How much evidence do the data give us about one hypothesis versus another? The standard way to measure evidence is still the p-value, despite a myriad of problems surrounding it. In this talk I will provide a gentle introduction to the e-value (wikipedia), a recently popularised notion of evidence which overcomes some of these issues.
- Speaker
- Date
- Thursday 16 Oct 2025, 12:00 - 13:00
- Type
- Seminar
- Room
- ET-14
- Location
- Campus Woudestein
In particular, they effortlessly deal with optional continuation: with e-value based tests and the corresponding anytime valid (AV) confidence intervals, one can always gather additional data, while keeping statistically valid conclusions. Until 2019, publications on e-values were few and far between: the concept did not even have a name. Then, in the course of a few months, four papers by different research groups appeared on arXiv that firmly established them as an important statistical concept.
The first of these was Safe Testing (see below) ; another one was Shafer's testing by betting (Shafer's "bets" are the same as e-values). By now, there are 100s of papers and there have been three international workshops on e-values. Allowing for optional continuation is just one way in which e-values provide more flexibility than p-values – they also allow to set a type of significance/confidence level alpha after seeing the data, which - despite being unconsciously done all the time - is a mortal sin in classical testing. In this talk I will introduce e-values, e-processes and AV confidence intervals, and discuss in detail the relation to Bayesian approaches, which are in some ways related and in others completely different.
Main literature:
- G., De Heide, Koolen. Safe Testing. Journal of the Royal Statistical Society Series B, 2024 (first version appeared on arXiv 2019).
- G. Beyond Neyman-Pearson: e-values enable hypothesis testing with a data-driven alpha. Proceedings National Academy of Sciences of the USA (PNAS), 2024.
- G. The E-Posterior. Phil. Trans. Roy. Soc. London Series A, 2023.
Bio
Peter Grünwald is founder and former head of the machine learning group at CWI in Amsterdam, the Netherlands. Currently a member of CWI Management, he is also full professor of statistics at the mathematical institute of Leiden University. He holds an ERC Advanced Grant (2024) for designing a flexible theory of statistics, based on e-processes. From 2018-2022 Peter served as President of the Association for Computational Learning, the organization running COLT, the world’s prime annual conference on machine learning theory, which he chaired in 2015, having earlier chaired UAI, another major ML conference. He is editor of Foundations and Trends in Machine learning, author of the book (and standard reference) The Minimum Description Length Principle (MIT Press, 2007), and co-recipient of the Van Dantzig prize, the highest Dutch award in statistics and operations research. He has frequently appeared in Dutch national media commenting, e.g., about statistical issues in court cases.
See also
- More information
Do you want to know more about the event? Contact the secretariat Econometrics at eb-secr@ese.eur.nl.
