National Small-Area Population Estimation Modelling Using Partial Coverage Health Intervention Campaign Data

Small-area population data underpin efficient resource allocation but are often unavailable where censuses are outdated or absent. We present a generalizable geostatistical model-based computational technique for estimating population counts from limited, imperfectly observed health intervention campaign data. The approach embeds a robust data cleaning strategy while integrating geolocated population enumerations with satellite-derived building footprints and geospatial covariates within a Bayesian Hierarchical modelling framework. Using real data application and a simulation study implemented over a range of biasedness and missingness scenarios, we evaluated three data cleaning strategies and showed that the approach involving exclusion of biased samples from the training set, produced the smallest estimation errors in both simulation studies (1.5% to 65.7% reductions in relative mean absolute error (RMAE)) and when used in prediction modelling from recent geolocated malaria bednet campaign data in Nigeria (17.1% reduction in RMAE). This approach which enabled us to generate gridded and administrative unit-level population estimates (total national estimate ~ 237.4M; 95%CI 233.6M – 243.7M people) from imperfect operational data across Nigeria is transferable to other data-scarce settings.