Thank you for your reply.
I am aware of the good reasons not to use the deviance estimate in
binomial, Poisson, and gamma families.
However, for the inverse Gaussian, the choice seems to me less clear
cut. So I just wanted to compare two different options.
I have used the dispersion parameter to compute the standardized
deviance residuals:
summary(model.gamma)$deviance.resid /(summary(model.gamma)$dispersion *
(1 - hatvalues(model.gamma)))^0.5
I noticed differences with Genstat which outputs these stand. dev.
residuals directly, and they are explained by the automatic use of
deviance instead of Pearson.
Kind regards,
Robin Smit
-----Original Message-----
From: Prof Brian Ripley [mailto:ripley@stats.ox.ac.uk
]
Sent: dinsdag 12 juli 2005 15:12
To: Smit, R. (Robin)
Cc: r-help@stat.math.ethz.ch
Subject: : Re: [R] Dispersion in glm (was (no subject))
Actually, glm() does not estimate the dispersion at all, so you will
need to be more specific.
For example, summary.glm() and predict.glm() use the Pearson statistic
if dispersion=NULL (the default) for most families. You can supply any
other value you choose, and the MASS package makes use of this for ML
estimation of the dispersion parameter (related to the shape) of the
gamma family.
There are rather good reasons (serious bias) not to use the deviance
estimate in the binomial and Poisson families (see the example plots in
MASS4), and good reasons not to use either in the gamma family. As the
Pearson and deviance estimates agree for the gaussian, that does leave
begging the question of why you want to do this. Further, McCullagh &
Nelder have general arguments why the Pearson estimate might always be
preferred to the deviance one. So that `another statastical package'
appears to need justification for its choice.
On Mon, 11 Jul 2005, Smit, R. (Robin) wrote:
> The estimate of glm dispersion can be based on the deviance or on the
> Pearson statistic.
> I have compared output from R glm() to another statastical package and
> it appears that R uses the Pearson statistic.
> I was wondering if it is possible to make use R the deviance instead
> by modifying the glm(...) function?
> Thanks for your attention.
--
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
This e-mail and its contents are subject to the DISCLAIMER at http://www.tno.nl/disclaimer/email.html
[[alternative HTML version deleted]]