Abstract
In a recent wave of economic development and environmental program evaluations, the outcome of interest is remotely sensed rather than directly measured, through e.g., satellite or text data. An intuitive approach is to predict the outcome of interest from its remotely sensed variable (RSV), then to proceed as usual. We prove that this common practice is biased under reasonable assumptions. Our solution is simple: by noting that causal effects are channeled by the treatment and outcome to the remotely sensed variable, we provide a simple set of non-parametric identification results for the average treatment effects. Our results do not require researchers to know or consistently estimate the relationship between the outcome, treatment, and RSV, which is typically mis-specified with unstructured data.
By leveraging existing literature on methods of moments, we show how estimation and valid asymptotic inference can be easily implemented using a data-driven representation of RSVs. As an empirical illustration, we re-evaluate the efficacy of a large-scale public program in India, showing that the program's measured effects on local consumption and poverty can replicated using satellite imagery alone.
Local Organizer: Giovanni Angelini