Abstract

Consider estimating the average treatment effect (ATE) by imputing the missing potential outcomes. In this talk I will show that (a) such imputations are all intrinsically estimating the covariate density ratio between treated and control, or equivalently, the propensity score; (b) combining imputation with a type of bias correction due to Rubin (1973) and Abadie and Imbens (2011) yields doubly robust and semiparametrically efficient ATE estimators; and (c) a double machine learning version exists; it produces similar theoretical guarantees under arguably milder conditions.