Module 09 Summary

What this module established

Every slope in a multiple regression is conditional on the other predictors. An unadjusted training coefficient of 853 became 52 once experience entered — and the true value was 50.

Carry forward

• A factor contributes one coefficient fewer than it has levels, all relative to the reference. Changing the reference changes every number and none of the fit.
• R-squared cannot detect the wrong shape. The curved model had the HIGHER R-squared; only its residuals, swinging +4.60, -9.83, +5.23, gave it away.
• One row in 41 changed a slope from 0.655 to 2.192, with a Cook's distance of 29.22 against a next-largest of 0.068.
• Twenty pure-noise predictors reached an R-squared of 0.457 on n = 60. Adjusted R-squared, at 0.179, is the honest one.

Before moving on

Move on when your residual plot is clean, your influential points are investigated, and no predictor entered because of its p-value.

Mark complete