Power-Enhanced Two-Sample Mean Tests for High-Dimensional Compositional Data with Application to Microbiome Data Analysis

Danning Li, Lingzhou Xue, Haoyi Yang, Xiufan Yu

Abstract

Testing differences in mean vectors is a fundamental task in the analysis of high-dimensional compositional data. Existing methods may suffer from low power if the underlying signal pattern is in a situation that does not favor the deployed test. In this work, we develop two-sample power-enhanced mean tests for high-dimensional compositional data based on the combination of $p$-values, which integrates strengths from two popular types of tests: the maximum-type test and the quadratic-type test. We provide rigorous theoretical guarantees on the proposed tests, showing accurate Type-I error rate control and enhanced testing power. Our method boosts the testing power towards a broader alternative space, which yields robust performance across a wide range of signal pattern settings. Our theory also contributes to the literature on power enhancement and Gaussian approximation for high-dimensional hypothesis testing. We demonstrate the performance of our method on both simulated data and real-world microbiome data, showing that our proposed approach improves the testing power substantially compared to existing methods.

Introduction

In microbiome research, identifying the drivers of health and disease often involves comparing microbial profiles between two groups—such as patients with a specific condition versus healthy controls. However, microbiome data presents unique statistical challenges. It is inherently high-dimensional, often involving thousands of species, and "compositional," meaning the data represents relative abundances that must sum to 100%.

This compositional nature places the data within the Aitchison simplex, creating a singular covariance matrix that invalidates many standard statistical tools. Furthermore, researchers face a "sparsity dilemma": they rarely know if a biological difference is driven by a few "smoking gun" taxa (a sparse signal) or by subtle, ecosystem-wide shifts across many taxa (a dense signal). Existing tests typically excel at detecting one pattern while failing the other. This paper introduces a robust framework designed to maintain high power regardless of the underlying signal structure.

Methodology

The Compositional Framework

To address the constraints of the Aitchison simplex, the authors utilize the Centered Log-Ratio (CLR) transformation. This maps the compositional data from the constrained simplex into a standard Euclidean space, allowing for the application of high-dimensional mean testing machinery while accounting for the relative nature of the data.

Power Enhancement via Combination

The core of the proposed method lies in combining two distinct statistical approaches:

  1. Maximum-type Statistics ($M_{n1,n2}$): These are highly sensitive to sparse signals where only a few components of the mean vector differ significantly.
  2. Quadratic-type Statistics ($Q_{n1,n2}$): These are more effective at detecting dense signals where many components show small, cumulative differences.

Theoretical Breakthrough: Asymptotic Independence

The primary theoretical contribution of this work is the rigorous proof that these two statistics are asymptotically independent under the null hypothesis. This independence is a non-trivial finding in high-dimensional settings. It allows the researchers to combine the $p$-values from both tests—using Fisher’s method or the Cauchy combination method—into a single, unified test statistic without "double-counting" evidence or inflating the Type-I error rate.

Empirical Results

Simulation Studies

The researchers conducted extensive simulations using Gaussian and Gamma distributions across various correlation structures (e.g., AR(1) and sparse block matrices). The results demonstrated that:

  • In sparse settings (e.g., only 1% of species changing), the combined test matched the high performance of the Maximum-type test.
  • In dense settings (e.g., 50% of species changing), the combined test remained competitive with the Quadratic-type test, whereas the Maximum-type test’s power dropped significantly.
  • The method consistently maintained the target Type-I error rate (e.g., $\alpha = 0.05$), ensuring that the gain in power did not lead to increased false positives.

Application to Microbiome Data

The framework was applied to a gut microbiome study involving patients with inflammatory bowel disease (IBD). Analyzing 226 genera, the authors identified cases where individual tests disagreed. For instance, when analyzing the impact of antibiotics, the Maximum-type test failed to find a significant difference, but the Quadratic-type and the proposed combined tests successfully identified a significant, ecosystem-wide shift. This highlights the practical utility of the power-enhanced approach as an "insurance policy" against unknown signal patterns.

Discussion and Limitations

While the proposed method offers a robust solution for high-dimensional mean testing, certain challenges remain. The authors noted that power can decrease if there is a severe sample size imbalance between the two groups. Additionally, while the current framework focuses on two-sample mean tests, future work could extend this to more complex metadata adjustments (e.g., age, diet, or geography) and further optimize computational efficiency for datasets exceeding tens of thousands of features.

Conclusion

By bridging the gap between maximum-type and quadratic-type tests, this research provides a versatile tool for microbiome analysis. The power-enhanced framework ensures that researchers do not have to gamble on the nature of their data's signal, offering a statistically rigorous and biologically sensitive method for discovering meaningful differences in high-dimensional compositional datasets.

Summary updated

Abstract

Testing differences in mean vectors is a fundamental task in the analysis of high-dimensional compositional data. Existing methods may suffer from low power if the underlying signal pattern is in a situation that does not favor the deployed test. In this work, we develop two-sample power-enhanced mean tests for high-dimensional compositional data based on the combination of $p$-values, which integrates strengths from two popular types of tests: the maximum-type test and the quadratic-type test. We provide rigorous theoretical guarantees on the proposed tests, showing accurate Type-I error rate control and enhanced testing power. Our method boosts the testing power towards a broader alternative space, which yields robust performance across a wide range of signal pattern settings. Our theory also contributes to the literature on power enhancement and Gaussian approximation for high-dimensional hypothesis testing. We demonstrate the performance of our method on both simulated data and real-world microbiome data, showing that our proposed approach improves the testing power substantially compared to existing methods.
BibTeX
@inproceedings{li2025power,
  author = {Danning Li and Lingzhou Xue and Haoyi Yang and Xiufan Yu},
  title = {{Power-Enhanced Two-Sample Mean Tests for High-Dimensional Compositional Data with Application to Microbiome Data Analysis}},
  booktitle = {Biometrics},
  year = {2025},
  doi = {10.1093/biomtc/ujaf034},
}
Figure