On Sun, 22 Apr 2001, Rosalind Renfrew wrote:
> I have a question for Splus experts:
>
> Burnham and Anderson, in their book "Model selection and inference" about
> Akaike Information Criteria (AIC), said that (as of 1989) there are no
1989 is a very long time ago: S-PLUS 2.0 as I recall.
> statistical software packages that compute QAIC, the specific form of AIC
> calculated when using the quasi function in generalized linear models. When
> I use Splus 2000 to use quasi, it gives me an AIC value under the "step"
> function, but can I assume this is the correct QAIC value? Or do I need to
No, not even for a binomial or Poisson model. It's the wrong definition
applied to the linear approximation at the fitted values. That's why there
is stepAIC in library MASS.
> calculate it separately by hand? If I need to calculate it, I need the
> log-likelihood value for the model, which I need to divide by the dispersion
> parameter (c-hat). Is there a straightforward way of obtaining the
> log-likelihood in the output?
A quasi model does not have a log-likelihood, and does not even necessarily
have a quasi-likelihood, although for the variance families implemented in
glm() in S one is assumed to exist. So what exactly is the definition
of `quasi AIC'?
The issue seems to be to do with using binomial/Poisson model with
overdispersion, from what you say. In that case stepAIC will do
what you seem to want from your `manual' description.
--
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 272860 (secr)
Oxford OX1 3TG, UK Fax: +44 1865 272595
|