Frailty Models and Survival Analysis in Cancer Research

描述(由申请人提供)：该提案的目标是开发实用的生存分析工具，用于肿瘤学和健康科学中的其他慢性疾病的临床、流行病学和基础科学研究。在目标1中，一个主要动机是艾滋病临床试验替代主要终点，如病毒失败可能被信息性退出审查，朴素的COX模型分析可能具有误导性。我研究了代理端点的时间依赖回归模型和时间依赖模型，包括保守的敏感性分析，它们解释了相依截尾。这些分析可以检测到治疗效果和审查机制的细微时间变化，并是对艾滋病和其他慢性病研究中单纯回归分析的有益补充，在这些研究中，辍学是有问题的。在目标2中，我调查了多变量竞争风险数据中的关联，这是基于人群的遗传流行病学生存研究中的一个重要主题，比如卡奇县的老龄化研究，在该研究中，人们对痴呆症等慢性病的家族发病关联感兴趣。标准的删失数据关联分析没有解决发病年龄可能受死亡制约的问题。我将扩展对病因特定风险和累积关联函数的经典单变量分析，以获得新的时变关联度量和检验。这些方法将提供有关家族性疾病关联的基本知识，这些知识可能无法通过更简单的参数方法检测到。在癌症试验中，比如国家外科辅助乳肠项目的试验，在测试协变量效应时，通常会使用特别的方法，其中一些变化是非正式比较的，多个测试问题可能会被忽略。这种“作弊”夸大了第一类错误率，并可能给出误导性的结果。目的3在制定生存终点的临床风险指数时，使用参数协变量变换提出协变量的最佳推断。构建的测试可能比带有固定转换的朴素测试更强大。这些结果为分析家在癌症预后的探索性亚组分析中提供了重要指导。在目标4中，我研究了竞争风险数据的非参数分位数推断。累积发病率估计经常在癌症试验中报告，例如，放化疗联合应用的局部区域复发率。对非局部区域事件的相依筛选使生存分析中常用的总结分位数定义复杂化。独立审查数据的分位数被广泛使用Kaplan-Meier曲线报告，但不适合竞争风险。拟议的方法学将广泛适用于癌症应用，解决癌症研究中的一个关键方法学空白。对于每个目标，用户友好的软件都将公开可用。主要科学期刊上的说明性论文将向影响较大的主题受众传播这一方法。公共卫生相关性：这笔赠款的目标是开发事件发生时间终点的统计方法，这些方法将广泛应用于肿瘤学的临床、流行病学和基础科学研究。这些方法将有助于识别家庭和环境风险因素，这些因素对于正确评估未受影响个人未来的癌症风险以及为癌症患者制定有效的预防和治疗干预措施至关重要。目前的统计方法不充分，阻碍了能够改善癌症预后和治疗的研究的设计和分析。

DESCRIPTION (provided by applicant): The proposal's objective is to develop practicable survival analysis tools for clinical, epidemiologic, and basic science studies in oncology and other chronic diseases in the health sciences. In Aim 1, a main motivation is AIDS clinical trials surrogate primary endpoints, like viral failure may be censored by informative dropout and naive Cox model analyses may be misleading. I investigate time-dependent regression models for the surrogate endpoints and time-dependent dependence models, including conservative sensitivity analyses, which account for dependent censoring. The analyses may detect subtle temporal changes in treatment efficacy and the censoring mechanism and are useful complements to naive regression analyses in AIDS other chronic disease studies where dropout is problematic. In Aim 2, I investigate associations in multivariate competing risks data, an important topic in population based genetic epidemiologic survival studies, like the Cache County Study of Aging, where familial onset associations for chronic diseases, like dementia, are of interest. Standard censored data association analyses do not address that the onset ages may be dependently censored by death. I will extend classic univariate analyses of cause-specific hazard and cumulative incidence functions to obtain novel time-varying association measures and tests. The methods will provide fundamental knowledge about familial disease assocations which may not be detected by simpler parametric methods. In cancer trials, like those at the National Surgical Adjuvant Breast and Bowel Project, ad hoc approaches are often used when testing covariate effects, where a few transformations are compared informally and multiple testing issues may be ignored. Such "cheating" inflates the type I error rate and may give misleading results. Aim 3 proposes optimal inferences for covariates using parametric covariate transformations when developing clinical risk indices for survival endpoints. Tests are constructed which may be more powerful than naive tests with fixed transformations. These results provide critical guidance to analysts in exploratory subgroup analyses for cancer prognosis. In Aim 4, I study nonparametric quantile inference for competing risks data. Cumulative incidence estimates are often reported in cancer trials, for example, rates of locoregional recurrence with combined radiation and chemotherapy. The dependent censoring from non-locoregional events complicates quantile definition, a commonly used summary in survival analysis. Quantiles for independently censored data are widely reported using Kaplan-Meier curves, but are not appropriate with competing risks. The proposed methodology will be broadly applicable in cancer applications, addressing a key methodologic gap in cancer research. For each aim, user friendly software will made publicly avaiable. Expository papers in leading scientific journals will disseminate the methodology to high impact subject matter audiences. PUBLIC HEALTH RELEVANCE: The goal of this grant is to develop statistical methods for time-to-event endpoints which will be widely applicable in clinical, epidemiologic, and basic scientific research in oncology. The methods will be useful in identifying familial and environmental risk factors which are critical to correctly assessing future cancer risk in unaffected individuals and to developing effective preventive and therapeutic interventions in cancer patients. Current statistical methods are inadequate and have hindered the design and analysis of studies which could bring about improvements in cancer prognosis and treatment.