Modeling Breast Cancer Recurrence Using New Statistical Methods for Semi-Markov P

描述(申请人提供)：纵向研究使我们能够调查慢性病的自然病史。对于一些疾病，这一过程的演变特征是一系列健康状态之间的转变。例如，在癌症中，受试者可能从无癌状态转变为早期疾病，然后转变为晚期疾病。半马尔可夫过程为多状态疾病过程的建模提供了一个灵活的框架。然而，将半马尔可夫过程应用于多态疾病研究的估计方法有限。现有的估计方法需要对这一过程强加许多假设，这在生物学上对许多疾病来说可能是不可信的。对于在小组观察下收集的数据，缺乏估计方法尤为严重。在这项建议中，我们开发了灵活的和生物学上合适的方法来估计纵向多态疾病研究中的半马尔可夫过程。我们提出了一类时间转换函数，它允许我们连接半马尔可夫过程和马尔可夫过程，以便利用可用于马尔可夫过程的直接估计方法。然后，我们建议使用混合模型，结合一组生物学上合理的假设，来扩展这些方法，以允许在小组观察下进行估计。这些方法将被应用于估计乳腺癌复发率和发生乳腺癌后的第二原发乳腺癌的比率。利用乳腺癌监测联盟13年的纵向数据，我们将展示我们的新方法相对于马尔可夫过程模型和比例风险生存模型在估计乳房事件状态之间的转移率方面性能有所改善。

DESCRIPTION (provided by applicant): Longitudinal studies allow us to investigate the natural history of chronic diseases. For some diseases the evolution of the process is characterized by transitions between a series of health states. For example, in cancer, subjects may transition from a cancer-free state to early stage disease, and subsequently to late stage disease. Semi-Markov processes provide a flexible framework for modeling multi-state disease processes. However, limited estimation methods are available for the application of Semi-Markov processes to the study of multi-state disease. Existing estimation methods require the imposition of numerous assumptions on the process, which may not be biologically plausible for many diseases. The lack of estimation methods is particularly acute for data collected under panel observation. In this proposal we develop flexible and biologically appropriate methods for estimating Semi-Markov processes in longitudinal multi-state disease studies. We propose a class of time transformation functions that allows us to bridge Semi-Markov and Markov processes in order to harness the straightforward estimation methods available for Markov processes. We then propose to use mixture models, coupled with a set of biologically reasonable assumptions, to extend these methods to allow for estimation under panel observation. These methods will be applied to the estimation of breast cancer recurrence rates and rates of second primary breast cancers subsequent to an incident breast cancer. Using 13 years of longitudinal data from the Breast Cancer Surveillance Consortium, we will demonstrate improved performance of our novel methods for estimating transition rates among breast event states relative to Markov process models and proportional hazards survival models.