Session summary
This GeoField 2026 methods session introduces core research design concepts for using Earth observation data in credible impact evaluation. The session focuses on counterfactual thinking, causal inference, spatial difference-in-differences, and randomized controlled trial design.
Moderator: Katherine Nolan
Causal Inference and Counterfactuals in Earth Observation Research
Anthony D’Agostino presents an introduction to causal inference and counterfactuals for Earth observation research. The chapter explains why correlation is not causation and introduces the central evaluation question: what would have happened to a person, plot, community, or country in the absence of an intervention? The presentation situates Earth observation within standard evaluation methods, including randomized controlled trials, matching, difference-in-differences, regression discontinuity, and synthetic control, while emphasizing that satellite data can serve as an outcome, treatment measure, covariate, or primary data source.
Spatial Differences-in-Differences
Ajay Shenoy presents a chapter on spatial difference-in-differences. The presentation explains how researchers can estimate treatment effects when policies, shocks, or events affect some places but not others, using data before and after the intervention. It covers the logic of parallel trends, spatial comparison groups, treatment geography, spillovers, spatial dependence, placebo tests, and spatial inference. A running example on regional development policy in India shows how nighttime lights and spatial treatment definitions can be used in a practical evaluation workflow.
Integrating Remote Sensing and Randomized Controlled Trials
Kendra Walker presents a chapter on designing randomized controlled trials with remote sensing data. The chapter explains how satellite data can strengthen RCTs by increasing statistical power, expanding measurable outcomes, detecting spillovers, and extending observation beyond survey rounds. It also cautions that remote sensing can weaken a study if satellite-derived measures are noisy, poorly validated, or biased. Examples from crop residue burning in Punjab show why ground truth, validation data, and plot-level linkage data should be planned from the start.
Together, the session provides a methodological foundation for geospatial impact evaluation. The presentations emphasize that Earth observation is most powerful when paired with clear counterfactual logic, well-defined treatment and comparison groups, appropriate spatial units, careful measurement, and field data collected with geospatial analysis in mind.