VIF issues

Good afternoon,

I ran a hierarchical regression model, (main effects in step 1, two-way in step 2 and three-way in step 3). My VIF in step 1 are fine, around 1, however, in step 2, for a few of the two-way interactions, the VIF are quite large 16-19 range (same with one of the three-way interactions' VIFs). The thing is, I ran a correlation analysis thinking this might be due to two of my variables explaining the same variance, but the highest correlation is .15 (among the IVs). It seems like it is not an issue of two variables explaining the same variance as none of them are correlated. I would highly appreciate your advise on what the issue might be and how can I fix it.

Thank you,


TS Contributor
this is natural, for interactions, as a product is not independent of its terms. You could try normalizing the IVs, this sometimes helps. Also, if your effect is significant the VIFs are less of a worry, imho.



New Member
Optionally, you can center your predictors -- this is an efficient way to get rid of high VIFs when interactions are present (Note, once you center your predictors, you also have to recalculate your interaction terms).
Thank you for your responses, I have centered my continuous variables prior to calculating the two-way and three-way interactions. It might be the issue of one of my dichotomous variables (participant gender) being unequal. I have 70% females.
If this is the issue, would you consider the high VIF not to be an issue any longer?
Alternatively, I could do select cases for each gender and ran the analysis again (without participant gender as one of the variables). This way the unequal number of females vs. males won't be a factor any longer. Indeed after doing so, the VIF are back to normal. The only question now is, is this legit to do?


TS Contributor
the VIF is basically extending your confidence interval. If your result is significant even with the extended confidence interval then you should not worry about the VIF. You risk missing effects but you do not risk having false effects.