Response-guided knockoffs for prespecified directional FDR control in linear models

arXiv:2608.29083 · stat.ME · Submitted 2026-08-29 · Read on arXiv

Listen

Radio episode about this paper

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: "Response-guided knockoffs for prespecified directional FDR control in linear models".

Jane: Here's the gist of the paper. Response-guided knockoffs for directional FDR control in linear models introduces a response-guided knockoff filter that leverages a noise-perturbed version of the response to guide knockoff construction toward features likely to have the target sign, while provably controlling the directional FDR <ref:2608.29083#pg12>.

Tom: Problem Statement The paper considers feature selection in linear models with finite-sample control of the false discovery rate (FDR) where the goal is to select features whose effects act in a pre-specified direction d <ref:2608.29083#pg3>.

Jane: Existing knockoff-based methods control the directional FDR but are response-agnostic and do not target discoveries in a pre-specified direction <ref:2608.29083#pg12>.

Tom: The authors introduce the response-guided knockoff filter, which uses a noise-perturbed response to focus knockoff construction on features likely to have the target sign, while provably controlling the directional FDR <ref:2608.29083#pg12>.

Paper summary: Tom: So to recap this paper, they're looking at feature selection in linear models where you care about features pointing a specific direction d, like positive or negative. They introduce the response-guided knockoff filter which uses a noise perturbed response to guide knockoff construction toward features likely having that target sign.

Jane: The main claim is that this method provably controls the directional false discovery rate, and it does this under a weaker sample size requirement of n greater than p plus two, compared to existing fixed X generators that need n greater than or equal to two p.

Lu: It matters because previous knockoff methods were response-agnostic; they didn't target discoveries in a pre-specified direction, so this new approach incorporates response information directly into the construction to focus on features likely matching the target sign.

Meng: So it's not just controlling the FDR anymore; it’s actively trying to find features that are pointing in a specific way, which is more relevant for targeted applications like drug resistance studies.

Tom: Exactly. The empirical results show power gains over existing methods, and simulations demonstrate that this response-guided knockoff filter can be substantially better than what's currently out there, especially when you have sample sizes between p plus two and two times p.

Jane: It seems like the key takeaway is that by using a noise perturbed response to guide the knockoff construction, we can get better directional control with less data than previously required.

Lu: The paper also mentions it extends to controlling other error criteria like the false discovery exceedance FDX and the k family wise error rate k-FWER, which broadens its applicability beyond just directional FDR control.

Meng: That's good for deployment because it means we can apply this same framework to monitor other kinds of errors, not just one specific type of false discovery.

Tom: So it’s about making the knockoff process smarter by using response information to focus the search space on what actually matters in a directional setting.

Jane: And that leads us into the next part, where we discuss what this means for feature selection in general and how it sets up future research.

Conclusion: Tom: Looking at the full title, Response-guided knockoffs for prespecified directional FDR control in linear models tells you exactly what this is about: it’s a selective inference framework that uses response information to guide how you build knockoffs when you only care about features pointing a specific direction.

Jane: It's important to remember that while the method offers better sample size requirements, we need to consider the practical trade-off they mention regarding noise perturbation epsilon tilde.

Lu: So, from a research standpoint, the main challenge is figuring out how to adapt this response information—the guide—to different model assumptions while maintaining that finite sample error control.

Meng: For me in production, if we can stabilize the noise selection process and get a reliable estimate for epsilon tilde, that would make this framework much more deployable than just a theoretical tool.

Tom: So the implication is that for researchers interested in one-sided feature selection, this gives them a way to achieve stronger statistical control with less data by making their knockoff construction smarter.

Jane: It gives us a practical direction to explore how incorporating response information into the knockoff process can actually boost power when the scientific question is inherently one-sided.

School of Mathematical and Physical Sciences, Macquarie University · School of Mathematics and Statistics, University of Sydney

stat.ME

Submitted: 2026-08-29

Updated: 2026-11-06

Code: https://github.com/freejstone/response_

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: Response-guided knockoffs for directional FDR control in linear models introduces a response-guided knockoff filter that leverages a noise-perturbed version of the response to guide knockoff…

Key concepts

Directional FDR Control
This is a statistical goal where the method aims to ensure that the proportion of false discoveries among features selected in a specific direction (e.g., positive effect) remains below a predetermined threshold. The paper introduces new methods to achieve this control specifically for pre-specified directions.
Response-Guided Knockoff Filter
This is the core method where noise is added to the actual response data to create a perturbed response. This perturbed data helps identify features that are both relevant and likely to have the desired sign, thereby focusing the knockoff generation process on promising candidates.
Noise Perturbation ($ ilde{y}$)
The original response data ($y$) is intentionally corrupted by adding noise to create a new, perturbed response ($ ilde{y}$). This perturbed version is used as a guide to determine which features are likely relevant and possess the target sign, steering the subsequent knockoff construction.
Model-X Assumptions
These are assumptions about the underlying linear model structure that allow for certain statistical guarantees. The paper notes that while it controls directional FDR under these assumptions, adapting this framework to different model structures is a future research direction.

Terminology

Summary

Response-guided knockoffs for directional FDR control in linear models introduces a response-guided knockoff filter that leverages a noise-perturbed version of the response to guide knockoff construction toward features likely to have the target sign, while provably controlling the directional FDR <ref:2608. p. 0

Where the paper says it

While existing knockoff-based methods control the directional FDR, which penalises incorrect sign estimates, they do not target discoveries in a pre-specified direction, and their knockoff constructions are entirely response-agnostic. We introduce the response-guided knockoff filter, which leverages a noise-perturbed version of the response to guide knockoff construction toward features likely to have the target sign, while provably controlling the directional FDR.

29083#pg12>.

Problem Statement

The paper considers feature selection in linear models with finite-sample control of the false discovery rate (FDR) where the goal is to select features whose effects act in a pre-specified direction d <ref:2608.29083#pg3>. Existing knockoff-based methods control the directional FDR but are response-agnostic and do not target discoveries in a pre-specified direction <ref:2608. p. 0

Where the paper says it

Response-guided knockoffs for directional FDR control in linear models Jack Freestone1*, Garth Tarr2 , Samuel Muller1 , and Uri Keich2 1School of Mathematical and Physical Sciences, Macquarie University 2School of Mathematics and Statistics, University of Sydney Abstract We consider the problem of feature selection in linear models with finite-sample control of the false discovery rate (FDR). While existing knockoff-based methods control the directional FDR, which penalises incorrect sign estimates, they do not target discoveries in a pre-specified direction, and their knockoff constructions are entirely response-agnostic.

29083#pg12>. The authors introduce the response-guided knockoff filter, which uses a noise-perturbed response to focus knockoff construction on features likely to have the target sign, while provably controlling the directional FDR <ref:2608. p. 0
Where the paper says it

While existing knockoff-based methods control the directional FDR, which penalises incorrect sign estimates, they do not target discoveries in a pre-specified direction, and their knockoff constructions are entirely response-agnostic. We introduce the response-guided knockoff filter, which leverages a noise-perturbed version of the response to guide knockoff construction toward features likely to have the target sign, while provably controlling the directional FDR.

29083#pg12>.

Methodology Overview

The response-guided knockoff filter begins by adding noise to the response y to create a perturbed response, denoted as y˜ <ref:2608.29083#pg12>. Using the perturbed response y˜ and X we can estimate which features are both relevant to the response and have the appropriate sign <ref:2608. p. 8

Where the paper says it

Our method, which we call the response-guided knockoff filter (Algorithm 3), begins by adding noise to the response y to create a perturbed response, denoted as y˜. Using the perturbed response y˜ and X we can estimate which features are both relevant to the response and have the appropriate sign.

29083#pg12>. The method involves two stages: Stage 1, where a subset Sˆ 0 of features is identified by regressing y˜ onto X <ref:2608. p. 8
Where the paper says it

Concretely, our approach (developed in Section 2. 6) is to assign nonzero separation to a set of features Sˆ 0 by regressing y˜ on X.

29083#pg12>, and Stage 2, where knockoff construction is applied only to the features in Sˆ 0 <ref:2608. p. 5
Where the paper says it

Barber and Cand`es [2015] presented SSS+ (Algorithm 1) as a more general procedure for multiple hypothesis testing which is applied as part of the knockoff filter with c = 1/2, where c ∈ (0, 1) is any constant such that P(Wj > 0) ≤ c for all j ∈ H0. It was shown that under the linear model assumption, reporting the features in SˆKF under Conditions 1–3 controls the FDR at level α.

29083#pg12>. The construction of the knockoffs must satisfy Condition 4, which involves a perturbed response feasibility constraint <ref:2608. p. 9
Where the paper says it

To account for peeking at the response through y˜, our knockoff construction must satisfy the following constraint, which extends Condition 1. Condition 4 (Perturbed response feasibility).

29083#pg12>.

Theoretical Guarantees

The signed-knockoff filter is shown to control the directional FDR when targeting discoveries with sign d by defining a score Sj = sign(Xj − Xe j)⊤y · sign (Wj) and applying SSS+ to features where Sj = d <ref:2608.29083#pg12>. The response-guided knockoff filter is proven to control the directional FDR under a milder sample-size requirement n > p + 2 compared to existing fixed-X generators <ref:2608. p. 0

Where the paper says it

We introduce the response-guided knockoff filter, which leverages a noise-perturbed version of the response to guide knockoff construction toward features likely to have the target sign, while provably controlling the directional FDR. The method operates under a weaker sample-size requirement n > p + 2, compared to n ≥ 2p required by existing fixed-X generators.

29083#pg12>. Theorem 1 establishes that the response-guided knockoff filter Sˆ RGKF controls the directional FDR in (4) when y ∼ N (Xβ, σ squared In) with n > p + 2 <ref:2608. p. 8
Where the paper says it

Assume y ∼ N (Xβ, σ2 In) and n ≥ 2p. Then the signed-knockoff filter Sˆ signed-KF controls the directional FDR in (4), i.

29083#pg12>. Furthermore, the framework extends to controlling other error criteria like the false discovery exceedance (FDX) and k-family-wise error rate (k-FWER) <ref:2608.29083#pg12>.

Empirical Performance

Simulations demonstrate that the response-guided knockoff filter variants are more powerful than existing methods, often substantially so, with the RG(reg-signed, ME) variant clearly performing best <ref:2608.29083#pg12>. The empirical directional FDR of all methods remains at or below the nominal level α in simulations and real-data experiments <ref:2608. p. 5

Where the paper says it

Section 2 shows why naive adaptations of the knockoff filter fail to control the directional FDR when targeting discoveries in a specific direction, and then develops two principled methods: (i) the signed-knockoff filter, a modification of the knockoff filter that selects features in a target direction with directional FDR control, and (ii) the response-guided knockoff filter, which is designed to increase power by incorporating response information into the knockoff construction. Section 3 reports simulation and real-data experiments.

29083#pg12>. In simulations where p + 2 < n < 2p, the response-guided knockoff filter outperforms the signed-knockoff filter, often substantially so, with the RG(reg-signed, ME) variant clearly performing best <ref:2608. p. 9
Where the paper says it

With the knockoffs constructed as above, we continue by applying the remaining steps of the signed-knockoff filter (Algorithm 2). Our approach clearly violates the canonical knockoff construction where the knockoffs are constructed without looking at the response.

29083#pg12>.

Future Directions

The paper notes that a crucial step is to explore adaptive ways of selecting the size of the perturbation ε˜ that is added to the response y <ref:2608.29083#pg12>. The authors suggest that a smaller perturbation makes y˜ a sharper guide and yields a more accurate screened set Sˆ 0 and allows us to optimize the separation parameters sj for j ∈ Sˆ 0, but because the inference is made invariant to y˜ (Condition 4), it attenuates the signal left for the statistics W <ref:2608.29083#pg12>. There is also an important direction for future work in exploring alternative ways for constructing response-guided knockoffs under the model-X assumptions <ref:2608. p. 1

Where the paper says it

The goal in such settings is not simply to identify which features are relevant to a response, but to select those whose effects act in a pre-specified direction d ∈ {±1}, while controlling the rate of false discoveries among these directional claims. We work in the classical linear model, where we observe n data points from Y = X p j=1 Xjβj + ε, (1) with unknown coefficients β ≡ (β1, .

29083#pg12>.

Closed Form Solutions

The closed-form solution to the equi-correlated response-guided knockoffs is given by s∗ = 2λmin Σ(e) Sˆ 0Sˆ 0 − Σ(e) Sˆ 0S c 0 Σ(e) S c 0S c 0 −1 Σ(e) S c 0Sˆ 0 <ref:2608.29083#pg12>. This solution is derived by replacing the matrix Σ(e) in the ordinary equi optimisation problem with a modified form that incorporates the structure of the response-guided construction <ref:2608.29083#pg12>. The proof relies on Schur complements and establishes this closed-form solution <ref:2608. p. 0

Where the paper says it

Simulations and HIV drug resistance experiments demonstrate power gains over existing methods. arXiv:2608.

29083#pg12>.

Algorithm Summary

The paper presents several algorithms, including Algorithm 1 Selective SeqStep+ (SSS+) which is used in the signed-knockoff filter <ref:2608.29083#pg12>, and Algorithm 3 JS which controls the k-FWER <ref:2608. p. 5

Where the paper says it

Barber and Cand`es [2015] presented SSS+ (Algorithm 1) as a more general procedure for multiple hypothesis testing which is applied as part of the knockoff filter with c = 1/2, where c ∈ (0, 1) is any constant such that P(Wj > 0) ≤ c for all j ∈ H0. It was shown that under the linear model assumption, reporting the features in SˆKF under Conditions 1–3 controls the FDR at level α.

29083#pg12>. The response-guided knockoff filter is implemented via Algorithm 3 Response-guided knockoff filter, which incorporates steps for constructing the perturbed response y˜ and solving for separation parameters s <ref:2608. p. 8
Where the paper says it

Instead, we propose to use part of the information in y to guide the construction of the knockoffs, so we can focus on the more promising features. Our method, which we call the response-guided knockoff filter (Algorithm 3), begins by adding noise to the response y to create a perturbed response, denoted as y˜.

29083#pg12>. The algorithms are designed to control directional FDR, FDX, and k-FWER <ref:2608. p. 5
Where the paper says it

Algorithm 1 Selective SeqStep+ (SSS+) [Barber and Cand`es, 2015] Input: W ≡ (Wi) m i=1 the list of statistics; c ∈ (0, 1) such that P(Wj > 0 | j ∈ H0) ≤ c; α ∈ (0, 1) the FDR threshold Output: A discovery list SˆKF 1: At ← #{i ∈ [m] : Wi ≤ −t} 2: Rt ← #{i ∈ [m] : Wi ≥ t} 3: τ ← inf{t > 0 : At+1 Rt∨1 · c 1−c ≤ α} 4: return SˆKF ← {j ∈ [p] : Wj ≥ τ} The remainder of this paper is organized as follows. Section 2 shows why naive adaptations of the knockoff filter fail to control the directional FDR when targeting discoveries in a specific direction, and then develops two principled methods: (i) the signed-knockoff filter, a modification of the knockoff filter that selects features in a target direction with directional FDR control, and (ii) the response-guided knockoff filter, which is designed to increase power by incorporating response information into the knockoff construction.

29083#pg12>.

Data Analysis Details

The analysis is performed on HIV drug resistance data from the Stanford HIV Drug Resistance Database <ref:2608.29083#pg12>. The data was pre-processed according to the tutorial at https://web. p. 7

Where the paper says it

Figure C1 confirms FDR control can be compromised with this approach. 2 Signed-knockoff filter for directional FDR control Our first procedure for controlling the directional FDR when targeting discoveries with sign d, the signed-knockoff filter, consists of pre-processing the scores according to a sign estimate different from (12) and running the knockoff filter.

stanford.edu/group/candes/knockoffs/software/knockoffs/tutorial-4-r. p. 0
Where the paper says it

Response-guided knockoffs for directional FDR control in linear models Jack Freestone1*, Garth Tarr2 , Samuel Muller1 , and Uri Keich2 1School of Mathematical and Physical Sciences, Macquarie University 2School of Mathematics and Statistics, University of Sydney Abstract We consider the problem of feature selection in linear models with finite-sample control of the false discovery rate (FDR). While existing knockoff-based methods control the directional FDR, which penalises incorrect sign estimates, they do not target discoveries in a pre-specified direction, and their knockoff constructions are entirely response-agnostic.

A ground truth was approximated as those mutations that did not clearly indicate a decrease in resistance for that drug <ref:2608. p. 1
Where the paper says it

1 Introduction In many scientific applications, the goal of feature selection is inherently one-sided. For example, when studying HIV drug resistance, investigators may seek genetic mutations that increase resistance to a particular drug; mutations that decrease resistance or have no effect are not of primary interest.

29083#pg12>. The dimensions of the HIV drug resistance data for each PI drug are summarized in Table C2 <ref:2608. p. 0
Where the paper says it

Simulations and HIV drug resistance experiments demonstrate power gains over existing methods. arXiv:2608.

29083#pg12>.

Conclusion

The response-guided knockoff filter is a selective inference framework that incorporates response information into knockoff construction while provably controlling the directional FDR <ref:2608.29083#pg12>. It can be implemented when p + 2 < n < 2p, and it controls other error rates such as the directional FDX and k-FWER <ref:2608. p. 1

Where the paper says it

The knockoff filter of Barber and Cand`es [2015] controls the FDR in finite samples by constructing artificial negative controls, called knockoffs, that mimic the correlation structure of the original features. In their follow up work, Barber and Cand`es [2019] further showed that the knockoff filter controls a stronger error metric called directional FDR, which penalises both irrelevant discoveries and incorrect sign estimates: FDRdir ≡ E(FDPdir) ≤ α, FDPdir ≡ |{j ∈ Sˆ : sign( \βj ) ̸= sign (βj )}| |S| ∨ ˆ 1 , (3) where sign( \βj ) is an estimate of the sign of βj .

29083#pg12>. The empirical results demonstrate that the response-guided knockoff filter can lead to substantial power gains over existing knockoff generators <ref:2608. p. 0
Where the paper says it

Simulations and HIV drug resistance experiments demonstrate power gains over existing methods. arXiv:2608.

29083#pg12>. The proof machinery relies heavily on the fact that the data is a linear model <ref:2608. p. 1
Where the paper says it

The goal in such settings is not simply to identify which features are relevant to a response, but to select those whose effects act in a pre-specified direction d ∈ {±1}, while controlling the rate of false discoveries among these directional claims. We work in the classical linear model, where we observe n data points from Y = X p j=1 Xjβj + ε, (1) with unknown coefficients β ≡ (β1, .

29083#pg12>. The response-guided knockoff filter cannot be easily adapted to the model-X assumptions while maintaining its finite-sample error control <ref:2608. p. 0
Where the paper says it

Response-guided knockoffs for directional FDR control in linear models Jack Freestone1*, Garth Tarr2 , Samuel Muller1 , and Uri Keich2 1School of Mathematical and Physical Sciences, Macquarie University 2School of Mathematics and Statistics, University of Sydney Abstract We consider the problem of feature selection in linear models with finite-sample control of the false discovery rate (FDR). While existing knockoff-based methods control the directional FDR, which penalises incorrect sign estimates, they do not target discoveries in a pre-specified direction, and their knockoff constructions are entirely response-agnostic.

29083#pg12>. The code for reproducing the analyses is available at https://github.com/freejstone/response-guided knockoffs code <ref:2608. p. 0
Where the paper says it

Response-guided knockoffs for directional FDR control in linear models Jack Freestone1*, Garth Tarr2 , Samuel Muller1 , and Uri Keich2 1School of Mathematical and Physical Sciences, Macquarie University 2School of Mathematics and Statistics, University of Sydney Abstract We consider the problem of feature selection in linear models with finite-sample control of the false discovery rate (FDR). While existing knockoff-based methods control the directional FDR, which penalises incorrect sign estimates, they do not target discoveries in a pre-specified direction, and their knockoff constructions are entirely response-agnostic.

29083#pg12>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Response-guided knockoff filter can increase power for discoveries in a pre-specified direction by using a noise perturbed version of the response to guide knockoff construction toward features likely to have the target sign.

  2. The response-guided knockoff filter can be implemented when n is only moderately larger than p, specifically when p + 2 = 2p with heuristics that no longer guarantee finite-sample FDR control.

  3. The framework extends naturally to other error criteria such as the false discovery exceedance (FDX) and k-family-wise error rate (k-FWER) by replacing the SSS+ procedure in Algorithm 3 with appropriate procedures designed to control these specific error rates.

  4. The response-guided knockoff filter can be implemented using a two stage procedure to construct separation parameters s, where the first stage screens for promising features and the second stage constructs knockoffs only for those features, leading to a trade-off between informing construction and inference.

  5. The framework allows for an adaptive way of selecting the size of the perturbation epsilon with respect to the response y, which can be optimized to yield further power gains by balancing informative screening against signal attenuation during inference.

Abstract

Conventional false discovery rate (FDR) control limits the expected proportion of selected features that are unrelated to the response. Directional FDR extends this criterion to also account for incorrect sign estimates. In some applications, however, discoveries are sought only in a direction specified in advance. We therefore introduce the prespecified directional FDR, which counts as errors both null effects and effects opposite to the target direction in context of the fixed-design Gaussian linear model with unknown noise variance. A straightforward adaptation of the knockoff filter provides prespecified directional FDR control. To improve power, we develop the response-guided knockoff filter, which uses a noise-perturbed response to guide knockoff construction in favour of features likely to have the target sign, while retaining finite-sample prespecified directional FDR control. Notably, our improved filter maintains this guarantee with a sample size n>p+2 in contrast to standard constructions that require n at least2p. Simulations across a range of settings demonstrate power gains of the response-guided knockoff filter. An application to HIV drug-resistance data illustrates improved recovery of externally annotated resistance-associated mutations.

Sources

Related papers