Skip to content

Multiple causal estimators and bootstrap inference

MultiEstimator runs several estimators on the same dataset in one pass — optionally with bootstrap standard errors and common-support filtering.

from CausalEstimate import MultiEstimator
from CausalEstimate.estimators import IPW, AIPW, TMLE

ipw = IPW(effect_type="ATE", treatment_col="treatment", outcome_col="outcome", ps_col="ps")
aipw = AIPW(
    effect_type="ATE",
    treatment_col="treatment",
    outcome_col="outcome",
    ps_col="ps",
    probas_t1_col="predicted_outcome_treated",
    probas_t0_col="predicted_outcome_control",
)
tmle = TMLE(
    effect_type="ATE",
    treatment_col="treatment",
    outcome_col="outcome",
    ps_col="ps",
    probas_t1_col="predicted_outcome_treated",
    probas_t0_col="predicted_outcome_control",
)

multi_estimator = MultiEstimator([ipw, aipw, tmle])

results = multi_estimator.compute_effects(
    df,
    n_bootstraps=50,                # > 1 triggers bootstrapping
    apply_common_support=True,
    common_support_threshold=0.05,
    return_bootstrap_samples=True,  # optionally return raw bootstrap estimates
)
print(results)

Output format

results is a dictionary keyed by estimator class name ("IPW", "AIPW", "TMLE"). With bootstrapping (n_bootstraps > 1), each entry contains:

  • "effect" — the mean effect across bootstrap samples
  • "std_err" — the standard deviation of the bootstrap estimates
  • "CI95_lower" / "CI95_upper" — the 95% confidence interval (percentile method)
  • "n_bootstraps" — the number of bootstrap samples
  • "bootstrap_samples" — the raw bootstrap estimates, if return_bootstrap_samples=True

Without bootstrapping (n_bootstraps=1), each entry is the estimator's own output with "n_bootstraps" set to 0. For IPW, AIPW and TMLE that includes analytic "std_err", "CI95_lower" and "CI95_upper" from the influence curve (Wald intervals, effect ± 1.96·SE; for risk ratios, the SE is on the log scale and the interval is exponentiated).

Common-support filtering

apply_common_support=True restricts estimation to the region where treated and control propensity distributions overlap, trimming observations outside the shared support (controlled by common_support_threshold). This guards against extrapolation into regions where one group has no comparable units. The underlying function is also available directly as CausalEstimate.filter_common_support — see the API Reference.