# Re: [R] Problem with multinom ?

From: John Fox <jfox_at_mcmaster.ca>
Date: Sun 12 Jun 2005 - 07:17:52 EST

Dear Marc,

> -----Original Message-----
> From: Marc Girondot [mailto:marc.girondot@ese.u-psud.fr]
> Sent: Saturday, June 11, 2005 2:16 PM
> To: John Fox
> Cc: R-help@r-project.org
> Subject: RE: [R] Problem with multinom ?
>
> >Dear Marc,
> >
> >I get the same results -- same coefficients, standard errors, and
> >fitted probabilities -- from multinom() and glm(). It's true
> that the
> >deviances differ, but they, I believe, are defined only up
> >
> >> predict(dt.b, type="response")
> > 1 2 3
> >0.4524946 0.5032227 0.5538845
> >
> >> predict(dt.m, type="probs")
> > 1 2 3 4 5 6
> >0.4524948 0.4524948 0.5032229 0.5032229 0.5538846 0.5538846
> >
> >These fitted probabilities *are* correct: Since the members of each
> >pair (1,2), (3,4), and (5,6) have identical values of factor
> they are
> >identical fitted probabilities.
>
> I expected rather to obtain (note 1- before some terms):
> > 1 2 3 4 5 6
> 0.4524948 1-0.4524948 0.5032229 1-0.5032229 0.5538846 1-0.5538846
>
> ...

But the fitted probabilities are for each observation in the data set; in your data, these have identical values of factor for each pair and hence identical fitted probabilities. When the response is dichotomous, you get only one of the two fitted probabilities for each observation; for a polytomous response, you get a matrix of fitted probabilities which sum to 1 row-wise.

Regards,
John

>
> Marc
> --
>
> __________________________________________________________
> Marc Girondot, Pr
> Laboratoire Ecologie, Systématique et Evolution Equipe de
> Conservation des Populations et des Communautés CNRS, ENGREF
> et Université Paris-Sud 11 , UMR 8079 Bâtiment 362
> 91405 Orsay Cedex, France
>
> Tel: 33 1 (0)1.69.15.72.30 Fax: 33 1 (0)1 69
> 15 56 96 e-mail: marc.girondot@ese.u-psud.fr
> Web: http://www.ese.u-psud.fr/epc/conservation/Marc.html
> Skype: girondot
> Fax in US: 1-425-732-6934

R-help@stat.math.ethz.ch mailing list