WP5·Spatial modelling and data integration real data synthetic

Where little green meets low income.

Street features against CBS income across 360 buurten. Then the protocol's step model, tested on synthetic data with an answer planted in advance.

Start ↓Explore the map →
Synthetic data
01 / 06real

360 buurten, by income.

Share of low-income households, CBS 2023: from under 1% to a third.

02 / 06real

Green in view.

The mean green view of each buurt's sampled streets, from WP3.

03 / 06real

Where little green clusters.

Spatial clusters of green view (LISA). Outlined: low-green clusters among the most low-income households.

21priority buurten for WP6
04 / 06real

Do features follow income?

Barely, for the low-income share (|ρ| ≤ 0.13). With mean income: richer buurten show less green but more crossings and seating, and the team's perception model rates them more walkable, pleasant and safe.

05 / 06synthetic

A planted effect, found?

Steps made twice: no effect (A), and 0.14 planted (B). Adding park-stay minutes pulls both down. A DAG names the variables, not their form.

06 / 06synthetic

What the trees learned.

XGBoost with SHAP. A should be flat, and is not: SHAP shows what a model learned, not what causes steps.

Explore

The buurten, one layer at a time.

Model observations of 2025 street views (real) and CBS 2023 income, for 360 buurten with at least ten sampled panoramas. Click a buurt for its values.

Details
The model, and what else the synthetic run shows

steps ~ z_green + hour + weekend [+ activity minutes] + (1 | person), glmmTMB negative binomial, 4,496 valid half-hours of 200 synthetic persons.

Moderation by income stratum. None was generated, yet the terms come out at 0.072 (middle, p = 0.015) and 0.092 (low, p = 0.002). Moderation tests need a pre-specified hypothesis, a multiple-testing correction and the same sensitivity analyses.

Precision by sample size (B, walking and cycling minutes, 20 subsets each). Synthetic data are less noisy than real data; the protocol's own simulation gives 96.6% power with 180 persons.

ST-DBSCAN (50 m, 10 min, 10 fixes): 881 clusters; 98.2% of fixes correctly stationary or moving. The 5 m rule removes almost all indoor fixes, so the clusters left are outdoor places.

Correlations and spatial autocorrelation

Spearman ρ with 95% bootstrap intervals. Global Moran's I with six nearest neighbours, 999 permutations.