📊 statistics

Beyond False Discovery Rate: A Stepdown Group SLOPE Approach for Grouped Variable Selection

This paper introduces the Group Stepdown SLOPE, a unified optimization framework that integrates Lehmann-Romano stepdown rules into SLOPE to provide finite-sample guarantees for controlling k-FWER and FDP in grouped variable selection, achieving superior statistical power compared to existing methods across both orthogonal and general design settings.

Original authors: Xuelin Zhang, Jingxuan Liang, Xinyue Liu, Hong Chen, Biqin Song

Published 2026-03-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xuelin Zhang, Jingxuan Liang, Xinyue Liu, Hong Chen, Biqin Song

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a massive mystery. You have a room full of 1,000 suspects (features), but you know that only a handful of them are actually guilty (relevant variables). Your job is to pick out the guilty ones while leaving the innocent ones alone.

The problem? If you accuse too many innocent people just to be safe, you create a "False Discovery" mess. If you are too strict, you might let the real criminals go free (low power).

This paper introduces a new, smarter way to run this investigation, specifically when the suspects come in groups (like families or gangs) and when you need to be extremely precise about how many mistakes you make.

Here is the breakdown of their new method, Group Stepdown SLOPE, using simple analogies:

1. The Old Way: The "One-Size-Fits-All" Net

Previously, detectives used a method called SLOPE. Imagine SLOPE as a fishing net with holes of different sizes. It casts a wide net to catch the "big fish" (important variables).

The Good: It's great at controlling the average number of innocent people you accidentally catch (called the False Discovery Rate or FDR).
The Bad: It doesn't care if you accidentally catch too many innocent people in a single batch. It also struggles if the suspects are standing in groups (e.g., a whole family is either guilty or innocent, not just one person).

2. The New Problem: The "Group" and the "Strict Boss"

In the real world (like in medical research or genetics), variables often come in groups. You either select the whole group or none of it.
Furthermore, sometimes a "Strict Boss" (like a government regulator or a medical board) says:

"I don't just want the average mistakes to be low. I want to guarantee that the chance of making more than 5 mistakes is almost zero." (This is k-FWER control).
"I also want to guarantee that the percentage of mistakes in your final list never exceeds 10%." (This is FDP control).

The old SLOPE net couldn't promise these strict guarantees, especially for groups.

3. The Solution: The "Stepdown" Ladder

The authors built a new tool called Group Stepdown SLOPE. Think of this as a smart, multi-level security checkpoint.

Instead of casting one big net, imagine a ladder with rungs. You start at the top (the most suspicious suspects) and work your way down.

The Stepdown Logic: You check the most suspicious person. If they pass the test, you move to the next. But here is the trick: as you go down the ladder, the test gets harder.
The "Group" Twist: If a whole group of suspects is standing together, this new method checks the entire group at once. If the group looks suspicious, you check the whole gang. If they look innocent, you let the whole gang go.

4. Why is this a Big Deal? (The "Magic" of the Math)

The paper proves two amazing things:

The Guarantee: They mathematically proved that if you use their specific "ladder rules" (regularization sequences), you can guarantee that you will never accidentally arrest more than $k$ innocent people (k-FWER) or that your mistake rate will never exceed a certain percentage (FDP). It's like having a legal contract that says, "We promise we won't mess up more than this."
The Power: Usually, when you make a rule stricter to avoid mistakes, you end up missing the real criminals (low power). But this new method is so efficient that it actually catches more guilty suspects than the old methods while still keeping the mistake rate low. It's like having a super-accurate metal detector that doesn't beep when you walk past a belt buckle.

5. Real-World Test: The Alzheimer's Case

To prove it works, they tested it on real data from the Alzheimer's Disease Neuroimaging Initiative (ADNI).

The Setup: They looked at brain scans where features (like brain thickness or volume) naturally come in groups (e.g., all measurements from the "frontal lobe").
The Result: Their new method found the most relevant brain features with fewer errors than the old methods. It was faster and more accurate, proving that their "smart ladder" works even when the data is messy and complex.

Summary Analogy

Old Method (SLOPE): A sieve that catches big fish but lets some small fish slip through and sometimes catches too many rocks (mistakes) on average.
New Method (Group Stepdown SLOPE): A high-tech, tiered security scanner. It checks people in groups. It has a strict rulebook that guarantees you won't let more than 5 innocent people through the gate, and it guarantees that less than 10% of the people you stop are actually innocent. Best of all, it's so good at its job that it catches the real bad guys more often than the old scanner did.

In short: This paper gives scientists a new, mathematically proven "super-tool" to find important patterns in huge, messy data sets without making too many embarrassing mistakes.

1. Problem Statement

High-dimensional feature selection often requires balancing statistical power with strict control over multiple error metrics. While existing methods like SLOPE (Sorted L-One Penalized Estimation) and Group SLOPE (g-SLOPE) effectively control the False Discovery Rate (FDR), they face two critical limitations:

Limited Error Metrics: They primarily focus on FDR (the expected proportion of false discoveries) but lack rigorous guarantees for stricter metrics like the $k$ -Family-Wise Error Rate ( $k$ -FWER) (probability of $k$ or more false rejections) and the False Discovery Proportion (FDP) (probability that the proportion of false discoveries exceeds a threshold $\gamma$ ).
Lack of Group Structure Integration: Existing stepdown procedures for controlling $k$ -FWER and FDP rely on $p$ -values, which are computationally challenging in high-dimensional, non-linear, or complex regression settings. Furthermore, standard SLOPE does not natively handle grouped variable selection under these stricter error criteria.

The paper addresses the need for a unified framework that performs grouped variable selection while providing finite-sample guarantees for $k$ -FWER and FDP control, without relying on complex $p$ -value calculations.

2. Methodology

The authors propose a unified optimization procedure called Group Stepdown SLOPE, which integrates the Lehmann–Romano stepdown rules into the SLOPE framework. The methodology is divided into single-regressor and grouped settings.

A. Core Concept: Stepdown Integration

Instead of using the Benjamini-Hochberg (BH) procedure (used for FDR control), the authors replace the regularization sequence with thresholds derived from the stepdown procedure. This allows the convex optimization problem to inherently control stricter error rates.

B. Single-Variable Selection (k-SLOPE and F-SLOPE)

For selecting individual features, the authors define two new convex optimization problems:

k-SLOPE: Controls $k$ -FWER.
- Objective: $\min_{\beta} \frac{1}{2}\|y - X\beta\|^2 + \sigma \sum \lambda_{k\text{-FWER}(i)} |\beta|_{(i)}$
- Regularization Sequence: Derived from Lehmann-Romano thresholds:
  $\lambda_{k\text{-FWER}(i)} = \Phi^{-1}\left(1 - \frac{k\alpha}{2(m+k-i)}\right) \quad (\text{for } i > k)$
- Guarantee: Proven to control $k$ -FWER $\leq \alpha$ under orthogonal designs.
F-SLOPE: Controls FDP.
- Objective: $\min_{\beta} \frac{1}{2}\|y - X\beta\|^2 + \sigma \sum \lambda_{\text{FDP}(i)} |\beta|_{(i)}$
- Regularization Sequence: Derived to bound $P(\text{FDP} > \gamma) \leq \alpha$ :
  $\lambda_{\text{FDP}(i)} = \Phi^{-1}\left(1 - \frac{(\lfloor \gamma i \rfloor + 1)\alpha}{2(m + \lfloor \gamma i \rfloor + 1 - i)}\right)$

C. Grouped Variable Selection (gk-SLOPE and gF-SLOPE)

The authors extend the above to Group SLOPE, where features are selected or discarded as whole groups.

Objective: Minimizes the loss plus a penalty on the sorted $\ell_2$ norms of group coefficients.
gk-SLOPE: Controls group $k$ -FWER ($gk$-FWER), ensuring the probability of selecting $k$ or more false groups is below $\alpha$ .
gF-SLOPE: Controls group FDP ( $g$ FDP).
Orthogonal vs. General Designs:
- Orthogonal: Closed-form regularization sequences are derived using the cumulative distribution function of the $\chi$ distribution (scaled by group size).
- General (Non-Orthogonal): A data-driven, calibrated sequence is proposed. It uses a Gaussian approximation and Monte-Carlo correction to adjust the penalty parameters, preserving convexity and scalability even when the design matrix $X$ is not orthogonal.

3. Key Contributions

New Algorithms: Introduction of k-SLOPE, F-SLOPE, gk-SLOPE, and gF-SLOPE. These are the first methods to combine SLOPE's adaptive penalization with stepdown rules for $k$ -FWER and FDP control.
Theoretical Guarantees:
- Proven finite-sample guarantees for $k$ -FWER and FDP control under orthogonal designs.
- Extension of these guarantees to grouped structures ($gk$-FWER and $g$ FDP).
- Demonstration that the proposed methods avoid the complex $p$ -value calculations required by traditional stepdown procedures.
Generalization: Development of a heuristic, data-driven regularization sequence for non-orthogonal (general) designs, ensuring the methods remain applicable to real-world correlated data.
Comprehensive Comparison: A detailed analysis (summarized in Table 1 of the paper) showing that the proposed methods are the only ones capable of simultaneously handling Group-wise selection, FDR, FDP, and $k$ -FWER control.

4. Empirical Results

Extensive simulations and real-world experiments were conducted to validate the theory.

Orthogonal Designs:
- Error Control: Both k-SLOPE and F-SLOPE successfully maintained error rates (FDP, $k$ -FWER) below the nominal levels ( $\alpha=0.1$ ).
- Power: The proposed methods achieved significantly higher statistical power (true positive rate) compared to competing stepdown procedures and standard SLOPE, particularly in sparse regimes.
Correlated/General Designs:
- In scenarios with correlated test statistics (simulating multiple mean testing) and Gaussian designs ( $X \sim N(0, 1/n)$ ), the data-driven sequences preserved error control while maintaining high power.
- Grouped Settings: gk-SLOPE and gF-SLOPE outperformed standard g-SLOPE in controlling group-level errors ($gk$-FWER and $g$ FDP) while maintaining comparable or superior power.
Real-World Application (ADNI Data):
- Applied to Alzheimer's Disease Neuroimaging Initiative data with known group structures (cortical thickness, volume, etc.).
- Results: The proposed methods achieved lower Mean Squared Error (MSE) and strictly controlled error rates (e.g., gk-FWER dropped from 0.82 for g-SLOPE to 0.12 for gk-SLOPE) compared to Group Lasso and standard g-SLOPE, without significant loss in computational efficiency.

5. Significance

This paper represents a significant advancement in high-dimensional statistics and machine learning by:

Broadening the Scope of SLOPE: Moving beyond FDR to support stricter, more practical error metrics ( $k$ -FWER and FDP) which are crucial in fields like genomics and neuroimaging where false positives are costly.
Unifying Group and Stepdown Methods: Successfully integrating group sparsity constraints with stepdown error control, a combination previously unexplored.
Practical Applicability: Providing a computationally feasible, convex optimization framework that works for both orthogonal and general correlated designs, making it ready for real-world deployment in complex datasets.

In summary, the authors have developed a robust, theoretically grounded, and empirically superior framework for controlled feature selection that addresses the limitations of current FDR-only methods and $p$ -value-dependent stepdown procedures.