A Rule-of-Thumb for the Variable Bandwidth Selection in Kernel Hazard Rate Estimation

Abstract

In nonparametric curve estimation the decision about the type of smoothing parameter is critical for the practical performance. The nearest neighbor bandwidth as introduced by Gefeller and Dette 1992 for censored data in survival analysis is specified by one parameter, namely the number of nearest neighbors. Bandwidth selection in this setting is rarely investigated although not linked closely to the frequently studied fixed bandwidth. We introduce a selection algorithm in the hazard rate estimation context. The approach uses a newly developed link to the fixed bandwidth which identifies the variable bandwidth as additional smoothing step. The procedure gains further data-adaption after fixed bandwidth smoothing. Assessment by a Monte Carlo simulation and a clinical example demonstrate the practical relevance of the findings.