Statistical models for clustered survival data and multivariate recurrent events

在科学和生物医学研究中，由于聚类的事实，所收集的数据可能不是独立的。例如，在多中心临床试验中，来自同一医院/临床中心的患者可能具有一些共同的风险因素。在统计模型中不考虑人口异质性会导致统计结果的偏倚和模型效率的损失。我的研究计划是在个体观察不独立的情况下，开发具有预测性生物标志物变量的生存和复发事件数据的统计方法。**在某些类型的癌症中，如霍奇金淋巴瘤，患者的预期寿命很长，在治疗期间完全缓解的个体与没有完全缓解的个体相比，其生存分布可能不同。因此，重要的是使用早期反应数据来预测以后的生存分布，这可以通过研究治疗反应与生存之间的潜在关联来实现。我将为集群数据开发短期反应结果和长期生存结果的联合模型。提出的联合模型可以进一步扩展，以适应不同类型的结果，如连续变量和计数变量。预测性生物标志物是作为正常生物过程指标测量的患者特征，可用于预测哪些患者将或不会从新疗法中受益。我将开发新的程序来评估聚类数据中生物标志物定义的子集效应。我计划在集群生存数据中使用灵活的统计模型来预测生物标志物。当阈值参数位于生物标志物变量的边界附近时，我将研究现有的测试程序在这种情况下的表现，并为生物标志物阈值的存在开发新的测试。在许多临床试验和流行病学研究中经常出现复发事件数据。在某些情况下，反复发生的事件可能因失败事件（如死亡）而终止。我将提出多变量随机效应模型来研究在依赖终止和聚类的情况下风险因素对多类型复发事件的影响。**我将通过让不同层次的学生（本科生、理科硕士和博士生）参与到高质量的人才培养中来，发展与本提案相关的新颖的统计方法和理论。我将与硕士学生合作开发假设检验问题的计算程序，如联合模型中的人口异质性，评估现有方法对集群生存和复发事件数据的性能。与我的博士生一起，我将研究新的方法来同时对标志物反应和生存结果进行统计推断，开发灵活的统计方法来估计生物标志物与治疗的相互作用，并探索聚类多元复发事件的新模型。**提出的研究将推进新的统计方法和理论来处理聚类生存数据和复发事件数据。这些方法可以减少估计的偏差，提高统计模型的效率，解决复杂数据结构的计算挑战。从拟议的研究中开发的软件将有利于加拿大的统计科学、工程和可靠性研究以及生物医学研究界。

In scientific and biomedical research, data being collected may not be independent due to the fact of clustering. For example, in multiple centre clinical trials, patients from the same hospital / clinical centre may share some common risk factors. Fail to take the population heterogeneity into account in the statistical model could lead to biased results and loss of model efficiency. My research plan for this discovery grant is to develop statistical methods for survival and recurrent events data with predictive biomarker variables when individual observations are not independent. **In certain types of cancer, such as Hodgkin lymphoma, patients have long life expectancy, and individuals who experience complete remission during the treatment period may have different survival distribution compared to those who do not. It is then important to use the early response data to predict the survival distribution later, and this can be achieved by studying the potential association between the treatment response and survival. I will develop joint models of short term response outcome and long term survival outcome for clustered data. The proposed joint model can be further extended to accommodate different types of outcomes such as continuous variables and counting variables. Predictive biomarkers are patient characteristics that are measured as an indicator of normal biological processes, which can be used to predict which patients will or will not benefit from a new therapy. I will develop new procedures to evaluate biomarker defined subset effect for clustered data. I plan to use flexible statistical models for predictive biomarker in clustered survival data. When the threshold parameter is located near the boundary of the biomarker variable, I will investigate how existing test procedures perform under this circumstance and develop new test for the existence of the biomarker threshold. Recurrent event data occur frequently in many clinical trial and epidemiologic studies. In some situations, the recurrent event may be terminated by a failure event such as death. I will propose multivariate random effects models to investigate the effect of risk factor to multi-type recurrent events in the present of dependent termination and clustering.**I will contribute to high quality personnel training by involving different level of students (undergraduate, Master of Sciences and doctoral students) to develop novel statistical methodologies and theories related to this proposal. I will work with M. Sc. Students to develop computational procedures for hypothesis testing problems such as population heterogeneity in joint model, evaluating performance of existing methods for clustered survival and recurrent events data. Together with my Ph. D students, I will investigate new methodologies to make simultaneous statistical inferences for marker response and survival outcomes, develop flexible statistical methods to estimate biomarker-treatment interaction and explore new models for clustered multivariate recurrent events.**The proposed research will advance new statistical methodologies and theories to deal with clustered survival data and recurrent event data. These methods can reduce bias of the estimation, improve the efficiency of the statistical model and address the computational challenges for complex data structures. Software developed from the proposed research will benefit statistical science, engineering and reliability research and the biomedical research community in Canada.