To analyze the impact of traffic composition on freeway safety deeply， the roadway， traffic and crash data on Kaiyang Freeway in Guangdong Province in 2014 were collected． The vehicles were classified into five categories according to the toll standard． A Bayesian hierarchical model and a conditional autoregressive ( CAＲ) model were developed to correlate roadwayrelated and trafficrelated attributes with crash frequency on freeway segments． Bayesian methods were used to estimate the parameters and to compare the models． The results of model comparison show that the CAＲ model，which accounts for the spatial correlation across adjacent freeway segments，outperforms the Bayesian hierarchical model． The parameter estimates in the CAＲ model suggest that there are 15. 5% and 24. 4% decreases in expected crash frequency on the freeway per 1% increase of Categories 1 ( e. g. ，automobile) and 3 ( e. g. ，medium coach and medium truck) vehicles， respectively． Moreover，crash frequency is found higher on longer freeway segments with more averaged daily traffic，bigger curvature and steeper grade．