Assessing and Improving Convergence of the Markov Chain
Emmanuel M. E. H. Lesaffre, Andrew B Lawson · 2012
The generality of the Markov chain Monte Carlo (MCMC) sampling procedures comes with a cost: the sampled values are not immediately taken from the posterior distribution and assessing convergence of the Markov chain is more difficult than for the likelihood case. When convergence is slow, one can either let the MCMC procedure run longer or one could try to accelerate the algorithm. This chapter focuses on a variety of graphical and formal diagnostics to assess convergence of the chain. It explores the general purpose techniques to speed up the MCMC sampling procedure. The chapter looks at the Bayesian implementation of data augmentation. It reviews some simple tricks to speed up convergence. Controlled Vocabulary Terms convergence of random variables; Markov chain monte carlo