As much as we love Markov Chain Monte Carlo as a flexible method for estimating all sorts of statistical models even when old-fashioned likelihood-based estimators aren’t available, nobody likes waiting till next Christmas for it to converge, having to throw away most of their massively auto-correlated steps on Boxing Day, or scratching his or her head at 2 a.m. when yet another set of initial values fails.
In recent years there has been a flurry of activity devising better algorithms that explore the parameter space efficiently and give you posterior distributions. One such is Hamiltonian Monte Carlo, and now Andrew Gelman and colleagues have released version 1 of new software that provides us with the first off-the-peg tool to try this technique out for ourselves! It is called Stan and its homepage is here. I wonder if that logo is inspired by Professor Gelman’s journey to work every morning… There is also an R interface called RStan with a useful quick start guide here.
Excuse me, does this go to 116th Street?
Yes, with probability 1 as time tends to infinity.
Now, I haven’t tried this out yet but initial reports say “reliable” and “very fast”. These are words I like to hear!