New Evidence on the Finite Sample Properties of Propensity Score Reweighting and Matching Estimators

TLDR

Frölich (2004) compared finite‑sample properties of reweighting and matching estimators of average treatment effects and found reweighting to perform far worse than even the simplest matching estimator. This paper argues that Frölich’s conclusion is unjustified and investigates empirical and hypothetical data‑generating processes, including misspecification effects. The authors extend Frölich’s analysis by evaluating both empirical and hypothetical DGPs and examining how misspecification influences reweighting and matching performance. The study finds that neither approach uniformly dominates; reweighting competes with the best matching estimators when overlap is good, whereas matching may be more effective when overlap is sufficiently poor.

Abstract

Frölich (2004) compares the finite sample properties of reweighting and matching estimators of average treatment effects and concludes that reweighting performs far worse than even the simplest matching estimator. We argue that this conclusion is unjustified. Neither approach dominates the other uniformly across data-generating processes (DGPs). Expanding on Frölich's analysis, this paper analyzes empirical as well as hypothetical DGPs and also examines the effect of misspecification. We conclude that reweighting is competitive with the most effective matching estimators when overlap is good, but that matching may be more effective when overlap is sufficiently poor.