Stuart's τ-c (rectangular ordinal tables)

Coming soon

Advanced & specialized

This test is implemented and is currently going through StatMinds’ production verification: every statistic is independently checked against a trusted reference (scipy / R), locked with regression tests, and the screen is exercised across assumption-met/violated and significant/non-significant scenarios before it opens up.

See what’s live now

Rectangular-table ordinal association.

τ-c = 2m(C − D) / (n²(m − 1)), with m = min(r, c). The key advantage over τ-b: |τ-c| can reach ±1 for ANY r × c ordinal table; τ-b's maximum is bounded below 1 whenever r ≠ c, so reading τ-b on a rectangular table understates the strength of the actual association. Reports τ-c, τ-c-max-attainable (the table-shape ceiling), and the ratio of |τ-c| to that ceiling.

Worked example

How associated are two ordinal variables in a rectangular table?

Kendall's tau-c measures ordinal association and, unlike tau-b, can reach ±1 in non-square tables.

Result

The variables were moderately associated, τc = .34, p = .002.

How you'd report it (APA)

Kendall's tau-c indicated a moderate positive ordinal association, τc = .34, p = .002.

When to use it

  • Two ordinal variables with unequal numbers of categories
    A consumer-research firm cross-tabulates customer satisfaction (1–5 Likert, 5 categories) against intent-to-recommend bracket (Detractor / Passive / Promoter, 3 categories).
  • Ordered severity categories on rectangular grids
    A clinician cross-tabulates BMI category (4 levels: underweight, normal, overweight, obese) against blood-pressure category (3 levels: normal, pre-hypertensive, hypertensive).

When NOT to — use instead

Hypotheses

H₀: τ-c = 0 — no monotonic ordinal association between x and y.
Hₐ: τ-c ≠ 0 — non-zero monotonic association (two-sided default).

Parameter tested: population Stuart's τ-c

Assumptions (and what to do if they fail)

The contingency table has interpretable row and column counts (r, c); m = min(r, c) is the smaller dimension.medium

Check: Report number of unique levels on each axis.

If violated: When either variable has many unique levels (essentially continuous), τ-c loses its interpretation as a bounded-table coefficient.

τ-c, τ-b, γ, and Somers' D all agree in sign.medium

Check: All measures should share the same sign.

If violated: Sign disagreement suggests a non-monotone pattern.

Ready to run a Stuart's τ-c (rectangular ordinal tables) on your own data?

Guided setup, automatic assumption checks, effect sizes, figures and an APA write-up.

Run this test →