Is this separate from HSD?

Usually it is an alternative name for the same pairwise procedure.

Why does the number of groups matter?

More means create more opportunities for an extreme difference.

Does the procedure rank all options with certainty?

No; estimated rankings can be uncertain.

AI Glossary / Statistics and probability / Intermediate

Tukey's range test

Definition

Tukey's range test is a common name for Tukey's honestly significant difference procedure, which compares group means using the studentized range distribution. Its key idea is to account for searching across all pairs of means. Under the classical balanced normal equal-variance model, it supports simultaneous pairwise inference with a specified familywise error rate.

What it means

The largest difference among several observed means tends to look impressive even when all population means are equal. It was selected from many possible differences, so an ordinary test for one preselected pair does not provide the same protection.
The range distribution calibrates that search. Here, range refers to the spread of the group means after scaling by estimated uncertainty. It is not simply a test based on the largest and smallest raw observations pooled across all groups.

How it works

Start with a defined collection of k groups and an outcome. In a balanced one-way design, each group contains n independent observations, and the usual exact theory assumes normal errors with a common variance.
Calculate the group means and the pooled within-group residual mean square, MSE. The error degrees of freedom are N-k, where N is the total number of observations in the simple model.
For the full observed range of group means, form q = (largest mean - smallest mean)/sqrt(MSE/n). The square-root denominator estimates the standard deviation of each group mean under the balanced model.
Compare with a studentized-range critical value based on k and the residual degrees of freedom. The distribution reflects the range among all k means rather than the uncertainty of a single fixed difference.
For individual pairs, compare each absolute difference with q-critical times sqrt(MSE/n). The same critical difference yields simultaneous confidence intervals around signed pairwise differences. This is the familiar Tukey HSD calculation under its alternative name.
When sample sizes differ, the Tukey-Kramer form uses pair-specific standard errors. Do not apply the balanced formula using an arbitrary average sample size without checking the justified method. Unequal variances or dependent measurements require a suitable alternative or model-based treatment.
Keep the comparison family explicit. Adding groups changes the set of comparisons and the critical value. A familywise guarantee for one outcome does not automatically cover searching across many outcomes, subgroups, and alternative preprocessing choices.
Read the findings in practical units. The range statistic describes the separation of estimates relative to uncertainty, while confidence intervals show plausible effect magnitudes under the model. Neither establishes which option is best after accounting for cost, feasibility, or other outcomes.

Three practical examples

A laboratory compares mean measurement results from several independently tested settings and wants all pairwise contrasts.
A product team evaluates multiple interface variants and recognizes that the most extreme pair was selected after seeing the data.
A reviewer checks whether a report's Tukey range test is the HSD procedure or a different Tukey-named method.

Think of it this way

Think of taking many photographs and choosing the most dramatic one. Its appearance must be judged with awareness that it was selected from a collection. Tukey's range-based calibration accounts for the opportunity to find an unusually large separation among several means.

Worked example

Consider an original three-group example. Group A contains -2, 2, 4, 6, 10; group B contains -1, 3, 5, 7, 11; group C contains 3, 7, 9, 11, 15. Each group has five observations, and their means are 4, 5, and 9.
Within each group, squared deviations from its mean sum to 80. The pooled residual sum of squares is therefore 240. With 15 observations and three means, residual degrees of freedom are 12, so MSE = 240/12 = 20.
The scaling factor is sqrt(20/5) = 2. The range of group means is 9 - 4 = 5, giving q = 5/2 = 2.5. At the 5% familywise level, the three-group, 12-degree-of-freedom critical value is approximately 3.773. The HSD threshold is 3.773 x 2 = 7.546.
Even the largest difference, 5, does not exceed that threshold, so none of the three pairwise differences is rejected by this calculation. The interval for C minus A is approximately 5 plus or minus 7.546, or [-2.546, 12.546]. This is inconclusive about a zero difference; it does not establish equality or practical equivalence.

Illustrative case study

Fictional case: Windward Design ranks four interface variants by observed mean completion time. The team initially highlights the fastest and slowest variants using an ordinary two-group test as though the pair had been specified beforehand.
An analyst identifies the selection across all four means and reports simultaneous comparisons instead. The revised report explains the broader uncertainty and avoids treating the observed ranking as a definitive ordering of the population means.

Common misconceptions and mistakes

The range is over means, not all individual measurements. Tukey's range test and Tukey's HSD usually name the same method, while Tukey's B and Tukey's additivity test are distinct. The method does not require every pair to differ when the overall range is large. Failure to reject all pairs does not prove equal means. A familywise error rate is not the probability that an individual conclusion is wrong.

Business use

The procedure helps organizations compare multiple options while accounting for the fact that the most striking difference may have been selected from many pairs. Reports should include sample sizes, residual variability, the defined family, and simultaneous intervals. If the decision concerns only comparisons with a baseline, a method tailored to that family may be more appropriate.

Check your understanding

What is the range of the means in the example?

Answer: Five.

What is the studentized range statistic?

Answer: 2.5.

Why is no pair rejected?

Answer: The largest difference is below the common threshold of about 7.546.

Does that show all effects are negligible?

Answer: No.

Related terms

Tukey's HSD test
Multiple comparisons
studentized range distribution
Tukey-Kramer method
simultaneous inference
selection effect

Learn next

Compare the maximum pairwise difference with the range of group means in a balanced design. Then explore how increasing residual variability widens every simultaneous interval.

Sources

NIST/SEMATECH, Tukey's method and studentized range: https://www.itl.nist.gov/div898/handbook/prc/section4/prc471.htm SciPy, tukey_hsd: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.tukey_hsd.html MedCalc, studentized-range critical-value table: https://www.medcalc.org/en/manual/tukey-q-table.php

Previous: Tukey's HSD test Next: Tukey's test of additivity