Null and Residual Deviance
What You'll Learn
You'll learn how to calculate and interpret null versus residual deviance to assess whether your GLM explains meaningful variation beyond a simple baseline.
The Core Idea
Think of deviance as a "badness-of-fit" score—higher values mean worse fit. Null deviance measures how poorly an intercept-only model (just predicting the overall mean/rate) fits your data. Residual deviance measures how poorly your fitted model (with all predictors) fits.
The comparison tells you: Did adding predictors actually help?
Null Deviance: The Baseline
The null model includes only the intercept—no predictors. It's like guessing the average outcome for everyone, ignoring all information. Null deviance quantifies this baseline lack-of-fit.
Residual Deviance: Your Model's Performance
After fitting your model with predictors, residual deviance measures the remaining unexplained variation. Lower residual deviance means your predictors captured more structure in the data.
The Gap Matters
The difference (null deviance − residual deviance) represents the variation your model explained. A large drop suggests your predictors add real value. A tiny drop? Your model barely improves on the baseline.
Interpreting the Numbers
- Large null deviance, small residual deviance: Great! Your model explains much more than the intercept-only baseline.
- Small drop: Your predictors may not be useful—the intercept-only model was nearly as good.
- Compare the difference to a chi-squared distribution (degrees of freedom = number of predictors added) to test statistical significance.
Key Takeaway: Null deviance measures baseline model badness-of-fit; residual deviance measures your fitted model's badness-of-fit. A large reduction indicates your predictors meaningfully improve predictions beyond just using the intercept.