File size: 5,183 Bytes
269e92b 54ea286 269e92b 54ea286 da4347e 54ea286 da4347e 54ea286 da4347e 269e92b da4347e 54ea286 da4347e 269e92b da4347e 54ea286 da4347e 54ea286 269e92b 54ea286 269e92b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 | ---
license: mit
pretty_name: populace-us
tags:
- policyengine
- microsimulation
- synthetic-population
- tax-benefit
- united-states
---
# populace-us
The **populace-built US population**: a calibrated synthetic microdataset for
[PolicyEngine-US](https://github.com/PolicyEngine/policyengine-us), built by the
[`populace`](https://github.com/PolicyEngine/populace) stack **entirely from
primary sources** — the enhanced CPS appears only as the benchmark this file is
scored against, never as a build input. It loads anywhere the enhanced CPS
loads (an API-compatible alternative population), with its own calibrated
weights — and its own strengths and gaps, both documented below.
## Load it
```bash
pip install 'populace-data[us]'
```
```python
from policyengine_us import Microsimulation
from populace.data import load
sim = Microsimulation(dataset=load("us", 2024))
sim.calculate("household_net_income", 2024).sum()
```
Or grab the H5 directly:
```python
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="policyengine/populace-us",
filename="populace_us_2024.h5",
repo_type="dataset",
)
```
## How it is built
One HDF5 `USSingleYearDataset` per year. Every layer comes from a primary
survey or administrative source:
| source | provides |
| --- | --- |
| Census CPS ASEC | household structure, demographics, incomes, benefits, tenure, hours, occupation flags, health coverage at interview, retirement distributions (DST codes), childcare, prior-year income (longitudinal PERIDNUM join) |
| IRS SOI Public Use File 2015 (uprated) | tax detail: capital gains, dividends, interest, itemized-deduction inputs, QBI/SSTB components, partnership self-employment, estates, tuition |
| Fed SCF 2022 | wealth: bank/stock/bond assets, net worth, mortgage balance hints |
| Census SIPP | tip income for tipped occupations; household vehicles (count and value) |
| CPS-ORG | hourly wage, paid-hourly status, union coverage |
| MEPS-IC parameters | employer-sponsored insurance premiums |
| Census ACS 2022 | rent for renter households |
Imputations use weight-aware quantile-forest models fit on each donor's own
records, and every imputed value is clipped to **that donor's** realized range
(the support guard) — nothing is anchored to the enhanced CPS. The result is
calibrated to PolicyEngine's administrative target surface (3,704 IRS/Census/
program targets, **plus a signed net short-term capital gains target** so the
optimizer cannot silently drive a net-negative aggregate to extremes) with a
hard per-record weight bound (`max_weight_ratio=50`), so no aggregate leans on
a handful of super-weighted records.
## Acceptance gates
The build refuses to publish unless every gate passes; this file passed all
of them:
- **Parity 0**: every PolicyEngine input layer the enhanced CPS populates
non-degenerately, this file's simulation populates (169 reference layers
checked at simulation level).
- **Exported-nonzero**: all 308 stored columns carry signal — no all-zero
scaffolding that would silently mask engine formulas or defaults.
- **Calibration**: 94.66% of 3,704 targets within 10% (loss 0.022); max
household weight 379,623 with **zero records above 500k** (the enhanced CPS
ships 21, max 1.05M).
- **Smoke aggregates** through `Microsimulation`: 332.8M people, $98.0B SNAP,
$175.5T net worth (Fed Z.1 ≈ $169T), net short-term capital gains
**−$77.4B** against the −$76.8B PUF-anchored target, tips $53.1B, rent
$759.7B.
## Validation
Scored by the sound comparison — matched samples (41,314 households),
symmetric weight refit on the full administrative target surface, held-out
targets never seen by either side's refit. Lower is better.
| metric | populace-us | enhanced CPS |
| --- | --- | --- |
| training loss (2,965 targets) | **0.176** | 1.089 |
| held-out loss (739 unseen targets) | **0.037** | 0.317 |
| full-surface loss (3,704 targets) | **0.213** | 1.406 |
Per individual target the incumbent still wins more often (2,528 of 3,704 to our 1,127, 49 ties): populace wins big where it wins and loses narrowly where it loses. Both facts are the story.
## Known gaps
We publish the misses with the hits:
- **Net worth runs ~4% above Fed Z.1** ($175.5T vs ≈ $169T): the calibration
target ($160T) sits below Z.1 and the achieved total lands between them.
- **Investment interest expense is thin** ($7.2B against IRS SOI ≈ $24B):
the PUF-residual rule populates the layer conservatively; a dedicated SOI
calibration target is the roadmap item.
- **Per-target wins vs the incumbent**: see Validation — aggregate losses
are what the comparison gates on, but per-target patterns differ between
the two populations. Results are not interchangeable.
The dashboard at [populace.dev/dashboard](https://populace.dev/dashboard)
shows the full per-family calibration fit, the worst-fit targets by name, the
weight distribution, and a live strip while a build chain runs. Methodology
and evidence: [populace.dev](https://populace.dev); loader and registry:
[github.com/PolicyEngine/populace](https://github.com/PolicyEngine/populace)
(`packages/populace-data`).
|