Speaker
Description
Granular and repeated measurements of socio-economic outcomes are essential for understanding and improving economic livelihoods, but detailed survey data is scarce, especially across low-and-middle income countries. Recent machine learning approaches successfully leverage non-traditional data such as remote sensing data to predict consumption expenditures and asset wealth across locations but not their variation over time. This work systematically investigates the potential of public satellite imagery, combined with environmental and infrastructure data to predict levels and fluctuations in consumption expenditures and asset wealth and their use for policy evaluation. We present a framework that integrates spatial and temporal models to predict socioeconomic outcomes. Models trained on panel data from five African countries between 2007 and 2021 explain up to 73% of spatial variation but struggle to explain temporal variation (R$^2\approx$0.14). Utilizing ML-predicted outcomes, we are able to recover the sign, though not the magnitude of treatment estimates in an exemplary policy evaluation in Nigeria. These results underscore the potential of non-traditional data for policy evaluation while highlighting its current limitations.
Keyword | Research Methods |
---|