Standardized and Studentized Residuals
What you'll learn: How to scale residuals to make outliers easier to spot and compare across different leverage points.
The Problem with Raw Residuals
You've been working with raw residuals (observed - predicted), but not all residuals are created equal. Some observations have more "influence" on the fitted line than others—especially points far from the mean of X (high leverage). A residual of +5 might be normal for one point but alarming for another.
Standardized Residuals
Standardized residuals divide each residual by an estimate of its standard deviation:
Standardized residual = residual / (RSE × √(1 - leverage))
Where RSE is the residual standard error. This accounts for the fact that points with high leverage (far from X̄) naturally have smaller residuals because they "pull" the line toward themselves.
Analogy: It's like adjusting test scores for difficulty—a 70% on a hard exam might be better than 80% on an easy one.
Standardized residuals typically fall between -3 and +3. Values beyond ±2 or ±3 signal potential outliers.
Studentized Residuals (Deleted Residuals)
Studentized residuals go further: they refit the model without that specific observation and see how much it differs:
Studentized residual = residual / (s₍₋ᵢ₎ × √(1 - leverage))
Where s₍₋ᵢ₎ is the RSE calculated without observation i. This gives each point a "fair trial" by not letting it influence its own error estimate.
Why This Matters
Both scalings put residuals on a comparable scale (roughly standard normal if assumptions hold), making outlier detection consistent across all points regardless of their position or leverage.
Key Takeaway: Standardized and studentized residuals rescale raw residuals to account for leverage, making outliers identifiable on a common scale—typically flagging values beyond ±2 or ±3 as suspicious.