Set-based differential covariance testing for high-throughput data

TL;DR

This paper introduces a flexible, high-dimensional framework for testing the association between sets of features' covariance matrices and experimental variables, applicable to both discrete and continuous data.

Contribution

It proposes a novel, general framework with four summary statistics, including a new connectivity statistic, for covariance testing in high-dimensional settings, extending beyond two-sample problems.

Findings

The approach is applicable when p >> n.

Asymptotic p-values are derived for several statistics.

The methods show improved power in simulations and real data analysis.

Abstract

The problem of detecting changes in covariance for a single pair of features has been studied in some detail, but may be limited in importance or general applicability. In contrast, testing equality of covariance matrices of a {\it set} of features may offer increased power and interpretability. Such approaches have received increasing attention in recent years, especially in the context of high-dimensional testing. These approaches have been limited to the two-sample problem and involve varying assumptions on the number of features $p$ vs. the sample size $n$. In addition, there has been little discussion of the motivating principles underlying various choices of statistic, and no general approaches to test association of covariances with a continuous outcome. We propose a uniform framework to test association of covariance matrices with an experimental variable, whether discrete or continuous. We describe four different summary statistics, to ensure power and flexibility under various settings, including a new "connectivity" statistic that is sensitive to changes in overall covariance magnitude. The approach is not limited by the data dimensions, and is applicable to situations where $p >> n$. For several statistics we obtain asymptotic $p$-values under relatively mild conditions. For the two-sample special case, we show that the proposed statistics are permutationally equivalent or similar to existing proposed statistics. We demonstrate the power and utility of our approaches via simulation and analysis of real data.