Bandwidth Selection for Semiparametric Estimators Using the m-out-of-n Bootstrap

This paper considers a class of semiparametric estimators that take the form of density-weighted averages. These arise naturally in a consideration of semiparametric methods for the estimation of index and sample-selection models involving preliminary kernel density estimates. The question considered in this paper is that of selecting the degree of smoothing to be used in computing the preliminary density estimate. This paper proposes a bootstrap method for estimating the mean squared error and associated optimal bandwidth. The particular bootstrap method suggested here involves using a resample of smaller size than the original sample. This method of bandwidth selection is presented with specific reference to the case of estimators of average densities, of density-weighted average derivatives and of density-weighted conditional covariances.