Course contentsShow
← Data Science
Lesson 583 of 2,14515. Regression DiagnosticsFree lesson

Condition Number and Eigenvalues

Use condition number and eigenvalue analysis as alternative diagnostics for detecting multicollinearity.

Condition Number and Eigenvalues

You'll learn an alternative mathematical approach to detecting multicollinearity using condition number and eigenvalue analysis.

What Are Condition Number and Eigenvalues?

While VIF measures multicollinearity one variable at a time, the condition number gives you a single overall score for how badly your predictor variables are correlated. It's derived from eigenvalues, which are numbers that reveal the underlying structure of your correlation matrix.

Think of eigenvalues like measuring the width of an ellipse: if your predictors are well-behaved, the ellipse is nicely rounded. When multicollinearity exists, the ellipse becomes squashed—almost flat in one direction. The eigenvalues tell you how wide or narrow each dimension is.

How It Works

The condition number is calculated as:

Condition Number = sqrt(λ_max / λ_min)

Where λ_max is the largest eigenvalue and λ_min is the smallest eigenvalue from your predictor correlation matrix.

Interpreting the condition number:

  • 1 to 10: No multicollinearity problem
  • 10 to 30: Moderate multicollinearity—watch carefully
  • 30+: Severe multicollinearity—action needed

When one eigenvalue becomes very small (close to zero), it signals that your predictors lie nearly on a flat plane—meaning they're highly correlated. The condition number amplifies this signal: as the smallest eigenvalue shrinks, the condition number explodes.

Why Use This Instead of VIF?

The condition number gives you a global snapshot in one number, while VIF requires checking each variable. Small eigenvalues also directly point to which combinations of variables are causing trouble—helpful when diagnosing complex multicollinearity patterns.

Key Takeaway: The condition number and eigenvalue analysis provide a matrix-wide diagnostic for multicollinearity; condition numbers above 30 signal severe problems requiring intervention.