xcausal.defaults.default_methods

xcausal.defaults.default_methods(outcome_type='continuous')[source]

Return a default set of causal estimators for the given outcome type.

Requires: econml, doubleml, causalml, stochtree, sklearn. Install via pip install xcausal[all] for the full default library.

The continuous pool spans EconML, DoubleML, CausalML, and stochtree. The binary pool targets the risk-difference estimand and uses probability-capable outcome models, with regressors in independent effect-model slots. The generated component reference lists the exact rosters, counts, model recipes, internal folds, and clipping settings, including inherited upstream defaults.

Configure member models through CausalEnsemble(methods=[...]). EconML DR’s model_final defaults to an unregularized linear effect stage, while X’s cate_models=None inherits fresh clones of the outcome-model family. CausalML R’s effect_learner is independent of its outcome and propensity learners. Supervised aggregation’s nuisance arguments configure only weight learning.

Components dropped on binary:

  • DoubleMLPLR: IRM is the canonical DoubleML choice for binary Y; PLR adds an awkward linear-probability fit without diversity.

  • EconML T/X-Learners: no discrete_outcome flag, .predict() would return hard labels. The CausalML Base{T,X}Classifier siblings stand in.

  • BaseRRegressor: replaced by BaseRClassifier.

  • BaseDRClassifier: like BaseDRRegressor, needs an externally computed propensity passed via fit(..., p=...); without it, predictions are on a degenerate scale. EconML’s DRLearner covers the DR family in the binary pool. Users who want it can wrap it explicitly with their own propensity wiring.

  • Stochtree BCF: Gaussian outcome model only.

Parameters:

outcome_type (str)

Return type:

list