Poverty Mapping in the Age of Machine Learning

Recent years have witnessed considerable methodological advances in poverty mapping,much of which has focused on the application of modern machine-learning approaches to remotely sensed data. Povertymaps produced with these methods generally share a common validation procedure, which assesses model performance bycomparing subnational machine-learning-based poverty estimates with survey-based, direct estimates. Althoughunbiased, survey-based estimates at a granular level can be imprecise measures of true poverty rates, meaning that it isunclear whether the validation procedures used in machine-learning approaches are informative of actual modelperformance. This paper examines the credibility of existing approaches to model validation by constructing apseudo-census from the Mexican Intercensal Survey of 2015, which is used to conduct several design-based simulationexperiments. The findings show that the validation procedure often used for machine-learning approaches can be misleadingin terms of model assessment since it yields incorrect information for choosing what may be the best set ofestimates across different methods and scenarios. Using alternative validation methods, the paper shows thatmachine-learning-based estimates can rival traditional, more data intensive poverty mapping approaches. Further, theclosest approximation to existing machine-learning approaches, using publicly available geo-referenced data,performs poorly when evaluated against “true” poverty rates and fails to outperform traditional poverty mapping methodsin targeting simulations.