R devel - Jan 2003 - help page for anova.glm/variation between S-PLUS and R behavior

When using test="F" in stat.anova() / anova.glm(), R uses the assumed 
dispersion parameter for the specified family (e.g. scale=1 for binomial), 
while S-PLUS automatically uses the estimated dispersion parameter 
(residual deviance/residual df).  I think there are good reasons for the 
behavior in R -- it fits with the "you get what you actually asked
for"
philosophy -- and there are hints in the documentation (anova.glm(): "for 
[models] with dispersion estimated by moments ... the F test is most 
appropriate"; family/binomial(): "The behaviour of S-PLUS is closer to
the
quasi- variants").
   Would R-core consider it reasonable to add yet another sentence to the
anova.glm() documentation, e.g. after the "... F test is most
appropriate"
sentence above, adding: "(Note: unlike R, S-PLUS automatically uses the
moment-estimated dispersion for F-tests even for the binomial and Poisson
families:  see ?quasibinomial.)"  ?

  Ben Bolker

-- 
318 Carr Hall                                bolker@zoo.ufl.edu
Zoology Department, University of Florida    http://www.zoo.ufl.edu/bolker
Box 118525                                   (ph)  352-392-5697
Gainesville, FL 32611-8525                   (fax) 352-392-3704

Not really!

I don't think it is our job to describe the inconsistencies of S-PLUS,
especially since that might change from version to version. It would be
really messy, as S-PLUS ignores the scale for the Chisq test (which seems
perverse), and lacks the dispersion argument.

On Mon, 27 Jan 2003, Ben Bolker wrote:
> 
>    When using test="F" in stat.anova() / anova.glm(), R uses the
assumed
> dispersion parameter for the specified family (e.g. scale=1 for binomial), 
> while S-PLUS automatically uses the estimated dispersion parameter 
> (residual deviance/residual df).  I think there are good reasons for the 
> behavior in R -- it fits with the "you get what you actually asked
for"
> philosophy -- and there are hints in the documentation (anova.glm():
"for
> [models] with dispersion estimated by moments ... the F test is most 
> appropriate"; family/binomial(): "The behaviour of S-PLUS is
closer to the
> quasi- variants").
>    Would R-core consider it reasonable to add yet another sentence to the
> anova.glm() documentation, e.g. after the "... F test is most
appropriate"
> sentence above, adding: "(Note: unlike R, S-PLUS automatically uses
the
> moment-estimated dispersion for F-tests even for the binomial and Poisson
> families:  see ?quasibinomial.)"  ?
> 
>   Ben Bolker
> 
> 
-- 
Brian D. Ripley,                  ripley@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