When R-Squared is Not Enough
What you'll learn: Why a high R-squared value doesn't automatically mean you have a reliable or useful regression model.
The Seductive Trap of High R-Squared
You've learned to calculate and interpret R-squared—it tells you what proportion of variance your model explains. It's tempting to chase high R-squared values and declare victory when you hit 0.85 or 0.95. But here's the critical reality: a high R-squared can mask serious problems.
Think of R-squared like a student's test score. A 95% sounds impressive, but what if they cheated? What if the test only covered trivial questions? The number alone doesn't tell the full story.
Why R-Squared Can Mislead
Non-linear relationships: R-squared measures linear fit. Your data might have a clear curved pattern that simple linear regression misses entirely, yet still produce a moderate R-squared.
Residual patterns: A high R-squared can coexist with systematic patterns in your residuals—violations of the core assumptions that make your predictions unreliable.
Domain validity: Your model might fit your training data beautifully (high R-squared) but make nonsensical predictions outside the observed range. A regression predicting negative house prices or 200% success rates is mathematically valid but practically useless.
Outliers: A few extreme points can artificially inflate R-squared while distorting your entire model.
What You Should Do Instead
Always complement R-squared with:
- Residual plots to check for patterns and violations
- Residual standard error (RSE) to understand prediction accuracy in original units
- Domain knowledge to verify predictions make real-world sense
- Visual inspection of your fitted line against the actual data
Key Takeaway: R-squared is a useful summary statistic, but never rely on it alone—always validate your model through residual diagnostics and domain reasonableness before trusting its predictions.