How to Interpret R² Without Overselling It
How to Interpret R² Without Overselling It
Plain-Language Definition
R² (the coefficient of determination) tells you what proportion of the variation in your dependent variable is explained by your independent variable(s). An R² of 0.66 means 66% of the variation in your outcome is explained by your model; the remaining 34% is explained by something else, or by randomness.
Why It Matters
Students often treat a high R² as automatic proof of a great model, and a low R² as automatic proof of a bad one. Neither is reliably true. Recent methodological research has specifically shown that R² and adjusted R² can meaningfully overstate a model’s real explanatory power — especially with smaller sample sizes — which means the tone you use to describe your R² matters as much as the number itself.
How to Interpret It Correctly
- State the number, then translate it into plain language — “66% of variance explained” is more useful to a reader than the raw statistic alone.
- Avoid absolute language — “this proves the model is reliable” oversells what R² can tell you. “This suggests a moderately strong fit” is more defensible.
- Consider it alongside sample size — a high R² from a small sample is less trustworthy than the same R² from a large one.
- Remember R² doesn’t tell you about causation, only about explained variance — a high R² can come from spurious relationships just as easily as genuine ones.
Worked Example (Fictitious Company)
Ferngate Retail Group’s regression produces an R² of 0.66.
Overselling it (avoid): “An R² of 0.66 proves that training hours are the primary driver of complaint rates.”
Correctly calibrated: “The model’s R² of 0.66 indicates that approximately 66% of the variance in complaint rate is explained by training hours, suggesting a moderately strong relationship. However, this leaves roughly a third of the variation unexplained, and R² alone shouldn’t be treated as a complete measure of the model’s real-world reliability.”
Key Takeaways
- R² measures explained variance, not proof of a strong or trustworthy model on its own
- Always translate the R² value into a plain-language percentage for your reader
- Pair R² with cautious, calibrated language rather than absolute claims
- A high R² doesn’t establish causation any more than a strong correlation does
Related Content
- C207 Course Roadmap
- C207 QUM3 Task 1 Guide: Linear Regression Analysis
- Correlation vs. Regression: When to Use Which
References & Further Reading
- Qi, J. (2023). How Much Should We Trust R2 and Adjusted R2: Evidence from Regressions in Top Economics Journals and Monte Carlo Simulations. Journal of Applied Economics, 26(1), Article 2207326. — A recent peer-reviewed paper demonstrating that R² and adjusted R² can overstate a model’s true explanatory power, the direct basis for this page’s guidance on calibrated language.