Built Environment and Metro Ridership in Beijing: Spatially Validated Prediction and Causal Caution

Integrating land use with metro systems is central to transit-oriented approaches to sustainable urban mobility, yet station-level prediction and association do not show whether planning changes generate ridership or wider sustainability benefits. We evaluate 292 Beijing metro stations using weekday mean morning- and evening-peak entries and exits (12–16 October 2020). Nine locally executable machine-learning models are compared across 27 zone-specific buffer combinations using out-of-sample prediction, degree of spatial interpretability (DSI), and geographic-block holdout. The selected common scale is 800/800/1800 m, although rankings vary by outcome, DSI definition, and spatial transfer. Across three prespecified scales, 43 of 52 mean adjusted contrasts (MACs) retain the same sign. In the primary causal-forest specification, six contrasts are nominally significant, including +1780 passengers (95% CI 329 to 3230) and −1248 (95% CI −2119 to −377), but none survives global false discovery rate (FDR) correction. An alternative partially linear analysis yields specification-dependent findings: 16 of 52 contrasts pass global FDR at a 20 km Conley cutoff, compared with 8 of 52 under a geographic-block score bootstrap. These results show that predictive fit and adjusted cross-sectional contrasts are not interchangeable with evidence that built-environment interventions cause ridership changes. The study contributes a spatially explicit evaluation framework for sustainable station-area planning and demonstrates why predictive accuracy, residual spatial structure, geographic transfer, and causal assumptions must be assessed separately. Stronger longitudinal or quasi-experimental evidence is required to establish intervention effects and sustainability outcomes.