Standardized and Studentized Residuals
What you'll learn: How to transform residuals into standardized scales that make it easy to identify unusual observations using simple thresholds.
Why Scale Residuals?
Raw residuals vary in magnitude depending on your data's scale. A residual of "5" might be huge in one dataset and tiny in another. Standardized and studentized residuals convert raw residuals into a common scale (similar to z-scores), making it easy to spot outliers regardless of your data's original units.
Standardized Residuals
Standardized residuals divide each residual by an estimate of its standard deviation:
Standardized residual = residual / (standard error of residuals)
This transforms them into a scale where most values should fall between -2 and +2 if the model fits well. Think of it like converting different currencies into a single standard currency for easier comparison.
Studentized Residuals
Studentized residuals (also called "deleted" or "externally studentized") go one step further. They account for the fact that each observation has slightly different variance depending on its leverage. This makes them even better at detecting outliers because they adjust for an observation's influence on the model.
Using Thresholds to Flag Outliers
The practical rule:
- |standardized residual| > 2: potentially unusual observation
- |standardized residual| > 3: likely outlier requiring investigation
These thresholds work because standardized and studentized residuals approximately follow a normal distribution, where ~95% of values fall within ±2 and ~99.7% within ±3.
Key Takeaway: Standardized and studentized residuals transform raw residuals into a common scale with clear thresholds (typically ±2 or ±3) that help you objectively identify outliers without needing to understand the original data's units or scale.