We develop a novel methodology that uses machine learning to produce accurate estimates of consumption per capita and poverty in 10x10km cells in sub-Saharan Africa over time. Using the new data, we revisit two prominent papers that examine the effect of institutions on economic development, both of which use “nightlights” as a proxy for development. The conclusions from these papers are reversed when we substitute the new consumption data for nightlights. We argue that the different conclusions about institutions are due to a previously unrecognized problem that is endemic when nightlights are used as a proxy for spatial economic well-being: nightlights suffer from nonclassical measurement error. This error will typically lead to biased estimates in standard statistical models that use nightlights as a spatially disaggregated measure of economic development. The bias can be either positive or negative, and it can appear when nightlights are used as either a dependent or an independent variable. Our research therefore underscores an important limitation in the use of nightlights, which has become the standard measure of spatial economic well-being for studies focusing on developing parts of the world. It also demonstrates how machine learning models can generate a useful alternative to nightlights, with important implications for the conclusions we draw from the analyses in which such data are employed.
