I am using 10 independent variables in building logistic regression model. I am sure that some of these variable are correlated. Can anybody tell me how to check for multicollinearity among independent variables in this case. Thanks!
$\begingroup$
$\endgroup$
1
-
$\begingroup$ related question, Logistic Regression - Multicollinearity Concerns/Pitfalls $\endgroup$– Andy WCommented Apr 10, 2012 at 12:00
Add a comment
|
4 Answers
$\begingroup$
$\endgroup$
2
You can use whatever method you would use for ordinary regression. The dependent variable is irrelevant to multicollinearity issues, so it doesn't matter if you used logistic regression or regular regression or whatever.
answered Apr 10, 2012 at 11:09
-
5$\begingroup$ One potential exception here is the interpretation of VIF does not hold for logistic regression, as there are glm weights in the variance. The VIF is still useful but is not an actual variance inflation factor in glms. $\endgroup$ Commented Apr 10, 2012 at 11:17
-
$\begingroup$ Thanks! but out of 10, 6 of my independent variables are "nominal". For building a logistic regression model is it necessary to find whether these variables are correlated or not? How can I find whether these nominal variables are correlated or not? $\endgroup$ Commented Apr 15, 2012 at 8:21
$\begingroup$
$\endgroup$
You can take the reference of condition index as well. a value greater than 30 indicates there is a near dependency in most cases. you can then go by either the correlation matrix or durbin watson test.
$\begingroup$
$\endgroup$
1
You could construct a correlation matrix and look for high values. An alternative would indeed be the VIF values as already mentioned. Both are quite arbitrary and rely on rules of thumb. For example what's the threshold for a correlation to be 'dangerous'? You could try to use factor scores on the correlated variables and check whether your results (estimates) are robust/sensitive to this issue. Good luck!
-
$\begingroup$ I have read that you cannot make a correlation matrix with nominal variables. As six of my independent variables are nominal in nature, how can I find if these variables are correlated in any sense or not? Is it necessary to test for multicollinearity among nominal variables for building a logistic regression model? $\endgroup$ Commented Apr 15, 2012 at 8:26
$\begingroup$
$\endgroup$
Examining a correlation matrix is helpful, but it is not a sufficient check since variables may be correlated when taken together but not pairwise. I recommend examining tolerance or Variance Inflation Factor diagnostics in regression using a weighted regression where the weights are set to be equal to phat x (1-phat) where phat are the predicted values obtained from the logistic regression model fit with the same variables.
answered Nov 12, 2014 at 6:42