Equivalence Test for Correlated Bivariate Binary Observation

TL;DR

This paper introduces a new equivalence test called the margin test for correlated bivariate binary data, comparing it to the McNemar test, and demonstrates its advantages in terms of acceptance region size and power.

Contribution

The paper develops a novel margin test based on the joint distribution of discordant observations, providing a more accurate null hypothesis testing method for correlated binary data.

Findings

The margin test has a larger acceptance region than the McNemar test.

The size and power of the margin test are favorable under certain conditions.

Real-world examples illustrate different decision outcomes between the tests.

Abstract

Under the null hypothesis, the marginal probability of the positive response is symmetric at any specified correlated coefficient, and the discordance probability is also symmetric to the positive response probability. The marginal distribution function of the discordant observation is monotonically decreasing with the increase of the discordance probability.And the minimum point of the distribution function is determined by the correlated coefficient.Based on the joint distribution of the two discordant observations, a confidence region of the possible values of two discordant variables is proposed, which deduces an equivalence test with the marginal distribution of the discordance observation, called the margin test.For a specified level of significance, the acceptance region of the McNemar test compares to the corresponding domain of the margin test for different sample sizes. The shape of the two kinds of acceptance regions is similar, except that the acceptance region of the margin test is slightly larger than the corresponding region of the McNemar test. The size and power of the McNemar test compare to the corresponding values of the margin test at a specified level of significance for different sample sizes. The risk of the type I error in both methods increases for a larger sample size or a smaller correlation coefficient, and the range of parameters where the margin test correctly accepts the null hypothesis is significantly larger than the corresponding range of the McNemar test.Two real-world examples demonstrate how to understand the different decisions from the McNemar test and the margin test, where the observed data is on the boundary of the rejection regions.