Primary Submission Category: Machine Learning and Causal Inference
Partial Identification with Unobserved Confounding Using the Rashomon Effect
Authors: Srikar Katta, Jon Donnelly, Emanuele Borgonovo, Cynthia Rudin,
Presenting Author: Srikar Katta*
Many quantities of scientific interest often take the form of non–pathwise differentiable or higher-order functionals for which standard semiparametric inference tools fail. We introduce a general strategy for statistical inference based on Rashomon sets—the set of all machine learning models compatible with the data distribution. We provide uniform coverage guarantees even when the estimand is non-smooth and lacks an influence-function representation, often found in machine learning tasks focused on model auditing. Importantly, our inferential framework easily handles settings in which identifying assumptions are violated, such as no unobserved confounding. Through semi-synthetic experiments, we demonstrate that our bounds achieve nominal coverage and remain informative in realistic settings with unmeasured confounding. We conclude with an application to credit risk assessment, illustrating how our framework enables principled inference on feature relevance despite both model uncertainty and omitted variables.