[R-sig-ME] MCMCglmm zero-altered

Jarrod Hadfield j.hadfield at ed.ac.uk
Tue Jul 31 19:52:57 CEST 2012


To use a ZAP model to test whether there is any zero inflation or  
deflation effects you want to hold the parameters constant across the  
Poisson and zero-altered part and compare them to a model in which  
they vary.

For the random effects this comparison would be random=~colony versus  
something more complex (idh(trait):colony or us(trait):colony). For  
the fixed effects you want to compare ~1 versus ~trait and  
percent.grass2 versus trait:percent.grass2 etc.

  For the overdispersion term MCMCglmm will not allow you to have the  
same "residual" for both parts, but ~trait:units allows the  
"residuals" for both parts to have the same distribution (although  
information regarding its variance only comes from the Poisson part).   
I believe this still allows valid testing of whether there is any  
zero-alteration or not (but as always, could be wrong). In the  
trait:units model you do NOT want to fix the variance: in your second  
prior you had fix=2 despite estimating a single variance(V=diag(1)) so  
it was probably ignored anyway.  I thought I had implemented MCMCglmm  
so it would generate an error if fix>nrow(V) - did you not get this?



Quoting "Brickhill, Daisy" <r01db11 at abdn.ac.uk> on Tue, 31 Jul 2012  
16:18:50 +0100:

> Hi,
> I am currently modelling the effect of different habitat variables  
> on the numbers of tipulid larvae found in soil cores using MCMCglmm.  
> The data is slightly zero inflated so I am trying a zero-altered  
> model (among others). I have used the following priors and model:
> prior1ZA = list(R = list(V=diag(2), n=0.002, fix=2), G = list(G1 =  
> list(V=diag(2), n=0.002)))
> model1ZA <- MCMCglmm(no._tips ~trait*(percent.grass2 + mean.veg.ht +  
> mean.soil.moisture + juldate + year),random = ~  
> idh(trait):colony,rcov = ~ idh(trait):units,
> family = "zapoisson", data = data, prior = prior1ZA, burnin = 3000,  
> nitt = 1003000, thin=1000)
> However I have read in a previous post by the immensely helpful  
> Jarrod Hadfield that "It is usual in zero-altered models to have the  
> zero bit and the truncated poisson bit have the same  
> over-dispersion. You do this by fitting the  interaction  
> rcov=~traits:units."
> I thought that ensuring the poisson and the zero process have the  
> same over-dispersion would require priors and model of the form:
> prior1ZA = list(R = list(V=diag(1), n=0.002, fix=2), G = list(G1 =  
> list(V=diag(1), n=0.002)))
> model1ZA <- MCMCglmm(no._tips ~trait*(percent.grass2 + mean.veg.ht +  
> mean.soil.moisture + juldate + year),random = ~ trait:colony, rcov =  
> ~ trait:units,
> family = "zapoisson", data = data, prior = prior1ZA, burnin = 3000,  
> nitt = 1003000, thin=1000)
> But looking at other posts I am beginning to think I am missing  
> something and that I *can* use my priors and model (with different  
> variances for the zero and poisson parts of the model). Is this  
> true? Can anyone tell me which of the two residual variance and  
> random effect structures is most advisable?
> Many thanks,
> Daisy
> The University of Aberdeen is a charity registered in Scotland, No SC013683.
> 	[[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.

More information about the R-sig-mixed-models mailing list