Measures of dependence between random vectors and tests of independence.\n Literature review

Julie Josse, Susan HolmesarXiv 2013 · DOI · arXiv

Abstract

Simple correlation coefficients between two variables have been generalized to measure association between two matrices in many ways. Coefficients such as the RV coefficient, the distance covariance (dCov) coefficient and kernel based coefficients have been adopted by different research communities. Scientists use these coefficients to test whether two random vectors are linked. If they are, it is important to uncover what patterns exist in these associations. We discuss the topic of measures of dependence between random vectors and tests of independence and show links between different approaches. We document some of the interesting rediscoveries and lack of interconnection between bodies of literature. After providing definitions of the coefficients and associated tests, we present the recent improvements that enhance their statistical properties and ease of interpretation. We summarize multi-table approaches and provide scenarii where the indices can provide useful summaries of heterogeneous multi-block data. We illustrate these different strategies on several examples of real data and suggest directions for future research.

Introduction

In modern data science, analysis rarely focuses on a single variable. Instead, researchers often encounter "multi-block" data—sets of variables grouped into matrices, such as genetic expression data paired with clinical symptoms. Measuring the association between these entire tables is a fundamental challenge. Historically, different research communities—ranging from ecologists to machine learners—have developed their own tools for this task, often in isolation. This paper serves as a "Rosetta Stone," linking these disparate methods and providing a comprehensive roadmap for measuring dependencies between random vectors.

Methodological Framework

The authors categorize the primary measures of dependence into three main classes, highlighting their mathematical interconnections:

The RV Coefficient

A generalization of the Pearson correlation for matrices, the RV coefficient measures the similarity between the internal structures (cross-product matrices) of two datasets. While robust, it is strictly a linear measure and may fail to detect more complex relationships.

Distance Covariance (dCov)

A more recent innovation, dCov utilizes Euclidean distances between observations. Its primary advantage is its ability to detect non-linear patterns; it equals zero if and only if the two sets of variables are truly independent. Interestingly, the authors demonstrate that dCov with an exponent of $\alpha=2$ is mathematically equivalent to the RV coefficient, positioning dCov as a more flexible evolution of the same underlying logic.

Kernel and Graph-based Measures

The paper links these statistical measures to machine learning via the Hilbert-Schmidt Independence Criterion (HSIC). HSIC is essentially a kernelized version of the logic used in dCov and RV. The authors also explore graph-based approaches, such as $k$-nearest neighbor graphs, which can be particularly effective for complex, high-dimensional data.

The High-Dimensional Trap and Debiasing

A critical contribution of this work is the emphasis on "double centering"—a mathematical transformation ensuring distance matrices are statistically valid for comparison. Furthermore, the authors address the "High-Dimensional Trap": in settings where variables outnumber samples, association coefficients naturally drift toward 1.0, even in purely random data. To combat this, the authors advocate for "Modified" or "Adjusted" RV coefficients that subtract the expected value under the null hypothesis, providing a more accurate reflection of true association.

Experimental Evaluation

The effectiveness of these methods was evaluated across several real-world datasets:

  • Sensory Analysis: "Napping" data where panelists categorized wines based on taste similarity.
  • Microarray Data: A study of 43 brain tumors described by transcriptome (356 variables) and genome (76 variables) data.
  • Morphology: Landmark data from Macaca monkey skulls used to study growth patterns.
  • Chemometry: Spectroscopic measurements of beers to verify authenticity.

Key Results

In simulations, while the RV test showed slightly more power for purely linear data, the dCov test was vastly superior at detecting non-linear (e.g., logarithmic) dependencies where the RV test failed entirely. In the high-dimensional brain tumor example, the raw RV coefficient was 0.34, but debiasing revealed a much lower "true" association of 0.18, illustrating the risk of overstating results without proper adjustment.

Discussion and Insights

  • Significance vs. Pattern: A significant p-value confirms an association exists but does not describe its nature. The authors argue that independence tests must be followed by visualization techniques like STATIS or Multiple Factor Analysis (MFA) to map the relationship.
  • Critique of the Mantel Test: Widely used in ecology, the Mantel test is criticized for lacking double centering, which can lead to high false-positive rates and a failure to detect certain linear relationships.
  • Computational Efficiency: Because permutation tests are computationally expensive, the authors suggest moment-matching approximations (like the Pearson Type III distribution) as faster alternatives for large-scale data.

Conclusion and Future Work

While dCov and its relatives are powerful for detecting dependencies, they often act as "black boxes" regarding the specific variables driving the association. Future research is needed to improve the interpretability of non-linear links and to create a unified framework that seamlessly handles mixed-type (continuous and categorical) data. Nonetheless, the integration of these cross-disciplinary tools provides a more rigorous foundation for multi-block data analysis.

Summary updated

Abstract

Simple correlation coefficients between two variables have been generalized\nto measure association between two matrices in many ways. Coefficients such as\nthe RV coefficient, the distance covariance (dCov) coefficient and kernel based\ncoefficients have been adopted by different research communities. Scientists\nuse these coefficients to test whether two random vectors are linked. If they\nare, it is important to uncover what patterns exist in these associations.\n We discuss the topic of measures of dependence between random vectors and\ntests of independence and show links between different approaches. We document\nsome of the interesting rediscoveries and lack of interconnection between\nbodies of literature. After providing definitions of the coefficients and\nassociated tests, we present the recent improvements that enhance their\nstatistical properties and ease of interpretation. We summarize multi-table\napproaches and provide scenarii where the indices can provide useful summaries\nof heterogeneous multi-block data.\n We illustrate these different strategies on several examples of real data and\nsuggest directions for future research.\n
BibTeX
@misc{josse2013measures,
  author = {Julie Josse and Susan Holmes},
  title = {{Measures of dependence between random vectors and tests of independence.\textbackslash\{\}n Literature review}},
  year = {2013},
  eprint = {1307.7383},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1307.7383},
}
Figure