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. Poverty maps produced with these methods generally share a common validation procedure, which assesses model performance by comparing sub-national poverty estimates with survey-based, direct estimates. While unbiased, direct estimates can be imprecise measures of true poverty rates, meaning that it is unclear whether these validation procedures are informative of actual model performance. In this paper, we use a rich dataset from Mexico to provide a more rigorous assessment of the modern approach to poverty mapping by evaluating its performance against a credible ground truth. We find that the modern method under-performs relative to benchmark traditional methods, largely because of the limited predictive capacity of remotely-sensed covariates. For a given covariate set, we also find that machine learning produces more biased poverty estimates than the traditional procedures, particularly for the poorest geographic areas.