Generalized geographically and temporally weighted regression

Abstract

This paper proposes Generalized Geographically and Temporally Weighted Regression (GGTWR) to address the limitations of Geographically and Temporally Weighted Regression (GTWR). The proposed GGTWR framework encompasses various generalized linear models, e.g. Poisson regression, negative binomial regression, and other models of the exponential distribution family. The paper also shows the classic GTWR bandwidth search algorithm is not suitable for GGTWR and proposes a new bandwidth search algorithm for GGTWR. Several simulation experiments are used to prove that GGTWR can effectively capture spatiotemporal non-stationary. The GGTWR framework enables the estimation of varying regression coefficients that capture spatial and temporal heterogeneity for generalized linear relationships, providing a comprehensive understanding of how predictor variables influence the response variable across different locations and time periods. An application to interprovincial population migration in China using 2005–2020 census data demonstrates the interpretability of the GGTWR framework. GGTWR provides a flexible modeling approach that more accurately explains real-world phenomena.