[R-sig-ME] best practices and methods for fitting poisson/negative binomial glmm in R

J. Aaron Hogan jamesaaronhogan at gmail.com
Mon Aug 28 17:02:10 CEST 2017

Dear R-sig mixed modelers,

It is a pleasure to be a part of this group.  I have a few questions
regarding fitting a glmm where the data are poisson/ negative binomially

The data are count data from a factorial block design experiment.  3
blocks, 4 treatments (one of which is a control).

The data may or may not be zero inflated; I'm not sure if it matters for
the discussion.

I've got 5 fixed factors, 6 interactions between them and three random
factors; one for block, one for plot and one for a species effect (all
random intercepts).

I fit a model using the lme4 package using the glmer.nb() function and got
some failure to converge" errors. Those errors seemed relatively benign,
after reading some of Ben Bolkers github
and checking the Hessian gradient with:

relgrad <- with(model.final at optinfo$derivs,solve(Hessian,gradient))
max(abs(relgrad))  # returns a very small number

Everything seems alright with that model (based on my understanding); I can
predict the fit back to simulated data. I get reasonable coefficients,
deviance etc.

My question is more of a philosophical one.  How do I know I have a solid
model?  Given the vast number of glmm packages, options etc, when do I know
that it's okay to stop modeling and start inferring on the model?

Should I try to fit the model in a different package, like 'glmmadmb' using
MCMC methods?
What is the best package and/or practice for going through the process of
fitting a glmm for poisson/negative binomially distributed count data?

What other suggestions or recommendations do the more experienced modelers
have for this case?

I do appreciate your time and help,

All the best,

J. Aaron Hogan M.Sc.
PhD Student
Florida International University
(970) 485-1412


	[[alternative HTML version deleted]]

More information about the R-sig-mixed-models mailing list