Sensitivity analysis of Bayesian nonparametric spatial crash frequency models

Abstract

This study focused on the sensitivity analysis of Bayesian nonparametric spatial models which combine the strengths of spatially structured random effects and the Dirichlet mixture to account for the unobserved heterogeneity of crash counts. Various evaluation criteria were employed to compare the performance of models with varying spatial weight matrices and precision parameters (alpha). The results demonstrate that there exists strong correlation among the posterior number of clusters, alpha, the fraction of variation explained by the spatial random effect, and different evaluation criteria. Even though the increased upper bound value of alpha does not necessarily lead to the enhanced model performance, the models with the full flexibility to choose the desirable amount of clustering tend to perform better than those with limited flexibility due to smaller allowable mass components. Compared with the precision parameter, no obvious trend is illustrated for the different evaluation criteria along the varying spatial weight matrices.