Replication Data for: Integrating Earth Observation Data into Causal Inference: Challenges and Opportunities

10.7910/DVN/QLCSVRJerzak, ConnorConnorJerzak0000-0003-1914-8905University of Texas at AustinReplication Data for: Integrating Earth Observation Data into Causal Inference: Challenges and OpportunitiesHarvard Dataverse2023Computer and Information ScienceEarth and Environmental SciencesSocial SciencesCausal inference; Observational studies; Computer visionJerzak, ConnorConnorJerzakUniversity of Texas at Austin2023-09-122025-01-132301.129852004048102125738634209619253636699941054868922application/zipapplication/zipapplication/zipapplication/ziptext/markdowntext/csv4.1CC0 1.0Abstract: Observational studies of global development require adjustment for confounding factors correlated with both treatment and outcome. Many methods have been developed for the setting where the confounders are tabular quantities such as neighborhood income. However, community-level data for global analysis are often scarce. In this context, satellite imagery can play an important role, proxying for confounding variables that would otherwise remain unobserved. We study confounder adjustment in this setting, where patterns contained in satellite images contribute to the confounder bias. We formalize challenges of observational inference at a global scale with satellite images---what conditions are sufficient to identify effects, how to perform estimation, and how to quantify which aspects of images most contribute to confounding. Via simulation, we explore sensitivity of satellite-based inference to image resolution, model misspecification, and assumption violations. Finally, we demonstrate our framework by estimating the impact of anti-poverty interventions in African communities using satellite imagery. GitHub: https://github.com/cjerzak/causalimages-software