Statistics Seminar: "Gibbs posterior distributions"
Speaker: Nicholas Syring, Washington University in Saint Louis
Abstract: Gibbs posterior distributions are the natural analog to Bayesian posteriors when the parameter of interest is defined via a loss function rather than a likelihood. One construction of Gibbs posteriors is given in Bissiri et al. (2016) using decision theory and motivated by a coherence principle. Besides being a principled method of updating beliefs Gibbs posteriors can exhibit favorable frequentist properties under general conditions. We will explore some sufficient conditions for posterior concentration and review examples.
Host: Likai Chen