---

name: causal-estimation description: "Classical CausalML causal estimators covering meta-learners, TMLE, IV/DRIV, serialization, and estimator API contracts." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

causal-estimation

Use this sub-skill for classical causal effect estimation with CausalML 0.17.0: S/T/X/R/DR meta-learners, TMLE ATE estimation, instrumental-variable estimators, propensity-score contracts, argument-order migration, DataFrame/Polars handling, and joblib-backed learner serialization.

Route by question type

• Meta-learners and CATE/ATE: use references/meta-learners.md. This includes BaseSRegressor, BaseSClassifier, BaseTRegressor, BaseTClassifier, BaseXRegressor, BaseXClassifier, BaseRRegressor, BaseRClassifier, BaseDRRegressor, BaseDRClassifier, LRSRegressor, XGBTRegressor, XGBTClassifier, MLPTRegressor, XGBRRegressor, XGBRClassifier, and XGBDRRegressor.
• TMLE, IV, DRIV, and persistence: use references/iv-tmle-serialization.md. The current IV entry point is IVRegressor; the current DRIV entry points are BaseDRIVLearner, BaseDRIVRegressor, and XGBDRIVRegressor.
• Input contracts and migration details: use references/api-contracts.md before writing code that passes treatment, y, p, assignment, w, or pZ.
• Error diagnosis: use references/troubleshooting.md.

Cross-route to sibling sub-skills

• Use ../tree-models/ for CausalTreeRegressor, CausalRandomForestRegressor, UpliftTreeClassifier, UpliftRandomForestClassifier, tree fill/prune, and tree visualization.
• Use ../deep-models/ for TensorFlow, Torch/Pyro, or JAX neural estimators such as DragonNet and CEVAE.
• Use ../analysis-and-decision/ for AUUC/Qini/gain curves, validation scoring, sensitivity analysis, feature selection, policy learning, and counterfactual value optimization.
• Use ../data-preparation/ for synthetic data, propensity-model construction, matching, feature utilities, and matched-data QA.

Always apply these defaults

1. Prefer keyword arguments for all public estimator calls: fit(X=X, treatment=treatment, y=y, ...), not positional fit(X, treatment, y, ...).
2. Check control_name and treatment labels before fitting. Output columns are ordered as learner.t_groups, which is the sorted set of non-control treatment groups.
3. Treat p as a strict probability contract: values must be inside (0, 1), arrays/Series are only for a single non-control group, and multi-treatment runs need dictionaries keyed by treatment group.
4. Use fit_predict only on learners that expose it. TMLELearner exposes estimate_ate but not fit_predict; IVRegressor exposes fit and predict; DRIV learners expose both fit_predict and estimate_ate.
5. Keep persistence on fitted learners: learner.save(path), ClassName.load(path), or generic load_learner(path). Do not assume save_model or load_model functions exist for these classical estimators.