Estimation of fine-scale spatial heterogeneity in treatment effects in a large cluster-randomized trial

Treatment effects in randomized trials may vary geographically, and estimating this variation is a broadly generalizable problem for large-scale, cluster-randomized trials. We conducted a secondary analysis of AVENIR, a cluster-randomized, placebo-controlled trial of biannual azithromycin distribution among children aged 1-59 months in Dosso, Niger (2020-2023). We assessed whether the survival benefit of mass azithromycin distribution varied geographically using Poisson models to estimate spatially continuous child mortality and average treatment effects (ATE). Among 1390 villages (276,452 person-years; 3635 deaths), the overall mortality rate was 13.1 and the ATE was -2.6 deaths per 1000 person-years. Mortality was highest in the south, whereas treatment benefit was concentrated in the southeast, following a northwest-southeast gradient. Aggregation of data to 160 primary health catchment areas preserved broad spatial patterns in mortality but attenuated local heterogeneity, suggesting that health catchment monitoring is sufficient for broad mortality surveillance. Examining unit-level, spatially predicted treatment effects as a function of mortality rate demonstrated that underlying mortality captured much, but not all, of the heterogeneity in treatment effects. The analysis demonstrates how to estimate localized treatment effects in large cluster-randomized trials and provides new insights into spatial heterogeneity in mortality and benefits of azithromycin distribution in Niger.