antecedent-estimate

Estimators for identified causal functionals. Estimators consume an IdentifiedEstimand — they never choose confounders or assert identifiability.

Frequentist surface: linear and GLM adjustment, g-computation, propensity methods (IPW, matching, AIPW), instrumental variables (Wald, 2SLS), front-door two-stage, sharp regression discontinuity, and temporal adjustment, mediation, and prediction. Continuous causal responses (Kennedy-style doubly robust curves, derivatives, elasticities) live in response. Bayesian estimation covers g-computation, HMC GLMs, prior transfer, and graph-by-effect posterior envelopes.

Observation mechanisms (observation), trial-to-target transport (transport), and randomized interference (interference) are explicit stage APIs: they change what identifies the estimand and are never inferred from the data.

Standard errors are analytic or bootstrap; overlap diagnostics and clipping policies are reported, not silent. See docs/capabilities.md for the full estimator inventory.