I am using R's glmmTMB for modeling negative binomial mixed effects. In the output, I see the following line : Overdispersion parameter for nbinom2 family (): 9.28e+06.

How do I interpret such a large overdispersion? Please help.

  • $\begingroup$ did you ever find more information about this? I'm facing a similar issue for a glmmTMB model. I only get this extreme overdispersion when I add an additional random intercept to my model, however. $\endgroup$
    – m.evans
    Commented May 6, 2019 at 13:12
  • $\begingroup$ I didn't find any further information on this. Not sure what to make of this and how to address it. $\endgroup$
    – SanMelkote
    Commented May 7, 2019 at 14:06
  • $\begingroup$ I'm having the same problem with glmmTMB, I was wondering how much overdisperion you got when you calculated it? I used this formula from Zuur, Savaliev and Ieno (2012): “by calculating the residuals, taking its sum of squares, and dividing by N - p, where N is the sample size and p the number of parameters (i.e. regression parameters and parameters in the random part of the model)”. res <- residuals(M1) p1 <- length(fixef(M1))+1 # +1 due to random intercept variance overdisp1 <- sum(res^2)/(nrow(data1)- p1) overdisp1 [1] 2.942775 Where you using a regular NB regression or a zero inflated mo $\endgroup$
    – Julie.gjm
    Commented May 17, 2019 at 13:41

2 Answers 2


I think a value of

Overdispersion parameter for nbinom2 family (): 9.28e+06

actually means no overdispersion. This is the theta parameter of a NB2 model, see also

What is theta in a negative binomial regression fitted with R?

That is, the right thing in this case would be fitting a Poisson mixed model.


I'm not 100% but I don't think it's a normal value even with a negative binomial distribution. And I think it requires model improvement. I had a similar problem and managed to almost solve it.

The final code gave me a much lower overdispersion, though it was still overdispersed when checked with the DHARMa test for dispersion (p < 0.05). Here it is:

glmmTMB(count ~ distance_to_pond * rainfall + distance_to_river * rainfall + (1|cell) * (1|date) + offset(log(area)), ziformula = ~1, family = nbinom2)

I initially had the random effects as nested so (1|date/cell) and shifting to crossed effect almost solved my issue. These random effects are for mitigating spatial and temporal autocorrelation (and it worked according to DHARMa tests for autocorrelation, though it is mentionned in the guidelines not to fully rely on them and perform further tests).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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