📊 statistics

Robust confidence intervals for generalized linear models

This paper proposes a robust method for constructing confidence intervals in generalized linear models by inverting sign-flipped hypothesis tests, ensuring reliable coverage even when variance assumptions are violated, as demonstrated through simulations and RNA-sequencing data analysis.

Original authors: Andrea Panarotto, Riccardo De Santis, Livio Finos

Published 2026-05-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Andrea Panarotto, Riccardo De Santis, Livio Finos

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 mystery using a set of clues. In the world of statistics, these clues are data points, and the "mystery" is figuring out how one thing (like a drug or a gene) affects another (like a patient's health or a cell's behavior).

To solve this, statisticians use a tool called a Generalized Linear Model (GLM). Think of a GLM as a sophisticated map that helps you draw a line through your data points to see the trend. But a map is only useful if you know how accurate it is. That's where Confidence Intervals come in.

The Problem: The "Perfect World" Map

Usually, when statisticians draw these maps, they assume the world is "perfect." They assume the noise in the data (the messy, unpredictable parts) behaves in a very specific, neat way. It's like assuming every car on the road drives at exactly the same speed and never swerves.

In reality, especially in biology and medicine, the world is messy. Data is often "over-dispersed" (too much variation) or "heteroskedastic" (the noise gets louder or quieter depending on the situation). When the real world doesn't match the "perfect world" assumption, the standard maps break.

The paper explains that the traditional way of drawing these confidence intervals (the "Wald" method) is like trusting a map that assumes perfect weather. If a storm hits (variance misspecification), the map tells you the destination is safe, but you might actually be walking off a cliff. The "confidence" you feel is fake because the interval is too narrow, missing the true answer far too often.

The Solution: The "Sign-Flip" Compass

The authors propose a new, more robust way to draw these intervals. Instead of trusting a pre-drawn map based on perfect assumptions, they use a method called Sign-Flipping.

Here is the analogy:
Imagine you have a group of friends giving you directions to a hidden treasure.

The Old Way: You ask them, "Is the treasure to the left?" They say "Yes." You trust them blindly because they usually tell the truth. But if they are all confused by a storm (variance issues), they might all be wrong, and you walk in the wrong direction.
The New Way (Sign-Flipping): You take their directions and play a game. You randomly flip the sign of their answers. Sometimes you pretend they said "Left" when they said "Right," and vice versa. You do this thousands of times.
- If the treasure is really on the left, flipping the signs will make the group look very confused and inconsistent.
- If the treasure is actually in the middle, the flipping won't change the overall pattern much.

By seeing how the group reacts to this "chaos" (the sign flips), you can figure out exactly where the treasure is, regardless of whether the weather is stormy or calm. This method doesn't care if the data is messy; it just looks at the structure of the evidence.

The Challenge: The "Staircase" Problem

The authors noticed a tricky problem with this new compass. Because you are flipping signs a finite number of times (you can't flip them infinite times in real life), the results don't change smoothly like a ramp. They change like a staircase.

If you try to find the exact edge of the confidence interval by walking up this staircase, you might get stuck on a step that isn't the right one. The paper introduces a clever "Bisection Algorithm" (a search method) to navigate this staircase. It's like a game of "Hot and Cold" where you keep halving your search area to find the exact step where the answer changes, ensuring you don't miss the true boundary.

The Proof: Simulations and Real Data

The authors tested this new compass in two ways:

Simulations: They created fake data where they knew the "perfect" answer. They intentionally broke the rules (added stormy weather/over-dispersion).
- Result: The old method (Wald) kept pointing to the wrong place, missing the true answer almost half the time. The new method (Sign-Flip) kept hitting the target, even in the storm.
Real Data: They applied this to a real cancer study involving RNA sequencing (looking at gene activity in liver cancer).
- Result: The old method produced very narrow, confident intervals that likely missed the truth. The new method produced slightly wider intervals, but they were reliable. They were "stable," meaning they gave consistent answers even when the underlying math model was slightly wrong.

The Bottom Line

This paper offers a new tool for scientists. It says: "Don't just trust your map because it looks neat. If the data is messy (which it usually is in biology), use this 'Sign-Flip' compass."

It ensures that when a scientist says, "We are 95% confident the effect is between X and Y," they actually mean it, even if the data doesn't follow the perfect rules they were taught in school. It trades a little bit of precision (wider intervals) for a lot more reliability (not being wrong).

Technical Summary: Robust Confidence Intervals for Generalized Linear Models

Problem Statement
Reliable uncertainty quantification is a critical challenge in modern biomedical data analysis, particularly within Generalized Linear Models (GLMs). Standard confidence intervals for regression parameters in GLMs are inherently parametric, relying on the correct specification of the mean–variance relationship. In practice, assumptions regarding homoskedasticity and the absence of overdispersion are frequently violated due to complex biological variability, heteroskedasticity, or unobserved sources of variation. When these assumptions fail, classical Wald-type intervals often suffer from severe undercoverage, leading to inaccurate scientific inference. While robust covariance estimators (e.g., sandwich estimators) offer partial remedies, they often exhibit slow convergence to nominal coverage levels in finite samples, particularly in settings with small sample sizes or specific distributional misspecifications.

Methodology
The authors propose a semi-parametric framework for constructing confidence intervals in GLMs that remains valid under general variance misspecification. The core of the methodology involves the inversion of hypothesis tests based on the sign-flip score test (Hemerik et al., 2020; De Santis et al., 2025).

Sign-Flip Score Test: The approach utilizes the effective score statistic, $S(\beta, \gamma)$ , derived from the likelihood function. Under the null hypothesis $H_0: \beta = \beta_0$ , the test statistic is computed using the null model (where $\beta_0$ acts as an offset). To generate the null distribution, the signs of the individual score contributions are randomly flipped using diagonal matrices $F_g$ containing $-1$ and $1$. The test statistic is standardized to improve convergence properties.
Test Inversion: Confidence intervals are defined as the set of parameter values $\beta_0$ $β_{0}$ that are not rejected by the level- $\alpha$ $α$ sign-flip test. To construct the interval, the authors invert one-sided tests:
- For $\beta_0 < \hat{\beta}_{obs}$ , they test $H_0: \beta = \beta_0$ vs. $H_1: \beta > \beta_0$ .
- For $\beta_0 > \hat{\beta}_{obs}$ , they test $H_0: \beta = \beta_0$ vs. $H_1: \beta < \beta_0$ .
Monotonicity and Bisection Algorithm: A significant technical hurdle addressed is the potential non-monotonicity of p-value functions derived from permutation-based tests, which can complicate the inversion process. The authors establish theoretical results (Theorems 2 and 3) demonstrating that, under suitable conditions (including linear models and asymptotic regimes for GLMs), the p-value functions are monotonic. Leveraging this, they implement a bisection algorithm to efficiently search for the interval bounds. The algorithm explicitly handles the discreteness of p-values and ensures the retrieval of conservative bounds by searching for the infimum/supremum of the non-rejected region.
Interval Types: The framework supports both "equitailed" intervals (finding separate lower and upper bounds at $\alpha/2$ ) and "symmetric" intervals (finding a symmetric distance $\delta$ from the estimate such that the sum of p-values equals $\alpha$ ).

Key Contributions

Theoretical Guarantees: The paper provides finite-sample and asymptotic monotonicity results for p-value functions arising from sign-flip score tests in the presence of nuisance confounders. This justifies the use of bisection algorithms for interval construction.
Robustness to Variance Misspecification: Unlike standard Wald intervals or even sandwich-based intervals in some finite-sample scenarios, the proposed method maintains nominal coverage even when the dispersion parameters or variance structure are incorrectly specified.
Algorithmic Implementation: The authors introduce a practical bisection search algorithm that navigates the discrete nature of permutation-based p-values, ensuring the resulting confidence sets are connected intervals with correct asymptotic coverage.
Software Availability: The methodology is implemented in the R package flipscores, allowing users to obtain robust confidence intervals via standard confint methods on model objects.

Results
The authors evaluate the method through simulation studies and a real-world application:

Simulation Studies: Across six settings (linear, logistic, Poisson, Negative Binomial, and heteroskedastic models), the proposed flip-based intervals consistently achieved nominal coverage (0.95) even under severe variance misspecification (e.g., overdispersed Poisson data fitted with a Poisson model, or heteroskedasticity in linear models). In contrast, classical Wald intervals showed substantial undercoverage (e.g., dropping below 0.5 in overdispersed Poisson cases), and sandwich-based intervals often failed to reach nominal levels in small samples ( $N=25, 50$ ). While parametric methods produced narrower intervals when assumptions held, they failed to provide reliable inference when assumptions were violated.
Real Data Application: The method was applied to RNA-sequencing data from The Cancer Genome Atlas (TCGA-LIHC) to analyze differential gene expression regarding tumor stage. The study compared intervals derived from Poisson and Negative Binomial models. The flip-based intervals demonstrated greater stability and coherence across different model specifications compared to sandwich-based intervals. Specifically, the flip-based method produced wider intervals that ensured nominal coverage, whereas model-based Wald intervals (especially under Poisson assumptions) yielded overly optimistic, narrow intervals that likely underestimated uncertainty.

Significance and Claims
The paper claims to offer a "practical and robust alternative" to conventional quasi-likelihood or Wald-based methods for interval estimation in GLMs. The primary significance lies in its ability to provide reliable uncertainty quantification in high-throughput biomedical applications where variance structures are complex and difficult to specify correctly. By decoupling the validity of the confidence interval from the correct specification of the variance (while retaining the mean structure), the method reduces the risk of false positives and misleading conclusions driven by modeling errors. The authors position this as a solution for scenarios where "modeling assumptions are difficult to verify," such as genomics, allowing researchers to move beyond point estimation to robust interval inference. The work is modest in scope, focusing on single scalar parameters and avoiding multiplicity issues, while identifying extensions to mixed models, zero-inflated models, and simultaneous inference as future directions.