Model 5: alternative-specific constants (rejected)
Model 4a: alternative-specific parameters plus two constants, asc_street and asc_os. A constant (ASC) is a fixed extra value for each type, which shows whether a type is preferred in general. Building is the reference (\(ASC_\text{bldg} = 0\)). In total, 15 parameters:
| 4a | 5 (+ ASC) | |
|---|---|---|
| Accuracy | 0.717 | 0.728 |
| CV accuracy (grouped) | 0.707 ± 0.069 | 0.717 ± 0.068 |
| LR vs 4a (df=2) | 29.6, p < 0.0001 | |
| Max VIF | 10.69 | 39.76 |
| Condition number | 775 | 37,583 |
On paper, the constants are significant. But the gain on unseen data is smaller than the normal spread between test sets. And asc_street and f_directional_street move almost together (r = 0.96). The reason: f_directional hardly changes between street candidates, so it acts almost like a constant for streets. The model is asked to measure the same "street-ness" twice. As a result, the direction weights and asc_street get standard errors of about 4.5 and are no longer significant, and their values change by 60-65% between test sets.
With a 15 m search distance it gets worse: the condition number reaches 92,891, and in some test sets the weights run off to values in the thousands. Tuning α adds two more pairs that move almost together (asc_t and f_distance_t, r = 0.98).
Rejected, because its weights cannot be trusted. One case of Overfitting signature. Part of Calibration.