Novel and Rigorous Statistical Learning and Inference for Comparative Effectiveness Research with Complex Data

项目摘要 比较有效性研究（CER）通常用于发现和提供信息， 替代药物或治疗方法在有效性和安全性方面可能存在的差异。等 信息，如果可靠和准确，可以帮助患者，临床医生和其他医疗保健利益相关者， 更明智的医疗决策，并改善医疗服务和结果。然而，图纸有效 观察性研究中关于治疗效果的相关推论涉及到来自 主题研究人员和统计学家。一方面，因果推理依赖于结构性推理， 选择。这类假设的两个突出类别是不混淆或工具变量（IV） 选择。另一方面，假定结构性假设，因果推断也需要统计建模 以及从经验数据中估计种群特性和关联。统计学习的问题 和推理可能具有挑战性，同时允许大量候选回归变量，如主效应 和协变量的相互作用。我们研究的目标是开发、评估和传播一种新的 一套理论上严格的，数值自动化的，实际上有用的统计学习方法， 在CER中使用复杂、高维数据估计治疗效果的推断。三个具体目标 是（1）在非控制下关于群体和亚群体平均治疗效果的高维推断- 多值处理的成立性，（2）关于局部平均处理效果的高维推断 和IV依赖的平均治疗效果的治疗与多值工具和治疗，和 (3)关于平均治疗效果的高维推断，例如生存率和风险之间的对比 纵向和生存数据的概率。我们将研究新方法的应用， 比较有效性和安全性研究，包括最近关于比较治疗策略的研究， 精神分裂症和一个正在进行的项目，以评估同类药物的治疗交换，例如， 房颤、房扑或二肽基肽酶-4患者中充足的直接口服抗凝剂 2型糖尿病患者中的抑制剂，同时利用由医疗保险前， 处方药的好处。我们将开发和公开发布用户友好的计算机软件，包括透明的 直接实施新方法的文件。

Project Summary Comparative effectiveness research (CER) in medicine is commonly conducted to discover and provide infor- mation on possible differences between alternative drugs or treatments in their effectiveness and safety. Such information, if reliable and accurate, can help patients, clinicians, and other healthcare stakeholders to make better-informed healthcare decisions and improve healthcare delivery and outcomes. However, drawing valid and relevant inferences about treatment effects from observational studies involves effort and expertise from both subject-matter researchers and statisticians. On one hand, causal inference relies on structural assump- tions. Two prominent classes of such assumptions are unconfoundedness or instrument variable (IV) assump- tions. On the other hand, granted the structural assumptions, causal inference also requires statistical modeling and estimation of population properties and associations from empirical data. The problem of statistical learning and inference can be challenging, while allowing a large number of candidate regressors such as main effects and interactions of covariates. The objective of our research is to develop, evaluate, and disseminate a new set of theoretically rigorous, numerically automated, and practically useful methods of statistical learning and inference for estimating treatment effects in CER with complex, high-dimensional data. Three specific aims are (1) high-dimensional inference about population and subpopulation average treatment effects under uncon- foundedness with multi-valued treatments, (2) high-dimensional inference about local average treatment effects and IV-dependent average treatment effects on the treated with multi-valued instruments and treatments, and (3) high-dimensional inference about average treatment effects such as contrasts between survival and hazard probabilities with longitudinal and survival data. We will investigate applications of the new methods to several comparative effectiveness and safety studies including a recent study on comparative treatment strategies in schizophrenia and an ongoing project to evaluate the therapeutic exchangeability of same-class drugs, for ex- ample, direct oral anticoagulants among patients with atrial fibrillation or atrial flutter or dipeptidyl peptidase-4 inhibitors among patients with type 2 diabetes, while exploiting IVs created by the design of the Medicare pre- scription drug benefit. We will develop and publicly release user-friendly computer software including transparent documentation for direct implementation of the new methods.