Bayesian Bivariate Semiparametric Spatial Models for Ozone and PM2.5 Emissions

Abstract

This study focused on the development of Bayesian bivariate semiparametric models for ozone and PM2.5 emissions. The semiparametric models rely on a Dirichlet mixture, which accounts for unobserved heterogeneity by employing random distribution for the intercept of each entity. The models with and without spatially structured random effects were also evaluated. Overall, four models were developed: three semiparametric with a flexible intercept and one parametric to serve as a reference. The significance of simultaneous estimation of PM2.5 and ozone was demonstrated by selection of common contributing factors as well as the statistically significant bivariate error term. In terms of relationship between dependent and independent variables, the areas with higher population and poverty, along with lesser educated residents, were observed to experience higher ozone and PM2.5 concentration. Also, the entities with higher vehicular traffic density and larger geographic areas were observed to be positively correlated with ozone and PM2.5 concentration. The underlying influential factor for both such variables may be the vehicular emissions, which is directly associated with traffic density and indirectly with land area as larger entities may tend to have higher traffic activity. The superiority of the semiparametric models, compared with parametric, signified the advantage associated with the flexible Dirichlet approach. The models with and without spatial correlation structures illustrated the mixed performance across the cases of parametric and semiparametric bivariate ones. The lack of consistent advantages associated with the inclusion of spatial effects may be due to the fact that the spatial correlation was not observed to be significant for the current dataset. Other spatial structure specifications may lead to different performance results.