8.5 Chapter summary

AIC-based model selection assesses which model or models best explain the variation in the outcome, but with penalization for complexity. This guards against overfitting and identifies the models most likely to perform well in future samples. Model selection can be used to test multiple non-null hypotheses against one another. It can also be used where there are many candidate models and the researcher wishes to know which one or ones are likely to be the best. Model averaging allows the analyst to obtain parameter estimates which come from the average of several different model specifications, providing a specification-neutral estimate of the impact of a predictor.