Normality of Residuals
What you'll learn: How to check whether ANOVA residuals (not your raw data) are normally distributed—a key assumption for valid F-tests.
Why Residuals, Not Raw Data?
In ANOVA, we don't actually care if your raw observations are perfectly normal. What matters is whether the residuals—the differences between each observation and its group mean—follow a normal distribution. Think of residuals as the "leftover noise" after accounting for group differences. If this noise is bell-shaped, the F-test's probability calculations remain trustworthy.
Analogy: Imagine you're measuring heights of plants given different fertilizers. The raw heights might look skewed, but that could just be because one fertilizer works better. What you really need is for the random variation within each group (the residuals) to be normally distributed.
How to Test Normality of Residuals
Visual Methods
Use the tools you've already learned, but apply them to residuals:
- Histogram of residuals: Should look roughly bell-shaped
- Q-Q plot (quantile-quantile plot): Residuals should fall along a straight diagonal line
Statistical Tests
- Shapiro-Wilk test: Tests the null hypothesis that residuals are normally distributed
- Anderson-Darling test: Another formal test for normality
What to Do with Results
If your residuals pass these checks (visual patterns look good, statistical tests don't reject normality), you're cleared to trust your ANOVA F-test results. If not, you may need to transform your data or use a non-parametric alternative.
Key Takeaway: ANOVA assumes residuals are normally distributed—always check residuals, not raw data, using both visual plots and statistical tests.