Estimation of severe crash frequency using two surrogates

Abstract: This thesis is concerned with the estimation of crash frequency based on the bivariate modeling of surrogate measures of safety (SMoS), which serve as indicators for traffic risk. Using the SMoS, any traffic conflict between two road users can be described by their proximity together with their hypothetical consequence. We quantify traffic conflicts of different severity as random vector of proximity SMoS and consequential SMoS, and define the traffic risk as the probability measure over the random vector of SMoS pair. We use EVT both in its bivariate context and in approximating the marginal distribution of proximity SMoS, which is combined with copula, to compute the probability of severe collision. The 10-year frequencies of severe collision are also computed based on the fitted models. From a methodological point of view, the copula approach with EV margin is more favorable than bivariate EV models, as collisions of lower severity can also be computed. From an implementation point of view, the bivariate EV model is more favorable, as the assumptions on the marginal distribution are defined by the model. The new approach that combines EV distributions and copula is found to have the most accurate estimated crash frequency given that the police report was used a reference, for our data set.