Relationship Between Speed Metrics And Crash Frequency And Severity PDF Download

Roadway safety is a high priority among transportation agencies domestically and internationally. In order to improve traffic safety, the research community has been working to develop and apply more advanced statistical modeling methods in an effort to identify relationships between crash frequency or severity and various driver, roadway, and vehicle characteristics. For example, crash frequency models have evolved from Poisson and negative binomial (NB) regression models, to models that handle excess zero counts (e.g., zero-inflated Poisson or NB models), and to models that address unobserved heterogeneity (e.g., latent class and random parameters models). Examples of crash severity models include multinomial, ordered, and nested logit models. These statistical models usually contain assumptions that are not straight-forward to assess, or put limitations on variables and data. To alleviate these concerns, researchers have recently introduced data mining techniques into traffic safety research. Examples of these methods include classification and regression trees (CART), random forests (RF), support vector machines (SVM) and neural networks (NN). Each has shown promising results to predict the frequency or severity of traffic crashes, which is essential to managing the safety performance of a roadway network.Most safety studies that have applied data mining algorithms have focused on model outcomes of specific crash types and have not addressed parameter sensitivity and its influence on model performance. Furthermore, the application of more sophisticated algorithms has been limited in the published traffic safety literature. In light of this, the purpose of the research is to fill these gaps by systematically evaluating the predictive power of CART and RF algorithms in crash severity analysis and the MOB algorithm in a crash frequency context, and by exploring the sensitivity of key parameters that affect model performance within each algorithm. In addition, guidelines for applying select data mining models are generalized for application purposes. Empirical data collected from Pennsylvania are used to accomplish the research tasks. The major findings from this research are as follows: 1. The RF model produced better predictive power than the CART model, while the CART model performed slightly better than the binary logit model, in the crash severity context. 2. When increasing the parameter values in the CART algorithm (i.e., minimum node size and prior probability of fatal and injury crashes), the sensitivity (i.e., prediction accuracy) of fatal and injury crashes increases and the precision decreases.3. The RF model is robust to parameter tuning. The evaluation metrics are stabilized when more than three variables are considered at each node splitting. The model can produce satisfactory performance with limited number of trees (i.e., far less than the default value). The variable importance ranking is not sensitive to different parameter settings.4. The MOB-NB model shows better data fitness than traditional NB models, in terms of log-likelihood and AIC, in the crash frequency context.5. The estimated confidence intervals and elasticities of independent variables suggest that the MOB-NB model can efficiently identify variable effect heterogeneity under different subgroup patterns in the dataset. In the crash frequency context, the existence of passing zones and the posted speed limit were identified as subgroups in the present study.6. The MOB-NB model produces the minimum mean square error among candidate models, including standard NB model and adjusted NB models, which incorporate splitting variables identified by MOB-NB model. The difference in prediction accuracy among the models is relatively small.