3
$\begingroup$

I have a Poisson glmm (using glmer) that is slightly over-dispersed at 1.854. I tried re-fitting it as a negative binomial model (using glmer.nb). On inspection of some of the diagnostic plots, I'm stuck with choosing which model is the better fit.

The qq-normal plot for the Poisson model (bottom left) looks better than that for the NB model (top left), while the fitted vs. residual plot looks somewhat better for the NB model (top right). I think that the structure that is appearing in the fitted vs. residual plot for the Poisson model is due to the fact that the data set is somewhat zero-inflated. enter image description here

My gut tells me to go with the slightly over-dispersed Poisson. Can anyone suggest a way to choose? Or a way to adjust for the over-dispersion in the Poisson model?

The data are from a paired design.

$\endgroup$
  • $\begingroup$ Why are you making a normal QQ plot of the residuals from a Poisson or a negative binomial model??? $\endgroup$ – GoF_Logistic May 3 '17 at 18:34
  • $\begingroup$ stackoverflow.com/questions/22842017/… $\endgroup$ – JKO May 3 '17 at 18:35
  • $\begingroup$ Having variance larger than the mean (the definition of overdispersion in a Poisson model) has absolutely nothing to do with the residuals (the pearson residuals I guess?) being normally distributed. $\endgroup$ – GoF_Logistic May 3 '17 at 18:46
  • $\begingroup$ Right. I am trying to assess which model is a better fit. If it is NB, then I don't need to worry about the dispersion parameter. If it is Poisson, then I need to correct for the over-dispersion. Thus, what I need is a way to distinguish between the two models. $\endgroup$ – JKO May 3 '17 at 20:00
4
$\begingroup$

I like the visual interpretation of how well a model fits using rootograms. The countreg package in R provides an implementation.

This paper has a nice description: https://arxiv.org/pdf/1605.01311.pdf

(I found the paper on cross-validated, but can't remember the post for further information)

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.