Economic development in pixels
(with John D. Huber)
We describe a novel and easily implementable methodology for generating estimates of per capita consumption in 10x10 km cells over time. The first step in this methodology is based on a statistical framework for translating asset indices from surveys into measures of consumption. The framework allows us to develop a training variable for machine learning models, which we use in a second step to create estimates of consumption for 42 sub-Saharan African countries. These new data allow us to highlight a fundamental problem with using “nightlights” – a de facto proxy for local economic well-being – in statistical models. We do so by revisiting two prominent papers that examine the effect of institutions on economic development, both of which rely on nightlights. The conclusions from these papers are reversed when we substitute our consumption-based measure for nightlights, and we show that this reversal is due to the nonclassical measurement error that is endemic to nightlights in regions where large swaths of territory appear unlit, as is common in rural Africa. This error can introduce unpredictable biases – either attenuating or amplifying – into statistical models that use nightlights as a measure of spatial economic performance. In contrast to nightlights, our methodology makes it possible to remove non-classical measurement error from the consumption estimates, and it produces estimates denominated in a metric that is well-understood. The approach described here therefore provides a promising way forward in the study of a wide range of questions and policies that have so far relied on nightlight data to track economic progress.
