Simple and Bias-Corrected Matching Estimators for Average Treatment Effects [report]

Alberto Abadie, Guido Imbens
2002 unpublished
In this paper we analyze large sample properties of matching estimators, which have found wide applicability in evaluation research despite that fact that their large sample properties have not been established in many cases. We show that standard matching estimators have biases in large samples that do not vanish in the standard asymptotic distribution if the dimension of the covariates is at least four, and in fact dominate the variance if the dimension of the covariates is at least five. In
more » ... at least five. In addition, we show that standard matching estimators do not reach the semiparametric efficiency bound, although the efficiency loss is typically small. We then propose a bias-corrected matching estimator that has no asymptotic bias. In simulations the bias-corrected matching estimator performs well compared to simple matching estimators and to regression estimators in terms of bias and root-mean-squarederror.
doi:10.3386/t0283 fatcat:s4qsjmgvvbhfrahhr2otol5bhm