Quotient correlation: A sample based alternative to Pearson’s correlation

Zhengjun ZhangThe Annals of Statistics 2008 · DOI

TL;DR

This paper introduces "quotient correlation" as a novel statistical measure and test for detecting nonlinear and tail dependence between variables, especially when traditional Pearson correlation fails. A key innovation is a new "gamma test" for independence with a limiting gamma distribution, offering improved power in certain scenarios and a new tool for analyzing extreme events.

What Problem Does This Paper Address?

Measuring the relationship between two variables is fundamental in statistics. Pearson's correlation coefficient is widely used for linear relationships, but it's often insufficient when dealing with nonlinear dependencies or when the "tail behavior" of the data (i.e., extreme values) is crucial. Traditional methods for testing independence, like Fisher's Z-transformation, can struggle with these complex relationships. Furthermore, in extreme value theory, there's a lack of robust correlation-like measures and efficient tests for tail independence, which is critical for understanding phenomena like financial crashes or extreme weather events. This paper aims to fill these gaps by proposing new correlation measures and a powerful statistical test.

Methods

The core contribution is the introduction of quotient correlation, defined as a ratio of maximum quotients of paired observations. This measure is designed to be more sensitive to nonlinear dependencies and tail behavior than Pearson's correlation.

Key methodological advancements include:

  • Quotient Correlation ($q_n$): A new statistic calculated from sample quotients (e.g., $Y_i/X_i$ and $X_i/Y_i$) that captures the agreement in the magnitudes of two variables.
  • Rank-Based Quotient Correlation ($q_{R,n}$): An adaptation of quotient correlation that uses ranks, making it analogous to Spearman's rank correlation.
  • Gamma Test for Independence: A novel hypothesis test for independence where the test statistic follows a limiting gamma distribution. This is a departure from traditional tests that rely on normal or chi-squared distributions. A crucial theoretical result is that this gamma distribution holds even when using estimated marginal distributions, making the test robust to practical data transformations.
  • Tail Quotient Correlation ($q_{u,n}$): An extension of quotient correlation specifically designed to measure tail dependence by focusing on values above a certain threshold.
  • Gamma Test for Tail Independence: A new test statistic derived from the tail quotient correlation, which also follows a gamma distribution, specifically tailored to detect tail independence.

Datasets

The paper uses simulated data for evaluation. Specific examples include:

  • Bivariate random samples from independent unit Fréchet random variables.
  • Bivariate random samples from Gumbel and survival Gumbel copulas.
  • Bivariate random samples from Example 2.1 (multivariate maxima).
  • Bivariate random samples from a uniform distribution.
  • Bivariate normal random samples.
  • Bivariate random samples from a specific construction involving exponential and Fréchet variables.
  • Bivariate random samples from Student's t-distribution.

A real-world dataset of internet traffic variables (size of response, time duration of response, and throughput) from the University of North Carolina is also analyzed.

Key Insights and Ideas

  • Beyond Linearity: Pearson's correlation is blind to many forms of nonlinear dependence. Quotient correlation offers a way to quantify these relationships, especially when the relative magnitudes of variables matter.
  • The Power of Tails: Extreme events (tail behavior) are often the most critical in risk management and natural phenomena. Quotient correlation and its tail-specific version provide tools to analyze these extreme dependencies, where traditional methods often fall short.
  • A New Distribution for Testing: The gamma distribution for the new test statistic is a significant theoretical contribution. It offers a different statistical framework for hypothesis testing, potentially providing more power in specific scenarios where other tests fail.
  • Robustness to Data Transformation: The gamma test's limiting distribution is invariant to the choice of marginal distribution estimation, making it practical for real-world data where true distributions are often unknown and must be estimated.
  • Bridging Extreme Value Theory and Correlation: The paper introduces concepts that can serve as analogues to correlation coefficients in extreme value theory, a field that has historically lacked such measures.

Key Results

  • The gamma test demonstrates superior power compared to Fisher's Z-transformation test when detecting nonlinear dependence, particularly in cases where Pearson's correlation is zero but a nonlinear relationship exists (e.g., $Y=X^2$).
  • The gamma test exhibits a low Type I error rate, indicating its efficiency in correctly retaining the null hypothesis of independence when it holds.
  • The tail quotient correlation and its associated gamma test are shown to effectively detect tail dependencies in simulated examples, outperforming traditional methods in identifying extreme co-movements.
  • An application to internet traffic data revealed tail dependence between response size and throughput, a finding not previously reported in the literature, highlighting the practical utility of the new methods.
  • The power of the gamma test for tail independence was found to be high (88%) across various simulated scenarios.

Limitations and Future Work

  • Smoothness of Power Curves: The authors note that power curves in their simulations are not always smooth and may not be perfectly accurate, suggesting the need for more refined estimation methods (e.g., monotone smoothing).
  • Generalization of Tail Dependence Measures: While the paper introduces tail quotient correlation, its properties as a tail dependence measure require further study. A generalization based on max-stable processes is suggested for future investigation.
  • Practical Implementation of Thresholds: The choice of the threshold value ($u$) for tail analysis is a parameter that can be explored further. The paper suggests that while increasing $u$ can increase the test statistic, the cutoff value for rejection also changes, indicating a complex interplay.
  • Broader Applications: The authors believe the methodology can be applied to many areas beyond internet traffic, including financial time series analysis, environmental science, and geophysics, suggesting these as avenues for future research.
  • Comparison with Existing Tail Dependence Measures: The paper acknowledges that other tail dependence measures exist and suggests that further research could explore how the proposed measures compare and complement these existing ones.

Summary updated

Abstract

The quotient correlation is defined here as an alternative to Pearson’s correlation that is more intuitive and flexible in cases where the tail behavior of data is important. It measures nonlinear dependence where the regular correlation coefficient is generally not applicable. One of its most useful features is a test statistic that has high power when testing nonlinear dependence in cases where the Fisher’s Z-transformation test may fail to reach a right conclusion. Unlike most asymptotic test statistics, which are either normal or χ2, this test statistic has a limiting gamma distribution (henceforth, the gamma test statistic). More than the common usages of correlation, the quotient correlation can easily and intuitively be adjusted to values at tails. This adjustment generates two new concepts—the tail quotient correlation and the tail independence test statistics, which are also gamma statistics. Due to the fact that there is no analogue of the correlation coefficient in extreme value theory, and there does not exist an efficient tail independence test statistic, these two new concepts may open up a new field of study. In addition, an alternative to Spearman’s rank correlation, a rank based quotient correlation, is also defined. The advantages of using these new concepts are illustrated with simulated data and a real data analysis of internet traffic.
BibTeX
@article{zhang2008quotient,
  author = {Zhengjun Zhang},
  title = {{Quotient correlation: A sample based alternative to Pearson’s correlation}},
  journal = {The Annals of Statistics},
  year = {2008},
  doi = {10.1214/009053607000000866},
}
Figure