[R-sig-ME] Model selection for mgcv::gamm (using PQL)

[Cornelia Oedekoven]

I am aware this topic is not necessarily lmer() related but was hoping I could
get some expert oppinion on this here nonetheless.

I am trying to do model selection for quasipoisson GAMM using PQL and have the
three models specified below.  What method should I use to select the best
model (note that “Phase” has two levels)?

gamm2<-gamm(response~s(SST)+as.factor(Month) +
s(Long,Lat,by=as.factor(Phase))+offset(logArea),random=list(transect=~1),
data=all.data,family=quasipoisson, niterPQL=40)

gamm1<-gamm(response~
s(SST)+as.factor(Month)+s(Long,Lat)+as.factor(Phase)+offset(logArea),
random=list(transect=~1), data=all.data,family=quasipoisson,niterPQL=40)

gamm0<- gamm(response~s(SST)+as.factor(Month)+s(Long,Lat)+offset(logArea),
random=list(transect=~1),data=all.data,family=quasipoisson,niterPQL=40)

Each model has an $lme and a $gam object where the former is fitted using nlme.
I have considered to compare AIC values from the $lme outputs but since the
log-likelihood is not from the fitted GAMM I assume this is not the appropriate
method.
Crossvalidation would take too long for this study (simulation study with a
large number of large data sets to which the models are fitted).
The function gamm4::gamm4 uses lme4 instead of nlme (and avoids PQL). It does,
however, not allow for specifying quasi-families and the data is overdispersed.
I am aware that QAIC has been used for overdispersed GLM models but I cannot
find anything that says this it is appropriate for overdispersed GLMM or GAMM.
Many thanks and any help is much appreciated.

Cheers, Cornelia


<>< <>< <>< <>< <>< <>< <><
Cornelia Oedekoven
CREEM
University of St Andrews
cornelia at mcs.st-and.ac.uk
www.creem.st-and.ac.uk
<>< <>< <>< <>< <>< <>< <><

The University of St Andrews is a charity registered in Scotland : No SC013532





------------------------------------------------------------------
University of St Andrews Webmail: https://webmail.st-andrews.ac.uk