Stabilizing Propensity Score Weighting Estimators: An Evaluation of Truncation, Trimming, and Shrinkage Approaches

Le résumé fourni par la source

Propensity score weighting is widely used to estimate the average treatment effect (ATE) in observational studies. Two common weighting estimators are inverse probability weighting (IPW) and weighted regression. Both methods can become unstable when estimated propensity scores approach 0 or 1, producing excessively large weights. To address this problem, researchers often process estimated propensity scores using truncation, trimming, or shrinkage. This paper reviews these approaches, including strategies for selecting the degree of processing, and evaluates their performance in two simulation studies. The results suggest that trimming may be less suitable for ATE estimation because it can induce substantial bias under treatment effect heterogeneity and reduce efficiency through sample loss. By contrast, the performance of truncation and shrinkage depends critically on the degree of processing. Sample-size-adjusted truncation provides the most stable performance across settings, whereas no single shrinkage level performs uniformly well. In weighted regression, propensity score processing reduces variance without adding bias when the outcome regression model is correctly specified, but this advantage disappears under misspecification. An empirical example illustrates the practical implications of these findings.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.