Evaluating The Performance Of Non-Experimental Estimators: Evidence From A Randomized Ui Program

One of the lessons of the treatment effects literature is the lack of consensus about the ability of statistical and econometric methods to replicate experimental estimates. In this paper, we provide new evidence using an unusual unemployment insurance experiment that allows the identification of discontinuities in the assignment mechanism. In particular, we use a set of regression functions and matching estimators based on kernel methods with mixed categorical and continuous data. A crucial issue with the kernel approach is the choice of the smoothing parameters. We develop a leave-one-out cross-validation algorithm that minimizes the mean square error of the average treatment effect on the treated weighting each comparison unit according to their distribution of covariates in the support region. Two main findings emerge. First, local constant and nearest-neighbor matching on kernel-based propensity score with mixed categorical and continuous data produces a closer approximation to the experimental estimates than traditional parametric propensity score models do. Second, the regression-discontinuity design emerges as a promising method for solving the evaluation problem. When restricted to sample observations in the neighborhood of the discontinuity points, the estimates are close approximation to the experimental estimates and are robust across different subsamples and estimators.