the most general form — regresses $Y$ on treatment, pre-experiment covariates, and their interactions. Lin (2013) showed that the fully interacted regression estimator is at least as efficient as the simple difference-in-means and is consistent even if the regression model is misspecified. Modern ex