A1506
Title: Semiparametric efficiency and flexible estimation of weighted average treatment effects under two-phase sampling
Authors: Kazuharu Harada - Tokyo Medical University (Japan) [presenting]
Masataka Taguri - Tokyo Medical University (Japan)
Abstract: Two-phase sampling is widely used in observational studies when important covariates are expensive to collect, but valid causal inference requires careful adjustment for the sampling design. Weighted average treatment effects under two-phase sampling are studied, including the average treatment effect, the average treatment effect on the treated or untreated, and the average treatment effect in the overlap population. The semiparametric efficiency bound for this broad class of target parameters is first derived and it is shown how it decomposes into the full-data efficiency component and an additional term induced by phase-2 subsampling. Several consistent estimators are then considered with focus on enriched estimators that incorporate information available for all phase-1 individuals. In particular, the enriched doubly robust estimator is shown to attain the semiparametric efficiency bound under suitable conditions, clarifying when phase-1 information can substantially improve efficiency. Design implications are also discussed, including how the benefit of enrichment depends on the phase-1 variables and the sampling mechanism. Finally, a data-adaptive implementation based on cross-fitting and double machine learning is outlined, which enables flexible nuisance estimation while preserving the main efficiency and robustness properties. The results provide practical guidance for both the design and analysis of observational studies with two-phase sampling.
