13.14
Test de Kruskal-Wallis
Le test de Kruskal-Wallis, également connu sous le nom de test H de Kruskal-Wallis, sert d'alternative non paramétrique au ANOVA à un facteur, offrant…
Le test de Kruskal-Wallis, ou test H, est une alternative non paramétrique à l’ANOVA à un facteur. Il teste efficacement l’hypothèse nulle selon laquelle trois populations ou plus ont des médianes identiques.
Prenons l'exemple de l'ensemble de données sur les ventes de raisins rouges, verts et noirs d'un supermarché.
Comment peut-on déterminer si les trois échantillons proviennent de populations ayant les mêmes ventes médianes ?
Pour tester cette affirmation, tous les échantillons sont initialement fusionnés pour attribuer des rangs. Ensuite, la somme des rangs de chaque échantillon est calculée indépendamment.
La statistique de test, H, est dérivée à l’aide d’une formule qui quantifie la variance entre les sommes de rang.
Dans ce scénario, la valeur H suit la distribution du khi-deux, les degrés de liberté étant inférieurs d’un au nombre total d’échantillons, qui est de deux.
À ce niveau de degrés de liberté, la valeur critique est déterminée à partir de la table correspondante à un niveau de signification choisi, tel que 0,05.
Étant donné que la valeur H calculée est inférieure à la valeur critique, l’hypothèse nulle, suggérant que les échantillons proviennent de populations ayant des médianes équivalentes, ne peut être rejetée.
View the full transcript and gain access to JoVE Core videos
Q1: What is the Kruskal-Wallis test used for?
The Kruskal-Wallis test, or H test, is a nonparametric alternative to one-way ANOVA that tests whether three or more independent groups have identical medians. It's particularly valuable when data doesn't meet the normal distribution assumption required by parametric tests, making it ideal for ordinal data or when variances across groups are unequal.
Q2: How does the Kruskal-Wallis test calculate its test statistic?
The test merges all samples and assigns ranks to each data point across the entire dataset. The sum of ranks for each sample is calculated independently, then used in a formula to derive the H statistic, which quantifies variance among rank sums. This H value follows the chi-square distribution with degrees of freedom equal to one less than the number of groups.
Q3: What are the key assumptions of the Kruskal-Wallis test?
The Kruskal-Wallis test requires independent observations, identical distribution shapes and scales across groups, and ordinal or continuous data. Unlike parametric tests such as one-way ANOVA, it doesn't assume normal distribution or equal variances, making it more robust for non-normally distributed datasets and reducing the impact of outliers.
Q4: When should you use the Kruskal-Wallis test instead of ANOVA?
Use the Kruskal-Wallis test when comparing medians across three or more groups with non-normal data, unequal variances, or ordinal measurements. It's advantageous in psychology for comparing therapeutic approaches, in medicine for evaluating non-normally distributed treatment outcomes, and in environmental science for comparing ecological measurements across different sites.
Q5: How do you interpret the Kruskal-Wallis test results?
Compare the calculated H statistic to the critical value from the chi-square distribution table at your chosen significance level, such as 0.05. If H is lower than the critical value, you cannot reject the null hypothesis that samples originate from populations with equivalent medians. If H exceeds the critical value, reject the null hypothesis.
Q6: Why does the Kruskal-Wallis test focus on ranks instead of means?
By using ranks rather than actual data values, the Kruskal-Wallis test reduces the impact of outliers and extreme values that could distort parametric analyses. This rank-based approach makes the test robust against non-normal distribution shapes and violations of homogeneity assumptions, providing valid conclusions even when data doesn't conform to traditional statistical requirements.
Q7: What is the relationship between the Kruskal-Wallis test and other nonparametric tests?
The Kruskal-Wallis test is part of a broader family of nonparametric methods used when parametric assumptions fail. For paired or matched data, the friedman two way analysis of variance by ranks serves as an alternative. These tests collectively provide researchers with flexible tools for analyzing diverse data types without stringent distributional assumptions.