Simpson’s Paradox for Kendall’s Rank Coefficient
摘要
This note revisits Simpson’s paradox and discusses confounding effects of hidden covariates on Kendall’s tau. As a result, observed correlation may vanish or even revert. More specifically, a decomposition of Kendall’s tau in the presence of subgroups is established, a formal definition of Simpson’s paradox for Kendall’s tau is given and some simple examples of paradoxical situations in the insurance domain are provided. Finally, necessary and sufficient conditions for a Simpson’s paradox to occur are studied.