In United States, there is an urgent need to identify the causes of racial disparity, with infant mortality a prime example. Mediation analysis was designed to evaluate such causal questions. Two main approaches have been promoted. The model-based approach has been adapted to implement counterfactual modeling in which a system of equations is used to partition causal effects and resolve the portion of effects attributable to direct and indirect effects. The outcome-based approach relies on direct Bayesian estimation of the conditional probabilities which are then used to estimate mediation effects. The objective of this study was to compare and evaluate differences between model-based and outcome-based mediation modeling. The motivating example models a binary outcome as a function of a binary cause and a binary mediator. The dataset was a random sample of 100,000 births from 2003 that have been made readily available to illustrate mediation modeling. Infant mortality was the outcome of interest. Race was the cause of interest, and the potential mediators were maternal smoking, low birthweight, and teenage maternity. The model-based and outcome-based models were implemented, along with a hybrid or combination model. The hybrid model started as a model-based approach with estimation of regression parameters. From the regression parameters, reverse transformation was used to estimate potential outcomes that were then used in mediational analysis. The model-based or counterfactual approach performed poorly relative to the Bayesian-implemented outcome-based approach. Evaluating the hybrid model traced the error to the set of equations recommended to estimate the mediation effects. The Bayesian estimation of potential outcomes provided a simple and intuitive approach to mediation modeling. Bayesian modelers are encouraged to take a lead in identifying causes of racial disparity, along with a myriad of other mediation questions.
James A. Thompson (Thu,) studied this question.