Investigation of a New Approach for Representing Traffic Volumes in Highway Crash Analysis and Forecasting

University Transportation Centers Program2008PDFResearch PaperIvan, John N.Zhang, ChenUniversity of Connecticut. Connecticut Transportation InstituteNew England University Transportation CenterUnited States. Department of Transportation. University Transportation Centers (UTC) ProgramUnited States. Department of Transportation. Region I University Transportation CenterUnited States. Department of Transportation. Region I University Transportation CenterNew EnglandCrash analysisCrash causesCrash characteristicsCrash exposureCrash risk forecastingFrontal crashesHighway safetyTraffic safetyRear end collisionsTraffic volumeTraffic crashesTraffic crash victimsCrash victimsCasualtiesTraffic exposureCrash causalitiesContributing factorsRear-end crashesHead-on crashesUCNR17-7DTRS99-G-0001Accident prediction modeling studies are good for identifying correlations between crash risk and explanatory factors, but cannot give definite safety effects of countermeasures related to the significant covariates. Causality can only be proven using crash reconstruction methods or carefully constructed before-after studies, both of which require special data observation or analysis time, making them impractical to apply on large study data sets representing the variation of characteristics in the road network. This report proposes two advances in crash modeling: (1) a collision type categorization based on factors contributing to the occurrence of collisions to support estimation of prediction models that can better identify crash causality, and (2) definition of crash exposure to consider the traffic flow situation that is necessary for specific collision types to occur. Generalized linear modeling assuming a negative binomial or scaled Poisson distribution is used to estimate prediction models. Evaluations of model goodness-of-fit and diagnostics results are discussed in comparing different model outcomes and assessing the effectiveness of adopting the newly defined collision type categories and crash exposure. Three studies are carried out to evaluate the new exposure definitions and the collision categorizing method based on contributing factors. In study 1, a new exposure, crash opportunities, defined as the number of times vehicles traveling in opposite directions meet, is proposed for opposite-direction collisions including head-on and sideswipe. However, models using crash opportunities do not perform better than those using traffic volumes. Study 2 uses K-means cluster analysis to categorize collisions into categories with similar patterns of contributing factors. The resulting categories are (1) rear-end collisions, (2) other same-direction collisions including sideswipe and turning, (3) intersecting-direction collisions including turning opposite-direction, turning same-direction, intersecting path and angle, and (4) segment collisions including head-on, opposite-direction sideswipe and single vehicle. Study 3 then proposes vehicle time spent following as a new exposure definition for same-direction collisions and evaluates this new exposure for the collision categories from study 2. The new exposure is found to have a linear relationship with the total same-direction crashes, especially rear-end crashes. This linearity shows that vehicle time spent following can well represent traffic flow intensity and state for rear-end crashes to occur. Also, this finding further indicates the value of categorizing collisions by their contributing factors instead of types of crashes.127