[R-sig-ME] About glmer with a quasi-Poisson

[Jarrod Hadfield]

Hi,

This model can be fitted using MCMCglmm (and others I expect) if it is  
coded correctly. As Doug mentioned in an earlier email, calling  
different points in different sites the same name is not really  
necessary any more. If they are called different names then:

random=~site+point

is the same as the glmer model you use. If they are called the same thing then

random=~site+site:point

is the same as the glmer model below, as can be seen in the glmer summary.

Coded this way many packages can probably fit the type of model you envisage.

Cheers,

Jarrod





Quoting Strubbe Diederik <diederik.strubbe at ua.ac.be>:

> Hey Mike,
>
> I am working on a similar dataset as you. I have point counts of   
> birds in 26 sites, in total 57 points [number of point counts per   
> site varying from 1 to 5]. These points were visited 5 times ( in   
> the same year). Most of the counts are 0 (> 90%), and the dataset is  
>  thus zero-inflated - which, as far as I know is a special case of   
> overdispersion. There  are 3 continuous explanatory variables under   
> consideration. I have been looking at different packages to fit a   
> model with
> 1)a  zero-inflated poison distribution
> 2)a nested random effect to account for the fact that I have points   
> located within sites
> 3)a random effect to account for the 5 visits (repeated measures)
>
> However, none of the packages that I tried is able to do this ( I   
> tried zeroinfl {pscl}, fmr {gnlm}, MCMCglmm {MCMCglmm} and   
> glmm.ADMB{glmmADMB}). The problem seems to be the nested random   
> effect (1|site/point in my case).
>
> The only function that I found able to handle this nested random   
> design and a distribution that comes close to be 'good' for my data   
> is glmer with a quasi-poisson distribution. My  model syntax is:
>
> test <-   
> glmer(data$abundance~data$pca1+data$pca2+data$Gynoxis+(1|site/point)  
>  + (1|visit),family=quasipoisson(link="log"))
>
> and yields the folowing output
>
> Generalized linear mixed model fit by the Laplace approximation
> Formula: data$abundance ~ data$pca1 + data$pca2 + data$Gynoxis + (1   
> |      site/point) + (1 | visit)
>    AIC   BIC logLik deviance
>  158.1 187.1 -71.05    142.1
> Random effects:
>  Groups     Name        Variance  Std.Dev.
>  point:site (Intercept) 0.0468101 0.216356
>  site       (Intercept) 0.0000000 0.000000
>  visit      (Intercept) 0.0012396 0.035208
>  Residual               0.0187386 0.136889
> Number of obs: 276, groups: point:site, 57; site, 26; visit, 5
>
> Fixed effects:
>              Estimate Std. Error t value
> (Intercept)  -3.20340    0.10701 -29.936
> data$pca1    -0.02629    0.03100  -0.848
> data$pca2    -0.15055    0.02907  -5.180
> data$Gynoxis -0.02246    0.01320  -1.702
>
> Correlation of Fixed Effects:
>             (Intr) dt$pc1 dt$pc2
> data$pca1   -0.239
> data$pca2   -0.098 -0.071
> data$Gynoxs -0.878  0.330  0.205
>
> I was not aware of the comments of Ben Bolker on glmer and   
> quasi-Poisson, and am thus also interested in any responses on this.  
>  Am I right to conclude that, ecept glmer, none of the packages   
> mentioned above is capable to handle nested random effects and   
> zero-inflated Poisson data?
>
> Cheers and many thanks,
>
> Diederik
>
>
>
> Diederik Strubbe
> Evolutionary Ecology Group
> Department of Biology, University of Antwerp
> Universiteitsplein 1
> B-2610 Antwerp, Belgium
> http://webhost.ua.ac.be/deco
> tel : 32 3 820 23 85
>
>
> 	[[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
>



-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.