McFadden's pseudo R-squared measures how much a fitted logistic regression improves on an intercept-only null model when predicting a binary outcome. Unlike linear-regression R-squared, it is likelihood-based rather than a proportion of variance explained. section 5.2.3
Compute it as 1 − (model log-likelihood ÷ null-model log-likelihood). The fitted model and null-model log-likelihoods are available as model attributes and appear in the regression table. Larger values indicate greater improvement over the baseline, although log-likelihoods and pseudo R-squared values are most useful when comparing models fitted to the same data. section 5.2.3
Use the metric as one part of logistic-regression evaluation:
- Compare competing models using their likelihood-based improvement over the same intercept-only baseline. section 5.2.3
- Pair it with the likelihood-ratio test p-value, predictor p-values, and classification measures such as accuracy and ROC area. section 5.3
- Do not interpret it as the percentage of binary-outcome variance explained, as with ordinary linear-regression R-squared. section 5.2.3
