Multicollinearity using dummy variables

[amanstha]

Hello, 

I want to have an opinion on conducting a multicollinearity test using proc reg. I have been using proc reg with VIF/TOL to test for multicollinearity even for logistic regression. What would be the best approach - use the original variables? (OR) use the dummy variables? While original variables only demonstrate the VIF value for the particular variable, dummy variables include detail for the response items for each variable.

PS: I do not include the reference variable whenever I use dummy variables to check for multicollinearity.

Example of code: 

proc reg data=abuse2;

     model rec_abuse=   rural age /*continuous*/ female signwprtnr numchild

                                       infedu fedu  pens agri othocc indincome nuclear spouse daughter relative faminc

                                        upwork  spcomu nsprvis nsprcal smknvrfmr alcnvrfmr dyexer ntrheal neveal

                                        smorbid mmorbid nhins nhcacc idep modep sedep

                                      /vif tol;

run;

quit;

Checking Multicollinearity in Logistic Regression model 

[StatDave · Accepted solution]

To evaluate your model for collinearity, you need to use the same form of the model that you want to fit. If you have categorical variables that you intend to represent with dummy variables in your model, then you need to use the same dummy variables (and coded the same way). The collinearity statistics are only useful for evaluating the form of the model that you specify in REG. But note that if your actual model is a logistic, or some other generalized linear model, then you need to use appropriate weights when evaluating collinearity. This is discussed and illustrated in this note.