variance structure

An embedded and charset-unspecified text was scrubbed...
Name: not available
URL: <https://stat.ethz.ch/pipermail/r-sig-mixed-models/attachments/20110405/23ab09ed/attachment.pl>
Hello,

I'm running a model where the response is normally distributed and has an excess 
of zeros. Because I was familiar with the lmer package I used it to run a GLMM 
on this response, and addressed the problem of the excess of zeros by running 
two models: one with the response as a binary one, using the complete data set, 
and another excluding all the zeros and using the remaining response values in a 
Gaussian model. This seemed to work fine. However, a reviewer suggested using 
the whole data set and a zero-inflated poisson error structure in the MCMCglmm 
package. I don???t know if this is appropriate as my response are rates (grams of 
meat consumed per hr of observation) and not discrete values.

I would give advice on how to get zero-inflated models working with
MCMCglmm (mainly, see Ch 5 of the "CourseNotes" vignette that comes with
MCMCglmm, but I think the reviewer is wrong to think that you could use
a zero-inflated Poisson.  I think the way you did it is fine.  However,
if you *wanted* to be fancy you might (after careful reading of Ch 5,
and thought) be able to set up a zero-inflated normal model in MCMCglmm
in a way analogous to the way zero-inflated Poissons are set up.

  Ben Bolker
Hi Cristina,

I agree with Ben. In addition,  MCMCglmm will not fit ZIP models to  
these data (because the data are not integers) and ZIG (zero-inflated  
Gaussian) models are not implemented. In fact, I can't really see what  
the ZIG likelihood would look like, but anyway ...

If the non-zero data are well separated from the zero's (i.e. if  
pnorm(0, mean(y[which(y!=0)]), sd(y[which(y!=0)]))  is small) then  
fitting a bivariate binary/gaussian model is one option, but perhaps  
more complex than your problem requires.

Cheers,

Jarrod

On 11-04-05 06:30 PM, Cristina Gomes wrote:
Hello,

I'm running a model where the response is normally distributed and  
has an excess
of zeros. Because I was familiar with the lmer package I used it to  
run a GLMM
on this response, and addressed the problem of the excess of zeros  
by running
two models: one with the response as a binary one, using the  
complete data set,
and another excluding all the zeros and using the remaining  
response values in a
Gaussian model. This seemed to work fine. However, a reviewer  
suggested using
the whole data set and a zero-inflated poisson error structure in  
the MCMCglmm
package. I don???t know if this is appropriate as my response are  
rates (grams of
meat consumed per hr of observation) and not discrete values.

 I would give advice on how to get zero-inflated models working with
MCMCglmm (mainly, see Ch 5 of the "CourseNotes" vignette that comes  
with
MCMCglmm, but I think the reviewer is wrong to think that you could  
use
a zero-inflated Poisson.  I think the way you did it is fine.   
However,
if you *wanted* to be fancy you might (after careful reading of Ch 5,
and thought) be able to set up a zero-inflated normal model in  
MCMCglmm
in a way analogous to the way zero-inflated Poissons are set up.

 Ben Bolker

_______________________________________________
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.