Discussion questions

1. You run several model selection methods on the same dataset and they don’t agree:

• cross-validation favors one model

• AIC suggests a slightly more complex one

• BIC prefers a simpler alternative How do you decide which result to trust?

2. We say we want the most parsimonious model, but what does that mean in practice?

• Is the most parsimonious model always the one with the fewest predictors?

• Is it the model that generalizes best to new data?

• Can a more complex model ever be more parsimonious if it captures the underlying structure better?

• How would your answer change depending on whether your goal is prediction vs explanation?