Sequential Model Building Strategy
What you'll learn: How to systematically add, test, and refine predictors to arrive at the best Generalized Linear Model (GLM) specification.
The Core Idea
Building a good GLM isn't about throwing all variables in at once and hoping for the best. A sequential model building strategy is a disciplined, step-by-step approach where you start simple, add predictors one at a time (or in logical groups), test their significance, and decide whether to keep them. Think of it like constructing a recipe: you taste after each ingredient to see if it improves the dish.
The Typical Workflow
- Start with a null or simple baseline model (intercept-only or a single strong predictor).
- Add candidate predictors one at a time or in meaningful blocks (e.g., demographics first, then behaviors).
- Test significance using likelihood ratio tests, Wald tests, or AIC/BIC comparisons—tools you've already learned.
- Evaluate improvement: Does deviance drop meaningfully? Does AIC/BIC improve? Are coefficients interpretable?
- Drop non-significant predictors if they don't contribute beyond noise.
- Check for interactions or polynomials only after main effects are stable.
- Repeat until adding more predictors doesn't improve the model or starts overfitting.
Why This Matters
Random variable selection leads to overfitting, multicollinearity, and uninterpretable models. A sequential strategy respects parsimony (Occam's Razor), keeps your model interpretable, and ensures every predictor earns its place by improving fit or explanatory power.
Key Takeaway: Sequential model building is a systematic, iterative process—start simple, add predictors deliberately, test rigorously, and refine until you reach a parsimonious, well-fitting GLM.