Residual Plots: Fitted Values vs. Residuals
What you'll learn: How to construct and interpret plots of residuals against fitted values to verify that your regression model meets key assumptions.
Why Plot Residuals?
After fitting a regression model, you need to check whether it's actually appropriate for your data. Two critical assumptions are linearity (the relationship is truly linear) and homoscedasticity (constant variance of errors). A residual plot is your diagnostic tool to verify both.
Constructing the Plot
On the horizontal axis, place the fitted (predicted) values ŷ from your regression model. On the vertical axis, place the residuals (the differences between actual and predicted values: y - ŷ). Each point represents one observation.
What to Look For
Ideal pattern: A random scatter of points around the horizontal line at zero, with no discernible pattern. Think of it like static on an old TV—pure noise with no structure.
Red flags:
- Curved pattern (like a U-shape or inverted-U): Suggests the relationship isn't linear. Your straight-line model is missing a curve in the data.
- Funnel shape (spreading out or narrowing): Indicates heteroscedasticity—the variability of errors changes as predictions change. Imagine a megaphone shape where residuals fan out.
- Clusters or groupings: May suggest missing categorical variables or interaction effects.
The Analogy
Think of residuals as "leftovers" after your model has explained what it can. If the leftovers form a pattern, your model missed something systematic. Random leftovers mean your model captured the important structure.
Key Takeaway: A good residual plot shows random scatter around zero with constant spread—any pattern suggests your model violates linearity or homoscedasticity assumptions and needs revision.