The development of aggregated macro-level safety prediction models

This dissertation explored the prediction of safety outcomes at the aggregate macro-level. The objective of this dissertation is to further the knowledge in the field of accident prediction by investigating and modeling the phenomenon of crash occurrence at the macro-level. Aggregate macro-level models lend themselves to prediction of safety at the planning stage and for the prediction of safety in the absence of detailed information of individual network elements and characteristics. The frequency of crashes, levels of crash severity and crash types (including vehicles, pedestrians, and deer) were modeled using Poisson and negative binomial (NB) regression. Models were evaluated across different levels of aggregation (census block group to tract, census tract to county, traffic analysis zone (TAZ) to county) and across jurisdiction (state and metropolitan planning organization (MPO) level). Findings from the study suggest that the aggregate macro-level modeling process requires particular attention to transferability the impact of particular crash types on aggregated crash counts (such as deer crashes) and correlation between independent variables at different levels of aggregation and across jurisdictions. Spatial autocorrelation did not significantly impact parameter estimates. The dissertation presents two major recommendations. A reduction in maximum correlation coefficients is recommended between independent variables in the same regression model to 0.5 to address concerns regarding multicollinearity at sub-regional and jurisdictional level. Model transferability was tested across jurisdictions and different levels of aggregation within jurisdictions. Findings suggest that the impact of regional and aggregation differences on aggregate safety outcomes is substantial. The census block group was found to be the most ideal study area unit for macro-level safety prediction modeling. The block group level allows for richness of model specification and produces stable parameter estimates and superior model performance. This research adds to the understanding of aggregate level correlational relationships between socio-demographics and roadway characteristics on safety across different boundary definitions levels of aggregation and jurisdictions. The findings and recommendations provide direction in terms of dataset development and approaches to aggregate modeling.