Collaborative Research: Theory of Causal Learning

University Of Washington, Seattle WA

Investigators

Abstract

How can we interpret results from complex machine learning algorithms? How can we mitigate the risks associated with using such models for policy decisions? This project addresses fundamental challenges in deriving valid, reliable, and interpretable causal conclusions from complex data using modern machine learning tools. As machine learning becomes increasingly integral to disciplines such as medicine, economics, education, and the social sciences, the demand for causal insight --- beyond predictive accuracy --- has become more pressing. Yet many machine learning algorithms function as “black boxes”, offering limited transparency and lacking rigorous frameworks for replicability and uncertainty quantification. This project aims to establish a theoretical foundation for causal learning that makes outputs from machine learning explainable, statistically sound, and actionable in real-world decision-making. The work is complemented by educational and outreach activities that promote understanding of causal reasoning among students and the broader public. Planned efforts include public lectures, collaborations with K–12 educators, and integration of research findings into university curricula. Collaborative partnerships with institutions such as Microsoft, Eli Lilly, and the Fred Hutchinson Cancer Center will help translate methodological advances into impactful scientific and societal applications. Technically, the project advances causal learning through three interrelated aims. (1) It develops methods for imputing unobserved counterfactual outcomes --- the hypothetical “what if” scenarios that form the core of causal reasoning --- by integrating flexible machine learning models with statistical principles to preserve both interpretability and rigor. (2) It promotes design-based approaches for quantifying uncertainty, particularly in settings where treatments are assigned randomly or pseudo-randomly via permutations. These methods isolate uncertainty from treatment allocation mechanisms, complementing model-based inference. (3) The project builds a statistical framework for finite-population inference, extending traditional inference techniques beyond super-population assumptions. By drawing on tools from empirical process theory and random matrix theory, the framework provides robust inferential guarantees in realistic data settings where independence and large-sample assumptions fail. Together, these contributions will advance the theory and practice of causal learning, bridging machine learning and statistics to improve both scientific understanding and data-informed decision-making. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.