The Binary Expansion Randomized Ensemble Test (BERET)

Duyeol Lee, Kai Zhang, Michael R. KosorokarXiv 2019 · DOI · arXiv

Abstract

Recently, the binary expansion testing framework was introduced to test the independence of two continuous random variables by utilizing symmetry statistics that are complete sufficient statistics for dependence. We develop a new test based on an ensemble approach that uses the sum of squared symmetry statistics and distance correlation. Simulation studies suggest that this method improves the power while preserving the clear interpretation of the binary expansion testing. We extend this method to tests of independence of random vectors in arbitrary dimension. Through random projections, the proposed binary expansion randomized ensemble test transforms the multivariate independence testing problem into a univariate problem. Simulation studies and data example analyses show that the proposed method provides relatively robust performance compared with existing methods.

Introduction

Determining whether two variables are independent is a foundational challenge in statistical inference. While classical methods, such as Pearson’s correlation, are highly effective at identifying linear relationships, they often fail to detect complex nonlinear patterns like circles, sine waves, or checkerboards. Modern nonparametric tests have been developed to address these limitations, but they frequently suffer from two primary issues: nonuniform consistency and the "black box" problem.

Nonuniform consistency refers to a test's tendency to be exceptionally powerful for certain data structures while remaining surprisingly weak for others (e.g., a test that catches a checkerboard pattern but misses a simple linear one). The "black box" problem refers to the lack of interpretability; most multivariate tests provide a p-value indicating that a relationship exists, but offer little insight into the nature of that relationship. The Binary Expansion Randomized Ensemble Test (BERET) is designed to overcome these hurdles by bridging the gap between linear and complex nonlinear dependency detection.

Methodology

BERET enhances the existing Binary Expansion Testing (BET) framework through three core components:

1. Binary Expansion (The "Bit" Approach)

Continuous variables are transformed into binary expansions, effectively looking at the "bits" of a number. By examining the interactions between these bits, the test can identify dependencies at various scales—from broad global trends to intricate local micro-patterns. This allows for a multi-resolution view of the data.

2. The Ensemble Strategy

To ensure the test does not lose power on simple linear relationships while searching for complex ones, BERET combines the sum of squared symmetry statistics with Distance Correlation (dCor). This ensemble approach allows the test to remain sensitive to both simple and highly complex patterns, providing a "best-of-both-worlds" statistical tool.

3. Random Projections for Multivariate Data

Testing independence between high-dimensional vectors is computationally and statistically difficult. BERET addresses this by using one-dimensional random projections. By projecting multivariate vectors onto random lines, the framework transforms a complex high-dimensional problem into a series of simpler univariate problems, which are then aggregated to form the final test statistic.

Experimental Evaluation

The authors evaluated BERET using a comprehensive battery of synthetic and real-world datasets:

  • Simulated Scenarios: Testing included challenging patterns such as Linear, Parabolic, Circular, Sine, Checkerboard, Local, Spherical, and a 3D "Double Helix."
  • Real-World Data:
    • WHO Life Expectancy: Analyzing geographic coordinates vs. life expectancy across 189 countries.
  • CIA Mortality Rate: Investigating the relationship between income/birth rates and mortality across 225 regions.
  • UCI House Prices: Mapping geographic location against real estate prices in Taipei.

Key Insights and Results

Superior Statistical Power

In simulation studies, BERET demonstrated remarkably robust performance. It was the only test among those compared that did not exhibit a "power pit" (a scenario where power drops significantly). It consistently ranked in the top tier across all patterns, outperforming established methods like the Hilbert-Schmidt Independence Criterion (HSIC) and various Mutual Information tests.

Interpretability and Symmetry Statistics

Unlike many nonparametric tests, BERET produces "symmetry statistics." These statistics can be mapped back to the original data to visualize exactly where the dependency lies. For example, in the Taipei house price analysis, BERET did not just indicate that location matters; it identified that prices were specifically clustered around the intersection of two major roads in the city center.

Uncovering Hidden Relationships

In the analysis of CIA mortality data, standard linear tests failed to reject the null hypothesis of independence. However, BERET identified a strong quadratic relationship: mortality rates are high in countries with very low birth rates (often due to aging populations) and very high birth rates (often due to public health challenges). This highlights BERET's ability to find meaningful patterns that traditional methods overlook.

Conclusion and Future Work

BERET provides a powerful, interpretable, and robust framework for independence testing. By ensembling binary expansion with distance correlation and utilizing random projections, it maintains high sensitivity across diverse data structures and dimensions.

While the test is currently more computationally intensive than simple correlation coefficients due to its reliance on permutation methods for p-value calculation, it offers a significant step forward in making nonparametric testing more transparent. Future research may focus on refining the random projection approach to automatically identify the most informative directions, further increasing the test's efficiency in high-dimensional spaces.

Summary updated

Abstract

Recently, the binary expansion testing framework was introduced to test the independence of two continuous random variables by utilizing symmetry statistics that are complete sufficient statistics for dependence. We develop a new test based on an ensemble approach that uses the sum of squared symmetry statistics and distance correlation. Simulation studies suggest that this method improves the power while preserving the clear interpretation of the binary expansion testing. We extend this method to tests of independence of random vectors in arbitrary dimension. Through random projections, the proposed binary expansion randomized ensemble test transforms the multivariate independence testing problem into a univariate problem. Simulation studies and data example analyses show that the proposed method provides relatively robust performance compared with existing methods.
BibTeX
@misc{lee2019binary,
  author = {Duyeol Lee and Kai Zhang and Michael R. Kosorok},
  title = {{The Binary Expansion Randomized Ensemble Test (BERET)}},
  year = {2019},
  eprint = {1912.03662},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1912.03662},
}
Figure