Course contentsShow
← Data Science
Lesson 570 of 2,14515. Regression DiagnosticsFree lesson

Q-Q Plots vs Formal Normality Tests: When Visual Checks Matter

Compare visual Q-Q plot interpretation with Shapiro-Wilk and other formal tests for practical diagnostic decisions.

Q-Q Plots vs Formal Normality Tests: When Visual Checks Matter

You'll learn when to trust your eyes versus when to rely on statistical tests for checking normality of residuals.

The Tension Between Seeing and Testing

You've learned to create Q-Q plots and interpret their patterns. But Python and R also offer formal normality tests like the Shapiro-Wilk test. These tests give you a p-value—seemingly objective evidence. So why bother with visual checks?

Here's the key insight: formal tests answer a different question than you actually care about.

What Formal Tests Really Tell You

The Shapiro-Wilk test asks: "Is there any detectable deviation from perfect normality?" With large sample sizes (n > 200), it will almost always reject normality—even when deviations are trivial and won't affect your regression conclusions. The test becomes too sensitive.

With small samples (n < 30), the opposite problem occurs: the test has low power and may fail to detect serious departures from normality that would matter.

What Q-Q Plots Tell You

A Q-Q plot shows how and how much your residuals deviate from normality. Minor wiggles near the tails? Probably fine—regression is robust to slight non-normality. Systematic curves or dramatic outliers? That's worth investigating, regardless of the p-value.

The Practical Balance

Use both tools together:

  • Q-Q plot first: Does the pattern look problematic for your purposes?
  • Formal test second: Does it confirm major concerns, or is it just picking up minor noise?

Think of it like checking if a wall is straight: a laser level (formal test) detects microscopic imperfections, but your eye (Q-Q plot) tells you if it actually looks crooked.

Key Takeaway: Q-Q plots show practical departures from normality; formal tests show statistical departures. For regression diagnostics, visual judgment of practical impact usually matters more than p-values.