JiscMail Logo
Email discussion lists for the UK Education and Research communities

Help for BUGS Archives


BUGS Archives

BUGS Archives


BUGS@JISCMAIL.AC.UK


View:

Message:

[

First

|

Previous

|

Next

|

Last

]

By Topic:

[

First

|

Previous

|

Next

|

Last

]

By Author:

[

First

|

Previous

|

Next

|

Last

]

Font:

Proportional Font

LISTSERV Archives

LISTSERV Archives

BUGS Home

BUGS Home

BUGS  2002

BUGS 2002

Options

Subscribe or Unsubscribe

Subscribe or Unsubscribe

Log In

Log In

Get Password

Get Password

Subject:

MCMC Follow-up question(s)

From:

"Paul, David A" <[log in to unmask]>

Reply-To:

Paul, David A

Date:

Mon, 17 Jun 2002 15:23:58 +0100

Content-Type:

text/plain

Parts/Attachments:

Parts/Attachments

text/plain (175 lines)

After coding our Metropolis-Hastings algorithm and testing it on some
simple linear regression problems, we have validated that it works, at least
for those sorts of things.

My earlier post had to do with a complex spatial model in which the
autoregressive terms that had dnorm( ) priors that had to be
iteratively updated.  The iterative nature of the problem had to do with
the fact that neighboring lattice grid squares define the mean of the
autoregressive term in the current grid square.  I have pasted another
copy of our code at the end of this message.

In this formulation, we essentially had 2184+8 total parameters, subject
to the autoregressive constraints, effectively reducing the dimension of
the problem somewhat.  That is to say, there were  2184 autoregressive
terms,
unique to each grid cell but stochastically dependent on neighboring
autoregressive
terms, and 8 model parameters.  We have 2434 total data points.

Therefore, the dimension of the space that we intend to span using MCMC
methods is of VERY high dimension, even if we do have an estimable model.
Asking MCMC to adequately cover these dimensions in incredible!  With this
many dimensions, there is no possible way to generate enough points from
*any*
candidate function in order to properly span the parameter space, unless
there
is a VERY smart and efficient way of choosing them within the interesting
portions of parameter space.  This leads me to my questions:

(1) Does anyone know of some *really* good references on how to implement
an adaptive phase within the Metropolis-Hastings algorithm?  WinBUGS needs
4500 iterations to properly calibrate its Metropolis algorithm whenever we
fit
this model (I am *assuming* that it is using a Metropolis-Hastings
algorithm...
but, having been wrong before, I may be making an erroneous assumption.)
 Since our primary goal is to validate the answers that WinBUGS is
giving us, having some idea of what WinBUGS is doing when it goes into its
"adaptive" mode of sampling would be helpful.

(2) Our understanding of the "tuning" that is possible during an MCMC has to
do with (a) the size of the steps that the sampler makes when it is drawing
values from the candidate function, (b) the shape of the candidate function
itself, and (c) the decision rule that tells the MCMC algorithm whether or
not
to keep a drawn value.  Is our understanding correct?

(3) Are there alternative MCMC methods to Metropolis-Hastings that sample
more efficiently for very high dimensional problems?  Gibbs won't work
because
it's impossible to write down the full conditionals....

(4) Because our model conforms generally to the SG model presented in Noel
Cressie's textbook ("Statistics for Spatial Data", Chapter 6, page 406,
equation
6.3.9) we believe it to be estimable.  Furthermore, there are more data
points
than parameter dimensions.  My question:  At what point does the
dimensionality
of a Bayesian problem become so great that even if it is theoretically
estimable, the
computational aspects prevent a solution from being found?  I ask this
because it
would be helpful to have some sort of guideline to know whether or not a
Bayesian
model should even be attempted.

Any input would be greatly appreciated!!

 David Paul, Ph.D.
  Battelle Memorial Institute
  505 King Avenue
  Columbus, OH  43201
  614.424.3176

model
{

#First, we define the errors in the SAR model structure for the
#data.  The term 'SAR' is used by Ripley in his Spatial Statistics textbook
#and denotes a 'simultaneous autoregressive' model.  The implementation
#here is slightly different from that of Ripley and Cressie.  The
#autoregressive terms are coded very similarly to that found in Peter
#Congdon's new book on Bayesian Statistical Modelling.  The application
#in question deals with a regular lattice, which makes a SAR model
attractive.

for(i in 1:Num.Total.LatticeGrids)
{

e[i] ~ dnorm(e.bar[i],tau.Error)

#The following line specifies that the mean of the autoregressive "errors"
#of all neighboring sites should be multiplied by rho.  This quanity forms
#the mean of the distribution of the
#autoregressive "error" at the current Lattice Grid Cell.

e.bar[i] <- rho*mean( e.neigh[
(cumulative.neighbor.number[i]+1):cumulative.neighbor.number[i+1] ] )

}

#This line causes WinBUGS to associate the numbering of the spatial
#autoregressive "errors" with the numbering of the Grid Cells in
#the Regular Lattice.

for(m in 1:Num.Total.Neighbors)
{
        e.neigh[m] <- e[neighbor[m]]
}

#Now we define the likelihood of the data.  We are going to impose a SAR
#structure on the Predictive errors.

for(i in 1:Num.Total.Grids)
{

#X[i,1] represents the latitude of an observation and X[i,2] represents the
#longitude of an observation, so mu.Trend is a simple trend surface with a
#spatial autoregressive term added to the end.

mu.Trend[i]<-Beta[1] + Beta[2]*X[i,1] + Beta[3]*X[i,2] + e[i]

#There are two sets of data that we are using to obtain estimates of the
#posterior spatial surface.  The variable "cumulative.observations"
#is an index for the total number of observations within a Grid Cell on
#the Regular Lattice.  It is assumed that one of the data sets is biased
#upwards from the truth, so that mu.Trend needs to be appended with a
constant
#term, a.  The function equals( ) is designed to decide if a particular
#observation belongs to the data set needing the adjustment. It is also
assumed
#that the precision associated with each data set is different, so that tau
#changes as well.

for(j in (cumulative.observations[i]+1):cumulative.observations[i+1])
{

logic[j] <- a*equals(indicator[j],2)
mu.Observation[j] <- logic[j] + mu.Trend[i]

Y.Combined[j] ~ dnorm(mu.Observation[j],tau[indicator[j]])

}

}

#PRIORS, not including the autoregressive terms - they are defined above

tau.Error ~ dgamma(0.0001,0.0001)

for(i in 1:2)
{
tau[i] ~ dgamma(0.0001,0.0001)
}

Beta[1:3] ~ dmnorm(Beta.Mean[] , Beta.R[,])
Beta.R[1:3,1:3] ~ dwish(Beta.Omega[,] , 3)

a ~ dnorm(0.0 , 0.001)

rho ~ dunif(0,1)
}

-------------------------------------------------------------------
This list is for discussion of modelling issues and the BUGS software.
For help with crashes and error messages, first mail [log in to unmask]

To mail the BUGS list, mail to [log in to unmask]
Before mailing, please check the archive at www.jiscmail.ac.uk/lists/bugs.html
Please do not mail attachments to the list.

To leave the BUGS list, send LEAVE BUGS to [log in to unmask]
If this fails, mail [log in to unmask], NOT the whole list

Top of Message | Previous Page | Permalink

JiscMail Tools


RSS Feeds and Sharing


Advanced Options


Archives

March 2024
January 2024
December 2023
August 2023
March 2023
December 2022
November 2022
August 2022
May 2022
March 2022
February 2022
December 2021
November 2021
October 2021
September 2021
July 2021
June 2021
May 2021
April 2021
March 2021
February 2021
January 2021
December 2020
October 2020
September 2020
August 2020
July 2020
June 2020
May 2020
April 2020
March 2020
February 2020
January 2020
December 2019
November 2019
October 2019
September 2019
August 2019
July 2019
June 2019
May 2019
April 2019
March 2019
February 2019
January 2019
November 2018
October 2018
September 2018
August 2018
July 2018
June 2018
May 2018
April 2018
March 2018
February 2018
January 2018
December 2017
November 2017
October 2017
September 2017
August 2017
July 2017
May 2017
April 2017
March 2017
February 2017
January 2017
December 2016
November 2016
October 2016
September 2016
August 2016
July 2016
June 2016
May 2016
April 2016
March 2016
February 2016
January 2016
December 2015
November 2015
October 2015
September 2015
August 2015
July 2015
June 2015
May 2015
April 2015
March 2015
February 2015
January 2015
December 2014
November 2014
October 2014
September 2014
August 2014
July 2014
June 2014
May 2014
April 2014
March 2014
February 2014
January 2014
December 2013
November 2013
October 2013
September 2013
August 2013
July 2013
June 2013
May 2013
April 2013
March 2013
February 2013
January 2013
December 2012
November 2012
October 2012
September 2012
August 2012
July 2012
June 2012
May 2012
April 2012
March 2012
February 2012
January 2012
December 2011
November 2011
October 2011
September 2011
August 2011
July 2011
June 2011
May 2011
April 2011
March 2011
February 2011
January 2011
December 2010
November 2010
October 2010
September 2010
August 2010
July 2010
June 2010
May 2010
April 2010
March 2010
February 2010
January 2010
December 2009
November 2009
October 2009
September 2009
August 2009
July 2009
June 2009
May 2009
April 2009
March 2009
February 2009
January 2009
December 2008
November 2008
October 2008
September 2008
August 2008
July 2008
June 2008
May 2008
April 2008
March 2008
February 2008
January 2008
December 2007
November 2007
October 2007
September 2007
August 2007
July 2007
June 2007
May 2007
April 2007
March 2007
February 2007
January 2007
2006
2005
2004
2003
2002
2001
2000
1999
1998


JiscMail is a Jisc service.

View our service policies at https://www.jiscmail.ac.uk/policyandsecurity/ and Jisc's privacy policy at https://www.jisc.ac.uk/website/privacy-notice

For help and support help@jisc.ac.uk

Secured by F-Secure Anti-Virus CataList Email List Search Powered by the LISTSERV Email List Manager