The f_distance² artifact
With α = 1, a squared distance term (f_distance²) looked clearly significant (t = 3.81) in the Interaction term screen (α = 1). After Alpha calibration set α to 2.4297, the same test, run again from the start, gives only t = 0.50.
So the squared term was not finding a real curve in the data. It was making up for a distance score that fell too fast because α was too small. It also caused most of the instability of Model 6.
| Screen term | t at α = 1 | t at α = 2.4297 |
|---|---|---|
| f_distance² | 3.81 | 0.50 |
| σ² | 2.23 | 2.47 |
| f_distribution × σ | -0.15 | 0.32 |
| f_distance × f_directional | 0.11 | 1.73 |
| f_distribution × share_type | 5.58 | 4.91 |
Lesson: after changing α, choose the extra terms again from the start. Keeping the old terms and only changing α doubled the VIF warning values (condition number from 20,695 to 48,275).
One case of Overfitting signature. Next: Model 6, re-derived at calibrated alpha. Part of Calibration.