API reference

One page per public class and function, generated from docstrings via sphinx.ext.autosummary. Each ## section below is a real import path (metacausal, metacausal.aggregation, …); each item is listed once, under the module it’s defined in.

Many aggregation-strategy and adapter classes (e.g. Median, CausalMLAdapter) are re-exported at the top level for convenience – from metacausal import Median works even though it’s documented under metacausal.aggregation below. See the Mixed-framework method list or metacausal.__all__ for the exact top-level surface.

metacausal

Top-level namespace: the ensemble class, its estimate/result types, and outcome-type detection.

metacausal.CausalEnsemble	Ensemble of causal estimators with robust aggregation.
metacausal.AteEstimate	Result from the ensemble of causal estimators.
metacausal.CateEstimate	Ensemble CATE result.
metacausal.ComponentAteEstimate	Result from a single causal estimator.
metacausal.ComponentCateEstimate	CATE result from a single component method.
metacausal.infer_outcome_type	Detect the outcome type from the values of Y.

Warnings

metacausal.MetaCausalWarning	Umbrella class for all metacausal-emitted warnings.
metacausal.ComponentWarning	Base class for warnings indicating an operating-ensemble pool reduction.
metacausal.ComponentFailureWarning	A component raised an exception during fit, cross-fitting, prediction, or as part of nuisance/weight-fitting machinery, and was silently excluded from the operating ensemble.
metacausal.ComponentExclusionWarning	A component was excluded by configuration mismatch.
metacausal.BootstrapWarning	A bootstrap inference call hit a reliability concern.

metacausal.aggregation

Aggregation strategies (combine component ATE/CATE predictions into an ensemble estimate) and the result types they produce.

metacausal.aggregation.AggregationStrategy	Abstract base class for all aggregation strategies.
metacausal.aggregation.PointwiseStrategy	Base class for strategies that apply a fixed statistical rule pointwise.
metacausal.aggregation.AgreementStrategy	Base class for strategies that select/weight models by inter-model agreement.
metacausal.aggregation.SupervisedStrategy	Base class for outcome-supervised aggregation strategies.
metacausal.aggregation.Mean	Pointwise mean aggregation.
metacausal.aggregation.Median	Pointwise median aggregation.
metacausal.aggregation.TrimmedMean	Pointwise trimmed-mean aggregation.
metacausal.aggregation.CBA	Consensus Based Averaging (Machluf et al., 2024).
metacausal.aggregation.CausalStacking	Causal Stacking (Han & Wu, 2022).
metacausal.aggregation.QAggregation	Doubly Robust Q-Aggregation (Lan & Syrgkanis, 2024).
metacausal.aggregation.RStacking	R-Stacking (Nie & Wager, 2021).
metacausal.aggregation.Select	Select the single component model minimizing a pseudo-outcome risk.
metacausal.aggregation.EnsembleWeights	Result of weight computation for any non-pointwise aggregation.
metacausal.aggregation.BootstrapResult	Bootstrap inference results for both ATE and CATE.

Cross-fitting

metacausal.aggregation.CrossFitSplit	Q-fold cross-fitting split.
metacausal.aggregation.TrainAvgSplit	Simple train / averaging-set split.
metacausal.aggregation.FoldSpec	Result of a data split: absolute indices into the original array per fold.

Nuisance estimation

metacausal.aggregation.NuisanceEstimates	Out-of-fold nuisance model predictions.
metacausal.aggregation.fit_nuisance	Fit nuisance models via cross-fitting and return out-of-fold predictions.
metacausal.aggregation.dr_pseudo_outcome	Compute the DR/AIPW pseudo-outcome for each observation.
metacausal.aggregation.robinson_residuals	Compute Robinson decomposition residuals.

metacausal.adapters

Wrappers presenting each supported library’s estimators (EconML, CausalML, DoubleML, stochtree) through one common interface, plus generic adapters for wrapping arbitrary user-supplied estimators.

metacausal.adapters.CausalEstimator	Protocol that all causal estimator adapters must satisfy.
metacausal.adapters.EconMLAdapter	Wrap an initialized-but-unfitted EconML estimator.
metacausal.adapters.CausalMLAdapter	Wrap a CausalML estimator as a metacausal causal estimator.
metacausal.adapters.DoubleMLAdapter	Wrap a DoubleML estimator class for use in CausalEnsemble.
metacausal.adapters.StochtreeAdapter	Wrap stochtree's BCFModel for use in CausalEnsemble.
metacausal.adapters.GenericATEAdapter	Wrap a callable fn(X, T, Y) -> float as an ATE-only causal estimator.
metacausal.adapters.GenericCATEAdapter	Wrap callables as a CATE-capable causal estimator.
metacausal.adapters.INNER_WORKER_ENV	Sentinel env var set inside every MetaCausal parallel worker.

metacausal.plots

Requires the plots extra (pip install "metacausal[plots]"). Each function is also available as a thin method on the class it plots – see the corresponding class above (e.g. BootstrapResult.forest()).

metacausal.plots.forest	Forest plot of per-component and ensemble ATEs.
metacausal.plots.weights	Bar chart of aggregation weights.
metacausal.plots.cate_profile	Ensemble CATE along one covariate, with optional bootstrap CI.
metacausal.plots.disagreement	Pairwise disagreement between component CATEs evaluated on X.

metacausal.datasets

metacausal.datasets.load_lalonde

Load the Lalonde job training dataset.