Primary Submission Category: Sensitivity Analysis
Sensitivity Analysis for the Attributable Fraction in Stratified Observational Studies
Authors: Zhong Zheng, Iris Horng, Dylan Small,
Presenting Author: Zhong Zheng*
The attributable fraction measures the proportion of observed outcomes that can be causally attributed to an exposure. In stratified observational studies, such quantities are often estimated from collections of 2×2 tables that adjust for measured covariates through stratification. Because treatment assignment is not randomized, inference depends on the assumption of no unmeasured confounding. We develop a sensitivity analysis for attributable fractions in stratified observational studies. Existing sensitivity methods for stratified designs are typically based on Mantel–Haenszel or chi-squared statistics and are not well suited for attributable effects, which are nonlinear functionals of stratum-specific risks and exposure prevalences. Our approach evaluates the worst-case departure from the null within a sensitivity model that bounds treatment odds across individuals. The procedure can be formulated as an optimization problem involving the conditional mean and variance of a stratified test statistic. We derive computational methods to evaluate the objective efficiently under the conditional randomization distribution and provide algorithms that scale to many strata.