STOR Colloquium: Robin Gong, Harvard University

Bayes is sensitive. Is imprecise probability more sensible?

Bayes is prized as principled and coherent, but its quality of inference is sensitive to prior and model misspecifications. Imprecise probability (IP) allows for the flexible expression of partially deficient probabilistic information. In our quest for minimal-assumption inference, is IP a more promising alternative to Bayes?

In this talk, I showcase the power of IP with an application of the Dempster-Shafer theory of belief functions to the prior-free estimation of infection time in acute HIV-1 patients. I discuss the non-trivial choice of IP updating rules (generalized Bayes, Dempster’s and Geometric) in relation to prominent behavioral features that do not occur in precise model updating, including dilation, contraction and sure loss. The results reinforce a “no free lunch” principle. IP abandons the prior at the expense of abandoning the Bayes rule, introducing a new kind of sensitivity that manifests the trade-off between inference risk and payoff.

No prior knowledge of IP is required of the audience, but a willingness to contribute opinions on whether it is more sensible, in the name of conducting scientifically defensible inference, to trade a choice of priors for a choice of rules in a series of hypothetical and real examples.

Bio: Ruobin (Robin) Gong is a PhD student in statistics at Harvard University, advised by Xiao-Li Meng and Arthur Dempster. Her research focuses on the theoretical, methodological and computational aspects of statistical modeling with imprecise probabilities, random sets, and the Dempster-Shafer theory of belief functions. Robin also has a broad interest in scientific applications in bioinformatics, astrostatistics, neuroscience, and education. She holds a B.Sc. in cognitive psychology from the University of Toronto. Webpage: https://scholar.harvard.edu/rgong
Refreshments will be served at 3:00pm in the 3^rd floor lounge of Hanes Hall