Information Theory and Propensity Score Matching

Abstract

This paper develops a generalized maximum entropy (GME) approach to propensity score matching (PSM). A GME discrete choice model is used to develop propensity scores and estimate treatment effects in a set of Monte Carlo simulations. The GME PSM is compared to a more traditional logit PSM. Sample sizes and common support regions are varied across simulations to reflect common problems in the program evaluation literature. The GME PSM exhibits bias levels that are comparable to the logit PSM, but it provides substantial improvements in the precision of treatment effect estimates (including lower standard deviation of the treatment effect estimate and lower RMSE).