Does the test compare means?

Its intended purpose is spread comparison, but location differences can affect it.

Is it identical to the Mann-Whitney test?

It uses a related rank-sum calculation after a different ranking transformation, so the target interpretation differs.

Implementations may apply tie adjustments, but exact p-value availability and conventions vary.

Should median adjustment always be enabled?

No; the decision depends on the scientific question and the chosen method's assumptions.

AI Glossary / Statistics and probability / Advanced

Siegel - Tukey test

Definition

The Siegel-Tukey test is a nonparametric procedure for comparing the spread of two independent samples using specially assigned ranks. It gives small ranks to observations near the extremes of the pooled sample and larger ranks nearer the center, then compares the groups' rank sums. The intended interpretation concerns relative variability under appropriate location and distributional assumptions. It is not a general-purpose test that isolates variance differences regardless of differences in the groups' centers or shapes.

What it means

Ordinary ranks run from the smallest value to the largest and are naturally sensitive to location. Siegel-Tukey ranks instead work inward from the pooled extremes so that a more dispersed group tends to receive smaller ranks. This turns a spread comparison into a rank-sum problem. However, a shift in location can also place one group disproportionately near an extreme. The rank construction therefore does not magically separate every possible location, shape and scale difference.

How it works

Check that the two groups represent independent units and clarify whether a common-location scale comparison is appropriate. Pool and sort the observations, then assign ranks in the prescribed alternating pattern from the extremes towards the center. Handle ties and any odd pooled sample size according to the chosen implementation. Sum ranks for a group and compare the statistic with an appropriate rank-sum null distribution. If medians differ, examine the issue explicitly; automatic median adjustment is a method choice whose calibration and interpretation need justification.

Three practical examples

Example 1: Engineers compare the variability of output from two processes whose target location is comparable. Example 2: An analyst investigates whether independent delivery-time samples differ in spread after examining their centers and shapes. Example 3: An AI evaluation team compares the dispersion of a performance measure across independently sampled cases for two settings, where the sampling structure supports the comparison. Paired outputs on the same cases require a method that recognizes pairing rather than this simple independent-sample procedure.

Think of it this way

Think of assigning the first tickets to people at the two ends of a line and continuing towards its middle. A group occupying the ends receives many early tickets. The tickets reveal where that group sits in the combined arrangement, but they cannot by themselves explain why. A group shifted to one side may also receive early tickets, so the investigator must distinguish a location shift from a spread difference.

Worked example

Take sample A as 1,2,7,8 and sample B as 3,4,5,6. Both sample medians are 4.5. The pooled values are 1 through 8. Assigning ranks from the extremes in the Siegel-Tukey sequence gives ranks 1,4,5,8,7,6,3,2 in ascending-value order. A receives ranks 1,4,3,2, with sum ten, the smallest possible sum for four distinct ranks. Under an exchangeable continuous null, there are eight choose four = 70 equally likely assignments of four ranks to A. Exactly one has sum ten and exactly one has the symmetric maximum sum 26. Thus the exact doubled-tail two-sided p-value is 2/70, approximately 0.02857. The sample variances are 37/3 for A and 5/3 for B, consistent with the visible spread contrast, but the test calculation uses the special ranks rather than their variance ratio.

Illustrative case study

In this fictional case study, Valley Calibration compares two production settings. Initial output plots show both a location shift and a possible spread change. Rather than treating a significant Siegel-Tukey result as pure evidence about variance, the analyst reviews whether the intended scale model is defensible and compares suitable alternatives. The final report separates center and spread summaries, states the method's assumptions and presents uncertainty. It avoids claiming that a rank-based procedure is automatically immune to differences unrelated to the target question.

Common misconceptions and mistakes

Common mistakes include using ordinary ascending ranks, treating the test as assumption-free, and overlooking differences in medians. Another error is equating its null with equality of finite variances for arbitrary distributions: its scale interpretation needs additional structure. Ties and implementation conventions can affect exact calculations. A small p-value does not quantify the size or practical importance of a variability difference. A nonsignificant result does not demonstrate equal spread, particularly with small samples or limited power.

Business use

Businesses may use the test as one option when comparing process consistency without relying on a normal-theory variance test. The choice should follow the data structure and the specific definition of variability that matters. Reports should include understandable spread measures, plots and relevant operational tolerances alongside any test. For AI applications, verify whether outputs are independent or paired and whether changes in location or case mix could explain the result. Method selection should not be driven solely by which test produces significance.

Check your understanding

Which group receives the smaller rank sum in the worked example?

Answer: A, because it occupies the pooled extremes.

Why is the exact two-sided p-value 2/70?

Answer: One assignment reaches each equally extreme end of the rank-sum distribution.

Does a location shift matter?

Answer: Yes; it can affect the test and complicate a pure spread interpretation.

Would paired model outputs satisfy the independent-group design automatically?

Answer: No.

Related terms

Ansari-Bradley test
Levene test
Related terms include scale test
rank-sum test
Mood test
Brown-Forsythe test and exchangeability

Learn next

Learn next how the rank-sum null distribution is obtained by assigning group labels to pooled ranks. Compare location and scale alternatives on plots before choosing a test. Review tie handling and the implementation's treatment of odd sample sizes, then practice reporting a spread estimate and its practical meaning separately from a p-value.

Sources

Sources: DescTools authors, SiegelTukeyTest documentation, https://andrisignorell.github.io/DescTools/reference/SiegelTukeyTest.html; DescTools reference manual, https://stat.ethz.ch/CRAN/web/packages/DescTools/DescTools.pdf. These primary software references describe the procedure, median sensitivity and implementation options and cite Siegel and Tukey's 1960 paper. The eight-observation example and fictional calibration case are original, with the exact rank allocation independently enumerated.

Previous: Sideways move Next: Sigmoid