Abstrakt
Context: Mediation analysis examines the role of intermediate variables that are thought to lie in the causal path between the regressor and the outcome. Thus it allows the researchers to study whether the causal effect of the regressor on the outcome is transmitted through a mediator.
Methods: The use of latent factors identified in a factor analysis as mediators and/or outcomes in subsequent analyses is common practice. We propose a complex Bayesian model in which the latent factors are treated as augmented data and are used as outcomes and/or mediators in a mediation model at the same time.
In case some of the variables are ordinal, we assume they are determined from corresponding latent continuous variables. Results: We fit a Bayesian mediation model which allows for multiple and not directly observable mediators.
Furthermore, the ordinal nature of the measurements of the depression symptoms is accounted for through the use of latent continuous variables. The dependence structure among the measurements of depressive symptoms was also considered.
On top of that, we reduce the dimensionality of the depressive symptoms by incorportating a factor analysis model among latent mediators. Conclusions: Bayesian methods enable researchers to carry out factor analysis jointly with a mediation model where the factors act as outcomes and/or mediators.