Figure 2.

Sensitivity analysis of the causal mediation analysis. Causal mediation analysis is performed under the assumption of sequential ignorability, stating that there was no unmeasured confounder that is related to both the mediator and the outcome. Sensitivity parameter $\mathrm{\rho }$ is the residual correlation and a large value means that there is a large unmeasured confounder. The dashed line indicates the causal mediation effect for the control (upper panel) and treatment groups (lower panel) under assumption that there is no unmeasured confounder (ignorability assumption: ρ = 0). It appears that a $\mathrm{\rho }$ of 0.3 is required to modify the sign of the mediation effect. Judged by subject-matter knowledge, a $\mathrm{\rho }$ value of 0.3 is very large and it is unlikely that the mediation effect doesn’t exist.