Abstract:

The classic papers by Newey and West (1987) and Andrews (1991) spurred a large body of work investigating methods for improved heteroskedasticity- and autocorrelation-robust (HAR) inference in time series regression. Empirical practice, however, has not kept up with these developments. The goal of this paper is to draw on these developments to make a recommendation for practitioners about how to compute standard errors in regression settings with heteroskedasticity and serial correlation. A general conclusion from this literature is that the large size distortions of conventional HAR tests can be greatly reduced by using longer time-domain truncation parameters and modified “fixed-b” asymptotic critical values. In general, fixed-b distributions are nonstandard, however a class of tests (using basis-function estimators of the long-run variance) has standard t and F fixed-b asymptotic distributions. We draw on recent theoretical results to organize extensive Monte Carlo experiments that reflect both the benchmark designs used in the literature and data-based designs using U.S. macroeconomic time series. Theory and Monte Carlo results suggest that little is lost by restricting attention to tests with fixed-b t- and F-distributions. We conclude with specific recommendations for kernel choice and bandwidth choice.

Notes: