causal-treatment-models

Run the training CLI with a CSV containing subject, binary treatment, outcome, and pretreatment covariate columns, selecting a model such as doubly_robust or causal_forest. It produces cross-fitted held-out CATE predictions, fold assignments, and metrics in the output directory.

The S-learner fits one outcome model with treatment as a feature, the T-learner fits separate models for treated and control groups, and the X-learner imputes individual effects and blends them using propensity scores. All three are available as model options in this CLI.

Yes, the causal_forest model requires the optional econml package, which provides the CausalForestDML backend. Install it with pip install econml; the other estimators only need numpy, pandas, scipy, scikit-learn, joblib, and torch.

No, the current CLI only supports binary 0/1 treatments. Multi-arm or continuous treatment regimes are not supported, and only pretreatment covariates should be included in the feature matrix.

Causal interpretation requires consistency, positivity, no unmeasured confounding, and defensible temporal ordering, none of which predictive accuracy establishes. Negative controls, sensitivity analyses, and randomized validation remain separate requirements.