Course contentsShow
← Data Science
Lesson 560 of 2,14515. Regression DiagnosticsFree lesson

Scale-Location Plot (Spread-Location Plot)

Using the square root of standardized residuals to better visualize changes in variance across fitted values.

Scale-Location Plot (Spread-Location Plot)

What you'll learn: How to use the square root of standardized residuals to better spot changes in variance across your fitted values.

Why Transform Residuals?

In the previous lesson, you learned to detect heteroscedasticity (non-constant variance) using residuals vs fitted plots. But sometimes patterns in variance are subtle and hard to see when residuals bounce above and below zero.

The Scale-Location plot solves this by plotting the square root of the absolute value of standardized residuals against fitted values. This transformation does two clever things:

  1. Makes everything positive – no more negative values, so you're only looking at spread magnitude
  2. Emphasizes smaller values – the square root compresses large values and stretches small ones, making patterns clearer

How to Read It

Think of it like measuring the "spread radius" at each point along your predictions:

  • Horizontal trend line? Good news—variance is roughly constant (homoscedasticity)
  • Upward or downward slope? Warning—variance is changing systematically as fitted values increase
  • Funnel or cone shape? Variance increases or decreases with predictions

The y-axis shows √|standardized residual|, so you're visualizing the scale or spread of errors at different prediction levels—hence the name!

The Benefit

By squaring the residuals (making them positive) and then taking the square root, you get a clearer visual signal of whether the "scatter width" stays constant or changes as you move along the x-axis. It's easier to spot trends in variability than in raw residual plots.

Key Takeaway: The Scale-Location plot transforms residuals to √|standardized residual| so patterns in variance become more visible—look for a flat horizontal line to confirm constant variance.