New designs for Bayesian adaptive cluster-randomized trials

Abstract: Adaptive approaches, allowing for more flexible trial design, have been proposed for individually randomized trials to save time or reduce sample size. However, adaptive designs for cluster-randomized trials in which groups of participants rather than individuals are randomized to treatment arms are less common. Motivated by a cluster-randomized trial designed to assess the effectiveness of a machine-learning based clinical decision support system for physicians treating patients with depression, two Bayesian adaptive designs for cluster-randomized trials are proposed to allow for early stopping for efficacy at pre-planned interim analyses. The difference between the two designs lies in the way that participants are sequentially recruited. Given a maximum number of clusters as well as maximum cluster size allowed in the trial, one design sequentially recruits clusters with the given maximum cluster size, while the other recruits all clusters at the beginning of the trial but sequentially enrolls individual participants until the trial is stopped early for efficacy or the final analysis has been reached. The design operating characteristics are explored via simulations for a variety of scenarios and two outcome types for the two designs. The simulation results show that for different outcomes the design choice may be different. We make recommendations for designs of Bayesian adaptive cluster-randomized trial based on the simulation results.