Jackknife empirical likelihood methods for Gini correlations and their equality testing

Yongli Sang, Xin Dang, Yichuan ZhaoJournal of Statistical Planning and Inference 2018 · DOI · arXiv

Abstract

Gini correlations provide a robust measure of dependence, particularly suited for heavy-tailed distributions where traditional Pearson correlation may falter. However, the asymmetric nature of Gini correlations and their representation as ratios of U-statistics make standard error estimation mathematically intensive. This paper proposes a Jackknife Empirical Likelihood (JEL) framework to facilitate statistical inference. By leveraging pseudo-values and proving a version of Wilks’ Theorem for Gini-based functionals, the authors provide a method for constructing accurate confidence intervals and testing symmetry without explicit variance estimation.

Introduction

In statistical analysis, Pearson’s correlation is the standard for measuring linear dependence but is notoriously sensitive to outliers and heavy-tailed data. The Gini correlation serves as a robust alternative, capturing dependencies that Pearson might miss. A distinctive feature of the Gini correlation is its asymmetry—the correlation of $X$ given $Y$ is not necessarily equal to the correlation of $Y$ given $X$. Testing this equality is crucial in fields like economics for assessing exchangeability and portfolio risk. Historically, the complexity of calculating the asymptotic variance for these "ratio of U-statistics" functionals has hindered their widespread application.

Methodology

The authors address these challenges using Jackknife Empirical Likelihood (JEL). Empirical Likelihood (EL) is a powerful nonparametric approach that allows for likelihood-based inference without assuming a specific underlying distribution. To handle the non-linear constraints inherent in Gini correlations, the authors incorporate the Jackknife method, which generates pseudo-values by systematically omitting data points.

The core innovation lies in treating the Gini correlation as a functional of multiple U-statistics. The authors established that the resulting likelihood ratio statistic follows a standard chi-square distribution (Wilks’ Theorem), simplifying the calculation of p-values and confidence intervals. For small sample sizes (e.g., $n=20$), an Adjusted JEL (AJEL) was introduced to further refine coverage accuracy.

Simulation Results

The performance of JEL and AJEL was evaluated against traditional methods across various distributions. In bivariate normal settings, JEL methods demonstrated coverage probabilities near the nominal levels (90% and 95%) with significantly shorter interval lengths compared to standard Jackknife-only (VJ) methods.

Table 1: Coverage probabilities and interval lengths for bivariate normal distributions.

Robustness to Heavy Tails

The advantages of the Gini-JEL approach are most pronounced in heavy-tailed scenarios, such as the $t$-distribution. While Pearson-based intervals ($\rho_p$) often exhibit over-coverage and excessively large lengths in these cases, the JEL-based Gini intervals remain stable and precise.

Table 2: Performance comparison under heavy-tailed distributions, highlighting the stability of JEL methods.

Real-World Application: Banknote Authentication

The methodology was applied to the UCI Banknote Authentication Dataset. By analyzing features like variance, skewness, and kurtosis of banknote images, the authors used Gini correlations to distinguish between genuine and forged notes. The results indicated that the Gini correlations between features were significantly different for the two classes, and the lack of symmetry ($\gamma(X, Y) \neq \gamma(Y, X)$) served as a key indicator of non-exchangeability in the data structure.

Conclusion

The proposed JEL and AJEL methods provide a powerful, nonparametric toolkit for Gini correlation inference. By bypassing the need for complex variance derivations, these methods offer a practical solution for robust dependence testing. Future research may focus on relaxing moment assumptions and extending the framework to high-dimensional data.

Summary updated

BibTeX
@article{sang2018jackknife,
  author = {Yongli Sang and Xin Dang and Yichuan Zhao},
  title = {{Jackknife empirical likelihood methods for Gini correlations and their equality testing}},
  journal = {Journal of Statistical Planning and Inference},
  year = {2018},
  doi = {10.1016/j.jspi.2018.05.004},
}
Figure