Identifying the risk factors affecting crash severity at intersections with considering crash characteristics and signal configuration using an ordered logistic model Ameera Alkhlaifi Department of Engineering Systems and Management Khalifa University of Science and Technology Masdar City, UAE Email: aalkhlaifi@masdar.ac.ae Abdulla Galadari Department of Engineering Systems and Management Khalifa University of Science and Technology Masdar City, UAE Email: agaladari@masdar.ac.ae Abstract—Intersections are critical places which experience high levels of accidents due to the availability of vehicle move- ments from different directions. Therefore, there is a need to understand the factors that significantly contribute to injuries at such places. These factors can fall under different dimensions in traffic safety such as environmental conditions, road user behavior, traffic police enforcement, road design and crash characteristics. The paper will analyze the factors related to crash characteristics and traffic signal operation that affect the likelihood of accident severity located at intersections. The data for intersection accidents in Abu Dhabi from 2013 to 2016 were used in this analysis. Ordinal logistic model was considered for the analysis to account for the ordinal nature of severity levels. Several diagnostics tests of the model were preformed such as parameters evaluation, overall model evaluation and prediction accuracy. For parameter evaluation, out of 11 independent variables, 6 were non-significant and dropped from the model. Most of the non-significant variables were related to the driver- at-fault details. The results of the final model showed an overall good fit based on a p-value of less than 0.05, as well as a good accuracy of prediction 84.8%. Finally, odds ratios were estimated to interpret the final results of the model. Keywords—Ordinal Logistic Mode, Accident severity, Intersec- tions. I. I NTRODUCTION Many countries such as United States found that 43% of crashes occurred at intersections[1]. Oman also reported a percentage of 47% of intersection related accidents[2]. Such lo- cations do not only experience high accident rates, in fact, most of sever accidents occur at intersections[3], which requires continuous efforts from various parties in traffic safety system. In consideration of that, most traffic safety engineers rely on assessing the safety level of the road using the final outcomes of the road safety system; which are expressed as accidents per unit of exposure. Although these rates provide a solid fact about the general safety conditions of the road, they fail in addressing the significant details behind accidents that might help in mitigating the problem in the future. For this reason, previous researchers expanded their efforts in understanding the relationship between risk factors and crash severity[4], [5], [6]. Even though the work of previous researchers were carried out in different countries and gave an idea about the factors that are influencing accident severity, the same factors might not be applicable for different regions. For this reason, the study will implement the accident data observed at Abu Dhabi intersections. The behavior of the driver is associated with accident’s occurrence. However, when it comes to accident severity, the case might not be the same. The factors that are influencing crash severity may be different from the factors affecting crash occurrence. [7]. Thus, the study will include the factors related to driver-at-fault behavior during the accident. Additionally, the study aims to contribute to the understanding of the ordered logistic modeling methodology and the diagnostics testing associated with measuring the reliability of the model. II. MODEL SELECTION The basic and most well-known model is the linear regres- sion model which deals with analyzing and fitting the data linearly when the dependent variable is continuous, however, when dealing with a dichotomous dependent variable, the case might be different. Such a case violates the assumptions of linear model and the inferences of results will be invalid.[8]. As a solution to such a problem, multinomial regression was recommended to account for the non-interval nature of the dependent variable as it uses a logit function to model the probabilities of an event.The simple form of the logistic model is given as the following: ln( P 1 − P )= β 0 + β 1 x 1 + ... + β k x k (1) where, P is the probability of an event occurrence P(Y=1), β 0 is the intercept, x k is the exploratory variable β k is the coefficient of the exploratory variable. Previous researchers analyzed accident severity data using a multinomial logistic model, considering that accident data are categorical in nature. Severity levels is categorical and ordinal (ranging from low to high). multinomial Regression does not