Cluster trials inference with CARE

We show that public health cluster-randomised trials (CRTs) feature highly heterogeneous clusters — an underreported issue exacerbated by pragmatic design approaches. This heterogeneity is neglected in simulations and unacknowledged in practice, compromising the reliability of statistical inferences. Through simulations — both fully artificial and placebo regressions using real trial data — we demonstrate that, alarmingly, in the settings commonly used, up to 62% of CRTs report false-positive findings and that small-sample corrections deemed appropriate in the trial literature are not improving the quality of inference. To address this, we introduce the CARE (Clarify, Apply, Refine, Evaluate) protocol, which leverages partial identification principles to provide reliable statistical conclusions and a structured, assumption-aware framework for safe data exploration. The protocol preserves the convenience of pragmatic trial designs while ensuring their statistical validity and also maximises the scientific value of often costly trials by integrating robust, frequentist, and efficient (Bayesian) methods.