Bayesian Analysis of Competing Risks

竞争风险的贝叶斯分析摘要研究者开发了新的统计方法，并显著地改进了目前可用的方法，用于分析具有竞争风险的删失生存数据，其中几个潜在的“风险”或失败原因对一个受试者或一个单元起作用。 在这种情况下出现的故障时间数据往往被掩盖的故障原因是不确切知道，但只能缩小到一个子集的所有潜在原因。研究者（a）使用原因特异性模型为掩盖的竞争风险数据开发贝叶斯框架，并将其与潜在方法进行比较，（B）为此类数据开发灵活的半参数模型，并将其与参数模型进行比较，（c）为参数和半参数模型开发贝叶斯模型选择方法，并将其用于比较模型拟合和预测能力。研究者研究的推理方法，用于估计总体和特定原因的生存概率，从部分掩蔽的数据;用于估计诊断失败的概率，从一个特定的原因，给定一个掩蔽的子集的原因;并纳入协变量的信息，通过回归在设盲的生存data.Masked竞争的风险数据经常出现在生物医学设置，临床试验和工程应用。例如，作用于被诊断患有前列腺癌的患者的竞争性风险可能是前列腺癌;其他疾病，如高血压、心血管疾病和糖尿病;或者仅仅是老年（因为前列腺癌是一种生长缓慢的疾病）。在工程环境中，系统（如计算机）的故障可能是由于可能无法准确识别的特定组件的故障。本研究的目标是开发新的统计方法，并显着推进目前可用的方法来分析这些数据。所开发的计算方法将公开提供，以便任何对分析此类竞争风险数据感兴趣的人都可以使用这些方法。

DMS-0306416Sanjib BasuBayesian analysis of competing risksAbstractThe investigator develops new statistical methodology and significantly advances currently available methodology for analyzing censored survival data with competing risks where several potential "risks" or causes of failure are operating on a subject or a unit. The failure time data arising in such contexts are often masked where the cause of failure is not exactly known, but can only be narrowed down to a subset of all potential causes. The investigator (a) develops a Bayesian framework for masked competing risks data using the cause-specific model and compares it to the latent approach, (b) develops flexible semiparametric models for such data and compares them to the parametric models, and (c) develops Bayesian model selection methodology for both parametric and semiparametric models and uses it to compare both model fit and predictive power. The investigator studies inferential methods for estimation of overall and cause-specific survival probabilities from the partially masked data; for estimation of the diagnostic probability of failure from a specific cause, given a masked subset of causes; and for incorporating covariate information through regressors in the setting of masked survival data.Masked competing risks data arise frequently in biomedical settings, clinical trials, and engineering applications. For example, the competing risks acting on a patient diagnosed with prostate cancer could be prostate cancer; other diseases like hypertension cardiovascular disease, and diabetes; or simply old age (since prostate cancer is a slow-growing disease). In an engineering setting, the failure of a system (such as a computer) could be due to failure of a specific component which may not be exactly identified. The goal of this research is to develop new statistical methodology and significantly advance currently available methodology for analyzing such data. The developed computational methods will be made publicly available so that anyone interested in the analysis of such competing risks data can use these methods.