[R-sig-ME] Glmm for positive count data

[Ben Bolker]

  (1) Are the issues with convergence generally the occurrence of
singular fits? (glmmTMB handles these less gracefully than glmer, since
it fits the RE variances on a logarithmic scale and relies more heavily
on Wald approximations in reporting

  (2) How slow is very slow? How big is your data set?

All other things being equal it's better to use the most realistic model
you can, but the biases due to model misspecification might not be too
huge.  How big are the differences you observe in the coefficients that
you're interested in, between a truncated-NB and a regular NB fit?

  If you have overdispersion you should *not* rely on a regular Poisson
fit (but if you've done something like add observation-level random
effects then you should be more or less OK, although arguably NB could
be better)

  Ben Bolker

On 2019-01-07 1:29 p.m., Naima M. wrote:
> 
> Dear all
> 
> I  have a positive count data response which I fitted with glmer and Poisson distribution (all worked well with no convergence problem in lme4).
> Recently, I realized that Truncated Poisson distribution may be more appropriate for my data since there is no zeros in my responses and tried to fit Truncated negative count distribution (I used truncated negative binomial because my models were overdispersed), but I have many issues with convergence and the run time is very slow using glmmTMB or glmmADMB.
> I'm wondering if I can keep my first models (glmer with Poisson), even if my data is positive?
> 
> Best,
> 
> Naïma
> 
> 	[[alternative HTML version deleted]]
> 
> _______________________________________________
> R-sig-mixed-models using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>