Model 4a: alternative-specific parameters
In this model, every score gets its own weight for each place type (building, street, open space). This is called "alternative-specific". sigma keeps one shared weight. In total, 13 parameters:
| Model 4a (R=10 m, α=1) | |
|---|---|
| ρ² vs LL(C) | 0.265 |
| LR vs 3a (df=6) | 114.15, p < 1e-16 |
| Accuracy | 0.717 |
| Building / street / open space | 0.902 / 0.443 / 0.466 |
| CV accuracy (grouped) | 0.707 ± 0.069 |
| Condition number | 775 |
The LR test compares this model with model 3a, which has shared weights. The result is very clear (p < 1e-16): separate weights per type fit much better.
For the first time, all three types improve together (building +3 percentage points, street +8.1, open space +6.8). The highest VIF is 10.69, for f_directional_street. The strong overlap between inputs of the same type is explained in Multicollinearity is structural.
f_distance_bldg and f_distance_street are not significant on their own (t = 0.29 and 0.34). Once weights are split by type, part of the distance information is picked up by share and sigma.
All later models build on this one. Next: Model 5: alternative-specific constants (rejected), Interaction term screen (α = 1). α was tuned on this model: Alpha calibration. Part of Calibration.