Rahul writes:

Some psychologists were asking me about testing for mediation in a multi-level model. Maybe you have some ideas. They have a 3 level model, repeated measurements over time per person, and each person is one half of a couple. The level 1 dependent variable Y, is predicted by variables X and M which are also measured at level 1. And M is predicted by X. X is actually an interaction between a level 2 variable A and a level 1 variable B.

Y ~ X + M + other level 1 variables
M ~ X
where X=A*B

Their question is whether M mediates the effect of X on Y, and what the effect of X is on Y directly, and indirectly through M. Is there something in your book that would help with this? I’ve been trying to code it in proc mixed in SAS.

My reply:

We discuss mediation in chapter 9 of our new book. The short answer is that you can’t say much about mediation unless you have some experiments that directly manipulate M.