Traffic fatalities and injuries on urban roads especially at urban intersections constitute a growing problem in China. This study aims at researching urban intersection crashes in China and drawing conclusions by using hierarchical structured data with reference to Bayesian network (BN). On the basis of 3584 recorded crashes collected from the urban intersections of Changshu, China, a BN topological structure is developed to reflect the hierarchical characteristic of crash variables. The parameter learning process is completed with Dirichlet prior distribution. Junction tree engine is used to make inference on crash types at urban intersections with two respective given evidences, i.e. human factor and vehicle type. Parameter learning results suggest the efficacy of BN approach in the prediction accuracy. The average learned probability of illegal driving is 40.83%, which is much higher than other learned probabilities of human factors. The inferred probabilities of frontal collision at urban intersection crashes involving bicycles and electric bikes are 43.16% and 40.44% respectively, which is higher than the probabilities involving small cars and heavy vehicles. However, heavy vehicles have a higher inferred probability in side collision than light vehicles, whose inferred side collision probability is 41.02%. This study has a good potential in traffic safety discipline to reveal the correlation exists in traffic risk factors. By means of BN, researchers can make an intensive study on the hierarchical traffic crash data, determine the key risk factors and then propose corresponding and appropriate improvement measures. First published online: 16 Oct 2013