Contamination-Resilient Cluster Randomized Trial Design under Spatial Interference: A Simulation Study of Contamination Control

Between-arm contamination, cross-boundary mixing of infection risk, biases cluster randomized trials in spatially-heterogeneous settings toward the null. We define the estimand under spatial interference, the boundary-insulated effect, and rank allocations on a composite of balance, topology, and contamination; simulations use a two-criterion reduction (flux with balance), showing balance alone does not control contamination. On real Ugandan Malaria Atlas Project survey data, in simulated trials allocating individual sites at moderate contamination, two analyses that ignore spillover, naive and boundary-censored (most-exposed sites dropped), are biased, so their coverage, how often 95% intervals contain the true effect, is low; the two-criterion design reduces their bias, lifting boundary-censored coverage from 26–29% under simple randomization to 28–48%, across eight candidate pools. A third, spillover-adjusted analysis has about 95% coverage under every design if its mobility kernel is correct, and the design cuts its root-mean-square error (RMSE) 36%; with a wrong kernel its coverage drops to 51–89%, while the design's bias reduction holds. These figures assume no between-site variance; at intracluster correlation 0.02–0.05, binomial intervals are too narrow even without contamination, site-level analysis stays near 95%, cluster-level analysis comes close, and site-allocation boundary-censored coverage gains persist; RMSE gains shrink at 0.05. Whole-cluster allocation raises coverage, but removes the two-criterion design's advantage over uniform randomization, and the ten-metric composite beats same-resolution random allocations only at strong contamination without between-cluster variance. Site-level findings recur on Yemen cholera and Louisiana COVID-19 geographies. We claim reduced exposure and measured residual bias, not the absence of interference. Preprint; not peer reviewed. Code and data: https://doi.org/10.17605/OSF.IO/XPA4S