MerQuR+Statistics Glossary

Statistics Glossary → r effect size from Z/√N

r effect size from Z/√Nconcept

An effect size for rank-based tests, obtained by dividing the test's standard normal statistic Z by the square root of the sample size, and read on the correlation scale.

Tests such as the Mann-Whitney U or the Wilcoxon signed-rank test produce a Z value through a normal approximation. Z grows with the sample size, whereas r = Z/√N is an effect measure freed from sample size (Rosenthal, 1991). It lies between −1 and 1, and Cohen's benchmarks for correlations apply to its absolute value: about 0.1 small, 0.3 medium, 0.5 large.

What N stands for depends on the design and should be reported. For independent groups it is the total number of observations; for paired designs some authors use the number of pairs and others the total number of observations, a choice that can change r by a factor of √2. Because Z depends on tie and continuity corrections, r is affected by them too. The rank-biserial correlation is a direct, clearly defined alternative.

An everyday example

A shop compares customer satisfaction with its redesigned window display, using 20 customers in each condition. Satisfaction is measured on a 1-5 scale, so a Mann-Whitney U test is used and Z = 2.5. r = 2.5/√40 ≈ 0.40, an effect between medium and large.

In MerQur

In the Mann-Whitney U test, the 'Effect sizes: Z / √N' checkbox gives r = Z/√N, and the Wilcoxon signed-rank test writes r = Z/√n' in every run. In the Kruskal-Wallis test, 'Extra outputs: Detailed pairwise table (+ r)' gives r = |z|/√N for the Dunn comparisons, and in the Friedman test the 'Pairwise detail: W, Z, r' table gives r = |Z|/√n for pairwise Wilcoxon comparisons; in Friedman the denominator is the number of blocks.

Analyses where this concept appears

Mann-Whitney UWilcoxon Signed-RankKruskal-WallisFriedman

Related concepts

Rank-biserial correlationEffect size