Adjusted R-squared (R²_adj)
Adjusted Coefficient of Determination · Also known as: Adjusted R², R²_adj
Adjusted R² is a corrected version of the coefficient of determination that accounts for the number of predictors in a regression model. Introduced by Henri Theil in 1961, it addresses the fundamental limitation of standard R²: the tendency to increase whenever any predictor is added, regardless of whether that predictor contributes meaningfully to explaining the target variable.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use adjusted R² whenever you need to compare regression models with different numbers of predictors. It is particularly valuable for feature selection: as you add variables, adjusted R² will guide you toward the optimal model. Adjusted R² is more reliable than standard R² on small samples where overfitting risk is high. Like standard R², it assumes linear relationships and is sensitive to outliers.
Strengths & limitations
- Directly penalizes additional predictors, preventing overfitting incentive
- Enables fair model comparison across different numbers of variables
- Maintains interpretability while addressing a key R² limitation
- Simple adjustment formula, easy to compute and explain
- Still assumes linear relationships; inappropriate for nonlinear models
- Penalty may be too lenient or too strict depending on sample size and context
- Does not account for correlation among predictors (multicollinearity)
- Does not measure out-of-sample prediction accuracy
Frequently asked
When should I use adjusted R² instead of AIC or BIC?
Adjusted R² is simpler and more interpretable, making it good for exploratory analysis and teaching. AIC and BIC are information-theoretic measures that also penalize model complexity but can be more appropriate for non-nested model comparison and prediction-focused tasks.
Can adjusted R² be negative?
Yes, adjusted R² can be more negative than standard R² because the penalty factor amplifies poor fit. A highly negative adjusted R² signals a severely misspecified model.
How much improvement in adjusted R² justifies adding a new predictor?
There is no fixed threshold. A rule of thumb is to add a predictor if adjusted R² increases by at least 0.01-0.02. However, statistical significance tests (F-tests) and domain knowledge should also guide decisions.
Does adjusted R² work for logistic or non-linear regression?
The formula is specific to linear regression with ordinary least squares. For logistic regression, use alternatives like McFadden's pseudo-R², and for non-linear models, prefer information criteria (AIC, BIC) or cross-validation metrics.
Why does adjusted R² sometimes go down when I add a significant variable?
This can happen if sample size is very small relative to the number of predictors, or if the improvement in fit is modest. The penalty term (n-1)/(n-p-1) becomes large when p is close to n, making the adjustment severe.
Sources
- Theil, H. (1961). Economic Forecasts and Policy. Amsterdam: North-Holland Publishing Company. link ↗
- Ezekiel, M. (1930). Methods of Correlation Analysis. New York: John Wiley & Sons. link ↗
- Judge, G. G., Griffiths, W. E., Hill, R. C., Lütkepohl, H., & Lee, T. C. (1985). The Theory and Practice of Econometrics. New York: John Wiley & Sons. ISBN: 978-0471050773
How to cite this page
ScholarGate. (2026, June 3). Adjusted Coefficient of Determination. ScholarGate. https://scholargate.app/en/model-evaluation/adjusted-r-squared
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Akaike Information CriterionModel Evaluation↔ compare
- Bayesian Information CriterionModel Evaluation↔ compare
- Mean Squared ErrorModel Evaluation↔ compare
- R-squaredModel Evaluation↔ compare
- Root Mean Squared ErrorModel Evaluation↔ compare