Spatial heterogeneity in machine learning-based poverty mapping: Where do models underperform?

• Machine learning-based poverty mapping underperforms due to spatial heterogeneity. • In interpolation, geographically weighted models reveal variation across regions. • In extrapolation, models overestimate welfare in poor, rural, single-sector regions. • Spatial models yield limited gains in underperforming areas. • Unbiased poverty maps require improved training data and remote sensing proxies. Accurately locating poor populations is increasingly urgent as global poverty reduction has stalled under the combined pressures of conflicts, climate shocks, rising food prices, pandemics, and growing inequality. Recent studies harnessing geospatial big data and machine learning (ML) have significantly advanced poverty mapping, enabling granular and timely welfare estimates in traditionally data-scarce regions. While much of the existing research has focused on overall out-of-sample predictive performance, there is a lack of understanding regarding where such models underperform and whether key spatial relationships might vary across places. This study investigates spatial heterogeneity in ML-based poverty mapping in East Africa, testing whether spatial regression and ML techniques produce more unbiased predictions. We find that extrapolation into unsurveyed areas suffers from biases that spatial methods do not resolve; welfare is overestimated in impoverished regions, rural areas, and single sector-dominated economies, whereas it tends to be underestimated in wealthier, urbanized, and diversified economies. Even as spatial models improve overall predictive accuracy, enhancements in traditionally underperforming areas remain marginal. This underscores the need for more representative training datasets and better remotely sensed proxies, especially for poor and rural regions, in future research related to ML-based poverty mapping. For development agencies, the findings caution against treating ML-based outputs as neutral or universally reliable, highlighting instead the need to pair technical advances with investments in inclusive data collection, integration of spatial theory, and institutional strategies that address structural data inequalities.