Speaker: Fucheng Warren Zhu

Paper: proceedings.mlr.press/v275/zhu25a.html // [.bib]

Transcript:

Hello, everyone. This is the official video on encoding multi-level dynamics in effect heterogeneity estimation. My name is Warren from Harvard College, and this is joint work with Connor Jerzak and Adel Daoud.

Introduction

Context & Motivation

Earth observation data is increasingly used in policy analysis. Researchers often link RCT (Randomized Controlled Trial) data to satellite (SAL) imagery around households and villages to improve estimation of conditional average treatment effects (CATEs). However, most existing algorithms in these CATE estimation procedures use images at a single, fixed scale. This approach struggles to balance the fine-grained, individual-level details (e.g., what an individual household looks like) with broader contextual information (e.g., environmental or neighborhood-level characteristics).

The key question we address is: How can we effectively incorporate multi-level dynamics in causal effect estimation? In other words, how do we capture both the local (household) context and the broader neighborhood context when estimating treatment effects?

Main Contribution

We propose a method called multiscale concatenation. This method:

1. Is a composable procedure that transforms single-scale algorithms into multiscale ones.
2. Is easy to use—no deep architectural changes are required.
3. Improves the performance of CATE estimation compared to single-scale models.

Intuition Behind Multiscale Representations

Consider two images of the Longfellow Historic Site in Cambridge, Massachusetts:

• A small-scale image focusing closely on the house, capturing fine-grained details about the structure.
• A large-scale image showing the surrounding neighborhood, including a river, roads, trees, and other aspects of the environment.

When estimating causal effects (e.g., how a certain intervention might affect households in an area), it’s not clear whether the small-scale or the large-scale image provides the most relevant information—often, both are important.

• Small-scale images capture fine-grained details but miss broader patterns.
• Large-scale images capture broader patterns but lose local details.

A single-scale analysis often fails to capture these multi-level dynamics. Our solution is to combine representations from different scales so we can preserve both local and contextual information.

The Multiscale Concatenation Procedure

Overview

1. Notation:
   • Let each unit (e.g., a household) have a geographic location.
   • We have an image extraction function that, given geo-coordinates and a specified image size (scale), outputs an image.
   • A representation function, ϕ, takes in images and produces latent features.
   • A CATE estimation function outputs the conditional average treatment effect given those latent features.
   • Finally, a heterogeneity quantification function, measuring the quality of the CATE estimates (we use the rate ratio as our metric).
2. Goal:
   • Use a representation function and CATE estimation function to maximize the heterogeneity signal detected in the images by choosing the best image sizes within these two scales.
   • Compare this to a baseline that uses a single image scale.

How It Works

• In a standard single-scale approach, you might feed one image of a fixed size image into a feature extractor to get features used by the CATE estimator.
• In multiscale concatenation, you instead take two (or more) images of different scales and pass them through two pre-trained single-scale feature extractors. You then concatenate these feature representations, so the CATE estimation model sees both local (small scale) and contextual (large scale) information.

This approach approximates having an extremely large image that captures all possible signals but avoids the pitfalls of a single-scale model.

Simulations

Evaluating causal effect estimation models is challenging due to the lack of ground truth data. Hence, we use simulation:

• We introduce perturbations at multiple scales (e.g., images of size 32×32 vs. 256×256) to test whether our model can capture signals at both the local and global levels.
• We compare multiscale concatenation to baselines that only use one scale.

Findings:

• Multiscale concatenation acts as both a robust estimator and a signal amplifier.
• It can recover most of the information when there is a signal at only one scale, and it can capture combined signals at different scales even more effectively.

Real-World Data

We also test multiscale concatenation on data from two RCTs:

1. Peru: A multifaceted anti-poverty intervention (2007–2015).
2. Uganda: A randomized business grant program for young people (2008–2012).

We link each RCT unit’s geographic location to satellite imagery. For example, in Peru, we use varying image sizes and observe how combining them affects the ability to detect heterogeneity.

Results:

• Uganda shows strong evidence of multi-level dynamics.
• Peru shows moderate evidence.
• The best scales differ by context and by the choice of representation function (i.e., which base feature extractor is used).

An important takeaway: a larger image is not always better. Sometimes a smaller-scale image more strongly captures fine-grained heterogeneity signals (e.g., in Uganda). Larger images can have extensive overlap between nearby units, making it harder for the model to learn individual-level differences.

Implications & Conclusion

1. Balancing Local and Contextual Information
   Our results highlight the need to carefully balance local detail with broader contextual cues.
2. Simplicity & Robustness
   The proposed multiscale concatenation approach is easy to implement (minimal changes to architecture) and provides robustness to both local and global information.
3. Future Work
   • Extending to observational (non-RCT) settings.
   • Integrating multiscale concatenation with dimensionality reduction methods to handle the increased representation dimensionality.
   • Developing interpretability techniques to better understand how small- vs. large-scale features drive heterogeneity in causal effects.

Thank you for watching. For more details, please visit planetarycausalinference.org.

References

@article{zhu2024encoding,
  title={Encoding Multi-level Dynamics in Effect Heterogeneity Estimation},
  author={Zhu, Fucheng Warren and Connor T. Jerzak and Adel Daoud},
  journal={Forthcoming in Proceedings of the Fourth Conference on Causal Learning and Reasoning (CLeaR), Proceedings of Machine Learning Research (PMLR)},
  year={2024},
  volume={},
  number={},
  pages={}
}

–

Related Work

@article{sakamoto2025scoping,
  title={A Scoping Review of Earth Observation and Machine Learning for Causal Inference: Implications for the Geography of Poverty},
  author={Sakamoto, Kazuki and Connor T. Jerzak and Adel Daoud},
  journal={Hall, Ola and Ibrahim Wahab (eds.), Geography of Poverty},
  year={2025},
  volume={},
  pages={},
  publisher={Edward Elgar Publishing (Cheltenham, UK)}
}

@article{daoud2024planetary,
  title={Planetary Causal Inference: Understanding Society, Economy, and Environment through Satellite Images},
  author={Daoud, Adel and Connor T. Jerzak},
  journal={A Book Project},
  year={2024},
  volume={},
  pages={},
  publisher={}
}

@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={}
}

–

Back to Home

• GitHub
• YouTube
• Bluesky
• LinkedIn
• Instagram
• TikTok
• X
• Mail