Study on Spatially Nonstationary Impact on Catering Distribution: A Multiscale Geographically Weighted Regression Analysis Using POI Data

Lu Tan, Xiaojun BuISPRS Int. J. Geo Inf 2025 · DOI

TL;DR

This paper tackles the challenge of understanding how various urban factors influence the location of restaurants, acknowledging that these relationships aren't uniform across a city. By combining GeoDetector and Exploratory Factor Analysis with a novel Multiscale Geographically Weighted Regression (MGWR) model, they reveal that factors like commuting, shopping, tourism, and gathering activities have spatially varying impacts on catering distribution, offering a more nuanced view than traditional methods.

What Problem Does This Paper Address?

The distribution of catering establishments, a significant economic sector, is influenced by a complex interplay of urban factors. While previous research has identified factors like population density, traffic, and land use, these studies often assume a uniform relationship across the entire study area. However, the reality is that these relationships are often "spatially nonstationary," meaning they vary depending on location. For instance, the impact of population density might be different in a bustling city center compared to a suburban area. Traditional regression models struggle to capture this spatial heterogeneity. Existing methods like Geographically Weighted Regression (GWR) attempt to address this by allowing parameters to vary spatially, but they assume a single spatial scale for all relationships. This paper aims to address this limitation by investigating the spatially nonstationary impacts of various factors on catering distribution using a multiscale approach, recognizing that different factors might operate at different spatial scales.

Methods

The study employs a multi-stage methodological framework:

  1. GeoDetector Analysis: This spatial statistical method was used to identify "effective variables" that significantly contribute to catering distribution and to understand the interactions between these variables. It quantifies the explanatory power of each variable and their synergistic effects.
  2. Exploratory Factor Analysis (EFA): To combat multicollinearity (where independent variables are highly correlated) and reduce the number of variables, EFA was applied. It groups correlated variables into a smaller set of interpretable "latent factors." This step helps in creating more robust and independent predictors for the regression models.
  3. Multiscale Geographically Weighted Regression (MGWR): This is the core modeling technique. MGWR is an extension of GWR that allows for different bandwidths (representing spatial scales) for each independent variable. This is crucial because different factors influencing catering distribution might operate at different spatial scales (e.g., commuting might be a local factor, while tourism might have a broader regional influence). The study compares MGWR with traditional Ordinary Least Squares (OLS) and standard GWR to demonstrate its superiority in capturing spatial nonstationarity.

Datasets

The primary dataset used in this study is Point of Interest (POI) data collected from the Baidu Map API in August 2023 for the city of Nanjing, China. This dataset includes:

  • Catering Points: Approximately 26,648 identified catering establishments (restaurants, cafes, tea houses, etc.). This serves as the dependent variable.
  • Candidate Explanatory Variables: POIs from eight categories: Transportation, Residential, Office, Commerce, Education, Healthcare, Tourism, and Venues. These were aggregated into density measures within a 2 km x 2 km grid system.

The study does not introduce a new benchmark dataset but leverages existing POI data, which is highlighted as a more precise and readily available alternative to traditional government statistics or survey data.

Key Insights and Ideas

  • POI Data as a Rich Source: The study effectively demonstrates the utility of POI data for fine-grained urban analysis, offering higher precision and spatial resolution than traditional data sources.
  • GeoDetector + EFA Synergy: Combining GeoDetector for initial variable selection and interaction analysis with EFA for dimensionality reduction and multicollinearity handling provides a robust approach to preparing data for spatial regression.
  • The Power of Multiscale: The core insight is that catering distribution is not uniformly influenced by urban factors. MGWR's ability to assign different spatial scales to different factors is key to understanding these nuanced relationships.
  • Factor Interpretation: The identified latent factors (commuting, shopping, tourism, gathering) provide a more abstract yet interpretable understanding of the underlying drivers of catering distribution.
  • Agglomeration Effects: The finding that residential, office, and transportation POIs together form a significant factor (F1) highlights how the co-occurrence and agglomeration of certain POI types drive catering demand, challenging the notion of treating each factor in isolation.

Key Results

  • Model Performance: MGWR significantly outperformed both OLS and GWR, achieving the highest adjusted R² (0.903) and the lowest AICc (1099.686), indicating a better fit and more retained information.
  • Effective Variables: All eight candidate POI categories were found to be effective variables influencing catering distribution. Healthcare and Education had the highest explanatory power (q-values), but also showed high multicollinearity.
  • Latent Factors: EFA identified four interpretable latent factors:
  • F1 (Commuting Activities): Primarily driven by residential, office, and transportation POIs.
  • F2 (Shopping Activities): Strongly associated with commerce POIs.
  • F3 (Tourism Activities): Linked to tourism POIs.
  • F4 (Gathering Activities): Associated with venues POIs.
  • Spatially Varying Impacts: MGWR parameter estimates revealed significant spatial variations in the impact of each factor.
  • F1 (Commuting): Strong positive impact in the central urban area, suggesting daily consumption drives catering density there.
  • F2 (Shopping) & F3 (Tourism): Showed stronger positive impacts in peripheral and suburban areas, indicating their greater role in driving catering demand outside the city center.
  • F4 (Gathering): Exhibited complex spatial relationships with both positive and negative impacts in different subregions.
  • Scale Differences: The bandwidths assigned by MGWR indicated that tourism activities (F3) operated at a broader spatial scale, while commuting (F1) and shopping (F2) were more local.

Limitations and Future Work

  • Statistical Unit Scale: The choice of grid size (2 km x 2 km) for analysis might influence results. Future work could explore using community-level units.
  • POI Data Limitations: The study acknowledges potential sampling bias in POI data and the exclusion of informal food services, which could lead to underestimation.
  • Attribute Richness: The analysis relied on POI density, not on attributes like scale, type, or building height, which could provide more accurate insights.
  • Temporal Dynamics: The study is cross-sectional. Future research should investigate both spatial and temporal nonstationarity to capture the dynamic nature of catering distribution.

Summary updated

BibTeX
@article{tan2025study,
  author = {Lu Tan and Xiaojun Bu},
  title = {{Study on Spatially Nonstationary Impact on Catering Distribution: A Multiscale Geographically Weighted Regression Analysis Using POI Data}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2025},
  doi = {10.3390/ijgi14030119},
}
Figure