MCMC Convergence Criteria

JPalka

Hello!

I am looking for confirmation, or an article I can reference, on some technical details of Bayesian estimation (e.g., t-tests) in JASP:

1. What MCMC sampling method does JASP use as default? Is it Gibbs?
2. What are the default MCMC settings used (i.e., with regards to number of chains, adaptive phase, iterations, thinning, etc.).
3. What is the default convergence criteria? (i.e., how do I know convergence has been achieved)? Is the Gelman-Rubin diagnostic criterion used in some way?

Any help is greatly appreciated! Thank you!

EJ

Hi JPalka,

What is done depends on the test. We try with all of our might to avoid MCMC whenever we can. So we try to use analytical expressions, or use very good approximations to those expressions. When we use a numerical approximation we usually print a "% error" that gives an idea of how reliable the result is. For any particular test, you can go to the output, expand the small triangle, and this gets you to "copy citations" -- pasting this will show you which manuscripts describe the test we use (or which R packages).

E.J.