Module 10 Summary

What this module established

A probability cannot be negative, and a straight line has no way of knowing that — a linear model on a churn indicator predicted -0.065 and -0.471. Logistic regression models the unbounded log-odds instead.

Carry forward

• Coefficients are log-odds; exponentiated they are odds ratios. With a churn rate of 0.533, an odds ratio of 2.58 badly overstates the risk ratio.
• Build prediction intervals on the link scale and convert afterwards, so they never leave [0, 1].
• On a 5% positive rate the model scored 0.949 accuracy while catching 0 of 102 positives — identical to a rule that always says no.
• Calibration is checkable; the threshold is not a statistical question. The same model caught 989 or 43 positives depending on the cut-off.

Before moving on

Move on when your report gives predicted probabilities rather than coefficients, and your threshold has a stated justification.

Mark complete