Adaptive Bandwidth Choice for Kernel Regression

Abstract

A data-based procedure is introduced for local bandwidth selection for kernel estimation of a regression function at a point. The estimated bandwidth is shown to be consistent and asymptotically normal as an estimator of the (asymptotic) optimal value for minimum mean square estimation. Simulation studies indicate satisfactory behavior of the new bandwidth estimator in finite samples. The findings are improvements over a global bandwidth estimator. The same methodology works for local linear regression and extends easily to weighted local polynomial fits.