ABSTRACT
This chapter has several rather practical purposes related to applied MCMC work: to
introduce formal convergence diagnostic techniques, to provide tools to improve mixing
and coverage, and to note a number of challenges that are routinely encountered. This
is a stark contrast to the last chapter, which was concerned with theoretical properties
of Markov chains and Markov chain Monte Carlo. Since applied work is generally done
computationally through the convenient programs BUGS (in any of the versions) and JAGS,
or by writing source code in R, C, or even Fortran, practical considerations are important
to getting reliable inferences from chain values. Most of the concern centers on assessing
convergence, but the speed of the sampler, and its ability to thoroughly explore the sample
space are also important issues to be concerned with. This chapter also describes the two
very similar R packages for analyzing MCMC output and evaluating convergence: BOA and
CODA. These are merely convenient functional routines, and users will often want to go
beyond their capabilities, particularly in graphics. However, the purpose here is mainly to
understand the key workings of these tools rather than to function as a detailed description
of the syntax of software. See Albert (2009) or Ntzoufras (2009) for recent book-length
works with very detailed R and BUGS code description.