A Two-Step Bayesian Hierarchical Modelling Approach for Estimating Population Density when Settlement Data Are Partially Observed

Chibuzor Christopher Nnanatu , University of Southampton
Amy Bonnie, University of Southampton
Josiah Joseph, National Statistical Office
Ortis Yankey, University of Southampton
Duygu Cihan, University of Southampton
Assane Gadiaga, University of Southampton
Mercedita Tia, United Nations Population Fund
Attila Lazar, University of Southampton
Andrew J. Tatem, University of Southampton
Marielle Sander, United Nations Population Fund

The demand for modelled population estimates has seen a considerable surge over the recent years because they provide rapid and up-to-date population estimates at finer spatial scales (which are often required to meet several humanitarian and sustainable development needs) than census projections. Model-based population estimation methods (PEMs) leverage advances in satellite imagery to produce small area estimates of population numbers by integrating satellite-based settlement data with other multiple geospatial data sources using advanced statistical techniques. However, satellite-based settlement datasets are susceptible to systematic bias which could lead to incorrect estimates because they are AI/machine learning classifications of imagery of where human settlements are detected, which means that under tree canopy and/or cloud cover settlement structures will be missed. The key motivation of this study is the need to produce reliable small area population estimates in Papua New Guinea (PNG), where many of the rural settlements are under tree canopy cover. Here, we present a novel two-step model solution which corrects for the potential bias in the settlement data in the first step, and estimates population numbers in the second step using the bias-corrected data. Simulation study results show that the two-step modelling solution caused a reduction in relative bias which varied between ~63% and ~88%. Moreover, the two-step model solution which was successfully applied to obtain subnational estimates of population in PNG and caused ~33% reduction in relative bias, provides an important framework for model-based bias correction within the contexts of population estimation and could be extended to other settings.

See extended abstract

 Presented in Session P3. Migration, Economics, Policies, History