"Two methods for mediation analysis with sequential mediators were compared using multiple simulation scenarios. The performances of each method were assessed using three key metrics: relative bias, root mean square error, and coverage. The methods shared both similarities and key differences and some modification and adjustment were necessary to perform comparable simulations across the scenarios. Overall performance was assessed primarily using relative bias, where each simulated effect estimate was compared to a "true" effect generated by simulating from a theoretical super population. Simulation scenarios included correctly specified models using both methods and various mis-specified estimation models by incorrectly specifying a critical parameter in the model to assess the performance and robustness of each mediation analysis method. The results of the simulations suggest that one method was particularly more resilient to mis-specification of the model over the other, and that proper specification of the marginal structural model is also critical to minimizing bias and maximizing coverage."--Page ii.
Read
