A0171
Title: A modified algorithm for MCMC in large dimensional models
Authors: Michael Pitt - Kings College London (United Kingdom) [presenting]
Abstract: Large-dimensional models are of interest in classical and Bayesian inference. Bayesian approaches, based upon Markov chain Monte Carlo (MCMC), can run into difficulties when the parameter dimension becomes large. The purpose is to introduce a simple but effective remedy by modifying existing algorithms. This allows MCMC to work effectively in large dimensions. Asymptotic results are provided, which simplify the calculation and optimization of the integrated autocorrelation time (IACT). These asymptotic results and associated central limit theorems (CLTs) are illustrated throughout with examples. It can be seen that the results are robust in that they hold even when the number of parameters is modest (at say 30).
