Centering Predictors for Polynomials
What You'll Learn
You'll learn why centering your predictor variables before creating polynomial terms reduces multicollinearity and makes your regression coefficients easier to interpret.
The Problem with Raw Polynomials
When you create polynomial terms like x, x², and x³, these variables become highly correlated with each other. For example, if x ranges from 100 to 200, then x² ranges from 10,000 to 40,000—they move together predictably. This creates severe multicollinearity, inflating variance inflation factors (VIFs) and making coefficient estimates unstable and hard to interpret.
The Solution: Centering
Centering means subtracting the mean from your predictor before creating polynomial terms. Instead of using x, you use (x - x̄), where x̄ is the mean of x.
When you center first:
- The linear, quadratic, and cubic terms become much less correlated
- VIF values drop dramatically
- The intercept now represents the fitted value at the mean of x (a meaningful, observed point)
- Lower-order coefficients (like the linear term) retain clearer interpretations
Think of It Like This
Imagine measuring temperature's effect on ice cream sales. Raw temperatures (70°F, 80°F, 90°F) and their squares (4900, 6400, 8100) track closely. But if you center around the mean (say 80°F), you get deviations (-10, 0, +10) and their squares (100, 0, 100)—now they're measuring different things: level versus curvature.
When to Center
Always center continuous predictors before creating polynomial features, especially in polynomial regression with degree 2 or higher. This is standard practice and should be automatic in your workflow.
Key Takeaway: Center predictors by subtracting their mean before creating polynomial terms to reduce multicollinearity, improve numerical stability, and make coefficients interpretable at meaningful values.