Abstract:
|
Elucidating CR among variables of interests is critical to epidemiological studies. Such studies often involve MM, where a mediator Z may explain the CR of X on Y. Appropriate statistical tools are needed to help investigators identify a mediator and quantify its contribution to a CR. However, the two currently used techniques--multi-step regression and SEM analysis--are limited to linear MM. Pearl (Causality,Cambridge Press) derived mathematical language for causal inference, but did not address the issues of parameterizing, estimating and testing causal effects. In an earlier work, we investigated Bayesian modeling technique for linear MM. Here, we propose a general class MM that unify and generalize the current methods for MM. The mediation effects are defined to allow for a detailed accounting of the CR of X on Y. Both asymptotic and Bayesian approaches are proposed to estimate and test mediation effects, and both simulation and case studies using national representative data will be presented. We found that the Bayesian approach along with the MCMC technique is statistically satisfactory and outperforms the the MLE-based asymptotic approach with regard to accuracy and efficiency.
|