Is REGWQ identical to Tukey's method?

No; both concern multiple mean comparisons, but their decision procedures differ.

Does it work for every unbalanced dataset?

Not without checking the implementation and assumptions.

AI Glossary / Statistics and probability / Intermediate

Ryan procedure (REGWQ)

Definition

The Ryan procedure (REGWQ), more fully the Ryan - Einot - Gabriel - Welsch Q procedure, is a stepwise multiple-comparison method for investigating differences among group means. It uses ordered means and studentized-range comparisons with adjusted significance levels to control familywise Type I error under its model conditions. The conventional procedure is especially associated with balanced, equal-sample-size comparisons following an analysis-of-variance framework.

What it means

When several group means are compared, running many ordinary pairwise tests can create an excessive chance of at least one false positive. REGWQ addresses the comparison family through a structured sequence rather than treating each test as unrelated.
The procedure sorts the observed means and tests ranges spanning different numbers of ordered means. Its name distinguishes it from the simpler Student - Newman - Keuls procedure, whose error-control properties are different. The Q refers to the studentized-range basis, not to a general measure of quality.

How it works

Define the family of means and comparisons before inspecting the results. Include the groups relevant to the scientific question; changing the family after observing which differences look promising can undermine the intended error control.
Fit the appropriate analysis model and obtain its pooled residual mean square and error degrees of freedom. In the simple balanced one-way setting, the standard range scale is sqrt(MSE/n), where n is the number of independent observations in each group.
Order the group means from smallest to largest. For a contiguous subset spanning r ordered means, calculate the difference between its largest and smallest means and compare the resulting studentized range with the applicable critical value.
Begin with the widest range and follow the procedure's stepwise rules for smaller ranges. Decisions are constrained by the testing sequence; a list of isolated comparisons using arbitrary cutoffs is not the REGWQ procedure.
In a common formulation with k total means, the significance level is alpha for ranges containing k or k−1 means. For smaller ranges it is 1−(1−alpha)^(r/k). Critical-value adjustments and the algorithm's stopping rules must also be implemented correctly.
Check assumptions. The classical model uses independent observations, normally distributed errors for exact small-sample calibration, and a common error variance. Repeated measurements, clustered data, or severe unequal variances require an analysis that addresses those features.
Treat unequal sample sizes carefully. Some software supplies adaptations, but these are not interchangeable with the conventional balanced procedure. Consult the particular implementation's documentation and error-control conditions rather than assuming that any available option preserves every guarantee.
Interpret results as evidence about mean differences, not proof of equivalence among non-significant groups. Report the comparison method, familywise level, sample sizes, error model, and meaningful effect sizes. A lettering display can aid reading, but shared letters indicate a failure to detect a difference under the method, not equality established beyond doubt.

Three practical examples

Process trials: an analyst compares average output from several equally replicated settings.
Interface research: a balanced experiment compares mean task scores across multiple designs using a prespecified comparison family.
AI workflow assessment: researchers compare several independently assigned workflow variants while accounting for the many mean comparisons.

Think of it this way

Think of examining an ordered row of group averages through windows of different widths. The method first considers the widest separation, then moves through permitted smaller windows using calibrated rules. Looking through many windows without such rules would create more opportunities to mistake noise for a meaningful difference.

Worked example

Suppose a fictional balanced study has four group means: 10, 11, 14, and 18. Each group has n = 4 observations, the pooled residual MSE is stipulated as 4, and the one-way error degrees of freedom are 4×(4−1) = 12.
The range scale is sqrt(4/4) = 1. Thus the full-range statistic is (18−10)/1 = 8. The range from 10 to 14 spans three ordered means and has statistic 4; the adjacent range from 10 to 11 has statistic 1.
At familywise alpha = 0.05, the four- and three-mean ranges use level 0.05 in the stated formulation. A two-mean range uses 1−0.95^(2/4), approximately 0.02532.
These calculations show the ingredients and the adjusted level. They deliberately do not declare which comparisons are significant: that requires the studentized-range critical values and the complete stepwise decision rules, not just these arithmetic values.

Illustrative case study

Fictional case: Beacon Operations tests four standardized task procedures with equal numbers of independently assigned workers. The analyst plans REGWQ comparisons before viewing the scores and checks the common-variance model.
When presenting results, the team reports estimated mean differences alongside the procedure's decisions. Two procedures sharing a grouping letter are described as not distinguished by this analysis; they are not marketed as proven interchangeable. An operationally meaningful equivalence question would require a different explicit analysis.

Common misconceptions and mistakes

REGWQ is not simply several unadjusted t tests.
A significant overall ANOVA result does not mean that every pair of groups differs.
Non-significance does not establish equal means or equivalent business performance.
Using a balanced-design formula with markedly unequal group sizes is not automatically justified.
Familywise error control concerns the stated family under assumptions; it does not guarantee that any particular reported difference is true or practically important.

Business use

Use the procedure when the study design and comparison objective fit its conditions. For a business report, explain the comparison family and distinguish statistical evidence from the size and cost of the observed differences. Select an alternative when only comparisons against a single control, unequal variances, or complex dependence structures define the real problem.

Check your understanding

Why not test every pair at an ordinary 5% level?

Answer: The chance of at least one false positive can exceed the intended familywise level.

In the example, does a range statistic of 4 alone establish significance?

Answer: No; its critical value and position in the procedure matter.

Related terms

Multiple comparisons
Familywise error rate
Analysis of variance
Pooled variance
studentized range
Tukey procedure
Student–Newman–Keuls test

Learn next

Inspect a documented REGWQ implementation using a balanced example. Trace the ordered ranges, adjusted levels, and stopping rules, then compare its conclusions with a prespecified alternative multiple-comparison method.

Sources

Howell, Statistical Methods for Psychology glossary: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html mutoss, regwq primary package documentation: https://search.r-project.org/CRAN/refmans/mutoss/html/regwq.html The numerical ingredients and fictional case are original illustrations.

Previous: RVQ Next: S chart