Posted on Thursday, October 6, 2022
The New England Journal of Statistics in Data Science (NEJSDS)
invites submissions to a special issue of “Game-theoretic Statistics and Safe Anytime-Valid Inference”.
This is an exciting new subfield that focuses on developing statistical
inference procedures (like hypothesis tests and confidence sets) for
experimental settings where the data may be continuously monitored as it
arrives, the experiment may be adaptively stopped or extended for any reason,
and yet the procedures must yield statistically valid inference at arbitrary
data-dependent stopping times. This area has implications for automated
experimentation in the IT industry, reproducibility of individual scientific
studies as well as meta-analyses.
“E-values” (or their sequential analog, “e-processes”) have emerged as a central
concept in this area. They provide an alternative to p-values and other
traditional notions of evidence for testing, with “confidence sequences” being
their natural twin concept for estimation. These are intimately tied to
nonnegative “test (super)martingales”, which can be viewed as wealth processes
in certain games that the statistician plays against nature. In this world view,
estimation and testing reduce to game and strategy design. This game-theoretic
approach yields a philosophical and methodological middle ground between
classical Bayesian and frequentist approaches, but also gives rise to new
data-adaptive procedures with interesting properties - not just in a sequential
setting but also, for example, in multiple testing problems.
For the unfamiliar, an overview paper from July’22 has recently been posted on arXiv: https://arxiv.org/abs/2210.01948. Also, several talk slides, tutorials and paper links can be found at https://www.stat.cmu.edu/~aramdas/SAVI/savi20.html, from a week-long workshop on the topic that was organized by the guest editors in June’22.
The tentative topics include (but are not limited to):
- Optimality or admissibility of e-values or e-processes
- Sequential tests of power one and e-processes for nonparametric problems, or challenging parametric problems (eg: high-dimensional settings, nuisance parameters, or complex likelihoods)
- Confidence sequences for functionals of distributions
- Causal inference (observational and randomized settings)
- Meta-analysis (fixed/random effects, live, heterogeneity)
- Decision theoretic implications of e-values
- Software, packages, empirical studies
- Connections to information theory (universal coding and minimum description length), online learning, game-theoretic probability, etc.
- Connections between game-theoretic statistics and frequentist, Bayesian, fiducial or other paradigms
- Multiple testing and post-selection inference with e-values or confidence sequences
- Multi-armed bandits, sequential experimentation (A/B/n testing), etc.
- Applications to new scientific or technological problems
- Drawbacks of SAVI approaches and an elaboration of open problems
- Extensions to other statistical problems (model selection, change detection, forensic statistics, and so on)
Submission Guidelines: Prospective authors submit their papers via the NEJSDS website （https://journal.nestat.org/submission). Using the NEJSDS format is preferable, but the first submission can also be in a standard arXiv article format (single column, single spaced, 1in margins). The main paper should be at most 25pgs in length (supplement is unlimited). Everything can be submitted as a single PDF file if desired.
Reviewing: All submissions will follow the same reviewing process according to the schedule below. We will follow the standard journal review process, soliciting at least two expert reviewers per paper. Each author of a submitted paper may be requested to review a paper that they are not in conflict with. This will ensure timely reviews from people who are familiar with the field. The main reviewing criteria would be quality, novelty, relevance to the topic, and whether the paper can be accepted with relatively minor modifications (papers that are good in principle but require major revisions may be rejected due to the strict timeline).
Key Dates:Papers should be submitted according to the following schedule:
Manuscript Due: Anytime before March 1, 2023
Initial Decision Notification: Jul 1, 2023
Rebuttal Deadline: Oct 1, 2023
Final Decision Notification: Nov 1, 2023
Proofs to Publisher: Dec 1, 2023
Expected Publication: March 1, 2024
Extensions may be granted in extenuating circumstances by emailing the guest editors.
Editors-in-Chief: Ming-Hui Chen, University of Connecticut
and Min-ge Xie, Rutgers University
Lead guest editors: Aaditya Ramdas, Carnegie Mellon University
and Peter Grünwald, CWI Amsterdam