|
Abstract:
|
Given a set of survival data, the problem of hazard estimation and inference with little prior knowledge of structure is common. Standard estimators in the non- or semi-parametric category are often employed as a first step to identifying a final model which fits the data--perhaps too well, resulting in overfitting. On the other hand, the flexibility of fully nonparametric estimators may be more than necessary, given a premise of underlying continuity in the hazard under investigation. As a remedy, a method of hazard estimation is described which is related to kernel smoothing and features automatic bandwidth selection. Cross-validated modified likelihood (CVML) is invoked, with adaptations for the problem of discrete hazard estimation in potentially large data sets, using the stratified bootstrap. Applications include CRC incidence in the Minnesota Colon Cancer Control Study.
|
