Consistent bandwidth selection for kernel binary regression

The use of nonparametric regression techniques for binary regression is a promising alternative to parametric methods. As in other nonparametric smoothing problems, the choice of smoothing parameter is critical to the performance of the estimator and the appearance of the resulting estimate. In this paper, we discuss the use of selection criteria based on estimates of squared prediction risk and show consistency and asymptotic normality of the selected bandwidths. The usefulness of the methods is explored on a data set and in a small simulation study.