Page 1 of 1

Q2 for Interaction Terms in Moderation Models

Posted: Thu Jun 20, 2019 10:14 am
by Alexg
Hello Everyone,

I am in the process of evaluating the effect size Q2 for a moderation model. I created the Interaction terms through the two-stages approach. In total I have two interaction terms on two different paths.

Hair et al. 2017, suggest that to evaluate the predictive relevance of any exogenous constructs the Blindfolding analysis should be run first with the exogenous construct in the model and then without it. Does this apply also to the Interaction terms?

Should I remove one interaction term per time and check what is their predictive relevance?

Thank you in advance for your reply!

Kind regards!


Re: Q2 for Interaction Terms in Moderation Models

Posted: Tue Jul 02, 2019 12:24 pm
by jmbecker
First, you could test the interaction effect itself for its predictive relevance by including and excluding it. You would do that one at a time to know the effect for each interaction.

Second, if you test one of the predictors that are also included in the interaction effect for their predictive relevance, you would also need to remove the interaction effect. But again you would usually test each predictor separately and not together.

Re: Q2 for Interaction Terms in Moderation Models

Posted: Tue Jul 02, 2019 2:57 pm
by Alexg
Dear Dr. Becker

thank you for your reply!

I already computed the Q2 of each interaction term by removing them one at the time as you suggested.

However, as you said, when I remove the predictor I automatically remove also the interaction term. Somehow, could this provide incorrect results of the Q2 effect for that specific predictor?

I have the following Model: A and B exogenous, C endogenous, M moderator, A*M interaction term 1 and B*M interaction term 2.

\ /
/ \

Also, what about the Moderator variable (M) that I actually used to create the interaction terms (M*A and M*B)? When I remove it, I am removing all the interaction effects on all paths. Should I compute the Q2 effect also for the M variable or it does not make sense given that the path M-->C indicates just the simple effect when the moderator variable is equal to its mean value?

Look forward to hearing from you.


Re: Q2 for Interaction Terms in Moderation Models

Posted: Fri Jul 12, 2019 11:46 am
by jmbecker
First, I'm not a big fan of the q² effect size and we do not provide automatic q² exactly because of these complexities. It needs judgment on what you want to test when using the q² effect size.

The interpretation of the interaction terms is straight forward I would think. You get a direct estimate of the additional predictive power when including or removing them.
For the moderator and predictors it is more complex. You might be interested in the overall effect assuming moderations (i.e., including the interaction term) or you might be interested in the predictors (or moderators) own predictive power without the interaction term.

Thus you could compute two q² for A.
1) Compare Q² without A against a model with A and A*M.
2) Compare Q² without A against a model with A but without A*M.

Similarly for B and M.

The first is something like the overall effect of A and the second the own effect of A.
NOTE: I have made up these terms because I think they facilitate understanding, but this is nothing you find in the literature.

Looking at both you might get additional interest insights about the nature of your interaction.