Robust Nonparametric Bias-Corrected Inference in the Regression Discontinuity Design

Abstract

In the regression discontinuity design units are assigned treatment based on whether their value of an observed covariate exceeds a known cutoff. Local polynomial estimators are now routinely employed to construct confidence intervals for treatment effects in this design. The performance of these confidence intervals in applications, however, may be seriously hampered by their sensitivity to the specific bandwidth employed. Available bandwidth selectors typically yield a “large” bandwidth, leading to data-driven confidence intervals that may be severely biased, with empirical coverage well below their nominal target. In this paper, we show that these interval estimators may be improved by employing bias-corrected local polynomial estimators together with a novel standard-error estimator. Our innovation is to consider an alternative distributional approximation for the bias-corrected estimator that explicitly accounts for the (possibly first-order) contribution of the bias-estimate to the variability of the statistic. This alternative asymptotics lead to a new standard-error formula which is used for studentization purposes. We also propose a standard-error estimator that does not require an additional bandwidth choice, and develop valid bandwidth choices compatible with our asymptotic theory. In a