Abstract:
|
In a cluster randomized trial (CRT), there are several individuals per cluster, and clusters are the unit of randomization. CRT are commonly used to evaluate interventions that can only be administered at a facility level. An example is the LARK trial, which was designed to compare the use of smartphone notifications for notifying new patients in cardiovascular care about their upcoming appointments. The trial evaluated two sequential outcomes: linkage to care, indicated by patients attending their first visit, and 12 month systolic blood pressure. The primary outcome was SBP; a secondary analysis examined the mediating effect of linkage to care. In this paper, we describe key ways in which cluster-randomized designs can affect the way mediation analyses are carried out. We propose a Bayesian approach to mediation analysis where the distributions of both the mediator and the outcome are essentially nonparametric, and cluster variation is captured using random effects. We illustrate the methods by presenting an analysis of the LARK study.
|