Sensitivity Analyses for Unmeasured Confounding: This Is the Way

Abstract

One important assumption for valid causal inference when comparing outcomes from two or more groups based on observational (non-randomized) data is the assumption of ‘no unmeasured confounders’. While researchers have long been aware of the potential bias from unmeasured confounders, recently many new approaches have been proposed to quantitatively assess the robustness of research to the potential for unmeasured confounding. These include a growing set of literature proposing methods that are applicable in a broad number of settings including when researchers can not even identify a specific potential unmeasured confounder. In this chapter, we review existing sensitivity analysis approaches and argue for a growing consensus of best practice principles. Broadly applicable methods, such as the E-value, simulation framework, or omitted variables approach are well suited as foundational steps in any quantitative assessment of robustness to unmeasured confounding. We also argue that given these broadly applicable methods and published R-packages to remove implementation complexity, consistent application of quantitative sensitivity analyses is now practical.