A Multiscale Spatially Varying Coefficient Model for Regional Analysis of Topsoil Geochemistry

• Published in: Journal of agricultural, biological, and environmental statistics Vol. 25; no. 1; pp. 74 - 89
• Main Authors: Kim, Keunseo, Kim, Hyojoong, Kim, Vinnam, Kim, Heeyoung
• Format: Journal Article
• Language: English
• Published: New York Springer Science + Business Media 01.03.2020; Springer US; Springer; Springer Nature B.V
• Subjects: Agriculture; Arsenic; Autoregressive models; Biostatistics; Coefficients; Correlation analysis; Environmental health; Gaussian process; Geochemistry; Hazardous materials; Health Sciences; Levels; Mathematics and Statistics; Medicine; Monitoring/Environmental Analysis; Regional analysis; Regional development; Regional planning; River sediments; Sediments; Statistics; Statistics for Life Sciences; Topsoil; Toxic substances; Toxicity; Bayesian hierarchical modeling; Spatially varying coefficient process; Arsenic; Multiscale modeling
• ISSN: 1085-7117; 1537-2693
• DOI: 10.1007/s13253-019-00379-x

A motivating example for this paper is to study a topsoil geochemical process across a large region. In regional environmental health studies, ambient levels of toxic substances in topsoil are commonly used as surrogates for personal exposure to toxic substances. However, toxicity levels in topsoil are usually sparsely measured at a limited number of point locations. Consequently, topsoil measurements only provide highly localized regional information and cannot be representative of the surrounding area. Instead, it is standard practice to use point-referenced measurements of stream sediments, because they are widely available across a region and are correlated with topsoil measurements at nearby locations. For more effective regional modeling of topsoil geochemistry, we develop a spatially varying coefficient model that integrates point-level topsoil and point-referenced area-level stream sediment data. The proposed model incorporates two spatial characteristics: the local spatial autocorrelation in the latent topsoil process and the spatially varying relationship between the latent topsoil and stream sediment processes. The former is modeled indirectly via a conditional autoregressive model for the stream sediment process, and the latter is modeled by spatially varying coefficients that follow a multivariate Gaussian process. We apply the proposed model to a real dataset of arsenic concentration and demonstrate better performance than competing models.