[R-sig-ME] Overdispersion in Poisson mixed-model?

[Marie-Hélène Hachey]

Hi,

I would like to know, in lmer, how to compute...:

1-Overdispersion of the model (estimated scale parameter)
2-% of variation explained by a model


1- For overdispersion, I found (on this mailing list) thant you can do
sqrt(sum(c(model at resid, model at u)^2)/length(model at resid)) = 1.29

(what is "u", by the way?)

But I also found #lme4:::sigma(model) = 1

Both dont give the same result. Which one should I use? And can I use this value
to say that my residuals are no longer overdispersed, due to the addition of random effects? 
When I had a Poisson glm model, the res. deviance/res. DF was 4.09.

2-For the % of variation explained, I used
(null model @deviance["wrss"] - model at deviance["wrss"])/null model at deviance["wrss"]
(also found on that mailing list, except the value was divided by the model deviance, instead of
the null model deviance... I thought it was the other way around...)
Do you think this is right to comment on how my model explained the variation observed? And why do 
we use weighted r sum of squared instead of, say, ML deviance?

Last question: if a model only explains 5% of the variation, should I not use it at all? I' m sorry, I know
it's a stats question but it's been bugging me.

Thank you in advance

Marie-Helene Hachey
M. Sc. Student