[R-sig-ME] ZIP MCMCglmm AND separate random est. for levels of fixed factor

[Jarrod Hadfield]

Dear Evelien,

In answer to your questions:

1) Is it conceptually a good idea? I am new with MCMCglmm but I
couldn't find any posts on similar problems and wondered whether this
should be a warning sign.

Yes, I think the model makes sense, and you seem to have a lot of data  
replicated at the correct level.

2) And pragmatically? Am I bound to get severe convergence issues
(just asking for an educated guess)?

ZIP models are not well implemented in MCMCglmm and so tend to  
converge/mix poorly, particularly if the probability of zero-inflation  
is extreme (0 or 1). Hurdle and zero-altered models are may be a  
better alternative.

3) What would the syntax of such a random structure be?

I think it should be idh(at.level(trait,1):cln):spp

4) and should I be aware of any specific considerations regarding the
prior?

You should probably fix the residual variance for the zero-inflation  
at something (e.g. 1) because it cannot be estimated from the data.  
The results can be sensitive to the prior, and if you don't have good  
prior information you should keep nu small (<<1) or use  
parameter-expanded priors (I tend to use these more and more). Always  
check to see how influential the prior is.

Cheers,

Jarrod



Quoting Evelien Jongepier <evelien_jongepier at hotmail.com>:

>
> Dear all,
>
> I would appreciate some help on the following:
> I am trying to run MCMCglmm with zipoisson distribution AND separate  
> random effects for a fixed binary factor
>
> Some background:
> I want to test for the effect of clonality (cln -> binary (yes or  
> no)), environmental factors (e.g. fire and rain) and their  
> interaction on species abundances (ab), where I use sampled location  
> (pl.id) as random factor.
>
> my data looks smth like this:
> pl.id  - spp - ab - cln - fire.....
> 001 - 01 - 28 - n - 10
> .....
> 001 - 35 - 11 - y - 10
> .....
> .....
> 770 - 01 - 10 - n - 17
> .....
> 770 - 35 - 12 - y - 17
> [so I've got 770 plots and 35 species]
>
> Species abundance is poisson distributed but severely zero inflated  
> as each species only occurs in a subset of plots.
> Moreover, I have reason to believe that clonality (a species level  
> trait) further contributes to zero inflation (so I add trait:cln).
> Otherwise I am not interested in the effect of environmental vars or  
> random factors on the zero-inflation process (so I use  
> at.level;(trait,1):...).
>
> which comes down to [please correct me if I am wrong]:
>
> mdl <- MCMCglmm(
> ab ~ trait - 1 + trait:cln + at.level(trait,1):(cln*(fire + rain)),
> random=~idh(at.level(trait,1)):pl.id,
> rcov = ~idh(trait):units,
> prior=pr1, family = "zipoisson",
> .....data=dat)
>
> where pr1 is e.g. a flat prior for B and half-Cauchy prior for G
>
> However, I would like to account for variation explained by species  
> for clonals and non-clonals separately.
> Without zero-inflation this would look something like:
>
> random=~pl.id + idh(cln):spp
>
> But now I want to include this structure in "mdl", where the random  
> effects are fitted to test only for their potential effects on the  
> poisson processes
> Intuitively I would suspect a syntax like this:
>
>
> random=~idh(at.level(trait,1)):pl.id +  
> idh(at.level(trait,1)):idh(cln):spp or perhaps
> random=~idh(at.level(trait,1)):pl.id + idh(at.level(trait,1)):cln:spp
>
>
> I fiddled around quite a bit but generally got this message:
>
> Error in buildZ(rmodel.terms[r], data = data) : object (idh(cln) and  
> pl.id) not found
>
> So my questions are:
> 1) Is it conceptually a good idea? I am not overly familiar with  
> MCMCglmm but I couldn't find any posts on similar problems and  
> wondered whether this should be a warning sign.
> 2) Am I bound to get severe convergence issues (just asking for an  
> educated guess)?
> 3) What would the syntax of such a random structure be.
> 4) and should I be aware of any specific considerations regarding the prior?
>
> Would be great if anyone could enlighten me a bit or perhaps share  
> experiences if you are trying to fit a similar model
>
> Thanks!
> Evelien Jongepier
>
> 	[[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.