I ran a logistic regression moderation analyses and I noticed that with the addition of the interaction term, one of the predictors flipped signs. The variable (X) was previously -.015 and became .006 when the interaction term was added. I suspected this might be an issue with multicollinearity but the VIF for each of the variables is below 1.5. I understand that suppressor effects can occur upon adding predictors.

However, can someone tell me whether they think the sign switching upon adding an interaction term is a problem? Would it make sense to correlate the variables in the interaction term?


Once you add an interaction the meaning of the main effect changes.

It's often regarded as good practice to center terms before multiplying them, to avoid colinearity problems, but this doesn't seem to be an issue here. It appears not to be a problem with coliearity, but with how you are interpreting the results.

Flipping a sign when you add an interaction just means that the main effect of X when the other term in the interaction is 0 is positive. But the main effect of X when the other variable is not 0 will be different (that's what an interaction is).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.