Course contentsShow
← Data Science
Lesson 647 of 2,14517. Categorical Predictors and InteractionsFree lesson

Impact on Model Results and Reporting

Why changing reference doesn't affect predictions or model fit, only coefficient interpretation and presentation.

Impact on Model Results and Reporting

You'll learn: How changing the reference category in categorical encoding affects what you see in reports, but not the underlying model quality or predictions.

The Core Idea

When you change which category serves as your reference in dummy variable encoding, the numbers in your coefficient table change dramatically—but your model's predictions, R-squared, residuals, and overall fit remain exactly the same.

Think of it like describing directions to a landmark. You can say "the cafe is 3 blocks north of the library" or "the cafe is 2 blocks south of the park." The cafe's actual location hasn't moved—only your reference point changed, so your description changed too.

What Changes

  • Coefficient values: Each dummy variable coefficient represents the difference from the reference, so new reference = new differences
  • Coefficient p-values: Statistical significance of specific dummies shifts because you're testing different comparisons
  • Interpretation: You'll describe group differences using different baseline language

What Stays the Same

  • Predictions: Every observation gets the identical predicted value
  • R-squared and adjusted R-squared: Model fit is unchanged
  • Residuals: Prediction errors remain identical
  • Overall model F-test: The global significance doesn't change

Why This Matters for Reporting

Choose your reference category strategically for clearer communication. If you're studying a medical treatment, making "no treatment" the reference makes coefficients easy to interpret as "treatment effects." But you're not changing the underlying mathematical model—just how you present it to your audience.

Key Takeaway: Changing the reference category is purely a presentation choice that affects coefficient interpretation, not model performance, predictions, or fundamental conclusions.