4
$\begingroup$

My data is cross-sectional macroeconomics data. I have six independent variables (x1,x2,x3,x4,x5,x6) plus 2 dummies (d1,d2) plus 2 interactions terms (d1*x1,d2*x1).I am testing my data for multicollinearity using VIF and condition indices(CI)

The t-test : seven statistical significant variables F: statistical significant overall

VIF&CI

Mean VIF : 10.63 (with very high R-square (>85%) in all dummies and interaction terms) CI : 48.3

When I remove dummies and interactions from the model the results are much more better (Mean VIF : 1.62 , CI: 19.34 R-square <50%).

I am expecting -due to the nature of dummies and interaction terms- that my results would present multicollinearity.

Are the above results serious evidence for multicollinearity in my model?

Improve this question
$\endgroup$
3
  • 2
    $\begingroup$ Multicollinearity may or may not be a problem, depending on the objective of your study. Can you describe a bit more about the objectives of your data analysis? Are you simply trying to predict the outcome variable or you concerned with the parameter estimates themselves? $\endgroup$
    – StatsStudent
    Commented Mar 10, 2015 at 14:59
  • $\begingroup$ hi, thank you for your answer. My data in cross-sectional, my dependent variable is the average growth rate per capita gdp, the x1 is the average ratio of public debt to gdp, the d1 is a dummy variable which describes high-income countries and d2 is dummy on high debt countries. The object is: does high-income countries has higher growth rate than other countries? does high-debt countries have higher growth rate of gdp than other countries? also the interactions of them. $\endgroup$
    – Ant
    Commented Mar 10, 2015 at 16:20
  • $\begingroup$ Then yes, you'll need to address the issue. Take a look at this article, which should help you. blog.minitab.com/blog/adventures-in-statistics/… $\endgroup$
    – StatsStudent
    Commented Mar 10, 2015 at 16:24

1 Answer 1

Reset to default
0
$\begingroup$

It really depends on the specifics of your study. But generally, you're right that dummy variables can lead to higher VIF as can the inclusion of interaction terms.

The p-value for the interaction term will not be affected by multicollinearity. You might lower the VIF by centering x1 and re-fitting your model. This should not change the p-values but might reduce the VIF for the interaction terms.

For the dummy variables: If your reference category has far fewer cases in it, then this will yield a high VIF. If this is the case, you can just switch your reference category to the one with the greatest number of cases.

Rather than looking at the mean VIF alone, you should also consider the VIF for each individual predictor. If the VIF for the dummy variables and the interaction terms are contributing to a higher mean VIF, then there might be nothing to worry about.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thank you for your comment. What is your opinion about condition indeces for detecting multicollinearity? some people believe that it is a better measure than VIF $\endgroup$
    – Ant
    Commented Mar 10, 2015 at 16:23
  • $\begingroup$ I'd say that you shouldn't depend on any one measure. The wikipedia entry gives some useful things to do to check for multicollinearity: en.wikipedia.org/wiki/… $\endgroup$
    – Jordan Collins
    Commented Mar 11, 2015 at 17:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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