Bandwidth selection for local linear regression (Q1979106)

A new method (PURE) of bandwidth selection for nonparametric regression estimation is proposed. In PURE the bandwidth is selected as the minimizer of an estimator for mean integrated squared error (MISE). To estimate MISE the authors plug a pilot regression estimator into the bias term of the bias-variance decomposition of MISE. A minimizer of the least squares cross-validation (LSCV) score is used as a pilot estimator. The performance of PURE is compared with LSCV, direct and interactive asymptotic plug-in and global refined bandwidth selection procedures via simulation studies. The authors conclusion is that PURE selectors ``even though not always the best, compare favorably with other selectors in virtually all cases''.