Methods for generalizing inferences from cluster randomized controlled trials to target populations

项目总结/摘要 当干预措施最好应用于组水平时， 一个人的暴露可能会影响同一集群中其他人的结果。群集试验是 越来越多地嵌入到大型医疗保健系统中，允许使用常规收集的数据， 提高研究效率。然而，有人担心-这项建议提供了支持性证据- 随机分组不能代表常规护理中的目标人群。当治疗 影响试验参与的因素不同，试验的治疗效果不能直接 适用于具有实质性利益的现实目标人群。因此，即使在设计良好的群集试验中， 选择性参与可能导致在对目标人群进行因果推断时产生偏差。鉴于 越来越多的集群试验正在进行，研究人员需要严格的方法来概括 针对目标人群的群集试验的结果，解决了选择性参与偏倚，并可以解释 多种数据科学挑战，包括同一聚类中观测值之间的随机依赖性; 仅从少数几个群集或从样本量相对较小的群集中获得随机试验数据; 缺乏对试验参与和结局的预测因素的了解，当候选协变量通常超过 可用集群的数量，需要使用灵活的机器学习方法;以及缺少 结果数据。为了响应特别关注通知NOT-LM-19-003，我们提出了新的，领域- 独立的、可重复使用的因果和统计方法来解决这些数据科学挑战， 通过消除选择性偏倚，提高集群试验为临床和政策决策提供信息的能力。 估计平均治疗效应和估计最佳协变量依赖性时的参与 治疗策略我们将评估的方法在现实的模拟研究和实证分析使用 来自美国养老院流感疫苗接种策略的3项大规模群集试验的数据。

PROJECT SUMMARY/ABSTRACT Cluster trials are the study design of choice when interventions are best applied at the group level and when exposure of one individual may affect the outcomes of other individuals in the same cluster. Cluster trials are increasingly embedded within large health care systems, allowing the use of routinely collected data to increase research efficiency. There is concern, however – and this proposal provides supportive evidence – that randomized clusters are not representative the target populations seen in routine care. When treatment effects vary over factors that influence trial participation, treatment effects from the trial cannot be directly applied to real-world target populations of substantive interest. Thus, even in well-designed cluster trials, selective participation can lead to bias in drawing causal inferences about the target population. Given the increasing number of cluster trials being conducted, investigators need rigorous methods for generalizing findings from cluster trials to target populations that address selective participation bias and can account for multiple data science challenges, including stochastic dependence among observations in the same cluster; availability of randomized trial data from only a few clusters or from clusters with relatively small sample sizes; lack of knowledge of predictors of trial participation and the outcome, when candidate covariates often exceed the number of available clusters and necessitate the use of flexible machine learning approaches; and missing outcome data. In response to Notice of Special Interest NOT-LM-19-003, we propose novel, domain- independent, reusable causal and statistical methods to address these data-science challenges and to increase the ability of cluster trials to inform clinical and policy decisions by eliminating bias due to selective participation when estimating average treatment effects and when estimating the optimal covariate-dependent treatment strategy. We will evaluate the methods in realistic simulation studies and in empirical analyses using data from 3 large-scale cluster trials of influenza vaccination strategies in U.S. nursing homes.