Distinguishing Between Cross- and Cluster-Level Mediation Processes in the Cluster Randomized Trial

For the cluster randomized trial, the treatment, implemented at the cluster level, may be hypothesized to impact a final outcome of interest via a mediating variable, with the mediator and outcome being measured typically at the participant level. Two opposing views—one that allows the mediation process to flow through a participant-level mediator and one that does not—have been presented in the literature. We integrate these perspectives and argue that it is the theoretical model and the nature of the mediating variable that determines whether the treatment effect may be mediated by a participant-level or only a cluster-level mediator in this design. An accompanying simulation study shows that when the mediation process is unnecessarily restricted to the cluster level, the power to detect the presence of mediation is substantially reduced.