Modeling the effects of AADT on the prediction of multiple-vehicle crashes for urban signalized intersections

The dissertation aims to address two key issues in predicting crash frequency: Safety Functional Functions (SPFs) for Annual Average Daily Traffic (AADT) and potential correlation among multiple years of data. AADT is often considered as a main factor for predicting crash frequency at urban and suburban intersections. A linear functional form is typically assumed for the Safety Performance Functions (SPFs) to describe the relationship between the logarithm of mean crash frequency aid logarithms of AADTs. However, such a linear relationship assumption has been questioned by many researchers. This study applies Generalized Additive Models (GAMS) with a bivariate (isotropic) smoother, Generalized Additive Mixed Model (GAMMs) and Piecewise Linear Negative Binomial (PLNB) regression models to analyze intersection crash data. The modeling results suggest that a nonlinear functional form may be more appropriate. Sometimes it is preferable to model such a nonlinear relationship by PLNB regression, since the result can be interpreted more easily. The results also show that it is important to consider the joint effects of multiple covariates, suggesting that the impacts of different covariates are not independent. In addition, the dissertation adopts fully specified GAMMs to account for potential unobserved heterogeneity in intersection crash data. The GAMM result indicates that the previously observed nonlinear functional forms are unlikely to be caused by model miss-specification.;When modeling intersection safety, it is common to collect multiple years of crash data in order to reduce data collection cost. To account for potential correlation among multiple years of crash data, Gaussian copula regression model is introduced in this research. Compared to traditional methods (e.g., Generalized Estimated Equation (GEE)) for modeling panel data, the proposed Gaussian copula model has the advantages of allowing more convenient model comparison and selection. For instance, models fitted using the GEE method cannot be compared based on conventional goodness-of-fit measures. This research fits various Gaussian copula models. The best-performing model's results indicate that it fits panel crash data well.;Keywords: Generalized additive model, Generalized additive mixed model, Generalized linear model, Piecewise linear negative binomial regression, Safety performance function, Isotropic smoother, Bivariate smoother, Gaussian copula.