Statistical Assessment of Mediational Effects for Non-Linear Mediational Models

Elucidating causal relationship among variables of interests is critical to epidemiological studies. Such studies often involve mediational models, and require appropriate statistical assessment of mediational effects. However, the current statistical methodologies used to assess mediational effect are limited to linear association. Here, we propose a new definition of mediational effects for a class of generalized mediational models, and estimate the mediational effects through Bayesian modeling and Markov chain Monte Carol (MCMC) techniques. In an earlier study, we showed that, under linear association, our method comes to the same results as the existing methods. Here, we focus on demonstrating how one can apply our method to a generalized mediational model using logistic mediational model as an example. Simulation studies are used to compare the asymptotic and the non-informative Bayesian estimates for the mediation effects. Case studies are presented to illustrate the application of the proposed method to public health research using a nationally representative database.