Robust Two-Sample Inference under Serial Dependence

Abstract

We propose robust inference for two-sample comparison with time-series data under serial dependence and heterogeneous long-run variances. Standardizing with orthonormal-basis (series HAR) projections, we develop two-sample t-tests and, for joint hypotheses on a vector of means, a series HAR Wald statistic. Because the increasing-K chi-square limit tends to over-reject, we propose Welch-type fixed-K t- and F-approximations with adjusted degrees of freedom. We further develop a series HAR wild bootstrap that reproduces serial dependence without resampling blocks. The framework nests difference-in-differences and Diebold-Mariano testing. Simulations and two empirical applications show accurate size control and competitive power, a tuning-free alternative to cluster-based inference.