Semiparametric Bayesian Survival Analysis

描述（由申请人提供）：通常在癌症和其他医学研究中，主要反应变量是某些特定感兴趣事件发生的时间（例如癌症复发，患者死亡等）。这种数据被称为生存数据。本研究的目标是发展和扩展半参数贝叶斯模型以及相关的全贝叶斯和经验贝叶斯方法，用于分析从各种生物医学研究中获得的生存数据。半参数模型和相关方法将足够复杂，足以处理存在复杂审查和不规则数据收集监测方案的生存数据，存在错过的门诊就诊，受试者面临不同类型复发事件的风险以及不同原因导致的失败。半参数模型是过于严格的参数模型和过于缺乏信息的非参数模型之间的一种流行折衷。半参数模型有非参数部分（未知函数，如基线危险或强度函数）和参数部分，包括一些参数，如解释变量的回归系数。非参数部分的先验信息可以概括为一个随机过程，称为先验过程。参数部分的可用先验信息将被建模为先验分布。在本项目中开发的方法将有助于分析复发事件数据、具有竞争性失败原因的生存数据、来自同一受试者的多个事件状态的研究的生存数据、通过结果依赖的不定期诊所访问测量的纵向数据以及来自多个生活质量事件的生存数据。模型的开发、相关的数据分析工具、相关的计算机程序和统计包代码、广泛的模拟研究以及用于验证关键建模假设的诊断工具将在每个具体目标中发挥核心作用。新的和现有的方法将进行评估和比较，主要使用来自已发表文献和其他来源（如MUSC Hoolings癌症中心）的癌症研究数据进行分析。

DESCRIPTION (provided by applicant): Often in cancer and other medical studies, the primary response variable is the time to the occurrence of some particular event of interest (e.g. relapse of cancer, death of the patient, etc.). This kind of data is called survival data. The goals of this proposed research are to develop and extend semiparametric Bayesian models and associated full Bayes and empirical Bayes methodologies for the analysis of survival data obtained from various biomedical studies. The semiparametric models and associated methods will be sophisticated enough to deal with survival data in the presence of complex censoring and irregular data- collection monitoring schemes, in the presence of missed clinic visits and with subjects at the risk of recurrent events of different types as well as failure from different causes. Semiparametric models are a popular compromise between the too restrictive parametric and too non-informative nonparametric models. A semiparametric model has a nonparametric part (an unknown function such as a baseline hazard or an intensity function) as well as a parametric part involving a few parameters, such as regression coefficients for explanatory variables. The available prior information on the nonparametric part will be summarized as a stochastic process, called a prior process. The available prior information on the parametric part will be modeled as a prior distribution. The methodology developed during this project will be useful for the analysis of recurrent events data, survival data with competing causes of failures, survival data from the studies with multiple event states par subject, longitudinal data measured via outcome-dependent irregular clinic visits and survival data from multiple quality of life events. Development of models, associated data analytic tools, related computer programs and codes for statistical packages, extensive simulation studies, and diagnostic tools for verifying the key modeling assumptions will play central roles in each specific aim. New and existing methods will be evaluated and compared using analysis of mainly cancer studies data from the published literature and from other sources such as the MUSC Hoolings Cancer Center.