A0764
Title: How does a better design improve the OLS regression?
Authors: Junho Choi - Seoul National University (Korea, South) [presenting]
Abstract: The aim is to demonstrate how balancing makes the inference of the OLS regression robust to model misspecification. First, a general representation of the average treatment effect is derived, which involves the OLS estimand. It decomposes their difference into three components, each bearing on the degree of self-selection, model misspecification, and imbalance in the distribution of control variables between treatment arms. It yields a useful outer bound for the bias of the OLS estimator, whose length is shown to be effectively invariant to misspecification once a better balance is attained. In this sense, better design leads are argued to the robustness of the OLS estimate. Lastly, the findings are extended to the staggered DiD settings.