CMStatistics 2021: Start Registration
View Submission - CMStatistics
Title: Cluster randomized trials: Assumptions, estimands, and estimation in the presence of post-randomization selection Authors:  Georgia Papadogeorgou - University of Florida (United States) [presenting]
Fan Li - Yale University (United States)
Fan Li - Duke University (United States)
Abstract: In cluster-randomized trials, treatment is assigned randomly at the level of the cluster, all units within a cluster receive that treatment level, and estimands generally represent contrasts of potential outcomes at the level of the individual. We will address causal inference for cluster randomized trials. We will discuss that cluster-level randomization does not necessarily imply individual-level randomization, and formalize an assumption under which it does. In pragmatic CRTs, individuals are recruited in the study after the treatment is assigned at the cluster level, and individual recruitment can differ between treated and control clusters. What's more, data are only available among the subset of individuals that recruited. In the presence of post-randomization selection for cluster randomized trials, we will formalize causal estimands among those that recruited and in the overall population. We will link the post-randomization recruitment to covariate imbalance, and introduce the assumption of non-differential recruitment based on which we can draw causal inferences on the recruited population. Under the stronger assumption of ignorable missingness, we will show that the causal effect among the recruited control population corresponds to the causal effect among a specific tilted version of the overall population. Lastly, I will discuss sensitivity analysis for inferences on the recruited population, and the always-recruited overall population.