Automated Bandwidth Selection for Inference in Linear Models With Time‐Varying Coefficients

ABSTRACT The problem of selecting the smoothing parameter, or bandwidth, for kernel‐based estimators of time‐varying coefficients in linear models with possibly endogenous explanatory variables is considered. We examine automated bandwidth selection by means of cross‐validation, a nonparametric variant of Akaike's information criterion, and bootstrap procedures based on wild bootstrap and dependent wild bootstrap resampling schemes. Our simulations show that data‐driven selectors based on cross‐validation and the dependent wild bootstrap are the most successful overall in a variety of settings that are relevant in econometrics. Empirical examples illustrate the practical use of the automated procedures.