Confidence intervals for Kendall's tau

A simulation study was conducted to examine the performance of several confidence intervals (CIs) for Kendall's tau (t(xy)) under a variety of population conditions. Two normal population variables (N = 10,000) were transformed to have tau correlations, tau = 0,.19,.41, or .71. Samples (n = 10, 50, 200) were drawn from the transformed populations 2000 times under each level of correlation, and accompanying CIs were computed on each sample. The results show that the CI for tau based on a consistent estimate of the variance of t(xy) has the best coverage and power among a number of alternatives. Kendall's t(xy) is unaffected by non-normality induced by monotonic transformations and with its consistent variance estimated from the sample, performs well under a wide range of conditions.