Confounder selection strategies targeting stable treatment effect estimators

摘要: Inferring the causal effect of a treatment on an outcome in observational study requires adjusting for observed baseline confounders to avoid bias. However, all covariates, when only subset are interest, is known yield potentially inefficient and unstable estimators effect. Furthermore, it raises risk finite-sample bias due model misspecification. For these stated reasons, confounder (or covariate) selection commonly used determine available covariates that sufficient confounding adjustment. In this article, we propose strategy focuses stable estimation particular, propensity score already includes adjust confounding, then addition associated with either or alone, but not both, should systematically change estimator. The proposal, therefore, entails first prioritizing inclusion model, using change-in-estimate approach select smallest adjustment set yields estimate. ability proposal correctly confounders, ensure valid inference following data-driven covariate selection, assessed empirically compared existing methods simulation studies. We demonstrate procedure three different publicly datasets inference.