Sensitivity analysis for unmeasured confounding
Mediation methods are developed under strict no-unmeasured confounding assumptions (22, 23). To examine the robustness of causal effects to unmeasured confounders, we estimated the E-value (24, 25). The E-value provides an estimate (on the excess risk scale) of the smallest excess risk of the unmeasured confounder(s), beyond that accounted for the adjusted confounders, to draw the observed excess toward the null. In addition, the E-value provides an indication of the smallest excess risk of the unmeasured confounder to move the lower 95% CI of the observed excess risk to cross the null.
All analyses were conducted using SAS, version 9.4 (SAS Institute, Cary, NC).