Course contentsShow
← Data Science
Lesson 559 of 2,14515. Regression DiagnosticsFree lesson

Detecting Heteroscedasticity (Non-Constant Variance)

Spotting funnel shapes or other variance patterns that violate the constant variance assumption.

Detecting Heteroscedasticity (Non-Constant Variance)

What you'll learn: How to spot violations of the constant variance assumption by recognizing characteristic patterns in residual plots.

The Core Idea

Homoscedasticity (constant variance) means the spread of residuals should remain roughly the same across all fitted values. Heteroscedasticity is when this assumption breaks down—when residuals become more (or less) spread out as predictions change.

When plotting residuals vs fitted values, heteroscedasticity reveals itself through distinctive patterns rather than the random cloud you want to see.

The Classic Funnel Shape

The most common heteroscedasticity pattern is a funnel (or cone):

  • Expanding funnel: Residuals start tight on the left and fan out wider on the right. This means the model's prediction errors grow larger as fitted values increase.
  • Contracting funnel: The opposite—wide on the left, narrow on the right.

Think of it like a megaphone: the sound (variance) gets louder (bigger) as you move away from the source.

Other Variance Patterns

Heteroscedasticity isn't always funnel-shaped:

  • Double funnel: Variance is small in the middle but large at both extremes
  • Clustering: Different groups have noticeably different spreads
  • Curved patterns with changing spread: Variance changes alongside non-linearity

Why This Matters

Heteroscedasticity doesn't bias your coefficient estimates, but it does invalidate standard errors, confidence intervals, and hypothesis tests. Your predictions become unreliable, especially where variance is high.

What to Look For

Scan your residuals vs fitted plot for any systematic change in vertical spread from left to right. If the scatter tightens or widens, you've detected heteroscedasticity.

Key Takeaway: Heteroscedasticity appears as funnel shapes or other systematic changes in residual spread across fitted values—any pattern where variance isn't constant violates a core regression assumption.