On Jul 4, 2008, at 7:51 AM, Jeremy Kedziora wrote:
> Adam,
>
> Based on our conversation about my last project, it's probably an
> identification
> problem; you might be able to figure out which are the problematic
> parameters by
> plotting each time series of draws from the posterior distribution.
> If there is
> an identification problem, then what you'll probably observe are two
> of the time
> series moving exactly opposite to one another over the course of the
> algorithm.
In a perfect world we wouldn't be guessing as to the possibility of an
identification problem, but...
So to follow up on the previous suggestion: pairwise trace plots or
scatterplots can sometimes help.
Also, my experience is that you can sometimes need long runs of the
MCMC algorithm (or several long runs, say from different starting
values) to get a handle on subtle identification issues. It can be
difficult because its not always obvious that (even if a particularly
set of parameters is unidentified) the (marginal) posterior you're
getting back is not that much different from the prior, etc (i.e., you
sometimes need a very long run even to assess that). Put differently:
the problem with slow mixing Markov chains is that they are slow mixing.
Finally, there is also the KISS strategy. What happens when you knock
certain parameters out of the model or set them to constants?
Sometimes pinning down one part of a complicated model can help you
get a handle on what is working, and it isn't. Again, long runs may
be needed even for this to be helpful. Going the other way can help
too: i.e., try running your sampler on a model or sub-model you *know*
to be unidentified, so you get a feel for what "broken" looks like.
Its tough to say anything more specific without knowing more about
your model etc (not that I'm volunteering... ;-)
-- Simon
>
>
> Jeremy
>
>> Hi all,
>>
>>
>>
>> I have been running some Bayesian HLM's with multiple latent
>> variables
>> lately and seem to have hit a wall. Virtually all of the time
>> series of the
>> Markov chains look well-behaved and seem to converge after a
>> reasonable
>> burn-in, however one vector of coefficients wobbles and bobbles all
>> over the
>> parameter space, regardless of how many iterations I allow. I have
>> checked
>> my code over and over again, but it looks right. I tried different
>> starting
>> values. All of this to no avail. Has anyone had this kind of
>> problem
>> before and, if so, any remedies?
>>
>>
>>
>> Thanks,
>>
>>
>>
>> Adam
>>
>>
>>
>> P.S. I didn't use WinBUGS, so I can't post the code easily. I wrote
>> the
>> sampler myself in C++.
>>
>>
>>
>> Adam Ramey
>>
>> Department of Political Science
>>
>> Harkness Hall 338
>>
>> University of Rochester
>> Rochester NY 14627-0146
>> Phone: 585-273-1678
>>
>> Fax: 585-271-1616
>>
>> E-mail: <mailto:[log in to unmask]> [log in to unmask]
>>
>>
>>
>>
>> **********************************************************
>> Political Methodology E-Mail List
>> Editors: Melanie Goodrich, <[log in to unmask]>
>> Delia Bailey, <[log in to unmask]>
>> **********************************************************
>> Send messages to [log in to unmask]
>> To join the list, cancel your subscription, or modify
>> your subscription settings visit:
>>
>> http://polmeth.wustl.edu/polmeth.php
>>
>> **********************************************************
>
> Jeremy Kedziora
>
> Department of Political Science
>
> Harkness Hall 315a
> University of Rochester
> Rochester NY 14627-0146
> Phone: 585-273-1710
> Fax: 585-271-1616
>
> "iacta alea est."
>
> **********************************************************
> Political Methodology E-Mail List
> Editors: Melanie Goodrich, <[log in to unmask]>
> Delia Bailey, <[log in to unmask]>
> **********************************************************
> Send messages to [log in to unmask]
> To join the list, cancel your subscription, or modify
> your subscription settings visit:
>
> http://polmeth.wustl.edu/polmeth.php
>
> **********************************************************
>
Professor Simon Jackman,
Depts of Political Science & (by courtesy) Statistics,
Stanford University, Stanford, CA 94305-6044, USA.
http://jackman.stanford.edu
Director, Political Science Computational Lab. http://pscl.stanford.edu
Director, Methods of Analysis Program in the Social Sciences, http://mapss.stanford.edu
cell: +1 (650) 387 3019 fax: +1 (650) 724-9095
**********************************************************
Political Methodology E-Mail List
Editors: Melanie Goodrich, <[log in to unmask]>
Delia Bailey, <[log in to unmask]>
**********************************************************
Send messages to [log in to unmask]
To join the list, cancel your subscription, or modify
your subscription settings visit:
http://polmeth.wustl.edu/polmeth.php
**********************************************************
|