Semiparametric regression analysis of clustered survival data with semi-competing risks
Semiparametric regression analysis of clustered survival data with semi-competing risks
复制标题
DOI：
10.1016/j.csda.2018.02.003
复制
发表时间：
2018-08
期刊：
影响因子：
--
通讯作者：
Mengjiao Peng;Liming Xiang;Shanshan Wang
中科院分区：
文献类型：
--
作者：
Mengjiao Peng;Liming Xiang;Shanshan Wang
半竞争风险数据的分析在医学研究中变得越来越重要，其中受试者可能会经历非终止事件和终止事件，并且中间非终止事件（例如疾病发作）的时间受到终止事件（例如死亡）的依赖性审查，但反之则不然。通常，两种类型的事件都是相关的。在许多应用中，受试者也可能嵌套在集群中，例如多中心研究中的患者，由于受试者之间未观察到的共享因素，导致事件时间之间可能存在关联。为了将聚类内的依赖性和两种类型的事件时间之间的关联结合起来，我们提出了一种新的灵活的半参数建模框架，其中采用 copula 模型来表示非终止事件和终止事件的联合分布，并且它们的边际分布由具有随机效应的 Cox 比例风险模型建模。非参数最大似然估计程序是通过蒙特卡罗 EM 算法开发和实现的。所提出的估计器也被证明具有理想的渐近特性。大量模拟研究的结果表明，所提出的方法在有限样本中表现非常好，并且对于随机效应分布的错误指定特别稳健。我们通过分析多机构乳腺癌研究的数据进一步说明该方法的实用性。
Analysis of semi-competing risks data is becoming increasingly important in medical research in which a subject may experience both nonterminal and terminal events, and the time to the intermediate nonterminal event (e.g. onset of a disease) is subject to dependent censoring by the terminal event (e.g. death) but not vice versa. Typically, both two types of events are dependent. In many applications, subjects may also be nested within clusters, such as patients in a multi-center study, leading to possible association among event times due to unobserved shared factors across subjects. To incorporate dependency within clusters and association between two types of event times, we propose a new flexible semiparametric modeling framework where a copula model is employed for the joint distribution of the nonterminal and terminal events, and their marginal distributions are modeled by Cox proportional hazards models with random effects. A nonparametric maximum likelihood estimation procedure is developed and implemented through a Monte Carlo EM algorithm. The proposed estimator is also shown to enjoy desirable asymptotic properties. Results from extensive simulation studies indicate that the proposed method performs very well in finite samples and is especially robust against misspecification of the random effects distribution. We further illustrate the practical utility of the method by analyzing data from a multi-institutional study of breast cancer.