From ripley at stats.ox.ac.uk Tue May 9 15:30:33 2000
Date: Tue, 9 May 2000 06:29:10 +0100 (BST)
Following p.206 of "Statistical Models in S", I wish to change
the code for summary.glm() so that it estimates the dispersion
for binomial & poisson models when the parameter dispersion is
set to zero. The following changes [insertion of ||dispersion==0
at one point; and !is.null(dispersion) at another] will do the trick:
I know S does that, but R is not documented to do so (so your example does
works as documented). I think this is at best confusing. Once phi is
estimated, they are quasi-likelihood models not binomial nor Poisson and in
particular have no likelihood, so e.g. drop1 becomes inappropriate. And it
is all too easy to have different treatments of dispersion in different
ancilliary functions (as S managed for many years).
That is why I did not submit a bug report. The problem is that in
many application areas phi is much greater than one.
My preference is to treat such models as quasi models. If enough
people really want them as `binomial' or `poisson' then we need to
make much wider changes to ensure consistency (and incompitibility
with S).
I agree with the sentiments. So would it be feasible to have
quasi-poisson and quasi-binomial errors? Would an immediate recourse
be to create functions summary.quasi and predict.quasi? Or perhaps
summary.phi, etc.