A1504
Title: Average treatment effect estimation under poor overlap via weighted estimands
Authors: Shunichiro Orihara - Tokyo Medical University (Japan) [presenting]
Sho Komukai - Tokyo Medical University (Japan)
Abstract: The average treatment effect (ATE) is commonly considered as the causal effect of interest. In situations with poor overlap, where the covariate support between the treatment and control groups is limited, the inverse probability weighting (IPW) estimator may have large variance. To address this issue, weighted ATE (WATE) estimands, such as the ATE for the overlap population (ATO), are sometimes considered. However, these estimands may differ from the original target of interest such as the ATE. A novel estimation procedure is proposed that uses a class of WATEs based on the beta weight family, including ATO, and estimates the ATE via extrapolation. Specifically, the following steps are considered: 1) estimating IPW estimators for several WATE estimands; 2) regressing the IPW estimates on the hyperparameter of the WATEs; and 3) setting the hyperparameter to 0 (extrapolation). Theoretical justifications for the procedure and its properties, such as asymptotic normality, are discussed. Additionally, since the proposed procedure is easy to implement, an explanation of how to conduct the analysis in R with a simple programming is provided.
