Model 6, re-derived at calibrated alpha
Model 6 built again with the tuned α (2.4297). It starts from Model 4a: alternative-specific parameters and adds the two terms that still passed the test with the new α: σ² and f_distribution × share_type. In total, 15 parameters:
\[V_t = w_{1,t} f_\text{distribution} + w_{2,t} f_\text{distance} + w_{3,t} f_\text{directional} + w_{4,t}\,share_t + w_5\,\sigma + w_6\,\sigma^2 + w_7\,(f_\text{distribution}\times share_\text{type})\]
| 4a at α* | Model 6 re-derived | |
|---|---|---|
| LL(b) | -0.6701 | -0.6520 |
| LR vs 4a at α* (df=2) | 32.72, p < 1e-7 | |
| Accuracy | 0.723 | 0.723 |
| CV accuracy (grouped) | 0.717 ± 0.065 | 0.710 ± 0.075 |
| Condition number | 2,338 | 14,851 |
| Max VIF | 18.38 (σ²) |
The fit is significantly better, but the model becomes much less stable (condition number from 2,338 to 14,851). σ² again adds little. Next: Stepwise VIF elimination. Part of Calibration.