Nested spatial processes in urban food delivery: How scale and boundary choice shape built-environment relationships in New Delhi, India

Abstract

Platform-mediated food delivery demand concentrates last-mile freight, yet planning aggregates it into zonal units, introducing Modifiable Areal Unit Problem (MAUP) distortion. Even though increasing literature on MAUP in freight literature exists, its impact on digital delivery has not been explored before. The current paper aims to provide a diagnosis of MAUP in ODFD estimation for New Delhi, using street view built environment variables and night light variables within hexagonal grids of different sizes (250 m – 1000 m), aggregating them to ward and Traffic Analysis Zones (TAZ). The Negative Binomial model used to explain global relationship, the MGWR – local relationship, while the hierarchical multilevel regression will consider nested variance. The Decision Score considering prediction error, information loss, spatial dependence, and coefficient instability will be used as an indicator of scale suitability. Delivery point density (β = 1.52), road density (β = 0.29) and night light density (β = 0.20) at 250 m were significantly related to the ODFD. The 500 m grid proved to be a more stable scale under MGWR. Since delivery point density is highly correlated with ODFD, the Decision Score is recomputed using all predictors except the delivery point density: the global model scale shifts from 250 m to 750 m, and the MGWR scales from 500 m to the 500–750 m range; the impact of PCA vs. equal weighting on the sensitivity ranking is greatly reduced when delivery point density is removed from consideration. Aggregation by wards produces a negative coefficient for parking length (p = 0.042), which is corroborated by pooled cross-scale tests for both ward and TAZ boundaries (p = 0.038 and p = 0.001, respectively); while road density switches significance (p = 0.079), pedestrian count changes sign but remains insignificant. Aggregation by TAZ eliminates most built environment variables, except for delivery point density and Distance to CBD. ICC declines from 0.348 at 250 m to 0.119 at 1000 m, which confirms information loss with increasing scale. In addition to reducing model fit, incorrect scale choice can change sign of some significant coefficients, which infrastructure planning depends on. The Decision Score provides planners with a reproducible way to select an appropriate scale.

The authors' abstract, as published at the source. Applied Geography, 2026 · DOI ↗