BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Transdimensional sampling algorithms for Bayesian variable selecti
 on in classification problems with many more variables than observations -
  Jim Griffin\, University of Kent
DTSTART:20081014T133000Z
DTEND:20081014T143000Z
UID:TALK10785@talks.cam.ac.uk
CONTACT:Nikolaos Demiris
DESCRIPTION:One problem in microarray analysis is the identification of a 
 (small) collection of genes that discriminate between two classes (such as
  diseased and not diseased). Variable selection in probit regression is on
 e statistical approach to this problem. We perform a Bayesian analysis usi
 ng Markov chain Monte Carlo methods with standard data augmentation techni
 ques. The main issue is how to efficiently move between different models (
 combination of variables) using Reversible Jump MCMC. In this talk\, I wil
 l discuss the construction and efficiency of different proposals. High bet
 ween model acceptance rates are one distinctive feature of many data sets 
 that we consider and this suggests using ideas motivated by Metropolis-Has
 tings random walk samplers to build proposals that have (near) optimal mix
 ing. The methods will be illustrated on several gene expression data sets.
  This is joint work with Demetris Lamnisos and Mark Steel. A technical rep
 ort is available at\n"http://www.kent.ac.uk/ims/personal/jeg28/trans.pdf":
 http://www.kent.ac.uk/ims/personal/jeg28/trans.pdf\n
LOCATION:Large Seminar Room\, 1st Floor\, Institute of Public Health\, Uni
 versity Forvie Site\, Robinson Way\, Cambridge
END:VEVENT
END:VCALENDAR