Dept. of Statistics, Harvard University
Large-scale Bayesian inference for GWAS with coupled Markov chain Monte Carlo
This talk will cover Markov chain Monte Carlo (MCMC) methods with “couplings” as a tool in large-scale Bayesian computation. Determining the burn-in period of MCMC is a recurring concern for practitioners of Bayesian inference in settings such as GWAS. We explain how couplings can diagnose convergence and give practical burn-in guidance for MCMC algorithms. The first half of the talk will focus on a simpler, pedagogical example. The second half will introduce coupled MCMC methods for high-dimensional Bayesian regression. We will conclude with a surprise: even in GWAS settings, with dimension ~100000, our MCMC algorithms can converge in only ~750 steps.