| Sumario: | This article introduces the multiscale geographically weighted zero-inflated negative binomial regression (MGWZINB), a novel local regression model. Geographically weighted regression (GWR) and its multiscale extension, multiscale geographically weighted regression (MGWR), have been extensively applied but are ill-suited for count data with an excessive number of zeros. The geographically weighted zero-inflated negative binomial regression (GWZINB), proposed by Da Silva and De Sousa (2023), tries to tackle the zero-inflation problem but suffers from two notable limitations: It uses single-scale bandwidths for both the logit and negative binomial components, wrongly assuming all parameters' spatial heterogeneity occurs at the same scale, and the interdependence between these two parts during bandwidth optimization leads to suboptimal bandwidth selection. MGWZINB addresses these issues by enabling parameter-specific bandwidths and presents a new bandwidth optimization algorithm. A simulation experiment shows MGWZINB outperforms GWZINB in estimating process scale, capturing spatial heterogeneity, achieving goodness of fit, and replicating the response variable. In an empirical study of skilled internal migrations in China, MGWZINB reveals that variables like population, wage, and living expense proportion have distinct effects and spatial patterns in determining whether the number of skilled migrants is zero (logit part) and the number when nonzero (negative binomial part). Despite its complex calibration process and high computational requirements, MGWZINB is a powerful tool for analyzing zero-inflated count data with spatial heterogeneity and scale, offering novel insights for research in geography, sociology, and other related disciplines.
|
|---|