How to determine the length of the required burn-in until convergence in MCMCglmm package or another package
Hi Euis, In an earlier post you said you were fitting zero-inflated models (zipoisson)? Is it possible you a) forgot to fix the non-identifiable residual variance for the zero-inflation process at some value (e.g. 1)? b) that the data are not zero-inflated but just over-dispersed so the zero-inflation parameters are heading off towards -Infinity? Cheers, Jarrod
On 27/03/2017 18:40, Ben Bolker wrote:
[please keep r-sig-mixed in Cc:]
To repeat what I said below, the general brute-force strategy would be
N=2 (or 10 or something)
run MCMCglmm with some reasonably optimistic default settings such that
the final sample size (nitt-nburn)/thin is 1000
while (convergence not satisfactory)
nitt = N*nitt
thin = N* thin
re-try MCMCglmm
This brute force strategy will fail if something is wrong with your
model (e.g. underdetermined). Strengthening priors may help. Other
than that, without more information, we really can't help you more.
On 17-03-27 11:19 AM, Euis Aqmaliyah wrote:
Thank you for your reply.
I'm sorry if my subject mail or my question is not clear.
Actually, i have understood that diagnostic convergence can use
potential scale reduction, potential scale reduction factor, or use
trace plot or another graphic (i use potential scale reduction and
trace plot). But, in MCMCglmm Tutorial that i read, if convergence
hasn't reached, we can increase length of chain, or length of burn-in,
or thinning interval. So, it is that i ask.
Oh yes, i also have apply raftery.diag(). The output show sample size
that i need. So, i combine chain length, burn-in length, and thinning
interval so that yield sample size like in that output. But, it is still
doesn't convergence.
Regards
Pada tanggal 27 Mar 2017 21.16, "Ben Bolker" <bbolker at gmail.com
<mailto:bbolker at gmail.com>> menulis:
We would probably need more information to help you.
Some quick thoughts:
- MCMCglmm usually burns in very quickly. I would guess that either
(1) your problem/data are really pathological; (2) you're confusing
"burn-in" with "mixing"; if your chain reaches the stationary state
quickly but samples it slowly, then you're having a burn-in rather
than a mixing problem. In general PRSF is meant to diagnose
convergence, not just burn-in. (Although now that I read your
question, it sounds like it's only the title that's specific to
burn-in ...)
- I think what most people do is brute-force (increase length of
chain, increasing thinning at the same time so that the number of
samples remains constant, until traceplots look OK/PRSF looks OK).
- setting more informative priors may be helpful/necessary
- the coda package has other diagnostics, in particular the
Raftery-Lewis (raftery.diag()), which is supposed to estimate the
chain length required for convergence. You should be able to apply it
to the components of an MCMCglmm fit ($Sol, $VCV, etc.), which are
mcmc objects
On Mon, Mar 27, 2017 at 5:16 AM, Euis Aqmaliyah
<aqmalsaepul at gmail.com <mailto:aqmalsaepul at gmail.com>> wrote:
> Hi,
>
> I stil try fit linear mixed model. I use Potencial Scale Reduction
(PSR) to
> check convergence. But, it still dosn't convergence. Is there any
function
> that can i use to determine length of chains, length of burn-in, or
> thinning interval?
>
> Thank you.
>
> [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org
<mailto:R-sig-mixed-models at r-project.org> mailing list
_______________________________________________ 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.