Test of Bivariate Independence Based on Angular Probability Integral Transform with Emphasis on Circular-Circular and Circular-Linear Data

TL;DR

This paper introduces a novel independence test for circular and circular-linear data based on the angular probability integral transform, utilizing NNTS distributions to evaluate test power against alternatives close to uniformity.

Contribution

It proposes a new independence testing method using the angular probability integral transform and NNTS distributions, emphasizing circular-circular and circular-linear data.

Findings

The test effectively detects independence in circular data.

NNTS distributions provide a flexible framework for power evaluation.

The method performs well against alternatives near the null hypothesis.

Abstract

The probability integral transform (PIT) of a continuous random variable $X$ with distribution function $F_X$ is a uniformly distributed random variable $U=F_X(X)$. We define the angular probability integral transform (APIT) as $\theta_U = 2 \pi U = 2 \pi F_{X}(X)$, which corresponds to a uniformly distributed angle on the unit circle. For circular (angular) random variables, the sum of absolutely continuous independent circular uniform random variables is a circular uniform random variable, that is, the circular uniform distribution is closed under summation, and it is a stable continuous distribution on the unit circle. If we consider the sum (difference) of the angular probability integral transforms of two random variables, $X_1$ and $X_2$, and test for the circular uniformity of their sum (difference), this is equivalent to the test of independence of the original variables. In this study, we used a flexible family of nonnegative trigonometric sums (NNTS) circular distributions, which include the uniform circular distribution as a member of the family, to evaluate the power of the proposed independence test; we complete this evaluation by generating samples from NNTS alternative distributions that may be at a closer proximity with respect to the circular uniform null distribution.