Assessing the accuracy of census-independent small area modelled population datasets

Abstract Small area population estimates are fundamental for effective decision-making across government, humanitarian, development and commercial sectors. Many countries face growing uncertainty in their population figures due to delayed or incomplete censuses – an issue highlighted during the 2020 round, with many censuses delayed and 15% of countries ultimately not undertaking one. To address this gap, methods to provide census-independent or “bottom-up” modelled population estimates have been developed that combine geospatial data derived from satellite imagery with population data samples within a statistical modelling framework. Such methods have been applied in collaboration with governments in more than ten countries to date, typically in data-scarce settings where opportunities for external validation are limited. An exception is Zambia, where census-independent modelled population estimates were developed in 2020, and a national census subsequently conducted in 2022. Using data from the Zambia 2022 census, this study presents the first national comparison of census-independent modelled estimates with recent census enumeration. We directly compare the modelled estimates with census counts across geographic scales and explore subsets of estimates considered to most plausibly represent residential populations. Our findings show that the modelled estimates are strongly correlated with the 2022 census (R>0.95 for wards and districts) but tend to exceed census counts. Accounting for residential census household locations reduces these disparities considerably, highlighting the need for improved stratification of built settlement types in model input data. This study strengthens the evidence base for bottom-up population modelling as an option where there are barriers to conducting a census, whilst also identifying priorities for methodological development.