The role of sensitivity analysis in estimating causal pathways from observational data

Causal inference in observational data is a nearly alchemic task because parameter estimates depend on the model being correctly specified. Researchers strive to include all potential confounders in their models, but this assumption cannot be directly tested. Further complications arise in causal mediation analyses where the decomposition to direct and indirect effects is of interest. We argue that sensitivity analysis is an effective method for probing the plausibility of this nonrefutable assumption. The goal of sensitivity analysis in the context of causal mediation is to quantify the degree to which the key assumption of no unmeasured confounders must be violated for a researcher's original conclusion regarding the decomposition to direct and indirect effects to be reversed. Three general scenarios where the assumption of no unmeasured confounders is violated will be discussed, and results derived from sensitivity analyses appropriate for each scenario will be presented.