|
Abstract:
|
Causal mediation analysis is essential for investigating the mechanisms through which an intervention operates. Causal inference regarding the hypothesized mediation mechanism might be invalidated in the presence of unmeasured confounders of the treatment-mediator, treatment-outcome, or mediator-outcome relationships. We develop simulation-based sensitivity analysis strategies to numerically and graphically evaluate the sensitivity of causal mediation analysis results to the presence of an unmeasured pretreatment confounder of the mediator-outcome relationship under treatment randomization. The methods have three primary advantages over the existing methods. First, using the conditional association of the unmeasured confounder with the mediator and that with the outcome as sensitivity parameters, the methods enable applied researchers to intuitively evaluate sensitivity parameters in reference to prior knowledge about the strength of a potential unmeasured confounder. Second, the methods accurately reflect the influence of unmeasured confounding on the estimation efficiency of the causal effects. Third, the methods reduce the reliance on model-based assumptions.
|
