Lane-changing is a complicated task and has a high probability of accident occurrence. Although a large body of literature has used vehicle trajectories to microscopically understand and model lane-changing behavior, most of these studies focus on lane-changing decision making and lane changing's impacts on surrounding vehicles, not on traffic safety. The contributing factors to lane-changing risks have not been fully explored from the perspective of microscopic behavior using vehicle trajectory data. This study investigates the contributing factors to accident risks in different lane-changing patterns with taking unobserved heterogeneity into account. A vehicle trajectory dataset, HighD is used and 4842 lane-changing vehicle groups are extracted for analysis. These vehicle groups are divided into sixteen patterns according to the vehicle type, and three major patterns are examined. A lane-changing risk index (LCRI) is proposed to evaluate the risk level of each vehicle group. Two methods are developed and compared for exploring lane-changing risks of the three patterns including (1) establishing the random parameters fractional logit models; and (2) classifying LCRI by k-means algorithm and establishing random parameters ordered logit models with heterogeneity in means and variances. The modeling results show that the latter method performs better and the risk level of the vehicle group is strongly associated with (1) the mean and standard deviation of the gap distance between vehicles; (2) the longitudinal velocities and acceleration of vehicles; and (3) the lane-changing direction and duration. However, different patterns are found to have different contributing variables and effects. The effects of gap distances vary considerably across different vehicle groups and the longitudinal velocity of vehicles are associated with the means of random parameters for gap distance.

变道是一项复杂的任务，并且发生事故的可能性很高。尽管大量文献使用车辆轨迹来微观地理解和模拟车道改变行为，但是这些研究大多数集中在车道改变决策和车道改变对周围车辆的影响上，而不是对交通安全的影响。从微观行为的角度出发，未使用车辆轨迹数据对导致变道风险的因素进行了充分的研究。本研究在考虑到未观察到的异质性的情况下，调查了不同换道模式下事故风险的成因。使用了车辆轨迹数据集HighD，并提取了4842个变道车辆组进行分析。这些车辆组根据车辆类型分为16种模式，并考察了三种主要模式。提出了变道风险指数（LCRI）以评估每个车辆组的风险水平。开发和比较了两种方法来探索三种模式的变道风险，包括：（1）建立随机参数分数对数模型；（2）利用k-means算法对LCRI进行分类，建立均值和方差均异的随机参数有序logit模型。建模结果表明，后一种方法效果更好，且车辆组的风险水平与（1）车辆之间的间隙距离的均值和标准差密切相关；（2）车辆的纵向速度和加速度；（3）变道方向和持续时间。但是，发现不同的模式具有不同的贡献变量和效果。