Normality of Residuals
What you'll learn: Why assuming residuals follow a normal distribution matters for valid statistical inference in linear regression.
When you fit a linear regression model, you're not just drawing a line—you're making statistical claims about the relationship between variables. To trust confidence intervals, p-values, and hypothesis tests about your coefficients (β₀ and β₁), the residuals should be approximately normally distributed.
The Core Idea
Remember that residuals are the differences between observed values and your fitted line—the "errors" your model makes. The assumption isn't that your data must be normal, but that these residuals (the unexplained variation) should scatter around zero in a bell-shaped pattern.
Think of it like this: if you're shooting arrows at a target, you want your misses to be random noise—sometimes a bit left, sometimes a bit right, mostly near the center, with extreme misses being rare and symmetric. Non-normal residuals suggest something systematic is going wrong (like a bent bow).
When It Matters Most
Normality of residuals is most critical for:
- Small sample sizes (n < 30): Your confidence intervals and p-values rely heavily on the normality assumption
- Inference: When you need trustworthy hypothesis tests and prediction intervals
- Extreme predictions: When estimating values far from your data's center
For large samples, the Central Limit Theorem provides some cushion—your coefficient estimates remain reliable even with mild non-normality. But for small datasets or precise inference, violating this assumption can lead to misleading conclusions about statistical significance.
Key Takeaway: Residuals should be normally distributed primarily to ensure valid p-values and confidence intervals, especially in small samples—though the model can still make decent predictions even when this assumption is violated.