Planetary Causal Inference (PCI) represents an emerging field that combines satellite imagery and planetary-scale data sources with localized studies, especially RCTs, to derive insights about causality.
Methodologies developed in this space enable researchers to understand cause-and-effect relationships across multiple scales, from household to village and beyond. The Planetary Causal Inference book is now forthcoming (Cambridge University Press).
[Aims] [Jobs] [PCI Book]
[Concept Video] [FAQ]
[Tutorial@IC2S2] [Gallery]
[Bibliography]
Research Directions: Exploring Environment, Society, and Economy Through Earth Observation
Treatment Effect Heterogeneity
While traditional approaches to understanding intervention effects rely on demographic variables such as age and gender, satellite imagery provides new insights into effect heterogeneity. This enriched understanding of treatment effect variation across different contexts can improve our ability to predict intervention effects in new areas. [1] [2] [3] [Tutorial]
Transportability and Generalizability
Satellite imagery also facilitates the modeling of effect transportability across different geographical contexts. This analysis can be conducted both conditionally—for specific subpopulations—and unconditionally—for entire regions. Such capabilities enable researchers to predict intervention effectiveness in new areas using new, unified information sources. [1] [2]
Confounder Modeling
Remote sensing data serves as a proxy for confounding variables within social, environmental, and economic systems. This approach allows researchers to control for factors that correlate with both interventions (e.g., aid programs) and outcomes (e.g., development levels). Satellite imagery provides a holistic view of these confounding factors that might otherwise be difficult to measure directly. [1] [2] [3] [Tutorial]
Natural Experiments and Earth Observation
Our ongoing research explores how features derived from Earth observation data can enhance the reliability and robustness of natural experiments. This integration of satellite data with experimental design strengthens causal inference methodologies.
Outcome Generation Using Satellite Data
Earth observation data enables the generation of outcome metrics for downstream causal analysis. For instance, researchers can leverage satellite imagery to assess conflict outcomes, providing valuable input for subsequent causal analyses, alongside measurement error adjustments. [1] [2] [3] [Tutorial]
Posts
Scaling Laws for Multi-scale Dynamics
Policymakers and researchers analyzing randomized controlled trials (RCTs) increasingly harness satellite imagery to capture the geographical, socioeconomic, or environmental covariates believed to moderate an intervention’s impact. Some of those moderating factors might be locally confined—such as...Continue reading
New Directions in Planetary Causal Inference
Planetary causal inference (PCI) is a field where we draw on experimental, observational, or natural experimental data from social, natural, and environmental sciences and link those sources to satellite imagery or other Earth Observation (EO) data....Continue reading
Earth Observation Meets Adam Smith
Satellite imagery is contributing to how we measure and understand global poverty. From miles above Earth, high-resolution photographs capture details of villages, farms, and cities that conventional surveys might have a hard time capturing. By analyzing these...Continue reading
Introducing the PCI RCT Repo
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....Continue reading
About
PlanetaryCausalInference.org was created by Connor T. Jerzak to showcase the work of researchers, many in the AI & Global Development Lab, involved in work related to Planetary Causal Inference, as well as to provide information about a book project (with Adel Daoud) of the same name.
Planetary Causal Inference (PCI) FAQ
Answers about planetary causal inference—the intersection of causal inference, Earth Observation (EO) / satellite imagery, and AI systems.
Jump to:
- What is PCI?
- What is EO data?
- PCI vs. prediction vs. standard causal inference
- When should you use satellite imagery?
- Do you need an RCT?
- ATE vs. CATE (treatment effect heterogeneity)
- Transportability & generalizability
- Image-based confounding
- Avoiding post-treatment leakage
- What is a Remote Audit?
- What are multi-scale representations?
- Where to find data, code, and tutorials
- The PCI book
- Collaboration & contributing experiments
What is Planetary Causal Inference (PCI)?
Planetary causal inference (PCI) combines causal inference with Earth Observation (EO) data—especially satellite imagery—to estimate cause-and-effect across space and time.
It links planetary-scale measurements to localized studies (including RCTs, quasi-experiments, and observational designs) to explain what caused what, where, and why.
- Common synonyms: EO-based causal inference, remote sensing causal inference, geospatial causal inference.
- Typical outputs: average effects, heterogeneous effects, and mechanism-oriented evidence.
What counts as “Earth Observation (EO) data” in PCI?
EO data includes satellite imagery (e.g., Landsat/Sentinel-style sensors) and other planetary-scale geospatial products aligned to time and place.
In PCI, EO data can be used as pre-treatment covariates, outcomes, or mechanism proxies—but only when the timing matches the causal design.
- Why EO helps: consistent coverage across regions, years, and (often) countries.
- What EO doesn’t replace: identification assumptions and research design.
How is PCI different from remote sensing prediction (and from “standard” causal inference)?
Remote sensing prediction maps correlations, while PCI uses causal designs to estimate counterfactual effects and their variation across space.
Compared with tabular-only causal inference, PCI adds high-dimensional, spatially aligned EO signals that capture historical and geographic context.
| Approach | Main goal | Typical output | Common pitfall |
|---|---|---|---|
| Remote sensing prediction | Forecast / map outcomes | Risk maps, predicted indicators | Correlation mistaken for causation |
| Standard causal inference | Estimate causal effects | ATE/CATE estimates, uncertainty | Missing or biased covariates |
| Planetary causal inference (PCI) | Causal effects + mechanisms across space | Effects + heterogeneity/transportability evidence | Leakage, spatial dependence, domain shift |
When should you use satellite imagery in a causal study—and when shouldn’t you?
Use EO imagery when units are geolocated and the “context” that drives outcomes or treatment assignment is visible on the landscape.
Avoid imagery when you can’t time-align it cleanly (pre vs. post), or when imagery is likely to encode post-treatment effects that would bias identification.
- Best for: heterogeneity, generalizability across places, and confounder measurement at scale.
- Not best for: designs where the only relevant confounders are purely social/organizational and leave no spatial trace.
Do you need a randomized controlled trial (RCT) to do planetary causal inference?
No—PCI includes RCTs, natural experiments, quasi-experiments, and observational studies that can defend identification assumptions.
RCTs are a clean starting point (especially for effect heterogeneity and transportability), but PCI is ultimately “design-first,” not “data-first.”
- What matters most: a credible causal estimand + identification strategy + robustness checks.
What’s the difference between ATE and CATE (treatment effect heterogeneity) in PCI?
ATE is the average impact of an intervention, while CATE asks how that impact changes by context (treatment effect heterogeneity).
PCI uses EO imagery to measure context (infrastructure, land use, settlement patterns, environmental conditions) that can explain where effects are larger, smaller, or reversed.
Example estimand: \text{ATE}=\mathbb{E}[Y(1)-Y(0)]
What does “transportability” / “generalizability” mean in PCI?
Transportability is predicting how an intervention’s effect changes when you move from the studied sites to new places.
EO data helps because it provides consistent covariate measurements in both the original study region and the target region—supporting effect prediction and uncertainty quantification.
- Practical question: “Will this program work in District B if it worked in District A?”
What is image-based confounding (sometimes called “image confounding”)?
In many observational settings, treatment assignment depends on local context that isn’t well measured in surveys, and EO imagery can proxy parts of that context.
This can reduce confounding, but it never removes the need to justify identification assumptions and run sensitivity/robustness checks.
- Use case: controlling for hard-to-measure neighborhood or environmental conditions in aid/policy allocation studies.
How do you avoid post-treatment leakage when using satellite imagery?
Decide your pre-treatment image windows before analyzing outcomes, and keep post-treatment imagery out of the covariate set.
If the images were captured after treatment rollout, treat them as outcomes/mechanisms (or auxiliary evidence), and validate using spatial/temporal splits where relevant.
- Red flag: “It predicts too well” can mean leakage or dependence, not true causal signal.
What is a “Remote Audit” for RCTs (and why would you run one)?
A Remote Audit uses pre-treatment satellite imagery plus machine learning to test whether treatment assignment is predictable—flagging potential randomization problems early.
It’s inexpensive, globally scalable, and complements (not replaces) traditional balance checks by providing a design-based diagnostic before costly field data collection.
What are “multi-scale representations” in satellite-based causal inference?
Multi-scale representations combine information from multiple spatial footprints (local and neighborhood context) instead of relying on a single image crop size.
This matters because real causal drivers often operate at different scales, and multi-scale features can improve heterogeneity discovery and generalization.
Where can I find PCI datasets, code, and tutorials?
Start with the site’s Research, Data+Code, RCT Repository, and Video Library pages for curated entry points and links to official resources.
For software, the CausalImages R package and related open tooling are the most direct way to run PCI workflows end-to-end.
What is the Planetary Causal Inference book, and when is it coming out?
Planetary Causal Inference: Understanding the Environment, Society, and Economy through Earth Observation and AI Systems is under contract with Cambridge University Press and is listed as coming in 2026/2027.
The book develops a unified framework for combining EO data, AI methods, and causal designs—emphasizing reliability, reproducibility, and responsible use.
How can I collaborate—or contribute an experiment/dataset to PCI resources?
If you have a geo-linked RCT, quasi-experiment, or dataset suitable for linking with EO data, the fastest path is to contact the project leads via the About/Team pages.
Strong candidates include studies with village/ward/block-level coordinates, documented designs, and enough geographic spread to support EO integration.