Flexible Bayesian approaches to causal inference with multilevel survival data and multiple treatments

项目摘要 将比较有效性研究（CER）与传播和实施研究相结合， 在公共卫生和医疗保健服务中发挥更大的作用，允许从业人员做出知情决定 关于治疗和改善循证实践的采用。在CER质疑 不适合直接实验或在实施试验中不完全采用- 当发生干预时，建议使用“现场数据”的因果推断工具来评估治疗效果。的 大型国家电子卫生数据库日益复杂，对统计分析构成挑战， 需求方法超越传统的因果推理技术，传统上侧重于双， 没有治疗。考虑到大规模数据中捕获的丰富信息， 仅用两种治疗方法来定义。这些数据通常来自全国各地的治疗机构 在制度效果上有相当大的差异。虽然已经确定，流行的工具，双- 单一处理不适用于多处理设置，忽略多级数据结构 可能会使治疗效果的估计产生偏差，因此很少有替代方法被提出来处理这两种情况。 并发症同时我们提出的项目的第一个目标是开发一个新颖的和可扩展的贝叶斯 用聚类数据估计多种治疗对生存率的因果影响的方法。然后，我们充分 调查我们提出的方法在各种模拟场景和对比的操作特性 这是实践中经常使用的方法。为了使因果估计无偏，研究人员通常会 无不可测混杂（UMC）的假设。虽然强烈建议使用二元治疗， 对于多治疗和多水平生存敏感性分析，没有已知的实现或框架 数据我们的项目的第二个目标是开发和应用一种灵活和可解释的贝叶斯方法， 评估因果估计对可能偏离无UMC假设的敏感性， 集群和个人级别。这种方法能够测量所需的未观察到的混杂因素的数量 改变观察到的治疗效果的方向我们的项目将首先应用开发的方法， 两个目标是一个大的代表性高风险局限性前列腺癌人群，从de-identified艾德 国家癌症数据库，以评估三种流行治疗方案对生存率的平均因果影响 并评估未测量的混杂因素如何改变因果关系的结论。我们还将评估治疗效果- 并确定治疗有效或有害的不同患者亚组。我们的方法 将建立有效性部分，并为建立成本效益模型奠定基础， 并为进一步调查干预实施和修改的变化提供证据， 导致不同患者结果的治疗建议。为了便于传播我们的工作， 我们将通过R包共享底层的统计代码。

Project Summary Combining comparative effectiveness research (CER) and dissemination and implementation research is playing an increased role in public health and health care service by allowing practitioners to make informed decisions about treatments and improving adoption of evidence-based practices. In circumstances where CER questions do not lend themselves to direct experimentation or in implementation trials where incomplete adoption of in- tervention occurs, causal inference tools for “field data” are recommended for evaluating treatment effects. The increased complexities in large national electronic health databases pose challenges for statistical analyses and demand approaches beyond conventional causal inference techniques, which have traditionally focused on bi- nary treatment. Given the wealth of information captured in large-scale data, it is rare that treatment regimens are defined in terms of two treatments only. The data are typically pooled from treating facilities across the nation with considerable variability in the institutional effect. Although it has been established that popular tools for bi- nary treatment are inappropriate for the multiple treatment setting, and that ignoring the multilevel data structure can bias the estimate of the treatment effect, few alternative methods have been proposed to deal with both complications simultaneously. The first aim of our proposed project is to develop a novel and flexible Bayesian approach to estimating the causal effects of multiple treatments on survival with clustered data. We then fully investigate the operating characteristics of our proposed method in a variety of simulated scenarios and contrast it with approaches often used in practice. For causal estimates to be unbiased, researchers commonly make the assumption of no unmeasured confounding (UMC). Though highly recommended with binary treatment, there is no known implementation or framework for sensitivity analysis with multiple treatments and multilevel survival data. The second aim of our project is to develop and apply a flexible and interpretable Bayesian approach to assessing the sensitivity of causal estimates to possible departures from the assumption of no UMC, at both cluster- and individual-level. This approach is capable of gauging the amount of unobserved confounding needed to change the direction of the observed treatment effects Our project will apply the developed methods in the first two aims to a large representative high-risk localized prostate cancer population, drawn from the de-identified National Cancer Data Base, to evaluate the average causal effects of three popular treatment options on survival and evaluate how unmeasured confounding might alter causal conclusions. We also will estimate treatment het- erogeneity and identify distinct subgroups of patients for which a treatment is effective or harmful. Our methods will establish the effectiveness component and lay the groundwork for building the cost-effectiveness models, and provide evidence for further investigations of variations in intervention implementation and modifications in recommendations for treatments leading to different patient outcomes. To facilitate the dissemination of our work, we will share the underlying statistical code via an R package.