Is this the same as the Wilcoxon rank-sum test on raw values?

No; it uses squared ranks of within-group absolute deviations.

Does it test equality of means?

No; its intended focus is dispersion.

No; rank assignment and the reference calculation need to handle them.

Is it automatically exact for any sample size?

No; the selected procedure and its assumptions determine the inference.

AI Glossary / Statistics and probability / Advanced

Squared ranks test

Definition

The squared ranks test, often called the Conover squared ranks test, is a rank-based procedure for comparing dispersion across independent groups. In a common formulation, observations are converted to absolute deviations from their own group means, those deviations are ranked jointly, and the ranks are squared. Differences in the resulting group scores provide evidence about unequal spread under the procedure's assumptions. It is distinct from tests that rank the original observations to compare location.

What it means

A group with more dispersed values tends to have larger deviations from its center. Ranking those deviations reduces reliance on their numerical magnitudes, while squaring the ranks gives greater emphasis to relatively large deviations. The procedure is commonly described as a test for equal variances, but interpretation requires attention to the distributional framework, group centering and sample sizes. Rank-based does not mean assumption-free, and arbitrary differences in distribution shape can complicate a simple variance-only explanation.

How it works

For each group, calculate its sample mean and take the absolute distance of each observation from that mean. Pool all these deviations, rank them from smallest to largest and use an appropriate average-rank convention for ties. Square the ranks and form the group score sums. The two-group and multiple-group procedures use these scores with specified standardization and reference distributions. Follow a documented implementation rather than mixing formulas from different versions. Small samples, extensive ties and estimated group centers deserve particular attention when selecting and interpreting inference.

Three practical examples

Example 1: A process analyst compares variation in output from independent production settings. Example 2: An AI team studies whether response-time dispersion differs across independently sampled operating conditions. Example 3: A research experiment compares the spread of measurement errors under several procedures. In each case, the relevant observations must match the independent-group design. Repeated readings from the same unit cannot automatically be treated as independent, and a statistically detectable difference in spread may still be too small to matter operationally.

Think of it this way

Think of comparing how far darts land from each group's own center. Instead of comparing the exact distances, arrange all distances in order and assign larger scores to those near the top of the ranking. The group accumulating larger scores tends to be more spread out. The analogy does not remove the need to account for how centers were estimated or how the observations were collected.

Worked example

Consider group A values (-1,0,1) and group B values (-3,0,3). Both means are zero. Their absolute deviations are (1,0,1) and (3,0,3). In the pooled list, the two zeros receive average rank 1.5, the two ones rank 3.5 and the two threes rank 5.5. Group A's squared-rank sum is 3.5 squared + 1.5 squared + 3.5 squared = 26.75. Group B's sum is 5.5 squared + 1.5 squared + 5.5 squared = 62.75. Their sample variances are one and nine, consistent with the larger deviation scores for B. This tiny example demonstrates score construction only. It does not supply a valid significance conclusion merely from comparing the sums, and it does not justify a large-sample approximation for three observations per group.

Illustrative case study

In this fictional case study, Flint Systems compares the variability of processing times under several configurations. Analysts verify that runs were independently sampled and inspect whether distributions differ in shape as well as spread. They calculate a documented squared-ranks procedure and report descriptive dispersion measures alongside its result. When a configuration shows greater variability, the team investigates long-tail behavior rather than assuming the test identifies a cause. The report keeps practical service requirements separate from statistical significance.

Common misconceptions and mistakes

Common mistakes include ranking the raw observations instead of their within-group absolute deviations, forgetting to square the ranks, and applying a two-sample formula to a multiple-group problem. Another error is interpreting nonsignificance as proof of equal variances. The use of ranks does not make extreme observations irrelevant, because group means themselves can be sensitive to extremes. Do not assume that permuting already-centered deviations provides an exact finite-sample test for every original-data null; the centering step and exchangeability assumptions matter.

Business use

Businesses can use dispersion comparisons when consistency matters alongside average performance. The test should complement plots, variance or robust spread summaries, and an understanding of measurement quality. A decision about changing a process needs the size and consequences of the spread difference, not only a p-value. If groups have markedly different shapes or dependence structures, consider whether another analysis more directly addresses the intended question. Preserve the chosen centering, tie handling and reference method in the analysis record so the result can be reproduced.

Check your understanding

What is group A's squared-rank sum in the example?

Answer: 26.75.

Why do both zeros receive rank 1.5?

Answer: They occupy the first two pooled positions and share their average.

Which group has the larger score sum?

Answer: B.

Why is no p-value reported for the example?

Answer: The example demonstrates the transformation and does not establish a suitable small-sample reference calculation.

Related terms

Fligner-Killeen test
Homogeneity of variance
Related terms include Conover scores
Levene's test
Brown-Forsythe test
pooled ranks and dispersion

Learn next

Learn next how location removal changes a dispersion comparison and why ties alter rank-based calculations. Compare squared-rank scores with absolute-deviation and normal-score procedures. Then examine robustness to distribution shape and the distinction between a meaningful spread difference and an assumption check used before another analysis.

Sources

Sources: NIST Dataplot, Squared Ranks Test, https://www.itl.nist.gov/div898/software/dataplot/refman1/auxillar/squarank.htm; Conover Scores, https://www.itl.nist.gov/div898/software/dataplot/refman2/auxillar/conovers.htm; SAS, Scores for Linear Rank and One-Way ANOVA Tests, https://support.sas.com/documentation/cdl/en/statug/65328/HTML/default/statug_npar1way_details08.htm. These primary implementation references describe the scores and procedures. The tied-rank example is original and independently checked; the case is fictional.

Previous: Squared hinge loss Next: SSS*