| Summary: | 碩士 === 國立臺灣大學 === 流行病學研究所 === 88 === Abstract
Recurrent event data are commonly encountered in longitudinal studies. Such data arises in various areas such as reliability、medicine、economics、sociology and etc... For examples, in a clinical study people with cancer may experience multiple tumor recurrences and in industrial studies the breakdown of a machine is also a recurrent event.
In this study, we review various regression models to describe the event recurrences related to various factors. A conditional hazards model generalized from Cox’s semiparametric hazards models is considered in the study. This model includes two types of effects, global common and episode-specific effects, and the hazards are assumed to be the same for each episode of events. The aim of the study is to develop a more efficient estimating method for the global effects and estimation for the cumulative common baseline hazard function. Under this conditional model, there are two methods to estimate the effects. The first method is based on partial likelihood, which is stratified by episodes of events. The second method is from an unstratified profile likelihood by pooling all of events. But these have their own advantage, the first ( stratified) method has smaller bias and the second ( unstratified) method has smaller variance. In order to get more efficient estimator, we consider a new method, compromise method, to balance the advantages of these two methods. The comparisons of these estimating methods are illustrated by simulation studies.
In the analysis of real data, we can apply the plot of the estimated cumulative episode-specific baseline hazard functions against time to select more appropriate compromise method.
|
|---|
