Date : 21.01.00 (jj.mm.aa)
Suppose you want to estimate a density fX from a sample that has been contaminated by random noise. In this talk we consider the so-called deconvolving kernel estimator introduced by Stefanski and Carroll (1990), which requires the choice of a smoothing parameter h (the bandwidth). Although deconvolution estimation by kernel methods has received considerable attention in the literature, most of the papers deal only with theoretical aspects of the estimation, and few focus attention on the yet important practical choice of the bandwidth. In this talk we study several practical methods of selection of the smoothing parameter when we have errors in variables : a method based on asymptotics, a cross validation method, and a bootstrap method.