I have a dependent variable made up of 3 categories and 14 binary predictor variables.

I have tried using mlogit and nnet/multinom packages in R.

Is there a better approach than multinomial logistic regression for this particular scenario?

  • 2
    $\begingroup$ how many data points do you have? also you may have separation if the implicit contingency table has 0 or 100% observed counts. this leads to infinite mles and ill conditioning in the newton rhapson scheme $\endgroup$ Jun 29, 2012 at 8:31
  • $\begingroup$ @probabilityislogic - I have a few hundred data points. If I understand your contingency table comment, I think this possible issue was noted as having been checked in my original post. $\endgroup$
    – Aengus
    Jun 30, 2012 at 22:05

1 Answer 1


Given your description of the situation, you are using the right model. There is no problem with having discrete IVs with multinomial logistic regression; MLR does not make any assumptions about the nature or distribution of the IVs. However, I wonder if your IVs are not orthogonal. It's hard to tell, but you may be describing some effects of multicollinearity.

I'm not sure what happened with R, you would need to show your code and data and perhaps the error messages for someone to help you figure that out. Questions about those sorts of issues should be asked on Stack Overflow or the R-help mailing list, though; they are off topic here.

  • $\begingroup$ In reference to @gung wondering about if your IV's are not orthogonal: given that you have 14 different variables, if the lengths of their vectors (given all the same lengths) are less than 14, wouldn't that necessarily make them non-orthogonal? E.g. [1, 0, 0] and [0, 1, 0] and [0, 0, 1] are orthogonal. Any other vector of length three, now say, [1, 1, 1] would share dependence with at least one of the former three vectors. $\endgroup$
    – tkl
    Jun 29, 2012 at 5:02
  • $\begingroup$ This is not an answer to the question; it is a new question. However, it cannot be answered well in comments. You should post it as a new question and let us help you properly. For the record, it is possible to have ($p$) orthogonal binary predictor variables, so long as $N>2^p$. $\endgroup$ Jun 29, 2012 at 13:36
  • $\begingroup$ @gung - I've addressed the issue with R; there is a small, but critical difference in how multinom and mlogit interpret formula coding for the basic regression. The major question was regarding the modeling framework and that seems to be okay. Multicollinearity was/remains a concern. Thanks for the suggestions. $\endgroup$
    – Aengus
    Jun 30, 2012 at 23:42
  • $\begingroup$ With respect to my point that MLR does not make assumptions about the dist of the IVs / that discrete IVs are fine, it might be helpful to read this question: what-if-residuals-are-normally-distributed-but-y-is-not & my answer there. Although it was written in the context of OLS regression, the situation is the same for both OLS reg & MLR in this respect. $\endgroup$ Aug 22, 2012 at 17:16

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.