ABSTRACT

This talk is about an Annals of Statistics paper describing the thesis
of Leif Johnson.  The random-walk Metropolis algorithm (RWMA) as implemented
in the CRAN package mcmc, is a "foolproof" way to simulate distributions of
arbitrary continuous random vectors.  Its only well-understood problem
is that it is not always geometrically ergodic.  Change-of-variable can
usually fix this.  This fix is implemented in the morph.metrop function
in the mcmc package, which is almost as easy to use as the metrop function
that does ordinary RWMA.

This talk is intended to be understandable by all statisticians,
even first-year graduate students.  (Places where something is presented that
the audience doesn't absolutely need to follow are marked!)  We give some
idea how easy it is to use this stuff and also explain what theory users need
to know in order to justify their use (which is simple real analysis, mostly
tail behavior of functions).