3.5.2 Kruskal-Wallis

Kruskal-Wallis Introduction The Kruskal-Wallis test is a nonparametric statistical test used to compare three or more independent groups on a continuous or ordinal outcome. It is often used as a substitute for one-way ANOVA when data do not meet ANOVA’s assumptions, especially normality and equal variances. Kruskal-Wallis is based on ranks rather than raw data, making it robust against outliers and non-normal distributions. It tests whether the groups come from the same population distribution. --- Purpose of the Kruskal-Wallis Test When to Use Kruskal-Wallis Use Kruskal-Wallis when: - Objective: Compare the central tendency of 3+ independent groups. - Data type: Ordinal or continuous data that may be skewed or non-normal. - Assumptions for parametric ANOVA are doubtful or clearly violated. - Design: One factor (one-way), with independent samples. Typical examples: - Comparing customer satisfaction scores (Likert scales) across several branches. - Comparing cycle times across several process designs when data are skewed. - Comparing defect detection scores across more than two training methods. Relationship to Other Tests - Alternative to one-way ANOVA: - ANOVA compares means assuming normality and equal variances. - Kruskal-Wallis compares distributions using ranks and is less sensitive to these assumptions. - Extension of Mann-Whitney: - Mann-Whitney (Wilcoxon rank-sum) compares 2 independent groups. - Kruskal-Wallis generalizes this concept to 3 or more independent groups. --- Assumptions and Data Requirements Core Assumptions For valid results, Kruskal-Wallis requires: - Independent groups: - Each observation belongs to only one group. - No subject appears in more than one group. - Independent observations: - Measurements from different participants or items are not correlated. - Ordinal or higher data: - You must be able to rank the data meaningfully. - Interval and ratio data are acceptable; nominal data are not. - Similar distribution shapes: - Groups should have similarly shaped distributions. - If this holds, Kruskal-Wallis mainly compares medians. - If not, it signals differences in overall distributions, not just central tendency. When Not to Use Kruskal-Wallis Avoid Kruskal-Wallis when: - You have paired or repeated measures data (within-subjects design); use a repeated-measures nonparametric method instead. - Groups are not independent (clustered in ways that create correlation). - Data are purely categorical without an inherent order. --- Conceptual Basis: Ranking and H Statistic Ranking the Data Kruskal-Wallis converts raw data to ranks: - Combine all observations from all groups into a single list. - Rank them from smallest (rank 1) to largest (rank N, where N is total sample size). - For ties, assign each tied value the average of the ranks they would occupy. This rank transformation: - Reduces the impact of non-normality and outliers. - Makes the test robust when distributions are skewed. The H Test Statistic Kruskal-Wallis uses a test statistic called H. Conceptually: - It compares the sum of ranks in each group to what would be expected if all groups came from the same population. - If group medians (or distributions) are similar, rank sums should be similar. - Larger differences in rank sums lead to larger H values. H is approximately chi-square distributed under the null hypothesis when sample sizes are not too small. --- Hypotheses, Interpretation, and Decision Rules Hypotheses - Null hypothesis (H₀): - All groups come from the same population distribution. - Equivalently, there is no difference in distribution location (medians), assuming similar shapes. - Alternative hypothesis (H₁): - At least one group’s distribution differs from at least one other group. Kruskal-Wallis is a global test: it tells you that differences exist but not which groups differ. Significance and Decision After calculating H (and possibly a tie-corrected version): - Determine the degrees of freedom: - df = k − 1, where k is the number of groups. - Obtain the p-value using the chi-square distribution with df = k − 1. - Decision rule: - If p-value ≤ α (commonly 0.05): reject H₀; at least one group differs. - If p-value > α: fail to reject H₀; no statistically significant difference detected. The test is directional only in the sense of “some difference exists,” not “which is higher.” --- Manual Calculation Steps Step 1: Organize the Data - List all observations with their group labels. - Let: - k = number of groups. - nᵢ = sample size of group i. - N = total sample size = n₁ + n₂ + … + n_k. Step 2: Rank All Observations - Sort all N values from smallest to largest. - Assign ranks 1 through N. - For ties: - Compute the average of the ranks that would be assigned if there were no ties. - Assign that average rank to each tied observation. Step 3: Compute Rank Sums and Means per Group For each group i: - Sum of ranks: Rᵢ. - Mean rank: R̄ᵢ = Rᵢ / nᵢ. These rank sums and mean ranks show which groups tend to have higher or lower values. Step 4: Compute the H Statistic (No Tie Correction) If there are no ties, use: H = [12 / (N(N + 1))] × Σ [Rᵢ² / nᵢ] − 3(N + 1) where the sum Σ is over all groups i = 1 to k. Step 5: Correct for Ties (When Present) Ties affect the distribution of ranks. Use a correction factor C: - Let: - For each tied group of values j, tⱼ is the number of tied observations. - Compute: C = 1 − [Σ (tⱼ³ − tⱼ)] / [N³ − N] Then use the tie-corrected H: H_corrected = H / C Use H_corrected when comparing to the chi-square distribution. Step 6: Obtain p-Value and Compare to Chi-Square - Degrees of freedom: df = k − 1. - Use the chi-square distribution with df = k − 1 to get the p-value for H_corrected. - Compare p-value to the chosen significance level α. For typical process improvement applications, the chi-square approximation is adequate when all group sizes are reasonably large (commonly all nᵢ ≥ 5, preferably ≥ 7–10). --- Post-Hoc Analysis and Multiple Comparisons Need for Post-Hoc Tests If Kruskal-Wallis is significant: - You know at least one group differs. - To identify which groups differ, perform pairwise comparisons. Kruskal-Wallis alone does not specify the pattern of differences. Common Nonparametric Pairwise Methods For pairwise follow-up comparisons: - Mann-Whitney (Wilcoxon rank-sum) on each pair of groups: - Compare Group A vs Group B, Group A vs Group C, etc. - Use the same rank-based logic as Kruskal-Wallis for two groups. - Dunn’s test (if available in software): - Specifically designed as a post-hoc test for Kruskal-Wallis. - Uses a standardized difference in mean ranks, with adjustment for multiple comparisons. Multiple Comparison Adjustments Because multiple pairwise tests inflate Type I error: - Adjust p-values or significance levels using methods such as: - Bonferroni adjustment: divide α by the number of comparisons. - Other corrections may be available in software interfaces. Key idea: ensure that your overall error rate remains controlled when testing several pairs. --- Effect Size and Practical Significance Why Effect Size Matters A statistically significant Kruskal-Wallis result does not automatically imply a large or meaningful difference. Effect size helps quantify how substantial the difference is. Common Effect Size Measures For Kruskal-Wallis, you can use: - Eta-squared (η²) approximation: - η² ≈ (H − k + 1) / (N − k) - Interpretation (rough guidance, not strict rules): - Around 0.01: small effect - Around 0.06: medium effect - Around 0.14: large effect - Epsilon-squared (ε²): - ε² = (H − k + 1) / (N² − 1) - Similar interpretation, but slightly different scaling. Choosing between η² and ε² is often dictated by convention or software output. The key is to interpret the magnitude in context of the process. --- Assumption Checks and Practical Diagnostics Distribution Shape Similarity Kruskal-Wallis is often described as comparing medians, but this is precise only when distributions have similar shapes and spreads. To check this: - Compare: - Group medians - Spreads (e.g., interquartile ranges) - Shapes (e.g., skew direction) If group distributions differ strongly in shape (not just location), then: - A significant Kruskal-Wallis result indicates at least one group distribution is different overall. - Interpretation should emphasize distributional differences, not just medians. Sample Size Considerations - Larger sample sizes: - Make the chi-square approximation more accurate. - Provide higher power to detect meaningful differences. - Very small sample sizes: - The approximation may be less accurate. - Exact methods may be needed, but are often handled automatically by software. --- Practical Implementation with Statistical Software Typical Software Inputs Most statistical packages or tools require you to specify: - Response variable: - The measurement you are comparing (e.g., time, score, rating). - Factor or group variable: - The categorical variable defining the groups (e.g., method, location, treatment). Typical Outputs Expect to see: - H statistic (or sometimes just labeled as test statistic). - Degrees of freedom (df). - p-value. - Mean ranks per group or rank sums. - Post-hoc test results (if requested). Focus interpretation on: - Whether p-value ≤ α. - Direction and magnitude of differences in mean ranks. - Effect size and practical impact on the process or system being studied. --- Common Pitfalls and Misinterpretations Misusing Kruskal-Wallis Avoid these errors: - Using Kruskal-Wallis with paired data: - It is not designed for repeated measures or matched pairs. - Treating nominal categories as ordered: - If categories have no natural order, do not treat them as ordinal. - Ignoring distribution shape differences: - Assuming it compares medians without checking whether shapes are similar. Over-Reliance on p-Values - A significant p-value: - Indicates evidence of a difference. - Does not describe how large or important the difference is. Combine Kruskal-Wallis results with: - Effect size estimates. - Contextual knowledge of the process. - Visual data summaries where possible (e.g., boxplots). --- Summary Kruskal-Wallis is a nonparametric, rank-based test used to compare three or more independent groups on an ordinal or continuous outcome when assumptions of one-way ANOVA are not met. It: - Uses ranks to reduce sensitivity to non-normality and outliers. - Tests the null hypothesis that all groups come from the same distribution. - Produces an H statistic, approximately chi-square distributed with df = k − 1. - Requires independent groups, independent observations, and at least ordinal data. - Is most interpretable as a test of median differences when group distributions share similar shapes. - Signals the presence of differences but requires post-hoc pairwise tests to identify which groups differ. - Benefits from effect size measures to assess practical significance. Mastering Kruskal-Wallis involves understanding when to use it, how it is calculated and interpreted, how to conduct and adjust post-hoc analyses, and how to connect statistical findings to meaningful conclusions about group differences.