R_buffer: 10 m vs 15 m
Which search distance (R_buffer) is better for finding candidates around a stay? 10 m beats 15 m for every model tested, in accuracy, in fit (log-likelihood) and in stability. The gap grows as the models get larger:
| Model | Accuracy R=10 m | Accuracy R=15 m | Δ |
|---|---|---|---|
| 1 Base | 0.518 | 0.509 | +0.9 pp |
| 2 Share-type + σ | 0.668 | 0.654 | +1.4 pp |
| 4a Alternative-specific | 0.717 | 0.698 | +1.9 pp |
| 5 + ASC | 0.728 | 0.709 | +1.9 pp |
| 6 + interactions | 0.724 | 0.699 | +2.5 pp |
(pp = percentage points.) The result held again with Grouped cross-validation on 21-07-2026. From the directional test onward, only 10 m was used.
Two cautions. First, ρ² cannot be compared between the two distances, because its baseline changes when the candidates change (Model evaluation metrics). Second, at 15 m, the model with ASC terms failed completely in cross-validation (Model 5: alternative-specific constants (rejected)).
A wider search distance brings in more street pieces that are not really relevant. Street accuracy suffers most at 15 m (Model 1: base).
Part of Calibration. How candidates are found: Candidate selection.