Course contentsShow
← Data Science
Lesson 564 of 2,14515. Regression DiagnosticsFree lesson

What to Do When Residual Plots Show Problems

Deciding between transformations, adding terms, robust methods, or collecting more data based on diagnostic findings.

What to Do When Residual Plots Show Problems

You'll learn: Once your residual plots reveal issues like non-linearity, heteroscedasticity, or outliers, you need a systematic strategy to fix them.

The Decision Framework

Think of regression diagnostics like a car inspection. When the mechanic finds problems, they don't just shrug—they recommend specific repairs. Similarly, when residual plots show violations of assumptions, you have four main repair options:

1. Transform Your Variables

When you see non-linearity or heteroscedasticity (funnel patterns), transforming Y (log, sqrt) or X can often restore linearity and constant variance. This is your first line of defense for curved patterns or spread that changes with fitted values.

2. Add Terms to Your Model

If transformations don't help or aren't appropriate, consider that your model might be too simple. The relationship might need polynomial terms or interaction effects to capture its true shape. A curved residual pattern suggests you're missing a systematic component.

3. Use Robust Methods

When you detect influential outliers or heavy-tailed residuals, robust regression techniques that downweight extreme observations can provide more reliable estimates than ordinary least squares.

4. Collect More or Better Data

Sometimes the problem isn't fixable with clever techniques—you simply need more observations, a wider range of X values, or better measurement quality. If patterns persist despite transformations, data quality might be the root issue.

Choosing Your Strategy

Start simple: try transformations first. If residual plots improve, you're done. If not, consider whether you're modeling the right relationship (add terms). Only move to robust methods when you've confirmed outliers are the main issue. Collecting new data should be a last resort, but sometimes it's the only honest solution.

Key Takeaway: When residual plots reveal problems, systematically try transformations, model complexity adjustments, robust methods, or new data collection—in that order—until diagnostic plots show healthy patterns.