[R-sig-ME] Fwd: model check for negative binomial model

[Ben Bolker]

  Short answer: if emergence success gets much above 50%, then the
approximation you're making (Poisson + offset for binomial, or NB +
offset for negative binomial) doesn't work well.  You might try a
beta-binomial (with glmmTMB) or a binomial + an observation-level random
effect.

  (On the other hand, your intercept is -0.3, which corresponds to a
baseline emergence of 0.42 - not *very* high (but some beaches and years
will be well above that ...)

  Beyond that, are there any obvious patterns of mis-fit in the
predicted values ... ?

On 2020-02-11 8:09 p.m., Alessandra Bielli wrote:
> Dear list
> 
> I am fitting a poisson model to estimate the effect of a treatment on
> emergence success of hatchlings. To estimate emergence success, I use
> number of emerged and an offset(log(total clutch).
> 
> However, overdispersion was detected:
> 
>> overdisp_fun(m.emerged) #overdispersion detected
> 
>       chisq       ratio         rdf           p
> 3490.300836    5.684529  614.000000    0.000000
> 
> Therefore, I switched to a negative binomial. I know overdispersion is not
> relevant for nb models, but the model plots don't look too good. I also
> tried to fit a poisson model with OLRE, but still the  plots don't look
> good.
> How do I know if my model is good enough, and what can I do to improve it?
> 
>> summary(m.emerged)
> Generalized linear mixed model fit by maximum likelihood (Laplace
> Approximation) ['glmerMod']
>  Family: Negative Binomial(7.604)  ( log )
> Formula: Hatched ~ Relocation..Y.N. + SP + offset(log(Total_Clutch)) + (1
> |Beach_ID) + (1 | Year)
>    Data: main
> 
>      AIC      BIC   logLik deviance df.resid
>   6015.6   6042.2  -3001.8   6003.6      614
> 
> Scaled residuals:
>     Min      1Q  Median      3Q     Max
> -2.6427 -0.3790  0.1790  0.5242  1.6583
> 
> Random effects:
>  Groups   Name        Variance Std.Dev.
>  Beach_ID (Intercept) 0.004438 0.06662
>  Year     (Intercept) 0.001640 0.04050
> Number of obs: 620, groups:  Beach_ID, 8; Year, 5
> 
> Fixed effects:
>                   Estimate Std. Error z value Pr(>|z|)
> (Intercept)       -0.29915    0.04055  -7.377 1.62e-13 ***
> Relocation..Y.N.Y -0.16402    0.05052  -3.247  0.00117 **
> SPL               -0.08311    0.04365  -1.904  0.05689 .
> ---
> Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
> 
> Correlation of Fixed Effects:
>             (Intr) R..Y.N
> Rlct..Y.N.Y -0.114
> SPL         -0.497 -0.054
> 
> 
> Thanks for your help,
> 
> Alessandra
> 
> 
> _______________________________________________
> R-sig-mixed-models using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>