Tyranny-of-the-minority regression adjustment in randomized experiments

Abstract

Regression adjustment is widely used for the analysis of randomized experiments to improve the estimation efficiency of the treatment effect. This paper reexamines a weighted regression adjustment method termed as tyranny-of-the-minority (ToM), wherein units in the minority group are given greater weights. We demonstrate that the ToM regression adjustment is more robust than Lin 2013's regression adjustment with treatment-covariate interactions, even though these two regression adjustment methods are asymptotically equivalent in completely randomized experiments. Moreover, we extend ToM regression adjustment to stratified randomized experiments, completely randomized survey experiments, and cluster randomized experiments. We obtain design-based properties of the ToM regression-adjusted average treatment effect estimator under such designs. In particular, we show that ToM regression-adjusted estimator improves the asymptotic estimation efficiency compared to the unadjusted estimator even when the regression model is misspecified, and is optimal in the class of linearly adjusted estimators. We also study the asymptotic properties of various heteroscedasticity-robust standard error estimators and provide recommendations for practitioners. Simulation studies and real data analysis demonstrate ToM regression adjustment's superiority over existing methods.