Two weights make a wrong: Cluster randomized trials with variable cluster sizes and heterogeneous treatment effects

In cluster randomized trials (CRTs), the hierarchical nesting of participants (level 1) within clusters (level 2) leads to two conceptual populations: clusters and participants. When cluster sizes vary and the goal is to generalize to a hypothetical population of clusters, the unit average treatment effect (UATE), which averages equally at the cluster level rather than equally at the participant level, is a common estimand of interest. From an analytic perspective, when a generalized estimating equations (GEE) framework is used to obtain averaged treatment effect estimates for CRTs with variable cluster sizes, it is natural to specify an inverse cluster size weighted analysis so that each cluster contributes equally and to adopt an exchangeable working correlation matrix to account for within-cluster correlation. However, such an approach essentially uses two distinct weights in the analysis (i.e. both cluster size weights and covariance weights) and, in this article, we caution that it will lead to biased and/or inefficient treatment effect estimates for the UATE estimand. That is, two weights "make a wrong" or lead to poor estimation characteristics. These findings are based on theoretical derivations, corroborated via a simulation study, and illustrated using data from a CRT of a colorectal cancer screening program. We show that, an analysis with both an independence working correlation matrix and weighting by inverse cluster size is the only approach that always provides valid results for estimation of the UATE in CRTs with variable cluster sizes.