Dear all,
The deviance for a binary logistic model is a function of the data through
its parameters (beta_hat) alone, which means that it cannot provide
information about how well the model fits. Is this true for other GLM
binomial models with non-canonical links? I am particularly interested in
using -- link = log --, as this model seems to be of standard use in
epidemiology for the estimation of risk (prevalence).
An immediate alternative to this model would be a GLM Poisson model with
log link (with robust variance estimation). Would the deviance for this
second model suffer from the same limitations as in the first case?
References and/or comments will be much appreciated.
Thanks,
Rafael
|