97
52

On the power of Chatterjee rank correlation

Abstract

Recently, Chatterjee (2020) introduced a new rank correlation that attracts many statisticians' attention. This paper compares it to three already well-used rank correlations in literature, Hoeffding's DD, Blum-Kiefer-Rosenblatt's RR, and Bergsma-Dassios-Yanagimoto's τ\tau^*. Three criteria are considered: (i) computational efficiency, (ii) consistency against fixed alternatives, and (iii) power against local alternatives. Our main results show the unfortunate rate sub-optimality of Chatterjee's rank correlation against three popular local alternatives in independence testing literature. Along with some recent computational breakthroughs, they favor the other three in many settings.

View on arXiv
Comments on this paper