Why can it give more separations?

Some comparison thresholds can be lower than HSD's.

Does that automatically make it preferable?

No; the error-control properties and analysis objective must be considered.

AI Glossary / Statistics and probability / Intermediate

Tukey's B method

Definition

Tukey's B method is a multiple-comparison procedure for comparing group means that combines features of Tukey's honestly significant difference procedure and the Student-Newman-Keuls procedure. Its critical value is the average of their corresponding studentized-range critical values. It is a distinct procedure and should not be presented as another name for Tukey's HSD.

What it means

After comparing several group means, an analyst may want to identify which differences deserve further attention. Testing every pair with an unadjusted test increases the opportunity for false positive findings.
Tukey's B is one historical response to that multiple-comparison problem. It uses ordered means and thresholds related to the range of means being compared. Its compromise construction does not automatically inherit the familywise error guarantee of Tukey's HSD.

How it works

Begin with the appropriate analysis-of-variance model and inspect its assumptions. In the familiar one-way setting, observations are independent and the model uses a common within-group variance. Exact normal-theory interpretations additionally rely on normally distributed errors.
Calculate group means, the residual mean square, and residual degrees of freedom. Order the group means so the span between a pair can be identified. The number of means in that span can affect the Student-Newman-Keuls component of the threshold.
For the balanced setting, let k denote the total number of groups and r the number of ordered means spanned by the comparison, including its endpoints. Tukey's HSD uses a studentized-range critical value based on k; the corresponding Student-Newman-Keuls value uses r.
Tukey's B averages those two critical values at the chosen probability level and residual degrees of freedom. With equal group sizes n, multiplying the averaged q value by the square root of MSE/n gives the corresponding critical difference on the outcome scale.
The complete procedure also requires its ordering and stepwise comparison rules. A single threshold calculation does not reproduce an entire software output or establish which homogeneous subsets will be displayed. Unequal sample sizes require attention to the implementation's conventions rather than casually substituting one group's size.
When r equals k, the two component critical values coincide. For a smaller span, the averaged value can be below the HSD threshold. This explains why the methods may reach different decisions without any arithmetic error.
State the error-control objective explicitly. A threshold positioned between two procedures is not, by itself, a proof of strong familywise error control. For a project requiring that guarantee, select and justify a method that provides it under the relevant assumptions.
Interpret grouped output carefully. Means that are not separated by the procedure are not thereby proven equal or practically equivalent. Overlapping subsets can occur, and statistical nonsignificance is not a transitive equivalence relation. Report estimated differences and context rather than reducing conclusions to a sequence of letters.

Three practical examples

An analyst reads an older report that used Tukey's B and checks why its groupings differ from HSD output.
A training evaluation compares several group means and documents the exact post-hoc method rather than saying only that a Tukey test was used.
A quality team reproduces a legacy analysis, preserving its method label while reassessing the error-control requirement.

Think of it this way

Think of choosing a threshold halfway between two existing thresholds. The arithmetic is simple, but the midpoint does not automatically acquire every protection associated with either endpoint. Its interpretation still depends on the procedure that produced it.

Worked example

Consider an original arithmetic illustration for one comparison in a balanced design. Suppose each group has n = 9 observations and the fitted model's residual mean square is MSE = 9. The studentized-range scaling factor is square root of 9/9 = 1.
For this illustration only, suppose the corresponding HSD critical value supplied to the calculation is 4.0 and the Student-Newman-Keuls critical value is 3.0. These are stipulated demonstration inputs, not claimed table values for a particular alpha or degrees of freedom.
Tukey's B averages them: (4.0 + 3.0)/2 = 3.5. Its critical difference for this calculation is 3.5 x 1 = 3.5. A mean difference of 3.6 exceeds that threshold but does not exceed the HSD threshold of 4.0.
This shows how the arithmetic can yield different threshold comparisons. It does not supply a complete stepwise test, a p-value, or a valid 5% conclusion for an actual dataset. A real analysis must obtain the appropriate distribution quantiles and apply all comparison rules.

Illustrative case study

Fictional case: Lakeshore Training receives a legacy report dividing five course formats into homogeneous subsets using Tukey's B. A new analyst initially labels the output as Tukey HSD and assumes the same familywise protection.
During review, the team identifies the exact software option and corrects the method description. It reproduces the historical output for traceability, then performs a separately specified analysis consistent with its current error-control objective. The report distinguishes changed methodology from changed underlying observations.

Common misconceptions and mistakes

Do not confuse the letter B with the B labels used to mark groups in a compact-letter display. Tukey's B is not HSD, Tukey-Kramer, or a Bayesian method. Averaging critical values does not mean averaging p-values. A significant omnibus ANOVA does not make every subsequent comparison immune to multiplicity. Failure to separate two means does not demonstrate equivalence.

Business use

The method is especially relevant when interpreting or reproducing older multi-group analyses. Record its full name, implementation, assumptions, and comparison family. Explain any differences from a modern reanalysis rather than silently changing procedures. Business choices should consider the magnitude and consequences of differences, not only post-hoc grouping labels.

Check your understanding

What is the average of the illustrative critical values 4 and 3?

Answer: It is 3.5.

Does the calculation establish a real 5% test?

Answer: No, because no actual quantiles, degrees of freedom, or complete stepwise analysis were supplied.

Related terms

Multiple comparisons
Familywise error rate
Tukey's HSD
Student-Newman-Keuls test
studentized range
homogeneous subsets

Learn next

Compare the critical-value logic of HSD and Student-Newman-Keuls. Then inspect an actual software report and identify exactly which procedure and sample-size convention it used.

Sources

IBM SPSS Statistics, One-Way ANOVA Post Hoc Tests, Tukey's b definition: https://www.ibm.com/docs/en/spss-statistics/30.0.0?topic=anova-one-way-post-hoc-tests University of Michigan, author lecture notes on contrasts and post-hoc tests, Tukey B discussion: https://websites.umich.edu/~gonzo/coursenotes/file3.pdf

Previous: Tukey - Duckworth test Next: Tukey's HSD test