Bayesian Computation with R Gibbs Sampling推荐.pdfVIP

10 Gibbs Sampling 10.1 Introduction One attractive method for constructing an MCMC algorithm is Gibbs sam- pling, introduced in Chapter 6. To slightly generalize our earlier discussion, suppose that we partition the parameter vector of interest into p components θ = (θ , ..., θ ), where θ may consist of a vector of parameters. The MCMC 1 p k algorithm is implemented by sampling in turn from the p conditional posterior distributions [θ |θ , ..., θ ], ..., [θ |θ , ..., θ ]. 1 2 p p 1 p −1 Under general regularity conditions, draws from this Gibbs sampler will con- verge to the target joint posterior distribution [θ , ..., θ ] of interest. 1 p For a large group of inference problems, Gibbs sampling is automatic in the sense that all conditional posterior distributions are available or easy to simulate using standard probability distributions. There are several attractive aspects of “automatic” Gibbs sampling. First, one can program these simu- lation algorithms with a small amount of R code, especially when one can use vector and matrix representations for parameters and data. Second, un- like the more general Metropolis-Hastings algorithms described in Chapter 6, there are no tuning constants or proposal densities to define. Last, these Gibbs sampling algorithms provide a nice introduction to the use of more sophisticated MCMC algorithms in Bayesian fitting. We illustrate the use of R to write Gibbs sampling algorithms for several popular inferential models. We revisit the robust modeling example of Sec- tion 6.8, where we applied various computational algorithms to summarize the exact posterior distribution. In Section 10.2, we illustrate a