Normality assumption: The data in each group should be approximately normally distributed for ANOVA to be valid.
- Calculate descriptive statistics for each group
- Construct normal probability plots (Q-Q plots)
- Apply formal tests (Shapiro-Wilk for small samples)
- Examine skewness and kurtosis
- Visually inspect histograms
- Make decision about normality
Method A: Mean = 85.6, SD = 2.28, n = 5
Method B: Mean = 79.0, SD = 2.24, n = 5
Method C: Mean = 91.2, SD = 1.92, n = 5
Method A: W = 0.956, p-value ≈ 0.75
Method B: W = 0.875, p-value ≈ 0.28
Method C: W = 0.964, p-value ≈ 0.80
Method A: Skewness = -0.28, Kurtosis = -1.2
Method B: Skewness = 0.0, Kurtosis = -1.2
Method C: Skewness = -0.28, Kurtosis = -1.2
Values near 0 indicate normality
Points should fall approximately along a straight line
Visual inspection shows reasonable linearity
All p-values > 0.05, so we fail to reject H₀
There is insufficient evidence to conclude non-normality
All three groups appear to satisfy the normality assumption
The normality assumption is satisfied for all three groups. The Shapiro-Wilk tests yield p-values greater than 0.05 for each group, indicating no significant departure from normality.
• Shapiro-Wilk test: For small samples (n < 50)
• Decision rule: p-value > α → fail to reject normality
• Visual confirmation: Q-Q plots support statistical tests
• Shapiro-Wilk p-values: A=0.75, B=0.28, C=0.80
• All values > 0.05
• Skewness and kurtosis near zero
Equal variances assumption: The population variances of all groups should be equal (homogeneity of variance).
Method A: s²_A = 5.2
Method B: s²_B = 5.0
Method C: s²_C = 3.7
Levene's test uses absolute deviations from group medians
For each observation: |x_ij - median_j|
Method A median = 86
Deviations: |85-86|=1, |87-86|=1, |82-86|=4, |88-86|=2, |86-86|=0
Average deviation = 1.6
Method B median = 79
Deviations: 0, 1, 3, 3, 0 → Average = 1.4
Method C median = 91
Deviations: 1, 1, 3, 0, 2 → Average = 1.4
ANOVA on deviations: F = 0.04
p-value ≈ 0.96
Since p-value (0.96) > α (0.05), we fail to reject H₀
Variances are not significantly different
Bartlett's test: χ² = 0.25, df = 2, p-value ≈ 0.88
Same conclusion: fail to reject equal variances
The equal variances assumption is satisfied. Levene's test yields a p-value of 0.96, which is greater than 0.05, indicating no significant difference in variances across groups.
• Levene's test: More robust than Bartlett's for non-normal data
• Decision rule: p-value > α → fail to reject equal variances
• Robustness: ANOVA is somewhat robust to mild violations
• Levene's test: p = 0.96 > 0.05
• Bartlett's test: p = 0.88 > 0.05
• Variances: 5.2, 5.0, 3.7 (similar)
Independence assumption: Observations within and between groups must be independent of each other.
Students randomly assigned to instructors
Each instructor teaches separate class
Students don't interact during study period
Random assignment minimizes correlation between students
Separate classes prevent interaction during study
Within-group independence appears satisfied
Different instructors teach different classes
No communication between groups during study
Between-group independence appears satisfied
Classroom environment differences
Time of day effects
These are controlled through design
Violations inflate Type I error rate
Standard errors become underestimated
Confidence intervals become too narrow
Random assignment helps ensure independence
Blinding reduces potential correlations
Control for confounding variables
The independence assumption appears satisfied. Random assignment and separate classes minimize correlations between observations, meeting the independence requirement for ANOVA.
• Random assignment: Helps ensure independence
• Design features: Separate classes prevent contamination
• Consequences: Violations affect Type I error rate
• Random assignment used
• Separate classes taught independently
• No interaction between groups
Normality assumption: The residuals (errors) in each group are normally distributed.
Equal variances (homogeneity): The population variances of all groups are equal.
Independence: Observations are independent of each other within and between groups.
Random sampling: Data comes from random samples from each population.
Linearity: The relationship between dependent and independent variables is linear.
Homoscedasticity: Constant variance across all levels of the independent variable.
- Study design review: Assess how data was collected
- Normality check: Use tests (Shapiro-Wilk) and plots (Q-Q)
- Equal variances: Apply Levene's or Bartlett's test
- Independence: Evaluate study design and data collection
- Residual analysis: Examine residuals for patterns
- Outlier detection: Identify influential observations
- Decision making: Determine if assumptions are satisfied
- Alternatives: Consider if assumptions violated
• Normality: Shapiro-Wilk, Anderson-Darling, Q-Q plots, histogram
• Equal variances: Levene's test, Bartlett's test, Brown-Forsythe
• Independence: Study design evaluation, Durbin-Watson for time series
• Use non-parametric alternatives (Kruskal-Wallis)
• Apply data transformations (log, square root)
• Use Welch's ANOVA for unequal variances
• Increase sample size to invoke central limit theorem
Assumption violations: When the data fails to meet one or more of ANOVA's underlying assumptions.
Group 1: Mean = 28.3, SD = 35.1, Median = 15
Group 2: Mean = 25.0, SD = 3.8, Median = 25
Group 3: Mean = 35.0, SD = 3.8, Median = 35
Group 1: Value 100 is clearly an outlier
It's more than 2.5 times the interquartile range above Q3
Group 1: Shapiro-Wilk test likely shows significant departure from normality
Due to the outlier, normality assumption is violated
Group 1 variance (1232) much larger than Groups 2 and 3 (14.4)
Levene's test would show significant violation of equal variances
The outlier in Group 1 violates both normality and equal variance assumptions
Could lead to inflated Type I error
Option 1: Remove the outlier if it's a data entry error
Option 2: Apply log transformation to reduce impact
Option 3: Use non-parametric Kruskal-Wallis test
Both normality and equal variances assumptions are violated due to the outlier (100) in Group 1. Consider removing the outlier, applying transformations, or using non-parametric methods.
• Outlier impact: Single outliers can severely violate assumptions
• Remedies: Removal, transformation, or non-parametric alternatives
• Verification: Always check for outliers before ANOVA
Formal testing: Using statistical tests to rigorously evaluate whether assumptions are satisfied.
Group 1: Mean = 14.0, SD = 1.10, n = 8
Group 2: Mean = 19.0, SD = 1.31, n = 8
Group 3: Mean = 23.0, SD = 1.31, n = 8
Group 4: Mean = 17.0, SD = 1.31, n = 8
Shapiro-Wilk tests (for n=8):
Group 1: W = 0.94, p = 0.62
Group 2: W = 0.96, p = 0.78
Group 3: W = 0.96, p = 0.78
Group 4: W = 0.94, p = 0.62
Levene's test: F = 0.12, df1 = 3, df2 = 28, p = 0.95
Groups are independent samples
No apparent correlation between groups
Normality: All p-values > 0.05 → Satisfied
Equal variances: p = 0.95 > 0.05 → Satisfied
Independence: Design-based → Satisfied
All ANOVA assumptions are satisfied
Proceed with parametric ANOVA
All ANOVA assumptions are satisfied. Proceed with the parametric ANOVA analysis. All groups show normality (p > 0.05) and equal variances (p = 0.95).
• Formal testing: Use statistical tests with decision rules
• Multiple tests: Check all assumptions systematically
• Decision rules: Compare p-values with α level
Parametric tests: Statistical tests that make assumptions about population parameters and distributions.
Non-parametric alternatives: Distribution-free tests that don't require normality assumptions.
Robustness: The ability of a statistical procedure to perform well even when assumptions are violated.
Power: The probability of correctly rejecting a false null hypothesis.
Type I error: Incorrectly rejecting a true null hypothesis (false positive).
Type II error: Failing to reject a false null hypothesis (false negative).
- Design review: Examine how data was collected and organized
- Descriptive analysis: Calculate means, variances, and identify outliers
- Visual inspection: Create histograms, boxplots, Q-Q plots
- Formal testing: Apply statistical tests for each assumption
- Residual analysis: Examine residuals for patterns
- Sensitivity analysis: Test robustness to violations
- Alternative methods: Consider non-parametric tests if needed
- Reporting: Document all assumption checks
• Robustness principle: ANOVA is relatively robust to minor violations
• Sample size effect: Larger samples provide more robustness
• Unequal variances: More problematic with unequal sample sizes
• Visual methods: Complement formal tests for better insight
• Practical significance: Consider severity of violations
• ANOVA is more sensitive to outliers than to normality
• Balanced designs (equal sample sizes) are more robust
• Consider bootstrap methods for non-standard situations
• Multiple imputation for missing data issues