Reconstructing subdistrict-level population denominators in Yemen after six years of armed conflict and forced displacement.

Introduction: Yemen has experienced widespread insecurity since 2014, resulting in large-scale internal displacement. In the absence of reliable vital events registration, we tried to reconstruct the evolution of Yemen's population between June 2014 and September 2021, at subdistrict (administrative level 3) resolution, while accounting for growth and internal migration.

Methods: We reconstructed subdistrict-month populations starting from June 2014 WorldPop gridded estimates, as a function of assumed birth and death rates, estimated changes in population density, net internal displacement to and from the subdistrict and assumed overlap between internal displacement and WorldPop trends. Available displacement data from the Displacement Tracking Matrix (DTM) project were subjected to extensive cleaning and imputation to resolve missingness, including through machine learning models informed by predictors such as insecurity. We also modelled the evolution of displaced groups before and after assessment points. To represent parameter uncertainty, we complemented the main analysis with sensitivity scenarios.

Results: We estimated that Yemen's population rose from about 26.3 M to 31.1 M during the seven-year analysis period, with considerable pattern differences at sub-national level. We found that some 10 to 14 M Yemenis may have been internally displaced during 2015-2016, about five times United Nations estimates. By contrast, we estimated that the internally displaced population had declined to 1-2 M by September 2021.

Conclusions: This analysis illustrates approaches to analysing the dynamics of displacement, and the application of different models and data streams to supplement incomplete ground observations. Our findings are subject to limitations related to data quality, model inaccuracy and omission of migration outside Yemen. We recommend adaptations to the DTM project to enable more robust estimation.