John M. Castelloe and Dale L. Zimmerman
Technical Report #313, February 2002
Department of Statistics and Actuarial Science, University of Iowa
Abstract: A new method for assessing convergence of a reversible jump MCMC sampler is presented, along with examples demonstrating both its sensitivity and specificity. The method is an extension of the popular technique of
Gelman and Rubin (1992) in two ways. First, the one-way ANOVA paradigm applied to chains is extended to a two-way paradigm by treating the sampler's parameter space as a factor, an idea similar to one explored by
Brooks and Giudici (2000) but addressed more appropriately as an unbalanced ANOVA. Second, the technique is extended from univariate to multivariate following the suggestions of
Brooks and Gelman (1998): several parameters can be monitored simultaneously and conservatively via worst-case scalar functions.
Keywords: analysis of variance, convergence, convergence diagnostic, Markov chain Monte Carlo, mixing
Full manuscript: PDF, postscript, zipped postscript