Cluster trials inference with CARE

Author-formatted postprint of the article published in Statistics in Medicine. The authoritative version of record is available at: https://doi.org/10.1002/sim.70610 This postprint incorporates the final author corrections and is archived as the final version of the manuscript. We show that cluster-randomised trials, especially pragmatic designs, often exhibit substantial but underappreciated heterogeneity in cluster sizes and structures, distorting inference. Our simulations, reassigning treatment in real data and varying imbalance in synthetic data, show that currently recommended methods such as targeted maximum likelihood estimation and small-sample corrected generalised estimating equations are not optimised to this challenge. We propose the CARE (Clarify, Apply, Refine, Evaluate) protocol, which anchors inference in a design-based benchmark and provides a principled pathway for incorporating assumption-rich methods, making trial analysis more credible, transparent, and directly comparable across studies.