为识别高速公路路侧事故的主要诱导因素,分析路侧事故起数与道路线形和交通条件之间的关系,以广珠(广州-珠海)东线高速公路3年中发生的178起路侧事故为基础,分别采用定长法和不定长法划分路段单元,从道路线形和交通条件中选取11个自变量,采用零堆积负二项回归模型建立路侧事故起数预测模型。选择Vuong检验统计量、对数似然值和AIC信息准则3个指标进行模型的拟合优度检验,选择相对误差和累积残差2个指标进行模型的拟合准确性检验;通过对比分析负二项回归模型和零堆积负二项回归模型的拟合优度和拟合准确性检验结果判断其优劣性,并采用弹性分析确定较优模型中显著自变量对因变量的影响程度。研究结果表明:无论采用定长法划分还是不定长法划分路段单元,零堆积负二项回归模型构建的路侧事故起数预测模型明显优于负二项回归模型;采用零堆积负二项回归模型构建的路侧事故起数预测模型,其定长法划分的路段单元模型的拟合准确性优于不定长法;对于定长法划分的路段单元,基于零堆积负二项回归模型的路侧事故起数预测模型有5个自变量对路侧事故起数均有显著影响,影响程度大小依次为车道数、曲率变化率、曲线比例、曲度和平均纵坡坡度。 In order to identify the main inducing factors of expressway roadside accidents, the relationship between the number of roadside accidents and road alignment and traffic conditions was analyzed. Taking the 178 roadsides occurred in the 3-year expressway of Guangzhou-Zhuhai Expressway Based on the accident, the fixed-length method and indefinite length method are respectively used to divide the road section units. Eleven independent variables are selected from the road shape and traffic conditions, and the zero-binning negative binomial regression model is used to establish the roadside accident count prediction model. Vuong test statistic, log likelihood ratio and AIC information criterion were selected to test the goodness-of-fit of the model. The relative error and the cumulative residual error were selected to test the fitting accuracy of the model. By contrasting the negative Binomial regression model and the zero-binomial negative binomial regression model and the fitting accuracy of the test results to determine its advantages and disadvantages, and use flexible analysis to determine the influence of the significant independent variables on the dependent variable. The results show that no matter whether the fixed-length method or the indefinite length method is used to divide the road segment elements, the forecasting model of roadside accidents with zero-accumulation negative binomial regression model is obviously better than the negative binomial regression model. The prediction model of roadside accidents is more accurate than the indefinite method for the lengthwise method of roadside unit model. For the roadway units classified by fixed length method, the road based on zero-stacking negative binomial regression model There are five independent variables on side accident prediction model have significant impact on the number of roadside accidents, the impact of the size of the number of lanes, curvature rate of change, curve ratio, curvature and average longitudinal gradient.