A0657
Title: Statistical learning under causal (path-specific) effect constraints
Authors: Razieh Nabi - Emory University (United States) [presenting]
David Benkeser - Emory University (United States)
Abstract: The estimation of a function-valued parameter defined as the minimizer of a risk criterion is studied, subject to constraints on one or more real-valued parameters of the observed data distribution, typically requiring them to be zero or bounded. A central focus is on counterfactual and causal constraints, particularly those derived from path-specific effects, which have gained prominence as formal tools for expressing fairness in decision-making. In many cases, closed-form solutions exist for the optimal function under these constraints, shedding light on how causal considerations shape predictive models. These solutions also motivate natural estimators for the constrained parameter, often constructed by combining estimates of unconstrained components from the data-generating process. As a result, constrained learning procedures can be readily implemented seamlessly with existing statistical or machine learning methods using off-the-shelf software.
