Pseudo R-Squared Measures
You'll learn: How to quantify the explanatory power of GLMs using pseudo R-squared statistics when traditional R² doesn't apply.
In linear regression, R-squared tells us the proportion of variance explained by our model. But in GLMs like logistic or Poisson regression, we don't have variance in the same sense—our outcomes are binary, counts, or other non-continuous responses. We need alternative measures of "fit quality."
What Are Pseudo R-Squared Measures?
Pseudo R-squared statistics attempt to mimic the intuition of R² by comparing your fitted model to a baseline (usually the null model with only an intercept). They range from 0 to 1, where higher values suggest better fit—but unlike true R², they don't represent "variance explained."
McFadden's Pseudo R-Squared
The most common pseudo R² is McFadden's R²:
McFadden's R² = 1 - (log-likelihood of fitted model / log-likelihood of null model)
Think of it like measuring improvement: if your model's log-likelihood is much better (less negative) than the null model's, you get a higher pseudo R². Values between 0.2–0.4 are often considered "good" in practice—much lower benchmarks than linear regression!
Other Pseudo R-Squared Variants
- Cox & Snell R²: Based on likelihood ratios but capped below 1
- Nagelkerke R²: Adjusts Cox & Snell to range fully from 0 to 1
These all use deviance and likelihood concepts you've already learned, giving different flavors of the same idea: "How much better is my model than guessing the mean?"
Key Takeaway: Pseudo R-squared measures like McFadden's quantify GLM fit by comparing deviance to a null model; they guide model comparison but shouldn't be interpreted as variance explained.