[R] Multinomial Logit Models

[ripley@stats.ox.ac.uk]

On Fri, 24 Jan 2003, L.E.Gross wrote:

> I am wanting to fit some multinomial logit models (multinom command in
> package nnet)
> 
> Is it possible to do any model checking techniques on these models 
> e.g. residual, leverage etc. I cannot seem to find any commands that 
> will allow me to do this.

residuals() should work: it does for me.

Leverage is not AFAIK well-defined: remember you have a multivariate
response, and for binomial logit models `leverage' is in the linear
approximation to the glm.  A multinomial logit model is not a glm, and
there is no such linear approximation in use.  Also, I have doubts about
the usefulness of leverage even for binomial logit models: it's a local
approximation when the only possible changes (0 <--> 1) are not local.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595