COMPoisson family for underdispersed count data

Lugz

New Member
Hello,

I am trying to test whether the species richness of butterflies in a natural reserve differs betweeen habitats and years separately (not enough statistical replicates to make intercation). The monitoring of butterflies started in 2002. There is 3 different habitats : xerophytic and mesophytic calcareous grasslands and edge. My data looks like that (I take the example of generalists species):
...... unitl 2018

Normally i should use GLMM with Transects as random factor (paired data). Because of the poor transect number per habitats I can't. So my model is the following :

Code:
Call :

Deviance Residuals:
Min        1Q    Median        3Q       Max
-1.44563  -0.33385   0.01237   0.36883   1.18072
[...]
Null deviance: 108.922  on 169  degrees of freedom
Residual deviance:  45.301  on 144  degrees of freedom
AIC: 773.95

Number of Fisher Scoring iterations: 4
I observe underdispersion. I make the ratio between the Residual devience and degrees of freedom to check it : 45.301/144 = 0,31

So, i tried COMPoisson but standard glm function do not recognize this family.
I used fitme function from spaMM package.
Code:
> (a=fitme(Rgeneralists~Habitats/Transects+Years,data=abvalb,family=COMPoisson()))
formula: RL1 ~ Habitats/Transects + Years
Estimation of fixed effects by ML.
Estimation of COMP_nu by 'outer' ML, maximizing p_v.
Family: COMPoisson(nu=3.979) ( link = loglambda )
[...]
------------- Likelihood values  -------------
logLik
p(h)   (Likelihood): -309.1393
I also tried glm.cmp function but i really don't understand what are the type of argument that should be used.

Do someone already used COMPoisson ?

Thanks a lot !

katxt

Active Member
By the look of it, you could probably just assume the numbers are normal.

Lugz

New Member
By the look of it, you could probably just assume the numbers are normal.
I checked the residuals of the model, it's normal distribution ... thank you.