Stepwise multinomial logistic regression

Dear all, I want to implement a stepwise multinomial logistic regresssion. I could not find any inbuild function for this procedure in multinomial logistic regressions in particular.

I have run mnrfit over my data (with 3 categories and 44 predictors). I do not know how I can compute a maximum likelihood estimator from the output or something equivalent which I can use for -LL-ratio testing to reduce the number of predictors. I am also not sure what algorithm with respect to reduction and stopping criteria would be the best.

Is there any function that does what I want, if not how can I compute ML for each of my logistic regressions and what comparison criterion (-LL, AIC, BIC) would you recommend in this case, and when should I stop reducing parameters?

Best wishes and many thanks in advance!

Daniel

[Erik]

Are you still having this question? If not: how did you solve your problem? I partially have the same question: I'm looking for a model to predict a multinomial response based on many possibly features. I want to do feature selection based on stepwise regression. Using

[beta,dev,stats] = mnrfit(...)

you can get the model deviance

dev

which is a function of log likelihood of the model with respect to best achievable log likelihood. Am I correct that deviance is zero if the model fits the data perfectly? See here for more info: link.

To implement your own stepwise regression algorithm could be possible. A function like

stepwiseglm

implements stepwise regression for models very similar to multinomial regression, e.g. binomial regression. For that function deviance is the default parameter on which the decision which model is best is based. AIC, BIC, etc. can be calculated (also manually) and used instead. In the documentation the stepwise regression algorithm is described.

Answer 1

[Thomas Gorman]

I have the same question. Did you ever find a solution?