Mediation analysis with weird pattern of results

#1
The total effect in mediation is equal to the direct effect plus the indirect effect. What if my direct effect is negative (non-significant), my indirect effect is positive (significant) and my total effect is negative (significant)? Is this still a case of inconsistent mediation?

Also, the negative total effect means in my case that more cortisol at time 6 leads to less intense reactions to angry faces (while more cortisol at a later time, time 7, leads to more intense reactions to angry faces). Can anyone explain this?
 

Karabiner

TS Contributor
#2
Not sure what situation this is. Research question, research design, variables and their measurements, sample size, or size of the coefficients are not described. Have you considered a suppression effect?

With kind regards

Karabiner
 
#3
The data are from a study by Karin Roelofs and Patricia Bakvis from Clinical and Health
Psychology. The design of the study is rather complex, but we will look at only three
variables within one subgroup of patients. The data are in the file Cortisol.sav.
After a lot of preliminary measurements, all subjects in our patient group were subjected
to a stressor. After that, three measures of cortisol level (a physiological indicator of stress)
were taken, of which we will look at the last two (variables cort6 and cort7). Almost
immediately after measuring cort7, subjects' reaction to subliminally presented angry faces
was measured, with higher scores indicating more intense negative reactions to these angry
faces (variable angryfac). So the temporal order of measurement of our three variables is:
cort6 -- cort7 -- angryfac.


This was the information we got. I did three regressionanalyses (cort6 on angryfac, cort6 on cort7 and cort6 + cort7 on angryfac). My results are:

In the first step of the mediation model, the regression of the independent variable cort6 on the dependent variable angryfac, ignoring the mediator, was not significant, R² = .021, F(1, 16) = .344, p = .566. The standardized regression weight for cort6 was negative and not significant, ßYX = -.145, t(16) = .-587, p = .566. The total proportion of angrfac variance explained by cort6 was Rtot2 = rYX2 = (-.145)² = .021.
The second step showed that the effect of the independent variable cort6 on the mediator cort7 was highly significant, R² = .689, F(1,16) = 35.421, p < .001. The standardized regression weight for cort6 was positive and significant, ßZX = .830, t(16) = 5.952, p < .001. That is, more cortisol at time 6 leads to more cortisol at time 7. The unique part of the direct effect was rY(X.Z)2 = (-.561)² = .315.
In the final regression analysis, the predictors cort6 and cort7 were regressed on angryfac, the dependent variable. This analysis was significant, R² = .356, F(2, 15) = 4.139, p = .037. Step 3 revealed that the effect of cort7 on angryfac, controlling for cort6, was significant, ßYZ.X = 1.037, t(15) = 2.791, p = .014. So, more cortisol at time 7 means more intense reactions to angry faces. In the fourth step, after controlling for cort7, the effect of cort6 on angryfac was significant, ßYZ.X = -1.006, t(15) = -2.707, p = .016. That is, cortisol at time 6 leads to less intense reactions to negative faces when taking into account cortisol at time 7. The unique part of explained variance by the indirect effect was Rind2 = rY(M.X)2 rMX2 = .578*.830 = .480. The overlap part was Roverlap2 = rYX2 - rY(X.Z)2 - rY(M.X)2 rMX2 = (-.145)² – (-.561)² - (.578 * .830) = .106.

How can it be explained that my direct effect of cort6 on angryfac is not significant, but my total effect of cort6 on angryfac is significant?