I have 3 IVs (X1,X2,X3) and 1 DV (Y), the DV being categorical so am using logistic regression for analysis purpose. There is 1 moderating variable (X4), apart from the 3 IVs.
Now normally I put all IVs and moderating variable in one block, followed by interaction term for all IVs with moderating variable in the next block (separately for each IV ofcourse, like X1*X4, X2*X4, X3*X4). This is in SPSS.
I got comments that I should separately test each IV and its moderating effect rather than putting them all together to test each hypothesis. Implication was that when X1*X4 is being tested, X2 and X3 will not feature anywhere. My earlier approach of testing all interaction effects in the same block was not welcomed.Basically the observation was that putting them all together does not tell you the exact effect of each and may give rise to type II error.
What exactly is to be done in this case?
Hi Rohit,
of course, your model only tests for interactions while controlling for all other main effects and interactions. If you have three distinct hypotheses about the moderating effects of this one moderator on the association of the respective IV and the DV, then you could model them seperately in a first step.
But finally, you should end up with your current model, because using seperate models results in an alpha error accumulation. Furthermore, the correlations between the IVs are not taken into account.
Concerning the rise of type II error: Adding predictors to a model only decreases statistical power, if they do not explain variance in the criterion.
Another comment: When using interactionterms, the variables should be mean-centered to avoid multicollinearity. Maybe this could also improve the interpretation of this model with all variables in it, because then you don't have an interaction effect on condition that all other variables are zero, but on condition that they have average values.
Best,
Brian
If you want to know if a given interaction is significant, you need to test that separately. If you put all interactions in one block, and that model has a significantly better fit than the model without interactions, you won't know which interaction (if any) is independently significant.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS