--- pretty_name: Statistical Inference Arena - Proximal Causal Inference language: - en tags: - tabular - synthetic - causal-inference - proximal-causal-inference - 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 · Proximal Causal Inference Given an observational study with **500 observations**, observed confounding proxies and one independent covariate query, estimate the **conditional average treatment effect** and return a two-sided 95% interval: $$\tau(x^*)=E[Y(1)-Y(0)\mid X=x^*].$$ Both classical estimators and tabular foundation models can use the inputs. [Arena leaderboard](https://statsinference.github.io/#domain=proximal) · [Official ProximalFM model](https://huggingface.co/christophemuller/ProximalFM) · [DGP equations and priors](DGP.md) This is a **data repository**, following the official-input/starter separation in the [mediation data release](https://huggingface.co/datasets/shibo769/StatsInference-Arena-Mediation). It does not contain a new trained model or claim official Arena endorsement. | What do you want to do? | Tables | Ground truth | Download | |---|---:|---|---| | Try observed-data loading, predictions and local scoring | 48 train, 16 dev, 32 test | Public | [Starter kit](proximal-starter-v1.zip) | | Evaluate a fixed method on the recorded benchmark inputs | 2,000 | Separate maintainer scoring materials | [Official inputs and submission CSV](proximal-official-inputs-v1.zip) | No historical ProximalFM pretraining, adapter-fitting or fine-tuning data are redistributed. The starter's `train` split is a small **new public illustration**, not data previously used to train the official model. The evaluated 2,000-table collection has **no ID/OOD split**. ## Target and observations Every official `tables/.csv` contains columns `X1,X2,Z,W,A,Y`: | Columns | Meaning | Type | |---|---|---| | X1, X2 | Baseline covariates defining the CATE subgroup | Continuous | | Z | Treatment-confounding proxy | Continuous | | W | Outcome-confounding proxy | Continuous | | A | Treatment | Binary 0/1 | | Y | Outcome | Continuous | `queries.csv` supplies one `X1,X2` query per `table_id`. Query Z/W/A/Y are not observed or required. U is latent and is never included in the observed inputs. This is a query-point population CATE, not ATT or a within-table average effect. The target averages over the natural `W|X=x*` distribution under the stated proximal assumptions. These assumptions require more than correlation between the proxies and the hidden confounder. ## Run the public starter Extract `proximal-starter-v1.zip` and run from its directory: ```sh python -m pip install numpy python baseline.py --tables dev/tables --queries dev/queries.csv --output dev_predictions.csv python score.py --truth dev/truth.csv --predictions dev_predictions.csv ``` The illustrative baseline is a linear S-learner on X/A, with treatment-by-X interactions and B=50 row-bootstrap percentile intervals. It **ignores proxies** and has no proximal identification guarantee. It is not the evaluated RF/GMM comparator. On the 16 development tables its saved example has bias **0.637534**, RMSE **1.024240**, coverage **62.5%**, and mean length **1.911591**. Those are interface/scoring illustrations, not leaderboard results or tuned performance. Replace that estimator, retain the four output columns and use the scorer. Tune on `train`/`dev`; keep the public `test` split for a final local check. These 96 fresh tables, their parameter draws and queries are separate from the frozen 2,000 official tables. Each starter configuration contributes **one table**. ## Sampling and file structure **Public starter.** A family is sampled uniformly from Cui linear bridge, Qi heterogeneous linear, Qi quadratic and Qi trigonometric; then source scenario and continuous parameters are drawn, followed by 500 observations and an independent query. Train/dev/test use distinct split-specific seed namespaces. There is one new parameterized mechanism per table, not paired replicates. `catalog/{train,dev,test}.csv` contains public labels and table paths within the kit. The Dataset Viewer calls `dev` the `validation` split. A generator is included for independent illustrations. [Equations and sources](DGP.md). **Official inputs.** The collection preserves all 2,000 frozen tasks evaluated on 2026-11-16, n=500. Family counts are 496, 516, 510 and 478 in the order above. Each task has a separate continuously sampled configuration and query. Unlike the mediation release's paired design, this collection does **not** contain 1,000 configurations with two repeated tables. Its error variance is across tasks, not within-configuration sampling variance. The archive provides observed CSVs, queries, variable types, the saved original five treatment-stratified folds and a blank prediction template. Opaque IDs replace internal task indices. Row order, query values and original fold labels are preserved. CSV uses round-trip float64 precision, without standardization, trimming, rounding to float32 or changing the observations. Direct table-level official true effects, generator assignments, parameters, seeds, original IDs and reference model outputs are **not included in this public input release**. They are retained in a separate maintainer package. This is a release boundary, **not a cryptographic secrecy guarantee**: anyone with other matching labeled artifacts or generation provenance may be able to link or reconstruct tasks. Published aggregate results and analytic equations are not withdrawn. Do not train or tune on references obtained elsewhere and claim an unseen held-out evaluation. ## Evaluate and return predictions Freeze your method/settings before evaluating the official inputs. If your method uses cross-fitting or context partitions, use `folds/.csv` aligned to the observed rows. Each contains one `fold` column, labels 0-4, 100 observations per fold. Fold labels are evaluator metadata, not features. Bootstrap refits must apply the declared resplitting procedure; original labels are not prescribed as labels for every resample. Fill `predictions_template.csv` with exactly one row per official table: ```csv table_id,estimate,ci_lower,ci_upper ``` All numeric values must be finite and `ci_lower <= ci_upper`. Return effects in **original outcome units**. The scoring reference uses the same query for all methods on each task. Maintainers can compute signed Bias, sample error variance with denominator 1,999, RMSE, 95% interval coverage and mean length. They can separately standardize per-task error/length by the original table's sample SD(Y); normalized and raw metrics must not be mixed. Runtime is not inferred from the four-column file: submit hardware and warm point-plus-interval timing, including all RF bootstrap work, separately when available. Uploading this dataset does not submit a method or automatically update the leaderboard. Share the resulting dataset URL with the Arena maintainers. No private scoring files need to be uploaded or posted in a public repository. ## Verification Run `python verify_release.py` in a complete download of this repository. It checks SHA256SUMS, both archives, all 2,000 input tables and folds, 96 labeled starter tables and template IDs. No model or official truth is required. The original formal evaluation inputs/results are unchanged; only public file format, table IDs and release organization are adapted. No new official-model evaluation or recalibration was performed for this data release. ## References and scope - Cui et al. (2024). *Semiparametric Proximal Causal Inference*. [DOI](https://doi.org/10.1080/01621459.2023.2191817), §5.1, equations (13)-(14). - Qi et al. (2024). *Proximal Learning for Individualized Treatment Regimes Under Unmeasured Confounding*. [DOI](https://doi.org/10.1080/01621459.2022.2147841), §6.1, Tables 3-4. - Muller et al. (2026). *ProximalFM: Amortized Proximal Causal Inference under Hidden Confounding*. [Paper](https://arxiv.org/abs/2610.08078v2) · [Official code](https://github.com/ChristopheMuller/ProximalFM). Four mechanism families share a Gaussian proxy skeleton and are explicit continuous-prior adaptations, not complete reproductions of all source tasks. Original scientific-review status remains `needs_scientific_review`; finite numerical and engineering checks do not prove universal identification or interval calibration. No license for third-party models, papers or code is inferred from the mediation example; their original terms remain applicable.