Trading-off Bias and Variance in Stratified Experiments and in Matching Studies, Under a Boundedness Condition on the Magnitude of the Treatment Effect

Abstract: I consider estimation of the average treatment effect (ATE), in a population composed of $S$ groups or units, when one has unbiased estimators of each group's conditional average treatment effect (CATE). These conditions are met in stratified experiments and in matching studies. I assume that each CATE is bounded in absolute value by $B$ standard deviations of the outcome, for some known $B$. This restriction may be appealing: outcomes are often standardized in applied work, so researchers can use available literature to determine a plausible value for $B$. I derive, across all linear combinations of the CATEs' estimators, the minimax estimator of the ATE. In two stratified experiments, my estimator has twice lower worst-case mean-squared-error than the commonly-used strata-fixed effects estimator. In a matching study with limited overlap, my estimator achieves 56\% of the precision gains of a commonly-used trimming estimator, and has an 11 times smaller worst-case mean-squared-error.