I have summarized the replies to my queries regarding MCMC tutorials
and code for Splus/R. It appears that this is not a well-developed area,
so that much information about MCMC techniques and their implementation
needs to be read in the journals. However, see responses #1 and #5 below.
I have suppressed the email addresses of the people who have replied to
me, as they did not post their replies to the entire S-news listserv
originally.
If Dave Fournier gives me permission, I will pass his email address on to
the
entire S-news listserv (since he states that he is "looking for new markets"
for some software that he has written.
Regards,
David Alan Paul, Ph.D.
Battelle Memorial Institute - SDAS
paulda@battelle.org
614-424-3176
614-424-4611 (fax)
http://www.battelle.org/statistics
######################################################
ORIGINAL QUESTION:
I have been a WinBUGS user for a while now, and enjoy
its ease of use very much. However, I now need to
code some MCMC routines of my own (presumably in Splus)
for various reasons (missing values in multivariate
distributions, multivariate distributions other than the MVN,
confirmatory analyses for the purposes of publishing some
papers, etc.).
To get up to speed quickly, it would be helpful to
have an available resource that gives some examples of
detailed code along with the statistical setting the code
is designed to address. Does anyone know of a book, .pdf file,
manuscript, or journal article that does this? It would be
especially helpful if there was a document that built up from
simple univariate examples to complex non-conjugate
multivariate ones.
######################################################
1. Dear David,
I regard as a good book also containing pseudocode
for many algorithms: Robert, C.P., Casella, G. (1999)
"Monte Carlo Statistical Methods," Springer, New York.
There are many other good books in general but not
containing code. Finally, when dealing with missing data
see Schafer, J.L. (1997) "Analysis of Incomplete Multivariate
Data," Chapman and Hall, London.
-Susanne
2. Look perhaps at the "Missing Values Module" that comes
with Splus 6.0. There is no printed documentation, but
there is a very nice PDF of the manual. I know it is
available for the Windows version. It has some MCMC routines
already built for missing values. Perhaps there are some
coding ideas you can get from the Insightful crew. The
manual is excellent too.
-Marc
3. Hi Paul,
I've written a java library for doing MCMC called
'Hydra' that was intended to help people do general MCMC
algorithms. You can check it out at
http://www.warnes.net/Hydra/.
As far as using, S-Plus, it is pretty simple to write
Metropolis-Hastings type algorithms. You may find, however,
that performance a huge problem because almost all MCMC
algorithms require explicit looping, for which S-Plus is
poorly designed. You might find it worthwhile to use R for
this, since its looping performance is considerably better.
Insightful has a pre-release library for doing MCMC. I
have a copy but haven't had a chance to look at it yet.
Drop an email to David Smith [dsmith@insightful.com] if
you are interested in looking at it.
-Gregory
4. Hi,
I have made a software package for nonlinear parameter
estimation, AD Model Builder which is widely used in natural
resource management esp. fisheries but not much anywhere else.
I am interested in trying to develop new markets. It may be
ideal for your problem. It is aimed at general nonmlinear
parameter estimation problems where the posterior distribution
has a mode. (ie includes maximum likelihood estimation).
It finds the mode, calculates the hessian, and uses this info
for a multivariate problem it can be hundreds of times faster
than Splus. Once you create a model you like with my software
you can easily turn it into a DLL that you can call from,
say, Splus.
-dave fournier
5. Good question, and I share your interest - from a web search:
http://www.ms.uky.edu/~viele/sta695s01/mcmc/mcmc.html
http://citeseer.nj.nec.com/cowles96markov.html
(Apparently the "CODA" is a set of SPLUS functions, thought
I thought that was part of BUGS) And some code in SPLUS here:
http://tamarama.stanford.edu/mcmc/
Have fun with this!
-Christopher
|