Anchored multiple testing: a transparent use of e-closure to improve FDR procedures
summary
The gist
The recent e-closure method can recover every procedure that controls FDR (and other expectation losses).
In short
The e-closure method recovers procedures controlling False Discovery Rate (FDR) but yields self-referential results that offer no improvement guidance. This work introduces an 'anchoring' technique that exploits existing proof looseness to enlarge a baseline procedure's local e-value. This new approach transparently retains all baseline discoveries while maintaining FDR control under the same conditions.
Key concepts
- e-closure
- A method used to recover every procedure that controls FDR (False Discovery Rate) and other expectation losses from a given multiple testing procedure. The recovered set is self-referential, meaning it doesn't easily show how to make the original procedure better.
- anchoring principle
- A generic technique that uses the looseness in existing FDR proofs. It defines a local e-value based on a baseline rejection set and an intersection e-value formula. This allows every measurable certified superset of the original set to control FDR at alpha and contain all its discoveries.
- local e-value
- A specific type of e-value derived from the anchoring principle. It is calculated using a baseline rejection set and a formula involving FDP and FS. Crucially, any certified superset obtained via e-closure using this local value satisfies FDR control at alpha.
- dominance (e.g., Anchored-BH dominates BH)
- A relationship indicating that one new procedure (like Anchored-BH) performs better than another established procedure (like BH) under specific conditions, such as PRDS. This shows the new method is superior or at least equally good for certain testing scenarios.
Terminology used across episodes
This episode discusses
- Anchored multiple testing: a transparent use of e-closure to improve FDR procedures · Paper Radio
- Controlling the false discovery rate via knockoffs
- A frequentist local false discovery rate
- Adaptive procedures for boundary FDR control
- Generalized Boundary FDR Control under Arbitrary Dependence: An Approach on Closure Principle
- QuTE: decentralized multiple testing on sensor networks with false discovery rate control
- Conditional calibration for false discovery rate control under dependence
- Controlling the false discovery rate under a non-parametric graphical dependence model
- Dependence-Aware False Discovery Rate Control in Two-Sided Gaussian Mean Testing
- A Uniform Improvement of the Benjamini-Hochberg Procedure via e-Closure
- Royen's proof of the Gaussian correlation inequality
- Maximum likelihood estimation in Gaussian models under total positivity
- Boosting e-BH via conditional calibration · Paper Radio
- Power of Ordered Hypothesis Testing
- Accumulation tests for FDR control in ordered hypothesis testing
- Improving knockoffs with conditional calibration
- A simple proof of the Gaussian correlation conjecture extended to multivariate gamma distributions
- The FDR-Linking Theorem
- Admissibility and Complete Classes for False Discovery Rate Control with E-values
- Bringing Closure to False Discovery Rate Control: A General Principle for Multiple Testing
The paper
Anchored multiple testing: a transparent use of e-closure to improve FDR procedures · Read on arXiv
Stanford University
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Anchored multiple testing".
Jane: Here's the gist of the paper. The recent e-closure method can recover every procedure that controls FDR (and other expectation losses).
Tom: But the recovered e-collection is “self-referential” and gives no insight on how to improve the procedure (if improvable), and some recent improvements have been somewhat opaque.
Jane: We introduce an elementary new technique called anchoring that exploits looseness in existing FDR proofs to enlarge a baseline multiple-testing procedure’s self-referential local e-value.
Tom: The resulting e-closure thus transparently retains every baseline discovery (and usually adding more) and controlling the false discovery rate under the same conditions as the baseline.
Paper summary: Tom: So, we're looking at this paper titled "Anchored multiple testing: a transparent use of e-closure to improve FDR procedures." Basically, the researchers are taking this existing e-closure method—which they say can recover every procedure that controls false discovery rate—and they found it’s kind of self-referential.
Jane: Self-referential how does that affect things? I mean, if you're trying to improve a testing procedure, and the method you use just tells you what you already rejected, how can it tell you what to try next?
Lu: The paper introduces this new technique called anchoring. It exploits some looseness in existing FDR proofs to make the self-referential local e-value bigger for a baseline multiple testing procedure.
Meng: Okay, so they're trying to enlarge the local e-value, right? But what does that actually do for us when we're thinking about real deployments? Does it just make the math look nicer, or is there something tangible happening?
Lalam: The main claim is that this new e-closure method transparently keeps every baseline discovery, and usually adds more discoveries while still controlling the false discovery rate under the same conditions as the original procedure.
Tom: So, it’s not just recovering old methods; it’s giving us a way to get better results without having to completely rebuild our entire testing pipeline from scratch. The paper shows this anchoring principle works across a lot of different procedures.
Jane: That's what's interesting because they test it against several things, like Anchored-BH dominating the Benjamini-Hochberg procedure under PRDS, but being incomparable to Goeman’s recent closed-BH method.
Lu: And they also show it works for the Benjamini-Yekutieli procedure under arbitrary dependence, and that Anchored-BY is incomparable to closed BY.
Meng: In terms of deployment, that comparison between their anchored version and the existing canonical versions—like closed BH—is important because it shows where we actually get an improvement in practice.
Lalam: They also tested it on two-sided Gaussian p-values under appropriate covariance conditions, where Anchored-2BH dominates running BH twice at half the level on two one-sided p-values <ref:2610.11279#pg2>.
Tom: It sounds like they're showing that this anchoring isn't just some theoretical curiosity; it’s actually producing concrete dominance in several key scenarios. The computational complexity for these anchored methods is generally in sorting or quadratic time, which is pretty reasonable for large datasets.
Jane: So, the paper suggests that this technique is broadly applicable and doesn't just work for one specific algorithm; it seems to exploit a general property of FDR proofs related to how much slack there is in the budget.
Lu: The anchoring principle itself is generic: they define a local e-value as E A S equals FDP S alpha equals A intersection S alpha A or one and they show that any measurable certified superset of A controls FDR at alpha and contains all of A’s discoveries <ref:2610.11279#pg2>.
Meng: That definition of the local e-value is key because it shows how you can define a set that guarantees control over the false discovery rate at alpha while making sure you don't miss anything from your initial set A.
Lalam: And they use this to show that in certain regimes, like when there are more non-nulls, the gains from anchoring are more evident because of that positive term one minus m0 m E F I0.
Tom: So, it seems like the real insight here is how anchoring acts as a generic form of adaptivity to the unknown null proportion, even when dealing with dependence in the data.
Jane: It's about taking something self-referential that gives you no clear path for improvement and turning it into something transparent that retains discoveries while controlling the FDR under the same conditions.
Lu: The paper also looks at how Anchored e-BH compares to closed e-BH, showing they are incomparable, which is a pretty strong finding given their transparency.
Meng: For someone actually running models, that means you can pick a procedure you already trust and apply this anchoring technique to get a more robust version without having to completely overhaul the underlying framework.
Tom: So, if we're talking about what this means for the world right now, it’s not necessarily about inventing a totally new statistical method; it’s about making existing methods more reliable and transparent when we can't fully trust their internal workings.
Jane: It suggests that even in complex areas like multiple testing, we can use simple structural insights—like exploiting looseness in the proofs—to build better, more understandable tools without getting bogged down in overly complicated adaptive corrections.
Lu: The authors conclude that while these new anchored procedures aren't necessarily admissible, they can safely replace earlier methods whenever those methods were employed because they offer simplicity and generality.
Meng: And for production environments, the computational complexity being sorting or quadratic time is what makes them deployable instead of something that would take forever to run.
Lalam: So, the practical implication is that we have a new set of procedures we can use with confidence when our current ones feel opaque or hard to upgrade. That’s the core idea behind this paper.
Conclusion: Tom: So we’re wrapping up on "Anchored multiple testing: a transparent use of e-closure to improve FDR procedures." Essentially, these authors have taken an existing method, e-closure, that can recover any procedure controlling false discovery rate, and they've added this new technique called anchoring.
Jane: It sounds like they’re making the recovery process more transparent. They argue that instead of getting a self-referential set of results that doesn't tell you much on how to improve things, this anchoring gives you a method that keeps every discovery from your baseline procedure while still controlling the false discovery rate under the same conditions.
Lu: The authors show this works across several procedures, like Anchored-BH dominating the standard Benjamini-Hochberg under PRDS, and they are comparing it against closed versions of those procedures to show where it actually gains an edge.
Meng: From a deployment standpoint, what I’m hearing is that you can take a procedure you already use and apply this anchoring technique to get a more robust version without having to completely overhaul your entire pipeline. That’s something practical.
Lalam: The paper shows that this anchoring principle is generic; it defines a local e-value in a way that any certified superset of your original set controls the false discovery rate at alpha and keeps all of your initial discoveries inside it.
Tom: So the main contribution here isn't just inventing a new testing algorithm, but using this anchoring idea to give existing methods more clarity and better performance against known alternatives.
Jane: It seems like the authors are arguing that while these new procedures might not be strictly admissible in every single case, they offer a simpler and more general way to use multiple testing tools when you have uncertainty about the exact null distribution.
Lu: They’re focusing on exploiting some inherent looseness in existing FDR proofs to enlarge that local e-value, which is a clever way to get more discoveries without breaking the control guarantees.
Meng: It’s interesting that they show this works with things like arbitrary dependence and even two-sided Gaussian p-values, meaning this isn't just a theoretical exercise for clean data.
Lalam: They also show that the computational cost for these anchored methods is generally in sorting or quadratic time, which makes them feasible to run on large datasets.
Tom: So the big picture is that we’ve got a new way to use e-closure that's more useful because it doesn't just tell you what happened, but also gives you a path forward for improvement, even if it’s just an enlarged set of discoveries.
Jane: And this leads us into how these improvements translate into real-world modeling and what the authors suggest for future work.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates