One-Way Analysis of Variance (ANOVA)

Unit Overview

Course: MPC-006 Statistics in Psychology · Block 3 · Unit 3
Pages: 77–95 · 📄 Download PDF
Prerequisite: t-Test for Independent & Correlated Samples | Significance of Difference Between Two Means

---

Why ANOVA? The Problem with Multiple t-Tests

Suppose you are comparing the effectiveness of three teaching methods across three groups of students. Your instinct might be to run multiple t-tests: Method A vs B, Method B vs C, Method A vs C — three tests in total. With eight groups this balloons to 28 comparisons. Two serious problems emerge:

1. Inflated Type I Error (Alpha Inflation)
Each t-test carries a 5% chance of a false positive at α = .05. Running three tests means the overall probability of making at least one false rejection is approximately:

$1 - (1 - 0.05)^3 = 1 - 0.857 = 0.143 \approx 14\%$

The error rate has nearly tripled. With 28 tests it approaches 76% — virtually guaranteeing a spurious finding. This phenomenon is called familywise error rate (FWER) inflation.

2. t-Test Ignores Within-Group Variance
Consider this dataset from the PDF (p. 78):

Group	Scores	Sum	Mean
Girls	15, 20, 5, 10, 35	85	17 ms
Boys	20, 15, 20, 20, 10	85	17 ms

Both groups have identical means (17 ms). A t-test would find no significant difference. Yet girls show extreme variability (range 5–35) while boys are tightly clustered (range 10–20). The groups differ meaningfully, but the t-test misses this because it only compares means — not the spread within each group.

ANOVA solves both problems by testing all group differences simultaneously, using the ratio of between-group variance to within-group variance as its test statistic — the F-ratio.

Historical Note

ANOVA was invented by Sir Ronald A. Fisher (1890–1962), widely regarded as the father of modern statistics. He first published the method in 1923, originally applying it to agricultural field experiments. The "F" in F-ratio stands for Fisher. His foundational text Statistical Methods for Research Workers (1925) transformed hypothesis testing across the sciences. Wikipedia: Ronald Fisher

---

The Logic of ANOVA

ANOVA partitions the total variance in a dataset into two independent components:

$SS_{Total} = SS_{Between} + SS_{Within}$

Between-groups variance (SS_A) captures differences among group means — this reflects the treatment effect plus random error.

Within-groups variance (SS_W) captures variability inside each group — this is pure random/error variance (individual differences, measurement error).

If the treatment truly has an effect, the between-groups variance will be large relative to within-groups variance. The F-ratio quantifies this:

$F = \frac{MSS_A \text{ (Between-Groups Mean Square)}}{MSS_W \text{ (Within-Groups Mean Square)}}$

When $F \approx 1$, between-group variation is no larger than random noise — no treatment effect. When $F \gg 1$, the groups differ beyond what chance can explain.

---

Variance: The Foundation

Before computing ANOVA, you must have a solid grasp of variance — the measure at the heart of the analysis.

Standard Deviation (σ) measures spread as average distance from the mean, in the same units as the data. Variance (σ²) is its square:

$\sigma^2 = \frac{\sum(X - M)^2}{N}$

Or using degrees of freedom (sample variance):

$s^2 = \frac{\sum(X - M)^2}{N - 1}$

Key Characteristics of Variance

Property	Explanation
Always positive (≥ 0)	Squaring eliminates negative deviations
Units are squared	Variance in ms² if data is in ms
Unaffected by adding/subtracting a constant	$Var(X + c) = Var(X)$
Additive for independent components	$Var(A + B) = Var(A) + Var(B)$
σ = √σ²	Standard deviation recovers original units

SD vs Variance

SD has direction (like length): it tells you how far scores typically stray from the mean.
Variance is an area (length²): it lives on the normal curve as a surface, not a distance. ANOVA works with variances — not SDs — precisely because variances add (SDs do not).

---

Step-by-Step ANOVA Procedure

Notation

Symbol	Meaning
$k$	Number of groups
$n$	Cases per group (equal groups)
$N$	Total cases = $n \times k$
$\sum X_a, \sum X_b ...$	Group sums
$\sum X^2$	Sum of all squared scores
$M_a, M_b ...$	Group means
$Cx$	Correction term

The 8-Step Formula Sequence

Step 1: Correction Term $Cx = \frac{(\sum X)^2}{N}$

Step 2: Total Sum of Squares $SS_T = \sum X^2 - Cx$

Step 3: Between-Groups Sum of Squares $SS_A = \frac{(\sum X_a)^2}{n_a} + \frac{(\sum X_b)^2}{n_b} + \cdots + \frac{(\sum X_k)^2}{n_k} - Cx$

Step 4: Within-Groups Sum of Squares $SS_W = SS_T - SS_A$

Step 5: Between-Groups Mean Square $MSS_A = \frac{SS_A}{k - 1}$

Step 6: Within-Groups Mean Square $MSS_W = \frac{SS_W}{N - k}$

Step 7: F-Ratio $F = \frac{MSS_A}{MSS_W}$

Step 8: ANOVA Summary Table

Source	df	SS	MSS	F
Between Groups	$k-1$	$SS_A$	$MSS_A$	$F$
Within Groups (Error)	$N-k$	$SS_W$	$MSS_W$	—
Total	$N-1$	$SS_T$	—	—

---

Worked Example 1: Intelligence Across Streams

Source: PDF p. 83–85

Three groups of 5 students each (Arts, Commerce, Science) took an intelligence test. Do the streams differ?

Data:

S.No.	Arts ($X_1$)	Commerce ($X_2$)	Science ($X_3$)
1	15	12	12
2	14	14	15
3	11	10	14
4	12	13	10
5	10	11	10
Σ	62	60	61
M	12.40	12.00	12.20

$k = 3$, $n = 5$, $N = 15$
$H_0$: $\mu_{Arts} = \mu_{Commerce} = \mu_{Science}$

Step 1: $Cx = \frac{(62+60+61)^2}{15} = \frac{183^2}{15} = \frac{33489}{15} = 2232.60$

Step 2: $\sum X^2 = (225+196+121+144+100) + (144+196+100+169+121) + (144+225+196+100+100)$
$= 786 + 730 + 765 = 2281$
$SS_T = 2281 - 2232.60 = 48.40$

Step 3: $SS_A = \frac{62^2}{5} + \frac{60^2}{5} + \frac{61^2}{5} - 2232.60$
$= \frac{3844+3600+3721}{5} - 2232.60 = 2233.00 - 2232.60 = 0.40$

Step 4: $SS_W = 48.40 - 0.40 = 48.00$

Step 5: $MSS_A = \frac{0.40}{3-1} = \frac{0.40}{2} = 0.20$

Step 6: $MSS_W = \frac{48.00}{15-3} = \frac{48.00}{12} = 4.00$

Step 7: $F = \frac{0.20}{4.00} = 0.05$

Summary Table:

Source	df	SS	MSS	F
Between Groups	2	0.40	0.20	0.05
Within Groups	12	48.00	4.00	—
Total	14	48.40	—	—

Decision: $F_{critical}$ for df₁ = 2, df₂ = 12 at .05 level = 3.88. Obtained F = 0.05 < 3.88.
Retain H₀. The three streams do not differ significantly in intelligence.

---

Worked Example 2: Four Drugs & Physical Growth (with Post-Hoc)

Source: PDF p. 85–88

20 rats randomly assigned to 4 drug groups (5 each). Gain in weight (oz) after 1 month:

S.No.	Drug P	Drug Q	Drug R	Drug S
1	4	9	2	7
2	5	10	6	7
3	1	9	6	4
4	0	6	5	2
5	2	6	2	7
Σ	12	40	21	27
M	2.40	8.00	4.20	5.40

$k = 4$, $n = 5$, $N = 20$
$H_0$: $\mu_P = \mu_Q = \mu_R = \mu_S$

Cx $= \frac{(12+40+21+27)^2}{20} = \frac{100^2}{20} = \frac{10000}{20} = 500$

$SS_T$ $= (46+334+105+167) - 500 = 652 - 500 = 152$

$SS_A$ $= \frac{12^2+40^2+21^2+27^2}{5} - 500 = \frac{144+1600+441+729}{5} - 500 = \frac{2914}{5} - 500 = 582.80 - 500 = 82.80$

$SS_W$ $= 152 - 82.80 = 69.20$

$MSS_A$ $= \frac{82.80}{4-1} = \frac{82.80}{3} = 27.60$

$MSS_W$ $= \frac{69.20}{20-4} = \frac{69.20}{16} = 4.325$

$F$ $= \frac{27.60}{4.325} = 6.38$

Source	df	SS	MSS	F
Between Groups	3	82.80	27.60	6.38
Within Groups	16	69.20	4.325	—
Total	19	152.00	—	—

Decision: $F_{.01}$ for df₁ = 3, df₂ = 16 = 5.29. Obtained F = 6.38 > 5.29.
Reject H₀ at .01 level. The four drugs are not equally effective.

Post-Hoc Comparisons (Paired t-Tests)

Since F is significant, we use t-tests to identify which drugs differ:

$t = \frac{M_1 - M_2}{SE_{DM}}$

$SE_{DM} = \sqrt{MSS_W} \cdot \sqrt{\frac{1}{n_1} + \frac{1}{n_2}} = \sqrt{4.325} \cdot \sqrt{\frac{2}{5}} = 2.079 \times 0.632 = 1.31$

Comparison	Difference	t-value	df	Significant?
P vs Q	5.60	4.28	16	✅ .01 level
P vs R	1.80	1.37	16	❌ n.s.
P vs S	3.00	2.29	16	✅ .05 level
Q vs R	3.80	2.90	16	✅ .05 level
Q vs S	2.60	1.98	16	❌ n.s.
R vs S	1.20	0.92	16	❌ n.s.

Conclusion: Drug Q (mean = 8.00) is significantly superior to P and R. Drug Q and Drug S are approximately equivalent. Drug P is least effective.

---

ANOVA Assumptions

ANOVA is a parametric test built on four assumptions. Violating them can invalidate conclusions:

What happens when assumptions are violated?

Assumption	Violation Effect	Remedy
Normality	F-test becomes inaccurate with small N	Use Welch ANOVA or non-parametric Kruskal-Wallis
Homogeneity	F inflated or deflated	Welch's correction; Brown-Forsythe ANOVA
Independence	Severely biased results	Redesign study; use repeated measures ANOVA

ANOVA is generally robust to mild normality violations when group sizes are equal and large (N ≥ 30 per group), due to the central limit theorem. Homogeneity of variance is more critical — especially with unequal sample sizes.

Wikipedia: Homoscedasticity

---

F-Test and t-Test: The Mathematical Relationship

The F-test and t-test are not independent tests — they are mathematically linked:

$F = t^2 \quad \text{and} \quad t = \sqrt{F}$

This means when you have exactly two groups, the one-way ANOVA F-test gives identical results to the independent samples t-test. The F-ratio with df₁ = 1 equals the square of the t-value.

Why use ANOVA over multiple t-tests then?

Feature	Multiple t-Tests	One-Way ANOVA
Number of tests	$\binom{k}{2}$	1
Type I error control	Poor (inflated)	Excellent (familywise)
Within-group variance	Ignored	Explicitly modeled
Efficiency	Low	High
Post-hoc flexibility	N/A	Many options (Tukey, Scheffé, Bonferroni)

The relationship $F = t^2$ also confirms that within-group variance is embedded in the F-test in a way it never is in simple t-tests — that is why F is the preferred approach when comparing 3+ groups.

---

Using the F-Table

F-tables have two df dimensions:

• df₁ (columns) = $k - 1$ = between-groups df = "greater mean squares"
• df₂ (rows) = $N - k$ = within-groups df = "smaller mean squares"

Most tables show two values per cell: the normal-print value is for .05 level and the bold/dark-print value is for .01 level.

Reading the table (Example 1): df₁ = 2, df₂ = 12
→ Find column 2, row 12 → F₀₅ = 3.88, F₀₁ = 6.93
→ Obtained F = 0.05 → much less than 3.88 → not significant

Reading the table (Example 2): df₁ = 3, df₂ = 16
→ F₀₅ = 3.24, F₀₁ = 5.29
→ Obtained F = 6.38 → exceeds F₀₁ → significant at .01 level

Exam Shortcut

df₁ = (number of groups – 1)
df₂ = (total N – number of groups)
df_total = N – 1 = df₁ + df₂

---

Merits and Limitations

✅ Merits

• Simultaneous comparison of 3+ groups without inflating Type I error
• Evaluates both between-group and within-group variance — more complete picture than t-test
• Economical: one overall F-test replaces $\binom{k}{2}$ pairwise tests
• Foundation for complex designs: two-way ANOVA, repeated measures, MANOVA all build on one-way ANOVA
• Versatile: applicable to experimental (random assignment) and quasi-experimental designs
• Supports interaction effects in factorial extensions

⚠️ Limitations

• Assumption-sensitive: normality and homogeneity of variance required
• Omnibus test only: significant F tells you that groups differ, not which pairs differ — post-hoc tests needed
• Time-consuming with large k: post-hoc comparisons multiply rapidly
• Requires F-table: cannot interpret results without critical value reference
• Cannot be used when groups are severely unequal in size without adjustment
• Sensitive to outliers, which can distort the F-ratio substantially

---

Post-Hoc Tests: Beyond the F-Ratio

When F is significant, you need post-hoc (after the fact) comparisons to pinpoint which groups differ. The PDF uses simple paired t-tests (the Least Significant Difference or LSD approach). Modern alternatives with better error control include:

Post-Hoc Test	Error Control	Best When
LSD (t-tests)	None (liberal)	All pairs planned a priori
Bonferroni	Adjusts α by m = number of comparisons	Few planned comparisons
Tukey's HSD	Familywise .05	All pairwise comparisons
Scheffé	Most conservative	Complex contrasts
Games-Howell	Unequal variances	Heteroscedastic data

Wikipedia: Post-hoc analysis

---

Indian Research Context

ANOVA is among the most commonly used statistical techniques in Indian psychological research. Some applications:

Educational Psychology: ANOVA studies comparing academic performance across government, aided, and private school students are a staple of M.Ed. and M.A. (Education) dissertations across Indian universities. Studies regularly examine whether stream selection (Science/Arts/Commerce) interacts with socioeconomic background in predicting academic self-efficacy.

Clinical Psychology: Indian researchers have used one-way ANOVA to compare depression and anxiety scores across diagnostic groups (clinical, at-risk, control) in hospitals affiliated with NIMHANS (Bengaluru) and AIIMS (New Delhi).

Occupational Psychology: ANOVA comparing job satisfaction scores across public sector, private sector, and self-employed workers has been a recurrent theme in Indian organizational research, particularly in states with diverse employment landscapes (Maharashtra, Tamil Nadu, Kerala).

Agrarian Research Heritage: Fisher developed ANOVA for agriculture — and Indian agricultural universities (e.g., IARI, ICAR-affiliated colleges) have used it extensively since the 1950s for comparing crop yield across fertilizer treatments, mirroring the original Fisherian applications.

---

Real-World Applications in Psychology

• Clinical trials: Comparing outcomes across a control group, low-dose, and high-dose drug groups — standard in RCTs
• Educational interventions: Testing whether three instructional formats (lecture, flipped classroom, peer learning) differ in learning outcomes
• Neuropsychology: Comparing reaction times across healthy controls, MCI patients, and Alzheimer's patients
• Consumer psychology: Measuring purchase intent across five advertisement formats
• Occupational psychology: Comparing burnout scores across three shift-work schedules

---

Memory Aids

🧠 The ANOVA Acronym: B-W-F

Between ÷ Within = F-ratio
"Between and Within give you F"

🎵 The Steps Rhyme

"Correct, Square Total, Subtract Among,
Within's what's left when variance belongs.
Mean-Squares by df, then F you see,
Check the table — reject or keep H₀ free."

🔢 The df Triangle

     N - 1  (total)
    /        \
  k - 1    N - k
(between)  (within)

Total df splits into between-groups df and within-groups df — they always add up.

⚡ F vs t Quick Rule

If you have 2 groups → t-test
If you have 3+ groups → ANOVA (F-test)
After a significant F with 3+ groups → Post-hoc t-tests
Always: $F = t^2$

---

External Resources

📖 Essential Reading

• Wikipedia: Analysis of Variance — comprehensive overview with history and extensions
• Wikipedia: F-test — mathematical foundations of the F-distribution
• Statistics How To: One-Way ANOVA — step-by-step tutorials with examples

🎓 Video Lectures

• Crash Course Statistics #33: ANOVA — conceptual introduction, highly accessible
• StatQuest: One-Way ANOVA (Josh Starmer) — visual explanation of between vs within variance
• Khan Academy: ANOVA — interactive problem sets with F-ratio interpretation

📚 Research Papers (2020–2025)

• Blanca, M.J., et al. (2023). ANOVA versus Kruskal-Wallis: Which test to use when the normality assumption is violated? Psicothema, 35(1). — examines robustness across conditions
• Delacre, M., et al. (2019). Taking parametric assumptions seriously. International Review of Social Psychology, 32(1). [doi:10.5334/irsp.198] — argues for Welch's ANOVA as default
• Sullivan, G.M., & Feinn, R. (2012). Using effect size — or why the P value is not enough. Journal of Graduate Medical Education, 4(3), 279–282. — on reporting eta-squared (η²) alongside F

🛠 Online Tools

• Social Science Statistics: One-Way ANOVA Calculator — free browser-based calculator
• Real Statistics: ANOVA in Excel — includes Excel formulas and interpretation guidance

---

Self-Assessment Questions

🟢 Knowledge (Recall)

1. Define variance and state how it differs from standard deviation. Why is variance (not SD) the preferred measure in ANOVA?

2. State all four assumptions underlying one-way ANOVA. What test can be used to verify the homogeneity of variance assumption?

3. Complete the ANOVA summary table for a study with k = 4 groups and N = 24 total observations, given SS_A = 36 and SS_T = 108.
   (Calculate df, SS_W, MSS_A, MSS_W, and F.)

🟡 Comprehension (Understanding)

4. Explain why ANOVA is preferred over multiple t-tests when comparing more than two group means. Use the concept of familywise error rate in your answer.

5. In the drugs example (Example 2), the F-ratio was 6.38 and significant at .01 level. But post-hoc comparisons showed that Groups B and D (Drugs Q and S) did not differ significantly from each other. How is it possible that the overall F is significant but some pairs are not?

6. What does an F-ratio of 1.0 indicate? What does an F-ratio of 15.0 suggest about between-group vs within-group variance?

🔴 Application (Problem-Solving)

7. Solve this problem: Four methods of stress management were tested on 4 groups of 6 participants each. The group anxiety scores (post-treatment) were:

Method A	Method B	Method C	Method D
12, 10, 14, 13, 11, 12	8, 7, 9, 8, 6, 10	15, 14, 16, 15, 13, 14	10, 9, 11, 10, 8, 10

Apply one-way ANOVA and determine whether the methods differ significantly in reducing anxiety (use F-table if available, or compare your calculated F with F₀₅ for appropriate df).

---

Summary

Concept	Key Point
ANOVA vs t-test	ANOVA tests 3+ groups simultaneously; avoids alpha inflation
F-ratio	$MSS_A / MSS_W$ — between ÷ within variance
SS partition	$SS_T = SS_A + SS_W$ (always)
df partition	$df_T = df_A + df_W$ = $(k-1) + (N-k)$
Significant F	Reject H₀ → proceed to post-hoc comparisons
Non-significant F	Retain H₀ → stop; no pair differences are significant
F and t relationship	$F = t^2$; for two groups, results are identical
Assumptions	Normality, homogeneity of variance, independence, random sampling
Invented by	Sir Ronald Fisher (1923) — originally for agricultural experiments

---

Source PDFs:

• 📄 Block-3/Unit-3.pdf — Pages 77–95
• 📚 MPC-006 Statistics in Psychology, IGNOU