A1153
Title: Modeling zero-inflated spatial proportion data with a structured beta spatial regression model
Authors: Yi-Ting Hwang - National Taipei University (Taiwan) [presenting]
Abstract: Proportion-type data are central to urban spatial analysis and mobility studies. These data are bounded between 0 and 1, often exhibit spatial dependence, and frequently contain many zeros. While zero-inflated Beta regression models address zero inflation, extending them with spatial random effects requires estimating high-dimensional covariance structures, leading to heavy computational costs and limiting large-scale applications. A structured Beta spatial regression model is proposed that integrates a zero inflation mechanism with thin plate spline (TPS) basis functions. Basis decomposition reduces the dimensionality of spatial random effects, alleviating computational burden while preserving multi-scale spatial variation. The model employs a two-part mixture structure to jointly characterize zero occurrence probability, the mean of non-zero proportions, and the precision parameter. Each component incorporates spatial smoothing and explanatory variables, enabling a comprehensive representation of the data-generating process. Parameters are estimated via maximum likelihood with an EM algorithm, treating spatial random effects as latent variables. Monte Carlo simulations validate the estimation, and an empirical analysis using Banqiao District mobility data demonstrates the framework. Results show comparable performance to logistic and Beta regression models, with the added advantage of simultaneously modeling zero inflation, non-zero structure, and spatial effects.
