vaccine doses, and again, \nthis is largely explained by the unequal distribution of health care resources. These results suggest that \nextant disparities in service provision are key to understanding racial-ethnic inequality in an acute crisis \nlike the COVID-19 pandemic.\nKeywords\nCOVID-19, health care, race-ethnicity, residential segregation, vaccines\n\nAnderson and Ray-Warren\t\n473\nsafety protocols. For example, unpunctured vials of \nthe vaccines must be kept between \u201390\u00b0C and \n\u201360\u00b0C for the Pfizer-BioNTech vaccine and between \n\u201350\u00b0C and \u201315\u00b0C for the Moderna vaccine, well \nbelow the temperature of a standard freezer. An \nobvious question that has emerged is how to do this \nequitably, especially when vaccine supply is low. \nRoughly following guidelines from the Centers for \nDisease Control and Prevention (CDC; 2020a), in \nthe case of Texas, which is focus of this analysis, \nPhase 1A of the rollout was reserved for frontline \nhealth care workers, residents of long-term nursing \nfacilities for the elderly, and workers in those same \nfacilities. Phase 1B included people ages 65 and \nabove and people ages 16 and above with at least \none comorbidity defined by the state as being par-\nticularly vulnerable for the disease, such as cancer, \ntype 2 diabetes, obesity, chronic heart disease, and \nso on. Although the goal was to equitably distribute \nvaccine doses to reach the vulnerable, the question \nremains whether this goal was met.\nThis is the overarching question that we tackle \nwith this analysis. We focus only on the Phase 1A \nand 1B periods to examine who got early access \nwhile vaccine supply and eligibility were still lim-\nited. Obviously, many considerations are at play for \nhow any individual can get the vaccine, and early \nmedia reports indicate that access to reliable Internet \nand transportation play a major role in people\u2019s abil-\nity to find an appropriate appointment time and \ntravel to the location (Menchaca and Agnew 2021; \nPrescott and Prescott 2021; Stone 2021). Here, we \nfocus on the structural element of this story to exam-\nine where these vaccine allocations went and \nwhether certain types of neighborhoods were advan-\ntaged or disadvantaged in the rollout. We ask: What \nneighborhood characteristics are related to the dis-\ntribution of COVID-19 vaccines? Specifically, are \nracial-ethnic minority communities less likely to \nhave vaccine supply? Furthermore, how do these \npatterns relate to extant inequalities in health care \nprovision? We posit that racial-ethnic minority areas \nwill be less likely to receive vaccine doses and that \nthis will be explained away or mediated by existing \ndisparities in service provision across neighbor-\nhoods given the difficult infrastructure requirements \nof distributing the two vaccines.\nBackground\nSegregation and Access to Health Care\nResearch has long documented important health and \nhealth care outcomes related to racial-ethnic residential \nsegregation for Black and Latino communities in the \nUnited States. For instance, analysts have documented \na relationship between segregation and racial differ-\nences in mortality for several causes of death (Collins \n1999; Collins and Williams 1999; Hart et\u00a0al. 1998), \nincluding infant mortality (Grady 2006; McFarland and \nSmith 2011). In addition, other studies have examined \noverall physical health, mental health, and functional \ndisability and found that minority residents of racially \nsegregated neighborhoods are more likely to report \nexperiencing overall poorer physical and mental health \nand disability (Acevedo-Garcia 2000; Acevedo-Garcia \net\u00a0al. 2003; Anderson and Fullerton 2014; Lee 2009). \nFinally, a few studies on access to health care demon-\nstrate an association between racial-ethnic segregation \nand diminished access to health care coverage, having  \na personal physician, and health care utilization \n(Anderson and Fullerton 2014; Gaskin et\u00a0 al. 2009; \nRodriguez et\u00a0al. 2007).\nAlthough there is robust literature demonstrat-\ning the association between segregation and several \nhealth and health care outcomes, less attention has \nbeen given to understanding and testing the mecha-\nnisms that could link these two. A notable exception \nis Williams and Collins\u2019s (2001) piece in which \nthey describe racial-ethnic minority segregation as \na \u201cfundamental cause\u201d of health and detail how this \nform of segregation can be fundamentally linked to \nhealth outcomes. They suggest that racial-ethnic \nminority segregation, with its intimate ties to socio-\neconomic status and overall life chances, can be \nthought of as a fundamental cause of poor health/\nhealth care outcomes in these communities (Williams \nand Collins 2001). Williams and Collins (2001) \ndescribe several specific mechanisms by which \nracial-ethnic minority segregation may limit life \nopportunities that impact health/health care out-\ncomes. One such mechanism argues that racial-eth-\nnic minority residential segregation may be related \nto health outcomes because it can constrain access \nto a variety of key community resources, including \ndaily necessities such as food, recreation, and pub-\nlic services, and critical resources, such as health \ncare. This perspective focuses on the unequal distri-\nbution of resources throughout urban areas and how \nthat inequity may lead to diminished health out-\ncomes for these residents. This is the approach uti-\nlized in the study presented here.\nRecently, there has been an explosion of interest \nin this mechanism, and such work has shown how \ndifferent neighborhood characteristics, mainly dif-\nferentiated by race and class, relate to service pro-\nviders and community organizations. The bulk of \nthis research has focused on food resources, known \n\n474\t\nJournal of Health and Social Behavior 63(4)\nas the \u201cfood deserts\u201d or \u201cfood swamps\u201d literature \n(Beaulac, Kristjansson, and Cummins 2009; \nCooksey-Stowers, Schwartz, and Brownell 2017). \nThis literature has demonstrated that disproportion-\nately poor and minority neighborhoods lack food \nresources and have demonstrated important differ-\nences in quality and price across neighborhoods \n(Beaulac et\u00a0al. 2009; Cooksey-Stowers et\u00a0al. 2017; \nMoore and Diez Roux 2006; Walker, Keane, and \nBurke 2010). Although food is the focus of much of \nthis literature, other work has shown a similar pat-\ntern for physical fitness centers, park and green \nspace, retail establishments, nonprofit associations, \nand social services (Allard 2009; Anderson 2017; \nGordon-Larsen et\u00a0al. 2006; Marwell and Gullickson \n2013; Small and McDermott 2006).\nWhereas the work on food and retail has been \nwell studied, there are relatively few sociological \nstudies on the case of health care specifically, which \nis most relevant to our present study on the COVID-\n19 vaccine distribution. The studies that exist on \nhealth care have generally shown a similar pattern \nin that Black and Latino segregated areas are less \nlikely to have a wide variety of health care estab-\nlishments in terms of density or distance to facilities \n(Anderson 2017; Dai 2010; Dinwiddie et\u00a0al. 2013; \nGaskin et\u00a0al. 2012; Ko et\u00a0al. 2014; Rodriguez et\u00a0al. \n2007). Specifically, this work has demonstrated that \nracial-ethnic minority segregation and socioeco-\nnomic variables are related to a lower incidence of \nphysicians\u2019 offices, primary care providers, mental \nhealth practitioners, urgent care facilities, auxiliary \nhealth care practitioners, surgical centers, and dial-\nysis facilities (Anderson 2017; Dai 2010; Dinwiddie \net\u00a0 al. 2013; Gaskin et\u00a0 al. 2012; Ko et\u00a0 al. 2014; \nRodriguez et\u00a0al. 2007).\nSome limited work has also linked the distribu-\ntion of health care locations to health and health \ncare outcomes. For example, Dai (2010), in a study \nof Detroit area neighborhoods, found that Black \nresidents of segregated neighborhoods had fewer \nfacilities that provided mammography services and \nconsequently had higher rates of late-stage breast \ncancer diagnosis. Another study found that racial-\nethnic residential segregation was linked to a \ngreater likelihood of patients seeing nonpsychia-\ntrists for mental health needs (Dinwiddie et\u00a0 al. \n2013). A study of the Phoenix area found that the \nlack of health care service provision in Latino-\nsegregated neighborhoods was linked to a lower \nlikelihood of seeing a personal physician for pediat-\nric care and greater use of clinics (Anderson 2020). \nChan et\u00a0al. (2012) also found that although Black \nand Latinos living in highly segregated areas had \nadequate spatial access to facilities, they were less \nlikely to receive certain services, especially special-\nist care.\nThis work provides some initial evidence for a \ndisparity by racial-ethnic residential segregation in \nhealth care service provision. In this analysis, we \naim to extend this literature to examine how this \nlack of adequate health care provision relates to the \nCOVID-19 pandemic and the early rollout of the \nvaccine. There may be myriad reasons that affect \nwhere and how people choose to procure provider \nservices, with physical proximity/convenience \nbeing only one of them that may be important to \npeople of limited means, who lack of access to \ntransportation or are elderly/homebound and may \nfind travel difficult. Distance from providers does \nnot create an impenetrable barrier to access (in this \ncase, vaccine access), but it can increase the \u201cfric-\ntion of distance,\u201d which may be difficult to over-\ncome absent other kinds of individual or household \nresources (Tobler 1970). Therefore, with this analy-\nsis, we are not assuming that people can only or \nwill only seek vaccine access within their local \ncommunities but, rather, that having access in close \nproximity would facilitate their ability to get it, \nespecially for those who are more vulnerable.\nSegregation and the COVID-19 \nPandemic\nAt this stage, limited work has been conducted in a \nsystematic fashion on the COVID-19 pandemic \nand, more specifically, the vaccine rollout. However, \nthe extant work published over the past year, includ-\ning media reporting, has demonstrated vast inequal-\nities by race-ethnicity throughout the pandemic. \nThis work has shown that racial-ethnic minority \npopulations have been more vulnerable to COVID-\n19 disease in terms of rates of infection and mortal-\nity (Millett et\u00a0al. 2020; Ramos and Zamudio 2020). \nAs it relates to segregation specifically, new work \nhas shown that areas with higher numbers of racial-\nethnic minorities were more likely to have higher \nrates of infections and that this pattern was further \nexacerbated where residential segregation was also \nhigh (Yang et\u00a0al. 2021). Various news reports have \nalso indicated that minority-segregated areas lack \nvital resources to contend with the spread of the \nvirus, such as testing facilities and health care \nresources (Garnam and Cai 2021; Godoy and Wood \n2020; Martinez 2021; McMinn et\u00a0al. 2020).\nMoreover, what is happening now is clearly part \nof a broader pattern of inequality in infectious \n\u00addisease as it relates to racial-ethnic residential \n\nAnderson and Ray-Warren\t\n475\nsegregation. Previous work has shown important \ndisparities by racial-ethnic residential segregation  \nin terms of infectious disease, such as tuberculosis \nand sexually transmitted diseases, among others \n(Acevedo-Garcia 2000, 2001; Biello et\u00a0 al. 2012; \nStrully 2011). The limited literature that examines \npast pandemics and social inequality has focused \nprimarily on race-ethnicity and socioeconomic sta-\ntus (\u00d8kland and Mamelund 2019; Roberts and \nTehrani 2020; Strully 2011).\nFrom this limited early work on the topic at hand \nand the more extensive literature on broader inequal-\nities in health care by segregation, we expect these \ninequalities by segregation to carry over to the early \nCOVID-19 vaccine rollout. Throughout fall 2020, \nwhen the possibility of a vaccine was becoming a \nreality, public health and medical practice journals \npublished numerous editorials with recommenda-\ntions for how to effectively and equitably roll out  \nthe vaccine (Persad, Peek, and Emanuel 2020). \nHowever, states must work within the existing public \nhealth and health care infrastructure to provide the \nvaccine in a manner that fulfills more stringent \nrequirements than most vaccines. Given the existing \nunequal infrastructure, it would be difficult to imag-\nine that the vaccine distribution could be done equi-\ntably. The first weeks of vaccine distribution saw \nreporting from media outlets questioning the equity \nof allocation locations (Harper 2021; Oladipo 2021). \nThis remains an empirical question, though, because \nno systematic work has been conducted on this topic \nconcerning the COVID-19 pandemic.\nConceptual Framework and Hypotheses\nFrom the theory and extant work, we thus have sev-\neral hypotheses for this analysis. Given the docu-\nmented inequalities in community establishments, \nwe expect that the gap in service provision in segre-\ngated communities will extend to the case of the \nCOVID-19 vaccine rollout. We examine this in two \nways\u2014the number of vaccination sites and the allo-\ncation of doses of vaccine. We also examine this at \nthe neighborhood level by examining the geographic \nclustering of groups in space (details in the follow-\ning). Thus, we have the following first hypothesis:\nHypothesis 1: A higher degree of racial-ethnic \nminority clustering across urban areas will be \nrelated to fewer vaccination sites and fewer \nvaccine doses.\nHowever, we also expect that this will be related to \nthe extant lack of health care provision within these \nneighborhoods. Thus, we expect that the density of \nexisting health care organizations within urban \nneighborhoods will attenuate this association. This \nleads us to our second hypothesis:\nHypothesis 2: The density of health care service \nproviders across urban areas will attenuate \nthe association between the racial-ethnic \nminority clustering and vaccine provision.\nWe test these two hypotheses in a study of early vac-\ncine allocations over a 9-week period in the five \nlargest urban counties in Texas.\nData and Methods\nData\nTo examine the association between racial-ethnic \nminority clustering and the vaccine rollout, we com-\nbined several sources of area-level data measured at \nthe ZIP (Zone Improvement Plan) area unit of analy-\nsis (N = 431). First, for data on vaccine allocations, \nwe used data from the Texas Department of State \nHealth Services. Since the first shipment of the \nPfizer-BioNTech vaccine on December 14, 2020, \nthe state of Texas has released information on a \nweekly basis on the location and quantity of doses \nallocated across the state by county. We collected \nthese data each week for the first 10 weeks of the \nvaccine rollout (December 14, 2020, to February 15, \n2021). However, we decided to exclude the first \nweek from our analysis because these allocations \nwent only to major hospital sites and were exclu-\nsively distributed to first-tier health care workers \nunder Texas\u2019s Phase 1A of the vaccine distribution. \nWe chose these weeks for several reasons. We \nwanted to focus on the early vaccine rollout when \nsupplies were limited and access was difficult to \nexamine which areas got early access to this privi-\nlege. Second, we chose this specific set of weeks due \nto several circumstances in the weeks that followed. \nIn the subsequent week (Week 11 of the rollout), \nmost of the state of Texas experienced a catastrophic \nwinter storm that affected vaccine distributions both \nin terms of the ability to make shipments to vaccine \nsites and the ability to keep those doses properly \nchilled because of widespread power outages. \nFollowing this event, in March, the state began to \nopen up eligibility, and the federal government allo-\ncated more funding to the vaccine rollout that dra-\nmatically increased vaccine supplies across the state.\nTo analyze the spatial distribution of these sites, \nan address location was provided for each site, and \n\n476\t\nJournal of Health and Social Behavior 63(4)\nwe geocoded these to a point-level location in \nESRI\u2019s ArcMap. Furthermore, because we were \nprimarily interested in how these inequalities relate \nto urban dynamics, not differences between urban \nand rural locales, we focused on the five largest \ncounties in the state, which are also the core coun-\nties of the five largest cities in the state. These \nincluded Bexar County (San Antonio), Dallas \nCounty (Dallas), Harris County (Houston), Tarrant \nCounty (Fort Worth), and Travis County (Austin). \nWe used counties rather than cities or metropolitan \nareas because we paired the vaccine data with data \non infection rates by ZIP code, which were pro-\nvided by county health departments.\nWe recognize that other states have provided \nsimilar data about their vaccine allocations. \nHowever, we chose to focus on a single state \nbecause each state has implemented its own criteria \nfor how vaccines should be allocated and who \nshould be eligible to receive them. Texas, in partic-\nular, allocated vaccines to sites based on two pieces \nof information: an advisory panel, which made rec-\nommendations for allocations, and a registration \nsystem for vaccine providers to determine eligibil-\nity and feasibility. Texas appointed a team of sub-\nject matter experts onto an Expert Vaccine Advisory \nPanel to develop vaccine allocation strategies as \nrecommendations to the Texas Commissioner of \nHealth. Information from the CDC and an \nappointed Advisory Committee on Immunization \nPractices helped develop these strategies. Vaccine \nequity was one of the stated goals of this panel, but \nthe exact algorithm used by the panel is not publicly \navailable. On the supply side of this question, vac-\ncine provider registration data aided in determining \nthe physical locations and quantities for distribu-\ntion. According to the CDC provider agreement, \nregistration data included licensure information, \npatient population numbers, and other logistical \ndetails required to ensure each facility\u2019s ability to \nstore and administer the vaccines.\nAlthough we limited our analysis to this one \nstate, we argue that Texas serves as a good test case. \nPart of this rationale is practical\u2014Texas provided \ndetailed information on the location and quantities \nof vaccines in a publicly available format. It is also \nthe only state with that many major U.S. cities (5 in \nthe top 20 largest cities in the United States) and \nthat also has a high degree of racial-ethnic diversity \nacross all three of the largest racial-ethnic groups in \nthe United States. Although Texas cities do not \nhave the highest rates of residential segregation, \ncertain Texas cities are fairly segregated, and there \nis quite a bit of variation across cities to allow for \ncomparison. Full descriptive statistics for variables \nused in our analysis and the racial-ethnic break-\ndown for each county (divided by county) can be \nfound in Appendix A in the online version of the \njournal for reference.\nWe combined these data with sociodemographic \ndata and data on establishments from two census \nproducts. First, for sociodemographic data, includ-\ning our measures of racial-ethnic clustering, we \nused the 2014 to 2018 American Community \nSurvey 5-year estimates at the ZIP code tabulation \narea level. The Census Bureau only provides data at \nthis small geographic unit using 5-year aggregates \nbecause the data are not representative for small \nunits of analysis like the ZIP code for a single year. \nWe also combined this with data on establishments \nfrom the 2016 County Business Patterns (CBP) ZIP \nCode Industry Detail File. The CPB uses IRS tax \nrecords to provide counts of establishments by ZIP \ncode and by industry type using the North American \nIndustry Classification System (NAICS). We \nincluded several industry codes related to health \ncare, which we hypothesized may be related to \nhigher vaccine allocations (details in the follow-\ning). Although 2016 is not the most recent year of \ndata publicly available (2018), starting in 2017, the \nCensus Bureau stopped releasing counts of estab-\nlishments where the count for the ZIP code was less \nthan three to further deidentify the information. \nBecause we included several industry classifica-\ntions that are relatively rare, such as hospitals, we \nopted to use the older version of the data to get \n\u00adbetter estimates of the available resources in \nneighborhoods.1\nAcross all models, we also included a control \nvariable for the cumulative number of infections in \nthe ZIP code for all five of our counties. These \ncame from publicly available sources from each \ncounty\u2019s public health department. These were col-\nlected in each of the five counties on February 9, \n2021, which was the second to last week of our vac-\ncine rollout time frame such that it could have \naffected allocation decisions. We included this as a \ncovariate in the event that state health officials were \nusing the infection rate as the basis by which vac-\ncines were being allocated by factoring in local \nvulnerability.\nWe included two dependent variables, each \nmeant to capture a different facet of the vaccine \nrollout across urban areas in the state. First, we \nexamined the number of vaccine sites per 100,000 \npeople in a ZIP code. This came from the geocoded \nlocations for vaccine allocations and reflected a \nsimple count of the locations providing COVID-19 \n\nAnderson and Ray-Warren\t\n477\nvaccines. We also divided this number by the popu-\nlation size and multiplied by 100,000 to derive a \nrate per 100,000 people. Beyond the mere availabil-\nity of a vaccine site, we also examined the distribu-\ntion of vaccine doses per 10,000 people to each site \nfor each week of the 9-week period. The state pro-\nvided data on the week-by-week allocations to each \nof the geocoded locations. To derive a rate, we \ndivided this number by the population and multi-\nplied by 10,000.\nOur main independent variables in this analysis \nincluded a set of variables for racial-ethnic cluster-\ning. Although much of the literature on racial- \nethnic segregation is focused on measuring global \nsegregation across a large area, such as the county \nor metropolitan area (Massey and Denton 1988), in \nthis study, we were interested in coding for which \nareas within the county have disproportionately \nhigh numbers of certain groups. Typically, studies \nthat examine racial-ethnic concentration at a \nsmaller geographic unit of analysis use composition \nscores, which reflect simply the percentage of a \ngroup over a certain area.2 However, this approach \nis aspatial and ignores the role of geographic clus-\ntering across space and how adjacent areas may \ninfluence each other (Reardon and O\u2019Sullivan \n2004; Roberto 2018). Thus, for this analysis, we \nused a geographic clustering score that considered \ntwo pieces of information: the concentration of a \ngroup in an area (ZIP codes) and the extent to which \nthat group is geographically clustered. To be more \nprecise, we refer to this as clustering rather than \nsegregation throughout the discussion of this analy-\nsis, and we conceptualized this as a neighborhood-\nlevel measure of segregation. We used the following \nformula:\nC\nx\nw x\ni\ni\nj\nj i\nn\nij\nj\n=\n=\n\u2260\n\u2211\n1,\n,\nwhere xi is the variable for ZIP code i, x j is the vari-\nable for ZIP code j, and wij is the spatial weight \nbetween ZIP codes i and j (Anderson 2017). The \nmeasure is essentially the product of the percentage \nof a certain group in a ZIP code and the spatial \nweight of its neighborhoods (row standardized). We \nused a queen contiguity matrix to calculate the spa-\ntial weight. This produces a theoretical range of 0 to \n10,000. For example, a ZIP code could have a score \nof 10,000 if that ZIP code contained 100% of its \nresidents from a certain group and all adjacent \nneighborhoods also had a population composition \nof 100% of the same group. In practice, no ZIP code \nin Texas has this high of a score for any group, and \nthe scores vary considerably depending on the \ncounty and group in question. For this reason, we \nalso group mean centered (to the county) each of \nthese scores to make them relative to the population \nsizes of the county. Otherwise, these measures \nmight reflect differences in the relative sizes of \nthese groups across areas rather than differences \nwithin a particular county in terms of how these \ngroups are spatially patterned. For example, San \nAntonio is 64.2% Latino, whereas Austin is only \n33.9% Latino. In each of these contexts, what might \nbe considered a disproportionately Latino commu-\nnity would be different, and group mean centering \nwould contextualize this difference. This is an \napproach used in previous work examining neigh-\nborhoods (Sampson, Raudenbush, and Earls 1997). \nFor this study, we included three scores: clustering \nmeasure for percentage Black (non-Latino), cluster-\ning measure for percentage Latino (of any race), \nand clustering measure for percentage Asian \n(non-Latino).3\nTo examine whether these patterns relate to \nexisting disparities in resource distribution, we also \nincluded four variables for counts of organizations \nthat are more likely to receive vaccine allocations. \nThese variables came from the CBP data set, which \nclassifies establishments by industry code using \nNAICS codes. In this analysis, we included counts \nby ZIP code of general hospitals (622///), physi-\ncians\u2019 offices (6211//), pharmacies (446110), and \nretirement and assisted living communities for the \nelderly (6233//).4\nWe also included several control variables to \naccount for populations that are eligible to receive \nthe vaccine earlier on the priority list, vulnerability \nto the disease, and other sociodemographic factors. \nThese included population density, percentage age \n65 and above, percentage of people employed in \nservice occupations, median family income, per-\ncentage of people with a bachelor\u2019s degree or \nhigher, percentage of households with no private \nvehicle, and cumulative number of infections in a \nZIP code. Descriptive statistics for all variables can \nbe found in Table 1.\nMethods\nTo model these two different dependent variables, \nwe present two sets of models, one for each out-\ncome. First, for the number of vaccine distribution \nsites, we estimated a series of spatial error models. \nWe calculated univariate global Moran\u2019s I statistics \n\n478\t\nJournal of Health and Social Behavior 63(4)\nfor our key independent variables that suggested \nthat spatial autocorrelation was a problem and \nwould therefore not meet the assumptions of ordine-\nary least squares (OLS) regression. Furthermore, \nthe LaGrange multiplier statistics indicated that the \nspatial error model was the most appropriate method \nto contend with this autocorrelation (Anselin, \nFlorax, and Rey 2004). Specifically, we used a \nqueen spatial weight matrix because this was found \nto best maximize global Moran\u2019s I for each of the \nkey variables used in the models (Anselin 1995; \nAnselin, Florax, and Rey 2004). We also included \nKelejian and Prucha (2010) robust standard errors \nto account for significant heteroscedasticity. \nHowever, of note, the term for lambda is not signifi-\ncant in the models presented in Table 2, suggesting \nthat correlated errors in omitted variables may not \nbe particularly a problem here. The OLS results \nwere also virtually identical to what is presented \nhere in terms of the sign, significance, and relative \neffect sizes for each of our variables. Because spa-\ntial autocorrelation was significant in our prelimi-\nnary analyses, though, we chose to present the \nresults as spatial error models. These results can be \nfound in Table 2.\nFor the second set of models for the week-by-\nweek allocation of vaccine doses, we present a series \nof hierarchical linear growth models (Raudenbush \nand Bryk 2002; Singer and Willett 2003) with a cor-\nrection for spatial dependency using a queen conti-\nguity spatial weight matrix (Savitz and Raudenbush \n2009). Given the nested structure of the data, \nweekly allocations per ZIP code, we used this \napproach to model the cumulative change over time \nin the number of vaccine doses per ZIP code. \nHowever, because the data were still organized by \nphysically adjacent spatial units of analysis at Level \n2, we used Savitz and Raudenbush\u2019s (2009) routine \nto account for spatial dependency using HLM 8.1, \nwhich is an approach used in similar health research \nTable 1.\u2002 Descriptive Statistics for Variables Used in Statistical Models.\nVariable Name\nMean\nSD\nRange\nDescription\nDependent variables\nVaccine sites\n7.92\n17.95\n0 to 169.71\nNumber of vaccine sites per \n100,000 people in a ZIP code\nVaccine doses\n1,199.01\n8,295.32\n0 to 238,415.58\nNumber of vaccine doses allocated \nto ZIP code per 10,000 people\nIndependent variables\nBlack clustering\n0\n648.99\n\u2212700.47 to 4,150.46\nClustering measure of % Black\nLatino clustering\n0\n1,611.93\n\u22122,986.80 to 5,851.91 Clustering measure of % Latino\nAsian clustering\n0\n84.01\n\u221273.08 to 812.42\nClustering measure of % Asian\nPopulation density\n3,291.17\n2,359.18\n6.83 to 16,811.43\nPopulation density per square mile\nPopulation age 65+\n10.94\n4.13\n0 to 36.60\n% of population age 65 and above\n% Service work\n17.01\n6.43\n0 to 43.40\n% of employed population in \nservice occupations\nHousehold income\n68,094.12\n31,467.35\n17,798 to 240,417\nMedian household income\n% Bachelor\u2019s degree\n32.96\n21.30\n2.60 to 90.00\n% of population over 25 with at \nleast a bachelor\u2019s degree\n% No vehicle\n6.10\n5.38\n0 to 39\n% of households with no private \nvehicle\nInfections\n2,127.59\n1,652.52\n0 to 9,906\nNumber of positive SARS-CoV-2 \ninfections\nHospitals\n.61\n1.46\n0 to 15\nNumber of general hospitals\nPhysicians\u2019 offices\n26.84\n41.99\n0 to 294\nNumber of physicians\u2019 offices\nPharmacies\n3.89\n3.39\n0 to 18\nNumber of pharmacies\nRetirement \ncommunities\n1.41\n1.88\n0 to 14\nNumber of retirement and \nassisted living facilities for the \nelderly\nNote: N = 431. Data come from the Texas Department of State Health Services, the 2014\u20132018 American \nCommunity Survey, and the 2016 County Business Patterns.\n\nAnderson and Ray-Warren\t\n479\n(O\u2019Connell 2015). These results can be found in \nTable 3.\nFor each set of models, we first present a model \nwith only the clustering scores and ZIP-code-level \ncontrol variables included (Model 1). Then, we add \nthe health care resource variables one by one in the \nmodel (Models 2\u20135) to avoid the problems of mul-\nticollinearity and to isolate the effect of each orga-\nnizational type as health care organizations tend to \nagglomerate. To our knowledge, no formal media-\ntion test exists that can account for the spatial \ndependencies in the models and the multilevel \nstructure of the second set of models. Therefore, \nwe used an informal approach and examine change \nin the effect sizes with the inclusion of certain \nvariables.\nResults\nFrom the results in Table 2, we can see that the \nracial-ethnic clustering scores have a significant \nrelationship to the density of vaccine sites. \nSpecifically, for Black and Latino clustering, these \nscores are significant and negative, meaning that as \nthe clustering of these two groups increases, the \nnumber of vaccine sites per 100,000 people \ndecreases. Essentially, the higher the concentration \nand clustering of these two groups, the lower the \nTable 2.\u2002 Coefficients and Z-Ratios from Spatial Error Models of Vaccination Sites per 100,000 People.\nVariable Name\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nBlack clusteringa\n\u2212.535***\n(\u20133.299)\n\u2212.503**\n(\u20133.044)\n\u2212.461**\n(\u20132.819)\n\u2212.516**\n(\u20133.091)\n\u2212.547***\n(\u20133.326)\nLatino clusteringa\n\u2212.137*\n(\u20132.185)\n\u2212.101\n(\u20131.530)\n\u2212.104\n(\u20131.625)\n\u2212.120\u2020\n(\u20131.793)\n\u2212.149*\n(\u20132.315)\nAsian clusteringa\n.474\n(.383)\n.198\n(.226)\n.100\n(.097)\n.230\n(.187)\n.565\n(.446)\nPopulation density\n\u2212.001**\n(\u20132.729)\n\u2212.001**\n(\u20132.687)\n\u2212.002**\n(\u20133.033)\n\u2212.002**\n(\u20132.845)\n\u2212.001**\n(\u20132.680)\nPopulation age 65 and up\n.099\n(.353)\n.173\n(.594)\n.085\n(.291)\n.105\n(.377)\n.204\n(.691)\n% Service work\n\u2212.426\u2020\n(\u20131.767)\n\u2212.391\u2020\n(\u20131.697)\n\u2212.406\u2020\n(\u20131.711)\n\u2212.418\u2020\n(\u20131.723)\n\u2212.413\u2020\n(\u20131.713)\nMedian household incomea\n\u2212.073\n(\u20131.380)\n\u2212.029\n(\u2013.629)\n\u2212.053\n(\u20131.137)\n\u2212.068\n(\u20131.318)\n\u2212.081\n(\u20131.482)\n% Bachelor\u2019s degree\n.172\u2020\n(1.826)\n.102\n(1.206)\n.100\n(1.069)\n.164\u2020\n(1.757)\n.187\u2020\n(1.897)\n% No vehicle\n2.193***\n(4.175)\n2.014***\n(3.726)\n2.040***\n(3.887)\n2.180***\n(4.122)\n2.153***\n(4.132)\nInfectionsa\n.004\n(.100)\n\u2212.015\n(\u2013.386)\n\u2212.024\n(\u2013.598)\n\u2212.030\n(\u2013.754)\n.022\n(.533)\nHospitals\n3.290**\n(2.866)\n\u2002\nPhysicians\u2019 offices\n.097*\n(2.383)\n\u2002\nPharmacies\n.419\u2020\n(1.677)\n\u2002\nRetirement communities\n\u2212.719*\n(\u20132.387)\nLambda\n\u2212.935\n(\u2013.765)\n\u22121.263\n(\u20131.000)\n\u2212.740\n(\u2013.603)\n\u2212.775\n(\u2013.640)\n\u2212.660\n(\u2013.540)\nPseudo R2\n.345\n.408\n.384\n.349\n.350\nNote: N = 431. Data come from the Texas Department of State Health Services, the 2014\u20132018 American \nCommunity Survey, and the 2016 County Business Patterns.\naCoefficient multiplied by 100 for the ease of presentation.\n\u2020p < .1, *p < .05, **p < .01, ***p < .001 (for two-tailed test).\n\n480\t\nJournal of Health and Social Behavior 63(4)\nnumber of vaccine sites. These effect sizes are also \nlarge. We discuss these coefficients in terms of stan-\ndard deviation changes because the scale of each \nvariable is so different, although the original coeffi-\ncients are available in the tables. In the case of Black \nclustering, a 1 SD (648.99) increase in Black clus-\ntering relates to 3.47 decrease in the number of vac-\ncine sites per 100,000 people. Given that the average \nnumber of sites per 100,000 people is only 7.92, this \nis a notable change. For Latino clustering, a 1 SD \n(1,611.93) increase in Latino clustering is related to \na 1.59 decrease in the number of vaccine sites per \n100,000 people. These are both sizable coefficients \nand indicate that Black- and Latino-clustered areas \nare less likely to have vaccine sites.\nSeveral of the other area-level coefficients are \nsignificant as well. Population density is significant \nand negative, the percentage of households with no \ncar is significant and positive, and two others, per-\ncentage college educated and percentage in service \nwork, are significant at the .1 level. The largest \neffect comes from the percentage of households \nwith no car, where a 1 SD increase (5.38) relates to \nan increase of 11.8 vaccine sites. This seems to run \ncounter to what we would expect from the literature \nbecause private vehicle ownership may relate to \nTable 3.\u2002 Coefficients and T-Ratios from Linear Growth Models of Weekly Vaccination Allocations per \n10,000 People.\nVariable Name\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nFixed effects\nWeek\n312.927***\n(8.553)\n312.927***\n(8.553)\n312.927***\n(8.553)\n312.927***\n(8.553)\n312.927***\n(8.553)\nBlack clustering\n\u22121.014*\n(\u20131.984)\n\u2212.820\u2020\n(\u20131.785)\n\u2212.593\n(\u20131.215)\n\u2212.907\u2020\n\u20131.771)\n\u22121.036*\n(\u20132.029)\nLatino clustering\n\u2212.424\u2020\n(\u20131.657)\n\u2212.204\n(\u2013.886)\n\u2212.250\n(\u20131.026)\n\u2212.345\n(\u20131.337)\n\u2212.465\u2020\n(\u20131.810)\nAsian clustering\n4.731\n(1.256)\n3.212\n(.949)\n2.669\n(.745)\n3.597\n(.947)\n4.956\n(1.317)\nPopulation density\n\u2212.376*\n(\u20132.355)\n\u2212.347*\n(\u20132.415)\n\u2212.467**\n(\u20133.072)\n\u2212.437**\n(\u20132.687)\n\u2212.352*\n(\u20132.190)\nPopulation age 65 and up\n\u2212141.990\u2020\n(\u20131.841)\n\u221290.095\n(\u20131.298)\n\u2212150.614*\n(\u20132.060)\n\u2212140.289\u2020\n(\u20131.825)\n\u2212109.474\n(\u20131.364)\n% Service work\n\u2212154.071\u2020\n(\u20131.941)\n\u2212138.248\u2020\n(\u20131.939)\n\u2212138.666\u2020\n(\u20131.842)\n\u2212148.326\u2020\n\u20131.873)\n\u2212147.523\u2020\n(\u20131.858)\nMedian household income\n\u2212.038*\n(\u20132.109)\n\u2212.012\n(\u2013.752)\n\u2212.027\n(\u20131.608)\n\u2212.035*\n(\u20131.970)\n\u2212.040*\n(\u20132.250)\n% Bachelor\u2019s degree\n42.995\n(1.348)\n2.743\n(.095)\n6.362\n(.207)\n39.659\n(1.245)\n48.578\n(1.513)\n% No vehicle\n485.960***\n(6.382)\n384.743***\n(5.568)\n401.661***\n(5.487)\n478.015***\n(6.288)\n471.639***\n(6.149)\nInfections\n\u2212.223\n(\u20131.132)\n\u2212.323\u2020\n(\u20131.826)\n\u2212.354\u2020\n(\u20131.891)\n\u2212.371\u2020\n(\u20131.759)\n\u2212.156\n(\u2013.774)\nHospitals\n1855.981***\n(10.099)\n\u2002\nPhysicians\u2019 offices\n50.823***\n(6.954)\n\u2002\nPharmacies\n192.471\u2020\n(1.908)\n\u2002\nRetirement communities\n\u2212241.504\n(\u20131.440)\nRandom effects\nLevel 2 error variance\n30,132,845.834\n23,546,583.226\n26,689,874.927\n29,920,526.251\n30,046,250.421\nLevel pseudo R2\n.146\n.333\n.244\n.152\n.149\nNote: Level 1 N = 3,879. Level 2 N = 431. Data come from the Texas Department of State Health Services, the \n2014\u20132018 American Community Survey, and the 2016 County Business Patterns. The Level 2 pseudo R2 is calculated \nfrom the proportional reduction in error variance from a model with no Level 2 variables (\u03c4 = 35,296,518.409).\n\u2020p < .1, *p < .05, **p < .01, ***p < .001 (for two-tailed test).\n\nAnderson and Ray-Warren\t\n481\nsocioeconomic status, but it may also be an indica-\ntion of central city location where residents may \nperceive less of a need to own a vehicle. Taken \ntogether, from the pseudo R2 value, these sociode-\nmographic characteristics of ZIP codes explain \n34.5% of the variation in vaccine sites per 100,000 \npeople, which is sizable.\nIn Models 2 to 5, when we add the health care \nestablishment variables, this pattern changes some-\nwhat. First, when we include two such health care \nestablishments, the number of hospitals and the \nnumber of physicians\u2019 offices, these are both sig-\nnificant and positive. This indicates that having \nmore health care establishments in a ZIP code \nrelates to a greater number of vaccine distribution \nsites. This is a notable increase as well, where each \nadditional hospital is related to an increase of 3.29 \nvaccination sites and each additional physician\u2019s \noffice is related to a .10 increase in sites. This is \nexpected because hospitals and clinics are where \nmost of the vaccine doses were allocated.\nThe addition of these two variables also reduces \nthe size of the coefficients for the clustering vari-\nables, suggesting that the lower number of vaccine \nsites in these areas is a function of their lack of \nhealth care resources. In the case of Latino cluster-\ning, the coefficients drop to nonsignificance with \nthe inclusion of hospitals and physicians\u2019 offices, \nand these changes relate to a 26.48% and a 23.99% \nreduction, respectively, in the size of the coeffi-\ncients for Latino clustering. This percentage change \nin the coefficient size is somewhat smaller for \nBlack clustering at 9.09% and 13.77%, respec-\ntively. This is reflected in their pseudo R2 values as \nwell, where hospitals alone explain an additional \n6.33% of the variation in vaccine sites and physi-\ncians\u2019 offices explain 3.96%. This suggests that \nexisting health care resources in areas may explain \nsome part of the negative association between \nminority clustering in ZIP codes and the number \nof vaccine sites. However, these findings are only \nsignificant at the .1 level for pharmacies and signif-\nicant and negative for retirement communities. \nTherefore, the bulk of this effect seems to be driven \nby hospitals and physicians\u2019 offices rather than \nother kinds of facilities that may offer the vaccine.\nWe also illustrate these patterns graphically in a \nseries of maps for each county. These can be found \nin Figures 1 to 5. In the background of each map is \na choropleth quintile map of the two clustering \nscores with the location of the vaccine sites overlaid \non top. Choropleth maps use shading to map pat-\nterns across a polygonal area. Here, we use a quin-\ntile map, meaning that the range of mapped values \nin the polygons include an equal number of ZIP \ncodes for each shade. Note that because these are \nevenly distributed quintile maps, the scale is differ-\nent for each group and location. From these maps, it \nis clear that minority clustered areas are less likely \nto have vaccine sites, but these patterns differ some-\nwhat by county, with more stark patterns in certain \ncounties over others and depending on the group in \nquestion. For example, Harris and Travis counties \neach have a clearer clustering pattern to them where \nthe vaccine sites appear to be less likely to be \nlocated in racial-ethnic minority areas. By contrast, \nTarrant County (Fort Worth) has a more scattered \npattern to the location of the facilities.\nIn Table 3, we present the results for the number \nof vaccine allocations over the 9-week period using \na series of linear growth models. The results with \nthis dependent variable reflect a similar pattern to \nthe one described previously with some notable dif-\nferences. Again, in Model 1, the coefficient for \nBlack clustering is significant and negative, indicat-\ning that the higher the degree of concentration and \nclustering of Blacks across these five counties, the \nsmaller the vaccine allocations week to week. \nSpecifically, every 1 SD increase in Black cluster-\ning (648.99) relates to 658.08 fewer vaccine doses \nper 10,000 people in a ZIP code. The same figure \nfor a 1 SD change in Latino clustering (1,611.93) is \n492.66 vaccine doses per 10,000 people, which is \nnot an inconsequential amount, although this coef-\nficient is only significant at the .1 level for Latino \nclustering. In all subsequent models, the results for \nLatino clustering are not significant. Concerning \nthe control variables, the results are similar to pre-\nvious models except that median household income \nis significant and negative here.\nWhen we add the different types of health care \nestablishments to the baseline model, the results are \nsimilar in pattern to the results previously described. \nOnce again, the distribution of hospitals and physi-\ncians\u2019 offices is significant and positive, meaning \nthat their presence in an area is related to a higher \nallocation of vaccine doses per 10,000 people. \nMoreover, the inclusion of these variables weakens \nthe association between the clustering scores and \nthe number of vaccine doses per 10,000 people, \nespecially in the case of Black clustering. Black \nclustering is only significant in Model 1 and Model \n5, and the inclusion of hospitals and physicians\u2019 \noffices rather substantially reduces the size of those \ncoefficients. These reductions are percentage \nchanges of 19.13%, 41.52%, and 10.55% for hospi-\ntals, physicians\u2019 offices, and pharmacies, respec-\ntively. Hospitals and physicians\u2019 offices appear to \n\n482\t\nJournal of Health and Social Behavior 63(4)\nexplain away a substantial portion of the differen-\ntial allocations across areas related to Black cluster-\ning. We also see changes in the size of the \ncoefficients for Latino clustering, but this score was \nonly significant at the .1 level in Model 1.\nDiscussion\nIn this study, we aim to understand how neighbor-\nhood sociodemographic characteristics relate to the \nearly COVID-19 vaccine rollout. We hypothesize \nthat racial-ethnic minority clustered areas will be \nless likely to have vaccination sites, and we surmise \nthat this is primarily a function of a lack of key \nhealth care sites prior to the pandemic. We test these \nhypotheses across the five largest urban counties in \nthe state of Texas during the first 10 weeks of the \nvaccine rollout in Texas (excluding the first week). \nMoreover, we test this using two ways of capturing \naccess to the vaccine for neighborhoods\u2014the num-\nber of vaccine sites and the number of doses allo-\ncated to each site.\nFor the first outcome, examining the number of \nvaccine sites across these urban counties, we find \nthat a higher concentration and clustering of Black \nand Latino residents in ZIP codes is associated with \nfewer vaccine sites per 100,000 people. These \nresults provide support for Hypothesis 1 for Black \nFigure 1.\u2002 Vaccine Distribution Sites over Racial-Ethnic Clustering Scores in Bexar County (San \nAntonio), Texas.\nNote: Data come from the Texas Department of State Health Services and the 2014\u20132018 American Community \nSurvey.\n\nAnderson and Ray-Warren\t\n483\nand Latino clustering (not Asian clustering). Moreover, \nthis negative association is partially explained by \nthe existing distribution of health care resources \nacross communities, especially hospitals and physi-\ncians\u2019 offices. The presence of these establishments \nis associated with a higher number of vaccine sites, \nso the lack of these sites in segregated communities \nmeans fewer vaccine sites for these areas. However, \nwe do not observe the same association for pharma-\ncies and retirement communities. Indeed, the asso-\nciation for retirement communities is actually \nsignificant and negative and appears to have a sup-\npression effect with the racial-ethnic clustering \nscores. We speculate that this may be because these \ngroups are less likely to use retirement homes for \nelderly relatives as other work suggests (Dilworth-\nAnderson, Williams, and Gibson 2002), although \nwe cannot directly test this assertion with the data \nhere. These findings provide partial support for \nHypothesis 2, with stronger evidence for the case of \nLatino clustering.\nFor the second outcome, we find a similar pat-\ntern, although with somewhat weaker effects. Black \nclustering is again significant and negatively asso-\nciated with the number of vaccine doses per 10,000 \npeople in a ZIP code over a 9-week period (although \nthis is only significant for Latino clustering at the .1 \nlevel), meaning that minority segregated communi-\nties were less likely to receive doses of vaccine over \nthis period. This provides further partial support for \nFigure 2.\u2002 Vaccine Distribution Sites over Racial-Ethnic Clustering Scores in Dallas County (Dallas), \nTexas.\nNote: Data come from the Texas Department of State Health Services and the 2014\u20132018 American Community \nSurvey.\n\n484\t\nJournal of Health and Social Behavior 63(4)\nHypothesis 1 in the case of Black clustering. And, \nagain, these results are somewhat attenuated with \nthe inclusion of the number of hospitals and physi-\ncians\u2019 offices. Both these types of establishments \nare strongly associated with a higher number of \nvaccine doses, and they reduce the size of the coef-\nficient for Black clustering. Thus, we find partial \nsupport for Hypothesis 2.\nThese findings are fitting with previous litera-\nture on segregation and the distribution of resources \nmore broadly that demonstrates poor access to a wide \nvariety of community establishments that would sup-\nport well-being (Anderson 2017; Dinwiddie et\u00a0al. \n2013; Gaskin et\u00a0al. 2012; Ko and Ponce 2013). We \nfind this for Black and Latino clustering but not for \nAsian clustering. However, the existing work on \nAsian segregation is limited and has not found the \nsame inequalities as Black and Latino segregations \n(Anderson 2017). The findings presented here also \npoint to a much stronger race story than one about \nsocioeconomic dynamics, which is a prominent \ntheme in much of the previous literature on resource \nallocation across neighborhoods (Beaulac et\u00a0 al. \n2009; Ko et\u00a0 al. 2014). In our study, area-level \nmedian income was not significant in the first anal-\nysis, and in the second analysis, it was significant \nand negative, meaning that higher income areas \nwere less likely to receive doses. Thus, the story \nFigure 3.\u2002 Vaccine Distribution Sites over Racial-Ethnic Clustering Scores in Harris County (Houston), \nTexas.\nNote: Data come from the Texas Department of State Health Services and the 2014\u20132018 American Community \nSurvey.\n\nAnderson and Ray-Warren\t\n485\nseems to be more about race specifically rather than \nclass, which limited research has also shown \n(Anderson 2017; Small and McDermott 2006).\nGiven the lack of detail provided by the state in \nterms of how specifically the vaccine allocation \nalgorithm was designed, we are not necessarily sug-\ngesting that this is an act of blatant and purposive \nracism on the part of the advisory board responsible \nfor vaccine allocation decision-making. Rather, \nbecause the infrastructure in terms of health care \nresources already disadvantages racial-ethnic minor-\nity communities, this represents a case of structural \ndiscrimination. Thus, we extend this literature by \ndemonstrating yet another empirical instance of a \nracial disparity, especially for a hugely impactful \nevent like the current COVID-19 crisis. Some lim-\nited findings on the COVID-19 pandemic, including \nmedia sources, has revealed important disparities by \nrace-ethnicity in infection rates and access to medi-\ncal services, such as testing sites and ICU facilities \n(CDC 2020b; McMinn et\u00a0al. 2020; Miller, Peek, and \nParker 2020; Millett et\u00a0 al. 2020; Ross 2021). The \nvaccine rollout appears to fit within this pattern. This \nis relevant to the current inequities we have seen \nthroughout the pandemic and has implications for \nthe next crisis.\nFigure 4.\u2002 Vaccine Distribution Sites over Racial-Ethnic Clustering Scores in Tarrant County (Fort \nWorth), Texas.\nNote: Data come from the Texas Department of State Health Services and the 2014\u20132018 American Community \nSurvey.\n\n486\t\nJournal of Health and Social Behavior 63(4)\nDespite these advances to the literature, the \nstudy has several limitations. First, the study only \nrelates the sociodemographic characteristics of \nneighborhoods with the distribution of vaccine \nsites and doses and not who is being vaccinated at \nthese sites. Media reporting on access to the vac-\ncine has highlighted how differential access to the \nInternet to search for vaccine appointments, time \nto spend searching for open appointments, and \nadequate transportation have led to important dis-\nparities in early access to the vaccine, even for eli-\ngible populations (Garnam and Cai 2021; Harper \n2021; Menchaca and Agnew 2021; Oladipo 2021; \nRoss 2021; Stone 2021). This study only factors in \ndisparities in the built environment in terms of the \nlocation of vaccine sites without accounting for \nwho is going to these sites. Second, the data from \nthe Texas Department of Health Services only \nincludes address locations for where vaccine doses \nwere shipped. Although the vast majority of ship-\nping locations were the same as where the vaccines \nwere administered, this approach fails to consider \nmobile units that some city and county health \ndepartments have utilized to reach special popula-\ntions, like mobility-challenged patients in elder \ncare facilities. These types of activities were not \npossible to systematically track over time. Future \nwork should consider these limitations.\nFigure 5.\u2002 Vaccine Distribution Sites over Racial-Ethnic Clustering Scores in Travis County (Austin), \nTexas.\nNote: Data come from the Texas Department of State Health Services and the 2014\u20132018 American Community \nSurvey.\n\nAnderson and Ray-Warren\t\n487\nWith this study, we contribute to the literature \non the unequal distribution of resources across \nneighborhoods, particularly by racial-ethnic minor-\nity clustering. Moreover, we demonstrate that these \npatterns are not a neutral fact. Not having health \ncare infrastructure in place means that when con-\nfronted with a public health catastrophe, the exist-\ning inequalities in our health care system are \ndeepened. The COVID-19 pandemic is, for many \npeople, one of the most disruptive and challenging \npublic health events of our lifetimes. Vaccination in \nthis context represents a lifeline to spare further \nhuman suffering and loss of life and a potential \nreturn to normalcy. However, this valuable resource \nwas not distributed evenly across urban areas, with \nlimited access to populations already at risk for \ncomplications from the virus. Although state public \nhealth officials implemented eligibility systems that \nprioritized health care workers, the elderly, and \nthose with medical comorbidities, the geographic \ncomponent of the vaccine rollout and allocation of \nthe vaccine over time has not been equal. This high-\nlights the necessity of creating more equitable \naccess to care broadly so that in crisis times, the \ninfrastructure is available to equitably meet the \nneeds of the affected communities.\nORCID iD\nKathryn Freeman Anderson \n https://orcid.org/0000- \n0002-7425-8923\nSupplemental Material\nAppendix A is available in the online version of the \narticle.\nNotes\n1.\t\nWe also ran the same models using the 2018 version \nof the data, which is the most recent version avail-\nable but likely has undercounts of establishments in \nmany areas. The results generally followed the same \npattern as what is presented here but with different \neffect sizes in some cases. As an additional sensitiv-\nity check, we also checked previous waves of the \ndata to see how correlated the counts are over time \nto justify the use of an earlier date. Using the 2012 \nCBP file (being four years before our data, which \nis the same distance from 2016 to 2020), we found \nthat all four of our organizational types were highly \ncorrelated. From 2012 to 2016, hospitals, physi-\ncians\u2019 offices, pharmacies, and retirement com-\nmunities had a correlation of .93, .98, .86, and .86, \nrespectively, indicating that these organizational \nresources are relatively stable over time, especially \nhealth care provision.\n2.\t\nOur analysis focused on these clustering scores as \na more geographically informed way of examining \nthe problem. To relate these findings to the broader \nliterature that typically uses composition scores, we \nran the same models using percentage non-Latino \nBlack, percentage Latino, percentage non-Latino \nAsian, and another version of these models\u2019 vari-\nables that also included their spatial W lags (which \nis the other term in the clustering equation). The \nresults were similar to what we found in the pres-\nent study, with some differences in the effect sizes \n(results available on request). We chose to present \nthe results with the clustering score as indicated by \nthe formula because we think this best captures the \nspatial dynamics of neighborhood-level segregation \nin a manner that accounts for both the composition \nof groups in an area and the extent to which they are \nspatially clustered, which is one of the main ways \nthat segregation is theorized and conceptualized in \nthe literature (Massey and Denton 1988).\n3.\t\nThe distribution of Whites also plays a role in the \nsegregation level of an area. However, due to multi-\ncollinearity, we could not include all four clustering \nscores for each of these groups in a single model. \nInstead, we focused on the distribution of racial-\nethnic minority populations across these counties. \nAs a check on this choice of approach, we also ran \nthe same models using the clustering score for per-\ncentage White. This score for Whites was not sig-\nnificant in any of the models, suggesting that the \ndistribution of Whites across areas is unrelated to \nthe distribution of vaccine sites and allocations.\n4.\t\nWe also tested several other organizational types \nthat might have received vaccine allocations, \nincluding general stores (452///), supermarkets \n(4451//), freestanding ambulatory care facilities \n(621493), and other (nonelderly) nursing and resi-\ndential care facilities (623///). We only present the \nresults for those that were most relevant and had \nsignificant results.\nReferences\nAcevedo-Garcia, Dolores. 2000. \u201cResidential Segregation \nand the Epidemiology of Infectious Diseases.\u201d Social \nScience & Medicine 51(8):1143\u201361.\nAcevedo-Garcia, Dolores. 2001. \u201cZIP Code-Level Risk \nFactors for Tuberculosis: Neighborhood Environ\u00ad\nment and Residential Segregation in New Jersey, \n1985\u20131992.\u201d American Journal of Public Health \n91(5):734\u201341.\nAcevedo-Garcia, Dolores, Kimberly Lochner, Theresa \nOsypuk, and S. V. Subramanian. 2003. \u201cFuture \nDirections in Residential Segregation and Health \nResearch: A Multilevel Approach.\u201d American Journal \nof Public Health 93(2):215\u201321.\nAllard, Scott W. 2009. Out of Reach: Place, Poverty, and \nthe New American Welfare State. New Haven, CT: \nYale University Press.\n\n488\t\nJournal of Health and Social Behavior 63(4)\nAnderson, Kathryn Freeman. 2017. \u201cRacial Residential \nSegregation and the Distribution of Health-Related \nOrganizations in Urban Neighborhoods.\u201d Social \nProblems 64(2):256\u201376.\nAnderson, Kathryn Freeman. 2020. \u201cResidential Segre\u00ad\ngation, Neighborhood Health Care Organizations, \nand Children\u2019s Health Care Utilization: The Case \nof the Phoenix Urbanized Area.\u201d City & Community \n19(3):771\u2013801.\nAnderson, Kathryn Freeman, and Andrew Fullerton. \n2014. \u201cResidential Segregation, Health, and Health \nCare: Answering the Latino Question.\u201d Race and \nSocial Problems 6(3):262\u201379.\nAnselin, Luc. 1995. \u201cLocal Indicators of Spatial \nAssociation\u2014LISA.\u201d Geographical Analysis 27(2): \n93\u2013115.\nAnselin, Luc, Raymond Florax, and Sergio Rey. 2004. \nAdvances in Spatial Econometrics: Methodology, \nTools and Applications. Berlin: Springer-Verlag.\nBeaulac, Julie, Elizabeth Kristjansson, and Steven \nCummins. 2009. \u201cA Systematic Review of Food \nDeserts, 1966\u20132007.\u201d Preventing Chronic Disease \n6(3):1\u201310.\nBiello, Katie, Trace Kershaw, Robert Nelson, Matthew \nHogben, Jeannette Ickovics, and Linda Niccolai. \n2012. \u201cRacial Residential Segregation and Rates \nof Gonorrhea in the United States, 2003\u20132007.\u201d \nAmerican Journal of Public Health 102(7):1370\u201377.\nCenters for Disease Control and Prevention. 2020a. \u201cCDC\u2019s \nCOVID-19 Vaccine Rollout Recommendations.\u201d \nhttps://www.cdc.gov/coronavirus/2019-ncov/ \nvaccines/recommendations.html.\nCenters for Disease Control and Prevention. 2020b. \n\u201cCOVID-19 in Racial and Ethnic Minority Groups.\u201d \nhttps://www.cdc.gov/coronavirus/2019-ncov/need-\nextra-precautions/racial-ethnic-minorities.html.\nChan, Kitty, Darrell Gaskin, Gniesha Dinwiddie, and \nRachael McCleary. 2012. \u201cDo Diabetic Patients \nLiving in Racially Segregated Neighborhoods \nExperience Different Access and Quality of Care?\u201d \nMedical Care 50(8):692\u201399.\nCollins, Chiquita. 1999. \u201cRacism and Health: Segregation \nand Causes of Death Amenable to Medical Inter\u00ad\nvention in Major U.S. Cities.\u201d Annals of the New \nYork Academy of Sciences 896:396\u201398.\nCollins, Chiquita, and David Williams. 1999. \u201cSegre\u00ad\ngation and Mortality: The Deadly Effects of \nRacism?\u201d Sociological Forum 14(3):495\u2013523.\nCooksey-Stowers, Kristen, Marlene Schwartz, and Kelly \nBrownell. 2017. \u201cFood Swamps Predict Obesity \nRates Better Than Food Deserts in the United States.\u201d \nInternational Journal of Environmental Research \nand Public Health 14(11):1366\u201386.\nDai, Dajun. 2010. \u201cBlack Residential Segregation, \nDisparities in Spatial Access to Health Care \nFacilities, and Late-Stage Breast Cancer Diagnosis in \nMetropolitan Detroit.\u201d Health & Place 16(5):1038\u201352.\nDilworth-Anderson, Peggye, Ishan Canty Williams, and \nBrent Gibson. 2002. \u201cIssues of Race, Ethnicity, and \nCulture in Caregiving Research: A 20-Year Review \n(1980\u20132000).\u201d The Gerontologist 42(2):237\u201372.\nDinwiddie, Gniesha, Darrell Gaskin, Kitty Chan, \nJanette Norrington, and Rachel McCleary. 2013. \n\u201cResidential Segregation, Geographic Proximity and \nType of Services Used: Evidence for Racial-Ethnic \nDisparities in Mental Health.\u201d Social Science & \nMedicine 80:67\u201375.\nGarnam, Juan Pablo, and Mandi Cai. 2021. \u201cAdvocates \nWorry Vaccines Will Be out of Reach for Black and \nHispanic Neighborhoods Devastated by COVID-\n19.\u201d The Texas Tribune, January 9. https://www \n.texastribune.org/2021/01/09/texas-coronavirus- \nvaccine-racial-inequality/.\nGaskin, Darrell J., Gniesha Dinwiddie, Kitty Chan, and \nRachael McCleary. 2012. \u201cResidential Segregation \nand the Availability of Primary Care Physicians.\u201d \nHealth Services Research 47(6):2353\u201376.\nGaskin, Darrell J., Adrian Price, Dwayne Brandon, and \nThomas LaVeist. 2009. \u201cSegregation and Disparities \nin Health Services Use.\u201d Medical Care Research and \nReview 66(5):578\u201389.\nGodoy, Maria, and Daniel Wood. 2020. \u201cWhat Do \nCoronavirus Racial Disparities Look like State by \nState?\u201d National Public Radio, May 30. https://www \n.npr.org/sections/health-shots/2020/05/30/865413079/\nwhat-do-coronavirus-racial-disparities-look-like-state-\nby-state.\nGordon-Larsen, Penny, Melissa Nelson, Phil Page, and Barry \nPopkin. 2006. \u201cInequality in the Built Environment \nUnderlies Key Health Disparities in Physical Activity \nand Obesity.\u201d Pediatrics 117(2):417\u201324.\nGrady, Sue. 2006. \u201cRacial Disparities in Low Birthweight \nand the Contribution of Residential Segregation: A \nMultilevel Analysis.\u201d Social Science & Medicine \n63(12):3013\u201329.\nHarper, Karen Brooks. 2021. \u201cAs Vaccine Eligibility \nWidens, Some Vulnerable Texans Are Still Fighting for \nAccess.\u201d The Texas Tribune, March 31. https://www \n.texastribune.org/2021/03/31/vaccine-eligibility- \ntexas/.\nHart, Kevin, Stephen Kunitz, Ralph Sell, and Dana \nMukamel. 1998. \u201cMetropolitan Governance, Resi\u00ad\ndential Segregation, and Mortality among African \nAmericans.\u201d American Journal of Public Health 88(3): \n434\u201338.\nKelejian, Harry, and Ingmar Prucha. 2010. \u201cSpecification \nand Estimation of Spatial Autoregressive Models with \nAutoregressive and Heteroskedastic Disturbances.\u201d \nJournal of Econometrics 157(1):53\u201367.\nKo, Michelle, Jack Needleman, Kathryn Pitkin Derose, \nMiriam \nLaugesen, \nand \nNinez \nPonce. \n2014. \n\u201cResidential Segregation and the Survival of U.S. \nUrban Public Hospitals.\u201d Medical Care Research \nand Review 71(3):243\u201360.\nKo, Michelle, and Ninez Ponce. 2013. \u201cCommunity \nResidential Segregation and the Local Supply of \nFederally Qualified Health Centers.\u201d Health Services \nResearch 48(1):253\u201370.\n\nAnderson and Ray-Warren\t\n489\nLee, Min-Ah A. 2009. \u201cNeighborhood Residential \nSegregation and Mental Health: A Multilevel \nAnalysis on Hispanic Americans in Chicago.\u201d Social \nScience & Medicine 68(11):1975\u201384.\nMartinez, A. 2021. \u201cThe Racial Disparities, Systemic \nRacism Behind Who Has Received Vaccines.\u201d \nNational Public Radio, March 18. https://www.npr \n.org/2021/03/18/978496045/the-racial-disparities-\nsystemic-racism-behind-who-has-received-vaccines.\nMarwell, Nicole, and Aaron Gullickson. 2013. \u201cInequality \nin the Spatial Allocation of Social Services: \nGovernment Contracts to Nonprofit Organizations \nin New York City.\u201d Social Service Review 87(2): \n319\u201353.\nMassey, Douglas, and Nancy Denton. 1988. \u201cThe Dimen\u00ad\nsions of Residential Segregation.\u201d Social Forces 67(2): \n281\u2013315.\nMcFarland, Michael, and Cheryl Smith. 2011. \u201cSegregation, \nRace, and Infant Well-Being.\u201d Population Research \nand Policy Review 30(3):467\u201393.\nMcMinn, Sean, Audrey Carlsen, Bret Jaspers, Ruth Talbot, \nand Stephanie Adeline. 2020. \u201cIn Large Texas Cities, \nAccess to Coronavirus Testing May Depend on Where \nYou Live.\u201d National Public Radio, May 27. https://www \n.npr.org/sections/health-shots/2020/05/27/862215848/\nacross-texas-black-and-hispanic-neighborhoods-have-\nfewer-coronavirus-testing-sit.\nMenchaca, Megan, and Duncan Agnew. 2021. \u201cTexas\u2019 \nDecentralized Internet Reliant System for Vaccine \nAppointments Leaves Many Eligible People Unable \nto Access a Shot.\u201d The Texas Tribune, March 20. \nhttps://www.texastribune.org/2021/03/20/texas-\ncoronavirus-vaccine-difficult/.\nMiller, William, Monica Peek, and William Parker. \n2020. \u201cScarce Resource Allocation Scores Threaten \nto Exacerbate Racial Disparities in Health Care.\u201d \nCHEST 158(4):1332\u201334.\nMillett, Gregorio, Austin Jones, David Benkeser, \nStefan Baral, Laina Mercer, Chris Beyrer, Brian \nHonermann, et\u00a0 al. 2020. \u201cAssessing Differential \nImpacts of COVID-19 on Black Communities.\u201d \nAnnals of Epidemiology 47:37\u201344.\nMoore, Latetia, and Ana Diez Roux. 2006. \u201cAssociations \nof Neighborhood Characteristics with the Location \nand Type of Food Stores.\u201d American Journal of \nPublic Health 96(2):325\u201331.\nO\u2019Connell, Heather. 2015. \u201cWhere There\u2019s Smoke: \nCigarette Use, Social Acceptability, and Spatial \nApproaches to Multilevel Modeling.\u201d Social Science \n& Medicine 140:18\u201326.\nOladipo, Gloria. 2021. \u201cHow Chicago\u2019s Vaccine Rollout \nIs Inhibited by Longstanding Inequality.\u201d The \nGuardian, February 5. https://www.theguardian.com/\nus-news/2021/feb/05/chicago-blacks-latinos-vaccine-\ndistribution?utm_term=2bd88cc02b2faffa886b221bd\nd480dc4&utm_campaign=GuardianTodayUS&utm_\nsource=esp&utm_medium=Email&CMP=GTUS_email.\n\u00d8kland, Helene, and Svenn-Erik Mamelund. 2019. \u201cRace \nand 1918 Influenza Pandemic in the United States: \nA Review of the Literature.\u201d International Journal \nof Environmental Research and Public Health \n16(14):2487\u2013505.\nPeek, Monica, Russell Simons, William Parker, David \nAnsell, Selwyn Rogers, and Brownsyne Tucker \nEdmonds. 2021. \u201cCOVID-19 among African Americans: \nAn Action Plan for Mitigating Disparities.\u201d American \nJournal of Public Health 111(2):286\u201392.\nPersad, Govind, Monica Peek, and Ezekiel Emanuel. \n2020. \u201cFairly Prioritizing Groups for Access to \nCOVID-19 Vaccines.\u201d JAMA 324(16):1601\u20131602.\nPrescott, Gina Marie, and William Allan Prescott. 2021. \n\u201cHealth Information Technology Utilization and \nImpact on COVID-19 Vaccination.\u201d Journal of the \nAmerican Pharmacists Association 61(40):E230\u201332.\nRamos, Elliot, and Mar\u00eda In\u00e9s Zamudio. 2020. \u201cIn \nChicago, 70% of COVID-19 Deaths Are Black.\u201d \nWBEZ Chicago, April 5. https://www.wbez.org/\nstories/in-chicago-70-of-covid-19-deaths-are-black/\ndd3f295f-445e-4e38-b37f-a1503782b507.\nRaudenbush, Stephen, and Anthony Bryk. 2002. \nHierarchical Linear Models: Applications and Data \nAnalysis Methods. Thousand Oaks, CA: SAGE \nPublications.\nReardon, Sean, and David O\u2019Sullivan. 2004. \u201cMeasures \nof Spatial Segregation.\u201d Sociological Methodology \n34:121\u201362.\nRoberto, Elizabeth. 2018. \u201cThe Spatial Proximity and \nConnectivity Method for Measuring and Analyzing \nResidential Segregation.\u201d Sociological Methodology \n48(1):182\u2013224.\nRoberts, Jennifer, and Shadi Tehrani. 2020. \u201cEnvironments, \nBehaviors, and Inequalities: Reflecting on the Impacts \nof the Influenza and Coronavirus Pandemics in the \nUnited States.\u201d International Journal of Environmental \nResearch and Public Health 17(12):4484\u2013511.\nRodriguez, Rudolph, Saunak Sen, Kala Mehta, Sandra \nMoody-Ayers, Peter Bacchetti, and Ann O\u2019Hare. \n2007. \u201cGeography Matters: Relationships among \nUrban Residential Segregation, Dialysis Facilities, \nand Patient Outcomes.\u201d Annals of Internal Medicine \n146(7):493\u2013501.\nRoss, Janell. 2021. \u201cWhen a Texas County Tried to \nEnsure Racial Equity in COVID-19 Vaccinations, \nIt Didn\u2019t Go as Planned.\u201d Time Magazine, March 2. \nhttps://time.com/5942884/covid-19-vaccine-racial-\ninequity-dallas/.\nSampson, Robert, Stephen Raudenbush, and Felton \nEarls. 1997. \u201cNeighborhoods and Violent Crime: A \nMultilevel Study of Collective Efficacy.\u201d Science \n277(5328):918\u201324.\nSavitz, Natalya Verbitsky, and Stephen Raudenbush. \n2009. \u201cExploiting Spatial Dependence to Improve \nMeasurement of Neighborhood Social Processes.\u201d \nSociological Methodology 39(1):151\u201383.\nSinger, Judith, and John Willett. 2003. Applied \nLongitudinal Data Analysis: Modeling Change \nand Event Occurrence. New York, NY: Oxford \nUniversity Press.\n\n490\t\nJournal of Health and Social Behavior 63(4)\nSmall, Mario, and Monica McDermott. 2006. \u201cThe \nPresence of Organizational Resources in Poor \nUrban Neighborhoods: An Analysis of Average and \nContextual Effects.\u201d Social Forces 84(3):1697\u2013724.\nStone, Will. 2021. \u201c\u2018Just Cruel\u2019: Digital Race For \nCOVID-19 Vaccines Leaves Many Seniors Behind.\u201d \nNational Public Radio, February 4. https://www.npr.\norg/sections/health-shots/2021/02/04/963758458/\ndigital-race-for-covid-19-vaccines-leaves-many-\nseniors-behind.\nStrully, Kate. 2011. \u201cHealth Care Segregation and Race \nDisparities in Infectious Disease: The Case of Nursing \nHomes and Seasonal Influenza Vaccinations.\u201d \nJournal of Health and Social Behavior 52(4):510\u201326.\nTobler, Waldo. 1970. \u201cA Computer Movie Simulating \nUrban Growth in the Detroit Region.\u201d Economic \nGeography 46(Suppl. 1):234\u201340.\nWalker, Renee, Christopher Keane, and Jessica Burke. \n2010. \u201cDisparities and Access to Healthy Food in the \nUnited States: A Review of Food Deserts Literature.\u201d \nHealth & Place 16(5):876\u201384.\nWilliams, David, and Chiquita Collins. 2001. \u201cRacial \nResidential Segregation: A Fundamental Cause of \nRacial Disparities in Health.\u201d Public Health Reports \n116(5):404\u201316.\nYang, Tse-Chuan, Seung-won Emily Choi, and Feinuo \nSun. 2021. \u201cCOVID-19 Cases in U.S. Counties: \nRoles of Racial-Ethnic Density and Residential \nSegregation.\u201d Ethnicity & Health 26(1):11\u201321.\nAuthor Biographies\nKathryn Freeman Anderson is an associate professor in the \nDepartment of Sociology at the University of Houston. Her \nresearch focuses on understanding the social sources of health \ndisparities in the United States. In particular, she examines the \nrole of race-ethnicity and urban neighborhood dynamics to \nanalyze how these factors may affect individual health. Her \nrecent work has been published in Social Problems, City & \nCommunity, and Race and Social Problems.\nDarra Ray-Warren is a graduate student in the Department \nof Sociology at the University of Houston. Her research \ninterests are in the socioeconomic determinants affecting \nmental and physical health, social status, and the public \npolicies that may impact human flourishing. Her current \nresearch focuses on criminal law and punishment ratio-\nnale, emerging progressive prosecutorial and indigent \ndefense styles, and the social factors that drive criminal \ncase outcomes.\n\n\n    Scientific Research Findings:", "answer": "\u2022 Black and Latino segregated neighborhoods were less likely to have COVID-19 vaccination facilities.\n\u2022 Black neighborhoods also received fewer doses to those facilities.\n\u2022 These two patterns were largely explained by the existing lack of hospitals and physicians\u2019 offices in those areas. Because most COVID-19 vaccine distribution sites were set up in existing health care facilities, these neighborhoods were less likely to have the necessary infrastructure to be able to set up distribution sites and receive large shipments of doses.", "id": 73}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/0022146520979664 \nJournal of Health and Social Behavior\n2021, Vol. 62(1) 2\u00ad\u201318\n\u00a9 American Sociological Association 2021\nDOI: 10.1177/0022146520979664 \njhsb.sagepub.com\nArticle\nThe connections between education and health are \none of the most widely studied and empirically robust \nin the field of medical sociology (Adler et\u00a0al. 1994; \nConti, Heckman, and Urzua 2010). Education has \nbeen associated with assorted morbidities, mortality \nrisk, disability, general health, mental health, and an \narray of health-related behaviors linked to various \nhealth outcomes (Adams 2002; Arendt 2005; Cutler \nand Lleras-Muney 2010; Lleras-Muney 2005; Turner \nand Lloyd 1999). One important question about the \nlinks between education and health that remains \nunanswered, however, is whether these links are \nrobust for individuals who pursued schooling at a \nstage of the life course when education has norma-\ntively been completed (as reflected in formal defini-\ntions of nontraditional students, as well as \ninstitutional expectations and structures): prior to \nhaving children (Macmillan 2005; Marini1984). This \nquestion of whether \u201cout-of-sequence\u201d education can \npromote greater health is particularly salient for \nwomen with children, who (based on estimates from \n2015 to 2016), represent over 15% of all students \nenrolled in college and 70% of all student parents \n(Cruse et\u00a0al. 2019).\nThis question is also relevant to current public \nhealth policy given that scholars from across disci-\nplines have argued that one of the most powerful \npolicy levers for reducing population-level health \ndisparities is promoting higher education (Frieden \n2010; Hahn and Truman 2015). This argument has \nbeen especially directed toward lower-educated \n979664 HSBXXX10.1177/0022146520979664 Journal of Health and Social BehaviorAugustine\nresearch-article2021\n1University of South Carolina, Columbia, SC, USA\nCorresponding Author:\nJennifer March Augustine, Department of Sociology, \nUniversity of South Carolina, 911 Pickens St., Columbia, \nSC 29208, USA. \nEmail: augustjm@mailbox.sc.edu\nMothers\u2019 Out-of-Sequence \nPostsecondary Education \nand Their Health and Health \nBehaviors\nJennifer March Augustine1\nAbstract\nAmple research suggests that the links between higher education and heath are robust and growing in \nstrength. This research, however, tends to assume education was completed prior to assuming other adult \nroles. Importantly, the life course framework raises the question of whether \u201cout-of-sequence\u201d college \ncompletion conveys similar health returns. I investigate this question among a population for whom out-\nof-sequence schooling has grown more common: lower-educated mothers. This focus is also important \ngiven the growing education gap in women\u2019s health and the links between maternal and child health. Data \ncome from the Fragile Families and Child Wellbeing Study (N = 4,898). Analyses involve random intercept \nand fixed effects models and diverse health measures. Findings suggest that postsecondary education \ndoes not improve mother\u2019s health, except for reduced smoking among mothers with high school degrees \nor less that earned bachelor\u2019s degrees. These findings inform health policy debates and theories linking \neducation to health.\nKeywords\nmaternal education, maternal health, stratification\n\nAugustine\t\n3\nwomen, who have recently experienced greater risks \nto their health than less educated men (Montez and \nZajacova 2013) and whose lower levels of education \nand health have been implicated in the low levels of \nhealth and development of their children and inter-\ngenerational transmission of inequality more \nbroadly (Gardner, Brooks-Gunn, and Chase-\nLansdale 2017; Kaushal 2014; Turney 2011). It \nremains unknown, however, whether promoting \nmothers\u2019 school reentry can remediate these issues. \nThis uncertainty is further underscored by a key \ntenet of the life course perspective (Elder 1984), \nwhich suggests that out-of-sequence life course pro-\ngressions can create challenges and conditions that \nundermine the efficacy of individuals in their roles \nand statuses (Pearlin et\u00a0 al. 2005) and reduce the \nrewards and resources that such roles and statues are \nsupposed to generate (Furstenberg 2005; also see \nHogan 1978 for a key theoretical contribution). \nThus, this question also raises the broader theoreti-\ncal issue of whether out-of-sequence schooling cre-\nates challenges that diminish the well-documented \nreturns to higher education (e.g., better labor market \nopportunities, wider social networks, increased cog-\nnitive skills; Mirowsky and Ross 2003; Zajacova \nand Lawrence 2018) for women with children.\nTo investigate these substantive, theoretical, \nand policy-relevant issues, I draw on data from a \nnational sample of largely lower-educated mothers \n(i.e., without postsecondary degrees at the baseline \ndata collection) participating in the Fragile \nFamilies and Child Wellbeing Study (FFCW). Not \nonly does the FFCW focus on this study\u2019s popula-\ntion of interest, it includes repeated measures of \nmaternal education and health subsequent to the \nbirth of child that allows for a dynamic investiga-\ntion of the connections among these two life course \npathways. As part of this investigation, I attend to \nissues of heterogeneity connected to the type of \npostsecondary degrees mothers completed (e.g., \nvocational, associate\u2019s, bachelor\u2019s) and issues of \nselection that might confound the results, and I \nexplore a range of maternal health outcomes that \nhave also been connected to children\u2019s health  \nand development (Garbarski 2014; Hofhuis, de \nJongste, and Merkus 2003; Turney 2011). The \nresults of this investigation will provide greater \nunderstanding of how sequencing education vis-\u00e0-\nvis a key life course event, the birth of a child, is \nassociated with various dimensions of health, \nextend research on the \u00adeducation\u2013health gradient, \nand inform policy debates aimed at promoting the \nwell-being of disadvantaged families.\nBackground\nEducation and the Health Gradient\nIn the United States, as in other parts of the world, \nindividuals with postsecondary levels of education \nhave a substantial health advantage. This health \nadvantage spans various health domains (e.g., phys-\nical health, mental health, health behaviors; \nMirowsky and Ross 2003), and has grown over the \npast several decades (Meara, Richards, and Cutler \n2008; Mirowsky and Ross 2008). The general \nexplanations for these patterns are that higher edu-\ncation (1) promotes the accumulation of human \ncapital, which increases access to better paying and \nless physically and emotionally demanding sectors \nof the labor market; (2) enhances cognitive skills \n(e.g., critical thinking) and capacities (e.g., self-\nesteem, stress management), which improve one\u2019s \nability to manage one\u2019s health in both the short and \nlong terms; and (3) conveys socioeconomic advan-\ntages that yield access to more supportive and stable \nrelationships, higher quality neighborhoods, and \nbetter health care and buffers against exposure to \nstressful environments and experiences (Link and \nPhelan 1995; Mirowsky and Ross 2003; Zajacova \nand Lawrence 2018).\nMoreover, these effects of education on health \nare considered to be causal (McFarland and Wagner \n2015). Although some scholars have challenged \nthis causal interpretation of the education\u2013health \ngradient (Duke and Macmillan 2016; Lynch and \nvon Hippel 2016), the prevailing view of most \nhealth scholars is that increases in one\u2019s education \nwill result in better health outcomes. This general \nnotion has inspired many scholars to argue for pub-\nlic health interventions that target the education of \ntraditionally disadvantaged groups (Frieden 2010; \nHahn and Truman 2015). Nevertheless, numerous \nquestions about the links between education and \nhealth remain, including whether an increase in \nmothers\u2019 education\u2014which would occur at a life \ncourse stage in which schooling has traditionally \nbeen completed (Marini 1984)\u2014results in an \nimprovement in their health outcomes. There is a \nstrong theoretical and empirical basis for pursuing \nthis specific question.\nChanges in the Structure of the Life \nCourse\nIn the past, the life course was more likely to follow \na \u201cnormative\u201d sequence of role transitions (Marini \n1984) and experiences. Prior research linking higher \n\n4\t\nJournal of Health and Social Behavior 62(1)\neducation to health, which tended to view education \nas a static status, reflected this understanding. \nToday, however, the life course has grown more \n\u201cdestandardized\u201d (MacMillan 2005). This phenom-\nenon is especially true for women (vs. men; Braboy \nJackson and Berkowitz 2005; Br\u00fcckner and Mayer \n2005), whose trajectories into and through adult-\nhood reflect greater diversity in the sequencing of \ntransitions into (and out of) various social roles, \nincluding the student and parent roles (Augustine \n2016). Indeed, as mentioned previously, among the \n22% of U.S. college students with dependents \nenrolled in 2015 to 2016, 70% were women (Cruse \net\u00a0 al. 2019). In the aggregate, nearly 2.7 million \nmothers were enrolled in college in 2015 to 2016 \n(Cruse et. al. 2019).\nSeveral factors explain the large number of \nmothers pursuing postsecondary schooling. The \nfirst are demographic factors. Increases in mothers\u2019 \nlabor force participation, decreases in men\u2019s labor \nforce participation, and changes in family structure, \nincluding the rise in nonmarital fertility and greater \nunion instability, have brought about a rise in the \nshare of women who are unmarried parents and the \nsole or primary breadwinners (Bianchi, Robinson, \nand Milkie 2006). A second set of economic factors \nincludes greater instability in the labor market, \nstagnating wages for lower-status work, and the \nrelative lack of adequate paying jobs for lower-\u00ad\neducated women (compared to men, who are more \nlikely to work in trades or manual labor). Thus, \ngreater education is also needed for women to \naccess stable work and adequate wages (Kalleberg \n2011). Third, there has been an expansion of the \nhigher education system\u2014which includes a rise in \nfor-profit schools, which have targeted their recruit-\nment efforts at lower-income mothers (Cottom \n2017)\u2014as well as federal student aid programs \nover the past few decades. These factors have made \nit easier for adults with fewer resources (e.g., moth-\ners with lower levels of education) to return to \nschool (Rosenbaum 2001). Outside of these large-\nscale factors, disadvantaged mothers, who are often \ntheir children\u2019s sole caregiver, also report returning \nto school in a renewed effort to improve their chil-\ndren\u2019s opportunities for mobility (Edin and Kefalas \n2011; Katz 2013).\nThe Life Course Paradigm and Key \nConceptual Principles\nDespite trends in mothers\u2019 schooling, scholars have yet \nto investigate whether mothers\u2019 postnatal college \ncompletion can result in an improvement in their health. \nThis lack of research is surprising given how lower-\neducated mothers have become a primary focus of two-\ngeneration policies and interventions, which are aimed \nat simultaneously improving the well-being of parents \nand children (Gardner et\u00a0al. 2017; Kaushal 2014). The \npromise of these interventions reflects the robust inter-\ndisciplinary body of research linking one\u2019s health and \nwell-being with one\u2019s education (Hout 2012) as well as \na key tenet of the life course: that of interlocking trajec-\ntories (Elder 1984). This tenet highlights the interde-\npendencies among parent\u2019s and children\u2019s experiences \n(Cooksey, Menaghan, and Jekielek, 1997; Yuan 2008) \nand the potential for increases in mothers\u2019 education to \nimprove the health and development of their children. \nThe concept of interlocking trajectories also highlights \nthe potential for, and importance of, examining whether \nan increase in a mother\u2019s education can improve her \nown health outcomes.\nAt the same time, an additional life course \ntenet suggests an alternate possibility that runs \ncounter to much thinking about the links between \nhigher education and health (Frieden 2010; Hahn \nand Truman 2015). The tenet of role sequencing \nhighlights how \u201cthe demands and conditions \nencountered in out-of-sequence or off-time transi-\ntions interfere with the achievement and enact-\nment of other roles and statuses\u201d (Pearlin et\u00a0 al. \n2005:212). This disruption has been called a life \ncourse penalty. Importantly, a break in schooling \nassociated with the transition to parenthood or \nneed to work is argued to carry among the greatest \nlong-term life course penalties. The tenet of role \nsequencing, therefore, suggests that mothers who \nattain higher education after having a child may \nhave difficulty realizing the returns to their degree \nand experiencing an improvement in their health. \nThere is also substantive research to support this \nexpectation.\nResearch on Out-of-Sequence Schooling \nand Maternal Health Outcomes\nThe most direct evidence comes from a handful of \nstudies in medical sociology that found that nonnor-\nmative life course sequences are linked to worse \nhealth outcomes than traditional sequences. For \nexample, one study found that adults who married \nbefore completing college had a 50% higher proba-\nbility of being obese than those who married after \ncompleting college (Miech et\u00a0al. 2015). Two other \nstudies exploring a range of life course trajectories \nfound that nonnormative trajectories were associated \n\nAugustine\t\n5\nwith lower self-reported health, higher rates of \nsmoking and drinking (Barban 2013), and worse \nmental health (Braboy Jackson 2004). A fourth study \nfound that the sequencing of motherhood prior to \nmarriage was associated with worse health at midlife \n(Williams et\u00a0al. 2011).\nResearch examining the mechanisms linking \neducation and health also indicate fewer health \nreturns to out-of-sequence schooling. Research in \nthe human capital tradition indicates that postsec-\nondary educational attainment in adulthood is \nassociated with fewer labor market opportunities \nand lower levels of wages compared to postsec-\nondary education completed earlier in the life \ncourse (Elman and O\u2019Rand 2004). Research on \ntime use and role strain finds that college student \nmothers have less time to devote to schooling than \ntheir peers without dependents and greater stress \nand fatigue than similar mothers who were not in \nschool (Augustine, Prickett, and Negraia 2018). \nResearch on social networks and support suggests \nthat student mothers have less support from family \nand friends than traditional students, less social \nengagement with their college peers, and less mal-\nleable social networks because many of their prin-\nciple relationships (e.g., their child\u2019s parent, \nneighbors) have already been established (Carney-\nCrompton and Tan 2002; Laumann 1973). Lastly, \nresearch in education suggests that student mothers \nmay be more oriented toward receiving credentials \nnecessary for specific jobs than the experiential \naspects of schooling (an emphasis of traditional stu-\ndents; McDonough 1997) or may attend lower-qual-\nity educational institutions (Seefeldt 2016). Such \nfactors may mitigate the nonpecuniary returns to \neducation, for example, its impact on lifestyle \nchoices related to health (Cockerham 2005) or \nhealth management skills.\nFinally, there is indirect evidence from studies \nexamining the link between increases in mothers\u2019 edu-\ncation and their children\u2019s outcomes (e.g., cognitive \nskills, behavior), which find no significant positive \nassociations, irrespective of potentially moderating \nfactors, such as degree type or the child\u2019s age \n(Augustine and Negraia 2018; Gardner, Martin, and \nPetitclerc 2019; Harding 2015). These findings sup-\nport the suggestion that the health returns to maternal \neducation\u2014which largely include the same set of \nhuman capital, socioeconomic, and social psychologi-\ncal factors shown to promote children\u2019s development \n(Augustine and Negraia 2018; for a detailed concep-\ntual framework, see Harding, Morris, and Hughes \n2015)\u2014may not accrue to mothers who pursed their \neducation after having children.\nCurrent Study\nGiven such factors, this study tests the hypothesis \nthat increased maternal education will not result in a \nsignificant improvement to mothers\u2019 health. In \ndoing so, I examine four maternal health measures: \nself-reported health, global mental health, parenting \nstress, and smoking. These measures represent an \narray of health domains, follow an educational gra-\ndient (Cooper et\u00a0al. 2009; Cutler and Lleras-Muney \n2010; Mirowsky and Ross 2003), and have implica-\ntions for public health as well as the health and well-\nbeing of children and the two-generation policy \nframework more broadly (Beck et\u00a0al. 2010; Kaushal \n2014; Turney 2011).\nMore specifically, self-rated health provides a \nholistic way of capturing population health dispari-\nties that encapsulates a broad range of health-related \nfactors, from disease to lifestyle (Parrish 2010). \nGlobal mental health provides a broad assessment of \npsychological functioning that reflects a variety of \npsychosocial mechanisms linked to education, such \nas social isolation or nutrition (Mirowsky and Ross \n2003). Parenting stress captures a dimension of \nstress\u2014how mothers manage and respond to daily \nparenting challenges in concert with other con-\nstraints, such as work and income (Cooper et\u00a0 al. \n2009)\u2014that has been associated with psychological \ndistress, depression, and mothers\u2019 physical health \n(Crnic, Gaze, and Hoffman 2005; Eisenhower, \nBaker, and Blatcher 2009; Gershoff et\u00a0 al. 2007). \nFinally, smoking represents an important maternal \nhealth behavior, public health indicator, and poor \nhealth risk factor (Bandeira et\u00a0 al. 2011; Hofhuis \net\u00a0al. 2003).1\nTo test these associations, I adhere to the life \ncourse principle of interlocking trajectories by \nexamining the linkages between maternal education \nand maternal health dynamically, drawing on six \npanels of data that span 15 years after the birth of a \nchild and random intercept models. I also examine \npatterns by institution type (i.e., vocational, associ-\nate\u2019s, bachelor\u2019s) and the robustness of any signifi-\ncant findings to unobserved heterogeneity using a \nfixed effects modeling approach.\nData and Methods\nData\nData came from the Fragile Families and Child \nWellbeing Study: a national birth cohort study of 4,898 \nchildren born between 1998 and 2000 in 20 U.S. cities. \nMothers were recruited from hospitals immediately \nafter the study child\u2019s birth and interviewed about \n\n6\t\nJournal of Health and Social Behavior 62(1)\nfamily background and demographic characteristics, \nincluding highest level of schooling (for information \nabout the sample design, see Reichman et\u00a0al. 2001). \nFollow-up interviews conducted one (Wave 2), 3 \n(Wave 3), 5 (Wave 4), 9 (Wave 5), and 15 (Wave 6) \nyears after the birth using telephone and in-home sur-\nveys provided data on maternal health factors and \nmothers\u2019 schooling since the last interview. The avail-\nability of repeated measurements of health and school-\ning across six data collections spanning 15 years make \nthe FFCW data particularly well suited to this study. A \nsecond advantage is the FFCW largely sampled \nwomen with lower levels of education at the baseline \nwave (89% was without a bachelor\u2019s degree) who \nwere more likely to pursue higher education subse-\nquent to a birth. The FFCW sample is also regionally, \nracially, and ethnically diverse, although not represen-\ntative of the entire U.S. population. More information \nabout the sociodemographic characteristics of the \nsample appear in Table 1 (presenting baseline/Wave 2 \nreports) and Appendix A in the online version of the \narticle (reports from each wave).\nThe analytic sample for the study was based on \nthe full sample of families enrolled in the FFCW  \n(N = 4,898), although modeling procedures \n(explained in the following) limited the \u00admultivariate \nanalyses to mothers with at least one health report \n(ns = 4,710 for models predicting overall health; \n4,697 for \u00adparenting stress; 4,709 for global mental \nhealth; and 4,689 for smoking). Missing data impu-\ntation techniques (described in the \u00adfollowing as \nwell) were used to retain mothers with missing \n\u00advalues on independent variables.\nVariables\nMaternal education.\u2002 Mothers\u2019 reports of baseline \neducation were coded into one of four categories: \nless than a high school degree, high school degree/\nGeneral Education Development (GED), some col-\nlege, bachelor\u2019s or higher. Unfortunately, the base-\nline reports did not differentiate mothers with \u201csome \ncollege\u201d from those with an associate\u2019s degree. It \ndid distinguish mothers with a technical/trade \ndegree, although this group was small (6% of the \nsample). Thus, mothers with technical/trade degrees \nwere included in the \u201csome college group,\u201d but I did \nuse this information in models that restrict the sam-\nple based on mothers\u2019 baseline education (explained \nin the following). At each follow-up wave, women \nreported whether they had earned a higher educa-\ntion degree among a variety of institution types \n(e.g., \ntechnical \ndegree, \nsecretarial \nschool, \nassociate\u2019s degree, etc.) since the last interview. \nUsing this information, I created three time-varying, \nbinary variables for completed a vocational pro-\ngram, completed an associate\u2019s degree, and com-\npleted a bachelor\u2019s degree. Once mothers reported \ndegree completion, mothers were assigned a value \nof 1 on that measure for all subsequent waves. This \ncoding technique captured a pooled estimate of the \nassociation between additional education and health \nthat used all available data and provided an average \nof mothers\u2019 health before versus after completing \nadditional education. Mothers were also asked if \nthey had enrolled in school since the last interview, \nbut only where they had enrolled during Waves 2 to \n5. I used this information to create a parallel set of \nbivariate measures of enrolled in a vocational pro-\ngram, enrolled in an associate\u2019s program, and \nenrolled in a bachelor\u2019s degree program for ancil-\nlary analyses exploring the impact of school enroll-\nment (restricted to Waves 2\u20135).\nMaternal health.\u2002 Mothers\u2019 health reflected four \ndomains. Overall health was based on a measure of \nself-reported health in which respondents described \ntheir health on a scale ranging from 1 (poor health) \nto 5 (excellent health). Global mental health was \ncaptured by mothers\u2019 reports of whether they felt \nsad, depressed, or anxious for a two-week period \nover the past year (1 = yes, 0 = no). Unfortunately, \nthe FFCW did not include more nuanced measures \nlike depressive symptomatology based on the Cen-\nter for Epidemiological Studies\u2013Depression Scale. \nIt did include an indicator of depression based on \nthe Composite International Diagnostic Indicator\u2013\nShort Form (CIDI\u2013SF), but the CIDI\u2013SF is a fairly \nspecific tool designed to capture one\u2019s probability of \nmeeting clinical levels of depression. The global \nindicator used here is seen as a less restrictive mea-\nsure that taps a wider range of mental health strug-\ngles. Parenting stress, which was based on the \naverage of mothers\u2019 responses (on a scale of 1 to 4 \ntapping different levels of agreement) to four ques-\ntions (e.g., I often feel tired, worn out, or exhausted \nfrom raising a family), was a shortened version of \nthe original Parenting Stress Index (see Abidin \n1983), as used in other national studies (e.g., the \nPanel Study of Income Dynamics). Smoking was \nbased on self-reports of whether the mother smoked \nin the last month (1 = yes, 0 = no). All four measures \nwere assessed at Waves 2 to 6. More details about \nthe univariate properties of these variables as well \nas the time-varying education variables appear in \nTable 2.\n\nAugustine\t\n7\nCovariates.\u2002 To account for sources of confound-\ning among mothers\u2019 education and health, I included \nseveral covariates. The first set of factors (subse-\nquently referred to as Set 1) included background \nfactors exogenous to mothers\u2019 baseline education. \nThese included the mothers\u2019 race-ethnic background \n(dummy coded white, Black, Hispanic, and other), \nage at first birth (continuous), current age (continu-\nous), her mother\u2019s education (dummy coded as less \nthan high school, high school degree, more than high \nschool), immigrant status (1 = immigrant, 0 = native \nborn), whether she grew up with both parents in the \nhome (1 = yes, 0 = no), and whether the study child \nwas her first birth (1 = yes, 0 = no). The second set \n(Set 2) included time-invariant factors taken from \nthe baseline or, when not available at baseline, Wave \n2, potentially exogenous to mothers\u2019 school reentry. \nThese included family structure (dummy coded \n\u00admarried, cohabiting, and single), family income-to-\nneeds (continuous), number of children in the home \nTable 1.\u2002 Means/Frequencies of Background Characteristics and Selected Baseline Characteristics of \nMothers in the Fragile Families and Child Wellbeing Study Sample (N = 4,898).\nMeans/Frequencies\nDemographic characteristics\n\u2003 Maternal age\n25.27 (6.04)\n\u2003 Age at first birth\n21.56 (5.24)\n\u2003 Race\n\u2003 \u2003 White\n21%\n\u2003 \u2003 Black\n48%\n\u2003 \u2003 Hispanic\n27%\n\u2003 \u2003 Other race\n4%\n\u2003 Mother immigrant status\n17%\nBackground characteristics\n\u2003 Grandmother education\n\u2003 \u2003 Less than high school\n21%\n\u2003 \u2003 High school\n54%\n\u2003 \u2003 More than high school\n25%\n\u2003 Mother lived with both parents\n43%\nFamily/child factors\n\u2003 Focal child mother\u2019s first birth\n39%\n\u2003 Family structure\n\u2003 \u2003 Married\n24%\n\u2003 \u2003 Single\n40%\n\u2003 \u2003 Cohabiting\n36%\n\u2003 Child low birth weight\n10%\nFamily resources\n\u2003 Family income-to-needs\n2.24 (2.41)\n\u2003 Mother employment\n\u2003 \u2003 Not employed\n14%\n\u2003 \u2003 Part-time work\n32%\n\u2003 \u2003 Full-time work\n46%\n\u2003 Welfare receipt\n36%\n\u2003 Health insurance\n71%\n\u2003 Emergency child care\n92%\n\u2003 Can borrow $200\n90%\n\u2003 Owned home\n34%\nNote: Due to missing data, sample sizes range between 4,842 and 4,897 for all variables, with the exception of age \nat first birth (N = 4,311), grandmother\u2019s education (N = 3,993), and health insurance (N = 4,337), which were taken \nfrom the Wave 2 reports in which 534 cases were not in the wave.\n\n8\t\nJournal of Health and Social Behavior 62(1)\n(continuous), maternal employment status (dummy \ncoded not working, working part-time, and working \nfull-time), child low birth weight (0 = no, 1 = yes), \nthe family received public assistance (0 = no, 1 = \nyes), the mother had insurance coverage (0 = no, 1 = \nyes), the mother owned her home (0 = no, 1 = yes), \nand two indicators of social support based on moth-\ners\u2019 responses to whether (a) she had friends or fam-\nily available that she could borrow $200 from (0 = \nno, 1 = yes) and (b) an emergency child care option \n(0 = no, 1 = yes). A third set of factors (Set 3) included \ntime-varying factors that may be exogenous to moth-\ners\u2019 additional education but also endogenous to it. \nThese factors were included to assess the robustness \nof the links between mothers\u2019 postsecondary degree \ncompletion and health. They included the child\u2019s \nhealth (range = 1\u20135, 1 = poor, 5 = excellent); whether \nthe mother had a birth between waves (0 = no, 1 = \nyes), coresided with a parent or in-law (0 = no, 1 = \nyes), or moved since the previous interview (0 = no, \n1 = yes); and time-varying versions of Set 2 variables \n(except child low birth weight and home ownership, \nwhich was a proxy for residential stability and cap-\ntured by the measure of having recently moved).\nAnalysis Plan\nThe main analytic technique involved random inter-\ncept models. These models took advantage of the \nlongitudinal data by incorporating all six waves, \naccounted for the hierarchical structure of the data in \nwhich individual observations are nested within \ntime, adjusted for nonindependence among the \nrepeated observations, allowed for the modeling of \ntime as a fixed effect (thereby capturing time trends), \nmodeled time-invariant key variables, like maternal \neducation at the baseline wave, and accounted for \nrandom variability in mothers\u2019 health (i.e., the ran-\ndom intercept; Laird and Ware 1982). This model in \nits unconditional form is expressed as follows:\nTable 2.\u2002 Descriptive Statistics for Main Study Variables by Wave and Pooled across Waves, Fragile \nFamilies and Child Wellbeing Study.\nMean/% (n)\nWithin\nBetween\n\u2002\nWave 2\nWave 3\nWave 4\nWave 5\nWave 6\nSD/%\nSD/%\nHealth measures\n\u2003 Self-rated health\n3.77 (1.06)\n3.75 (1.05)\n3.64 (1.03)\n3.55(1.04)\n3.42 (1.05)\n.61\n.81\n\u2003 \u2003 n\n4,362\n4,225\n4,131\n3,512\n3,144\n\u2002\n\u2003 Global mental health\n20%\n25%\n22%\n22%\n24%\n46%\n50%\n\u2003 \u2003 n\n4,357\n4,215\n4,125\n3,497\n3,135\n\u2002\n\u2003 Parenting stress\n2.18 (.67)\n2.25 (.67)\n2.18 (.69)\n2.03 (.69)\n2.06 (.70)\n.44\n.54\n\u2003 \u2003 n\n3,787\n4,205\n4,100\n3,616\n3,145\n\u2002\n\u2003 Smoking\n27%\n24%\n31%\n30%\n23%\n72%\n38%\n\u2003 \u2003 n\n4,361\n3,129\n4,131\n1,032\n732\n\u2002\nWave-specific education\n\u2003 Vocational degree\n5%\n5%\n7%\n13%\n11%\n31%\n26%\n\u2003 \u2003 n\n4,360\n4,221\n4,126\n3,509\n3,571\n\u2002\n\u2003 Associate\u2019s degree\n1%\n1%\n2%\n4%\n6%\n26%\n10%\n\u2003 \u2003 n\n4,360\n4,221\n4,126\n3,509\n3,571\n\u2002\n\u2003 Bachelor\u2019s degree\n1%\n1%\n1%\n2%\n4%\n26%\n7%\n\u2003 \u2003 n\n4,360\n4,221\n4,126\n3,509\n3,571\n\u2002\nCumulative education\n\u2003 Vocational degree\n5%\n9%\n15%\n26%\n33%\n59%\n26%\n\u2003 \u2003 n\n4,360\n4,221\n4,126\n3,509\n3,571\n\u2002\n\u2003 Associate\u2019s degree\n1%\n2%\n4%\n8%\n13%\n48%\n10%\n\u2003 \u2003 n\n4,360\n4,221\n4,126\n3,509\n3,571\n\u2002\n\u2003 Bachelor\u2019s degree\n1%\n2%\n3%\n6%\n9%\n47%\n7%\n\u2003 \u2003 n\n4,360\n4,221\n4,126\n3,509\n3,571\n\u2002\nNote: Within percentages describes among mothers who ever reported a 1 on a variable (e.g., low mental health) the \npercentage of reports in which they reported a value of 1. The between percentages captures the percentage of all \nsample reports that were a 1 on that measure.\n\nAugustine\t\n9\nYij =\n0i\ni\n1i\nj\nij\n+ U\n+ \nwave + \n\u03b2\n\u03b2\n\u03b5\n(\n)\n(\n)\n\u03b20 represents the intercept for respondent i at Wave \n2, and \u03b21 is the coefficient for study wave j ranging \nfrom 2 to 6. The \u03b2 coefficients reflect the fixed \neffects, or parameters of central values (i.e., popula-\ntion averages) observed for each mother. Ui repre-\nsents the random intercept, which captures \nindividual-level deviations from the central values. \nRandom effects are assumed to be normally distrib-\nuted, with a mean of zero (Ghisletta et\u00a0al. 2015). The \n\u03b5ij term is the random within-person error of predic-\ntion for respondent i at time j. This term is assumed \nto be reflective of the random effects, which are \nassumed to be independent of the vector of covari-\nates (Singer and Willet 2003).\nAs an initial step, I explored variability in mater-\nnal health across waves. First, for all four outcomes, \nI calculated the within- and between-person vari-\nance components. This step revealed that variances \nfor overall health and parenting stress were similar \nin magnitude within mothers as between mothers \n(results in Table 2; estimates reflected standard \ndeviations). For global mental health, the within \nestimate suggested that of the mothers who reported \na mental health struggle, mothers reported a mental \nhealth struggle in 46% of their reports. Among \nmothers who smoked, this number was 72%. \nSecond, transition probabilities provided a different \nperspective on these two categorical outcomes. \nThey revealed that 22% of women who smoked at \none wave did not smoke at a later wave, and 52% of \nmothers who reported a mental health struggle the \npast year did not do so in a later wave. These mod-\nels indicated the potential for health to change \nacross time and provided the foundation for the \nsubsequent modeling steps.\nThese steps involved (1) providing evidence of \nthe link between a static measure of maternal educa-\ntion (baseline education) and maternal health as \nfound in previous research and (2) building on this \nmodel by examining whether, net of this association, \na postsecondary degree completed after the birth of \nthe focal child is associated with more optimal \nhealth outcomes. Thus, I first extended the uncondi-\ntional model to incorporate parameters for baseline \nmaternal education (\u03b23 Educi) and X\u2032i, which is a \nvector of the Set 1 background factors (i.e., those \nthat may select mothers into different baseline levels \nof education). This step laid the foundation for the \nmain analysis, which incorporated the time-varying \nmeasures for increased education (i.e., \u03b24 Additional \nEducationij).\nFor this step, I estimated a different model for each \nadditional education measure (e.g., vocational, associ-\nate\u2019s), predicting the four health outcomes, for a total \nof 12 models. Model coefficients for the additional \neducation variables are interpreted as the difference in \nthe average of mothers\u2019 health across all waves in \nwhich mothers had completed an additional degree \ncompared to mothers\u2019 health in waves for which no \nadditional education was reported. To ensure I was \ncapturing an educational increase (e.g., some mothers \nwith a vocational degree completed a second voca-\ntional degree; see Table 3), I made restrictions based \non the mothers\u2019 baseline education. For models esti-\nmating the \u201ceffect\u201d of a postnatal vocational degree, \nthe sample was restricted to mothers with a high \nschool degree or less at the baseline wave. Models \nincluding associate\u2019s degree completion were restricted \nto mothers with a technical/trade degree or less. I \nmade this restriction, rather restricting the sample to \nmothers\u2019 with some college, because this group \nincluded mothers with associate\u2019s degrees at baseline, \nas explained previously. I did examine models using \nthis alternative restriction, however. Models measur-\ning bachelor\u2019s degree completion were restricted to \nmothers with some college (which includes those with \nassociate\u2019s degrees) or less. All sample sizes are \nreflected in the tables (see Table 6). Here, I also added \nin Set 2 covariates to address sources of selection into \nadditional degree completion.\nTo examine the robustness of any significant asso-\nciations between maternal health and the time-varying \nmeasures of mothers\u2019 postnatal college attainment to \nunobserved, fixed sources of heterogeneity (e.g., \n\u00adpersonality trait, genetic factor), as well as potentially \nconfounding time-varying factors, I turned to fixed \neffects modeling techniques. This approach provided \nestimates of within-person residualized change, \nthereby addressing a key limitation of random intercept \n\u00admodels\u2014the assumption that all unobserved fixed \nsources of heterogeneity were orthogonal to the predic-\ntors (Allison 2009). As a class of hierarchical models, \nthey also accommodated the longitudinal structure of \nthe data, including time-varying Set 3 confounds. As a \nfinal analysis step, I conducted a series of robustness \nanalyses that addressed the potential for population \nheterogeneity based on mothers\u2019 race-ethnicity, parity \n(i.e., number of live births) at baseline, and education at \nbaseline; that better tease out issues of lagged effects \nand bidirectionality; and that explore additional health \noutcomes and the potential for indirect effects.\nModels were estimated using Stata 15.1. To \naccount for modest missingness on the independent \nvariables (see Appendix A in the online version of \nthe article), multiple imputation (with 20 iterations) \n\n10\t\nJournal of Health and Social Behavior 62(1)\nvia chained equations was performed. Dependent \nvariables were part of the imputation model but not \nimputed (von Hippel 2007). To analyze the multi-\nply imputed data, the Stata mi estimate suite of \ncommands was used. Random intercept models \npredicting continuous outcomes (overall health, \nparenting stress) were estimated using the xtreg, re \nprocedure. The xtlogit, re procedure was used for \ncategorical outcomes (global mental health, smok-\ning). Assuming data are missing at random (Allison \n2012), these procedures also helped address cases \nwith missingness on the dependent variable (i.e., \nmaternal health) in which there was at least one \nobservation through the use of maximum likelihood \nestimation. For the fixed effects models, the xtreg, \nfe (for continuous outcomes) and clogit (for cate-\ngorical) procedures were used.\nResults\nA Descriptive Overview of Maternal \nEducation and Its Linkage to Maternal \nHealth\nTable 3 provides a description of mothers\u2019 education \nat the baseline and postnatal college completion \nbetween Waves 1 and 6. At the baseline wave, 40% \nof the FFCW sample did not have a high school \ndegree, 25% reported a high school degree or GED, \n24% reported some college (i.e., a vocational degree, \nassociate\u2019s degrees, or college schooling that did not \nresult in a degree), and 11% had a bachelor\u2019s degree. \nBetween the baseline interview and sixth wave, 33% \n(n = 1,099) of the Wave 6 sample (n = 3,146 \u00admothers) \nhad completed a vocational degree, 13% (n = 417) \nhad completed an associate\u2019s degree, and 9%  \n(n = 295) had completed a bachelor\u2019s degree. Among \nmothers who completed a vocational degree, 67% of \ncases (n = 733) resulted in an increase in education. \nAmong mothers who completed an associate\u2019s \ndegree, 54% of cases (n = 224) resulted in an observ-\nable increase in education (this estimate is likely \nconservative because of how some college is coded). \nAmong mothers who completed a bachelor\u2019s degree, \nthere was an increase in education in 85% of cases (n \n= 251). More information about the number and per-\ncentage of mothers (based on the Wave 6 sample) \nwho earned various degrees by their starting level of \neducation (thus providing details on cell sizes) is \nalso presented in Table 3.\nTo provide a preliminary picture of the link \nbetween maternal health and education, Table 4 \nTable 3.\u2002 Frequencies and Percentages of Maternal Education at Baseline and Increases in Education \nthrough Wave 6 in the Fragile Families and Child Wellbeing Study Sample.\nBaseline Sample\nCompletion by Wave 6a\n\u2002\n(n = 4,898)\n(n = 3,286)\n(n = 3,182)\n(n = 3,162)\n\u2002\nVocational Degree\nAssociate\u2019s Degree\nBachelor\u2019s Degree\n1. Frequencies\nEducation at baseline\n\u2003 Less than high school\n1,949\n413\n94\n27\n\u2003 High school degree\n1,237\n320\n130\n53\n\u2003 Some college\n1,189\n313\n185\n171\n\u2003 College or more\n524\n53\n8\n43\nTotal\n4,898\n1,099\n417\n295\n2. Percentagesb\nEducation at baseline\n\u2003 Less than high school\n40%\n13%\n3%\n<1%\n\u2003 High school degree\n25%\n10%\n4%\n<2%\n\u2003 Some college\n24%\n10%\n6%\n5%\n\u2003 College or more\n11%\n<2%\n<1%\n<2%\nTotal\n33%\n13%\n9%\naWave 6 measure based on an aggregate of all reports of additional education, within degree types, across waves \namong mothers in Wave 6. Thus, sample sizes vary slightly across columns because of differential rates of dropout \nand reentry into the sample across waves among mothers who completed different degrees.\nbReflects percentages of mothers who earned each degree by their baseline education among mothers remaining in \nthe sample at Wave 6.\n\nAugustine\t\n11\npresents the bivariate associations between baseline \nmaternal education and the four maternal health \noutcomes. Pairwise estimates of statistical signifi-\ncance between different levels of education were \nperformed using one-way analysis of variance. \nMaternal health reflects an average of the pooled \nestimates across the five study waves. Results dem-\nonstrate the educational gradient in maternal health. \nAs mothers\u2019 baseline level of educational attain-\nment increased, their overall self-rated health \nincreased, and their reports of parenting stress, hav-\ning poor mental health, and having recently smoked \ndecreased. The rate of smoking was four times \ngreater for mothers with less than a high school \ndegree than for mothers with a bachelor\u2019s degree \n(36% vs. 8%). Differences in mental health were \nalso noteworthy, albeit smaller (26% vs. 14%). \nPatterns for overall health represented a 55% of a \nstandard deviation (overall SD = 1.05) difference \nbetween mothers with the highest and lowest levels \nof education. Patterns for parenting stress amount \nto a 15% of a standard deviation (overall SD = .69) \ndifference. Notably, differences among those with a \nhigh school degree and some college were not as \npronounced.\nMultivariate Analysis of Maternal \nEducation and Health\nFor the multivariate analyses, I first estimated an \nunconditional (i.e., without predictors) random \nintercept model to assess the functional form of \nmaternal health (Model 1). This step revealed a time \ntrend in three of four health outcomes that was lin-\near in form for global mental health (i.e., increased \nover time) and quadratic in form for overall health \nand parenting stress (both increased over time, but \nthe rate of increase decelerated at later waves). This \nquadratic term was thus retained in subsequent \nmodels. These coefficients, along with other param-\neters of interest (e.g., variance in within person \nresiduals, intraclass correlation), appear in Appendix \nB in the online version of the article.\nNext, I added baseline maternal education to the \nmodel, with the high school degree/GED as the ref-\nerence category, along with the Set 1 background \ncovariates (see Table 5, Model 2). These results \nwere similar to the patterns presented in the bivari-\nate analyses. Compared to mothers with high school \ndegrees or GEDs, mothers with bachelor\u2019s degrees \nhad higher overall health and were less likely to \nreport a mental health issue or smoking. Conversely, \nTable 4.\u2002 Pooled Means/Frequencies and Contrasts of Maternal Health Outcomes by Maternal \nEducation at Baseline based on One-Way ANOVA with Data from Fragile Families and Child Wellbeing \nStudy, Waves 2 through 6.\nMeans (SEs)/Frequencies of Maternal Health Outcomes\n\u2002\nSelf-Rated Overall \nHealth\nParenting Stress\nGlobal Mental \nHealth\nSmoking\nBaseline education level\n\u2003 Less than high school\n3.49c\n(.02)\n2.22a\n(.01)\n27%a\n36%a\n\u2003 High school degree\n3.63b\n(.01)\n2.11bc\n(.01)\n21%b\n27%b\n\u2003 Some college\n3.68b\n(.01)\n2.07c\n(.01)\n21%b\n22%c\n\u2003 College degree\n4.07a\n(.02)\n2.12b\n(.01)\n14%c\n8%d\nFull sample (SDs)\n3.64\n(1.05)\n2.14\n(.69)\n22%\n27%\nn observations\n19,374\n18,868\n19,329\n18,275\nNote: Estimates based on pooled values across Waves 2 through 6. Different superscripts indicate statistically \nsignificant differences in health (with a p value of at least .05) between education levels. For example, for self-rated \noverall health, mothers with college degrees (superscript \u201ca\u201d) have significantly different levels of overall health \ncompared to mothers with some college education (superscript \u201cb\u201d), but mothers with some college education do \nnot have significantly different levels compared to the high school group (also superscript \u201cb\u201d). The \u201ca\u201d superscript \nindicates the largest value. The \u201cd\u201d superscript indicates the smallest value.\n\n12\t\nJournal of Health and Social Behavior 62(1)\nmothers without a high school diploma had worse \noverall health and higher rates of poor mental \nhealth and smoking. They also had more parenting \nstress. Differences among the high school and some \ncollege groups were nonsignificant, with the excep-\ntion of smoking, with lower rates of smoking \namong the some college group. Education also did \nnot significantly differentiate parenting stress levels \namong mothers with a high school degree/GED or \nhigher, although patterns across all outcomes are \nconsistent with the gradient.\nExamining the Health Impact \nof Mothers\u2019 Higher Educational \nAttainment\nHaving established these patterns, I examined \nwhether, net of these associations, postsecondary \neducation acquired after a birth can improve moth-\ners\u2019 average levels of health. In doing so, I added \nthe Set 2 covariates (e.g., Wave 1 family income), \nwhich were excluded from prior analysis because \nthey were likely to be endogenous to mothers\u2019 base-\nline education. The results of this modeling step \n(see Table 6, Models 3\u20135) revealed few significant \nassociations between the time-varying measures \nfor postsecondary education and maternal health \n(10 of 12 models). There were two exceptions. \nAmong mothers with a technical/trade degree or \nless at baseline, those who earned an associate\u2019s \ndegree reported significantly higher levels of over-\nall health (B = .11, SE = .05; these patterns were the \nsame when restricting the sample to mothers in the \nsome college category). Second, mothers with \nsome college or less at the baseline wave who sub-\nsequently completed a bachelor\u2019s degree reported \nsignificantly lower likelihoods of smoking (B = \n\u2013.62, SE = .26; odds ratio = .54). These estimates, \nhowever, remain vulnerable to omitted variable \nbias.\nThus, as a final step, I used fixed effects to tease \nout time-invariant unobserved sources of heteroge-\nneity. For this analysis, I followed an iterative pro-\ncess that incorporated the time-varying covariates \nTable 5.\u2002 Random Intercept Models Predicting Maternal Health by Maternal Education at Baseline with \nData from Fragile Families and Child Wellbeing Study, Waves 2 through 6.\nB (SE) of Maternal Health Outcomes\n\u2002\nSelf-Rated Overall \nHealth\nParenting Stress\nGlobal Mental \nHealth\nSmoking\nModel 2\nMain parameters\n\u2003 Intercept\n3.89***\n(.10)\n2.16***\n(.07)\n\u20132.17***\n(.25)\n\u20132.85***\n(.58)\n\u2003 Wave\n.04\n(.03)\n.05**\n(.02)\n.04\n(.02)\n.97***\n(.13)\n\u2003 Wave2\n\u2212.02***\n(.00)\n\u2212.01***\n(.00)\n\u2014\n\u2212.12***\n(.01)\nMother education baselinea\n\u2003 Less than high school\n\u2212.15***\n(.03)\n.13***\n(.02)\n.56***\n(.09)\n1.85***\n(.18)\n\u2003 Some college\n.03\n(.03)\n\u2212.04\n(.02)\n\u2212.04\n(.10)\n\u2212.82***\n(.19)\n\u2003 Bachelor\u2019s degree\n.42***\n(.05)\n.02\n(.03)\n\u2212.69***\n(.14)\n\u22124.00***\n(.35)\nn observations\n19,374\n18,868\n19,329\n18,275\nn groups\n4,710\n4,697\n4,709\n4,689\nNote: Models include random intercept and control for maternal race, mother age, mother age at first birth, mother \nimmigrant status, whether the current birth was the mother\u2019s first, whether the mother lived with both parents \nthrough age 15, and the education of the mother\u2019s mother. Model 1 results available in the online version of the \narticle. Intraclass correlation, variance of within person residuals, and variance of residuals are not reported.\naReference group = high school/GED.\n***p < .001.\n\nAugustine\t\n13\nin a step-wise fashion. Factors less likely to be \nmechanisms of mothers\u2019 postsecondary attainment \n(e.g., child health, access to emergency childcare) \nwere entered first. Factors more likely to be mecha-\nnisms (e.g., family income, maternal paid work) \nwere entered last. Following this approach and \nspecifying fixed effects, the inclusion of the first \ncovariate, child overall health, attenuated the asso-\nciation between mothers\u2019 associate\u2019s degree com-\npletion and overall health and between mothers\u2019 \nbachelor degree completion and reduced smoking \nto nonsignificance.\nAdditional Robustness Analyses\nGiven the lack of significant findings, I performed \nseveral auxiliary analyses to determine whether a \nsignificant association may emerge under alternate \nconditions. These included (1) among certain sub-\npopulations, including mothers for whom the focal \nchild was their first birth and mothers who were not \nracial-ethnic minorities; (2) for other health out-\ncomes, including the CIDI\u2013SF measure of depres-\nsion, reports of heavy drinking, and body mass \nindex; or (3) using measures of school enrollment. \nNone of the models produced significant findings.\nI also examined (4) heterogeneity in the \u201ceffects\u201d \nof a postsecondary degree based on mothers\u2019 base-\nline levels of education by adding interactions \nbetween baseline education and the time-varying \nadditional education measures in the fixed effects \nmodels. These analyses did not reveal heterogeneity \nbased on baseline education, with one exception. \nThere was a marginally significant association \nbetween bachelor\u2019s completion and reduced smok-\ning among mothers with both a high school degree/\nGED and less than a high school degree at baseline. \nGiven that these were small groups (ns = 27 and 53, \nrespectively) and low in statistical power, I reesti-\nmated the model for just this group. This model sug-\ngested that a bachelor\u2019s degree for this group was \nsignificantly associated (p < .01) with a 16% reduc-\ntion in the odds of smoking. Due to limited cell \nsizes, I could not examine the cumulative effect of \nmultiple postnatal degrees.\nAdditionally, I considered (5) the potential for a \n\u201clagged\u201d effect by regressing health at Wave 6 on \nmeasures on additional education measures from \nWave 5 net of the Set 1 and 2 covariates. These \nmodels did not yield significant results. Finally, I \nexplored (6) the potential for indirect effects, focus-\ning on income, which is likely to be a main driver in \nmothers\u2019 decisions to return to school. These analy-\nses used mediation tests in structural equation mod-\neling that extended the lagged models. In short, \nwhile income at Wave 6 was significantly associ-\nated with all four concurrent health outcomes, it did \nnot significantly mediate the link between voca-\ntional degree or associate\u2019s degree attainment and \nmaternal health. It did significantly mediate the \nTable 6.\u2002 Random Intercept Models, Predicting Maternal Health by Postsecondary Attainment, Net \nof Baseline Education and Restricted by Education at Birth with Data from Fragile Families and Child \nWellbeing Study, Waves 2 through 6.\nB (SE) of Maternal Health Outcomes\n\u2002\nSelf-Rated \nOverall Health Parenting Stress\nGlobal Mental \nHealth\nSmoking\nDegree completion\nModel 4: Vocational degree\n\u2212.01\n(.03)\n\u2212.03\n(.02)\n.01\n(.09)\n.28\n(.14)\n\u2003 n groups =\n3,050\n3,039\n3,049\n3,035\nModel 5: Associate\u2019s degree\n.10*\n(.05)\n\u2212.04\n(.03)\n\u2212.19\n(.16)\n\u2212.23\n(.25)\n\u2003 n groups =\n3,197\n3,186\n3,196\n3,183\nModel 6: Bachelor\u2019s degree\n.08\n(.05)\n.02\n(.03)\n\u2212.16\n(.17)\n\u2212.62*\n(.27)\n\u2003 n groups =\n4,202\n4,189\n4,201\n4,182\nNote: Models 4 through 6 control for all variables in Model 3 as well as baseline/Wave 2 measures of family structure, \nfamily income-to-needs, work status, child birth weight, insurance coverage, emergency child care, access to financial \nsupport, owned home, and number of children in the home.\n*p < .05.\n\n14\t\nJournal of Health and Social Behavior 62(1)\nassociation with Wave 5 bachelor\u2019s degree attain-\nment, but the size of the indirect effects were small \n(see Appendix C in the online version of the arti-\ncle). Taken in combination with the small coeffi-\ncient size and insignificant direct effect for \nbachelor\u2019s degree completion, these results suggest \nthat this small positive mediating effect of greater \nincome is likely offset by other negative mediating \nfactors (for a methodological discussion, see Zhao, \nLynch, and Chen 2010).\nDiscussion\nThis study aimed to understand the health impact of \na particular nonnormative process (in terms of how \neducation has conventionally been, and still consid-\nered to be, completed and reflected in how educa-\ntional institutions are structured and traditional \nstudents are defined): women\u2019s pursuit of higher \neducation after the birth of a child. A rich body of \nscholarship linking education to health suggests that \nmothers who return to school to complete a higher \neducation degree should experience improved \nhealth outcomes, yet this question has not been \ndirectly investigated. Moreover, life course theories \nthat point to the \u201clife course penalties\u201d associated \nwith off-time schooling suggest that mothers may \nnot experience the same health returns to higher \neducation as women who completed higher educa-\ntion prior to having children. As such, I tested the \nexpectation that mothers\u2019 postnatal schooling will \ndo little to alter their health. Doing so has theoreti-\ncal relevance to understanding a key social deter-\nminant of health, education, as well as tremendous \npolicy importance given the increasing health pol-\nicy emphasis on investments in higher education \nand two-generation policies that target lower edu-\ncated mothers (Gardner, Brooks-Gunn, and Chase-\nLansdale 2017; Kaushal 2014).\nThe results widely supported my expectation. \nOverall, mothers\u2019 pursuit of higher education did \nnot improve their health. This pattern held true \nacross a variety of outcomes that reflected physical, \nmental, and behavioral aspects of health; popula-\ntion subgroups; and measures of higher education \nthat reflected increased years and degrees and insti-\ntution types. Although these findings stand in con-\ntrast with the general consensus of much of the \nresearch and theory linking education to health, \nthey are consistent with several recent studies that \nhave explored the connections between nonnorma-\ntive educational trajectories and health outcomes \n(e.g., Barban 2013; Miech et al 2015). This bur-\ngeoning literature highlights the need for more \nnuanced treatments of the linkage between educa-\ntion and health and raises critical questions for the \nunderstanding of theories informing this body of \nresearch. Specifically, why does sequencing the \ncompletion of schooling after the transition to \nmotherhood diminish the power of educational \nattainment to affect one\u2019s health?\nI offer some tentative explanations. First, the \nsocial aspects of mothers\u2019 lives\u2014where they live, \nwho they interact with, whether they are married\u2014\nare \u201csticky\u201d features of people\u2019s lives that may \nremain largely unchanged after having children \n(Laumann 1973). Second, the circumstances \naround mothers\u2019 births may have enduring nega-\ntive health effects; for example, women with unin-\ntended births experience higher risks for depression \nacross the life course (Herd et\u00a0al. 2016). Third, the \nnonpecuniary returns to higher education\u2014the \nskills and capacities that help one manage their \nhealth (see Mirowsky and Ross 2003)\u2014may be \nless for mothers if they have less time to devote to \nschool work; have less time for and interest in \nextracurricular activities (Bloom 2007); suffer \nlower levels of emotional well-being due to the \nchallenges of balancing work, family, and school-\ning (Augustine et\u00a0al. 2018); and receive less sup-\nport from family and friends (Carney-Crompton \nand Tan 2002). Fourth, the labor market returns for \nmothers may be small because they are penalized \nfor their age or lack of work experience, lack \naccess to the types of networks that connect them \nto good jobs, or had enrolled in institutions like \nfor-profit schools that yield fewer labor market \nopportunities and result in greater debt than non-\nprofit schools (Deming, Goldin, and Katz 2013; \nElman and O\u2019Rand 2004; Felmlee 1988). Lastly, \nmothers who earn a bachelor\u2019s degree may experi-\nence increased income, as suggested by the media-\ntion models, but these benefits may be offset by \nother factors, such as the loss of social support as a \nresult of mothers\u2019 changed socioeconomic position \n(Wilson 2008), heightened time crunches due to \nwork demands, and many of the factors described \npreviously, such as student debt. Adjudicating \namong these explanations is beyond the scope of \nthe current study but are important topics for future \nresearch. A distinct explanation that should also be \nconsidered reflects the etiology of health and the \nquestion of whether and to what extent health is set \nearly in the life course (Conti et\u00a0al. 2010; Harris \net\u00a0al. 2013).\nAs to social policy, these findings suggest that \nefforts aimed at improving maternal health may be \nmore efficacious if they target women prior to \n\nAugustine\t\n15\ntransitioning into adult roles. This does not mean \nthat additional education cannot help some mothers. \nRather, providing mothers with educational oppor-\ntunities may not lead to reductions in population-\nlevel disparities in maternal health\u2014at least not \nuntil more knowledge is acquired to help scholars \nidentify and intervene in the factors that explain \nwhy the health benefits of additional education for \nmothers are low. By the same token, there was one \nexception to the pattern of results. Completion of a \nbachelor\u2019s degree among mothers who had no col-\nlege schooling at baseline led to a decrease in smok-\ning. Although speculative, this pattern may be \nexplained by an increased appraisal of the \u201copportu-\nnity costs\u201d to smoking for expected work outcomes, \neither in terms of physical health or coworker evalu-\nations (Cowell 2006), the prevalence of smoking \nbans and education campaigns on college campuses \n(Halperin and Rigotti 2003), and lower rates of reg-\nular smoking among traditional-aged college stu-\ndents, who student mothers may view as a standard \nof comparison (Centers for Disease Control and \nPrevention 2019). Of course, future research must \nalso aim to assess the validity of this finding, partic-\nularly given the small size of this subsample.\nThis study also had several limitations. First, the \nsample overrepresented minority women in urban \nmetros. Although I did not find evidence of a health \nbenefit to postnatal college completion among white \nmothers\u2014for whom there may be a larger health \nbenefit to higher education than minority mothers \n(Kimbro et\u00a0al. 2008)\u2014this possibility still needs to \nbe investigated because the sample of white mothers \nprovided limited power to detect significant effects. \nGiven subsample size limitations, I also did not \ninvestigate the moderating role of other demo-\ngraphic factors, such as marital status or work expe-\nrience (see Rindfuss, Swicegood, and Rosenfeld \n1987), which future research should explore, nor did \nI explore circumstances under which some mothers \nmay benefit from completing a college degree (see \nFurstenberg 2005), for example, in families with \naccess to high-quality child care (Harding 2015).\nFurthermore, I could not differentiate among the \n\u201cquality\u201d of mothers\u2019 education, for example, \nwhether the coursework was online or in person, \nthe type of major, or whether mothers earned \ndegrees from a for-profit institution. This latter \nissue is particularly noteworthy given that enroll-\nment in for-profit schools is substantially higher \namong mothers than other groups (Cruse et\u00a0 al. \n2019) and leads to fewer labor market payoffs and \ngreater debt (Deming et\u00a0al. 2013). At the same time, \nenrollment in for-profit schools, as well as among \nstudent mothers, has declined over the past several \nyears (Cruse et\u00a0 al. 2019). Given these trends, as \nwell as recent increases in tuition costs and turbu-\nlence in the economy (which recently included his-\ntorically low and high levels of unemployment), it \nalso remains unclear whether mothers who com-\npleted postsecondary education subsequent to the \nFFCW sample would have experienced different \nreturns to postsecondary schooling.\nIn sum, this study focused on an emerging  \nyet understudied phenomenon among health \nresearchers\u2014mothers\u2019 return to school to increase \ntheir education\u2014to investigate the policy-relevant \nissue of whether mothers\u2019 reentry into school can \nimprove their health outcomes. Contrary to much \nresearch and theory, the findings suggest that in \ngeneral, the answer is no. This finding points to an \nimportant source of heterogeneity in the health \nreturns to education\u2014the timing of degree attain-\nment vis-\u00e0-vis the timing of children\u2019s births\u2014that \nfurther stratifies the education\u2013health gradient in an \noften unacknowledged way. It also speaks to the \nlimits of current public health and two-generation \npolicy approaches, points to alternative prescrip-\ntions for social policy, and highlights a new research \nfrontier that explores the social and structural \n\u00adconstraints that limit the health returns to additional \neducation among women with children.\nAuthor\u2019s Note\nPrevious versions of this article were presented at the 2018 \nannual meeting for the Interdisciplinary Association for \nPopulation Health Science and the 2019 annual meeting \nfor the Population Association of America.\nORCID Id\nJennifer March Augustine \n https://orcid.org/0000- \n0003-1446-8723\nNote\n1.\t\nI also included three additional measures\u2014heavy \ndrinking, clinical depression, and body mass \nindex\u2014in auxiliary analyses, as reported in the \nRobustness Analyses section, but not as focal mea-\nsures for the following reasons. Heavy drinking \nwas more common among higher-educated mothers \nthan lower-educated mothers and thus did not con-\nform to the \u201ceducational gradient.\u201d It was also low \nin variability. Body mass index was only reported in \nWaves 3 to 5 and only among a subset of mothers. \nClinical depression, as explained in the Methods \nsection, was a more restrictive measure than the \nGlobal Mental Health measure that was also lower \nin variability.\n\n16\t\nJournal of Health and Social Behavior 62(1)\nReferences\nAbidin, Richard. 1983. Parenting Stress Index. Charlottesville, \nVA: Pediatric Psychology Press.\nAdams, Scott. J. 2002. \u201cEducational Attainment and \nHealth: Evidence from a Sample of Older Adults.\u201d \nEducation Economics 10(1):97\u2013109.\nAdler, Nancy E., Thomas Boyce, Margaret A. Chesney, \nSheldon Cohen, Susan Folkman, Robert L. Kahn, and \nSyme S. Leonard. 1994. \u201cSocioeconomic Status and \nHealth: The Challenge of the Gradient.\u201d American \nPsychologist 49(1):15\u201324.\nAllison, Paul D. 2009. Fixed Effects Regression Models. \nVol. 160. Newbury Park, CA: SAGE Publications.\nAllison, Paul. 2012. \u201cHandling Missing Data by Maximum \nLikelihood.\u201d Paper 312-2012. SAS Global Forum. \nhttp://www.statisticalhorizons.com/wp-content/ \nuploads/MissingDataByML.pdf.\nArendt, Jacob N. 2005. \u201cDoes Education Cause Better \nHealth? A Panel Data Analysis Using School Reforms \nfor Identification.\u201d Economics of Education Review \n24(2):149\u201360.\nAugustine, Jennifer M. 2016. \u201cExploring New Life \nCourse Patterns of Women\u2019s Secondary and Post-\nSecondary Schooling.\u201d Population Research and \nPolicy Review 35(6):727\u201355.\nAugustine, Jennifer M., and Daniela Negraia. 2018. \n\u201cCan Increased Educational Attainment among \nLower Educated Mothers Reduce Inequalities in \nthe Skill Development of Children?\u201d Demography \n55(1):59\u201382.\nAugustine, Jennifer March, Kate C. Prickett, and Daniela \nV. Negraia. 2018. \u201cDoing It All? Mothers\u2019 College \nEnrollment, Time Use, and Affective Well-Being.\u201d \nJournal of Marriage and Family 80(4):963\u201374.\nBandeira, Frank C., Amanda Kalaydjian Richardson, \nDavid J. Lee, Jian-Ping He, and Kathleen R. \nMerikangas. 2011. \u201cSecondhand Smoke Exposure \nand Mental Health among Children and Adolescents.\u201d \nArchives of Pediatrics & Adolescent Medicine 165(4): \n332\u201338.\nBarban, Nicola. 2013. \u201cFamily Trajectories and Health: \nA Life Course Perspective.\u201d European Journal of \nPopulation 29(4):357\u201385.\nBeck, Audrey, Carey E. Cooper, Sara McLanahan, and \nJeanne Brooks-Gunn. 2010. \u201cPartnership Transitions \nand Maternal Parenting.\u201d Journal of Marriage and \nFamily 72(2):219\u201333.\nBianchi, Suzanne M., John P. Robinson, and Melissa \nA. Milkie. 2006. The Changing Rhythms of \nAmerican Family Life. New York, NY: Russell Sage \nFoundation.\nBloom, Janice. 2007. \u201c(Mis)reading Social Class in the \nJourney towards College: Youth Development in \nUrban America.\u201d Teachers College Record 109(2): \n343\u201368.\nBraboy Jackson, Pamela. 2004. \u201cRole Sequencing: Does \nOrder Matter for Mental Health?\u201d Journal of Health \nand Social Behavior 45(2):132\u201354.\nBraboy Jackson, Pamela, and Alexandra Berkowitz. \n2005. \u201cThe Structure of the Life Course: Gender \nand Racioethnic Variation in the Occurrence and \nSequencing of Role Transitions.\u201d Advances in Life \nCourse Research 9:55\u201390.\nBr\u00fcckner, Hannah, and Karl Ulrich Mayer. 2005. \n\u201cDe-standardization of the Life Course: What It \nMight Mean? And if It Means Anything, Whether \nIt Actually Took Place?\u201d Advances in Life Course \nResearch 9:27\u201353.\nCarney-Crompton, Shawn, and Josephine Tan. 2002. \n\u201cSupport Systems, Psychological Functioning, and \nAcademic Performance of Nontraditional Female \nStudents.\u201d Adult Education Quarterly 52(2):140\u201354.\nCenters for Disease Control and Prevention. 2019. \u201cCurrent \nSmoking among U. S. Adults Aged 18 and Older.\u201d \nhttps://www.cdc.gov/tobacco/campaign/tips/resources/\ndata/cigarette-smoking-in-united-states.html.\nCockerham, William C. 2005. \u201cHealth Lifestyle Theory \nand the Convergence of Agency and Structure.\u201d \nJournal of Health and Social Behavior 46(1):51\u201367.\nConti, Gabriella, James Heckman, and Sergio Urzua. 2010. \n\u201cThe Education-Health Gradient.\u201d The American \nEconomic Review 100(2):234\u201338.\nCooksey, Elizabeth C., Elizabeth G. Menaghan, and \nSusan M. Jekielek. 1997. \u201cLife-Course Effects of \nWork and Family Circumstances on Children.\u201d \nSocial Forces 76(2):637\u201365.\nCooper, Carey E., Sara S. McLanahan, Sarah O. \nMeadows, and Jeanne Brooks-Gunn. 2009. \u201cFamily \nStructure Transitions and Maternal Parenting Stress.\u201d \nJournal of Marriage and Family 71(3):558\u201374.\nCottom, Tressie McMillan. 2017. Lower Ed: The \nTroubling Rise of For-Profit Colleges in the New \nEconomy. New York City, NY: New Press.\nCowell, Alexander J. 2006. \u201cThe Relationship between \nEducation and Health Behavior: Some Empirical \nEvidence.\u201d Health Economics 15(2):125\u201346.\nCrnic, Keith A., Catherine Gaze, and Casey Hoffman. \n2005. \u201cCumulative Parenting Stress across the \nPreschool Period: Relations to Maternal Parenting \nand Child Behavior at Age 5.\u201d Infant and Child \nDevelopment: An International Journal of Research \nand Practice 14(2):117\u201332.\nCruse, Lindsey Reichlin, Tessa Holtzman, Barbara Gault, \nDavid Croom, and Portia Polk. 2019. Parents in \nCollege by the Numbers. Washington, DC: Institute \nfor Women\u2019s Policy Research.\nCutler, David M., and Adriana Lleras-Muney. 2010. \n\u201cUnderstanding Differences in Health Behaviors by \nEducation.\u201d Journal of Health Economics 29(1):1\u201328.\nDeming, David, Claudia Goldin, and Lawrence Katz. \n2013. \u201cFor-Profit Colleges.\u201d The Future of Children \n21(1):137\u201363.\nDuke, Naomi, and Ross Macmillan. 2016. \u201cSchooling, \nSkills, and Self-Rated Health: A Test of Conventional \nWisdom on the Relationship between Educational \nAttainment and Health.\u201d Sociology of Education \n89(3):171\u2013206.\n\nAugustine\t\n17\nEdin, Kathryn, and Kefalas Maria 2011. Promises I Can \nKeep: Why Poor Women Put Motherhood before \nMarriage. Berkeley, CA: University of California \nPress.\nEide, Eric R., and Mark H. Showalter. 2011. \u201cEstimating \nthe Relation between Health and Education: What \nDo We Know and What Do We Need to Know?\u201d \nEconomics of Education Review 30(5):778\u201391.\nEisenhower, Abbey Severance, Bruce L. Baker, and Jan \nBlacher. 2009. \u201cChildren\u2019s Delayed Development \nand Behavior Problems: Impact on Mothers\u2019 \nPerceived Physical Health across Early Childhood.\u201d \nSocial Science & Medicine 68(1):89\u201399.\nElder, G. H. 1984 \u201cFamilies, Kin, and the Life Course: \nA Sociological Perspective.\u201d Pp. 80\u2013136 in Review \nof Child Development Research: The Family, edited \nby R. Parke. Chicago: University of Chicago Press.\nElman, Cheryl, and Angela M. O\u2019Rand. 2004. \u201cThe \nRace Is to the Swift: Socioeconomic Origins, \nAdult Education, and Wage Attainment.\u201d American \nJournal of Sociology 110(1):123\u201360.\nFelmlee, Diane H. 1988. \u201cReturning to School and \nWomen\u2019s Occupational Attainment.\u201d Sociology of \nEducation 61(1):29\u201341.\nFrieden, Thomas R. 2010. \u201cA Framework for Public \nHealth Action: The Health Impact Pyramid.\u201d \nAmerican Journal of Public Health 100(4):590\u201395.\nFurstenberg, Frank F. 2005. \u201cNon-normative Life Course \nTransitions: Reflections on the Significance of \nDemographic Events on Lives.\u201d Advances in Life \nCourse Research 10:155\u201372.\nGarbarski, Dana. 2014. \u201cThe Interplay between Child \nand Maternal Health: Reciprocal Relationships \nand Cumulative Disadvantage during Childhood \nand Adolescence.\u201d Journal of Health and Social \nBehavior 55(1):91\u2013106.\nGardner, Margo, Jeanne Brooks-Gunn, and P. Lindsay \nChase-Lansdale. \n2017. \n\u201cThe \nTwo-Generation \nApproach to Building Human Capital: Past, Present, \nand Future.\u201d Pp. 330\u201361 in The Wiley Handbook of \nEarly Childhood Development Programs, Practices, \nand Policies, edited by Edited by E. Votruba-Drzal \nand E. Dearing. Hoboken, NJ: Wiley.\nGardner, Margo, Anne Martin, and Am\u00e9lie Petitclerc. \n2019. \u201cMothers\u2019 Postsecondary Entry during \nEarly Childhood: Short- and Long-Term Effects \non Children.\u201d Journal of Applied Developmental \nPsychology 62:11\u201325.\nGershoff, Elizabeth T., J. Lawrence Aber, C. Cybele \nRaver, and Mary Clare Lennon. 2007. \u201cIncome Is \nNot Enough: Incorporating Material Hardship into \nModels of Income Associations with Parenting and \nChild Development.\u201d Child Development 78(1):70\u2013\n95.\nGhisletta, Paolo, Olivier Renaud, Nad\u00e8ge Jacot, and \nDelphine Courvoisier. 2015. \u201cLinear Mixed-Effects \nand Latent Curve Models for Longitudinal Life \nCourse Analyses.\u201d Pp. 155\u201378 in A Life Course \nPerspective on Health Trajectories and Transitions, \nedited by C. Burton-Jeangros, S. Cullati, A. Sacker, \nand D. Blane. Springer: Cham, the Netherlands.\nHahn, Robert A., and Benedict I. Truman. 2015. \n\u201cEducation Improves Public Health and Promotes \nHealth Equity.\u201d International Journal of Health \nServices 45(4):657\u201378.\nHalperin, Abigail C., and Nancy A. Rigotti. 2003. \u201cU. S. \nPublic Universities\u2019 Compliance with Recommended \nTobacco-Control Policies.\u201d Journal of American \nCollege Health 51(5):181\u201388.\nHarding, Jessica F. 2015. \u201cIncreases in Maternal \nEducation and Low-Income Children\u2019s Cognitive and \nBehavioral Outcomes.\u201d Developmental Psychology \n51(5):583\u201399.\nHarding, Jessica F., Pamela A. Morris, and Diane \nHughes. 2015. \u201cThe Relationship between Maternal \nEducation and Children\u2019s Academic Outcomes: A \nTheoretical Framework.\u201d Journal of Marriage and \nFamily 77(1):60\u201376.\nHarris, Kathleen Mullan, Carolyn Tucker Halpern, Jon \nHussey, Eric A. Whitsel, Ley Killeya-Jones, Joyce \nTabor, and Glen Elder. 2013. \u201cSocial, Behavioral, \nand Genetic Linkages from Adolescence into \nAdulthood.\u201d American Journal of Public Health \n103(1):S25\u201332.\nHerd, Pamela, Jenny Higgins, Kamil Sicinski, and Irina \nMerkurieva. 2016. \u201cThe Implications of Unintended \nPregnancies for Mental Health in Later Life.\u201d \nAmerican Journal of Public Health 106(2):421\u201329.\nHofhuis, Ward, Johan C. de Jongste, and Peter Merkus. \n2003. \u201cAdverse Health Effects of Prenatal and \nPostnatal Tobacco Smoke Exposure on Children.\u201d \nArchives of Disease in Childhood 88(12):1086\u201390.\nHogan, Dennis P. 1978. \u201cThe Variable Order of Events \nin the Life Course.\u201d American Sociological Review \n43(4):573\u201386.\nHout, Michael. 2012. \u201cSocial and Economic Returns to \nCollege Education in the United States.\u201d Annual \nReview of Sociology 38:379\u2013400.\nKalleberg, Arne L. 2011. Good Jobs, Bad Jobs: The Rise \nof Polarized and Precarious Employment Systems \nin the United States, 1970s-2000s. New York, NY: \nRussell Sage Foundation.\nKatz, Sheila. 2013. \u201c\u2018Give Us a Chance to Get an \nEducation\u2019: Single Mothers\u2019 Survival Narratives \nand Strategies for Pursuing Higher Education on \nWelfare.\u201d Journal of Poverty 17(3):273\u2013304.\nKaushal, Neeraj. 2014. \u201cIntergenerational Payoffs of \nEducation.\u201d The Future of Children 24(1):61\u201378.\nKimbro, Rachel Tolbert, Sharon Bzostek, Noreen \nGoldman, and Germ\u00e1n Rodr\u00edguez. 2008. \u201cRace, \nEthnicity, and the Education Gradient in Health.\u201d \nHealth Affairs 27(2):361\u201372.\nLaird, Nan M., and James H. Ware. 1982. \u201cRandom-\nEffects Models for Longitudinal Data.\u201d Biometrics \n38(4): 963\u201374.\nLaumann, Edward O. 1973. Bonds of Pluralism: The \nForm and Substance of Urban Social Networks. New \nYork, NY: Wiley-Interscience.\n\n18\t\nJournal of Health and Social Behavior 62(1)\nLink, Bruce G., and Jo Phelan. 1995. \u201cSocial Conditions \nas Fundamental Causes of Disease.\u201d Journal of \nHealth and Social Behavior 1995:80\u201394.\nLleras-Muney, Adriana. 2005. \u201cThe Relationship \nbetween Education and Adult Mortality.\u201d U.S. \nReview of Economic Studies 72(1):189\u2013221.\nLynch, Jamie L., and Paul T. von Hippel. 2016. \u201cAn \nEducation Gradient in Health, a Health Gradient \nin Education, or a Confounded Gradient in Both?\u201d \nSocial Science & Medicine 154:18\u201327.\nMacmillan, Ross. 2005. \u201cThe Structure of the Life \nCourse: Classic Issues and Current Controversies.\u201d \nAdvances in Life Course Research 9:3\u201324.\nMarini, Margaret Mooney. 1984. \u201cAge and Sequencing \nNorms in the Transition to Adulthood.\u201d Social Forces \n63(1):229\u201344.\nMcDonough, Patricia M. 1997. Choosing Colleges: How \nSocial Class and Schools Structure Opportunity. \nAlbany, NY: SUNY Press.\nMcFarland, Michael, and Brandon G. Wagner. 2015. \n\u201cDoes a College Education Reduce Depressive \nSymptoms in American Young Adults?\u201d Social \nScience & Medicine 146:75\u201384.\nMeara, Ellen R., Seth Richards, and David M. Cutler. \n2008. \u201cThe Gap Gets Bigger: Changes in Mortality \nand Life Expectancy, by Education, 1981\u20132000.\u201d \nHealth Affairs 27(2):350\u201360.\nMiech, Richard A., Michael Shanahan, Jason Boardman, \nand Shawn Bauldry. 2015. \u201cThe Sequencing of a \nCollege Degree during the Transition to Adulthood: \nImplications for Obesity.\u201d Journal of Health and \nSocial Behavior 56(2):281\u201395.\nMirowsky, John, and Catherine E. Ross. 2008. \u201cEducation \nand Self-Rated Health: Cumulative advantage and \nIts Rising Importance.\u201d Research on Aging 30(1): \n93\u2013122.\nMontez, Jennifer Karas, and Anna Zajacova. 2013. \n\u201cTrends in Mortality Risk by Education Level and \nCause of Death among White Women in the United \nStates from 1986 to 2006.\u201d American Journal of \nPublic Health 103(3):473\u201379.\nMirowsky, John, and Catherine Ross. 2003. Education, \nSocial Status, and Health. New York, NY: Aldine \nDe Gruyter.\nParrish, R. Gibson. 2010. \u201cMeasuring Population Health \nOutcomes.\u201d Preventing Chronic Disease 7(4):1\u201311.\nPearlin, Leonard, Scott Schieman, Elena M. Fazio, and \nStephen Meersman. 2005. \u201cStress, Health, and \nthe Life Course: Some Conceptual Perspectives.\u201d \nJournal of Health and Social Behavior 46(2):205\u201319.\nReichman, Nancy E., Julien O. Teitler, Irwin Garfinkel, \nand Sara S. McLanahan. 2001. \u201cFragile Families: \nSample and Design.\u201d Children and Youth Services \nReview 23(4/5):303\u201326.\nRindfuss, Ronald R., C. Gray Swicegood, and Rachel \nA. Rosenfeld. 1987. \u201cDisorder in the Life Course: \nHow Common and Does It Matter?\u201d American \nSociological Review 512(6):785\u2013801.\nRosenbaum, James E. 2001. Beyond College for All: \nCareer Paths for the Forgotten Half. New York, NY: \nRussell Sage Foundation.\nSeefeldt, Kristin S. 2016. Abandoned Families: Social \nIsolation in the Twenty-First Century. New York, \nNY: Russell Sage Foundation.\nSinger, Judith D., and John B. Willett. 2003. Applied \nLongitudinal Data Analysis: Modeling Change and \nEvent Occurrence. Oxford, UK: Oxford University \nPress.\nTurner, R. Jay, and Donald A. Lloyd. 1999. \u201cThe Stress \nProcess and the Social Distribution of Depression.\u201d \nJournal of Health and Social Behavior 40(4):374\u2013404.\nTurney, Kristin. 2011. \u201cMaternal Depression and \nChildhood Health Inequalities.\u201d Journal of Health \nand Social Behavior 52(3):314\u201332.\nVon Hippel, Paul T. 2007. \u201cRegression with Missing Ys: \nAn Improved Strategy for Analyzing Multiply Imputed \nData.\u201d Sociological Methodology 37(1):83\u2013117.\nWilliams, Kristi, Sharon Sassler, Adrianne Frech, Fenaba \nAddo, and Elizabeth Cooksey. 2011. \u201cNonmarital \nChildbearing, Union History, and Women\u2019s Health \nat Midlife.\u201d American Sociological Review 76(3): \n465\u201386.\nWilson, Kristin B. 2008. \u201cInstitutional Influences \nAffecting the College-Going Decisions of Low-\nIncome Mothers Attending a Rural Midwestern \nCommunity College.\u201d PhD dissertation, Department \nof Sociology, University of Missouri.\nYuan, Anastasia S. 2008. \u201cExploring the Changes in \nEconomic Hardship and Children\u2019s Well-Being over \nTime: The \u2018Linked Lives\u2019 of Parents and Children.\u201d \nAdvances in Life Course Research 13:321\u201341.\nZajacova, Anna, and Elizabeth M. Lawrence. 2018. \u201cThe \nRelationship between Education and Health: Reducing \nDisparities through a Contextual Approach.\u201d Annual \nReview of Public Health 39:273\u201389.\nZhao, Xinshu, John G. Lynch, and Qimei Chen. 2010. \n\u201cReconsidering Baron and Kenny: Myths and Truths \nabout Mediation Analysis.\u201d Journal of Consumer \nResearch 37(2):197\u2013206.\nAuthor Biography\nJennifer March Augustine is an associate professor in \nthe Department of Sociology at the University of South \nCarolina. Her research applies life course and ecological \napproaches to studying the connections among parents\u2019 \nsocioeconomic circumstances, their health and well-being, \nand the health and development of their children. She is \nparticularly interested in understanding the significance of \nparents\u2019 education, its implication for other related life \ncourse outcomes, and the combined impact of these factors \non various dimensions of family life.\n\n\n    Scientific Research Findings:", "answer": "\u2022 Mothers\u2019 postsecondary degree completion after having children does not improve their health, including their self-rated health, mental health, levels of parenting stress, or reports of smoking.\n\u2022 These patterns were true irrespective of whether mothers completed a vocational, associate\u2019s, or bachelor\u2019s degree and were robust across population groups like White women, first-time mothers, and mothers without high school degrees as well as alternate analysis methods.\n\u2022 The one exception to this overall pattern was that smoking rates were reduced among mothers who did not have any college schooling and subsequently earned a bachelor\u2019s degree.", "id": 74}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465211003232\nJournal of Health and Social Behavior\n2021, Vol. 62(2) 120\u00ad\u2013135\n\u00a9 American Sociological Association 2021\nDOI: 10.1177/00221465211003232\njhsb.sagepub.com\nArticle\nResearchers have established the presence of bias in \nAmerican health care with considerable work focus-\ning on provider bias within the clinical encounter. \nIndeed, medical providers demonstrate bias on par \nwith the general population, even among those who \nself-identify as holding nondiscriminatory beliefs \n(Burgess et\u00a0al. 2007; van Ryn and Fu 2003; Zestcott, \nBlair, and Stone 2016). Understanding such bias is \nimportant because it contributes to the profound \nhealth disparities seen in the United States. For \ninstance, providers counsel and treat patients differ-\nently depending on patient characteristics, which \noften disproportionately affects marginalized popu-\nlations, who consistently experience worse health \noutcomes (see Bao, Fox, and Escarce 2007; \nSchulman et\u00a0 al. 1999; van Ryn et\u00a0 al. 2006). Yet \nbeyond recognizing the presence of bias in provid-\ners\u2019 patient counseling and decision-making, mini-\nmal research has explored how providers then \nnavigate bias in their care to minimize its effects. By \nexploring this process, we demonstrate the ways that \nvarious provider approaches to reduce bias reflect \nand reproduce this inequality in care. Elucidating \nthese coping strategies simultaneously reveals not \nonly providers\u2019 individually held biases but also the \nsocial forces that engender such discrimination and \ndisparities within their clinical encounters.\nTo combat bias, researchers have recommended \nnumerous interventions, such as patient-centered \ncare, cultural competency training, organizational \nreform, and diversifying medical schools and \n1003232 HSBXXX10.1177/00221465211003232Journal of Health and Social BehaviorManzer and Bell\nresearch-article2021\n1University of Delaware, Newark, DE, USA\nCorresponding Author:\nJamie L. Manzer, Department of Sociology and Criminal \nJustice, University of Delaware, 325 Smith Hall, \nNewark, DE, 19716, USA. \nEmail: jmanzer@udel.edu\n\u201cWe\u2019re a Little Biased\u201d: \nMedicine and the Management \nof Bias through the Case of \nContraception\nJamie L. Manzer1\n and Ann V. Bell1\nAbstract\nThere is a wealth of literature demonstrating the presence of bias throughout the American health care \nsystem. Despite acknowledging such presence, however, little is known about how bias functions within \nmedical encounters, particularly how providers grapple with bias in their patient counseling and decision-\nmaking. We explore such processes through the case of contraceptive counseling, a highly raced, classed, \nand gendered context. In-depth interviews with 51 health care providers reveal that providers use four \nprimary strategies to navigate and minimize bias in their care\u2014using scientific rationale, employing \u201csafe\u201d \nbiases, standardizing counseling, and implementing patient-centered care. Paradoxically, using these \nstrategies can exacerbate rather than resolve bias. Understanding these bias management strategies reveals \nprovider-held biases, how they manifest within appointments, and the potential consequences for patients\u2019 \nhealth autonomy. Such knowledge informs interventions that promote contraceptive use among women in \nthe United States, addresses bias in health care broadly, and thus ultimately helps combat health disparities.\nKeywords\nbias, clinicians, contraception, health care, reproduction\n\nManzer and Bell\t\n121\nreforming their curriculum (Beach et\u00a0 al. 2005; \nBetancourt et\u00a0al. 2003; Higgins, Kramer, and Ryder \n2016; Metzl and Hansen 2014). However, few, if \nany, are evidence based, and most address provider-\nheld biases while neglecting the structural nature of \nthe inequality that informs such exchanges (Metzl \nand Hansen 2014; Zestcott et\u00a0al. 2016). Moreover, \nbecause little is known about how providers man-\nage this bias within their patient interactions, rec-\nommendations for improving these encounters may \ninevitably fall short. Indeed, the patient\u2013provider \nrelationship is central to the provision of health \ncare. Patients are, overall, more active consumers \nthan in the past, and the provider\u2013patient relation-\nship has become less paternalistic as a result. \nHowever, because of how U.S. health care is struc-\ntured and the rise of disease nationwide, the clinical \nencounter is pivotal to health outcomes because \nproviders guide downstream access to care (Boyer \nand Lutfey 2010). As such, the quality of these \nexchanges must be continuously assessed.\nTo do so, we explore how providers navigate \nand minimize bias in health care through the case of \ncontraceptive counseling. As scholars emphasize, \nproviders\u2019 contraceptive decision-making is an \nimportant area in which to study bias because these \nencounters involve the consideration of medical, \ncultural, and social factors, such as patients\u2019 sexual \nbehaviors, partner histories, and pregnancy desires \n(see Dehlendorf et\u00a0al. 2010; Stevens 2015). Moral \njudgments and cultural norms around these issues, \nwhether implicit or explicit, can influence provid-\ners\u2019 prescribing patterns and the substance of their \ncounseling (Metzl and Hansen 2014). A growing \nbody of research has identified possible bias in \nthese areas (Borrero et\u00a0al. 2009; Dehlendorf et\u00a0al. \n2010; Gomez and Wapman 2017; Yee and Simon \n2011). For instance, some clinicians are more likely \nto encourage women of color as well as young or \npoor women to choose sterilization or long-acting \nreversible contraception (LARC; Borrero et\u00a0 al. \n2009; Dehlendorf et\u00a0al. 2010; Gomez, Fuentes, and \nAllina 2014; Gomez and Wapman 2017; Yee and \nSimon 2011). Patients are attuned to this bias, \nsometimes feeling implicit pressure to use methods \nthey do not desire (Downing, LaVeist, and Bullock \n2007; Gomez and Wapman 2017). Beyond limiting \naccess to women\u2019s preferred contraceptives, such \ncounseling practices can cause women to desist \nfrom contraceptive use, creating gaps in coverage, \nor result in some women forgoing follow-up care \naltogether, jeopardizing their overall health (Gomez \nand Wapman 2017).\nAdditionally, the range of contraceptive methods \navailable and the consideration of patient prefer-\nences make it a particularly complex area of medi-\ncine. Consequently, providers\u2019 biases may be more \ninfluential within contraceptive counseling contexts \nrelative to other areas of medicine where broad clin-\nical consensus for appropriate treatment leaves less \nroom for subjective decision-making (Dehlendorf \net\u00a0 al. 2010). Therefore, to better understand how \nbias functions within contraceptive counseling, we \nask: How do providers rationalize, cope with, and \nmanage bias within their patient interactions and \ndecision-making processes? Using a sociological \nlens to answer such a question allows us to see not \nonly the existence of bias among providers but also \nhow it manifests and persists. Doing so highlights \nthe relevant structural features of medicine along \nwith the broader social forces that produce bias in \nhealth care. Understanding this link offers opportu-\nnities for multilevel solutions to reduce bias in med-\nicine and the disparities it perpetuates.\nBackground\nHealth Care and Bias\nResearchers have extensively documented health \ndisparities in the United States, noting that relative \nto more privileged patients, marginalized popula-\ntions experience worse health outcomes and have \nless access to quality health care (Adler and \nNewman 2002; McLeod 2015; Smedley, Stith, and \nNelson 2002). Variables such as race, class, gender, \nand ethnicity become factors in care because of the \nprovider\u2013patient interaction as well as the structural \ncontexts in which these interactions occur. Like \nwith other institutions, medical organizations are \nnot value neutral. Providers apply cultural knowl-\nedge and stereotypes in these professional spaces \njust as they do in lay settings to make, often uncon-\nscious, moral evaluations about patients as they \ndiagnose and treat them (Spencer and Grace 2016).\nResearchers note that patient characteristics \nconsistently predict providers\u2019 decisions even when \naccess to and use of medical care and presenting \nsymptoms are controlled (see Metzl and Hansen \n2014; Spencer and Grace 2016). Some of this varia-\ntion in care is related to provider bias that persists \ndespite substantial efforts within medicine to \naddress both these disparities and the bias that con-\ntributes to them (Chapman, Kaatz, and Carnes \n2013).1 Addressing provider bias is particularly \nchallenging because racism and other forms of dis-\ncrimination and privilege now operate in less overt \n\n122\t\nJournal of Health and Social Behavior 62(2)\nways within the clinical encounter. For instance, as \nracism and the effects of racialized structures \nbecome less explicit, it is easier for health care pro-\nviders and lay people to deny racial stratification in \nthe United States as well as individually held racial \nbias (Bonilla-Silva 2010). Further complicating \nefforts to reduce bias, providers can simultaneously \nhold explicit egalitarian beliefs about patient groups \nbut still harbor implicit forms of bias that interfere \nwith their interactions, affect, and evaluation of \npatients (Chapman et\u00a0al. 2013). Therefore, under-\nstanding how providers manage bias within their \ncare is a key step in identifying both how bias per-\nsists and how to address it.\nIncreasingly, scholars note that although pro-\nvider bias is a significant factor in predicting dispa-\nrate care, more needs to be done to understand the \nmechanisms that underlie these patterns. For \ninstance, Metzl and Hansen (2014) critiqued cul-\ntural competency training, a popular intervention \nused nationwide to combat bias by encouraging \nproviders to be reflexive about their personal \nworldview and develop competence toward patients \nwith diverse backgrounds. Instead, they proposed \nstructural competency training, noting that cultur-\nally sensitive clinicians do little to mitigate the \neffects of the structural barriers patients experience \nwhen accessing care.\nConversely, cultural competency training and \nother similar interventions may, instead, prompt \nproviders to assume these structural barriers, such \nas challenges accessing medication, reliable trans-\nportation, and lack of healthy foods, to be the result \nof patients\u2019 cultural preferences that can imply \neither a moral failing or moral superiority (Metzl \nand Hansen 2014). In other words, structural-level \nfactors inform clinical interactions, diagnosis, and \ntreatment when subsequent assumptions about cul-\ntural groups become embedded in providers\u2019 lan-\nguage and beliefs to the benefit of some and harm \nto others (Metzl and Hansen 2014). Exploring how \nproviders manage and attempt to minimize bias \nreveals how this language and entrenched beliefs \naffect providers\u2019 actions within medical contexts \nand subsequently, patients\u2019 experiences, outcomes, \nand autonomy.\nIn this way, provider interactions and decision-\nmaking have multiple spheres of influence beyond \nstrictly physical manifestations of illness that need \nto be better understood (see Spencer and Grace \n2016). For instance, providers triage, diagnose, and \ntreat patients in emergency rooms differently based \non a range of patient characteristics (e.g., cleanli-\nness, insurance status, and perceived morality) that \nhave little to do with their presenting symptoms \n(Lara-Mill\u00e1n 2014). This occurs because institu-\ntions like medicine produce and reproduce cultural \nvalues that result in disparate care by creating a \n\u201ccommonsense\u201d medicine where stereotypes about \nsocial groups serve as explanations for their cir-\ncumstances that, over time, gain medical legitimacy \n(Bonilla-Silva and Baiocchi 2001; Metzl and \nHansen 2014). Such findings are particularly poi-\ngnant for reproductive health care given the United \nStates\u2019s continued use of contraception to encour-\nage the births of more privileged women over those \nof marginalized women (Colen 1986; Gomez et\u00a0al. \n2014; Roberts 1997).\nIndeed, medicine is a racialized system where \none\u2019s health status is predetermined by and predi-\ncated on other racialized structures, such as educa-\ntion, economic policy, and the law (Bonilla-Silva \nand Baiocchi 2001). These structures are also gen-\ndered and classed in ways that are often invisible to \nproviders. This inequality has been found to limit \naccess to resources and health care, ultimately \naffecting what needs patients present with and how \nproviders then perceive these patients (see Spencer \nand Grace 2016). Importantly, institutionalized \nnorms and practices of predominantly white, \nWestern nations manifest in these structures, where \nassumptions and stigma about individual patients \nare not produced in individual encounters but are, \ninstead, enacted and normalized within these \nexchanges (Carmichael 2003). Indeed, how provid-\ners manage bias within their care helps to reveal \nthese often insidious biases and elucidates the \nsocial forces that engender them within patient\u2013\nprovider interactions. Doing so supports necessary, \nmultilevel interventions to address both provider \nbiases and broader health care inequity.\nContraceptive Counseling and Bias\nSocial justice efforts such as the reproductive justice \n(RJ) movement, begun by a group of black women \nin the mid-1990s, emphasize this nexus between \nintersecting social identities and the effects of struc-\ntural inequalities on health\u2014specifically, women\u2019s \nright to have children (or not) and to parent their \nchildren in a safe and healthy environment (Luna \nand Luker 2013). A key tenet of the RJ movement is \nidentifying contraceptive equity as a human right. \nThrough this RJ lens, women\u2019s contraceptive \nchoices should be fully informed and fully volun-\ntary, and these methods should be readily available \n(Gomez et\u00a0al. 2014; Gubrium et\u00a0al. 2016; Roberts \n1997). Women should be able to change or remove \n\nManzer and Bell\t\n123\nmethods when they desire without undue provider \ninfluence as well as feel respected when they elect \nnot to use contraception (Higgins et\u00a0 al. 2016). \nHowever, provider bias and the structural inequality \nthat shapes these exchanges have been found to \nundermine the quality of marginalized women\u2019s \ncare and their contraceptive decision-making, which \nhave broad implications for how women manage \ntheir fertility (Gomez et\u00a0 al. 2014; Gomez and \nWapman 2017; Gubrium et\u00a0al. 2016; Higgins et\u00a0al. \n2016).\nIndeed, as RJ advocates and scholars empha-\nsize, contraception has historically been used as a \nmechanism to limit the births of socially marginal-\nized women in the United States through intentional \ninvoluntary sterilization and eugenics programs \n(Dehlendorf et\u00a0al. 2018; Gomez et\u00a0al. 2014; Gomez \nand Wapman 2017; Gubrium et\u00a0al. 2016; Higgins \net\u00a0al. 2016). Some researchers caution that provid-\ners may, often inadvertently, carry on these tradi-\ntions by persuasively counseling or implicitly \npressuring marginalized women to use contracep-\ntion more so than their privileged patients\u2014 \nparticularly LARC methods, such as the intrauter-\nine device (IUD) or subdermal arm implant \n(Borrero et\u00a0al. 2009; Dehlendorf et\u00a0al. 2010; Gomez \nand Wapman 2017; Gubrium et\u00a0al. 2016; Yee and \nSimon 2011). Moreover, providers may be more \nhesitant to remove a LARC when requested by a \nwoman of lower socioeconomic status (SES) or \nwoman of color, citing perceptions of pregnancy \nrisk related to patient age, gender, perceived sexual \npromiscuity, and readiness to parent (Gomez et\u00a0al. \n2014; Henderson et\u00a0al. 2011; Kalmuss et\u00a0al. 1996; \nLittlejohn and Kimport 2017; Stevens 2015). \nBorrero and colleagues (2009) noted that this \nunevenness in contraceptive counseling is not \npatient initiated; rather, unconscious beliefs about \nwho should reproduce may affect providers\u2019 care.\nImportantly, providers recognize their influen-\ntial position (Henderson et\u00a0 al. 2011). Henderson \nand colleagues (2011) found that most clinicians \nbelieved they could exert influence on their \npatients\u2019 contraceptive choices through how they \npresented contraceptive information to them. \nClinicians described emphasizing different benefits \nand side effects of methods to influence patient \nchoice and made efforts to make an emotional \nimpression that might motivate behavior change. \nProviders also attempted to use fear and shame to \nencourage women to adopt healthy sexual behav-\niors. Given that providers recognize their influence \nwithin appointments and that biased care is an \nongoing issue in U.S. health care, there is a need to \nbetter understand how providers cope with such \nbias within their appointments and its potential \neffects on patients\u2019 well-being.\nData and Methods\nTo gain such an understanding, we drew on 51 semi-\nstructured interviews with health care providers. \nThis research was part of a larger evaluation of a \nstatewide initiative focused on improving access to \nand delivery of contraceptives, particularly LARCs.\nWe recruited provider participants through \nphone, email, and letters to their workplace. Twenty-\nthree (45%) were physicians, 19 (37%) were nurse \npractitioners, 8 (16%) were nurse midwives, and 1 \n(2%) participant was a nurse. These providers spe-\ncialized in pediatrics, family medicine, obstetrics \ncare, gynecological care, and obstetrics/gynecology. \nForty (78%) participants were from private clinic \nsites, and 11 (22%) were from Title X-serving clinic \nsites. The average age of respondents was 46 years \nold, ranging from 26 to 74. On average, providers \npracticed for 16 years, with a range between 2 and \n47 years in practice. Thirty-eight (74%) providers \nidentified as white/white non-Hispanic, 9 (18%) as \nAfrican American or black, 3 as Indian/Asian (6%), \nand 1 (2%) as Hispanic. Our sample was relatively \nracially diverse compared to the U.S. physician pop-\nulation, where only 6% of physicians are black \n(Xierali and Nivet 2018).\nMost interviews were held in private offices at \nthe clinic sites, and participants provided informed \nconsent prior to interview initiation. Interviews \nlasted, on average, 71 minutes, with a range \nbetween 38 minutes and nearly 2 hours. The inter-\nview inquired about providers\u2019 history practicing \nmedicine, their contraceptive prescribing patterns, \nand details about how they engage in dialogue with \ntheir patients about contraception. At the comple-\ntion of the interviews, the providers received $40 \ngrocery store gift cards.\nA diverse research team of five conducted and \nanalyzed the interviews, collaborating to construct a \nrigorous analytical process. Codes were developed \ninductively, similar to a grounded theory approach, \nthrough an iterative process of reading transcripts to \nidentify patterns and themes, many of which cen-\ntered around questions that we posed to participants, \nsuch as if and how they counsel patients from differ-\nent backgrounds about contraception (Charmaz \n2014). We then entered codes such as \u201cspecific con-\ntraceptive for specific patient\u201d into a qualitative \nresearch software program and further refined them \ninto hierarchical coding schemes using line-by-line \n\n124\t\nJournal of Health and Social Behavior 62(2)\ncoding. Given the semistructured nature of the inter-\nviews, findings for this article around providers\u2019 \nefforts to minimize and manage bias emerged \nunprompted during the interviews when providers \nwere asked about their approaches to patients. We \nthen ensured that all examples fit the codes and \n\u00adsubcodes. To maintain consistency, researchers \nreviewed transcripts in pairs for intercoder reliabil-\nity and continuously memoed throughout the coding \nprocess about relevant themes. All quotes used in \nthis article were the most illustrative examples \namong many of the participants. Data were only \nincluded that were commonplace across partici-\npants. This research was reviewed and approved by \nstate and university Institutional Review Boards. All \nnames are pseudonyms.\nResults\nAs prior research has found, provider bias continues \nto be a poignant issue in contemporary American \nmedicine and within contraceptive counseling (i.e., \nBorrero et\u00a0al. 2009; Dehlendorf et\u00a0al. 2010; Gomez \nand Wapman 2017; Gubrium et\u00a0al. 2016; Yee and \nSimon 2011). For instance, Michelle, an African \nAmerican nurse practitioner, reflects, \u201cYou have a \nHispanic population who they pretty much pop kids \nout left and right.\u201d In addition to ethnicity, providers \nfrequently distinguish patients based on their SES. \nPrya, an Indian physician, explains:\nI mean we have patients who have been popping \nbabies out every year\u2013year and a half and\u2009.\u2009.\u2009.\u2009don\u2019t \nwant birth control [yet] are the ones from a lower \nsocioeconomic status with much lesser education \nwhere they are not thinking about, \u201cHey, I have \nto take care of this child.\u201d\u2009.\u2009.\u2009.\u2009They don\u2019t care if \nthe child finishes school.\nTo better understand what drives this persistent \nbias, how it manifests within the clinical encounter, \nand its potential consequences for women\u2019s repro-\nductive autonomy, we explore how providers nego-\ntiate and manage bias within their contraceptive \ndecision-making and counseling. In doing so, we \nobserve four primary strategies providers imple-\nment to justify, rationalize, minimize, and/or deny \nbias: first, couching their bias within scientific ratio-\nnale; second, employing \u201csafe\u201d biases to minimize \ndifferential counseling; third, standardizing their \napproaches to patients as a means of denying differ-\nential care; and finally, using the tenets of patient-\nled care to distance themselves from responsibility \nfor patients\u2019 final contraceptive decisions.\nCouching Assumptions in Scientific \nRationale\nProviders frequently evade claims of bias in their \ncare by utilizing data, especially around pregnancy \nrisk, to rationalize their differential contraceptive \ncounseling. These data function to produce and \nreinforce biases about patients that then inform pro-\nvider practices. Scientific arguments have long been \nused to encourage and validate eugenicist beliefs \nabout who should reproduce, especially regarding \nwomen of color (Roberts 1997). In its more modern \nand implicit form, such claims hinge on references \nto aggregated data about contraceptive efficacy and \nwomen most at risk for unintended pregnancies. \nThese data prompt providers to anticipate certain \npatient risks and contraceptive needs. Kathleen, a \nwhite nurse midwife, explains:\nYoung women I am more likely to dissuade from \nusing pills [b]ecause of research.\u2009.\u2009.\u2009.\u2009Like I have \nheard anything from 84 to 89% [efficacy]\u2009.\u2009.\u2009.\u2009I \nmean I have been here for a long time and at \ntimes it really feels Third World and (pauses) just \nit seems like especially with teenagers\u2009.\u2009.\u2009.\u2009where I \nworked before, that [pregnancy] was almost \ncelebrated. It wasn\u2019t frowned upon or necessarily \nconsidered (pauses) bad.\nKathleen cites statistical evidence of the pill\u2019s effi-\ncacy among teenagers to express why she discour-\nages pill use among these patients. However, she \nfollows that evidence with subjective, value-laden \nassertions about teen pregnancy being implicitly \n\u201cThird World\u201d and \u201cbad,\u201d thus requiring teens to be \ncounseled against using less effective methods. In \nrooting this decision-making in scientific rationale \nvis-\u00e0-vis the data she cites, Kathleen overlooks and \njustifies her persuasive counseling approach without \nacknowledging its inherent bias around pregnancy \nintendedness among teenagers that may also moti-\nvate her decision-making. Furthermore, Kathleen\u2019s \nreference to the \u201cThird World\u201d also speaks to class- \nand race-based assumptions about \u201cbad\u201d nonnorma-\ntive pregnancies being antithetical to Western, \u201cFirst \nWorld\u201d childbearing expectations.\nPublic health data that identify some women as \nmost \u201cat risk\u201d for unplanned pregnancies prime \nproviders to preemptively conclude which contra-\nceptives may benefit specific patients. Robert, a \nwhite nurse practitioner, loosely references these \ndata on pregnancy risk and contraceptive compli-\nance to explain why he is more likely to steer young \npatients of low SES toward LARCs. He remarks:\n\nManzer and Bell\t\n125\n[M]y appointment would say \u201cTitle X\u201d and then \n\u201cbirth control\u201d would be the thing that\u2019s after \nit\u2009.\u2009.\u2009.\u2009and then I would steer them.\u2009.\u2009.\u2009.\u2009[W]e\u2019re a \nlittle biased towards trying to get one of \nthe\u2009.\u2009.\u2009.\u2009LARCs in\u2009.\u2009.\u2009.\u2009our younger Title Xs just \nbecause\u2009.\u2009.\u2009.\u2009they are statistically at a higher risk \nfor becoming pregnant\u2009.\u2009.\u2009.\u2009or for not understanding \nand following through on their birth control.\nRobert conducts a form of \u201csocial triaging,\u201d a practice \ncommon across participants, where he draws preemp-\ntive conclusions about patient risk and contraceptive \nappropriateness based on a patient\u2019s demographic \ntraits relative to national trends on pregnancy risk and \ncontraceptive use. These data become institutional-\nized knowledge or a \u201ccommonsense\u201d way of evaluat-\ning patients from specific backgrounds.\nIn this way, Robert appears to use what he \nbelieves to be data-driven decision-making about \nwho is \u201cstatistically\u201d most at risk for unintended \npregnancy as rationale to \u201csteer\u201d some patients to \nuse LARC methods rather than prioritizing his \npatients\u2019 preferences. By justifying his counseling \napproach using statistics, Robert validates his dif-\nferential counseling without having to confront the \ninequity of this approach that could limit some \nwomen\u2019s pregnancies. In doing so, Robert does not \nquestion why some patients may be less compliant \nor knowledgeable about contraception, neglecting \nprofound structural factors that may be causing \nthese differences. Moreover, Robert\u2019s comments \nhave an implicitly racial undertone. Title X family \nplanning clinics disproportionately serve women of \ncolor, suggesting that such steering for his \u201cTitle \nXs\u201d may disproportionately affect racial minorities. \nIn these ways, interpersonal interactions within the \nclinical encounter can reflect and produce systemic, \ngendered racism around reproductive equity and \ncontraceptive access.\nWhereas Robert loosely cites statistical evi-\ndence that some women are at greater risk of unin-\ntended pregnancy, most providers first acknowledge \nunintended pregnancy can happen to anyone and \nthus remark that all reproductive-age women bene-\nfit from contraceptives. However, turning to demo-\ngraphic information that differentiates pregnancy \nrisk levels across patient populations allows provid-\ners like Peter, a white physician, to justify some \nwomen\u2019s greater need for contraception. He states:\nI think unintended pregnancy breaks all socio\u00ad\neconomic barriers, but it probably is more \nprevalent in the more populated, inner city, \nlower SES neighborhoods than it is in [names \ntwo \nwealthy, \nlargely \nwhite \nsuburban \nneighborhoods]\u2009.\u2009.\u2009.\u2009I think people who are in \nfamilies where career and education are valued \nvery high, I think there is a lot more education \nabout \nunintended \npregnancy \nand \nsexual \npromiscuity.\u2009.\u2009.\u2009.\u2009It doesn\u2019t mean [unintended \npregnancy] doesn\u2019t happen. It just means it\u2019s \nprobably less typical.\nOn the one hand, Peter acknowledges that unin-\ntended pregnancy affects women from all back-\ngrounds but infers that unintended pregnancies are \nnonnormative and more prevalent among poor \nurban women whose family values around educa-\ntion, career, and sexual behaviors are implicitly less \nsocially acceptable. Peter\u2019s sense of who is more \nlikely to experience an unplanned pregnancy is \nfounded in cultural assessments of patient behavior \nthat his use of data-oriented terms like prevalence \nallows him to mask. Importantly, Peter\u2019s experience \nshows just how cultural competency, in this case, \nthe ability to acknowledge that some families value \nneoliberal goals of wealth and career while others \nmay not, can reinforce stigma toward patients while \nneglecting the structural limitations that create gaps \nin social, economic, and health outcomes across \npopulations (Metzl and Hansen 2014).\nWhen asked how to reduce unintended preg-\nnancy rates, Peter reemphasizes the differences \namong populations affected by unintended preg-\nnancy and who should ostensibly be targeted for \ninterventions that encourage LARC methods, \nsaying:\nPut contraception in the drinking water, put it \ninto the ventilation systems.\u2009.\u2009.\u2009.\u2009If that\u2019s not \nconsidered ethically appropriate, then [do] \nsomething like making it a statewide [LARC] \ninitiative\u2009.\u2009.\u2009.\u2009that\u2019s based on the literature and the \nresearch they\u2019ve done. If you knew that 100% of \nunintended pregnancies took place in [these two \nlow SES, urban zip codes], then yes, I would say \ntarget [those two zip codes].\u2009.\u2009.\u2009.\u2009But it doesn\u2019t \nmean we shouldn\u2019t be educating everybody\u2009.\u2009.\u2009.\u2009but \nthe way you deliver that message may be \ndifferent. The way you sell alcohol to inner cities \nand the way you market and sell alcohol to the \nrich is different. It\u2019s about what attracts their \nattention.\nPeter suggests that everyone benefits from informa-\ntion about LARCs, but in this case, by citing evi-\ndence of pregnancy risk from the medical literature, \nhe can ignore his use of coded assertions about for \n\n126\t\nJournal of Health and Social Behavior 62(2)\nwhom and when pregnancy is socially acceptable. \nPeter also reflects rhetoric common in LARC pro-\nmotion efforts that encourage providers to empha-\nsize the most effective methods based on presumed \npregnancy risk rather than patient preferences or the \npatients\u2019 personal risk of side effects related to con-\ntraceptive use (Bertotti, Mann, and Miner 2020; \nGomez et\u00a0al. 2014; Littlejohn 2013; Littlejohn and \nKimport 2017).\nInterestingly, Peter likens medical advice to sell-\ning a product such as alcohol. Ironically, reflecting \nresearch (e.g., Moore, Williams, and Qualls 1996) \nregarding alcohol propaganda, these marketing \nstrategies are not only rooted in disparate stereo-\ntypes, but they also further perpetuate such images \nand inequalities. Placing his reasoning for the vari-\nous emphases in evidence, however, is a mecha-\nnism that Peter uses to support and justify his biased \ncommunication and advice. Similarly, although \nPeter\u2019s reference to putting contraception in the \ndrinking water or ventilation system is likely face-\ntious, such comments reflect the near universal \nappeal some providers see in limiting unintended \nbirths among marginalized women without suffi-\ncient recognition of how such measures could \nundermine their patients\u2019 reproductive autonomy. \nIndeed, although explicit eugenicist policies have \nbeen abandoned, the logic that fostered them can \nunconsciously persist in some providers\u2019 attitudes \nand subsequent counseling about which contracep-\ntive methods are most appropriate for which patient \npopulation (Luna and Luker 2013).\nFor instance, to limit the number of unintended \npregnancies, providers like Mark, a white physi-\ncian, do not explicitly tell their marginalized \npatients not to get pregnant; rather, they more \nstrongly encourage LARC use among those popula-\ntions. Mark explains:\nThe other thing that really frustrates the crap out of \nme\u2014excuse me\u2014is the patient who comes in and \nsays, \u201cNo, I don\u2019t want to be pregnant, but I don\u2019t \nuse any birth control\u201d\u2009.\u2009.\u2009.\u2009you want to take that \nperson and shake them. Some of it is ignorance, \nsome of it is cultural\u2009.\u2009.\u2009.\u2009I\u2019m not telling you not to \nget pregnant. I\u2019m telling you, \u201cIf you don\u2019t want to \nbe pregnant, you should do something to prevent \nit\u201d\u2009.\u2009.\u2009.\u2009I hate to say this, but I think it\u2019s a fact that \nthere is a strong correlation between African \nAmerican and low socioeconomic status\u2009.\u2009.\u2009.\u2009I think \nthe white population seems to be overall more \nlikely to be knowledgeable, compliant, etc. And \nit\u2019s not meant to be a racist statement (laughs).\nClearly, Mark is frustrated when patients, particu-\nlarly black women of low SES, state that they do not \nwant children yet do not use contraception to pre-\nvent such a pregnancy. The emotion evident in \nMark\u2019s quote alongside its content reveals the \nstigma that he holds about black low-SES women \nand runs counter to traditional views of medical pro-\nfessionals as unemotional, objective, and value neu-\ntral. In effect, Mark debates whether to blame the \nindividual women and their cultural preferences for \ntheir inconsistent contraceptive use or recognize \nthat rather than being a moral failing, their \u201cigno-\nrance\u201d about birth control is rooted in structural \ninequality around education and health care access.\nIn situating these women\u2019s attributes in statistics \n(i.e., \u201ccorrelation\u201d), Mark applies race and class \nassumptions without the implication of being \nbiased. Indeed, he explicitly rejects the notion that \nhe made a \u201cracist statement.\u201d As Metzl and Hansen \n(2014) cautioned, teaching providers to be cultur-\nally competent may promote stereotypes about \nwhole demographic groups when providers fail to \nlink structural barriers faced by these populations to \ntheir health outcomes and decision-making.\nDifferentiating Patient Care Using  \nSafe Biases\nIn addition to couching bias within scientific ratio-\nnale, providers employ what we term safe biases as \na minimizing strategy. Safe biases are like coded \nlanguage in that both convey multiple layers of \nmeaning and target a specific group of people to \nwhom the term becomes associated, such as thug or \ncriminal. In this context, safe biases include using \nmedically acceptable, often professional terms like \nMedicaid, at risk, insured, or underserved to safely \nexpress race-, class-, or gender-based beliefs about \npatients and their behaviors.\nFor example, Kevin, a white physician, uses \nsafe biases to implicitly link pregnancy risk to \nwomen of low SES. He explains, \u201cEducation is \nhugely important. You need to care for yourself \nbefore you can care for somebody else.\u2009.\u2009.\u2009.\u2009They\u2019re \njust completely not educated [about birth con-\ntrol]\u2009.\u2009.\u2009.\u2009neighborhoods where four generations and \nthree generations have been on public assistance.\u201d \nKevin reflects common, though contested, asser-\ntions that unintended pregnancies more frequently \noccur among less educated women, have intergen-\nerational effects, and bear social costs related to \ninevitable welfare expenditures\u2014a frequent justifi-\ncation for LARC interventions (see Gubrium et\u00a0al. \n\nManzer and Bell\t\n127\n2016; Thomas and Monea 2011). Kevin\u2019s use of \n\u201cpublic assistance\u201d has negative connotations that \nare classed and raced. The phrase is often code for \nlow-income communities and communities of color \npresumed deficient or unproductive (Gold 2014). In \nthis case, Kevin associates public assistance with a \nneed for contraceptive education, explicitly stating \nthat these women cannot care for themselves or \ntheir children. By neglecting the structural inequali-\nties that force families onto welfare rolls, Kevin \ninvokes political and public health discourses that \njustify state-sponsored efforts to limit reproduction \nthrough contraception and welfare reform (Gold \n2014). By using safe biases like uneducated and \npublic assistance, Kevin can openly discuss the \nmerits of limiting the births of welfare recipients \nwithout taking responsibility for the prejudicial \nthinking that undergirds such beliefs. His statement \nalso demonstrates how readily discriminatory \nassumptions about patients can subtly pervade con-\ntraceptive counseling practices.\nStephanie, a white nurse, further elaborates on \nthese assumptions when differentiating patients by \ninsurance type\u2014a safe and common medical refer-\nence used to distinguish patients by social class:\nAnd we cater to a specific population because we \nare private insurance only.\u2009.\u2009.\u2009.\u2009Because we don\u2019t \nhave your Medicaids or your more underserved \npopulations [but] I did some volunteering \nrecently [and] the clientele are\u2009.\u2009.\u2009.\u2009people who \ndon\u2019t have private insurance.\u2009.\u2009.\u2009.\u2009Well, your \nunderserved population. Yeah. I don\u2019t like the \nword \u201cpoor\u201d (laughing)\u2009.\u2009.\u2009.\u2009They\u2019re young and \ntypically if you\u2019ve had one unplanned pregnancy, \nyou\u2019re more likely to have another. So it\u2019s just \nkind of the way the cycle goes.\u2009.\u2009.\u2009.\u2009But if they\u2019re in \nthe underserved populations, it\u2019s not going to \nhelp anybody to have multiple children [and] \nthey\u2019re not going to buy condoms, they\u2019re still \ngoing to have sex.\u2009.\u2009.\u2009.\u2009And so it would make the \nmost sense to put themselves in the easiest \nsituation possible for long term where they don\u2019t \nhave to worry about having a baby [by using] your \nMirenas, your ParaGards [IUD brands]\u2009.\u2009.\u2009.\u2009which \nis what I would recommend.\nStephanie not only discusses SES indirectly by ref-\nerencing Medicaid versus privately insured patients, \nshe also suggests that poor young women have \nunprotected sex that can result in multiple, likely \nunplanned, children. Stephanie makes a semantic \nmove by substituting the more loaded term poor for \nsafer ones like underserved and Medicaids to \ndescribe these women (Teal et\u00a0al. 2010). Yet despite \ndoing so, Stephanie is still more likely to prescribe \nan IUD to a poor woman because shifting her lan-\nguage simply obscures the bias rather than confronts \nit. Using safe biases allows for this more explicit \nbias to be managed, justified, and discreet in her \nassertions that IUDs are ideal for women of low \nSES. Moreover, recommending LARCs to help put \nwomen in the \u201ceasiest position possible\u201d neglects \nthe structural factors that put these women in such \ndifficult, marginalized positions in the first place.\nProviders often use these safe biases to con-\nstruct and explain their different communication \napproaches with patients. Rather than candidly dis-\ncussing issues of race or class with the interviewer, \nDrew, a white physician, explains these different \napproaches by first using patients\u2019 education levels \nand later more explicitly tying education to SES \nand culture to race:\nI mean the \u201cless educated\u201d person\u2009.\u2009.\u2009.\u2009One\u2009.\u2009.\u2009.\u2009 \nthey\u2019re not exposed to the information as much\u2009.\u2009.\u2009.\u2009 \nI find that the less educated people\u2009.\u2009.\u2009.\u2009don\u2019t \nunderstand well when I try and explain \n[contraception]\u2009.\u2009.\u2009.\u2009whereas somebody who\u2019s a \nhigher-educated person\u2009.\u2009.\u2009.\u2009they obviously get [it]. \nSo it makes it easier to give a scientific answer to \nthem whereas for the lower socioeconomic, less \neducated people, it is a little harder\u2009.\u2009.\u2009.\u2009I speak in a \nvery different way\u2009.\u2009.\u2009.\u2009at least our African American \npopulation is probably a lower socioeconomic \nstatus than our average Caucasian. So it\u2019s hard to \ntell whether it\u2019s the cultural part or the \nsocioeconomic part\u2014and\u2009.\u2009.\u2009.\u2009I try not to bring my \nmorals into when I counsel people. But we had \nsingle parents\u2009.\u2009.\u2009.\u2009that can\u2019t take any more kids \nand that the government is paying for \nMedicaid\u2009.\u2009.\u2009.\u2009and [the kids are] not getting raised \nwell.\u2009.\u2009.\u2009.\u2009Sometimes people need help, but we \ncan\u2019t enable them for so long that they can\u2019t help \nthemselves.\nDrew\u2019s statements are demonstrative of color-blind \nracism, where individuals develop verbal strategies \nto conceal directly racist statements (Bonilla-Silva \n2002). By virtue of using safe biases common \nwithin the medical profession like education and \nMedicaid with coded language such as cultural, \nDrew attempts to avoid making directly racist, \n\u00adclassist, and moral statements about patients. \nParadoxically, Drew demonstrates reflexivity by \nsaying he does not want his personal morality to \nshape his contraceptive counseling. However, he \nargues that single, presumably black, parents with \n\n128\t\nJournal of Health and Social Behavior 62(2)\npublicly funded insurance neither need additional \nchildren nor parent their current children well. As \nRJ advocates and scholars caution, providers may \nprescribe contraception based on a patient\u2019s per-\nceived ability to parent in culturally acceptable \nways\u2014limiting the number of births to marginal-\nized women by recommending LARCs or steriliza-\ntion (Higgins et\u00a0 al. 2016; Roberts 1997). Such \nbeliefs, no matter how implicitly held, can creep \ninto contraceptive counseling practices that affect \nthe reproductive autonomy of marginalized groups \nwhile privileging white patients of high SES.\nStandardizing Contraceptive Counseling \nacross Patients\nMedical care has become increasingly standardized \nin response to nationwide reports of inconsistent \ntreatment and outcomes (Smedley et\u00a0al. 2002). This \nstandardization may encourage objectivity and neu-\ntralize biased clinical encounters by offering similar \ncontraceptive counseling to each patient. However, \nas reflected through Mark\u2019s quote, this approach \nmay actually promote prejudicial thinking. He says:\nThen it\u2019s an issue of again getting back to the one \nkey question: Do you want to have a baby in the \nnext year or two?\u2009.\u2009.\u2009.\u2009And there are a lot of women \nwho are single and who want to have a baby and \nagain, some of them are teenagers who am I to \nsay they shouldn\u2019t, but they probably shouldn\u2019t. \nSome of them are professional women who, \n\u201cYeah, yeah, I want to have a baby. And I don\u2019t \ncare who the father is.\u201d\nMark references an effort to reduce unintended \npregnancy, in which more than 30 U.S. states train \ntheir providers to ask women ages 18 to 50 about \ntheir pregnancy desires at each appointment (Power \nto Decide 2019). Despite using this standard \napproach, Mark\u2019s statement reveals his personal \nbeliefs about the acceptability of pregnancies \namong single mothers, comparing \u201cprofessional\u201d \nwomen to teens. However, asking all women about \ntheir pregnancy desires allows Mark to feel he is \ninquiring objectively about women\u2019s family plan-\nning despite their background.\nOften it is the providers who most decisively \ndescribe their standardized approaches who are \nmore likely to unconsciously encourage \u201cat-risk\u201d \npatients to use LARCs relative to their more privi-\nleged patients. This is reflected in the experiences \nof Vanessa, a white physician:\nBut I think if you are standard in the way that you \nask everybody, it\u2019s for me personally the best \nway to try to eliminate any bias.\u2009.\u2009.\u2009.\u2009If I ask every \nkid the same thing in the same hopefully sensitive \nway, then I am not\u2014I am hopefully trying to \neliminate any personal biases or perceptions\u2009.\u2009.\u2009.\u2009a \nlot of medicine is trying to recognize quickly \ncertain things or trying to think about what your \npatients are most at risk for\u2009.\u2009.\u2009.\u2009I really think almost \nevery kid would benefit from a LARC\u2009.\u2009.\u2009.\u2009so I \nreally try to recommend them\u2009.\u2009.\u2009.\u2009[they] are really \nthe way to go.\nAs evinced above, providers are sensitive to bias \nand seek to manage it by standardizing patient care. \nIndeed, Vanessa suggests that all patients engaging \nin certain behaviors would benefit from LARCs. \nHowever, health care constructions of \u201crisky behav-\niors\u201d are inherently raced and classed rather than \nunbiased, as this provider suggests (Metzl and \nKirkland 2010). As other research affirms, when \ncontraceptive counseling efforts go beyond ensur-\ning access for all women to then promoting LARC \nuse among specific populations, marginalized \nwomen are more likely to experience restricted con-\ntraceptive options that can serve to undermine \nreproductive autonomy (Gomez et\u00a0al. 2014).\nYet other providers describe their contraceptive \ncounseling as consistent across all patients and \nactively deny race- or class-based discrimination. \nDayna, a white physician, says:\nI try and give the same exact word-for-word to \neverybody\u2009.\u2009.\u2009.\u2009I don\u2019t say, \u201cOh, you\u2019re black. \nLet\u2019s just talk about this.\u201d Or, \u201cOh, you\u2019re white, \nwe\u2019re just going to talk about this.\u201d\u2009.\u2009.\u2009.\u2009[A] good \nportion of [my current patients] have a college \neducation\u2009.\u2009.\u2009.\u2009so I will speak to them a little bit \nmore medically\u2009.\u2009.\u2009.\u2009I don\u2019t want to say \u201cbigger \nwords\u201d\u2014but \nI\u2019ll \nuse \nmore \ndata\u2009.\u2009.\u2009.\u2009in \nresidency\u2009.\u2009.\u2009.\u2009that was more like an inner city \noffice, usually lower socioeconomic status, \nmaybe not high school education, a little bit \nmore of like an urban population\u2009.\u2009.\u2009.\u2009I don\u2019t know \nif this is maybe like a weird prejudice, but I\u2019ll try \nto use examples with like that more urban kind \nof undereducated [patient]. I think I had a \n19-year-old girl a few months ago who was like, \n\u201cI just don\u2019t care if I get pregnant.\u201d I was like, \n\u201cThat\u2019s an awful attitude.\u201d\nDespite claims to the contrary, Dayna does adjust \nher counseling strategies based on patient race and \n\nManzer and Bell\t\n129\nclass. Indeed, she uses implicitly racialized lan-\nguage like urban and inner city, two terms that are \ncode for black, poor patients. Dayna linguistically \nenacts color-blind racism both by patronizing black \npatients through her differential counseling and by \nreferencing race indirectly when speaking about \nthese exchanges (Bonilla-Silva 2002). By asserting \nher color-blind, standardized approach, Dayna\u2019s \nbiased counseling can continue unabated without \nintervention. Moreover, calling her teen patient\u2019s \nattitude about pregnancy ambivalence \u201cawful\u201d \nreveals just how strongly some providers feel about \ndiscouraging nonnormative pregnancies. Such sen-\ntiments may affect providers\u2019 contraceptive coun-\nseling and subsequently, women\u2019s contraceptive \nchoice.\nFacilitating Choice with Patient-\nCentered Care\nPatient-centered care, like standardization, is a \nwidely used model designed to improve health care \noutcomes (Epstein and Street 2011). Patient cen-\nteredness encourages patients to guide clinical \nexchanges and strives for providers to support \npatient decision-making with a sense of empathy \nand collaboration (Epstein and Street 2011; Stewart \net\u00a0al. 2000). Most providers inherently support these \ngoals. Nicholas, a white physician, explains, \u201cI \nalways tell people, \u2018I am a democracy\u2009.\u2009.\u2009.\u2009not a dicta-\ntorship.\u2019\u201d However, unacknowledged biases under-\nmine patient-centered care. In fact, most providers \nimplement tenets of patient-centered care by advo-\ncating for women\u2019s contraceptive choice. However, \nthe belief that they are patient centered in their care \noften prevents providers from confronting their per-\nsonally held biases and recognizing how such biases \ninfluence their counseling practices (Stewart et\u00a0al. \n2000). Nicholas continues saying:\nIf [the patient] asks me my personal preference, \nI will always tell them\u2009.\u2009.\u2009.\u2009LARCs.\u2009.\u2009.\u2009.\u2009But I would \nnever force somebody\u2009.\u2009.\u2009.\u2009I would say\u2009.\u2009.\u2009.\u2009\u201cWell, if \nyou\u2019re letting me pick, I want you to have a \nMirena [a brand of IUD],\u201d but I\u2019m giving you all \nthe options so you can make an informed \ndecision about what\u2019s going to work best for \nyou, and your lifestyle, and be the most \neffective\u2009.\u2009.\u2009.\u2009I\u2019m going to show you that I am \nreally looking out for your best interest.\nPatient-centered approaches may prioritize patient-\ndirected contraceptive decision-making; however, in \nimplementation, providers\u2019 beliefs about who should \nuse LARCs subvert such outcomes. Nicholas clearly \ndescribes exerting his clinical authority over patients\u2019 \ndecisions by actively encouraging the use of a LARC \nfor women whose \u201clifestyles\u201d he believes warrant \nLARC use. However, Nicholas can deny unduly \ninfluencing the contraceptive decisions of specific \npatient populations by arguing that his care is patient \ncentered and that he is a \u201cdemocracy.\u201d\nFurthermore, Nicholas\u2019s reference to lifestyle \nevokes a similar meaning as culture, especially \nwhen racially connoted. Culture is often a proxy for \nrace, and lifestyle is a common proxy for risky \nbehaviors attributed to various racial or social \ngroups. Shim (2005:415) argues that these \u201ccultural \nconstructions of race resonate particularly well with \nhealth promotion and risk reduction mandates. \nViewing race through the prism of culture provides \nan extremely convenient entry point for targeting \nrisky lifestyle behaviors,\u201d which in this case refers \nto unprotected sex and pregnancy risk. Out of con-\ncern for women\u2019s pregnancy risk, their lifestyle \nchoices, and best interest, Nicholas is more likely to \ncounsel these women to use LARCs.\nEven providers who openly advocate for \npatients to obtain their contraceptive choice use \npersuasive counseling techniques to influence those \ndecisions. For instance, Beverly, a white nurse \npractitioner, reflects, \u201cI read the consultation notes \nof the patients that have babies\u2009.\u2009.\u2009.\u2009to make sure that \nthey got their birth control of choice.\u201d When asked \nto describe her counseling approach in these cir-\ncumstances, Beverly explains:\nWell, if they have multiple kids at home, I really \ntry to get them to get their tubes tied\u2009.\u2009.\u2009.\u2009I\u2019m like, \n\u201cLook, \nnobody \nlikes \nkids \nthat \nmuch \n(laughs).\u201d\u2009.\u2009.\u2009.\u2009Or if they won\u2019t do that, I really try \nto get them to do a LARC\u2009.\u2009.\u2009.\u2009so that they don\u2019t \nhave to keep running in here\u2009.\u2009.\u2009.\u2009you can just tell \nthe poverty.\u2009.\u2009.\u2009.\u2009You don\u2019t need any more kids.\nBeverly is adamant that her patients make contra-\nceptive decisions, a core tenet of patient-centered \ncare. However, Beverly is more likely to recom-\nmend sterilization or LARC methods to impover-\nished women with multiple children. By claiming to \nsupport patient choice, Beverly does not confront \nhow her beliefs may subsequently undermine this \nchoice. Through such examples, it becomes appar-\nent that efforts to be patient centered are impaired \nby persistent provider bias. It is also evident that \nclaims of patient centeredness may skew providers\u2019 \nperceptions of their own care and rather than \nimprove health care equity, actually serve as a \n\n130\t\nJournal of Health and Social Behavior 62(2)\nmeans through which providers can deny, and thus \nperpetuate, biased care.\nProviders do recognize the complex factors that \ncan affect not only how they make prescribing deci-\nsions but also if decisions can realistically be \npatient centered. For instance, structural issues \nwithin medicine can dictate what methods patients \ncan access and which providers can prescribe. Ruth, \na white nurse practitioner, shares:\nFor the most part, I mean nobody pushes any \nservice or anything on a person. We\u2019re not in that \nday and age nowadays. But at the same point in \ntime, if here are your choices: X, Y, Z. You\u2019ve \ngot to choose X, Y, Z because that\u2019s what we \nhave available to you. That\u2019s the thing that\u2019s \nmost easy for you to get X, Y, Z. \u201cOh, you want \nH? Oh, then we have to do a little bit more work \non that. Hm. I have 15 minutes.\u201d So and it all \nplays a role. So when you\u2019re looking especially \nMedicaid and what Medicaid will pay for.\nRuth describes how the structure of medicine can \nlimit providers\u2019 ability to enact patient-centered \ncontraceptive care due to time constraints, over-\nscheduling, stocking issues, and insurance-related \nrestrictions. She also notes that patients of lower \nSES who have Medicaid may be particularly \nimpacted by these constraints. Angela, a white phy-\nsician, adds:\nIt\u2019s really easy to see things in black and white \nand then as soon as you start to see the gray\u2009.\u2009.\u2009.\u2009it \nbecomes harder to navigate the patient \nencounters\u2009.\u2009.\u2009.\u2009my goal is always to assess what \nthe patient\u2019s desires are\u2009.\u2009.\u2009.\u2009helping them decide \nwhat\u2019s right for them. That\u2019s the ideal. Sometimes \nit\u2019s easier said than done.\u2009.\u2009.\u2009.\u2009Some of it are issues \nwith the patient\u2009.\u2009.\u2009.\u2009and sometimes it\u2019s the \nsystem\u2009.\u2009.\u2009.\u2009ideally we would have plenty of time \nto go through all of the methods, whatever the \nwoman picked that day, we could then provide \nher with. But there are lots of barriers to doing \nthat.\nImportantly, as other researchers have noted, \nstressed providers in ambiguous or \u201cgray\u201d situations \namid rushed interactions are more likely to rely on \nstereotypes when making clinical decisions \n(Burgess et\u00a0 al. 2007). As such, although Angela \ndoes not describe explicitly biased beliefs, she and \nRuth describe an environment in which such biases \nthrive.\nAs Ruth suggests, times have indeed changed; \nproviders may be less likely to explicitly coerce \nwomen to use contraceptives. However, we find \nthat this coercion has simply taken on a more sub-\ntle, \u201cgray\u201d form, easily couched in scientifical ratio-\nnale, made implicit using safe biases, and rationalized \nthrough standardized or patient-centered approaches \nto care.\nDiscussion\nRecognizing how health care providers grapple with \nbias offers opportunities to better understand the \nbiases they hold, how bias manifests within appoint-\nments, and how patients\u2019 autonomy and outcomes \nmay be affected. Elucidating these management \nstrategies simultaneously reveals not only provid-\ners\u2019 biases but also the social forces that engender \nsuch discrimination and inequalities within the clin-\nical encounter.\nMost providers aspire to offer equal access to all \ncontraceptive varieties in alignment with their \npatients\u2019 preferences. To do so, they employ a range \nof strategies to navigate and minimize potential \nbias that may undermine this goal\u2014the most salient \nof which we found to be couching bias within sci-\nentific rationale, speaking about these risks using \nsafe biases that avoid the explicitly raced and \nclassed language, adopting standardized approaches, \nand employing patient-centered care. However, we \nfind that providers\u2019 efforts to provide equitable care \nparadoxically perpetuate this bias.\nBiased care is challenging to confront, espe-\ncially, as researchers note, providers are less \nexplicit in their biases than in prior decades (Blair, \nSteiner, and Havranek 2011; Bonilla-Silva 2002; \nZestcott et\u00a0al. 2016). The insidiousness of bias \nwithin health care contexts, a setting where \n\u00addecision-making is often lauded for being objec-\ntive, makes intervening in health care contexts \ncomplicated (Burgess, Fu, and van Ryn 2004; Teal \net\u00a0al. 2010). Providers, like the general population, \nare now more likely to use subtle biases. Moreover, \nstereotyping may be exacerbated by the fast-paced, \nhigh-stress environment in which providers work \nthat all but necessitates clinicians to make general-\nizations about various groups because of time con-\nstraints (Burgess et\u00a0al. 2004). These generalizations \npersist even as providers acknowledge that unmet \nsocial needs result in worse health for all Americans \nand are aware that structural factors, such as restric-\ntive insurance policies, undermine health outcomes \n(Metzl and Hansen 2014).\n\nManzer and Bell\t\n131\nProviders may be less likely to acknowledge \npersonal biases because they expect themselves to \nbe unaffected by patients\u2019 sociodemographic char-\nacteristics and often have a difficult time accepting \nthat they harbor prejudicial attitudes (van Ryn and \nFu 2003). Indeed, as the findings demonstrate, pro-\nviders deny overtly coercing women to use contra-\nception. For example, most participants clearly \nreject assertions of biased contraceptive counseling \nbecause they believe in women\u2019s right to choose \nand feel they have found unbiased ways to facilitate \nthis choice within their appointments.\nMedicine lends itself to denying individual pro-\nvider bias as well as systemic discrimination. As a \nprofession and scientifically oriented discipline, \nmedicine garners much of its prestige from the \nbelief that it is both an objective science as well as \nan altruistic profession (Freidson 1970). As our \nfindings demonstrate, health care providers can \ndeny prejudicial counseling by citing documented \ntrends in contraceptive use, pregnancy risk, and the \nmeasurable consequences for unplanned pregnan-\ncies among various sociodemographic groups that \njustify targeted counseling behaviors. This target-\ning disproportionately affects black and Latinx \nwomen as well as poor, young, and unmarried \nwomen, who are most often cited as at risk for such \npregnancies. Indeed, in other areas of medicine, \nproviders similarly rely on objective measurements \nand medical terminology such as obese quantified \nby a patient\u2019s body mass index rather than the more \nsubjective term fat to disguise what researchers \nhave established are often moral judgments about \npatients\u2019 individual behaviors (Murray 2008). \nDoing so allows such judgments to operate without \nquestion and often lends them medical legitimacy.\nAs Zestcott and colleagues (2016) find, addi-\ntional work is needed to understand the conditions \nin which provider bias is likely to affect communi-\ncation with patients and subsequent prescribing \ndecisions. The participants elucidated several perti-\nnent structural limitations within medicine, like \nshort appointments and increasingly large patient \nrosters. These conditions require clinicians to con-\nduct \u201csocial triaging,\u201d assessing risk based on infor-\nmation drawn from provider-initiated questions and \npatient sociodemographic traits rather than from \nquality dialogue. However, doing so produces and \nreinforces stereotypes about patients that become \n\u201ccommonsense\u201d ways of prescribing contraception \nto various demographic groups. For instance, as in \nprior research, providers described a preference for \nrecommending LARCs to younger women, women \nof color, and women of lower SES out of concern \nfor their pregnancy risk (Borrero et\u00a0 al. 2009; \nDehlendorf et\u00a0al. 2010). Rather than acknowledge \ntheir individually held biases, providers evoked data \nor used safe biases to disregard and/or justify their \nown assumptions about pregnancy that influenced \nmany of their prescribing and counseling decisions. \nThis finding held across approaches\u2014whether stan-\ndardizing or implementing patient-centered care.\nImportantly, by ascribing risk to behaviors, pro-\nviders fail to position women\u2019s behaviors within \nthe structural contexts in which they are situated. \nHowever, making such connections is challenging. \nLike Chapman and colleagues (2013) find, provid-\ners with insufficient time may resort to filling in \nknowledge gaps about their patients with stereotype- \nbased information like race or relationship status \nwhen evaluating pregnancy risk and contraceptive \nappropriateness. Such findings have significant \nconsequences for women\u2019s reproductive auton-\nomy, especially that of women of color and lower \nSES women.\nIndeed, continued bias among providers who \nprescribe contraception results in patient profiling \nin the promotion of LARCs as a first-line or pre-\nferred method (Dehlendorf et\u00a0al. 2018; Gomez et\u00a0al. \n2014). However, denying bias also neglects the \nwell-documented history of reproductive abuses \nperpetrated against marginalized women in the \nUnited States. We observed that well-intentioned \npractitioners, some of whom even cited this history, \nstill held deep-seated judgments about whether \n(more) children are acceptable for certain women. \nImportantly, many of these stereotypes about who \nshould reproduce are artifacts of the United States\u2019s \nracialized and class-based policies and practices \nthat have stratified reproduction for decades (Colen \n1986; Luna and Luker 2013; Roberts 1997).\nRather than assuming these biases do not exist, \nclinicians would benefit from acknowledging biases \nand identifying techniques to actively challenge and \nundermine dominant stereotypes. Providers deserve \ntools to help them respond when patients make dif-\nferent choices than they might prefer. One such \nstrategy is to do more within contraceptive initia-\ntives to explicitly address individually held biases \naround reproduction and connect them to the social \nforces that engender them. Additionally, researchers \nshould reflect epistemologically on the impact of \ntheir findings about pregnancy risk and contracep-\ntive use on stoking and legitimizing provider biases \nabout who is fit to parent and which contraceptive is \npreferred. Contextualizing their findings, for \ninstance, may mitigate such effects (Dehlendorf \net\u00a0al. 2018).\n\n132\t\nJournal of Health and Social Behavior 62(2)\nThe persistence and prevalence of bias in health \ncare, and in contraceptive counseling more specifi-\ncally, has significant implications for women\u2019s \nreproductive health and the provider\u2013patient rela-\ntionship. Like other research has found, we note that \nprovider biases affect some women\u2019s health-seeking \nbehaviors over the life course. Indeed, many women \ninitiate contraception relatively young. Adverse \nhealth care experiences during adolescence and \nyoung adulthood can negatively impact future \nhealth-seeking behavior, with important implications \nfor health later in life (Gomez and Wapman 2017).\nNotably, the findings presented here are limited \nto the sample from which they are drawn and can-\nnot be used to generalize about provider bias in \nother settings or among other provider groups. \nAdditionally, it is beyond the scope of this article to \nmore systematically explore the role of specific \npatient or provider sociodemographic characteris-\ntics in determining provider counseling and contra-\nceptive prescribing. However, patient race-ethnicity, \nSES, age, and relationship status frequently emerge \nas key justifications for provider decision-making \nthat warrant ongoing attention.\nMoreover, as evinced in the results, we found \nbiased care across all providers despite specializa-\ntion, age, race, or gender. This may be due to the rel-\native homogeneity among clinicians in terms of \nupbringing and class status (see Alsan, Garrick, and \nGraziani 2019; Carrese and Rhodes 1995). Providers \nfrom higher SES backgrounds with similar medical \nschool education may hold shared beliefs about \nparental readiness, parenting, and the appropriate-\nness of specific contraceptives for specific women \n(Stevens 2015). Moreover, providers\u2019 use of preg-\nnancy risk data in their clinical decision-making may \nfurther reinforce these personal beliefs. We also note \nthat white providers appeared to discuss their strate-\ngies to avoid biased care more explicitly relative to \nproviders of color, although all providers did discuss \nthese strategies. The emphasis on race-based bias \ntrainings may hold some explanatory value. White \nproviders learning about discrimination against \nracial minorities may be more inclined to reflect on \ntheir own approaches and more consciously develop \ncoping strategies with which to navigate and mini-\nmize bias. Both findings should be explored further \nin future research to expand on work that currently \nexplores the relationship between race concordance \nand patient care outcomes (see Alsan et\u00a0 al. 2019; \nCarrese and Rhodes 1995).\nProvider bias remains a poignant issue in \nAmerican health care as well as within contracep-\ntive counseling. As RJ activists and scholars \ncontinue to underscore, bias and the social forces \nthat engender it persist and continue to dispropor-\ntionately affect the reproductive autonomy of mar-\nginalized women as well as their health care \ngenerally. Understanding how providers manage \nthis bias is key not only to minimizing differential \ntreatment within medicine but also to reducing \nrelated health disparities and achieving reproduc-\ntive justice for all women.\nAcknowledgments\nWe extend our thanks to Brianna VanArsdale and Virginia \nBerndt for their research assistance and discussion as we \nconceptualized this article. We also thank the anonymous \nreviewers and Miranda Waggoner for their thoughtful and \ninstructive comments on previous versions of the article.\nFunding\nThe authors disclosed receipt of the following financial \nsupport for the research, authorship, and/or publication of \nthis article: This work was supported by a research grant \nfrom a private philanthropic foundation that had no \ninvolvement in the analysis and interpretation of the data \nor the decision to submit the article for publication.\nORCID iD\nJamie L. Manzer \n https://orcid.org/0000-0002-3111- \n5578\nNote\n1.\t\nWe define implicit bias as a set of social stereotypes \nabout groups of people formed and held outside of \none\u2019s awareness, whereas explicit biases are those \nknowingly held (see Chapman et\u00a0al. 2013).\nReferences\nAdler, Nancy E., and Katherine Newman. 2002. \n\u201cSocioeconomic Disparities in Health: Pathways and \nPolicies.\u201d Health Affairs 21(2):60\u201376.\nAlsan, Marcella, Owen Garrick, and Grant C. Graziani. \n2019. \u201cDoes Diversity Matter for Health? Experimental \nEvidence from Oakland.\u201d American Economic \nReview 109(12):4071\u2013111.\nBao, Yuhua, Sarah Fox, and Jose Escarce. 2007. \n\u201cSocioeconomic and Racial-Ethnic Differences in \nthe Discussion of Cancer Screening: \u2018Between-\u2019 \nVersus \u2018within-\u2019 Physician Differences.\u201d Health \nServices Research 42(3):950\u201370.\nBeach, Mary Catherine, Ebony Price-Haywood, Tiffany L.  \nGary-Webb, Karen A. Robinson, Aysegul Gozu, \nAna Palacio, Carole Smarth, et al. 2005. \u201cCultural \nCompetence: A Systematic Review of Healthcare \nProvider Educational Interventions.\u201d Medical Care \n43(4):356\u201373.\n\nManzer and Bell\t\n133\nBertotti, Andrea M., Emily S. Mann, and Skye A. Miner. \n2021. \u201cEfficacy as Safety: Dominant Cultural \nAssumptions and the Assessment of Contraceptive \nRisk.\u201d Social Science & Medicine 270:113547. \ndoi:10.1016/j.socscimed.2020.113547\nBetancourt, Joseph R., Alexander A. Green, J. Emilio \nCarrillo, and Owusu Ananeh-Firempong. 2003. \n\u201cDefining \nCultural \nCompetence: \nA \nPractical \nFramework for Addressing Racial-Ethnic Disparities \nin Health and Healthcare.\u201d Public Health Reports \n18(1):1\u201318.\nBlair, Irene V., John Steiner, and Edward Havranek. \n2011. \u201cUnconscious (Implicit) Bias and Health \nDisparities: Where Do We Go from Here?\u201d The \nPermanente Journal 15(2):71\u201378.\nBonilla-Silva, Eduardo. 2002. \u201cThe Linguistics of Color \nBlind Racism: How to Talk Nasty about Blacks \nwithout Sounding \u2018Racist.\u2019\u201d Critical Sociology \n28(1/2):41\u201364.\nBonilla-Silva, Eduardo. 2010. Racism without Racists: \nColor-Blind Racism and the Persistence of Racial \nInequality in the United States. 3rd ed. Lanham, MA: \nRowman & Littlefield.\nBonilla-Silva, Eduardo, and Gianpaolo Baiocchi. 2001. \n\u201cAnything but Racism: How Sociologists Limit the \nSignificance of Racism.\u201d Race and Society 4(2): \n117\u201331.\nBorrero, Sonya, Eleanor B. Schwarz, Mitchell Creinin, \nand Said Ibrahim. 2009. \u201cThe Impact of Race and \nEthnicity on Receipt of Family Planning Services \nin the United States.\u201d Journal of Women\u2019s Health \n18(1):91\u201396.\nBoyer, Carol A., and Karen E. Lutfey. 2010. \u201cExamining \nCritical Health Policy Issues within and beyond the \nClinical Encounter: Patient\u2013Provider Relationships \nand Help-Seeking Behaviors.\u201d Journal of Health & \nSocial Behavior 1:80\u201393.\nBurgess, Diana J., Steven S. Fu, and Michelle van Ryn. \n2004. \u201cWhy Do Providers Contribute to Disparities \nand What Can Be Done about It?\u201d Journal of General \nInternal Medicine 19(11):1154\u201359.\nBurgess, Diana J., Michelle van Ryn, John Dovidio, \nand Somnath Saha. 2007. \u201cReducing Racial Bias \namong Health Care Providers: Lessons from Social\u2013\nCognitive Psychology.\u201d Journal of General Internal \nMedicine 22(6):882\u201387.\nCarmichael, Stokely. 2003. \u201cBlack Power, a Critique \nof the System of International White Supremacy \nand International Capitalism.\u201d Pp. 151\u201374 in The \nDialectics of Liberation, edited by D. Cooper. New \nYork, NY: Penguin.\nCarrese, Joseph A., and Lorna A. Rhodes. 1995. \n\u201cWestern Bioethics on the Navajo Reservation: \nBenefit or Harm?\u201d Journal of the American Medical \nAssociation 274(10):826\u201329.\nChapman, Elizabeth N., Anna Kaatz, and Molly Carnes. \n2013. \u201cPhysicians and Implicit Bias: How Doctors \nMay Unwittingly Perpetuate Healthcare Disparities.\u201d \nJournal of General Internal Medicine 28(11):1504\u201310.\nCharmaz, Kathy. 2014. Constructing Grounded Theory. \nLondon, UK: Sage.\nColen, Shellee. 1986. \u201c\u2018With Respect and Feelings\u2019: \nVoices of West Indian Child Care and Domestic \nWorkers in New York City.\u201d Pp. 46\u201370 in All \nAmerican Women: Lives That Divide, Ties That Bind, \nedited by J. B. Cole. New York, NY: Free Press.\nDehlendorf, Christine, Reiley Reed, Edith Fox, Dominika \nSeidman, Cara Hall, and Jody Steinauer. 2018. \n\u201cEnsuring Our Research Reflects Our Values: The \nRole of Family Planning Research in Advancing \nReproductive Autonomy.\u201d Contraception 98(1):4\u20137.\nDehlendorf, Christine, Rachel Ruskin, Kevin Grumbach,  \nEric Vittinghoff, Kirsten Bibbins-Domingo, Dean  \nSchillinger, and Jody Steinauer. 2010. \u201cRecommen\u00ad\ndations for Intrauterine Contraception: A Randomized \nTrial of the Effects of Patients\u2019 Race-Ethnicity \nand Socioeconomic Status.\u201d American Journal of \nObstetrics and Gynecology 203(4):1\u20138.\nDowning, Roberta A., Thomas A. LaVeist, and Heather \nE. Bullock. 2007. \u201cIntersections of Ethnicity \nand Social Class in Provider Advice Regarding \nReproductive Health.\u201d American Journal of Public \nHealth 97(10):1803\u2013807.\nEpstein, Ronald M., and Richard L. Street. 2011. \u201cThe \nValues and Value of Patient-Centered Care.\u201d Annals \nof Family Medicine 9(2):100\u2013103.\nFreidson, Eliot. 1970. Profession of Medicine: A Study \nof the Sociology of Applied Knowledge. New York, \nNY: Dodd, Mead.\nGold, Rachel. 2014. \u201cGuarding Against Coercion while \nEnsuring Access: A Delicate Balance.\u201d Guttmacher \nPolicy Review. https://www.guttmacher.org/gpr/ \n2014/09/guarding-against-coercion-while-ensuring-\naccess-delicate-balance.\nGomez, Anu Manchikanti, Liza Fuentes, and Amy \nAllina. 2014. \u201cWomen or LARC First? Reproductive \nAutonomy and the Promotion of Long-Acting \nReversible Contraceptive Methods.\u201d Perspectives on \nSexual and Reproductive Health 46(3):171\u201375.\nGomez, Anu Manchikanti, and Mikaela Wapman. 2017. \n\u201cUnder (Implicit) Pressure: Young Black and Latina \nWomen\u2019s Perceptions of Contraceptive Care.\u201d \nContraception 96(4):221\u201326.\nGubrium, Aline C., Emily S. Mann, Sonya Borrero, \nChristine Dehlendorf, Jessica Fields, Arline T. \nGeronimus, Anu M. G\u00f3mez, et al. 2016. \u201cRealizing \nReproductive Health Equity Needs More Than \nLong-Acting Reversible Contraception (LARC).\u201d \nAmerican Journal of Public Health 106(1):18\u201319.\nHenderson, Jillian T., Tina Raine, Amy Schalet, Maya \nBlum, and Cynthia C. Harper. 2011. \u201c\u2018I Wouldn\u2019t \nBe This Firm if I Didn\u2019t Care\u2019: Preventive Clinical \nCounseling for Reproductive Health.\u201d Patient \nEducation and Counseling 82(2):254\u201359.\nHiggins, Jenny A., Renee D. Kramer, and Kristin \nM. Ryder. 2016. \u201cProvider Bias in Long-Acting \nReversible Contraception (LARC) Promotion and \nRemoval: Perceptions of Young Adult Women.\u201d \n\n134\t\nJournal of Health and Social Behavior 62(2)\nAmerican Journal of Public Health 106(11): \n1932\u201337.\nKalmuss, Debra, Andrew R. Davidson, Linda F. Cushman, \nStephen Heartwell, and Marvin Rulin. 1996. \n\u201cDeterminants of Early Implant Discontinuation \namong Low-Income Women.\u201d Family Planning \nPerspectives 28(6):256\u201360.\nLara-Mill\u00e1n, Armando. 2014. \u201cPublic Emergency Room \nOvercrowding in the Era of Mass Imprisonment.\u201d \nAmerican Sociological Review 79(5):866\u201387.\nLittlejohn, Krystale E. 2013. \u201c\u2018It\u2019s Those Pills That Are \nRuining Me\u2019: Gender and the Social Meanings of \nHormonal Contraceptive Side Effects.\u201d Gender & \nSociety 27(6):843\u201363.\nLittlejohn, Krystale E., and Katrina Kimport. 2017. \n\u201cContesting and Differentially Constructing Uncertainty: \nNegotiations of Contraceptive Use in the Clinical \nEncounter.\u201d Journal of Health and Social Behavior \n58(4):442\u201354.\nLuna, Zakiya, and Kristin Luker. 2013. \u201cReproductive \nJustice.\u201d Annual Review of Law and Social Science \n9(1):327\u201352.\nMcLeod, Jane. 2015. \u201cWhy and How Inequality Matters.\u201d \nJournal of Health and Social Behavior 56(2):149\u201365.\nMetzl, Jonathan M., and Helena Hansen. 2014. \n\u201cStructural Competency: Theorizing a New Medical \nEngagement with Stigma and Inequality.\u201d Social \nScience & Medicine 103:126\u201333.\nMetzl, Jonathan M., and Anna Kirkland, eds. 2010. Against \nHealth: How Health Became the New Morality. New \nYork, NY: New York University Press.\nMoore, David J., Jerome D.  Williams, and William Qualls. \n1996. \u201cTarget Marketing of Tobacco and Alcohol-\nRelated Products to Ethnic Minority Groups in the \nUnited States.\u201d Ethnicity & Disease 6(1-2):83\u201398.\nMurray, Samantha. 2008. \u201cPathologizing \u2018Fatness\u2019: \nMedical Authority and Popular Culture.\u201d Sociology \nof Sport Journal 25(1):7\u201321.\nPower to Decide. 2019. \u201cOne Key Question Overview.\u201d \nhttps://powertodecide.org/what-we-do/information/\nresource-library/one-key-questionr-overview.\nRoberts, Dorothy. 1997. Killing the Black Body: Race, \nReproduction, and the Meaning of Liberty. New \nYork, NY: Pantheon Books.\nSchulman, Kevin, Jesse A. Berlin, William Harless, \nJon F. Kerner, Shyrl Sistrunk, Bernard J. Gersh, \nRoss Dube, et\u00a0 al. 1999. \u201cThe Effect of Race and \nSex on Physicians\u2019 Recommendations for Cardiac \nCatheterization.\u201d The New England Journal of \nMedicine 340(8):618\u201326.\nShim, Janet K. 2005. \u201cConstructing \u2018Race\u2019 across \nthe Science-Lay Divide: Racial Formation in the \nEpidemiology and Experience of Cardiovascular \nDisease.\u201d Social Studies of Science 35(3):405\u201336.\nSmedley, Brian D., Adrienne Y. Stith, and Alan R. Nelson. \n2002. Unequal Treatment: Confronting Racial and \nEthnic Disparities in Healthcare. Washington, DC: \nThe National Academies Press.\nSpencer, Karen Lutfey, and Matthew Grace. 2016. \n\u201cSocial Foundations of Health Care Inequality \nand Treatment Bias.\u201d Annual Review of Sociology \n42(1):101\u201320.\nStevens, Lindsay. 2015. \u201cPlanning Parenthood: Health \nCare Providers\u2019 Perspectives on Pregnancy Intention, \nReadiness, and Family Planning.\u201d Social Science & \nMedicine 139:44\u201352.\nStewart, Moira, Judith Belle Brown, Allan Donner, Ian R. \nMcWhinney, Julian Oates, Wayne W. Weston, and \nJohn Jordan. 2000. \u201cThe Impact of Patient-Centered \nCare on Outcome.\u201d The Journal of Family Practice \n49(9):796\u2013804.\nTeal, Cayla R., Rachel E. Shada, Anne C. Gill, Britta \nM. Thompson, Ernest Frug\u00e9, Graciela B. Villarreal, \nand Paul Haidet. 2010. \u201cWhen Best Intentions \nAren\u2019t Enough: Helping Medical Students Develop \nStrategies for Managing Bias about Patients.\u201d \nJournal of General Internal Medicine 25(2):115\u201318.\nThomas, Adam, and Emily Monea. 2011. \u201cThe High  \nCost of Unintended Pregnancy.\u201d Brookings Institute.  \nhttps://www.brookings.edu/wpcontent/uploads/ \n2016/06/07_unintended_pregnancy_thomas_monea \n.pdf.\nvan Ryn, Michelle, Diana Burgess, Jennifer Malat, and \nJoan Griffin. 2006. \u201cPhysicians\u2019 Perceptions of \nPatients\u2019 Social and Behavioral Characteristics and \nRace Disparities in Treatment Recommendations \nfor Men with Coronary Artery Disease.\u201d American \nJournal of Public Health 96(2):351\u201357.\nvan Ryn, Michelle, and Steven S. Fu. 2003. \u201cPaved with \nGood Intentions: Do Public Health and Human \nService Providers Contribute to Racial-Ethnic \nDisparities in Health?\u201d The American Journal of \nPublic Health 93(2):248\u201355.\nXierali, Imam M., and Marc A. Nivet. 2018. \u201cThe Racial \nand Ethnic Composition and Distribution of Primary \nCare Physicians.\u201d Journal of Health Care for the \nPoor and Underserved 29(1):556\u201370.\nYee, Lynn, and Melissa Simon. 2011. \u201cPerceptions \nof Coercion, Discrimination and Other Negative \nExperiences in Postpartum Contraceptive Counseling \nfor Low-Income Minority Women.\u201d Journal of  \nHealth Care for the Poor and Underserved 22(4): \n1387\u2013400.\nZestcott, Colin, Irene V. Blair, and Jeff Stone. 2016. \n\u201cExamining the Presence, Consequences, and \nReduction of Implicit Bias in Health Care: A \nNarrative Review.\u201d Group Processes & Intergroup \nRelations 19(4):528\u201342.\nAuthor Biographies\nJamie L. Manzer is a PhD candidate in the Department of \nSociology & Criminal Justice at the University of \nDelaware and holds a master\u2019s of public policy degree. \nHer research explores the intersections of reproduction, \nhealth, inequality, and policy. She has published in \n\nManzer and Bell\t\n135\njournals such as Qualitative Health Research and Journal \nof Interpersonal Violence.\nAnn V. Bell is an associate professor of sociology at the \nUniversity of Delaware. Her research, centering on the \nintersection of gender and health, specifically examines \nprocesses, inequalities, and constructions of reproductive \nhealth. Her work has examined the experiences of infertil-\nity, abortion, and contraception and how those experiences \nare shaped by class- and race-based social understandings \nof reproduction.\n\n\n    Scientific Research Findings:", "answer": "\u2022 Providers believe prejudicial contraceptive counseling was problematic but also described counseling their patients differently based on the patients\u2019 social and demographic backgrounds.\n\u2022 Providers used four primary strategies to rationalize, cope with, and manage this differential counseling: using scientific research to explain differences in counseling approaches, using \u201csafe\u201d medical jargon to distinguish between patients, attempting to standardize care across patients to avoid accusations of bias, and identifying their counseling approach as \u201cpatient-centered\u201d rather than biased.\n\u2022 Ironically, using these strategies exacerbated bias by allowing providers to deny or justify bias rather than confront and address it.", "id": 75}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465211052568\nJournal of Health and Social Behavior\n2021, Vol. 62(4) 460\u00ad\u2013476\n\u00a9 American Sociological Association 2021\nDOI: 10.1177/00221465211052568\njhsb.sagepub.com\nArticle\nMass incarceration has increasingly been recog-\nnized as a pressing public health challenge. Nearly \n6.5 million adults\u20141 in 40\u2014are under correctional \nsupervision in the United States (Maruschak and \nMinton 2020). More than 2.1 million are currently \nserving sentences in prisons and jails. An additional \n4.9 million individuals in the population were for-\nmerly imprisoned (Shannon et\u00a0 al. 2017). This \ndecades-long expansion of the penal system has \noccurred without the structures in place to facilitate \nreintegration. Although the policies of mass incar-\nceration have been far-reaching, their imprint has \nbeen concentrated among marginalized groups. \nBlack adults account for 33% of individuals incar-\ncerated in prisons\u2014almost triple their share of the \nadult population\u2014and they are incarcerated at rates \nthat are three to five times higher than whites \n(Carson 2020; Zeng 2020). Noting these disparities, \nsome scholars have argued that incarceration is so \nconcentrated among disadvantaged groups that it \n\u201chas become a normal stage of the life course\u201d \n(Wildeman and Wang 2017:1465).\nThe deleterious effects of incarceration for \n\u00adindividual well-being have been the subject of a \ngrowing body of interdisciplinary research at the \nintersection of medicine and the social sciences. \n1052568 HSBXXX10.1177/00221465211052568Journal of Health and Social BehaviorBerg et al.\nresearch-article2021\n1University of Iowa, Iowa City, IA, USA\n2University of Georgia, Athens, GA, USA\nCorresponding Author:\nMark T. Berg, Department of Sociology and \nCriminology, University of Iowa, 400H North Hall, Iowa \nCity, IA 52242, USA. \nEmail: mark-berg@uiowa.edu\nLosing Years Doing Time: \nIncarceration Exposure and \nAccelerated Biological  \nAging among African  \nAmerican Adults\nMark T. Berg1\n, Ethan M. Rogers1, Man-Kit Lei2,  \nand Ronald L. Simons2\nAbstract\nResearch suggests that incarceration exposure increases the prevalence of morbidity and premature \nmortality. This work is only beginning to examine whether the stressors of the incarceration experience \nbecome biologically embedded in ways that affect physiological deterioration. Using data from a longitudinal \nsample of 410 African American adults in the Family and Community Health Study and an epigenetic index \nof aging, this study tests the extent to which incarceration accelerates epigenetic aging and whether \nexperiences with violence moderate this association. Results from models that adjust for selection effects \nsuggest that incarceration exposure predicted accelerated aging, leaving formerly incarcerated African \nAmerican individuals biologically older than their calendar age. Direct experiences with violence also \nexacerbated the effects of incarceration. These findings suggest that incarceration possibly triggers a stress \nresponse that affects a biological signature of physiological deterioration.\nKeywords\nAfrican Americans, epigenetic aging, health, incarceration\n\nBerg et al.\t\n461\nPrisons and jails are known vectors for the trans-\nmission of chronic infectious diseases, such as \ntuberculosis and pneumonia (Hammett, Harmon, \nand Rhodes 2002). Another line of work developed \nin the psychoimmunology literatures maintains that \nincarceration triggers an upregulated, acute stress \nresponse that leads to physiological deterioration \n(Massoglia and Pridemore 2015). Empirical research \nindeed finds that formerly incarcerated persons \nhave elevated levels of morbidity, especially for \nstress-related illnesses, and heightened risk for \n\u00adpremature mortality (e.g., Boen 2020).\nDespite important advances in recent years, this \nbody of literature remains in a nascent stage owing \nto several substantive and methodological chal-\nlenges (Massoglia and Remster 2019). First, the \npossibilities of selection bias and statistical con-\nfounding complicate efforts to isolate the causal \neffects of incarceration exposure on health out-\ncomes. Second, the extent to which physical health \nis affected by witnessing or experiencing interper-\nsonal violence while detained has received scant \nattention in the quantitative empirical research. \nThis is an important omission because periods of \nconfinement often involve unwanted, uncontrolla-\nble, and threatening experiences that conceivably \ngovern the effects of incarceration on postrelease \nhealth (Porter 2019). As such, prior work has \nlargely considered incarceration a homogeneous \nexperience, thus ignoring the effects of varied expe-\nriences. It is therefore important to assess whether \nspecific stress-inducing experiences during incar-\nceration are differentially related to poor health.\nThird, with few exceptions, most studies examine \nself-reported health outcomes, including perceived \nhealth, illnesses, and chronic conditions, rather than \nobjective biomarkers obtained from independent \nassessments (Wildeman and Wang 2017:1468). \nAlthough useful, such indices of self-reported health \nprovide limited information on the biological pro-\ncesses that precede the manifestations of disease. \nAsymptomatic conditions go unmeasured. Furthermore, \nsubjective reports of health confound \u201cphysiological \nwith psychological and emotional well-being\u201d \n(Harris and Schorpp 2018:362). Because of the reli-\nance on self-report instruments, the extent to which \nincarceration exposure becomes biologically embed-\nded or \u201cgets under the skin\u201d is not well understood. \nAlthough objective biomarkers have been featured in \nsome studies of incarceration and health, this work \nremains scarce (e.g., Boen 2020).\nAccelerated aging is a \u201cpredominant risk factor \nfor most diseases and conditions that limit health \nspan\u201d (Kennedy et\u00a0al. 2014:709). The basis of healthy \naging is a focus of modern medicine (Sinclair 2019). \nNewly developed epigenetic indices (\u201cclocks\u201d) of \nbiological aging have been shown to be powerful pre-\ndictors of premature mortality and age-related dis-\neases (Ryan et\u00a0al. 2020). How social adversities affect \nbiomarkers of accelerated aging has been the subject \nof burgeoning research on the global burden of age-\nrelated disease (Moffitt et\u00a0al. 2017; Sinclair 2019). \nDespite its promise for understanding the biology of \ndisease and illnesses, this work has not been inte-\ngrated into the research on incarceration.\nUsing a longitudinal sample of African American \nadults from the Family and Community Health \nStudy (FACHS), this study examines whether incar-\nceration exposure and experiences with violence \nwhile incarcerated predict accelerated biological \naging. Note that because the sample is limited to a \nsample of African American adults, the study does \nnot assess whether incarceration contributes to \nracial disparities in aging. We posit that individuals \nexposed to incarceration will be biologically older \nthan their numerical age and even more so among \nthose who experienced violence while serving time. \nThe study will inform efforts to understand the \nextent to which incarceration becomes biologically \nembedded to affect physiological changes impli-\ncated in the onset of chronic illness and disease.\nBackground\nIncarceration and Stressors\nAlthough relative differences in mortality between \nblack and white individuals is significantly lower in \nprison compared to the differences in the general \npopulation\u2014a pattern largely attributable to deaths \nfrom firearms and motor vehicle accidents (Patterson \n2010)\u2014formerly incarcerated individuals are at \ngreater risk of dying and suffering from disease fol-\nlowing release. Incarceration leaves a negative \nimprint on the physical health of this population. The \ntotal health effects of incarceration are a product of \nthe transition from arrest to confinement, the time \nspent in prison, and the period following release \n(Massoglia and Pridemore 2015). The initial transi-\ntion from the community to prison is a disruptive \nevent partly because individuals must learn to assim-\nilate into the prison environment (Irwin 1980). \nMoreover, confinement removes individuals from \ntheir primary social networks, who are sources of \ncritical emotional and psychological support. During \nspells of incarceration, personal relationships often \nbecome strained and difficult to nurture from a \ndistance.\n\n462\t\nJournal of Health and Social Behavior 62(4)\nAdding to the stressors of adjustment, individu-\nals face a variety of harsh and traumatic experi-\nences during incarceration (Massoglia and Remster \n2019). Direct experiences with violent events are \nnot uncommon in jails and prisons (Labrecque, \nScherer, and McCafferty 2018). For instance, \nprison surveys have found that upward of 21% of \nboth men and women report attacks by other incar-\ncerated individuals, and nearly 25% and 8%, \nrespectively, report attacks by correctional staff \n(Wolff et\u00a0al. 2007). National estimates indicate that \napproximately 13% of incarcerated individuals  \nare injured in a violent encounter (Teasdale et\u00a0al. \n2016). The loss of personal security\u2014one \u201cpain of \n\u00adimprisonment\u201d\u2014is exacerbated by these distinct \nmoments of danger (Maier and Ricciardelli 2019; \nSykes 1958). Secondary, or witnessed, violence \nalso increases fear, hypervigilance, and general \nunease with the social environment (Novisky and \nPeralta 2020). Concerns about safety and hostile \ninteractions with other incarcerated individuals and \ncorrectional staff create an action readiness of being \n\u201con point\u201d that contributes to an \u201coverwhelming \nsense of anxiety\u201d (Porter 2019:5). Altogether, \nencounters with violence may \u201ccontribute to the sus-\ntained experience of emotional distress\u2009.\u2009.\u2009.\u2009up to and \nbeyond release from prison and jail\u201d (Boxer, \nMiddlemass, and Delorenzo 2009:802). Studies have \nreported associations between violent experiences \nand the worsening of psychological health during \nincarceration and after release (Wolff and Shi 2009; \nZweig et\u00a0al. 2015).\nSeveral well-documented barriers to reintegra-\ntion add to the burden of stressors associated with \nincarceration (Semenza and Link 2019). Many \nindividuals reenter society with limited resources \nand narrow employment prospects because of \nexclusionary hiring practices and weak ties to con-\nventional networks (Petersilia 2003). Safe and \nsecure housing is also a \u201churdle in the steeplechase \nof reentry\u201d because some are barred from renting \nbecause of their criminal records (LeBel 2017:889). \nFor instance, in a sample of formerly incarcerated \nindividuals, only half were unemployed 6 months \nafter release, and more than one-third stayed in \nmarginal or temporary housing (Western et\u00a0 al. \n2015). Harding and colleagues (2014:450) found \nmore than one-quarter of formerly incarcerated per-\nsons shifted between \u201cextreme desperation and sur-\nvival when they were not in custody.\u201d These \ndifficulties are compounded by substance abuse \nand mental health problems (Tyler and Brockmann \n2017). Furthermore, even with the implementation \nof the Affordable Care Act, formerly incarcerated \nindividuals are more likely to be uninsured and less \nlikely to utilize preventive health care relative to the \ngeneral population (Widdowson and Fisher 2020; \nWinkelman, Choi, and Davis 2017).\nIncarceration, Stress Response,  \nand Physical Health\nAccording to the neuroendocrine and immunology \nliteratures, the stressors resulting from social and \nenvironmental adversities, such as the incarceration \nexperience, trigger the circuitry of the autonomic \nand parasympathetic nervous systems in addition to \nthe hypothalamic\u2013pituitary\u2013adrenal axis (Miller, \nChen, and Parker 2011; Snyder-Mackler et\u00a0 al. \n2020). Activation of each system is necessary for \nsurvival and reproductive success. However, the \nprolonged and elevated stimulation of these systems \nhas a gradual \u201cwear and tear\u201d effect on the body\u2019s \nadaptive capacities, producing impairments that are \nexpressed in surrogate end-point biomarkers (Lu \net\u00a0 al. 2019; McEwen and Gianaros 2011). \nPhysiological adaptations in response to adversity \nmay also contribute to the programming of the \nstress response via alternations in histone acetyla-\ntion (Meaney and Szyf 2005). Moreover, phases of \nneuroinflammation disrupt the brain\u2019s capacity to \nmonitor and control pathogen invasion, which is \ndetrimental to cerebral functioning (Danese and \nBaldwin 2017). Geronimus and colleagues (2006) \nargue that the relatively extensive social adversities \nexperienced by African Americans promotes a type \nof \u201cbiological weathering.\u201d Specifically, the weath-\nering work contends that African Americans are \nburdened by material hardships and marginalization \nencountered in a race stratified society (Geronimus \net\u00a0al. 2006). The accumulation of these adverse con-\nditions contributes to physiological deterioration, \nincluding accelerated aging (Simons et\u00a0al. 2021).\nLife course theories of the stress process \n(Pearlin et\u00a0al. 2005), the psychoimmunology litera-\ntures (Miller et\u00a0al. 2011), and the biological weath-\nering model (Geronimus et\u00a0 al. 2006) all would \ncontend that incarceration engenders a harmful \nstress response. Empirical research has indeed \nshown that incarceration exposure is associated \nwith stress-related illnesses, such as hypertension, \ndiabetes, obesity, heart problems, and functional \nlimitations (Baquero et\u00a0 al. 2020; Houle 2011; \nMassoglia 2008; Schnittker and John 2007). \nHowever, a variety of substantive and methodologi-\ncal issues complicate efforts to fully understand the \nnature of the relationship between incarceration \nexposure and morbidity.\n\nBerg et al.\t\n463\nFirst, selection bias and confounding threaten the \nvalidity of causal estimates of the effects of incar-\nceration exposure on health. Because adverse physi-\ncal health conditions and criminal justice contact \nshare many common correlates\u2014social class, child-\nhood adversities, and self-control\u2014it is plausible \nthat preexisting risk factors confound the effects  \nof incarceration exposure on health outcomes \n(Schnittker, Massoglia, and Uggen 2012). Any \nunique effect attributable to incarceration might \ninstead result from the endowments that individuals \nbring to prison. This possibility requires adjustments \nfor preexisting confounds, including adverse child-\nhood experiences and childhood disadvantage, \nwhich are reliable determinants of adult morbidity \nand mortality (Miller et\u00a0al. 2011). Due to preexisting \ntendencies to have poor health and to engage in \nbehaviors that result in confinement, it is also neces-\nsary for observational research to adjust for selec-\ntion effects (Massoglia and Pridemore 2015). \nProspective longitudinal data are a powerful tool in \nthis regard because they allow for the implementa-\ntion of pretreatment (i.e., incarceration) measures.\nSecond, there is a shortage of research on the \nextent to which violent and hostile social interac-\ntions during incarceration contribute to the health \nof released individuals. As noted earlier, these \nexperiences are not uncommon based on survey \nand qualitative accounts. The psychological burden \nof these experiences could encumber the process of \nreintegration, engendering additional or secondary \nstressors that are expressed in the postrelease set-\nting. There are good theoretical reasons to expect \nthat exposure to violence, whether direct or indi-\nrect, might regulate the effects of incarceration on \nphysical health outcomes. In this way, incarceration \nis not a homogeneous experience; rather, the extent \nto which incarceration affects the health of released \nindividuals is determined by their differential expe-\nriences with social stressors during the period of \nconfinement.\nThird, most research on postrelease morbidity \nexamines measures of general health, chronic con-\nditions, illnesses, and diseases ascertained from \nsurvey self-reports. Self-report measures provide \nvaluable information about one\u2019s perceived health \nburden, and some predict mortality (e.g., Jylh\u00e4 \n2009), but these tools are known to suffer from \nthreats of validity and reliability, especially among \nmarginalized sociodemographic groups (Assari, \nLankarani, and Burgard 2016; Zajacova and Dowd \n2011). Importantly, self-reports of various health \nconditions are unlikely to capture cases in which \nunderlying illnesses have not yet manifested as \nobvious symptoms (Boen 2020; Morton and \nFerraro 2020). For instance, respondents who report \nno evidence of illness but have undetected pathol-\nogy would be considered healthy based on self-\nreports when they are not, thereby producing \nmisclassification errors. The possibility of these \ntypes of errors is heightened among younger popu-\nlations. Should formerly incarcerated individuals \nnot report symptoms of illness or disease on self-\nreports, despite underlying pathology, this would \nthus bias estimates of the incarceration\u2013health asso-\nciation. Incarcerated persons may also be better \naware of their own health conditions than if they \nhad not been incarcerated because of their prison \nmedical assessments. If so, this could increase the \nassociation between incarceration and self-reported \nhealth conditions (Massoglia and Remster 2019). \nObjective biomarkers are not vulnerable to such \nsubjective recall and reporting errors and, further-\nmore, they capture predisease pathology regardless \nof whether individuals exhibit symptoms.\nAs such, there have been recent calls to integrate \nbiomarkers into research on incarceration and \nhealth to understand how incarceration exposure \nbecomes biologically embedded (see Massoglia \nand Remster 2019; Wildeman and Wang 2017). \nNotably, some researchers have modeled objective \nbiomarkers, but there have been mixed results. \nUsing longitudinal data from the Coronary Artery \nRisk Development in Young Adults study, Wang \nand colleagues (2009) found that prior incarcera-\ntion was associated with hypertension and left ven-\ntricular hypertrophy among adults ages 23 to 35. \nMore recently, scholars have leveraged the National \nLongitudinal Study of Adolescent to Adult Health \n(Add Health) study to investigate the association \nbetween incarceration and health biomarkers \namong young adults (Boen 2020; Esposito et\u00a0 al. \n2017). For instance, Esposito and colleagues (2017) \nfound no effect of incarceration on hypertension \nafter adjusting for selection effects. Similarly, Boen \n(2020) found a relationship between incarceration \nand inflammation, but this effect was reduced after \napplying treatment weights to adjust for selection. \nAdditionally, incarceration was only associated \nwith inflammation in that study when the period of \nconfinement was a year or longer. Other work with \nAdd Health has found associations between incar-\nceration and cardiovascular disease risk\u2014an index \nderived from objective and subjective self-report \nitems\u2014using models that did not adjust for threats \nof selection bias (e.g., Semenza et\u00a0al. 2021), which \nraises questions about the internal validity of the \nfindings.\n\n464\t\nJournal of Health and Social Behavior 62(4)\nIncarceration and Epigenetic Aging\nResearch on the social determinants of health has \nincreasingly focused on the speed of cellular aging. \nThis work is consequential because aging is the \n\u201cmost important risk factor for many human dis-\neases\u201d (Ryan et\u00a0al. 2020:481). Conditions including \ncognitive decline, heart disease, and cancers are doc-\numented products of biological aging (Sinclair \n2019:67). Epigenetic aging involves the \u201caccumula-\ntion of damage\u201d responsible for cellular and organis-\nmal decline (Song and Johnson 2018:1). Researchers \nhave studied mechanisms of aging with measures of \ntelomere length and pubertal timing. Since the 1960s, \nresearchers have been aware of the strong association \nbetween age and deoxyribonucleic acid (DNA) \nmethylation (Koch and Wagner 2011). DNA methyl-\nation is a form of epigenetic modification\u2014or chemi-\ncal alteration to the genome that affects gene \nactivity\u2014whereby a methyl group attaches to a seg-\nment of DNA at a CpG site, which can inhibit gene \nexpression. Through this process, the activity of a \nDNA segment can change without an adjustment to \nits base sequence. Many of these sites are positioned \nnear genes that trigger age-related conditions, includ-\ning cancer, oxidative stress, respiratory diseases, and \nAlzheimer\u2019s disease (Hannum et\u00a0al. 2013; Yoon, Jin, \nand Sin 2019). Some have argued that cellular aging \nis caused by epigenetic signalers\u2014which repair dam-\naged DNA\u2014becoming overtaxed from cellular \ninsults, including those present in the external envi-\nronment (Sinclair 2019:47\u201349). Site-specific patterns \nin DNA methylation are integral to the aging process \namong humans (Ryan et\u00a0al. 2020).\nSuch findings have prompted the development \nof various epigenetic indices or \u201cclocks\u201d of biologi-\ncal aging (Murabito et\u00a0al. 2018). These clocks assess \nthe discrepancy between a person\u2019s chronological \nage and the age predicted by the epigenetic index, \nwhich indicates whether a person is aging at an \naccelerated or decelerated pace. A recently devel-\noped clock known as the GrimAge index examines \nmethylation at 1,030 sites and has a superior ability \nto predict time to mortality and age-related illnesses \n(Lu et\u00a0al. 2019) relative to the other available epi-\ngenetic clocks (see McCrory et\u00a0al. 2021). Research \nindicates that accelerated GrimAge is associated \nwith social adversities, including material hardship, \ndiscrimination (Simons et\u00a0 al. 2021), and trauma \n(Yang et\u00a0al. 2020). \nThe extent to which incarceration exposure \n\u00adpredicts accelerated biological aging has not been \nexamined. As noted previously, the substantive liter-\nature implies that formerly incarcerated persons con-\nceivably age at a faster pace than nonincarcerated \npersons. If so, this would provide an important clue \nabout how incarceration might give rise to health \ndisparities.\nCurrent Study\nAlthough there is consistent evidence that incarcer-\nation exposure undermines the physical health of \nindividuals, important gaps remain in this research. \nUsing data on a sample of African American adults, \nthe present study addresses two objectives. First, \nusing the newly developed GrimAge index, it \nassesses the extent to which incarceration exposure \npredicts accelerated biological aging. Second, the \nstudy examines whether experiences with violence, \nboth direct and secondary, during confinement \nexacerbate the effect of incarceration on accelerated \nbiological aging.\nData and Methods\nSample\nWe tested hypotheses for the current study by \nemploying data from the Family and Community \nHealth Study (FACHS). FACHS is designed to \ninvestigate the factors associated with the health and \nwell-being of African American adults (see Lei \net\u00a0al. 2018; Simons et\u00a0al. 2019). The FACHS sam-\npling strategy includes African American families \nwith a child who was in fifth grade at the time of \nrecruitment. Families were intentionally recruited \nfrom neighborhoods that varied on demographic \nand economic characteristics based on 1990 census \ndata. For 259 census block groups (115 in Georgia \nand 144 in Iowa), households were randomly \nselected from rosters of fifth graders in the public \nschool system. Households were then contacted \nuntil the required number of households were \nrecruited (84% response rate).\nThe data for the first wave of FACHS were col-\nlected in 1997 and 1998 (for a description of FACHS \nwaves, see Appendix A in the online version of the \narticle). The sample consisted of 889 African \nAmerican fifth-grade children who were an average \nage of 10.6 years and resided in Iowa (n = 467) or \nGeorgia (n = 422). The second through sixth waves \nof data were collected between 1999 and 2012, cap-\nturing information when the target children were \n12.5, 15.7, 18.8, 21.6, and 23.6 years. The seventh \nwave of data was completed between 2015 and \n2016 when the target children had an average age of \n28.8 years. By Wave 6, approximately 49% of \nrespondents resided in Georgia, 40% in Iowa, and \nthe remaining 10% were located in other states. \n\nBerg et al.\t\n465\nOnly members of the sample residing in Georgia, \nIowa, or a contiguous state were identified as \n\u00adeligible for the blood draws needed for the current \nresearch due to the logistical challenges of schedul-\ning home visits by phlebotomists. After also \n\u00adexcluding persons who were deceased (n = 12), \nincarcerated at the time (n = 15), hard refusals \n(n = 19), unidentifiable phone number (n = 46), or \notherwise unreachable (n = 99), 556 respondents \nagreed to be interviewed, 470 of whom successfully \n\u00adprovided blood. Of these respondents, successful \nassaying for methylation was achieved for 449 \n\u00adindividuals. In the current study, analyses were based \non the 410 African American respondents who were \nsuccessfully assayed and for whom data were avail-\nable on all study measures\u201413 \u00adrespondents were \nmissing data on incarceration \u00adhistory (2.9%), and an \nadditional 26 respondents were missing data on the \ncovariates in the model (5.8%).1\nProcedures\nQuestions were administered to families in their \nhomes using audio-enhanced, computer-assisted, and \nself-administered interviews. Respondents answered \nquestions on a computer as they were presented visu-\nally on the screen. In Wave 7, participants were also \nasked to provide a blood sample. The phlebotomist \ndrew four tubes of blood (30 mL) from each partici-\npant; these were shipped on the same day to a labora-\ntory at a public university for preparation. Blood \nsamples were assessed for genome-wide methylation \nstatus using standard protocols. The Illumina \nInfinium HumanMethylationEPIC 850 BeadChip \nwas used to assay genome-wide DNA methylation. \nParticipants were randomly assigned to 16 sample \n\u201cslides/chips\u201d with groups of eight slides being bisul-\nfite converted in a single plate, resulting in two \n\u201cbatches/plates.\u201d A replicated sample of DNA was \nincluded in each plate to aid in assessment of batch \nvariation. Methylation data were filtered according to \nthe following guidelines: (a) Samples containing 1% \nof CpG sites with detection p < .05 were removed, (b) \nsites were removed if a bead count of <3 was present \nin 5% of samples, and (c) sites with a detection p < \n.05 in 1% of samples were removed.\nMeasures\nAccelerated aging.\u2002 To calculate the speed of aging, \nwe used an epigenetic clock known as the GrimAge \nindex (Horvath and Raj 2018). The index is called \nGrimAge\u2014after the Grim Reaper\u2014because an \naccelerated score is grim news. This measure of \naging has been shown to be a more robust predictor \nof morbidity and mortality among different racial \ngroups than other epigenetic clocks currently avail-\nable (Lu et\u00a0 al. 2019; McCrory et\u00a0 al. 2021). The \nindex was developed by identifying a set of plasma \nprotein predictors of mortality and then using these \nprotein predictors to identify CpG sites that could \npredict time to death. As expected, the GrimAge \nindex has a weak to moderate correlation with \nchronological age in the FACHS sample (r = .22), \nindicating a discrepancy between biological and \ncalendar ages. Following standard procedures, we \nregressed GrimAge on chronological age to create \na speed of aging score or \u201caccelerated aging.\u201d The \nresulting speed of Grim aging measure was thus \nadjusted to correlate with chronological age at 0. A \npositive value on this variable indicates, in years, \naccelerated epigenetic aging, whereas a negative \nvalue indicates, in years, decelerated aging.\nIncarceration exposure.\u2002 To assess the prevalence \nof incarceration, we used reports from Waves 5, 6, \nand 7. If respondents reported being arrested (since \nage 18 in Wave 7), respondents were asked to report \ntheir frequency of being to jail and their frequency of \nbeing to prison (since age 18 in Wave 7). If respon-\ndents reported being in jail or prison at least once, the \nmeasure was coded as 1 = formerly incarcerated. If \nthe respondent never reported being in jail or prison \nat Waves 5, 6, or 7, the measure was coded as 0 = \nnever incarcerated. \u201cDon\u2019t know\u201d responses were \ntreated as missing. Although most studies assess \nincarceration using a lifetime prevalence measure, it \nmay be that the experience in itself does not fully \ncapture how incarceration exposure is associated \nwith individual health. For this reason, it is important \nto conduct comprehensive tests of the varieties of \nexposure. To probe the nature of exposure effects, \nwe also measured the timing and frequency of incar-\nceration. Past 4 years was coded as 1 if the respon-\ndent reported being in jail or prison in the past 4 \nyears at Wave 7 (average age at Wave 7 was 28.8 \nyears). This 4-year measure was included in the \nWave 7 survey to capture spells of incarceration \nsince Wave 6, which was administered approxi-\nmately 4 years earlier. The measure of incarceration \nin the past 4 years is intended to assess whether the \neffect of incarceration exposure on accelerated aging \nextends beyond lifetime prevalence to recent spells \nof incarceration. More than once was coded as 1 if \nthe respondent reported being in jail or prison two or \nmore times. A frequency measure of incarceration is \nintended to assess the effect of exposure to multiple \nspells of incarceration on accelerated aging.\n\n466\t\nJournal of Health and Social Behavior 62(4)\nViolence during incarceration.\u2002 We created two \nmeasures to examine experiences with violence dur-\ning incarceration. To measure direct violence, \nrespondents were asked to report during Waves 6 \nand 7 how often the following experiences happened \nduring jail or prison: property was taken from them \nusing force or intimidation, they were threatened \nwith violence, they were assaulted with an object \nused as a weapon, or they were involved in physical \nfights with others incarcerated. If respondents \nreported experiencing any of these violent events, \ndirect violence was coded as 1 (0 otherwise). To \nmeasure secondary violence, respondents reported \nhow often they witnessed other incarcerated indi-\nviduals being assaulted with a weapon or involved in \na physical fight (Waves 6 and 7). If respondents \nreported experiencing either of these events, second-\nary violence was coded as 1 (0 otherwise).\nConfounds.\u2002 We controlled for the following \ndemographic and socioeconomic characteristics at \nWave 7: male, high school education, income, and \nmarried or cohabiting (for description of the mea-\nsurement of the confounds, see Appendix B in the \nonline version of the article). These measures were \nconcurrent to the outcome. Income was assessed by \nasking respondents to report their income in the past \nyear and was log-transformed to reduce skew. The \naverage income in the analytical sample was \n$21,474.26. We also adjusted for past-year health \ninsurance at Wave 7. To account for preincarcera-\ntion health and health risk behaviors, we included \nmeasures of smoking, drinking, diet, exercise, and \nBMI. This class of measures was included either on \nor before the wave of the incarceration measure \nbased on their availability in the data. Smoking was \nmeasured as the prevalence of cigarette smoking in \nthe past year, and binge drinking was measured as \nthe frequency of consuming three or more drinks of \nalcohol during the past year (Waves 1 to 4). Two \nitems indicating the frequency of eating fruits and \nvegetables were used to measure a healthy diet \n(Waves 4 to 5), and two items indicating the fre-\nquency of physical activity for at least 30 minutes \nwere used to measure exercise (Waves 3 to 5). BMI \nwas calculated using the respondents\u2019 recorded \nheight and weight (Waves 1 to 2, 4).\nAdditional preincarceration conditions were \nincluded as potential confounds (see Appendix B \nin the online version of the article). To measure \ncommunity disorder (Waves 1 to 4), participants \nwere asked about the frequency of criminal acts in \ntheir neighborhood. Mental health is an important \nconfound because it is related to both criminal \nbehavior and physical health (Beyers and Loeber \n2003). A battery from the Diagnostic Interview \nSchedule for Children, Version 4 (DISC-IV), was \nused to measure depression (Waves 1 to 4; Shaffer \net\u00a0al. 1993). We summed all responses to create a \ncount variable scored from 0 (no symptoms) to 22 \n(all symptoms). A measure of self-control (Waves \n1 to 2, 4) was measured using an 11-item scale \nfrom the Kendall and Wilcox (1979) inventory of \nself-constraint (e.g., \u201cyou usually think before you \nact\u201d). We also adjusted for a measure of delin-\nquency (Waves 1 to 4) using the conduct disorder \nsection of the DISC-IV. We summed responses to \npast-year involvement in 15 acts to create a count \nvariable scored from 0 (no acts) to 15 (all acts).\nTwo measures accounted for childhood charac-\nteristics implicated in criminal justice contact and \npoor health (see Appendix B in the online version \nof the article). To assess childhood adversities, \nrespondents were asked at Wave 1 to report whether \nthey experienced various negative events during the \npast year (see Berg et\u00a0 al. 2020). Each item was \nsummed to create a total childhood adversity score \nranging from 0 to 14. Childhood socioeconomic \nstatus was assessed by summing caregivers\u2019 reports \nof education and income and averaging the reports \nacross Waves 1 and 2.\nAnalytical Strategy\nWe first conducted descriptive analyses by examin-\ning the means and percentages of all variables for \nthe total sample and by incarceration exposure. We \nthen used a series of ordinary least squares regres-\nsion models to test our hypothesis regarding the \nassociation between incarceration and our continu-\nous speed of aging variable, GrimAge, adjusting \nfor a series of key confounds. Given that respon-\ndents were not randomly assigned to incarceration, \nwe then estimate the analysis using an inverse prob-\nability of treatment weighting (IPTW) scheme to \nadjust for possible threats of selection bias. For the \nIPTW procedures, we first created propensity \nscores by calculating the probabilities of incarcera-\ntion conditional on our confounds (Rosenbaum and \nRubin 1983). We then reran our regression analysis \nweighted by the propensity scores. By doing so,  \nthe IPTW method balances the distribution of \n\u00adconfounds across the treated and untreated groups \nand allows for less biased estimates of the potential \ncausal effect of incarceration (Robins, Hern\u00e1n, and \nBrumback 2000). For both the treatment and \u00adoutcome \nmodels, we used the same set of measures. For the \nraw and weighted mean scores and standardized \n\nBerg et al.\t\n467\ndifferences between the formerly and never incar-\ncerated samples, see Appendix C in the online ver-\nsion of the article. We also conducted sensitivity \nanalyses often recommended for the proper use of \nIPTW: restricting to common support and trimming \n(Cole and Hern\u00e1n 2008). It is worth noting that the \nquality of estimates\u2014the extent to which they min-\nimize selection bias\u2014derived from propensity \nscore-based methods depends on the extent to \nwhich the observed variables not only exhaust \nsources of possible selection but also if the vari-\nables have good validity and reliability.\nFinally, we applied an internal moderator \napproach to assess if the timing, frequency, and \nexposure to violence determine the health effects of \nincarceration on GrimAge. Timing, frequency, and \nexposure to violence during incarceration are condi-\ntionally relevant measures\u2014they were only asked \nof individuals who have been incarcerated. The \ninternal moderator approach explicitly considers \nthat these hypothesized moderators only apply to the \nrespondents who experienced incarceration, provid-\ning a \u201ccontingent\u201d interaction effect (Mirowsky \n2013). In other words, this approach allowed us to \ncompare those who were formerly incarcerated with \nthose who were not while simultaneously assessing \nif the qualities of the situation\u2014the incarceration \nspell\u2014determine the effects of being in that situa-\ntion (Ross and Mirowsky 1992).\nResults\nMeans and percentages for the total sample and by \nincarceration exposure are displayed in Table 1. \nOver half of this African American sample reported \nbeing formerly incarcerated (52.2%), over one-third \nwere incarcerated in the past 4 years (35.0%), and \ntwo-thirds were incarcerated more than once \n(65.9%). These descriptive results are consistent \nwith literature indicating the mass incarceration of \nyoung low-income black adults (see Wildeman and \nWang 2017).2 Among the formerly incarcerated indi-\nviduals, 12.0% experienced direct violence during \nincarceration, and 23.4% experienced secondary \nviolence. GrimAge scores indicate that, on average, \nformerly incarcerated individuals experience accel-\nerated aging (x = 1.179), whereas those who were \nnever incarcerated experience decelerated aging (x = \n\u20131.349, p < .001). Table 1 also reveals differences by \nincarceration exposure across demographic charac-\nteristics, preincarceration health behaviors, and \nchildhood adversity.\nIncarceration Exposure and Accelerated \nBiological Aging\nRegression models predicting the speed of Grim \naging are shown in Table 2. Model 1 presents the esti-\nmates of the association between incarceration expo-\nsure and the speed of aging net of the full array of \nconfounds. The results show that incarceration is \nassociated with accelerated epigenetic aging in this \nAfrican American sample, with exposure predicting a \ngreater accelerated aging score by about 2 years (b = \n1.970; p \u2264 .001). After applying IPTW (Model 2) to \nthe model, the association between incarceration \nexposure and the speed of aging is slightly reduced \nbut remains statistically significant. Exposure to \nincarceration is associated with a 1.719 difference \n(approximately 1 year and 9 months) in accelerated \naging (p \u2264 .001). This finding is illustrated in  \nFigure 1\u2014formerly incarcerated individuals are \nroughly 11 months (12 \u00d7 .902) older biologically than \nexpected based on their chronological (or calendar) \nage, whereas those who have never been incarcerated \nare 10 months younger biologically (12 \u00d7 \u2013.816) than \nexpected based on their chronological age. In Model \n3, we restricted the sample to respondents with pro-\npensity scores that fall within the range of propensi-\nties calculated for both the incarcerated and never \nincarcerated respondents. Similarly, in Model 4, we \nexclude all respondents with atypical propensities\u2014\npropensity scores within the fifth percentiles. Both \nmodels reveal a stronger association between incar-\nceration exposure and the speed of Grim aging, indi-\ncating a conservative estimate in Model 2.3 Together, \nthe evidence from Table 2 indicates that among the \nAfrican American respondents in our sample, incar-\nceration exposure is associated with an acceleration \nin biological aging by 21 to 28 months.\nTo probe the effects of incarceration exposure, \nwe examine whether the timing and frequency of \nincarceration operate as internal moderators of the \nassociation between incarceration exposure and the \nspeed of aging (see Models 1 and 2 of Table 3). \nResults suggest that the timing of incarceration may \nnot moderate the association\u2014formerly incarcer-\nated individuals did not experience greater acceler-\nated aging if they were incarcerated in the past 4 \nyears (Model 1; see also, Boen 2020). Results also \nshow that the effects of incarceration exposure did \nnot vary by the frequency of incarceration spells \n(Model 2).4 These findings align with past studies in \nwhich incarceration exposure or prevalence mat-\ntered more for health outcomes than dosage (e.g., \nSchnittker and John 2007; but see Boen 2020).\n\n468\t\nJournal of Health and Social Behavior 62(4)\nViolence During Incarceration\nNext, in Models 3 and 4 of Table 3, we examine \nwhether violence moderates the association \nbetween incarceration exposure and the speed of \nGrim aging. As shown in Model 3, direct violence \nintensifies the association between incarceration \nexposure and accelerated aging in this African \nAmerican sample (b = 1.388; p < .05). Figure 2 \nillustrates these patterns. As shown, formerly incar-\ncerated individuals who did not report direct vio-\nlence are roughly 9 months (12 \u00d7 .776) older \nbiologically than expected based on their chrono-\nlogical age. Those who did report direct violence \nduring incarceration are approximately 2 years and \n2 months (12 \u00d7 2.164) older biologically than \nexpected based on their chronological age. These \nresults indicate that the direct experiences with vio-\nlence during incarceration moderate the association \nbetween incarceration exposure and the speed of \naging. Secondary violence, however, does not mod-\nerate the association (see Model 4), suggesting that \nwitnessing violence may not exacerbate the already \nstressful incarceration experience for African \nAmerican adults.\nSupplementary Analyses\nThe literature remains mixed regarding gender dif-\nferences in the physical health impacts of incarcera-\ntion. Some studies find that the association between \nincarceration and mortality is stronger for women \nthan men (e.g., Massoglia et\u00a0 al. 2014; Patterson \n2010), whereas studies of incarceration and morbid-\nity tend to find little variation (e.g., Baquero et\u00a0al. \n2020; Schnittker and John 2007; Wang et\u00a0al. 2009). \nFew studies, however, have examined gender differ-\nences among a sample of black individuals (cf. \nSemenza et\u00a0al. 2021). In supplementary analyses, \nwe explore whether the gender of respondents mod-\nerates the association between incarceration expo-\nsure and the speed of Grim aging (not shown in \ntabular form). Approximately 66.9% of males and \n43.4% of females were formerly incarcerated. \nTable 1.\u2002 Means (Standard Deviations) and Proportions of Study Variables Stratified by Incarceration \nExposure.\nVariable\nTotal Sample\n(n = 410)\nNever \nIncarcerated\n(n = 196)\nFormerly \nIncarcerated\n(n = 212)\nt Test\nAccelerated GrimAge\n\u2013.030 (4.202)\n\u20131.349 (3.708)\n1.179 (4.272)\n\u20136.373***\nIncarceration exposure\n.522\n\u2014\n\u2014\n\u2014\n\u2003 Past 4 years\n.183\n\u2014\n.350\n\u2014\n\u2003 More than once\n.344\n\u2014\n.659\n\u2014\n\u2003 Direct violence\n.063\n\u2014\n.120\n\u2014\n\u2003 Secondary violence\n.122\n\u2014\n.234\n\u2014\nMale\n.376\n.260\n.481\n\u20134.731***\nHigh school education\n.915\n.959\n.874\n3.119**\nIncome\n7.908 (4.195)\n8.446 (3.780)\n7.415 (4.494)\n2.502*\nMarried or cohabiting\n.300\n.306\n.294\n.258\nHealth insurance\n.834\n.867\n.804\n1.732\nSmoking\n.067 (.149)\n.039 (.115)\n.093 (.171)\n\u20133.716***\nBinge drinking\n1.235 (.394)\n1.162 (.306)\n1.301 (.451)\n\u20133.632***\nHealthy diet\n2.912 (.931)\n2.906 (.908)\n2.917 (.954)\n\u2013.124\nExercise\n2.653 (.790)\n2.615 (.782)\n2.688 (.796)\n\u2013.939\nBMI\n24.452 (5.239)\n25.047 (5.814)\n23.906 (5.239)\n2.090*\nDepression\n.309 (.730)\n.343 (.748)\n.278 (.714)\n.896\nCommunity disorder\n1.264 (.247)\n1.240 (.237)\n1.287 (.255)\n\u20131.951\nSelf-control\n1.660 (.249)\n1.662 (.266)\n1.658 (.233)\n.124\nDelinquency\n.255 (.510)\n.184 (.387)\n.320 (.595)\n\u20132.725**\nChildhood adversity\n2.844 (2.735)\n2.546 (2.632)\n3.117 (2.805)\n\u20132.120*\nChildhood socioeconomic status\n\u2013.032 (1.449)\n.138 (1.438)\n\u2013.189 (1.444)\n2.293*\nNote: Data from Family and Community Health Study, Waves 1997\u20132016.\n*p < .05, **p < .01, ***p < .001, two-tailed.\n\nBerg et al.\t\n469\nSupplementary models\u2014available on request\u2014\nrevealed that gender does not significantly moderate \nthe association between incarceration exposure and \naccelerated aging in this African American sample \n(b = .303, SE = .900; p = .736). Exposure to incar-\nceration is associated with a 1.904 difference \n(approximately 1 year and 11 months) in acceler-\nated aging for African American men (p \u2264 .001) and \na 1.602 difference (approximately 1 year and 7 \nmonths) in accelerated aging for African American \nwomen (p \u2264 .001).\nDiscussion\nA growing literature traces the consequences of mass \nincarceration for dimensions of individual-level \nhealth and well-being. Some have argued that the \nTable 2.\u2002 Ordinary Least Squares Regression Models Predicting Accelerated GrimAge.\nModel 1a\nModel 2b\nModel 3c\nModel 4d\nVariable\nb\nSE\nb\nSE\nb\nSE\nb\nSE\nIncarceration exposure\n1.970***\n(.408)\n1.719***\n(.415)\n1.846***\n(.429)\n2.339***\n(.446)\nMale\n2.041***\n(.458)\n2.186***\n(.501)\n2.142***\n(.515)\n1.981***\n(.577)\nHigh school education\n\u2013.666\n(.841) \u20131.016\n(.772)\n\u2013.976\n(.786)\n.069\n(1.215)\nIncome\n\u2013.086\n(.057)\n\u2013.063\n(.058)\n-.058\n(.057)\n\u2013.103\n(.065)\nMarried or cohabiting\n\u2013.627\n(.417)\n\u2013.592\n(.443)\n\u2013.500\n(.434)\n\u2013.402\n(.471)\nHealth insurance\n.399\n(.578)\n.219\n(.593)\n.186\n(.594)\n.310\n(.639)\nSmoking\n2.967*\n(1.450)\n4.856*\n(1.906)\n5.827**\n(2.097)\n3.503\n(1.885)\nBinge drinking\n.179\n(.576)\n1.029\n(.555)\n.684\n(.695)\n.527\n(.764)\nHealthy diet\n\u2013.197\n(.204)\n.003\n(.215)\n\u2013.068\n(.216)\n\u2013.154\n(.248)\nExercise\n\u2013.105\n(.266)\n\u2013.028\n(.276)\n\u2013.029\n(.277)\n\u2013.295\n(.331)\nBMI\n.086*\n(.034)\n.074*\n(.036)\n.083*\n(.037)\n.070*\n(.039)\nDepression\n.148\n(.306)\n\u2013.234\n(.256)\n\u2013.060\n(.268)\n.101\n(.346)\nCommunity disorder\n\u2013.405\n(.883)\n\u2013.922\n(.902)\n\u2013.807\n(.929)\n\u2013.092\n(1.012)\nSelf-control\n\u2013.380\n(.838)\n\u2013.356\n(.838)\n\u2013.206\n(.831)\n\u2013.211\n(.920)\nDelinquency\n\u2013.334\n(.482)\n\u2013.426\n(.475)\n\u2013.433\n(.495)\n\u2013.571\n(.502)\nChildhood adversities\n\u2013.064\n(.076)\n\u2013.010\n(.081)\n\u2013.031\n(.088)\n\u2013.043\n(.087)\nChildhood socioeconomic \nstatus\n\u2013.040\n(.161)\n\u2013.008\n(.152)\n.097\n(.151)\n.094\n(.183)\nConstant\n\u2013.989\n(2.282) \u20131.641\n(2.217) \u20131.767\n(2.286) \u20132.235\n(2.619)\nNote: Data from Family and Community Health Study, Waves 1997\u20132016. Unstandardized coefficients (b) and robust \nstandard errors (SE) are presented.\naOrdinary least squares regression (n = 410).\nbOrdinary least squares regression with inverse probability of treatment weighting (n = 410).\ncOrdinary least squares regression with inverse probability of treatment weighting restricted to common support  \n(n = 402).\ndOrdinary least squares regression with inverse probability of treatment weighting trimmed at the fifth percentiles  \n(n = 313).\n*p < .05, **p < .01, ***p < .001, two-tailed.\n-2\n-1.5\n-1\n-0.5\n0\n0.5\n1\n1.5\n2\n2.5\n3\n3.5\nNever Incarcerated\nFormerly Incarcerated\nla\ncig\nolo\nn\no\nr\nh\nC\nm\no\nrf\ne\ng\nA\nM\nI\nR\nG\nf\no\nn\noitaiv\ne\nD\nAge in Years\nFigure 1.\u2002 Speed of GrimAge by Incarceration \nExposure.\nNote: Data from Family and Community Health Study, \nWaves 1997\u20132016. Marginal effects estimates derived \nfrom regression Model 2 of Table 2. Vertical bars \nindicate the 95% confidence interval; n = 410.\n\n470\t\nJournal of Health and Social Behavior 62(4)\npervasiveness and concentration of incarceration \namong disadvantaged groups contributes to widening \nhealth inequalities (Wildeman and Muller 2012:12). \nAlthough research has reported associations between \nincarceration exposure and various negative health \noutcomes, important substantive and methodological \ngaps remain in this body of work. These gaps reflect \nspecial challenges in determining whether and why \nformerly incarcerated individuals report an elevated \nburden of chronic illnesses and diseases. The goal of \nthe current study was to extend this area of research \nby examining whether incarceration exposure is \nrelated to the pace of biological aging and if experi-\nences with violence during incarceration intensify \nthis association. To do so, the study employed data on \na sample of African American adults along with mea-\nsures of several relevant confounds from different life \ncourse domains, and it implemented adjustments for \nselection.\nThe results of the analysis support two key sub-\nstantive conclusions. First, incarceration exposure \nsignificantly predicts accelerated biological aging \namong African American adults, as assessed in the \nGrimAge measure. Study participants who reported \na history of incarceration were, on average, biologi-\ncally older than expected given their numerical age. \nNotably, the estimates were robust to a host of \npotential confounding factors obtained earlier in the \nlife course, including adverse childhood experi-\nences, social class, depression, and delinquency. \nEven after implementing statistical adjustments,  \nthe findings remained stable. This is noteworthy \nbecause the proposition that any observed effects of \nincarceration on physical health are due to selec-\ntion, confounding, or both has a strong logical basis \n(Schnittker et\u00a0al. 2012). The results therefore pro-\nvide reasonable evidence that the stressors of incar-\nceration exposure appear to contribute to the pace \nof epigenetic aging.\nThis evidence is important for understanding the \nphysical health toll of incarceration among African \nAmericans because accelerated aging is a funda-\nmental predisease process. This process reliably pre-\ndicts mortality risk and the onset of chronic and \nacute morbidities (Sinclair 2019). Research has \nTable 3.\u2002 Ordinary Least Squares Regression Models Predicting Accelerated GrimAge from \nIncarceration Timing, Frequency and Exposure to Violencea\nModel 1\nModel 2\nModel 3\nModel 4\nVariable\nb\nSE\nb\nSE\nb\nSE\nb\nSE\nIncarceration exposure\n1.504***\n(.452)\n1.194*\n(.546)\n1.593***\n(.424)\n1.616***\n(.441)\nInternal moderators:\n\u2002\n\u2003 \u00d7 past 4 years\n.693\n(.613)\n\u2014\n\u2014\n\u2014\n\u2002\n\u2003 \u00d7 more than once\n\u2014\n.854\n(.624)\n\u2014\n\u2014\n\u2002\n\u2003 \u00d7 direct violence\n\u2014\n\u2014\n1.388*\n(.701)\n\u2014\n\u2002\n\u2003 \u00d7 secondary violence\n\u2014\n\u2014\n\u2014\n.488\n(.625)\nNote: Data from Family and Community Health Study, Waves 1997\u20132016. Unstandardized coefficients (b) and robust \nstandard errors (SE) are presented.\naOrdinary least squares regression with inverse probability of treatment weighting (n = 410). Confound estimates \nnot shown in the table: male, high school education, income, married or cohabiting, health insurance, smoking, binge \ndrinking healthy diet, exercise, BMI, depression, community disorder, self-control, delinquency, childhood adversities, \nchildhood socioeconomic status.\n*p < .05, *** p < .001, two-tailed.\n-2\n-1.5\n-1\n-0.5\n0\n0.5\n1\n1.5\n2\n2.5\n3\n3.5\nNever Incarcerated\nNo Direct Violence\nDirect Violence\n e\ng\nA\n la\ncig\nolo\nn\no\nr\nh\nC\n \nm\no\nrf e\ng\nA\n \nm\nir\nG\n f\no n\noitaiv\ne\nD\nin Years\nFigure 2.\u2002 Speed of GrimAge by Incarceration \nExposure and Direct Violence during \nIncarceration.\nNote: Data from Family and Community Health Study, \nWaves 1997\u20132016. Marginal effects estimates derived \nfrom multivariate regression model in Model 3 of  \nTable 3. Vertical bars indicate the 95% confidence \ninterval; n = 410.\n\nBerg et al.\t\n471\nshown that accelerated aging is among the most \nimportant drivers of adverse health conditions \nalready linked to incarceration, including hyperten-\nsion and cancer (Massoglia 2008; Wang et\u00a0al. 2009), \nall of which diminish total health span (Murabito \net\u00a0al. 2018; Yoon et\u00a0al. 2019). Epigenetic aging may \nbe a biological pathway whereby the hardships of \nincarceration exposure affect the onset of acute ill-\nnesses and chronic conditions. Furthermore, the cur-\nrent findings add to the burgeoning body of work \nfocused on how incarceration may \u201cget under the \nskin\u201d to leave an imprint on bodily systems (e.g., \nBoen 2020; Esposito et\u00a0al. 2017). That we uncov-\nered evidence of accelerated aging in a sample of \nAfrican American adults who were approximately \n30 years of age is worth noting\u2014this is a period in \nthe life course when age-related chronic diseases \nand emerging pathology typically do not yet mani-\nfest and register clinically. As Moffitt and colleagues \n(2017) note, accelerated aging is not simply an \naffliction of the elderly but a lifelong biological pro-\ncess that begins early in the life span.\nSecond, exposure to violence during incarcera-\ntion exacerbates the degree to which incarceration \nexposure hastens the pace of Grim aging among \nAfrican American adults. Those who experienced \nviolence were more than 2 years older than their cal-\nendar age. This effect, however, was exclusive to \ndirect incidents of violent threats, assaults with and \nwithout weapons, and robbery. Interpersonal vio-\nlence likely engenders a stress response that chroni-\ncally activates immunological and nervous system \ncapacities. This evidence aligns with the literature \non prison violence and mental health (Zweig et\u00a0al. \n2015). Moreover, it accords with research indicating \nthat threatening social environments trigger adverse \nphysiological responses (Miller et\u00a0al. 2011; Simons \net\u00a0al. 2021). Perhaps violent victimization is espe-\ncially stressful in the confined spaces of the prison \ncontext where hostile interpersonal relations are par-\nticularly difficult to circumvent. As one formerly \nincarcerated individual explained: \u201cThe whole men-\ntal thing of me\u2009.\u2009.\u2009.\u2009trying to think of who I could have \na problem with or who might have a problem with \nme all the time\u2009.\u2009.\u2009.\u2009is extremely stressful\u201d (Porter \n2019:9). Other potentially traumatic prison and jail \nconditions are deserving of additional research on \nthe biological consequences of incarceration, \nincluding discriminatory treatment and placement in \nsolitary confinement, the latter of which has been \nlinked to mental health (Strong et\u00a0 al. 2020). \nSimilarly, researchers should study conditions dur-\ning postrelease that might buffer the link between \nincarceration and adverse physiological deteriora-\ntion, including relational and community supports.\nThe present study did not examine how incar-\nceration contributes to racial disparities in aging \ngiven that the sample was limited to African \nAmerican individuals. But this work is needed. \nContact with the criminal justice system\u2014stops, \narrests, and jail and prison stays\u2014is disproportion-\nately felt by African Americans owing to a legacy \nof discrimination and marginalization in a racial-\nized society. Whether the linkage between incar-\nceration and adverse health outcomes differs by \nrace is the focus of ongoing research (Boen 2020; \nSemenza et\u00a0al. 2021). The findings from this study \nsuggest ways forward with research on health dis-\nparities, particularly biological markers. As noted, \nepigenetic aging indices are thought to be direct \nmeasures of the biological weathering conceptual-\nized in Geronimus and colleagues\u2019 (2006) model of \nhealth inequalities (see Simons et\u00a0 al. 2021). \nPerhaps, then, epigenetic aging serves as a biologi-\ncal mechanism through which disparities in incar-\nceration contribute to racial differences in chronic \nmorbidity and excess mortality. Note, however, that \nwe are unable to test this possibility given our reli-\nance on an African American sample. Future \nresearch should examine the validity of this \nassumption given that it might hold important clues \nabout how mass incarceration and other race-\nrelated social adversities combine to affect health \ninequalities. Modern medicine often considers dif-\nferences in health risk behaviors (e.g., diet) as the \nmajor cause of class and race disparities in chronic \nillness (Milani and Lavie 2015). Some argue that \nthis view perhaps exaggerates the importance of \nhealth risk behaviors as a cause of health inequali-\nties. Our findings show that regardless of diet, exer-\ncise, and smoking, incarceration exposure is a \nsignificant social determinant of accelerated aging.\nThe current study is not without limitations. \nDespite utilizing a series of preincarceration mea-\nsures of likely confounds, the analysis did not assess \nthe GrimAge of respondents prior to incarceration. \nRather, GrimAge was obtained at a single period\u2014\nWave 7. Assessing a biomarker of aging at multiple \ntime points would allow for a more compelling \nassessment of how incarceration affects changes \nacross the life course in the pace of biological aging. \nThis question should be a priority for future \nresearch. Readers should also use caution when gen-\neralizing our findings to other health biomarkers, \nincluding objective markers of inflammation and \ncardiovascular disease risk. Furthermore, the IPTW \n\n472\t\nJournal of Health and Social Behavior 62(4)\nprocedure is intended to minimize threats of selec-\ntion based on differences in observables, which pro-\nvides some confidence in our estimates. However, \nas noted earlier, the quality of these estimates is \ndependent on the reliability and validity and scope \nof the observed measures. Furthermore, the esti-\nmates remain vulnerable to unobserved sources of \nselection that are not controlled in the current IPTW \nmethods but may be addressed through fixed effects \nestimators in longitudinal data.\nAdditionally, as noted, our sample consists of \nAfrican American adults under age 30, which limits \nthe generalizability of our findings to other groups. \nAlthough an exclusive focus on an African American \nsample is a strength given their high level of expo-\nsure to incarceration and worse health profile, future \nresearch should explore the extent to which incar-\nceration accelerates aging among other racial and \nethnic groups. Similarly, researchers should con-\nsider the possibility that our findings could vary in a \nsample of older adults with a more substantial health \nburden. A sample that included older adults could \nalso allow for a detailed assessment of how the tim-\ning of incarceration may affect the speed of aging. \nIn addition, because the majority of FACHS respon-\ndents reside in two states, this limits the generaliz-\nability of the findings to individuals located in other \nstates. The characteristics of prisons and jails and \ncorrectional policies, including reentry program-\nming, in each of these states might uniquely affect \nthe biological health of the formerly incarcerated in \nsuch a way that similar findings may not be observed \namong those in other states.\nThe consequences of mass incarceration for \nhealth and well-being have become a topic of inter-\nest among policymakers and academics from mul-\ntiple scientific disciplines for good reason: With so \nmany Americans cycling through prisons and jails \neach year, the public health toll of the carceral state \nis likely to mount, creating challenges for commu-\nnities already beleaguered by accumulated hard-\nships. The findings would seem to have important \nintervention and policy implications. At the very \nleast, they provide further support for decreasing \nthe scale of the carceral state and providing sound \nalternatives to punitive penal policies for the sake \nof reversing health inequalities.\nFunding\nThe authors disclosed receipt of the following financial sup-\nport for the research, authorship, and/or publication of this \narticle:This research was supported by the National Heart, \nLung, Blood Institute (R01 HL118045), the National Institute \non Child Health and Human Development (R01 HD080749), \nthe National Institute on Aging (R01 AG055393), the \nNational Institute on Drug Abuse (P30 DA027827), the \nNational Institute of Mental Health (R01 MH62699), and the \nCenter for Disease Control (1 R49 CE003095-01).\nORCID iDs\nMark T. Berg \n https://orcid.org/0000-0003-2709-0950\nRonald L. Simons \n https://orcid.org/0000-0002- \n4902-0462\nSuppplemental Material\nAppendices A through C are available in the online ver-\nsion of the article.\nNotes\n1.\t\nWe reestimated our analyses using a multiple impu-\ntation strategy to recover missing values from the \nconfounding measures\u2014this step increased the \nsample size to 436 respondents. The results did not \nsubstantively depart from the main findings and \nconclusions.\n2.\t\nFurthermore, Iowa\u2014where many in the Family \nand Community Health Study (FACHS) sample \nreside\u2014incarcerates black residents at a rate that \nexceeds most other states (Beck and Blumstein \n2018). Compared to other states, Georgia\u2019s rate of \nincarceration (1,110 per 100,000 adult residents) is \nhigher than the average rate (750 per 100,000 adult \nresidents) in the United States (Maruschak and \nMinton 2020). Iowa\u2019s rate of incarceration (580 per \n100,000 adult residents) is lower than the average in \nthe United States.\n3.\t\nIn an additional sensitivity analysis, we use a regres-\nsion adjustment estimator in addition to inverse \nprobability of treatment weighting (see Boen 2020). \nWhen using a regression adjustment, we find a simi-\nlar treatment effect (b = 1.836, SE = .386, p = .000).\n4.\t\nAll formerly incarcerated individuals in the sample \nspent time in jail; 25 (6.1%) spent time in prison in \naddition to jail. There were no cases where individ-\nuals spent time in prison but not jail. Supplemental \nanalyses revealed that exposure to prison did not \nhave a unique effect on the speed of aging among \nthose exposed to incarceration (b = \u2013.348, SE = \n.906, p = .701).\nReferences\nAssari, Shervin, Maryam Moghani Lankarani, and Sarah \nBurgard. 2016. \u201cBlack\u2013White Differences in Long-\nTerm Predictive Power of Self-Rated Health on \nAll-Cause Mortality in the United States.\u201d Annals of \nEpidemiology 26(2):106\u201314.\nBaquero, Mar\u00eda, Kimberly Zweig, Sonia Y. Angell, and \nSharon B. Meropol. 2020. \u201cHealth Behaviors and \nOutcomes Associated with Personal and Family \n\nBerg et al.\t\n473\nHistory of Criminal Justice System Involvement, \nNew York City, 2017.\u201d American Journal of Public \nHealth 110(3):378\u201384.\nBeck, Allen J., and Alfred Blumstein. 2018. \u201cRacial \nDisproportionality in U.S. State Prisons: Accounting \nfor the Effects of Racial and Ethnic Differences in \nCriminal Involvement, Arrests, Sentencing, and \nTime Served.\u201d Journal of Quantitative Criminology \n34(3):853\u201383.\nBerg, Mark T., Man-Kit Lei, Steven R. Beach, Ronald \nL. Simons, and Leslie Gordon Simons. 2020. \n\u201cChildhood \nAdversities \nas \nDeterminants \nof \nCardiovascular Disease Risk and Perceived Illness \nBurden in Adulthood: Comparing Retrospective and \nProspective Self-Report Measures in a Longitudinal \nSample of African Americans.\u201d Journal of Youth and \nAdolescence 49(6):1292\u2013308.\nBeyers, Jennifer M., and Rolf Loeber. 2003. \u201cUntangling \nDevelopmental Relations between Depressed Mood \nand Delinquency in Male Adolescents.\u201d Journal of \nAbnormal Child Psychology 31(3):247\u201366.\nBoen, Courtney E. 2020. \u201cCriminal Justice Contacts \nand Psychophysiological Functioning in Early \nAdulthood: Health Inequality in the Carceral State.\u201d \nJournal of Health and Social Behavior 61(3): \n290\u2013306.\nBoxer, Paul, Keesha Middlemass, and Tahlia Delorenzo. \n2009. \u201cExposure to Violent Crime during Incarceration: \nEffects on Psychological Adjustment Following \nRelease.\u201d Criminal Justice and Behavior 36(8): \n793\u2013807.\nCarson, E. Ann. 2020. Prisoners in 2018. Washington, \nDC: U.S. Department of Justice.\nCole, Stephen R., and Miguel A Hern\u00e1n. 2008. \n\u201cConstructing Inverse Probability Weights for \nMarginal Structural Models.\u201d American Journal of \nEpidemiology 168(6):656\u201364.\nDanese, Andrea, and Jessie R. Baldwin. 2017. \u201cHidden \nWounds? Inflammatory Links between Childhood \nTrauma and Psychopathology.\u201d Annual Review of \nPsychology 68:517\u201344.\nEsposito, Michael H., Hedwig Lee, Margret T. Hicken, \nLauren C. Porter, and Jerald R. Herting. 2017. \u201cThe \nConsequences of Contact with the Criminal Justice \nSystem for Health in the Transition to Adulthood.\u201d \nLongitudinal Life Course Studies 8(1):57\u201374.\nGeronimus, Arline T., Margaret Hicken, Danya Keene, \nand John Bound. 2006. \u201c\u2018Weathering\u2019 and Age \nPatterns of Allostatic Load Scores among Blacks and \nWhites in the United States.\u201d American Journal of \nPublic Health 96(5):826\u201333.\nHammett, Theodore M., Mary Patricia Harmon, and \nWilliam Rhodes. 2002. \u201cThe Burden of Infectious \nDisease among Inmates of and Releasees from US \nCorrectional Facilities, 1997.\u201d American Journal of \nPublic Health 92(11):1789\u201394.\nHannum, Gregory, Justin Guinney, Ling Zhao, Li Zhang, \nGuy Hughes, SriniVas Sadda, Brandy Klotzle, \net\u00a0 al. 2013. \u201cGenome-Wide Methylation Profiles \nReveal Quantitative Views of Human Aging Rates.\u201d \nMolecular Cell 49(2):359\u201367.\nHarding, David J., Jessica J. B. Wyse, Cheyney Dobson, \nand Jeffrey D. Morenoff. 2014. \u201cMaking Ends \nMeet after Prison.\u201d Journal of Policy Analysis and \nManagement 33(2):440\u201370.\nHarris, Kathleen Mullan, and Kristen M. Schorpp. 2018. \n\u201cIntegrating Biomarkers in Social Stratification \nand Health Research.\u201d Annual Review of Sociology \n44:361\u201386.\nHorvath, Steve, and Kenneth Raj. 2018. \u201cDNA \nMethylation-Based Biomarkers and the Epigenetic \nClock Theory of Ageing.\u201d Nature Reviews Genetics \n19(6):371\u201384.\nHoule, \nBrian. \n2011. \n\u201cObesity \nDisparities \namong \nDisadvantaged Men: National Adult Male Inmate \nPrevalence Pooled with Non-incarcerated Estimates, \nUnited States, 2002\u20132004.\u201d Social Science & Medicine \n72(10):1667\u201373.\nIrwin, John. 1980. Prisons in Turmoil. Boston, MA: \nLittle, Brown.\nJylh\u00e4, Marja. 2009. \u201cWhat Is Self-Rated Health and Why \nDoes It Predict Mortality? Towards a Unified Conceptual \nModel.\u201d Social Science & Medicine 69(3):307\u201316.\nKendall, Philip C., and Lance E. Wilcox. 1979. \u201cSelf-\nControl in Children: Development of a Rating \nScale.\u201d Journal of Consulting & Clinical Psychology \n47(6):1020\u201329.\nKennedy, Brian K., Shelley L. Berger, Anne Brunet, Judith \nCampisi, Ana Maria Cuervo, Elissa S. Epel, Claudio \nFranceschi, et al. 2014. \u201cGeroscience: Linking Aging \nto Chronic Disease.\u201d Cell 159(4):709\u201313.\nKoch, Carmen M., and Wolfgang Wagner. 2011. \n\u201cEpigenetic-Aging-Signature to Determine Age in \nDifferent Tissues.\u201d Aging 3(10):1\u201310.\nLabrecque, Ryan M., Heidi Scherer, and James \nT. McCafferty. 2018. \u201cReducing Violence in \nCorrectional Institutions: Revalidation of the \nInmate Risk Assessment for Violent, Nonsexual \nVictimization.\u201d Violence and Victims 33(1):126\u201341.\nLeBel, Thomas P. 2017. \u201cHousing as the Tip of the \nIceberg in Successfully Navigating Prisoner Reentry.\u201d \nCriminology & Public Policy 16(3):889\u2013906.\nLei, Man-Kit, Steven R. H. Beach, and Ronald L. Simons. \n2018. \u201cBiological Embedding of Neighborhood \nDisadvantage and Collective Efficacy: Influences on \nChronic Illness via Accelerated Cardiometabolic Age.\u201d \nDevelopment and Psychopathology 30(5):1797\u2013815.\nLu, Ake T., Austin Quach, James G. Wilson, Alex P. \nReiner, Abraham Aviv, Kenneth Raj, Lifang Hou, et\u00a0al. \n2019. \u201cDNA Methylation GrimAge Strongly Predicts \nLifespan and Healthspan.\u201d Aging 11(2):303\u201327.\nMaier, Katharina H., and Rosemary Ricciardelli. 2019. \n\u201cThe Prisoner\u2019s Dilemma: How Male Prisoners \nExperience and Respond to Penal Threat While \nIncarcerated.\u201d Punishment & Society 21(2):231\u201350.\nMaruschak, Laura M., and Todd D. Minton. 2020. \nCorrectional Populations in the United States, 2017-\n2018. Washington DC: U.S. Department of Justice.\n\n474\t\nJournal of Health and Social Behavior 62(4)\nMassoglia, Michael. 2008. \u201cIncarceration as Exposure: \nThe Prison, Infectious Disease, and Other Stress-\nRelated Illnesses.\u201d Journal of Health and Social \nBehavior 49(1):56\u201371.\nMassoglia, Michael, Paul-Philippe Pare, Jason Schnittker, \nand Alain Gagnon. 2014. \u201cThe Relationship between \nIncarceration and Premature Adult Mortality: \nGender Specific Evidence.\u201d Social Science Research \n46:142\u201354.\nMassoglia, Michael, and William Alex Pridemore. \n2015. \u201cIncarceration and Health.\u201d Annual Review of \nSociology 41:291\u2013310.\nMassoglia, Michael, and Brianna Remster. 2019. \n\u201cLinkages Between Incarceration and Health.\u201d \nPublic Health Reports 134(Suppl. 1):8S\u201314S.\nMcCrory, Cathal, Giovanni Fiorito, Belinda Hernandez, Silvia \nPolidoro, Aisling M. O\u2019Halloran, Ann Hever, Cliona Ni \nCheallaigh, et\u00a0al. 2021. \u201cGrimAge Outperforms Other \nEpigenetic Clocks in the Prediction of Age-Related \nClinical Phenotypes and All-Cause Mortality.\u201d The \nJournals of Gerontology, Series A: Biological Sciences \nand Medical Sciences 76(5):741\u201349.\nMcEwen, Bruce S., and Peter J. Gianaros. 2011. \u201cStress- \nand Allostasis-Induced Brain Plasticity.\u201d Annual \nReview of Medicine 62:431\u201345.\nMeaney, Michael J., and Moshe Szyf. 2005. \u201cEnvironmental \nProgramming of Stress Responses Through DNA \nMethylation: Life at the Interface Between a Dynamic \nEnvironment and a Fixed Genome.\u201d Dialogues \nClinical Neuroscience 7(2):103\u201323.\nMilani, Richard V., and Carl J. Lavie. 2015. \u201cHealthcare \n2020: Reengineering Health Care Delivery to \nCombat Chronic Disease.\u201d The American Journal of \nMedicine 128(4):337\u201343.\nMiller, Gregory E., Edith Chen, and Karen J. Parker. \n2011. \u201cPsychological Stress in Childhood and \nSusceptibility to the Chronic Diseases of Aging: \nMoving towards a Model of Behavioral and \nBiological Mechanisms.\u201d Psychological Bulletin \n137(6):959\u201397.\nMirowsky, John. 2013. \u201cAnalyzing Associations between \nMental Health and Social Circumstances.\u201d Pp. 143\u2013\n65 in Handbook of the Sociology of Mental Health, \nedited by C. S. Aneshensel, J. C. Phelan, and J. \nMirowsky. Dordrecht: Springer Netherlands.\nMoffitt, Terrie E., Daniel W. Belsky, Andrea Danese, \nRichie Poulton, and Avshalom Caspi. 2017. \u201cThe \nLongitudinal Study of Aging in Human Young \nAdults: Knowledge Gaps and Research Agenda.\u201d \nJournals of Gerontology, Series A: Biological \nSciences and Medical Sciences 72(2):210\u201315.\nMorton, Patricia M., and Kenneth F. Ferraro. 2020. \n\u201cEarly Social Origins of Biological Risks for Men \nand Women in Later Life.\u201d Journal of Health and \nSocial Behavior 61(4):503\u201322.\nMurabito, Joanne M., Qiang Zhao, Martin G. Larson, Jian \nRong, Honghuang Lin, Emelia J. Benjamin, Daniel \nLevy, and Kathryn L. Lunetta. 2018. \u201cMeasures of \nBiologic Age in a Community Sample Predict Mortality \nand Age-Related Disease: The Framingham Offspring \nStudy.\u201d Journal of Gerontology, Series A: Biological \nSciences and Medical Sciences 73(6):757\u201362.\nNovisky, Meghan A., and Robert L. Peralta. 2020. \n\u201cGladiator School: Returning Citizens\u2019 Experiences \nwith Secondary Violence Exposure in Prison.\u201d \nVictims & Offenders 15(5):594\u2013618.\nPatterson, Evelyn J. 2010. \u201cIncarcerating Death: Mortality \nin U.S. State Correctional Facilities, 1985\u20131998.\u201d \nDemography 47(3):587\u2013607.\nPearlin, Leonard I., Scott Schieman, Elena M. Fazio, and \nStephen C. Meersman. 2005. \u201cStress, Health, and \nthe Life Course: Some Conceptual Perspectives.\u201d \nJournal of Health and Social Behavior 46(2):205\u201319.\nPetersilia, Joan. 2003. When Prisoners Come Home: \nParole and Prisoner Reentry. Oxford, UK: Oxford \nUniversity Press.\nPorter, Lauren C. 2019. \u201cBeing \u2018on Point\u2019: Exploring \nthe Stress-Related Experiences of Incarceration.\u201d \nSociety and Mental Health 9(1):1\u201317.\nRobins, James M., Miguel Angel Hern\u00e1n, and Babette \nBrumback. 2000. \u201cMarginal Structural Models and \nCausal Inference in Epidemiology.\u201d Epidemiology \n11(5):550\u201360.\nRosenbaum, Paul R., and Donald B. Rubin. 1983. \u201cThe \nCentral Role of the Propensity Score in Observational \nStudies for Causal Effects.\u201d Biometrika 70(1):41\u201355.\nRoss, Catherine E., and John Mirowsky. 1992. \n\u201cHouseholds, Employment, and the Sense of \nControl.\u201d Social Psychology Quarterly 55(3):217\u201335.\nRyan, Joanne, Jo Wrigglesworth, Jun Loong, Peter D. \nFransquet, and Robyn L. Woods. 2020. \u201cA Systematic \nReview and Meta-analysis of Environmental, \nLifestyle, and Health Factors Associated with \nDNA Methylation Age.\u201d Journals of Gerontology, \nSeries A: Biological Sciences and Medical Sciences \n75(3):481\u201394.\nSchnittker, Jason, and Andrea John. 2007. \u201cEnduring \nStigma: The Long-Term Effects of Incarceration \non Health.\u201d Journal of Health and Social Behavior \n48(2):115\u201330.\nSchnittker, Jason, Michael Massoglia, and Christopher \nUggen. 2012. \u201cOut and Down: Incarceration and \nPsychiatric Disorders.\u201d Journal of Health and Social \nBehavior 53(4):448\u201364.\nSemenza, Daniel C., and Nathan W. Link. 2019. \u201cHow Does \nReentry Get under the Skin? Cumulative Reintegration \nBarriers and Health in a Sample of Recently \nIncarcerated Men.\u201d Social Science & Medicine \n243:112618. doi:10.1016/j.socscimed.2019.112618.\nSemenza, Daniel C., Alexander M. Testa, Dylan B. Jackson, \nand Michael G. Vaughn. 2021. \u201cIncarceration and \nCardiovascular Health: Multiple Mechanisms within an \nIntersectional Framework.\u201d Journal of Criminal Justice \n73:101759. doi:10.1016/j.jcrimjus.2020.101759.\nShaffer, David, Mary Schwab-Stone, Prudence Fisher, \nPatricia Cohen, John Placentini, Mark Davies, C. Keith \nConners, and Darrel Regier. 1993. \u201cThe Diagnostic \nInterview Schedule for Children\u2013Revised Version \n\nBerg et al.\t\n475\n(DISC-R): I. Preparation, Field Testing, Interrater \nReliability, and Acceptability.\u201d Journal of the \nAmerican Academy of Child & Adolescent Psychiatry \n32(3):643\u201350.\nShannon, Sarah K. S., Christopher Uggen, Jason \nSchnittker, Melissa Thompson, Sara Wakefield, and \nMichael Massoglia. 2017. \u201cThe Growth, Scope, and \nSpatial Distribution of People with Felony Records \nin the United States, 1948\u20132010.\u201d Demography \n54(5):1795\u2013818.\nSimons, Ronald L., Man-Kit Lei, Steven R. H. Beach, \nLeslie Gordon Simons, Ashley B. Barr, Frederick X. \nGibbons, and Robert A. Philibert. 2019. \u201cTesting Life \nCourse Models Whereby Juvenile and Adult Adversity \nCombine to Influence Speed of Biological Aging.\u201d \nJournal of Health and Social Behavior 60(3):291\u2013308.\nSimons, Ronald L., Man-Kit Lei, Eric Klopack, Steven R. H. \nBeach, Frederick X. Gibbons, and Robert A. Philibert. \n2021. \u201cThe Effects of Social Adversity, Discrimination, \nand Health Risk Behaviors on the Accelerated Aging of \nAfrican Americans: Further Support for the Weathering \nHypothesis.\u201d Social Science & Medicine. 282:113169. \ndoi:10.1016/j.socscimed.2020.113169.\nSinclair, David. 2019. Lifespan: Why We Age\u2014And \nWhy We Don\u2019t Have to. New York, NY: Simon & \nSchuster.\nSnyder-Mackler, Noah, Joseph Robert Burger, Lauren \nGaydosh, Daniel W. Belsky, Grace A. Noppert, \nFernando A. Campos, Alessandro Bartolomucci, \net\u00a0 al. 2020. \u201cSocial Determinants of Health and \nSurvival in Humans and other Animals.\u201d Science \n368(6493):eaax9553. doi:10.1126/science.aax9553.\nSong, Shufei, and F. Brad Johnson. 2018. \u201cEpigenetic \nMechanisms Impacting Aging: A Focus on Histone \nLevels and Telomeres.\u201d Genes 9(201):1\u201327.\nStrong, Justin D., Keramet Reiter, Gabriela Gonzalez, \nRebecca Tublitz, Dallas Augustine, Melissa Barragan, \nKelsie Chesnut, Pasha Dashtgard, Natalie Pifer, and \nThomas R. Blair. 2020. \u201cThe Body in Isolation: \nThe Physical Health Impacts of Incarceration in \nSolitary Confinement.\u201d PLoS One 15(10):e0238510. \ndoi:10.1371/journal.pone.0238510.\nSykes, Gesham M. 1958. The Society of Captives: A \nStudy of a Maximum Security Prison. Princeton, NJ: \nPrinceton University Press.\nTeasdale, Brent, Leah E. Daigle, Shila R. Hawk, and \nJane C. Daquin. 2016. \u201cViolent Victimization in the \nPrison Context: An Examination of the Gendered \nContexts of Prison.\u201d International Journal of \nOffender Therapy and Comparative Criminology \n60(9):995\u20131015.\nTyler, Elizabeth Tobin, and Bradley Brockmann. 2017. \n\u201cReturning Home: Incarceration, Reentry, Stigma, \nand the Perpetuation of Racial and Socioeconomic \nHealth Inequity.\u201d The Journal of Law, Medicine, & \nEthics 45(4):545\u201357.\nWang, Emily A., Mark Pletcher, Feng Lin, Eric \nVittinghoff, Stefan G. Kertesz, Catarina I. Kiefe, \nand Kirsten Bibbins-Domingo. 2009. \u201cIncarceration, \nIncident Hypertension, and Access to Health \nCare: Findings from the Coronary Artery Risk \nDevelopment in Young Adults (CARDIA) Study.\u201d \nArchives of Internal Medicine 169(7):687\u201393.\nWestern, Bruce, Anthony A. Braga, Jaclyn Davis, and \nCatherine Sirois. 2015. \u201cStress and Hardship after \nPrison.\u201d American Journal of Sociology 120(5):1512\u201347.\nWiddowson, Alex O., and Benjamin W. Fisher. 2020. \n\u201cMass Incarceration and Subsequent Preventive Health \nCare: Mechanisms and Racial-Ethnic Disparities.\u201d \nAmerican Journal of Public Health 110(S1):S145\u201351.\nWildeman, Christopher, and Christopher Muller. 2012. \n\u201cMass Imprisonment and Inequality in Health and \nFamily Life.\u201d Annual Review of Law and Social \nScience 8(1):11\u201330.\nWildeman, Christopher, and Emily A. Wang. 2017. \n\u201cMass Incarceration, Public Health, and Widening \nInequality in the USA.\u201d Lancet 389(10077): \n1464\u201374.\nWinkelman, Tyler N. A., HwaJung Choi, and Matthew \nM. Davis. 2017. \u201cThe Affordable Care Act, Insurance \nCoverage, and Health Care Utilization of Previously \nIncarcerated Young Men: 2008\u20132015.\u201d American \nJournal of Public Health 107(5):807\u201311.\nWolff, Nancy, Cynthia L. Blitz, Jing Shi, Jane Siegel, and \nRonet Bachman. 2007. \u201cPhysical Violence inside \nPrisons: Rates of Victimization.\u201d Criminal Justice \nand Behavior 34(5):588\u201399.\nWolff, Nancy, and Jing Shi. 2009. \u201cContextualization \nof Physical and Sexual Assault in Male Prisons: \nIncidents and Their Aftermath.\u201d Journal of \nCorrectional Health Care 15(1):58\u201377.\nYang, Ruoting, Gwyneth W. Y. Wu, Josine E. Verhoeven, \nAarti Gautam, Victor I. Reus, Jee In Kang, Janine \nD. Flory, et\u00a0 al. 2020. \u201cA DNA Methylation Clock \nAssociated with Age-Related Illnesses and Mortality \nIs Accelerated in Men with Combat PTSD.\u201d Molecular \nPsychiatry. doi:10.1038/s41380-020-0755-z.\nYoon Young Soon, Minhee Jin, and Don D. Sin. 2019. \n\u201cAccelerated Lung Aging and Chronic Obstructive \nPulmonary Disease.\u201d Expert Review of Respiratory \nMedicine 13(4):369\u201380.\nZajacova, Anna, and Jennifer Beam Dowd. 2011. \n\u201cReliability of Self-Rated Health in US Adults.\u201d \nAmerican Journal of Epidemiology 174(8):977\u201383.\nZeng, Zhen. 2020. Jail Inmates in 2018. Washington, \nDC: U.S. Department of Justice.\nZweig, Janine M., Jennifer Yahner, Christy A. Visher, \nand Pamela K. Lattimore. 2015. \u201cUsing General \nStrain Theory to Explore the Effects of Prison \nVictimization Experiences on Later Offending and \nSubstance Use.\u201d The Prison Journal 95(1):84\u2013113.\nAuthor Biographies\nMark T. Berg is a professor of sociology and criminology \nand director of the Crime and Justice Policy Research \nProgram at the University of Iowa, where he is also a \nCollegiate Scholar. His current research examines conflict \n\n476\t\nJournal of Health and Social Behavior 62(4)\nescalation and the life course determinants of biological \naging and cardiovascular disease risk.\nEthan Rogers is an assistant research scientist in the \nCrime and Justice Policy Research Program at the \nUniversity of Iowa Public Policy Center. His research uti-\nlizes quantitative analyses to explore interpersonal conflict \nand violence, the spatial distribution of crime across com-\nmunities, trends in violence, and the health consequences \nof incarceration.\nMan-Kit Lei is an assistant professor of sociology and \naffiliated with the Center for Family Research at the \nUniversity of Georgia. His research utilizes quantitative \nmethods to examine health-related outcomes such as \nchronic inflammation and rate of biological aging. Recent \narticles have investigated the impact, for better or worse, \nof the social environment on biomarkers of health, espe-\ncially among minority populations.\nRonald L. Simons is Regent\u2019s professor in the Department \nof Sociology and codirector of the Center on the Biological \nEmbedding of Social Events and Relationship at the \nUniversity of Georgia. He has been funded continuously \nby National Institutes of Health for over 25 years. His \nresearch utilizes longitudinal data to investigate avenues \nwhereby social events and circumstances such as racial \ndiscrimination, economic hardship, neighborhood disor-\nder, and relationship conflict become biologically embed-\nded and influence speed of aging, health, and mortality.\n\n\n    Scientific Research Findings:", "answer": "\u2022 Incarceration exposure predicts accelerated biological aging among African American adults. Formerly incarcerated individuals had, on average, a faster pace of biological aging than those who had never been incarcerated.\u2022 Exposure to violence during incarceration amplifies the association between incarceration exposure and the speed of biological aging. Those who experienced violence while incarcerated were more than 2 years older than their calendar age.\u2022 The timing and frequency of incarceration spells did not moderate the effects of incarceration exposure on the speed of epigenetic aging. Nor did these effects vary by gender.", "id": 76}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/0022146519899115\nJournal of Health and Social Behavior\n2020, Vol. 61(1) 4\u00ad\u201323\n\u00a9 American Sociological Association 2020\nDOI: 10.1177/0022146519899115\njhsb.sagepub.com\nOriginal Article\nThe mental health of racial-ethnic minorities some-\ntimes presents a paradox. Despite expectations to \nthe contrary, minorities report mental health that is \neither somewhat better or no worse than that of the \nnon-Hispanic white population (Williams and Earl \n2007). The 12-month prevalence of most psychiat-\nric disorders tends to be lower among minorities \n(Breslau et\u00a0al. 2005; Miranda et\u00a0al. 2008; Vilsaint \net\u00a0al. 2019), including for depression (Hasin et\u00a0al. \n2018). Other patterns are consistent with this \n\u00adobservation. Death rates from suicide among non-\u00ad\nHispanic whites are two to three times higher than \namong blacks, Asians, and Latinos (Miranda et\u00a0al. \n2008). Once adjustments for socioeconomic status \nare included, African Americans almost always \nreport better mental health on dimensional measures \nof depression (Barnes and Bates 2017). Hispanics, \ntoo, tend to report a lower prevalence of psychiatric \ndisorders (Breslau et\u00a0al. 2005), and among Asian \npopulations, the prevalence of psychiatric disorders \nis especially low (Takeuchi et\u00a0al. 2007). To be sure, \nrelative to whites, African Americans sometimes \nreport worse mental health on cognitive measures of \n899115 HSBXXX10.1177/0022146519899115Journal of Health and Social BehaviorSchnittker and Do\nresearch-article2020\n1University of Pennsylvania, Philadelphia, PA, USA\nCorresponding Author:\nJason Schnittker, Department of Sociology, University \nof Pennsylvania, 3718 Locust Walk, Philadelphia, PA \n19104-6299, USA. \nEmail: jschnitt@ssc.upenn.edu\nPharmaceutical Side Effects \nand Mental Health Paradoxes \namong Racial-Ethnic Minorities\nJason Schnittker1 and Duy Do1\nAbstract\nSociologists have long struggled to explain the minority mental health paradox: that racial-ethnic minorities \noften report better mental health than non-Hispanic whites despite social environments that seem less \nconducive to well-being. Using data from the 2008\u20132013 Medical Expenditure Panel Survey (MEPS), this \nstudy provides a partial explanation for the paradox rooted in a very different disparity. Evidence from \nMEPS indicates that non-Hispanic whites consume more pharmaceuticals than racial-ethnic minorities for \na wide variety of medical conditions. Moreover, non-Hispanic whites consume more pharmaceuticals that \nalthough effective in treating their focal indication, include depression or suicide as a side effect. In models \nthat adjust for the use of such medications, the minority advantage in significant distress is reduced, in \nsome instances to statistical nonsignificance. Although a significant black and Hispanic advantage in a \ncontinuous measure of distress remains, the magnitude of the difference is reduced considerably. The \nrelationship between the use of medications with suicide as a side effect and significant distress is especially \nlarge, exceeding, for instance, the relationship between poverty and significant distress. For some minority \ngroups, the less frequent use of such medications is driven by better health (as in the case of Asians), \nwhereas for others, it reflects a treatment disparity (as in the case of blacks), although the consequences \nfor the mental health paradox are the same. The implications of the results are discussed, especially with \nrespect to the neglect of psychological side effects in the treatment of physical disease as well as the \nproblem of multiple morbidities.\nKeywords\ncomorbidity, disparities, pharmaceuticals, psychological distress, race-ethnicity\n\nSchnittker and Do\t\n5\nwell-being, such as life satisfaction (Hughes and \nThomas 1998), though they usually have a lower \nrisk of clinically significant distress (Breslau et\u00a0al. \n2005; Riolo et\u00a0al. 2005).\nDifferences of this sort confound easy explana-\ntions. For one, the apparent racial-ethnic advantage \nin mental health does not align with racial-ethnic \ndisparities in physical health. On most indicators, \nAfrican Americans report worse physical health \nthan non-Hispanic whites (Williams 2005). The \nHispanic population between the ages of 15 and 34 \nreports worse health, though overall the health of \nHispanics is somewhat better than whites (Williams \n2005). Even when racial-ethnic minorities maintain \na clear relative advantage in health, as with the \nAsian population (Acciai, Noah, and Firebaugh \n2015), racial-ethnic minorities are often exposed to \nmore stress, especially when considering race-spe-\ncific stressors (Williams 2018). Black and Hispanic \npopulations, for instance, report higher levels of \ndiscrimination along with lower socioeconomic sta-\ntus (Kessler, Mickelson, and Williams 1999; P\u00e9rez, \nFortuna, and Alegr\u00eda 2008). Asian populations, too, \nreport high levels of discrimination, and some \n\u00adspecific Asian populations (e.g., Hmong and \nBurmese) also report lower socioeconomic status \nthan whites (Yip, Gee, and Takeuchi 2008).\nIn light of these facts, sociologists interested in \nexplaining the racial-ethnic minority advantage \nhave focused less on stress exposure and more on \nprotective factors. These efforts, however, have not \nbeen entirely successful. Research has uncovered \nsome important protective factors, to be sure, but \nthese factors are not always relevant or suitable to \nexplaining the minority paradox. They either apply \nonly to coping with certain forms of stress, have \nhighly complex and contingent associations with \nmental health, or are not uniformly more common \namong racial-ethnic minorities (Ellison, Musick, \nand Henderson 2008; Jackson, Knight, and Rafferty \n2010; Odom, Garrett-Peters, and Vernon-Feagans \n2014; Phinney and Haas 2003). As an alternative \nexplanation for the paradox, some scholars have \nturned to measurement biases in common measures \nof mental health (Sue et\u00a0 al. 2012). Yet here, too, \nsuch explanations are limited, at least for purposes \nof explaining the paradox. Measurement differ-\nences between groups are generally not large or \nconsistent enough to explain the paradox altogether \n(Alegr\u00eda and McGuire 2003). Mixed findings of \nthis sort have prompted some to argue that the liter-\nature on minority mental health, no matter how well \ndeveloped, still has more questions than answers \n(Williams and Earl 2007).\nThis study seeks to push the literature in a dif-\nferent direction and provides a different answer for \nthe paradox. In particular, this study explores \nwhether the paradox is a function of the side effects \nof medications frequently prescribed for common \nphysical conditions. This study presents evidence \nthat much of the paradox in mental health can be \nexplained by the more frequent use of medications \nthat have either depression or suicide as a side \neffect among non-Hispanic whites. In this way, this \nstudy argues that the resolution of one paradox \nmight rest on another: The elusive protective factor \napparent among racial-ethnic minorities might be, \nin part, in their relatively low levels of medical \nintervention.\nBackground\nPatterns in racial-ethnic minority mental health are \nconfounding in large part because most sociological \ntheory points to the many ways that minorities are \ndisadvantaged (Williams 2018). For good reason, \nmost explanations for minority mental health begin \nwith socioeconomic status (Williams et\u00a0al. 1997). At \nleast in the context of physical health, socioeco-\nnomic status explains some though not all of the \nblack disadvantage, including with respect to life \nexpectancy (Geruso 2012; Hayward et\u00a0 al. 2000). \nSocioeconomic status also affects mental health, \nlargely through chronic stress (Thoits 2010), but \ngiven the relatively low socioeconomic status of \nblacks, socioeconomic status only suppresses the \nblack\u2013white difference in depression. For some \ndimensional measures of depression specifically, \nAfrican Americans report less depression even \nbefore adjustments for socioeconomic status, \nthough such adjustments further inflate the advan-\ntage (Barnes and Bates 2017). A similar pattern is \nfound for other groups and is sometimes even more \npronounced. The Mexican immigrant disadvantage \nin socioeconomic status, for instance, is quite large \neven as the 12-month prevalence of most psychiat-\nric disorders is very low (Vega et\u00a0al. 1998; Vilsaint \net\u00a0al. 2019).\nAs research on minority mental health has \nevolved, it has increasingly focused on uncovering \nother sources of stress. This focus has done much to \nforefront previously neglected race-based stressors \n(Williams 2018). Perhaps the most prominent of \nsuch stressors is discrimination, which is both com-\nmon among racial-ethnic minorities and conse-\nquential for mental health. Blacks, Hispanics, and \nAsians all report more discrimination than the white \npopulation, including major lifetime experiences \n\n6\t\nJournal of Health and Social Behavior 61(1)\nwith unfair treatment, such as losing a job, as  \nwell as everyday experiences of discrimination  \nand \u00addisrespect, such as inferior service (American \nPsychological Association 2016). Discrimination \nhas, in turn, been related to subsequent psycho\u00ad\nlogical distress and an increased risk of major \n\u00adpsychiatric disorders (Brown et\u00a0 al. 2000; Lewis, \nCogburn, and Williams 2015; Williams 2018).\nAlthough this research is revealing, it only deep-\nens the paradox of minority mental health. In effect, \nit shows that once stress is measured more compre-\nhensively, racial-ethnic minorities report even more \nstress than the white population. In this light, some \nscholars have shifted their focus from documenting \nrisks to uncovering protective factors among \nminorities. A variety of factors have been put for-\nward. Much of this work has focused on social sup-\nport and religious involvement, and both are \ndemonstrably important for coping with stress \n(Brody et\u00a0al. 2006; Ellison et\u00a0al. 2008). In a similar \nvein, though positing something more novel, Keyes \n(2009) pointed to the possibility that blacks are bet-\nter able to create positive meaning from negative \nevents. A somewhat smaller literature has taken a \ndifferent tact and attempted to identify risk factors \nthat are more common among whites. For instance, \nsome studies point to inflexible coping styles \namong non-Hispanic whites, including a belief in \nmeritocracy, which can lead to more self-blame in \nsituations of low socioeconomic status (Kwate and \nMeyer 2010). Other research has explored the ways \nin which whiteness can sometimes diminish health \n(Malat, Mayorga-Gallo, and Williams 2018). Yet \neven as research identifies these racial-ethnic dif-\nferences, it generally does not address the mental \nhealth paradox directly. Moreover, research rarely \nfinds coping resources that are uniformly more \ncommon among racial-ethnic minorities or signifi-\ncant risk factors that are plainly more frequent \namong whites. Although racial-ethnic minorities \nexperience more discrimination, for instance, they \nare not necessarily better able to cope with that dis-\ncrimination than non-Hispanic whites (Phinney and \nHaas 2003). Even among populations that expect \ndiscrimination, discrimination remains detrimental \nto their mental health (Kessler et\u00a0al. 1999). Neither \nfamily relationships nor relationships of choice can \nexplain the paradox, though both are important for \nwell-being (Mouzon 2013, 2014). Although blacks \nreport considerably more religious involvement, \nreligion, too, fails to explain the paradox (Mouzon \n2017). In this light, recent research has framed its \nfindings in a different and perhaps narrower fash-\nion, arguing only for a set of offsetting mechanisms, \nas when studies find evidence for higher self-\nesteem among blacks relative to whites but greater \nexposure to trauma (Louie and Wheaton 2019).\nExplaining the minority mental health paradox \nrequires something more: It requires thinking about \nrisk factors that both are more common in the most \nadvantaged population and have less contingent \nrelationships with mental health. Very little research \nhas focused on precisely this set of empirical issues, \nthough a different and emerging literature could \nprovide clues (Qato, Ozenberger, and Olfson 2018).\nTreatment Disparities in Health Care\nA very well-developed literature has focused on dis-\nparities in the receipt and quality of health care \n(Institute of Medicine 2003). African Americans \ntend to receive less medical care than whites, and \nwhen they do receive care, it tends to be of lower \nquality. Some of this difference is driven by access \nto care, including the availability of health insur-\nance (Sohn 2017). Yet even when adjusting for \nhealth insurance, African Americans still report \nlower quality care. Consistent with this, some evi-\ndence links treatment disparities to physicians\u2019 \nimplicit preference for white patients over patients \nof color (Sabin et\u00a0al. 2009; van Ryn et\u00a0al. 2011). The \nsame pattern is found in prescription medications, \nthe most common form of medical intervention. \nAfrican Americans receive fewer prescription medi-\ncations, and even among insured patients with the \nsame condition, black and Hispanic patients use sig-\nnificantly fewer medications that white patients \n(Briesacher, Limcangco, and Gaskin 2003). The \ninterpretation of this difference is not always consis-\ntent. In some instances, the difference reflects over-\ntreatment on the part of non-Hispanic white patients, \nas in the case of the overuse of antibiotics (Goyal \net\u00a0al. 2017). Much of the evidence, though, points to \nsignificant \nundertreatment \namong \nminorities, \nincluding with respect to the use of new essential \ndrugs (Wang et\u00a0al. 2006) and the use of analgesics in \nthe treatment of pain (Meghani, Byun, and Gallagher \n2012). Pharmaceutical innovations, like medical \ninnovations more generally, are usually given to the \nmost advantaged patients first, producing disparities \neven as the medical armamentarium improves.\nTreatment disparities of this sort have a number \nof consequences, including higher mortality among \nAfrican American patients who are not adequately \ntreated (Morris et\u00a0al. 2010). Yet some treatment dis-\nparities might have the paradoxical effect of reduc-\ning the burden of depression among minority \npopulations, who either take fewer medications \n\nSchnittker and Do\t\n7\nbecause they are prescribed fewer medications at a \ngiven level of illness (i.e., blacks and Hispanics) or \ntake fewer medications because they have better \nhealth (i.e., Asians). Prescription drug use is, in gen-\neral, very common, representing the average \nAmerican\u2019s most routine medical treatment. In 2012, \nnearly 60% of adults reported the use of at least one \nprescription medication, and 15% reported using five \nor more (Kantor et\u00a0al. 2015). Such medications are \noften highly effective in treating the condition they \nwere designed to treat or in reducing a known risk \nfactor. More directly relevant to mental health, \nthough, evidence points to a significant increase in \nthe use of medications with depression or suicide as \na significant side effect (Qato et\u00a0al. 2018). Indeed, at \npresent, about 37% of U.S. adults take at least one \nsuch medication, and 15% take three or more (Qato \net\u00a0 al. 2018). The same data show a significant \nincrease in the risk of major depression associated \nwith using such medications, especially when taking \nmore than one (Qato et\u00a0al. 2018).\nTo date, no study has attempted to link such \nmedications to racial-ethnic differences in depres-\nsion. It remains unclear whether side effects of this \nsort are relevant to understanding racial-ethnic \n\u00addifferences in depression, both because racial-\u00ad\nethnic differences in the use of such medications \nare not well understood and because few studies \nhave explored the implications of such medications \nfor population mental health. The present study \naddresses these issues by using the Medical \nExpenditure Panel Survey (MEPS), a high-quality \nnationally representative survey useful for assess-\ning medication use.\nData and Methods\nData came from the 2008\u20132013 MEPS (see Cohen, \nCohen, and Banthin 2009), including the file made \navailable at the Integrated Public Use Microsample \nrepository at the University of Minnesota (Blewett \net\u00a0al. 2018). The MEPS is a nationally representa-\ntive probability survey of the civilian noninstitu-\ntionalized U.S. population, drawn from a subsample \nof households that participated in the prior year\u2019s \nNational Health Interview Survey (NHIS). Both \nMEPS and NHIS include oversamples: Black, \nHispanic, and low-income respondents are overs-\nampled to improve the precision of estimates for \nthese subgroups. The overall MEPS response rate \n(after taking into account the nonresponse rate from \nthe NHIS) ranges from 46% to 71%. All the \u00adanalyses \npresented here employed survey weights to account \nfor oversampling and nonresponse.\nThe data are organized as annual files, though \ndata collection occurred through a series of overlap-\nping household panels. MEPS respondents enter the \nsurvey annually as members of a panel. A new panel \nis selected each year, and data are collected over five \nrounds in two calendar years. The data are then orga-\nnized into annual files, each of which will include \nmultiple panels. For each calendar year, data from \ninterview rounds one, two, and three are included for \nindividuals in their first of two years, and data from \ninterview rounds three, four, and five are included \nfor individuals in their second of two years. \nCollectively, respondents from the two overlapping \npanels in each year provide nationally representative \nestimates for that year. More information is available \nat https://meps.ahrq.gov/survey_comp/hc_data_col-\nlection.jsp. We began with 2008 because it was the \nfirst year to include comprehensive chronic disease \nindicators. Listwise deletion was used to eliminate a \nsmall number of missing cases, less than 3% for the \ndependent variable and an additional 1% across the \nset of remaining covariates.\nNonspecific Psychological Distress\nThe MEPS survey includes the Kessler K6, a well-\nvalidated instrument assessing nonspecific psycho-\nlogical distress (Kessler et\u00a0al. 2003). The K6 was \nnot fielded in every round: In each calendar year, \nrespondents in the first year of their panel were \ngiven the K6 in the second round of interviews, and \nrespondents in the final year of their panel were \ngiven the K6 in the fourth round. The K6 asks about \nthe frequency of six symptoms in the past 30 days. \nRespondents were asked: \u201chow often did you feel so \nsad that nothing could cheer you up; nervous; rest-\nless or fidgety; hopeless; that everything was an \neffort; and worthless,\u201d with response categories \nranging from none of the time (0) to all of the time \n(4). The six items were combined to create a sum-\nmary measure with scores ranging from 0 to 24. In \naddition to a continuous score, some models \nemployed a categorical measure of significant dis-\ntress, based on a threshold score of 13 or higher, as \nused in previous studies (Reeves et\u00a0al. 2011).\nMedication Use\nAmong the most valuable features of MEPS is its \ncollection of detailed medication use. MEPS \nincludes the names of all the drugs taken by the \nrespondent. Although MEPS collected information \non drugs in every round, we did not include drugs \nreported in rounds after respondents reported their \n\n8\t\nJournal of Health and Social Behavior 61(1)\nmental health to ensure proper time ordering. Drug \ninformation was then linked to information on \nwhether the specific drug taken by the respondent \ninvolved suicide or depression as a side effect (in \nseparate categories). This linkage was based on pro-\ncedures described in detail in Qato et\u00a0al. (2018) and, \nin particular, on lists provided in an online appendix \nto that article. In that study, Qato and colleagues \n(2018) linked information on medications, as \nreported in the National Health and Nutritional \nExamination Study (NHANES), to information on \nadverse effects provided in Micromedex, a proprie-\ntary database created by Truven Health Analytics, \nwhich, in turn, is based on information provided by \nthe U.S. Food and Drug Administration (Barrons \n2004; Cheng et\u00a0al. 2010). After using the informa-\ntion provided in Qato et\u00a0al. (2018) we created two \nvariables, the first corresponding to the number of \ndrugs the respondent was taking with depression as \na side effect and the second corresponding to the \nnumber of drugs the respondent was taking with sui-\ncide as a side effect. In each case, we divided the \nsum into four categories, consistent with the Qato \net\u00a0 al. (2018) analysis of the NHANES: no drugs \nwith such a side effect (the reference category), one \ndrug, two drugs, or three or more drugs. In some of \nthe descriptive information presented in our study, \nwe refer to actual counts rather than categories. In \naddition, in some analyses, we removed antidepres-\nsants from the suicide side-effect count (a sensitiv-\nity test discussed in more detail shortly). The list of \ndrugs included both prescription and over-the-\u00ad\ncounter medications, though most were prescrip-\ntion. The Qato et\u00a0al. (2018) analysis was based on \nreports of prescription medications in NHANES, \nthough some prescription medications have over-\nthe-counter formulas.\nSociodemographic Variables\nRace-ethnicity was divided into four groups: non-\nHispanic white, non-Hispanic black, Hispanic, and \nAsian. Less than 2% of the sample consisted of an \n\u201cother\u201d race-ethnicity group, and because of their \nsmall numbers, they were dropped from the analy-\nsis. The models also included controls for the sex of \nthe respondent, marital status (married; widowed, \ndivorced, or separated; never married), age (18\u2013to \n39, 40\u201364, 65\u201379, and 80 and over), education (less \nthan high school, high school or equivalent, college \nor higher), insurance status, and poverty (below \n100% of the federal poverty line for household \nincome). All the models also controlled for the year \nof the survey using a series of dummy variables.\nChronic Conditions\nMany of the medications with either depression or \nsuicide as a side effect are used in the treatment of \nchronic conditions (as opposed to prevention). To \ncontrol for physical health comorbidities, which \nthemselves could play a powerful role in depres-\nsion, the models included controls for whether the \nrespondent was ever told they had each of 11 condi-\ntions: angina pectoris, arthritis, asthma, coronary \nheart disease, diabetes, high cholesterol, emphy-\nsema, a heart attack, heart disease or other heart \ncondition, hypertension, or a stroke. These condi-\ntions were entered into the regression model as \nseparate covariates, and in some specifications, \ninteractions among all pair-wise conditions were \nexplored to assess the influence of comorbidity.\nResults\nThe results begin by describing basic differences in \nthe use of medications with depression and suicide \nas side effects, as presented in Tables 1, 2, and 3. \nTable 1 presents basic demographic information, \nTable 2 presents drug use information, and Table 3 \npresents means for the distress measures arrayed \nover race-ethnicity and drug use. Table 1 shows, \namong other things, few significant differences in \ndistress on average. There is no black-white or \nHispanic-white paradox, though in basic multivari-\nate models (see Table 3), a significant difference \nemerges. Table 2 shows large racial-ethnic differ-\nences in the use of medications with side effects. \nAbout 60% of non-Hispanic whites do not use med-\nications with depression as a side effect, but this \npercentage is much higher among minorities, espe-\ncially Asians and Hispanics. About 5% of whites \nuse three or more such medications, relative to \nslightly over 4% of blacks, 1% of Asians, and 2% of \nHispanics. The same pattern is found for drugs \ninvolving suicide as a side effect, though the racial-\nethnic differences are, if anything, larger. About \n72% of whites use no such medications, relative to \n85% of blacks, 91% of Asians, and 87% of \nHispanics. The white population uses three or more \nsuch medications at levels that are at least two-to-\none relative to racial-ethnic minorities. These differ-\nences are not driven by any particular class of \nmedication. For most specific subclasses, the same \npattern is evident, with the notable exception of \nanalgesics with depression as a side effect, where \nblacks are slightly more likely to take such medica-\ntions than non-Hispanic whites. The single largest \ncontributor to the overall count is antidepressants \nwith suicide as a side effect. These medications are \n\nSchnittker and Do\t\n9\ntaken by nearly 16% of non-Hispanic whites but \nfewer than 7% of racial-ethnic minorities.\nTable 2 also presents the prevalence of specific \nchronic conditions, once again arrayed over race-\nethnicity. In this case, the racial-ethnic differences \nare somewhat different. In general, many chronic \nconditions are more common among non-Hispanic \nwhites than minorities, though there are notable \nexceptions. Blacks report a considerably higher \nprevalence of diabetes, hypertension, stroke, and \nasthma. For many chronic conditions, Asians and \nHispanics report a considerably lower prevalence \nthan non-Hispanic whites.\nTable 3 presents means for the distress variables \narrayed by race-ethnicity and drug use. It also \nshows two patterns in particular that set the stage \nfor the multivariate models. In general, the use of \nmedications with side effects increases the preva-\nlence of significant distress and the K6 score. \nFurthermore, racial-ethnic differences in distress \nare different within the rows denoting the use of no \ndrugs with side effects. Within this group, the \nprevalence of significant distress is generally higher \namong blacks than among whites (the difference is \nstatistically significant for three of four compari-\nsons). Similarly, the prevalence of distress is higher \namong Hispanics (for three of four comparisons).\nThe regression models presented in Table 4 \nexplore the consequences of such medications for \nthe minority advantage in mental health. Pairs of \nmodels are presented for each model specification, \nthe first using linear regression for the K6 summary \nscore and the second using logit regression for the \ncategorical coding of the K6. The use of control \nvariables is critical both for understanding the mag-\nnitude of the paradox and for contextualizing the \ninfluence of pharmaceutical use. The baseline \nmodel controls for socioeconomic status (as well as \nother variables) and shows a minority advantage, \nboth with respect to a dimensional measure of dis-\ntress and a categorical one. The magnitude of the \ndifference is large, with minorities showing at least \na 20% reduction in the odds of reporting significant \ndistress. The second pair of models inserts controls \nTable 1.\u2002 Descriptive Statistics for Socioeconomic and Demographic Variables.\nRace and Ethnicity\n\u2002\nNon-Hispanic \nWhite\nNon-Hispanic \nBlack\nNon-Hispanic \nAsian\nHispanic\nSignificant distress (K6 \u2265 13) (%)\n4.95\n5.36\n3.09***\n5.35\nK6 score (continuous, minimum = 0, \nmaximum = 24)\n3.38\n3.29\n2.87***\n3.34\nFemale (%)\n51.60\n54.94***\n53.16*\n49.30***\nAge (%)\n\u2003 18\u201344\n34.13\n43.71***\n45.09***\n53.68***\n\u2003 45\u201364\n45.13\n43.67*\n41.90**\n37.48***\n\u2003 65\u201379\n15.18\n9.82***\n10.14***\n6.93***\n\u2003 80+\n5.56\n2.80***\n2.87***\n1.90***\nMarital status (%)\n\u2003 Married\n57.96\n34.19***\n62.02**\n48.82***\n\u2003 Widowed, divorced, separated\n20.27\n23.44***\n11.54***\n16.45***\n\u2003 Never married\n21.78\n42.37***\n26.44***\n34.73***\nEducational attainment (%)\n\u2003 Less than high school\n9.90\n18.38***\n13.03***\n38.30***\n\u2003 High school graduate\n29.79\n34.02***\n16.36***\n27.44***\n\u2003 College or above\n60.31\n47.60***\n70.61***\n34.26***\nLess than 100% of poverty guideline (%)\n8.16\n20.70***\n9.73*\n19.38***\nHas any type of insurance (%)\n89.28\n81.50***\n85.30***\n63.81***\nNumber of observations\n58,508\n24,899\n9,796\n34,055\nNote: All estimates are weighted using survey weights. Data from Medical Expenditure Panel Survey 2008\u20132013.\n*p < .05, **p < .01, ***p < .001 (for test of significant differences between non-Hispanic whites and other racial/ethnic \ngroups).\n\n10\t\nJournal of Health and Social Behavior 61(1)\nTable 2.\u2002 Descriptive Statistics for Medications with Depression or Suicide as a Potential Side Effect.\nRace and Ethnicity\n\u2002\nNon-Hispanic \nWhite\nNon-Hispanic \nBlack\nNon-Hispanic \nAsian\nHispanic\nTaking any drugs with depression side effects (%)\n\u2003 None\n60.14\n66.85***\n79.44***\n75.52***\n\u2003 1 drug\n25.20\n20.80***\n15.43***\n16.77***\n\u2003 2 drugs\n9.79\n8.10***\n3.84***\n5.48***\n\u2003 3+ drugs\n4.88\n4.25**\n1.29***\n2.23***\nTaking any drugs with depression side effects by subclass (%)\n\u2003 Analgesics\n14.62\n15.52*\n5.82***\n11.05***\n\u2003 Antihypertensives\n11.53\n9.79***\n6.45***\n5.46***\n\u2003 Corticosteroids\n5.50\n3.63***\n2.03***\n2.48***\n\u2003 Gastrointestinal agents\n9.20\n6.98***\n4.13***\n5.56***\n\u2003 Hormones/hormone modifiers\n9.05\n4.89***\n4.47***\n4.09***\n\u2003 Respiratory agents\n.54\n.65\n.24***\n.48\n\u2003 Others\n5.18\n4.63*\n1.94***\n2.65***\nTaking drugs with suicide side effects (%)\n\u2003 None\n72.45\n84.98***\n91.05***\n86.82***\n\u2003 1 drug\n16.38\n9.57***\n6.42***\n8.49***\n\u2003 2 drugs\n6.51\n3.36***\n1.50***\n2.65***\n\u2003 3+ drugs\n4.66\n2.09***\n1.02***\n2.04***\nTaking any drugs with suicide side effects by subclass (%)\n\u2003 Analgesics\n2.67\n2.51\n.72***\n1.83***\n\u2003 Anticonvulsants\n7.96\n5.00***\n2.61***\n3.86***\n\u2003 Anxiolytics, hypnotics, sedatives\n7.28\n2.90***\n1.87***\n3.27***\n\u2003 Antidepressants\n15.85\n6.17***\n3.34***\n5.98***\n\u2003 Gastrointestinal agents\n.38\n.43\n.22*\n.31\n\u2003 Hormones/hormone modifiers\n1.04\n.47***\n.55***\n.38***\n\u2003 Respiratory agents\n1.74\n1.18***\n.92***\n.88***\n\u2003 Others\n5.65\n3.61***\n2.19***\n3.21***\nChronic conditions (%)\n\u2003 Angina pectoris\n3.04\n2.14***\n1.43***\n1.58***\n\u2003 Arthritis/gout/lupus/fibromyalgia\n28.94\n24.27***\n12.04***\n13.88***\n\u2003 Asthma\n9.66\n11.07***\n5.42***\n6.76***\n\u2003 Coronary heart disease\n6.52\n4.35***\n2.84***\n3.62***\n\u2003 High cholesterol\n33.90\n26.25***\n27.35***\n23.80***\n\u2003 Diabetes\n8.53\n12.13***\n8.88\n9.32*\n\u2003 Emphysema\n2.82\n1.30***\n.49***\n.55***\n\u2003 Heart attack\n4.30\n3.14***\n1.56***\n2.01***\n\u2003 Heart disease (exclude coronary heart \ndisease, angina, heart attack)\n13.06\n8.62***\n4.27***\n4.53***\n\u2003 Hypertension\n34.02\n40.67***\n24.17***\n23.59***\n\u2003 Stroke\n3.85\n4.53**\n1.57***\n1.95***\nNumber of observations\n58,508\n24,899\n9,796\n34,055\nNote: All estimates are weighted using survey weights. Data from Medical Expenditure Panel Survey 2008\u20132013.\n*p < .05, **p < .01, ***p < .001 (for test of significant differences between non-Hispanic whites and other racial-ethnic \ngroups).\n\nSchnittker and Do\t\n11\nfor each of the chronic health conditions. As \nexpected, this adjustment alters the magnitude of \nthe racial-ethnic differences, though in ways spe-\ncific to each racial-ethnic group. The black-white \ndifference is suppressed somewhat by physical \nhealth, such that the black advantage would be even \nlarger\u2014about 7% larger for the K6 score\u2014were it \nnot for a black disadvantage in physical health. For \nboth Hispanics and Asians, however, the inclusion \nof controls for chronic conditions explains some of \nthe apparent mental health advantage. The Asian \nadvantage for the K6 is reduced by about 67%, \nwhereas the Hispanic advantage is reduced by 46%. \nThe difference between Hispanics and whites in \nsignificant distress, meanwhile, is reduced to statis-\ntical nonsignificance.\nThe next two models introduce controls for the \nuse of pharmaceuticals involving depression and \nsuicide as side-effects. In the first pair of models \n(Model 3), the count of medicines involving suicide \nincludes antidepressants, but in the second pair \n(Model 4), the count excludes them. The inclusion \nof the side-effect variables explains what remains \nof much of the racial-ethnic advantage. Although \nthe black-white difference in the K6 summary \nremains significant and shows a black advantage, it \nis reduced a great deal, by 60%, relative to Model 1. \nThe difference in significant distress, meanwhile, is \nreduced to statistical nonsignificance. A similar pat-\ntern is found for the Asian-white difference, and in \nfact, the Asian advantage in the K6 turns into a sig-\nnificant Asian disadvantage. Similarly, the Hispanic \nadvantage with respect to significant distress in \nModel 1 is reversed in Model 4, turning into a sig-\nnificant Hispanic disadvantage. Formal tests of \nmediation using structural equations and the contin-\nued K6 outcome provide further evidence for medi-\nation. In Model 3, the indirect effect of black was \n\u2013.359 (p < .001), and the direct effect was \u2013.214  \n(p < .001). The parallel effects for Asians and \nHispanics were, respectively, \u2013.375 (p < .001) and \n.235 (p < .001) and \u2013.222 (p < .001) and \u2013.031  \n(p = .583). In Model 4, similar patterns of signifi-\ncance among direct and indirect effects were found \nTable 3.\u2002 Descriptive Statistics for Nonspecific Psychological Distress by Race-Ethnicity and the Use of \nMedications with Depression and Suicide Side Effects.\nRace and Ethnicity\n\u2002\nNon-Hispanic \nWhite\nNon-Hispanic \nBlack\nNon-Hispanic \nAsian\nHispanic\nSignificant distress (K6 \u2265 13), % (SE)\n\u2003 No drugs with depression side effects\n3.26 (.13)\n3.96 (.20)**\n2.29 (.23)**\n3.80 (.17)** \n\u2003 1 drug with depression side effects\n5.46 (.23)\n6.05 (.33)\n4.49 (.66)\n7.77 (.59)*** \n\u2003 2 drugs with depression side effects\n9.33 (.45)\n10.03 (.73)\n10.06 (1.86)\n12.17 (1.19)* \n\u2003 3 drugs with depression side effects\n14.35 (.78)\n15.16 (1.30)\n15.00 (3.75)\n22.93 (2.60)*** \n\u2003 No drugs with suicide side effects\n2.46 (.11)\n3.55 (.17)*** \n2.20 (.22)\n3.21 (.15)*** \n\u2003 1 drug with suicide side effects\n6.52 (.27)\n10.21 (.78)*** \n9.05 (1.34)*\n12.39 (.90)*** \n\u2003 2 drugs with suicide side effects\n12.39 (.69)\n19.45 (1.62)*** 14.79 (3.75)\n25.41 (1.80)*** \n\u2003 3 drugs with suicide side effects\n27.67 (1.16)\n34.04 (2.27)** \n28.15 (6.27)\n41.10 (2.60)*** \nK6 score, mean (SE)\n\u2003 No drugs with depression side effects\n2.86 (.03)\n2.81 (.05)\n2.62 (.07)**\n2.87 (.05)\n\u2003 1 drug with depression side effects\n3.62 (.05)\n3.73 (.09)\n3.40 (.13)\n4.27 (.10)*** \n\u2003 2 drugs with depression side effects\n4.62 (.09)\n4.57 (.14)\n5.02 (.37)\n5.36 (.21)** \n\u2003 3 drugs with depression side effects\n5.99 (.14)\n6.30 (.22)\n5.68 (.65)\n7.16 (.39)** \n\u2003 No drugs with suicide side effects\n2.63 (.03)\n2.79 (.04)**\n2.61 (.06)\n2.80 (.05)**\n\u2003 1 drug with suicide side effects\n4.16 (.06)\n4.89 (.13)*** \n4.74 (.28)*\n5.52 (.16)*** \n\u2003 2 drugs with suicide side effects\n5.90 (.11)\n7.12 (.29)*** \n6.37 (.56)\n8.14 (.30)*** \n\u2003 3 drugs with suicide side effects\n8.74 (.17)\n9.97 (.32)** \n9.06 (1.00)\n10.91 (.39)*** \nNumber of observations\n58,508\n24,899\n9,796\n34,055\nNote: All statistics are weighted using survey weights. Data from Medical Expenditure Panel Survey 2008\u20132013.\n*p < .05, **p < .01, ***p < .001 denote whether the estimates are significantly different between non-Hispanic whites \nand other racial-ethnic groups.\n\n12\t\nJournal of Health and Social Behavior 61(1)\nTable 4.\u2002 Linear Regression and Logistic Regression Analysis of K6 Score and Significant Distress (K6 \u2265 13).\nModel 1\nModel 2\nModel 3\nModel 4\n\u2002\n(Basic Controls)\n(Model 1 + Chronic \nConditions)\n(Model 2 + Drugs \nwith Depression/\nSuicide Side Effects)\n(Model 2 + Drugs with \nDepression/Suicide \nSide Effects, Excluding \nAntidepressants)\n\u2002\nK6 Score\n(K6 \u2265 \n13)\nK6 Score\n(K6 \u2265 \n13)\nK6 Score\n(K6 \u2265 13)\nK6 Score\n(K6 \u2265 13)\n\u2002\nCoeff. (SE)\nOR (SE)\nCoeff. (SE)\nOR (SE)\nCoeff. (SE)\nOR (SE)\nCoeff. (SE)\nOR (SE)\nRace and ethnicity (non-Hispanic white is reference)\n\u2003 Non-Hispanic \nblack\n\u2013.534***\n(.057)\n.779***\n(.039)\n\u2013.573***\n(.057)\n.754***\n(.040)\n\u2013.214***\n(.052)\n1.019\n(.059)\n\u2013.362***\n(.053)\n.900\n(.050)\n\u2003 Non-Hispanic \nAsian\n\u2013.424***\n(.068)\n.654***\n(.056)\n\u2013.140*\n(.066)\n.822*\n(.069)\n.235***\n(.065)\n1.158\n(.099)\n.103\n(.065)\n1.020\n(.087)\n\u2003 Hispanic\n\u2013.468***\n(.066)\n.795***\n(.043)\n\u2013.254***\n(.061)\n.940\n(.050)\n\u2013.031\n(.057)\n1.139*\n(.062)\n\u2013.122*\n(.058)\n1.051\n(.056)\nWhether taking drugs with depression side effects (none is reference)\n\u2003 1 drug\n.266***\n(.033)\n1.166***\n(.050)\n.348***\n(.034)\n1.254***\n(.053)\n\u2003 2 drugs\n.543***\n(.061)\n1.358***\n(.079)\n.725***\n(.063)\n1.535***\n(.090)\n\u2003 3+ drugs\n.860***\n(.103)\n1.363***\n(.100)\n1.156***\n(.110)\n1.590***\n(.123)\nWhether taking drugs with suicide side effects (none is reference)\n\u2003 1 drug\n1.381***\n(.048)\n2.600***\n(.119)\n\u2002\n\u2003 2 drugs\n2.840***\n(.093)\n4.502***\n(.280)\n\u2002\n\u2003 3+ drugs\n5.258***\n(.141)\n10.295***\n(.735)\n\u2002\nWhether taking drugs with suicide side effects (exclude antidepressant drugs; none is reference)\n\u2003 1 drug\n1.447***\n(.055)\n2.386***\n(.105)\n\u2003 2 drugs\n2.625***\n(.114)\n3.832***\n(.267)\n\u2003 3+ drugs\n5.081***\n(.186)\n7.916***\n(.640)\nFemale (0/1)\n.473***\n(.032)\n1.237***\n(.044)\n.451***\n(.032)\n1.200***\n(.046)\n.160***\n(.028)\n.974\n(.037)\n.258***\n(.029)\n1.051\n(.040)\nAge\n\u2003 18\u201339 (reference)\n\u2003 40\u201364\n.345***\n(.045)\n1.582***\n(.081)\n\u2013.309***\n(.043)\n.932\n(.055)\n\u2013.365***\n(.039)\n.873*\n(.050)\n\u2013.351***\n(.040)\n.884*\n(.051)\n\u2003 64\u201379\n\u2013.272***\n(.062)\n.968\n(.075)\n\u20131.825***\n(.069)\n.346***\n(.032)\n\u20131.679***\n(.066)\n.380***\n(.035)\n\u20131.762***\n(.068)\n.363***\n(.034)\n\u2003 80+\n.015\n(.101)\n.974\n(.089)\n\u20131.754***\n(.102)\n.334***\n(.032)\n\u20131.435***\n(.098)\n.423***\n(.042)\n\u20131.653***\n(.102)\n.371***\n(.036)\nMarital status (married is reference)\n\u2003 Widowed, \ndivorced, \nseparated\n1.035***\n(.056)\n1.944***\n(.099)\n.812***\n(.055)\n1.692***\n(.088)\n.657***\n(.050)\n1.552***\n(.081)\n.725***\n(.053)\n1.621***\n(.085)\n\u2003 Never married\n.438***\n(.050)\n1.371***\n(.077)\n.499***\n(.048)\n1.445***\n(.077)\n.424***\n(.044)\n1.377***\n(.073)\n.439***\n(.046)\n1.393***\n(.075)\n(continued)\n\nSchnittker and Do\t\n13\nModel 1\nModel 2\nModel 3\nModel 4\n\u2002\n(Basic Controls)\n(Model 1 + Chronic \nConditions)\n(Model 2 + Drugs \nwith Depression/\nSuicide Side Effects)\n(Model 2 + Drugs with \nDepression/Suicide \nSide Effects, Excluding \nAntidepressants)\n\u2002\nK6 Score\n(K6 \u2265 \n13)\nK6 Score\n(K6 \u2265 \n13)\nK6 Score\n(K6 \u2265 13)\nK6 Score\n(K6 \u2265 13)\n\u2002\nCoeff. (SE)\nOR (SE)\nCoeff. (SE)\nOR (SE)\nCoeff. (SE)\nOR (SE)\nCoeff. (SE)\nOR (SE)\nEducational attainment (less than high school is reference)\n\u2003 High school \ngraduate\n\u2013.435***\n(.064)\n.776***\n(.036)\n\u2013.341***\n(.061)\n.826***\n(.039)\n\u2013.359***\n(.055)\n.802***\n(.038)\n\u2013.347***\n(.056)\n.816***\n(.039)\n\u2003 College or \nabove\n\u2013.943***\n(.064)\n.475***\n(.027)\n\u2013.707***\n(.060)\n.557***\n(.031)\n\u2013.726***\n(.055)\n.531***\n(.029)\n\u2013.709***\n(.058)\n.546***\n(.031)\nUnder federal \npoverty \nguideline\n1.600***\n(.068)\n2.522***\n(.107)\n1.379***\n(.063)\n2.188***\n(.095)\n1.189***\n(.058)\n1.979***\n(.085)\n1.228***\n(.059)\n2.007***\n(.088)\nHas any insurance\n\u2013.005\n(.061)\n1.022\n(.053)\n\u2013.191**\n(.058)\n.870*\n(.047)\n\u2013.415***\n(.056)\n.687***\n(.037)\n\u2013.370***\n(.055)\n.728***\n(.039)\nChronic \nconditionsa\nNo\nNo\nYes\nYes\nYes\nYes\nYes\nYes\nYear fixed effects\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nNumber of \nobservations\n127,258\n127,258\n127,258\n127,258\n127,258\n127,258\n127,258\n127,258\nNote: All analyses are weighted using survey weights. Data from Medical Expenditure Panel Survey 2008\u20132013. OR = \nodds ratio; coeff. = coefficient.\naChronic conditions include a set of binary variables that indicate whether a person has ever been told by a doctor \nor health professional that he or she has the following conditions: angina pectoris, arthritis, asthma, coronary heart \ndisease, high cholesterol, diabetes, emphysema, heart attack, heart condition/disease (excluding coronary heart \ndisease, angina, and heart attack), hypertension, and stroke.\n*p < .05, **p < .01, ***p < .001.\nTable 4.\u2002 (continued)\nwith the exception of the direct effect of Asian was \nno longer significant (b = .103, p = .115) and the \ndirect effect of Hispanic emerged as borderline sig-\nnificant (b = \u2013.122, p = .037).\nThese patterns are due in part to the particularly \npowerful relationship between drugs with suicide \nas a side effect and the risk of significant distress, \nespecially when using three or more such medica-\ntions. Taking three or more medications with sui-\ncide as a side effect results in an increase in the \nodds of significant distress by a factor of more than \n10. Even one such medication more than doubles \nthe odds of significant distress. Drugs with depres-\nsion as a side effect are consequential, too, though \nthe relationship is considerably weaker than it is for \nsuicide. To be sure, some of the relationship \nbetween drugs with suicide as a side effect and sig-\nnificant distress could be driven by the inclusion of \nantidepressants in the count. Such drugs involve \nsuicide as a side effect and therefore are relevant to \nconsidering nonspecific psychological distress, but \nsuch drugs obviously do not involve depression as a \nside effect. The next pair of models eliminates anti-\ndepressants from the count. This reduces the rela-\ntionship between drugs with suicide as a side effect \nand the risk of significant distress, though it does \nnot reduce the relationship by much. For instance, \nin the K6 models, the coefficient for three or more \nsuch drugs is reduced by only 3% between Models \n3 and 4. Moreover, pharmaceutical side effects con-\ntinue to explain the racial-ethnic minority advan-\ntage with respect to significant distress. The \nmediating effect for the black-white difference in \nthe K6 score is weaker than for Model 3, though \nside effects still explain 32% of the difference.\nTable 5 employs several sensitivity tests. The \n\u00adprimary threat to the models presented in Table 4 is \nthat the relationship between the use of pharma\u00ad\nceuticals with side effects and the K6 reflects \n\u00adunobserved \u00adheterogeneity rather than side effects \n\n14\nTable 5.\u2002 Sensitivity Analysis of K6 Score and Significant Distress (K6 \u2265 13) Regression Models.\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nModel 6\n\u2002\nAll Adults 18+\nAll Adults 18+\nAll Adults 18+\nAll Adults 18+\nYoung Adults 18\u201330\nYoung Adults 18\u201330\n\u2002\nK6 Score (K6 \u2265 13) K6 Score (K6 \u2265 13) K6 Score (K6 \u2265 13) K6 Score (K6 \u2265 13) K6 Score (K6 \u2265 13) K6 Score\n(K6 \u2265 13)\n\u2002\nCoeff. (SE) OR (SE) Coeff. (SE) OR (SE) Coeff. (SE) OR (SE) Coeff. (SE) OR (SE) Coeff. (SE) OR (SE) Coeff. (SE) OR (SE)\nRace and ethnicity (non-Hispanic white is reference)\n\u2003 Non-Hispanic black\n\u2212.364***\n(.053)\n.904\n(.050)\n\u2212.350***\n(.053)\n.903\n(.050)\n\u2212.707***\n(.094)\n.767*\n(.083)\n\u2212.463***\n(.087)\n.977\n(.106)\n\u2212.500***\n(.104)\n.811\n(.093)\n\u2212.276**\n(.104)\n.979\n(.115)\n\u2003 Non-Hispanic Asian\n.115\n(.065)\n1.042\n(.088)\n.128\n(.065)\n1.033\n(.088)\n\u2212.245\n(.129)\n.552*\n(.130)\n.018\n(.129)\n.696\n(.163)\n\u2212.395**\n(.131)\n.589*\n(.121)\n\u2212.165\n(.133)\n.699\n(.149)\n\u2003 Hispanic\n\u2212.123*\n(.058)\n1.061\n(.057)\n\u2212.103\n(.058)\n1.061\n(.057)\n\u2212.447***\n(.097)\n.684***\n(.069)\n\u2212.260**\n(.097)\n.818\n(.084)\n\u2212.514***\n(.087)\n.779*\n(.077)\n\u2212.315***\n(.086)\n.912\n(.096)\nWhether taking drugs with depression side effects (none is reference)\n\u2003 1 drug\n.339***\n(.034)\n1.223***\n(.053)\n.284***\n(.034)\n1.204***\n(.052)\n.413***\n(.077)\n1.317**\n(.111)\n.262**\n(.089)\n1.058\n(.100)\n\u2003 2 drugs\n.729***\n(.064)\n1.507***\n(.090)\n.656***\n(.065)\n1.468***\n(.090)\n.665***\n(.135)\n1.702***\n(.198)\n.433**\n(.166)\n1.513*\n(.264)\n\u2003 3+ drugs\n1.172***\n(.110)\n1.585***\n(.121)\n1.094***\n(.111)\n1.527***\n(.120)\n1.089***\n(.287)\n1.768**\n(.375)\n1.373***\n(.387)\n1.948*\n(.503)\nWhether taking drugs with suicide side effects (exclude antidepressant drugs; none is reference)\n\u2003 1 drug\n1.440***\n(.055)\n2.358***\n(.104)\n1.406***\n(.055)\n2.320***\n(.099)\n1.466***\n(.123)\n2.624***\n(.248)\n1.236***\n(.166)\n2.006***\n(.269)\n\u2003 2 drugs\n2.609***\n(.114)\n3.777***\n(.265)\n2.584***\n(.114)\n3.722***\n(.255)\n2.600***\n(.247)\n3.949***\n(.567)\n2.580***\n(.365)\n3.501***\n(.813)\n\u2003 3+ drugs\n5.071***\n(.183)\n7.796***\n(.621)\n5.043***\n(.185)\n7.684***\n(.612)\n5.959***\n(.456)\n13.792***\n(2.298)\n5.367***\n(.624)\n8.440***\n(2.158)\n(continued)\n\n15\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nModel 6\n\u2002\nAll Adults 18+\nAll Adults 18+\nAll Adults 18+\nAll Adults 18+\nYoung Adults 18\u201330\nYoung Adults 18\u201330\n\u2002\nK6 Score (K6 \u2265 13) K6 Score (K6 \u2265 13) K6 Score (K6 \u2265 13) K6 Score (K6 \u2265 13) K6 Score (K6 \u2265 13) K6 Score\n(K6 \u2265 13)\n\u2002\nCoeff. (SE) OR (SE) Coeff. (SE) OR (SE) Coeff. (SE) OR (SE) Coeff. (SE) OR (SE) Coeff. (SE) OR (SE) Coeff. (SE) OR (SE)\nWhether taking drugs without depression/suicide side effects (none is reference)\n\u2003 1 drug\n.344***\n(.041)\n1.180**\n(.067)\n.507***\n(.073)\n1.419**\n(.153)\n\u2003 2 drugs\n.322***\n(.049)\n1.266***\n(.086)\n.600***\n(.112)\n1.593***\n(.218)\n\u2003 3+ drugs\n.330***\n(.056)\n1.243**\n(.084)\n.690***\n(.147)\n1.336*\n(.195)\nChronic conditionsa\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nComorbidityb\nYes\nYes\nNo\nNo\nNo\nNo\nNo\nNo\nNo\nNo\nNo\nNo\nYear fixed effects\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nYes\nNumber of observations\n127,258\n127,258\n127,258\n127,258\n27,494\n27,494\n27,494\n27,494\n30,841\n30,841\n30,841\n30,841\nNote: All analyses are weighted using survey weights. All analyses control for gender, age, marital status, education, poverty status, and insurance status as in Table 3. Data from \nMedical Expenditure Panel Survey 2008\u20132013. OR = odds ratio; coeff. = coefficient.\naChronic conditions include a set of binary variables that indicate whether a person has ever been told by a doctor or health professional that he or she has the following conditions: \nangina pectoris, arthritis, asthma, coronary heart disease, high cholesterol, diabetes, emphysema, heart attack, heart condition/disease (excluding coronary heart disease, angina, and \nheart attack), hypertension, and stroke.\nbComorbidity includes all pairwise interactions between chronic conditions.\n*p < .05, **p < .01, ***p < .001.\nTable 5.\u2002 (continued)\n\n16\t\nJournal of Health and Social Behavior 61(1)\nspecifically, most likely with respect to poor physi-\ncal health. The models presented in Table 4 already \ncontrol for a broad set of chronic conditions, \nthough there are other possibilities that can be \naddressed empirically. The first model of Table 5 \npresents a regression model identical to Model 3 of \nTable 4 but includes all possible two-way interac-\ntions between the chronic conditions dummy vari-\nables (coefficients not shown). This model, in \neffect, tests whether patterns in comorbidity are \nimportant in a nonadditive fashion. The inclusion \nof these interactions does little to change the \nparameters for race-ethnicity or the parameters for \npharmaceutical side effects.\nThere are other ways to explore unobserved het-\nerogeneity with respect to health. It is possible that \ndrugs involving depression or suicide as a side \neffect have no more of a relationship with the K6 \nthan drugs that do not have depression or suicide as \na side effect\u2014that the entire relationship between \npharmaceutical use and the K6 is driven by under-\nlying treated illness. Other studies have explored \nthis possibility and found that drugs without depres-\nsion or suicide as a side effect are statistically unre-\nlated to depression (Qato et\u00a0al. 2018), though it is \nimportant to see if the same holds true in the present \nstudy, which uses both a different data set and a dif-\nferent outcome. Model 2 suggests other drugs have \na significant relationship with the K6, too, though \nthe magnitude of the relationship is much weaker \nthan it is for drugs involving suicide as a side effect. \nFor each pair-wise coefficient comparison (i.e., one \nmedication with suicide as a side effect is equal to \none drug with neither depression nor suicide as a \nside effect), the difference is statistically significant \nat p < .001.\nA narrower sensitivity test is to estimate the \nsame models among those with only one chronic \ncondition, thereby eliminating respondents with \ncomplex comorbidities altogether. Model 3 pres-\nents a baseline model, including all the controls \npresented in Model 1 of Table 4 but limiting the \nsample to those with a single condition. Model 4 \nthen introduces controls for medications with \ndepression and suicide as side effects. The same \npattern is evident for this subsample as it is for the \nfull sample. All the coefficients indicating a racial-\nethnic minority advantage are eliminated once con-\ntrols for medications with depression or suicide as a \nside effect are introduced.\nThe final two models provide an even more \nstringent test. Pharmaceuticals are taken for many \nreasons, including the treatment of disease but also \nits prevention and the treatment of risk rather than \ndisease itself. Younger people are, on average, \nhealthier than older people and so consume phar-\nmaceuticals under different conditions. Models 5 \nand 6 limit the sample to those between the ages of \n18 and 30. Although the magnitude of the racial-\nethnic difference in distress itself differs in this sub-\nsample, the influence of pharmaceuticals with side \neffects is very similar to that estimated in the full \nsample, greatly reducing or eliminating the minor-\nity advantage.\nDiscussion\nThis study used nationally representative data to \nexplore the consequences of pharmaceutical side \neffects for racial-ethnic differences in distress. The \nresults indicate that medications play a significant \nrole in shaping the apparent mental health advan-\ntage among racial-ethnic minorities, though there is \nsome variation between groups and between the \ncontinuous and categorical codings of the K6. The \nblack-white difference in significant distress is \nreduced to statistical nonsignificance in a model \ncontrolling for side effects. The black-white differ-\nence in the continuous distress score remains sig-\nnificant, though the difference, as estimated with \nbasic controls, is reduced by 32% in a model that \nexcludes antidepressants (Model 4) and by 60% in a \nmodel that includes them (Model 3). The advantage \nin significant distress for Asians is reduced to statis-\ntical nonsignificance, and at least in Model 3, the \nadvantage for significant distress turns into a sig-\nnificant disadvantage among Hispanics. In the case \nof Asians, the advantage for continuous distress is \nreduced to nonsignificance in Model 4. Although \ndifferences between whites and both Hispanics and \nblacks in continuous distress remain in Model 4, \npharmaceutical side effects play an important role in \nexplaining the minority mental health paradox.\nThese patterns are determined by two things. \nFirst, racial-ethnic differences in the use of phar-\nmaceuticals are large. For virtually every class of \n\u00admedication, non-Hispanic whites consume more \nmedications with side effects than do racial-ethnic \nminorities. The Asian-white difference and Hispanic-\nwhite difference is larger than the black-white differ-\nence, though all three differences are statistically \nsignificant. The fact that medications with side \neffects cut across disease categories suggests that the \nrole of medications is not limited to domains where \ntreatment disparities are well established or where \nmedications are especially consequential for health, \nincluding, for instance, the more aggressive treat-\nment of cardiovascular disease among non-Hispanic \n\nSchnittker and Do\t\n17\nwhites (e.g., Davis et\u00a0al. 2007). For this reason, the \nrole of pharmaceutical side effects in the minority \nmental health advantage is likely to be preserved \neven if some especially prominent treatment dis-\nparities are eliminated. Second, the pattern is \nshaped by the large impact of these pharmaceutical \non psychological distress, a pattern found in other \ndata sets with different measures of depression \nand, of course, a pattern anticipated based on the \nobservation of significant side effects in drug trials \nand postmarketing surveillance (Qato et\u00a0al. 2018). \nIn our study, drugs with suicide as a side effect are \nespecially consequential. A single such medication \nmore than doubles the odds of significant distress. \nRelative to other factors implicated in minority \nmental health, this relationship is large. For \ninstance, it is as large as the relationship between \npoverty and significant distress and much larger \nwhen it involves three or more such medications. \nFurthermore, the percentage of the population \nexposed to this risk is not small. About 28% of \nnon-Hispanic whites take at least one such medica-\ntion, though no more than 15% of racial-ethnic \nminorities do. To be sure, the reason minority \npatients consume fewer such pharmaceuticals dif-\nfers between racial-ethnic groups. Among African \nAmericans, the lower consumption of such phar-\nmaceuticals likely reflects a treatment disparity, \nwhereas among Asians, it likely reflects better \nhealth. For this reason, controls for comorbid med-\nical conditions play a different role between these \ngroups, suppressing the difference for the former \nand reducing it for the latter. Furthermore, the \nmediating effect of drugs with side effects is \nsmaller for the black-white difference than for \nother race-ethnic differences. Nonetheless, the fact \nthat the non-Hispanic white population consumes \nmore medications with significant side effects \nplays an important role in undermining their men-\ntal health relative to other racial-ethnic groups.\nMore research is needed on the social process \nthat produces between-group differences in the use \nof pharmaceuticals. In this regard, there are at least \ntwo distinct topics to address: first, the process that \nleads to a race-ethnic difference in prescription \ndrug use overall and second, the process that leads \nto the frequent use of medications with depression \nor suicide as a side effect. The relatively low levels \nof pharmaceutical use on the part of African \nAmericans and Hispanics likely reflects a combina-\ntion of patient and provider influences. On the one \nhand, racial-ethnic minorities might be less \ndemanding with respect to requesting prescriptions. \nSome evidence indicates that black and Hispanic \npatients express greater concern than white patients \nregarding the quality-of-life effects of pharmaceuti-\ncals, including concerns about side effects (Adams \net\u00a0al. 2018; Huang et\u00a0al. 2009). It is unclear where \nthis concern comes from, though it need not reflect \na misapprehension of the risks and benefits of phar-\nmaceuticals. Indeed, recent evidence indicates that \nnon-Hispanic white populations may be, if any-\nthing, too credulous of pharmaceuticals, rarely dis-\ncussing alternative forms of treatment (Adams et\u00a0al. \n2018) or assuming that the use of a pharmaceutical \nis the privilege and right of a properly treated \n\u201cpharmaceutical person\u201d (Ballantyne et\u00a0al. 2018). \nIn addition, evidence indicates that the relationship \nbetween physical health and mental health is gener-\nally stronger for whites, perhaps increasing the \ndemand for treatment irrespective of the taste for \npharmaceuticals (Assari and Lankarani 2016). On \nthe other hand, though, there is evidence that differ-\nences in prescription drug use emerge further \nupstream and involve the behavior of physicians. \nPhysicians may hold beliefs that lead to significant \nundertreatment among racial-ethnic minorities, \nincluding the belief that African Americans tolerate \npain better than non-Hispanic whites and therefore \nrequire fewer analgesics (Hoffman et\u00a0al. 2016) or \nthat white patients are more \u201cmedically coopera-\ntive\u201d and therefore are more likely to follow \nthrough with treatment recommendations (Oliver \net\u00a0 al. 2014). Understanding the white advantage \nwith respect to pharmaceuticals requires under-\nstanding the reasons they are prescribed medica-\ntions at higher rates as well as the reasons they \nmight persist in taking such medications even when \nthose medications present significant side effects.\nIt is also important to understand whether and \nhow physicians consider psychological side effects \nwhen treating physical disease. One possibility is \nthat physicians largely ignore the risk of depres-\nsion as a side effect, either because the medications \ninvolved have proven value in treating illnesses of \nputatively greater significance or because physi-\ncians infer that a patient\u2019s depression is caused by \ndisease. Many drugs with depression or suicide as \na side effect are effective in treating the conditions \nthey were designed to treat, to the point of being \nessential. For instance, some common antihyper-\ntensives involve depression as a side effect, though \nthey also reduce blood pressure quickly and reduce \nthe risk of cardiac events (Beers and Passman \n1990). When presented with a significant risk fac-\ntor like hypertension, a physician might ignore \npotential psychological side effects in favor of \neffectively treating the risk at hand. At the same \n\n18\t\nJournal of Health and Social Behavior 61(1)\ntime, physicians might disregard depression in \nthose patients who present with other significant \nillnesses. Oncologists\u2019 sensitivity to depression in \ntheir patients, for instance, can be as low as 6% \n(Newell et\u00a0 al. 1998). There are several explana-\ntions for why physicians overlook depression, but \none is a straightforward causal inference: that phy-\nsicians expect depression in cases of severe physi-\ncal illness and that any depression will resolve if \nand when the illness that occasioned it is treated. \nWhatever the explanation, it is almost certainly the \ncase that physicians either disregard the psycho-\nlogical side effects of medications or regard them \nas less clinically significant.\nIn general, the results of this study speak to the \nproblem of multiple morbidity, hinting at some of its \ncomplexities for understanding population health. \nThe use of multiple pharmaceuticals has increased \nover time, in tandem with the increasing intensity of \nmedical intervention (Aronowitz 2012; Swinglehurst \nand Fudge 2017), a pattern some refer to as \u201cpharma-\nceuticalization\u201d in reference to the parallel concept \nof medicalization (Abraham 2010). Polypharmacy \nitself has been linked to adverse events and nonad-\nherence, and there has been some discussion on how \nto assist physicians in managing the practice (Grando \net\u00a0al. 2012). Yet little is known about multiple mor-\nbidity per se or the ways in which the treatment of \none disease has consequences for another, in large \npart because medical science continues to be orga-\nnized around discrete disease categories. It is \nunlikely that medical treatment always involves \ncomplementarities between diseases. Some evidence \npoints to a rise, for instance, in emergency room vis-\nits for adverse effects of medical treatment (Bernstein \net\u00a0al. 2003). But a lack of complementarities is per-\nhaps especially likely when mental illness has con-\nsistently played a secondary role to physical illness. \nThe scope of the side-effects issue is broad, and \nfuture research should consider the role of pharma-\nceutical side effects in other psychiatric disorders, \nincluding side effects related to anxiety, confusion, \nand psychotic syndromes.\nAs implied in the introduction, the results do not \nspeak directly to any theory in medical sociology \nfor understanding the mental health paradox given \nhow the literature has mostly centered on adum-\nbrating risks among the disadvantaged, though the \nresults are at least thematically consistent with cer-\ntain facets of the literature. Research on fundamen-\ntal causes, for instance, has occasionally uncovered \ncountervailing mechanisms, providing some degree \nof advantage for an otherwise disadvantaged popu-\nlation (Lutfey and Freese 2005). Mechanisms of \nsuch sort are not unanticipated when disease is \ncaused by a complex social process involving a  \nmix of intentional and unintentional influences. \nSubsequent refinements of the fundamental cause \ntheory have explicitly called for a deeper consider-\nation of such countervailing mechanisms and has \ncast such mechanisms either in terms of well-\nresourced people pursuing multiple goals, some of \nwhich occasionally cut against health, or in terms of \nimperfect knowledge regarding how to pursue bet-\nter health (Phelan, Link, and Tehranifar 2010). Yet \nPhelan and colleagues (2010) cautioned scholars to \navoid using the concept of countervailing mecha-\nnisms to provide only a post hoc justification for \nresults that are inconsistent with fundamental cause \ntheory. They further suggested that some behaviors \ncan be well accommodated in the theory, as when \nthe pursuit of status undermines health. The results \nof the current study provide a different way for-\nward. They suggested that countervailing mecha-\nnisms can appear even in the context of pursuing \nbetter health, in this case, poor mental health \nemerging in pursuit of optimal physical health. The \nresults also suggest that countervailing mechanisms \ncan be born of something other than imperfect \nknowledge or asymmetric goals. In the present \nstudy, the lack of consideration of side effects likely \nreflects a structured neglect of certain side effects \non the part of both patients and providers, rather \nthan simple ignorance or lack of awareness.\nMore generally, the results speak to the potential \nadverse consequences of exceptionally high levels \nof confidence in medicine. Americans maintain \nenormous confidence in their personal physicians \nand in the medical treatment of disease. As medical \nauthority continues to expand over a wider set of \nconditions and the pharmaceutical armamentarium \ncontinues to grow (Clarke et\u00a0al. 2003), the apparent \nbenefits of pharmaceuticals might loom so large as \nto overwhelm serious consideration of their numer-\nous side effects. To be sure, the neglect of side \neffects might be especially prominent with respect \nto mental health given that patients and physicians \nmight be willing to trade some measure of emo-\ntional well-being for an improvement in disease, \nphysical functioning, or longevity. But the tendency \nto neglect side effects is more general, and other \nside effects are almost certainly overlooked as well. \nAdapting theories regarding the minority mental \nhealth advantage only requires a more complete \naccounting of risks and resources, arrayed over a \nsomewhat more complex ledger. It also requires \nintegrating sociology\u2019s understanding of the causes \nof mental health, informed by fundamental cause \n\nSchnittker and Do\t\n19\ntheory, with its understanding of how patients inter-\nface with medicine.\nLimitations\nThis study has several limitations on both the inde-\npendent and dependent variable sides. Although the \nmedication data provided in MEPS is among the \nbest among nationally representative surveys for \nevaluating health care, there are several reasons to \nsuspect the effect of medications with depression \nand suicide as side effects, as estimated here, is \nunderstated. The analyses are premised on drugs \nthat formally list depression or suicide as a side \neffect, though there are likely other medications that \ninvolve such side effects. It is not uncommon, for \ninstance, for new side effects to become apparent \nonly in Phase IV of drug development, during post-\nmarketing surveillance, when drugs are effectively \nevaluated in a naturalistic setting (Suvarna 2010). \nAt the same time, respondents may neglect to report \nall the medications they are taking, especially over-\nthe-counter medications taken on an occasional \nbasis. Most medications with significant side effects \nare prescription medications, though a number of \nover-the-counter medications also involve depres-\nsion as a side effect, including proton-pump inhibi-\ntors, used to reduce stomach acid production \n(Laudisio et\u00a0al. 2018). It is unlikely, however, that a \nperfectly comprehensive assessment of medication \nuse would substantially change the findings with \nrespect to race-ethnicity and mental health. Racial-\nethnic differences in the use over-the-counter medi-\ncations are similar to those in prescription \nmedications, suggesting a general reluctance to use \ndrugs and not a specific reluctance to use what a \ndoctor prescribes (Fillenbaum et\u00a0al. 1993).\nThe dependent variable used in this study is a \nwell-validated measure of psychological distress, \nand empirically, the models produce the pattern \nthan research has tried to explain. Yet much of the \nresearch has focused on major depression, for \nwhich the K6 cannot provide a direct analogue. \nFurthermore, some reviews of the racial-ethnic dif-\nferences in major depression and distress find that \nthe black-white difference in the K6 is closer to par-\nity than a black advantage (Barnes and Bates 2017). \nFurthermore, the impact of pharmaceutical side \neffects among blacks differs between the categori-\ncal and dimensional coding of the K6. Other depen-\ndent variables are likely to yield similar patterns to \nthose found here, though the magnitude of the \nminority advantage as well as the mediating effects \nof pharmaceutical side effects are likely to differ.\nAlthough we performed numerous sensitivity \nanalyses with respect to the relationship between \ndrug use and depression, there remains the possibil-\nity of unobserved confounding, perhaps premised on \nprodromal rather than manifest disease and therefore \nbased on diseases not reported by the respondent. \nThe relationship between physical disease and \ndepression is not limited to the most significant ill-\nnesses, the kind that are usually measured in omni-\nbus health surveys. The risk of suicide, for instance, \nincreases among those who have been hospitalized \nfor any illness (Qin et\u00a0al. 2013). At least some of the \nrelationship between the use of pharmaceuticals and \ndepression might be due to unmeasured morbidity, \nespecially among those taking multiple medications. \nBut given the breadth of controls for chronic disease \nemployed in this study as well as specifications that \nspecifically focused on comorbidity, it is unclear \nwhat the relevant set of other morbidities might be. \nFurthermore, the evidence linking pharmaceuticals \nto depression and suicide is strong. A significant \nupward bias in the coefficients seems unlikely.\nConclusion\nThe racial-ethnic minority advantage in mental health \nhas long puzzled social scientists. Although there is \nmuch to be gained from thinking about protective fac-\ntors among minorities, the present study suggests it is \nstill worth considering risk factors that are more com-\nmon among non-Hispanic whites. Such risk factors \nare not easy to identify, given social scientists\u2019 interest \nin uncovering stressors among racial-ethnic minori-\nties. And this interest is theoretically well founded\u2014\nthe idea of fundamental causes suggests multiple \noverlapping risk factors flowing from one fundamen-\ntal cause such that any one risk can be replaced by \nanother and still preserve the effects of the cause (Link \nand Phelan 1995). To explain paradoxes, though, a \ndifferent perspective is required. Identifying relevant \nrisk factors among non-Hispanic whites might require \nconsidering the shadow side of otherwise desirable \ntherapeutic interventions. And explaining disparities \nin one domain of health might require considering dis-\nparities in a seemingly very different one.\nReferences\nAbraham, John. 2010. \u201cPharmaceuticalization of Society \nin Context: Theoretical, Empirical and Health Dimen\u00ad\nsions.\u201d Sociology 44(4):603\u201322.\nAcciai, Francesco, Aggie J. Noah, and Glenn Firebaugh. \n2015. \u201cPinpointing the Sources of the Asian Mortality \nAdvantage in the USA.\u201d Journal of Epidemiology \nand Community Health 69(10):1006\u201311.\n\n20\t\nJournal of Health and Social Behavior 61(1)\nAdams, Crystal, Anwesa Chatterjee, Brittany M. Harder, \nand Liza Hayes Mathias. 2018. \u201cBeyond Unequal \nAccess: Acculturation, Race, and Resistance to \nPharmaceuticalization in the United States.\u201d SSM - \nPopulation Health 4:350\u201357.\nAlegr\u00eda, Margarita, and Thomas McGuire. 2003. \n\u201cRethinking a Universal Framework in the Psychiatric \nSymptom\u2013Disorder Relationship.\u201d Journal of Health \nand Social Behavior 44(3):257\u201374.\nAmerican Psychological Association. 2016. Stress in \nAmerica: The Impact of Discrimination. Washington, \nDC: American Psychological Association.\nAronowitz, Robert A. 2012. \u201cThe Rise and Fall of the \nLyme Disease Vaccines: A Cautionary Tale for Risk \nInterventions in American Medicine and Public \nHealth.\u201d The Milbank Quarterly 90(2):250\u201377.\nAssari, Shervin, and Maryam Moghani Lankarani. 2016. \n\u201cChronic Medical Conditions and Negative Affect: \nRacial Variation in Reciprocal Associations over \nTime.\u201d Frontiers in Psychiatry 7:140. doi:10.3389/\nfpsyt.2016.00140\nBallantyne, Peri J., Pauline Norris, Venkata Praveen \nParachuru, and W. Murray Thomson. 2018. \n\u201cBecoming a \u2018Pharmaceutical Person\u2019: Medication \nUse Trajectories from Age 26 to 38 in a Representative \nBirth Cohort from Dunedin, New Zealand.\u201d SSM - \nPopulation Health 4:37\u201344.\nBarnes, David M., and Lisa M. Bates. 2017. \u201cDo Racial Patterns \nin Psychological Distress Shed Light on the Black\u2013White \nDepression Paradox? A Systematic Review.\u201d Social \nPsychiatry and Psychiatric Epidemiology 52(8):913\u201328.\nBarrons, Robert. 2004. \u201cEvaluation of Personal Digital \nAssistant Software for Drug Interactions.\u201d American \nJournal of Health-System Pharmacy 61(4):380\u201385.\nBeers, Mark H., and Leigh J. Passman. 1990. \u201cAnti\u00ad\nhypertensive Medications and Depression.\u201d Drugs \n40(6):792\u201399.\nBernstein, Amy, E. Hing, A. J. Moss, A. B. Siller, and \nR. B. Tiggle. 2003. Health Care in America: Trends \nin Utilization. Hyattsville, MD: National Center for \nHealth Statistics.\nBlewett, Lynn A., Julia A. Rivera Drew, Risa Griffin, Kari \nC. W. Williams, and Daniel Backman. 2018. IPUMS \nHealth Surveys: Medical Expenditure Panel Survey, \nVersion 1.0 [Dataset]. Minneapolis: University of \nMinnesota. doi:http://doi.org/10.18128/D071.V1.0.\nBreslau, Joshua, Kenneth S. Kendler, Maxwell Su, \nSergio Gaxiola-Aguilar, and Ronald C. Kessler. \n2005. \u201cLifetime Risk and Persistence of Psychiatric \nDisorders across Ethnic Groups in the United States.\u201d \nPsychological Medicine 35(3):317\u201327.\nBriesacher, Becky, Rhona Limcangco, and Darrell \nGaskin. 2003. \u201cRacial and Ethnic Disparities in \nPrescription Coverage and Medication Use.\u201d Health \nCare Financing Review 25(2):63\u201376.\nBrody, Gene H., Yi-Fu Chen, Velma McBride Murry, \nXiaojia Ge, Ronald L. Simons, Frederick X. Gibbons, \nMeg Gerrard, and Carolyn E. Cutrona. 2006. \n\u201cPerceived Discrimination and the Adjustment of \nAfrican American Youths: A Five-Year Longitudinal \nAnalysis with Contextual Moderation Effects.\u201d Child \nDevelopment 77(5):1170\u201389.\nBrown, Alan S., Jim van Os, Corine Driessens, Hans W. \nHoek, and Ezra S. Susser. 2000. \u201cFurther Evidence \nof Relation between Prenatal Famine and Major \nAffective Disorder.\u201d American Journal of Psychiatry \n157(2):190\u201395.\nCheng, Christine M., B. Joseph Guglielmo, Judy Maselli, \nand Andrew D. Auerbach. 2010. \u201cCoverage of FDA \nMedication Boxed Warnings in Commonly Used \nDrug Information Resources.\u201d Archives of Internal \nMedicine 170(9):831\u201333.\nClarke, Adele E., Janet K. Shim, Laura Mamo, Jennifer \nRuth Fosket, and Jennifer R. Fishman. 2003. \n\u201cBiomedicalization: Technoscientific Trans\u00adformations \nof Health, Illness, and U.S. Biomedicine.\u201d American \nSociological Review 68(2):161\u201394.\nCohen, Joel W., Steven B. Cohen, and Jessica S. Banthin. \n2009. \u201cThe Medical Expenditure Panel Survey: A \nNational Information Resource to Support Healthcare \nCost Research and Inform Policy and Practice.\u201d \nMedical Care 47(7):S44\u201350.\nDavis, Andrew M., Lisa M. Vinci, Tochi M. Okwuosa, \nAyana R. Chase, and Elbert S. Huang. 2007. \n\u201cCardiovascular Health Disparities.\u201d Medical Care \nResearch and Review 64(5 Suppl.):29S\u2013100S.\nEllison, Christopher G., Marc A. Musick, and Andrea \nK. Henderson. 2008. \u201cBalm in Gilead: Racism, \nReligious Involvement, and Psychological Distress \namong African-American Adults.\u201d Journal for the \nScientific Study of Religion 47(2):291\u2013309.\nFillenbaum, Gerda G., Joseph T. Hanlon, Elizabeth H. \nCorder, Thandi Ziqubu-Page, William E. Wall, \nJr., and Dwight Brock. 1993. \u201cPrescription and \nNonprescription Drug Use among Black and White \nCommunity-Residing Elderly.\u201d American Journal of \nPublic Health 83(11):1577\u201382.\nGeruso, Michael. 2012. \u201cBlack-White Disparities in \nLife Expectancy: How Much Can the Standard SES \nVariables Explain?\u201d Demography 49(2):553\u201374.\nGoyal, Monika K., Tiffani J. Johnson, James M. \nChamberlain, T.Charles Casper, Timothy Simmons, \nEvaline A. Alessandrini, Lalit Bajaj, Robert W. \nGrundmeier, Jeffrey S. Gerber, Scott A. Lorch, \nand Elizabeth R. Alpern. 2017. \u201cRacial and \nEthnic Differences in Antibiotic Use for Viral \nIllness in Emergency Departments.\u201d Pediatrics \n140:e20170203. doi: 10.1542/peds.2017-0203\nGrando, Adela, Susan Farrish, Cynthia Boyd, and Aziz \nBoxwala. 2012. \u201cOntological Approach for Safe and \nEffective Polypharmacy Prescription.\u201d AMIA Annual \nSymposium Proceedings 2012:291\u2013300.\nHasin, Deborah S., Aaron L. Sarvet, Jacquelyn L. \nMeyers, Tulshi D. Saha, W. June Ruan, Malka Stohl, \nand Bridget F. Grant. 2018. \u201cEpidemiology of Adult \nDSM-5 Major Depressive Disorder and Its Specifiers \n\nSchnittker and Do\t\n21\nin the United States.\u201d JAMA Psychiatry 75(4):336\u2013\n46.\nHayward, Mark D., Toni P. Miles, Eileen M. Crimmins, \nand Yu Yang. 2000. \u201cThe Significance of \nSocioeconomic Status in Explaining the Racial \nGap in Chronic Health Conditions.\u201d American \nSociological Review 65(6):910\u201330.\nHoffman, Kelly M., Sophie Trawalter, Jordan R. Axt, \nand M. Norman Oliver. 2016. \u201cRacial Bias in Pain \nAssessment and Treatment Recommendations, and \nFalse Beliefs about Biological Differences between \nBlacks and Whites.\u201d Proceedings of the National \nAcademy of Sciences 113(16):4296\u2013301.\nHuang, Elbert S., Sydney E. S. Brown, Nidhi Thakur, \nLisabeth Carlisle, Edward Foley, Bernard Ewigman, \nand David O. Meltzer. 2009. \u201cRacial/Ethnic \nDifferences in Concerns about Current and Future \nMedications among Patients with Type 2 Diabetes.\u201d \nDiabetes Care 32(2):311\u201316.\nHughes, Michael, and Melvin E. Thomas. 1998. \u201cThe \nContinuing Significance of Race Revisited: A Study \nof Race, Class, and Quality of Life in America, 1972 \nto 1996.\u201d American Sociological Review 63(6):785\u2013\n95.\nInstitute of Medicine. 2003. Unequal Treatment: \nConfronting Racial and Ethnic Disparities in Health \nCare, edited by B. D. Smedley, A. Y. Stith, and A. R. \nNelson. Washington, DC: The National Academies \nPress.\nJackson, James S., Katherine M. Knight, and Jane A. \nRafferty. 2010. \u201cRace and Unhealthy Behaviors: \nChronic Stress, the HPA Axis, and Physical and \nMental Health Disparities over the Life Course.\u201d \nAmerican Journal of Public Health 100(5):933\u201339.\nKantor, Elizabeth D., Colin D. Rehm, Jennifer S. Haas, \nAndrew T. Chan, and Edward L. Giovannucci. \n2015. \u201cTrends in Prescription Drug Use among \nAdults in the United States from 1999\u20132012.\u201d JAMA \n314(17):1818\u201330.\nKessler, Ronald C., Peggy R. Barker, Lisa J. Colpe, \nJoan F. Epstein, Joseph C. Gfroerer, Eva Hiripi, \nMary J. Howes, Sharon-Lise T. Normand, Ronald \nW. Manderscheid, Ellen E. Walters, and Alan M. \nZaslavsky. 2003. \u201cScreening for Serious Mental \nIllness in the General Population.\u201d Archives of \nGeneral Psychiatry 60(2):184\u201389.\nKessler, Ronald C., Kristin D. Mickelson, and David R. \nWilliams. 1999. \u201cThe Prevalence, Distribution, and \nMental Health Correlates of Perceived Discrimination \nin the United States.\u201d Journal of Health and Social \nBehavior 40(3):208\u201330.\nKeyes, Corey L. M. 2009. \u201cThe Black\u2013White Paradox in \nHealth: Flourishing in the Face of Social Inequality \nand Discrimination.\u201d Journal of Personality 77(6): \n1677\u2013706.\nKwate, Naa Oyo A., and Ilan H. Meyer. 2010. \u201cThe \nMyth of Meritocracy and African American Health.\u201d \nAmerican Journal of Public Health 100(10):1831\u201334.\nLaudisio, Alice, Raffaele Antonelli Incalzi, Antonella \nGemma, Silvia Giovannini, Maria Rita Lo Monaco, \nDavide L. Vetrano, Luca Padua, Roberto Bernabei, \nand Giuseppe Zuccal\u00e0. 2018. \u201cUse of Proton-Pump \nInhibitors Is Associated with Depression: A Population-\nBased Study.\u201d International Psychogeriatrics 30(1): \n153\u201359.\nLewis, Ten\u00e9 T., Courtney D. Cogburn, and David R. \nWilliams. 2015. \u201cSelf-Reported Experiences of \nDiscrimination and Health: Scientific Advances, \nOngoing Controversies, and Emerging Issues.\u201d \nAnnual Review of Clinical Psychology 11(1):407\u201340.\nLink, Bruce G., and Jo Phelan. 1995. \u201cSocial Conditions \nas Fundamental Causes of Disease.\u201d Journal of \nHealth and Social Behavior 35(Special Issue):80\u201394.\nLouie, Patricia, and Blair Wheaton. 2019. \u201cThe \nBlack\u2013White Paradox Revisited: Understanding \nthe Role of Counterbalancing Mechanisms During \nAdolescence.\u201d Journal of Health and Social Behavior  \n60(2):169\u201387.\nLutfey, Karen, and Jeremy Freese. 2005. \u201cToward \nSome Fundamentals of Fundamental Causality: \nSocioeconomic Status and Health in the Routine \nClinic Visit for Diabetes.\u201d American Journal of \nSociology 110(5):1326\u201372.\nMalat, Jennifer, Sarah Mayorga-Gallo, and David R. \nWilliams. 2018. \u201cThe Effects of Whiteness on the \nHealth of Whites in the USA.\u201d Social Science & \nMedicine 199:148\u201356.\nMeghani, Salimah H., Eeeseung Byun, and Rollin M. \nGallagher. 2012. \u201cTime to Take Stock: A Meta-\nAnalysis and Systematic Review of Analgesic \nTreatment Disparities for Pain in the United States.\u201d \nPain Medicine 13(2):150\u201374.\nMiranda, Jeanne, Thomas G. McGuire, David R. \nWilliams, and Philip Wang. 2008. \u201cMental Health in \nthe Context of Health Disparities.\u201d American Journal \nof Psychiatry 165(9):1102\u20131108.\nMorris, Arden M., Kim F. Rhoads, Steven C. Stain, and \nJohn D. Birkmeyer. 2010. \u201cUnderstanding Racial \nDisparities in Cancer Treatment and Outcomes.\u201d \nJournal of the American College of Surgeons 211(1): \n105\u201313.\nMouzon, Dawne M. 2013. \u201cCan Family Relationships \nExplain the Race Paradox in Mental Health?\u201d \nJournal of Marriage and Family 75(2):470\u201385.\nMouzon, Dawne M. 2014. \u201cRelationships of Choice: Can \nFriendships or Fictive Kinships Explain the Race \nParadox in Mental Health?\u201d Social Science Research \n44:32\u201343.\nMouzon, Dawne M. 2017. \u201cReligious Involvement and \nthe Black\u2013White Paradox in Mental Health.\u201d Race \nand Social Problems 9(1):63\u201378.\nNewell, Sallie, Robert W. Sanson-Fisher, Afaf Girgis, and \nAnthony Bonaventura. 1998. \u201cHow Well Do Medical \nOncologists\u2019 Perceptions Reflect Their Patients\u2019 \nReported Physical and Psychosocial Problems?\u201d Cancer \n83(8):1640\u201351.\n\n22\t\nJournal of Health and Social Behavior 61(1)\nOdom, Erika C., Patricia Garrett-Peters, and Lynne \nVernon-Feagans. 2014. \u201cRacial Discrimination as \na Correlate of African American Mothers\u2019 Emotion \nTalk to Young Children.\u201d Journal of Family Issues \n37(7):970\u201396.\nOliver, M. Norman, Kristen M. Wells, Jennifer A. Joy-\nGaba, Carlee Beth Hawkins, and Brian A. Nosek. \n2014. \u201cDo Physicians\u2019 Implicit Views of African \nAmericans Affect Clinical Decision Making?\u201d The \nJournal of the American Board of Family Medicine \n27(2):177\u201388.\nP\u00e9rez, Debra Joy, Lisa Fortuna, and Margarita Alegr\u00eda. \n2008. \u201cPrevalence and Correlates of Everyday \nDiscrimination among U.S. Latinos.\u201d Journal of \nCommunity Psychology 36(4):421\u201333.\nPhelan, Jo C., Bruce G. Link, and Parisa Tehranifar. \n2010. \u201cSocial Conditions as Fundamental Causes of \nHealth Inequalities: Theory, Evidence, and Policy \nImplications.\u201d Journal of Health and Social Behavior \n51(Supplement):S28\u201340.\nPhinney, Jean S., and Kumiko Haas. 2003. \u201cThe Process \nof Coping among Ethnic Minority First-Generation \nCollege Freshmen: A Narrative Approach.\u201d The \nJournal of Social Psychology 143(6):707\u201326.\nQato, Dima Mazen, Katharine Ozenberger, and \nMark Olfson. 2018. \u201cPrevalence of Prescription \nMedications with Depression as a Potential Adverse \nEffect among Adults in the United States: Prevalence \nof Use of Prescription Medication with Depression \nas a Potential Adverse Effect.\u201d JAMA 319(22): \n2289\u201398.\nQin, Ping, Roger Webb, Nav Kapur, and Henrik Toft \nS\u00f8rensen. 2013. \u201cHospitalization for Physical \nIllness and Risk of Subsequent Suicide: A \nPopulation Study.\u201d Journal of Internal Medicine \n273(1):48\u201358.\nReeves, William C., Tara W. Strine, Laura A. Pratt, \nWilliam Thompson, Indu B. Ahluwalia, Satvinder \nS. Dhingra, Lela R. McKnight-Eily, Leslie Harrison, \nDenise V. D\u2019Angelo, Letitia Williams, Brian \nMorrow, Deborah Gould, and Marc A. Safran. 2011. \n\u201cMental Illness Surveillance among Adults in the \nUnited States.\u201d Morbidity and Mortality Weekly \nReport 60(3):1\u201329.\nRiolo, Stephanie A., Tuan Anh Nguyen, John F. Greden, \nand Cheryl A. King. 2005. \u201cPrevalence of Depression \nby Race/Ethnicity: Findings from the National Health \nand Nutrition Examination Survey III.\u201d American \nJournal of Public Health 95(6):998\u20131000.\nSabin, Janice, Brian A. Nosek, Anthony Greenwald, and \nFrederick P. Rivara. 2009. \u201cPhysicians\u2019 Implicit \nand Explicit Attitudes about Race by MD Race, \nEthnicity, and Gender.\u201d Journal of Health Care for \nthe Poor and Underserved 20(3):896\u2013913.\nSohn, Heeju. 2017. \u201cRacial and Ethnic Disparities in \nHealth Insurance Coverage: Dynamics of Gaining \nand Losing Coverage over the Life-Course.\u201d \nPopulation Research and Policy Review 36(2): \n181\u2013201.\nSue, Stanley, Janice Ka Yan Cheng, Carmel S. Saad, \nand Joyce P. Chu. 2012. \u201cAsian American Mental \nHealth: A Call to Action.\u201d American Psychologist \n67(7):532\u201344.\nSuvarna, Viraj. 2010. \u201cPhase IV of Drug Development.\u201d \nPerspectives in Clinical Research 1(2):57\u201360.\nSwinglehurst, Deborah, and Nina Fudge. 2017. \u201cThe \nPolypharmacy Challenge: Time for a New Script?\u201d \nThe British Journal of General Practice: The Journal \nof the Royal College of General Practitioners \n67(662):388\u201389.\nTakeuchi, David T., Nolan Zane, Seunghye Hong, \nDavid H. Chae, Fang Gong, Gilbert C. Gee, Emily \nWalton, Stanley Sue, and Margarita Alegria. 2007. \n\u201cImmigration-Related Factors and Mental Disorders \namong Asian Americans.\u201d American Journal of \nPublic Health 97(1):84\u201390.\nThoits, Peggy A. 2010. \u201cStress and Health: Major \nFindings and Policy Implications.\u201d Journal of Health \nand Social Behavior 51(Supplement):S41\u201353.\nvan Ryn, Michelle, Diana J. Burgess, John F. Dovidio, \nSean M. Phelan, Somnath Saha, Jennifer Malat, Joan \nM. Griffin, Steven S. Fu, and Sylvia Perry. 2011. \n\u201cThe Impact of Racism on Clinician Cognition, \nBehavior, and Clinical Decision Making.\u201d Du Bois \nReview: Social Science Research on Race 8(1):199\u2013\n218.\nVega, William A., Bohdan Kolody, Sergio Aguilar-\nGaxiola, Ethel Alderete, Ralph Catalano, and Jorge \nCaraveo-Anduaga. 1998. \u201cLifetime Prevalence of \nDSM-III-R Psychiatric Disorders among Urban and \nRural Mexican Americans in California.\u201d Archives of \nGeneral Psychiatry 55(9):771\u201378.\nVilsaint, Corrie L., Amanda NeMoyer, Mirko Fillbrunn, \nEkaterina Sadikova, Ronald C. Kessler, Nancy A. \nSampson, Kiara Alvarez, Jennifer Greif Green, Katie \nA. McLaughlin, Ruijia Chen, David R. Williams, \nJames S. Jackson, and Margarita Alegr\u00eda. 2019. \n\u201cRacial/Ethnic Differences in 12-Month Prevalence \nand Persistence of Mood, Anxiety, and Substance Use \nDisorders: Variation by Nativity and Socioeconomic \nStatus.\u201d Comprehensive Psychiatry 89:52\u201360.\nWang, Junling, Jason M. Noel, Ilene H. Zuckerman, \nNancy A. Miller, Fadia T. Shaya, and C. Daniel \nMullins. 2006. \u201cDisparities in Access to Essential \nNew Prescription Drugs between Non-Hispanic \nWhites, Non-Hispanic Blacks, and Hispanic Whites.\u201d \nMedical Care Research and Review 63(6):742\u201363.\nWilliams, David R. 2005. \u201cThe Health of U.S. Racial and \nEthnic Populations.\u201d Journal of Gerontology: Series \nB 60B(Special Issue II):53\u201362.\nWilliams, David R. 2018. \u201cStress and the Mental Health of \nPopulations of Color: Advancing Our Understanding \nof Race-Related Stressors.\u201d Journal of Health and \nSocial Behavior 59(4):466\u201385.\nWilliams, David R., and Tara R. Earl. 2007. \u201cCommentary: \nRace and Mental Health\u2014More Questions Than \nAnswers.\u201d International Journal of Epidemiology \n36(4):758\u201360.\n\nSchnittker and Do\t\n23\nWilliams, David R., Yu Yan, James S. Jackson, and \nNorman B. Anderson. 1997. \u201cRacial Differences \nin Physical and Mental Health: Socio-economic \nStatus, Stress and Discrimination.\u201d Journal of Health \nPsychology 2(3):335\u201351.\nYip, Tiffany, Gilbert C. Gee, and David T. Takeuchi. \n2008. \u201cRacial Discrimination and Psychological \nDistress: The Impact of Ethnic Identity and Age \namong Immigrant and United States-Born Asian \nAdults.\u201d Developmental Psychology 44(3):787\u2013800.\nAuthor Biographies\nJason Schnittker is professor of sociology at the \nUniversity of Pennsylvania. His research focuses on the \nsocial, cultural, and biological determinants of health, with \na particular focus on mental health.\nDuy Do is a graduate student in demography at the \nUniversity of Pennsylvania. He is interested in health pol-\nicy, health economics, and the social and medical determi-\nnants of health.\n\n\n    Scientific Research Findings:", "answer": "\u2022 We found that non-Hispanic whites take more pharmaceuticals than minorities, and furthermore, they take more pharmaceuticals with depression as a side effect across a wide variety of specific drug indications and classes.\n\u2022 When adjusting for the use of such medications, the minority paradox is reduced. ", "id": 77}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/0022146520924810\nJournal of Health and Social Behavior\n2020, Vol. 61(2) 134\u00ad\u2013152\n\u00a9 American Sociological Association 2020\nDOI: 10.1177/0022146520924810\njhsb.sagepub.com\nOriginal Article\nChildhood diagnoses of attention-deficit/hyperactiv-\nity disorder (ADHD) have increased 41% in the \nUnited States over the past decade alone, with 6.4 \nmillion (11%) American children ages 4 to 17 having \nbeen diagnosed as of 2016 (Xu et\u00a0al. 2018). ADHD \nis today\u2019s most commonly diagnosed childhood \nmental health disorder, surpassing anxiety, depres-\nsion, oppositional defiant disorder, conduct disorder, \nand autism spectrum disorders (Centers for Disease \nControl and Prevention 2019). Without proper treat-\nment, ADHD can lead to more behavior problems, \npoorer social relationships, and lower academic per-\nformance (Hinshaw and Scheffler 2014).\nIncreasing medicalization through ADHD diag-\nnoses may partly reflect parents\u2019 and educators\u2019 \nawareness of the importance of behavioral/noncog-\nnitive skills for long-term success (Conrad 1975). \nAn ADHD diagnosis can connect children to posi-\ntive resources like proper treatment that effectively \nmanages inattention and hyperactivity/impulsivity \n(Molina et\u00a0al. 2009; Swanson, Baler, and Volkow \n2010) and to educational accommodations in \nschools like extra testing time (Gius 2007).\nHowever, the effect of an ADHD diagnosis may \nnot be uniformly positive (Owens and Jackson \n2017). Situating ADHD within the broader literature \non negative stereotypes surrounding children with \ndisabilities (Pescosolido et\u00a0 al. 2008), diagnosed \nchildren may experience negative feelings of being \nunlike their peers (Hinshaw 2005) and increased \nscrutiny and lower expectations from teachers \n(Eisenberg and Schneider 2007). Depending on how \n924810 HSBXXX10.1177/0022146520924810Journal of Health and Social BehaviorOwens\nresearch-article2020\n1Brown University, Providence, RI, USA\nCorresponding Author:\nJayanti Owens, Department of Sociology and Watson \nInstitute for International and Public Affairs, Brown \nUniversity, 108 George St., Box 1916, Providence, RI \n02912, USA. \nEmail: jayanti_owens@brown.edu\nSocial Class, Diagnoses of \nAttention-Deficit/Hyperactivity \nDisorder, and Child Well-Being\nJayanti Owens1\nAbstract\nAttention-deficit/hyperactivity disorder (ADHD) is the most commonly diagnosed mental health disorder \namong U.S. children. Diagnosis can bring positives, like proper treatment, extra testing time, and social \nsupport, but may also trigger negatives, like stigmatization. Although rates of diagnosis are high across \nsocioeconomic status (SES) groups, the balance of positive and negative consequences of diagnosis may \ndiffer by SES. In high-SES communities, mental health diagnoses are less stigmatized and parents have \ngreater ability to connect children to support resources, suggesting greater positive effects of diagnosis \nfor high-SES children. Alternatively, the greater academic pressure present in high-SES communities may \namplify the negative effects of mental health stigma, suggesting larger negative diagnostic effects. Using \nthe Early Childhood Longitudinal Study\u2013Kindergarten Cohort of 1998\u20131999, I found that diagnosed and \nmedicated high-SES but not low-SES children exhibit significantly poorer future self-competence and \nteacher-rated school behaviors than undiagnosed matches. Findings suggest that diagnosis may not always \nbe a net positive.\nKeywords\nADHD, children, education, inequality, well-being\n\nOwens\t\n135\nthese positives and negatives balance out, processes \nlike labeling and stigma can produce negative per-\nceived self-competence and ultimately, higher levels \nof the school behavior problems diagnosis is \nintended to address.\nRates of childhood ADHD diagnoses are high \nacross social class groups, ranging from 13% \namong children from families earning below the \nfederal poverty level to 10% of those earning 1 to \n3.9 times the poverty level and 9% of those earn-\ning at least 4 times the poverty level (Xu et\u00a0al. \n2018). In spite of relatively high rates of diagno-\nsis across social class groups, our understanding \nof the relationship between social class privilege \nand diagnosed children\u2019s well-being remains \nlimited.\nVariation across social class contexts can lead to \ndifferences in the effects of a diagnosis on child \nsocioemotional well-being (i.e., children\u2019s expo-\nsure to the positive or negative consequences of \ndiagnosis noted previously). On one hand, research \nin the tradition of Lareau (1989) would suggest that \nhigh-socioeconomic-status (SES) parents\u2019 ability to \nactivate social capital and other resources for edu-\ncational intervention is predominantly beneficial \nfor children. For children with disabilities, high-\nSES parents likewise may be better at navigating \nschools and the health care system to connect their \nchild to proper treatment and other supports while \nalso systematically avoiding negative exposures, \nlike stigma (Blum 2015; Lareau 1989). For exam-\nple, high-SES families benefit from less mental \nhealth stigma in high-SES communities compared \nto less advantaged communities, including many \ncommunities of color (Blum 2015; Gary 2005). \nHigh-SES children may also experience earlier \ndetection and greater medication treatment and \nadherence, strategic medication use, and educa-\ntional accommodations like extra testing time com-\npared to low-SES children (King, Jennings, and \nFletcher 2014). As such, one prediction is that high-\nSES children will exhibit greater positive effects of \ndiagnosis than their less advantaged peers.\nOn the other hand, the negatives associated with \nan ADHD diagnosis may be larger for high-SES \nthan low-SES children. High-SES children might \nsuffer from the negative effects of mental health \nstigma associated with ADHD diagnosis even if \noverall levels of such stigma are lower in high-SES \nthan less advantaged communities (Francis 2012). \nFor example, whereas low-SES children are more \nlikely to experience multiple other negative stigmas \n(e.g., from school suspension or family instability), \nthe stigma associated with an ADHD diagnosis may \nbe one of the few stigmas experienced by high-SES \nchildren. Additionally, if high-SES families are bet-\nter able to access or adhere to medication treatment \nfor ADHD, any negative side effects of treatment \ncombined with the rise of polypharmacy might also \ndisproportionately negatively affect high-SES chil-\ndren (Bussing and Winterstein 2012; Comer, \nOlfson, and Mojtabai 2010). Thus, a competing \nprediction is that the marginal negative effect of \ndiagnosis may actually be larger for high-SES than \nlow-SES children even if overall levels of stigma \nare lower in high-SES than less advantaged com-\nmunities (Francis 2012).\nWhether and under what conditions a childhood \nADHD diagnosis is associated with better or worse \nbehavioral outcomes remains an open empirical \nquestion. This study helps reconcile these conflict-\ning expectations by investigating the following \nquestions:\nResearch Question 1: How does ADHD diagnosis \naffect future perceived self-competence and \nultimately, teacher-rated social and academic \nbehaviors at school?\nResearch Question 2: Does an ADHD diagnosis \naffect future perceived self-competence and \nschool behaviors differently for children from \nhigh-SES as opposed to low-SES backgrounds?\nThe increase in diagnoses of many childhood \nmental health disorders is motivated by the belief \nthat positive consequences, like proper treatment \nthat effectively manages symptoms, outweigh neg-\native consequences, like social stigma. However, \nempirically testing this belief requires disentan-\ngling the implicated social, psychological, and \nmedical processes. This study introduces a general-\nized framework for contextualizing these poten-\ntially positive and negative influences on children\u2019s \nwell-being and how they might differ by social \nclass. I then apply this framework to the case of \nchildhood diagnoses of ADHD. Using matching \ntechniques among children in the Early Childhood \nLongitudinal Study\u2013Kindergarten Cohort of 1998\u2013\n1999, I find that diagnosed and medicated middle- \nand upper-SES\u2014but not lower-SES\u2014children \nexhibit significantly poorer future self-competence \nand teacher-rated school behaviors than undiag-\nnosed matches. Findings carry important implica-\ntions for the reproduction of privilege within the \nlarger process of the medicalization of childhood \nbehaviors.\n\n136\t\nJournal of Health and Social Behavior 61(2)\nBackground\nTheoretical Framework: Disentangling \nSocial, Psychological, and Medical \nInfluences on Children\u2019s Well-Being\nI interpret the effects of ADHD diagnosis on chil-\ndren\u2019s socioemotional well-being in the context of a \nframework that encompasses broader interrelated \nsocial, psychological, and medical factors. In the \nfollowing, I note how these factors can unfold in \nfamilies and schools in ways that have both positive \nand negative influences on diagnosed children and \nhow these may differ by social class.\nPositive social, educational, and medical processes \nfor diagnosed children compared to similar undiag-\nnosed children.\u2002 Children may experience benefits \nfrom a mental health diagnosis compared to their \nsimilar undiagnosed counterparts. Diagnosis can \nbring access to appropriate medications that help \ncontrol behaviors, reduce anxiety, and improve \nmood. In schools, diagnosis may also bring positive \neducational and social supports, such as extra test-\ning time through a 504 Plan, or connection to peers \nwho have similar mental health challenges, through \nan Individualized Education Program (IEP) and \nspecial education classes (Gius 2007). That is, diag-\nnosis can help \u201clevel the playing field\u201d compared to \notherwise similar undiagnosed children.\nThe education and disability literatures help to \nexplain why these positive effects may be larger for \nchildren from high-SES families. High-SES parents \nare more able to mobilize social capital and other \nresources to successfully navigate school and medi-\ncal bureaucracies in order to connect their child to \npositive resources at the first signs of difficulty \n(Lareau 1989). For children with disabilities, Ong-\nDean (2009:4) argued that \u201cit is precisely because \nschools often try to manage students with disabili-\nties in the easiest, cheapest ways that parents\u2019 advo-\ncacy is so important,\u201d but \u201cparents\u2019 resources affect \nhow far they can shoulder this burden of being \ninvolved in identifying and accommodating their \nchildren\u2019s disabilities\u201d (p. 4). For example, Liu, \nKing, and Bearman (2010) show that residential \nproximity to other children diagnosed with autism \nshapes a child\u2019s own subsequent diagnosis due to \nthe spread of information about how to secure diag-\nnoses and the benefits of doing so. Blum (2015) \nalso shows that social capital is one resource that \nhelps high-SES parents connect children to the best \nteachers and proper treatments and educational \naccommodations. Finally, mental health conditions \nare also less stigmatized in high-SES than low-SES \ncommunities, which can lead to better medication \nadherence (Ohan et\u00a0 al. 2013; Pescosolido et\u00a0 al. \n2008).\nIn addition to being better positioned to inter-\nvene at school, high-SES parents have greater \nknowledge of the behaviors and skills rewarded in \nschools and are more likely to support development \nof these skills through a style of parenting at home \nknown as \u201cconcerted cultivation\u201d (Lareau 2003). \nConcerted cultivation is oriented around fostering \nchildren\u2019s talents and ability to navigate social insti-\ntutions through participation in organized activities, \nincluding education-enhancing activities like tutor-\ning. Together, early detection, greater medication \nadherence and social support, and less stigma in \nhigh-SES compared to low-SES communities may \nlead to even greater positive effects of diagnosis for \nchildren from more advantaged than less advan-\ntaged backgrounds.\nNegative social and psychological processes for \ndiagnosed children compared to similar undiagnosed \nchildren.\u2002 Mental health diagnoses can also instigate \nnegative social and psychological processes. Litera-\nture focusing especially on labeling and stigma \nhelps to explain potential negative consequences \nassociated with diagnosis (Goffman 1963; Scheff \n1974). Negative social feedback from teachers and \npeers may even further exacerbate negative diag-\nnostic effects on the child, especially when future \nschool behaviors are rated by teachers (Rist 1977). \nThe child may experience a sense of inadequacy  \nor low self-esteem based in beliefs that they have \nless academic ability than their peers, or that their \n\u00adteachers and parents do not believe they can succeed \nat a high level (Eisenberg and Schneider 2007). \nMoreover, if children internalize this shift in self-\u00ad\ncompetence as a result of exposure to negative soci-\netal stereotypes about individuals with disabilities, \nstigma can instigate a self-fulfilling prophecy even \nin cases where teachers and peers do not know about \nthe child\u2019s diagnosis, resulting in poorer behavioral \nand socioemotional outcomes (Link et\u00a0al. 1989).\nDiagnosis may also have a larger negative effect \non high-SES than low-SES children. In the case of \nlearning disabilities (LD), Hale (2011) argues that \ndisabilities are socially constructed in that \u201cthe \nmore contextual structures demand academic skills \nand dispositions, the more children with learning \nimpairments become \u2018learning disabled\u2019\u201d (p. 4). In \nhigh-SES contexts where parents activate social \ncapital for school intervention (Lareau 1987) and \nreflect school values in their home environment \n(Lareau 2003), academic excellence is often not \n\nOwens\t\n137\nonly expected, but expected to come easily (Mueller \nand Abrutyn 2016).\nMoreover, Francis (2012) found that children \nwith \u201cinvisible disabilities,\u201d like ADHD and LD, \nexperienced higher levels of stigma than those with \n\u201cvisible disabilities,\u201d like intellectual disability, \nbecause their communities tended to accept invisible \ndiagnoses as legitimate while holding diagnosed \nchildren and their families accountable to the same \nhigh expectations of normally functioning children. \nWith the academic demands present in high-SES \ncontexts, even subtle negative feedback from teach-\ners, peers, or family, may lead diagnosed high-SES \nchildren to believe that they fail to meet academic \nexpectations. Given that high-SES children are also \nless likely to have as many other negative exposures \nas low-SES children, the marginal effect of a mental \nhealth label may loom even larger for high-SES than \nlow-SES children. High-SES children might believe \nthemselves to be even more \u201clearning disabled\u201d than \ntheir low-SES counterparts (Hale 2011) while also \nbeing more likely to tie their sense of self-worth to \ntheir school performance (Conner and Pope 2013).\nIn sum, through the label that accompanies their \ndiagnosis, high-SES children might experience \nlarger negative effects of stigma, even though abso-\nlute levels of mental health stigma are lower than in \nlow-SES communities. As a result, diagnosed chil-\ndren may experience even poorer self-competence \nand, through a self-fulfilling prophecy, poorer \nschool behaviors, than diagnosed low-SES children.\nThe Case of Childhood ADHD \nDiagnosis and Child Well-being\nMost prior research on children with ADHD has \nexamined the consequences associated with having \nADHD versus not having ADHD based on a com-\nparison of outcomes of children with versus without \nhigh levels of behavioral problems prior to a diag-\nnosis (path a of Figure 1). This prior work finds that \nhaving high levels of behavior problems is associ-\nated with a range of poorer social, educational, and \nwell-being outcomes (Charach et\u00a0al. 2013; Currie \nand Stabile 2006). By contrast, only one prior study \nknown to the author isolates the role of the ADHD \ndiagnosis (path d of Figure 1), as opposed to that of \nthe underlying behavior problems (path a) or other \nfactors (paths b\u2013c). In that study, Owens and \nJackson (2017) compare diagnosed and undiag-\nnosed children who had the same underlying pre-\ndiagnosis behavioral problems. They find that, even \namong diagnosed and medicated children who had \nmild behavior problems prior to diagnosis, an early \nelementary school ADHD diagnosis is associated \nwith roughly .30 standard deviation (SD) lower \nmath and reading scores in 8th grade. Owens and \nJackson (2017) focus on long-term academic out-\ncomes and as such are unable to shed light on the \nimplicated social and psychological outcomes in \naddition to medical processes.\nThe present study builds upon that prior work to \napply the theoretical framework elaborated above to \nthe case of childhood ADHD diagnosis. I disentangle \nFigure 1.\u2002 Framework for Assessing the Relationship between Social Class, ADHD Diagnosis, \nMedication Treatment, and Children\u2019s Future Socioemotional Well-being.\n\n138\t\nJournal of Health and Social Behavior 61(2)\nimplicated social, psychological, and medical pro-\ncesses and examine potentially heterogeneous diag-\nnostic effects on children\u2019s subsequent perceived \nself-competence and teacher-rated school behaviors \nby social class.\nSocial Class Differences in the Positive Effects  \nof Diagnosis Compared to Similar Undiagnosed \n\u00adChildren.\u2002 For all children, an ADHD diagnosis can \nlead to the social and educational supports dis-\ncussed above (Gius 2007). Moreover, in high-SES \nfamilies, parents are more likely to believe that an \nADHD diagnosis will be beneficial. Combined \nwith the fact that high-SES families are more able \nto activate their social capital to intervene in \nschools and the medical system, high-SES fami-\nlies are consequently more inclined and able to \nseek out diagnosis at the early signs of behavioral \ndifficulties. Thus:\nEven without medication, diagnosis is associated \nwith positive effects on future self-competence \nand school behaviors. Moreover, diagnosed and \nunmedicated children from high-SES backgrounds \non average exhibit better future self-competence \nand school behaviors than diagnosed and \nunmedicated children from low-SES backgrounds, \neach relative to their undiagnosed counterparts \n(hypothesis 1).\nIn terms of medical processes, the stimulant \nmedications commonly prescribed for ADHD  \n(e.g., Ritalin and Adderall) have been shown to effec-\ntively reduce inattention and hyperactivity in children \n(Molina et\u00a0al. 2009; Swanson et\u00a0al. 2010). Greater \nparental social capital and economic resources com-\nbined with lesser stigma surrounding diagnosis may \nmean greater medication access, adherence, and stra-\ntegic use among high-SES children (Hinshaw and \nScheffler 2014). Indeed, research shows that chil-\ndren from higher-SES backgrounds are more likely \nthan children from lower-SES backgrounds to \nincrease the use of the stimulant medications to \nimprove attention and concentration during months \nwith high levels of academic pressure, such as dur-\ning finals periods, and limit use during holiday \nbreaks (King et\u00a0al. 2014). Thus:\nDiagnosed and medicated children from high-\nSES backgrounds on average exhibit even better \nfuture self-competence and school behaviors \nthan diagnosed and medicated children from \nlow-SES backgrounds (hypothesis 2).\nSocial Class Differences in the Negative Effects  \nof Diagnosis Compared to Similar Undiagnosed \n\u00adChildren.\u2002 The rise in ADHD diagnoses is driven by \nan assumption about diagnosis\u2019 positive behavioral \neffects, but an ADHD diagnosis might instead insti-\ngate negative social feedback from teachers and \npeers. Its status as an \u201cinvisible disability\u201d may lead \ndiagnosed children and family members to internal-\nize a sense of culpability for the child\u2019s difficulties, \nwhich is consistent with experiences of labeling and \nstigma. For children from high-SES backgrounds in \nparticular, parents\u2019 ability to activate social capital \nfor school intervention (Lareau 1987) and reflect \nschool values in their home environment (Lareau \n2003) combined with fewer other negative expo-\nsures may further magnify diagnosed children\u2019s \nsense of being \u201cdisabled\u201d and unable to meet high \nacademic expectations. Thus:\nWithout medication, diagnosis has negative effects \non future self-competence and school behaviors. \nMoreover, diagnosed and unmedicated children \nfrom high-SES backgrounds on average exhibit \nworse future self-competence and school behaviors \nthan diagnosed and unmedicated children from \n\u00adlow-SES backgrounds, each relative to their \nundiagnosed counterparts (hypothesis 3).\nThe benefits of proper medication discussed previ-\nously may help effectively manage symptoms, and \nchildren from high-SES families may have greater \nmedication access or adherence. However, greater \nmedication access and adherence may also lead \nhigh-SES children to be more likely to experience \nnegative side-effects of treatment or polypharmacy. \nThus:\nDiagnosed and medicated children from high-\nSES backgrounds on average exhibit comparable \nfuture self-competence and school behaviors as \ndiagnosed and medicated children from low-\nSES backgrounds (hypothesis 4).\nThese hypotheses are summarized in Figure 2. \nThe y-axis displays expected differences in the \nfuture school behaviors of diagnosed vs. otherwise \ncomparable undiagnosed children. Positive bars \nrepresent better future teacher-rated school behav-\niors among diagnosed compared to undiagnosed \nchildren, while negative bars represent poorer \nfuture behaviors, and the dotted line represents no \ndifference between diagnosed and undiagnosed \nchildren.\n\nOwens\t\n139\nData And Methods\nResearch Design: Disentangling the \nEffects of Behavior, Diagnosis, and \nMedication\nData and sample.\u2002 To estimate differences between \nthe teacher-rated school behaviors and perceived self-\ncompetence outcomes of diagnosed and undiagnosed \nchildren, this study draws on the restricted-use Early \nChildhood Longitudinal Study-Kindergarten Cohort \nof 1998 (ECLS-K), an initially nationally representa-\ntive sample of kindergarteners followed through mid-\ndle school. The analytic sample used here consisted of \nthe 7,330 children who remained in the study through \n5th grade, were not missing data on either the out-\ncomes or ADHD diagnosis, and were either diag-\nnosed (N = 380) or plausible undiagnosed matches \nwho had comparable levels of early ADHD-related \nbehavioral problems in spite of not having been diag-\nnosed (N = 6,950). Details on the longitudinal sample, \nattrition, and trimming are detailed in the Online \nAppendix; cell sizes were rounded to the nearest 10 \nper the restricted-use data agreement.\nThese data offered several substantial strengths for \nisolating the net \u201cmarginal effect\u201d of an ADHD diag-\nnosis (path d of Figure 1)\u2014defined as the average dif-\nference between the outcomes of diagnosed and \notherwise comparable undiagnosed children who had \nthe same observed pre-diagnosis behavioral problems, \ncognitive skills, family/demographic contexts and \noverall propensities to be diagnosed even though one \nchild was diagnosed and the other was not. First, given \nthat ADHD diagnoses require functional impairment \nfrom inattention and/or hyperactivity/impulsivity in at \nleast two settings (American Psychiatric Association \n2013), both teachers and parents rated children\u2019s \nbehaviors. Moreover, ratings occurred in children\u2019s \nnaturalistic settings (school and home), independent \nof diagnostic evaluation and less subject to reporting \nbias.\nSecond, to directly measure the core behaviors \nof ADHD, access to certain individual-level behav-\nior items were made available through copyright \napproval from the publisher, allowing diagnosed \nand undiagnosed children to be matched on the indi-\nvidual subscales of inattention and hyperactivity/\nimpulsivity. Although imperfect, these reports \nresemble those used in ADHD screeners such as the \nConnor\u2019s, SNAP-IV, and DSM-IV (limitations are \ndiscussed in the Discussion section) (Currie and \nStabile 2006; Swanson et\u00a0 al. 2010). Moreover, to \naddress potential confounding from conduct/opposi-\ntional defiance disorder and depression, the two \nmost common co-morbid conditions of ADHD \n(Hinshaw and Scheffler 2014), I also matched on \nrelevant pre-diagnosis anti-social behaviors and \ninternalizing behaviors in addition to inattention and \nhyperactivity. Third, ratings were for all sample \nchildren, including the roughly 95% of children with \ndevelopmentally normal behaviors who are never \ndiagnosed, thus providing a wide range of suitable \nundiagnosed comparison cases, or \u201cmatches.\u201d\nFourth, behavior ratings were collected in the \nwave prior to diagnosis with ADHD, offering a \nsubstantial strength over measuring pre-diagnosis \nbehaviors in a static period. For example, exclu-\nsively measuring behaviors in kindergarten would \nunderstate behavior problems for children whose \nFigure 2.\u2002 Hypothesized Differences in the Future Self-concept and Behavior of Diagnosed vs. \nUndiagnosed Children, by Social Class and Whether Receiving Medication\n\n140\t\nJournal of Health and Social Behavior 61(2)\nproblems increase after school entry and before \ndiagnosis. Conversely, drawing exclusively on third \ngrade measures (the end of the diagnostic observa-\ntion period) could introduce reverse causality issues \nif diagnosis prompts behavior change. For undiag-\nnosed children, behavior problems come from first \ngrade, the near mid-point of the Kindergarten \nthrough third grade diagnostic observation period. \nSince behavior problems typically decrease with \nage, the use of kindergarten behaviors for undiag-\nnosed children would likely overstate problems, \nwhile third grade measures would understate them.\nTreatment of missing data.\u2002 Of the working ana-\nlytic sample of 7,330 children, 15.4% (N = 1,130) \nlacked complete information on matching variables. \nItem-missingness on these covariates was more \ncommon among boys, African-Americans, and chil-\ndren of mothers with less than a college degree. \nMultiple imputation of twenty datasets was used to \naddress item-missingness on matching variables. \nADHD diagnosis and the outcomes are included in \nthe imputation equation but children originally \nmissing on these outcomes are excluded from all \nanalyses, per Von Hippel (2007).\nAnalytic Strategy: Matching for Sample \nBalance\nBecause ADHD diagnosis is not randomly assigned, \nthere are a number of potential child, family, and \nclassroom/school confounders that may influence \nboth diagnosis and outcomes. These include ADHD-\nrelated behaviors prior to diagnosis, cognitive skills, \nsex, age-for-grade (whether they are older or younger \nthan their classmates) (Layton et\u00a0al. 2018), internaliz-\ning behavior problems score (a common \u00adco-\u00admorbidity \nof ADHD), insurance coverage status, race/ethnicity \n(Morgan et\u00a0al. 2013), parents\u2019 concerted cultivation \nparenting styles, type of kindergarten classroom, and \naverage classroom behavior based on teacher ratings \non a scale from zero (\u201cstudents are typically extremely \npoorly behaved\u201d) to four (\u201cstudents are typically \nextremely well behaved\u201d), which might help shape \nreference group effects. If unobserved, these factors \nmay lead to biased estimates of the marginal effect of \ndiagnosis.\nTo help address some of these potential con-\nfounders, standard multiple regression techniques \nwould introduce controls to adjust for consequential \nobservable differences between diagnosed and undi-\nagnosed children. However, standard controls within \nOLS regressions may not be adequate if there is \ninsufficient overlap, or balance, in the distributions \nof characteristics between diagnosed and undiag-\nnosed children (Imbens and Rubin 2015).\nResearchers have turned to matching techniques \nto help achieve sample balance on key variables for \nwhich OLS models may yield estimates that lack \nsample support/balance (Gangl 2010). This study \nused two types of matching to obtain estimates that \nare \u201cdoubly robust\u201d to confounding between diag-\nnosed and undiagnosed children: coarsened exact \nmatching (CEM) and propensity score matching \n(PSM) (Stuart et\u00a0al. 2009). First, I use CEM to pre-\nprocess the data and ensure sample balance on three \nkey factors that shape both diagnosis and the out-\ncomes: family social class group (given the study\u2019s \ntheoretical motivation), severity quartile of pre-\ndiagnosis behavioral problems (a key confounder \nof diagnostic effects), and child sex (since sex dif-\nferences in children\u2019s presentation of ADHD-\nrelated behaviors could produce improper matches \ngiven greater diagnosis of the hyperactive subtype \namong boys).\nThe advantages of CEM include that it is not  \na method of estimation but, rather, a way to pre-\u00ad\nprocess a data set. As a result, estimation based on \nthe matched data set will be less model-dependent \n(i.e., less a function of indefensible modeling deci-\nsions) than when based on the original full data set \n(Iacus, King, and Porro 2011). Because the degree \nof model dependence is bounded ex ante by the \nresearcher, estimates are less vulnerable to mis-\nspecification of functional form. Second, CEM bal-\nances the nonlinearities and interactions between \nvariables that exist in the data. Third, CEM restricts \nthe matched data to areas of common empirical sup-\nport, helping to ensure that the researchers are not \nextrapolating beyond the data. Finally, the trimming \ndescribed earlier leads to lower bias during subse-\nquent estimation, and (by removing heterogeneity) \nincreased efficiency (Iacus et\u00a0 al. 2011). Once the \nCEM samples have been generated, subsequent data \nanalysis resumes use of the original values of all \ncovariates (King et\u00a0al. 2010).\nNonetheless, CEM is limited by the number of \nexact matching dimensions it can accommodate. \nEven after CEM, still there may be additional \nobservable child, family, and classroom/school dif-\nferences between diagnosed and undiagnosed chil-\ndren even within coarsened groups (Imbens and \nRubin 2015). Thus, PSM was used second, after \nCEM, to help address additional possible confound-\ning by ensuring that diagnosed and undiagnosed \nchildren were comparable, or balanced, on 37 \npotential child, family, and classroom matching \nvariables included in the PSM equation: all six \n\nOwens\t\n141\nteacher- and parent-rated behavioral dimensions \nand 31 other variables mentioned above and \ndetailed in Table 1 and the Online Appendix.\nAltogether, matching allowed me to isolate the \nmarginal effects of diagnosis by comparing diag-\nnosed and undiagnosed children with comparable \nunderlying pre-diagnosis behavior problems, cog-\nnitive skills, sex, and other child, family, and class-\nroom characteristics. I examined variation by \nfamily social class. Finally, I examined the sensitiv-\nity of results to modeling approach (see the Online \nAppendix).\nModels\nFirst, differences by family social class in the mar-\nginal effects of diagnosis were estimated overall and \nthen separately by whether or not the child subse-\nquently received medication following diagnosis. In \nthe latter analyses, diagnosed children were further \nclassified as those who \u201cadditionally receive medi-\ncation\u201d and those who \u201cdo not receive medication,\u201d \nboth relative to undiagnosed children. This modera-\ntion analysis (first by family social class and then \nalso by medication status) occurred after CEM but \nbefore PSM (Stuart et\u00a0al. 2009). To guard against \nreverse-causality issues, sensitivity analyses relied on \nmedication receipt reported only in third grade; result-\ning estimates remained substantively unchanged. \nBaseline models pooling across social class are shown \nin Appendix Table A.1.\nThe CEM package in Stata 14 was used to inde-\npendently and ex ante ensure that diagnosed and \nundiagnosed children were \u201cexact matches\u201d on \nthree characteristics on which there are well-\u00ad\nestablished differences in diagnosis and future \nbehaviors: (1) family social class (3 groups); (2) \nquartile of pre-diagnosis behavioral problems based \non parent- and teacher-rated subscales for inatten-\ntive or hyperactive/impulsive behavioral type (16 \ngroups); and (3) child sex (2 groups). The continu-\nous variables for social class and pre-diagnosis \nADHD-related behaviors were first temporarily \n\u201ccoarsened\u201d into the above categories so that match-\ning occurred within the 96 broad groups above (3 \u00d7 \n16 \u00d7 2 = 96). All 380 diagnosed children were suc-\ncessfully matched to otherwise comparable undiag-\nnosed children. By contrast, 340 (.5% of) \nundiagnosed children were pruned for reasons of \nachieving balance described in the Online Appendix.\nFor the pre-diagnosis ADHD-related behaviors \nmentioned above, there was substantial overlap in \nthe distributions of both diagnosed and undiag-\nnosed children: for both groups, pre-diagnosis \nbehavioral ratings spanned the full range from \n\u201cnever\u201d to \u201calmost always.\u201d As mentioned earlier, \nwith 5.2% of sample children ages 4 to 10 diag-\nnosed with ADHD, nearly 95% of children are not \ndiagnosed, offering a large number of potential \nmatches who were not diagnosed but nonetheless \nhad the same propensity to be.\nPSM occured after the CEM and is detailed in \nthe Online Appendix (balance statistics are in \nAppendix Table A.2). This approach does not \nremove the risk of bias from omitted variables and \nmeasurement error. The term marginal \u201ceffect\u201d is \nused with this important caveat.\nKey Measures\nTable 1 displays all variables used in the analyses.\nDependent variables: child perceived self-\u00adcompetence \nand teacher-rated school behaviors in fifth grade.\u2002 \nTeacher-rated \u201cpositive approaches to learning\u201d\u2014 \nor attentiveness, task persistence, eagerness to learn, \nlearning independence, flexibility, and organiza-\ntion\u2014and teacher-rated \u201cexternalizing behavior \nproblems\u201d\u2014or social problems like the frequency \nof arguing, fighting, getting angry, acting impul-\nsively, and disturbing ongoing activities\u2014were sub-\nscales from the Social Rating Scale; the child-rated \noutcome was from the Self-Description Question-\nnaire (Tourangeau et\u00a0al. 2009). Each scale averaged \nacross items for behavioral frequency that ranged \nfrom 0 = \u201crarely\u201d to 3 = \u201calways.\u201d1\nI focused on fifth grade behavioral outcomes, \nfirst, because I anticipated the largest diagnostic \neffects here and, second, because I believed there \nwas a reciprocal relationship between perceived \nself-competence and teacher-rated school behav-\niors. Although student self-reported measures were \navailable in eighth grade, teacher ratings of school \nbehaviors were not. Relying on the eighth grade \nmeasures would have introduced a (false) assumed \ncausal ordering to the outcomes.\nPrimary independent variables.\u2002 ADHD diagnosis \nbetween kindergarten and third grade (\u201ctreat-\nment\u201d) was identified based on when the parent \nanswered \u201cyes\u201d to all three of the following ques-\ntions in a given wave: (1) \u201cHas the child been evalu-\nated by a professional in response to a problem in \npaying attention, learning, behaving, or in activity \nlevel?\u201d (2) \u201cHas the child received a diagnosis by \nthis professional?\u201d and; (3) \u201cWas the diagnosis for \nADHD, ADD, or hyperactivity?\u201d (see Morgan et\u00a0al. \n[2013] and Tourangeau et\u00a0al. [2009] for advantages \n\n142\t\nJournal of Health and Social Behavior 61(2)\nTable 1.\u2002 Means and Standard Deviations (or Proportions) for Variables Used in the Main Text (N = 7,330).\nUpper SES \n(N = 1,850)\nMiddle SES \n(N = 3,630)\nLower SES \n(N = 1,850)\nOverall \nMin\nOverall \nMax\n\u2002\nMean\nSD\nMean\nSD\nMean\nSD\n\u2002\nOutcomes (5th Grade)\n\u2003 Positive approaches to learning behaviors \nscore (teacher report)\u20135th grade\n2.29b\n.69\n2.10a,b\n.75\n1.91a\n.76\n0\n3\n\u2003 Negative externalizing behavior problems \nscore (teacher report)\u20135th Grade\n.42b\n.57\n.56a,b\n.65\n.63a\n.69\n0\n3\n\u2003 Perceived self-competence score\u20135th grade \n(child report)\n2.19b\n.77\n2.05a,b\n.80\n2.01a\n.81\n0\n3\nADHD Diagnosis (K\u20133rd Grades) and Possible Medication Receipt (3rd\u20135th Grades)\n\u2003 Diagnosed with ADHD\n.04\n.05\n.04\n0\n1\n\u2003 Receiving medication in 3rd\u20135th Grades  \n(of those diagnosed)\n.80\n.80\n.75\n\u2002\nPre-diagnosis ADHD-related and Co-occurring Behaviors (Measured in the Wave Prior to Diagnosis \nor 1st Grade for Undiagnosed)\n\u2003 Inattentive behaviors score (teacher report)\n\u2013.18b\n.80\n\u2013.01a\n.86\n.18a,b\n.90\n-1\n2\n\u2003 Hyperactivity behaviors score (teacher report)\n\u2013.09b\n.63\n.01a\n.69\n.07a,b\n.74\n-1\n2\n\u2003 ODD or CD behaviors score (teacher report)\n\u2013.07b\n.65\n\u2013.01a\n.66\n.09a,b\n.68\n-1\n2\n\u2003 Internalizing problems score (teacher report)\n\u2013.08b\n.61\n\u2013.01a\n.61\n.09a,b\n.64\n-1\n2\n\u2003 Inattentive behaviors score (parent report)\n\u2013.11b\n.66\n\u2013.01a\n.69\n.11a,b\n.71\n-1\n2\n\u2003 Hyperactivity behaviors score (parent report)\n\u2013.07b\n.57\n\u2013.01a\n.65\n.10a,b\n.73\n-1\n2\n\u2003 ODD or CD behaviors score (parent report)\n\u2013.07b\n.42\n\u2013.04a\n.46\n.14a,b\n.57\n-1\n2\nEarly Cognitive Development (Kindergarten)\n\u2003 Average reading achievement scores (std.)\n.50b\n.97\n\u2013.06a,b\n.91\n\u2013.68a\n.89\n-3\n4\n\u2003 Average math achievement scores (std.)\n.54b\n.89\n.04a,b\n.87\n\u2013.65a\n.91\n-3\n4\n\u2003 Fine motor skills score\n6.45b\n1.75\n6.10a,b\n1.94\n5.37a\n2.14\n0\n9\nEarly School Context (Kindergarten)\n\u2003 Average behavior of students in kindergarten \nclass (teacher report)\n2.60b\n.75\n2.47a,b\n.80\n2.33a\n.80\n0\n4\n\u2003 Child received any special education services \nin kindergarten\n.08b\n.10b\n.11\n0\n1\n\u2003 School in state with consequential educational \naccountability standards\n.70\n.68a\n.73a\n0\n2\nEarly \u201cConcerted Cultivation\u201d at Home and Parental Involvement in Education/Schooling \n(Kindergarten)\n\u2003 Parental participation in educational \ninstitutions\n4.87b\n1.28\n4.20a,b\n1.50\n2.96a\n1.59\n0\n7\n\u2003 Child activities and leisure time\n3.76b\n1.83\n2.76a,b\n1.70\n1.71a\n1.53\n0\n10\n\u2003 Parent perceptions of responsibilities towards \nchild cognitive and social development\n11.05b\n2.71\n10.41a,b\n2.95\n9.45a\n3.43\n0\n18\n\u2003 Number of books at home\n109.31b\n58.25\n82.15a,b 56.76\n41.82a\n45.09\n0\n200\n\u2003 Parent educational expectations for child\n1.93b\n.30\n1.69a,b\n.58\n1.44a\n.78\n0\n2\n\u2003 Child age (months) at kindergarten entry\n65.49\n4.24\n65.65a\n4.34\n65.28a\n4.44\n37\n81\nFamily Demographic and Other Child Characteristics (Kindergarten)\n\u2003 Age of social mother\n36.68b\n4.97\n33.72a,b\n5.94\n31.87a\n7.42\n18\n80\n\u2003 Mother has CES-D Score >9 (clinically \ndepressive symptoms)\n.10b\n.17a,b\n.25a\n0\n1\n\u2003 Child born weighing less than 5.5 lbs (low \nbirthweight)\n.06b\n.08a,b\n.10a\n0\n1\n\u2003 Child not covered by insurance\n.15\n.15a\n.20a\n0\n1\n\u2003 Number of other children in household\n1.51b\n.97\n1.44a,b\n1.06\n1.79a\n1.32\n0\n11\n\u2003 Child been in childcare outside home\n.53\n.53a\n.44a\n0\n1\n\u2003 Male\n.52\n.51\n.49\n0\n1\n\u2003 Black\n.03b\n.10a,b\n.20a\n0\n1\n(continued)\n\nOwens\t\n143\nof this conservative approach to identifying diag-\nnosed children). Children whose parents answered \n\u201cno\u201d to any of these questions were coded as \u201cnot \ndiagnosed with ADHD\u201d between respective waves. \nDiagnosis during the kindergarten through third \ngrade diagnostic observation period was confirmed \nusing parent report of \u201cyear of first [ADD/ADHD] \ndiagnosis.\u201d The 14% of diagnosed children who \nwere diagnosed by kindergarten were retained and \ncoded \u201c1\u201d (diagnosed), but sensitivity analyses \nexcluding these children for whom measurement of \nADHD-related behaviors did not precede diagnosis \nyielded consistent substantive results. Children first \ndiagnosed with ADHD after third grade were also \nretained but were coded \u201c0\u201d (undiagnosed). Their \ninclusion as potential \u201cundiagnosed matches\u201d \nyielded conservative (lower-bound) estimates of the \neffect of diagnosis as they may have had undiag-\nnosed ADHD during the kindergarten through third \ngrade diagnostic observation period.\nMedication treatment receipt between 3rd and \n5th grades (\u201ctreatment moderator\u201d) was ascer-\ntained based on parent report of whether the child \nwas \u201ctaking medication to control his/her behav-\nior\u201d in third or fifth grade. In fifth grade, parents \nwere asked to name these medications. 90% of the \nchildren \u201creceiving medication\u201d were taking one  \nof three medications frequently prescribed for \nADHD: Ritalin, Adderall, or Concerta. Of the 50 \nchildren diagnosed by kindergarten, 40 (80%) were \nreceiving medication by fifth grade. Of the 210 \nadditional children first diagnosed after third \ngrade, 150 (71%) were taking medication in fifth \ngrade.\nFamily social class background in kindergarten \n(\u201cpredictor/moderator\u201d) was categorized, follow-\ning Cheadle (2008), using a composite, standard-\nized scale consisting of mother/female guardian\u2019s \nand father/male guardian\u2019s educational attainment, \nhousehold income, and parental occupational pres-\ntige (see chapter 7, pages 8\u201311 of Tourangeau et\u00a0al. \n[2009]) and conceptualized the bottom quartile as \n\u201clower SES,\u201d the middle two quartiles as \u201cmiddle \nSES,\u201d and the top quartile as \u201cupper SES\u201d. There \nare, of course, multiple approaches for measuring \nsocial class. The composite gives precedence to \nfamily income, although I note that the results were \nsimilar when using maternal educational level \n(mom has a bachelor\u2019s degree or higher versus \nmom has less than a bachelor\u2019s degree).\nOther variables in the PSM equation. To help \nensure that diagnosed and undiagnosed children \nwere as comparable as possible on observable char-\nacteristics, a number of other key measures, includ-\ning pre-diagnosis ADHD-related behaviors and \ncommonly co-occurring (\u201ccomorbid\u201d) behaviors, \nwere included as PSM variables and are described \nin the Online Appendix.\nUpper SES \n(N = 1,850)\nMiddle SES \n(N = 3,630)\nLower SES \n(N = 1,850)\nOverall \nMin\nOverall \nMax\n\u2002\nMean\nSD\nMean\nSD\nMean\nSD\n\u2002\n\u2003 Hispanic\n.07b\n.13a,b\n.36a\n0\n1\n\u2003 White\n.79b\n.67a,b\n.32a\n0\n1\n\u2003 Other race/ethnicity\n.11\n.10\n.12\n0\n1\n\u2003 Single parent household\n.07b\n.14a,b\n.29a\n0\n1\n\u2003 Social father present in household\n.01b\n.07b\n.07\n0\n1\n\u2003 Other family type in household\n.04b\n.05a,b\n.08a\n0\n1\n\u2003 Two biological parents in household\n.88b\n.75a,b\n.56a\n0\n1\n\u2003 Lives in Midwest\n.31\n.32a\n.18a\n0\n1\n\u2003 Lives in West\n.20b\n.18a,b\n.27a\n0\n1\n\u2003 Lives in Northeast\n.22b\n.19a,b\n.13a\n0\n1\n\u2003 Lives in South\n.26b\n.30a,b\n.40a\n0\n1\nSource: Early Childhood Longitudinal Study-Kindergarten Cohort, 1998\u20131999.\nNote: ADHD = attention-deficit/hyperactivity disorder; CD = compulsive disorder; ODD = oppositional defiant \ndisorder; SD = standard deviation; SES = socioeconomic status.\naSignificant difference between lower SES and middle SES at p < .05. All significant/non-significant differences between \nlower SES and middle SES also apply to the difference between lower SES and upper SES (except child age, which is \nnot significant between lower SES and upper SES).\nbSignificant difference between middle SES and upper SES at p < .05.\nTable 1.\u2002 (continued)\n\n144\t\nJournal of Health and Social Behavior 61(2)\nResults\nThe primary goals of this analysis are to: (1) deter-\nmine whether differences in the future perceived \nself-competence and teacher-rated school behaviors \nof diagnosed and undiagnosed children are located \namong those from middle- or upper-SES rather than \nlower-SES backgrounds; (2) examine whether any \nsocial class differences hold even for students who \nreceive medication treatment following diagnosis, \nand; (3) explore variation in these relationships by \nchild sex.\nDescriptive Differences by Family  \nSocial Class\nTo understand the differing contexts within which a \nchildhood ADHD diagnosis occurs in this sample, \nTable 1 displays descriptive statistics for analyzed \nvariables by family social class. There is a statisti-\ncally significant social class gradient on all three \noutcomes: lower-SES children exhibit the poorest \npositive approaches to learning, perceived self-\ncompetence, and negative externalizing problems, \nfollowed by middle- and then upper-SES children.\nHowever, on diagnosis and medication use as \nwell as many predictors, middle-SES and upper-\nSES children are quite similar. Consistent with \nnational statistics interpolated for children ages five \nto nine, roughly 4% to 5% of lower-SES, middle-\nSES, and upper-SES children are each diagnosed \nwith ADHD (see Appendix Table A.3 for counts) \n(Xu et\u00a0al. 2018). 80% of both diagnosed middle- \nand upper-SES children receive medication follow-\ning diagnosis, compared to 75% of diagnosed \nchildren from lower-SES families. Additional \ndescriptive differences are discussed in the Online \nAppendix.\nSocial Class Differences in the \n\u201cMarginal Effect\u201d of an ADHD \nDiagnosis\nGiven this study\u2019s theoretical emphasis on moderat-\ning differences by family social class, Table 2 pres-\nents estimates of differences in outcomes between \ndiagnosed children and their undiagnosed matches \nof the same social class group. For clarity, results \nare presented sequentially for each social class \ngroup before making explicit comparisons across \nsocial class groups.\nOverall, results are consistent with hypothesis 3 \nand inconsistent with hypotheses 1 and 2: an ADHD \ndiagnosis has a negative \u201cmarginal effect\u201d on future \nchild self-competence and teacher-rated school \nbehaviors, but only among children from middle- \nand upper-SES families (Table 2). Both diagnosed \nchildren from upper-SES and middle-SES back-\ngrounds exhibit .36 points (or .36 / .69 = .52 SD) \nstatistically significantly lower future positive \nlearning-related behaviors in fifth grade compared \nto their undiagnosed counterparts who had similar \npropensities for diagnosis but were not diagnosed \nTable 2.\u2002 Average Marginal Relationships between an ADHD Diagnosis and Future Social and Academic \nBehaviors and Child\u2019s Perceived Self-competence, by Family Social Class (N = 7,330).\nPositive Approaches to Learning \n(Teacher Report) \u2013 5th Grade\nExternalizing Behavior Problems \n(Teacher Report) \u2013 5th Grade\nPerceived Self-competence \n(Child Report) \u2013 5th Grade\n\u2002\n(1)\n(2)\n(3)\n(4)\n(5)\n(6)\n(7)\n(8)\n(9)\n\u2002\nUpper\nSES\n(N = 1,850)\nMiddle\nSES\n(N = 3,630)\nLower\nSES\n(N = 1,850)\nUpper\nSES\n(N = 1,850)\nMiddle\nSES\n(N = 3,630)\nLower\nSES\n(N = 1,850)\nUpper\nSES\n(N = 1,850)\nMiddle\nSES\n(N = 3,630)\nLower\nSES\n(N = 1,850)\nDiagnosed \nwith ADD/\nADHD\n\u2013.36***a\n(.08)\n\u2013.36***b\n(.06)\n\u2013.10a,b\n(.09)\n.25***\n(.07)\n.18*\n(.06)\n.08\n(.08)\n\u2013.25**a\n(.09)\n\u2013.27***b\n(.06)\n.08a,b\n(.09)\nSource: Early Childhood Longitudinal Study-Kindergarten Cohort, 1998\u20131999.\nNote: Displaying propensity score matching estimates of the \u201caverage marginal effect\u201d of ADHD diagnosis with \ncoarsened exact matching to preprocess the data (see text). Standard errors in parentheses. ADD = attention deficit \ndisorder, ADHD = attention-deficit/hyperactivity disorder, SES = socioeconomic status.\naSignificant difference between lower SES and upper SES at p < .05.\nbSignificant difference between lower SES and middle SES at p < .05.\ncSignificant difference between middle-SES and upper-SES at p < .05.\n*p < .05, **p < .01, ***p < .001.\n\nOwens\t\n145\n(Models 1\u20132). Diagnosed upper-SES children also \nexhibit .25 points (.25 / .57 = .44 SD) statistically \nsignificantly higher externalizing problems in fifth \ngrade than their undiagnosed counterparts (Model \n4). Diagnosed middle-SES children likewise exhibit \nstatistically significant .18 points (.18 / .65 =  \n.28 SD) higher externalizing problems than their \nundiagnosed counterparts (Model 5). Finally, diag-\nnosed upper-SES children report statistically sig-\nnificantly lower perceived self-competence by .25 \npoints (.25 / .77=.32 SD). Diagnosed middle-SES \nchildren likewise report .27 points (.27 / .80 = .34 SD)  \nstatistically significantly lower perceived self-com-\npetence relative to undiagnosed matches who had a \nsimilar propensity to be diagnosed (Models 7\u20138). \nConsistent with prior qualitative work (Lareau \n2003), differences in outcomes between diagnosed \nand undiagnosed middle- versus upper-SES children \nare not statistically significant from one another.\nOne explanation for the poorer future school \nbehaviors and perceived self-competence of diag-\nnosed middle- and upper-SES (but not lower-SES) \nchildren is that diagnosis triggers a social process \nthat objectively leads to poorer behaviors in diag-\nnosed middle- and upper-SES children versus their \nundiagnosed matches. Alternatively, teachers may \nalso have biased reports, for example from aware-\nness of a child\u2019s diagnosis. Although this study is \nunable to distinguish between these processes, even \nsubjectively rated behaviors have important conse-\nquences for self-fulfilling prophecies as discussed \npreviously.\nMoreover, the lower perceived self-competence \nof diagnosed middle- and upper-SES children but \nnot diagnosed lower-SES children points to the \nsalience of an internalized psychological process \nconsistent with labeling. Internalization may be one \nkey mechanism underlying poorer future school \nbehaviors, as explored in the Online Appendix.\nResults in Table 3 differentiate between diag-\nnosed upper-SES children who did versus did not \nreceive medication following diagnosis. Results are \nconsistent with hypothesis 3 on two of the three \noutcomes and run counter to hypothesis 4. Among \ndiagnosed upper-SES children, ADHD diagnosis is \nassociated with poorer approaches to learning and \nexternalizing problems irrespective of medication \nuse, but is only significantly tied to poorer aca-\ndemic self-competence for those who are receiving \nmedication. Among middle-SES children, diagno-\nsis is tied to poorer outcomes on all three measures \nirrespective of medication receipt. Specifically, \nboth \u201cdiagnosed and medicated\u201d and \u201cdiagnosed \nand unmedicated\u201d upper-SES children respectively \nexhibit .46 points (.67 SD) and .30 points (.43 SD) \nsignificantly poorer positive approaches to learning \nin fifth grade than their undiagnosed counterparts \n(model 1). They also respectively exhibit .18 points \n(.32 SD) and .38 points (.67 SD) significantly \npoorer externalizing problems in fifth grade (model \n4). Similarly, both \u201cdiagnosed and medicated\u201d and \n\u201cdiagnosed and unmedicated\u201d middle-SES children \nrespectively exhibit .31 points (.41 SD) and .42 points \n(.56 SD) significantly poorer positive approaches to \nlearning (model 2) and .17 points (.26 SD) and .20 \n(.31 SD) significantly worse externalizing problems \nthan their undiagnosed matches (model 5).\nWhen it comes to perceived self-competence, \nboth \u201cdiagnosed and medicated\u201d and \u201cdiagnosed \nand unmedicated\u201d middle-SES children again score \nsignificantly worse than their undiagnosed matches, \nby .22 points (.28 SD) and .36 points (.45 SD), \nrespectively (model 8). However, among upper-\nSES children, only those who are diagnosed and \nmedicated score significantly worse than their undi-\nagnosed matches, by .41 points (.53 SD) (model 7). \nFor diagnosed upper-SES children not receiving \nmedication, the relationship between diagnosis and \nperceived self-competence is directionally similar \nand statistically non-significantly smaller (\u2013.24 \npoints or .31 SD) compared to upper-SES children \nreceiving medication (\u2013.41 points) (model 7). But \nthis .24 points poorer reported self-competence is \nnot itself statistically significantly different from 0. \nLikewise, none of the other differences between \n\u201cdiagnosed and medicated\u201d and \u201cdiagnosed and \nunmedicated\u201d children are significantly different \nwithin social class groups either.2\nBy contrast, for children in lower-SES families, \nADHD diagnosis is not significantly tied to any of \nthe outcome variables. That is, diagnosed lower-\nSES children have statistically indistinguishable \nlevels of the fifth grade outcomes as their undiag-\nnosed counterparts (Models 3, 6, and 9 of Table 2). \nSubstantively, the estimated effects of diagnosis are \nat least 50% to 66% smaller for lower-SES com-\npared to middle- and/or upper-SES children. On \nperceived self-competence and positive approaches \nto learning, differences are statistically significantly \nsmaller among lower-SES compared to both \u00admiddle- \nSES and upper-SES children (model 3 vs. 1 and 2, \nand model 9 vs. 7 and 8 of Table 2). On future \nteacher-rated externalizing problems, the differ-\nences between lower SES and middle SES or upper \nSES are not statistically significant but follow in the \nexpected direction (model 6 vs. 4 and 5).\nMedication receipt does not further moderate \ndifferences in the marginal effects of diagnosis for \n\n146\nTable 3.\u2002 Average Marginal Relationships between an ADHD Diagnosis and Future Social and Academic Behaviors and Child\u2019s Perceived Self-competence, by \nFamily Social Class and Medication Treatment Status (N = 7,330).\nPositive Approaches to Learning  \n(Teacher Report) \u2013 5th Grade\nExternalizing Behavior Problems  \n(Teacher Report) \u2013 5th Grade\nPerceived Self-competence (Child Report) \u2013  \n5th Grade\n\u2002\n(1)\n(2)\n(3)\n(4)\n(5)\n(6)\n(7)\n(8)\n(9)\n\u2002\nUpper\nSES\n(N = 1,850)\nMiddle SES\n(N = 3,630)\nLower\nSES\n(N = 1,850)\nUpper\nSES\n(N = 1,850)\nMiddle\nSES\n(N = 3,630)\nLower\nSES\n(N = 1,850)\nUpper\nSES\n(N = 1,850)\nMiddle SES\n(N = 3,630)\nLower\nSES\n(N = 1,850)\nDiagnosed with \nADD/ADHD, \nreceiving \nmedication\n\u2013.46***a\n\u2013.31***\n\u2013.18a\n.18*\n.17*\n.04\n\u2013.41**a\n\u2013.22**\n\u2013.04a\n\u2002\n(.10)\n(.08)\n(.10)\n(.08)\n(.08)\n(.04)\n(.15)\n(.07)\n(.12)\nDiagnosed with \nADD/ADHD, \nnot receiving \nmedication\n\u2013.30***\n\u2013.42***\n\u2013.06\n.38***a\n.20**\n.11a\n\u2013.24\n\u2013.36**b\n.16b\n\u2002\n(.04)\n(.11)\n(.20)\n(.08)\n(.07)\n(.09)\n(.17)\n(.11)\n(.17)\nSource: Early Childhood Longitudinal Study-Kindergarten Cohort, 1998\u20131999.\nNote: Displaying propensity score matching estimates of the \u201caverage marginal effect\u201d of ADHD diagnosis with subsequent medication treatment/non-treatment (each compared to \nundiagnosed children) with coarsened exact matching to preprocess the data (see text). Standard errors in parentheses. ADD = attention deficit disorder; ADHD = attention-deficit/\nhyperactivity disorder, SES = socioeconomic status.\naSignificant difference between lower SES and upper SES at p < .05.\nbSignificant difference between lower SES and middle SES at p < .05.\ncSignificant difference between middle SES and upper SES at p < .05.\nNone of the within-model (i.e., internal moderation) differences by medication treatment/non-treatment status reach statistical significance at p < .05.\n*p < .05, **p < .01, ***p < .001 (two-tailed t-tests).\n\nOwens\t\n147\nlower-SES children (Table 3). The magnitudes of \nthe differences in outcomes between \u201cdiagnosed \nand medicated\u201d and undiagnosed lower-SES chil-\ndren are likewise much smaller than the analogous \ndifferences for middle- and upper-class children \nand, in fact, are statistically significantly smaller on \nfuture positive approaches to learning and per-\nceived self-competence compared to upper-SES \nchildren (model 3 vs. 1 and model 9 vs. 7 of  \nTable 3). Without medication receipt, \u201cdiagnosed \nand unmedicated\u201d children from lower-SES back-\ngrounds report significantly better future perceived \nself-competence than \u00admiddle-SES children (.16 vs. \n\u2013.36) and exhibit \u00adsignificantly better future teacher-\nrated school externalizing problems than upper-\nSES children (.11 vs. .38), each relative to their \nundiagnosed counterparts. Between \u201cdiagnosed and \n\u00adunmedicated\u201d lower-SES vs. higher-SES children, \nhowever, \u00adneither differences in self-competence \n(.16 vs. \u2013.24) nor positive approaches to learning \n(e.g., \u2013.06 vs. \u2013.42) differ significantly, though each \nfollows the expected pattern. This may be due to \ncell size limitations for \u201cdiagnosed and unmedi-\ncated\u201d children, per endnote 3. Together, overall \nresults indicate that diagnosed and medicated \nupper-SES and middle-SES children\u2014but not \n\u00addiagnosed and medicated lower-SES children\u2014\nexperience significantly worse future outcomes, \neach relative to their otherwise comparable undiag-\nnosed counterparts.\nThe Overall Winners and Losers: An \nExamination of Predicted Scores\nFigure 3 displays predicted outcome scores for \ndiagnosed and undiagnosed children by family \nsocial class and medication treatment status. For \nundiagnosed children on all three outcomes (panels \na through c), there is a social class gradient that most \nadvantages the upper SES (white bars), followed by \nmiddle SES (grey bars), then lower SES (black \nbars).\nThe most important finding from Figure 3 is \nthat, on all three outcomes, \u201cdiagnosed and medi-\ncated\u201d upper-SES and middle-SES children fare \ncomparably to undiagnosed lower-SES children. \n\u201cDiagnosed and medicated\u201d upper- and middle-\nSES children also fare slightly worse than their \n\u201cdiagnosed and unmedicated\u201d counterparts on \nfuture positive approaches to learning and per-\nceived self-competence. The opposite is true for \nfuture externalizing problems.\nThat \u201cdiagnosed and medicated\u201d middle- and \nupper-SES children fare comparably to undiagnosed \nlower-SES children is surprising. One interpretation is \nthat medication more effectively improves the hyper-\nactivity/impulsivity that largely comprises externaliz-\ning problems than it does the inattention that partly \nconstitutes approaches to learning. Alternatively, med-\nication may have larger negative side effects for atten-\ntion and concentration than hyperactivity, as is \nindicated for depression medications.\nA third possibility is that medications may effec-\ntively act on the nervous system to improve both \ninattention \nand \nhyperactivity/impulsivity, \nas \ndesigned, but that there may be an additional psy-\nchological mechanism in addition to the biologic \nones. Even if fully effective at their stated aim, \nmedications are not designed to interrupt the psy-\nchological effects of feeling \u201cmarked,\u201d which may \nhave their own independent effects on children\u2019s \nperceived self-competence and school behaviors. \nMedication may even slightly increase a child\u2019s \nawareness of being \u201cmarked,\u201d or affect the way she \nor he is treated by others.\nFinally, on future perceived self-competence \n(panel c), \u201cdiagnosed and unmedicated\u201d lower-SES \nchildren are predicted to fare as well as undiag-\nnosed upper-SES children and slightly higher than \nundiagnosed middle-SES children. For these lower-\nSES children, diagnosis may provide a \u201clegitimate\u201d \n(i.e., medical) explanation for the child\u2019s difficul-\nties and ease some of the burden posed by other \nstructural barriers to school success. The lack of \nmedication treatment may diminish any effects of \nstigma experienced by the child or family. Together, \nthe result may be the high subjective levels of per-\nceived self-competence observed.\nRobustness Checks\nThe lower perceived self-competence and poorer \nschool-related behaviors of \u201cdiagnosed and medi-\ncated\u201d children from middle-SES and upper- \nSES backgrounds appears robust and systematic. \nNonetheless, I additionally examine whether results \nare driven by differential selection into: evaluation, \ndiagnosis by age, medication treatment, schools, \nand remission of symptoms between the third and \nfifth grades. Results discussed in the Online \nAppendix and displayed in Appendix Tables A.4 \nthrough A.8 lend confidence that selection along \nthese dimensions does not drive results. Moreover, \nestimates would need to be biased by more than \n57% in order to invalidate findings.\nFor example, if the lower medication rate among \ndiagnosed lower-SES than higher-SES children reflects \nunder-medication rather than less severe clinical \n\n148\t\nJournal of Health and Social Behavior 61(2)\nsymptoms (for example due to financial barriers or lack \nof insurance coverage), \u201cdiagnosed and unmedicated\u201d \nlower-SES children actually may be more similar to \n\u201cdiagnosed and medicated\u201d lower-SES children than is \nthe case among higher-SES children. This similarity \ncould lead to an underestimate of the moderating effect \nUpper SES (N=1,850)\nMiddle SES (N=3,630)\nLower SES (N=1,850)\n0\n0.5\n1\n1.5\n2\n2.5\n3\nUndiagnosed\nDiagnosed, Receiving\nMedica\u001fon\nDiagnosed, Not Receiving\nMedica\u001fon\nPredicted Perceived Self-Competence \nScore\n Perceived Self-Competence (Child Report) - 5th Grade\n0\n0.5\n1\n1.5\n2\n2.5\n3\nUndiagnosed\nDiagnosed, Receiving\nMedica\u001fon\nDiagnosed, Not Receiving\nMedica\u001fon\nPredicted Externalizing Behavior \nProblems Score\nExternalizing Behavior Problems (Teacher Report)-5th Gr.\n(Higher is \nr\ne\nh\ngi\nH\n(\n)r\ne\ntt\ne\nb\n is worse)\n(Higher is be\u001eer)\n0\n0.5\n1\n1.5\n2\n2.5\n3\nUndiagnosed\nDiagnosed, Receiving\nMedica\u001fon\nDiagnosed, Not Receiving\nMedica\u001fon\nPredicted Posi\u001fve Approaches to \nLearning Score\n Posi\u001fve Approaches to Learning (Teacher Report) - 5th Gr.\nB\nA\nC\nFigure 3.\u2002 Predicted Future School Behaviors and Perceived Self-competence Between \u2018Undiagnosed,\u2019 \n\u2018Diagnosed and Medicated,\u2019 and \u2018Diagnosed and Unmedicated\u2019 Children, by Family Social Class.\n\nOwens\t\n149\nof medication status among lower-SES children. \nHowever, I calculate that the difference between \u201cdiag-\nnosed and unmedicated\u201d and \u201cdiagnosed and medi-\ncated\u201d lower-SES children would have to be at a \nminimum 2.05 times (105%) greater (for perceived \nself-competence) and 3.67 times (267%) greater (for \npositive approaches to learning) for there to be a signifi-\ncant difference between \u201cdiagnosed and unmedicated\u201d \nand \u201cdiagnosed and medicated\u201d lower-SES children at \np < .05 (two-sided t-test).\nI also investigate whether results are driven by \nboys or external visibility of the diagnosis by teach-\ners/peers (rather than internalized stigma by the \nchild). Results discussed in the Online Appendix \nare inconsistent with both claims.\nDiscussion\nThis study shifts our understanding of the relation-\nship between social class privilege, including the \nability of high-SES parents to activate the types of \nsocial capital rewarded by schools, and children\u2019s \nwell-being and educational outcomes. Using the \ncase of childhood ADHD diagnoses, this study \noffers a conceptual framework for disentangling the \nconditions under which the interrelated social, psy-\nchological, and medical factors associated with a \nchildhood mental health diagnosis can trigger posi-\ntive, neutral, or negative effects on children\u2019s later \nwell-being and school behaviors. The use of match-\ning strategies helps to disentangle the effects of high \nlevels of behavior problems from those of the diag-\nnosis itself or the social class factors that may drive \ndifferential diagnosis.\nThis study uncovers three major findings. First, \nADHD diagnosis is tied to worse approaches to \nlearning, more behavior problems, and poorer \n\u00adacademic self-competence in fifth grade, but only \nfor children in upper- and middle-SES families. \nSecond, for children in low-SES families, ADHD \ndiagnosis is not significantly tied to any of the out-\ncomes considered here. Third, when differentiating \nbetween diagnosed upper-SES children who are \nversus are not receiving medication following diag-\nnosis, ADHD diagnosis is significantly tied to \npoorer learning approaches and greater externaliz-\ning problems regardless of medication receipt. By \ncontrast, ADHD diagnosis is only significantly tied \nto academic self-competence for those upper-SES \nchildren who are receiving medication.\nWhen it comes to the relationship between \nsocial class privilege and children\u2019s well-being, \nwork in the tradition of Lareau (1989) may be taken \nto assume that high-SES parents\u2019 activation of \nsocial capital and other resources for school inter-\nvention is predominantly beneficial to their chil-\ndren. However, the present findings indicate that \nthere is also a need to consider that there could be a \nnegative side to the consequences of this social cap-\nital and privilege in schools through children\u2019s \nincreased susceptibility to negative social labeling. \nThe ability of parents to convert social capital may \nbe highly influenced by the ways in which broader \nsocial structures (e.g., schools) differentiate and \nstigmatize disability differently among high-SES \nand low-SES families. For example, the blame \nassigned to parents and children with disability may \nbe heightened in the high-SES schools attended by \nmany high-SES children.\nThe findings also contribute to our understand-\ning of social labeling processes. While key litera-\ntures in sociology of education and medical \nsociology have highlighted potential negatives \nassociated with labeling (Link et\u00a0 al. 1989; Rist \n1977), the present study extends this work by show-\ning how parental social capital (e.g., school and \nmedical intervention) can have negative conse-\nquences for children\u2019s well-being.\nAlthough this study cannot be certain of the \nmechanisms, one interpretation is that in the high-\npressure environments of many middle- and upper-\nSES children, negative stereotypes applied to \nmembers of advantaged groups may elicit even \ngreater self-stigma, perhaps due to the greater aca-\ndemic demands and expectations in their local con-\ntexts.3 This internalization can produce poorer \nsubjective ratings of self-competence, as observed. \nEven subtle negative feedback from teachers and \npeers may exacerbate this effect. As with internalized \nstigma, lower perceived self-competence may result \nin a self-fulfilling prophecy that leads to poorer \nschool-related behaviors, instigating a \u00adreciprocal \nprocess (Mueller and Abrutyn 2016). Crucially, this \ninterpretation would suggest that ADHD diagnoses \nwould continue to have a large negative effect even \nif diagnoses become more common and normalized \nin high-SES communities.\nFinally, this study shows that the observed \npoorer outcomes of diagnosed middle- and upper-\nSES children can manifest at an earlier age than \npreviously documented. Here, patterns of negative \neffects of diagnosis appear similar for both middle- \nand upper-SES children, and quite distinct from \nlower-SES children, consistent with the patterns of \n\u201cconcerted cultivation\u201d Lareau (2003) observed \namong middle- and upper-SES parents. Altogether, \nfor medical sociology, education, and social stratifi-\ncation scholars, these findings add nuance to a large \n\n150\t\nJournal of Health and Social Behavior 61(2)\nbody of evidence finding that social stigmas have \ngreater effects among disadvantaged populations, \nsuch as the mark of a criminal record and hiring \n(Pager 2003) or a school suspension (Owens and \nMcLanahan Forthcoming).\nAt first glance, one might think that the results \ncould be an artifact of class-specific reference group \neffects: perhaps middle- and upper-SES children \nhave better-behaved classmates than low-SES chil-\ndren such that slight deviations in behavior may lead \nto artificially inflated teacher and child ratings of \nproblem behaviors. However, these results appear \nrobust and systematic even among teachers who \nsimilarly rate the average behavior of students in \ntheir classrooms and when using fifth grade teacher \nratings or an average across elementary school.\nAdditionally, descriptive patterns suggest that \nresults are not driven by differences in the schools \nattended by diagnosed and undiagnosed children: \nmean levels of the outcomes are similar across \ndiagnosed and undiagnosed children in the same \nschools. To help address the possibility that ADHD-\nrelated behavior problems may have worsened \nleading up to diagnosis, pre-diagnosis behavior \nmeasures come from the wave just prior to diagno-\nsis, offering a substantial advantage over measuring \npre-diagnosis behaviors at a static time point, such \nas kindergarten entry.\nThis study also has a number of limitations and \npossible extensions. The ADHD-related behav-\nioral scales (and co-morbid internalizing and \noppositional defiant behavioral scales) used here \nmeasure frequency of behaviors, not other rele-\nvant factors like their intensity or duration, and \nalso do not perfectly map on to DSM criteria. \nThese unobserved clinical factors, if correlated \nwith both diagnosis and outcomes, could lead to \nomitted variables bias. Given the null results \namong low-SES children, there would have to be \ndifferential omitted variables between diagnosed \nand undiagnosed children by family social class in \norder to yield biased estimates among the middle- \nand upper-SES children for whom negative effects \nare estimated. Additionally, the sample balance \nachieved on key concerted cultivation measures, \nearly parental education expectations, and child \ncognitive skills, helps protect against such class-\nspecific omitted variables. Nonetheless, measure-\nment error in pre-diagnosis behavioral problems \nmay be larger for children from lower-SES than \nhigher-SES families; adjusting for parenting styles \nmay not entirely address this issue. Depending on \nthe outcome, estimated diagnostic effects would \nhave had to be 9% to 566% larger among lower-\nSES children in order for there to have been \nstatistically significant diagnostic effects among \nlower-SES children. Given this range, an alterna-\ntive interpretation of results is that, rather than \nself-stigma concentrated among higher-SES chil-\ndren, the diagnosis and clinical care of ADHD \namong lower-SES children may be less consistent \nand reflective of actual symptoms.\nThese data also lack a direct measure of stigma \nand/or internalized shame, as mentioned above. \nHowever, prior work finds that children diagnosed \nwith ADHD experience substantial labeling and \nstigma (Pescosolido et\u00a0al. 2008). This prior work \nsuggests that stigma may be a reasonable mecha-\nnism underlying observed relationships. Future \nexperimental or qualitative work should directly \nexamine this possible mechanism. Finally, this \nstudy is limited by the small sample of unmedicated \ndiagnosed children (100 children total).\nThis study also carries practical implications. It \nmight caution against parents, educators, and medi-\ncal providers considering an ADHD diagnosis for \nmiddle- and upper-SES children at the first signs of \nbehavioral difficulties. For example, prior research \npoints to potential negative diagnostic effects on \nlater school behaviors and academic achievement \namong children who had only mild pre-diagnosis \nbehavioral problems (Owens 2020; Owens and \nJackson 2017). Further research is needed to under-\nstand diagnostic effects on academic outcomes \namong high-SES children, but these studies together \nsuggest that both potential positive and potential \nnegative consequences of diagnosis should be con-\nsidered. Findings should not dissuade ADHD diag-\nnosis for high-SES children who have a clear-cut \nneed for diagnosis.\nAcknowledgments\nI thank three anonymous reviewers, Editor Amy Burdette, \nJordan Conwell, Laura Doering, Heide Jackson, Margot \nJackson, Susanna Loeb, John Logan, John Mullahy, \nChandra Muller, Zhenchao Qian, Paul Rathouz, Emily \nRauscher, Stephanie Robert, Kelley Smith, and Michael \nWhite for invaluable feedback during the preparation of \nthis manuscript. Earlier versions of this manuscript were \npresented at the Robert Wood Johnson Foundation Health \nand Society Scholars Program Annual Meeting in 2015, \nthe University of Wisconsin\u2013Madison in 2015, the \nAmerican Sociological Association Annual Meeting in \n2019, and the Population Association of America Annual \nMeeting in 2019.\nFunding\nThe author disclosed receipt of the following financial \n\u00adsupport for the research, authorship, and/or publication of \n\nOwens\t\n151\nthis article: The author is grateful for support from the \nBrown Population Studies and Training Center, which \nreceives funding from the NICHD (P2CHD041020); \nBrown University; the National Academy of Education/\nSpencer Foundation Postdoctoral Fellowship Program;  \nthe Robert Wood Johnson Foundation Health and Society \nScholars Program; and the University of Wisconsin\u2013Madison.\nSupplemental Material\nSupplemental material for this article is available online.\nNotes\n1.\t\nThe data provided only averaged scales in 5th grade, \npreventing standardization of summed indexes. \nRegardless, standardizing was less preferred because \nit unreasonably assumes that teachers follow a stable \nlogic across contexts when rating behaviors.\n2.\t\nThe lack of significant internal moderation by \nmedication status may reflect small sample sizes \nand large standard errors among \u201cdiagnosed and \nunmedicated\u201d children. Replication analyses for \nthe two available outcomes using the more recent \nECLS-K:2010\u201311, which includes more \u201cdiagnosed \nand unmedicated\u201d children (n = 70 for upper-SES \nand n = 80 for lower-SES children, up from n = 20 \nand n = 30, respectively, in the ECLS-K: 1998\u201399 \ncohort), reveal descriptively (and in some cases, \nsignificantly) larger negative effects of diagnosis \nfor middle- and upper-SES children and \u201cdiagnosed \nand unmedicated\u201d lower-SES children, but not for \n\u201cdiagnosed and medicated\u201d lower-SES children.\n3.\t\nWhile stereotype threat faced by students of color or \nof women in STEM may seem like counter-exam-\nples, these negative stereotypic \u201cmarks\u201d only target \nstructurally disadvantaged groups.\nReferences\nAmerican Psychiatric Association. 2013. \u201cDiagnostic and \nStatistical Manual of Mental Disorders: DSM-5.\u201d \nArlington: American Psychiatric Publishing.\nBlum, Linda M. 2015. Raising Generation Rx: Mothering \nKids with Invisible Disabilities in an Age of Inequality. \nNew York: NYU Press.\nBussing, Regina, and Almut G. Winterstein. 2012. \n\u201cPolypharmacy in Attention Deficit Hyperactivity \nDisorder Treatment: Current Status, Challenges \nand Next Steps.\u201d Current Psychiatry Reports 14(5): \n447\u201349.\nCenters for Disease Control and Prevention. 2019. \n\u201cData and Statistics on Children\u2019s Mental Health.\u201d \nWashington, D.C.: CDC.\nCharach, Alice, Patricia Carson, Steven Fox, Muhammad \nUsman Ali, Julianna Beckett, and Choon Guan \nLim. 2013. \u201cInterventions for Preschool Children at \nHigh Risk for ADHD: A Comparative Effectiveness \nReview.\u201d Pediatrics 131(5):1\u201321.\nCheadle, Jacob E. 2008. \u201cEducational Investment, Family \nContext, and Children\u2019s Math and Reading Growth \nfrom Kindergarten through the Third Grade.\u201d \nSociology of Education 81(1):1\u201331.\nComer, Jonathan S., Mark Olfson, and Ramin Mojtabai. \n2010. \u201cNational Trends in Child and Adolescent \nPsychotropic Polypharmacy in Office-based Practice, \n1996\u20132007.\u201d Journal of the American Academy of \nChild & Adolescent Psychiatry 49(10):1001\u201310.\nConner, Jerusha O., and Denise C. Pope. 2013. \u201cNot Just \nRobo-students: Why Full Engagement Matters and \nHow Schools Can Promote It.\u201d Journal of Youth and \nAdolescence 42(9):1426\u201342.\nConrad, Peter. 1975. \u201cThe Discovery of Hyperkinesis: \nNotes on the Medicalization of Deviant Behavior.\u201d \nSocial Problems 23(1):12\u201321.\nCurrie, Janet, and Mark Stabile. 2006. \u201cChild Mental \nHealth and Human Capital Accumulation: The Case of \nADHD.\u201d Journal of Health Economics 25(6):1094\u2013118.\nEisenberg, Daniel, and Helen Schneider. 2007. \u201cPerceptions \nof Academic Skills of Children Diagnosed with \nADHD.\u201d Journal of Attention Disorders 10(4):390\u201397.\nFrancis, Ara. 2012. \u201cStigma in an Era of Medicalisation \nand Anxious Parenting: How Proximity and \nCulpability Shape Middle-class Parents\u2019 Experiences \nof Disgrace.\u201d Sociology of Health & Illness \n34(6):927\u201342.\nGangl, Markus. 2010. \u201cCausal Inference in Sociological \nResearch.\u201d Annual Review of Sociology 36:21\u201347.\nGary, Faye A. 2005. \u201cStigma: Barrier to Mental Health \nCare Among Ethnic Minorities.\u201d Issues in Mental \nHealth Nursing 26(10):979\u201399.\nGius, Mark P. 2007. \u201cThe Impact of the Americans with \nDisabilities Act on Per-student Public Education \nExpenditures at the State Level: 1987\u20132000.\u201d American \nJournal of Economics and Sociology 66(5):925\u201336.\nGoffman, Erving. 1963. Stigma: Notes on the Management \nof Spoiled Identity. New York: Simon and Schuster.\nHale, Chris. 2011. From Exclusivity to Exclusion: The LD \nExperience of Privileged Parents. Rotterdam: Sense \nPublishers.\nHinshaw, Stephen P. 2005. \u201cThe Stigmatization of Mental \nIllness in Children and Parents: Developmental Issues, \nFamily Concerns, and Research Needs.\u201d Journal of \nChild Psychology and Psychiatry 46(7):714\u201334.\nHinshaw, Stephen P., and Richard M. Scheffler. 2014. \nThe ADHD Explosion: Myths, Medication, Money, \nand Today\u2019s Push for Performance. Oxford: Oxford \nUniversity Press.\nIacus, Stefano M., Gary King, and Giuseppe Porro. 2011. \n\u201cMultivariate Matching Methods that are Monotonic \nImbalance Bounding.\u201d Journal of the American \nStatistical Association 106(493):345\u201361.\nImbens, Guido W., and Donald B. Rubin. 2015. Causal \nInference in Statistics, Social, and Biomedical \nSciences. Cambridge: Cambridge University Press.\nKing, Gary, Matthew Blackwell, Stefano Iacus, and \nGiuseppe Porro. 2010. \u201cCEM: Coarsened Exact \nMatching in Stata.\u201d Stata Journal 9(4):524\u201346.\n\n152\t\nJournal of Health and Social Behavior 61(2)\nKing, Marissa D., Jennifer Jennings, and Jason M. Fletcher. \n2014. \u201cMedical Adaptation to Academic Pressure: \nSchooling, Stimulant Use, and Socioeconomic Status.\u201d \nAmerican Sociological Review 79(6):1039\u201366.\nLareau, Annette. 1989. Home Advantage: Social Class \nand Parental Intervention in Elementary Education. \nNew York: Falmer Press.\nLareau, Annette. 2003. Unequal Childhoods: Class, Race, \nand Family Life. Berkeley: University of California \nPress.\nLayton, Timothy J., Michael L. Barnett, Tanner R. \nHicks, and Anupam B. Jena. 2018. \u201cAttention \nDeficit\u2013Hyperactivity Disorder and Month of School \nEnrollment.\u201d New England Journal of Medicine \n379(22):2122\u201330.\nLink, Bruce G., Francis T. Cullen, Elmer Struening, \nPatrick E. Shrout, and Bruce P. Dohrenwend. 1989. \n\u201cA Modified Labeling Theory Approach to Mental \nDisorders: An Empirical Assessment.\u201d American \nSociological Review 54(3):400\u201323.\nLiu, Ka-Yuet, Marissa King, and Peter S. Bearman. \n2010. \u201cSocial Influence and the Autism Epidemic.\u201d \nAmerican Journal of Sociology 115(5):1387\u20131434.\nMolina, Brooke S. G., Stephen P. Hinshaw, James M. \nSwanson, L. Eugene Arnold, Benedetto Vitiello, \nPeter S. Jensen, Jeffery N. Epstein, Betsy Hoza, \nLily Hechtman, Howard B. Abikoff, Glen R. Elliott, \nLaurence L. Greenhill, Jeffrey H. Newcorn, Karen \nC. Wells, Timothy Wigal, Robert D. Gibbons, Kwan \nHur, Patricia R. Houck, The MTA Cooperative Group. \n2009. \u201cThe MTA at 8 Years: Prospective Follow-up \nof Children Treated for Combined-type ADHD in a \nMultisite Study.\u201d Journal of the American Academy \nof Child & Adolescent Psychiatry 48(5):484\u2013500.\nMorgan, Paul L., Jeremy Staff, Marianne M. Hillemeier, \nGeorge Farkas, and Steven Maczuga. 2013. \n\u201cRacial and Ethnic Disparities in ADHD Diagnosis \nfrom Kindergarten to Eighth Grade.\u201d Pediatrics \n132(1):85\u201393.\nMueller, Anna S., and Seth Abrutyn. 2016. \u201cAdolescents \nUnder Pressure: A New Durkheimian Framework \nfor Understanding Adolescent Suicide in A Cohesive \nCommunity.\u201d American Sociological Review 81(5): \n877\u201399.\nOhan, Jeneva L., Troy A. Visser, Rachael G. Moss, and \nNicholas B. Allen. 2013. \u201cParents\u2019 Stigmatizing \nAttitudes Toward Psychiatric Labels for ADHD \nand Depression.\u201d Psychiatric Services 64(12): \n1270\u201373.\nOng-Dean, Colin. 2009. Distinguishing Disability: \nParents, Privilege, and Special Education. Chicago: \nUniversity of Chicago Press.\nOwens, Jayanti. Forthcoming. \u201cRelationships between \nan ADHD Diagnosis and Future School Behaviors \namong Children with Mild Behavioral Problems.\u201d \nSociology of Education Online First: https://doi.\norg/10.1177/0038040720909296.\nOwens, Jayanti, and Heide Jackson. 2017. \u201cAttention-\nDeficit/Hyperactivity Disorder Severity, Diagnosis, \n& Later Academic Achievement in a National \nSample.\u201d Social Science Research 61:251\u201365.\nOwens, Jayanti, and Sara McLanahan. Forthcoming. \n\u201cUnpacking the Drivers of Racial Disparities in \nSchool Suspension and Expulsion in the U.S.\u201d Social \nForces Online First: https://academic.oup.com/sf/\narticle/doi/10.1093/sf/soz095/5521044.\nPager, Devah. 2003. \u201cThe Mark of a Criminal Record.\u201d \nAmerican Journal of Sociology 108(5):937\u201375.\nPescosolido, Bernice A., Peter S. Jensen, Jack K. Martin, \nBrea L. Perry, Sigrun Olafsdottir, and Danielle \nFettes. 2008. \u201cPublic Knowledge and Assessment \nof Child Mental Health Problems: Findings from \nthe National Stigma Study\u2013Children.\u201d Journal \nof the American Academy of Child & Adolescent \nPsychiatry 47(3):339\u201349.\nRist, Ray C. 1977. \u201cOn Understanding the Processes of \nSchooling: The Contributions of Labeling Theory.\u201d \nPp. 292\u2013305 in Power and Ideology in Education, \nedited by J. Karabel and A.H. Halsey. New York: \nOxford University Press.\nScheff, Thomas J. 1974. \u201cThe Labelling Theory of \nMental Illness.\u201d American Sociological Review \n39(3):444\u201352.\nStuart, Elizabeth A., Sue M. Marcus, Marcela V. \nHorvitz-Lennon, Robert D. Gibbons, Sharon Lise T. \nNormand, and C. Hendricks Brown. 2009. \u201cUsing \nNon-experimental Data to Estimate Treatment \nEffects.\u201d Psychiatric Annals 39(7):719\u201328.\nSwanson, James, Ruben D. Baler, and Nora D. Volkow. \n2010. \u201cUnderstanding the Effects of Stimulant \nMedications on Cognition in Individuals with \nAttention-deficit Hyperactivity Disorder: A Decade of \nProgress.\u201d Neuropsychopharmacology 36(1):207\u201326.\nTourangeau, Karen, Christine Nord, Thanh L\u00ea, Alberto \nG. Sorongon, and Michelle Najarian. 2009. \u201cEarly \nChildhood Longitudinal Study, Kindergarten Class \nof 1998\u201399 (ECLS-K): Combined Users Manual \nfor the ECLS-K Eighth-grade and K\u20138 Full Sample \nData Files and Electronic Codebooks. NCES \n2009-004.\u201d Washington, DC: National Center for \nEducation Statistics.\nVon Hippel, Paul T. 2007. \u201cRegression with Missing Ys: \nAn Improved Strategy for Analyzing Multiply Imputed \nData.\u201d Sociological Methodology 37(1):83\u2013117.\nXu, Guifeng, Lane Strathearn, Buyun Liu, Binrang \nYang, and Wei Bao. 2018. \u201cTwenty-year Trends in \nDiagnosed Attention-deficit/Hyperactivity Disorder \namong US Children and Adolescents, 1997\u20132016.\u201d \nJAMA Network Open 1(4): 1-9. doi: 10.1001/\u00adjamane\ntworkopen.2018.1471\nAuthor Biography\nJayanti Owens is the Mary Tefft and John Hazen White, \nSr. Assistant Professor of Sociology and International and \nPublic Affairs at Brown University. Her research interests \ninclude sociology of education, social stratification, social \ndemography, organizations, and population health.\n\n\n    Scientific Research Findings:", "answer": "\u2022 ADHD diagnosis is tied to worse approaches to learning, more behavior problems, and poorer academic self-competence in fifth grade but only for children in upper- and middle-SES families.\n\u2022 For children in low-SES families, ADHD diagnosis is not significantly tied to any of the outcomes considered here.\n\u2022 For upper-SES children, ADHD diagnosis is significantly tied to poorer learning approaches and greater externalizing problems regardless of whether they are receiving medication for ADHD. By contrast, ADHD diagnosis is only tied to academic self- competence for those upper-SES children who are receiving medication.\n\u2022 Across all three outcomes (child-perceived self-competence and teacher-rated positive approaches to learning and externalizing problems), diagnosed and medicated upper-SES and middle-SES children fare comparably to undiagnosed lower-SES children.", "id": 78}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/0022146520939514\nJournal of Health and Social Behavior\n2020, Vol. 61(3) 275\u00ad\u2013289\n\u00a9 American Sociological Association 2020\nDOI: 10.1177/0022146520939514\njhsb.sagepub.com\nOriginal Article\nAn abundance of evidence indicates that the drug \noverdose epidemic in the United States is a \nnational public health emergency (Gomes et\u00a0 al. \n2018; Hedegaard, Mini\u00f1o, and Warner 2018; \nKariisa et\u00a0al. 2019). In 2017, 70,237 drug overdose \ndeaths occurred in the United States, with opioids \ninvolved in 67.8% of these fatal poisonings (Scholl \net\u00a0 al. 2019). The U.S. drug-related mortality rate \n(age-adjusted) increased from 6.1 per 100,000 peo-\nple in 1999 to 21.7 in 2017 (Hedegaard et\u00a0al. 2018). \nFrom 1999 to 2006, the average annual increase in \nthe drug-related mortality rate was 10%, and that \nrate has risen over time. From 2006 to 2014, the \naverage increase was 3%, which subsequently \njumped to 16% from 2014 to 2017 (Hedegaard et\u00a0al. \n2018).\nAlthough opioids are often recognized as the \nmajor contributor to drug-related mortality, it is \nimportant to note that in 2017, cocaine and other \npsychostimulants were involved in one-third of the \ndrug overdose deaths in the United States (Kariisa \net\u00a0al. 2019). Three-fourths of the cocaine-involved \ndeaths and one-half of the psychostimulant-\ninvolved deaths also included an opioid. Since \n2013, drug overdoses involving cocaine and psy-\nchostimulants have increased across all demo-\ngraphic groups and U.S. census regions (Kariisa \net\u00a0al. 2019). Thus, the current drug overdose epi-\ndemic in the United States appears to be an evolv-\ning one that is increasingly characterized by \n939514 HSBXXX10.1177/0022146520939514Journal of Health and Social BehaviorThombs et al.\nresearch-article2020\n1Boston College, Chestnut Hill, MA, USA\n2University of North Texas, Fort Worth, TX, USA\n3Northeastern University, Boston, MA, USA\nCorresponding Author:\nRyan P. Thombs, Department of Sociology, Boston \nCollege, 426 McGuinn Hall, 140 Commonwealth \nAvenue, Chestnut Hill, MA 02467, USA. \nEmail: thombs@bc.edu\nWhat Is Driving the Drug \nOverdose Epidemic in the \nUnited States?\nRyan P. Thombs1, Dennis L. Thombs2, Andrew K. Jorgenson1,  \nand Taylor Harris Braswell3\nAbstract\nThe demand-side perspective argues that the drug overdose epidemic is a consequence of changes in the \neconomy that leave behind working-class people who lack a college education. In contrast, the supply-side \nperspective maintains that the epidemic is primarily due to changes in the licit and illicit drug environment, \nwhereas a third, distinct perspective argues that income inequality is likely a key driver of the epidemic. To \nevaluate these competing perspectives, we use a two-level random intercept model and U.S. state-level \ndata from 2006 to 2017. Contrary to the demand-side approach, we find that educational attainment is \nnot associated with drug-related mortality. In support of the supply-side approach, we provide evidence \nindicating that opioid prescription rates are positively associated with drug-related mortality. We also \nfind that income inequality is a key driver of the epidemic, particularly the lack of resources going to the \nbottom 20% of earners. We conclude by arguing that considerations of income inequality are an important \nway to link the arguments made by the demand-side and the supply-side perspectives.\nKeywords\ndeaths of despair, drug overdose epidemic, income inequality, mortality, population health\n\n276\t\nJournal of Health and Social Behavior 61(3) \npolysubstance use (Jones, Einstein, and Compton \n2018; McCall Jones, Baldwin, and Compton 2017).\nAlthough drug-related mortality has been \nincreasing for nearly 20 years, there is no consensus \nabout its causes (Case and Deaton 2020; Monnat \n2018, 2019; Ruhm 2019). The debate emerges into \ntwo broad schools of thought\u2014the demand side \nversus the supply side. The demand-side perspec-\ntive, most recently framed as the \u201cdeaths of despair\u201d \nnarrative, argues that the current drug epidemic is a \nproduct of macroeconomic changes over the past \nhalf century that leave behind vulnerable members \nof the working class who lack a college education \n(Case and Deaton 2020). In contrast, the supply-\nside perspective argues that the increase in drug-\nrelated mortality is a product of drugs, particularly \nopioids, becoming more readily available and \naffordable through both licit and illicit means (Lin, \nLiu, and Ruhm 2020; Ruhm 2019; Singhal, Tien, \nand Hsia 2016). Although these two perspectives \nare often positioned as competing paradigms, a \nthird approach\u2014the income inequality perspective\u2014\nprovides a way to incorporate insights from both \napproaches because it explains how macrolevel \nchanges in the political economy directly affect \nindividuals and the ways by which they relate to \none another. The role of income inequality in the \ncurrent drug overdose epidemic is given negligible \ndiscussion in both the demand-side and the supply-\nside approaches. However, we argue that drug-\nrelated mortality is likely a product of both financial \nopportunities cultivated by wealthy and powerful \nelites as well as the vulnerabilities present in the \nworking class, particularly nonmobile and dis-\nplaced workers.\nIn the current study, we evaluated these three \napproaches to understanding the drug overdose epi-\ndemic. To do so, we used a panel data set of the 50 \nU.S. states and the District of Columbia covering \nthe 2006 to 2017 period. We analyzed these data \nusing a two-level random intercept model. In con-\ntrast to fixed-effects models that estimate within \neffects, multilevel models allow for the simultane-\nous estimation of within and between effects, which \npermitted us to evaluate how the drivers of the epi-\ndemic are associated with drug-related mortality \nthrough time and across states.\nBefore reporting the results of the analysis, we \ndiscuss the key arguments made by the demand-\nside and the supply-side approaches concerning the \ncurrent drug overdose epidemic and how insights \nfrom the income inequality and health literature can \npotentially serve as a link between the two. Our \nfindings support the arguments made by the \nsupply-side perspective and the inequality\u2013health \nliterature more so than the demand-side approach. \nHowever, the effect of income inequality on drug-\nrelated mortality is complex. Our findings indicate \nthat the share of income going to the top 5%, the top \n20%, and the Gini coefficient are not associated \nwith drug-related mortality but that the share of \nincome of the bottom 20% is associated with drug-\nrelated mortality. In other words, and in the context \nof income inequality, the lack of resources going to \nthe earners at the bottom of the income distribution \nis driving the drug overdose epidemic. We conclude \nthat a broader discussion of inequality provides a \nway to integrate the demand-side and the supply-\nside approaches to enhance our understanding of \ndrug-related mortality.\nBackground\nDivergent Perspectives on the Drivers of \nthe Drug Overdose Epidemic\nThe demand-side perspective.\u2002 The demand-side per-\nspective argues that the current drug overdose epi-\ndemic is a product of macroeconomic changes in the \nUnited States beginning in the 1970s (Case and \nDeaton 2015, 2017, 2020; Dasgupta, Beletsky, and \nCiccarone 2017; Monnat 2018). Case and Deaton \n(2015, 2017, 2020) showed that in some segments \nof the U.S. population, particularly among middle-\naged white men with a high school education or \nless, there is an association between diminished \neconomic opportunity and increased overall mortal-\nity. The phenomenon, recently described as deaths \nof despair, includes deaths related to drugs, suicide, \nand alcohol. They argue that this pattern of mortal-\nity is a consequence of structural changes in the U.S. \neconomy that have greatly reduced the number of \nwell-paying jobs (particularly in the manufacturing \nsector), which many once had access to regardless \nof their level of education (Case and Deaton 2017; \nCharles, Hurst, and Schwartz 2018).\nScholars working within this framework argue \nthat a variety of processes are behind these jobs \n\u00addisappearing\u2014production moving overseas, the \nweakening of labor unions, and computer and \nrobotic technology replacing skilled workers \n(Desilver 2017; Pierce and Schott 2020). These \neconomic changes led to the rise of a postindustrial \neconomy, causing elevated hopelessness and \ndespair (Case and Deaton 2017, 2020). According \nto Case and Deaton (2020), deaths of despair are \nprimarily restricted to those without a college edu-\ncation regardless of race or gender. Although it is \n\nThombs et al.\t\n277\ntrue that deaths of despair are highest among white \nmen without a college degree, deaths of despair \nhave also increased for women and African \nAmericans without a college education (Case and \nDeaton 2020). Thus, the demand-side approach \nplaces significant emphasis on differences in educa-\ntional attainment to explain the drug overdose \nepidemic.1\nThe supply-side perspective.\u2002 The second para-\ndigm used to explain the drug overdose epidemic is \nthe supply-side perspective, which argues that drugs, \nparticularly opioids, have become more readily \navailable since the 1990s (Paulozzi, Mack, and \nHockenberry 2014; Pezalla et\u00a0al. 2017; Ruhm 2019). \nThis perspective argues that although there may be a \nlink between macroeconomic conditions and drug-\nrelated mortality, the association disappears once \nsupply-side characteristics are accounted for (Ruhm \n2019). Moreover, this approach suggests that the epi-\ndemic is not due to changes in the macroeconomic \nenvironment for two reasons. First, states with rela-\ntively strong economies (e.g., Massachusetts) have \nexperienced high rates of drug-related mortality, and \nsecond, whites have a higher drug-related mortality \nrate than minority groups who have long lived in \ngreater economic precarity (Ruhm 2019).\nMoreover, scholars working within this approach \npoint to the increase in drug-related mortality \noccurring at the same time that opioids started to \nbecome more readily prescribed by physicians \n(Ruhm 2019). The supply-side paradigm highlights \nthe role that pharmaceutical companies play in \nincreasing opioid availability by encouraging phy-\nsicians to prescribe opioid analgesics (Hadland \net\u00a0 al. 2018; Makary, Overton, and Wang 2017). \nAlthough the number of prescriptions have declined \nsince 2010, the supply of illicit fentanyl and other \nopioids, such as heroin, have markedly increased in \nresponse (Ruhm 2019). For example, illicit fen-\ntanyl, produced in China, has dramatically \nincreased over the past decade (Drug Enforcement \nAgency 2016; Suzuki and El-Haddad 2017). Thus, \nthe supply-side perspective is primarily focused on \nthe role of specific sectors in the economy\u2014the \nhealth care/pharmaceutical industry and the illicit \ndrug industry\u2014rather than changes in the economy \nas a whole.\nThe income inequality\u2013health relationship: a frame-\nwork to integrate the demand-side and supply-side per-\nspectives.\u2002 Both the demand-side and supply-side \nperspectives downplay the role of income inequality \nin the drug overdose epidemic. Case and Deaton \n(2020) suggested that the rise in inequality is a prod-\nuct of the aforementioned macroeconomic changes, \nbut they did not believe it influences drug-related \nmortality to any discernable degree. Relying on cor-\nrelations, they showed that highly unequal states, \nlike California and New York, have relatively low \nrates of drug-related mortality, whereas more equal \nstates, like New Hampshire, have high mortality \nrates (Case and Deaton 2020). Similarly, from the \nsupply-side perspective, Ruhm (2019) argued that \nalthough demand-side conditions and inequality \nmay explain some portion of the drug epidemic, the \noverall expanded supply and availability of drugs \nprovides a much stronger explanation for the over-\ndose epidemic.\nA large body of research in the social sciences \nsuggests that inequality is a key driver of a range of \nhealth-related outcomes (see Pickett and Wilkinson \n2015; Wilkinson and Pickett 2010, 2019). Recent \nstudies that used relatively more sophisticated sta-\ntistical modeling techniques have provided empiri-\ncal evidence of the detrimental health impacts of \nmacro levels of income inequality, especially \nreductions in country-level and U.S. state-level \naverage life expectancy (e.g., Curran and Mahutga \n2018; Hill and Jorgenson 2018; Jorgenson et\u00a0 al. \n2020). A number of studies have also observed \ninequality to be associated with adult and infant \nmortality, obesity, HIV infections, mental illness, \nand homicides (Buot et\u00a0 al. 2014; Daly 2016; \nRibeiro et\u00a0 al. 2017; Torre and Myrskyl\u00e4 2014; \nWilkinson and Pickett 2010, 2019). Overall, the \nmajority of these studies suggest that inequality is \nharmful to human health, well-being, economies, \nand social cohesion.\nAlthough the demand-side and the supply-side \napproaches underplay the role of inequality in the \ndrug overdose epidemic, multiple theoretical per-\nspectives that can be applied to drug use help \nexplain why income inequality is related to various \npopulation health outcomes (Hill, Jorgenson et\u00a0al. \n2019; Hill and Jorgenson 2018; Jorgenson et\u00a0 al. \n2020). The psychosocial and social capital perspec-\ntives take a micro point of view to the inequality\u2013\nhealth relationship, whereas neomaterialism takes a \nmacro perspective. The psychosocial perspective \nsuggests that the stress of relative deprivation, from \nthe unequal distribution of income, contributes to \nlow self-esteem, emotional distress, and risky cop-\ning behaviors, such as drug use (Wilkinson and \nPickett 2010).\nA similar framework\u2014the social capital \n\u00adperspective\u2014argues that income inequality facili-\ntates widespread status competition, which tends to \n\n278\t\nJournal of Health and Social Behavior 61(3) \nundermine social cohesion and interpersonal trust \nand as a consequence, reduces collective political \nefforts to support vulnerable populations (Kawachi \net\u00a0 al. 1997; Truesdale and Jencks 2016). A third \napproach, the neomaterialist perspective, suggests \nthat income inequality concentrates wealth and \npower among elites and weakens broader commit-\nments to the general interests of society. These con-\nditions create political pressure to cut taxes, \nderegulate industry, and limit investments in public \nresources and social services that promote public \nhealth, all of which disproportionately impact those \nin lower income groups (Neumayer and Pl\u00fcmper \n2015).\nAlthough the inequality\u2013health link is a well-\nestablished body of research, few empirical investi-\ngations have studied the relationship between \nincome inequality and the drug epidemic, and those \nthat have often did so indirectly. Three important \nstudies that either directly or tangentially study the \nrelationship are Monnat (2018, 2019) and Peters \net\u00a0 al. (2019). Monnat (2018) found that income \ninequality (operationalized as the Gini coefficient) is \nassociated with higher rates of drug-related mortal-\nity. Monnat (2019), although not explicitly investi-\ngating inequality, found that more economically \ndistressed counties have higher drug-related mortal-\nity rates. Peters et\u00a0 al. (2019), whom also did not \ninvestigate inequality explicitly, found that places \nhit hardest by the prescription-opioid epidemic are \nthose that have been economically left behind. \nThese studies illustrate that economic distress plays \na significant role in the drug epidemic and that \n\u00adsupply-side factors also matter, but they only tan-\ngentially linked their findings to the unequal distri-\nbution of resources and power in society.\nThe income inequality\u2013health literature and the \naforementioned studies by Monnat (2018, 2019) \nand Peters et\u00a0al. (2019) also highlight that different \nparts of the income distribution may affect drug-\nrelated mortality more so than others. The inequality\u2013\nhealth literature has generally focused on inequality \nmeasures that quantify the concentration of income at \nthe top of the distribution or the Gini coefficient that \ntakes into account the entire distribution while giving \nless attention to the bottom of the income distribution \n(Hill and Jorgenson 2018; Pickett and Wilkinson \n2015; Wilkinson and Pickett 2010). However, the \nstudies by Monnat (2018, 2019) and Peters et\u00a0 al. \n(2019) suggest that economic distress, or the lack of \nresources going to earners at the bottom of the income \ndistribution, is driving the epidemic. Thus, whether \ninequality is associated with drug-related mortality \nmay depend on how inequality is measured.\nEven though it has been given limited attention \nin the drug epidemic literature, we argue that the \nincome inequality\u2013health paradigm can serve as a \nbridge between the demand-side and supply-side \nperspectives for four reasons. First, the resources in \na society have to be distributed in some way. When \nthey are concentrated at the top of the income distri-\nbution, the rich not only have more resources, but \nalso they have the power to influence political deci-\nsions (Cole 2018; Saez and Zucman 2019). As the \nneomaterialist perspective argues, the elite try to \nactively preserve that power by undermining social \nwelfare programs and legislation that benefit the \nworking class and poor (Hill and Jorgenson 2018; \nNeumayer and Pl\u00fcmper 2015).\nSecond, because inequality is a measure of the \ndistribution of social power, higher inequality also \nmeans the elite have more economic power relative \nto everyone else. When the elite have more power, \nworking-class people have less, which is associated \nwith a weakening of labor unions and stagnant \nwages (Piketty 2014; Saez and Zucman 2019; \nStiglitz 2012). The economic distress and alienation \nthat this creates is a key aspect of the demand-side \nparadigm (Case and Deaton 2020).\nThird, from the supply side, higher inequality is \ndirectly tied to the rise and reproduction of monop-\nolistic and oligopolistic sectors (Piketty 2014; Saez \nand Zucman 2019). In the United States, a prime \nexample is the private health care/pharmaceutical \nindustry, which the supply-side approach criticizes \nfor their role in the drug overdose epidemic. \nHowever, it should be noted that this industry also \nproduces instability and hardship for people through \nthe very existence of private health care insurance \n(Saez and Zucman 2019). If we consider health care \ninsurance to be a tax on labor (because it is essen-\ntially mandatory), it increases the effective labor \ntax rate in the United States from 29% to 37%, \nwhich disproportionately burdens the poor and \nworking class (Saez and Zucman 2019). In contrast \nto a tax levied by the government, this money goes \nprimarily to industry executives, which further \nreproduces inequality.\nFourth, Case and Deaton (2020) suggested that \nin contrast to what inequality scholars argue, people \nonly compare themselves to those in their surround-\ning community\u2014not to the elite. However, a long \nline of research suggests that people do compare \nthemselves to the elite and that they do it more so in \nhighly unequal societies (e.g., the United States), \nwhich can lead to poor health outcomes (Bourdieu \n1984; Schor 1993; Veblen 1994; Wilkinson and \nPickett 2010, 2019).\n\nThombs et al.\t\n279\nOverall, the arguments made by the demand-\nside and supply-side approaches suggest that \nincome inequality is central to the underlying argu-\nments of each perspective even though it is given \nlimited recognition by both. We suggest that both \napproaches are concerned with how the unequal \ndistribution of resources and power drives drug-\nrelated mortality and as such should not be consid-\nered antithetical. Thus, we believe a focus on \nincome inequality is key to integrating the demand-\nside and the supply-side approaches.\nGiven the discussion of the three approaches to \nthe drug overdose epidemic\u2014the demand-side per-\nspective, the supply-side perspective, and the \nincome inequality\u2013health perspective\u2014we test the \nfollowing hypotheses:\nThe Demand-Side Hypothesis: Educational \nattainment is negatively associated with \ndrug-related mortality.\nThe Supply-Side Hypothesis: The opioid pre-\nscription rate is positively associated with \ndrug-related mortality.\nThe Income\u2013Inequality Hypothesis: The income \nshare of the top 5%, top 20%, and the Gini \ncoefficient are positively associated with \ndrug-related mortality, and the income share \nof the bottom 20% is negatively associated \nwith drug-related mortality.\nData and Methods\nSample\nWe analyzed state-level annual observations for the \ntemporal period 2006 to 2017 for the 50 U.S. states \nand the District of Columbia. The time period of this \nstudy corresponds to the first year in which the opi-\noid prescription data were made available by the \nCenters for Disease Control (CDC 2019) and the \nlast year of available mortality data.\nModel Estimation Technique: Two-Level \nRandom Intercept Model\nWe used a two-level random intercept model to test \nour hypotheses. Our model nested annual state-level \nobservations within states. In total, 611 observations \n(Level 1) were nested within the 50 states and the \nDistrict of Columbia (Level 2) from 2006 to 2017. \nThe two-level random intercept model, also known \nas the within-between random-effects model \n(REWB) or the hybrid model,2 allowed us to model \nthe within and between effects for each driver of \ndrug-related mortality. The model is written as \nfollows:\ny\nx\nx\nx\nu\ne\nij\nij\nj\nj\nj\nij\n=\n+\n\u2212\n+\n+\n+\n\u2212\n\u2212\n\u03b2\n\u03b2\n\u03b2\n0\n1\n2\n(\n)\n.\nb1 represents the within effects, which are estimated \nby group mean centering the variables, and b2 is the \ngroup mean, which represents the between effects. uj \nis the Level 2 error term, and eij is the Level 1 error \nterm.3 The primary advantage of this model is that it \nallows the researcher to obtain both the within and \nbetween effects simultaneously. Doing so is not pos-\nsible in the standard fixed-effects model, which relies \non within variance only, or the standard \u00adrandom-effects \nmodel, which uses a weighted average of within and \nbetween variance (Bell, Fairbrother, and Jones 2019). \nIn the context of this study, it allowed us to test \nwhether an increase in a driver within a state had the \nsame effect as cross-sectional differences (the average \nlevel of the driver) between states.\nBefore group mean centering the variables, we \ngrand mean centered all of the independent vari-\nables to provide a meaningful interpretation of the \nintercepts. The dependent variable and indepen-\ndent variables were converted to natural loga-\nrithms, making them equivalent to elasticity \nmodels. We logged the variables to (1) correct for \nskewness and (2) because we posited that the rela-\ntionship between drug-related mortality and its \ndeterminants are multiplicative in nature; that is, \nthe determinants are not independent of one \nanother but rather, act proportionally. To correct \nfor autocorrelation, each model was estimated with \nan exponential covariance structure, which models \nan autoregressive process (Rabe-Hesketh and \nSkrondal 2012). We also estimated robust standard \nerrors to correct for heteroskedasticity.\nDrug-Related Mortality Rate  \nper 100,000 People\nOur dependent variable was the annual drug-related \nmortality rate per 100,000 people by state. We \nobtained these data from CDC WONDER\u2019s (CDC \n2018) multiple cause of death database. As defined \nby the CDC, drug-related deaths are those that are \nunintentional (ICD-10 codes X40-X44), by suicide \n(ICD-10 codes X60-X64), by homicide (ICD-10 \ncode X85), undetermined (ICD-10 codes Y10-\nY14), and all other drug-induced causes (deaths not \n\n280\t\nJournal of Health and Social Behavior 61(3) \ncategorized in any of the aforementioned ICD-10 \ncodes).\nKey Independent Variables: Educational \nAttainment, Opioid Prescription Rates, \nand Income Inequality\nOur three main variables of interest were educational \nattainment (the percentage of the state population \nwith a bachelor\u2019s degree), the opioid prescription \nrate (per 100 people), and four different measures of \nincome inequality: the income share of the (1) top \n5%, (2) the top 20%, (3) the bottom 20%, and (4) the \nGini coefficient (range = 0\u20131). Educational attain-\nment (used to test the demand-side hypothesis) and \nthe inequality measures (used to test the income \ninequality hypothesis) were obtained from the \nAmerican Community Survey (ACS) 1-year esti-\nmates (U.S. Census Bureau 2018a), and the opioid \nprescription rates (used to test the supply-side \nhypothesis) were garnered from the CDC (2019).\nEach measure of income inequality provided a \nsufficiently different approach to understanding \nhow inequality is associated with drug-related mor-\ntality. The share of income going to the top 5%, top \n20%, and bottom 20% are measures of concentra-\ntion toward the tail ends of the income distribution, \nwhereas the Gini coefficient takes into account the \nentire income distribution but does not indicate spe-\ncifically where the inequality lies within the distri-\nbution (Burns, Tomita, and Lund 2017; Hill and \nJorgenson 2018; Jorgenson, Schor, and Huang \n2017). Therefore, the Gini coefficient provided a \ngeneral measure of how unequal a distribution is, \nwhereas income shares provided more specific \nmeasures of how resources are concentrated at spe-\ncific locations along the income distribution. The \nincome inequality measures were estimated from \nincome data that were calculated as the sum of \nwages net of all other forms of income, such as \ngovernment assistance, interest, and dividends \n(U.S. Census Bureau 2018b).\nAdditional Covariates\nFollowing the social determinants of health litera-\nture (e.g., Monnat 2018; Solar and Irwin 2010), we \nincluded additional covariates that controlled for the \nstructure of each state\u2019s economy and potential \nregional differences that could be driving the epi-\ndemic. Our models included two trend terms, one \nthat was centered on the year 2006 and a second one \nthat was the quadratic trend term, which we deemed \nappropriate based on the Bayesian information \ncriterion (BIC) statistic. In addition, we included the \nannual state-level median household income in \n2017 inflation-adjusted dollars and the percentage \nof the labor force in manufacturing, which con-\ntrolled for the affluence and structure of the econ-\nomy for each state, respectively. To control for \nregional differences, we included a set of indicator \nvariables denoting census region (1 = Northeast, \nreference group; 2 = Midwest; 3 = South; 4 = West). \nThe median household income data and the percent-\nage of the labor force in manufacturing were gath-\nered from the ACS 1-year estimates (U.S. Census \nBureau 2018a). The census regions followed the \nU.S. Census Bureau\u2019s (2015) categorization. We \nreport sensitivity analyses with additional covari-\nates in the Sensitivity Analysis section following the \nresults.\nResults\nDescriptive Statistics: Where Is the \nVariance?\nThe univariate, nonlogged descriptive statistics for \nthe dependent and independent variables are \nreported in Table 1. The table includes the mean and \nthe overall, within, and between standard deviation \n(SD) for each variable. For all of the variables, the \nvariance is greater between states rather than within \nthem over time. Particularly notable is the mean of \nthe drug-related mortality rate, which is 15.98 \ndeaths per 100,000 people. The mortality rate varies \nsubstantially within states (within SD = 4.80) and \nbetween states (between SD = 4.98). Figure 1 illus-\ntrates the average drug-related mortality rate from \n2006 to 2017, and Figure 2 shows the change in the \nmortality rate over the same period.\nWe also found that much of the variance for our \nthree main independent variables of interest (the \npercentage of the population with a bachelor\u2019s \ndegree, the opioid prescription rate, and income \ninequality) is between states rather than within \nthem. The mean percentage of people with a bache-\nlor\u2019s degree is 28.84 (between SD = 5.86%; within \nSD = 1.63%), and the mean opioid prescription rate \nis 79.41 opioids per 100 people (between SD = \n21.88; within SD = 8.14). Regarding income \ninequality, the mean share of income going to the \ntop 5% is 21.50% (between SD = 1.37%; within \nSD\u00a0= .68%), and the mean Gini coefficient is .46 \n(between SD = .02; within SD = .01). The average \nshare of income going to the top 20% is 49.37% \n(between SD = 1.86%; within SD = .70%), whereas \nthe mean share of income going to the bottom 20% \nis 3.46% (between SD = .41%; within SD = .15%).\n\nThombs et al.\t\n281\nRandom Intercept Model Results\nWe first estimated a model that included all of the \nwithin and between effects for each independent \nvariable (not reported here) and used Wald tests \n(Table 2) to determine whether the between and \nwithin effects for the continuous variables are statis-\ntically different from one another. Of the five con-\ntinuous independent variables included in the \nmodel, only the opioid prescription rate\u2019s within and \nbetween effects are statistically different at the .05 \nTable 1.\u2002 Descriptive Statistics: CDC Wonder and American Community Survey 1-Year Estimates, \n2006 to 2017.\nMeasurement\nMean (SD)\nBetween SD\nWithin SD\nn\nDrug-related \nmortality\nPer 100,000 \npeople\n15.98\n(6.89)\n4.98\n4.80\n612\n% Bachelor\u2019s degree\n%\n28.84\n(6.03)\n5.86\n1.63\n612\nOpioid prescription \nrate\nPer 100 people\n79.41\n(23.16)\n21.88\n8.14\n612\nTop 5%\n%\n21.50\n(1.52)\n1.37\n.68\n612\nTop 20%\n%\n49.37\n(1.97)\n1.86\n.70\n612\nBottom 20%\n%\n3.46\n(.43)\n.41\n.15\n612\nGini\nIndex\n.46\n(.02)\n.02\n.01\n612\nMedian household \nincome\n2017 inflation-\nadjusted $\n59,386.58\n(11,563.06)\n9,705.46\n6,419.06\n612\nManufacturing\n% of labor force\n10.22\n(3.98)\n3.99\n.57\n611\nFigure 1.\u2002 Average Drug-Related Mortality Rate, 2006 to 2017.\n\n282\t\nJournal of Health and Social Behavior 61(3) \nlevel. As such, we included the within and between \neffects for the opioid prescription rate in the models \nand report the random effects (because they are \nmore efficient than the within effects) for the other \nvariables.\nTable 3 reports the results of the two-level ran-\ndom intercept models for drug-related mortality by \nincome inequality measure, and Figure 3 visually \npresents the point estimates and confidence inter-\nvals for each key variable (the percentage of people \nwith a bachelor\u2019s degree, the opioid prescription \nrate, and each inequality measure). All of the two-\nlevel intercepts are statistically significant, indicat-\ning that the two-level model fits the data well and is \nsuperior to the linear model. The percentage of \npeople with a bachelor\u2019s degree is not statistically \nsignificant at the .05 level in any of the models, \nwhich does not support the demand-side hypothe-\nsis. The opioid prescription rate\u2019s within effects are \nnot statistically significant in any model, but all of \nthe between effects are, which supports the supply-\nside hypothesis. In other words, states with a higher \nrate of opioid prescriptions, on average, have a \nhigher drug-related mortality rate.\nThe results for the income inequality measures \nindicate that inequality has a complex association \nwith drug-related mortality. In contrast to what \nthe income inequality hypothesis expects, the \nshare of income of the top 5%, the top 20%, and \nGini coefficient are not associated with drug-\nrelated mortality, but the share of income of the \nbottom 20% is negatively associated with \nFigure 2.\u2002 Percentage Change in the Drug-Related Mortality Rate from 2006 to 2017.\nTable 2.\u2002 Wald Tests of the within and between Effects: CDC WONDER and American Community \nSurvey 1-Year Estimates, 2006 to 2017.\n(1)\nTop 5%\n(2)\nTop 20%\n(3)\nBottom 20%\n(4)\nGini\n% Bachelor\u2019s degree\n.19\n.03\n.05\n.02\nOpioid prescription rate\n22.15*\n22.43*\n22.66*\n22.53*\nIncome inequality\n1.65\n2.51\n3.24\n2.45\nMedian household income\n.50\n1.21\n1.70\n1.30\nManufacturing\n.21\n.16\n.00\n.12\nNote: The test statistics are \u03c72 values.\n*p < .05. H0: The coefficients are equivalent.\n\nThombs et al.\t\n283\ndrug-related mortality. Moreover, the values of \nthe BIC statistic for each model indicate that the \nbottom 20% model best fits the data. Based on \nRaftery\u2019s (1995) grades of evidence,4 there is \n\u201cstrong\u201d evidence that the bottom 20% model is \nbetter than the top 5% model and \u201cpositive\u201d evi-\ndence that it is better than the top 20% model and \nthe Gini model. Overall, these results indicate that \nthe lack of resources going to the bottom 20% of \nearners best explains the income inequality\u2013 \ndrug-related mortality relationship\u2014rather than \nthe concentration of resources at the top of the \nincome distribution.\nSensitivity Analysis\nAs a sensitivity analysis, we estimated models with \ntwo additional covariates that may be associated \nwith the drug epidemic (Ariizumi and Schirle 2012; \nCase and Deaton 2015; Ruhm 2005): (1) the per-\ncentage of the population that is white and (2) a \ndichotomous variable corresponding to the years of \nTable 3.\u2002 Two-Level Random Intercept Model for the Regression of the Drug-Related Mortality Rate, \n2006 to 2017 (N = 611): CDC WONDER and American Community Survey 1-Year Estimates.\n(1)\n(2)\n(3)\n(4)\n\u2002\nTop 5%\nTop 20%\nBottom 20%\nGini\n% Bachelor\u2019s degree\n.49\n(.26)\n.49\n(.26)\n.50\n(.26)\n.49\n(.26)\nOpioid prescription rate, \nwithin\n\u2212.05\n(.16)\n\u2212.05\n(.16)\n\u2212.04\n(.16)\n\u2212.05\n(.16)\nOpioid prescription rate, \nbetween\n1.15*\n(.24)\n1.17*\n(.24)\n1.23*\n(.24)\n1.18*\n(.24)\nIncome inequality\n.18\n(.22)\n.89\n(.63)\n\u2212.48*\n(.23)\n.83\n(.53)\nMedian household income\n.11\n(.07)\n.11\n(.07)\n.12\n(.07)\n.12\n(.07)\nManufacturing\n\u2212.27*\n(.11)\n\u2212.27*\n(.11)\n\u2212.25*\n(.10)\n\u2212.27*\n(.11)\nMidwest\n\u2212.32*\n(.12)\n\u2212.30*\n(.12)\n\u2212.28*\n(.11)\n\u2212.30*\n(.11)\nSouth\n\u2212.33*\n(.08)\n\u2212.35*\n(.08)\n\u2212.36*\n(.08)\n\u2212.35*\n(.08)\nWest\n\u2212.25*\n(.10)\n\u2212.23*\n(.10)\n\u2212.20*\n(.09)\n\u2212.22*\n(.10)\nTime\n\u2212.02\n(.02)\n\u2212.02\n(.02)\n\u2212.03\n(.02)\n\u2212.03\n(.02)\nTime2\n.01*\n(.00)\n.01*\n(.00)\n.01*\n(.00)\n.01*\n(.00)\nConstant\n3.04*\n(.70)\n3.05*\n(.69)\n3.02*\n(.70)\n3.04*\n(.69)\nRandom components\n\u2003 State-level intercept\n.04*\n(.01)\n.04*\n(.01)\n.04*\n(.01)\n.04*\n(.01)\n\u2003 Year intercept\n.05*\n(.01)\n.05*\n(.01)\n.05*\n(.01)\n.05*\n(.01)\n\u2003 Rho\n.76*\n(.06)\n.76*\n(.06)\n.76*\n(.06)\n.76*\n(.06)\n\u2003 BIC\n\u2212450.18\n\u2212452.51\n\u2212457.66\n\u2212453.18\nNote: Robust standard errors are in parentheses. The models are estimated with an exponential covariance structure \nto correct for autocorrelation. Census region reference group = Northeast. BIC = Bayesian information criterion.\n*p < .05.\n\n284\t\nJournal of Health and Social Behavior 61(3) \nthe Great Recession (1 = the years 2008 and 2009; \n0\u00a0= otherwise). Measures of the percentage of the \npopulation that is white were obtained from the \nACS 1-year estimates (U.S. Census Bureau 2018a), \nand the recessionary years were coded according to \nthe National Bureau of Economic Research (2010).5 \nThe results are reported in Appendix A available in \nthe online version of the article.\nThe inclusion of the additional variables makes \nthe bottom 20% slope coefficient only marginally \nstatistically significant (p = .06). However, neither \nthe percentage of the population that is white nor \nthe Great Recession dichotomous variable are sta-\ntistically significant in the models. Additionally, the \nmodel\u2019s BIC values are significantly larger than \nthe ones reported in the aforementioned results. \nThe difference between the BIC value for the bot-\ntom 20% model in the main results compared to the \nsensitivity analysis is 11.19, indicating \u201cvery \nstrong\u201d evidence that the main results fit the data \nbetter than the sensitivity analysis. Overall, these \nfindings suggest that the inclusion of the additional \nvariables is unwarranted and that their addition \nleads to inflated standard errors.\nAt a reviewer\u2019s request, we also interacted the \nGreat Recession variable with our main variables of \ninterest\u2014the percentage of people with a bachelor\u2019s \ndegree, the opioid prescription rate (between effect), \nand each inequality measure. The results are reported \nin the Appendix B available in the online version of \nthe article. The interaction is not statistically signifi-\ncant for the percentage of people with a bachelor\u2019s \ndegree or the opioid prescription rate, but it is for all \nof the inequality measures except for the share of the \ntop 5%.6 The results suggest that none of the \ninequality measures had an effect on drug-related \nmortality during the Great Recession but did so dur-\ning nonrecessionary years. The slope coefficient of \nthe top 20% during the Great Recession was \u2013.28 \n(p\u00a0= .77), compared to its nonrecession slope coeffi-\ncient of 1.13 (p = .04). The slope coefficient of the \nbottom 20% during the Great Recession was \u2013.08 \n(p\u00a0= .74), compared to its nonrecession slope coeffi-\ncient of \u2013.58 (p = .02), and the Gini\u2019s Great \nRecession slope coefficient was \u2013.10 (p = .75), com-\npared to its nonrecession slope coefficient of 1.06 \n(p\u00a0= .46). These findings align with prior research \nshowing that some population health characteristics \nFigure 3.\u2002 Key Results by Income Inequality Model.\nNote: The circle represents the point estimate, and the black bars correspond to the 95% confidence interval.\n\nThombs et al.\t\n285\nimprove during recessionary periods (Ariizumi and \nSchirle 2012; Ruhm 2000, 2005). Regardless, the \nprimary results reported in this study appear to be \nmore robust compared to alternative specifications \nand both opioid prescription rates and the income \nshare of the bottom 20% of earners are drivers of the \ndrug overdose epidemic.\nDiscussion\nThe results of this study have important implica-\ntions for understanding the drivers of the U.S. drug \noverdose epidemic and the policies needed to com-\nbat it. Regarding the drivers of drug-related mortal-\nity, the findings support the arguments made by the \nsupply-side perspective more so than the demand-\nside perspective. As we discussed, the supply-side \nperspective argues that changes in the drug environ-\nment are driving the epidemic. For the licit drug \nmarket, prior research points to the exploitive prac-\ntices of pharmaceutical companies that made opi-\noids more available to the general public (Ruhm \n2019). Even though the prescribing rates of licit opi-\noids have declined during the past several years, the \nillicit drug market has picked up where it left off by \nmaking fentanyl and heroin increasingly cheap and \nwidely available. Undoubtedly, the increase in \ncheap opioids (and other drugs) plays a significant \nrole in the epidemic.\nHowever, the arguments made by the supply-\nside approach have limited explanatory power \nwhen it comes to the underlying mechanism of why \nthe epidemic emerged. For example, the supply-\nside argument does not provide an explanation for \nwhat is driving people to use drugs, and it does not \noffer a compelling reason for why the epidemic is \ndisproportionately affecting white (and increas-\ningly black) working-class people without a college \ndegree (Case and Deaton 2020). Although the \ndemand side\u2019s emphasis on educational attainment \nis not supported in this study, we contend that \nincome inequality can act as a link between the \narguments made by the demand-side and the sup-\nply-side approaches. On one hand, increasing \ninequality both produces and is reproduced by the \nhealth care and pharmaceutical industries at the \ncenter of the epidemic. Americans are reporting \nhigher levels of pain than ever before, but the \nreported increase appears to be largely from social \nand economic distress rather than from physical ail-\nments (Case and Deaton 2020). Beginning in the \n1990s, the pharmaceutical industry quickly \u201cphar-\nmaceuticalized\u201d this phenomenon by treating it as a \nmedical issue to be addressed by pharmaceuticals \nand drugs rather than by policy interventions or \nsocial change (Abraham 2010; Bell and Figert \n2012). On the other hand, income inequality is also \nan important demand-side factor because it concen-\ntrates resources and power among a small number \nof elites, which increases alienation and under-\nmines the well-being of lower income groups \n(Neumayer and Pl\u00fcmper 2015; Wilkinson and \nPickett 2019). As we show, the best inequality pre-\ndictor of drug-related mortality is the lack of \nresources going to the poorest 20% of earners rela-\ntive to everyone else.\nBuilding on this discussion, the results indicate \nthat health policy should take on a wider set of mea-\nsures to combat the drug overdose epidemic. The \nactions taken by federal and state governments in \nthe United States have primarily focused on the \npractices of pharmaceutical companies and pre-\nscribing physicians (Gross and Gordon 2019). \nAlthough we find that addressing the supply side of \nthe epidemic is necessary\u2014it is an inadequate pre-\nvention response by itself (also see Monnat 2019; \nPeters et\u00a0al. 2019). Policymakers must also address \nstructural factors like economic inequality, which \nwill require implementing policies that redistribute \nincome and resources. Although wealth is not \nexplicitly discussed in this article, implementing a \nwealth tax may be an even more effective strategy \nbecause the wealthiest often structure their assets so \nthey have relatively low levels of taxable income \n(Saez and Zucman 2019). Furthermore, eliminating \nprivate insurance and moving to a single-payer sys-\ntem could potentially not only check the power of \npharmaceutical companies and limit the harmful \nprescribing practices of physicians but also serve as \na large pay increase for employees. Today, an \nemployee\u2019s health care costs an average of $13,000 \na year, and a single-payer system funded through \nprogressive taxation would shift the burden of pay-\ning for health care from workers to the rich (Saez \nand Zucman 2019).\nAlthough the findings for this study are rela-\ntively robust, they should be interpreted with two \nlimitations in mind. First, due to the availability of \nopioid prescription rates, our analysis dates back \nto 2006. However, as Case and Deaton (2015) \nshowed, drug-related mortality has been on the rise \nsince the beginning of the twenty-first century. This \nlonger-term trend is not captured in the present \nstudy. Second, there may also be important spatial \ndifferences within states, such as at the county or \ncity levels, that are not captured in this state-level \nstudy. Future research should evaluate whether this \nis the case.\n\n286\t\nJournal of Health and Social Behavior 61(3) \nWe conclude by reiterating that drug-related mor-\ntality is likely a deleterious \u201cdownstream\u201d conse-\nquence of changes in the U.S. economy over the past \nhalf-century that have led to increased income \ninequality and an exploitive private health care and \npharmaceutical industry. Our emphasis on the role of \nincome inequality in the current drug overdose epi-\ndemic is consistent with recent research demonstrat-\ning that growing income inequality is a key \ndeterminant of other health-related outcomes, includ-\ning overall life expectancy, crime, and mental illness \nas well as anthropogenic greenhouse gas emissions \nthat cause climate change and exacerbate air pollu-\ntion\u2019s impact on public health (Hill, Jorgenson et\u00a0al. \n2019; Hill and Jorgenson 2018; Jorgenson et\u00a0al. 2016, \n2017, 2020; Knight, Schor, and Jorgenson 2017; \nPickett and Wilkinson 2015). Therefore, addressing \nthe supply side of the drug overdose epidemic is cer-\ntainly warranted, but taking a more structural per-\nspective to the epidemic that involves reducing \nincome inequality would likely not only lead to \nreduced drug-related mortality but also have positive \neconomic and environmental benefits.\nAcknowledgments\nWe would like to thank Amy Burdette, editor of JHSB, \nand the three anonymous reviewers for their helpful \ncomments.\nNotes\n1.\t\nCase and Deaton\u2019s emphasis on education aligns \nwith sociological work that explains why educa-\ntion is important to health (e.g., Mirowsky and \nRoss 2015). In the postindustrial economy, educa-\ntional attainment is critical to accessing well-paying \njobs (Case and Deaton 2020). Mirowsky and Ross \n(2015:297) also argued that education is important \nfor individuals to be able to overcome the unhealthy \n\u201cdefault American lifestyle.\u201d\n2.\t\nSome researchers refer to it as the hybrid model, but \nthis term is misleading because it is a random-effects \nmodel\u2014not a combination of the fixed-effects and \nrandom-effects models (Bell and Jones 2015).\n3.\t\nAlthough the within effects are not subject to constant, \ntime-invariant unobserved heterogeneity bias, the \nbetween effects can be biased if time-invariant (Level \n2) variables are omitted from the model. However, the \nwithin effects are still subject to time-variant omitted \nvariable bias (also see Hill, Davis, et\u00a0al. 2019).\n4.\t\nRaftery\u2019s (1995) grades of evidence provide a way \nto compare how well different, nonnested models fit \nthe data.\n5.\t\nTechnically, the Great Recession began in December \n2007, but we code 2007 as a 0 because it is the last \nmonth of the year.\n6.\t\nThe percentage of the population that is white is \nexcluded because the BIC values of the models are \nsignificantly smaller when it is removed from the \nanalysis, and it is not statistically significant in any \nof the analyses.\nSupplemental Material\nAppendices A and B are available in the online version of \nthe article.\nReferences\nAbraham, John. 2010. \u201cPharmaceuticalization of Society \nin Context: Theoretical, Empirical and Health \nDimensions.\u201d Sociology 44(4):603\u201322. doi:10.1177/ \n0038038510369368.\nAriizumi, Hideki, and Tammy Schirle. 2012. \u201cAre \nRecessions Really Good for Your Health? Evidence \nfrom Canada.\u201d Social Science & Medicine 74(8):1224\u2013\n31. doi:10.1016/j.socscimed.2011.12.038.\nBell, Andrew, Malcolm Fairbrother, and Kelvyn \nJones. 2019. \u201cFixed and Random Effects Models: \nMaking an Informed Choice.\u201d Quality & Quantity \n53(2):1051\u201374. doi:10.1007/s11135-018-0802-x.\nBell, Andrew, and Kelvyn Jones. 2015. \u201cExplaining Fixed \nEffects: Random Effects Modeling of Time-Series \nCross-Sectional and Panel Data.\u201d Political Science \nResearch and Methods 3(1):133\u201353. doi:10.1017/\npsrm.2014.7.\nBell, Susan E., and Anne E. Figert. 2012. \u201cMedicalization \nand Pharmaceuticalization at the Intersections: \nLooking Backward, Sideways and Forward.\u201d Social \nScience & Medicine 75(5):775\u201383. doi:10.1016/j.\nsocscimed.2012.04.002.\nBourdieu, Pierre. 1984. Distinction: A Social Critique of \nthe Judgement of Taste. Cambridge, MA: Harvard \nUniversity Press.\nBuot, Max-Louis G., Jeffrey P. Docena, Brenda K. \nRatemo, Matthew J. Bittner, Jacob T. Burlew, Aziz R. \nNuritdinov, and Jennifer R. Robbins. 2014. \u201cBeyond \nRace and Place: Distal Sociological Determinants \nof HIV Disparities.\u201d PLoS ONE 9(4):e91711. \ndoi:10.1371/journal.pone.0091711.\nBurns, Jonathan K., Andrew Tomita, and Crick Lund. \n2017. \u201cIncome Inequality Widens the Existing \nIncome-Related Disparity in Depression Risk in Post-\nApartheid South Africa: Evidence from a Nationally \nRepresentative Panel Study.\u201d Health & Place 45:10\u2013\n16. doi:10.1016/j.healthplace.2017.02.005.\nCase, Anne, and Angus Deaton. 2015. \u201cRising Morbidity \nand Mortality in Midlife among White Non-Hispanic \nAmericans in the 21st Century.\u201d Proceedings of the \nNational Academy of Sciences 112(49):15078\u201383. \ndoi:10.1073/pnas.1518393112.\nCase, Anne, and Angus Deaton. 2017. \u201cMortality and \nMorbidity in the 21st Century.\u201d. https://www \n.brookings.edu/wp-content/uploads/2017/08/ \ncasetextsp17bpea.pdf.\n\nThombs et al.\t\n287\nCase, Anne, and Angus Deaton. 2020. Deaths of Despair \nand the Future of Capitalism. Princeton, NJ: \nPrinceton Press.\nCenters for Disease Control. 2018. \u201cCDC WONDER.\u201d \nhttps://wonder.cdc.gov/.\nCenters for Disease Control. 2019. \u201cU.S. State Prescribing \nRates.\u201d \nhttps://www.cdc.gov/drugoverdose/maps/\nrxstate2017.html.\nCharles, Kerwin Kofi, Erik Hurst, and Mariel Schwartz. \n2018. \u201cThe Transformation of Manufacturing and the \nDecline in U.S. Employment.\u201d National Bureau of \nEconomic Research, Working Paper 24468. https://\nwww.nber.org/papers/w24468.pdf.\nCole, Wade M. 2018. \u201cPoor and Powerless: Economic and \nPolitical Inequality in Cross-National Perspective, \n1981\u20132011.\u201d International Sociology 33(3):357\u201385. \ndoi:10.1177/0268580918760430.\nCurran, Michaela, and Matthew C. Mahutga. 2018. \n\u201cIncome Inequality and Population Health: A Global \nGradient?\u201d Journal of Health and Social Behavior \n59(4):536\u201353. doi:10.1177/0022146518808028.\nDaly, Martin. 2016. Killing the Competition: Economic \nInequality and Homicide. New Brunswick, NJ: \nTransaction.\nDasgupta, Nabarun, Leo Beletsky, and Daniel Ciccarone. \n2017. \u201cOpioid Crisis: No Easy Fix to Its Social \nand Economic Determinants.\u201d American Journal \nof Public Health 108(2):182\u201386. doi:10.2105/\nAJPH.2017.304187.\nDesilver, Drew. 2017. \u201cMost Americans Unaware That \nas U.S. Manufacturing Jobs Have Disappeared, \nOutput Has Grown.\u201d https://www.pewresearch.org/\nfact-tank/2017/07/25/most-americans-unaware-that-\nas-u-s-manufacturing-jobs-have-disappeared-out-\nput-has-grown/#:~:text=Most%20Americans%20\nunaware%20that%20as,have%20disappeared%2C%20\noutput%20has%20grown&text=Manufacturing%20\njobs%20in%20the%20United,the%20U.S.%20\n%E2%80%93%20has%20grown%20strongly.\nDrug Enforcement Agency. 2016. \u201cNational Heroin \nThreat Assessment Summary \u2013 Updated.\u201d https://\nwww.dea.gov/sites/default/files/2018-07/hq062716_\nattach.pdf.\nGomes, Tara, Mina Tadrous, Muhammad M. Mamdani, J. \nMichael Paterson, and David N. Juurlink. 2018. \u201cThe \nBurden of Opioid-Related Mortality in the United \nStates.\u201d JAMA Network Open 1(2). doi:10.1001/\njamanetworkopen.2018.0217.\nGross, Jacob, and Debra B. Gordon. 2019. \u201cThe \nStrengths and Weaknesses of Current US Policy to \nAddress Pain.\u201d American Journal of Public Health \n109(1):66\u201372. doi:10.2105/AJPH.2018.304746.\nHadland, Scott E., Magdalena Cerd\u00e1, Yu Li, Maxwell \nS. Krieger, and Brandon D. L. Marshall. 2018. \n\u201cAssociation of Pharmaceutical Industry Marketing of \nOpioid Products to Physicians with Subsequent Opioid \nPrescribing.\u201d JAMA Internal Medicine 178(6):861\u2013\n63. doi:10.1001/jamainternmed.2018.1999.\nHedegaard, Holly, Arialdi M. Mini\u00f1o, and Margaret \nWarner. 2018. \u201cDrug Overdose Deaths in the United \nStates, 1999\u20132017.\u201d NCHS Data Brief (329):1\u20138. \nhttps://www.cdc.gov/nchs/data/databriefs/db329-h.\npdf.\nHill, Terrence D., Andrew P. Davis, J. Micah Roos, \nand Michael T. French. 2019. \u201cLimitations of \nFixed-Effects Models for Panel Data.\u201d Sociological \nPerspectives. doi:10.1177/0731121419863785.\nHill, Terrence D., and Andrew Jorgenson. 2018. \u201cBring \nout Your Dead!: A Study of Income Inequality and \nLife Expectancy in the United States, 2000\u20132010.\u201d \nHealth \n& \nPlace \n49:1\u20136. \ndoi:10.1016/j.health \nplace.2017.11.001.\nHill, Terrence D., Andrew Jorgenson, Peter Ore, Kelly \nBalistreri, and Brett Clark. 2019. \u201cAir Quality and \nLife Expectancy in the United States: An Analysis \nof the Moderating Effect of Income Inequality.\u201d \nSSM \u2013 Population Health 7:100346. doi:10.1016/j.\nssmph.2018.100346.\nJones, Christopher M., Emily B. Einstein, and Wilson \nM. Compton. 2018. \u201cChanges in Synthetic Opioid \nInvolvement in Drug Overdose Deaths in the United \nStates, \n2010\u20132016.\u201d \nJAMA \n319(17):1819\u201321. \ndoi:10.1001/jama.2018.2844.\nJorgenson, Andrew, Terrence Hill, Brett Clark, Ryan \nThombs, Peter Ore, Kelly Balistreri, and Jennifer \nGivens. 2020. \u201cPower, Proximity, and Physiology: Does \nIncome Inequality and Racial Composition Amplify \nthe Impacts of Air Pollution on Life Expectancy in \nthe United States?\u201d Environmental Research Letters \n15:024013. doi:10.1088/1748-9326/ab6789.\nJorgenson, Andrew, Juliet Schor, and Xiaorui Huang. \n2017. \u201cIncome Inequality and Carbon Emissions \nin the United States: A State-Level Analysis, \n1997\u20132012.\u201d Ecological Economics 134:40\u201348. \ndoi:10.1016/j.ecolecon.2016.12.016.\nJorgenson, Andrew, Juliet Schor, Kyle Knight, and \nXiaorui Huang. 2016. \u201cDomestic Inequality and \nCarbon Emissions in Comparative Perspective.\u201d \nSociological Forum 31(S1):770\u201386. doi:10.1111/\nsocf.12272.\nKariisa, Mbabazi, Lawrence Scholl, Nana Wilson, Puja \nSeth, and Brooke Hoots. 2019. \u201cDrug Overdose \nDeaths Involving Cocaine and Psychostimulants \nwith Abuse Potential\u2014United States, 2003\u20132017.\u201d \nMorbidity and Mortality Weekly Report 68(17): \n388\u201395. https://www.cdc.gov/mmwr/volumes/68/\nwr/mm6817a3.htm.\nKawachi, Ichiro., Bruce. P. Kennedy, Kimberly \nLochner, and Deborah Prothrow-Stith. 1997. \n\u201cSocial Capital, Income Inequality, and Mortality.\u201d \nAmerican Journal of Public Health 87(9):1491\u201398. \ndoi:10.2105/AJPH.87.9.1491.\nKnight, Kyle W., Juliet B. Schor, and Andrew K. \nJorgenson. 2017. \u201cWealth Inequality and Carbon \nEmissions in High-Income Countries.\u201d Social Currents \n4(5):403\u201312. doi:10.1177/2329496517704872.\n\n288\t\nJournal of Health and Social Behavior 61(3) \nLin, Dajun, Siying Liu, and Christopher J. Ruhm. 2020. \n\u201cOpioid Deaths and Local Healthcare Intensity: A \nLongitudinal Analysis of the U.S. Population, 2003\u2013\n2014.\u201d American Journal of Preventive Medicine \n58(1):50\u201358. doi:10.1016/j.amepre.2019.09.008.\nMakary, Martin A., Heidi N. Overton, and Peiqi Wang. \n2017. \u201cOverprescribing Is Major Contributor to \nOpioid Crisis.\u201d BMJ 359:j4792. doi:10.1136/bmj.\nj4792.\nMcCall Jones, Christopher, Grant T. Baldwin, and \nWilson\u00a0 M. Compton. 2017. \u201cRecent Increases in \nCocaine-Related Overdose Deaths and the Role \nof Opioids.\u201d American Journal of Public Health \n107(3):430\u201332. doi:10.2105/AJPH.2016.303627.\nMirowsky, John, and Catherine E. Ross. 2015. \n\u201cEducation, Health, and the Default American \nLifestyle.\u201d Journal of Health and Social Behavior \n56(3):297\u2013306. doi:10.1177/0022146515594814.\nMonnat, Shannon M. 2018. \u201cFactors Associated \nwith County-Level Differences in U.S. Drug-\nRelated Mortality Rates.\u201d American Journal of \nPreventive Medicine 54(5):611\u201319. doi:10.1016/j \n.amepre.2018.01.040.\nMonnat, Shannon M. 2019. \u201cThe Contributions of \nSocioeconomic and Opioid Supply Factors to U.S. \nDrug Mortality Rates: Urban-Rural and Within-\nRural Differences.\u201d Journal of Rural Studies 68:319\u2013\n35. doi:10.1016/j.jrurstud.2018.12.004.\nNational Bureau of Economic Research. 2010. \u201cUS \nBusiness Cycle Expansions and Contractions.\u201d \nhttps://www.nber.org/cycles.html.\nNeumayer, Eric, and Thomas Pl\u00fcmper. 2015. \u201cInequalities \nof Income and Inequalities of Longevity: A Cross-\nCountry Study.\u201d American Journal of Public Health \n106(1):160\u201365. doi:10.2105/AJPH.2015.302849.\nPaulozzi, Leonard J., Karin A. Mack, and Jason M. \nHockenberry. 2014. \u201cVital Signs: Variation among \nStates in Prescribing of Opioid Pain Relievers and \nBenzodiazepines\u2014United States, 2012.\u201d Morbidity \nand Mortality Weekly Report 63(26):563\u201368. \nhttps://www.cdc.gov/mmwr/preview/mmwrhtml/\nmm6326a2.htm.\nPeters, David J., Shannon M. Monnat, Andrew L. \nHochstetler, and Mark T. Berg. 2019. \u201cThe Opioid \nHydra: Understanding Overdose Mortality Epidemics \nand Syndemics across the Rural-Urban Continuum.\u201d \nRural Sociology. doi:10.1111/ruso.12307.\nPezalla, Edmund, D. Rosen, Jg Erensen, David Haddox, \nand Tj Mayne. 2017. \u201cSecular Trends in Opioid \nPrescribing in the USA.\u201d Journal of Pain Research \n10:383\u201387. doi:10.2147/JPR.S129553.\nPickett, Kate E., and Richard G. Wilkinson. 2015. \u201cIncome \nInequality and Health: A Causal Review.\u201d Social \nScience & Medicine 128:316\u201326. doi:10.1016/j.socs \ncimed.2014.12.031.\nPierce, Justin R., and Peter K. Schott. 2020. \u201cTrade \nLiberalization and Mortality: Evidence from US \nCounties.\u201d American Economic Review: Insights \n2(1):47\u201364. doi:10.3386/w22849.\nPiketty, Thomas. 2014. Capital in the Twenty-First \nCentury. Cambridge, MA: The Belknap Press of \nHarvard University Press.\nRabe-Hesketh, Sophia, and Anders Skrondal. 2012. \nMultilevel and Longtiudinal Modeling Using Stata. \n3rd ed. Vol. 1. College Station, TX: Stata Press.\nRaftery, Adrian E. 1995. \u201cBayesian Model Selection \nin Social Research.\u201d Sociological Methodology \n25:111\u201363. doi:10.2307/271063.\nRibeiro, Wagner Silva, Annette Bauer, M\u00e1rio C\u00e9sar \nRezende Andrade, Marianna York-Smith, Pedro \nMario Pan, Luca Pingani, Martin Knapp, Evandro \nSilva Freire Coutinho, and Sara Evans-Lacko. 2017. \n\u201cIncome Inequality and Mental Illness-Related \nMorbidity and Resilience: A Systematic Review and \nMeta-analysis.\u201d The Lancet Psychiatry 4(7):554\u201362. \ndoi:10.1016/S2215-0366(17)30159-1.\nRuhm, Christopher J. 2000. \u201cAre Recessions Good for \nYour Health?\u201d The Quarterly Journal of Economics \n115(2):617\u201350. doi:10.1162/003355300554872.\nRuhm, Christopher J. 2005. \u201cHealthy Living in Hard \nTimes.\u201d Journal of Health Economics 24(2):341\u201363. \ndoi:10.1016/j.jhealeco.2004.09.007.\nRuhm, Christopher J. 2019. \u201cDrivers of the Fatal Drug \nEpidemic.\u201d Journal of Health Economics 64:25\u201342. \ndoi:10.1016/j.jhealeco.2019.01.001.\nSaez, Emmanuel, and Gabriel Zucman. 2019. The \nTriumph of Injustice: How the Rich Dodge Taxes and \nHow to Make Them Pay. New York: Norton.\nScholl, Lawrence, Puja Seth, Mbabazi Kariisa, Nana \nWilson, and Grant Baldwin. 2019. \u201cDrug and \nOpioid-Involved Overdose Deaths\u2014United States, \n2013\u20132017.\u201d MMWR. Morbidity and Mortality \nWeekly Report 67 (5152):1419\u20131427. https://www \n.cdc.gov/mmwr/volumes/67/wr/mm675152e1.htm.\nSchor, Juliet B. 1993. The Overworked American. New \nYork, NY: Basic Books.\nSinghal, Astha, Yu-Yu Tien, and Renee Y. Hsia. 2016. \n\u201cRacial-Ethnic Disparities in Opioid Prescriptions \nat Emergency Department Visits for Conditions \nCommonly Associated with Prescription Drug \nAbuse.\u201d PLoS ONE 11(8):e0159224. doi:10.1371/\njournal.pone.0159224.\nSolar, Orielle, and Alec Irwin. 2010. \u201cA Conceptual \nFramework for Action on the Social Determinants of \nHealth: Debates, Policy & Practice, Case Studies.\u201d \nhttps://www.who.int/sdhconference/resources/\nConceptualframeworkforactiononSDH_eng.pdf.\nStiglitz, Joseph. 2012. The Price of Inequality. New \nYork: W.W. Norton & Company.\nSuzuki, Joji, and Saria El-Haddad. 2017. \u201cA Review: \nFentanyl and Non-pharmaceutical Fentanyls.\u201d Drug \nand Alcohol Dependence 171:107\u201316. doi:10.1016/j \n.drugalcdep.2016.11.033.\nTorre, Roberta, and Mikko Myrskyl\u00e4. 2014. \u201cIncome \nInequality and Population Health: An Analysis of \nPanel Data for 21 Developed Countries, 1975\u20132006.\u201d \nPopulation Studies 68(1):1\u201313. doi:10.1080/003247\n28.2013.856457.\n\nThombs et al.\t\n289\nTruesdale, Beth C., and Christopher Jencks. 2016. \u201cThe \nHealth Effects of Income Inequality: Averages \nand Disparities.\u201d Annual Review of Public Health \n37(1):413\u201330. doi:10.1146/annurev-publhealth-032 \n315-021606.\nU.S. Census Bureau. 2015. \u201cCensus Regions and \nDivisions of the United States.\u201d https://www2 \n.census.gov/geo/pdfs/maps-data/maps/reference/\nus_regdiv.pdf.\nU.S. Census Bureau. 2018a. \u201cAmerican Community \nSurvey: Data Profiles.\u201d https://www.census.gov/acs/\nwww/data/data-tables-and-tools/data-profiles/2016/.\nU.S. Census Bureau. 2018b. \u201cAmerican Community \nSurvey and Puerto Rico Community Survey 2018 \nSubject \nDefinitions.\u201d \nhttps://www2.census.gov/ \nprograms-surveys/acs/tech_docs/subject_definitions/ \n2018_ACSSubjectDefinitions.pdf.\nVeblen, Thorstein. 1994. Theory of the Leisure Class. \nMineola, NY: Dover.\nWilkinson, Richard G., and Kate E. Pickett. 2010. The \nSpirit Level: Why Greater Equality Makes Societies \nStronger. New York: Bloomsbury Press.\nWilkinson, Richard, and Kate Pickett. 2019. The Inner \nLevel: How More Equal Societies Reduce Stress, \nRestore Sanity, and Improve Everyone\u2019s Well-being. \nNew York: Penguin.\nAuthor Biographies\nRyan P. Thombs is a PhD student in the sociology depart-\nment at Boston College. His primary research interests are \nin the areas of political economy, environmental sociol-\nogy, the structural determinants of health, and quantitative \nmethods. His published work appears in such journals as \nSociological Forum, Energy Research & Social Science, \nClimatic Change, and Sociology of Development.\nDennis L. Thombs, PhD, FAAHB, is professor and dean \nof the School of Public Health at the University of North \nTexas Health Science Center in Fort Worth. For the past \n30 years, his scholarship has focused on substance abuse \nprevention and policy. He has authored more than 100 \narticles in peer-reviewed national and international jour-\nnals and is the coauthor of Introduction to Addictive \nBehaviors, published by the Guilford Press, which is now \nin its fifth edition. Dr. Thombs\u2019s research has been sup-\nported by the National Institute on Alcohol Abuse and \nAlcoholism, the National Institute of Mental Health, and \nthe U.S. Department of Education.\nAndrew K. Jorgenson is professor and chair in the \nDepartment of Sociology at Boston College. He conducts \nresearch in the areas of environmental sociology, global polit-\nical economy, and sustainability science. His published work \nappears in such venues as the American Journal of Sociology, \nNature Climate Change, Social Forces, Environmental \nResearch Letters, Social Problems, and Sociological Theory, \nand he is the coauthor of Super Polluters: Tackling the \nWorld\u2019s Largest Sites of Climate-Disrupting Emissions, \nforthcoming from Columbia University Press. He is the 2020 \nrecipient of the Fred Buttel Distinguished Contribution \nAward from the American Sociological Association\u2019s \nSection on Environmental Sociology.\nTaylor Harris Braswell is a sociology PhD candidate at \nNortheastern University. He works at the intersection of \nurban and environment sociology, with a particular focus \non urbanization, energy systems, and spatial data science. \nHis ongoing dissertation research is about the relationship \nbetween energy governance and urbanization processes in \nthe United States. He has also been involved in a number \nof projects that use spatial methods to examine other \nissues related to urban geography, such as gentrification, \nvacancy, and urban governance.\n\n\n    Scientific Research Findings:", "answer": "\u2022 Of the four measures of income inequality that we used, the share of income going to the bottom 20% was the key inequality measure associated with drug-related mortality.\n\u2022 An increase in the share of income going to the bottom 20% of earners was associated with a reduction in drug-related mortality.\n\u2022 We also found that states with a higher opioid prescription rate, on average, had higher drug-related mortality.", "id": 79}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/0022146520970190\nJournal of Health and Social Behavior\n2020, Vol. 61(4) 398\u00ad\u2013417\n\u00a9 American Sociological Association 2020\nDOI: 10.1177/0022146520970190\njhsb.sagepub.com\nOriginal Article\nThe global COVID-19 pandemic presented a threat \nto rival the Spanish influenza pandemic, more than \na hundred years before (Sly 2020). In many nations, \npublic health measures intended to prevent the \nspread of the virus and \u201cflatten the curve\u201d in terms \nof the rate of transmission resulted in extreme \nchanges to norms of social contact (Lai 2020; \nMorgan 2020). In Canada, public gatherings were \nbanned and citizens were urged to stay at home as \nmuch as possible (Government of Alberta 2020; \nLoewen 2020; Public Health Agency of Canada \n2020). The purpose of the current study is to apply a \nsynthesis of Durkheimian and life course perspec-\ntives to examine whether the social estrangement \ncreated by these public health measures resulted in \nan increase in psychological distress in the Canadian \npublic, as well as whether social estrangement and \nconsequent psychological distress were more pre-\ndominant in older respondents.\nTo address these questions, we compare two \nnational probability samples of working Canadians \nfrom the Canadian Quality of Work and Economic \nLife Study. The first was gathered in late September \n2019; the second was gathered in 2020, from March \n17 to March 23, when social isolation measures were \nenacted in Canada. Comparison of these two sam-\nples in measures of feelings of isolation, community \ndistrust, and symptoms of psychological distress \nallow us to examine not only whether the average \nlevel of psychological distress increased in the \n970190 HSBXXX10.1177/0022146520970190Journal of Health and Social BehaviorBierman and Schieman\nresearch-article2020\n1University of Calgary, Calgary, AB, Canada\n2University of Toronto, Toronto, ON, Canada\nCorresponding Author:\nAlex Bierman, Department of Sociology, University of \nCalgary, 2500 University Drive NW, Calgary, AB T2N \n1N4, Canada. \nEmail: aebierma@ucalgary.ca\nSocial Estrangement and \nPsychological Distress before \nand during the COVID-19 \nPandemic: Patterns of Change \nin Canadian Workers\nAlex Bierman1\n and Scott Schieman2\nAbstract\nThis article argues that the COVID-19 pandemic and associated social distancing measures intended to \nslow the rate of transmission of the virus resulted in greater subjective isolation and community distrust, \nin turn adversely impacting psychological distress. To support this argument, we examine data from the \nCanadian Quality of Work and Economic Life Study, two national surveys of Canadian workers\u2014one from \nlate September 2019 (N = 2,477) and the second from mid-March 2020 (N = 2,446). Analyses show that \nsubjective isolation and community distrust increased between the two surveys, which led to a substantial \nrise in psychological distress. Increases in subjective isolation were stronger in older respondents, resulting \nin a greater escalation in psychological distress. These findings support a Durkheimian perspective on the \nharm to social integration and mental health caused by periods of rapid social change but also illustrate \nhow a life course context can differentiate individual vulnerability to disintegrative social forces.\nKeywords\nCOVID-19, life course perspective, mental health, psychological distress, social integration\n\nBierman and Schieman\t\n399\npopulation but also whether population differences \nin subjective social isolation and distrust explain the \nevolution in levels of distress. We therefore contrib-\nute to the research in the sociology of mental health \nby examining whether public health measures \nintended to stop the spread of the COVID-19 virus \nmay also have had substantial adverse consequences \nfor public mental health in North America.\nBackground\nSocial Integration and Mental Health \nduring the Pandemic\nThe primary basis of our study is in the fundamental \nDurkheimian insight that social integration provides \na binding influence on suicide (Durkheim [1897] \n1951). Subsequent theorizing has clarified this core \ninsight, emphasizing that assimilation into a larger \nwhole through a web of social attachments acts a \nbulwark against vulnerabilities that can provoke \nanxiety while also assuaging hopelessness and \n\u201cmetaphysical exigencies\u201d that can occur in states \nof individual atomization (Abrutyn and Mueller \n2014:334). These arguments cohere with a social-\npsychological perspective in which social discon-\nnectedness develops into perceptions of social \nisolation that stimulate negative cognitive processes \nby acting as aversive figural signals of social vul-\nnerability (Cacioppo and Cacioppo 2014; Dahlberg, \nAndersson, and Lennartsson 2018). Disintegratory \nconditions will therefore be reflected in a subjective \nstate of social isolation that arouses feelings of anxi-\nety and dread. Empirical research supports this gen-\neral \nperspective, \ndemonstrating \nthat \nsocial \ndetachment can spur perceptions of social isolation \nthat are substantially associated with psychological \ndistress (Cacioppo et\u00a0al. 2011; Swader 2019).\nA Durkheimian perspective further builds on \nthese insights to explicate the negative conse-\nquences of rapid social change for mental health \n(Lester 2001). From this perspective, times of \nsocial turbulence weaken the social bonds of soci-\nety (Berkman et\u00a0al. 2000), thereby creating condi-\ntions that deplete societal integration (Zhao and \nCao \n2010). \nSimilarly, \nexpanding \nfrom \na \nDurkheimian perspective, Abrutyn and Mueller \n(2016) argue that periods of social disruption can \nthreaten or sever meaningful social ties, in turn cre-\nating subsequent negative emotions. Rapid social \nchange that interferes with established patterns of \nsocial interactions therefore acts as a destructive \ninfluence on social integration, in turn enhancing \nfeelings of social isolation in the population that \ngive rise to psychological discomfort. From this \nperspective, then, even if policies of social distanc-\ning were necessary to slow the spread of the \nCOVID-19 virus, the degree to which these mea-\nsures interrupted established patterns of social \ninteraction was likely to create a disintegratory state \nthat heightened subjective social isolation, thereby \nincurring mental health damages through increased \npsychological distress.\nDespite empirical support linking individual \nlevels of social integration to mental health out-\ncomes (Turner and Turner 2013), as well as research \ndemonstrating how contextual levels of social inte-\ngration can influence individual outcomes such as \nsuicidal ideation (e.g., Maimon and Kuhl 2008; \nWinfree and Jiang 2010), there is much less direct \nevidence for the consequences of societal change \nfor integration and subsequent effects on mental \nhealth. One of the primary areas of empirical evi-\ndence comes from periods of economic turmoil \n(Cockerham 2017), as increasing rates of foreclo-\nsure and unemployment were associated with \nspikes in suicide rates following the Great \nRecession (Houle and Light 2014; Phillips and \nNugent 2014). Closer to individual outcomes, \nmeso-level changes in foreclosure rates following \nthe recession were also inversely associated with \nindividual mental health (Houle 2014; Settels \n2020). An additional line of research linking social \nchange to integration has argued that increases in \nbirth cohort size and births to unwed mothers are \ncausative agents in declining levels of social inte-\ngration that affect suicide rates (O\u2019Brien and \nStockard 2006; Stockard and O\u2019Brien 2002). It is \nnotable, however, that much of this research is con-\nducted purely at a contextual level and does not \nclearly tie rapid social change to individual experi-\nences of social isolation. The current study there-\nfore builds on this body of evidence to demonstrate \nwhether increases in perceptions of social isolation \ncontributed to a rise in psychological distress fol-\nlowing the onset of the pandemic.\nConsequences of the Pandemic for \nCommunity Distrust\nThe disintegratory conditions of the pandemic may \nhave had additional consequences for social estrange-\nment by resulting in decreasing social trust. A focus \non social trust is directed by Abrutyn\u2019s (2019) recent \nproposition that disintegrative forces that work \nagainst social solidarity can lead individuals to feel \nan increasing sense of threat. This argument reso-\nnates with a social-psychological perspective on trust \n\n400\t\nJournal of Health and Social Behavior 61(4) \nthat positions a willingness to make one\u2019s self vulner-\nable in uncertain situations as fundamental to a high \ndegree of social trust (Baumert et\u00a0al. 2017). Societal \natomization that enhances a sense of threat will deter \nindividual openness to vulnerability that is elemental \nin building trust. These general processes are highly \nrelevant in the context of the COVID-19 pandemic \nbecause the novel requirements to maintain social \ndistance signaled that the threat posed by members of \nthe community inherently could not be contained, \nthereby fomenting distrust of others.\nEvidence from prior pandemics supports these \narguments. In particular, evidence from the Spanish \ninfluenza pandemic suggests that pandemics can \nlower levels of social trust (Aassve et\u00a0al. 2020). For \nexample, Barry (2005:329) describes how fear \neroded social trust in Philadelphia during the \nSpanish influenza pandemic: \u201cFear began to break \ndown the community of the city. Trust broke down. \nSigns began to surface of not just edginess but \nanger, not just finger-pointing or protecting one\u2019s \nown interests but active selfishness in the face of \ngeneral calamity.\u201d Similarly, suspicion increased \nfollowing the H1N1 outbreak in the late 2000s, as \nindividuals infected with the virus were seen as put-\nting others at risk (Gilman 2010). Consequently, \nbecause public leaders addressing the COVID-19 \npandemic began to call for social distancing, and in \nparticular warn people to guard against interactions \nwith others outside of their homes and in the com-\nmunity, we expect distrust of others in the commu-\nnity to have increased as well.\nDecreasing trust in members of the community \nis in turn likely to have substantial consequences for \npsychological distress. Trust in members of one\u2019s \nown neighborhood is associated with better mental \nhealth (Murayama et\u00a0 al. 2015; Tomita and Burns \n2013; Wu et\u00a0al. 2018) even when additional aspects \nof social trust are taken into account (Carpiano and \nFitterer 2014). Trust in members of the community \ncan be important for mental health by increasing a \nsense of being accepted and facilitating social sup-\nport, as well as by reinforcing informal social con-\ntrol that serves to prevent harmful health behaviors \n(Fujiwara and Kawachi 2008; Glanville and Story \n2018). Conversely, feeling suspicion and needing to \nbe on guard of the people we come into contact with \noutside of our homes acts as a stressor that increases \npsychological distress (Ross 2011). In the context of \nincreased isolation associated with the COVID-19 \npandemic, neighbors may become a focus of social \nacceptance and sources of support, with the result \nthat declining trust in the member of one\u2019s commu-\nnity will act as a further stressor that elevates \npsychological distress. Within this research, we \ntherefore examine whether increases in psychologi-\ncal distress following the onset of the COVID-19 \npandemic were not only attributable to increases in \nsubjective social isolation, but a rise in community \ndistrust as well.\nIntegrating a Life Course Perspective\nDurkheimian perspectives on the consequences of \nsocietal change for social integration tend to frame \nthese processes expansively, focusing on broad \ndimensions of social change and their summative \nintegratory consequences. In the current research, \nthough, we integrate insights from a life course per-\nspective into the study of societal change and social \nintegration. We suggest a concept of \u201cintegratory \nvulnerability,\u201d in which a life course context\u2014and \nparticularly cohort membership and the timing of \nsocietal events in the life course\u2014critically differ-\nentiates individual vulnerability to disintegratory \nsocietal events.\nWe are guided to the pivotal role of integratory \nvulnerability by the strong emphasis of research in \nthe life course perspective on the differentiated \nramifications of large-scale economic events for \nindividual lives (Elder 1999). A key paradigmatic \nprinciple of a life course perspective is that histori-\ncal events can affect people differently depending \non the timing of these events in the life course \n(Elder, Johnson, and Crosnoe 2003). Recent \nresearch exemplifies these patterns, demonstrating \nthat the timing of the Great Recession as individuals \nentered the labor market had subsequent implica-\ntions for individual earning capabilities (Atherwood \nand Sparks 2019).\nThe question of timing is especially relevant to \nthe threat of the COVID-19 pandemic. The threat of \nserious adverse health consequences due to con-\ntraction of the virus are greater in older individuals \n(Heymann and Shindo 2020). Consequently, age \nmay have been critical in determining integratory \nvulnerability: Older individuals may have isolated \nto a greater degree as a result of the pandemic and \nalso may have experienced a greater distrust of \nmembers of their community due to their height-\nened vulnerability and subsequent fear of contract-\ning the virus. Greater increases in feelings of \nisolation and distrust would in turn lead to more \nsubstantial gains in psychological distress.\nYet, differentiation in the consequences of his-\ntorical events can also occur in part because the \ntimes in which individuals are born into and develop \nalter the resources and deficits that different birth \n\nBierman and Schieman\t\n401\ncohorts bring to bear in times of crisis (Elder 1994; \nKeyes et\u00a0al. 2010). Birth cohorts may be differently \nequipped to meet the challenges of historical turbu-\nlence, in turn altering the degree to which historical \nevents create negative repercussions across age \ncohorts. We suggest that we will observe differ-\nences in integratory vulnerability across cohorts \ndue to a \u201cdigital divide\u201d between cohorts in the \ncomfort and use of internet and other electronic \nmeans of communications (Friemel 2016). A com-\nmon characterization of the divide between cohorts \nis that younger cohorts, particularly those born \nbefore 1980, are \u201cdigital immigrants\u201d and those \nborn after the 1980s are \u201cdigital natives\u201d (Nevin \nand Schieman 2020; Prensky 2001). In support of \nthis characterization, research thoroughly docu-\nments that members of younger age cohorts have \ngreater comfort and facility in internet use (B\u00fcchi, \nJust, and Latzer 2016; Hargittai and Dobransky \n2017). Concomitantly, there tends to be greater reti-\ncence toward the use of social networking technol-\nogy with age (Yu et\u00a0al. 2016). Older users may find \nless social utility and fulfillment from social net-\nworking technology (L\u00fcders and Brandtz\u00e6g 2017) \nand instead prefer face-to-face interactions (Yuan \net\u00a0al. 2016).\nEvidence of cohort differences in comfort \nwith and utility in online social interactions sug-\ngests that we will observe that patterns of change \nin subjective social isolation and community \n\u00addistrust differ across age cohorts. Members of \nyounger cohorts may have been more able to gain \nsocial sustenance through online interactions and \nas a result felt less isolation in the wake of social \ndistancing measures. Similarly, younger cohorts \nmay have been more able to use social media and \nother electronic resources to gain information on \nlocal spread of the infection and means of mini-\nmizing risk of transmission, which may in turn \nserve to lessen generalized suspicion of one\u2019s \nneighbors.\nIt is critical to underscore that the sum of integra-\ntory vulnerability due to both age and cohort effects \nis that older individuals are likely to experience more \nsubstantial social estrangement. A lack of ability to \ndifferentiate between age and birth cohort in the cur-\nrent analyses is therefore not a substantial weakness \nbecause we expect moderation by both factors to be \nin the same direction. Within this research, we there-\nfore examine whether increases in subjective social \nisolation and community distrust are observed more \nprominently in older individuals, leading to stronger \nincreases in psychological distress.\nSummary of Expectations\nFigure 1 summarizes the primary expectations of \nthis article. First, the figure indicates that we expect \nto observe increased feelings of social isolation and \ncommunity distrust in 2020, following the outbreak \nof the COVID-19 pandemic. Second, we expect \nthat subjective social isolation and community dis-\ntrust will be associated with greater psychological \ndistress. Consequently, increases in subjective \nsocial isolation and community distrust will lead to \na rise in psychological distrust following the out-\nbreak of the pandemic. However, Figure 1 also \nshows positive paths between age and the paths \nbetween wave of survey and the measures of social \nestrangement, which illustrates that we expect \namplified increases in subjective social isolation \nand community distrust among older members of \nour study, in turn leading to greater increases in \npsychological distress.\nFigure 1.\u2002 Model of Social Estrangement and Psychological Distress Following the Outbreak of \nCOVID-19.\n\n402\t\nJournal of Health and Social Behavior 61(4) \nData and Methods\nData\nData were derived from two waves of the Canadian \nQuality of Work and Economic Life Study \n(C-QWELS), national surveys intended to examine \nsocial conditions and well-being among Canadians \nwho were currently employed. Data were gathered \nby the study authors in cooperation with the Angus \nReid Forum, a Canadian national survey research \nfirm that maintains an ongoing national panel of \nCanadian respondents. The C-QWELS I was gath-\nered from September 19 to September 24, 2019, and \nwas an online survey conducted among a represen-\ntative sample of 2,524 working Canadians. The \nresponse rate was 42%, but results were statistically \nweighted according to the most current education, \nage, gender, and region census data to ensure a sam-\nple representative of working Canadians. The \nC-QWELS II was conducted from March 17 to \nMarch 23, 2020 with another nationally representa-\ntive sample of 2,528 working Canadians. The \nresponse rate was 43%, and responses were simi-\nlarly weighted. Of the 5,052 total respondents, \n4,923 were retained in the analytic sample (2019 \nsample = 2,477; 2020 sample = 2,446), a retention \nrate of over 97%, suggesting little bias due to list-\nwise deletion.\nFocal Measures\nPsychological distress.\u2002 Psychological distress was \nmeasured using five common symptoms of nonspe-\ncific psychological distress (Kessler et\u00a0 al. 2002): \nfeel anxious or tense, feel nervous, feel restless or \nfidgety, feel sad or depressed, feel hopeless. \nRespondents indicated the frequency they experi-\nenced each symptom in the previous month, with \nresponse scales of all of the time, most of the time, \nsome of the time, a little of the time, and none of the \ntime. All responses were coded so that higher values \nindicated more frequent symptoms. Psychological \ndistress was measured as the mean of responses to \nthese five questions (Cronbach\u2019s \u03b1 = .877).\nCommunity distrust.\u2002 Similar to other studies (e.g., \nCarpiano and Fitterer 2014; Fujiwara and Kawachi \n2008), community distrust was measured using a sin-\ngle item that asked, \u201cThinking about the people in your \nneighbourhood\u2014that is, the local area in which you \nlive,\u201d how much do you agree or disagree with the \nstatement, \u201cMy neighbours can be trusted.\u201d Response \nchoices were strongly agree, somewhat agree, some-\nwhat disagree, strongly disagree. Responses were \ncoded so that higher values indicated greater disagree-\nment, thereby creating a measure of community dis-\ntrust. However, a small proportion of respondents \nindicated strong disagreement, and ancillary analyses \nshowed that standard errors were substantially inflated \ndue to the small number of these cases. We therefore \ncombined responses of strongly disagree and disagree \ninto an overall \u201cdisagree\u201d category.\nSubjective social isolation.\u2002 Subjective social isola-\ntion was measured using one item that asked respon-\ndents how often they felt \u201cisolated from other people\u201d \nin the previous month, with the same response cate-\ngories as the distress items.\nAge.\u2002 Age was measured in years of age, with a \ntop value of 74 to avoid the undue influence of \nsparse values of high age in tests of moderation. Age \nwas centered over a value of 40, the approximate \nmedian age in the sample, to provide clearer inter-\npretations of the interactions.\nWave of survey.\u2002 Membership in the surveys was \nindicated by a dichotomous variable in which a \nvalue of 0 indicated the respondent participated in \nthe September 2019 survey and a value of 1 indi-\ncated the respondent participated in the March 2020 \nsample. In the results, these were referred to as the \n2019 and 2020 samples, respectively.\nControl Measures\nGeneralized trust.\u2002 To take broader social trust into \naccount, respondents were asked a common survey \nquestion on social trust: \u201cGenerally speaking, would \nyou say that most people can be trusted, or that you \ncan\u2019t be too careful in dealing with people? Please \ntell me what you think, where 1 means you can\u2019t be \ntoo careful and 5 means most people can be trusted.\u201d \nIn the analyses, an indicator of the lowest level of \ntrust was contrasted with a set of dichotomous indi-\ncators for each of the other categories of trust. \nAncillary analyses showed that community and \ngeneralized trust were not substantially correlated, \nsuggesting that each were distinct indicators of \ntrust.\nEmployment conditions.\u2002 Because analyses were \nbased on two samples of working Canadians, employ-\nment conditions were taken into account to address \nthe degree to which occupational experiences contrib-\nuted to psychological distress as well as the extent to \nwhich individuals may have experienced changes in \nwork conditions and scheduling due to working at \n\nBierman and Schieman\t\n403\nhome. Occupational class was measured using a five-\ncategory classification\u2014professional, administrative, \nsales, clerical, and laborer\u2014with professional as ref-\nerence. Number of work hours in main job were con-\ntrolled using a set of dichotomous indicators, in which \npart-time (\u2264 29 hours or less) was contrasted with full-\ntime (30 \u2013 49) and extended hours (\u2265 50). Working \nmore than one job was controlled by a dichotomous \nindicator in which the higher value indicated that the \nrespondent worked more than one job. Degree of \nworking at home was taken into account using a mea-\nsure in which individuals who never worked at home \nwere contrasted to categories of a few times a year, \nabout once a month, about once a week, more than \nonce a week, and every day/mainly work at home. \nThe degree of control over work scheduling was taken \ninto account by asking respondents, \u201cHow much con-\ntrol do you have in scheduling your work hours?\u201d \nwith responses of none contrasted to very little, some, \na lot, and complete control.\nFamilial statuses.\u2002 Familial statuses that may pro-\nvide support and ward off social estrangement were \ntaken into account with a dichotomous indicator in \nwhich the higher value indicated that the respondent \nlived with a romantic partner and a dichotomous \nvariable in which the higher value indicated that the \nrespondent lived with at least one child under the \nage of 18.\nSocial and economic statuses.\u2002 Social and eco-\nnomic statuses that may contribute to both social \nestrangement and psychological distress were also \ncontrolled, including education, income, economic \nhardship, gender, and minority status. Education \nwas operationalized as a set of categories in which \nindividuals with a university degree were compared \nto a category of high school, some university or col-\nlege/trade school, and graduated from college or \ntrade school; because less than 2% of the weighted \nsample at each wave had less than a high school \ndegree, these respondents were grouped with those \nwith a high school degree. Income was measured as \na set of categories in which $150,000 or more in \nhousehold income was compared to less than \n$25,000, $25,000 to less than $50,000, $50,000 to \nless than $100,000, and $100,000 to less than \n$150,000. Because individuals who do not provide \nincome often reside in high-income categories and \ntaking nonresponse into account would help to con-\ntrol for biases in self-reports, missing income was \nconsidered as an additional analytic category. Fur-\nthermore, because the measure of income did not \naddress more proximal experiences of economic \ndeprivation that may have been associated with the \npandemic, we also included a commonly employed \nmeasure of economic hardship that has been shown \nto be a valid indicator of physical and mental health \n(e.g., Kahn and Pearlin 2006). Respondents were \nasked, \u201cHow do your finances usually work out by \nthe end of the month?,\u201d with responses of a lot of \nmoney left over used as a comparison group to not \nenough to make ends meet, barely enough to get by, \njust enough to make ends meet, and a little money \nleft over. Gender was coded as 0 = men, 1 = women. \nRacial and ethnic minority status is typically mea-\nsured in Canada using the designation of \u201cvisible \nminority\u201d (Little 2016), and in keeping with this \nconvention, visible minority status was measured \nby asking respondents, \u201cWould you say you are a \nmember of a visible minority here in Canada (in \nterms of your ethnicity/race)?\u201d Responses were \nindicated by a dichotomous variable in which the \nhigher value indicated visible minority.\nPlan of Analysis\nAll primary analyses were conducted in Stata 16.1. \nAnalyses were conducted in three stages. In the first \nstage, we examined bivariate differences between the \nfocal study measures in the two waves of the surveys. \nIn the second stage, we examined predictors of com-\nmunity distrust and subjective social isolation in mul-\ntiple regression models. Each outcome was examined \nusing two models. First, we examined between-wave \ndifferences in the outcome independent of the control \nvariables. Second, we tested whether between-wave \ndifferences in each measure differed by respondent\u2019s \nage by testing an interaction between wave of survey \nand age. Because both community distrust and sub-\njective social isolation were based on an ordinal \nresponse scale, we utilized ordinal logistic regression \nin the multiple regression analyses (Hoffmann 2016). \nOrdinal logistic regression models depend on an \nassumption that the change in risk based on a predic-\ntor is the same between each category of the depen-\ndent variable (Williams 2006), and preliminary \nanalyses that applied a Brant test (Brant 1990) sup-\nported this assumption for the association between \nwave of survey and both outcomes as well as for the \ninteraction term.\nIn a third stage of analyses, we used ordinary \nleast squares (OLS) regression to examine the asso-\nciation between wave of survey and psychological \ndistress. We first examined between-wave differ-\nences in psychological distress while holding con-\nstant all background controls. In additional models, \nwe sequentially controlled for community distrust \n\n404\t\nJournal of Health and Social Behavior 61(4) \nand subjective social isolation, which demonstrated \nthe extent to which each explained between-wave \ndifferences in psychological distress (MacKinnon \n2008). To account for the noncontinuous nature of \nthe measures of community distrust and subjective \nsocial isolation, each measure was entered into the \nregression model as a set of categorical indicators, \nwith strong trust or no sense of isolation as the ref-\nerence group, respectively. We then repeated this \nprocess by removing the measures of community \ndistrust and subjective social isolation and testing \nan interaction between wave of survey and age, \nwhich demonstrated the extent to which between-\nwave differences in psychological distress differed \nby age. A reintroduction of the measures of com-\nmunity distrust and subjective social isolation into \nthe model examined the extent to which these fac-\ntors explained age-based contingencies in between-\nwave changes in psychological distress.\nResults\nTable 1 displays the distribution of measures for \neach survey wave and for the combined sample. \nTable 1 shows a shift in both community distrust \nand subjective social isolation toward greater dis-\ntrust and isolation. Generalized trust appeared rela-\ntively stable, however, which reflects the importance \nof considering trust in specific targets with whom an \nindividual may interact rather than more diffuse per-\nceptions of trust. Table 1 also shows that mean lev-\nels of psychological distress increased between the \ntwo waves. Although the difference in distress is not \nstatistically significant, ancillary analyses showed \nthat the lack of statistical significance was largely \ndue to small compositional differences between the \ntwo surveys. For example, simply controlling for \ncompositional differences in age and presence of \nchildren led to an estimation of significant greater \ndistress in March compared to September. We there-\nfore turn to the multivariate analyses that examine \ndifferences between the two waves of surveys when \ntaking these compositional factors into account.\nMultiple Regression Analyses\nTable 2 displays the results of the ordered logistic \nregression analyses of community distrust and sub-\njective social isolation. Model 1 shows that, inde-\npendent of the controls, respondents in 2020 \nevidenced a significantly increased risk of commu-\nnity distrust. Being a respondent in the 2020 survey \nwas associated with almost 50% greater odds of \nreporting a higher level of distrust than being a \nrespondent in the 2019 sample. However, these \nbetween-wave differences did not vary by age; the \ninteraction between wave of survey and age in \nModel 2 is not significant.\nTurning to subjective social isolation, Model 3 \nshows that respondents in 2020 also had increased \nrisk of subjective social isolation. Being a respon-\ndent in the 2020 sample was associated with  \n36% greater odds of reporting a higher level of iso-\nlation than being a respondent in the 2019 sample. \nFurthermore, between-wave differences in subjec-\ntive social isolation differ by age; Model 4 shows \nthat the interaction between wave of survey and age \nis statistically significant.\nTo explicate this interaction, Figure 2 presents \nthe estimated odds ratios for between-wave differ-\nences in subjective social isolation across the range \nof ages in the survey. This figure shows that, for \nrespondents in their 20s and 30s, the odds ratios for \nbetween-wave differences in subjective social iso-\nlation are relatively small and are not statistically \nsignificant for those in their 20s. By age 40, how-\never, respondents in 2020 had 33% greater odds of \nreporting increased feelings of isolation, and this \ndifference was significant. The between-wave odds \nof subjective social isolation increased further in \nstrength in later age cohorts. Respondents at age 50 \nhad almost 50% greater odds of increased feelings \nof isolation in 2020, and by age 60, the odds were \n63% greater in 2020. In accordance with our expec-\ntations, then, the increased risk of subjective social \nisolation following the COVID-19 outbreak was \ngreater among older respondents.\nTable 3 presents the results of the OLS regres-\nsion analyses of psychological distress. Model 1 \nshows that between-wave increases in distress are \nsignificant, independent of controls. To demon-\nstrate the strength of this difference, we examined \nthe semistandardized difference, in which the met-\nric difference is divided by the standard deviation \nof distress (McClendon 1994), thereby expressing \nthis difference in units of standard deviations of dis-\ntress. When semistandardized, this difference was \n.069. It should be emphasized that this difference \nwas observed in the population of working adults in \na relatively short six-month period, and subsequent \nanalyses will demonstrate that this increase is a \ncombination of much stronger and much weaker \nage-variegated changes in distress.\nModel 2 controls for categories of community \ndistrust. When compared to respondents who \nreported strong agreement with trust in neighbors, \nrespondents in the combined disagreement category \nreported significantly higher levels of psychological \n\nBierman and Schieman\t\n405\nTable 1.\u2002 Sample Descriptives.\n2019 Survey\n2020 Survey\nMerged Surveys\np\nDistress\n2.357\n2.404\n2.380\n\u2002\nCommunity distrust\n\u2003 Strongly agree with trust in neighbors\n.348\n.265\n.307\n\u2002\n\u2003 Somewhat agree with trust in neighbors\n.502\n.550\n.526\n\u2002\n\u2003 Somewhat disagree/strongly disagree with \ntrust in neighbors\n.150\n.185\n.167\n***\nSubjective isolation\n\u2003 None of the time\n.401\n.345\n.373\n\u2002\n\u2003 A little of the time\n.263\n.245\n.254\n\u2002\n\u2003 Some of the time\n.209\n.249\n.229\n\u2002\n\u2003 Most of the time\n.104\n.126\n.115\n\u2002\n\u2003 All of the time\n.023\n.036\n.029\n***\nAge\n41.967\n41.914\n41.940\n\u2002\nGeneralized trust\n\u2003 You can\u2019t be too careful\n.122\n.101\n.112\n\u2002\n\u2003 2\n.167\n.178\n.172\n\u2002\n\u2003 3\n.400\n.390\n.395\n\u2002\n\u2003 4\n.243\n.265\n.254\n\u2002\n\u2003 Most people can be trusted\n.068\n.067\n.067\n\u2002\nOccupational class\n\u2003 Professional\n.398\n.404\n.401\n\u2002\n\u2003 Administrative\n.156\n.124\n.140\n\u2002\n\u2003 Sales\n.188\n.195\n.192\n\u2002\n\u2003 Clerical\n.181\n.170\n.175\n\u2002\n\u2003 Laborer\n.077\n.107\n.092\n***\nWork hours\n\u2003 Part-time\n.183\n.218\n.200\n\u2002\n\u2003 Full-time\n.682\n.657\n.670\n\u2002\n\u2003 Extended hours\n.135\n.125\n.130\n*\nWorking multiple jobs\n\u2003 One job\n.775\n.776\n.776\n\u2002\n\u2003 More than one job\n.225\n.224\n.225\n\u2002\nWork at home\n\u2003 Never\n.353\n.334\n.343\n\u2002\n\u2003 A few times a year\n.108\n.112\n.110\n\u2002\n\u2003 About once a month\n.068\n.081\n.074\n\u2002\n\u2003 About once a week\n.113\n.113\n.113\n\u2002\n\u2003 More than once a week\n.163\n.171\n.167\n\u2002\n\u2003 Every day/mainly work at home\n.195\n.190\n.193\n\u2002\nControl over work scheduling\n\u2003 None\n.158\n.141\n.149\n\u2002\n\u2003 Very little\n.197\n.200\n.198\n\u2002\n\u2003 Some\n.273\n.267\n.270\n\u2002\n\u2003 A lot\n.233\n.245\n.239\n\u2002\n\u2003 Complete control\n.140\n.147\n.144\n\u2002\nLiving with romantic partner\n\u2003 Partner\n.648\n.648\n.648\n\u2002\n\u2003 No partner\n.352\n.352\n.352\n\u2002\n(continued)\n\n406\t\nJournal of Health and Social Behavior 61(4) \n2019 Survey\n2020 Survey\nMerged Surveys\np\nAny children in household\n\u2003 No children\n.678\n.618\n.648\n\u2002\n\u2003 Children\n.322\n.382\n.352\n***\nEducation\n\u2003 High school\n.096\n.121\n.108\n\u2002\n\u2003 Some university or college/trade school\n.206\n.217\n.211\n\u2002\n\u2003 College/trade school\n.230\n.231\n.230\n\u2002\n\u2003 University degree\n.469\n.432\n.450\n*\nIncome\n\u2003 < $25,000\n.061\n.068\n.065\n\u2002\n\u2003 $25,000 to < $50,000\n.146\n.136\n.141\n\u2002\n\u2003 $50,000 to < $100,000\n.302\n.304\n.303\n\u2002\n\u2003 $100,000 to < $150,000\n.222\n.229\n.226\n\u2002\n\u2003 \u2265 $150,000\n.172\n.163\n.168\n\u2002\n\u2003 Missing income\n.096\n.100\n.098\n\u2002\nFinances at end of month\n\u2003 A lot of money left over\n.113\n.109\n.111\n\u2002\n\u2003 A little money left over\n.406\n.425\n.415\n\u2002\n\u2003 Just enough to make ends meet\n.256\n.249\n.253\n\u2002\n\u2003 Barely enough to get by\n.172\n.162\n.167\n\u2002\n\u2003 Not enough to make ends meet\n.053\n.056\n.054\n\u2002\nGender\n\u2003 Men\n.513\n.514\n.513\n\u2002\n\u2003 Women\n.487\n.486\n.487\n\u2002\nVisible minority\n\u2003 Not a visible minority\n.873\n.863\n.868\n\u2002\n\u2003 Visible minority\n.127\n.138\n.132\n\u2002\nNote: N = 4,923 (2019 sample = 2,477; 2020 sample = 2,446). Descriptives are weighted. Means are presented for \ncontinuous measures, proportions for categorical measures. Data are from the Canadian Quality of Work and \nEconomic Life Study.\n*p < .05, ***p < .001, two tailed.\nTable 1.\u2002 (continued)\ndistress. The difference in distress for individuals in \nthe somewhat agree category is weaker, however, and \nnot statistically significant. These results showed that \nit is marked distrust in neighbors\u2014as indicated by \ndisagreement with a statement of trust in neighbors\u2014 \nthat is the clear distressing aspect of community \n\u00addistrust. The difference in distress for individuals \nwho distrusted neighbors was also relatively strong, \nwith a semistandardized value of .178. Furthermore, \nthe between-wave difference in distress declined \nalmost 15% from the previous model, from .060  \nto .052, and reduced in significance from p < .01 to  \np < .05, indicating that increased community  \ndistrust contributed to explaining between-wave \n\u00addifferences in psychological distress (MacKinnon, \n2008). There was also a commensurate decline in the \nsemistandardized between-wave difference in \u00addistress, \nfrom .069 to .060.\nModel 3 introduces controls for categories of \nresponses to subjective social isolation, with no \n\u00adfeelings of isolation as the reference group. All cate-\ngories of feelings of isolation are significantly asso-\nciated with greater distress. Furthermore, these \ndifferences are quite substantial. Ancillary analyses \nshowed that the semistandardized coefficient feeling \nisolated some of the time was .835, whereas the \nsemistandardized coefficient for feeling isolated \nmost or all of the time was 1.33 and was 2.024 for all \nof the time. There is also a substantial decrease in the \nbetween-wave difference in psychological distress \nwhen feelings of isolation are taken into account, as \nthe between-wave increase in distress is entirely \n\n407\nTable 2.\u2002 Ordinal Logistic Regression Analyses of Community Distrust and Subjective Social Isolation.\nCommunity Distrust\nSubjective Social Isolation\u2002\n\u2002\nModel 1\nModel 2\nModel 3\nModel 4\n\u2002\nb\nSE\nexp(b)\np\nb\nSE\nexp(b)\np\nb\nSE\nexp(b)\np\nb\nSE\nexp(b)\np\nFocal predictors\n\u2003 Change from 2019 to 2020\n.391\n.060\n1.479\n***\n.379\n.062\n1.461\n***\n.311\n.057\n1.365\n***\n.299\n.058\n1.349\n***\n\u2003 Age\n\u2212.011\n.003\n.989\n***\n\u2212.013\n.004\n.987\n***\n\u2212.032\n.002\n.968\n***\n\u2212.037\n.003\n.964\n***\n\u2003 Survey \u00d7 Age\n.006\n.005\n1.006\n.009\n.005\n1.009\n*\nControl measures\n\u2002\n\u2003 Generalized trusta\n\u2002\n\u2003 \u2003 2\n\u2212.138\n.130\n.871\n\u2212.142\n.130\n.867\n\u2212.286\n.126\n.751\n*\n\u2212.291\n.127\n.748\n*\n\u2003 \u2003 3\n\u2212.495\n.116\n.609\n***\n\u2212.501\n.117\n.606\n***\n\u2212.505\n.115\n.603\n***\n\u2212.511\n.116\n.600\n***\n\u2003 \u2003 4\n\u22121.119\n.124\n.326\n***\n\u22121.123\n.124\n.325\n***\n\u2212.616\n.121\n.540\n***\n\u2212.618\n.121\n.539\n***\n\u2003 \u2003 Most people can be trusted\n\u22121.612\n.164\n.199\n***\n\u22121.617\n.164\n.198\n***\n\u2212.995\n.158\n.370\n***\n\u2212.999\n.158\n.368\n***\n\u2003 Occupational classb\n\u2002\n\u2003 \u2003 Administrative\n\u2212.080\n.096\n.923\n\u2212.081\n.096\n.922\n.130\n.090\n1.139\n.129\n.090\n1.137\n\u2002\n\u2003 \u2003 Sales\n.073\n.091\n1.076\n.076\n.091\n1.079\n.136\n.087\n1.146\n.139\n.087\n1.149\n\u2002\n\u2003 \u2003 Clerical\n.072\n.102\n1.075\n.072\n.103\n1.075\n\u2212.083\n.098\n.920\n\u2212.084\n.098\n.920\n\u2002\n\u2003 \u2003 Laborer\n.162\n.108\n1.176\n.163\n.108\n1.177\n\u2212.030\n.100\n.971\n\u2212.025\n.100\n.975\n\u2002\n\u2003 Work hoursc\n\u2002\n\u2003 \u2003 Full-time\n.093\n.085\n1.097\n.090\n.085\n1.094\n\u2212.120\n.084\n.887\n\u2212.125\n.084\n.883\n\u2002\n\u2003 \u2003 Extended hours\n\u2212.148\n.121\n.862\n\u2212.154\n.121\n.857\n\u2212.132\n.114\n.876\n\u2212.139\n.115\n.870\n\u2002\n\u2003 \u2003 Working multiple jobs\n\u2212.100\n.074\n.905\n\u2212.102\n.074\n.903\n.130\n.069\n1.139\n.128\n.069\n1.137\n\u2002\n\u2003 Work at homed\n\u2002\n\u2003 \u2003 A few times a year\n.019\n.104\n1.019\n.021\n.105\n1.021\n.095\n.099\n1.099\n.097\n.099\n1.102\n\u2002\n\u2003 \u2003 About once a month\n.064\n.115\n1.066\n.069\n.115\n1.071\n\u2212.065\n.117\n.937\n\u2212.057\n.117\n.944\n\u2002\n\u2003 \u2003 About once a week\n.007\n.109\n1.007\n.007\n.109\n1.007\n.233\n.100\n1.262\n*\n.232\n.100\n1.262\n*\n\u2003 \u2003 More than once a week\n\u2212.016\n.095\n.984\n\u2212.013\n.095\n.987\n.280\n.090\n1.323\n**\n.286\n.090\n1.331\n**\n\u2003 \u2003 Every day/mainly work at home\n\u2212.138\n.095\n.871\n\u2212.133\n.095\n.875\n.290\n.092\n1.336\n**\n.297\n.092\n1.346\n**\n\u2003 Control over work schedulinge\n\u2002\n\u2003 \u2003 Very little\n.081\n.102\n1.084\n.079\n.102\n1.083\n\u2212.074\n.094\n.929\n\u2212.078\n.094\n.925\n\u2002\n\u2003 \u2003 Some\n.049\n.099\n1.050\n.049\n.099\n1.050\n\u2212.209\n.093\n.811\n*\n\u2212.211\n.093\n.810\n*\n\u2003 \u2003 A lot\n\u2212.069\n.104\n.933\n\u2212.071\n.104\n.932\n\u2212.337\n.097\n.714\n**\n\u2212.342\n.097\n.711\n***\n\u2003 \u2003 Complete control\n\u2212.123\n.120\n.885\n\u2212.127\n.119\n.881\n\u2212.565\n.115\n.568\n***\n\u2212.576\n.115\n.562\n***\n\u2003 Not living with romantic partner\n.114\n.073\n1.121\n.119\n.073\n1.126\n.335\n.071\n1.398\n***\n.343\n.071\n1.409\n***\n(continued)\n\n408\nCommunity Distrust\nSubjective Social Isolation\u2002\n\u2002\nModel 1\nModel 2\nModel 3\nModel 4\n\u2002\nb\nSE\nexp(b)\np\nb\nSE\nexp(b)\np\nb\nSE\nexp(b)\np\nb\nSE\nexp(b)\np\n\u2003 Any children in household\n\u2212.176\n.067\n.839\n**\n\u2212.171\n.068\n.843\n*\n\u2212.029\n.063\n.971\n\u2212.020\n.064\n.981\n\u2002\n\u2003 Educationf\n\u2002\n\u2003 \u2003 High school\n.138\n.117\n1.148\n.146\n.116\n1.157\n.064\n.113\n1.066\n.079\n.113\n1.082\n\u2002\n\u2003 \u2003 Some university or college/trade school\n\u2212.033\n.088\n.968\n\u2212.033\n.088\n.968\n\u2212.202\n.084\n.817\n*\n\u2212.202\n.085\n.817\n*\n\u2003 \u2003 College/trade school\n\u2212.012\n.081\n.988\n\u2212.009\n.081\n.991\n\u2212.210\n.077\n.810\n**\n\u2212.206\n.077\n.814\n**\n\u2003 Incomeg\n\u2002\n\u2003 \u2003 < $25,000\n.272\n.187\n1.312\n.271\n.187\n1.311\n.375\n.161\n1.454\n*\n.375\n.161\n1.454\n*\n\u2003 \u2003 $25,000 to < $50,000\n.304\n.127\n1.355\n*\n.299\n.127\n1.348\n*\n.154\n.120\n1.167\n.147\n.120\n1.158\n\u2002\n\u2003 \u2003 $50,000 to < $100,000\n.281\n.096\n1.324\n**\n.276\n.096\n1.317\n**\n.104\n.088\n1.109\n.096\n.089\n1.101\n\u2002\n\u2003 \u2003 $100,000 < $150,000\n.052\n.095\n1.054\n.050\n.095\n1.051\n\u2212.014\n.089\n.986\n\u2212.015\n.089\n.985\n\u2002\n\u2003 \u2003 Missing income\n.207\n.127\n1.230\n.202\n.127\n1.224\n.091\n.118\n1.095\n.083\n.119\n1.086\n\u2002\n\u2003 Finances at end of monthh\n\u2002\n\u2003 \u2003 A little money left over\n.103\n.111\n1.109\n.107\n.111\n1.113\n.176\n.104\n1.193\n.183\n.104\n1.201\n\u2002\n\u2003 \u2003 Just enough to make ends meet\n.157\n.120\n1.171\n.160\n.120\n1.174\n.336\n.112\n1.399\n**\n.340\n.112\n1.406\n**\n\u2003 \u2003 Barely enough to get by\n.343\n.131\n1.409\n**\n.344\n.130\n1.411\n**\n.710\n.123\n2.035\n***\n.714\n.123\n2.043\n***\n\u2003 \u2003 Not enough to make ends meet\n.206\n.167\n1.229\n.204\n.167\n1.226\n.848\n.155\n2.336\n***\n.845\n.156\n2.328\n***\n\u2003 Women\n\u2212.045\n.064\n.956\n\u2212.045\n.064\n.956\n\u2212.011\n.061\n.990\n\u2212.012\n.061\n.989\n\u2002\n\u2003 Visible minority\n.230\n.089\n1.259\n*\n.229\n.089\n1.258\n*\n.274\n.085\n1.315\n**\n.272\n.085\n1.312\n**\nCut 1\n\u2212.985\n.213\n\u2212.994\n.214\n\u2212.648\n.199\n\u2212.654\n.199\n\u2002\nCut 2\n1.642\n.214\n1.634\n.215\n.518\n.198\n.513\n.198\n\u2002\nCut 3\n\u2014\n\u2014\n\u2014\n\u2014\n1.902\n.201\n1.897\n.201\n\u2002\nCut 4\n\u2014\n\u2014\n\u2014\n\u2014\n3.696\n.217\n3.689\n.217\n\u2002\nNote: N = 4,923. Data are from the Canadian Quality of Work and Economic Life Study.\naYou can\u2019t be too careful is reference.\nbProfessional is reference.\ncPart-time is reference.\ndNever is reference.\neNone is reference.\nfUniversity degree is reference.\ng\u2265$150,000 is reference.\nhA lot of money left over is reference.\n*p < .05, **p < .01, ***p < .001, two-tailed.\nTable 2.\u2002 (continued)\n\nBierman and Schieman\t\n409\nnegated and the difference is no longer statistically \nsignificant. Increasing levels of subjective social iso-\nlation from September 2019 to March 2020 therefore \nsubstantially explain increases in psychological \ndistress\nModel 4 removes the indicators of community \ndistrust and subjective social isolation but intro-\nduces an interaction between wave of survey and \nage. This interaction is significant, demonstrating \nthat between-wave differences in psychological \ndistress varied significantly by age cohort. Figure 3 \nclarifies the meaning of this interaction by depict-\ning the semistandardized coefficients for differ-\nences in psychological distress across values of age. \nFigure 3 shows that at younger ages, between-wave \ndifferences in psychological distress are not signifi-\ncant. By age 40, however, respondents in 2020 \nreported significantly higher mean levels of psy-\nchological distress. These differences increased in \nstrength at older ages. For respondents at age 50, \nthere was an increase of over a tenth of a standard \nbetween 2019 and 2020, and for respondents at age \n60, this increase was almost a sixth of a standard \ndeviation, indicating a change in population mental \nhealth that is relatively substantial, especially in the \nshort amount of time between waves.\nModel 5 introduces controls for categories of \ncommunity distrust. The interaction between wave \nof survey and age in Model 5 was almost entirely \nunchanged compared to the coefficient for the same \ninteraction in Model 4. Community distrust there-\nfore does not explain age cohort contingencies in \nbetween-wave differences in psychological distress, \nbut this is to be expected because age did not moder-\nate between-wave differences in risk of community \ndistrust. However, the ordered logistic regression \nanalyses did show significant age contingencies in \nbetween-wave differences in risk of subjective \nsocial isolation. Model 6 shows that controlling for \ncategories of subjective social isolation reduces the \nsize of the interaction by approximately a third, and \nthis interaction is no longer significant. Moreover, \nancillary analyses showed that even at age 60, \nbetween-wave differences in psychological distress \nwere no longer significant once subjective social \nisolation was taken into account. That older respon-\ndents were more vulnerable to an increased risk in \nsubjective social isolation between 2019 and 2020 \ntherefore explains why older respondents were more \nat risk for an increase in psychological distress \nbetween waves of the survey.\nDiscussion\nA central basis of Durkheimian theory is in the con-\nsequences of social integration for population \nhealth (Berkman et\u00a0al. 2000). Working from this \nperspective, theorists have hypothesized that soci-\netal instability can lead to a loss of social integra-\ntion, with subsequent emotional ill effects (Abrutyn \nand Mueller 2016). The outbreak of the COVID-19 \npandemic presents a rare opportunity for a natural \nexperiment that permits comparison of the popula-\ntion both shortly prior to and after the initiation of \nsocial distancing measures. Our study therefore \npermits the examination of how the wide-scale \nalteration of established patterns of social interac-\ntions were associated with changes in psychologi-\ncal distress, as well as two likely mechanisms for \nthese effects.\nOur comparisons of two surveys of the Canadian \nworking population\u2014one prepandemic in September \n2019 and another in mid-March 2020 as the pan-\ndemic accelerated\u2014demonstrates that when the two \nsamples were adjusted to be compositionally equal, \nthere was evidence of an increase in population levels \nof psychological distress during the pandemic. Two \nfactors indicative of a loss of social integration con-\ntributed to explaining the rise in psychological dis-\ntress. Both subjective social isolation and community \ndistrust increased substantially in the intervening six \nmonths, with the growth in subjective social isolation \nespecially explaining changes in the population level \nof psychological distress. However, subsequent anal-\nyses showed important life course contingencies in \nthese effects, with increases in subjective social isola-\ntion and consequent psychological distress far more \npervasive among middle-aged and older individuals.\nFigure 2.\u2002 Odds Ratios of Increased Risk in \nSubjective Isolation across Ages.\nNote: Dark bars indicate statistically significant odds \nratios, and light bars indicate nonsignificant odd ratios. \nDifference at age 30 is significant at p < .05; for later \nages, it is significant at p < .001.\n\n410\nTable 3.\u2002 Ordinary Least Squares Regression Analyses of Psychological Distress.\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nModel 6\n\u2002\nb\nSE\np\nb\nSE\np\nb\nSE\np\nb\nSE\np\nb\nSE\np\nb\nSE\np\nFocal predictors\n\u2003 Change from 2019 to 2020\n.060\n.023\n**\n.052\n.023\n*\n\u2212.015\n.019\n.052\n.024\n*\n.044\n.024\n\u2212.020\n.020\n\u2002\n\u2003 Age\n\u2212.021\n.001\n***\n\u2212.020\n.001\n***\n\u2212.014\n.001\n***\n\u2212.023\n.001\n***\n\u2212.022\n.001\n***\n\u2212.015\n.001\n***\n\u2003 Survey \u00d7 Age\n.004\n.002\n*\n.004\n.002\n*\n.002\n.001\n\u2002\n\u2003 Community distrusta\n\u2002\n\u2003 \u2003 Somewhat agree with trust in neighbors\n.047\n.026\n.011\n.022\n.046\n.026\n.011\n.022\n\u2002\n\u2003 \u2003 Somewhat disagree/strongly disagree with \ntrust in neighbors\n.153\n.039\n***\n.069\n.032\n*\n.152\n.039\n***\n.068\n.032\n*\n\u2003 Subjective isolationb\n\u2003 \u2003 A little of the time\n.366\n.024\n***\n.364\n.024\n***\n\u2003 \u2003 Some of the time\n.721\n.026\n***\n.720\n.026\n***\n\u2003 \u2003 Most of the time\n1.149\n.036\n***\n1.147\n.037\n***\n\u2003 \u2003 All of the time\n1.746\n.078\n***\n1.747\n.078\n***\nControl measures\n\u2003 Generalized trustc\n\u2003 \u2003 2\n\u2212.245\n.050\n***\n\u2212.241\n.050\n***\n\u2212.156\n.040\n***\n\u2212.247\n.050\n***\n\u2212.242\n.050\n***\n\u2212.156\n.040\n***\n\u2003 \u2003 3\n\u2212.376\n.046\n***\n\u2212.361\n.046\n***\n\u2212.232\n.036\n***\n\u2212.378\n.046\n***\n\u2212.363\n.046\n***\n\u2212.233\n.036\n***\n\u2003 \u2003 4\n\u2212.472\n.048\n***\n\u2212.444\n.048\n***\n\u2212.299\n.038\n***\n\u2212.473\n.048\n***\n\u2212.445\n.048\n***\n\u2212.299\n.038\n***\n\u2003 \u2003 Most people can be trusted\n\u2212.630\n.059\n***\n\u2212.594\n.060\n***\n\u2212.395\n.050\n***\n\u2212.630\n.059\n***\n\u2212.595\n.059\n***\n\u2212.395\n.050\n***\n\u2003 Occupational classd\n\u2003 \u2003 Administrative\n.057\n.038\n.059\n.037\n.039\n.031\n.056\n.037\n.058\n.037\n.038\n.031\n\u2002\n\u2003 \u2003 Sales\n.060\n.034\n.057\n.034\n.028\n.027\n.061\n.034\n.059\n.034\n.029\n.027\n\u2002\n\u2003 \u2003 Clerical\n\u2212.069\n.040\n\u2212.071\n.039\n\u2212.049\n.032\n\u2212.069\n.040\n\u2212.071\n.040\n\u2212.049\n.032\n\u2002\n\u2003 \u2003 Laborer\n.035\n.043\n.032\n.043\n.050\n.037\n.036\n.043\n.033\n.042\n.051\n.037\n\u2002\n\u2003 \u2003 Work hourse\n\u2003 \u2003 Full-time\n\u2212.025\n.033\n\u2212.027\n.033\n.002\n.027\n\u2212.027\n.033\n\u2212.029\n.033\n.001\n.027\n\u2002\n\u2003 \u2003 Extended hours\n\u2212.098\n.044\n*\n\u2212.097\n.044\n*\n\u2212.080\n.037\n*\n\u2212.102\n.045\n*\n\u2212.100\n.045\n*\n\u2212.083\n.037\n*\n\u2003 \u2003 Working multiple jobs\n.034\n.029\n.035\n.029\n.003\n.023\n.033\n.029\n.034\n.029\n.002\n.023\n\u2002\n\u2003 Work at homef\n\u2003 \u2003 A few times a year\n.051\n.039\n.050\n.038\n.034\n.031\n.051\n.039\n.050\n.038\n.034\n.031\n\u2002\n\u2003 \u2003 About once a month\n.081\n.047\n.081\n.047\n.092\n.039\n*\n.084\n.047\n.084\n.047\n.093\n.039\n*\n\u2003 \u2003 About once a week\n.102\n.042\n*\n.102\n.042\n*\n.065\n.032\n*\n.102\n.041\n*\n.102\n.041\n*\n.065\n.032\n*\n\u2003 \u2003 More than once a week\n.152\n.038\n***\n.153\n.038\n***\n.101\n.031\n**\n.154\n.038\n***\n.155\n.038\n***\n.102\n.031\n**\n\u2003 \u2003 Every day/mainly work at home\n.118\n.037\n**\n.122\n.037\n**\n.068\n.030\n*\n.122\n.037\n**\n.125\n.037\n**\n.070\n.030\n*\n\u2003 Control over work schedulingg\n\u2003 \u2003 Very little\n.005\n.039\n.004\n.039\n.028\n.033\n.003\n.039\n.003\n.039\n.027\n.033\n\u2002\n\u2003 \u2003 Some\n\u2212.003\n.038\n\u2212.003\n.038\n.046\n.032\n\u2212.003\n.038\n\u2212.003\n.038\n.046\n.032\n\u2002\n(continued)\n\n411\nModel 1\nModel 2\nModel 3\nModel 4\nModel 5\nModel 6\n\u2002\nb\nSE\np\nb\nSE\np\nb\nSE\np\nb\nSE\np\nb\nSE\np\nb\nSE\np\n\u2003 \u2003 A lot\n\u2212.068\n.039\n\u2212.066\n.038\n.001\n.033\n\u2212.069\n.038\n\u2212.068\n.038\n.000\n.033\n\u2002\n\u2003 \u2003 Complete control\n\u2212.164\n.047\n***\n\u2212.162\n.047\n**\n\u2212.057\n.039\n\u2212.168\n.047\n***\n\u2212.166\n.047\n***\n\u2212.060\n.039\n\u2002\n\u2003 Not living with romantic partner\n.050\n.028\n.048\n.028\n\u2212.031\n.023\n.053\n.028\n.050\n.028\n\u2212.030\n.023\n\u2002\n\u2003 Any children in household\n\u2212.104\n.026\n***\n\u2212.099\n.026\n***\n\u2212.092\n.022\n***\n\u2212.100\n.026\n***\n\u2212.096\n.026\n***\n\u2212.090\n.022\n***\n\u2003 Educationh\n\u2003 \u2003 High school\n.098\n.046\n*\n.093\n.046\n*\n.063\n.036\n.104\n.046\n*\n.099\n.046\n*\n.066\n.036\n\u2002\n\u2003 \u2003 Some university or college/trade school\n.023\n.034\n.023\n.034\n.060\n.028\n*\n.023\n.034\n.023\n.034\n.060\n.028\n*\n\u2003 \u2003 College/trade school\n\u2212.030\n.030\n\u2212.030\n.030\n.007\n.025\n\u2212.028\n.030\n\u2212.028\n.030\n.008\n.025\n\u2002\n\u2003 Incomei\n\u2003 \u2003 < $25,000\n.074\n.067\n.066\n.066\n\u2212.028\n.055\n.074\n.067\n.066\n.066\n\u2212.028\n.055\n\u2002\n\u2003 \u2003 $25,000 to < $50,000\n.129\n.046\n**\n.122\n.046\n**\n.084\n.038\n*\n.126\n.046\n**\n.119\n.046\n**\n.082\n.038\n*\n\u2003 \u2003 $50,000 to < $100,000\n.022\n.035\n.017\n.035\n\u2212.012\n.030\n.019\n.035\n.014\n.035\n\u2212.013\n.030\n\u2002\n\u2003 \u2003 $100,000 to < $150,000\n.044\n.034\n.043\n.034\n.034\n.029\n.043\n.034\n.042\n.034\n.033\n.029\n\u2002\n\u2003 \u2003 Missing income\n.005\n.047\n.001\n.047\n\u2212.025\n.040\n.002\n.047\n\u2212.002\n.047\n\u2212.027\n.040\n\u2002\n\u2003 Finances at end of monthj\n\u2003 \u2003 A little money left over\n.113\n.037\n**\n.113\n.038\n**\n.093\n.030\n**\n.116\n.037\n**\n.116\n.038\n**\n.095\n.030\n**\n\u2003 \u2003 Just enough to make ends meet\n.289\n.042\n***\n.287\n.042\n***\n.233\n.033\n***\n.291\n.042\n***\n.289\n.042\n***\n.234\n.033\n***\n\u2003 \u2003 Barely enough to get by\n.467\n.048\n***\n.462\n.048\n***\n.317\n.039\n***\n.468\n.048\n***\n.463\n.048\n***\n.318\n.039\n***\n\u2003 \u2003 Not enough to make ends meet\n.614\n.065\n***\n.611\n.066\n***\n.429\n.053\n***\n.613\n.065\n***\n.610\n.066\n***\n.428\n.053\n***\n\u2003 Women\n.123\n.025\n***\n.124\n.025\n***\n.129\n.020\n***\n.123\n.025\n***\n.124\n.025\n***\n.129\n.020\n***\n\u2003 Visible minority\n.052\n.036\n.047\n.036\n\u2212.023\n.028\n.052\n.036\n.046\n.036\n\u2212.023\n.028\n\u2002\nConstant\n2.395\n.075\n***\n2.338\n.077\n***\n1.898\n.064\n***\n2.400\n.075\n***\n2.343\n.077\n***\n1.902\n.064\n\u2002\nR2\n.249\n.252\n.480\n.250\n.253\n.481\n\u2002\nNote: N = 4,923. Data are from The Canadian Quality of Work and Economic Life Study.\naStrongly agree with trust in neighbors is reference.\nbNone of the time is reference.\ncYou can\u2019t be too careful is reference.\ndProfessional is reference.\nePart-time is reference.\nfNever is reference.\ngNone is reference.\nhUniversity degree is reference.\ni\u2265 $150,000 is reference.\njA lot of money left over is reference.\n*p < .05, **p < .01, ***p < .001, two-tailed.\nTable 3.\u2002 (continued)\n\n412\t\nJournal of Health and Social Behavior 61(4) \nFigure 3.\u2002 Standardized Difference in \nPsychological Distress across Ages.\nNote: Dark bars indicate statistically significant \ndifferences, and light bars indicate nonsignificant \ndifferences. Difference at age 40 is significant at p < .05 \nand is significant at p < .001 for later ages.\nThese findings are especially notable because \nthey provide an important qualification to the \nintended public health protections intended by the \nCOVID-19 pandemic social distancing measures. \nAlthough evidence suggests that social distancing \nmeasures are critical for helping limit the spread of \ninfection and strains on the health care system \n(Delen, Eryarsoy, and Davazdahemami 2020; \nLewnard et\u00a0al. 2020), it has also been suggested the \nsocial consequences of stay-at-home orders may \nhave had psychological costs (Douglas et\u00a0al. 2020; \nTull et\u00a0al. 2020). Our analyses provide support for \nthese concerns, suggesting that the necessity of \nsocial distancing was concomitant with a rise in \npsychological distress. The mental health costs of \nsocial distancing are especially important to take \ninto account because these measures also curtailed \nindividuals\u2019 abilities to seek out medical or thera-\npeutic assistance for increased distress. Advocates \nhave suggested that a critical response to social dis-\ntancing policies is the fortification of programs and \nmechanisms that will help to address a surge in \nmental health problems during the pandemic \n(Galea, Merchant, and Lurie 2020), and the results \nof the current research support these proposals.\nSome have also suggested that psychological \nvulnerability to the adverse effects of social dis-\ntancing may be differentially distributed in the pop-\nulation, with older adults especially at risk of social \nisolation and subsequent adverse consequences for \nmental health (Armitage and Nellums 2020). The \nresults of the current research support these con-\ncerns as well because respondents at older ages \nexperienced greater risk of increases in a sense of \nisolation, with subsequent heightened increases in \npsychological distress. Furthermore, the survey \ndata analyzed in this study are intended to be repre-\nsentative of Canadian workers, which underrepre-\nsents the larger population of older adults, many of \nwhom are retired. Because the working population \nwill tend to have at least some social contact \nthrough interwork relations, it is likely that this \nstudy minimizes the consequences of social dis-\ntancing measures for a sense of isolation among the \nlarger population of older adults. The risks to the \npsychological well-being of older adults as a result \nof COVID-19 social distancing measures are there-\nfore likely even stronger than those presented here.\nThese findings support a more nuanced perspec-\ntive on Durkheimian expectations regarding the \nconsequences of social change for social integra-\ntion. Fundamental to a Durkheimian perspective is \nthat \u201cthe social fabric is eroded by rapid social \nchange\u201d (Turner 2003:9), thereby emphasizing the \ndissolution of social integration in times of social \nchange. From a life course perspective, however, \nbirth cohorts are born with and acquire different \nresources and vulnerabilities as a consequence of \ntheir placement in historical time and place (Elder \n1994; Keyes et\u00a0al. 2010). An emphasis on cohort \nmembership is particularly relevant to the study of \ndisintegratory social change because some have \ncharacterized those born before 1980 as digital \nimmigrants and those born in 1980 and after as dig-\nital natives (Nevin and Schieman 2020; Prensky \n2001), with the result that older cohorts were likely \nto gain less social fulfillment from digital forms of \nsocial interactions that substituted for more conven-\ntional forms of interactions. In fact, our results fol-\nlowed this categorization. Respondents younger \nthan 40 (and therefore born after 1980) were much \nless at risk for increased feelings of social isolation \nthan those born after 1980. The current research \ntherefore suggests that cohorts may possess differ-\nent capabilities in conserving and maintaining \nsocial integration based not only on facilities with \ntechnology but also on ingrained patterns of social \npractices and expectations. For cohorts in which \npatterns of face-to-face contact and in-person meet-\nings are less common, there may be weaker suscep-\ntibility to reductions in social integration as a result \nof social change.\nA life course perspective invites further theoreti-\ncal refinement through its emphasis on timing \n(Elder 1999). A key tenet of a life course perspec-\ntive is that similar events and experiences can influ-\nence individuals differently depending on the \ntiming of these events in the life course (Elder et\u00a0al. \n2003), but Durkheimian perspectives have less \n\nBierman and Schieman\t\n413\nclearly articulated the degree to which social \nchange may vary in its influence on social integra-\ntion due to timing in the life course. It is likely that \nthe age-differentiated patterns we observed in the \ncurrent research are not only due to birth cohort dif-\nferences, but also to the degree to which older indi-\nviduals had more established patterns and bases of \nsocial interaction that could be disrupted by large-\nscale social change. Timing of the pandemic in the \nlife course as well as historical forces of cohort \nchange also likely shaped vulnerabilities to disinte-\ngratory influences of social distancing measures.\nA central theoretical contribution of the current \nresearch is therefore in the concept of integratory \nvulnerability\u2014in which a life course context differ-\nentiates vulnerability to disintegratory societal \nevents. The current research in fact underscores the \nimportance of considering life course context as a \nkey dimension of integratory vulnerability: Critical \ncontingencies in changes in subjective social isola-\ntion and psychological distress would have been \noverlooked in the absence of the insights provided \nby a life course perspective. The results of this \nresearch therefore suggest that attention to the life \ncourse context of integratory vulnerability will help \nboth theorists and empirical researchers to specify \nwhy and for whom the consequences of social \nchange on social integration and well-being are \nlikely to be especially pertinent. We especially wish \nto emphasize that both birth cohort as well as age \nare likely to contribute to integratory vulnerability. \nThe sociological study of life course contexts has a \npoor history of overemphasizing age to the neglect \nof birth cohort influences, and an appropriate level \nof theorizing and empirical study of life course con-\ntexts should take both of these dual influences into \naccount.\nAn emphasis on the contribution of birth cohort \nmembership to integratory vulnerability also extends \nprevious work framing cohort change as a disinte-\ngratory agent (O\u2019Brien and Stockard 2006; Stockard \nand O\u2019Brien 2002). This work essentially positions \nvariations in intercohort characteristics as proxies for \nsocial change, but a Durkheimian perspective often \nviews social change in broader terms, such as those \nof wide-scale economic collapse (Cockerham 2017). \nThe concept of integratory vulnerability extends \nthese ideas to propose that an understanding of the \nrole of birth cohort as an integratory influence is bet-\nter served through an intersectional emphasis: Social \nchange occurs at both a societal and cohort level, \nwith the result that the two intersect to influence indi-\nvidual outcomes. Thus, an important clarification \nand extension to a Durkheimian perspective in future \ntheorizing is to specify social change simultaneously \nat a societal and cohort level, with particular care to \nthe way the two forms of change may intersect to \nshape consequences for social integration.\nAlthough declining levels of social integration \ncan lead to individual feelings of isolation, more \nrecent theorizing has also linked loss of social inte-\ngration to an increased sense threat (Abrutyn 2019). \nRises in threat can in turn harm community trust \nbecause individuals become reticent to allow them-\nselves to be vulnerable in the context of trusting rela-\ntionships. Evidence from previous pandemics \nsupports these assertions, showing how fear can \nundermine levels of public trust and enhance selfish \nmotives (Barry 2005). We observed some evidence \nof this increase as well. Even in the short time \nbetween surveys, the odds of greater distrust in \nneighbors increased precipitously. Furthermore, we \nobserved an increasing risk of community distrust \neven though overall trust remained relatively consis-\ntent between waves of the survey. The increase spe-\ncifically in distrust of others in local surroundings \nsuggests that individuals began to look at one another \nmore suspiciously even if they did not perceive peo-\nple more generally as less trustworthy. Essentially, \nthe loss of bonds of integration and increasing threat \ninherent in social distancing measures led to less \ntrust in people with whom individuals were likely to \ncome into contact in the community.\nThere may be a hesitancy to attribute substantial \nmeaning to the loss of community trust because \ncommunity distrust only minorly explained between-\nwave differences in psychological distress. Yet, \nhealth is only one area of social life which is influ-\nenced by trust, as trust is a core dimension of human \nrelations. Without trust, individuals cannot engage in \nfundamental processes of reciprocity that serve to \nbuild equity in human relationships (Cialdini and \nGoldstein 2004). An erosion of trust could therefore \nlead to a loss of social order, as was observed in \nBarry\u2019s (2005:330) chronicle of Philadelphia during \nSpanish influenza pandemic, in which the city \n\u201cturned into itself. There was no trust, no trust, and \nwithout trust all human relations were breaking \ndown.\u201d Thus, increasing levels of distrust during the \nCOVID-19 pandemic could have even more sub-\nstantial consequences for the disturbance of social \norder as the pandemic continued beyond the early \nstages observed in the current study, especially if a \ngreater number of people shifted to more extreme \nlevels of distrust. These findings therefore also \nunderscore an important additional area for future \ntheoretical refinement. Although a primary emphasis \nin Durkheimian research is on the ramifications of \n\n414\t\nJournal of Health and Social Behavior 61(4) \nsocial integration for health, an important additional \ndirection of theoretical refinement is to consider how \ncontexts of decreasing social integration could also \ncreate more widespread harms to social order \nthrough a loss of social trust.\nSeveral limitations to this study should be noted. \nFirst, both community distrust and subjective social \nisolation were assessed with single-item measures. \nAlthough community distrust has previously been \nmeasured using single questions (e.g., Carpiano and \nFitterer 2014; Fujiwara and Kawachi 2008), single-\nitem measures typically have lower levels of reli-\nability than multiple-item scales. However, that we \nsee similar increases in both measures, as well as a \nreliable scale of distress, suggests that the changes \nobserved in these analyses are not simply due to \nrandom fluctuations caused by unreliability. In \naddition, it should be emphasized that although the \nchanges observed here are likely attributable to the \nCOVID-19 pandemic and associated social distanc-\ning measures, we cannot directly link any observed \nchanges to the COVID-19 pandemic. This does not \nweaken the underlying findings of this article that \nthe Canadian working population experienced a \ndramatic increase in social estrangement that led to \ngreater levels of psychological distress.\nOne additional theoretically motivated issue is \nthat we do not control for changes in religious \ninvolvement. Durkheim ([1897] 1951) emphasized \nthe importance of religious involvement as a source \nof social integration, and social distancing in \nresponse to COVID-19 necessitated the cessation of \nreligious attendance, thereby contributing to a loss \nof social integration. At the same time, additional \nresearch suggests that a key conduit for these effects \nis likely to be through an increase in subjective \nsocial isolation (Rote, Hill, and Ellison 2013), with \nthe implication that our focus on subjective social \nisolation absorbs the primary consequences of \ndeclines in religious attendance. Stress research has \nalso emphasized the importance of anticipatory \nstressors during times of social crisis (Pearlin and \nBierman 2013). During the COVID-19 pandemic, \nconcerns in regards to contracting the virus or future \neconomic hardship also likely exacted a toll on men-\ntal health as the pandemic continued and the econ-\nomy contracted beyond the time frame in the current \nstudy. Finally, and similar to the question of retired \nolder adults, these analyses focused on employed \nindividuals, and it is possible that unemployed indi-\nviduals may have also experienced an even greater \nincrease in subjective social isolation in the absence \nof social interactions with coworkers as well as a \ngreater burden due to economic stressors.\nConclusion\nThe COVID-19 pandemic represents a once-in-a-\nlifetime shock to social life across societies. From  \na Durkheimian perspective, it is unsurprising that \nwe observe consequences of such rapid and all-\u00ad\nencompassing social change for integration and, \nultimately, population health. In times of great \n\u00adturbulence and social disruption, it is critical to \nmaintain meaningful social ties, but sustaining \nthose bonds becomes increasingly more difficult, \nresulting in challenges to mental health. As the \n\u00adpandemic accelerated, we began to observe the \nexpected fallout for social life and mental health. \nHowever, the patterns are not equivalent across age. \nSubjective social isolation increased more dramati-\ncally during this period among older respondents, \nleading to a more substantial rise in psychological \ndistress during this period. This research therefore \npresents an important qualification to a Durkheimian \nperspective by demonstrating that a life course con-\ntext plays a crucial role in differentiating individual \nvulnerability to disintegrative large-scale social \nforces and their consequences for mental health.\nAcknowledgments\nWe thank Paul Glavin for his suggestions as this article \nwas being developed.\nDeclaration of Conflicting \nInterests\nThe authors declared no potential conflicts of interest with \nrespect to the research, authorship, and/or publication of \nthis article.\nFunding\nThe authors disclosed receipt of the following financial \nsupport for the research, authorship, and/or publication of \nthis article: Funding from the University of Toronto \nCOVID-19 Action Initiative 2020 and Tri-Council Bridge \nfunding supported this research (Scott Schieman, PI).\nORCID iDs\nAlex Bierman \n https://orcid.org/0000-0003-2573-0969\nScott Schieman \n https://orcid.org/0000-0003-2882-0417\nReferences\nAassve, Arnstein, Guido Alfani, Francesco Gandolfi, and Marco \nLe Moglie. 2020. \u201cEpidemics and Trust: The Case of the \nSpanish Flu.\u201d Working Papers 661 IGIER (Innocenzo \nGasparini Institute for Economic Research), Bocconi \nUniversity. http://www.igier.unibocconi.it/files/661.pdf.\n\nBierman and Schieman\t\n415\nAbrutyn, Seth. 2019. \u201cToward a General Theory of \nAnomie: The Social Psychology of Disintegration.\u201d \nEuropean Journal of Sociology 60(1):109\u201336.\nAbrutyn, Seth, and Anna S. Mueller. 2014. \u201cThe Socio\u00ad\nemotional Foundations of Suicide: A Microsocio\u00ad\nlogical View of Durkheim\u2019s Suicide.\u201d Sociological \nTheory 32(4):327\u201351.\nAbrutyn, Seth, and Anna S. Mueller. 2016. \u201cWhen \nToo Much Integration and Regulation Hurt: \nRe-envisioning Durkheim\u2019s Altruistic Suicide.\u201d \nSociety and Mental Health 6(1):56\u201371\nArmitage, Richard, and Laura B. Nellums. 2020. \n\u201cCOVID-19 and the Consequences of Isolating the \nElderly.\u201d Lancet Public Health 5(5):e256.\nAtherwood, Serge, and Corey S. Sparks. 2019. \u201cEarly-\nCareer Trajectories of Young Workers in the U.S. \nin the Context of the 2008\u201309 Recession: The \nEffect of Labor Market Entry Timing.\u201d PLos ONE \n14(3):e0214234.\nBarry, John M. 2005. The Great Influenza: The Story \nof the Deadliest Pandemic in History. New York: \nPenguin Group.\nBaumert, Anna, Anna Halmburger, Tobias Rothmund, \nand Christian Schemer. 2017. \u201cEveryday Dynamics \nin Generalized Social and Political Trust.\u201d Journal of \nResearch in Personality 69:44\u201354.\nBerkman, Lisa F., Thomas Glass, Ian Brissette, and \nTeresa E. Seeman. 2000. \u201cFrom Social Integration to \nHealth: Durkheim in the New Millennium.\u201d Social \nScience & Medicine 51(6):843\u201357.\nBrant, Rollin. 1990. \u201cAssessing Proportionality in the \nProportional Odds Model for Ordinal Logistic \nRegression.\u201d Biometrics 46(4):1171\u201378.\nB\u00fcchi, Moritz, Natascha Just, and Michael Latzer. 2016. \n\u201cModeling the Second-Level Digital Divide: A Five-\nCountry Study of Social Differences in Internet \nUse.\u201d New Media & Society 18(11): 2703\u201322.\nCacioppo, John T., and Stephanie Cacioppo. 2014. \n\u201cSocial Relationships and Health: The Toxic Effects \nof Perceived Social Isolation.\u201d Social and Personality \nPsychology Compass 8(2):58\u201372.\nCacioppo, John T., Louise C. Hawkley, Greg J. Norman, \nand Gary G. Berntson. 2011. \u201cSocial Isolation.\u201d \nAnnals of the New York Academy of Sciences \n1231(1):17\u201322.\nCarpiano, Richard M., and Lisa M. Fitterer. 2014. \n\u201cQuestions of Trust in Health Research on Social \nCapital: What Aspects of Personal Network Social \nCapital Do They Measure?\u201d Social Science & \nMedicine 116:225\u201334.\nCialdini, Robert B., and Noah J. Goldstein. 2004. \u201cSocial \nInfluence: Compliance and Conformity.\u201d Annual \nReview of Psychology 55:591\u2013621.\nCockerham, William C. 2017. Sociology of Mental \nDisorder. 10th ed. New York: Routledge.\nDahlberg, Lena, Lars Andersson, and Carin Lennartsson. \n2018. \u201cLong-Term Predictors of Loneliness in Old \nAge: Results of a 20-Year National Study.\u201d Aging & \nMental Health 22(2):190\u201396.\nDelen, \nDursun, \nEnes \nEryarsoy, \nand \nBehrooz \nDavazdahemami. 2020. \u201cNo Place Like Home: \nCross-National Data Analysis of the Efficacy of \nSocial Distancing during the COVID-19 Pandemic.\u201d \nJMIR Public Health and Surveillance 6(2):1\u201310.\nDouglas, Margaret, Srinivasa Vittal Katikireddi, Martin \nTaulbut, Martin McKee, and Gerry McCartney. \n2020. \u201cMitigating the Wider Health Effects of \nCOVID-19 Pandemic Response.\u201d BMJ 369:m1557.\nDurkheim, \u00c9mile. [1897] 1951. Suicide: A Study in \nSociology. Glencoe, IL: Free Press.\nElder, Glen H., Jr. 1994. \u201cTime, Human Agency, and \nSocial Change: Perspectives on the Life Course.\u201d \nSocial Psychology Quarterly 57(1):4\u201314.\nElder, Glen H., Jr. 1999. Children of the Great Depression: \nSocial Change in Life Experience. Boulder, CO: \nWestview Press.\nElder, Glen H., Monica K. Johnson, and Robert Crosnoe. \n2003. \u201cThe Emergence and Development of Life \nCourse Theory.\u201d Pp. 3\u201319 in Handbook of the Life \nCourse, edited by J. T. Mortimer and M. J. Shanahan. \nBoston, MA: Springer.\nFriemel, Thomas N. 2016. \u201cThe Digital Divide Has \nGrown Old: Determinants of a Digital Divide among \nSeniors.\u201d New Media & Society 18(2):313\u201331.\nFujirawa, T., and I. Kawachi. 2008. \u201cA Prospective \nStudy of Individual-Level Social Capital and \nMajor Depression in the United States.\u201d Journal of \nEpidemiology and Community Health 62(7):627\u201333.\nGalea, Sandro, Raina M. Merchant, and Nicole Lurie. \n2020. \u201cThe Mental Health Consequences of COVID-\n19 and Physical Distancing: The Need for Prevention \nand Early Intervention.\u201d JAMA Internal Medicine \n180(6):817\u201318.\nGilman, Sander L. 2010. \u201cMoral Panic and Pandemics.\u201d \nThe Lancet 375(9729):1866\u201367.\nGlanville, Jennifer L., and William T. Story. 2018. \n\u201cSocial Capital and Self-Rated Health: Clarifying the \nRole of Trust.\u201d Social Science Research 71:98\u2013108.\nGovernment \nof \nAlberta. \n2020. \n\u201cRestrictions \non \nGatherings and Businesses.\u201d https://www.alberta.ca/\nrestrictions-on-gatherings-and-businesses.aspx.\nHargittai, Eszter, and Kerry Dobransky. 2017. \u201cOld \nDogs, New Clicks: Digital Inequality in Skills and \nUses among Older Adults.\u201d Canadian Journal of \nCommunication 42(2):195\u2013212.\nHeymann, David L., and Nahoko Shindo. 2020. \u201cCOVID-\n19: What Is Next for Public Health?\u201d The Lancet \n395(10224):542\u201345.\nHoffmann, John P. 2016. Regression Models for \nCategorical, Count, and Related Variables: An \nApplied Approach. Oakland, CA: University of \nCalifornia Press.\nHoule, Jason N. 2014. \u201cMental Health in the Foreclosure \nCrisis.\u201d Social Science & Medicine 118:1\u20138.\nHoule, Jason N., and Michael T. Light. 2014. \u201cThe \nHome Foreclosure Crisis and Rising Suicide Rates, \n2005\u20132010.\u201d American Journal of Public Health \n104(6):1073\u201379.\n\n416\t\nJournal of Health and Social Behavior 61(4) \nKahn, Joan R., and Leonard I. Pearlin. 2006. \u201cFinancial \nStrain over the Life Course and Health among Older \nAdults.\u201d Journal of Health and Social Behavior \n47(1):17\u201331.\nKessler, Ronald C., Gavin Andrews, Lisa J. Colpe, Eva \nHiripi, Dan K. Mroczek, Sharon-Lise T. Normand, \nEllen E. Walters, and Alan M. Zaslavsky. 2002. \u201cShort \nScreening Scales to Monitor Population Prevalences \nand Trends in Non-specific Psychological Distress.\u201d \nPsychological Medicine 32(6):959\u201376.\nKeyes, Katherine M., Rebecca Utz, Whitney R. \nRobinson, and Guohua Li. 2010. \u201cWhat Is a Cohort \nEffect? Comparison of Three Statistical Methods \nfor Modeling Cohort Effects in Obesity Prevalence \nin the United States, 1971\u20132006.\u201d Social Science & \nMedicine 70(7):1100\u20131108.\nLai, K. K. Rebecca. 2020. \u201cAre Countries Flattening the \nCurve for Coronavirus.\u201d The New York Times, April \n3. https://nyti.ms/2UFD38P.\nLester, Bijou Yang. 2001. \u201cLearnings from Durkheim \nand Beyond: The Economy and Suicide.\u201d Suicide \nand Life-Threatening Behavior 31(1):15\u201331.\nLewnard, Joseph, Vincent X. Liu, Michael L. Jackson, \nMark A. Schmidt, Britta L. Jewell, Jean P. Flores, Chris \nJentz, Graham R. Northrup, Ayesha Mahmud, Arthur \nL. Reingold, Maya Petersen, Nicholas P. Jewell, Scott \nYoung, and Jim Bellows. 2020. \u201cIncidence, Clinical \nOutcomes, and Transmission Dynamics of Severe \nCoronavirus Disease 2019 in California and Washington: \nProspective Cohort Study.\u201d BMJ 369:m1923.\nLittle, William. 2016. Introduction to Sociology. 2nd \nCanadian ed. https://opentextbc.ca/introductionto \nsociology2ndedition/.\nLoewen, Claire. 2020. \u201cAs Quebec Bans Gatherings \nof 2 or More, Police Forces Given More Power to \nDisperse Groups.\u201d CBC News, March 23. https://\nwww.cbc.ca/news/canada/montreal/as-quebec-bans-\ngatherings-of-2-or-more-police-forces-given-more-\npower-to-disperse-groups-1.5506327.\nL\u00fcders, Marika, and Petter Bae Brandtz\u00e6g. 2017. \n\u201c\u2018My Children Tell Me It\u2019s So Simple\u2019: A Mixed-\nMethods Approach to Understand Older Non-users\u2019 \nPerceptions of Social Networking Sites.\u201d New Media \n& Society 19(2):181\u201398.\nMacKinnon, David P. 2008. Introduction to Statistical \nMediation Analysis. New York, NY: Taylor & Francis \nGroup.\nMaimon, David, and Danielle C. Kuhl. 2008. \u201cSocial Control \nand Youth Suicidality: Situating Durkheim\u2019s Ideas in a \nMultilevel Framework.\u201d American Sociological Review \n73(6):921\u201343.\nMcClendon, McKee J. 1994. Multiple Regression and \nCausal Analysis. Itasca, IL: F. E. Peacock.\nMorgan, Clancy. 2020. \u201cHow Long Will Social Distance \nLast? It\u2019s Complicated.\u201d Business Insider, March 30. \nhttps://www.businessinsider.com/why-social-distancing-\nmight-last-several-months-2020-3.\nMurayama, Hiroshi, Yu Nofuji, Eri Matsuo, Mariko \nNishi, Yu Taniguchi, Yoshinori Fujiwara, and \nShoji Shinkai. 2015. \u201cAre Neighborhood Bonding \nand Bridging Social Capital Protective against \nDepressive Mood in Old Age? A Multilevel \nAnalysis in Japan.\u201d Social Science & Medicine \n124:171\u201379.\nNevin, Andrew D., and Scott Schieman. 2020. \n\u201cTechnological Tethering, Digital Natives, and \nChallenges in the Work\u2013Family Interface.\u201d The \nSociological Quarterly, first published on January \n23, 2020. doi:10.1080/00380253.2019.1711264\nO\u2019Brien, Robert, and Jean Stockard. 2006. \u201cA Common \nExplanation for the Changing Age Distributions of \nSuicide and Homicide in the United States, 1930\u2013\n2000.\u201d Social Forces 84(3):1541\u201357.\nPearlin, Leonard I., and Alex Bierman. 2013. \u201cCurrent \nIssues and Future Directions in Research into the \nStress Process.\u201d Pp. 325\u201340 in Handbook of the \nSociology of Mental Health. 2nd ed, edited by C. S. \nAneshensel, J. C. Phelan, and A. Bierman. Dordrecht, \nthe Netherlands: Springer.\nPhillips, Julie A., and Colleen N. Nugent. 2014. \u201cSuicide \nand the Great Recession of 2007\u20132009: The Role \nof Economic Factors in the 50 US States.\u201d Social \nScience & Medicine 116:22\u201331.\nPrensky, \nMarc. \n2001. \n\u201cDigital \nNatives, \nDigital \nImmigrants, Part I.\u201d On the Horizon 9(5):1\u20136.\nPublic Health Agency of Canada. 2020. \u201cPhysical \u00addistancing \n(Fact Sheet).\u201d https://www.canada.ca/en/public-health/\nservices/publications/diseases-conditions/social-\u00ad\ndistancing.html.\nRoss, Catherine E. 2011. \u201cCollective Threat, Trust, and \nthe Sense of Personal Control.\u201d Journal of Health \nand Social Behavior 52(3):287\u201396.\nRote, Sunshine, Terrence Hill, and Christopher Ellison. \n2013. \u201cReligious Attendance and Loneliness in Later \nLife.\u201d The Gerontologist 53(1):39\u201350.\nSettels, Jason. 2020. \u201cChanges in City-Level Foreclosure \nRates and Home Prices through the Great Recession \nand Depressive Symptoms among Older Americans.\u201d \nSociety and Mental Health, first published January 2, \n2020. doi:10.1177/2156869319895568.\nSly, Timothy. 2020. \u201cCOVID-19 Has the Potential to \nBecome as Severe as the Spanish Flu.\u201d MacLean\u2019s, \nMarch 5. https://www.macleans.ca/opinion/covid-\n19-has-the-potential-to-become-as-severe-as-the-\nspanish-flu/.\nStockard, Jean, and Robert O\u2019Brien. 2002. \u201cCohort \nEffects on Suicide Rates: International Variations.\u201d \nAmerican Sociological Review 67(6):854\u201372.\nSwader, Christopher S. 2019. \u201cLoneliness in Europe: \nPersonal and Societal Individualism-Collectivism and \nTheir Connection to Social Isolation.\u201d Social Forces \n97(3):1307\u201336.\nTomita, Andrew, and Jonathan K. Burns. 2013. \u201cA \nMultilevel \nAnalysis \nof \nAssociation \nbetween \nNeighborhood Social Capital and Depression: \nEvidence from the First South African National \nIncome Dynamics Study.\u201d Journal of Affective \nDisorders 144(1\u20132):101\u2013105.\n\nBierman and Schieman\t\n417\nTull, Matthew T., Keith A. Edmonds, Kayla M. Scamaldo, \nJulia R. Richmond, Jason P. Rose, and Kim L. Gratz. \n2020. \u201cPsychological Outcomes Associated with Stay-\nat-Home Orders and the Perceived Impact of COVID-\n19 on Daily Life.\u201d Psychiatry Research 289:113098.\nTurner, Bryan S. 2003. \u201cSocial Capital, Inequality and \nHealth: The Durkheimian Revival.\u201d Social Theory \nand Health 1(1):4\u201320.\nTurner, J. Blake, and R. Jay Turner. 2013. \u201cSocial \nRelations, Social Integration, and Social Support.\u2019\u2019 \nPp. 341\u201356 in Handbook of the Sociology of Mental \nHealth. 2nd ed., edited by C. S. Aneshensel, J. C. \nPhelan, and A. Bierman. New York, NY: Springer.\nWilliams, Richard. 2006. \u201cGeneralized Ordered Logit/Partial \nProportional Odds Models for Ordinal Dependent \nVariables.\u201d The Stata Journal 6(1):58\u201382.\nWinfree, L. Thomas, Jr., and Shanhe Jiang. 2010. \u201cYouthful \nSuicide and Social Support: Exploring the Social \nDynamics of Suicide-Related Behavior and Attitudes \nwithin a National Sample of US Adolescents.\u201d Youth \nViolence and Juvenile Justice 8(1):19\u201337.\nWu, Yun-Hsuan, Kellee White, Nancy L Fleischer, Bo \nCai, Shing-Chia Chen, and Spencer Moore. 2018. \n\u201cNetwork-Based and Cohesion-Based Social Capital \nand Variations in Depressive Symptoms among \nTaiwanese Adults.\u201d International Journal of Social \nPsychiatry 64(8):726\u201336.\nYu, Rebecca P., Nicole B. Ellison, Ryan J. McCammon, \nand Kenneth M. Langa. 2016. \u201cMapping the Two \nLevels of Digital Divide: Internet Access and Social \nNetwork Site Adoption among Older Adults in the \nUSA.\u201d Information, Communication & Society \n19(10):1445\u201364.\nYuan, Shupei, Syed A. Hussain, Kayla D. Hales, and \nShelia R. Cotten. 2016. \u201cWhat Do They Like? \nCommunication Preferences and Patterns of Older \nAdults in the United States: The Role of Technology.\u201d \nEducational Gerontology 42(3):163\u201374.\nZhao, Ruohui, and Liqun Cao. 2010. \u201cSocial Change and \nAnomie: A Cross-National Study.\u201d Social Forces \n88(3):1209\u201329.\nAuthor Biographies\nAlex Bierman is associate professor in the Department of \nSociology at the University of Calgary. His research \nbroadly focuses on the degree to which social inequality \nand social change condition stress exposure and the conse-\nquences of stress exposure for well-being, particularly in \nthe context of aging.\nScott Schieman, professor of sociology at the University \nof Toronto, studies stress, work, and health in large \n\u00adpopulation-based studies. He is particularly interested in \nthe stress of higher status and its impact on broader social \ninequalities in health and well-being. He holds a Canada \nResearch Chair in the Social Contexts of Health.\n\n\n    Scientific Research Findings:", "answer": "\u2022 A sense of social isolation and lack of trust in neighbors both increased from September 2019 to March 2020.\n\u2022 Increases in feelings of isolation and lack of trust led to a rise in psychological distress in the population.\n\u2022 Increases in feelings of isolation were especially notable at later ages, which led to greater increases in distress.", "id": 80}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Sustainability\nnature sustainability\nhttps://doi.org/10.1038/s41893-025-01533-9\nArticle\nA full list of affiliations appears at the end of the paper. \n\u2009e-mail: shanshan.li@monash.edu\nRespiratory risks from wildfire-specific PM2.5 \nacross multiple countries and territories\nUnder a warming climate, wildfires are becoming more frequent and \nsevere. Multicountry studies evaluating associations between wildfire fine \nparticulate matter (PM2.5) and respiratory hospitalizations are lacking. \nHere we evaluate the short-term effects of wildfire-specific PM2.5 on \nrespiratory hospitalizations from 1,052 communities across Australia, \nBrazil, Canada, Chile, New Zealand, Vietnam, Thailand and Taiwan, during \n2000\u20132019. A 1\u2009\u00b5g\u2009m\u22123 increase in wildfire-specific PM2.5 was associated \nwith increased hospitalization risks for all-cause respiratory, asthma, \nchronic obstructive pulmonary disease, acute upper respiratory infection, \ninfluenza and pneumonia by 0.36%, 0.48%, 0.38%, 0.42%, 0.79% and \n0.36%, respectively. Higher risks were observed among populations \u226419 or \n\u226560\u2009years old, from low-income or high non-wildfire PM2.5 communities, \nand residing in Brazil, Thailand, Taiwan and Vietnam. Australia and New \nZealand exhibited a greater hospitalization risk for asthma associated with \nwildfire-specific PM2.5. Compared with non-wildfire PM2.5, wildfire-specific \nPM2.5 posed greater hospitalization risks for all respiratory diseases and \na greater burden of asthma. Wildfire-specific PM2.5 contributed to 42.4% \nof PM2.5-linked respiratory hospitalizations, dominating in Thailand. \nOverall, the substantial contribution of wildfire-specific PM2.5 to respiratory \nhospitalizations demands continued mitigation and adaptation efforts \nacross most countries. Intervention should be prioritized for influenza, \nchildren, adolescents, the elderly and populations in low-income or \nhigh-polluted communities.\nIn recent decades, wildfires, encompassing all types of fire in vegetated \nlandscapes, have been increasingly reported worldwide. This trend is \nparticularly evident in Australia, the Brazilian Amazon, Europe, Russia, \nCanada and the western United States1. Notably, Canada experienced \na record-breaking series of wildfires in 2023, which burned an esti-\nmated 185,000\u2009km2, impacting all provinces and territories2. Severe \nwildfires were also experienced in the United States and Europe dur-\ning the 2023 fire season3. Climate change is expected to increase the \nfrequency, duration and intensity of extreme wildfire events4. The \nglobal fire-prone area is projected to increase by 29%, by the end of \nthe twenty-first century5. This may result in amplified socioeconomic \nand health burdens6. Wildfires harm human health primarily through \nwildfire smoke (for example, toxic gases and particles), of which ~90% \nof the total particle mass is comprised of fine particulate matter with \ndiameters \u22642.5\u2009\u00b5m (PM2.5)7.\nCompared to urban-sourced PM2.5, the higher submicronic particle \ncontent8 and distinct chemical composition of wildfire-specific PM2.5 \nprobably contribute to its increased toxicity at equivalent doses1, \npotentially posing a greater threat to human health, with respira-\ntory complications as a particular concern6,9. Despite the substantial \nburden of wildfire-related health impacts on low- and middle-income \ncountries (LMICs)10, research investigating the health consequences \nReceived: 22 April 2024\nAccepted: 21 February 2025\nPublished online: xx xx xxxx\n Check for updates\nA list of authors and their affiliations appears at the end of the paper\n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\nRespiratory hospitalization risk and wildfire-specific PM2.5\nWildfire-specific PM2.5 generated persistently elevated risks in res-\npiratory hospitalizations on the current day and the previous day \nof wildfire-specific PM2.5 exposure (lag 0 and lag 1) (Extended Data \nFig. 1). The hospitalization risk of respiratory diseases peaked on the \nday of exposure (relative risk (RR)\u2009=\u20091.0030, 95% confidence interval \n(CI) 1.0027\u20131.0033), followed by a lag of 1\u2009day (RR\u2009=\u20091.0006, 1.0005\u2013\n1.0008), per 1\u2009\u00b5g\u2009m\u22123 increase in wildfire-specific PM2.5. The risks were \nmarkedly lower and the CIs contained 1, when longer lag day periods \nwere investigated (Extended Data Fig. 1). Therefore, in the final analyses, \nwe focused on cumulative RR over the lag of 0\u20131\u2009days, expressed as RR.\nWhen considering the pooled cumulative C\u2013R relationship \nbetween wildfire-specific PM2.5 and respiratory hospitalizations dur-\ning the lag 0\u20131\u2009days, we observed a monotonically increasing linear \nor supralinear curve for all categories of respiratory hospitalizations \nexamined. Overall, the RR of respiratory hospitalization from differ-\nent causes responded linearly to the elevating wildfire-specific PM2.5. \nThis finding was corroborated by community-specific quasi-Bayesian \ninformation criterion (qBIC) values. In >92% of the study communities, \nlinear models demonstrated a better fit than nonlinear models, as \nindicated by lower qBIC values (Supplementary Table 3). The strongest \nrelationship was observed for influenza, where the RR exhibited a nota-\nble initial increase before 20\u2009\u00b5g\u2009m\u22123, followed by a gentler incline up to \n50\u2009\u00b5g\u2009m\u22123. Similar C\u2013R patterns were observed for AURI, asthma and \npneumonia. The C\u2013R curve of COPD exhibited the most pronounced \ndegree of linearity (Fig. 2).\nOverall, wildfire-specific PM2.5 presented consistent positive \nassociations, for each of the respiratory hospitalizations examined. \nA 1\u2009\u00b5g\u2009m\u22123 increase in wildfire-specific PM2.5, during the lag 0\u20131\u2009days, \nwas significantly associated with elevated risks in all-cause respira-\ntory (RR\u2009=\u20091.0036, 1.0032\u20131.0038), asthma (RR\u2009=\u20091.0048, 1.0040\u2013\n1.0057), COPD (RR\u2009=\u20091.0038, 1.0032\u20131.0042), AURI (RR\u2009=\u20091.0042, \n1.0032\u20131.0050), influenza (RR\u2009=\u20091.0079, 1.0059\u20131.0096) and pneu-\nmonia (RR\u2009=\u20091.0036, 1.0032\u20131.0040) (Table 1). Sensitivity analyses \nconfirmed the robustness of our results (Supplementary Figs. 1 and 2 \nand Supplementary Table 8).\nWe found that age, community income level, community \nnon-wildfire PM2.5 level and country or territory significantly modi-\nfied the association between wildfire-specific PM2.5 and respiratory \nhospitalization risks (Fig. 3 and Supplementary Table 5). Specifically, \nindividuals \u226419 or \u226560\u2009years old presented a greater hospitalization \nrisk in all-cause respiratory disease and pneumonia, compared with \nthose aged 20\u201359\u2009years old. Greater risks were observed in communi-\nties with high non-wildfire PM2.5 levels for all-cause respiratory disease, \nCOPD and pneumonia. In contrast, low non-wildfire PM2.5 commu-\nnities presented a higher hospitalization risk for asthma. Further-\nmore, individuals in low-income communities experienced higher \nhospitalization risks from all-cause respiratory diseases, COPD and \npneumonia, compared to higher-income communities. Conversely, \ngreater hospitalization risks for asthma were observed in high-income \ncommunities. At the country and territory level, residents of Taiwan \ngenerally experienced higher cause-specific risks than other countries \nand territories. Whereas greater risks were observed, for all-cause \nrespiratory disease in Brazil, Thailand and Vietnam; asthma in New \nZealand and Australia; COPD, influenza and pneumonia in Brazil; and \nAURI in Canada (Supplementary Table 5).\nAnnual attributable burden\nOverall, 25,321 (20,478\u201330,114) all-cause respiratory hospitaliza-\ntions were attributed to wildfire-specific PM2.5 annually, equating to \nan attributable fraction (AF) of 1.42% (1.15\u20131.69%). For each year, on \naverage, up to 1.49% (0.86\u20132.10%) of asthma, 1.30% (0.89\u20131.70%) of \nCOPD, 1.77% (0.88\u20132.64%) of AURI, 2.84% (0.90\u20134.41%) of influenza and \n1.57% (1.24\u20131.89%) of pneumonia hospitalizations, were attributed to \nwildfire-specific PM2.5. When community income levels were examined, \nof wildfire-specific PM2.5 in these regions remains surprisingly limited. \nPrevious investigations have largely been conducted in high-income \ncountries, such as the United States and Canada, focusing on a specific \nlocal area or fire event9,11\u201317. This focus potentially hinders the generaliz-\nability of findings to other countries disproportionately burdened by \nwildfires, such as Chile and Brazil. Furthermore, studies in LMICs are \nimperative to inform targeted intervention and resource allocation \ndirected towards areas with the greatest need, ultimately contributing \nto global health equity. To date, only two multicountry studies have \ninvestigated the impacts of wildfire-specific PM2.5 on respiratory health. \nOne investigated a single outcome (acute respiratory infection) within \na specific population (children)18 and the other one focused solely on \nall-cause respiratory mortality19, without covering other major types \nof respiratory diseases.\nAdditionally, few multicountry studies have quantified the \nspatiotemporal burden of respiratory diseases attributable to \nwildfire-specific PM2.5, which limits their ability to account for the \ndynamic vulnerability of populations over time. Further, previous \ninvestigations used heterogeneous modelling approaches. This poses \nmethodological challenges in establishing a more representative \nassociation between short-term wildfire-specific PM2.5 exposure \nand respiratory diseases. Finally, few existing literature compared \nthe association of respiratory hospitalizations with wildfire and \nnon-wildfire PM2.5 exposure. Therefore, a systematic analysis of the \nrisks and burden of cause-specific respiratory diseases associated \nwith wildfire-specific PM2.5 is warranted, to provide more compara-\nble evidence, by applying a unified analytic framework to different \ncommunities.\nHere, by leveraging hospitalization data from eight countries and \nterritories, over 20\u2009years, we aimed to: (1) characterize the concen-\ntration\u2013response (C\u2013R) relationship between wildfire-specific PM2.5 \nand hospitalization risks for all-cause respiratory, asthma, chronic \nobstructive pulmonary disease (COPD), acute upper respiratory \ninfection (AURI), influenza and pneumonia; (2) identify vulnerable \npopulations; (3) quantify the spatiotemporal trends of respiratory \nhospitalizations attributable to wildfire-specific PM2.5; and (4) compare \nthe health impacts of wildfire and non-wildfire PM2.5 on respiratory \nhospitalizations.\nResults\nSummary statistics\nAnnually, 1,777,941 all-cause respiratory, 130,618 asthma, 223,661 \nCOPD, 97,275 AURI, 38,502 influenza and 692,691 pneumonia hospi-\ntalizations were included in our analyses. These data covered 1,052 \ncommunities, in eight countries and territories, during different \nperiods of 2000\u20132019. On average, there were approximately five \nall-cause respiratory hospitalizations, two pneumonia hospitaliza-\ntions and less than one hospitalization from asthma, COPD, AURI \nand influenza per day (Supplementary Table 1). Countries and ter-\nritories contributed to a mean of 16.5\u2009years (s.d.\u2009=\u20095.2) of the data. \nBrazil contributed the highest number of respiratory hospitaliza-\ntions, followed by Thailand (Supplementary Table 2). Overall, males, \nchildren and adolescents, the elderly \u226560\u2009years and individuals from \nlow-income communities or areas with higher non-wildfire PM2.5 lev-\nels bore a slightly larger proportion of respiratory hospitalizations \nrelative to their counterparts (Supplementary Table 2). The median \ndaily concentration of wildfire-specific PM2.5 and non-wildfire PM2.5 \nacross study communities was 1.4 (interquartile range (IQR) 3.6, \nrange 0.0\u2013169.2)\u2009\u00b5g\u2009m\u22123 and 13.4 (IQR 11.0, range 0.0\u2013198.4)\u2009\u00b5g\u2009m\u22123, \nrespectively. The median daily concentration of non-wildfire PM2.5 \nwas highest in Chile, followed by Thailand and Taiwan (Supplemen-\ntary Table 1). The average daily concentration of wildfire-specific \nPM2.5 was pronounced in communities of Brazil, Thailand, Chile, \nNew South Wales Australia, Vietnam and Northwest Canada (Fig. 1 \nand Supplementary Table 1).\n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\na larger attributable burden was observed in low-income communi-\nties for most respiratory diseases and in middle-income communi-\nties for asthma. When examined at the country and territory level, the \nhighest cause-specific AFs were found in Taiwan, ranging from 2.29% \n(1.82\u20132.76%) for all-cause respiratory diseases to 3.78% (1.65\u20135.86%) \nfor AURI. In contrast, a relatively lower AF was observed in Canada for \nasthma (0.45% (0.12\u20130.77%)), COPD (0.19% (0.04\u20130.33%)) and AURI \n(0.62% (0.18\u20131.05%)) (Table 2 and Supplementary Table 7).\nComparisons between wildfire-specific and non-wildfire PM2.5\nWildfires only constituted a small portion (17.1%) of ambient PM2.5, \ncompared to non-wildfire sources, with a median concentration of \n1.4 (IQR 3.6)\u2009\u00b5g\u2009m\u22123 and 13.4 (IQR 11.0)\u2009\u00b5g\u2009m\u22123, respectively. However, \nwildfire-specific PM2.5 was associated with a significantly greater risk \nand contributed to a substantial proportion of hospitalizations for \nall the major types of respiratory diseases examined, compared with \nnon-wildfire PM2.5 (Fig. 4 and Supplementary Tables 10 and 11).\nThe spatiotemporal pattern of annual AF of all-cause respiratory \nhospitalization due to wildfire-specific PM2.5 varied. We found an over-\nall increasing trend in annual AF. This pattern extended to Australia, \nCanada, Chile, New Zealand and Taiwan during 2000\u20132019. Taiwan \nexperienced an exceptional rise during the most recent 5\u2009years. In con-\ntrast, the annual AF in Brazil and Vietnam decreased over 2005\u20132019 \nand 2000\u20132019, respectively (Extended Data Fig. 2).\nTaiwan\nVietnam\nChile\nAustralia\nBrazil\n0\n5\n10\n15\n20\n25\nNo data\nCanada\nAverage daily wildfire-specific PM2.5 (\u00b5g m\u20133)\nThailand\nNew Zealand\nFig. 1 | Geographical context of the wildfire-specific PM2.5 and study communities. The spatial distribution of average daily wildfire-specific PM2.5 (\u00b5g\u2009m\u22123) across \n1,052 study communities in Canada, Brazil, New Zealand, Thailand, Australia, Chile, Vietnam and Taiwan during 2000\u20132019.\n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\nFor most study periods, wildfire-specific PM2.5 contributed to a \nsubstantial proportion of PM2.5-related respiratory hospitalizations, \nespecially in Brazil, Chile, New Zealand, Thailand and Taiwan. In Aus-\ntralia, the annual AF difference (wildfire AF minus non-wildfire AF) \nexhibited a continuous upward trend from 2005 to 2019, cumulating in \nwildfire very close to non-wildfire for the first time in 2015\u20132019. In Bra-\nzil, wildfire-specific PM2.5 carried a greater annual AF than non-wildfire \nPM2.5 in 2010\u20132014, with the annual AF difference declining over time. A \nnotable reduction was exhibited in the contribution of wildfire-specific \nPM2.5 to the PM2.5-related AF from 2010\u20132014 to 2015\u20132019. Declining \nannual AF of wildfire PM2.5 in Canada and New Zealand (2005\u20132019) \nwas accompanied by a growing annual AF difference, especially in \nNew Zealand, where wildfire smoke emerged as the dominant source \nof PM2.5-linked respiratory hospitalizations. From 2005 to 2014, Chile \nexperienced a rising trend in annual AF of wildfire PM2.5. In Taiwan, both \nannual AF of wildfire-specific PM2.5 and AF differences increased from \n2000 to 2019, respectively. In Vietnam, annual AF of wildfire-specific \nand non-wildfire PM2.5 increased from 2000 to 2019, with their differ-\nences decreasing (Extended Data Fig. 2).\nDiscussion\nTo the best of our knowledge, this is the largest study investigating \nthe association between acute exposure to wildfire-specific PM2.5 \nand cause-specific respiratory hospitalizations. Overall, short-term \nwildfire-specific PM2.5 exposure was linearly associated with elevated \nrisks in various respiratory hospitalizations, especially for influ-\nenza, individuals \u226419 or \u226560\u2009years old, populations in low-income or \nhigh-polluted communities and residents of Brazil, Vietnam, Thai-\nland and Taiwan. An estimated 1.42% (1.15\u20131.69%) of all-cause res-\npiratory hospitalizations were attributable to wildfire-specific PM2.5, \nincreasing overall, in Australia, Taiwan and Vietnam during various \nperiods between 2010 and 2019. Compared with non-wildfire PM2.5, \nwildfire-specific PM2.5 posed a greater hospitalization risk for all the \nmajor type of respiratory diseases. Wildfire emerged as a notable source \nof PM2.5-linked respiratory hospitalizations overall, in Brazil, Chile, New \nZealand, Thailand and Taiwan.\nOur findings of the linear C\u2013R relationship20,21 and the elevated \nrisks of respiratory hospitalizations with short-term exposure to \nwildfire-specific PM2.5 were broadly in line with previous studies11\u201319. \nHowever, they mostly focused on several cities or territories within a \nsingle country and reported highly heterogeneous risk estimates. For \nexample, a study in Darwin, Australia, found a 9.1% (RR\u2009=\u20091.091, 1.023\u2013\n1.163) increase in all-cause respiratory hospitalizations per 10\u2009\u00b5g\u2009m\u22123 \nincrease in PM2.5 during bushfire events, one day after exposure22. \nAnother study of the 2003 southern California wildfires found a 2.8% \n(RR\u2009=\u20091.028, 1.014\u20131.041) increase in respiratory hospitalizations for \n10\u2009\u00b5g\u2009m\u22123 increase in moving average PM2.5 exposure during the wildfire \nperiod23. Others found that a 10\u2009\u03bcg\u2009m\u22123 increase in wildfire-related PM2.5 \nwas associated with a 5.09% (RR\u2009=\u20091.0509, 1.0473\u20131.0544) increase in \nrespiratory hospital admissions over 0\u20131\u2009days after the exposure in \nBrazil during 2000\u2013201514. The effect varied across studies possibly \ndue to factors including study population and period, wildfire severity, \nexposure assessment and time window and modelling strategy. Our \nunified assessment approach across countries and territories ensured \nrobust comparability and avoids potential publication bias.\nThe underlying mechanism for the health effects of \nwildfire-specific PM2.5 remains unclear. Nonetheless, the potential \npathways by which PM2.5 affects the respiratory system may also apply \nto wildfire-specific PM2.5, including injury from free radical peroxida-\ntion, unbalanced intracellular calcium homeostasis and inflamma-\ntion24. Other mechanisms for influenza and asthma are provided in \nSupplementary Discussion 1. In particular, the greater susceptibility \nof influenza could be due to greater transmission and gene modifica-\ntion of the influenza virus. Specifically, ambient PM2.5 is suggested as \na direct transmission mode for influenza virus infection to the human \nalveolar epithelium, with nearly half (47%) of inhaled PM2.5 reaching the \nalveolar epithelium, the primary target site for influenza infection25. \nFurthermore, PM2.5 may not only act as a carrier but also influence the \nvirus itself. PM2.5 components potentially modify the influenza virus \ngenome25 and pre-exposure to PM2.5 may alter the antiviral response \nof bronchial epithelial cells, increasing their susceptibility to infec-\ntion26. The unique characteristics of wildfire-specific PM2.5, including \nb    Asthma\nc    COPD\na    All-cause respiratory\n1.4\n1.2\n1.0\n1.4\n1.2\n1.0\n0\n10\n20\n30\n40\n50\n0\n10\n20\n30\n40\n50\n0\n10\n20\nWildfire-specific PM2.5 (\u00b5g m\u20133)\nWildfire-specific PM2.5 (\u00b5g m\u20133)\nWildfire-specific PM2.5 (\u00b5g m\u20133)\n30\n40\n50\n2 \u00d7 10\n6\n1 \u00d7 10\n6\n0\n2 \u00d7 10\n6\n1 \u00d7 10\n6\n0\ne    Influenza\nf    Pneumonia\nd    AURI\nRR (95% CI)\nRR (95% CI)\nNo. of observations\nNo. of observations\nFig. 2 | The pooled C\u2013R curves for respiratory hospitalization risks from \nwildfire-specific PM2.5 exposure during the lag 0\u20131\u2009days. a\u2013f, C\u2013R curves for \nall-cause respiratory disease (a), asthma (b), COPD (c), AURI (d), influenza (e) \nand pneumonia (f). The centre of the error bands represents the mean effect \nestimates comparing days with lower versus higher wildfire-specific PM2.5 levels \nacross 1,052 study communities, with error bands indicating uncertainty ranges \nderived from the 95% CIs of 1,000 Monte Carlo samples. All the models were \ncontrolled for temperature, relative humidity, non-wildfire PM2.5, day of week, \nseasonality and long-term trend.\n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\nTable 1 | Cumulative relative risk of cause-specific respiratory hospitalizations (95% CI) associated with per 1\u2009\u00b5g\u2009m\u22123 increase \nin wildfire-specific PM2.5 exposure during the lag 0\u20131\u2009days across countries and territories\nSubgroups\nAll-cause respiratorya\nAsthmaa\nCOPDa\nAURIa\nInfluenzaa\nPneumoniaa\nOverall\n1.0025 (1.0023, \n1.0028)\n1.0035 (1.0028, \n1.0041)\n1.0027 (1.0023, \n1.0031)\n1.0029 (1.0024, \n1.0035)\n1.0056 (1.0042, \n1.0070)\n1.0025 (1.0022, 1.0028)\nCountry/territory\n\u2003 Australia\n1.0014 (1.0009, \n1.0019)a\n1.0063 (1.0049, \n1.0077)a\n1.0024 (1.0014, \n1.0034)a\n1.0001 (0.9988, \n1.0015)a\n1.0053 (1.0011, \n1.0095)a\n1.0007 (0.9997, 1.0016)a\n\u2003 Brazil\n1.0030 (1.0027, \n1.0034)a\n1.0032 (1.0023, \n1.0042)a\n1.0037 (1.0031, \n1.0044)a\n1.0039 (1.0027, \n1.0051)a\n1.0057 (1.0032, \n1.0082)a\n1.0029 (1.0025, 1.0033)a\n\u2003 Canada\n1.0006 (0.9999, \n1.0014)a\n1.0038 (1.0010, \n1.0066)a\n1.0015 (1.0003, \n1.0027)a\n1.0052 (1.0015, \n1.0089)a\n1.0037 (0.9899, \n1.0176)a\n1.0007 (0.9996, 1.0019)a\n\u2003 Chile\n1.0006 (1.0001, \n1.0010)a\n1.0001 (0.9973, \n1.0029)a\n1.0003 (0.9998, \n1.0007)a\n1.0000 (0.9981, \n1.0020)a\n1.0006 (0.9989, \n1.0024)a\n1.0005 (1.0000, 1.0011)a\n\u2003 New Zealand\n1.0005 (0.9994, \n1.0016)a\n1.0076 (1.0044, \n1.0108)a\n0.9957 (0.9928, \n0.9987)a\n1.0003 (0.9972, \n1.0033)a\n1.0040 (0.9867, \n1.0216)a\n0.9991 (0.9967, 1.0015)a\n\u2003 Thailand\n1.0025 (1.0022, \n1.0028)a\n1.0020 (1.0013, \n1.0027)a\n1.0019 (1.0014, \n1.0023)a\n1.0031 (1.0025, \n1.0037)a\n1.0049 (1.0040, \n1.0058)a\n1.0024 (1.0019, 1.0029)a\n\u2003 Taiwan\n1.0120 (1.0094, \n1.0146)a\n1.0766 (0.9985, \n1.1607)a\n1.0196 (1.0059, \n1.0335)a\n1.0200 (1.0084, \n1.0318)a\n1.0415 (0.8775, \n1.2363)a\n1.0165 (1.0135, 1.0195)a\n\u2003 Vietnam\n1.0024 (1.0009, \n1.0040)a\nNA\nNA\nNA\nNA\nNA\nA Wald-type test was used to test the significance of the effect modifications. NA is resulted from data paucity. aA two-sided P for difference across countries and territories <0.05.\nRR (95% CI)\na \nSex\nFemale\nMale\n0.998\n1.000\n1.002\n1.004\n1.006\n1.008\n1.010\n1.012\n1.014\nb\nAge (years)\n\u226419\n20\u201359\n\u226560\n0.998\n1.000\n1.002\n1.004\n1.006\n1.008\n1.010\n1.012\n1.014\nc\nNon-wildfire PM2.5\nLow\nHigh\nRR (95% CI)\nAll-cause\nAsthma\nAURI\nInfluenza\nPneumonia\nCOPD\n0.994\n0.996\n0.998\n1.000\n1.002\n1.004\n1.006\n1.008\n1.010\n1.012\n1.014\nd\n0.994\n0.996\n0.998\n1.000\n1.002\n1.004\n1.006\n1.008\n1.010\n1.012\n1.014\nAll-cause\nAsthma\nAURI\nInfluenza\nPneumonia\nCOPD\nIncome level\nLow\nincome\nMiddle\nincome\nHigh\nincome\n*\n*\n*\n*\n*\n*\n*\n*\n*\n*\nFig. 3 | Cumulative relative risk for respiratory hospitalizations associated \nwith a 1\u2009\u00b5g\u2009m\u22123 increase in wildfire-specific PM2.5 during lag 0\u20131\u2009days.  \na\u2013d, Relative risk modified by sex (a), age (b), community non-wildfire PM2.5 level \n(c) and community income level (d). The centre of the error bars represents \nthe mean effect estimates comparing days with lower versus higher wildfire-\nspecific PM2.5 levels across 1,052 study communities, with error bars indicating \nuncertainty ranges derived from the 95% CIs of 1,000 Monte Carlo samples.  \nAll the models were controlled for temperature, relative humidity, non-wildfire  \nPM2.5, day of week, seasonality and long-term trend. Income level was categorized \naccording to tertiles of the community-level GDP per capita: GDP per capita\u2009\u2264\u2009 \ntertile\u20091 as low income; tertile\u20091\u2009<\u2009GDP per capita\u2009\u2264\u2009tertile\u20092 as middle income; \nand GDP per capita\u2009>\u2009tertile\u20092 as high income. Non-wildfire PM2.5 level was \nclassified according to the 50th percentile of non-wildfire PM2.5 concentration \ndistributions (17.25\u2009\u00b5g\u2009m\u22123): average daily non-wildfire PM2.5 during the study \nperiod \u226417.25\u2009\u00b5g\u2009m\u22123 as low non-wildfire PM2.5 level; and >17.25\u2009\u00b5g\u2009m\u22123 as high non-\nwildfire PM2.5 level. A Wald-type test was used to test the significance of the effect \nmodifications. *A two-sided P for the difference is statistically significant (<0.05).\n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\nits chemical composition and long-range transport capability, might \nfurther facilitate this process, potentially increasing human suscepti-\nbility to influenza infection.\nAligned with a previous study in Southern California, 200323, \nwe found that individuals \u226419 or \u226560\u2009years old demonstrated a \ngreater vulnerability to respiratory hospitalization associated with \nwildfire-specific PM2.5. Such stronger associations among children \nand adolescents might be explained by environmental, behavioural \n(for example, prolonged outdoor activities) and physiological factors \n(for example, underdeveloped detoxification systems)27. However, \nprevious studies have yielded conflicting results. A meta-analysis \nsuggested that young people might be less vulnerable to adverse res-\npiratory effects from wildfire smoke exposure than are adults28. This \ndiscrepancy warrants further investigation.\nSocioeconomic status played a complex role in the association \nbetween wildfire-specific PM2.5 and respiratory diseases. Individuals \nfrom disadvantaged communities experienced substantially amplified \nrisks in respiratory hospitalization associated with wildfire-specific \nPM2.5, supported by a previous study in Northern California29. This \ngreater vulnerability may be multifactorial, possibly including more \nchildhood respiratory infections, lower housing conditions and indoor \nair quality, deficient nutrition and occupational exposures30. Further \nunravelling of how socioeconomic status modulates wildfire-specific \nPM2.5-related health effects demands use of multidimensional measures \nof socioeconomic status.\nNon-wildfire PM2.5 significantly modified the association between \nwildfire-specific PM2.5 and respiratory hospitalizations. Consistent \nfindings were observed across most respiratory diseases, indicating \nthat individuals residing in areas with higher non-wildfire PM2.5 levels \nwere more susceptible to wildfire smoke. This was probably due to \nimpaired lung function20 affected by chronic exposure to PM2.5 from \nother sources. This finding is in line with a recent study from North \nCarolina, United States, reporting that individuals residing in areas with \nhigher chronic PM2.5 exposure may exhibit heightened susceptibility \nto hospitalization during acute PM2.5 spikes31. Additionally, chronic \ninflammation resulting from regular exposures to higher levels of air \npollution31 may further exacerbate the population susceptibility, dur-\ning short-term increases in wildfire-specific PM2.5.\nWe observed significant spatiotemporal variations in hospital-\nization risks from respiratory diseases, related to wildfire-specific \nPM2.5. This may be attributed to differences in exposure levels of \nwildfire-specific and non-wildfire PM2.5, climate conditions, popula-\ntion susceptibility, basic health status and socioeconomic status (Sup-\nplementary Discussion 2). Specifically, as we discussed before, areas \nwith higher levels of air pollution may bear greater health risks from \nwildfire smoke due to impaired lung function and chronic inflamma-\ntion. This could contribute to the higher respiratory hospitalization \nrates in Taiwan, Brazil, Thailand and Vietnam, regions with historically \nhigher median concentrations of non-wildfire PM2.5. Additionally, the \npronounced annual concentrations of wildfire-specific PM2.5 may also \nplay a role, a finding recently reported31. It is impractical for residents \nto stay indoors or seek shelter for extended periods, during prolonged \nwildfire smoke pollution events, as would be feasible for shorter periods, \nsuch as in Brazil, Thailand and Vietnam. What is more, housing design \nalso varies across countries and territories. Hotter climates (tropical and \nsubtropical) tend to have more open housing, which offers less protec-\ntion from outdoor pollution. It is noteworthy that stronger associations \nof wildfire-specific PM2.5 with asthma hospitalization were found in \nAustralia and New Zealand. This may be explained by poorer resident \nadaptability, substantial asthma burdens32,33 and high concentrations of \nvarious allergens (for example, pollen, dust mites and fungal spores)34. \nFurthermore, wildfire-specific PM2.5 toxicity can vary by biomass type \nand fire intensity across communities35. This could also contribute to \nthe geographical disparities in the observed effect estimates.\nDuring the study period, the proportions of respiratory hospi-\ntalizations attributable to wildfire-specific PM2.5 increased overall, in \nAustralia, Taiwan and Vietnam. This rising burden could be attributable \nto increased population vulnerability and exposure to wildfire-specific \nPM2.5 (refs. 36,37). However, a decreasing trend was exhibited in coun-\ntries commonly considered wildfire-prone areas (for example, Brazil, \nCanada and Chile). This may be due to improved resident adaptation \nability, increased protective measures taken by residents and a long \nhistory of wildfire risk management38. Nonetheless, this does not dimin-\nish the severity or the need for addressing the wildfire-related health \nburden in these regions. While both Brazil and Chile have witnessed \ndeclines in the proportion of respiratory hospitalizations attributed \nto wildfire-specific PM2.5 exposure, wildfires remained a important \ncontributor to PM2.5-linked respiratory hospitalizations in Brazil and \nthe dominant source in Chile. This highlights a greater threat of PM2.5 \nfrom wildfire than non-wildfire sources.\nThe public health significance of PM2.5 from wildfire and \nnon-wildfire sources could be different. A previous study reported \nTable 2 | Annual attributable cases of cause-specific respiratory hospitalizations (N, 95% CI) associated with wildfire-specific \nPM2.5 exposure during the lag 0\u20131\u2009days by community income level and countries and territories\nSubgroups\nAll-cause respiratory\nAsthma\nCOPD\nAURI\nInfluenza\nPneumonia\nOverall\n25,321 (20,478, 30,114)\n1,949 (1,120, 2,739)\n2,913 (1,994, 3,811)\n1,724 (858, 2,566)\n1,083 (342, 1,682)\n10,866 (8,600, 13,105)\nIncome level\n\u2003 Low income\n16,676 (14,889, 18,423)\n1,171 (810, 1,528)\n2,133 (1,788, 2,468)\n1,162 (865, 1,453)\n786 (630, 937)\n6,736 (5,819, 7,643)\n\u2003 Middle income\n9,867 (8,320, 11,445)\n757 (538, 972)\n1,197 (864, 1,522)\n621 (429, 809)\n389 (170, 595)\n4,363 (3,485, 5,247)\n\u2003 High income\n872 (542, 1,200)\n227 (173, 280)\n188 (98, 277)\n73 (12, 133)\n188 (98, 277)\n140 (6, 276)\nCountry/territory\n\u2003 Australia\n485 (323, 646)\n186 (147, 224)\n142 (86, 197)\n4 (\u221236, 42)\n36 (8, 62)\n46 (\u221216, 106)\n\u2003 Brazil\n13,327 (11,924, 14,722)\n1,333 (946, 1,715)\n1,829 (1,513, 2,140)\n712 (504, 918)\n510 (297, 718)\n6,666 (5,777, 7,558)\n\u2003 Canada\n220 (\u221221,461)\n59 (16, 102)\n143 (30, 254)\n71 (21, 119)\n12 (\u221244, 53)\n56 (\u221235, 146)\n\u2003 Chile\n1,454 (323, 2,567)\n10 (\u2212226, 219)\n80 (\u221262, 219)\n7 (\u2212279, 280)\n22 (\u221239, 80)\n545 (\u221224, 1,085)\n\u2003 New Zealand\n61 (\u221275, 198)\n107 (63, 149)\n\u221285 (\u2212146, \u221226)\n4 (\u221242, 50)\n11 (\u221236, 57)\n\u221222 (\u221280, 35)\n\u2003 Thailand\n6,575 (5,733, 7,408)\n350 (227, 471)\n925 (712, 1,134)\n959 (781, 1,134)\n596 (495, 694)\n2,305 (1,876, 2,729)\n\u2003 Taiwan\n4,634 (3,675, 5,575)\n10 (0, 17)\n161 (49, 267)\n239 (104, 370)\n54 (\u2212236, 222)\n1,990 (1,646, 2,340)\n\u2003 Vietnam\n817 (322, 1,308)\nNA\nNA\nNA\nNA\nNA\nAll the models were controlled for temperature, relative humidity, non-wildfire PM2.5, day of week, seasonality and long-term trend. NA is resulted from data paucity.\n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\nthat wildfire-specific PM2.5 may pose greater health risks than PM2.5 \nfrom other sources39. This may be due to the potential heightened \ntoxicity of wildfire-specific PM2.5, accounting for an increased pres-\nence of smaller particles (for example, submicrometre particles \nand ultrafine particles) and a higher concentration of oxidative and \npro-inflammatory components (for example, polycyclic aromatic \nhydrocarbons and aldehydes)1. This coincided with our findings that \nwildfire-specific PM2.5 posed a greater hospitalization risk for various \nrespiratory diseases than did non-wildfire PM2.5. Therefore, continued \nmitigation efforts are warranted to attenuate and reverse the rising \ndifference in the proportion of respiratory hospitalizations due to \nwildfire and non-wildfire particles, especially in Chile, New Zealand, \nThailand, Taiwan and Vietnam.\nAs by far the largest study on the association between wildfire- \nspecific PM2.5 and respiratory hospitalizations, this study has several \nstrengths. First, we used an extensive multicountry dataset with high \nstatistical power and a unified well-established two-stage analytical \nframework, including both the exposure assessment method and risk \nassessment model. This ensured the robustness, generalizability and \ncomparison of the results across countries and territories. No previ-\nous study has done this comprehensive consistent assessment, for \ncause-specific respiratory hospitalization with wildfire-specific PM2.5. \nSecond, the communities included in this study were predominantly \nregions with a history of wildfires and thereby serve as areas suitable \nfor robust investigations on wildfire PM2.5 and respiratory health. Third, \nwe identified vulnerable populations from different causes, which can \ninform the development of targeted interventions and aid in addressing \nenvironmental injustice. Fourth, the spatiotemporal assessment of the \nrespiratory health burden attributable to wildfire-specific PM2.5 allows \nfor a better understanding of the extent to which wildfire air pollution \nhas affected different populations, in multiple countries and territories \nover varying periods. This may aid in resource allocation and more \ncost-effective policy-making. Finally, we compared the effect esti-\nmates, AF and spatiotemporal burden of respiratory hospitalizations \nattributable to wildfire and non-wildfire PM2.5 as few previous studies \nhave reported such findings40. This provided supportive quantitative \nevidence on the stronger detrimental health effects of PM2.5 from \nwildfire relative to other sources.\nSome limitations should be acknowledged. Despite extensive spa-\ntiotemporal coverage, our dataset was still incomplete. For example, \nsome country- and territory-specific effect estimates may not be fully \nnationally representative as a result of the inclusion of a part of the com-\nmunities. Specifically, Taiwan only had data from six municipalities; \nAustralia only had data from New South Wales; and data from some \ncommunities did not cover the full study period. Moreover, although \nPM2.5 is the key component of wildfire smoke mixtures, other air pol-\nlutants such as ozone were not considered in this analysis. However, \nour results changed minimally when we controlled for wildfire-specific \nozone in the model across all investigated causes. This suggests that \nthe presence of wildfire-specific ozone has a negligible influence on \nthe estimated effects of wildfire-specific PM2.5 on respiratory hos-\npitalizations. Additionally, community-level wildfire-specific PM2.5 \nexposures were used in the analysis, which may not fully reflect per-\nsonal exposures because of variations in housing quality and personal \nprotective behaviours. The grid resolution used may not fully capture \nspatial variability, potentially leading to exposure measurement errors. \nTo address this, we used population-weighted exposure. Overall, our \nresults should be conservative given that a coarser exposure assign-\nment may bias the effect estimates towards the null41. Meanwhile, the \npotential measurement error introduced by the uncertainties of the fire \nemission inventory, the GEOS-Chem simulations and machine learning \nmodels may also underestimate the effect estimates. Future studies \nwith expanded health data, coupled with broader spatiotemporal \ncoverage and more effect modifiers and air pollutants, could further \nadvance the assessment of the health impacts of wildfire-specific  \nair pollution.\nPublic health implications\nWildfire impacts are still likely to intensify as the global climate \ncontinues to warm1. In the context of the notable contribution of \nwildfire-specific PM2.5 exposure to respiratory burdens, prioritizing \nrobust mitigation and adaptation strategies across diverse countries \nand territories emerges as a critical public health imperative, especially \nfor diseases, populations and areas experiencing a greater suscepti-\nbility. Some solutions consist of raising awareness of wildfire-related \nhealth risks targeted at policy-makers, clinicians and the public, improv-\ning emission control, closely monitoring wildfire air pollution levels \nand strengthening community preparedness through resource plan-\nning and allocation.\nMethods\nHospitalization and socioeconomic data\nOn the basis of the international classification of diseases nineth (ICD-9)  \nrevision or ICD-10 codes, we obtained daily hospitalization data for \nall-cause respiratory diseases, asthma, COPD, AURI, influenza and \npneumonia across 1,052 communities in eight countries and territo-\nries, with a median area of 5,185.3\u2009km2 (IQR 11,664.2\u2009km2) and a median \npopulation of 139,256 (IQR 271,027). The hospitalization data for all \nAll-cause\nAsthma\nAURI\nCOPD\nInfluenza\nPneumonia\nb\nAttributable fraction (%, 95% CI)\n0\n2.50\n5.00\n7.50\na\nRR (95% CI)\n*\n*\n*\n*\n*\n*\n1.000\n1.002\n1.005\n1.008\n1.010\nWildfire-specific PM2.5\nNon-wildfire PM2.5\nWildfire-specific PM2.5\nNon-wildfire PM2.5\nFig. 4 | Comparisons in the health impacts of wildfire-specific and non-wildfire \nPM2.5 on respiratory hospitalization. a, The relative risks of hospitalizations for \ncause-specific respiratory disease per 1\u2009\u00b5g\u2009m\u22123 increase in wildfire-specific and \nnon-wildfire PM2.5. b, The AFs of hospitalizations for cause-specific respiratory \ndiseases associated with wildfire-specific and non-wildfire PM2.5. The centre of \nthe error bars represents the mean effect estimates comparing days with lower \nversus higher wildfire-specific PM2.5 levels across 1,052 study communities, \nwith error bars indicating uncertainty ranges derived from the 95% CIs of 1,000 \nMonte Carlo samples. Models for wildfire-specific PM2.5 were controlled for \ntemperature, relative humidity, non-wildfire PM2.5, day of week, seasonality and \nlong-term trend. Models for non-wildfire PM2.5 were controlled for temperature, \nrelative humidity, wildfire-specific PM2.5, day of week, seasonality and long-term \ntrend. A Wald-type test was used to test the significance of the difference between \neffect estimates of wildfire-specific and non-wildfire PM2.5. *A two-sided P for \ndifference <0.05.\n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\ncountries and territories fell within the period 2000 to 2019, with dif-\nfering start and/or end dates: for Australia, Brazil and New Zealand, \nhospitalization data were from 2000 to 2019; for Canada, data were \nfrom 2005 to 2019; for Chile, data were from 2001 to 2019; for Thai-\nland, data were from 2015 to 2019; for Taiwan, data were from 2000 \nto 2018; and for Vietnam, data were from 2002 to 2015. Additional \ndetails on data collection, ICD codes and data coverage are provided \nin Supplementary Methods 1.\nWe estimated the community-level gross domestic product (GDP) \nper capita, by dividing the total GDP by the population count in each com-\nmunity, using data from the global gridded GDP (10\u2009\u00d7\u200910\u2009km2) (ref. 42) \nand the gridded population of the world (GPW, v.4, 1\u2009\u00d7\u20091\u2009km2) (ref. 43). \nCommunity-specific population and total GDP were calculated by \naggregating the corresponding values of grid cells within each com-\nmunity. We then categorized community-specific socioeconomic \nstatus, based on tertiles of GDP per capita, across study communities: \nGDP per capita\u2009\u2264\u2009tertile\u20091 as low income; tertile\u20091\u2009<\u2009GDP per capita\u2009\u2264\u2009ter-\ntile\u20092 as middle income; and GDP per capita\u2009>\u2009tertile\u20092 as high income. \nWe also assigned the community non-wildfire PM2.5 level according to \nthe 50th percentile of non-wildfire PM2.5 concentration distribution \nacross study communities (\u226417.25\u2009\u00b5g\u2009m\u22123 as low level and >17.25\u2009\u00b5g\u2009m\u22123 \nas high level)44.\nFinally, we analysed a multicountry daily hospitalization dataset \nof cause-specific respiratory diseases, across 1,052 communities, in \neight countries and territories, during different periods of 2000\u2013\n2019, characterized by age, sex, community income level, community \nnon-wildfire PM2.5 level and country and territory. More information \non data collection and processing is in Supplementary Methods 1.\nWildfire-specific PM2.5 and meteorological data\nAs detailed in our previous work37, daily concentrations of \nwildfire-specific PM2.5 at a 0.25\u00b0\u2009\u00d7\u20090.25\u00b0 (~28\u2009\u00d7\u200928\u2009km2) spatial reso-\nlution during 2000\u20132019, were estimated using a chemical trans-\nport model (GEOS-Chem) and calibrated through a machine learning \napproach. Briefly, wildfire-specific PM2.5 was estimated as differences \nin GEOS-Chem-derived simulations for total PM2.5, under two sce-\nnarios with and without wildfire emissions. The GEOS-Chem-derived \nestimates of total PM2.5 from the original 2.0\u00b0\u2009\u00d7\u20092.5\u00b0 spatial resolution \n(~220\u2009\u00d7\u2009280\u2009km2) were interpolated to 0.25\u00b0\u2009\u00d7\u20090.25\u00b0 grid-cell resolution \nto match other auxiliary variables used in machine learning calibra-\ntion, using inverse distance weighted spatial interpolation10,37,45. Then \nwe trained a random forest machine learning model to calibrate the \nGEOS-Chem-derived estimates of total PM2.5 against ground-level \ndaily total PM2.5 measurements, with daily gridded (0.25\u00b0\u2009\u00d7\u20090.25\u00b0 \nresolution) meteorological parameters (mean/minimum/maximum \ntemperature, temperature variations, relative humidity, wind speed, \nprecipitation, air pressure and ultraviolet radiation) and calendar \nyear, calendar month, day of week, day of year, longitude and latitude \nas auxiliary variables37. According to tenfold cross-validation, the ran-\ndom forest-calibrated total PM2.5 concentrations accounted for 91.0% \n(root mean squared error\u2009=\u20098.5\u2009\u00b5g\u2009m\u22123) of the variability observed in \nground-level measurements37. Finally, the calibrated wildfire-specific \nPM2.5 concentrations were estimated by multiplying the calibrated total \nPM2.5 concentrations with the GEOS-Chem-derived wildfire-to-all ratio \nin each 0.25\u00b0\u2009\u00d7\u20090.25\u00b0 grid cell.\nWe obtained the daily ambient temperature and dewpoint tem-\nperature from the European Centre for medium-range weather fore-\ncasts reanalysis v.5 (ERA5)46, at the 0.25\u00b0\u2009\u00d7\u20090.25\u00b0 grid-cell resolution. \nThese data were used to calculate the daily relative humidity for each \ncommunity. Details on the collection and calculation of meteorologi-\ncal data are shown in Supplementary Methods 2. Subsequently, three \ncommunity-specific, daily data values were assigned to each study \ncommunity: the population-weighted wildfire-specific PM2.5 concen-\ntrations, temperature and relative humidity. For communities smaller \nthan the 0.25\u00b0\u2009\u00d7\u20090.25\u00b0 grid, the value assigned to such a community is \nthe value of the grid cell it falls entirely within. The value assigned to a \ncommunity exceeding the grid size is the average value across all grid \ncells covered at least partly by that community.\nStatistical analysis\nWe implemented a standard two-stage analytic framework, to estimate \nthe association between short-term exposure to daily wildfire-specific \nPM2.5 and cause-specific respiratory hospitalization. The two-stage \ndesign is a flexible and computationally efficient analytical frame-\nwork commonly used in environmental epidemiology to model large, \nheterogeneous data from multiple communities19,47\u201349. This approach \nallows for separate modelling of community-specific characteristics, \npreserving crucial local nuances and mitigating potential biases that \ncould arise from pooling diverse data into a single model (Supplemen-\ntary Methods 3).\nIn the first stage, we used a quasi-Poisson regression with a dis-\ntributed lag nonlinear model in each community, with the following \nequation:\nYit \u223cPoisson (\u03bcit) log (\u03bcit) = cb (Fire_PM2.5it, lag = 7)\n+ ns (Tempit, d.f. = 6) + ns (RHit, d.f. = 3) + nonfire_PM2.5it\n+ ns (Timeit, d.f. = 7/year) + \u03b2DOWit + \u03b1\nWhere Yit denotes the hospitalization counts in community i on day t; \ncb is a distributed lag model incorporated with a crossbasis47 function \nof a linear term over the current and previous seven (lag 0\u20137) days and \ntwo internal knots placed at equally spaced values in the log scale of \nlag days; ns is the natural cubic spline function; and d.f. is the degree of \nfreedom. To account for the confounding effects of temperature and \nrelative humidity (RH), we included moving averaged daily ambient \ntemperature (Temp) and RH during lag 0\u20137\u2009days in the model. Daily \nnon-wildfire PM2.5 concentrations were adjusted to control for the \npotential confounding effects of PM2.5 from other sources; time vari-\nable was fitted in an ns term of seven degrees of freedom per year to \naccount for long-term trends and seasonality; DOW is the categorical \nvariable for day of the week to control for weekly pattern and \u03b2 is its \ncoefficient; and \u03b1 is the intercept.\nIn the second stage, we pooled the community-specific cumula-\ntive effect estimates over the lag period within each country and ter-\nritory, using a random-effect meta-analytic model. This allowed us to \ngenerate the country- and territory-specific estimates, as per a 1\u2009\u00b5g\u2009m\u22123 \nincrement in wildfire-specific PM2.5. Our preliminary analyses revealed \nmoderate heterogeneity in effect estimates across communities for \ncause-specific respiratory hospitalization (I2 34\u201358%) (Supplementary \nTable 4). Consequently, we pooled the community-specific estimates \nto derive overall estimates.\nTo characterize the C\u2013R relationship between wildfire-specific \nPM2.5 and respiratory hospitalization, we fitted a B-spline function with \ntwo knots at the 25th and 75th percentiles of mean wildfire-specific PM2.5 \nconcentration distributions50\u201352. For each country and territory, indi-\nvidual community percentiles were first averaged to obtain country- \nand territory-level values, which were then used to compute an overall \naverage across all countries and territories50\u201352. To further test the \nlinearity of the C\u2013R relationships, we compared community-specific \nqBIC values of linear models with nonlinear models (Supplementary \nMethods 4). We conducted various sensitivity analyses to assess the \nrobustness of our results (Supplementary Methods 5).\nTo explore potential effect modifiers, we performed subgroup \nanalyses by age, sex, community income level and community \nnon-wildfire PM2.5 level. We tested the significance of the effect modi-\nfications using the Wald-type test53.\nUnder a causal assumption, we estimated attributable cases and AF \nof respiratory hospitalizations attributable to short-term exposure to \nwildfire-specific PM2.5, using the pooled country- and territory-specific \n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\neffect estimates, as detailed elsewhere54. Briefly, we calculated the \nrespiratory hospitalizations attributable to wildfire-specific PM2.5 in \neach community, using pooled country- and territory-specific effect \nestimates. Subsequently, we divided the total hospitalizations attrib-\nutable to wildfire-specific PM2.5 by the total hospitalizations across all \nthe study communities or communities within a specific country and \nterritory, to obtain the overall and country- and territory-specific AF. \nWe quantified the uncertainty in attributable cases and AF using 95% \nempirical CI, calculated with 1,000 Monte Carlo simulations54. We \nfurther explored the temporal trends of attributable burden, which \nwere first calculated on the basis of year-specific effect estimates and \nthen aggregated by four different periods: 2000\u20132004, 2005\u20132009, \n2010\u20132014 and 2015\u20132019.\nTo compare the health impacts of wildfire-specific PM2.5 and \nnon-wildfire PM2.5, we estimated the hospitalization risk for all respira-\ntory diseases included, as per 1\u2009\u00b5g\u2009m\u22123 increase in non-wildfire PM2.5. We \nquantified the proportion of respiratory hospitalizations attributable \nto non-wildfire PM2.5 during the study period. We measured the differ-\nence in the proportion of respiratory hospitalizations attributable to \nwildfire-specific and non-wildfire PM2.5, as AF of wildfire-specific PM2.5 \nminus AF of non-wildfire PM2.5 (termed AF difference), across different \nperiods of countries and territories.\nAll analyses were implemented by R software (v.4.0.5 for Windows) \nusing mixmeta53 and dlnm55 R packages. A two-sided P\u2009<\u20090.05 was deter-\nmined as statistically significant.\nInclusion and ethics\nThis study has been approved by the Monash University Human \nResearch Ethics Committee.\nReporting summary\nFurther information on research design is available in the Nature \nPortfolio Reporting Summary linked to this article.\nData availability\nPublicly available data are found here: the global gridded GDP (https://\nwww.nature.com/articles/sdata20184); the GPW (https://cmr.earth-\ndata.nasa.gov/search/concepts/C1597158029-SEDAC.html); surface \ntemperature and ambient dewpoint temperature from ERA5 (https://\nwww.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era5); \nmodern-era retrospective analysis for research and applications v.2 \n(MERRA-2) data, biomass burning emissions inventory of global fire \nemissions database v.4.1 (GFED v.4.1) data and anthropogenic emis-\nsions inventory of EDGAR v.4.2 data that support the GEOS-Chem \nmodel development and wildfire-specific PM2.5 simulation in this study \nare available from https://gmao.gsfc.nasa.gov/reanalysis/MERRA-2/, \nhttps://daac.ornl.gov/VEGETATION/guides/fire_emissions_v4_R1.html \nand http://edgar.jrc.ec.europa.eu/, respectively. For respiratory hos-\npitalization data, the authors are not permitted to share the data. \nInterested researchers should contact the corresponding author S.L. \n(shanshan.li@monash.edu) on a case-by-case basis. Source data are \nprovided with this paper.\nCode availability\nAnalysis codes are available from the corresponding authors on request \nand will be shared on https://github.com/Yiwen-Zhang-259/Multicoun-\ntry_Wildfire_Respiratory_Hospitalization.git.\nReferences\n1.\t\nXu, R. et al. Wildfires, global climate change, and human health. \nNew Engl. J. Med. 383, 2173\u20132181 (2020).\n2.\t\nFire Statistics (Canadian Interagency Forest Fire Centre, 2023); \nhttps://ciffc.net/statistics\n3.\t\nZhou, S., Xu, R., Chen, G., Yu, P. & Guo, Y. Where is the boundary of \nwildfire smoke? Innov. Med. 1, 100024 (2023).\n4.\t\nMorawska, L. et al. The state of science on severe air pollution \nepisodes: quantitative and qualitative analysis. Environ. Int. 156, \n106732 (2021).\n5.\t\nSenande-Rivera, M., Insua-Costa, D. & Miguez-Macho, G.  \nSpatial and temporal expansion of global wildland fire  \nactivity in response to climate change. Nat. Commun. 13,  \n1208 (2022).\n6.\t\nReid, C. E. et al. Critical review of health impacts of wildfire smoke \nexposure. Environ. Health Perspect. 124, 1334\u20131343 (2016).\n7.\t\nVicente, A. et al. Emission factors and detailed chemical \ncomposition of smoke particles from the 2010 wildfire season. \nAtmos. Environ. 71, 295\u2013303 (2013).\n8.\t\nMakkonen, U., Hell\u00e9n, H., Anttila, P. & Ferm, M. Size distribution \nand chemical composition of airborne particles in south-eastern \nFinland during different seasons and wildfire episodes in 2006. \nSci. Total Environ. 408, 644\u2013651 (2010).\n9.\t\nZhang, Y. et al. Health impacts of wildfire smoke on children \nand adolescents: a systematic review and meta-analysis. Curr. \nEnviron. Health Rep. 11, 46\u201360 (2024).\n10.\t Xue, T. et al. Associations between exposure to landscape fire \nsmoke and child mortality in low-income and middle-income \ncountries: a matched case-control study. Lancet Planet. Health 5, \ne588\u2013e598 (2021).\n11.\t\nReid, C. E. et al. Differential respiratory health effects from the \n2008 northern California wildfires: a spatiotemporal approach. \nEnviron. Res. 150, 227\u2013235 (2016).\n12.\t Matz, C. J. et al. Health impact analysis of PM2.5 from wildfire \nsmoke in Canada (2013\u20132015, 2017\u20132018). Sci. Total Environ. 725, \n138506 (2020).\n13.\t Liu, J. C. et al. Wildfire-specific fine particulate matter and risk of \nhospital admissions in urban and rural counties. Epidemiology 28, \n77 (2017).\n14.\t Ye, T. et al. Risk and burden of hospital admissions associated with \nwildfire-related PM2.5 in Brazil, 2000\u201315: a nationwide time-series \nstudy. Lancet Planet. Health 5, e599\u2013e607 (2021).\n15.\t Morgan, G. et al. Effects of bushfire smoke on daily mortality and \nhospital admissions in Sydney, Australia. Epidemiology 21, 47\u201355 \n(2010).\n16.\t Johnston, F. H. et al. Air pollution events from forest fires and \nemergency department attendances in Sydney, Australia 1996\u2013\n2007: a case-crossover analysis. Environ. Health 13, 105 (2014).\n17.\t Martin, K. L., Hanigan, I. C., Morgan, G. G., Henderson, S. B. & \nJohnston, F. H. Air pollution from bushfires and their association \nwith hospital admissions in Sydney, Newcastle and Wollongong, \nAustralia 1994\u20132007. Aust. NZ J. Public Health 37, 238\u2013243 (2013).\n18.\t Li, J. et al. Landscape fire smoke enhances the association \nbetween fine particulate matter exposure and acute respiratory \ninfection among children under 5 years of age: findings of a \ncase-crossover study for 48 low-and middle-income countries. \nEnviron. Int. 171, 107665 (2023).\n19.\t Chen, G. et al. Mortality risk attributable to wildfire-related PM2.5 \npollution: a global time series study in 749 locations. Lancet \nPlanet. Health 5, e579\u2013e587 (2021).\n20.\t Kelly, F. J. & Fussell, J. C. Air pollution and public health: emerging \nhazards and improved understanding of risk. Environ. Geochem. \nHealth 37, 631\u2013649 (2015).\n21.\t Wen, J., Heft-Neal, S., Baylis, P., Boomhower, J. & Burke, M. \nQuantifying fire-specific smoke exposure and health impacts. \nProc. Natl Acad. Sci. USA 120, e2309325120 (2023).\n22.\t Crabbe, H. Risk of respiratory and cardiovascular hospitalisation \nwith exposure to bushfire particulates: new evidence from \nDarwin, Australia. Environ. Geochem. Health 34, 697\u2013709 (2012).\n23.\t Delfino, R. J. et al. The relationship of respiratory and \ncardiovascular hospital admissions to the southern California \nwildfires of 2003. Occup. Environ. Med. 66, 189\u2013197 (2009).\n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\n24.\t Xing, Y.-F., Xu, Y.-H., Shi, M.-H. & Lian, Y.-X. The impact of PM2.5 on \nthe human respiratory system. J. Thorac. Dis. 8, E69 (2016).\n25.\t Hsiao, T.-C. et al. Interactions of chemical components in ambient \nPM2.5 with influenza viruses. J. Hazard. Mater. 423, 127243 (2022).\n26.\t Chiv\u00e9, C. et al. Exposure to PM2.5 modulate the pro-inflammatory \nand interferon responses against influenza virus infection in a \nhuman 3D bronchial epithelium model. Environ. Pollut. 348, \n123781 (2024).\n27.\t Goldizen, F. C., Sly, P. D. & Knibbs, L. D. Respiratory effects of air \npollution on children. Pediatr. Pulmonol. 51, 94\u2013108 (2016).\n28.\t Kondo, M. C. et al. Meta-analysis of heterogeneity in the effects of \nwildfire smoke exposure on respiratory health in North America. \nInt. J. Environ. Res. Public Health 16, 960 (2019).\n29.\t Reid, C. et al. Effect modification of the association between fine \nparticulate air pollution during a wildfire event and respiratory  \nhealth by area-level measures of socio-economic status, race/\nethnicity, and smoking prevalence. Environ. Res. Health 1, \n025005 (2023).\n30.\t Doiron, D. et al. Air pollution, lung function and COPD: results \nfrom the population-based UK Biobank study. Eur. Resp. J. 54, \n1802140 (2019).\n31.\t Cowan, K. N. et al. Effect measure modification of the association \nbetween short-term exposures to PM2.5 and hospitalizations \nby long-term PM2.5 exposure among a cohort of people with \nchronic obstructive pulmonary disease (COPD) in North Carolina, \n2002\u20132015. Environ. Health 22, 49 (2023).\n32.\t Chronic Respiratory Conditions (Australian Institute of \nHealth and Welfare, 2023); www.aihw.gov.au/reports/\nchronic-respiratory-conditions/chronic-respiratory-conditions/\ncontents/summary\n33.\t Asthma (Asthma and Respiratory Foundation of New \nZealand, 2023); www.asthmafoundation.org.nz/your-health/\nliving-with-asthma\n34.\t Gautier, C. & Charpin, D. Environmental triggers and avoidance in \nthe management of asthma. J. Asthma Allergy 10, 47\u201356 (2017).\n35.\t Kim, Y. H. et al. Mutagenicity and lung toxicity of smoldering vs. \nflaming emissions from various biomass fuels: implications for \nhealth effects from wildland fires. Environ. Health Perspect. 126, \n017011 (2018).\n36.\t Xu, R. et al. Climate change, environmental extremes, and human \nhealth in Australia: challenges, adaptation strategies, and policy \ngaps. Lancet Reg. Health Western Pacific 40, 100936 (2023).\n37.\t Xu, R. et al. Global population exposure to landscape fire air \npollution from 2000 to 2019. Nature 621, 521\u2013529 (2023).\n38.\t Tymstra, C., Stocks, B. J., Cai, X. & Flannigan, M. D. Wildfire \nmanagement in Canada: review, challenges and opportunities. \nProg. Disaster Sci. 5, 100045 (2020).\n39.\t Aguilera, R., Corringham, T., Gershunov, A. & Benmarhnia, T. \nWildfire smoke impacts respiratory health more than fine particles \nfrom other sources: observational evidence from southern \nCalifornia. Nat. Commun. 12, 1493 (2021).\n40.\t Jiao, A., Headon, K., Han, T., Umer, W. & Wu, J. Associations \nbetween short-term exposure to wildfire particulate matter and \nrespiratory outcomes: a systematic review. Sci. Total Environ. 907, \n168134 (2024).\n41.\t Smith, G. S. et al. Air pollution and pulmonary tuberculosis: a nested \ncase\u2013control study among members of a northern California health \nplan. Environ. Health Perspect. 124, 761\u2013768 (2016).\n42.\t Kummu, M., Taka, M. & Guillaume, J. H. A. Gridded global datasets \nfor Gross Domestic Product and Human Development Index over \n1990\u20132015. Sci. Data 5, 180004 (2018).\n43.\t Center For International Earth Science Information Network- \nCIESIN-Columbia University. Gridded Population of the World, \nVersion 4 (GPWv4): Population Density, Revision 11 (Version 4.11) \n[Data set]. Earth Data https://doi.org/10.7927/H49C6VHW (2017).\n44.\t Wang, Q. et al. Independent and combined effects of \nheatwaves and PM 2.5 on preterm birth in Guangzhou, China: \na survival analysis. Environ. Health Perspect. 128, 017006 \n(2020).\n45.\t Ye, T. et al. Short-term exposure to wildfire-related PM2.5 increases \nmortality risks and burdens in Brazil. Nat. Commun. 13, 7651 \n(2022).\n46.\t Hersbach, H. et al. The ERA5 global reanalysis. Q. J. R. Meteorol. \nSoc. 146, 1999\u20132049 (2020).\n47.\t Gasparrini, A. et al. Mortality risk attributable to high and low \nambient temperature: a multicountry observational study. Lancet \n386, 369\u2013375 (2015).\n48.\t Liu, C. et al. Ambient particulate air pollution and daily mortality \nin 652 cities. New Engl. J. Med. 381, 705\u2013715 (2019).\n49.\t Vicedo-Cabrera, A. M. et al. Short term association between \nozone and mortality: global two stage time series study in 406 \nlocations in 20 countries. Br. Med. J. 368, m108 (2020).\n50.\t Huang, W.-Z. et al. Improved morbidity-based air quality health \nindex development using Bayesian multi-pollutant weighted \nmodel. Environ. Res. 204, 112397 (2022).\n51.\t Chen, R. et al. Fine particulate air pollution and daily mortality. A \nnationwide analysis in 272 Chinese cities. Am. J. Resp. Crit. Care \nMed. 196, 73\u201381 (2017).\n52.\t Zhang, Y. et al. Short-term exposure to wildfire-specific PM2.5 \nand diabetes hospitalization: a study in multiple countries and \nterritories. Diabetes Care 47, 1664\u20131672 (2024).\n53.\t Sera, F., Armstrong, B., Blangiardo, M. & Gasparrini, A. An \nextended mixed-effects framework for meta-analysis. Stat. Med. \n38, 5429\u20135444 (2019).\n54.\t Gasparrini, A. & Leone, M. Attributable risk from distributed lag \nmodels. BMC Med. Res. Method. 14, 55 (2014).\n55.\t Gasparrini, A. Distributed lag linear and non-linear models in R: \nthe package dlnm. J. Stat. Softw. 43, 1\u201320 (2011).\nAcknowledgements\nThis study is supported by Australian Research Council \n(DP210102076) and Australian National Health & Medical Research \nCouncil (APP2000581). Yiwen Zhang is supported by the NHMRC \ne-Asia Joint Research Program Grant (GNT2000581). R.X. is \nsupported by Monash Faculty of Medicine Nursing and Health \nScience (FMNHS) Bridging Postdoctoral Fellowships 2022 and \nVicHealth Postdoctoral Research Fellowships 2022. W. Huang and \nT.Y. are supported by China Scholarship Council funds (W. Huang, \n202006380055; T.Y., 201906320051). S.L. was supported by an \nEmerging Leader Fellowship (GNT2009866) of the Australian \nNational Health and Medical Research Council. Y.G. was supported \nby Career Development Fellowship (GNT1163693) and Leader \nFellowship (GNT2008813) of the Australian National Health and \nMedical Research Council.\nAuthor contributions\nS.L. and Y.G. designed the study and are co-senior authors.  \nYiwen Zhang conducted the statistical analysis and took lead in \ndrafting the manuscript and interpreting the results. R.X. estimated \nthe wildfire-specific PM2.5 exposure data. R.X. and W. Huang provided \nsubstantial scientific input in interpreting the results and editing of \nthe manuscript. R.X., W. Huang, T.Y., P.Y., W.Y., Y.W., Y.L., Z.Y., B.W., K.J., \nJ.S., M.J.A., A.J., Y.G. and S.L. revised and edited the manuscript. Yiwen \nZhang, R.X., W. Huang, T.Y., P.Y., W.Y., Y.W., Y.L., Z.Y. and B.W. cleaned \nthe data. A.C., B.J., D.G., E.L., F.H.J., G.M., L.D.K., Ying Zhang, G.G.M., \nJ.H., J.A., Y.L.G., L.M., M.S.Z.S.C., P.H.N.S., P.M., P.B., S.H., W. Hu and \nD.P. provided the data and contributed to the submitted version of the \nmanuscript. The corresponding author attests that all listed authors \nmeet authorship criteria and that no others meeting the criteria have \nbeen omitted.\n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\nFunding\nOpen access funding provided by Monash University.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nExtended data is available for this paper at  \nhttps://doi.org/10.1038/s41893-025-01533-9.\nSupplementary information The online version  \ncontains supplementary material available at  \nhttps://doi.org/10.1038/s41893-025-01533-9.\nCorrespondence and requests for materials should be addressed to \nShanshan Li.\nPeer review information Nature Sustainability thanks Santu Ghosh, \nColleen Reid and Jennifer Stowell for their contribution to the peer \nreview of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher\u2019s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nOpen Access This article is licensed under a Creative Commons \nAttribution 4.0 International License, which permits use, sharing, \nadaptation, distribution and reproduction in any medium or format, \nas long as you give appropriate credit to the original author(s) and the \nsource, provide a link to the Creative Commons licence, and indicate \nif changes were made. The images or other third party material in this \narticle are included in the article\u2019s Creative Commons licence, unless \nindicated otherwise in a credit line to the material. If material is not \nincluded in the article\u2019s Creative Commons licence and your intended use \nis not permitted by statutory regulation or exceeds the permitted use, you \nwill need to obtain permission directly from the copyright holder. To view \na copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.\n\u00a9 The Author(s) 2025, corrected publication 2025\n1Climate, Air Quality Research Unit, School of Public Health and Preventive Medicine, Monash University, Melbourne, Victoria, Australia.  \n2Monash Biomedicine Discovery Institute, Department of Biochemistry and Molecular Biology, Monash University, Melbourne, Victoria, Australia.  \n3School of Public Health and Preventive Medicine, Monash University, Melbourne, Victoria, Australia. 4School of Population Health, the University of  \nNew South Wales, Sydney, New South Wales, Australia. 5School of Biological, Earth & Environmental Sciences, the University of New South Wales, \nSydney, New South Wales, Australia. 6School of Epidemiology and Public Health, University of Ottawa, Ottawa, Ontario, Canada. 7Environmental Health \nScience and Research Bureau, Health Canada, Ottawa, Ontario, Canada. 8Menzies Institute for Medical Research, University of Tasmania, Hobart, \nTasmania, Australia. 9Sydney School of Public Health, The University of Sydney, Sydney, New South Wales, Australia. 10Public Health Research Analytics \nand Methods for Evidence, Public Health Unit, Sydney Local Health District Camperdown, Sydney, New South Wales, Australia. 11School of Clinical \nMedicine, University of New South Wales, Sydney, New South Wales, Australia. 12Faculty of Health and Medical Sciences, University of Western Australia, \nCrawley, Western Australia, Australia. 13School of Earth, Atmosphere and Environment, Monash University, Melbourne, Victoria, Australia. 14Environmental \nand Occupational Medicine, National Taiwan University and National Taiwan University Hospital, Taipei, Taiwan. 15School of Earth and Atmospheric \nSciences, Queensland University of Technology, Brisbane, Queensland, Australia. 16Department of Pathology, School of Medicine, University of S\u00e3o \nPaulo, S\u00e3o Paulo, Brazil. 17School of Medicine, University of the Andes (Chile), Las Condes, Regi\u00f3n Metropolitana, Las Condes, Chile. 18School of Public \nHealth, The University of Adelaide, Adelaide, South Australia, Australia. 19Department of Public Health, University of Otago, Wellington, New Zealand. \n20School of Public Health & Social Work, Queensland University of Technology, Brisbane, Queensland, Australia. 21School of Public Health, University of \nQueensland, Brisbane, Queensland, Australia. 22These authors contributed equally: Yuming Guo, Shanshan Li. \n\u2009e-mail: shanshan.li@monash.edu\nYiwen Zhang\u2009\n\u200a\u20091, Rongbin Xu\u2009\n\u200a\u20091, Wenzhong Huang\u2009\n\u200a\u20091, Tingting Ye\u2009\n\u200a\u20091, Pei Yu\u2009\n\u200a\u20091, Wenhua Yu\u2009\n\u200a\u20091, Yao Wu\u2009\n\u200a\u20091, \nYanming Liu\u2009\n\u200a\u20091, Zhengyu Yang\u2009\n\u200a\u20091, Bo Wen\u2009\n\u200a\u20091, Ke Ju1, Jiangning Song\u2009\n\u200a\u20092, Michael J. Abramson\u2009\n\u200a\u20093, Amanda Johnson1, \nAnthony Capon\u2009\n\u200a\u20093, Bin Jalaludin\u2009\n\u200a\u20094, Donna Green5, Eric Lavigne6,7, Fay H. Johnston\u2009\n\u200a\u20098, Geoffrey G. Morgan\u2009\n\u200a\u20099, \nLuke D. Knibbs9,10, Ying Zhang\u2009\n\u200a\u20099, Guy Marks\u2009\n\u200a\u200911, Jane Heyworth12, Julie Arblaster\u2009\n\u200a\u200913, Yue Leon Guo\u2009\n\u200a\u200914, \nLidia Morawska\u2009\n\u200a\u200915, Micheline S. Z. S. Coelho\u2009\n\u200a\u200916, Paulo H. N. Saldiva16, Patricia Matus17, Peng Bi\u2009\n\u200a\u200918, Simon Hales\u2009\n\u200a\u200919, \nWenbiao Hu20, Dung Phung\u2009\n\u200a\u200921, Yuming Guo\u2009\n\u200a\u20091,22 & Shanshan Li\u2009\n\u200a\u20091,22\u2009\n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\nExtended Data Fig. 1 | The respiratory hospitalization risks associated with a \n1\u2009\u00b5g/m3 increase in wildfire-specific PM2.5 over the lag 0\u20137 days. The center \nof the error bars represents the mean effect estimates comparing days with \nlower versus higher wildfire-specific PM2.5 levels across 1052 study communities, \nwith error bars indicating uncertainty ranges derived from the 95% CIs of 1,000 \nMonte Carlo samples. The models controlled for temperature, relative humidity, \nnon-wildfire PM2.5, day of week, seasonality and long-term trend. Abbreviations: \nPM2.5, particulate matter with diameter \u2264 2.5\u2009\u00b5m; CI, confidence interval.\n\nNature Sustainability\nArticle\nhttps://doi.org/10.1038/s41893-025-01533-9\nExtended Data Fig. 2 | Spatiotemporal pattern of average annual attributable \nfraction (%) of all-cause respiratory hospitalizations, attributed to wildfire-\nspecific and non-wildfire PM2.5, during 2000\u20132019. 2000\u201304, years from \n2000 to 2004; 2005\u201309, years from 2005 to 2009; 2010\u201314, years from \n2010 to 2014; 2015\u201319, years from 2015 to 2019. The black points and line \ncurves represent the temporal trend of difference between the proportion of \nhospitalizations attributable to wildfire-specific PM2.5 and non-wildfire PM2.5, \nmeasured as AF of wildfire-specific PM2.5 minus AF of non-wildfire PM2.5. Models \nfor wildfire-specific PM2.5 were controlled for temperature, relative humidity, \nnon-wildfire PM2.5, day of week, seasonality and long-term trend. Models for \nnon-wildfire PM2.5 were controlled for temperature, relative humidity, wildfire-\nspecific PM2.5, day of week, seasonality and long-term trend. Abbreviations: AF, \nattributable fraction; AUS, Australia; BRA, Brazil; CAN Canada; CHL, Chile; NZL, \nNew Zealand; THA, Thailand; TWN, Taiwan; and VNM, Vietnam.\n\n\n    Scientific Research Findings:", "answer": "We found that wildfire-specific PM2.5 emerged as a significant source of respiratory hospitalization risks from short-term PM2.5 exposure. It posed a greater risk for all major types of respiratory diseases than non-wildfire PM2.5. Annually, around 25,321 respiratory hospitalizations were attributable to wildfire-specific PM2.5, with its proportion relative to total respiratory hospitalizations increasing in Australia, Vietnam and Taiwan during 2000\u20122019. Wildfire\u2019s substantial contribution to PM2.5-linked respiratory hospitalizations demands continued relief efforts across most countries and territories. Vulnerable disease types and populations needing targeted intervention were influenza, children and adolescents, older individuals, individuals in low-income and high-polluted communities, and residents of Brazil, Thailand, Taiwan and Vietnam. Our findings can be generalized to regions sharing similar socio-demographic characteristics with our study communities. Despite extensive spatiotemporal coverage, this study cannot be interpreted as a global representative because it only included seven countries and territories with the best data availability.", "id": 81}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465241265435\nJournal of Health and Social Behavior\n2024, Vol. 65(4) 468\u00ad\u2013488\n\u00a9 American Sociological Association 2024\nDOI: 10.1177/00221465241265435\njournals.sagepub.com/home/hsb\nOriginal Article\nUnderstanding variations in mental well-being has \nbecome a globally salient health challenge (Lesthaeghe \n1995; Van de Kaa 1987). The \u201csecond demographic \ntransition\u201d has led to significant shifts in work and \nfamily dynamics, creating complex life trajectories \nthat warrant an examination of their impacts on \nmental health, especially among women who face \nunique stressors due to societal gender roles and \nexpectations (Caballero et al. 2022; Zagel and Van \nWinkle 2020).\nIn industrial societies, demographic, economic, \nand cultural transformations have made employment \ntrajectories unpredictable, with critical factors such \nas occupational stress (Chandola et al. 2007), \nresource accumulation (Lantz et al. 2005; Read, \nGrundy, and Foverskov 2016), and social network \ndevelopment (Grundy and Sloggett 2003) impacting \nmental health. Concurrently, family formation has \nevolved with trends such as increasing cohabitation \n(Heuveline and Timberlake 2004), higher divorce \nrates (Schoen and Canudas-Romo 2006), remarriage \n(Coleman, Ganong, and Fine 2000), and the rise of \nsingle-parent families (Heuveline, Timberlake, and \nFurstenberg 2003) affecting mental well-being.\nThe relationship between work, family, and \nhealth necessitates understanding how life course \ntrajectories, particularly for women, influence later \nlife mental health (Chen et al. 2018; Dannefer \n2003). Women, especially those with social \n1265435 HSBXXX10.1177/00221465241265435Journal of Health and Social BehaviorAzar\nresearch-article2024\n1Purdue University, West Lafayette, IN, USA\nCorresponding Author:\nAriel Azar, Department of Sociology, Purdue University, \n100 University Street, Suite 1114, West Lafayette, IN \n47907, USA. \nEmail: arazar@purdue.edu\nWork\u2013Family Life Course \nTrajectories and Women\u2019s \nMental Health: The Moderating \nRole of Defamilization Policies \nin 15 European Territories\nAriel Azar1\nAbstract\nThis study employs multichannel sequence analysis of data from the Survey of Health, Ageing, and Retirement \nin Europe to explore variations in the association between work\u2013family life trajectories and women\u2019s \nmental health across European cohorts born between 1924 and 1965 within different policy contexts. \nIt finds that trajectories characterized by prolonged employment and delayed familial commitments are \ngenerally associated with increased depressive symptoms. Notably, the strength of this association varies \nsignificantly across cohorts and is notably moderated by defamilization policies. These policies, which aim \nto reduce dependency on family for managing social risks, buffer mental health challenges in traditional \nfamily roles but are less effective for women in trajectories with delayed family formation. This investigation \nhighlights the nuanced ways in which historical and cultural contexts alongside policy environments shape \nmental health outcomes at various life stages, offering valuable insights into our understanding of health \ndisparities across the life course, with an emphasis on exposure to changing institutions.\nKeywords\ndefamilization, life course, mental health, sequence analysis, work\u2013family conflict\n\nAzar\t\n469\ndisadvantages, often experience greater health \nrisks, including hypertension, depression, disabil-\nity, and heart disease (Anand, Esposito, and \nVillase\u00f1or 2018; Diaz-Toro et al. 2018; Engels et al. \n2019; Land\u00f6s et al. 2018; Leupp 2017; Lucumi  \net al. 2017).\nThis study investigates how gendered welfare \nstate policies, particularly defamilization policies, \naffect these dynamics. Defamilization policies, \nwhich aim to reduce individuals\u2019 dependency on \nfamily for welfare and economic security, may \ninfluence the mental health of women by facilitat-\ning or hindering their ability to balance work and \ncaregiving roles (Lohmann and Zagel 2016). I \nexamine the impact of such policies using \nSHARELIFE and country-year data covering 14 \nEuropean countries to understand their role in the \nlink between midlife work\u2013family formation pat-\nterns and mental health later in life.\nGiven Europe\u2019s varied welfare state models and \ndiverse cultural dynamics, this research provides \ninsights into how policy environments interact with \nlife trajectories to affect women\u2019s mental health. By \nanalyzing these relationships, the study seeks to \nunderstand how policy interventions might mitigate \nthe health impacts of work\u2013family conflicts. The \nsubsequent sections detail the theoretical frame-\nworks, working hypotheses, methodologies, and \nfindings, culminating in a discussion that contextu-\nalizes the results within the broader literature on \nhealth, life events, and welfare policy interaction.\nBackground\nWork, Family, and Mental Health\nResearch on disparities in mental health has increas-\ningly focused on the interrelation of work and fam-\nily arrangements, highlighting three main areas of \nimpact: the effects of work status and labor force \nattachment, changes in family structures, and the \ninteraction between work and family dynamics.\nSeveral studies underline a robust link between \nemployment status and mental health outcomes. \nJob security, workplace environments, and employ-\nment status significantly influence disparities in \nmental health (Coduti et al. 2015; Leupp 2017; \nLaMontagne et al. 2021; Llena-Nozal, Lindeboom, \nand Portrait 2004; Sverke, Hellgren, and N\u00e4swall \n2002; Witte 1999). Unemployment or precarious \nemployment often leads to increased stress, depres-\nsion, and anxiety, exacerbated by financial instabil-\nity and impaired self-esteem, where employment \nserves as a core part of personal identity (Brydsten, \nHammarstr\u00f6m, and San Sebastian 2018; Jahoda \n1982; Paul and Moser 2009; van der Noordt et al. \n2014). Consequently, stable employment provides \nfinancial security and a sense of purpose and \nbelonging, thereby bolstering mental health.\nShifts in family structures also significantly \nimpact mental health. Changes in marital status, \ncohabitation dynamics, parenting roles, and house-\nhold compositions have all been linked to mental \nhealth outcomes (Brown, Bulanda, and Lee 2005; \nDinh et al. 2017; Lawton, Moss, and Kleban 1984; \nMcDonald et al. 2022; Taylor and Ford 1983; \nWilliams and Umberson 2004). Life transitions, \nsuch as divorce, widowhood, or transitioning to \nsingle parenthood, can heighten stress and mental \nhealth risks, with stable family environments pro-\nviding crucial emotional and practical support \n(Cohen and Wills 1985). Conversely, family disrup-\ntions can result in loss of support and increased \nstress due to financial and emotional challenges, \nadversely affecting mental health (Gaydosh and \nHarris 2018).\nThe intersection of work and family life, partic-\nularly work\u2013family conflict, has been a growing \nfocus due to its profound impact on mental health. \nBalancing these duties often leads to stress, burn-\nout, and mental health issues, with theories like role \nstrain and role conflict suggesting that stress arises \nwhen simultaneous social roles conflict (Allen et al. \n2000; Greenhaus and Beutell 1985; Schieman and \nGlavin 2011). Inadequate workplace policies or \nfamily structures that do not support role integra-\ntion can exacerbate this stress, leading to psycho-\nlogical exhaustion and subsequent mental health \ndeclines over time.\nIn conclusion, the intricate interplay between \nwork and family dynamics significantly shapes \nmental health. This body of research has adopted a \nlife course perspective, examining how changes in \nwork and family influence health throughout an \nindividual\u2019s life, utilizing frameworks such as criti-\ncal periods and cumulative disadvantages to deepen \nunderstanding of these dynamics.\nLife Course Trajectories and Mental Health\nWork and family life significantly affect health and \nwell-being throughout different life stages. Long-\nterm labor market attachment is associated with bet-\nter mental health, whereas weaker attachments, such \nas part-time work or homemaking, can have nega-\ntive impacts (Madero-Cabib, Azar, and P\u00e9rez-Cruz \n2019; Montez et al. 2015; Ross and Mirowsky 2002; \nWahrendorf et al. 2013). This is attributed to the \n\n470\t\nJournal of Health and Social Behavior 65(4) \naccumulation of material resources, a sense of pur-\npose, and social connections. In the family domain, \nstable relationships, such as marriage and parent-\nhood, enhance health through stronger social support \n(Grundy and Tomassini 2010; Kravdal et al. 2012; \nLacey et al. 2016; Read, Grundy, and Wolf 2011). \nConversely, marital disruptions, such as divorce or \nwidowhood, often lead to psychological stress and \nwell-being decline (Liu 2012; Umberson, Thomeer, \nand Williams 2013), with significant impacts \nobserved particularly in women due to the loss of \nmarital benefits and the stress of marital dissolution \n(Amato 2000; Williams and Umberson 2004).\nThe timing and sequence of work and family \nexperiences have historically varied. In sociological \nlife course research, concepts such as differentia-\ntion, (de)standardization, and institutionalization \nexplain these changes. Differentiation (Br\u00fcckner \nand Mayer 2005; Mayer 1991) refers to the increas-\ning variety of life statuses individuals encounter, \nleading to greater life course variability. This is \nexemplified by more frequent job changes (Mertens \n1998) and higher rates of divorce, remarriage, and \nchildlessness, indicating diverse life trajectories \n(Br\u00fcckner and Mayer 2005). Standardization refers \nto the process where specific life transitions or tim-\nings become normative within a population. For \nexample, once highly standardized, retirement \nshows varied patterns with multiple transitions, \nindicating destandardization (Calvo, Madero-\nCabib, and Staudinger 2018). Institutionalization \nrelates to the regulation of life stages through \nnorms, laws, or organizational rules (Br\u00fcckner and \nMayer 2005). Whereas traditional marriage was \nonce a strictly institutionalized norm, the recogni-\ntion of diverse unions over time illustrates its partial \ndeinstitutionalization.\nI use two theoretical models to analyze the link \nbetween diversified work\u2013family trajectories and \nmental health: the cumulative (dis)advantage model \nand role enhancement/conflict theory. As defined \nby Merton (1968), cumulative advantages are the \nprocess in which early benefits increase inequalities \nover time. In the context of health research, the \naccumulation of (dis)advantages in areas such as \neducation, employment, and family ties leads to \ndivergent health trajectories over time (Cunningham \net al. 2018; Dannefer 2003; DiPrete and Eirich \n2006; Madero-Cabib, Undurraga, and Valenzuela \n2019; Singh et al. 2017). Consequently, individuals, \nparticularly women, with life paths marked by \nlong-term harmful exposure in these domains may \nexperience worsening mental health due to cumula-\ntive stress.\nYet concurrent roles within work and family \ndimensions are pivotal, with role enhancement the-\nory (Ahrens and Ryff 2006; Barnett and Hyde \n2001) arguing that multiple roles improve psycho-\nlogical well-being through empowerment, resource \navailability, and emotional satisfaction (Muller and \nLitwin 2011). Conversely, role conflict theory sug-\ngests that stress arises from the incompatible \ndemands of these roles, leading to adverse psycho-\nlogical effects (Gove 1984; Lahelma et al. 2002). \nThis is exemplified by findings that employment \nbenefits women\u2019s mental health unless offset by the \ndemanding care of young children, with such role \nconflict diminishing as children age (Leupp 2017; \nTosi and Grundy 2019).\nWork\u2013family trajectories impact mental health \ndifferently, affected by the stress of role conflict or \nthe rewards from work and family life. These expe-\nriences are seen as cumulative advantages or disad-\nvantages through a life course lens. Contextual \nfactors can modify these effects, making certain \ndisadvantages less impactful. Therefore, I argue \nthat a nuanced understanding of these dynamics \nrequires a comprehensive contextual framework.\nBringing the Welfare State into the Picture\nIn this study, I aim to contribute to the existing  \nliterature by assessing how accumulated social \nadvantages or disadvantages from work\u2013family tra-\njectories influence the mental health of European \nwomen over 50. I also examine the role of \ndefamilization policies in moderating this effect, a \nperspective that has not been extensively explored \nin previous research. Welfare states, as crucial pro-\nviders of welfare services and influencers of well-\nbeing (Bambra and Eikemo 2008; Esping-Andersen \n1990), shape individual opportunities and life tra-\njectories in both work and family spheres, thereby \nimpacting health outcomes.\nEsping-Andersen\u2019s (1990) influential work, The \nThree Worlds of Welfare Capitalism, provides a \nwidely recognized framework for categorizing wel-\nfare states, including examples from Europe, the \nUnited States, and Japan. This typology, which \nfocuses on reducing market dependency and social \nstratification and delineating the welfare responsibili-\nties between state, market, and family, informs much \ncomparative welfare state research and is founda-\ntional in understanding varying work and family \nexperiences (Corna 2013; Esping-Andersen 1997).\nResearch leveraging these typologies indicates \nthat welfare state configurations influence social \nstratification and its impact on health (Bambra, \n\nAzar\t\n471\nSmith, and Pearce 2019; Espelt et al. 2008; Sacker, \nWorts, and McDonough 2011). Notably, social-\ndemocratic welfare states mitigate the adverse \nhealth effects of specific old-age labor trajectories \nmore effectively than corporatist states (Madero-\nCabib, Corna, and Baumann 2019). However, the \nbroad scope of these typologies can obscure the \nspecific institutional configurations or policy areas \nthat moderate the relationship between life course \nconditions and health.\nRecognizing the gendered dimensions of social \nstratification, scholars have developed gender-\nfocused typologies based on \u201cdefamilization\u201d \n(Bambra 2007; Saxonberg 2013). Despite these \ndevelopments, the gender perspective remains \ninsufficiently explored in comparative health \nresearch (Uccheddu et al. 2019). This study aims to \nfill this gap by incorporating a gender perspective \nin the analysis of welfare states, particularly in the \ncontext of health disparities. Bambra et al. (2009) \nstress this importance, noting that although social \ndemocratic states have reduced income-related \nhealth disparities, these benefits are not consistently \nextended to women (Mackenbach 2012).\nFurthermore, viewing welfare states as evolving \nentities influenced by political and cultural dynam-\nics is essential (Abbott and DeViney 1992). A mere \nsnapshot of welfare state policies at a single point \ndoes not adequately capture the interplay between \nindividual life courses and welfare state dynamics \naffecting health. Understanding the impact of insti-\ntutional configurations on health requires an appre-\nciation of varied institutional setups encountered by \nindividuals across different life stages, shaped by \nthe timing of their birth and historical policy \nreforms.\nThis research aims to bridge political sociology \nand social epidemiology by evolving beyond static \ndefinitions of welfare states and shifting from \nbroad, ungendered interpretations of welfare policy \narrangements. Although general typologies help \nidentify overarching welfare state trends, they often \nfail to capture intracountry variations across policy \nsectors that could significantly influence health \noutcomes.\nBy focusing on Europe, this study provides a \nfoundation for exploring a dynamic, nuanced, and \ngendered perspective of the welfare state in life tra-\njectory and health research. The cross-national life \ncourse approach illuminates how diverse institu-\ntional configurations shape individual life paths and \nhealth disparities, enhancing our understanding of \nhow welfare states can promote health equity \nthrough equitable life trajectories.\nDefamilization and Health\nDefamilization is a crucial concept in understanding \nthe interplay between gender, work, and family. \nHistorically burdened with caregiving and house-\nhold duties, women\u2019s engagement in the labor market \nhas been significantly influenced by defamilization \npolicies. These policies aim to reduce the caregiving \nresponsibilities traditionally assigned to women, \nenhancing their autonomy and ability to participate \nfully in the labor force. Such measures help balance \nwork and family life and lessen financial depen-\ndence within families (Bambra 2007; Chau et al. \n2017; Lohmann and Zagel 2016).\nThe effectiveness of defamilization policies is \nevident in their capacity to redistribute paid and \nunpaid labor, which is vital for understanding how \nlife course trajectories in work and family contexts \nimpact health. For example, in environments with \nminimal defamilization support, women juggling \nintensive domestic and professional roles may face \ndifferent mental health challenges compared to \nthose in contexts with substantial state support for \nunpaid labor. Similarly, the mental health outcomes \nin traditional family settings with long-term labor \nforce attachment can vary significantly depending \non the level of defamilization.\nDespite the apparent links, few studies have \nthoroughly investigated the relationship between \ndefamilization and mental health, particularly through \nthe lens of gender disparities. Most research utilizes \ntraditional welfare state typologies, focusing on \ninstitutional frameworks without considering indi-\nviduals\u2019 dynamic exposure to these policies \nthroughout their life courses (Bambra et al. 2009; \nChung et al. 2013).\nThis study proposes that defamilization policies \nmitigate the stress associated with specific work\u2013\nfamily trajectories, influencing mental health out-\ncomes later in life. By examining the role that these \npolicies play in the association between life trajec-\ntories and mental health later in life, we can better \nunderstand how welfare states might alleviate the \nstress induced by particular life paths, providing \ncrucial insights into the role of policy in shaping \nhealth across different life stages.\nWorking Hypotheses\nBased on the comprehensive literature review that \nexplores the interplay between work\u2013family dynam-\nics and mental health over the life course and the \ninfluence of welfare states in shaping this relation-\nship, this study posits several hypotheses to clarify \nthe connections between life course trajectories, \n\n472\t\nJournal of Health and Social Behavior 65(4) \ndefamilization policies, and women\u2019s mental health \nin various European settings and across different \ncohorts:\nHypothesis 1: Life course trajectories and men-\ntal health: Women\u2019s mental health in later life is \nsignificantly influenced by their work and fam-\nily trajectories. Stable employment and family \nstructures throughout life are hypothesized to \nlead to better mental health outcomes than frag-\nmented or unstable work and family histories.\nHypothesis 2: Cohort differences in the impact \nof life trajectories on mental health: The impact \nof life course trajectories on later life mental \nhealth outcomes varies across birth cohorts. \nChanges in societal norms, employment pat-\nterns, family structures, and welfare policies \nover time may cause later cohorts to experience \ndifferent mental health impacts compared to ear-\nlier ones.\nHypothesis 3: Moderating role of defamilization \npolicies: The level of defamilization policies \nwithin a country is expected to moderate the \nimpact of work\u2013family life trajectories on men-\ntal health outcomes. Higher levels of defamiliza-\ntion are predicted to mitigate the adverse effects \nof challenging work and family trajectories, \nreflecting the policies\u2019 supportive role in enhanc-\ning autonomy and reducing dependency on fam-\nily roles for welfare.\nThese hypotheses aim to dissect the complex \ndynamics affecting women\u2019s mental health by con-\nsidering the roles of individual life trajectories, soci-\netal changes, and state policies. The study seeks to \nprovide an integrated understanding of how various \nfactors interact to influence health and well-being \namong women in diverse European environments.\nData And Methods\nData Source and Sample\nI utilized data from the Survey of Health, Ageing, \nand Retirement in Europe (SHARE), a panel survey \nrepresentative of individuals ages 50 and older \nacross various European countries. SHARE covers \nmany topics, including economic, demographic, \nhealth, and family matters (Schr\u00f6der 2011). \nSpecifically, I focused on the SHARELIFE data col-\nlected during the third and seventh waves of \nSHARE, which provides detailed life histories, \nincluding employment and family formation experi-\nences from childhood to advanced age, for women \nborn between 1924 and 1965. More details on the \ndata source are provided in Appendix A in the online \nversion of the article.\nI analyzed the relationship between life trajecto-\nries and the risk of depression by merging \nSHARELIFE data with information from Waves 1 \nthrough 7 of SHARE. This integration resulted in a \nfinal sample of 22,250 women who had compre-\nhensive life histories from ages 15 to 50 and data on \ndepressive symptoms. This sample was initially \nidentified from SHARELIFE and then supple-\nmented with the latest observations from other \nSHARE waves to include mental health data and \nother covariates. The structure of the analytical \ndataset is explained in detail in Appendix A in the \nonline version of the article.\nTo address potential selection biases due to \nhealth issues influencing work\u2013family trajectories \nand subsequent health (Case and Paxson 2011; \nHaas 2008; Haas, Glymour, and Berkman 2011), I \nconducted robustness checks, excluding those with \nsignificant early health issues, shown in Figure A.2 \nin the online version of the article.\nMeasurement\nResearch on life courses and health often overlooks \nmental health (Angelini, Howdon, and Mierau \n2019) despite evidence of it being affected in the \nlong term by life circumstances (Luo and Waite \n2005; Tani et al. 2016). The EURO-D scale, from \nSHARE, measures depressive symptoms for cross-\nnational comparisons in European populations, \nassessing 12 symptoms, such as depression, pessi-\nmism, and fatigue (for a detailed explanation of the \nEURO-D scale and its items, see Appendix B in the \nonline version of the article). Based on EURO-D\u2019s \nassociation with clinical depression, I used a thresh-\nold of four or more symptoms to indicate a high risk \nof depression.\nTo build work\u2013family life histories, participants \nwere prompted to recall significant milestones, \nsuch as job-related changes and key family events, \nannually. These included moving out of their par-\nent\u2019s home, cohabiting, marriage, childbirth, wid-\nowhood, and divorce. Leveraging these data, I \ncrafted a life history data set in which each individ-\nual is observed annually from age 15 up to their age \nat the time of the survey. Their recorded state \nremained unchanged until they indicated a change \nin status. This information was used as an input for \nsequence analysis to construct work\u2013family life \nhistories, as detailed in the empirical strategy sec-\ntion. For more information on the measurement of \nwork\u2013family statuses throughout the life course, \nsee Appendix B in the online version of the article.\n\nAzar\t\n473\nRegarding the moderating variable, defamiliza-\ntion, I used a country-year indicator elaborated by \nZagel and Van Winkle (2020), which relies on prin-\ncipal components analysis, applied to indicators for \nsix policy areas between 1924 and 2006 that seek to \nremove the burden of family work exclusively from \nwomen, as detailed in Appendix B in the online ver-\nsion of the article. Using the country-year index, I \nestimated an average for when individuals were 15 \nto 40 as an indicator of the level of exposure to \ndefamilization policies. For example, a woman \nborn in 1940 in Italy had a defamilization exposure \nlevel calculated as the average for all values in Italy \nbetween 1955 and 1990.\nFinally, all models included controls for demo-\ngraphic and health-risk-related variables. For \nsociodemographic indicators, all models included \nage at the time of the interview, educational level, \nincome, and occupation of the primary breadwinner \nat age 10 as a proxy for childhood socioeconomic \nstatus (SES). For health-related variables, current \nage and smoking status were considered as lifestyle \nindicators, childhood self-rated health was included \nto control for possible health selection effects into \nemployment and family trajectories, and self-rated \nhealth at the time of interview for physical health \nwas also included. Also, economic development, \nmeasured as the gross domestic product (GDP) per \ncapita averaged at the same years as the defamiliza-\ntion indices, was included at the country level.\nEmpirical Strategy\nMy analysis comprised two stages: starting with \napplying multichannel sequence analysis (MCSA) \nand cluster analysis to build simultaneous career \nand family formation trajectories and followed by \nlogistic regression models to examine the likelihood \nof being at high risk of depression across these tra-\njectories. MCSA, an advanced form of sequence \nanalysis (MacIndoe and Abbott 2011), facilitates the \ncomparison of life trajectories in multiple domains \nby using operations such as insertion/deletion and \nstate substitution for optimal matching. This led to \nvarious work\u2013family trajectories, with costs based \non the inverse of state probabilities and a constant \ninsertion/deletion cost.\nHierarchical agglomerative clustering via the \nWard algorithm was utilized to categorize these tra-\njectories, using Hubert\u2019s gamma, point biserial cor-\nrelation, average silhouette width, and Hubert\u2019s C \nfor selecting the best cluster solution, resulting in \nsix distinct types (Studer 2013).\nIn the second stage, I linked these trajectories \nwith the risk of depression variable, employing \nfixed effects logistic regression models with adjust-\nments for country and cohort. This approach facili-\ntated investigating variations in the relationship \nacross different defamilization index levels and \ncohorts, providing insights into the interaction \neffects and average marginal effects.\nFor more details on the empirical strategy, see \nAppendix C in the online version of the article.\nResults\nFigures 1 and 2 show each trajectory type\u2019s chrono-\ngram and sequence index plots. The chronogram \ndepicts the work and family status percentages for \nages 15 to 50, with employment on the left and fam-\nily trajectories on the right. The x-axis tracks ages \nfrom 15 to 50, and the legend lists 7 employment \nand 12 family statuses. The sequence index plot in \nFigure 2 illustrates individual trajectories with lines \nfor each person, where color changes indicate status \ntransitions, not group proportions.\nHere, I use the \u201ccareerist, late union\u201d trajectory \ntype, which comprises 8.1% of the sample, as an \nexample to understand Figures 1 and 2. The chro-\nnogram shows several statuses at age 15, the start-\ning point (education, employment, family work), \nwith a shift in the proportion from education to \nemployment as individuals age. The sequence \nindex plot in Figure 2 displays individual paths, \nhighlighting prolonged employment with intermit-\ntent family work for most women. This illustrates \nthat the chronogram\u2019s proportions reflect transi-\ntions over time, not constant roles, clarifying that \nthe observed employment or family work percent-\nages result from these individual trajectories.\nThe careerist, late union type involves women \nmainly employed, leaving parental homes unpart-\nnered, and then cohabiting and marrying without \nchildren. The \u201chomemaker archetype\u201d (29.6%) fea-\ntures women primarily in home/family work, tran-\nsitioning from parental home to marriage and \nmotherhood. \n\u201cEntrepreneur \nfamilist\u201d \n(7.4%) \nincludes women consistently in self-employment, \nwith family paths similar to homemakers.\n\u201cStable job, traditional family\u201d (40.3%) repre-\nsents the largest group with stable employment and \ntraditional family life. Stable job, family diverse \nshows similar employment but varied family paths, \nwith women approaching 50 in different family sta-\ntuses. \u201cJob flux, traditional family\u201d describes \nwomen with fluctuating employment but traditional \nfamily trajectories.\nFigure 3 depicts the distribution of the trajectory \ntypes by country on the left and by cohort on the \nright. Variability between countries is significant. In \n\n474\t\nJournal of Health and Social Behavior 65(4) \nall countries, the modal trajectory is either the \nhomemaker archetype or the stable job, traditional \nfamily. Southern European nations and the \nNetherlands predominantly display the former type.\nBy contrast, countries such as Sweden, \nDenmark, Eastern Germany, and the Czech Republic \nshow a higher prevalence of the latter. There is also \nnoteworthy variation across cohorts. Later cohorts \nlean toward a higher prevalence of stable job, tradi-\ntional family, whereas the homemaker archetype \nremains dominant in earlier cohorts. Notably, a sig-\nnificantly heightened prevalence of nontraditional \nFigure 1.\u2002 Chronogram Plots of Multichannel Trajectory Types from Ages 15 to 50 in the Work and \nFamily Dimensions, the Survey of Health, Ageing and Retirement in Europe (SHARE).\nNote: N\u2009=\u200926,274. \n\nAzar\t\n475\nfamily formation types also exists in nations and \ncohorts where stable job, traditional family emerges \nas the predominant trajectory.\nTable 1 details the variables used in the study, \nwith 30% reporting over three symptoms. The aver-\nage age of respondents is 62 (SD \u2248 9 years). \nSocioeconomically, half have upper secondary/\nvocational education, and the average household \nincome is around EUR 33,000. Predominantly, \nwomen\u2019s childhood households had blue-collar \nbreadwinners. Health-wise, 17% smoke; 40% drink \nweekly; current and childhood self-rated health \nFigure 2.\u2002 Index Plots of Multichannel Trajectory Types from Ages 15 to 50 in the Work and Family \nDimensions, the Survey of Health, Ageing and Retirement in Europe (SHARE).\nNote: N\u2009=\u200926,274.\n\n476\n7\n57\n6\n21\n7\n2\n7\n51\n10\n20\n5\n6\n13\n47\n8\n21\n7\n5\n10\n46\n12\n20\n4\n8\n8\n40\n8\n29\n5\n11\n9\n37\n6\n34\n11\n4\n14\n35\n8\n28\n11\n4\n8\n35\n9\n31\n15\n2\n10\n29\n7\n37\n13\n4\n7\n25\n7\n41\n15\n5\n4\n24\n18\n42\n9\n3\n9\n13\n6\n54\n17\n1\n8\n12\n5\n51\n24\n1\n7\n8\n4\n64\n16\n1\n4\n3\n2\n72\n18\n0\n0\n20\n40\n60\n80\n100\nIreland\nSpain\nNetherlands\nGreece\nItaly\nWest Germany\nSwitzerland\nAustria\nBelgium\nFrance\nPoland\nDenmark\nSweden\nEast Germany\nCzech Republic\n8\n18\n7\n49\n10\n7\n9\n20\n7\n42\n19\n4\n11\n22\n7\n42\n14\n5\n9\n23\n8\n39\n16\n5\n9\n26\n7\n43\n13\n2\n8\n29\n8\n41\n11\n4\n5\n34\n7\n38\n11\n5\n7\n36\n7\n35\n9\n5\n6\n42\n8\n32\n6\n6\n6\n44\n8\n28\n7\n7\n7\n47\n8\n25\n5\n8\n8\n47\n11\n23\n5\n6\n8\n52\n9\n22\n3\n6\n8\n48\n10\n22\n5\n7\n0\n20\n40\n60\n80\n100\n1963\u20131966\n1960\u20131962\n1957\u20131959\n1954\u20131956\n1951\u20131953\n1948\u20131950\n1945\u20131947\n1942\u20131944\n1939\u20131941\n1936\u20131938\n1933\u20131935\n1930\u20131932\n1927\u20131929\n1924\u20131926\n1\n2\n3\n4\n5\n6\nFigure 3.\u2002 Distribution of Trajectory Types across Countries and Cohorts, the Survey of Health, Ageing and Retirement in Europe (SHARE).\nNote: Percentages are shown over the bars. N\u2009=\u200926,274. Trajectory types correspond as follows: 1 = careerist, late union; 2 = homemaker archetype; 3 = entrepreneur familist; 4 = \nstable job, traditional family; 5 = stable job, family diverse; 6 = job flux, traditional family.\n\nAzar\t\n477\naverages are 2.97 and 3.26, respectively; and 10% \nhad an illness before age 50. The standardized aver-\nages for defamilization and economic development \nexposure are .17 and 10.35, respectively.\nTable 2 examines the link between trajectory \ntypes, covariates, and risk of depression. Women \nwith stable job, traditional family or entrepreneur \nfamilist trajectories have a lower risk of depression. \nBy contrast, careerist, late union individuals have a \nrelatively higher prevalence in the high-risk depres-\nsion group. The risk of depression correlates with \nlower defamilization, with women at higher risk \nhaving experienced significantly lower levels of \ndefamilization. Typically, women at higher risk of \ndepression are older, have lower education and \nincome, come from lower childhood SES, and rate \ntheir current and childhood health as poor, with a \nhigher prevalence of pre-50 illness. Notably, they \ndrink less, but this finding reverses and loses sig-\nnificance in the multivariate analysis.\nTo better understand these associations, I con-\nducted a logistic regression to predict the post-50 \nlikelihood of reporting a high risk of depression \nconsidering work/family trajectories, birth cohort, \ndefamilization, and other covariates. The analysis \nincludes interactions between trajectories, cohorts, \nand defamilization. Models feature country and \nyear-of-interview fixed effects, clustering standard \nerrors by country-cohort.\nTable 3 presents four logistic regression models \nanalyzing the likelihood of reporting a high risk of \ndepression based on work and family trajectories. \nModel 1 indicates that women in the entrepreneur \nfamilist trajectory have significantly lower odds of \nbeing at high risk of depression compared to those \nin the homemaker archetype. By contrast, stable \njob, family diverse women exhibit approximately \n16% higher odds. In Model 2, which adjusts for \naverage GDP per capita from ages 15 to 50, a 1 SD \nincrease in defamilization exposure is associated \nwith a 5.8 percentage point rise in the likelihood of \ndepression. Notably, this association is consider-\nable, especially when compared to the effect size of \neducational attainment; the difference in the pre-\ndicted probability of reporting more than three \ndepressive symptoms between women with less \nthan upper secondary education and those with ter-\ntiary education or more is 3.0 percentage points, as \nestimated from Model 2. Furthermore, the signifi-\ncant magnitude of this association appears even \nTable 1.\u2002 Descriptive Statistics, the Survey of Health, Ageing and Retirement in Europe.\nMean (SD) or %\nDepressive symptoms\n\u2003 \u22643\n69.03\n\u2003 >3\n30.97\nAge\n62.15 (8.89)\nEducational level\n\u2003 Less than upper secondary\n46.34\n\u2003 Upper secondary and vocational training\n34.00\n\u2003 Tertiary\n19.66\nTotal income (in thousands EUR)\n32.92 (47.63)\nBreadwinner occupation at age 10\n\u2003 White collar\n17.67\n\u2003 Blue collar\n81.44\n\u2003 Military\n.89\nSmokes (reference\u2009=\u2009doesn\u2019t smoke)\n17.35\nDrinks weekly (reference\u2009=\u2009doesn\u2019t drink weekly)\n39.18\nSelf-rated health (1\u20135)\n3.02 (1.06)\nIllness before age 50 (reference\u2009=\u2009didn\u2019t experience illness)\n10.59 (1.06)\nSelf-rated childhood health (1\u20135)\n3.26 (.99)\nDefamilization index exposure level (standardized)\n.17 (1.20)\nGDP per capita exposure level (standardized)\n10.35 (4.04)\nN\n21,743\nNote: GDP\u2009=\u2009gross domestic product.\n\n478\t\nJournal of Health and Social Behavior 65(4) \nwhen controlling for country-specific unobserved \nheterogeneity by including country fixed effects, \nindicating that the observed relationship is robust \nand sizable within countries.\nModel 3 reveals significant interactions between \ntrajectory types and cohorts, presented as average \nmarginal effects in Figure 4a. Specifically, the \nentrepreneur familist trajectory correlates with \nlower depression risk, but only in later cohorts. By \ncontrast, the stable job, family diverse trajectory is \nassociated with increased depressive symptoms \namong these cohorts. Intriguingly, the relationship \nbetween the careerist, late union trajectory and \ndepression risk reverses across cohorts: It decreases \nthe risk in earlier cohorts but increases it in later \nones. Adding this interaction to the model slightly \nenhances the model performance, increasing the \npseudo R2 by about 1%. Despite being small, fur-\nther analyses using a bootstrap likelihood ratio test \nwith 1,000 replications showed that this increase \nwas significant.\nModel 4 examines the interaction between \ndefamilization exposure and trajectory types, \nassessing their impact on mental health, with results \nshown in Figure 4b. This interaction is significant \nfor the careerist, late union and entrepreneur \nfamilist trajectories and significantly increases the \nfit of the model relative to Model 2 by around 9%. \nFurther analyses estimating a bootstrap likelihood \nratio test with 1,000 replications showed that the \ninclusion of the defamilization-trajectory interac-\ntion significantly enhanced the model performance. \nWomen in the careerist, late union trajectory expe-\nrience no differences in their risk of depression \nTable 2.\u2002 Bivariate Descriptive Statistics, the Survey of Health, Ageing and Retirement in Europe.\nLow Risk of \nDepression  \n(\u22643 Symptoms)\nHigh Risk of \nDepression  \n(>3 Symptoms)\np Value\nTrajectory type\n\u2003 Homemaker archetype\n68.17\n31.83\n.000\n\u2003 Careerist, late union\n66.08\n33.92\n\u2002\n\u2003 Entrepreneur familist\n70.14\n29.86\n\u2002\n\u2003 Stable job, traditional family\n72.08\n27.92\n\u2002\n\u2003 Stable job, family diverse\n68.23\n31.77\n\u2002\n\u2003 Job flux, traditional family\n62.77\n37.23\n\u2002\nAge\n61.92 (8.68)\n62.66 (9.32)\n.000\nEducational level\n\u2003 Less than upper secondary\n63.18\n36.82\n.000\n\u2003 Upper secondary and vocational training\n72.22\n27.78\n\u2002\n\u2003 Tertiary\n77.29\n22.71\n\u2002\nTotal income (in thousands EUR)\n35.43 (49.99)\n27.31 (41.36)\n.000\nBreadwinner occupation at age 10\n\u2003 White collar\n74.40\n25.60\n.000\n\u2003 Blue collar\n67.83\n32.17\n\u2002\n\u2003 Military\n72.16\n27.84\n\u2002\nSmokes (reference = doesn\u2019t smoke)\n16.63\n18.70\n.000\nDrinks weekly (reference = doesn\u2019t drink weekly)\n41.40\n33.81\n.000\nSelf-rated health (1\u20135)\n2.73 (.99)\n3.52 (1.02)\n.000\nIllness before age 50 (reference = didn\u2019t experience illness)\n8.53\n15.17\n.000\nSelf-rated childhood health (1\u20135)\n3.26 (.99)\n2.48 (1.01)\n.000\nDefamilization index exposure level (standardized)\n.20 (1.23)\n0.09 (1.13)\n.000\nGDP per capita exposure level (standardized)\n10.58 (3.94)\n9.85 (4.21)\n.000\nN\n15,009\n6,734\n\u2002\n%\n69.03\n30.97\n\u2002\nNote: The p values correspond to \u03c72 tests for categorical variables and t tests for continuous variables. Standard \ndeviations are in parentheses. For categorical variables, the conditional distribution of depressive symptomatology \nis shown. For continuous variables, the mean of the variable for each category of depressive symptomatology is \ndisplayed. GDP = gross domestic product.\n\n479\nTable 3.\u2002 Logistic Regression Models over the Probability of Reporting Depressive Symptomatology for Work and Family Trajectory Types, the Survey of \nHealth, Ageing and Retirement in Europe.\nModel 1\nModel 2\nModel 3\nModel 4\n\u2002\nOR\nSE\nOR\nSE\nOR\nSE\nOR\nSE\nTrajectory type (reference = homemaker archetype)\n\u2003 Careerist, late union\n1.030\n(.102)\n1.023\n(.100)\n.493***\n(.116)\n.995\n(.096)\n\u2003 Entrepreneur familist\n.780**\n(.081)\n.771**\n(.080)\n1.111\n(.280)\n.784**\n(.079)\n\u2003 Stable job, traditional family\n.966\n(.064)\n.961\n(.063)\n.886\n(.163)\n.958\n(.064)\n\u2003 Stable job, family diverse\n1.162*\n(.106)\n1.146\n(.103)\n.802\n(.190)\n1.146\n(.107)\n\u2003 Job flux, traditional family\n1.054\n(.114)\n1.051\n(.113)\n.933\n(.257)\n1.059\n(.118)\nCohort\n1.101\n(.090)\n1.079\n(.086)\n1.079\n(.087)\n1.081\n(.086)\nAge\n1.022\n(.027)\n1.041\n(.030)\n1.033\n(.029)\n1.041\n(.030)\nEducational level (reference = less than upper secondary)\n\u2003 Upper secondary and vocational training\n.899\n(.059)\n.904\n(.059)\n.898*\n(.058)\n.902\n(.059)\n\u2003 Tertiary\n.870\n(.066)\n.862*\n(.066)\n.858**\n(.064)\n.856**\n(.065)\nTotal income (logged)\n.953***\n(.013)\n.951***\n(.014)\n.954***\n(.013)\n.951***\n(.014)\nBreadwinner occupation at age 10 (reference = white collar)\n\u2003 Blue collar\n.972\n(.062)\n.971\n(.062)\n.971\n(.062)\n.968\n(.062)\n\u2003 Military\n.750\n(.174)\n.761\n(.174)\n.751\n(.173)\n.759\n(.174)\nSmokes (reference = doesn\u2019t smoke)\n1.088\n(.079)\n1.087\n(.080)\n1.096\n(.082)\n1.093\n(.081)\nDrinks weekly (reference = doesn\u2019t drink weekly)\n1.049\n(.063)\n1.046\n(.062)\n1.047\n(.061)\n1.045\n(.061)\nSelf-rated health (1\u20135)\n.470***\n(.018)\n.472***\n(.018)\n.470***\n(.018)\n.471***\n(.018)\nIllness before age 50 (reference = didn\u2019t experience illness)\n1.284***\n(.114)\n1.276***\n(.113)\n1.295***\n(.115)\n1.279***\n(.114)\nSelf-rated childhood health (1\u20135)\n1.077***\n(.024)\n1.083***\n(.024)\n1.082***\n(.025)\n1.085***\n(.024)\nDefamilization index average exposure\n1.350***\n(.116)\n1.250*\n(.147)\nGDP per capita average exposure\n1.033\n(.034)\n1.05\n(.035)\n1.035\n(.034)\nTrajectory type \u00d7 Cohort\n\u2003 Careerist, late union\n1.095***\n(.031)\n\u2002\n\u2003 Entrepreneur familist\n.953\n(.030)\n\u2002\n\u2003 Stable job, traditional family\n1.012\n(.024)\n\u2002\n(continued)\n\n480\nModel 1\nModel 2\nModel 3\nModel 4\n\u2002\nOR\nSE\nOR\nSE\nOR\nSE\nOR\nSE\n\u2003 Stable job, family diverse\n1.044\n(.029)\n\u2002\n\u2003 Job flux, traditional family\n1.018\n(.034)\n\u2002\nTrajectory type \u00d7 Defamilization\n\u2003 Careerist, late union\n1.281**\n(.141)\n\u2003 Entrepreneur familist\n.826\n(.112)\n\u2003 Stable job, traditional family\n1.089\n(.102)\n\u2003 Stable job, family diverse\n1.075\n(.115)\n\u2003 Job flux, traditional family\n1.048\n(.182)\nConstant\n.667\n(1.445)\n.171\n(.408)\n.242\n(.573)\n.168\n(.399)\nPseudo R2\n.1200\n.1214\n.1219\n.1225\n\u2002\nObservations\n22,100\nNote: GDP = gross domestic product. \n*p < 0.1. **p < 0.05. ***p < 0.01.\nTable 3.\u2002 (continued)\n\nAzar\t\n481\nrelative to the homemaker archetype at low \ndefamilization levels. However, at high defamiliza-\ntion levels, they have a notably higher chance of \nbeing at high risk of depression. Conversely, \nwomen in the entrepreneur familist trajectory show \nno difference in their risk of depression at low \ndefamilization levels but report fewer depressive \nsymptoms at higher levels.\nAddressing the potential selection bias of women \ninto life trajectories due to early or midlife health \nissues, all models adjust for self-reported illnesses \nbefore observing the trajectories. To ensure the \nrobustness of Model 4\u2019s results, I recalculated the \nmodel excluding women who reported illnesses \nbefore age 50, mitigating the bias from health-\ndriven self-selection into trajectories affecting later \nmental health. The Appendix A in the online version \nof the article shows these findings, confirming the \nresults remain consistent with a more conservative \nassessment, excluding these women.\n-.2\n-.1\n0\n.1\n.2\n1924\u20131926\n1927\u20131929\n1930\u20131932\n1933\u20131935\n1936\u20131938\n1939\u20131941\n1942\u20131944\n1945\u20131947\n1948\u20131950\n1951\u20131953\n1954\u20131956\n1957\u20131959\n1960\u20131962\n1963\u20131966\nCareerist, Late Union\n-.2\n-.1\n0\n.1\n.2\n1924\u20131926\n1927\u20131929\n1930\u20131932\n1933\u20131935\n1936\u20131938\n1939\u20131941\n1942\u20131944\n1945\u20131947\n1948\u20131950\n1951\u20131953\n1954\u20131956\n1957\u20131959\n1960\u20131962\n1963\u20131966\nEnterpreneur Familist\n-.2\n-.1\n0\n.1\n.2\n1924\u20131926\n1927\u20131929\n1930\u20131932\n1933\u20131935\n1936\u20131938\n1939\u20131941\n1942\u20131944\n1945\u20131947\n1948\u20131950\n1951\u20131953\n1954\u20131956\n1957\u20131959\n1960\u20131962\n1963\u20131966\nStable Job, Family Diverse\n-.4\n-.2\n0\n.2\n.4\n-1\n-.5\n0\n.5\n1\n1.5\n2\n2.5\n3\nDefamilization (std)\nCareerist, Late Union\n-.4\n-.2\n0\n.2\n.4\nEffects on Pr(Eurod3)\n-1\n-.5\n0\n.5\n1\n1.5\n2\n2.5\n3\nDefamilization (std)\nEnterpreneur Familist\nA\nB\nFigure 4.\u2002 Average Marginal Effect of Selected Work\u2013Family Trajectory Types on the Probability \nof Women Reporting More than Three Depressive Symptoms by Birth Cohort and Defamilization \nLevel Exposure, the Survey of Health, Ageing and Retirement in Europe (SHARE). (a) Cohort. (b) \nDefamilization Level Exposure.\nNote: The 95% confidence intervals are shown in gray. Average marginal effects are relative to following a homemaker \narchetype trajectory type. Defamilization levels exposure values are standardized. N\u2009=\u200922,100.\n\n482\t\nJournal of Health and Social Behavior 65(4) \nDiscussion\nThis study analyzed the life trajectories of European \nwomen born between 1924 and 1965, uncovering \nhow work and family histories relate to mental well-\nbeing (Cohen and Manning 2010). Employing life \ncourse and institutional frameworks, it explored the \nimpact of defamilization policies and generational \nshifts on mental health (Cornwell, Laumann, and \nSchumm 2008; Liefbroer 1999).\nFindings demonstrate a nuanced evolution in \nwork\u2013family patterns across different generations, \naffecting mental health. This aligns with established \nliterature that connects life course events with \nhealth through social roles, family, and work \nengagement (Grundy, Read, and V\u00e4is\u00e4nen 2020; \nRepetti, Taylor, and Seeman 2002; Simon and \nBarrett 2010; Zella and Harper 2020). This research \nsheds light on these dynamics and provides evi-\ndence in support of Hypothesis 1, demonstrating \nthat specific life patterns, such as entrepreneur \nfamilist, may lead to a lower risk of depression, \nreflecting how the unfolding of life trajectories in \nwork\u2013family dimensions can shape mental health \nlater in life. The accumulation of self-employment \nthroughout the life course together with the forma-\ntion of a traditional family can enhance mental \nhealth through the existence of emotional backing, \na feeling of safety, and hands-on help, all of which \nare essential for managing life challenges brought \nby self-employment (Cohen and Wills 1985). These \ninsights emphasize the importance of considering \nthe overall influence of these trajectories on mental \nhealth and how societal norms and institutions may \npromote specific life paths that could adversely \naffect individual well-being when not adhered to.\nThe main objective of this study, however, was \nto extend beyond observing work\u2013family trajecto-\nries in isolation and place them within historical \nand policy contexts. Results reveal the relationship \nbetween life trajectories and mental health across \ndifferent generations. This supports Hypothesis 2, \nwhich states that the association between work\u2013\nfamily life course trajectories and mental health \noutcomes in later life differ across birth cohorts. \nPrevious research indicates that the timing of mar-\nriage affects health, with early marriage being detri-\nmental (Dupre, Beck, and Meadows 2009; Grundy \nand Holt 2000) and late marriage sometimes being \nprotective (Dupre and Nelson 2016; Hughes and \nWaite 2009; McFarland, Hayward, and Brown \n2013). This study suggests that such relationships \nmay depend on the historical context in which life \ncourses unfold and how family interacts with labor \nforce participation. For example, earlier cohort \nwomen who marry late, engage in the labor market \nlong term, and have no children (the careerist, late \nunion type) are at less risk of depression. By con-\ntrast, among later cohorts, this trajectory type cor-\nrelates with a greater risk of depression. It is not the \nindividual choice to focus on a career and delay \nmarriage that inherently impacts personal fulfill-\nment and mental wellness but, rather, the cultural \nnorms and pressures that shape and sometimes \nchallenge these decisions, indicating that societal \nexpectations play a significant role in mental health \noutcomes when understanding them as the result of \nlife and family life paths.\nFor women who have combined long-term self-\nemployment with traditional family roles (entrepre-\nneur familist), those from more recent cohorts \nexperience less risk of depression in later life, a \nreflection of societal changes (Pampel 2011; \nParkinson et al. 2018) toward valuing flexibility \nand autonomy in the professional sphere while \nmaintaining traditional family roles (Scherger, \nNazroo, and May 2016). These shifts suggest a \ngrowing appreciation for balancing self-employ-\nment and family commitments, aligning with mod-\nern ideals of individual choice and work\u2013life \nharmony (Inglehart 1997). In this context, women \nmay achieve greater control and fulfillment in both \nprofessional and personal domains, contributing to \nimproved mental health outcomes.\nInterestingly, this study also highlights that tra-\nditional employment paired with nontraditional \nfamily formation tends to negatively impact mental \nhealth in later cohorts. This finding suggests a con-\nflict because the stability expected from traditional \njobs clashes with the fluidity of nontraditional fam-\nily setups despite growing societal acceptance. The \nresulting stress and mental health difficulties high-\nlight the challenges faced by those navigating non-\ntraditional life courses within rapidly changing \nsocietal norms. This underscores the ongoing need \nfor supportive structures tailored to unconventional \nfamily structures.\nBeyond historical changes, I use an institutional \ntheory perspective (Beckfield et al. 2015) to explore \nhow policy context, particularly defamilization poli-\ncies aimed at reducing caregiving burdens and pro-\nmoting work\u2013life balance (Bambra 2007; Chau et al. \n2017; Lohmann and Zagel 2016), might shape the \nrelationship between life trajectories and mental \nhealth, aligning with Hypothesis 3. These policies, \nwhich intend to support women\u2019s workforce involve-\nment and reduce family financial reliance, interact \nwith individual life trajectories and affect mental \n\nAzar\t\n483\nhealth outcomes (Bambra 2007; Chau et al. 2017; \nLohmann and Zagel 2016). On the one hand, women \nwho pursue nontraditional family trajectories like the \ncareerist, late union type face increased depression \nrisks in highly defamilized environments, indicating \na misalignment between policy support and their \nneeds or experiences. This phenomenon can be \nunderstood through role conflict theory, which sug-\ngests that stress from balancing personal and profes-\nsional life exacerbates mental health issues. Moen \n(2022) and Moen and Roehling\u2019s (2005) concept of \n\u201ccareer mystique plus\u201d further illustrates this, where \nwomen are expected to excel in both career and care-\ngiving roles, with significant stress and penalties for \nfailing to meet these dual expectations.\nOn the other hand, I also find that women adher-\ning to a traditional family path, who marry at norma-\ntive ages and have children, and who are long-term \nself-employed (entrepreneur familist) derive mental \nhealth benefits in high defamilization contexts. Role \nenhancement theory suggests that the positive iden-\ntities and support networks from their family roles \nand self-employment align well with defamilization \npolicies, which alleviate caregiving burdens and \nfacilitate less conflict between family and profes-\nsional roles, thereby improving mental well-being \n(Ahrens and Ryff 2006; Barnett and Hyde 2001). As \npart of these policies, flexible work initiatives sup-\nport better work\u2013family balance and enhance \nemployees\u2019 well-being (Moen et al. 2016).\nThe contrasting impacts of defamilization poli-\ncies on women with varied life trajectories and \nemployment types underscore the complexity of \nunderstanding how policy contexts interact with \ndifferent population segments. This highlights the \nnecessity of addressing how policy contexts engage \nwith women\u2019s diverse life courses. By adopting an \napproach that recognizes the pluralization of work\u2013\nfamily life courses, we can better understand how \ngendered welfare state policies influence mental \nhealth (Bambra 2007).\nThis study is not without limitations. It provides \nhistorical depth by focusing on 14 European coun-\ntries and birth years from 1924 to 1965. Still, it may \nnot fully capture the diversity of cultural and policy \nenvironments globally or later generational changes. \nAdditionally, relying on self-reported mental well-\nbeing data introduces the potential for subjective \nbiases, which may differ from clinical evaluations \n(Spitzer and Weber 2019). The cross-sectional \ndesign also limits the ability to draw causal connec-\ntions between work\u2013family dynamics and mental \nhealth. Moreover, although it examines the impact \nof defamilization policies on women\u2019s mental \nhealth, significant policies or societal factors may be \noverlooked. Future research should address these \nissues with longitudinal studies that cover a more \ncomprehensive geographic and temporal range.\nFuture research should explore how work\u2013family \ntrajectories influence mental health, focusing on \nstressors and protective factors (Carvalho et al. 2018; \nZhou et al. 2018). Broadening the scope to include \ndiverse populations and generations is vital for \nunderstanding the evolving interplay between work, \nfamily, and mental health. To understand the cultural \ncontexts and policy environments that affect these \nrelationships, conducting cross-cultural and cross-\nnational studies across diverse countries and cohorts \nis crucial. Such research is essential for examining \nhow the welfare state and social policies, including \nlabor market and health care regulations, influence \nwomen\u2019s mental health outcomes and interact with \nwork\u2013family dynamics (Bambra 2007).\nResearch should also consider cultural settings \nwith varying gender equality and family life views \nto dissect the complex interplay of life course, insti-\ntutional factors, and cultural influences on health \ndisparities (Perlman and Gleason 2007).\nThis study, focusing on European women, pro-\nvides systematic insights into how work\u2013family \ndynamics and defamilization policies relate to \ndepressive symptoms through a life course perspec-\ntive. It underscores the need for informed actions by \npolicymakers and researchers to facilitate work\u2013\nfamily integration, enhance mental health, and strive \nfor genuine gender equity in Europe (Moen 2003). \nThis endeavor aims to transform gender equity from \nan ideal into a tangible reality, motivating future \nresearch to further unravel these complex dynamics.\nAcknowledgments\nI am grateful for comments and suggestions from Linda \nWaite, Jason Beckfield, Jenny Trinitapoli, Ignacio Cabib, \nUPenn\u2019s PSC Colloquium assistants, and anonymous \nreviewers.\nFunding\nThe author disclosed receipt of the following financial \nsupport for the research, authorship, and/or publication of \nthis article: This work received support from the Institute \nof Humane Studies under Grant Nos. IHS017599 and \nIHS017396.\nORCID iD\nAriel Azar \n https://orcid.org/0000-0002-4868-8081\n\n484\t\nJournal of Health and Social Behavior 65(4) \nSUPPLEMENTAL MATERIAL\nAppendices A through C are available in the online ver-\nsion of the article.\nReferences\nAbbott, Andrew, and Stanley DeViney. 1992. \u201cThe \nWelfare State as Transnational Event: Evidence \nfrom Sequences of Policy Adoption.\u201d Social Science \nHistory 16(2):245\u201374.\nAhrens, Christina J. Chrouser, and Carol D. Ryff. 2006. \n\u201cMultiple Roles and Well-Being: Sociodemographic \nand \nPsychological \nModerators.\u201d \nSex \nRoles \n55(11):801\u201315.\nAllen, Tammy D., David E. L. Herst, Carly S. Bruck, and \nMartha Sutton. 2000. \u201cConsequences Associated with \nWork-to-Family Conflict: A Review and Agenda for \nFuture Research.\u201d Journal of Occupational Health \nPsychology 5(2):278\u2013308.\nAmato, Paul R. 2000. \u201cThe Consequences of Divorce \nfor Adults and Children.\u201d Journal of Marriage and \nFamily 62(4):1269\u201387.\nAnand, Paul, Lucio Esposito, and Adri\u00e1n Villase\u00f1or. \n2018. \u201cDepression and Economic Status: Evidence \nfor Nonlinear Patterns in Women from Mexico.\u201d \nJournal of Mental Health 27(6):529\u201351.\nAngelini, Viola, Daniel D. H. Howdon, and Jochen O. \nMierau. 2019. \u201cChildhood Socioeconomic Status \nand Late-Adulthood Mental Health: Results from the \nSurvey on Health, Ageing and Retirement in Europe.\u201d \nThe Journals of Gerontology, Series B: Psychological \nSciences and Social Sciences 74(1):95\u2013104.\nBambra, Clare. 2007. \u201cDefamilisation and Welfare State \nRegimes: A Cluster Analysis: Defamilisation and \nWelfare State Regimes.\u201d International Journal of \nSocial Welfare 16(4):326\u201338.\nBambra, Clare, and Terje A. Eikemo. 2008. \u201cWelfare \nState Regimes, Unemployment, and Health: A \nComparative Study of the Relationship between \nUnemployment and Self-Reported Health in 23 \nEuropean Countries.\u201d Journal of Epidemiology & \nCommunity Health 63(2):92\u201398.\nBambra, Clare, Katherine E. Smith, and Jamie Pearce. \n2019. \u201cScaling up: The Politics of Health and Place.\u201d \nSocial Science & Medicine 232:36\u201342.\nBambra, Clare, D. Pope, V. Swami, D. Stanistreet, A. \nRoskam, A. Kunst, and A. Scott-Samuel. 2009. \n\u201cGender, Health Inequalities, and Welfare State \nRegimes: A Cross-National Study of 13 European \nCountries.\u201d Journal of Epidemiology & Community \nHealth 63(1):38\u201344.\nBarnett, Rosalind Chait, and Janet Shibley Hyde. 2001. \n\u201cWomen, Men, Work, and Family: An Expansionist \nTheory.\u201d American Psychologist 56(10):781\u201396.\nBeckfield, Jason, Clare Bambra, Terje A. Eikemo, Tim \nHuijts, Courtney McNamara, and Claus Wendt. \n2015. \u201cAn Institutional Theory of Welfare State \nEffects on the Distribution of Population Health.\u201d \nSocial Theory & Health 13(3):227\u201344.\nBrown, Susan L., Jennifer Roebuck Bulanda, and Gary \nR. Lee. 2005. \u201cThe Significance of Nonmarital \nCohabitation: Marital Status and Mental Health \nBenefits among Middle-Aged and Older Adults.\u201d The \nJournals of Gerontology, Series B: Psychological \nSciences and Social Sciences 60(1):S21\u201329.\nBr\u00fcckner, Hannah, and Karl Ulrich Mayer. 2005. \n\u201cDe-standardization of the Life Course: What It \nMight Mean? And If It Means Anything, Whether \nIt Actually Took Place?\u201dAdvances in Life Course \nResearch 9:27\u201353.\nBrydsten, Anna, Anne Hammarstr\u00f6m, and Miguel San \nSebastian. 2018. \u201cHealth Inequalities between \nEmployed and Unemployed in Northern Sweden: A \nDecomposition Analysis of Social Determinants for \nMental Health.\u201dInternational Journal for Equity in \nHealth 17(1):59.doi:10.1186/s12939-018-0773-5.\nCaballero, Sandra Toribio, Violeta Cardenal Hern\u00e1ez, \nAlejandro \u00c1vila Espada, and Mar\u00eda Mercedes \nOvejero Bruna. 2022. \u201cGender Roles and Women\u2019s \nMental Health: Their Influence on the Demand for \nPsychological Care.\u201dAnales de Psicolog\u00eda/Annals of \nPsychology 38(1):7\u201316.\nCalvo, Esteban, Ignacio Madero-Cabib, and Ursula M. \nStaudinger. 2018. \u201cRetirement Sequences of Older \nAmericans: Moderately Destandardized and Highly \nStratified across Gender, Class, and Race.\u201d The \nGerontologist 58(6):1166\u201376.\nCarvalho, V\u00e2nia, Maria Chambel, Mariana Neto, and \nS\u00edlvia Lopes. 2018. \u201cDoes Work\u2013Family Conflict \nMediate the Associations of Job Characteristics \nwith Employees\u2019 Mental Health among Men \nand Women?\u201d Frontiers in Psychology 9:966.\ndoi:10.3389/fpsyg.2018.00966.\nCase, Anne, and Christina Paxson. 2011. \u201cThe Long \nReach of Childhood Health and Circumstance: \nEvidence from the Whitehall II Study.\u201d The \nEconomic Journal 121(554):F183\u2013204.\nChandola, Tarani, Jane Ferrie, Amanda Sacker, and \nMichael Marmot. 2007. \u201cSocial Inequalities in Self-\nReported Health in Early Old Age: Follow-up of \nProspective Cohort Study.\u201d BMJ 334(7601):990. \ndoi:10.1136/bmj.39167.439792.55.\nChau, Ruby C. M., Liam Foster, Sam W. K. Yu, and \nYuk Pun Yu. 2017. \u201cDefamilization/Familization \nMeasures and Women\u2019s Pension Income\u2014The Case \nof Taiwan.\u201d Asian Social Work and Policy Review \n11(2):116\u201323.\nChen, Ming, Huiyun Zhu, Yiqi Du, and Geliang Yang. \n2018. \u201cHow Does the Social Environment during Life \nCourse Embody in and Influence the Development \nof Cancer?\u201d International Journal of Public Health \n63(7):811\u201321.\nChung, Haejoo, Edwin Ng, Selahadin Ibrahim, Bj\u00f6rn \nKarlsson, Joan Benach, Albert Espelt, and Carles \nMuntaner. 2013. \u201cWelfare State Regimes, Gender, \nand Depression: A Multilevel Analysis of Middle \nand High Income Countries.\u201d International Journal \nof Environmental Research and Public Health \n10(4):1324\u201341.\n\nAzar\t\n485\nCoduti, Wendy A., Kristin Tugman, Susanne M. Bruyere, \nand Valerie Malzer. 2015. \u201cAging Workers: Work \nEnvironment as a Factor in Employee Mental \nHealth.\u201d International Journal of Disability Manage\u00ad\nment 10:e4. doi:10.1017/idm.2015.4.\nCohen, Jessica, and Wendy Manning. 2010. \u201cThe \nRelationship \nContext \nof \nPremarital \nSerial \nCohabitation.\u201d Social Science Research 39(5):766\u201376.\nCohen, Sheldon, and Thomas A. Wills. 1985. \u201cStress, \nSocial Support, and the Buffering Hypothesis.\u201d \nPsychological Bulletin 98(2):310\u201357.\nColeman, Marilyn, Lawrence Ganong, and Mark Fine. \n2000. \u201cReinvestigating Remarriage: Another Decade \nof Progress.\u201d Journal of Marriage and Family \n62(4):1288\u2013307.\nCorna, Laurie M. 2013. \u201cA Life Course Perspective on \nSocioeconomic Inequalities in Health: A Critical \nReview of Conceptual Frameworks.\u201d Advances in \nLife Course Research 18(2):150\u201359.\nCornwell, Benjamin, Edward Laumann, and L. Philip \nSchumm.2008. \u201cThe Social Connectedness of Older \nAdults: A National Profile.\u201d American Sociological \nReview 73(2):185\u2013203.\nCunningham, Solveig A., Shivani A. Patel, Gloria L. Beckles, \nLinda S. Geiss, Neil Mehta, Hui Xie, and Giuseppina \nImperatore. 2018. \u201cCounty-Level Contextual Factors \nAssociated with Diabetes Incidence in the United \nStates.\u201d Annals of Epidemiology 28(1):20\u201325.\nDannefer, \nDale. \n2003. \n\u201cCumulative \nAdvantage/\nDisadvantage and the Life Course: Cross-Fertilizing \nAge and Social Science Theory.\u201d The Journals of \nGerontology, Series B: Psychological Sciences and \nSocial Sciences 58(6):S327\u201337.\nDiaz-Toro, Felipe, Ignacio Madero-Cabib, Esteban \nCalvo, and Ursula Staudinger. 2018. \u201cHeartbreaking \nCareers in Old Age: Retirement Sequences as a \nNontraditional Risk Factor for Cardiovascular \nDiseases.\u201d Circulation 137(Suppl. 1):A020.doi:10 \n.1161/circ.137.suppl_1.020.\nDinh, Huong, Amanda R. Cooklin, Liana S. Leach, \nElizabeth M. Westrupp, Jan M. Nicholson, and \nLyndall Strazdins. 2017. \u201cParents\u2019 Transitions into \nand out of Work\u2013Family Conflict and Children\u2019s \nMental Health: Longitudinal Influence via Family \nFunctioning.\u201d Social Science & Medicine 194:42\u201350.\nDiPrete, Thomas A., and Gregory M. Eirich. 2006. \n\u201cCumulative Advantage as a Mechanism for \nInequality: A Review of Theoretical and Empirical \nDevelopments.\u201d Annual Review of Sociology \n32:271\u201397.\nDupre, Matthew E., Audrey N. Beck, and Sarah \nO. Meadows. 2009. \u201cMarital Trajectories and \nMortality among U.S. Adults.\u201d American Journal of \nEpidemiology 170(5):546\u201355.\nDupre, Matthew E., and Alicia Nelson. 2016. \u201cMarital \nHistory and Survival after a Heart Attack.\u201d Social \nScience & Medicine 170:114\u201323.\nEngels, Miriam, Simone Weyers, Susanne Moebus, Karl-\nHeinz J\u00f6ckel, Raimund Erbel, Beate Pesch, Thomas \nBehrens, Nico Dragano, and Morten Wahrendorf. \n2019. \u201cGendered Work\u2013Family Trajectories and \nDepression at Older Age.\u201d Aging & Mental Health \n23(11):1478\u201386.\nEspelt, Albert, Carme Borrell, Maica Rodriguez-\nSanz, Carles Muntaner, M. Isabel Pasarin, Joan \nBenach, Maartje Schaap, Anton E. Kunst, and \nVicente Navarro. 2008. \u201cInequalities in Health by \nSocial Class Dimensions in European Countries of \nDifferent Political Traditions.\u201d International Journal \nof Epidemiology 37(5):1095\u2013105.\nEsping-Andersen, G\u00f8sta. 1990. The Three Worlds of \nWelfare Capitalism. Cambridge, UK: Polity Press.\nEsping-Andersen, G\u00f8sta. 1997. \u201cHybrid or Unique?: \nThe Japanese Welfare State between Europe and \nAmerica.\u201d Journal of European Social Policy \n7(3):179\u201389.\nGaydosh, Lauren, and Kathleen Mullan Harris. 2018. \n\u201cChildhood Family Instability and Young Adult \nHealth.\u201d Journal of Health and Social Behavior \n59(3):371\u201390.\nGove, Walter R. 1984. \u201cGender Differences in Mental \nand Physical Illness: The Effects of Fixed Roles \nand Nurturant Roles.\u201d Social Science & Medicine \n19(2):77\u201384.\nGreenhaus, Jeffrey H., and Nicholas J. Beutell. 1985. \n\u201cSources of Conflict between Work and Family \nRoles.\u201d The Academy of Management Review \n10(1):76\u201388.\nGrundy, Emily, and Gemma Holt. 2000. \u201cAdult Life \nExperiences and Health in Early Old Age in Great \nBritain.\u201d Social Science & Medicine 51(7):1061\u201374.\nGrundy, Emily, and Andy Sloggett. 2003. \u201cHealth \nInequalities in the Older Population: The Role of \nPersonal Capital, Social Resources, and Socio-\nEconomic Circumstances.\u201d Social Science & \nMedicine 56(5):935\u201347.\nGrundy, Emily M. D., Sanna Read, and Heini V\u00e4is\u00e4nen. \n2020. \n\u201cFertility \nTrajectories \nand \nLater-Life \nDepression among Parents in England.\u201d Population \nStudies 74(2):219\u201340.\nGrundy, Emily M.D., and Cecilia Tomassini. 2010. \n\u201cMarital History, Health, and Mortality among Older \nMen and Women in England and Wales.\u201d BMC Public \nHealth 10:554.doi:10.1186/1471-2458-10-554.\nHaas, Steven A. 2008. \u201cTrajectories of Functional \nHealth: The \u2018Long Arm\u2019 of Childhood Health and \nSocioeconomic Factors.\u201d Social Science & Medicine \n66(4):849\u201361.\nHaas, Steven A., M. Maria Glymour, and Lisa F. \nBerkman. 2011. \u201cChildhood Health and Labor \nMarket Inequality over the Life Course.\u201d Journal of \nHealth and Social Behavior 52(3):298\u2013313.\nHeuveline, Patrick, and Jeffrey M. Timberlake. 2004. \n\u201cThe Role of Cohabitation in Family Formation: The \nUnited States in Comparative Perspective.\u201d Journal \nof Marriage and Family 66(5):1214\u201330.\nHeuveline, Patrick, Jeffrey M. Timberlake, and Frank \nF. Furstenberg, Jr. 2003. \u201cShifting Childrearing to \n\n486\t\nJournal of Health and Social Behavior 65(4) \nSingle Mothers: Results from 17 Western Countries.\u201d \nPopulation and Development Review 29(1):47\u201371.\nHughes, Mary Elizabeth, and Linda J. Waite. 2009. \n\u201cMarital Biography and Health at Mid-life.\u201d Journal \nof Health and Social Behavior 50(3):344\u201358.\nInglehart, Ronald. 1997. \u201cModernization, Postmod\u00ad\nernization, and Changing Perceptions of Risk.\u201d \nInternational Review of Sociology 7(3):449\u201359.\nJahoda, Marie. 1982. Employment and Unemployment: \nA Social-Psychological Analysis. New York, NY: \nCambridge University Press.\nKravdal, \u00d8ystein, Emily Grundy, Torkild H. Lyngstad, \nand Kenneth Aa Wiik. 2012. \u201cFamily Life History \nand Late Mid-life Mortality in Norway.\u201d Population \nand Development Review 38(2):237\u201357.\nLacey, Rebecca, Mai Stafford, Amanda Sacker, Anne \nMcMunn, and Ariel Azar. 2016. \u201cWork\u2013Family \nLife Courses and Subjective Well-Being in the \nMRC National Survey of Health and Development \n(the 1946 British Birth Cohort Study).\u201d Journal of \nPopulation Ageing 9(1):69\u201389.\nLahelma, Eero, Sara Arber, Katariina Kivel\u00e4, and Eva \nRoos. 2002. \u201cMultiple Roles and Health among \nBritish and Finnish Women: The Influence of \nSocioeconomic Circumstances.\u201d Social Science & \nMedicine 54(5):727\u201340.\nLaMontagne, Anthony Daniel, Lay San Too, Laura \nPunnett, and Allison Joy Milner. 2021. \u201cChanges in \nJob Security and Mental Health: An Analysis of 14 \nAnnual Waves of an Australian Working-Population \nPanel Survey.\u201d American Journal of Epidemiology \n190(2):207\u201315.\nLand\u00f6s, Aljoscha, Martina Arx, Boris Cheval, Stefan Sieber, \nMatthias Kliegel, Rainer Gabriel, Dan Orsholits, \nBernadette W. A. Linden, David Blane, and Matthieu \nP. Boisgontier. 2018. \u201cChildhood Socioeconomic \nCircumstances and Disability Trajectories in Older \nMen and Women: A European Cohort Study.\u201d \nEuropean Journal of Public Health 29(1):50\u201358.\nLantz, Paula M., James S. House, Richard P. Mero, and \nDavid R. Williams. 2005. \u201cStress, Life Events, and \nSocioeconomic Disparities in Health: Results from \nthe Americans\u2019 Changing Lives Study.\u201d Journal of \nHealth and Social Behavior 46(3):274\u201388.\nLawton, M.Powell, Miriam Moss, and Morton H. Kleban. \n1984. \u201cMarital Status, Living Arrangements, and the \nWell-Being of Older People.\u201d Research on Aging \n6(3):323\u201345.\nLesthaeghe, Ron. 1995. \u201cThe Second Demographic \nTransition in Western Countries: An Interpretation.\u201d \nPp. 17\u201362 in Gender and Family Change in \nIndustrialized Countries, edited by K. Oppenheim \nand A. Jensen. Oxford, UK: Clarendon Press.\nLeupp, Katrina. 2017. \u201cDepression, Work and Family \nRoles, and the Gendered Life Course.\u201d Journal of \nHealth and Social Behavior 58(4):422\u201341.\nLiefbroer, Aart C. 1999. \u201cFrom Youth to Adulthood: \nUnderstanding \nChanging \nPatterns \nof \nFamily \nFormation from a Life Course Perspective.\u201d Pp. 53\u2013\n85 in Population Issues: An Interdisciplinary Focus, \nThe Plenum Series on Demographic Methods and \nPopulation Analysis, edited by L. J. G. van Wissen \nand P. A. Dykstra. Dordrecht, Netherlands: Springer \nNetherlands.\nLiu, Hui. 2012. \u201cMarital Dissolution and Self-Rated \nHealth: Age Trajectories and Birth Cohort Variations.\u201d \nSocial Science & Medicine 74(7):1107\u201316.\nLlena-Nozal, Ana, Maarten Lindeboom, and France \nPortrait. 2004. \u201cThe Effect of Work on Mental \nHealth: Does Occupation Matter?\u201d Health Economics \n13(10):1045\u201362.\nLohmann, Henning, and Hannah Zagel. 2016. \u201cFamily \nPolicy in Comparative Perspective: The Concepts and \nMeasurement of Familization and Defamilization.\u201d \nJournal of European Social Policy 26(1):48\u201365.\nLucumi, Diego I., Amy J. Schulz, Ana V. Diez Roux, \nand \nAndrew \nGrogan-Kaylor. \n2017. \n\u201cIncome \nInequality and High Blood Pressure in Colombia: A \nMultilevel Analysis.\u201d Cadernos de Sa\u00fade P\u00fablica \n33(11):e00172316. \ndoi:10.1590/0102-311X00172 \n316.\nLuo, Ye, and Linda J. Waite. 2005. \u201cThe Impact of \nChildhood and Adult SES on Physical, Mental, and \nCognitive Well-Being in Later Life.\u201d The Journals of \nGerontology, Series B: Psychological Sciences and \nSocial Sciences 60(2):S93\u2013101.\nMacIndoe, Heather, and Andrew Abbott. 2011. \u201cSequence \nAnalysis and Optimal Matching Techniques for \nSocial Science Data.\u201d Pp. 387\u2013406 in Handbook of \nData Analysis, edited by M. Hardy and A. Bryman. \nLondon: Sage.\nMackenbach, Johan P. 2012. \u201cThe Persistence of \nHealth Inequalities in Modern Welfare States: \nThe Explanation of a Paradox.\u201d Social Science & \nMedicine 75(4):761\u201369.\nMadero-Cabib, Ignacio, Ariel Azar, and Pedro P\u00e9rez-\nCruz. 2019. \u201cAdvantages and Disadvantages across \nthe Life Course and Health Status in Old Age among \nWomen in Chile.\u201d International Journal of Public \nHealth 64(8):1203\u201314.\nMadero-Cabib, Ignacio, Laurie Corna, and Isabel \nBaumann. 2019. \u201cAging in Different Welfare \nContexts: A Comparative Perspective on Later-\nLife Employment and Health.\u201d The Journals of \nGerontology, Series B: Psychological Sciences and \nSocial Sciences 75(7):1515\u201326.\nMadero-Cabib, Ignacio, Rosario Undurraga, and Cristi\u00e1n \nValenzuela. 2019. \u201cHow Have Women\u2019s Employ\u00ad\nment Patterns during Young Adulthood Changed \nin Chile? A Cohort Study.\u201d Longitudinal and Life \nCourse Studies 10(3):375\u201397.\nMayer, Karl Ulrich. 1991. \u201cSoziale Ungleichheit Und \nDie Differenzierung von Lebensverl\u00e4ufen [Social \nInequality and the Differentiation of Life Courses].\u201d \nPp. 667\u201387 in Die Modernisierung moderner \nGesellschaften. \nVerhandlungen \ndes \nDeutschen \nSoziologentages in Frankfurt am Main 1990, edited \nby W. Zapf. Frankfurt: Camput.\n\nAzar\t\n487\nMcDonald, Andr\u00e9 J., Hayley A. Hamilton, Tara Elton-\nMarshall, Yeshambel T. Nigatu, Damian Jankowicz, \nSusan J. Bondy, Samantha Wells, and Christine \nM. \nWickens. \n2022. \n\u201cHousehold \nComposition \nand Anxiety Symptoms during the COVID-19 \nPandemic: A Population-Based Study.\u201d PLoS One \n17(11):e0277243. doi:10.1371/journal.pone.0277243.\nMcFarland, Michael J., Mark D. Hayward, and Dustin \nBrown. 2013. \u201cI\u2019ve Got You under My Skin: \nMarital Biography and Biological Risk.\u201d Journal of \nMarriage and Family 75(2):363\u201380.\nMertens, Antje. 1998. Labor Mobility and Wage \nDynamics: An Empirical Study for Germany in \nComparison with the United States. Aachen, \nGermany: Shaker Verlag.\nMerton, Robert K. 1968. \u201cThe Matthew Effect in \nScience.\u201d Science 159(3810):56\u201363.\nMoen, Phyllis, ed. 2003. It\u2019s about Time: Couples and \nCareers. Ithaca, NY: Cornell University Press.\nMoen, Phyllis. 2022. \u201cThe Uneven Stress of Social \nChange: Disruptions, Disparities, and Mental \nHealth.\u201d Society and Mental Health 12(2):85\u201398.\nMoen, Phyllis, Erin L. Kelly, Wen Fan, Shi-Rong Lee, \nDavid Almeida, Ellen Ernst Kossek, and Orfeu \nM. Buxton. 2016. \u201cDoes a Flexibility/Support \nOrganizational \nInitiative \nImprove \nHigh-Tech \nEmployees\u2019 Well-Being? Evidence from the Work, \nFamily, and Health Network.\u201d American Sociological \nReview 81(1):134\u201364.\nMoen, Phyllis, and Patricia Roehling. 2005. The Career \nMystique: Cracks in the American Dream. Oxford, \nUK: Rowman & Littlefield.\nMontez, Jennifer Karas, Pekka Martikainen, Hanna \nRemes, and Mauricio Avendano. 2015. \u201cWork\u2013\nFamily Context and the Longevity Disadvantage of \nU.S. Women.\u201d Social Forces 93(4):1567\u201397.\nMuller, Ziva, and Howard Litwin. 2011. \u201cGrandparenting \nand Psychological Well-Being: How Important Is \nGrandparent Role Centrality?\u201d European Journal of \nAgeing 8(2):109\u201318.\nPampel, Fred. 2011. \u201cCohort Change, Diffusion, and \nSupport for Gender Egalitarianism in Cross-National \nPerspective.\u201d Demographic Research 25(21):667\u201394.\nParkinson, Jane, Jon Minton, James Lewsey, Bouttell \nBouttell, and Gerry McCartney. 2018. \u201cDrug-\nRelated Deaths in Scotland 1979\u20132013: Evidence \nof a Vulnerable Cohort of Young Men Living in \nDeprived Areas.\u201d BMC Public Health 18:357.\ndoi:10.1186/s12889-018-5267-2.\nPaul, Karsten I., and Klaus Moser. 2009. \u201cUnemployment \nImpairs Mental Health: Meta-analyses.\u201d Journal of \nVocational Behavior 74(3):264\u201382.\nPerlman, Bruce J., and Gregory Gleason. 2007. \u201cCultural \nDeterminism versus Administrative Logic: Asian \nValues and Administrative Reform in Kazakhstan \nand Uzbekistan.\u201d International Journal of Public \nAdministration 30(12\u201314):1327\u201342.\nRead, Sanna, Emily Grundy, and Else Foverskov. 2016. \n\u201cSocio-economic Position and Subjective Health \nand Well-Being among Older People in Europe: A \nSystematic Narrative Review.\u201d Aging & Mental \nHealth 20(5):529\u201342.\nRead, Sanna, Emily Grundy, and Douglas A. Wolf. 2011. \n\u201cFertility History, Health, and Health Changes in \nLater Life: A Panel Study of British Women and Men \nBorn 1923\u201349.\u201d Population Studies 65(2):201\u201315.\nRepetti, Rena L., Shelley E. Taylor, and Teresa E. \nSeeman. 2002. \u201cRisky Families: Family Social \nEnvironments and the Mental and Physical Health of \nOffspring.\u201d Psychological Bulletin 128(2):330\u201366.\nRoss, Catherine E., and John Mirowsky. 2002. \u201cFamily \nRelationships, Social Support, and Subjective Life \nExpectancy.\u201d Journal of Health and Social Behavior \n43(4):469\u201389.\nSacker, Amanda, Diana Worts, and Peggy McDonough. \n2011. \u201cSocial Influences on Trajectories of Self-\nRated Health: Evidence from Britain, Germany, \nDenmark, and the U.S.A.\u201d Journal of Epidemiology \n& Community Health 65(2):130\u201336.\nSaxonberg, Steven. 2013. \u201cFrom Defamilialization to \nDegenderization: Toward a New Welfare Typology.\u201d \nSocial Policy & Administration 47(1):26\u201349.\nScherger, Simone, James Nazroo, and Vanessa May. \n2016. \u201cWork and Family Trajectories: Changes \nacross Cohorts Born in the First Half of the 20th \nCentury.\u201d Journal of Population Ageing 9:131\u201355.\nSchieman, Scott, and Paul Glavin. 2011. \u201cEducation and \nWork\u2013Family Conflict: Explanations, Contingencies, \nand Mental Health Consequences.\u201d Social Forces \n89(4):1341\u201362.\nSchoen, Robert, and Vladimir Canudas-Romo. 2006. \n\u201cTiming Effects on Divorce: 20th Century Experience \nin the United States.\u201d Journal of Marriage and \nFamily 68(3):749\u201358.\nSchr\u00f6der, Mathis. 2011. \u201cRetrospective Data Collection \nin the Survey of Health, Aging, and Retirement in \nEurope. SHARELIFE Methodology.\u201d Working Paper \nNo. 172-2008. Mannheim, Germany: Mannheim \nResearch Institute for the Economics of Aging.\nSimon, Robin W., and Anne E. Barrett. 2010. \u201cNonmarital \nRomantic Relationships and Mental Health in Early \nAdulthood: Does the Association Differ for Women \nand Men?\u201d Journal of Health and Social Behavior \n51(2):168\u201382.\nSingh, Gopal K., Gem P. Daus, Michelle Allender, \nChristine T. Ramey, Elijah K. Martin, Chrisp Perry, \nAndrew A. De Los Reyes, and Ivy P. Vedamuthu. \n2017. \u201cSocial Determinants of Health in the United \nStates: Addressing Major Health Inequality Trends \nfor the Nation, 1935\u20132016.\u201d International Journal of \nMCH and AIDS 6(2):139\u201364.\nSpitzer, Sonja, and Daniela Weber. 2019. \u201cReporting \nBiases in Self-Assessed Physical and Cognitive \nHealth Status of Older Europeans.\u201d PLoS One \n14(10):e0223526. doi:10.1371/journal.pone.0223526.\nStuder, Matthias. 2013. \u201cWeightedCluster Library \nManual: A Practical Guide to Creating Typologies of \nTrajectories in the Social Sciences with R.\u201d Working \n\n488\t\nJournal of Health and Social Behavior 65(4) \nPaper No. 24. Lausanne, Switzerland: LIVES, \nUniversity of Lausanne.\nSverke, Magnus, Johnny Hellgren, and Katharina N\u00e4swall. \n2002. \u201cNo Security: A Meta-analysis and Review of \nJob Insecurity and Its Consequences.\u201d Journal of \nOccupational Health Psychology 7(3):242\u201364.\nTani, Yukako, Takeo Fujiwara, Naoki Kondo, Hisashi \nNoma, Yuri Sasaki, and Katsunori Kondo. 2016. \n\u201cChildhood Socioeconomic Status and Onset of \nDepression among Japanese Older Adults: The \nJAGES Prospective Cohort Study.\u201d The American \nJournal of Geriatric Psychiatry 24(9):717\u201326.\nTaylor, R. C., and E. G. Ford. 1983. \u201cThe Elderly at Risk: \nA Critical Examination of Commonly Identified \nRisk Groups.\u201d The Journal of the Royal College of \nGeneral Practitioners 33(256):699\u2013705.\nTosi, Marco, and Emily Grundy. 2019. \u201cWork\u2013Family \nLifecourses and Later-Life Health in the United \nKingdom.\u201d Ageing and Society 41(6):1\u201327.\nUccheddu, Damiano, Anne H. Gauthier, Nardi Steverink, \nand Tom Emery. 2019. \u201cGender and Socioeconomic \nInequalities in Health at Older Ages across Different \nEuropean Welfare Clusters: Evidence from SHARE \nData, 2004\u20132015.\u201d European Sociological Review \n35(3):346\u201362.\nUmberson, Debra, Mieke Beth Thomeer, and Kristi \nWilliams. 2013. \u201cFamily Status and Mental Health: \nRecent Advances and Future Directions.\u201d Pp. 405\u2013\n31 in Handbook of the Sociology of Mental Health, \nHandbooks of Sociology and Social Research, edited \nby C. S. Aneshensel, J. C. Phelan, and A. Bierman. \nDordrecht, Netherlands: Springer Netherlands.\nVan de Kaa, Dirk J. 1987. \u201cEurope\u2019s Second Demographic \nTransition.\u201d Population Bulletin 42(1):1\u201359.\nvan der Noordt, Maaike, Helma IJzelenberg, Mari\u00ebl \nDroomers, and Karin I. Proper. 2014. \u201cHealth Effects \nof Employment: A Systematic Review of Prospective \nStudies.\u201d Occupational and Environmental Medicine \n71(10):730\u201336.\nWahrendorf, Morten, David Blane, Mel Bartley, Nico \nDragano, and Johannes Siegrist. 2013. \u201cWorking \nConditions in Mid-life and Mental Health in Older \nAges.\u201d Advances in Life Course Research 18(1): \n16\u201325.\nWilliams, Kristi, and Debra Umberson. 2004. \u201cMarital \nStatus, Marital Transitions, and Health: A Gendered \nLife Course Perspective.\u201d Journal of Health and \nSocial Behavior 45(1):81\u201398.\nWitte, Hans De. 1999. \u201cJob Insecurity and Psychological \nWell-Being: Review of the Literature and Exploration \nof Some Unresolved Issues.\u201d European Journal of \nWork and Organizational Psychology 8(2):155\u201377.\nZagel, Hannah, and Zachary Van Winkle. 2020. \n\u201cWomen\u2019s Family and Employment Life Courses \nacross Twentieth-Century Europe: The Role of \nPolicies and Norms.\u201d Social Politics: International \nStudies in Gender, State & Society 29(2):446\u201376.\nZella, Sara, and Sarah Harper. 2020. \u201cThe Impact of \nLife Course Employment and Domestic Duties on \nthe Well-Being of Retired Women and the Social \nProtection Systems That Frame This.\u201d Journal of \nAging and Health 32(5):285\u201395.\nZhou, Shiyi, Shuyang Da, Heng Guo, and Xichao \nZhang. 2018. \u201cWork\u2013Family Conflict and Mental \nHealth among Female Employees: A Sequential \nMediation Model via Negative Affect and Perceived \nStress.\u201d Frontiers in Psychology 9:544.doi:10.3389/\nfpsyg.2018.00544.\nAuthor Biography\nAriel Azar is an assistant professor of sociology at Purdue \nUniversity. His research investigates the interplay between \npolicy contexts, political systems, and broader social \nstructures, focusing on their influence on health disparities \nand trajectories throughout the life course, with special \nattention to migrant and LGBT populations. Azar also \nholds a visiting scholar position at the Population Research \nInstitute at Pennsylvania State University and is a faculty \nassociate at the Center on Aging and the Life Course at \nPurdue University.\n\n\n    Scientific Research Findings:", "answer": " \u2022 Women following traditional life paths (e.g., stable jobs combined with traditional family structures) reported better mental health outcomes compared to those with less conventional trajectories.\n\u2022 Delayed family formation combined with prolonged employment was associated with higher depressive symptoms, especially in contexts with high defamilization policies.\n\u2022 Defamilization policies moderated the impact of life trajectories on mental health, with greater policy support benefiting women in traditional roles but proving less effective for those following nontraditional life paths.", "id": 82}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    https://doi.org/10.1177/00221465211058153\nJournal of Health and Social Behavior\n2022, Vol. 63(1) 2\u00ad\u201318\n\u00a9 American Sociological Association 2021\nDOI: 10.1177/00221465211058153\njhsb.sagepub.com\nArticle\nHealth disparities between women and men in the \nUnited States are well documented. Women experi-\nence a higher incidence of chronic conditions, \n\u00adfunctional limitations, depression, and disability \ncompared to men (Case and Paxson 2005; Leveille \net\u00a0 al. 2000; Nolen-Hoeksema 2001), and rates of \nmaternal morbidity and mortality in the United \nStates have increased over the past several decades \n(Creanga et\u00a0al. 2014). Women also experience more \nbarriers to health care access (Ng et\u00a0al. 2010; Rustgi, \nDoty, and Collins 2009), which may exacerbate \ngender gaps in health. Women are more likely to be \nadversely affected by barriers to health care access, \nin part, because women have greater need for health \ncare services to manage chronic conditions and \nreproductive care (Case and Paxson 2005; Doyal \n2000; Owens 2008). Unmet health care need is \nassociated with higher rates of morbidity and mor-\ntality for women and higher infant mortality (Atrash \net\u00a0al. 2006; Kent, Patel, and Varela 2012).\nAvailability of care describes the volume and \nselection of existing services, including ability to \nget care when needed, consistency and timeliness \nof care, and choice of providers (Penchansky and \nThomas 1981). Women are more likely than men to \nreport unmet need for medical care (Long, Stockley, \nand Shulman 2011) and delays in receiving medical \ncare (Ng et\u00a0 al. 2010). Affordability barriers, \ndescribed as patient difficulty or inability to pay for \nhealth insurance, medications, medical tests, and \nother out-of-pocket health care costs (Penchansky \n1058153 HSBXXX10.1177/00221465211058153Journal of Health and Social BehaviorRapp et al.\nresearch-article2021\n1Roanoke College, Salem, VA, USA\n2North Carolina State University, Raleigh, NC, USA\n3George Mason University, Fairfax, VA, USA\nCorresponding Author:\nKristen Schorpp Rapp, Department of Sociology and \nPublic Health, Roanoke College, 221 College Lane, \nSalem, VA 24153, USA. \nEmail: schorpp@roanoke.edu\nState\u2013Level Sexism and \nGender Disparities in Health \nCare Access and Quality in the \nUnited States\nKristen Schorpp Rapp1\n, Vanessa V. Volpe2,  \nTabitha L. Hale3, and Dominique F. Quartararo1\nAbstract\nIn this investigation, we examined the associations between state-level structural sexism\u2014a multi\u00ad\ndimensional index of gender inequities across economic, political, and cultural domains of the gender \nsystem\u2014and health care access and quality among women and men in the United States. We linked \nadministrative data gauging state-level gender gaps in pay, employment, poverty, political representation, \nand policy protections to individual-level data on health care availability, affordability, and quality from the \nnational Consumer Survey of Health Care Access (2014\u20132019; N = 24,250). Results show that higher \nstate-level sexism is associated with greater inability to access needed health care and more barriers to \naffording care for women but not for men. Furthermore, contrary to our hypothesis, women residing in \nstates with higher state-level sexism report better quality of care than women in states with lower levels \nof sexism. These findings implicate state-level sexism in perpetuating gender disparities in health care.\nKeywords\ngender inequity, health care access, state policy, structural sexism, United States\n\nRapp et al.\t\n3\nand Thomas 1981), also disproportionately affect \nwomen. Women are more vulnerable to experienc-\ning gaps or inadequacies in health insurance cover-\nage and more often forgo medical care due to cost \n(Lavelle and Smock 2012; Rustgi et\u00a0al. 2009). One \nimportant aspect of quality of medical care is \npatient\u2013provider communication, characterized by \nattentive listening, answering patient questions, and \ninvolving patients in their own care (Jesus and Silva \n2016; Vowles and Thompson 2012). Some studies \nfind that women report higher satisfaction with pro-\nvider communication than men (e.g., Elliott et\u00a0al. \n2012), and others show no gender difference in \npatient satisfaction with provider communication \n(e.g., Hall and Roter 1995). The extent to which \nwomen face structural inequities may explain some \nof the inconsistencies in prior research.\nSexism, defined as gender inequity in power and \nresources that systematically privileges men and \ndisadvantages women (Ridgeway and Correll \n2004), is a significant determinant of gender gaps in \naccess to health care. The majority of past research \non sexism and women\u2019s health care implicates pro-\nvider gender bias and discrimination as important \nbarriers to quality care (Chapman, Kaatz, and \nCarnes 2013; Hamberg 2008; Marcum 2017). \nHowever, provider bias and discrimination do not \nfully explain gender gaps in the availability and \naffordability of health care, which may also be \nrooted in sociopolitical inequities that intersect with \nthe health care system, such as gender inequity in \npolitical representation, pay, and employment. In \naddition, the prevalence of provider bias and dis-\ncrimination within the U.S. health care system may \nbe influenced by sexism at the institutional level, \nsuch as the underrepresentation of women in policy-\nmaking and health care leadership. Literature exam-\nining racial disparities in health implicates structural \nracism as a determinant of poor quality of medical \ncare among patients of color (Feagin and Bennefield \n2014; Yearby 2018), but this rationale has not been \napplied to research exploring sexism as a determi-\nnant of women\u2019s health care access and quality.\nTo examine the role of sexism in U.S. health care \ninequities, the current investigation tested associa-\ntions between state-level structural sexism and dis-\nparities in health care access and quality among \nwomen and men. Homan (2019) defines structural \nsexism as gender inequity in power and resources \nacross institutional, interpersonal, and internalized \nlevels of the gender system that collectively shapes \ngender disparities in health. Research has largely \nexamined interpersonal and internalized levels of \nsexism rather than inequities institutionalized in the \nsystems of our society (e.g., employment, govern-\nmental, legal). Although research on the health \neffects of gender discrimination and other interper-\nsonal interactions (e.g., intimate partner violence) is \nimportant, a paucity of research at other levels may \nprevent us from understanding structural sexism. \nBecause theoretical developments in health dispari-\nties research articulate the role of institutional power \nstructures in shaping and perpetuating health inequi-\nties (Feagin and Bennefield 2014; Krieger 2001, \n2020), we examine state-level sexism as one macro-\nlevel dimension of the larger structural sexism \nsystem.\nBased on existing studies examining state-level \ngender inequities and health (e.g., Homan 2019; \nKawachi et\u00a0 al. 1999; Montez, Zajacova, and \nHayward 2016), we constructed a multidimensional \nindex of state-level sexism that measures gender \ninequities in wages and employment, policies that \ndiscriminate against women, underrepresentation of \nwomen in government, and restrictions on women\u2019s \nreproductive rights to investigate the role of institu-\ntional power structures in shaping access to quality \nhealth care among women and men. We examined \nassociations between state-level structural sexism \nand (1) inability to get needed health care, (2) barri-\ners in health care availability and affordability, and \n(3) quality of medical care received. We also tested \nwhether associations between state-level sexism and \nhealth care access and quality differ by gender. This \ninvestigation provides novel insights about the con-\nnections between sexism at the institutional level \nand individual health care experiences, which have \nimportant applications in policymaking and health \ncare reform.\nBackground\nSexism as a Social Determinant of \nHealth Care Access and Quality\nSexism is a significant determinant of women\u2019s \nhealth (Homan 2019; Molix 2014; Moss 2002). \nMost studies examining the impacts of sexism on \nhealth focus on the impacts of interpersonal dis-\ncrimination and sexual harassment (Harnois and \nBastos 2018; Krieger 2001; Molix 2014). Although \nthese investigations provide insight into the effects \nof sexism on health, studies have reframed sexism \nas a multilevel construct that is expressed and rein-\nforced within economic, political, and cultural insti-\ntutions (Homan 2019; Krieger 2020). For example, \nRisman (2004) conceptualizes gender as a social \nstructure, embedded in individual, interactional, and \n\n4\t\nJournal of Health and Social Behavior 63(1)\ninstitutional dimensions of society. In addition, \nKrieger\u2019s (2020) ecosocial theory of health exam-\nines how legacies of systematic oppression across \ninstitutions structure individual exposure to health \nrisks and limit availability of health-promoting \nresources for oppressed groups.\nResearchers examining the relationship between \nstructural sexism and health have proposed that \nhealth care access and quality are important path-\nways through which structural inequities shape indi-\nvidual health outcomes (e.g., Homan 2019; Wisdom, \nBerlin, and Lapidus 2005). Indeed, the efficacy of \nthe health care system in providing consistent and \naffordable medical care is highly contingent on eco-\nnomic and political contexts, such as the extent of \nincome inequality, social welfare spending, and \nhealth care legislation (Dickman, Himmelstein, and \nWoolhandler 2017). However, no studies to date \nhave examined the relationship between structural \nsexism and health care access and how this relation-\nship differs among women and men.\nWe argue that sexism across economic, labor \nforce, and political institutions acts as a fundamental \nbarrier to women\u2019s health care access through multi-\nple direct and indirect pathways. Labor markets that \nvalue the work of men over women systematically \nunderpay, underinsure, and underemploy women, \nincreasing the likelihood that women will be unable \nto afford medical care (Gijsbers Van Wijk, Van Vliet, \nand Kolk 1996). In addition, women\u2019s political rep-\nresentation influences health policies that shape \naccess to care. Women policymakers are more likely \nto pursue policies that invest in health care, social \nwelfare, and other determinants of public health,  \nand the extent of women\u2019s political representation \npredicts the implementation of these policies \n(Bolzendahl and Brooks 2007; Park 2017). Finally, \nmultiple studies document the impact of U.S. \n\u00adeconomic and health care legislation on barriers to \nwomen\u2019s health care access, including the lack of \npolicy mandating paid family leave (Gault et\u00a0 al. \n2014) and policies that restrict reproductive rights \n(Stevenson et\u00a0al. 2016).\nState-Level Sexism, Access to Health \nCare, and Women\u2019s Health\nIn the current study, we measure structural sexism at \nthe state level. State-level sexism encompasses the \ngendered distribution of power and resources within \neconomic, labor force, political, and cultural institu-\ntions among U.S. states. We posit that women\u2019s rel-\native lack of power within these institutions \nproduces barriers to health care access for women \nvia multiple direct and indirect mechanisms, includ-\ning cost barriers due to gender gaps in employment \nand wages, availability barriers due to women\u2019s \nunderrepresentation in political decision-making \nregarding health care, and overall access barriers \ndue to cultural norms that restrict women\u2019s repro-\nductive choice. We chose to examine sexism at the \nstate level because states vary widely in the extent \nof gender inequity (Hess et\u00a0al. 2015), demonstrating \nthe need to investigate how sexism at the state level \ncontributes to gender disparities in health care.\nState-level sexism is a robust predictor of wom-\nen\u2019s morbidity and mortality, implicating gender \ninequities across economic, political, and cultural \ninstitutions as pathways leading to poorer health for \nwomen. Recent research documents higher mortal-\nity rates for women and infants residing in U.S. \nstates with higher levels of economic inequity \n(Pabayo et\u00a0al. 2019), and studies using state-level \ncomposite scores of economic inequity (e.g., gen-\nder gaps in wages, employment, and poverty) find \nhigher mortality rates and poorer health for women \nresiding in more inequitable states (Homan 2019; \nMontez et\u00a0al. 2016). Within the political domain, \nunderrepresentation of women in state legislature is \nassociated with higher rates of female and infant \nmortality (Homan 2017; Kawachi et\u00a0al. 1999), and \nthe implementation of state policies designed to \npromote gender equity (e.g., paid family and mater-\nnal leave, gun ownership restrictions for domestic \nviolence offenders) predicts better health and lower \nmortality rates for women (Doran et\u00a0al. 2020; Lee \net\u00a0al. 2020; Wisdom et\u00a0al. 2005). Finally, state-level \npolicies regarding Medicaid eligibility and access \nto reproductive care have significant implications \nfor the health of women (Hawkins et\u00a0 al. 2020; \nJohnston et\u00a0al. 2018; Margerison et\u00a0al. 2020).\nHealth care access and quality could underlie the \nrobust associations between state-level sexism and \nwomen\u2019s health; however, existing literature has not \nadequately examined the relationship between state-\nlevel sexism and health care. Because the U.S. \nhealth care system relies heavily on private, \nemployer-based health insurance, systemic state-\nlevel gender gaps in wages and labor force partici-\npation may reduce both women\u2019s insurance \ncoverage and ability to afford health care. Indeed, \nwomen are more likely to depend on their spouses \nfor health insurance and spend a higher share of \ntheir income on health care than men (Patchias and \nWaxman 2007), supporting the notion that gender \ninequalities in pay and employment increase barri-\ners to care for women. In addition, the extent of \nwomen\u2019s political representation in U.S. state \n\nRapp et al.\t\n5\nlegislatures is associated with the proposal and \nimplementation of state policies investing in health \ncare and overall welfare spending (Berkman and \nO\u2019Connor 1993; Bratton and Haynie 1999). Such \npolicies have unique implications for women\u2019s \nhealth care because state-level policy changes in \nMedicaid eligibility and health care funding affect \nthe affordability of women\u2019s care and the availabil-\nity of screening for breast and cervical cancers \n(Daniel et\u00a0al. 2018; Johnston et\u00a0al. 2018).\nState social and economic policies may operate \nindirectly to shape health care or may signal broader \npolicymaking contexts that fail to address women\u2019s \nhealth needs. Although no studies have examined \nthe effect of state-mandated paid family and medical \nleave on women\u2019s health care access, the availability \nof paid family leave has been shown to increase \ncontinuity in women\u2019s labor force participation \n(Baum and Ruhm 2016), which has implications for \nwomen\u2019s wages and access to employment-based \nhealth insurance. State firearm laws may also be \nhealth-related policies that reflect the commitment \nof state legislatures to protect women\u2019s health and \nsafety. States with fewer laws prohibiting possession \nof firearms for domestic violence offenders have \nhigher rates of intimate partner homicide against \nwomen than states that have more of these laws \n(Sivaraman et\u00a0al. 2019). Although gun laws are not \nhealth care policies per se, the lack of legislation \nprotecting women from gun violence could signal \nthe failure of state policymaking contexts to priori-\ntize women\u2019s health and safety.\nFinally, state political contexts shape women\u2019s \neligibility for health care coverage and access to \nreproductive health care. Women residing in states \nthat expanded eligibility for Medicaid in response \nto the Affordable Care Act reported lower rates of \nuninsurance and better self-rated health compared \nto women in nonexpansion states (Johnston et\u00a0al. \n2018; Margerison et\u00a0al. 2020). Furthermore, imple-\nmentation of restrictive state-level abortion policies \nin several states has led to the closure of family \nplanning clinics that once provided a wide array of \nhealth care services to women, including primary \ncare, preventive screenings, and reproductive care \n(Lawrence and Ness 2017; Stevenson et\u00a0al. 2016).\nState-Level Sexism and Quality of \nPatient\u2013Provider Communication\nState-level sexism might also influence the quality \nof patient care. Patient\u2013provider communication is \nan important determinant of patient treatment out-\ncomes and disease management (Owens 2008; \nStreet et\u00a0al. 2009), but the ways in which institu-\ntional sexism shapes patient care have not been \nstudied. Existing literature that explicitly connects \ninstitutional sexism to the quality of interpersonal \ninteractions examines how organizational structures \nin the labor force relate to women\u2019s experiences of \ninterpersonal workplace discrimination (Bobbitt-\nZeher 2011; Ridgeway and Correll 2004). Aligned \nwith ecosocial perspectives of the gender system, \nthis body of research frames sexism as embedded in \ninstitutional structures in which systematic gender \nbias shapes women\u2019s experiences of mistreatment \nand discrimination (Ridgeway and Correll 2004). \nTherefore, we examine connections between state-\nlevel sexism and the quality of women\u2019s interac-\ntions with health providers, positing that women \nresiding in states with high levels of institutional \nsexism may also report poorer quality of communi-\ncation with health care providers.\nThe Present Study\nThis study examines the relationship between state-\nlevel structural sexism across economic, political, \nand cultural domains and gender gaps in health care \naccess and quality. We tested the associations \nbetween state-level sexism and multiple dimensions \nof health care access and quality, including inability \nto access health care, lack of health insurance, barri-\ners in the availability and affordability of care, and \nthe quality of patient\u2013provider communication \namong women and men who sought out health care \nwithin the past year. We also tested for gender dif-\nferences in each of these associations. It is important \nto note that race, education, income, and other \naspects of women\u2019s lived experiences play a role in \nhealth care. In the present study, we focus on eluci-\ndating differences as a function of gender, account-\ning for these other factors statistically. This study \ncontributes to existing literature by drawing novel \nconnections between macro-level institutional pro-\ncesses and individual health care experiences \namong women and men in the United States.\nData and Methods\nData\nState-level data were compiled from administrative \ndata sources (e.g., Bureau of Labor Statistics, \nCurrent Population Survey, Guttmacher Institute) to \ncapture gender inequity in economic standing, polit-\nical representation, policy protections, and repro-\nductive rights. Table 1 includes a full list of measures \nwith corresponding administrative data sources.\n\n6\t\nJournal of Health and Social Behavior 63(1)\nIndividual-level data came from the December \n2014 to January 2019 waves of the Association of \nAmerican Medical Colleges (AAMC) Consumer \nSurvey of Health Care Access, a repeat cross-sec-\ntional, online survey of adults age 18 and older in \nthe United States. Surveys were conducted by an \nexternal firm that maintains an active panel of \npotential study participants. Stratified sampling was \nused to collect data based on age and health insur-\nance status, with oversamples of various subpopu-\nlations of interest (minority, rural, Medicaid \nrecipients, etc.) in particular survey waves. U.S. \ncensus weights were available to account for non-\nprobability sampling procedures.\nOf the 25,267 eligible participants in the AAMC \nsurvey, 24,250 (96%) participants had complete \ndata for all variables in Stage 1 of the analysis \n(associations of state-level sexism with inability to \naccess care and uninsured). The latter stages of the \nanalysis were restricted to the 21,329 participants \nwho had at least one medical care visit in the past \nyear and had complete data for additional health \ncare access and quality variables. Supplemental \nanalysis revealed that participants who were \nexcluded from the analysis were more likely to be \nfemale, younger (ages 18\u201334), nonwhite, and \nunmarried. Excluded participants also had lower \nhousehold income on average, were less likely to \nhave a college education, and were more likely to \nlive in a rural residential setting. We did not use \nmultiple imputation to impute missing data because \nthe source of nearly all missing data was dependent \nvariables, and imputed values for dependent vari-\nables are not typically included in regression analy-\nses (von Hippel 2007).\nMeasures\nHealth care access.\u2002 Inability to access health care \nwas measured using the following item: \u201cThinking \nabout the times you needed medical care in the last \n12 months, how often were you able to get it?\u201d Par-\nticipants who reported always or sometimes being \nable to access care were considered \u201cable to get \ncare,\u201d and participants who reported never being \nable to get care were considered \u201cunable to get \ncare.\u201d Sensitivity analyses were conducted with \nalternative coding of inability to access care. Results \nwere not driven by choice of coding (see Appendix \nTable A in the online version of the article).\nNo health insurance was measured using a sin-\ngle item from the AAMC survey: \u201cWhat type of \nhealth insurance did you have the most recent time \nyou needed medical care?\u201d Participants who \nreported that they did not have health insurance \nwere considered uninsured.\nSix barriers to health care access were examined \namong participants who had at least one medical care \nTable 1.\u2002 State-Level Data and Descriptive Statistics, 2014\u20132018 (N = 51; Averaged across Years).\nMeasure\nData Source\nMean (SD)\nRange\nState-level sexism index\nMultiple sources\n.00 (1.00)\n\u20131.98\u20133.51\nEarnings ratio (M:W)\nBureau of Labor Statistics\n1.24 (.06)\n1.13\u20131.44\nLabor force ratio (M:W)\nBureau of Labor Statistics\n1.15 (.04)\n1.06\u20131.28\nPoverty ratio (W:M)\nIPUMS Current Population Survey\n1.05 (.05)\n.89\u20131.14\nProportion men in state legislature\nCenter for American Women in \nPolitics\n.75 (.07)\n.59\u2013.87\nNo paid family/medical leave policy\nNational Partnership for Women & \nFamilies\n.94 (.24)\n0\u20131\nNo state law restricting gun ownership \nfor domestic violence offenders\nState Firearms Laws Database\n.65 (.45)\n0\u20131\nProportion women without abortion \naccess\nGuttmacher Institute\n.46 (.26)\n0\u2013.96\nNo Medicaid expansion\nKaiser Family Foundation\n.28 (.45)\n0\u20131\nProportion of white residents\nU.S. Census Bureau\n.77 (.13)\n.25\u2013.95\nProportion of people residing in urban \nareas\nU.S. Census Bureau\n.74 (.15)\n.39\u20131.00\nProportion of people in poverty\nU.S. Census Bureau\n.13 (3.61)\n.07\u2013.22\nGini coefficient\nU.S Census Bureau\n.46 (.02)\n.42\u2013.53\nNote: M = men; W = women.\n\nRapp et al.\t\n7\nvisit in the past year, including (1) inconsistency in \nability to access care, (2) delay in accessing care, (3) \nlimited choice in care, (4) inability to fill a prescrip-\ntion due to out-of-pocket cost, (5) inability to com-\nplete a medical test or treatment due to cost, and (6) \ndifficulty paying medical bills. For a description of \nquestionnaire items and coding, see Appendix Table \nB in the online version of the article.\nPatient\u2013provider communication.\u2002 Three items were \nused to gauge the quality of patient\u2013provider com-\nmunication during the participants\u2019 most recent med-\nical care visit. Participants were asked whether \nproviders (1) \u201cexplain[ed] things in a way that was \neasy to understand,\u201d (2) \u201canswer[ed] questions\u201d to \nthe participant\u2019s satisfaction, and (3) \u201cspen[t] enough \ntime\u201d with the participant during the visit. For ques-\ntionnaire items and coding, see Appendix Table B in \nthe online version of the article.\nState-level sexism.\u2002 Aligned with other studies of \nstructural gender discrimination (Chen et\u00a0al. 2005; \nHoman 2019; Kawachi et\u00a0al. 1999), state-level sex-\nism was measured using publicly available data to \ngauge economic inequity, labor force inequity, polit-\nical inequity, and lack of reproductive rights. State-\nlevel economic and labor force measures included \nratios of men\u2019s to women\u2019s earnings, women\u2019s to \nmen\u2019s poverty, and men\u2019s to women\u2019s labor force \nparticipation. Political inequity was measured using \nthe proportion of state legislature seats occupied by \nmen. Policy-based inequity was measured based on \nthe absence of three state-level policies that dispro-\nportionately benefit women: paid family and medi-\ncal leave, Medicaid expansion, and gun ownership \nrestrictions for domestic violence offenders. Finally, \nlack of reproductive choice was measured using the \nproportion of women residing in counties without an \nabortion provider. We selected these measures \nbecause all were available annually from 2014 to \n2018, with the exception of abortion access (avail-\nable 2015, 2017). After using linear interpolation for \nabortion access missing data, we linked state-level \nmeasures to AAMC study participants based on year \nof participation. Participants in the final wave of the \nAAMC survey (January 2019) were linked to 2018 \nstate-level measures. Consistent with Homan (2019), \na continuous index of state-level sexism was created \nby standardizing state-level measures relative to the \nfull observation period and then summing standard-\nized scores to create a continuous index of state-level \nsexism (Cronbach\u2019s \u03b1 = .70). We then divided the \nindex by its standard deviation so a one-unit change \nin state-level sexism reflects a 1 SD difference. \nConfirmatory factor analysis suggested a one-factor \nstructure for the state-level sexism index (see Appen-\ndix C in the online version of the article). Supple-\nmentary analyses also ensured that results were not \ndriven by any single item in the index (see Appendix \nTable D in the online version of the article).\nCovariates.\u2002 We adjusted for four state-level \ncovariates, including the percentage of white resi-\ndents, percentage of people residing in an urban \narea, poverty rate, and Gini coefficient (to capture \nstate income inequity). Individual-level covariates \nincluded age group, race-ethnicity, marital status, \npresence of a child younger than18 years old in the \nhousehold, urbanicity, income category, educational \nattainment, and frequency of needing medical care \nover the past year.\nAnalytic Methods\nDescriptive statistics were computed for the state-\nlevel sexism measures and individual-level measures \nfrom the AAMC survey. Weighted individual-level \ndescriptive statistics were computed separately for \nmen and women, and Pearson\u2019s \u03c72 and Mann-\nWhitney U tests were conducted to test for gender \ndifferences.\nLogistic regression analyses were completed in \nthree stages. First, we tested associations of state-\nlevel sexism with inability to access health care \nand lack of health insurance. Second, we tested \nassociations between state-level sexism and spe-\ncific barriers to health care access among partici-\npants who accessed medical care at least once over \nthe past year. Third, we tested associations \nbetween state-level sexism and patient\u2013provider \ncommunication.\nFor all analyses, regressions were first run sep-\narately for women and men, then run using the \nentire sample and including an interaction term to \ntest for gender differences. All models adjusted for \ncovariates listed previously. All analyses were \nconducted using STATA and employed U.S. cen-\nsus weights. To account for the nested nature of \nthe data, multilevel analysis was employed by \nclustering by state.\nResults\nDescriptive Statistics of State-Level \nSexism\nTable 1 shows descriptive statistics for the state-\nlevel structural sexism measures averaged across \nstudy years. The three economic measures of \n\n8\t\nJournal of Health and Social Behavior 63(1)\nstate-level sexism are ratios, with values greater \nthan 1 indicating gender inequity that favors men. \nThe mean values for all three measures are greater \nthan 1, meaning that at the state level, men have \nhigher earnings, higher labor force participation, \nand lower poverty rates than women on average. \nAmong the political measures of state-level sexism, \nrepresentation of men in state legislature was calcu-\nlated as a proportion, with .5 indicating gender \nequality and higher proportions indicating more \nrepresentation of men in government relative to \nwomen. All states had higher representation of men \nin state government (range = .59\u2013.87). In addition, \nthe majority of states had no paid family/medical \nleave policy across the observation period (94%) \nand no policy prohibiting gun ownership for people \ncharged with domestic violence (65%). Although \nthe majority of states expanded Medicaid eligibil-\nity, 28% did not expand eligibility during the study \ntime frame. Finally, the number of women residing \nin counties without an abortion provider was mea-\nsured as a proportion, with higher proportions indi-\ncating lower abortion access. The percentage of \nwomen residing in counties without an abortion \nprovider varied from 0% to 96%.\nDescriptive Statistics of Health Care \nAccess and Quality among Women  \nand Men\nTable 2 shows weighted descriptive statistics for \nthe AAMC survey. Women were slightly more \nlikely to report inability to access care compared to \nmen (p = .034 for gender difference). There was no \nsignificant gender difference in being uninsured (p \n= .351) or inconsistent access to care (p = .208). \nHowever, women were more likely to report lim-\nited choice in care (p < .001), and men were more \nlikely to report delay in accessing care (p < .001) \nand affordability barriers to care (p = .005 for \nunable to pay medical bills; p < .001 for medical \ntests and prescriptions too expensive). Women \nwere also less likely to report that providers spent \nenough time with them during a recent visit (p = \n.045). Women and men in the sample also differed \nin most sociodemographic characteristics, includ-\ning age, race, and ethnicity. Men were also more \nlikely to have higher household income (p < .001), \nmore years of education (p < .001), be married (p < \n.001), have children (p = .002), and live in an urban \nresidential setting (p < .001). Finally, women \nreported significantly more frequent need for health \ncare than men (p < .001).\nRelationship between State-Level Sexism \nand Barriers in Access to Health Care\nTable 3 shows results from logistic regression mod-\nels of the association between state-level sexism, \ninability to access health care, and lack of health \ninsurance. State-level sexism was associated with \nsignificantly higher odds of inability to access \nhealth care for women (odds ratio [OR] = 1.86, 95% \nCI, 1.51\u20132.30) and marginally higher odds for men \n(OR = 1.32, 95% CI, 0.98 \u20131.78), meaning that a 1 \nSD increase in state-level sexism was associated \nwith 86% higher odds of inability to access care \namong women and 32% higher odds of inability to \naccess care among men. Results from pooled-sam-\nple models that included an interaction term for \nstate-level sexism and gender show no significant \ngender differences in the associations between \nstate-level sexism and inability to access care. \nHigher state-level sexism was also associated with \nhigher odds of being uninsured among both women \nand men (OR = 1.58, 95% CI, 1.25\u20132.00 for women; \nOR = 1.56, 95% CI, 1.14\u20132.12 for men). Pooled-\nsample models with the interaction term reveal no \nsignificant gender difference in this association.\nTable 4 shows logistic regression results for the \nassociations between state-level sexism and avail-\nability barriers to care (inconsistent access to health \ncare, limited choice in health care, and delay in \naccessing care) among men and women who \naccessed medical care at least once during the past \nyear. Overall, there was no gender difference in the \nassociations between state-level sexism and barri-\ners in availability to care. State-level sexism was \npositively associated with inconsistent access to \nhealth care among women (OR = 1.15, 95% CI \n1.02\u20131.29) but not among men (OR = 1.08, 95% CI \n0.91\u20131.29). However, pooled-sample models that \ninclude an interaction term for state-level sexism \nand gender show no significant gender difference in \nthese associations. State-level sexism was not asso-\nciated with limited choice in care or delay in access-\ning care for women or men.\nTable 5 shows logistic regression results for the \nassociations between state-level sexism and afford-\nability barriers to care (unable to pay medical bills, \nunable to complete medical test due to cost, and \nunable to full prescription due to cost) among men \nand women who accessed medical care at least \nonce during the past year. For all three affordability \nbarriers, an increase in state-level sexism predicted \na significant increase in the odds of experiencing a \nbarrier for women but not for men. Women residing \nin states higher in sexism reported 17% higher odds \n\nRapp et al.\t\n9\nTable 2.\u2002 Descriptive Statistics, Consumer Survey of Health Care Access (2014\u20132019).\nWomen  \n(n = 14,304)\nMen  \n(n = 9,946)\nGender Difference  \np Valuea\nBarriers to health care access\n\u2003 Unable to access care\n2.12%\n1.57%\n.034\n\u2003 No health insurance\n8.14%\n7.68%\n.351\nAvailability barriersb\n\u2003 \u2003 Inconsistent access to care\n10.54%\n9.81%\n.208\n\u2003 \u2003 Limited choice in care\n10.85%\n5.44%\n< .001\n\u2003 \u2003 Delay in accessing care\n23.13%\n33.54%\n< .001\n\u2003 Affordability barriersb\n\u2003 \u2003 Unable to pay medical bills\n36.08%\n38.61%\n.005\n\u2003 \u2003 Unable to complete medical test due to cost\n29.69%\n34.46%\n< .001\n\u2003 \u2003 Unable to fill prescription due to cost\n28.91%\n35.39%\n< .001\nPatient\u2013provider communicationb\n\u2003 Provider spent time\n91.04%\n92.11%\n.045\n\u2003 Provider explained\n96.45%\n95.85%\n.093\n\u2003 Provider answered questions\n93.02%\n93.21%\n.691\nControls\n\u2003 Age group\n\u2003 \u2003 18\u201324\n13.80%\n7.63%\n< .001\n\u2003 \u2003 25\u201334\n16.66%\n20.87%\n\u2003 \u2003 35\u201344\n15.11%\n19.81%\n\u2003 \u2003 45\u201354\n18.80%\n16.79%\n\u2003 \u2003 55\u201364\n16.69%\n14.52%\n\u2003 \u2003 65 +\n18.94%\n20.38%\n\u2003 Race-ethnicity\n\u2003 \u2003 Non-Hispanic white\n65.71%\n67.54%\n.017\n\u2003 \u2003 Non-Hispanic black\n10.78%\n10.53%\n\u2003 \u2003 Hispanic\n16.07%\n14.53%\n\u2003 \u2003 Asian\n3.94%\n4.54%\n\u2003 \u2003 Other\n3.50%\n2.86%\n\u2003 Marital status\n\u2003 \u2003 Single, never married\n25.69%\n27.67%\n< .001\n\u2003 \u2003 Married/cohabiting\n50.20%\n59.15%\n\u2003 \u2003 Widowed\n7.32%\n2.98%\n\u2003 \u2003 Divorced\n14.84%\n8.80%\n\u2003 \u2003 Separated\n1.95%\n1.40%\n\u2003 Child\n39.96%\n42.60%\n.002\n\u2003 Urbanicity\n\u2003 \u2003 Suburban\n45.46%\n39.46%\n< .001\n\u2003 \u2003 Urban\n31.25%\n44.32%\n\u2003 \u2003 Rural\n23.29%\n16.22%\n\u2003 Educational attainment\n\u2003 \u2003 Less than high school\n4.94%\n4.45%\n< .001\n\u2003 \u2003 High school degree or equivalent\n32.59%\n24.70%\n\u2003 \u2003 Some college\n35.81%\n31.15%\n\u2003 \u2003 College or more\n26.65%\n39.70%\n\u2003 Income\n\u2003 \u2003 Under $25,000\n25.71%\n14.42%\n< .001\n\u2003 \u2003 $25,000\u2013$49,000\n26.10%\n18.89%\n\u2003 \u2003 $50,000\u2013$74,999\n19.15%\n20.00%\n\u2003 \u2003 $75,000\u201399,999\n10.39%\n17.63%\n\u2003 \u2003 $100,000 +\n18.64%\n29.05%\n\u2003 Needed care 2+ times\n49.51%\n41.30%\n< .001\naPearson\u2019s \u03c72 tests were conducted for nominal variables. Mann-Whitney U tests were conducted for ordinal variables.\nbWomen\u2019s n = 12,448; men\u2019s n = 8,881.\n\n10\t\nJournal of Health and Social Behavior 63(1)\nTable 3.\u2002 Associations between State-Level Sexism, Inability to Access Care, and No Health Insurance \namong Women and Men; Odds Ratios, 95% Confidence Intervals (Consumer Survey of Health Care \nAccess, 2014\u20132019).\nWomen\n(n = 14,304)\nMen  \n(n = 9,946)\nFull Sample  \n(N = 24,250)\nUnable to access health care\nState-level sexism\n1.86***\n(1.51\u20132.30)\n1.32\n(.98\u20131.78)\n1.43**\n(1.14\u20131.79)\nFemale\n.81\n(.61\u20131.09)\nState-level sexism \u00d7 female\n1.26\n(.94\u20131.69)\n\u2002\nNo health insurance\nState-level sexism\n1.58***\n(1.25\u20132.00)\n1.56**\n(1.14\u20132.12)\n1.50**\n(1.15\u20131.97)\nFemale\n.85*\n(.74\u2013.99)\nState-level sexism \u00d7 female\n1.07\n(.95\u20131.20)\nNote: All models are logistic regressions and control for age, race-ethnicity, state Gini coefficient, state racial \ncomposition, state population density, state poverty rate, household income, education, marital status, children in the \nhousehold, urbanicity, and frequency of needing care.\n*p < .05, **p < .01, ***p < .001 (two-tailed tests).\nTable 4.\u2002 Associations between State-Level Sexism and Barriers in Availability of Health Care Access \namong Women and Men; Odds Ratios, 95% Confidence Intervals (Consumer Survey of Health Care \nAccess, 2014\u20132019).\nWomen  \n(n = 12,448)\nMen\n(n = 8,881)\nFull Sample  \n(N = 21,329)\nInconsistent access to health care\nState-level sexism\n1.15*\n(1.02\u20131.29)\n1.08\n(.91\u20131.29)\n1.08\n(.94\u20131.24)\nFemale\n.89\n(.78\u20131.02)\nState-level sexism \u00d7 female\n\u2002\n1.05\n(0.96\u20131.15)\nLimited choice in health care\nState-level sexism\n1.07\n(.98\u20131.17)\n1.03\n(.85\u20131.24)\n1.11\n(.96\u20131.28)\nFemale\n1.72***\n(1.53\u20131.93)\nState-level sexism \u00d7 female\n.94\n(.83\u20131.06)\nDelay in accessing care\nState-level sexism\n.95\n(.85\u20131.07)\n.96\n(.88\u20131.05)\n.93\n(.84\u20131.02)\nFemale\n.68***\n(.63\u2013.73)\nState-level sexism \u00d7 female\n1.05\n(.98\u20131.12)\nNote: All models are logistic regressions and control for age, race-ethnicity, state Gini coefficient, state racial \ncomposition, state population density, state poverty rate, household income, education, marital status, children in the \nhousehold, urbanicity, and frequency of needing care.\n*p < .05, ***p < .001 (two-tailed tests).\n\nRapp et al.\t\n11\nof being unable to pay medical bills compared to \nwomen in states with lower sexism scores (OR = \n1.17, 95% CI 1.08\u20131.27). In addition, women \nexposed to high state-level sexism had 12% higher \nodds of reporting that medical tests and prescrip-\ntions were too expensive (OR = 1.12, 95% CI 1.03\u2013\n1.22). State-level sexism was not associated with \naffordability barriers for men. Pooled-sample mod-\nels including interaction terms between state-level \nsexism and gender reveal that gender differences in \nthe associations between state-level sexism and \naffordability barriers were significant.\nBased on results from the pooled-sample interac-\ntion models, we plotted gender differences in the \npredicted probability of reporting each affordability \nbarrier by state-level sexism scores (Figure 1). The \npredicted probability of inability to pay medical \nbills was 49% among women who resided in states \nwith high sexism scores (3.5 SD above the mean), \ncompared to 30% among women in states with low \nsexism scores (1.5 SD below the mean). In addition, \nthe predicted probability of women reporting expen-\nsive medical tests was 40% in states high in sexism, \ncompared to 25% in states with low sexism scores. \nA similar pattern was observed for women\u2019s report-\ning of expensive prescriptions (38% in high-sexism \nstates compared to 23% in low-sexism states). \nAmong men, the predicted probability of reporting \naffordability barriers showed little change as state-\nlevel sexism increased, remaining within 3 percent-\nage points from the lowest state-level sexism score \nto the highest for each barrier. Supplemental analy-\nsis of average marginal effects of the interaction \nbetween gender and state-level sexism (see \nAppendix Table E in the online version of the arti-\ncle) shows that in states with low state-level sexism, \nthe predicted probability of reporting an affordabil-\nity barrier was significantly lower for women com-\npared to men. Furthermore, in states with above \naverage state-level sexism scores, the predicted \nprobability of reporting an affordability barrier was \nsignificantly higher for women compared to men.\nTable 5.\u2002 Associations between State-Level Sexism and Barriers in Affordability of Health Care Access \namong Women and Men; Odds Ratios, 95% Confidence Intervals (Consumer Survey of Health Care \nAccess, 2014\u20132019).\nWomen\n(n = 12,448)\nMen\n(n = 8,881)\nFull Sample\n(N = 21,329)\nUnable to pay medical bills\nState-level sexism\n1.17***\n(1.08\u20131.27)\n1.00\n(.91\u20131.10)\n1.03\n(.96\u20131.11)\nFemale\n.90*\n(.83\u2013.98)\nState-level sexism \u00d7 female\n1.14***\n(1.07\u20131.21)\n\u2002\nUnable to complete medical test due to cost\nState-level sexism\n1.12**\n(1.03\u20131.22)\n1.03\n(.94\u20131.12)\n1.03\n(.95\u20131.11)\nFemale\n.91\n(.82\u20131.01)\nState-level sexism \u00d7 female\n1.11*\n(1.02\u20131.21)\n\u2002\nUnable to fill prescription due to cost\nState-level sexism\n1.12**\n(1.03\u20131.22)\n1.02\n(.95\u20131.09)\n.98\n(.92\u20131.05)\nFemale\n.82***\n(.74\u2013.89)\nState-level sexism \u00d7 female\n1.17***\n(1.08\u20131.26)\nNote: All models are logistic regressions and control for age, race-ethnicity, state Gini coefficient, state racial \ncomposition, state population density, state poverty rate, household income, education, marital status, children in the \nhousehold, urbanicity, and frequency of needing care.\n*p < .05, **p < .01, ***p < .001 (two-tailed tests).\n\n12\t\nJournal of Health and Social Behavior 63(1)\nRelationship between State-Level \nSexism and Patient\u2013Provider \nCommunication\nWe tested associations between state-level sexism \nand patient\u2013provider communication during the \nrespondent\u2019s most recent health care visit (Table 6). \nState-level sexism was not associated with quality \nof patient\u2013provider communication for men. \nAmong women, the odds of reporting that providers \nexplained things and answered questions increased \nas state-level sexism increased. Among women, a 1 \nSD increase in state-level sexism was associated \nwith 19% higher odds that providers explained \nthings well (OR = 1.19, 95% CI 1.04\u20131.36) and \n14% higher odds that providers answered questions \n(OR = 1.14, 95% CI 1.05\u20131.24). However, pooled \nmodels showed no significant gender difference in \nthese associations.\nDISCUSSION\nSummary of Findings and Connections \nto Existing Research\nThe current study examined the associations \nbetween state-level sexism, barriers to health care \naccess, and health care quality among women and \nmen who reported needing medical care within the \npast year. This study is innovative in examining the \nconsequences of sexism at the structural level on \nhealth care experiences. Aligned with recent work \ninvestigating structural sexism as a determinant of \nhealth (Homan 2019), we used a measure of state-\nlevel sexism that encompassed economic, political, \nand cultural domains of the gender system. We also \nincorporated several measures of health care access \nand quality to assess ability to access health care, \nspecific barriers in the availability and affordability \nof care, and the quality of patient\u2013provider \ncommunication.\nConsistent with our hypotheses, as state-level \nsexism increased, the odds of being unable to get \nhealth care increased for women. This association \nwas robust after adjusting for covariates, suggesting \nthat state-level sexism may shape ability to access \ncare above and beyond other individual- and state-\nlevel determinants of access to care. Higher state-\nlevel sexism was associated with higher odds of \nbeing uninsured for both women and men. Higher \nlevels of sexism may be equally detrimental for \nwomen\u2019s and men\u2019s access to health insurance \nbecause health care policies in states with higher \nstate-level sexism (e.g., Medicaid eligibility) uni-\nformly restrict access to health insurance. Indeed, \nsupplemental analysis that tested the associations \nbetween Medicaid expansion and insurance cover-\nage showed that both women and men residing in \nstates that did not expand Medicaid during the \nstudy period were more likely to be uninsured. \nHowever, restrictions in Medicaid eligibility have \nbeen shown to disproportionately affect women\u2019s \nbarriers to health care access and subsequent health \n(Margerison et\u00a0 al. 2020; Stimpson, Pintor, and \nWilson 2019). We may not have observed such dif-\nference because men and women reported similar \nrates of being insured in the current sample.\nMore state-level sexism was associated with \nmore affordability barriers to care (unable to pay for \nmedical bills, medical tests, and prescriptions) for \nwomen but not for men. Although the mechanisms \nTest Due to Cost\nFigure 1.\u2002 Gender Differences in the Association between State-Level Sexism and Barriers in \nAffordability of Health Care Access (Consumer Survey of Health Care Access, 2014\u20132019).\n\nRapp et al.\t\n13\nlinking state-level sexism to individual-level afford-\nability of care cannot be determined in our analyses, \nwe found that associations between state-level sex-\nism and affordability barriers were not explained by \nindividual sociodemographic and health-related \ncharacteristics. These results suggest that state-level \nsexism may shape women\u2019s ability to afford health \ncare in ways that go beyond \u00adindividual-level socio-\neconomic standing. Future studies should investi-\ngate the ways in which state-level sexism shapes \ncommunity-level contexts that affect health care \naffordability for women.\nInterestingly, in plots showing gender differ-\nences in the association between state-level sexism \nand barriers to health care access, it appears that \nwomen residing in states with higher than average \nstate-level sexism reported more affordability barri-\ners to care compared to men, whereas women in \nstates with lower than average state-level sexism \nreported fewer barriers than men. The finding that \nwomen residing in states with lower sexism scores \nreported fewer barriers to care compared to men \nwarrants further research. Perhaps states with lower \nlevels of sexism have more comprehensive health \ncare legislation that specifically targets the health \ncare needs of women (e.g., expanded Medicaid eli-\ngibility for women of reproductive age or widely \navailable and affordable reproductive care), and \nthis legislation substantially reduces affordability \nbarriers to care for women but not for men. A grow-\ning literature has identified the benefits of health \ncare legislation for women\u2019s health care access \n(Johnston et\u00a0al. 2018; Stevenson et\u00a0al. 2016), but \nlittle research has examined the impacts of similar \nlegislation on men\u2019s access to care. Alternatively, \ngiven the greater health care needs of women on \naverage, state-level social and economic conditions \nmight have a more substantial impact on access to \naffordable care for women compared to men. \nFurther research is needed to identify strategies that \nboth improve women\u2019s access to care and promote \nequitable access for women and men.\nTable 6.\u2002 Associations between State-Level Sexism and Patient\u2013Provider Communication among \nWomen and Men; Odds Ratios, 95% Confidence Intervals (Consumer Survey of Health Care Access, \n2014\u20132019).\nWomen\n(n = 12,448)\nMen\n(n = 8,881)\nFull Sample\n(N = 21,329)\nProvider spent time\nState-level sexism\n1.09\n(1.00\u20131.18)\n.91\n(.79\u20131.03)\n.95\n(.85\u20131.07)\nFemale\n.95\n(.85\u20131.06)\nState-level sexism \u00d7 female\n1.12*\n(1.03\u20131.23)\n\u2002\nProvider explained\nState-level sexism\n1.19**\n(1.04\u20131.36)\n1.02\n(.84\u20131.25)\n1.06\n(.89\u20131.27)\nFemale\n1.21*\n(1.00\u20131.46)\nState-level sexism \u00d7 female\n1.10\n(.95\u20131.27)\n\u2002\nProvider answered questions\nState-level sexism\n1.14**\n(1.05\u20131.24)\n.99\n(.87\u20131.14)\n1.02\n(.90\u20131.14)\nFemale\n1.09\n(.93\u20131.27)\nState-level sexism \u00d7 female\n1.12\n(.97\u20131.30)\nNote: All models are logistic regressions and control for age, race-ethnicity, state Gini coefficient, state racial \ncomposition, state population density, state poverty rate, household income, education, marital status, children in the \nhousehold, urbanicity, and frequency of needing care.\n*p < .05, **p < .01 (two-tailed tests).\n\n14\t\nJournal of Health and Social Behavior 63(1)\nWe found no association of state-level sexism \nwith barriers in the availability of care for women \nand men. These results were unexpected given that \nwomen are more likely to report unmet health care \nneeds and delays in receiving care than men (Long \net\u00a0al. 2011; Ng et\u00a0al. 2010; Rustgi et\u00a0al. 2009) and \nthat state policies can affect the availability of care \nfor women (Johnston et\u00a0al. 2018; Stevenson et\u00a0al. \n2016). Because results could be a function of over-\nall measurement and/or availability of health care \nin the present sample, future research should test \nthis relationship using various measures of health \ncare availability in different samples.\nSurprisingly, across two of the three measures of \npatient\u2013provider communication, women who \nresided in states with higher levels of sexism reported \nmore positive interactions with their providers com-\npared to women in states with lower levels of sex-\nism. We propose several explanations. Because some \ngender stereotypes may entail benevolent sexism, \ndefined as seemingly innocuous and patronizing \nbehaviors that may appear complimentary to some \nindividuals (Glick and Fiske 1996), perhaps women \nin states high in structural sexism do not interpret \nsexist patient\u2013provider communication as negative.\nWomen may also be more likely to minimize inter-\npersonal experiences of sexism that are more ambigu-\nous as a coping mechanism (Ruggiero and Taylor \n1995, 1997), particularly in contexts where women \nfeel that they may not be believed or may be retaliated \nagainst (Kaiser and Miller 2001; Sechrist and Delmar \n2009). Perhaps women in states where sexism is the \nstructural norm utilize this coping mechanism more \nfrequently to preserve their psychosocial health \n(Bosson, Pinel, and Vandello 2010) in the otherwise \ntoxic state-level environment. Conversely, they may \nbe putting more effort into communication with pro-\nviders given the hostile structural-level environment \nand, in the process of doing so, getting more positive \noutcomes due to their own self-advocacy, in line with \nresearch on confronting sexism (Ayres, Friedman, and \nLeaper 2009; Good et\u00a0 al. 2019). Such propositions \nhighlight the importance of further research that will \nincorporate examination of the social psychological \naspects of gender and the gendered system, which we \nwere not able to do given data limitations.\nFinally, provider communication with patients \nmay not necessarily reflect institutional and cultural \ncontexts where they are currently practicing medi-\ncine but instead reflect communication skills devel-\noped through medical training from institutions in \nother geographic areas.\nSuggestions for Future Research\nAlthough this study enhances our understanding of \nthe relationship between structural sexism and \nhealth care access and quality, several limitations \nshould be addressed in future research. Because \ndata utilized in this study were observational and \ncross-sectional, our results do not imply causality. \nAlthough sampling was stratified by age and insur-\nance status and all analyses incorporated U.S. cen-\nsus weights, study participants were more likely to \nbe insured and were less likely to be unable to get \ncare compared to U.S. national averages (Azam and \nMoy 2016). Therefore, our estimates are likely to be \nconservative and do not necessarily generalize to \nthe U.S. population. In addition, although we utilize \nmultiple measures of state-level sexism that mirror \ndimensions of structural inequity explored in exist-\ning research (e.g., Homan 2019; Wisdom et\u00a0 al. \n2005), these measures are not exhaustive. State leg-\nislation relating to reproductive rights and spending \non public assistance programs were not included in \nthe present study due to gaps in the availability of \ndata during the study years. We could not examine \nthe relative importance of indicators of structural \nsexism given multicollinearity. Consequently, we \ncannot glean which dimensions are particularly \nimportant.\nAlthough we adjusted for potential state-level \nconfounders in the relationship between state-level \nsexism and barriers to care, other state-level charac-\nteristics may contribute to the results we observed. \nFuture research is needed to identify additional \nstate-level economic, political, and cultural factors \nthat may contribute to health care disparities. In \naddition, we limited our measurement of structural \nsexism to the state level and did not examine \ncounty-, neighborhood-, or household-level indica-\ntors of sexism. Policies and inequalities at the state \nlevel impact individual health and health care \n(Hawkins et\u00a0al. 2020; Lee et\u00a0al. 2020; Montez et\u00a0al. \n2016), but state-level measures do not account for \nwithin-state variation in economic, political, and \ncultural contexts.\nWe did not analyze relationships between state-\nlevel sexism and barriers to health care access \namong transgender and nonbinary populations due \nto data limitations. Growing literature documents \nextensive barriers to health care access for trans and \nnonbinary people due to discriminatory policies at \nthe federal and state levels (Bakko and Kattari \n2019; Goldenberg et\u00a0al. 2020). Future data collec-\ntion efforts must adopt inclusive measures of \n\nRapp et al.\t\n15\ngender identity to further this area of research using \nnational samples.\nConclusion\nState-level sexism appears to increase barriers to \nhealth care access among women in the United \nStates by affecting women\u2019s ability to afford medi-\ncal care. These results were robust after adjusting \nfor individual-level sociodemographic characteris-\ntics, suggesting that state-level sexism acts on envi-\nronmental or organizational contexts to directly \nshape women\u2019s access to health care. Our findings \nare consistent with existing research that documents \nthe negative impacts of structural gender inequities \non women\u2019s health and longevity (Homan 2019; \nMontez et\u00a0al. 2016; Wisdom et\u00a0al. 2005). Furthermore, \nin our analysis, state-level sexism appears to have \nno effect on men\u2019s access to health care. This coun-\nters some research that finds structural gender \ninequity to be universally harmful for both wom-\nen\u2019s and men\u2019s health (Homan 2019; Kawachi \net\u00a0al. 1999) and instead suggests that macro-level \nstructural sexism is uniquely detrimental for wom-\nen\u2019s health care access. Finally, women in states \nwith higher levels of structural sexism reported \nbetter quality of patient\u2013provider communication \ncompared to women in states with lower levels of \nsexism. Further research is needed to unpack these \nassociations.\nThese findings demonstrate the need to use a \nmultilevel approach to understand and act on the \nsocial determinants of women\u2019s health care access \nand quality. The implications of these findings are \nvital given that multiple states have enacted legisla-\ntion to restrict women\u2019s access to reproductive care \nand weakened antipoverty programs that provide \nneeded assistance to millions of women and moth-\ners (Kogan et\u00a0 al. 2019; McKernan and Ratcliffe \n2018; Reingold and Gostin 2019). Meanwhile, the \ngender gap in wages has remained stagnant in the \nUnited States since 2004, and women continue to \nbe severely underrepresented in federal, state, and \nlocal government (Homan 2017; Patton and \nFording 2020; U.S. Bureau of Labor Statistics \n2019). Findings from the current investigation \nimplicate sexism across economic, labor force, and \npolitical institutions as a key determinant of wom-\nen\u2019s barriers to accessing health care. As the future \nof U.S. health care reform continues to be debated, \nwe urge lawmakers to also consider social and eco-\nnomic policies as essential forms of health care \nlegislation.\nAcknowledgments\nThis material is based on data provided by the Association \nof American Medical Colleges (AAMC) to the Principal \nInvestigators (Rapp and Volpe) and their team. The views \nexpressed herein are those of the authors and do not neces-\nsarily reflect the position or policy of the AAMC.\nORCID iD\nKristen Schorpp Rapp \n https://orcid.org/0000-0003- \n4829-1612\nSupplemental Material\nAppendices A through E are available in the online version \nof the article.\nReferences\nAtrash, Hani K., Kay Johnson, Myron Adams, Jos\u00e9 F. \nCordero, and Jennifer Howse. 2006. \u201cPreconception \nCare for Improving Perinatal Outcomes: The Time to \nAct.\u201d Maternal and Child Health Journal 10(1):3\u201311.\nAyres, Melanie M., Carly K. Friedman, and Campbell \nLeaper. 2009. \u201cIndividual and Situational Factors \nRelated \nto \nYoung \nWomen\u2019s \nLikelihood \nof \nConfronting Sexism in Their Everyday Lives.\u201d Sex \nRoles 61(7\u20138):449\u201360.\nAzam, Irim, and Ernest Moy. 2016. National Healthcare \nQuality and Disparities Report: Chartbook on \nAccess to Health Care. Rockville, MD: Agency for \nHealthcare Research and Quality.\nBakko, Matthew, and Shanna K. Kattari. 2019. \n\u201cDifferential Access to Transgender Inclusive \nInsurance and Healthcare in the United States: \nChallenges to Health across the Life Course.\u201d \nJournal of Aging and Social Policy 33(1):67\u201381.\nBaum, Charles L., and Christopher J. Ruhm. 2016. \u201cThe \nEffects of Paid Family Leave in California on Labor \nMarket Outcomes.\u201d Journal of Policy Analysis and \nManagement 35(2):333\u201356.\nBerkman, Michael B., and Robert E. O\u2019Connor. 1993. \n\u201cDo Women Legislators Matter?: Female Legislators \nand State Abortion Policy.\u201d American Politics \nResearch 21(1):102\u201324.\nBobbitt-Zeher, Donna. 2011. \u201cGender Discrimination at \nWork: Connecting Gender Stereotypes, Institutional \nPolicies, and Gender Composition of Workplace.\u201d \nGender & Society 25(6):764\u201386.\nBolzendahl, Catherine, and Clem Brooks. 2007. \n\u201cWomen\u2019s Political Representation and Welfare \nState Spending in 12 Capitalist Democracies.\u201d Social \nForces 85(4):1509\u201334.\nBosson, Jennifer K., Elizabeth C. Pinel, and Joseph \nA. Vandello. 2010. \u201cThe Emotional Impact of \nAmbivalent \nSexism: \nForecasts \nVersus \nReal \nExperiences.\u201d Sex Roles 62(7\u20138):520\u201331.\nBratton, Kathleen A., and Kerry L. Haynie. 1999. \n\u201cAgenda Setting and Legislative Success in State \n\n16\t\nJournal of Health and Social Behavior 63(1)\nLegislatures: The Effects of Gender and Race.\u201d \nJournal of Politics 61(3):658\u201379.\nCase, Anne, and Christina Paxson. 2005. \u201cSex Differences \nin Morbidity and Mortality.\u201d Demography 42(2): \n189\u2013214.\nChapman, Elizabeth N., Anna Kaatz, and Molly \nCarnes. 2013. \u201cPhysicians and Implicit Bias: How \nDoctors May Unwittingly Perpetuate Health Care \nDisparities.\u201d Journal of General Internal Medicine \n28(11):1504\u201310.\nChen, Ying Yeh, S. V. Subramanian, Doloros Acevedo-\nGarcia, and Ichiro Kawachi. 2005. \u201cWomen\u2019s Status \nand Depressive Symptoms: A Multilevel Analysis.\u201d \nSocial Science and Medicine 60(1):49\u201360.\nCreanga, Andreea A., Cynthia J. Berg, Jean Y. Ko, \nSherry L. Farr, Van T. Tong, F. Carol Bruce, and \nWilliam M. Callaghan. 2014. \u201cMaternal Mortality \nand Morbidity in the United States: Where Are We \nNow?\u201d Journal of Women\u2019s Health 23(1):3\u20139.\nDaniel, Hilary, Shari M. Erickson, Sue S. Bornstein, \nGregory C. Kane, Heather E. Gantzer, Tracey L. \nHenry, Joshua D. Lenchus, et\u00a0al. 2018. \u201cWomen\u2019s \nHealth Policy in the United States: An American \nCollege of Physicians Position Paper.\u201d Annals of \nInternal Medicine 168(12):874\u201375.\nDickman, Samuel L., David U. Himmelstein, and Steffie \nWoolhandler. 2017. \u201cInequality and the Health-\nCare System in the USA.\u201d The Lancet 389(10077): \n1431\u201341.\nDoran, Elizabeth L., Ann P. Bartel, Christopher J. \nRuhm, and Jane Waldfogel. 2020. \u201cCalifornia\u2019s Paid \nFamily Leave Law Improves Maternal Psychological \nHealth.\u201d Social Science & Medicine 256:113003. \ndoi:10.1016/j.socscimed.2020.113003.\nDoyal, Lesley. 2000. \u201cGender Equity in Health: Debates \nand Dilemmas.\u201d Social Science & Medicine \n51(6):931\u201339.\nElliott, Marc N., William G. Lehrman, Megan K. Beckett, \nElizabeth Goldstein, Katrin Hambarsoomian, and \nLaura A. Giordano. 2012. \u201cGender Differences in \nPatients\u2019 Perceptions of Inpatient Care.\u201d Health \nServices Research 47(4):1482\u2013501.\nFeagin, Joe, and Zinobia Bennefield. 2014. \u201cSystemic \nRacism and U.S. Health Care.\u201d Social Science & \nMedicine 103:7\u201314.\nGault, Barbara, Heidi Hartmann, Ariane Hegewisch, \nJessica Milli, and Lindsey Reichlin. 2014. Paid \nParental Leave in the United States: What the \nData Tell Us about Access, Usage, and Economic \nand Health Benefits. Washington, DC: Institute for \nWomen\u2019s Policy Research.\nGijsbers Van Wijk, Cecile M. T., Katja P. Van Vliet, and \nAnnemarie M. Kolk. 1996. \u201cGender Perspectives and \nQuality of Care: Towards Appropriate and Adequate \nHealth Care for Women.\u201d Social Science & Medicine \n43(5):707\u201320.\nGlick, Peter, and Susan T. Fiske. 1996. \u201cThe Ambivalent \nSexism Inventory: Differentiating Hostile and \nBenevolent Sexism.\u201d Journal of Personality and \nSocial Psychology 70(3):491\u2013512.\nGoldenberg, Tamar, Sari L. Reisner, Gary W. Harper, \nKristi E. Gamarel, and Rob Stephenson. 2020. \u201cState \nPolicies and Healthcare Use among Transgender \nPeople in the U.S.\u201d American Journal of Preventive \nMedicine 59(2):247\u201359.\nGood, Jessica J., Julie A. Woodzicka, Kimberly A. \nBourne, and Corinne A. Moss-Racusin. 2019. \u201cThe \nDecision to Act: Factors That Predict Women\u2019s and \nMen\u2019s Decisions to Confront Sexism.\u201d Pp. 49\u201371 \nin Confronting Prejudice and Discrimination: The \nScience of Changing Minds and Behaviors, edited \nby R. K. Mallett and M. J. Monteith. Cambridge: \nAcademic Press.\nHall, Judith A., and Debra L. Roter. 1995. \u201cPatient \nGender and Communication with Physicians: Results \nof a Community-Based Study.\u201d Women\u2019s Health \n1(1):77\u201395.\nHamberg, Katarina. 2008. \u201cGender Bias in Medicine.\u201d \nWomen\u2019s Health 4(3):237\u201343.\nHarnois, Catherine E., and Jo\u00e3o L. Bastos. 2018. \n\u201cDiscrimination, \nHarassment, \nand \nGendered \nHealth Inequalities: Do Perceptions of Workplace \nMistreatment Contribute to the Gender Gap in Self-\nReported Health?\u201d Journal of Health and Social \nBehavior 59(2):283\u201399.\nHawkins, Summer Sherburne, Marco Ghiani, Sam \nHarper, Christopher F. Baum, and Jay S. Kaufman. \n2020. \u201cImpact of State-Level Changes on Maternal \nMortality: A Population-Based, Quasi-experimental \nStudy.\u201d American Journal of Preventive Medicine \n58(2):165\u201374.\nHess, Cynthia, Jessica Milli, Jeff Hayes, Ariane \nHegewisch, Yana Mayayeva, Stephanie Rom\u00e1n, \nJulie Anderson, and Justine Augeri. 2015. The Status \nof Women in the States: 2015. Washington, DC: \nInstitute for Women\u2019s Policy Research.\nHoman, Patricia. 2017. \u201cPolitical Gender Inequality and \nInfant Mortality in the United States, 1990\u20132012.\u201d \nSocial Science & Medicine 182:127\u201335.\nHoman, Patricia. 2019. \u201cStructural Sexism and Health \nin the United States: A New Perspective on Health \nInequality and the Gender System.\u201d American \nSociological Review 84(3):486\u2013516.\nJesus, Tiago Silva, and Isabel Lopes Silva. 2016. \u201cToward \nan Evidence-Based Patient\u2013Provider Communication \nin Rehabilitation: Linking Communication Elements \nto \nBetter \nRehabilitation \nOutcomes.\u201d \nClinical \nRehabilitation 30(4):315\u201328.\nJohnston, Emily M., Andrea E. Strahan, Peter Joski, Anne \nL. Dunlop, and E. Kathleen Adams. 2018. \u201cImpacts \nof the Affordable Care Act\u2019s Medicaid Expansion \non Women of Reproductive Age: Differences by \nParental Status and State Policies.\u201d Women\u2019s Health \nIssues 28(2):122\u201329.\nKaiser, Cheryl R., and Carol T. Miller. 2001. \u201cStop \nComplaining! \nThe \nSocial \nCosts \nof \nMaking \nAttributions to Discrimination.\u201d Personality and \nSocial Psychology Bulletin 27(2):254\u201363.\nKawachi, Ichiro, Bruce P. Kennedy, Vanita Gupta, and \nDeborah Prothrow-Stith. 1999. \u201cWomen\u2019s Status \n\nRapp et al.\t\n17\nand the Health of Women and Men: A View from \nthe States.\u201d Social Science & Medicine 48(1):21\u201332.\nKent, Jennifer A., Vinisha Patel, and Natalie A. Varela. \n2012. \u201cGender Disparities in Health Care.\u201d Mount \nSinai Journal of Medicine 79(5):555\u201359.\nKogan, Richard, Hannah Katch, Dorothy Rosenbaum, \nDouglas Rice, Kathleen Romig, Ife Floyd, and \nSharon Parrott. 2019. Cuts to Low-Income Assistance \nPrograms in President Trump\u2019s 2020 Budget Are \nWide-Ranging. Washington, DC: Center on Budget \nand Policy Priorities.\nKrieger, Nancy. 2001. \u201cTheories for Social Epidemiology \nin the 21st Century: An Ecosocial Perspective.\u201d \nInternational Journal of Epidemiology 30(4):668\u201377.\nKrieger, Nancy. 2020. \u201cMeasures of Racism, Sexism, \nHeterosexism, and Gender Binarism for Health \nEquity Research: From Structural Injustice to \nEmbodied Harm\u2014An Ecosocial Analysis.\u201d Annual \nReview of Public Health 41:37\u201362.\nLavelle, Bridget, and Pamela J. Smock. 2012. \u201cDivorce \nand Women\u2019s Risk of Health Insurance Loss.\u201d \nJournal of Health and Social Behavior 53(4):413\u201331.\nLawrence, Hal C., and Debra L. Ness. 2017. \u201cPlanned \nParenthood \nProvides \nEssential \nServices \nThat \nImprove Women\u2019s Health.\u201d Annals of Internal \nMedicine 166(6):443\u201344.\nLee, Bethany C., Sepideh Modrek, Justin S. White, \nAkansha Batra, Daniel F. Collin, and Rita Hamad. \n2020. \u201cThe Effect of California\u2019s Paid Family Leave \nPolicy on Parent Health: A Quasi-experimental \nStudy.\u201d Social Science & Medicine 251:112915. \ndoi:10.1016/j.socscimed.2020.112915.\nLeveille, Suzanne G., Brenda W. J. H. Penninx, David \nMelzer, Grant Izmirlian, and Jack M. Guralnik. \n2000. \u201cSex Differences in the Prevalence of Mobility \nDisability in Old Age: The Dynamics of Incidence, \nRecovery, and Mortality.\u201d The Journal of Gerontology, \nSeries B: Psychological Sciences and Social Sciences \n55(I):41\u201350.\nLong, Sharon K., Karen Stockley, and Shanna Shulman. \n2011. \u201cHave Gender Gaps in Insurance Coverage \nand Access to Care Narrowed under Health Reform? \nFindings from Massachusetts.\u201d American Economic \nReview 101(3):640\u201344.\nMarcum, James A. 2017. \u201cClinical Decision-Making, \nGender Bias, Virtue Epistemology, and Quality \nHealthcare.\u201d Topoi 36(3):501\u2013508.\nMargerison, Claire E., Colleen L. MacCallum, Jiajia \nChen, Yasamean Zamani-Hank, and Robert Kaestner. \n2020. \u201cImpacts of Medicaid Expansion on Health \namong Women of Reproductive Age.\u201d American \nJournal of Preventive Medicine 58(1):1\u201311.\nMcKernan, Signe-Mary, and Caroline Ratcliffe. 2018. \nPolicy Efforts to Reduce Material Hardship for Low-\nIncome Families. Washington, DC: Urban Institute.\nMolix, Lisa. 2014. \u201cSex Differences in Cardiovascular \nHealth: Does Sexism Influence Women\u2019s Health?\u201d \nAmerican Journal of the Medical Sciences 348(2): \n153\u201355.\nMontez, Jennifer Karas, Anna Zajacova, and Mark \nD. Hayward. 2016. \u201cExplaining Inequalities in \nWomen\u2019s Mortality between U.S. States.\u201d SSM - \nPopulation Health 2:561\u201371.\nMoss, Nancy E. 2002. \u201cGender Equity and Socioeconomic \nInequality: A Framework for the Patterning of \nWomen\u2019s Health.\u201d Social Science & Medicine 54(5): \n649\u201361.\nNg, Judy H., Shakeh J. Kaftarian, Wilma M. Tilson, Paul \nGorrell, Xiuhua Chen, Francis D. Chesley, and Sarah \nHudson Scholle. 2010. \u201cSelf-Reported Delays in \nReceipt of Health Care among Women with Diabetes \nand Cardiovascular Conditions.\u201d Women\u2019s Health \nIssues 20(5):316\u201322.\nNolen-Hoeksema, Susan. 2001. \u201cGender Differences in \nDepression.\u201d Current Directions in Psychological \nScience 10(5):173\u201376.\nOwens, Gary M. 2008. \u201cGender Differences in Health Care \nExpenditures, Resource Utilization, and Quality of \nCare.\u201d Journal of Managed Care Pharmacy 14(3):2\u20136.\nPabayo, Roman, Daniel M. Cook, Guy Harling, Anastasia \nGunawan, Natalie A. Rosenquist, and Peter Muennig. \n2019. \u201cState-Level Income Inequality and Mortality \namong Infants Born in the United States 2007\u20132010: \nA Cohort Study.\u201d BMC Public Health 19(1):1\u20139.\nPark, Soon Seok. 2017. \u201cGendered Representation and \nCritical Mass: Women\u2019s Legislative Representation \nand Social Spending in 22 OECD Countries.\u201d \nSociological Perspectives 60(6):1097\u2013114.\nPatchias, Elizabeth M., and Judy Waxman. 2007. Women \nand Health Coverage: The Affordability Gap. New \nYork, NY: The Commonwealth Fund.\nPatton, Dana, and Richard C. Fording. 2020. \u201cThe \nEffect of Women\u2019s Descriptive Representation on \nPremature Mortality.\u201d Social Forces 99(2):820\u201346.\nPenchansky, Roy, and J. William Thomas. 1981. \u201cThe \nConcept of Access: Definition and Relationship \nto Consumer Satisfaction.\u201d Medical Care 19(2): \n127\u201340.\nReingold, Rebecca B., and Lawrence O. Gostin. 2019. \n\u201cState Abortion Restrictions and the New Supreme \nCourt: Women\u2019s Access to Reproductive Health \nServices.\u201d Journal of the American Medical \nAssociation 322(1):21\u201322.\nRidgeway, Cecilia L., and Shelley J. Correll. 2004. \n\u201cUnpacking the Gender System: A Theoretical \nPerspective on Gender Beliefs and Social Relations.\u201d \nGender & Society 18(4):510\u201331.\nRisman, Barbara J. 2004. \u201cGender as a Social Construct: \nTheory Wrestling with Activism.\u201d Gender & Society \n18(4):429\u201350.\nRuggiero, Karen M., and Donald M. Taylor. 1995. \n\u201cCoping with Discrimination: How Disadvantaged \nGroup Members Perceive the Discrimination That \nConfronts Them.\u201d Journal of Personality and Social \nPsychology 68(5):826\u201338.\nRuggiero, Karen M., and Donald M. Taylor. 1997. \n\u201cWhy Minority Group Members Perceive or Do Not \nPerceive the Discrimination That Confronts Them: \n\n18\t\nJournal of Health and Social Behavior 63(1)\nThe Role of Self-Esteem and Perceived Control.\u201d \nJournal of Personality and Social Psychology \n72(2):373\u201389.\nRustgi, Sheila D., Michelle M. Doty, and Sara R. Collins. \n2009. Women at Risk: Why Many Women Are \nForgoing Needed Health Care. New York, NY: The \nCommonwealth Fund.\nSechrist, Gretchen B., and Courtney Delmar. 2009. \n\u201cWhen Do Men and Women Make Attributions to \nGender Discrimination? The Role of Discrimination \nSource.\u201d Sex Roles 61(9\u201310):607\u201320.\nSivaraman, Josie J., Shabbar I. Ranapurwala, Kathryn \nE. Moracco, and Stephen W. Marshall. 2019. \n\u201cAssociation of State Firearm Legislation with \nFemale Intimate Partner Homicide.\u201d American \nJournal of Preventive Medicine 56(1):125\u201333.\nStevenson, Amanda J., Imelda M. Flores-Vazquez, Richard \nL. Allgeyer, Pete Schenkkan, and Joseph E. Potter. \n2016. \u201cEffect of Removal of Planned Parenthood \nfrom the Texas Women\u2019s Health Program.\u201d New \nEngland Journal of Medicine 374(9):853\u201360.\nStimpson, Jim P., Jessie Kemmick Pintor, and Fernando A. \nWilson. 2019. \u201cAssociation of Medicaid Expansion \nwith Health Insurance Coverage by Marital Status \nand Sex.\u201d PLOS ONE 14(10):e0223556. doi:10.1371/\njournal.pone.0223556.\nStreet, Richard L., Gregory Makoul, Neeraj K. \nArora, and Ronald M. Epstein. 2009. \u201cHow Does \nCommunication Heal? Pathways Linking Clinician\u2013\nPatient Communication to Health Outcomes.\u201d \nPatient Education and Counseling 74(3):295\u2013301.\nU.S. Bureau of Labor Statistics. 2019. \u201cHighlights \nof Women\u2019s Earnings in 2018.\u201d BLS Reports \n1083(November):1\u2013113.\nvon Hippel, Paul T. 2007. \u201cRegression with Missing Ys: \nAn Improved Strategy for Analyzing Multiply Imputed \nData.\u201d Sociological Methodology 37(1):83\u2013117.\nVowles, Kevin E., and Miles Thompson. 2012. \u201cThe \nPatient\u2013Provider Relationship in Chronic Pain.\u201d \nCurrent Pain and Headache Reports 16(2):133\u201338.\nWisdom, Jennifer P., Michelle Berlin, and Jodi A. \nLapidus. 2005. \u201cRelating Health Policy to Women\u2019s \nHealth Outcomes.\u201d Social Science & Medicine \n61(8):1776\u201384.\nYearby, Ruqaiijah. 2018. \u201cRacial Disparities in Health \nStatus and Access to Healthcare: The Continuation \nof Inequality in the United States Due to Structural \nRacism.\u201d American Journal of Economics and \nSociology 77(3\u20134):1113\u201352.\nAuthor Biographies\nKristen Schorpp Rapp is an assistant professor in the \nDepartment of Sociology and Public Health at Roanoke \nCollege. Her research examines the social determinants \nof health disparities in the United States, with a particu-\nlar focus on the political, economic, and cultural condi-\ntions that shape health inequalities by gender and \nrace-ethnicity.\nVanessa V. Volpe is an assistant professor of psychology \nat North Carolina State University. Her applied health psy-\nchology research examines the health impacts of structural \nsystems of inequity and interpersonal experiences of dis-\ncrimination and the ways that minoritized individuals can \npreserve their health in the United States. Her program of \nresearch seeks to inform policy and practice focused on the \nelimination of racial-ethnic health disparities.\nTabitha L. Hale is a graduate student of public health in \nthe Department of Global and Community Health at \nGeorge Mason University. Her research focuses on the \neffects of endocrine disrupting compounds on the repro-\nductive health of women, specifically examining racial \ndisparities and impacts on vulnerable populations.\nDominique F. Quartararo is a recent graduate from the \nDepartment of Sociology and Public Health at Roanoke \nCollege. She is primarily interested in health policy, wom-\nen\u2019s access to reproductive care, and population health \ndisparities.\n\n\n    Scientific Research Findings:", "answer": "\u2022 Higher state-level sexism is associated with higher inability to access care and more barriers to affording care for women but not for men.\n\u2022 Women residing in states with higher state-level sexism report better quality of care than women in states with lower levels of sexism.\n\u2022 Individual-level factors, such as income, education, family structure, and demographic characteristics, do not explain these links.", "id": 83}
{"query": "You are given the full text of a scientific research paper.\n\n    Summarize the paper's key **research findings** in the style of a policy brief. Your response should:\n    - Focus on the major empirical or model-based results\n    - Present findings that are policy-relevant and supported by the study\n    - Use concise, non-technical language with essential technical terms as needed\n    - Avoid quoting directly or summarizing minor details\n\n    Write a short, clear paragraph appropriate for the \u201cScientific Research Findings\u201d section of a policy brief.\n\n    Scientific Research Paper:\n    Nature Sustainability | Volume 7 | December 2024 | 1592–1595\n1592\nnature sustainability\nBrief Communication\nhttps://doi.org/10.1038/s41893-024-01471-y\nPotential increase of neurotoxic mercury risk \nin global blue carbon nature-based solutions\nZhijia Ci \n  1 \n, Wenjie Shen2,3, Baowei Chen1, Yanbin Li \n  4, Yongguang Yin \n  5, \nXiaoshan Zhang5 & Yong Cai \n  6\nBlue carbon strategies are now being proposed as a promising, nature-based \nsolution for achieving multiple benefits. Here we provide field evidence \non the co-accumulation of organic carbon and neurotoxic mercury (Hg) \nin coastal environments. We estimate the global Hg stock in the top \nmetre of sediment of blue carbon ecosystems to be 21,306 to 125,018 Mg \n(mean = 73,162 Mg), highlighting that Blue Hg stock is an important, \ndynamic, reactive, but overlooked Hg pool in global Hg cycle and health risk.\nAnthropogenic emissions of both carbon dioxide (CO2) and mercury \n(Hg) are long-term stressors for the global environment and human \nsociety1. The global community is seeking common and coordinated \nactions, such as the Paris Agreement on Climate Change (https://unfccc.\nint/process-and-meetings/the-paris-agreement) and the Minamata \nConvention on Mercury (https://minamataconvention.org/en), to \nreduce their emissions and associated adverse effects, although both \ntreaties still are in the early stages of implementation.\nThe term ‘Blue C’ entered the lexicon in 2009 and refers to organic \nC (OC) that is captured and stored by vegetated coastal ecosystems \n(VCEs, including seagrass meadows, tidal marshes and mangrove for-\nests)2,3. Recently, tidal mudflats have also entered the Blue C discourse4. \nThe importance of implementation of Blue C strategies—expanding \nand restoring Blue C ecosystems—to cope with climate crisis, biodi-\nversity loss and natural hazards3 has now reached a consensus in the \ninternational community. However, there are few empirical studies on \nthe negative effects of Blue C strategies5.\nAs a persistent and ubiquitous global pollutant, Hg deposition \nof any species and concentration poses potential risks to humans \nand ecosystems1. Compared with terrestrial environments, coastal \nenvironments have many physical, ecological and biogeochemical \ncharacteristics that increase the accumulation of Hg and the produc-\ntion, release, bioaccumulation and risk of neurotoxic methylmercury \n(MMHg, the methylated form of Hg) (refs. 6–10). Therefore, Hg stored \nin coastal ecosystems poses more serious risks to human health and \necosystem safety10. Field measurements show that OC-rich coastal \nsediments are generally accompanied by elevated Hg concentrations \nsince sedimentary OC provides a large surface area and strong binding \nsites for divalent Hg (Hg(II)) (ref. 9). However, the sample size, spatial \ndistribution and ecosystem type of coastal sediments in existing stud-\nies are extremely limited9,11–13. More critically, no attempt has been \nmade to estimate the Hg stock in global coastal sediments, which is \none of the fundamental constraints on determining Hg risks in global \ncoastal ecosystems.\nIn this Brief Communication, we carry out a targeted survey along \nChinese coastlines to document the relationship between coastal habi-\ntat types and sediment Hg and OC contents (Fig. 1a and Methods). We \nfind that the sediment of VCEs not only is OC-rich, but also is a Hg-rich \nsystem (Fig. 1b), and that there is a strong and positive correlation \nbetween OC contents and Hg concentrations in coastal sediments \n(Fig. 1c). Notably, the Hg/OC ratio reported here (3.95 µg g–1) is ~2.5 \ntimes that of permafrost soils (1.6 µg g–1 (ref. 14)), indicating that coastal \nsediment is a more favourable environment for storing Hg under the \nsame OC sequestration potential as permafrost soils. The Hg release \nfrom anthropogenic activities around coastal regions could addition-\nally increase the Hg/soil OC ratio of coastal sediments9.\nWe combine the linear correlation between OC and Hg in all coastal \nsediments (N = 819) generated from our regional observational data \n(Fig. 1c) and the Blue C stock in the top metre of sediment in different \nBlue C ecosystems compiled by others to roughly estimate the Hg \nstock in global Blue C ecosystems (Methods). We find that over 95% \nof the Hg stock in global Blue C ecosystems is stored in VCEs (Fig. 1d). \nThe Hg stock in global Blue C ecosystems (mean = 73,162 Mg, mini-\nmum–maximum: 21,306–125,018 Mg) is ~15 times the Hg pool in the \nReceived: 15 May 2024\nAccepted: 28 October 2024\nPublished online: 29 November 2024\n Check for updates\n1Guangdong Provincial Key Laboratory of Marine Resources and Coastal Engineering, School of Marine Sciences, Sun Yat-sen University, Zhuhai, \nChina. 2School of Earth Sciences and Engineering, Sun Yat-sen University, Zhuhai, China. 3Guangdong Key Laboratory of Geological Process and \nMineral Resources Exploration, Zhuhai, China. 4Frontiers Science Center for Deep Ocean Multispheres and Earth System, and Key Laboratory of Marine \nChemistry Theory and Technology, Ministry of Education, and College of Chemistry and Chemical Engineering, Ocean University of China, Qingdao, \nChina. 5Research Center for Eco-Environmental Sciences, Chinese Academy of Sciences, Beijing, China. 6Department of Chemistry and Biochemistry and \nInstitute of Environment, Florida International University, Miami, FL, USA. \n e-mail: cizhj@mail.sysu.edu.cn\n\nNature Sustainability | Volume 7 | December 2024 | 1592–1595\n1593\nBrief Communication\nhttps://doi.org/10.1038/s41893-024-01471-y\nTM\n4,369 Mg\n3,405–5,532 Mg\n(5.79%)\nMF\n20,342 Mg\n7,505–33,180 Mg\n(27.80%)\nTF\n3,555 Mg\n(4.86%)\nSM\n44,895 Mg\n6,841–82,950 Mg\n(61.36%)\nTF\n(19)\nSM\n(10)\nTM\n(14)\nMF\n(15)\n0\n10\n20\n30\nSediment LOI\nSediment Hg\nSediment LOI (%)\n(181)\n(239)\n(137)\n(262)\n0\n200\n400\n600\nSediment Hg (ng g–1)\nTF\nTM\nSM\nMF\na\nb\nc\nd\n800\n600\n400\n200\n0\n0\n5\n10\n15\nSediment LOI (%)\nSediment Hg (ng g–1)\n20\n25\nSediment Hg (ng g–1) = 22.896 × LOI (%) + 21.96\nr2 = 0.62, P = 1.42 × 10\n–9\nTF\nSM\nTM\nMF\nFig. 1 | Estimate of Hg stock in global blue C ecosystems. a, Sampling sites \n(N = 58) in diverse coastal ecosystem locations of China. A jitter is used to offset \nclosely placed sites. b, Organic carbon contents (loss on ignition (LOI), a proxy for \norganic carbon) and Hg concentrations in the sediments of four coastal habitat \ntypes. Central box lines and squares show the median and mean, respectively; \nbox boundaries show the 25th and 75th percentiles; whiskers show 1.5 times the \ninterquartile range; triangles show the maximum and minimum. Text in brackets \nabove and below the panel show the number of sediment samples and sampling \nlocations in each coastal habitat type, respectively. c, Strong and positive linear \ncorrelation between OC and Hg in the sediments (N = 819) across four coastal \nhabitat types, P = 1.42 × 10–9 (two-sided Student’s t test). d, Hg stock in the top \nmetre of sediment in different Blue C ecosystems in the world. Data in d are \nreported as average and minimum–maximum (except tidal flat), and percentage \n(%) represents the relative contribution of each coastal habitat type to the \nHg stock in global blue carbon ecosystems. TF, tidal flat; TM, tidal marsh; SM, \nseagrass meadow; MF, mangrove forest. Data used to generate a–c are available \nin Supplementary Table 1. Basemap in a from Earthdata (https://www.earthdata.\nnasa.gov).\n–\n–\n–\nCO2 + Hg(0) + Hg(II)\nInput\nCO2, Hg(0), Hg(II)\nCO2, Hg(0), Hg(II)\nCO2, CH4, Hg(0)\nCO2, CH4, Hg(0)\nCO2, CH4, Hg(0)\nCO2, CH4, Hg(0)\nMMHg\nDOC, DIC, Hg, MMHg\nCO2\nSediment OC\nmineralization\nHg\nHg\nDOC\nDOC\nMMHg\nPOC\nCH4\nTidal marsh\nMangrove forest\nSeagrass meadow\nOutput\nCO2 + CH4 + Hg(0)\nFig. 2 | Close linkage between C and Hg biogeochemical cycles in coastal \necosystems. As a high primary production ecosystem, coastal vegetation \neffectively assimilates atmospheric CO2 into OC via photosynthesis, \naccompanied by the uptake of atmospheric Hg (Hg(0) and Hg(II)) into biomass, \nmainly in foliage20. Submerged canopies and fine root systems of VCEs greatly \nincrease sediment deposition and reduce sediment resuspension2,3. The \nlitterfall deposition and particles trapped from surrounding ecosystems lead to \nsynchronous and effective accumulation of OC and Hg in surface sediments2,13. \nThe water-logged anoxic environment favours the long-term storage of OC \nand Hg in the sedimentary system21. Both the sediments of VCEs and the \ncoastal water column are hotspots of MMHg production6–8,22, and the coastal \nseafood webs are hotspots of MMHg bioaccumulation, biomagnification and \nhuman exposure1,10. The release of C (including CO2, CH4, dissolved OC (DOC), \ndissolved inorganic C (DIC) and particulate OC (POC)) and Hg (including Hg(0), \ndissolved and particulate Hg and MMHg) to coastal waters via bioturbation and \nmultiple hydrological processes such as tidal pumping, submarine groundwater \ndischarge and diffusion and to the atmosphere via ebullition and diffusion \ngreatly affects the OC and Hg stock and influences the biogeochemistry and \ntoxicology of marine environments3,6,9,22. DOC is the dominant mobile carrier of \ndissolved Hg and MMHg and greatly regulates Hg speciation and bioavailability \nin coastal environments22. The sediment OC mineralization during early \ndiagenesis highlights that the upper part of coastal sediment (approximately \nBlue C sediment) is not a stable sink for C and Hg.\n\nNature Sustainability | Volume 7 | December 2024 | 1592–1595\n1594\nBrief Communication\nhttps://doi.org/10.1038/s41893-024-01471-y\npresent-day atmosphere (4,400 Mg (ref. 15)), ~25% of the Hg stored in \npresent-day ocean water (260,000 to 350,000 Mg) and ~80% of the Hg \nstored in the upper 1,000 m of ocean water (62,800 to 120,000 Mg) \n(refs. 16,17). Our methodology underestimates the Hg stock since we \ncalculate only the Hg stored in the top metre of sediment and do not \naccount for Hg stored in biomass, particularly in mangrove forests, \nwhich have large biomass.\nGlobal VCEs and tidal flats are estimated to annually bury \n80–220 Tg OC (ref. 18) and 6.8 Tg OC (ref. 4), respectively. Correspond-\ningly, 316 to 869 Mg (mean = 592 Mg) and 26.9 Mg Hg could be seques-\ntered by global VCEs and tidal flats annually, although the degradation \nof VCEs2 may lead to Hg release. This Hg burial flux is comparable to \n~25% of global anthropogenic Hg emissions of 2,500 Mg yr–1 (ref. 15), \nclose to half of the global river Hg discharge of 893–1,224 Mg yr–1 to \ncoastal seas19 and roughly equivalent to 70% of the annual oceanic Hg \nsequestration of 800 Mg (ref. 15).\nOur study provides field evidence of the co-accumulation of \nclimate-friendly OC and neurotoxic Hg in Blue C ecosystems. On the \nbasis of the strong commonality and close linkage between C and Hg \nstorage and cycling in coastal ecosystems (Fig. 2), by analogy to the \nterm ‘Blue C’, we propose the term ‘Blue Hg’ to refer to Hg captured \nand stored in coastal ecosystems, particularly in VCEs.\nOur study reveals that extremely limited extents of VCEs (~0.2% \nof the ocean surface18) and their substantial stock and burial flux of Hg \nhighlight the disproportionate importance of the VCEs to global Hg \nstorage and cycling. Blue Hg stock is therefore an important, dynamic, \nreactive but overlooked Hg pool in the global Hg cycle and health risk. \nBlue C ecosystems can provide a range of ecosystem services such as \nprotecting biodiversity and fisheries and increasing water quality. \nHowever, the co-accumulation and contrary ecological effects of OC \nand Hg in coastal ecosystems indicate that expanding Blue C ecosystems \nare very likely to add to the Blue Hg stock, consequently resulting in an \nincrease of Hg risk in global coastal seas (Fig. 2). Delay and failure to \nimplement the Blue C strategies may well accelerate the impending cli-\nmate catastrophe and other coastal and ecological disasters, inevitably \nresulting in substantial economic, ecological and human losses3. We \nrefer to increased Blue Hg stocks and the risk caused by expanding Blue \nC ecosystems to mitigate climate change and achieve co-benefits as the \n‘Blue Hg dilemma’. The close interactions between C and Hg in coastal \necosystems point to the importance of a holistic understanding of the \nbenefits, risks and trade-offs of Blue C sequestration strategies. We high-\nlight that the Blue Hg dilemma should be accounted for in future Blue \nC strategy assessments to inform policymakers and help them design \nbetter management and restoration policies across sectors and actors \nthat can balance economic, ecological, climatic and health interests for \nfuture generations, particularly in the context of the Paris Agreement \non Climate Change and the Minamata Convention on Mercury.\nMethods\nWe collected 819 sediment samples from 58 coastal wetlands, includ-\ning unvegetated (tidal flats, N = 19) and vegetated (seagrass meadows, \nN = 10; tidal marshes, N = 14; mangrove forests, N = 15) coastal ecosys-\ntems along Chinese coastlines from the cold temperate zone (~40° N) \nto the tropical zone (~19° N) (Fig. 1a) during 2010 to 2022. We measured \nsediment Hg concentrations using a direct Hg analyser (DMA–80, \nMilestone), following the US Environmental Protection Agency Method \n7473 (ref. 9). The limit of detection of the Hg analyser was 0.1 ng g−1. The \nperformance of the Hg analyser was evaluated by multiple measure-\nments of certified reference materials GBW07314 (marine sediment) \nand GBW07405 (soil). Results were not statistically different from the \ncertified values, with Hg concentrations of 47.1 ± 3.9 ng g−1 (1σ, N = 57, \ncertified value of 48 ± 12 ng g−1) for GBW07314 and 282.9 ± 12.5 ng g−1 (1σ, \nN = 16, certified value of 290 ± 40 ng g−1) for GBW07405. We measured \nLOI as the loss of weight of a sediment sample ignited at 550 °C for 8 h \nby a Muffle furnace9.\nWe multiplied a single general linear trend (Hg/OC ratio: \n3.95 µg g–1) between OC and Hg in all coastal sediments generated \nfrom our regional observational data (Fig. 1c) by the Blue C stock in \ndifferent Blue C ecosystems compiled by others to roughly estimate the \nHg stock in the top metre of sediment in global coastal environments \n(Blue Hg stock; Fig. 1d). The van Bemmelen factor (LOI/OC = 1.724) is \nused to convert sediment LOI to OC content. The Blue C stock in global \nmangrove forests (mean = 5,150 Tg OC, range = 1,900–8,400 Tg OC), \nseagrass meadows (mean = 11,366 Tg OC, range = 1,732–21,000 Tg OC) \nand tidal marshes (mean = 1,106 Tg OC, range = 862–1,350 Tg OC) was \nobtained from ref. 3 and tidal flats (900 Tg OC) from ref. 4. We applied \nthe identical method to estimate the Blue Hg fluxes. Since the huge \nuncertainties of current Blue C stocks and fluxes dominate the uncer-\ntainties of Blue Hg stock and flux estimates in the present study, we did \nnot consider the error propagation of sediment LOI and Hg determina-\ntion and the regression coefficient of OC and Hg in the calculation.\nReporting summary\nFurther information on research design is available in the Nature \nPortfolio Reporting Summary linked to this article.\nData availability\nThe data supporting the findings of this study are available within the \npaper and its Supplementary Table 1.\nReferences\n1. \nSchartup, A. T. et al. Climate change and overfishing increase \nneurotoxicant in marine predators. Nature 572, 648–650  \n(2019).\n2. \nDuarte, C. M., Losada, I. J., Hendriks, I. E., Mazarrasa, I. &  \nMarbà, N. The role of coastal plant communities for climate \nchange mitigation and adaptation. Nat. Clim. Change 3, 961–968 \n(2013).\n3. \nMacreadie, P. I., Costa, M. D. P., Atwood, T. B. & Friess, D. A. Blue \ncarbon as a natural climate solution. Nat. Rev. Earth Environ. 2, \n826–839 (2021).\n4. \nChen, Z. L. & Lee, S. Y. Tidal flats as a significant carbon reservoir \nin global coastal ecosystems. Front. Mar. Sci. 9, 900896 (2022).\n5. \nBoyd, P. W. et al. Potential negative effects of ocean afforestation \non offshore ecosystems. Nat. Ecol. Evol. 6, 675–683 (2022).\n6. \nBergamaschi, B. A. et al. Tidally driven export of dissolved \norganic carbon, total mercury and methylmercury from a \nmangrove-dominated estuary. Environ. Sci. Technol. 46, 1371–1378 \n(2012).\n7. \nMitchell, C. P. J. & Gilmour, C. C. Methylmercury production in \na Chesapeake Bay salt marsh. J. Geophys. Res. Biogeosci. 113, \nG00C04 (2008).\n8. \nLei, P., Zhong, H., Duan, D. & Pan, K. A review on mercury \nbiogeochemistry in mangrove sediments: hotspots of \nmethylmercury production? Sci. Total Environ. 680, 140–150 \n(2019).\n9. \nCi, Z., Yin, Y., Shen, W. & Chen, B. Non-conservative mixing \nbehaviors of mercury in subterranean estuary: coupling effect of \nhydrological and biogeochemical processes and implications for \nrapidly changing world. Water Res. 244, 120455 (2023).\n10. Médieu, A. et al. Evidence that Pacific tuna mercury levels are \ndriven by marine methylmercury production and anthropogenic \ninputs. Proc. Natl Acad. Sci. USA 119, e2113032119 (2022).\n11. \nFostier, A. H., Do, N., Costa, F. & Korn, M. D. G. A. Assessment \nof mercury contamination based on mercury distribution in \nsediment, macroalgae and seagrass in the Todos os Santos Bay, \nBahia, Brazil. Environ. Sci. Pollut. Res. 23, 19686–19695 (2016).\n12. Wolswijk, G. et al. Distribution of mercury in sediments, plant and \nanimal tissues in Matang Mangrove Forest Reserve, Malaysia.  \nJ. Hazard. Mater. 387, 121665 (2020).\n\nNature Sustainability | Volume 7 | December 2024 | 1592–1595\n1595\nBrief Communication\nhttps://doi.org/10.1038/s41893-024-01471-y\n13. Wang, Y., Wang, Z., Zheng, X. & Zhou, L. Influence of Spartina \nalterniflora invasion on mercury storage and methylation in the \nsediments of Yangtze River estuarine wetlands. Estuar. Coast. \nShelf Sci. 265, 107717 (2022).\n14. Schuster, P. F. et al. Permafrost stores a globally significant \namount of mercury. Geophys. Res. Lett. 45, 1463–1471 (2018).\n15. Outridge, P. M., Mason, R. P., Wang, F., Guerrero, S. & \nHeimbürger-Boavida, L. E. Updated global and oceanic mercury \nbudgets for the United Nations Global Mercury Assessment 2018. \nEnviron. Sci. Technol. 52, 11466–11477 (2018).\n16. Sunderland, E. M. & Mason, R. P. Human impacts on open ocean \nmercury concentrations. Glob. Biogeochem. Cycles 21, GB4022 \n(2007).\n17. Lamborg, C. H. et al. A global ocean inventory of anthropogenic \nmercury based on water column measurements. Nature 512, \n65–68 (2014).\n18. Spivak, A. C., Sanderman, J., Bowen, J. L., Canuel, E. A. & \nHopkinson, C. S. Global-change controls on soil-carbon \naccumulation and loss in coastal vegetated ecosystems.  \nNat. Geosci. 12, 685–692 (2019).\n19. Liu, M. et al. Rivers as the largest source of mercury to coastal \noceans worldwide. Nat. Geosci. 14, 672–677 (2021).\n20. Zhou, J., Obrist, D., Dastoor, A., Jiskra, M. & Ryjkov, A. Vegetation \nuptake of mercury and impacts on global cycling. Nat. Rev. Earth \nEnviron. 2, 269–284 (2021).\n21. Biswas, A., Blum, J. D., Bergquist, B. A., Keeler, G. J. & Xie, Z. \nNatural mercury isotope variation in coal deposits and organic \nsoils. Environ. Sci. Technol. 42, 8303–8309 (2008).\n22. Soerensen, A. L., Schartup, A. T., Skrobonja, A. & Björn, E. Organic \nmatter drives high interannual variability in methylmercury \nconcentrations in a subarctic coastal sea. Environ. Pollut. 229, \n531–538 (2017).\nAcknowledgements\nWe greatly appreciate Joel D. Blum of University of Michigan for \nproviding valuable comments on drafts of the paper. We thank \nYingying Liu, Zhilong Shi and Dongping Chen for sample collection \nand analysis, and Peng Wang and the volunteers who helped in the \nfield sampling. This study was supported by the National Natural \nScience Foundation of China (nos. 42477224 and 41773123 to  \nZ.C. and 21777198 to B.C.) and the Innovation Group Project of \nSouthern Marine Science and Engineering Guangdong Laboratory \n(Zhuhai) (no. 311022010 to W.S.).\nAuthor contributions\nZ.C. designed and performed research. W.S. and X.Z. contributed \nanalytic tools. Z.C., W.S. and B.C. acquired funding. Z.C. analysed  \ndata, generated figures and wrote the original draft of the paper.  \nAll co-authors contributed to the discussion, editing and approval of \nthe paper.\nCompeting interests\nThe authors declare no competing interests.\nAdditional information\nSupplementary information The online version contains \nsupplementary material available at  \nhttps://doi.org/10.1038/s41893-024-01471-y.\nCorrespondence and requests for materials should be addressed to \nZhijia Ci.\nPeer review information Nature Sustainability thanks Jean-Francois \nLapierre, Maodian Liu and Ke Pan for their contribution to the peer \nreview of this work.\nReprints and permissions information is available at  \nwww.nature.com/reprints.\nPublisher’s note Springer Nature remains neutral with regard to \njurisdictional claims in published maps and institutional affiliations.\nSpringer Nature or its licensor (e.g. a society or other partner) holds \nexclusive rights to this article under a publishing agreement with \nthe author(s) or other rightsholder(s); author self-archiving of the \naccepted manuscript version of this article is solely governed by the \nterms of such publishing agreement and applicable law.\n© The Author(s), under exclusive licence to Springer Nature Limited \n2024\n\n1\nnature portfolio  |  reporting summary\nApril 2023\nCorresponding author(s):\nZhijia Ci\nLast updated by author(s): Oct 18, 2024\nReporting Summary\nNature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency \nin reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.\nStatistics\nFor all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.\nn/a Confirmed\nThe exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement\nA statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly\nThe statistical test(s) used AND whether they are one- or two-sided \nOnly common tests should be described solely by name; describe more complex techniques in the Methods section.\nA description of all covariates tested\nA description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons\nA full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) \nAND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)\nFor null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted \nGive P values as exact values whenever suitable.\nFor Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings\nFor hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes\nEstimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated\nOur web collection on statistics for biologists contains articles on many of the points above.\nSoftware and code\nPolicy information about availability of computer code\nData collection\nData collection in this study does not rely on software.\nData analysis\nNo custom code or mathematical algorithm were used to perform the data analysis.\nFor manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and \nreviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.\nData\nPolicy information about availability of data\nAll manuscripts must include a data availability statement. This statement should provide the following information, where applicable: \n- Accession codes, unique identifiers, or web links for publicly available datasets \n- A description of any restrictions on data availability \n- For clinical datasets or third party data, please ensure that the statement adheres to our policy \n \nThe data supporting the findings of this paper are available within the paper and its Supporting Information file.\n\n2\nnature portfolio  |  reporting summary\nApril 2023\nResearch involving human participants, their data, or biological material\nPolicy information about studies with human participants or human data. See also policy information about sex, gender (identity/presentation), \nand sexual orientation and race, ethnicity and racism.\nReporting on sex and gender\nN/A\nReporting on race, ethnicity, or \nother socially relevant \ngroupings\nN/A\nPopulation characteristics\nN/A\nRecruitment\nN/A\nEthics oversight\nN/A\nNote that full information on the approval of the study protocol must also be provided in the manuscript.\nField-specific reporting\nPlease select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.\nLife sciences\nBehavioural & social sciences\n Ecological, evolutionary & environmental sciences\nFor a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf\nEcological, evolutionary & environmental sciences study design\nAll studies must disclose on these points even when the disclosure is negative.\nStudy description\nIn this study, we carried out a targeted survey along Chinese coastlines to document the relationship between coastal habit types \nand sediment Hg and OC contents. We then based on the correlation between OC and Hg in coastal sediments generated from our \nregional observational data and global blue carbon stock and flux data from literatures to estimate the Hg stock in the top meter \nsediment of global blue C ecosystems and annual burial flux of Hg to global blue C ecosystems. We also propose two terms, \"Blue Hg\" \nand \"Blue Hg dilemma\", to advance our understanding of the carbon-climate-Hg nexus.\nResearch sample\nSediment samples (n=819) collected from diverse coastal environments along Chinese coastlines were used in this study.\nSampling strategy\nSampling sites designed to encompass typical coastal habitats, including tidal flats, seagrass meadows, tidal marshes and mangrove \nforest.\nData collection\nData collection was mainly done by Zhijia Ci with the assistance of the co-workers and volunteers.\nTiming and spatial scale\nSediment samples were collected along Chinese coastlines from the cold temperate zone (~40 oN) to the tropical zone (~19 oN) \nduring 2010 to 2022.\nData exclusions\nNo data was excluded.\nReproducibility\nN/A. Well-established standardized methods were used for sample collections and analysis, as described in the Methods section.\nRandomization\nN/A. Our study is a field-based experimental design.\nBlinding\nN/A. Our study is a field based study with actual field data measurements.\nDid the study involve field work?\nYes\nNo\nField work, collection and transport\nField conditions\nSediment samples were collected from diverse coastal habitats, including unvegetated (tidal flats, n=19) and vegetated (seagrass \nmeadows, n=10; tidal marshes, n=14; mangrove forests, n=15) coastal ecosystems, under clam and slight ocean and weather \nconditions.\nLocation\nSediment samples were collected along Chinese coastlines from the cold temperate zone (~40 o) to the tropical zone (~19 oN).\n\n3\nnature portfolio  |  reporting summary\nApril 2023\nAccess & import/export\nAll sampling sites were publicly accessible during the sampling period of this study.\nDisturbance\nSediment samples were carefully collected by well-trained personnel or by volunteers under the guidance and/or supervision of well-\ntrained personnel. After sediment samples were collected, holes were backfilled if necessary, with minimal disturbance caused.\nReporting for specific materials, systems and methods\nWe require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, \nsystem or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. \nMaterials & experimental systems\nn/a Involved in the study\nAntibodies\nEukaryotic cell lines\nPalaeontology and archaeology\nAnimals and other organisms\nClinical data\nDual use research of concern\nPlants\nMethods\nn/a Involved in the study\nChIP-seq\nFlow cytometry\nMRI-based neuroimaging\nNovel plant genotypes\nDescribe the methods by which all novel plant genotypes were produced. This includes those generated by transgenic approaches, \ngene editing, chemical/radiation-based mutagenesis and hybridization. For transgenic lines, describe the transformation method, the \nnumber of independent lines analyzed and the generation upon which experiments were performed. For gene-edited lines, describe \nthe editor used, the endogenous sequence targeted for editing, the targeting guide RNA sequence (if applicable) and how the editor \nwas applied.\nSeed stocks\nReport on the source of all seed stocks or other plant material used. If applicable, state the seed stock centre and catalogue number. If \nplant specimens were collected from the field, describe the collection location, date and sampling procedures.\nAuthentication\nDescribe any authentication procedures for each seed stock used or novel genotype generated. Describe any experiments used to \nassess the effect of a mutation and, where applicable, how potential secondary effects (e.g. second site T-DNA insertions, mosiacism, \noff-target gene editing) were examined.\nPlants\n\n\n    Scientific Research Findings:", "answer": "This study focuses on the co-accumulation of sedimentary OC and Hg in coastal environments. We find that coastal sediment in VCEs is not only rich in OC, but also in Hg, and there is a strong and positive correlation between sediment OC and Hg across a variety of coastal habitats (Fig. 1a). The Hg to OC ratios in the coastal sediments is about 2.5 times that of permafrost soils, which harbour the largest stock of OC and Hg in terrestrial ecosystems. The Hg stored in blue carbon ecosystems ranges from 21,306 to 125,018 Mg with a mean of 73,162 Mg, and 95% of the Hg stock is stored in VCEs (Fig. 1b). The disproportionate contribution of the VCEs to global Hg storage, cycle and risk suggests that the carbon-Hg relationships, dynamics, feedbacks and trade-offs should be addressed in the carbon-focused NbS efforts.", "id": 84}