Generalized cross-validation for bandwidth selection of backfitting estimates in generalized additive models

This article presents a modified Newton method for minimizing multidimensional bandwidth selection for estimation in generalized additive models. The method is based on the generalized cross-validation criterion applied to backfitting estimates. The approach in particular is applicable to higher dimensional problems and provides a computationally efficient alternative to full grid search in such cases. The implementation of the proposed method requires the estimation of a number of auxiliary quantities, and simple estimators are suggested. Extensions to semiparamatric models and other bandwidth selections are discussed.