TL;DR
This paper introduces a novel framework for causal inference using satellite imagery, addressing the challenge of unobserved confounding in data-scarce regions. By treating satellite images as a proxy for latent confounders, the authors develop methods to estimate the causal effects of interventions, demonstrating their approach with an application to anti-poverty aid programs in Nigeria.
What Problem Does This Paper Address?
In social science research, particularly in developing regions, it's often difficult to obtain comprehensive data on factors that might influence both an intervention (treatment) and its outcome. These unobserved factors, known as confounders, can bias the results of observational studies. Traditional methods rely on tabular data (like income or demographics), but this information is frequently unavailable or unreliable in many parts of the world. This paper tackles this "data scarcity" problem by proposing the use of satellite imagery as a rich source of information that can act as a proxy for these unobserved confounders. The core challenge is to develop methods that can effectively adjust for confounding when the relevant information is embedded in complex, unstructured image data, rather than simple tabular features.
Methods
The authors propose a framework for causal inference that leverages satellite imagery to adjust for confounding. They formalize the problem by considering how patterns or objects within satellite images can represent latent confounders that influence treatment assignment and outcomes.
Key methodological contributions include:
- Image-based Confounding Model: They extend traditional causal graphs to incorporate image data (M) as a source of information for latent confounders (U). This includes both pixel-level and scene-level confounding structures.
- Identification Conditions: They establish conditions under which causal effects can be identified, particularly when the confounder is a deterministic function of the image (U = f(M)). This implies that the image itself can serve as an adjustment set.
- Estimation Framework: They propose using machine learning models, specifically convolutional neural networks, to learn representations from satellite images that can then be used for estimating treatment effects, often via Inverse Propensity Weighting (IPW).
- Interpretability Tools: They introduce salience maps to visualize which parts of the satellite image are most influential in predicting treatment assignment, offering insights into the confounding mechanism.
- Simulation Studies: They conduct simulations to assess the sensitivity of their approach to image resolution and model misspecification, providing guidance on practical implementation.
Datasets
- Landsat Satellite Imagery: Used for pre-treatment data in the Nigerian application. This data is described as orthorectified ETM+ pan-sharpened, with a resolution of 14.25 meters, capturing reflectance in green, near-infrared, and short-wave infrared bands.
- Demographic and Health Surveys (DHS): Used to derive the outcome variable, the International Wealth Index (IWI) for Nigerian households in 2018.
- AidData Dataset: Used to identify international aid programs in Nigeria after 2003, serving as the treatment variable.
Key Insights and Ideas
- Satellite Imagery as a Powerful Proxy: The paper strongly advocates for satellite imagery as a viable and often necessary data source for causal inference in data-scarce environments, bridging a critical gap where traditional data is unavailable.
- Image Patterns as Confounders: The core insight is that complex patterns and objects within images, not just simple pixel values, can encode crucial information about unobserved confounders that drive treatment decisions.
- Resolution Matters: The resolution of satellite imagery significantly impacts the ability to resolve confounder objects and, consequently, the accuracy of causal estimates. Lower resolution can lead to more residual confounding.
- Model Specification is Key: While image-based methods offer advantages, performance is sensitive to how the image-derived confounders are modeled and how the image processing pipeline is specified (e.g., kernel width).
- Interpretability Aids Understanding: Salience maps are crucial for understanding why the model makes certain predictions and for gaining insights into the underlying confounding mechanisms, which is vital for trusting the causal estimates.
Key Results
- Improved Causal Estimates: The proposed satellite image-based adjustment methods consistently outperform naive difference-in-means estimators, showing reduced bias and RMSE in simulations and empirical applications.
- Sensitivity to Model Parameters: Simulations demonstrate that estimation bias and RMSE are minimized when the estimating model's parameters (e.g., kernel width) align with the true confounder-generating process.
- Impact of Resolution: Lower image resolution generally leads to increased bias and RMSE, but even low-resolution images provide improvements over no adjustment.
- Empirical Application (Nigeria): The study successfully applied the framework to estimate the impact of local aid programs on neighborhood-level poverty in Nigeria. Adjusted estimates for treatment effects were positive but smaller in magnitude than naive estimates, suggesting that aid programs were indeed directed towards areas with higher potential for growth, a factor captured by the satellite imagery.
- Reduced Covariate Imbalance: After applying IPW weighting based on satellite image features, the imbalance in geographical coordinates (longitude/latitude) between treated and control groups was significantly reduced, indicating improved counterfactual comparisons.
Limitations and Future Work
- Unresolved Objects: The paper acknowledges that if confounder objects are too small to be resolved by the satellite imagery, residual confounding can persist. This highlights the need for higher-resolution sensors.
- Ambiguity in Scene Definition: Defining the "scene" or unit of analysis can be ambiguous in spatial data, leading to potential bias (modifiable areal unit problem). Future work could explore methods robust to scale.
- Connecting Image Patterns to Human Decision-Making: While machine learning models learn image patterns predictive of treatment, further research is needed to connect these learned patterns to the actual cognitive processes of decision-makers.
- Privacy and Fairness: The authors note that future work should consider privacy and fairness issues in causal analyses using passive sensor technologies.
- Stochastic Confounding: While the paper focuses on deterministic confounding from images, future work could explore scenarios where confounders are more stochastically related to image features, as explored in the appendix.
Summary updated