Course contentsShow
← Mathematics
Lesson 3091 of 3,12160. Statistics — ModelingFree lesson

Multicollinearity Diagnostics

Use variance inflation factors (VIF) and correlation matrices to detect and address multicollinearity among predictors.

Multicollinearity Diagnostics

What you'll learn: How to quantitatively detect and measure multicollinearity using variance inflation factors and correlation matrices.

The Core Problem

You've learned that multicollinearity occurs when predictor variables are highly correlated with each other. While you know what it is and its consequences (inflated standard errors, unstable coefficients), now you need tools to diagnose it systematically.

Two Essential Diagnostic Tools

Correlation Matrices

The simplest diagnostic is examining pairwise correlations between predictors. Create a correlation matrix showing all predictor-to-predictor correlations. Rules of thumb:

  • Correlations above |0.8| or |0.9| suggest problematic multicollinearity
  • However, this only catches pairwise relationships—not more complex dependencies involving three or more variables

Analogy: It's like checking if two team members are too similar; you might miss that three different people all bring redundant skills.

Variance Inflation Factor (VIF)

The VIF measures how much a predictor's variance is "inflated" due to correlation with other predictors. For predictor j:

  1. Regress predictor j on all other predictors
  2. Calculate R²ⱼ from this regression
  3. VIF = 1/(1 - R²ⱼ)

Interpretation:

  • VIF = 1: No multicollinearity
  • VIF between 5-10: Moderate concern
  • VIF > 10: Severe multicollinearity requiring action

The VIF catches any linear dependency, not just pairwise correlations—it's your most reliable diagnostic.

Addressing High VIF

When you detect multicollinearity:

  • Remove one of the correlated predictors
  • Combine correlated predictors (e.g., average them)
  • Use regularization methods (future topic)
  • Collect more data if possible

Key Takeaway: Use correlation matrices for quick pairwise checks and VIF values to comprehensively detect multicollinearity; VIF > 10 signals serious problems demanding action before trusting your regression coefficients.