E-values are currently a hot topic in mathematical statistics. They provide a notion of “evidence”, and as such an alternative for the p-value and the Bayes factor. Compared to these “old” methods, e-values are better suited for 21st century data, allowing for much more user flexibility: they are the basic building blocks of anytime-valid hypothesis tests and confidence sets : you may stop and continue gathering data at any time you like, while retaining error probability guarantees) Also, they allow for  post-hoc, data-dependent changes of the significance/confidence level. They underly some theoretical break-throughs in statistics, serving, for example, as a unifying principle for multiple testing. The e-value was introduced in modern form only in 2019 (among others by us: [GHK24], first version appeared on arXiv in 2019) - by now there are 100s of papers about it, by top statistics groups around the world (Berkeley, Stanford, CMU, Chicago, etc.), and it has been the subject of several international workshops.

At its core,  e-value theory is based on viewing statistical inference in terms of a betting game where you "bet against a null hypothesis". If you make a lot of money in this bet, you have evidence that the hypothesis is false.  This puts it in touch with the very foundations of probability theory - in the course will also discuss replacing the standard measure-theoretic foundations of probability by game-theoretic ones - the idea of 'probability as limiting frequency' gets replaced by 'probability as fair price', which puts a lot of existing results in a different light, and avoids hard to interpret 'almost-sure' statements.

The course will introduce e-values and their main interpretations, the main types of e-variables (log-optimal e-variables, mixture-method based, conditional), the most important examples (e-value based z-test, t-test, e-value for exponential family nulls, for group invariant nulls, for nonparametric nulls and so on); their connection to nonnegative discrete-time super-martingales and concentration inequalities such as Hoeffding and Berstein; their main applications (flexible tests, confidence intervals) and further fundamental results (construction of e-values by information projection; e-values as a unifying principle in multiple testing) . We will also review classical approaches to testing and confidence and discuss what each of them can and cannot achieve. These include Fisherian testing, Neyman-Pearson testing, Jeffreys-Bayesian (all from the 1930s), sequential testing (1940s) and pure likelihood-based (1960s) approaches.

Prerequisites: Prerequisites are basic probability theory (including laws of large numbers, central limit theorem) and statistics (maximum likelihood, least squares, hypothesis testing). Knowledge of machine learning or more advanced probability/statistics may be useful but is not essential. In particular, all stochastic process/martingale/game theory that is needed will be developed from scratch. Basic programming experience (Python or R or such) is also needed for the practical sessions. There are no dependencies on other mastermath courses.

References

[GKH24]  Safe Testing Peter Grünwald, Rianne de Heide and Wouter Koolen. Journal of the Royal Statistical Society Series B: Statistical Methodology, Volume 86, Issue 5, November 2024, Pages 1091–1128. Discussion paper. doi.org/10.1093/jrsssb/qkae011. The discussion meeting took place on January 24, 2024 (video).