Application of Model Selection for the Weight of BC Pink Salmon Using Bayes Factors Estimated by Thermodynamic Integration

• Activity Number: 536
• Type: Contributed
• Date/Time: Wednesday, August 4, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Jinhee Lim* and David A. Campbell+
• Companies: Simon Fraser University and Simon Fraser University
• Keywords: model selection ; bayes factors ; thermodynamic integration ; parallel tempering

Abstract

The weight of Pink Salmon off the BC coast is showing a declining trend but the shape of the trend is unclear. In this work we compare linear, quadratic and changepoint models using Bayes Factors (BF), the ratio of marginal likelihoods for the models, obtained by integrating over the model parameter space. MCMC approximates a large sample from the posterior distribution, however these samples produce unstable BF estimates. Therefore we use parallel tempering in which parallel MCMC chains are run, each approximating the posterior distribution through a sequence of power-posteriors. Parallel tempering improves Markov Chain convergence and reduces autocorrelation by permitting large moves across the parameter space. Furthermore, the samples obtained from the parallel chains are combined through Thermodynamic Integration to produce stable BF estimates.