How to test for multicollinearity in models with interaction terms and categorical variables

[Jeanette Atan]

Dear listers,

I am working on a model with interaction terms and factor variables. I am worried with potential collinearity. Due to the type of regressors, I am using the "perturb" prefix with the "collin" command to analyze VIF. I wrote the “perturb: collin” command using the guidance from an earlier post in 2014 about factor variables and “perturb” manual about interaction terms. I am getting the error message “interactions not allowed”.

The manual for "perturb" suggests that: "If a model contains interaction or nonlinear transformation then perturbations are only added to the main effects/untransformed variables". In the meantime 'search collin' within Stata does not yield any hits.

The model and the “collin” commands are below.

regress y X1 X2 i.X3#c.x4 i.mth c.X5 c. X5 #c. X5 c. X5 #c. X5 #c. X5

perturb: xi: collin X1 X2 i.X3#X4 i.mth X5 X5 #X5 X5 #X5 #X5, ///
poptions(pvars(X1 X2) prange(1 1 1) pfac(mth) pcnt(96) )

For information:
x1, X2 and X5 are continuous
Mth is a categorical variable, from 1 to 12, indicating the month of the year
X3 is a categorical variable indicating fico score

I would really appreciate your help if you could advise on the syntax of the code or perhaps suggest another code for multicollinearity analysis.

Thank you in advance for your help.

Best,

Jeanette

[Jeff Wooldridge]

You shouldn't be "testing" for multicollinearity. You've shown no results, so we can't know whether you have standard errors that are too large to be useful. And you need to be clear about what your questions is: Which coefficients and marginal effects are you interested in? You may have severe multicollinearity simply because of the way you parameterized the model. Without clearly stating what effect or effects your interested in, and then presenting estimates of those along with their standard errors, investigating collinearity is not even meaningful. In fact, in general it should be avoided.

[daniel klein]

perturb seems to be from SSC as you are asked to explain. It seems to be meant as an alternative to, not a prefix for collin (also from SSC, probably). If at all, you are supposed to use it with regress.

Jeff already covered the more basic question of whether the hole approach is useful at all.

Best
Daniel