[R-sig-ME] Questions about mix models

[Jarrod Hadfield]

Dear Alena,

As other's have said, its hard to assess the problem without more  
information. However, with regards to MCMCglmm, the model you  
specified is equivalent to the model you tried to fit in glmer (in  
terms of random effects) but models over-dispersion, which is  
important. However, you have fixed the residual variance to one which  
you should not do - this is only for categorical and ordinal responses.

If glmer is issuing warnings of this sort it often suggests there may  
be some problems with the model and so I would be very careful. If  
there is little replication for the random effects the priors you use  
will be quite informative. There is no perfect prior but I often find  
parameter expanded priors to work well:

prior=list(R=list(V=1, n=0), G=list(G1 = list(V =1,n=1, alpha.mu=0,  
alpha.V=1000),G2=list(V=1,n=1, alpha.mu=0,  
alpha.V=1000),G3=list(V=1,n=1, alpha.mu=0, alpha.V=1000)))

Cheers,

Jarrod


On 16 Aug 2010, at 10:58, Alena Drašnarová wrote:

> Dear all,
> I have so complicated data and I am trying to gain correct results  
> from them.
>
> I am interested in factors influencing density and diversity of the
> soil seed bank on alluvial meadows. I have nested design of my
> experiment: 35 meadows (mead=M1-M35), three transects on each meadow
> (trans=T1-T3) and  2 plots on each transect (top=A,B).
> I found out  a lot of information (about soil properties, moisure,
> litter, biomass, vegetation diversity and management).
> At first, I tried to use glmer, but sometimes there was error message:
>
>> a2<-glmer(number~top+depth+HPV+K+VVS+(1|mead/trans/ 
>> top),data=dat,family=poisson)
> Warning messages:
> 1: In mer_finalize(ans) :
> Cholmod warning 'not positive definite' at
> file:../Cholesky/t_cholmod_rowfac.c, line 432
> 2: In mer_finalize(ans) :
> Cholmod warning 'not positive definite' at
> file:../Cholesky/t_cholmod_rowfac.c, line 432
> 3: In mer_finalize(ans) : false convergence (8)
>
> So, I decided to use MCMCglmm, but I am not sure with fitting the
> model. I tried to fitt it by this way (example below is for one
> factor):
>
>> prior=list(R=list(V=1, n=0, fix=1), G=list(G1 = list(V  
>> =1,n=1),G2=list(V=1,n=1),G3=list(V=1,n=1)))
>> m1 <- MCMCglmm(number ~ as.factor(top), random=~mead+mead:trans 
>> +mead:trans:top, family = "poisson", data=dat,prior=prior)
> I am not sure with define prior and random effect.
>
> I will be very happy, if anybody write me own experiences with these
> models and similar data and help me which model is the best to use.
>
> With kind regards
> Alena Drašnarová
>
> _______________________________________________
> 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.