Re: Really Easy Question about Mediation (HELP!!)

Posted by Bruce Weaver on May 18, 2011; 9:07pm
URL: http://spssx-discussion.165.s1.nabble.com/Really-Easy-Question-about-Mediation-HELP-tp4405692p4407858.html

Gene M. has responded with some comments.  Here are some more from a slightly different perspective.  My training was in psychology, but that was in the days before mediation became the rage.  By that time, I was working in the area of health-related research & biostatistics.  One thing I find interesting is that the definitions of confounding and mediation are nearly identical, the difference being that the mediator is on a presumed causal chain between X and Y.  For some good discussion of this, see Mike Babyak's short article available here:

   http://ebmh.bmj.com/content/12/3/68.full

Given this similarity between confounding and mediation, I find it very interesting that the standard advice in the world of biostatistics is (or at least was a few years ago) that one should NOT perform a statistical test for confounding, whereas the mediation folks are going whole hog down the road of testing (e.g., developing standard errors).  

Here's an example of the biostats view from Bob Wolfe's (from University of Michigan) "classic lecture" series.  (Unfortunately, these lectures no longer seem to be available online.  Fortunately, I saved copies of some of them.)

--- start of excerpt ---

C. Testing for confounding (don't do it):

    When attempting to document the effect of a specific risk
    factor on an outcome, confounding factors should be
    controlled for even if they are not significantly related to
    the outcome in the analysis.  In this case, the objective of
    the analysis is to estimate the strength of the relationship
    between the risk factor and the outcome not explainable by
    confounding factors, and the strength of the relationship
    between the confounding factors and the outcome is not as
    important.

    When searching for a parsimonious model for predicting the
    outcome, it is useful to exclude unimportant factors from
    the model.  Some data analysts use statistical significance
    as a criterion for importance.  This approach may miss confounders.

--- end of excerpt ---

Finally, I should mention that the next section in that same lecture has the title, "D. Confounders and intervening variables:  Often a quandary."  Intervening variable is another name for mediator, of course.

HTH.

toph_bei_fong wrote

Hi everyone,

I am very, very new to mediation and could use some help. I know the steps to Baron and Kenny's method but am unsure of how to actually do them in SPSS.

I have two categorical IVs (which I have dummy coded), the interaction term of these categorical variables and three continuous mediators.

So, here are my questions:

1) If I'm wanting to look at mediation of one of my IV effects, would I enter that variable in by itself in Step 1 as a predictor and leave the other IV out? Or can I assess both at the same time by entering both as my IVs? Or should I just enter the full model (IV1, IV2, IV1*IV2)?

2) If I'm wanting to look at mediation at it pertains to my interaction term, do it enter the interaction term as the sole predictor in Step 1 or do I also enter IV1 and IV2 as covariates?

I just really need someone to walk me through these steps, but after that I'm golden.

Thank you!

tbf