[ADMB Users] MLE and SE are satisfactory, but Bayesian samples are NOT.

[Ian Taylor]

Hi Saang-Yoon,
I think some models are just poorly behaved and no improved algorithm will
solve their convergence problems. You might look at reparameterizing to
avoid some of the highest correlations, or fixing some of the parameters
that are the most autocorrelated.

If you want to try additional MCMC options in ADMB, a few to consider are
-mcgrope to fatten the tails of the jump function, -mcrb to reduce
correlations in the jump, and -hybrid (along with inputs for -hyeps and
-hynsteps to use an alternative algorithm). You can read a little
discussion about these options here:
https://groups.google.com/forum/#!topic/admb-users/QZhF0CaJLnw. -mcgrope
and -mcrb are also described (a little) in the user manual.
-Ian

On Fri, May 4, 2012 at 10:08 AM, Saang-Yoon <shyunuw at gmail.com> wrote:

> Hi, all.
> From my TPL and DAT files, the estimates (i.e., MLE) and its SE are
> reasonable.  However, Bayesian samples from the same codes are not
> satisfactory in terms of diagnosis.   I ran large MCMC runs (e.g., >
> 20 million runs), thinning samples at an enough interval (e.g., every
> 5000 runs).  Despite that, the posterior samples are not
> satisfactory.   I wonder what kind of additional commands I should
> try.   I used only the following commands: (i) -mcmc; (ii) -mcsave,
> and (iii) -mceval.  Thank you.
> Saang-Yoon
> _______________________________________________
> Users mailing list
> Users at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/users/attachments/20120504/671133e2/attachment.html>