Primary Submission Category: Sensitivity Analysis
A Unified Approach for Assessing Sensitivity to Violations of Causal Assumptions
Authors: Guilherme Duarte,
Presenting Author: Guilherme Jardim Duarte*
This paper introduces a general method for sensitivity analysis to assess the robustness of causal estimates when key assumptions are violated. Unlike prior sensitivity approaches that are developed on a case-by-case basis, it presents a unified framework that accommodates a wide variety of assumption types, including (1) functional-form violations, e.g. the presence of defiers in studies that assume monotonicity; (2) exclusion-restriction violations, e.g. an encouragement that directly affects the outcome in instrumental-variable studies; and (3) unconfoundedness violations, e.g. unobserved common causes in selection-on-observables studies. The key innovation is to allow for an assumption to be violated in a proportion of the data that is at most t; at t=0, the approach recovers the original estimates, and t=1 is equivalent to discarding the assumption entirely. By using recent developments in partial identification, the method derives sharp bounds representing the range of possible conclusions for a given t; if the best- and worst-case conclusions both have the same sign (e.g., both positive), then the original estimate is said to be t-robust. By varying t, researchers can determine the severity of the violation needed to reverse their causal estimates, expressed in a form that is easy to reason about with domain expertise. This method offers a widely applicable, easy-to-interpret tool for examining the implications of assumption violations on causal conclusions.