A0611
Title: Geographically weighted regression for compositional data
Authors: Takahiro Yoshida - The University of Tokyo (Japan) [presenting]
Abstract: Geographically weighted regression (GWR) is a widely used spatial data analysis technique across various fields. Additionally, the extension for non-Gaussian distributed data has been progressing. However, studies on the extension for compositional data are limited. Spatial regression model developments for compositional data are crucial topics in compositional data analysis (CoDA) literature. Geostatistical compositional models, such as the compositional kriging model, are popular approaches because CoDA has been historically developed in geosciences in which a continuous spatial process can be assumed. Other study approaches employ conditional autoregressive models or simultaneous autoregressive (spatial econometric) models. Although spatial autocorrelated errors are considered in these models, models for compositional data with spatial heterogeneity or spatially varying relationships remain limited. The objective is to build a GWR model for compositional data to consider both spatial heterogeneity and the constant-sum constraint. The GWR model and log-ratio techniques of CoDA are accommodated, and then the GWR model in the simplex space. This model is applied to analyze an actual data set from a US census survey data. The interpretational usefulness of the results of spatially varying compositional semi-elasticities was empirically verified.
