Does it require normal data?

The classical continuous exchangeable null does not require normality.

Does a nonsignificant result prove the distributions match?

No; the test uses limited information and may have low power for some differences.

AI Glossary / Statistics and probability / Intermediate

Tukey - Duckworth test

Definition

The Tukey-Duckworth test is a simple two-sample test that uses counts of nonoverlapping observations at opposite ends of the combined data. When one sample contains the overall highest value and the other contains the overall lowest, it adds the high-end count from one sample to the low-end count from the other. Large totals indicate separation between the samples.

What it means

The procedure was designed to be easy to remember and calculate with little equipment. It uses the ordering of observations rather than computing means, variances, or all pairwise ranks.
That simplicity comes with limitations. It focuses on particular tail-separation information and can miss differences visible elsewhere in the data. Under a common-shape location-shift interpretation, separation suggests a location difference; without that structure, a rejection should not automatically be described as a difference in means or medians alone.

How it works

Obtain two independent samples measured on a comparable ordered scale. The simplest distribution-free calibration assumes observations from the same continuous distribution under the null, so ties have probability zero and group labels are exchangeable.
Check which sample contains the overall minimum and maximum. The basic rule requires different samples to occupy these opposite extremes. If the same sample contains both extremes, the usual two-tail counting condition is not satisfied.
Suppose sample A contains the largest observation and B the smallest. Count A observations strictly above the maximum of B. Separately count B observations strictly below the minimum of A. Add those two positive counts.
Reverse the roles if B contains the largest observation and A the smallest. The resulting total is an undirected separation statistic. A one-sided question would require a direction specified in advance and the appropriate reference calculation.
The familiar pocket-test cutoffs of roughly 7, 10, and 13 correspond to approximate two-sided levels of 5%, 1%, and 0.1% in suitable settings. They are not exact universal p-value translations. The original paper provides refinements for sample sizes and imbalance.
For an exact small-sample calculation under exchangeability and no ties, enumerate label assignments with the observed sample sizes. Recompute the statistic for every assignment. One complete definition assigns zero when the two extreme observations belong to the same group, then counts assignments whose statistic is at least as large as observed.
Handle ties explicitly rather than silently treating equality as strict exceedance. The classical continuous-data tables do not automatically resolve every discrete or tied-data situation. A justified permutation approach or another test may be needed.
Interpret the result together with effect estimates and the full data display. A count-based p-value says little about the size or practical cost of a difference. Dependence, unequal measurement conditions, or selective sampling can invalidate the exchangeability assumption even when the arithmetic is correct.

Three practical examples

A field engineer makes a quick preliminary comparison of two independently collected measurement samples.
A statistics class contrasts a tail-count procedure with a rank-sum test that uses more of the ordering information.
A reviewer reproduces a historical analysis and checks whether its approximate cutoff was appropriate for the actual sample sizes.

Think of it this way

Think of laying two colors of beads in order of size. Count the uninterrupted stretch of one color at the small end and the opposite color at the large end. The count records how strongly the ends separate, while deliberately ignoring much of the arrangement in the middle.

Worked example

Consider original samples A = [5, 8, 9, 10, 11, 12, 13, 14] and B = [1, 2, 3, 4, 6, 7, 10.5, 11.5]. All 16 values are distinct. A contains the maximum, and B contains the minimum.
Three A observations, 12, 13, and 14, exceed B's maximum of 11.5. Four B observations, 1, 2, 3, and 4, fall below A's minimum of 5. The statistic is therefore 3 + 4 = 7.
For an exact two-sided permutation calculation, define the statistic as zero when the same group occupies both overall extremes. There are 16 choose 8 = 12,870 equally likely label allocations under the continuous exchangeable null. Exhaustive enumeration finds 508 allocations with a statistic of at least 7.
The exact p-value for this stated statistic and null is 508/12,870, approximately 0.03947. This is below 0.05, but it is not exactly 0.05 merely because the pocket cutoff is 7. The result does not establish a mean difference, a causal effect, or a practically important separation.

Illustrative case study

Fictional case: Ashford Instruments uses a quick tail-count comparison during an exploratory equipment check. A preliminary result suggests separation, but the team notices that several readings were taken repeatedly from the same device.
The analyst does not treat those repeated readings as independent samples. The team redesigns the comparison around independent units and reports the full measurements and an effect estimate alongside any test result.

Common misconceptions and mistakes

This is not Tukey HSD, Tukey's trend test, or a general test of every distributional difference. The values 7, 10, and 13 are approximate remembered thresholds, not p-values. Both opposite-end counts must be positive for the basic counting rule. A large numerical outlier alone does not automatically create a large count. Independence and exchangeability remain important despite the absence of a normality assumption.

Business use

The test is useful mainly for understanding simple historical or exploratory comparisons. Modern reporting can retain its transparent counting logic while using exact calibration and fuller summaries. For consequential decisions, compare methods appropriate to the design and effect of interest rather than selecting a test solely because it is easy to calculate.

Check your understanding

Why is the example's count seven?

Answer: Three A values exceed all B values, and four B values lie below all A values.

Why is its exact p-value not 0.05?

Answer: The finite-sample permutation distribution is discrete, and the remembered cutoff is approximate.

Related terms

Permutation test
Order statistic
Two-sample test
nonparametric inference
exchangeability
location shift
order statistics
rank-sum test

Learn next

Reproduce the opposite-end counts on small samples, then enumerate label assignments. Compare how the statistic responds to location shifts, spread differences, and tied observations.

Sources

Tukey, A Quick, Compact, Two-Sample Test to Duckworth's Specifications, original 1959 Technometrics paper: https://www.stat.cmu.edu/technometrics/59-69/VOL-01-01/v0101031.pdf Original article DOI: https://doi.org/10.1080/00401706.1959.10489847

Previous: Tucker decomposition Next: Tukey's B method