--- pretty_name: Statistical Inference Arena — ATE and CATE language: - en tags: - tabular - synthetic - causal-inference - ate - cate - benchmark configs: - config_name: starter_catalog default: true data_files: - split: train path: catalog/train.csv - split: validation path: catalog/dev.csv - split: test path: catalog/test.csv - config_name: official_inputs data_files: - split: test path: official_schema.csv --- # Statistical Inference Arena · ATE and CATE Estimate the population average treatment effect (ATE) and the conditional average treatment effect (CATE) at a supplied query point from a simulated study of 500 observations. For each target, return a point estimate and a two-sided 95% interval. The collection supports evaluation of classical estimators and tabular foundation models. [Release v1.2.2](https://huggingface.co/datasets/fuyao11/StatsInference-Arena-ATE-CATE/tree/v1.2.2) | What do you want to do? | Tables | Ground truth | Download | |---|---:|---|---| | Train, validate and test a method with local scoring | 48 train, 16 dev, 32 test | Public | [Starter kit](ate-cate-starter-v1.zip) | | Evaluate a method on the official test | 2,000 | Maintainers score predictions | [Test inputs and prediction templates](ate-cate-official-inputs-v1.zip) | ## Target and observations $$ \mathrm{ATE}=E[Y(1)-Y(0)],\qquad \mathrm{CATE}(x_\star)=E[Y(1)-Y(0)\mid X=x_\star]. $$ ATE averages over the population of a study's generating mechanism. CATE is the conditional mean effect at the supplied query. Inputs are baseline covariates $X$, binary treatment $T$ and outcome $Y$, together with $x_\star$ for CATE. Effects are measured in outcome units; binary outcomes use probability differences. Each study contributes one estimate for each target. ## Run your first study comparison Unzip the starter kit, open its directory and run: ```sh python -m pip install numpy python baseline.py --tables dev/tables --task ATE --output dev_predictions_ate.csv python score.py --truth dev/truth_ate.csv --predictions dev_predictions_ate.csv --timings dev_predictions_ate.csv.timings.json python baseline.py --tables dev/tables --task CATE --output dev_predictions_cate.csv python score.py --truth dev/truth_cate.csv --predictions dev_predictions_cate.csv --timings dev_predictions_cate.csv.timings.json ``` The example uses a linear T-learner: fit a separate outcome regression in each treatment arm, then contrast their predictions. ATE averages over the study's covariates; CATE uses the supplied query. Percentile intervals come from 200 whole-row bootstrap refits. Across the 16 development studies, ATE gives bias 0.2270, RMSE 0.2382, coverage 31.25% and mean interval length 0.3234. For CATE, the values are bias 0.1992, RMSE 0.3013, coverage 62.50% and mean interval length 0.7601. Example predictions, timing files and all six metrics are included in the kit. To compare another estimator, produce the same prediction columns and run the scorer. Choose settings on `dev/`, then evaluate the public `test/` split. ## Sampling and files **Public starter.** Two designs are adapted: Nie–Wager Setup A has five continuous covariates and a continuous outcome; the Benkeser et al. design has two covariates and a binary outcome. Train, dev and test have disjoint parameter configurations. Each configuration generates two independent studies and shares one CATE query. The kit includes the generator and public ATE/CATE labels. Its `starter_catalog` lists the table ID, preset, configuration, replicate, sample size, `truth_ate`, `truth_cate`, observed CSV path and query path. The `dev/` directory appears as `validation` in the viewer. [Mechanisms and sources](DGP.md). **Official test.** The 2,000 studies cover 1,000 configurations across 13 source families, with two draws per configuration. Each study has an opaque ID, 500 observations, 2–10 covariates and a continuous or binary outcome. The download contains observed CSVs (`tables/`), CATE queries (`queries/`), variable types (`official_schema.csv`), five treatment-stratified folds for each target (`folds/ATE/` and `folds/CATE/`), and two blank prediction templates. Columns are `X1` through `Xp`, `T` and `Y`. The `official_inputs` viewer shows table IDs, sample sizes, covariate counts and variable types. Generation parameters, replicate pairings and effect labels are excluded from this input download; earlier public releases included the labels. ## Submit a method Fix the estimator and settings before running the official test. When using cross-fitting, follow the target's supplied folds and preserve observation order. Complete `predictions_ate_template.csv` and `predictions_cate_template.csv`, each with the 2,000 table IDs and these columns: ```csv table_id,estimate,ci_lower,ci_upper ``` Use finite numeric values and ensure `ci_lower <= ci_upper`. Send both completed files to the benchmark maintainer for scoring against population ATE and fixed-query CATE. Bias, RMSE, coverage and mean interval length use equal study weights. Variance is computed within paired configurations; inference time is reported when supplied. Timing includes point and interval computation, including bootstrap, and excludes model loading and warm-up. Report the method, interval procedure, data exposure and hardware alongside predictions. ## References - Balazadeh et al. *CausalPFN: Amortized Causal Effect Estimation via In-Context Learning*. [Paper](https://arxiv.org/abs/2506.07918) · [Official checkpoint](https://huggingface.co/vdblm/causalpfn). - Nie and Wager. *Quasi-oracle estimation of heterogeneous treatment effects*. [Paper](https://doi.org/10.1093/biomet/asaa076). - Wager and Athey. *Estimation and Inference of Heterogeneous Treatment Effects using Random Forests*. [Paper](https://doi.org/10.1080/01621459.2017.1319839). - Athey, Tibshirani and Wager. *Generalized Random Forests*. [Paper](https://doi.org/10.1214/18-AOS1709). - Ye, Shao, Yi and Zhao. *Toward better practice of covariate adjustment in analyzing randomized clinical trials*. [Paper](https://doi.org/10.1080/01621459.2022.2049278). - Benkeser, Carone, van der Laan and Gilbert. *Doubly robust nonparametric inference on the average treatment effect*. [Paper](https://doi.org/10.1093/biomet/asx053).