Kendall’s tau: rank correlation
Kendall’s tau describes whether pairs of people have the same order on two measurements, with tau-b accounting for ties.
Fit
When to use it, and when not to
Use it when
- Each independent person has two measurements with a meaningful order.
- You want rank association based on concordant and discordant pairs, including tied ratings.
Not when
- You want a straight-line association in the original values. Use Pearson correlation. Read: Pearson correlation: linear association
- You want correlation of the ranks themselves. Use Spearman correlation. Read: Spearman correlation: rank association
- You want agreement between raters on an ordered scale. Use weighted kappa. Read: Weighted kappa: agreement on an ordered scale
Before you start
What you need
- Paired measurements and their orderKeep each person’s two values together. State which direction means more.
- Tau variant and tied valuesUse tau-b when ties need adjustment. Tau-a uses all pairs in its denominator without that adjustment.
- Null hypothesis and interval methodState whether testing independence or zero tau. An exact permutation test for independence need not test only zero association.
- People contribute independent observation pairs.How to check: Use a cluster-aware method for repeated visits or related people.
- Both scales have a meaningful order and each varies.How to check: Check category coding. Tau-b is undefined when either variable is constant.
- A trend in one direction is the association of interest.How to check: Plot the pairs. A trend that reverses direction can cancel in a single coefficient.
Plain language
How it works
Compare every pair of people. A pair is concordant when both measurements put them in the same order, and discordant when the orders oppose. Tau-b divides concordant minus discordant counts by a denominator adjusted for ties. With no ties, it is the difference between those two pair proportions.
Tau-b uses the square root of two factors: concordant plus discordant plus ties only on one variable, one factor for each variable. Pairs tied on both add to neither factor. Hutson and Yu distinguish independence tests from tests of zero ordinal association. Shih and Fay develop rank inference for clustered pairs. Neither rank scoring nor a software label makes inference assumption-free.
Worked example
A worked example
- Setting
- Five independent people have two ordered scores.
- Question
- Do higher scores on one scale tend to go with higher scores on the other?
- Approach
- x values: 1, 2, 2, 3, 4. y values: 1, 3, 2, 2, 4.
- Result
- Across 10 pairs, 7 are concordant, 1 is discordant, 1 ties only on x, and 1 ties only on y. Tau-b is 0.667.
- What it means
- The ordering tends to agree in this sample. This describes rank association, not whether the two scales give interchangeable scores.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Kendall’s tau measures rank association, not agreement or cause. State the tau variant and tie handling; inference must respect clustering and the null hypothesis.
- Reading tau as a percentage of agreement.Instead: Report tau-b and the tie pattern. Even equal ordering does not imply equal measured values.
- Using a no-ties denominator when the scales have ties.Instead: Name the variant and use its tie adjustment.
- Calling every exact test valid for a zero-tau null.Instead: Check the null and the sampling design. Permuting labels requires the relevant exchangeability assumption.
Inside Mallard
How Mallard uses this
Mallard offers this article through Learn search, related topics and the Atlas. It has no dedicated Kendall calculator or direct plan cue.
analysis.kendalls_tau, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Exact inference around ordinal measures of association is often not exact (opens in a new tab) · Comput Methods Programs Biomed · doi 10.1016/j.cmpb.2023.107725
- Pearson's chi-square test and rank correlation inferences for clustered data (opens in a new tab) · Biometrics · doi 10.1111/biom.12653