Institute of Transportation, MOTC

       This study concatenates three sets of questionnaire data describing driving behavior (i.e. behavior diagnosis test, hazard perception test, driver knowledge test), driving assessment and crash data which were derived from employees of a foreign company from 2016 to 2019. The data were analyzed for the relevance of the cause of crashes, and driver characteristics to clarify the factors that affect driving behavior. K-means clustering method was used to classify drivers into high, medium-low risks according to the total score of risk perception to analyze the relevance of driver factors and different risk groups. The crash types were merged into front-end, sideswipe, T-bone and others. The types of vehicles involved in the crash are divided into vehicle-vehicle, scooter-vehicle, pedestrians and other single-vehicle incidents. The variables that do not have collinearity are classified into variable types (binary variables, categorical variables, continuous variables) were established based on logit regression, probit regression, complementary log-log regression, multinomial logit regression, Poisson regression, negative binomial regression and linear regression to analyze the factor whether the driver is involved in a crash, crash severity (whether injured), the number of crashes, the type of crashes, the types of vehicles involved in a crash, driving risk and driving behavior. The stepwise regression is used to exclude insignificant variables. Finally, the ROC curve is used to select a binary classification model with higher discriminatory power. The results of the study found that age, education level, violation experience in three years, total stress score of behavior diagnosis test, urban roadside vehicle driving door of hazard perception test, mountain passers-by walking in the middle, and mountainous right road vehicle driving out and the technology knowledge score and the maintenance score of the driver knowledge test are all significant factors. Among them, the more violation experience the driver has within three years, the higher the risk (the model prediction results fall in the high risk group). This study also found that the coefficients of some significant variables are contrary to common sense. For example, drivers with higher safety perception scores are more likely to be involved in a vehicle-to-vehicle crash.