The Likelihood Ratio Test
What you'll learn: How to use differences in deviance to compare nested GLMs and decide if additional predictors significantly improve your model.
The Core Idea
You've learned that deviance measures how well a GLM fits your data. Now you can use that to answer: "Does adding more predictors meaningfully improve the model?"
The Likelihood Ratio Test compares two nested models—where one model (the simpler one) is a special case of the other (the more complex one). The test statistic is simply the difference in deviance between them.
How It Works
- Fit your null (simpler) model and record its residual deviance
- Fit your alternative (more complex) model and record its residual deviance
- Calculate: LR statistic = Deviance(null) - Deviance(alternative)
- Under the null hypothesis (extra predictors don't help), this statistic follows a chi-squared distribution with degrees of freedom equal to the difference in the number of parameters
If the LR statistic is large enough (low p-value), the additional predictors significantly improve fit.
An Analogy
Think of deviance as "unexplained error." The likelihood ratio test asks: "Did adding those extra predictors reduce the unexplained error by more than you'd expect from random chance?"
When to Use It
- Comparing nested GLMs (logistic regression, Poisson regression, etc.)
- Deciding whether interaction terms or polynomial features are worth including
- Formal hypothesis testing for model selection alongside AIC/BIC
Key Takeaway: The Likelihood Ratio Test uses the difference in deviance between nested models, compared against a chi-squared distribution, to formally test whether additional predictors significantly improve model fit.