Module 09 Activity

Scenario

Build and check a multi-predictor model, and be honest about what its coefficients mean.

What you build

A statsmodels multiple regression with diagnostics run and its interpretation qualified.

Steps

1. Decide the predictors from subject knowledge and the plan, not from p-values.
2. Set the reference level for every categorical with C(var, Treatment(reference=...)) and state it in the caption.
3. Report unadjusted and adjusted estimates side by side, and treat a large shift as evidence that more confounding may remain.
4. Add an interaction only if pre-specified, and report the group-specific slopes rather than the interaction coefficient.
5. Inspect residuals against fitted values, compute Cook's distance from get_influence(), and check VIF where predictors are related.
6. Compare models on adjusted R-squared or held-out performance, never on R-squared.

Evidence to hand in

• The predictor list with its justification.
• The coefficient table with its reference level in the caption.
• The unadjusted and adjusted estimates.
• The interaction, if any, as group-specific slopes.
• The residual inspection, the Cook's distances and the VIFs.
• The model comparison.

Review checklist

• No predictor was selected because of its p-value.
• The reference level appears in the table caption.
• Residuals were inspected and any pattern addressed.
• Influential points were investigated rather than deleted.
• Model comparison used adjusted R-squared or held-out data.

Mark complete