Skip to content
Prev 6991 / 20628 Next

mixed zero-inflated poisson regression

Hi Jarrod (and all other r-sig-mixed-members who answered me)

First, thanks to all who replied. I learned a lot from all answers, and 
it seems I have got a better understanding on the topic now.

Jarrod Hadfields suggestion about priors helped me a lot to get betters 
models. As you suggested my models did not run well at all and there 
were several issues about my code. There were a lot of serious problems 
associated with the models (autocorrelation, correlation between chains 
from different parameters, poor convergence). Refining my code according 
to suggestions makes the models now run much better. I also centralised 
the predictor variable, which helped a lot on the correlation between 
the chains of the parameter estimates for the fixed factors. It also 
appears that the models have to run for some time in order to converge, 
and that I need a quite high thinning to reduce autocorrelation. 
Finally, using the priors suggested by Jarrod (with the 
fix=2-adjustment) of course did the trick and made models more stable 
and reliable. Currently I am running 3 MCMCglmm for each model, with 
nitt=1000000, burn=500000 and thin=1000. This seems to do give 
reasonable output (Gelman diagnostics < 1.02 for all chains), which are 
consistent among similar runs, although it takes some time.

As suggested by other list members, it may be a good idea to verify 
results using a MLE-based approach, for instance using ADMB. A recent 
(future) book was also suggested (even as a Christmas present for 
myself. I was thinking more in line of a pair of new skis, but perhaps a 
book about zero-inflated models and generalized linear mixed models 
would be just as fun). Link to the book:  http://www.highstat.com/book4.htm

Again, thanks to all for the help.

Ivar
On 16.11.2011 21:43, Jarrod Hadfield wrote: