My research interests are in econometric and statistical theory, with a particular interest in hypothesis testing and applications of statistical decision theory.
Vox EU briefing here.
Offline training for improving online performance of a genetic algorithm based optimization model for hourly multi-reservoir operation, with Duan Chen, Arturo S. Leon, Claudio Fuentes, and Qiuwen Chen.
Abstract: When constructing Wald tests, consistency is the key property required for the variance estimator. This property ensures asymptotic validity of Wald tests and confidence intervals. Classical efficiency comparisons of hypothesis tests indicate all consistent variance estimators lead to equivalent Wald tests. This paper develops a simple relative efficiency measure which leads to several new conclusions. These include quantifying the power loss associated with using cluster-robust variance estimators when using overly coarse clusters, recommending particular kernels for estimating the asymptotic variance in quantile regression, and comparing the power of Anderson-Rubin tests to the standard Wald test. As a byproduct, the asymptotic distributions of several test statistics are derived under fixed alternatives. Simulation evidence indicates the new asymptotic efficiency measure provides good finite-sample predictions. In an application using data from the American Community Survey, it is demonstrated how to use the new approach for conducting power analysis when looking at the effect of minimum wage increases on employment.
Abstract: The t-test is a standard inferential procedure in economics and finance. When the data exhibit heavy tails, the t-test may have low power. This paper characterizes the rate at which power converges to 1 for data in a particular class of heavy tailed distributions. While classical results on the rate of convergence of power focus on exponential rates, we find the rate to be a much slower polynomial rate when the data have heavy tails. We compare these results with other results on the efficiency of the t-test in the literature, and use empirically-calibrated simulation evidence to demonstrate how our results make good finite-sample predictions