ML-based geographic sampling frames miss transitory populations in fragile regions

Survey research on conflict and displacement depends on the reliability of sampling frames. We show that common approaches to developing these frames may underrepresent populations central to this research: displaced people, returnees, and civilians exposed to violence. We develop a hybrid approach to sampling frame generation that combines machine-learning (ML)-generated building footprints with satellite imagery. We test the approach in displacement-affected Iraqi communities. Our approach achieved 87% residential accuracy overall, and our data reveal non-random omissions in the common ML-only approach. We find systematic coverage differences across frame-generation methods. Manual methods capture more rural internally displaced persons and urban returnees, due to informal shelters and wartime reconstruction. Our hybrid ML-and-satellite sampling can mitigate coverage error and improve inference about conflict and displacement.

Keywords: machine learning; building footprints; sampling frames; survey methodology; displacement; internally displaced persons; post-conflict; satellite imagery; Iraq; humanitarian assistance.