Statistics Glossary → Pseudo-R² (McFadden, Cox-Snell, Nagelkerke)
Pseudo-R² (McFadden, Cox-Snell, Nagelkerke)concept
Measures that, by analogy with R² in linear regression, summarize how much a model estimated by maximum likelihood, such as logistic regression, improves on the null model.
The classical R², based on the residual sum of squares, is not defined for logistic regression; the log-likelihood is used instead. McFadden's R² = 1 - LL_model / LL_null shows how much the model's log-likelihood improves on a model with only an intercept. Cox-Snell R² = 1 - exp(-LR / n) is based on the likelihood ratio statistic, but with a discrete outcome its maximum is below 1. Nagelkerke's R² divides the Cox-Snell value by that maximum to stretch it over the 0-1 range.
These measures should not be read as a 'proportion of variance explained', and they give noticeably different values for the same model, so it should be clear which one is reported. Values are generally lower than R² in linear regression; for McFadden, for example, 0.2-0.4 is considered a very good fit. Pseudo-R² is useful for comparing models fitted to the same data and outcome; a model's classification performance and calibration call for separate tools such as the ROC curve and the Hosmer-Lemeshow test.
An everyday example
A bank builds a logistic model predicting whether customers will repay their credit-card balance on time, using income and age. It finds McFadden R² = 0.18 and Nagelkerke R² = 0.29. The two values summarize the same model on different scales; both say the model is clearly better than a single 'same probability for everyone' guess, but that much of repayment behaviour depends on other factors.
In MerQur
In Logistic Regression, 'Extra output: Pseudo-R² (CS/Nagelkerke)' reports the McFadden, Cox-Snell and Nagelkerke values; with the Firth correction on, Cox-Snell and Nagelkerke are computed from the penalized likelihood ratio. Multinomial Logistic Regression also reports the McFadden, Cox-Snell and Nagelkerke R².
Analyses where this concept appears
Logistic RegressionMultinomial Logistic Regression
Related concepts
Coefficient of determination (R²) and adjusted R²Likelihood ratio test (LRT)