Methodological Issues in Event Histories Analysis

点击翻译按钮获取中文摘要

Event history studies are conducted by monitoring subjects or individuals over time and collecting information on the occurrence of certain events or conditions according to one or multiple dynamic processes. It is characterized by multiple events and longitudinal observations. In the last twenty years, considerable attention has been drawn towards methodology development for multivariate survival times, recurrent events and multistate models and more general event history processes for complex outcomes.******A typical prevalent cohort study only includes individuals being in certain states (e.g. alive with disease) at recruitment time and then at, most typically, planned follow-up times. If the disease survival time is of primary interest, ineligibility of those diseased but dead prior to recruitment results in selection bias in estimation. This is cast into a length biased problem in the context of ``single event" survival analysis and it has been well studied in the literature. However, little has been done for response dependent sampling in a general framework of multistate event history analysis. This research aims to quantify selection biases in the transition intensities estimation, and develop suitable methods for adjustment. Inference based on parametric, weakly parametric and non-parametric approaches will be considered. Methods to deal with intermittent assessment will be investigated.******Repeated or recurrent events often arise in event history studies. Models include Poisson processes based on rate functions and counts of events; renewal processes focus on the gaps or waiting times and hazard functions. Complex association structure arises when multiple event processes with recurrent episodes are involved. This research aims to develop a joint method using different random effects to character the association among gap times, including within-process association among duration of exacerbation and remission of the same type of event and between-processes association among duration of different types of events. Innovative approaches will be used to tackle the difficulty with constructing high-dimensional multivariate distributions. Joint methods based on composite likelihood will also be considered. ******Despite the advances in joint consideration of multiple events using multistate models or multivariate survival models, it is often appealing in practice to evaluate the treatment effect on a composite endpoint where the outcome of interest is the time to the first occurrence of any component endpoint. Complications arise in the analysis when a component (e.g. progression) is intermittently assessed. I plan to investigate the impact of right endpoint imputation, a standard approach to deal with interval censoring in practice, in the composite endpoint analysis under various modelling frameworks. Methods for sample size adjustment are also of interest. *** **