# How to interpret basic three variable mediation model?

I have run a multiple regression to determine which factors predict the willingness to pay for members of another group.

The multiple regression determined that symbolic threat (fear of a cultural and traditional attack on group values by others) was the only significant predictor of how much one would pay for members of another groups. I did however, in my research find that attitudes towards the outgroup in question entirely mediates this effect. The final results are as shown:

(I can't post a picture because my reputation isn't 10 yet so I'll write it as best I can, all ---- lines should be solid)

       Attitude toward outgroup
ㅅ                 \
/                   \
.633***/                     \.482***
/                       \
/                         v
Symbolic ------------------> Amount of Tax Willing
Threat   .495*** (.190 ns)    To pay on outgroup.


I am however having difficulty in interpreting what this mediation means.

• Does this mediation mean that without any attitudes towards the outgroup, symbolic threat would no longer predict the amount one would pay for the other group?
• Does this mean that existing attitudes combined with symbolic threat faithfully predict the amount one will pay for the other group, while separately no prediction can be made?

Any ideas or useful references would be really helpful thank you.

• It would be useful to know what kind of variables symbolic, attitude and amount are; binary, discrete, continuous? In general, you have to run the regression with and without the mediation to know what kind of difference it makes. It all depends on how those three variables are correlated. – Nameless Jul 15 '13 at 8:53
• Is there something wrong with your variables? Do higher scores on symbolic threat actually mean lower symbolic threat? Do higher scores on attitude represent more positive attitudes? The pattern of relationships looks to me like one of the variables is flipped. – Jeromy Anglim Jul 15 '13 at 9:56
• Jeromy you are correct. Higher scores on symbolic threat should actually mean lower symbolic threat. I think I should go reverse code that now. – richard Jul 16 '13 at 6:48
• And yes, higher scores on attitude represent more positive attitudes – richard Jul 16 '13 at 6:49

There are two general approaches to mediation. The older approach has a seminal article by Baron & Kenny . This views mediation as either present or absent. More recent work, by MacKinnon et al. (e.g. this book) views it as a continuum.

In either case, the usual view of mediation is that 1) In the absence of the mediator, there is a strong relationship between the independent and dependent variables. 2) In the presence of the mediator, that relationship changes (some would only call it mediation if the relationship becomes near 0). Baron & Kenny also require a relationship between the IV and the mediator.

So, at a minimum, you'd want to compare two regressions:

Tax ~ SymbolicThreat

and

Tax ~ SymbolicThreat + AttitudeToOutgroup

However, as others have commented, the latter regression might have problems with collinearity.

• All of the variables statistically correlate. I found two useful websites. One is from Bangor University explaining step by step how to perform the a mediation analysis in SPSS pages.bangor.ac.uk/~pes004/resmeth/mediation/mediation.htm. While the other provides a syntax which not only performs the mediation analysis but also Sobel, Goodman and Goodman II tests ats.ucla.edu/stat/spss/faq/mediation.htm. I can see doing it step by step and by using the syntax that the model is highly statistically significant. It would be safe to conclude that mediation has occurred – richard Jul 16 '13 at 6:50
• I would recommend these sites for any mediating beginners. Thank you everyone for your help. – richard Jul 16 '13 at 6:55
• Additionally, would it be strange to add the condition at which a participant was assigned as a predictor? – richard Jul 16 '13 at 8:23