Planetary Causal Inference

What will you get from this Planetary Causal Inference tutorial?

This R-based tutorial shows you how to combine satellite imagery, machine learning, and modern causal inference to answer social science questions at planetary scale. You will learn how to turn Earth observation (EO) data into usable variables and plug them into standard causal pipelines.

• Understand what Planetary Causal Inference (PCI) is and when to use it.
• Learn how to work with satellite images in R, even if you have never touched image data before.
• See how EO-based machine learning (EO-ML) can proxy missing confounders in observational and experimental studies.
• Explore real examples on poverty, inequality, infrastructure, conflict, and development programs.
• Get an introduction to the CausalImages R package and the Planetary Causal Inference book framework.

What is Planetary Causal Inference (PCI)?

Planetary Causal Inference (PCI) is a methodological framework for using satellite-based Earth observation data, together with tabular and text data, to study cause and effect at household, neighborhood, regional, and global scales. It treats images as rich measurements of how social and environmental processes play out on the surface of the Earth.

PCI uses EO-based machine learning (EO-ML) models to transform high-dimensional satellite imagery into low-dimensional features that can enter standard causal estimators. By combining insights from geography, global development, political economy, statistics, and computer science, PCI helps social scientists squeeze more information out of existing experiments and observational studies.

How does Earth observation data compare to traditional social science data?

Traditional data sources like surveys, registries, and national statistics are powerful but often expensive, slow to update, and limited in coverage. Without Earth observation data, many questions about spatial inequality, infrastructure, and environmental change remain under-measured or completely unmeasured.

Satellite-based EO data, by contrast, offers global coverage, repeated over time, at fine spatial resolution. It can capture patterns of urban growth, deforestation, coastal expansion, conflict scars, mining activity, and neighborhood wealth differences, even where no surveys exist.

Earth observation vs. traditional survey data

Who is this tutorial best for?

This tutorial is best for applied researchers who want to add satellite imagery to their causal inference toolkit without becoming full-time remote sensing experts. It assumes you care about causal questions first and want EO and machine learning to serve those questions.

• Social scientists in economics, political science, sociology, development studies, public health, or geography.
• Data scientists working on global development, climate, humanitarian aid, or policy evaluation.
• Students and researchers who already use R for empirical work and want to extend their pipelines with images.
• Practitioners running experiments or impact evaluations who want to generalize and audit their studies using EO data.

What will you learn in this R-based EO-ML causal inference tutorial?

The tutorial, originally delivered at IC2S2 2025, is organized in two main parts. Part 1 focuses on image data and Earth observation; Part 2 focuses on causal inference and how images enter causal pipelines.

What do you learn in Part 1: Working with satellite images in R?

Part 1 builds intuition for image data in computational social science and gives you hands-on experience with basic EO workflows in R. It is designed for participants who may have never worked with an image object before.

• What an image is in data terms: pixels, bands, channels, arrays, and how they relate to light.
• The difference between RGB bands and non-visible bands (e.g., shortwave infrared, night lights).
• How spatial resolution, temporal resolution, and patch size affect what you can and cannot measure.
• How to load and plot images in R using packages like raster, and how to think about memory and large datasets.
• Key EO platforms and sensors (e.g., Landsat, Sentinel) and the practical issues of cloud cover, missingness, and sensor failures.

What do you learn in Part 2: Causal inference with images?

Part 2 shows how image-derived features can stand in for unobserved confounders or outcomes, and how PCI connects predictive EO-ML with rigorous causal estimation. It uses both simulations and real-world applications.

• How to formalize causal questions with directed acyclic graphs (DAGs) and potential outcomes.
• How unobserved confounders can be proxied by image information, text, or other modalities.
• How EO features can help deconfound observational studies (e.g., roads near mines, aid placement, conflict exposure).
• How to use image-based predictions of poverty and wealth in downstream causal analyses.
• How EO can support randomized controlled trials via remote auditing, selection-bias checks, and transportability analyses.

Do you need R, Python, or both for Planetary Causal Inference?

The tutorial is delivered in R because R remains the dominant language for many applied social scientists. Most examples are built so that you can stay inside R for data handling, visualization, and causal analysis.

Under the hood, however, some large EO-ML models still rely on Python and libraries such as JAX. You will see how R can call Python when needed, so you can benefit from modern deep learning backends without fully switching your workflow. Without any Python at all, you can still follow the core ideas and many of the hands-on steps in R.

What is the role of the CausalImages package?

The CausalImages package provides an R interface to efficient image storage and modeling backends. It helps you:

• Serialize and store large image corpora in compact formats for faster I/O.
• Train or reuse EO-ML models (e.g., CNNs, transformers, embeddings) for prediction and confounding adjustment.
• Integrate image-derived features into propensity score models and treatment effect estimators.
• Explore treatment effect heterogeneity using image features in randomized or quasi-experimental designs.

Which research questions can you answer with PCI and EO-based machine learning?

PCI is designed to help with any causal question where geography, time, and human–environment interactions matter. It is particularly powerful when traditional data are missing, outdated, or too sparse.

• Measuring and mapping poverty, inequality, and wealth trends where surveys are rare or absent.
• Evaluating the impact of infrastructure and aid projects (e.g., World Bank and Chinese development programs in Africa).
• Studying deforestation, land-use change, and environmental degradation over decades.
• Analyzing urban growth, coastal expansion, and infrastructure connectivity at scale.
• Auditing and generalizing field experiments using image-based representations of sites and surrounding areas.

How is the Planetary Causal Inference book and framework structured?

The tutorial is based on the book-in-progress Planetary Causal Inference, which organizes the research program around "ingredients" and "recipes." Ingredients are the core components of planetary causal inference; recipes are reusable pipelines that combine them.

What are the main PCI ingredients?

• Research questions and theory: The causal system you care about and the estimand you want (e.g., ATE).
• Image and EO data: Satellite imagery (Landsat, Sentinel, commercial providers), bands, and temporal stacks.
• Causal inference: Identification strategies, confounding adjustment, uncertainty quantification.
• Machine learning and computer vision: Models that turn high-dimensional images into useful features.
• Hardware and compute: Practical constraints around RAM, GPUs, and large-scale training pipelines.

What types of PCI recipes does the book cover?

• Remote surveying: Using EO-ML to measure outcomes (e.g., poverty) where ground data are scarce.
• Exposure measurement: Deriving treatments and exposures from images (e.g., roads, mines, connectivity).
• Observational studies: Using image-based proxies to adjust for confounding in non-randomized settings.
• Field experiments and RCTs: Remote auditing, selection-bias checks, and transportability using image features.

What do you miss without Earth observation in your causal pipeline?

Without EO data, you may have strong causal designs on paper but weak measurement of the actual world where those designs operate. You risk ignoring geographic selection, unobserved infrastructure, or environmental changes that shape where treatments are placed and how they take effect.

PCI helps you make those hidden processes visible. By encoding the landscape around people, projects, and events, EO-ML features give your causal models more realistic information about the systems they are trying to describe.

How can you keep learning after the PCI tutorial?

After the tutorial, you can explore the Planetary Causal Inference ecosystem to go deeper. The instructors share code, slides, and updates through their lab and book channels so you can replicate and extend the examples.

• Visit the book site at PlanetaryCausalInference.org for materials and chapter updates.
• Follow the AI & Global Development Lab for repositories, EO datasets, and tutorials on CausalImages.
• Subscribe to the Planetary Causal Inference Substack for workshop notes, code walkthroughs, and new applications.
• Listen to related interviews on the Journeys of Scholars podcast for broader context on causal inference and multimodal data.

Frequently asked questions about the Planetary Causal Inference R tutorial