Cluster randomized trials (CRTs) involve randomizing groups of individuals to different interventions. While model-based methods are extensively studied for analyzing CRTs, there has been little reflection around the treatment effect estimands at the outset. In the first part of this presentation, we describe two relevant estimands that can be addressed through CRTs and point out that they can differ when the treatment effects vary according to cluster sizes. As a cautionary note, we demonstrate how choices between different analytic approaches can impact the interpretation of results by fundamentally changing the question being asked. In the second part, we revisit the linear mixed model as the most commonly used method for analyzing CRTs. The linear mixed model makes stringent assumptions, including normality, linearity, and typically a compound symmetric correlation structure, all of which may be challenging to verify. However, under certain conditions, we show that the linear mixed model consistently estimates the average causal effect under arbitrary misspecification of its working model. Under equal randomization, its model-based variance estimator, surprisingly, remains consistent under model misspecification, justifying the use of confidence intervals output by standard software. These results hold under both simple and stratified randomization, and serve as an important causal inference justification for linear mixed models. Caveats and extensions of our findings will also be mentioned.

For more information, visit https://prevention.nih.gov/education-training/methods-mind-gap/toward-causal-inference-cluster-randomized-trials-estimands-and-reflection-current-practice.