Program Evaluation With Remotely Sensed Variables

Economists often use remote sensing variables (RSVs), such as satellite imagery, to estimate treatment effects in experiments when direct economic measurements are lacking. The usual method of training a predictor based on an observational sample and using its predictions as outcome measures is biased when RSVs themselves are influenced by economic outcomes.

The authors show that the treatment effect can be identified nonparametrically if one assumes that the distribution of RSVs remains stable across samples, given the outcome and treatment. Their identification formula requires predictions for the outcome, treatment, and sample indicator, rather than just the outcome, thus enabling valid inferences without placing particular demands on predictive accuracy.

Finally, they demonstrate their approach using a re-evaluation of an Indian poverty reduction program with satellite imagery.

Davide Viviano is an Assistant Professor at the Department of Economics at Harvard. His research combines economics and data science to develop or justify statistical methods for social-science applications, with a focus on policy design and causal inference.