Linear regression adjustments for pre-treatment covariates are widely used in economics to lower the variance of treatment effect estimates when analyzing data from randomized experiments. This method is robust to misspecification, and delivers reliable confidence intervals even in relatively small samples. More flexible covariate adjustments, using nonlinear parametric or fully nonparametric methods, have the potential to improve efficiency. They are rather uncommon in practice, however, because they can introduce bias or require very large samples in order for asymptotic inference to be reliable. This paper shows that with a simple modification of the treatment effect estimator, it is possible to alleviate these issues substantially. For a large class of covariate adjustments, estimation and inference in randomized experiments is possible without sacrificing the robustness properties of linear regressions. Full efficiency can be achieved through nonparametric adjustments under minimal conditions, in particular without imposing high-order smoothness restrictions in settings with many covariates
Dieser Eintrag ist Teil der Universitätsbibliographie.
Search Authors in