|
Abstract:
|
Statistical methods for estimating population average treatment effects have been established in observational studies. In real-world problems, treatment effects are likely to be different across different subpopulations, which is more challenging than the whole population problem. First, the indicators for subpopulations with differential effects may not be known ahead of time. Second, even with known subpopulation indicators, it is not easy to balance covariate distributions to remove confounding. Third, subgroup effect measures must be carefully selected. We propose a matching-based strategy that first identify the subpopulation structure via tree-based methods, then estimate the subpopulation effects with multiplicity adjustment. Two tree-based methods, Classification and Regression Tree (CART) and Causal Inference Tree (CIT), are applied in this study to identify the subgroups. We have run extensive simulation studies using pre-identified potential subpopulation indicators and propensity scores to build the tree. It turns out that CART performs slightly better than CIT and the use of categorized propensity scores improves the accuracy of subpopulation classification.
|
