Dear S-plus users,
I'm new to logistic regression, and relatively new to statistics in general, so
I've got a few basic questions:
I'm looking at the success of native species sown in the field. I have 28
species, and have 15 traits for each species that I would like to use to create
a regression model that can be used to predict the probability of establishment.
I have presence/absence data (0/1) for 9 species in 56 quadrats, with the traits
(continuous) serving as the predictor variables.
1. I am using stepAIC in the MASS library in S-plus to create a smaller model
from a list of 15 predictor variables. There is some evidence of
overdispersion (scale parameter > 2), and I am wondering if this affects how
stepAIC chooses the model, or if I can control for this after getting the model
from stepAIC.
2. Related to question 1, I've read that you can control for overdispersion by
using a quasi-likelihood model. Can I take the model specified by stepAIC, then
use the quasi family, logit link, and mu(1-mu) variance function to create a glm
that avoids the overdispersion problem? Let me know if I'm way off
track.
3. When I create the suggested glm from stepAIC, the residuals show skewness
and outliers. I'm reluctant to throw out the outliers, and according to
Venables and Ripley (3rd ed.) residual plots are not very useful in logistic
regression. I would be delighted if I could just ignore the residual plots,
but I don't feel like this is a good idea. What are the best diagnostics to
evaluate the fit of the model?
Am I proceeding in the right direction? I fear I'm making some fatal errors
that will prevent me from passing my final exam in the near future.
I am using S-plus 7.0, Windows XP.
Thanks for any help you can provide,
Rachael Roberts
Department of Botany & Plant Pathology
Oregon State University
|