Robust Causal Comparisons of Nonrandomized Oncology Studies

肿瘤学研究中非随机分组的稳健因果比较 我们研究的目标是开发稳健的统计模型，用于非随机组的因果比较。与 提高真实世界数据(RWD)的可用性，这是一项比较新疗法与外部对照的试验设计 通过RWD(例如，EHR)获得的信息最近重新受到关注。该设计已在治疗中得到应用 发展，特别是在RCT不可行的罕见疾病中，例如比较三年无复发生存期(RFS) 局部治疗的高危眼黑色素瘤患者接受辅助联合免疫治疗与匹配的患者之间的比较 在审判之外获得的同时控制。那么挑战就是如何做出因果推论 治疗在延长时间方面是有效的。病人存活率。但是用哪种方法呢？已知的因果推理依赖于 各种假设。尽管在使各种因果推理方法稳健方面取得了进展，例如，最显著的是双重 稳健估计(DRE)现有的DRE仍然存在几个主要缺陷，例如，对温和模型过于敏感 规格错误。我们对增强的DRE的初步研究，表明了在进行因果关系时所需的稳健性 通过半参数模型实现对具有连续主要终点的试验的推理。建立在这个基础上 开发，该应用程序的目标是开发用于分析非随机临床试验的新的DRE方法 具有二元(如肿瘤学中的反应、流行病学中的发病率)和事件发生时间(如生存和进展 肿瘤学中的自由生存)结果并评估其统计特性，众所周知，这可能与 连续结果的原则性证明(目标1-2)。然后，这些方法将被应用于NCI赞助的试验， 以及一项人口科学研究和一项正在进行的免疫疗法试验。

Robust Causal Comparison of Nonrandomized Groups in Oncology Studies The goal of our research is to develop robust statistical models for causal comparison of nonrandomized groups. With the increasing availability of real world data (RWD), the trial design to compare a new therapy versus an external control obtained via RWD (e.g., EHR) has received renewed attention recently. This design has been utilized in therapeutic development especially in rare diseases when RCT is not feasible, e.g., comparing 3-year relapse-free survival (RFS) between locally treated high-risk ocular melanoma patients on adjuvant combination immunotherapy versus a matched contemporaneous control obtained outside of the trial. Then the challenge is at making causal inference on whether the treatment is efficacious in prolonging (e.g.) patient survival. But which method? Causal inference is known to depend on various assumptions. Despite advances in making various causal inference methods robust, e.g., most notably the doubly robust estimate (DRE) existing DREs continue to suffer several major drawbacks, e.g., being too sensitive to mild model misspecifications. Our preliminary studies on enhanced DREs that shows the needed robustness in making causal inference realized through semi-parametric models for trials limited with continuous primary endpoint. Built upon this development, the goal of this application is to develop the novel DRE approach for analyzing nonrandomized clinical trials with binary (such as response in oncology, incidence in epidemiology) and time to event (such as survival and progression free survival in oncology) outcomes and assess their statistical properties, which as is well known can be quite different from the proof of principle case with continuous outcome (Aims 1-2). The methods will then be applied to an NCI sponsored trial, and a population science study and one ongoing immunotherapy trial.