Model Building Strategies

Summary

Distinguishes three goals of regression modeling, description versus prediction versus causal inference, and matches model-building strategies to each, advocating purposeful selection guided by a directed acyclic graph for causal questions and warning against stepwise selection's inflated Type I error, unstable selection, and biased coefficients. Develops penalized regression for prediction including LASSO, ridge, and elastic net with cross-validated tuning, then model averaging as a way to acknowledge uncertainty in model choice. Closes with functional-form decisions including polynomials and restricted cubic splines, interactions, centering, and standardization conventions for interpretation.