ykleung posted on Friday, December 01, 2017 - 6:04 am
I am trying to test a mediation model with multiple IVs and DVs. To do so, I first compared the fit of multiple models (full mediation, partial mediation and no-mediation) and examined my direct and indirect effects afterward.
Results suggest that the partial mediation model (i.e., with direct path between IVs and DVs) is significantly better than the the full mediation (i.e., without direct path between IVs and DVs) and no mediation models.
Surprisingly, when examining the significant level of the partial mediation model, none of the direct path between IVs and DVs is significant. This seems to contradict with the results of the model fit which suggest the partial mediation model has a better fit than the full mediation model.
My question is am I correct that I should first select a model with best model fit before examining the paths when testing for mediation. And in my case should I conclude it is a partial or full mediation?
In the partial mediation model you want to use Model Test to test that all your direct effects are jointly zero. That would correspond to your chi-square diff test.
ykleung posted on Tuesday, December 12, 2017 - 3:40 pm
Dear Prof. Muthen,
Thanks for your reply. Following your instruction, I added a model test to test whether all the direct effects are jointly zero as showed below (abridged input)
Model: Y1 on M1 X1(a1); Y2 on M1 X1(a2); M1 on X1;
Model test a1=0; a2=0;
Just to make sure. Am I right that, when the Wald test is significant then it means that path X1 to Y1 (a1) and X1 to Y2(a2) are jointly different from zero and thus it should be interpreted as partial mediation?