Can satellite-based predictions substitute for traditional outcome measurements in program evaluation? Using forest cover data from the Democratic Republic of the Congo, Brazil, and Indonesia, we conduct semisynthetic simulations comparing estimation methods. The RSV estimator (Rambachan, Singh, and Viviano 2025) formalizes the postoutcome structure of remotely sensed data—changes in forest cover cause changes in satellite imagery—delivering approximately unbiased treatment effects with correct coverage. Some alternative approaches exhibit significant bias despite highly accurate pretrained predictors. When limited validation data are available, the RSV estimator efficiently incorporates observational samples while remaining robust to distribution shift, achieving meaningful reductions in standard errors.
AlSharif et al. (Fri,) studied this question.