Optimal zone for bandwidth selection in semiparametric models

Abstract

We study the general problem of bandwidth selection in semiparametric regression. By expanding the higher-order terms in the Taylor series for the asymptotic mean-squared error, we provide a theoretical justification for the earlier empirical observations of an optimal zone of bandwidths in the literature. Based on the idea of cross-validating parametrical estimates, we further introduce a novel bandwidth selector for semiparametric models. The method is demonstrated by numerical studies to be able to preserve the selected bandwidth within the optimal zone. This data-driven cross-validation method may also be applicable for model diagnosis and longitudinal data settings. Examples from two clinical trials are provided to illustrate the applications.