Experiments are powerful tools for learning about cause and effect. But researchers are usually constrained in the kind of data on people they can actually measure. These data commonly include features such as age and ethnicity. Rarely is longer-term historical data on subjects included. Neighborhood-level information is also hard to quantify by investigators in their manual data collection in the experiment. To make matters even more challenging, it is usually difficult to learn about how people in places outside the experimental context would have reacted to the intervention (such as an anti-poverty program) because we have little data about those non-experimental contexts, especially in much of the developing world where state capacity is generally low.
In this project, we introduce satellite images as a powerful tool for addressing these problems. Satellite images such as those captured by NASA‘s Landsat mission, contain rich information about neighborhood-level and geographic features of every place on earth. We can thus use them to model how people in different kinds of communities react to treatments such as anti-poverty programs or experiments.
Because satellite data are passively collected even in areas outside the experimental context, we can, with assumptions, perform a transportability analysis where we predict the treatment effects in the entire community, not just those places that actually received the experiment. Because satellite data are comprehensive going back to the 1970s, we can also decompose treatment responses on the basis of pre-experimental historical information as well—perhaps opening up new lines of inquiry into the realm of causal history.
The main model we propose for performing this image-based heterogeneity decomposition is Bayesian in nature and clusters the treatment response on the basis of the satellite images. These clusters reveal information about the places that are higher or lower responders.
We also develop probabilistic salience measures that capture the parts of the satellite images that are influencing the treatment cluster prediction. We conclude with reflections on how these image-based heterogeneity methods may be helpful in other domains—such as medicine—where image information is prevalent.
–
References
@article{jerzak2023image,
title={Image-based Treatment Effect Heterogeneity},
author={Jerzak, Connor T. and Fredrik Johansson and Adel Daoud},
journal={Forthcoming in Proceedings of the Second Conference on Causal Learning and Reasoning (CLeaR), Proceedings of Machine Learning Research (PMLR)},
year={2023},
pages={},
publisher={}
}–
Related Work
@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={}
}–
–
