Kendall's Tau

Kendall's tau, written with the Greek letter tau, is a statistic that measures how much two variables agree in their ordering. Rather than looking at the raw values, it works with ranks, and it does so by examining pairs of observations. Take any two data points and compare them on both variables. If the point that ranks higher on the first variable also ranks higher on the second, the pair agrees, and it is called concordant. If the point that ranks higher on the first ranks lower on the second, the pair disagrees, and it is called discordant. Kendall's tau is essentially the balance of these: concordant pairs minus discordant pairs, scaled so the result runs from minus one to plus one. A tau of plus one means the two orderings are identical, minus one means they are exact reverses, and zero means agreements and disagreements roughly cancel out.

This construction gives Kendall's tau an unusually clean interpretation. A positive tau means that, across all the pairs, agreements outnumber disagreements, so knowing that one point outranks another on the first variable makes it more likely to outrank it on the second as well. That is a statement about probability of concordance, which is easier to reason about than many correlation figures. Because tau depends only on the direction of each pairwise comparison, not on the size of the gaps, it is a measure of monotonic association, whether the variables move together in order, not whether they follow a straight line. It is also robust: outliers that would swing a value-based measure barely move tau, since flipping one point's rank changes only the pairs it belongs to.

Kendall's tau is one of three correlation coefficients people commonly reach for, and the differences matter. Pearson's correlation measures linear association between the actual values of two variables, and it assumes the relationship is roughly straight-line and the data reasonably well behaved. Spearman's rho and Kendall's tau are both rank correlations, so they measure monotonic association, whether one variable tends to rise as the other does, without assuming linearity, which makes them more robust to outliers and to non-straight relationships than Pearson. The first fork is therefore between Pearson, which uses values and assumes linearity, and the two rank measures, which use order and do not.

Between the two rank measures, Spearman and Kendall differ in how they compute agreement. Spearman's rho is, in effect, Pearson's correlation applied to the ranks, built from the squared differences in rank. Kendall's tau is built instead from counting concordant and discordant pairs. In practice, Kendall's tau usually comes out numerically smaller than Spearman's rho on the same data, so the two are not interchangeable numbers even when they agree in direction. Kendall's tau tends to handle tied ranks more gracefully and has a more direct probabilistic reading, the probability of concordance minus the probability of discordance, while Spearman is more widely known and quicker to compute by hand. Neither is simply better; they answer the same question, monotonic agreement of ranks, through different arithmetic, and the right choice depends on the data and what interpretation you want.

Using Kendall's tau well means reaching for it when you care about the agreement of orderings rather than a straight-line relationship between values. It suits ordinal data, ranked lists, and situations with outliers or non-linear but monotonic relationships, where Pearson would be misleading. Interpret the sign and size together: a positive tau means orderings tend to agree, a negative tau means they tend to reverse, and the magnitude, closer to one, means stronger agreement. Lean on its probabilistic meaning, since tau reads naturally as how much more likely two points are to agree than disagree in order. And remember it measures monotonic association only, so a strong non-monotonic pattern, like a U-shape, can produce a tau near zero even when the variables are clearly related.

The failures come from misapplying or misreading it. Treating Kendall's tau as if it were Pearson, and expecting it to capture the strength of a linear relationship or to match Pearson's number, confuses two different measures. Comparing a tau directly against a Spearman rho as if they were on the same scale misleads, because tau is typically the smaller of the two on identical data. Reading a tau near zero as proof of no relationship ignores that a non-monotonic pattern can hide from any rank correlation. And using tau on data where the values, not just the ranks, carry the meaning throws away information a value-based measure would use. The discipline is to match the coefficient to the question: Kendall's tau for robust, interpretable agreement of orderings, and to read its number in its own terms.

Kendall's tau — a rank correlation coefficient built from concordant and discordant pairs — measures monotonic agreement of orderings, distinct from Spearman's rho and Pearson's linear correlation.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where kendall's tau is a core concern: