Moderator and mediator variables
I want to take a few minutes on this page to talk about what kinds of variables can be moderators or mediators. In short, mediators are variables that are expected to change in relation to the other two variables involved, whereas moderators are not (see Holmbeck, 1997). Why would this be the case? The essence of mediation, as you will see from the example presented on another page, is that the mediator changes in regard to the IV, and in turn it affects another variable, the DV. Thus, the mediator is expected to be a "conductor" or mediator of the effect of the IV on the DV.
The moderator, in contradistinction, is not expected to necessarily be correlated with both the IV and the DV. Baron and Kenny note that it is best if the moderator is not correlated with the IV, although it should be noted that this is not a necessary precondition. The moderating variable, in contrast to the mediating variable, is expected to be a relatively stable variable like a demographic variable (e.g., gender, socio-economic status) or a personality trait (e.g., extraversion, internal locus of control).
Can a variable be both a moderator and a mediator? Holmbeck has criticized researchers in the area of stress and coping for considering coping strategies as mediators, and he favours treating them as moderators. His contention is that coping strategies do not always change in relation to an IV such as stress. However, Folkman and Lazarus (1988) have explicitly argued that coping mediates the effect of stress on adjustment. The literature includes examples of researchers treating coping in both fashions, and the argument does not seem to be resolved at this juncture. My own opinion is that coping strategies exhibit characteristics of being both a mediator (i.e., the changing part) and a moderator (i.e., the unchanging part). I think that a good rule of thumb is that a mediator is a variable that changes in relation to the other two variables, and a moderator is a variable that does not necessarily change in relation to the IV.
The bottom line for me is that a mediation model tested with concurrent data is potentially flawed as a model because assumptions about causal changes are made with data captured with a ‘snapshot’ at one point in time. The best mediation analyses are performed with longitudinal data and they are termed ‘longitudinal mediations’. The procedures for this approach include some additional variables in the model beyond those specified in the example presented in these pages, and one should learn these techniques in order to perform these analyses. Again, refer to MacKinnon’s book or my book for clear examples of how to conduct longitudinal mediation.