A possible approach to bandwidth selection for difference-based variance estimators in the nonparametric regression is proposed. The approach is based on the crossvalidation-type idea modified for the case of correlated data. The method is compared to an alternative plug-in style method. Difficulties in implementing the latter are highlighted and it is argued that the proposed method represents a better alternative. Simulations for several models are given that illustrate good practical performance of the proposed method.
M. Levine