Mediation analysis with continuous mediator?

Question

I would like to conduct a mediation analysis with a continuous mediator. I have a hard time finding comprehensive guides on how to do this.

1. From what I have read, bootstrap-based mediation is preferred, but if I understand correctly, the mediation package only allows for qualitative mediators. Any suggestion how to do it?
2. My second doubt concerns fitting a model for use in mediation. Should I use simple bivariate regression models or should I look for the best-fitting model and then calculate mediation? The latter approach can be problematic due to the number of variables and the equivalence of some models in terms of goodness of fit.

Answer 1

The mediation package can handle continuous mediators. See U. of Virginia Introduction to Mediation Analysis which also gives a pretty comprehensive look at how to do the whole analysis. There are also lots of references in other posts here. Look at the mediation tag.

For your second question: This is going to depend on all sorts of details of your data and problem. I would not do either of the things you suggest, at least, not as a general rule. Model building is part art.

As a general rule: First I would look at collinearity. There are various things to do about that, and many posts here about it. Second, I would not use some automatic method of model building. Third, I would include the mediation in all models.

But, if the mediation is not really imporotant to you, then I might do things differently. And if sample size is small, then you may be limited to rather simple models.