1
$\begingroup$

How can I look for collinearity diagnostics, particularly condition indexes and proportions of variance in Multinomial logistic regression?

$\endgroup$
1
  • 1
    $\begingroup$ Jochem, I've merged your two accounts (thanks to @Macro who noticed that). You still need to register your account once and for all. $\endgroup$ – chl May 18 '12 at 13:00
5
$\begingroup$

Since collinearity is a function only of the independent variables, you can get the condition indices and proportions of variance by pretending it's OLS regression. (e.g. the /COLLIN option in PROC REG of SAS, or however you usually do it). The difficulty comes in whether there are different levels of condition index that is problematic for multinomial logistic.

$\endgroup$
6
  • $\begingroup$ The corr() command does the same for R $\endgroup$ – gregmacfarlane May 18 '12 at 11:33
  • 2
    $\begingroup$ @gmacfarlane I think corr() just computes the correlations, which is not the same. A bit of searching found the perturb package, which seems useful in this context, especially the colldiag() function. $\endgroup$ – Peter Flom May 18 '12 at 11:47
  • 1
    $\begingroup$ @Peter Flom You are really providing great help to the OP on this problem. $\endgroup$ – Michael R. Chernick May 18 '12 at 12:03
  • 1
    $\begingroup$ (+1) Re: "The difficulty comes in whether there are different levels of condition index that is problematic for multinomial logistic." - this is very true. Can you comment on this from experience? Are different thresholds for "problematic" VIFs needed for the multinomial model? $\endgroup$ – Macro May 18 '12 at 12:20
  • 1
    $\begingroup$ I found a paper by Hendrickx, Belzer, te Grotenhuis, and Lammers entitled "collinearity involving ordered and unordered categorical variables" that looks helpful $\endgroup$ – Peter Flom May 18 '12 at 22:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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