Statistics Seminars: Problems with predictive distributions
22 September 2016 14:00 in CM219
This talk discusses difficulties with the use of the Bayesian posterior predictive distribution, or frequentist prediction intervals, for probability statements about new observations.
A new Bayesian approach to the prediction problem is described, and its properties illustrated with the prediction of new Bernoulli and normal observations.
The current posterior predictive distribution appears as the posterior mean of the new observation distribution, while the new approach give the full posterior distribution of the new observation.
This is joint work with Charles Liu (Cytel Inc., Boston).
Contact firstname.lastname@example.org for more information