On Mon, 31 Jan 2005, Nicola Koper wrote:
Hello everyone,
I am using S-plus 6.2 on Windows XP and Windows ME.
When using the drop-down menu to do a glm or logistic
regression, the output isn't giving me p-values at
all. I realize it is becoming fashionable not to
refer to p-values at all,
Is it?
but I still find them
useful! Not having had this problem with other modules
within S-plus that I've used in the past, including
glme which I would have thought would be similar to
glm, I'm at a loss for what to do. Please could
someone let me know how to get p-values for these
simple analyses.
It is unfortunately not at all simple.
I presume you are looking at a summary table that has `t value's? Those
`t' values are approximately normally distributed, so you can use pnorm()
to give you approximate p-values for them. However, there are at least
three good reasons why the authors of glm() might have decided not to have
done that for you.
1) They are rather approximate, as indeed they are in glme.
2) The Hauck-Donner effect means that these Wald tests for coefficients
can be very seriously misleading: a small t value can mean either a small
effect or a very significant effect. See the literature, e.g. MASS4, for
more details.
3) It is not clear in Poisson and binomial GLMs whether you want to allow
for over-dispersion or not, and if so what correction to use.
You would normally be much better off using anova() to test for terms in
the model by a likelihood-ratio test.
--
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
|