Household selection in the absence of a sampling frame in community-based surveys – comparing satellite image-based approaches for a survey on snakebite envenoming in rural Malawi

Abstract Background Household-level survey data are essential for assessing population health, but creating sampling frames remains challenging in rural areas without administrative data. In preparation for a large community-based snakebite survey, we compared satellite-based methods for generating household sampling frames in such settings. Methods In June 2023, we selected and classified six areas in Neno and Mwanza districts, Malawi, to create sampling frames using three approaches: manual building identification from satellite images, building extraction from OpenStreetMap and artificial intelligence Open Buildings. Between July and December 2024, we completed a building census in the six areas, grouping residential buildings into household compounds, and compared census-derived compounds with the three sampling frames. In January 2025, we repeated the manual building identification using updated satellite images and compared the resulting sampling frame with the census. Results Of 916 buildings, only 22.9% ( n = 210) were identified in all three sampling frames. The OpenStreetMap sampling frame performed well in Neno district, but poorly in Mwanza, identifying only 32 buildings. During the census, 1268 buildings were identified, grouped into 428 compounds. The manual sampling frame contained 80.4% (344/428) of compounds, followed by Open Buildings (76.4%, n = 327). The proportion of compounds differed across districts and areas, with OSM again performing better in Neno, where a recent, crowd-sourced mapping exercise had been conducted, than in Mwanza. Using the updated manual sampling frame, compound coverage rose from 80.4% to 92.5%, with the largest gains in areas with longer gaps between satellite images. Conclusion Our study demonstrates that manual digitization, OSM, and AI-generated databases can provide viable sampling frames for rural household surveys in LMICs. However, method selection should consider area, resources, and required coverage, with local ground validation recommended to ensure completeness and minimise household selection bias.