Jun 9, 2016 · 5:30 PM
This location is shown only to members
This session will focus on common problems when estimating Bayesian models using Markov chain Monte Carlo (MCMC) algorithms. We'll show examples of how MCMC can go wrong and how to proceed when it does. We'll fit a few models and use various visualizations of the Markov chains to diagnose problems with sampling. Once we've identified specific problems we'll look at different ways to proceed, in some cases solving the problem directly, in other cases finding a reparameterization that avoids the problem entirely.
We'll use Stan to fit the models, but most of the concepts also apply more generally. To follow along on your own laptop (optional) you'll need to have R installed, along with the 'rstan' and 'shinystan' R packages.
The presenter, Jonah Gabry, is a member of the Stan core development team, and a researcher in statistics at Columbia University working with Andrew Gelman. He is affiliated with the Columbia Applied Statistics Center, the Columbia Population Research Center, and the Institute for Social and Economic Research and Policy. Jonah is a co-author of the rstan and rstanarm interfaces to Stan, as well as the loo and shinystan R packages for model evaluation.