Does the word new mean this is a modern method?

No. It belongs to the established historical name.

Does alpha = 0.05 protect the entire family at 5%?

Not generally; this is a central limitation of the procedure.

Do identical grouping letters prove equal means?

No. They indicate that the procedure did not declare that comparison different under its rules.

AI Glossary / Statistics and probability / Advanced

Duncan's new multiple range test

Definition

Duncan's new multiple range test is a procedure for comparing ordered group means using a sequence of studentized-range comparisons. Its critical ranges depend on how many ordered means a comparison spans. It is historically common in analysis-of-variance applications, particularly agricultural experiments. A key limitation is that using a nominal level such as 0.05 does not provide general family-wise error control at 0.05 across all comparisons. The word new is part of its historical name, not a claim of recent development.

What it means

After a study produces several group means, analysts may want to identify which differ. Duncan's procedure orders the observed means and applies range-dependent thresholds rather than one identical cutoff to every comparison. It can produce more declared differences than procedures designed for stronger multiplicity control. That sensitivity comes with a different error-control objective and should not be described as an unconditional improvement. The choice of comparison procedure should follow the intended inferential claims and tolerance for false positives, rather than whichever method produces preferred groupings.

How it works

Fit an appropriate model and obtain group means, the residual mean square and error degrees of freedom. In a balanced independent-groups setting with n observations per group, the scale for a mean range is sqrt(MSE/n). For a span containing r ordered means, Duncan's nominal range level is commonly expressed as alpha_r = 1 - (1 - alpha)^(r - 1); the corresponding studentized-range reference and procedure determine a critical range. Apply the specified stepwise comparison rules, including the implementation's handling of unequal group sizes. Assumptions inherited from the classical model include suitable independence, error normality and common variance. Validate the design and implementation rather than treating a grouping table as self-explanatory.

Three practical examples

Historical study interpretation: understand a published treatment-mean grouping and its weaker multiplicity protection.
Experimental-method comparison: examine how different multiple-comparison procedures alter conclusions on the same fitted model.
Statistical education: connect ordered mean ranges, residual variability and changing range levels in a reproducible calculation.

Think of it this way

Think of comparing positions on a ranked ladder using rules that change with the number of rungs spanned. The procedure is not simply a collection of identical pairwise checks. Understanding those changing rules matters before interpreting the resulting separations as equally protected claims.

Worked example

Suppose a hypothetical balanced study has four ordered means 10, 12, 14 and 16, with five independent observations per group and residual mean square four. The range scale is sqrt(4/5), approximately 0.8944. The widest observed range is 16 - 10 = six, or about 6.7082 on the studentized-range scale. With base alpha = 0.05, Duncan's range levels are 0.05 for r = 2, 1 - 0.95 squared = 0.0975 for r = 3, and 1 - 0.95 cubed = 0.142625 for r = 4. These calculations show why the nominal range level changes with span. They are not a completed set of significance decisions: matching studentized-range quantiles, error degrees of freedom and the procedure's sequential rules are still needed. In particular, 0.142625 is not a universal exact family-wise error probability for this dataset.

Illustrative case study

Fictional case: Clover Process Research receives a legacy report using Duncan group letters for four treatments. Analysts reproduce the stated method and explain its error-control limitations before comparing results with a procedure aligned to the new project's inferential objective. They report estimated differences and uncertainty rather than presenting one letter arrangement as definitive. This invented case illustrates responsible interpretation of a historical method. It does not claim that replacing the method necessarily reverses a conclusion or that any particular comparison family is suitable for every experiment.

Common misconceptions and mistakes

Do not confuse Duncan's test with Dunnett's treatment-versus-control procedure or with the Waller-Duncan method. Do not report its nominal alpha as a general 5% probability of at least one false positive across the family. An overall significant ANOVA does not automatically supply strong family-wise protection for all subsequent Duncan comparisons. Shared grouping letters indicate lack of rejection under the implemented procedure, not demonstrated equivalence. Unequal group sizes and complex designs need appropriate treatment; simply substituting an arbitrary sample size into a balanced formula is not sufficient.

Business use

Businesses reviewing multi-group experiments should first define the comparisons they intend to claim. A historical Duncan analysis can be reproduced for transparency, while its limitations remain explicit. When simultaneous claims require family-wise control, choose and justify a method designed for that goal rather than assuming all post-analysis comparison tools offer the same protection. Report effect magnitudes and operational importance alongside testing results. More declared differences can increase follow-up costs and false leads, so apparent sensitivity is meaningful only in relation to the error criterion.

Check your understanding

Why is the four-mean range level larger than the two-mean level in the example?

Answer: Duncan's construction uses alpha_r = 1 - (1 - alpha)^(r - 1), which increases with the number of ordered means spanned. This changing level is part of the procedure and must not be mistaken for a fixed family-wise threshold.

Related terms

Multiple comparisons
Newman-Keuls test
Dunnett's test
Compact letter display
studentized range
ANOVA
family-wise error rate
ordered means
Tukey procedure
residual mean square

Learn next

identify the intended comparison family before selecting a procedure. Compare Duncan's range-level construction with procedures that control family-wise error, and reproduce a small example using documented software. Then interpret group letters alongside effect estimates, emphasizing that nonrejection is not equivalence.

Sources

SAS, Multiple Comparisons: https://support.sas.com/documentation/cdl/en/statug/67523/HTML/default/statug_glm_details29.htm SAS, Alternative Multiple Comparison Procedures: https://support.sas.com/documentation/cdl/en/statug/66103/HTML/default/statug_anova_examples02.htm agricolae authors, Duncan's new multiple range test: https://myaseen208.com/agricolae/reference/duncan.test.html

Previous: Dummy variable Next: Dunn-Sidák test