PEtests: Power-Enhanced (PE) Tests for High-Dimensional Data

Xiufan Yu, Danning Li, Lingzhou Xue, Runze LiOpenalex Unknown 2023 · DOI · OpenAlex

Abstract

Two-sample power-enhanced mean tests, covariance tests, and simultaneous tests on mean vectors and covariance matrices for high-dimensional data. Methods of these PE tests are presented in Yu, Li, and Xue (2022) doi:10.1080/01621459.2022.2126781; Yu, Li, Xue, and Li (2022) doi:10.1080/01621459.2022.2061354.

Introduction

In modern data science—spanning fields from genomics to high-frequency finance—researchers frequently encounter "high-dimensional" datasets where the number of features ($p$) significantly exceeds the sample size ($n$). Traditional statistical tests, such as Hotelling’s $T^2$, are inapplicable in these settings because they require the inversion of covariance matrices, which become singular when $p > n$.

Existing high-dimensional alternatives generally target specific signal structures: $l_2$-norm tests are optimized for "dense" signals (where many variables exhibit small changes), while $l_\infty$-norm tests excel at "sparse" signals (where only a few variables change significantly). Selecting the wrong test for the data's underlying structure leads to a substantial loss in statistical power. The PEtests R package addresses this challenge by implementing "Power-Enhanced" (PE) tests that maintain high sensitivity across sparse, dense, and intermediate signal regimes.

Methodology and the PE Component

The PEtests package implements several innovative approaches based on the research of Yu, Li, Xue, and Li (2022). The core methodology bridges different testing philosophies:

  1. Standard High-Dimensional Tests: The package includes established $l_2$-norm tests (e.g., Chen & Qin, 2010) and $l_\infty$-norm tests (e.g., Cai et al., 2014).
  2. P-value Aggregation: Utilizing Cauchy and Fisher combinations, the package aggregates p-values from different test types. Since $l_2$ and $l_\infty$ statistics are often asymptotically independent in high dimensions, their combination creates a robust test sensitive to diverse alternatives.
  3. Power-Enhancement (PE) Construction: The most significant innovation is the pe.comp method. A "Power-Enhancement" component ($J$) is constructed via a screening procedure and added to a standard $l_2$ test statistic. If no sparse signal exists, $J$ converges to zero, preserving the original test's validity. If a sparse signal is present, $J$ boosts the statistic, enhancing the power to reject the null hypothesis.
Parameters for the PE component construction, including the thresholding value delta.

As shown in the parameter definitions above, the PE component relies on a thresholding value, $\delta$. The package provides a data-driven default based on the log of the sample size and dimension to effectively separate signal from high-dimensional noise.

Key Insights

  • Asymptotic Independence: A theoretical pillar of the package is that in high-dimensional spaces, the "average" difference (captured by $l_2$) and the "maximum" difference (captured by $l_\infty$) provide non-overlapping information.
  • No-Free-Lunch Mitigation: While improving a test's sensitivity to one signal type often degrades it for another, the PE construction aims to be "power-enhanced" against sparse alternatives while remaining "power-consistent" against dense ones.
  • Simultaneous Inference: The simultest functions allow researchers to test for changes in both the mean (center) and covariance (shape) of the data distribution simultaneously, providing a more holistic view of group differences.

Results and Implementation

The package provides a unified interface for 16 different testing functions. Key technical features include:

  • Distributional Convergence: Under the null hypothesis, statistics converge to well-known distributions (Standard Normal, Gumbel, Cauchy, or Chi-squared), allowing for accurate asymptotic p-values without expensive permutations.
  • Application in Gene-Set Testing: While demonstrated with simulated data, these methods are primarily designed for real-world applications like gene-set testing, where researchers compare expression levels and interactions across thousands of genes.

Limitations

The validity of these tests relies on regularity conditions regarding the tail behavior and dependency structures of the data. Extremely heavy-tailed data or complex internal correlations may affect p-value accuracy. Additionally, while the package provides sensible defaults for the thresholding parameter $\delta$, performance may vary depending on this choice in specific practical scenarios.

Summary updated

Abstract

Two-sample power-enhanced mean tests, covariance tests, and simultaneous tests on mean vectors and covariance matrices for high-dimensional data. Methods of these PE tests are presented in Yu, Li, and Xue (2022) &lt;<a href="https://doi.org/10.1080%2F01621459.2022.2126781" target="_top">doi:10.1080/01621459.2022.2126781</a>&gt;; Yu, Li, Xue, and Li (2022) &lt;<a href="https://doi.org/10.1080%2F01621459.2022.2061354" target="_top">doi:10.1080/01621459.2022.2061354</a>&gt;.
BibTeX
@inproceedings{yu2023petests,
  author = {Xiufan Yu and Danning Li and Lingzhou Xue and Runze Li},
  title = {{PEtests: Power-Enhanced (PE) Tests for High-Dimensional Data}},
  booktitle = {Openalex Unknown},
  year = {2023},
  doi = {10.32614/cran.package.petests},
}
Figure