[R-sig-ME] GLMM with overdispersion and Wald test

[ONKELINX, Thierry]

Dear Thomas,

Removing the zero observations is IMHO a problem. Suppose one chick has
been fed and the other not. The data from the latter is missing from
your dataset. Then how can you see the difference between both chicks?

If I remember well, there were (are) some concerns about the quasi
family. The latest recommendation was to use the poisson family and add
an observation-level random effect. This random effect will have
variance when overdispersion is present.

Furthermore, you wrongly specified the nested random effect. It should
be (1|Nest/ChickID) instead of (1|ChickID/Nest).

Residiual degrees of freedom are not clearly defined in a mixed model.
Suppose you have N observations, f fixed effects df and one random
intercept with r levels. Then the residual degrees of freedom will range
from N - 1 - f - 1 (random effect takes one df) to N - 1 - f - r (random
effect takes r df)

Best regards,

Thierry

------------------------------------------------------------------------
----
ir. Thierry Onkelinx
Instituut voor natuur- en bosonderzoek
team Biometrie & Kwaliteitszorg
Gaverstraat 4
9500 Geraardsbergen
Belgium

Research Institute for Nature and Forest
team Biometrics & Quality Assurance
Gaverstraat 4
9500 Geraardsbergen
Belgium

tel. + 32 54/436 185
Thierry.Onkelinx at inbo.be
www.inbo.be

To call in the statistician after the experiment is done may be no more
than asking him to perform a post-mortem examination: he may be able to
say what the experiment died of.
~ Sir Ronald Aylmer Fisher

The plural of anecdote is not data.
~ Roger Brinner

The combination of some data and an aching desire for an answer does not
ensure that a reasonable answer can be extracted from a given body of
data.
~ John Tukey
  

> -----Oorspronkelijk bericht-----
> Van: r-sig-mixed-models-bounces at r-project.org 
> [mailto:r-sig-mixed-models-bounces at r-project.org] Namens 
> tmerklin at cict.fr
> Verzonden: vrijdag 7 januari 2011 21:24
> Aan: r-sig-mixed-models at r-project.org
> Onderwerp: [R-sig-ME] GLMM with overdispersion and Wald test
> 
> Dear list members,
> 
> This is my first post on the list. I hope that I will be clear.
> 
> I'm studying chick feeding in a bird species and I want to 
> know if parents feed their male chicks more than their female chicks.
> I have data concerning feeding bouts (count data) and I also 
> consider chick rank (1st born or 2nd born) and chick age (in 
> classes), because they can possibly influence feeding.
> 
> Firstly, I only consider feeding events, so that I don't have 
> any zero in my data. Observations have been made during 15 
> minutes three times a day. I calculate the number of feedings 
> for each chick per day, if a chick is not fed during an 
> observation it will not appear in the dataset. Is this a 
> problem if my data don't contain any zero, when Poisson or 
> quasipoisson is specified ?
> 
> I specified my model as following:
> model<-lmer(Number~SexChick*EggRank*AgeClass+(1|ChickID/Nest)+
> (1|Date),data=feeding,family="quasipoisson",REML=FALSE)
> 
> I assessed overdispersion using: (phi<-lme4:::sigma(model)), 
> and its values was between 1.5 and 4 depending on the analyses.
> 
> Is it correct ?
> 
> Then, to test the interactions and the variables I used the function
> anova(model1,model2) with both models differing from only one 
> term i.e. I used likelihood ratio test (LRT). However I 
> recently read in Bolker et al 2009 TREE that LRT are not the 
> best solution when sample size is small to moderate and 
> especially when data are oversdispersed.
> 
> They adviced to run F Wald test or to use QAIC.However the 
> problem seems to be that the software is unable to calculate 
> the residual degrees of freedom (rdf). I could not find a 
> method in R to assess these rdf.
> Have I missed something ?
> 
> Finally, assuming that I find a way to assess these rdf I'm 
> not use how to use F Wald test. There is a function in the 
> aod library called wald.test(). I don't understand if we have 
> to test each coefficients from the full model or if we need 
> to run models with and without a specified tems to test its 
> significance. Can you help concerning this function too ?
> 
> I don't want to use QAIC because I have other analyse that 
> I've made using p-values so I don't want to mix the two approaches.
> 
> I hope that I have been clear in the explanations of my 
> problem and I hope that someone will be able to help me !
> 
> Thank you by advance!
> Thomas Merkling
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list 
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>