[R-sig-ME] MCMC fitting in glmmADMB

[- -]

I received some feedback that made me aware that I didn’t explain my problem
correctly in my original request due to translation issues. Sorry for that. It
is the RAM and not the hard drive space that is reaching its capacity. After
running the MCMC chain for 24 hours (approximately 2000 iterations for the model
of my own data) the allocated memory is already about 6 GB.

Therefore I am wondering if there is a way to reduce the used memory during the
calculation of the model, as I don’t need all the information in the output at
the end. Or can I divide the MCMC chain in smaller parts and define the end
values from the part before as starting values for each subsequent part?

Best wishes,
Maren



> maren <maren.rebke at avitec-research.de> hat am 20. Oktober 2014 um 10:58
> geschrieben:
>
>
> Thank you very much for your suggestion, Ben. Saving only a thinned
> sequence will of course reduce the size of the stored data file. I was
> using that option already in the analysis of my own data, but maybe I
> just have to choose larger intervals.
>
> From your reply I assume that it is not easily possible to store only
> the samples of the estimates for certain parameters (i.e. don’t save
> u.01-u.27 in the owl example), define a burn-in period or play around
> with the jump size. Or were my questions not clear enough?
>
> Best wishes,
>
> Maren
>
>
> Am 18.10.2014 17:23, schrieb Ben Bolker:
> > You can use mcmcControl(mcsave=...), as illustrated below.
> >
> > library("glmmADMB")
> >
> > om <- glmmadmb(SiblingNegotiation~FoodTreatment*SexParent+
> > (1|Nest)+offset(log(BroodSize)),
> > zeroInflation=TRUE,family="nbinom",data=Owls,
> > mcmc=TRUE,
> > mcmc.opts=mcmcControl(mcmc=200))
> > nrow(om$mcmc) ## 20
> >
> > om2 <- glmmadmb(SiblingNegotiation~FoodTreatment*SexParent+
> > (1|Nest)+offset(log(BroodSize)),
> > zeroInflation=TRUE,family="nbinom",data=Owls,
> > mcmc=TRUE,
> > mcmc.opts=mcmcControl(mcmc=200,mcsave=20))
> > nrow(om2$mcmc) ## 10
> >
> >
> > On Thu, Oct 16, 2014 at 7:18 AM, maren <maren.rebke at avitec-research.de
> > <mailto:maren.rebke at avitec-research.de>> wrote:
> >
> > Hi,
> >
> > I fit a zero-inflated Poisson model with random effects using the
> > package glmmADMB, which worked perfectly well. Now I am trying to get
> > credible intervals by running a Markov chain using mcmc=TRUE,
> > which also
> > works fine in general.
> >
> > The problem is, that I have many parameters as well as several random
> > effects in my model and it seems that I need to run long chains to get
> > proper estimates. Therefore the automatically stored file eventually
> > gets very big and my computer cannot handle it anymore. Therefore I
> > would like to store only the samples of the estimates for the fixed
> > effects (only beta) and not the rest. Is that possible somehow?
> >
> > I am not sure, but would it help to specify parameters via mcmcpars? I
> > tried to include mcmcpars in the owl example in section 2.2 from the
> > vignette of the package
> > (http://glmmadmb.r-forge.r-project.org/glmmADMB.pdf):
> >
> > fit_zinbinom1_bs_mcmc <-
> > glmmadmb(NCalls~(FoodTreatment+ArrivalTime)*SexParent+BroodSize+(1|Nest),data=Owls,zeroInflation=TRUE,family="nbinom1",mcmc=TRUE,mcmc.opts=mcmcControl(mcmc=10,mcmcpars="beta"))
> >
> > But unfortunately, I get an error message stating "unused argument
> > (mcmcpars="beta")". As I wasn't sure if I have to state the fixed
> > effects by using "beta" or the names of the parameters directly, I
> > also
> > tried including mcmcpars="BroodSize" but got the same error.
> >
> > Is it not possible to define mcmcpars in glmmADMB? Is the
> > definition of
> > mcmcpars at all what I need and if so, how do I do it correctly?
> >
> > Otherwise, is it possible to state that only the samples after a
> > certain
> > burnin period should be saved? Or can I play around with the jump
> > sizes
> > to reach faster convergence? As far as I understood those are rescaled
> > depending on the acceptance rate at the moment. The automatic
> > rescaling
> > can be switched off by stating mcnoscale=TRUE, which is working. But I
> > am not sure how I can then adjust the jump size and what the
> > default is.
> >
> > Thank you very much for taking the time to read this long email.
> >
> > Best wishes,
> >
> > Maren Rebke
> >
> > [[alternative HTML version deleted]]
> >
> > _______________________________________________
> > R-sig-mixed-models at r-project.org
> > <mailto:R-sig-mixed-models at r-project.org> mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> >
> >
>
>
> [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
	[[alternative HTML version deleted]]