ANOVA vs. Kruskal-Wallis: A Direct Comparison

Overview

This unit examines how one-way ANOVA (parametric) and Kruskal-Wallis ANOVA (non-parametric) compare — not just conceptually, but computationally, including the remarkable mathematical equivalence between the F and H statistics when applied to the same data.

---

3.7 Full Comparison

Conceptual Comparison Table

Feature	One-Way ANOVA	Kruskal-Wallis ANOVA
What it compares	Group means	Group medians
Data type	Interval/ratio	Ordinal or above
Normality required	Yes	No
Homogeneity of variance	Required (explicit)	Implied by identical shapes
Test statistic	F ratio	H statistic
Critical value table	F distribution	Table H (small n) or χ² (large n)
Significance direction	F ≥ F-critical	H ≥ critical value
Post-hoc tests	Tukey, Scheffé, Bonferroni	Dunn's test, pairwise MW-U
Outlier sensitivity	High (mean-based)	Low (rank-based)
Special case at k = 2	Equivalent to independent t-test	Equivalent to Mann-Whitney U
Developed by	R.A. Fisher	Kruskal & Wallis (1952)

---

Worked Comparison: Same Data, Both Tests

Research Scenario

Three groups perform a task. We compare their scores using both tests.

Data:

Group 1	Group 2	Group 3
3	6	9
5	7	10
6	8	11
7	9	12
8	10	15
9	10	—
—	11	—

Sample sizes: $n_1 = 6$, $n_2 = 7$, $n_3 = 5$, $N = 18$

---

One-Way ANOVA Calculation

Step 1: Group Sums

$\Sigma X_1 = 38, \quad \Sigma X_2 = 61, \quad \Sigma X_3 = 57$

$\Sigma X_1^2 = 264, \quad \Sigma X_2^2 = 551, \quad \Sigma X_3^2 = 671$

Step 2: Sum of Squares

$SS_{total} = (264 + 551 + 671) - \frac{(38+61+57)^2}{18} = 1486 - 1352 = 134$

$SS_{between} = \frac{38^2}{6} + \frac{61^2}{7} + \frac{57^2}{5} - \frac{156^2}{18} = 70.038$

$SS_{within} = SS_{total} - SS_{between} = 63.962$

Step 3: ANOVA Table

Source	SS	df	MS	F
Between Groups	70.038	2	35.019	8.213
Within Groups	63.962	15	4.264	—
Total	134.000	17	—	—

Critical F at α = .05, df = (2, 15): 3.68

Since F = 8.213 > 3.68: Reject H₀ — the three groups differ significantly.

---

Kruskal-Wallis Calculation (Same Data)

Step 1: Rank all 18 scores

Group 1	Rank	Group 2	Rank	Group 3	Rank
3	1	6	3.5	9	10
5	2	7	5.5	10	13
6	3.5	8	7.5	11	15.5
7	5.5	9	10	12	17
8	7.5	10	13	15	18
9	10	10	13	—	—
—	—	11	15.5	—	—
T₁ = 29.5		T₂ = 68		T₃ = 73.5

Step 2: Compute H

$H = \frac{12}{18(19)} \left[ \frac{29.5^2}{6} + \frac{68^2}{7} + \frac{73.5^2}{5} \right] - 3(19)$

$H = \frac{12}{342} \left[ 145.04 + 660.57 + 1081.35 \right] - 57$

$H = 0.0351 \times 1886.96 - 57 = 66.22 - 57 = \mathbf{9.22}$

Critical χ² at α = .05, df = 2: 5.99

Since H = 9.22 > 5.99: Reject H₀ — the three groups differ significantly.

Both tests reach the same conclusion: Group 3 > Group 2 > Group 1.

---

The Mathematical Equivalence of F and H

Iman and Conover (1981) demonstrated that F and H are mathematically related through the rank transform principle:

$F = \left[ \frac{k-1}{N-k} \cdot \left( \frac{N-1}{H} - 1 \right) \right]^{-1}$

Verification with Our Example

We found F = 8.213 and H = 9.22.

$F = \left[ \frac{2}{15} \cdot \left( \frac{17}{9.22} - 1 \right) \right]^{-1}$

$F = \left[ 0.1333 \times (1.843 - 1) \right]^{-1} = \left[ 0.1333 \times 0.843 \right]^{-1} = 0.1124^{-1} \approx 8.90$

(Minor discrepancy due to rounding; both reject H₀ consistently.)

This equivalence illustrates that Kruskal-Wallis is literally ANOVA applied to ranks — the rank transform bridges the parametric and non-parametric worlds.

---

When the Two Tests Disagree

Both tests will usually agree in their conclusion. However, discrepancies can occur when:

1. Data is heavily skewed: ANOVA may be more influenced by extreme scores; Kruskal-Wallis is more robust
2. Outliers are present: ANOVA F can be inflated or deflated by outliers; H is resistant
3. Variance heterogeneity: ANOVA is more sensitive to unequal variances; H is less so
4. Very small samples: Exact small-sample corrections make a larger difference

---

Decision Framework: ANOVA or Kruskal-Wallis?

Indian Research Context

In Indian psychological research — particularly in clinical, educational, and rural community settings — Kruskal-Wallis is frequently the preferred choice because:

• Small institutional samples (hospital wards, school cohorts) rarely exceed 20–30 per group
• Likert-scale data is ubiquitous in personality, attitude, and wellbeing research
• Non-normal distributions are the rule rather than exception in applied settings
• Studies comparing groups such as income brackets, caste categories, or urban/rural/tribal populations often have naturally ordinal or skewed measures

---

Post-Hoc Testing After a Significant Kruskal-Wallis Result

A significant H tells you only that at least one pair of groups differs. To identify which pairs differ, use:

Method	Description	When to Use
Dunn's test	Pairwise comparisons with Bonferroni-style correction	Standard choice; available in SPSS
Pairwise Mann-Whitney U with Bonferroni	Each pair tested with U; α divided by number of comparisons	Simple and transparent
Steel-Dwass test	More powerful than Bonferroni correction	When many comparisons are needed

Bonferroni correction formula: Adjusted α = .05 / number of pairwise comparisons

For k = 3 groups: 3 pairs → adjusted α = .05/3 = .0167
For k = 4 groups: 6 pairs → adjusted α = .05/6 = .0083

---

Self-Assessment Questions

Level 1 — Recall

1. In what direction is the Kruskal-Wallis H test significant: when H is larger or smaller than the critical value?

2. What is the degrees of freedom for the chi-square approximation of H when k = 4?

3. State the mathematical relationship between F and H as demonstrated by Iman and Conover (1981).

Level 2 — Application

4. Using the formula $F = \left[\frac{k-1}{N-k} \cdot \left(\frac{N-1}{H} - 1\right)\right]^{-1}$, compute F if k = 3, N = 21, and H = 7.5. Is this F significant at α = .05 with df = (2, 18)? (Critical F ≈ 3.55)

5. A researcher runs a Kruskal-Wallis test across 4 groups and finds H = 10.2, df = 3. Is this significant at α = .05? (Critical χ² = 7.815.) If significant, what post-hoc test would you recommend, and what would the Bonferroni-adjusted α be?

Level 3 — Critical Thinking

6. Two researchers analyse the same dataset. Researcher A uses one-way ANOVA (F = 4.1, p = .031, significant). Researcher B uses Kruskal-Wallis (H = 3.8, p = .052, not significant). Both use α = .05. How is this possible? What does this reveal about the relationship between the two tests?

7. A colleague argues: "Kruskal-Wallis is always preferable to ANOVA because it makes fewer assumptions." Construct a counter-argument explaining when one-way ANOVA is the better choice. Include statistical power considerations in your answer.

---

🧠 Memory Aid

"HAWK soars above both F and H"

• H = H statistic (Kruskal-Wallis)
• A = ANOVA equivalent (when rank-transformed)
• W = Wallis (one of the developers)
• K = Kruskal (the other developer)

The F-H relationship: Think of it as "F on ranks ≈ H" — ranking the data first turns an ANOVA into a Kruskal-Wallis.

Post-hoc reminder: Significant K-W → use Dunn's test (named for the statistician Olive Jean Dunn — "Done with Dunn's test after K-W").

---

📚 External Resources

Foundational Reading

• 🌐 Wikipedia: Kruskal-Wallis test — Post-hoc analysis — Dunn's test and other follow-ups
• 🌐 Wikipedia: Rank-based statistics — Overview of the rank transform principle

Tutorials

• 📖 Statistics How To: Post Hoc Tests for Kruskal-Wallis — Dunn's test and Bonferroni correction explained
• 📖 SPSS: Kruskal-Wallis with Post-Hoc — Complete SPSS tutorial including pairwise comparisons
• 📖 UCLA OARC: Non-Parametric Analysis in SPSS — Decision guide

Videos

• 🎥 ANOVA vs. Kruskal-Wallis — When to Use Which — Conceptual comparison
• 🎥 Post-hoc Tests after Kruskal-Wallis in SPSS — Pairwise comparison tutorial

Research Articles

• 📄 Iman, R. L., & Conover, W. J. (1981). Rank transformations as a bridge between parametric and nonparametric statistics. The American Statistician, 35(3), 124–129. (F-H equivalence)
• 📄 Zimmerman, D. W., & Zumbo, B. D. (1993). The relative power of parametric and nonparametric statistical methods. A handbook for data analysis in the behavioral sciences, 1, 481–517.
• 📄 Dunn, O. J. (1964). Multiple comparisons using rank sums. Technometrics, 6(3), 241–252. (Original Dunn's test paper)

---

Source PDFs:

• 📄 Block-4/Unit-3.pdf — Pages 43–46
• 📚 MPC-006 Statistics in Psychology