Sensitivity Analysis Under the f -Sensitivity Model: A Distributional Robustness Perspective.

This article introduces the f-sensitivity model, a novel nonparametric framework for sensitivity analysis in causal inference with observational data subject to unmeasured confounding. Unlike traditional models that impose uniform bounds on selection bias, the f-sensitivity model characterizes the average impact of unobserved confounders via f-divergence constraints, allowing for large but rare confounding effects while providing more informative and interpretable bounds on treatment effects. The authors connect this model to a new class of distributionally robust optimization (DRO) problems, develop dual formulations enabling statistically valid and computationally efficient estimation and inference procedures, and propose a debiasing technique that achieves root-n convergence rates even under slow nuisance parameter estimation. Numerical experiments demonstrate the method’s validity, sharpness, and robustness, offering a flexible tool for assessing causal effect robustness in economics, healthcare, and policy analysis.