Causal inference using satellite data and artificial neural networks.

Abstract:

Satellite imagery offers researchers the unprecedented ability to measure outcomes on the ground at high spatial and temporal frequency. This paper develops and validates a set of methods which take advantage of spectral, temporal, and spatial variation to measure outcomes and infer unobserved outcomes. I show how traditional time series methods as well as neural networks with recurrent and convolutional structures can accurately classify pixels and objects. I show how these measurement strategies naturally fit into cutting edge causal inference methods for very high-dimensional data. I use double machine learning methods to model treatment propensity and outcomes as functions of geophysical and human processes. Combined with the enormous amount of data encoded in the satellite record these methods allow researchers to measure causal relationships in settings where experiments are impossible. Finally I demonstrate these methods by testing whether a land-titling reform program in Benin resulted in productivity-enhancing improvements and changes to nearby ecosystems.