AIC and BIC for Model Selection
You'll learn how to use information criteria to compare models fairly, even when they're not nested.
The Core Idea
You've already met AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion) individually. Now let's see how to apply them to real model selection problems.
Both AIC and BIC balance goodness-of-fit (how well the model explains your data) with complexity (how many predictors you're using). Lower values are better. The key advantage? You can compare any two models—even non-nested ones that don't share the same predictors.
When to Use AIC vs BIC
AIC is more forgiving about adding variables. It's better when you want a model that predicts well, even if it's slightly more complex.
BIC penalizes complexity more harshly (especially with larger datasets). It favors simpler, more parsimonious models—closer to Occam's Razor.
Think of it like shopping: AIC is willing to pay a bit more for extra features, while BIC insists on getting only what's truly necessary.
Practical Workflow
- Fit your candidate models (they can use completely different predictors)
- Calculate AIC and BIC for each
- Compare values: smaller = better
- A difference of 2+ points is generally considered meaningful
- Consider both criteria—if they disagree, think about your goal (prediction vs parsimony)
You're not doing hypothesis tests here—just ranking models by their information criteria. The "best" model has the lowest AIC or BIC, but models within ~2 units are often considered competitive.
Key Takeaway: AIC and BIC let you compare any models on fair ground by trading off fit against complexity; use AIC for prediction-focused work and BIC when you value simplicity.