[R] residuals in multinom

[Prof Brian Ripley]

A multinom model is not a glm, and its residuals() method follows the 
arguments of the generic and so is not otherwise documented. You got what 
?residuals tells you.

There is no analogue of `working': the model is not fitted by IWLS, and
the linear predictor is multidimensional.

One could calculate deviance and Pearson residuals, but I have never seen 
them used for multinom model.  It is not entirely clear how one would 
define Pearson residuals as there are K outcomes: do you want all K or 
only the one that happened?

On Mon, 17 May 2004, Schmidt.Matthias (FORST) wrote:

> is there a possibility to calculate the different "types" of residuals
> directly using the multinom function from MASS as it is possible for the
> functions gam, glm
> using type="deviance" or "working" or "pearson" or  "response"? I tried it
> but got always the "response" type, I guess. 

-- 
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