Spatial Heterogeneity in Machine Learning-Based Poverty Mapping: Where Do Models Underperform?

Recent studies harnessing geospatial big data and machine learning 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 machine learning-based poverty mapping, testing whether spatial regression and machine learning 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 machine learning-based poverty mapping.