See McCullagh and Nelder (1989) (and the source code).
On Fri, 31 Mar 2006, Chad Reyhan Bhatti wrote:
> Hello,
>
> I have a question about the add1() function and quasilikelihoods for GLMs.
> I am fitting quasi-Poisson models using glm(, family = quasipoisson).
> Technically, with the quasilikelihood approach the deviance does not have
> the interpretation as a likelihood-based measure of sample information.
> Functions such as stepAIC() cannot be used. The function add1() returns
> the change in the scaled deviance between the null model and the
> alternative model along with a chi-sq p-value.
Not true. You choose a suitable test, and for quasipoisson the F test is
the appropriate choice and that does not give `a chi-sq p-value'.
If the dispersion is estimated, it is estimated via the residual deviance
(as it says on the help page) and that has a degrees of freedom in the
model output. The test is done in exactly the same way as a lm fit, just
replacing sums of squares by deviances (and indeed, exactly as a Gaussian
glm fit).
> Agresti, (Categorical Data Analysis p.140 last sentence before Section
> 4.5.2), makes the statement "For some GLMs the scaled deviance has an
> approximate chi-squared distribution." with no details, no references,
and
> no degrees of freedom.
>
> Does anyone know:
>
> 1) How the p-value is computed in add1(), ie degrees of freedom and maybe
> a reference?
>
> 2) A reference for using the scaled deviance for model comparison.
>
>
> Thanks,
>
> Chad R. Bhatti
--
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