Sophisticated common support selection for treatment analyses

This paper examines the role of the region of common support (RCS) in causal inference using propensity score–based methods. Although techniques such as propensity score matching and inverse-probability weighting rely critically on overlap between treated and control units, the construction and restriction of the RCS are often treated as secondary implementation details. We show that inadequate handling of the RCS can lead to biased estimates or substantial losses in external validity. To address this issue, we introduce comsup, a Stata command that provides a flexible framework for estimating treatment models, assessing overlap, and restricting the RCS using trimming and pruning procedures. The command integrates lasso-based model selection for propensity score estimation and allows users to incorporate outcome information for more parsimonious specifications. Using Monte Carlo simulations, we demonstrate that proper RCS handling can meaningfully improve estimator performance, but that the benefits depend on the degree of treatment effect heterogeneity. In particular, strict RCS restrictions may worsen performance when heterogeneous treatment effects are strong, highlighting an important trade-off in applied work.