Integrating Earth Observation Data into Causal Inference

Connor Jerzak | Fredrik Johansson | Adel Daoud

Paper Code Data .bib

Motivation

In recent years, the technology world has experienced a space renaissance, with the number of satellites doubling since 2019 and with the space economy reaching $424 billion in economic value in 2020. Much of this innovation has come in the form of new earth observation technologies. About 1/3 of all satellites are earth observation in nature, meaning that these satellites are devoted to imaging the earth. Moreover, from firms such as Maxar, 15-by-15 cm resolution satellite images are now available to researchers for nearly every place on earth. This means that even small objects such can now be reliably discerned from space.

Despite this space renaissance, the social science communityhas as of yet not yet explored the full potential of earth observation data for improving the practice of quantitative social science. These earth observation data have important strengths for social science: they are not subject to missingness due to the action of the people being studied, they are available for nearly every place on earth, they can be obtained (in some cases) on a daily basis, a nearly unparalleled temporal resolution compared to many social science datasets that are accessible on a yearly basis at best.

Paper Contributions

In this paper, we explore these strengths of earth observation data for social science, especially social science using causal inference methods. We argue that the data stream can provide valuable information about neighborhood-level confounders that otherwise make the practice of causal inference difficult. These confounders are what separate correlation from causation. For example, communities that do receive an aide program in the African context may be at baseline different from those that do not. Satellite data can bring to bear significant and qualitatively new information to model—and adjust for—these kinds of differences.

To support our argument, our paper contains several broad lines of investigation. Analytically, we investigate the assumptions required to formally prove what is known as identification, which allows one to argue for causality when using satellite images in observational causal inference. Via simulation, we probe the role of resolution and model misspecification in affecting the quality of causal effect estimates when using earth observation data in the research pipeline. We discuss how to learn about what in the earth observation data is associated with different treatment decisions via the application of a computational tool known as automatic differentiation.

Finally, we apply earth observation data to the study of anti-poverty assistance in Nigeria, where we find that the satellite information is useful in separating out causation from correlation. Throughout the paper, we keep our eye to the future—to how this unique data stream may come to shape quantitative social science in the years to come.

References

Connor T. Jerzak, Fredrik Johansson, Adel Daoud. Integrating Earth Observation Data into Causal Inference: Challenges and Opportunities. ArXiv Preprint, 2023.
@article{jerzak2023integrating,
  title={Integrating Earth Observation Data into Causal Inference: Challenges and Opportunities},
  author={Jerzak, Connor T. and Fredrik Johansson and Adel Daoud},
  journal={ArXiv Preprint},
  year={2023},
  pages={},
  publisher={}
}
[Code][Data]

Related Work

Connor T. Jerzak, Fredrik Johansson, Adel Daoud. Image-based Treatment Effect Heterogeneity. Proceedings of the Second Conference on Causal Learning and Reasoning (CLeaR), Proceedings of Machine Learning Research (PMLR), 213: 531-552, 2023.
@article{jerzak2023image,
  title={Image-based Treatment Effect Heterogeneity},
  author={Jerzak, Connor T. and Fredrik Johansson and Adel Daoud},
  journal={Proceedings of the Second Conference on Causal Learning and Reasoning (CLeaR), Proceedings of Machine Learning Research (PMLR)},
  year={2023},
  volume={213},
  pages={531-552},
  publisher={}
}
[Overview]

Back to Research
Back to Home