Kruskal-Wallis Test: Formula, Examples, Calculator & Interpretation

What Is the Kruskal-Wallis Test?

Named for William Kruskal and W. Allen Wallis, who published it in 1952, the test works by replacing raw observations with their ranks within the combined dataset. Instead of asking "are the group means different?", it asks "are the rank distributions different?" That shift makes it useful for ordinal data, skewed distributions, and situations where small samples make normality assumptions difficult to verify.

One thing to keep in mind from the start: calling the Kruskal-Wallis test simply "the nonparametric ANOVA" is a common shorthand that can mislead. ANOVA compares means directly. The Kruskal-Wallis test compares rank distributions. A median interpretation is reasonable when the group distributions share a similar shape, but when shapes differ, a significant result may reflect a difference in spread or skewness rather than location. Parametric vs. nonparametric tests covers these distinctions in depth.

When to Use the Kruskal-Wallis Test

The Kruskal-Wallis test is appropriate when all of the following conditions point toward a rank-based approach:

The Kruskal-Wallis test does not have a universal minimum sample size that guarantees validity. For small samples where the chi-square approximation may be imprecise, exact methods exist and some software implementations offer them. For most practical applications, a few observations per group is sufficient for the asymptotic approximation to perform adequately, but researchers should verify this for their specific situation.

Kruskal-Wallis Test Hypotheses

The null hypothesis states that all groups share the same underlying distribution. When the distributions have similar shapes, this is equivalent to saying all groups have the same population median (or, more loosely, the same central tendency). The alternative states that at least one group differs, not that all groups are mutually distinct.

A significant Kruskal-Wallis result tells you that the groups do not all behave the same way statistically, but it does not tell you which ones differ. That requires post-hoc testing, covered below.

Kruskal-Wallis H Statistic Formula

After ranking all N observations across the combined dataset and calculating the rank sum Rᵢ for each group, the H statistic is:

The factor 12 / [N(N+1)] normalizes the rank sums so that H follows (approximately) a chi-square distribution with k−1 degrees of freedom when the null hypothesis is true and sample sizes are not extremely small. Larger H values indicate greater separation in the rank distributions across groups.

Tie Correction

When observations share the same value, they receive the average of the ranks they would occupy if they were distinct. This averaging reduces the variability of ranks and inflates H slightly relative to what the uncorrected formula gives. A tie correction factor C corrects for this:

The tie correction has minimal effect when the proportion of tied observations is small. Most statistical software — including SPSS, R's kruskal.test(), and scipy.stats.kruskal in Python — applies the tie correction automatically. When ties are extensive, they can indicate that a different measurement approach (capturing more granular data) might be worth considering.

Degrees of Freedom and P-Value

The degrees of freedom for the Kruskal-Wallis test are:

For four groups, df = 3. The p-value is the probability of observing H ≥ h (or a more extreme value) under the null hypothesis, evaluated against the chi-square distribution with df = k−1. When p < α, reject H₀. When p ≥ α, there is not sufficient evidence to conclude that the groups differ. Failure to reject H₀ is not the same as proving that the groups are identical.

Worked Examples — Step-by-Step Manual Calculations

Each example below follows a consistent structure and shows every calculation. Data values are kept small enough for manual verification. The methodology follows the approach described by NIST.

Example 1 — Basic Three-Group Comparison

Example 2 — Four Groups with Tied Ranks

Example 3 — Unequal Sample Sizes

Example 4 — Non-Significant Result

Example 5 — Medical Research Context

Kruskal-Wallis Test Calculator

Enter raw data values for each group (comma-separated). The calculator ranks all observations together, handles tied ranks with the tie correction, computes the H statistic, degrees of freedom, p-value, effect size (ε²), and optionally runs pairwise Dunn's test with Bonferroni or Holm correction.

Calculations run entirely in your browser. No data is transmitted or stored.

Post-Hoc Testing After a Significant Kruskal-Wallis Test

A significant Kruskal-Wallis result answers one question: do the groups differ? It does not answer which groups differ. For that, pairwise post-hoc comparisons are necessary. Skipping post-hoc analysis and concluding "Group A and Group B are different" from a significant omnibus test alone is a common statistical error.

Dunn's Test

Dunn's test (Dunn, 1964) is the most widely recommended post-hoc procedure following a significant Kruskal-Wallis test. It computes pairwise z-statistics using the same rank information as the omnibus test, comparing each pair of groups:

Multiple-Comparison Correction

When making multiple pairwise comparisons simultaneously, the probability of a false positive increases with each comparison. For k groups, there are k(k−1)/2 pairs. A multiple-comparison correction controls the family-wise error rate or false discovery rate.

The Holm correction is generally preferred over Bonferroni because it is uniformly more powerful while providing the same FWER control. The choice between FWER-controlling and FDR-controlling methods depends on the research context and the consequences of false positives versus missed effects.

Effect Size for the Kruskal-Wallis Test

A statistically significant p-value tells you the result is unlikely under H₀. It does not tell you how large or practically meaningful the difference is. Effect size addresses that question.

Epsilon Squared (ε²)

One commonly used effect size estimator for the Kruskal-Wallis test is epsilon squared:

Regardless of which convention you apply, report the effect size alongside the p-value. A large sample can yield statistical significance for a trivially small group difference, and an effect-size measure helps readers judge practical importance.

Does the Kruskal-Wallis Test Compare Medians?

This distinction matters for reporting. Rather than routinely saying "the test compared medians," state that the test compared rank distributions, then note the distributional assumption required for a median interpretation. If group distributions do appear similarly shaped (something box plots or violin plots can help assess visually), the median interpretation is reasonable and should be supported by reporting medians as the descriptive summary.

Kruskal-Wallis vs. Other Statistical Tests

Kruskal-Wallis vs. One-Way ANOVA

When ANOVA's assumptions hold, ANOVA has greater statistical power. When data are ordinal or clearly non-normal with smaller samples, the Kruskal-Wallis test is the more defensible choice. See the one-way ANOVA guide for the parametric approach.

Kruskal-Wallis vs. Mann-Whitney U

The Mann-Whitney U test is designed for exactly two independent groups. The Kruskal-Wallis test extends the rank-based comparison to three or more groups. Running multiple pairwise Mann-Whitney U tests in place of a Kruskal-Wallis omnibus test inflates the family-wise error rate and is generally not recommended as the primary analysis when comparing three or more groups simultaneously. After a significant Kruskal-Wallis result, Dunn's test (rather than raw Mann-Whitney comparisons) should be used for post-hoc pairwise testing.

Kruskal-Wallis vs. Friedman Test

Confusing these two tests is a common error. If the same participants appear in each condition, or if observations are matched (for example, litters of animals, matched patient pairs, or multiple measurements from the same individual), the Friedman test is appropriate and the Kruskal-Wallis test is not.

Kruskal-Wallis Test in SPSS, R, and Python

SPSS

In SPSS, the Kruskal-Wallis test is found under Analyze → Nonparametric Tests → Legacy Dialogs → K Independent Samples. Set the test variable (your dependent/outcome variable) and the grouping variable, then define the range of group codes. Under "Test Type," ensure Kruskal-Wallis H is selected. In newer SPSS versions (Statistics 21+), the same test is also accessible through Analyze → Nonparametric Tests → Independent Samples with a more automated interface.

The SPSS output includes the H statistic, degrees of freedom, and asymptotic p-value. Post-hoc pairwise comparisons are available through the newer interface under "Compare Groups." Check the "Mean ranks" table to see which groups had higher or lower average ranks.

R

# Basic Kruskal-Wallis test
result <- kruskal.test(score ~ group, data = mydata)
print(result)

# Returns: Kruskal-Wallis chi-squared, df, p-value

# Post-hoc Dunn's test with Holm correction
# Requires: install.packages("dunn.test") or rstatix
library(dunn.test)
dunn.test(mydata$score, mydata$group, method = "holm")

# With rstatix (tidyverse-friendly)
library(rstatix)
mydata %>%
  kruskal_test(score ~ group) # omnibus
mydata %>%
  dunn_test(score ~ group, p.adjust.method = "holm") # post-hoc

# Effect size (epsilon squared)
mydata %>%
  kruskal_effsize(score ~ group)

R's kruskal.test() applies the tie correction automatically. The result object contains the H statistic (labeled as statistic), degrees of freedom, and p-value. For full reproducibility, always report the R version and package versions used. Documentation for the base function is available from The R Foundation.

Python

from scipy import stats
import scikit_posthocs as sp  # pip install scikit-posthocs
import numpy as np

# Your group data as separate arrays
group_a = [65, 72, 68]
group_b = [80, 85, 78]
group_c = [55, 60, 58]

# Kruskal-Wallis test (tie correction applied automatically)
H, p = stats.kruskal(group_a, group_b, group_c)
print(f"H = {H:.4f}, p = {p:.4f}")

# Degrees of freedom
k = 3  # number of groups
df = k - 1

# Epsilon squared effect size
N = len(group_a) + len(group_b) + len(group_c)
epsilon_sq = (H - k + 1) / (N - k)
print(f"ε² = {epsilon_sq:.4f}")

# Post-hoc Dunn's test with Holm correction
data_all = [group_a, group_b, group_c]
posthoc = sp.posthoc_dunn(data_all, p_adjust='holm')
print(posthoc)

scipy.stats.kruskal accepts two or more array-like arguments and returns the tie-corrected H statistic and p-value. The scikit-posthocs library provides Dunn's test and several other post-hoc methods. Documentation for scipy.stats.kruskal is maintained at SciPy.org.

Excel

Excel does not include a built-in Kruskal-Wallis function. You can implement it manually using Excel's RANK.AVG function to assign average ranks to tied values, then apply the H formula with SUMPRODUCT. For regular use, dedicated statistical software (R, Python, SPSS, or JASP) is more practical and less error-prone. JASP is a free open-source alternative with a point-and-click interface that includes the Kruskal-Wallis test and Dunn's post-hoc analysis.

How to Report the Kruskal-Wallis Test

Report the H statistic with its degrees of freedom, the sample sizes per group, the p-value, and a measure of effect size. Include medians and interquartile ranges as descriptive statistics (rather than means and standard deviations) since the test is rank-based. If post-hoc comparisons were conducted, report those results with the correction method used.

Common Mistakes to Avoid

Real-World Applications

Practice Problems

Work through these problems to test your understanding. Answers follow each problem block.

Frequently Asked Questions

Formula Reference Sheet