False Discovery Rate Control for High-Dimensional Networks of Quantile Associations Conditioning on Covariates

TL;DR

This paper introduces a novel quantile-based statistical method for high-dimensional network analysis that controls the false discovery rate and is computationally efficient, with applications in gene co-expression networks related to cancer.

Contribution

It proposes a new sample quantile-based contingency statistic and a multiple testing procedure that handles high-dimensional data with complex associations, extending to cases where outcomes and covariates exceed sample size.

Findings

The method asymptotically controls the false discovery rate.

It outperforms existing methods in complex association scenarios.

Successfully applied to gene co-expression networks in gastric cancer.

Abstract

Motivated by the gene co-expression pattern analysis, we propose a novel sample quantile-based contingency (squac) statistic to infer quantile associations conditioning on covariates. It features enhanced flexibility in handling variables with both arbitrary distributions and complex association patterns conditioning on covariates. We first derive its asymptotic null distribution, and then develop a multiple testing procedure based on squac to simultaneously test the independence between one pair of variables conditioning on covariates for all $p(p-1)/2$ pairs. Here, $p$ is the length of the outcomes and could exceed the sample size. The testing procedure does not require resampling or perturbation, and thus is computationally efficient. We prove by theory and numerical experiments that this testing method asymptotically controls the false discovery rate (\FDR). It outperforms all alternative methods when the complex association panterns exist. Applied to a gastric cancer data, this testing method successfully inferred the gene co-expression networks of early and late stage patients. It identified more changes in the networks which are associated with cancer survivals. We extend our method to the case that both the length of the outcomes and the length of covariates exceed the sample size, and show that the asymptotic theory still holds.