A New Non-parametric Test for Multivariate Paired Data and Pair Matching

Jingru Zhang, Hao Chen, Xiao‐Hua ZhouarXiv 2020 · DOI · arXiv

Abstract

In paired design studies, it is common to have multiple measurements taken for the same set of subjects under different conditions. In observational studies, it is many times of interest to conduct pair matching on multiple covariates between a treatment group and a control group, and to test the treatment effect represented by multiple response variables on well pair-matched data. However, there is a lack of an effective test on multivariate paired data. The multivariate paired Hotelling's $T^2$ test can sometimes be used, but its power decreases fast as the dimension increases. Existing methods for assessing the balance of multiple covariates in matched observational studies usually ignore the paired structure and thus they do not perform well under some settings. In this work, we propose a new non-parametric test for paired data, which exhibits a substantial power improvement over existing methods under a wide range of situations. We also derive the asymptotic distribution of the new test and the approximate $p$-value is reasonably accurate under finite samples through simulation studies even when the dimension is larger than the sample size, making the new test an easy-off-the-shelf tool for real applications. The proposed test is illustrated through an analysis of a real data set on the Alzheimer's disease research.

Introduction

In clinical trials and observational studies, researchers frequently encounter "paired data," such as measurements taken from the same patient before and after treatment, or from a treated subject matched with a control subject based on similar characteristics. While standard tools like the paired t-test are effective for univariate data, modern datasets are increasingly multivariate, often involving dozens or hundreds of measurements per subject.

The traditional approach for multivariate paired data, Hotelling’s $T^2$ test, suffers from two significant limitations. First, it faces the "curse of dimensionality": its statistical power diminishes rapidly as the number of variables ($d$) increases relative to the sample size ($n$), and it fails entirely if $d > n$. Second, it relies on the assumption of multivariate normality, which is rarely satisfied in biological or behavioral data. This paper introduces a robust, graph-based non-parametric test designed to overcome these challenges.

Methodology: Graph-Based Non-parametric Testing

The proposed method utilizes a similarity graph, specifically a $k$-Minimum Spanning Tree ($k$-MST), to represent relationships between multivariate observations. In this graph, nodes represent subjects, and edges connect subjects that are "close" to each other in the multivariate space.

Figure 1: Illustration of different edge types in a similarity graph for two pairs (i, i*) and (j, j*). The test focuses on edges between different pairs to detect distributional differences.

The test statistic, $D$, is derived by counting "within-sample" edges—those connecting two observations from the same group. A higher-than-expected number of such edges suggests a significant difference between the groups. Crucially, the authors recognized that edges connecting members of the same pair (e.g., $i$ and $i^*$) provide no information about group differences under the null hypothesis. By focusing on edges between different pairs, the test achieves higher efficiency. To determine significance, a paired-comparison permutation is used, where labels are only swapped within each pair. The authors demonstrate that $D$ follows an asymptotic $\chi^2_2$ distribution, allowing for rapid $p$-value calculation.

Experimental Results and Applications

Simulation Studies

The test was evaluated across various high-dimensional scenarios ($d$ up to 1000) using distributions such as multivariate Normal, $t$, Log-normal, and Laplace. In nearly every case, the $D$ test outperformed Hotelling’s $T^2$ and other graph-based tests that do not account for the paired structure. Notably, the test is sensitive not only to shifts in the mean but also to differences in distribution shape and scale.

Alzheimer’s Disease Research

The practical utility of the test was demonstrated using data from the National Alzheimer’s Coordinating Center (NACC). The study analyzed 22 neuropsychological variables from participants at an initial visit and a five-year follow-up.

Figure 2: P-values for individual neuropsychological variables in the Alzheimer’s disease dataset analysis.

As shown in the table above, several individual variables such as NACCMMSE, DIGIF, and WAIS showed significant changes over time. However, when considering the full multivariate profile for a group with mild dementia, the standard Hotelling’s $T^2$ test failed to find significance ($p = 0.2121$). In contrast, the new $D$ test yielded a $p$-value of 0.0224, successfully identifying the cognitive decline that was obscured by high-dimensional noise in traditional methods.

Conclusion

The proposed non-parametric test offers a powerful, dimension-agnostic alternative for analyzing multivariate paired data. By leveraging similarity graphs and accounting for the inherent paired structure, it provides a robust tool for detecting complex distributional differences. Future work may explore the application of this test to non-Euclidean data types, such as images or network structures, and further optimize graph density parameters for diverse datasets.

Summary updated

Abstract

In paired design studies, it is common to have multiple measurements taken for the same set of subjects under different conditions. In observational studies, it is many times of interest to conduct pair matching on multiple covariates between a treatment group and a control group, and to test the treatment effect represented by multiple response variables on well pair-matched data. However, there is a lack of an effective test on multivariate paired data. The multivariate paired Hotelling's $T^2$ test can sometimes be used, but its power decreases fast as the dimension increases. Existing methods for assessing the balance of multiple covariates in matched observational studies usually ignore the paired structure and thus they do not perform well under some settings. In this work, we propose a new non-parametric test for paired data, which exhibits a substantial power improvement over existing methods under a wide range of situations. We also derive the asymptotic distribution of the new test and the approximate $p$-value is reasonably accurate under finite samples through simulation studies even when the dimension is larger than the sample size, making the new test an easy-off-the-shelf tool for real applications. The proposed test is illustrated through an analysis of a real data set on the Alzheimer's disease research.
BibTeX
@misc{zhang2020new,
  author = {Jingru Zhang and Hao Chen and Xiao‐Hua Zhou},
  title = {{A New Non-parametric Test for Multivariate Paired Data and Pair Matching}},
  year = {2020},
  eprint = {2007.01497},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.01497},
}

Loading related content...

Figure