KRUSKAL-WALLIS TEST group n, ti is the average observed rank sums for the
group, tj is the observed rank for an observation for the
corresponding group, and Ti is the observed total average The Kruskal-Wallis (Kruskal & Wallis, 1952) is a non- rank sums. parametric statistical test that assesses the differences Suppose we are interested in assessing whether the among three or more independently sampled groups on crime rankings from a random sample of 15 U.S. cities a single, non-normally distributed continuous variable. from the Northeast, West, and South differ. The original Non-normally distributed data (e.g., ordinal or rank data) data are as follows: Northeast (U.S. rankings: 2, 5, 89, 203, are suitable for the Kruskal-Wallis test. In contrast, 241), West (U.S. rankings: 1, 34, 50, 55, 292), and South the one-way analysis of variance (ANOVA), which is a (U.S. rankings: 6, 12, 19, 22, 356). The procedure reranks parametric test, may be used for a normally distributed the rank data and leaves us with new relative rankings continuous variable. The Kruskal-Wallis test is an exten- among these 15 cities: Northeast (2, 3, 11, 12, 13), West sion of the two-group Mann-Whitney U (Wilcoxon rank) (1, 8, 9, 10, 14), and South (4, 5, 6, 7, 15). The H statistic test. Thus, the Kruskal-Wallis is a more generalized form for the data above is nonsignificant (H(2) = 0.14, p = .93), of the Mann-Whitney U test and is the nonparametric and we accordingly fail to reject the null and conclude that version of the one-way ANOVA. the three regions do not differ on crime rankings. A typical application of the Kruskal-Wallis test is to assess whether three or more groups differ on a single REFERENCE variable that fails to meet the normality assumptions of ANOVA. Since the variable of interest does not meet Kruskal, W. H., & Wallis, W. A. (1952). Use of ranks in normality assumptions, we may not compare group means; one-criterion variance analysis. Journal of the American instead we compare ranks. The null hypothesis specifies Statistical Association, 47, 583–621. that the groups are subsets from the same population (i.e., H0 : (a, b, c, . . . , n) ⊆p). To test the null hypothesis, SUGGESTED READINGS we combine the groups into a single group and rank the variable of interest. The new rank scores are summed by Cohen, B. H. (2008). Explaining psychological statistics (3rd ed.). Hoboken, NJ: John Wiley & Sons. group (Ta , Tb , . . . , Tn ) and, along with group sample sizes, can be used to calculate the H statistic (see Equation 1). H Gibbons, J. D. (1993). Nonparametric statistics: An introduction reflects the variance in ranks between groups and closely (Quantitative Applications in the Social Sciences, 90). Newbury Park, CA: Sage. resembles the chi-square distribution. When testing the null, we can use H and refer to a chi-square table with Hollander, M., & Wolfe, D. A. (1999). Nonparametric statistical degrees of freedom equal to n (number of groups) minus methods (2nd ed.). New York: John Wiley & Sons. 1. If H exceeds a critical value, we may conclude that PATRICK E. MCKIGHT the groups do not come from the same population; many JULIUS NAJAB researchers often conclude that the test is an omnibus test George Mason University for differences between groups, but many statisticians object to this interpretation: See also: Mann-Whitney U Test; Nonparametric Statistical H = N − 1(gnn (ti − Ti ) /gsnn (tj − Ti ) ) 2 2 (1) Tests
where nn is the sample size of the corresponding group, g is
the sum of the group n, snn is the sum of the corresponding