Bayesian multimodel inference by RJMCMC: A Gibbs sampling approach

摘要

Purpose: To describe a version of reversible-jump Markov chain Monte Carlo (RJMCMC) that intuitively represents the process as Gibbs sampling with alternating updates of a categorical variable M (for model) and a 'palette' of parameters xp, from which any of the model-specific parameters can be calculated. Summary: Bayesian multimodel inference treats a set of candidate models as the sample space of a latent categorical random variable, sampled once; the data at hand are modeled as having been generated according to the sampled model. Model selection and model averaging are based on the posterior probabilities for the model set. RJMCMC extends ordinary MCMC methods to this meta-model. The representation in the article makes it clear as to how model-specific Monte Carlo outputs (analytical and numerical) can be post-processed to compute model weights or Bayes factors. Illustrative examples are given. (19 refs.)