Evaluating Spatially Varying Coefficient Models Across Structural Patterns of Spatial Heterogeneity

ABSTRACT

Spatial heterogeneity refers to spatial variation in relationships between dependent and independent variables across geographic locations, conceptualized differently depending on the analytical perspective, each highlighting distinct patterns of variation. Spatially varying coefficient (SVC) models are commonly employed to explore spatial heterogeneity. Because each SVC model utilizes a fundamentally distinct estimation approach, performance is anticipated to vary across patterns. However, prior comparative evaluations have remained fragmented within single structural perspectives. This study systematically evaluates SVC model performance across diverse structural patterns of spatial heterogeneity, conceptualized through continuity, directionality, and scale. Using a controlled simulation framework, 12 patterns are generated and evaluated for multiscale geographically weighted regression (MGWR), spatial generalized additive models (SGAM), geographical Gaussian process GAM (GGP‐GAM), and Moran eigenvector spatial filter‐based SVC (MESF‐SVC). The findings demonstrate that GAM‐based models consistently excel in capturing smooth, continuous spatial gradients at the global scale. In contrast, MESF‐SVC proves robust in resolving localized and discontinuous spatial fragmentation. Meanwhile, MGWR exhibits instability under high frequency local variations, frequently over‐smoothing abrupt boundaries due to its localized kernel weighting. An empirical analysis of Seoul metropolitan election data shows these distinctions persist under multivariate conditions. Based on these outcomes, this study provides practical guidelines for SVC model selection.