Adjusted R-Squared vs Other Criteria
You'll learn how adjusted R-squared compares to other model selection tools like AIC, BIC, and cross-validation.
The Comparison Landscape
You've learned that adjusted R-squared penalizes unnecessary complexity by accounting for the number of predictors. But it's not the only tool for choosing between models. Other criteria serve similar purposes but with different philosophies.
Four Approaches Compared
Adjusted R-Squared
Directly adjusts for model complexity using degrees of freedom. Best when you want an intuitive extension of R-squared that you can explain easily. It's descriptive and familiar.
AIC (Akaike Information Criterion)
Balances model fit against complexity using a different penalty structure rooted in information theory. Generally more lenient than BIC—prefers slightly more complex models. Think of it as optimizing prediction accuracy.
BIC (Bayesian Information Criterion)
Similar to AIC but applies a stronger penalty for adding predictors, especially with larger sample sizes. More conservative—prefers simpler models. Rooted in Bayesian statistical theory.
Cross-Validation
Directly measures predictive performance on unseen data by splitting your dataset. The gold standard for assessing how well your model generalizes, but computationally expensive and requires sufficient data.
When to Use What
Adjusted R-squared: Quick model comparisons, reporting to non-technical audiences, when interpretability matters most.
AIC/BIC: Formal model selection procedures, comparing non-nested models, automated selection algorithms.
Cross-validation: When prediction is paramount, sufficient data exists, and computational resources allow it.
Think of it like choosing navigation tools: adjusted R-squared is your compass (simple, intuitive), AIC/BIC are your GPS coordinates (precise, theory-based), and cross-validation is actually test-driving the route (real-world validation).
Key Takeaway: Adjusted R-squared is intuitive and descriptive, AIC/BIC offer theory-driven alternatives with different complexity penalties, and cross-validation directly tests real-world predictive performance—choose based on your goals and resources.