Colloquium Series: Giovanni Parmigiani

Our upcoming event for the Statistics Colloquium Series is scheduled for Monday, March 30th from 12:00 – 1:00pm (ET) and will be an in-person presentation at Maxwell-Dworkin 134A/B. Lunch will be provided to guests following the talk. This week's speaker will be faculty member Giovanni Parmigiani from our Statistics department.

Rationality or Hold-out Data Set

Hold-out data sets are data sets used in model evaluation and not in model training, even though they are available at the time of training. Using hold-out data is common practice when developing prediction models, it is often a recommended step before the implementation of machine learning and artificial intelligence systems in practice, and can be a regulatory requirement. On the other hand, from a Bayesian standpoint, there is no formal rational justification that I know of to not look at all the data available when training a model. Can it be rational to hold out data from training to do an out-of-sample evaluation?  In this presentation, I formally define a game that I hope may serve as the normative foundation for a Bayesian model developer's decisions about model construction and evaluation. In a simple setting, I present a conjecture that may provide an early view into the conditions for Bayesian rationality of hold-out sets.

Giovanni Parmigiani, PhD, is a professor of Biostatistics at Harvard TH Chan School of Public Health and Dana-Farber Cancer Institute and Associate Director for Population Sciences at the Dana-Farber/Harvard Cancer Center. He received his undergraduate degree in economics and social sciences at Università L. Bocconi, and a Masters and PhD in statistics from Carnegie Mellon University. He has held faculty positions at Carnegie Mellon, Duke and Johns Hopkins, before joining the faculty at the Harvard TH Chan School of Public Health in 2009 and becoming the Chair of the Department of Biostatistics and Computational Biology at the Dana-Farber Cancer Institute.Dr. Parmigiani is the recipient of numerous awards for his contributions to both science and teaching, including the Advising, Mentoring, and Teaching Recognition Award from the Johns Hopkins School of Public Health Student Assembly. He was named a Fellow of the American Statistical Association in 1999. While completing graduate studies at Carnegie Mellon, he received the Leonard J. Savage Dissertation Prize. His 2009 book on “Decision Theory” received the DeGroot prize. Dr. Parmigiani’s work has been published in the Journal of the American Medical Association, Science, Cancer Research, the Journal of the American Statistical Association, the Journal of Clinical Oncology, and the American Journal of Human Genetics.