Causal Inference with Predicted Outcomes: Correcting prediction error bias in satellite-based impact evaluation

Recent advances in Earth observation and machine learning open new frontiers in impact evaluation that appear well-suited for agricultural settings. We probe the nuances and limitations of these promising methods using satellite-predicted maize yields to measure the impact of Ethiopia’s Direct Seed Marketing (DSM) program, which was rolled out after 2011 to enhance farmer access to improved seed varieties. Machine learning-generated outcomes inevitably introduce prediction error, which attenuates coefficients and understates standard errors in downstream causal analysis. We find positive effects of DSM on crop cut maize yields, but weaker effects when using (naive) predicted yields. We demonstrate how ground-truth data can be leveraged to diagnose and correct this bias using tools from the prediction-powered inference literature — albeit with additional assumptions about unobservable prediction errors. Our corrected estimates suggest substantial DSM impacts on satellite-predicted maize yields in all producing areas and years.