🤖 AI Summary
This paper addresses inaccurate statistical inference in cluster-randomized trials with matched pairs. We propose an exact inference method that jointly accounts for intra-cluster correlation and pair-level matching structure. Unlike conventional mixed-effects models relying on strong distributional assumptions, our approach extends the randomization inference framework to matched-pair cluster trials for the first time, establishing a unified procedure comprising permutation testing, pair-constrained resampling, and within-cluster robust variance estimation. The method imposes no parametric assumptions on the intra-cluster correlation structure and accommodates diverse matching schemes, including size-based matching. Simulation studies and empirical applications demonstrate that the proposed method rigorously controls Type I error rates, achieves substantially higher statistical power, and reduces average estimation bias by 42% relative to existing methods.
