Machine Learning and Causal Inference for Extreme Event Attribution in Climate Science

Machine learning and causal inference for extreme event attribution in climate science
Cassandra C. Chou, Scott L. Zeger, Benjamin Q. Huynh 

Abstract
Extreme event attribution, the science of assessing the extent to which a disaster was caused by climate change, is crucial for informing climate adaptation policy. Machine learning is increasingly being used for extreme event attribution to model rare weather events otherwise too difficult or computationally intensive to model using traditional simulation methods. However, the validity of using machine learning for extreme event attribution remains unclear, particularly given concerns of distribution shift and algorithmic governance. We utilize California wildfire data from 2003-2021 to introduce a machine learning and causal inference framework for extreme event attribution, exploring the sensitivity of estimates to model design choices and demonstrating methods to improve model validity under distribution shift. We observe that individual predictions are highly sensitive to model specifications despite similar predictive performance, raising concerns about using these methods to guide policy. We further observe model performance under distribution shifts projected from climate change scenarios is highly variable. We provide recommendations to improve model robustness and causal validity to more accurately inform climate adaptation policy.