Thanks Jorge. In terms of SES, I think you mean that SES could in fact be one of the predictors of interest? If so, you are absolutely right, and often times both race and SES are of interest. Other approaches to consider is SES as an effect modifier and as a contextual factor. I know you couldn't be at the lecture today, but I really encourage you to watch it and see how the DAGs map on to these different modeling approaches. One other thing to keep in mind is that mediation can be partial - so the direct pathway doesn't have to be completely eliminated by the inclusion of the mediator.
For dementia, I think policy context with respect to access to in home services and other social services, as well as attitudes toward the elderly, would be important contextual factors. Also, age distribution within the community would have an effect. The importance thing to remember about multi level models is that you are including something that can't be just measured at the individual level, but is really about the context that the individual is in.