Research Semester Programme Machine Learning TheorySeminar++


Seminar++ meetings consist of a one-hour lecture building up to an open problem, followed by an hour of brainstorming time. The meeting is intended for interested researchers including PhD students. These meetings are freely accessible without registration. Cookies and tea will be provided in the half-time break.

This lecture is part of a series of 8.

Alexander Ly

Alexander Ly
Postdoc at the Centrum Wiskunde & Informatica.

On constructing e-values for statistical practice [slides]

Abstract: The safe anytime-valid inference framework based on e-values allows practitioners to adaptively design their experiments and draw more reliable conclusions compared to conventional p-value-based approaches. The presentation begins with a concise overview of the recently developed general theory of e-values and the various procedures to construct them. In order to distinguish among the different e-values, Gr├╝nwald, de Heide and Koolen (2019) introduced the GROW criterion along with a general procedure specifically designed to construct the optimal e-value according to this criterion. Practical considerations and the context of the statistical problem itself might lead us to deviate from recommending this so-called GROW e-value. We shed light on the choices we made when constructing e-values for fundamental classical inference problems, such as z-tests, t-tests, one-way ANOVAs, and (generalised) linear models. Our objective is to investigate the potential generalisability of these context-specific solutions to a broader range of inference problems in hope to expand the practicality and versatility of safe anytime-valid inference.