>On Sat, 11 Jun 2005, John Fox wrote:
>
>>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 to an additive
constant:
>
>Yes. There are many variations on the definition
>of (residual) deviance, but it compares -2 log
>likelihood with a `saturated' model. For
>grouped data you have a choice: a separate term
>for each group or for each observation. A
>binomial GLM uses the first but the second is
>more
>normal in logistic regression (since it has a
>direct interpretation via log-probability
>scoring).
>
>multinom() is support software for a book (which
>the R posting guide does ask you to consult):
>this is discussed with a worked example on pp
>203-4.
Dear Prof. Ripley,
I have your book... but I don't find the answer to my questions...
You propose that the difference in residual
deviance between two versions of the same model
(0.001841823 for glm and 106.2304 for
multinom()) is due to a difference in the
specification of the satured model. However, as
RD=-2 Ln L model+2 Ln L saturated and that -2 Ln
L model=11.1146... it seems impossible to me that
RD > -2 Ln L model ...
Marc Girondot
Sorry to be so close-minded !
--
__________________________________________________________
Marc Girondot, Pr
Laboratoire Ecologie, Syst??matique et Evolution
Equipe de Conservation des Populations et des Communaut??s
CNRS, ENGREF et Universit?? Paris-Sud 11 , UMR 8079
B??timent 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 at ese.u-psud.fr
Web: http://www.ese.u-psud.fr/epc/conservation/Marc.html
Skype: girondot
Fax in US: 1-425-732-6934
