Preference-Informed Cluster Randomized Design for Pragmatic Clinical Trials.

Cluster randomized trials (CRTs), in which entire clusters of subjects are randomized to treatment arms, are widely used in pragmatic trials to evaluate interventions under real-world conditions. However, CRTs are particularly vulnerable to treatment non-adherence, especially when cluster-level preferences lead subjects in clusters to deviate from their assigned treatment. Such deviations can reduce power, introduce bias, and compromise generalizability if not properly addressed. This research is directly motivated by a planned multi-center trial in Kawasaki Disease patients with high risk for coronary artery abnormalities, in which institutional treatment preferences influence both willingness to participate and adhere. To address this issue, we propose a Bayesian hierarchical model under a Preference-Informed Cluster Randomized Design (PICRD). This model explicitly incorporates cluster-level treatment switching into the analysis rather than excluding non-willing or non-adherent clusters. We conduct a simulation study to evaluate the performance of the PICRD model across a range of treatment effect sizes and switching proportions. Results demonstrate that the PICRD model consistently outperforms per-protocol analyses by maintaining higher power for the main treatment effect, producing narrower 95% credible intervals, and yielding more stable bias and root mean square error in the presence of substantial non-adherence. By explicitly modeling preference within a Bayesian hierarchical framework, the PICRD approach provides a flexible and robust solution for CRTs conducted in pragmatic settings when willingness to accept randomization assignment or adherence to randomization is often unrealistic.