Introducing the Planetary Experiments Repository
Randomized controlled trials (RCTs) long ago proved themselves as workhorses of rigorous social science, medical research, and policy evaluation. Yet, they can struggle to account for the spatial and contextual heterogeneity of real-world settings: there are so many factors that make each RCT setting unique. While each RCT is self-contained--focusing on carefully defined treatments, outcomes, and sampling frames--those who stand to gain the most from policy interventions often inhabit remote or under-documented regions. An essential question thus arises: can experimental results be scaled from a few local settings to a global framework? And can new data sources--like satellite imagery--refine inferences about who benefits most from what kinds of programs?
The repository detailed here is motivated by our Planetary Causal Inference (PCI) project. This project aims to link satellite and geospatial data with RCT-based causal inference to illuminate both local intervention effects and the broader contexts that shape their success.
To see why this matters, one must recognize the knowledge void that remains when only conventional data—surveys, administrative records—are employed. Village roads, seasonal cropping patterns, density of tree cover, or the shifting geometry of urban sprawl: these features are often left out of “standard” RCT covariate sets, yet they can condition whether an intervention succeeds or fails. Planetary Causal Inference places geo-located trials into conversation with Earth observation (EO) data, at scale, to estimate how treatments intersect with place and time.
The Problem of Missing Context
An RCT’s immediate strength is that it identifies an Average Treatment Effect (ATE) in a specific population. Run a test, randomize an intervention, measure the difference from control---and we're done. The trouble is that these experiments often capture only a narrow strip of a multidimensional reality. Ask yourself: do we truly know why certain households fail to benefit from a livelihood program, or why certain villages respond more robustly to an educational campaign? Without deeper context, researchers can at best guess: maybe it’s the soils, the roads, the climate, or a thousand other factors.
Here satellite data become surprisingly valuable. Thanks to continuous public missions (e.g., NASA–USGS Landsat, ESA’s Sentinel) and commercial providers, we have historical snapshots of the Earth’s surface for virtually any location. Think of it as a Library of Alexandria of the past: we can see rivers that might influence a sanitation study, farmland that might mediate a poverty-reduction intervention, forest density that might shape a cash-transfer’s impact on livelihood. If you geo-locate your RCT properly---so that each site or cluster is linked to its precise lat/long coordinates---these images become more than mere pictures: they form high-dimensional, quantitative records of environment and infrastructure. That in turn can help us understand and forecast heterogeneous treatment effects. [1]
Heterogeneity: Local and Planetary
Now, suppose your RCT involves a new microfinance scheme. In one location, poor roads might hamper the scheme’s benefits. In another, vibrant local markets amplify the returns. The satellite perspective lets us measure features like road density or market-town expansion, potentially distinguishing precisely which sites see bigger gains--and why.
But, for most policymakers and researchers too, there’s a greater ambition: generalizing beyond the sample. A typical RCT, however large, is physically local. You might want to know: would my “success” replicate in another district? Another country? The Planetary Causal Inference approach harnesses geospatial coverage that scales globally. With properly geo-located RCT units—where the data is aggregated at the village or ward level to avoid privacy leaks--one can train machine learning models that learn from the interplay between local contexts (visible in remote sensing) and measured outcomes. Then, in principle, these same models can be applied across entire countries or continents to produce predictions or “transported” effect estimates. It’s not a magic wand--external validity is never guaranteed--but the data framework is there.
The Repository: A Catalog of Experiments and Geo-Links
To help us and others towards these ends, we're assembling a repository of RCT studies that exemplify the PCI principles---where the location data is well-specified, with distinct treatment and control groups, and enough spatial spread to interface with satellite imagery. The examples range from Banerjee et al.’s multifaceted anti-poverty interventions across Honduras, Peru, and Ghana to the grant program from Blattman et al. in Uganda and more. Each experiment offers well-defined sites, which can then be matched with Landsat or other imagery. The notion is that, once these RCTs are pinned to their lat/long coordinates, researchers can systematically incorporate environmental or infrastructural data at scale.
The repository also outlines minimal criteria for inclusion: is the trial’s treatment geo-located? Are sample sizes nontrivial? Did the authors provide publicly or privately requestable data on village or ward boundaries? Satisfying these criteria means that the experiment can be prime material for bridging the gap between local randomization and planetary context.
Encoding Multi-Level Dynamics
As you start to work in the PCI framework, some complexities will come up. For example, the practical mechanics of linking RCT data with satellite imagery get complicated. As you scale from the ground-level (households) to neighborhoods, wards, or entire districts, the image patches that represent them can overlap substantially. That can be a nuisance (if the model sees repeated data) or a feature (if you believe that certain images represent truly shared contextual factors). The highlight is that you can glean both “close” and “far” scaled perspectives, capturing localized features (condition of a household’s roof, adjacency to farmland) and more global features (distance to bigger cities, connectivity by major roads). Multi-scale methods let you attach multiple concentric images to each RCT unit, thereby capturing local texture and broader environment. [2]
Toward a Broader Vision
All these methods target a deeper question: How do we unify separate RCTs, each with its own objectives, sample definitions, and logistical constraints, under one planetary or near-planetary lens? We don’t just want a mosaic of single-sample analyses; the aspiration is to see causal learning accumulate.
No single “silver bullet” solution exists. Different computing architectures---Vision Transformers, multi-scale CNNs, or even simpler random forest approaches---offer distinct trade-offs between interpretability and predictive capacity. Earth observation itself isn’t uniform: sentinel satellites differ from Landsat in their revisit frequencies, spectral bands, and resolution. The repository is, thus, a demonstration that we must carefully pick the right data for the right RCT, mindful of possible biases. [3]
What you see in this Planetary Experiments Repository is a first step: a collection of RCT data that meet threshold criteria for meaningful satellite linkage, plus references to the original authors’ data portals. These references show exactly where to obtain the data and how to request access (household-level coordinates are often restricted). Yet the real payoff is not merely in having a curated list—it’s in enabling new lines of research that combine the power of random assignment with the omnipresence of satellite records to parse out the how, where, and when of treatment effect variation.
Conclusion: Building a Planetary-Scale Evidence Base
Causal inferences can be strong at the local scale (the RCT’s classic domain) but then have difficulty in broader generalization. Satellite imagery and geospatial data, read with modern ML algorithms, close part of this gap by providing environmental “fingerprints” that capture the contexts in which interventions operate. So the repository described is an open invitation: we can do more than produce localized average effects. We can examine how programs might fare in contexts with similar or different terrain, climate, infrastructure, or land use patterns. That means scaling from a single site’s evidence to insights that stand a chance of traveling across regions, countries, or continents.
Such a planetary perspective on causal inference--blending rigorous experiment design with vantage points from space--represents a step toward a truly global evidence base, one that acknowledges local specificities, tracks changes over time, and recognizes that solutions to society’s pressing challenges need not remain locked within the confines of a single randomized trial. It’s about weaving local experiments into a planetary tapestry, so that policymakers, researchers, and citizens can jointly understand where an intervention works best, how it might improve, and why.
[RCT Repository Link]