Dear Splus users,
I fitted a GLM with a Poisson distribution. The model involves 3 factors
and all interactions. I am interested in testing effects involving 2
factors and their interaction, say A*B.
I carry out multiple comparisons on differences between adjusted means
on A*B (A+B+A:B), but I am interested by tests on the scale of the
original response.
My first question is : how to use predict.glm et multicomp for
calculating adjusted means and testing several contrasts involving
differences on these adjusted means?
If this implies using multicomp.default, what vmat and nn should be
specified ?
Second question:
May I use summary.aov on a GLM object (to have Type III sums of squares)
(are these computations appropriate for GLM?)
Third question:
Because I present GLM results along with ANOVA results, I would like to
report an equivalent of R2 for GLM. I was considering using :
(null deviance - residual deviance)/null deviance
Is this appropriate ? Is there a better measure ?
Thank you very much for your help,
--
Dominique Pelletier
Laboratoire MAERHA
IFREMER BP 21105
44311 Nantes Cedex 03
FRANCE
E-mail: dpellet@ifremer.fr
Tel: 33 2 40374164
Fax: 33 2 40374075
http://www.ifremer.fr/maerha/
|