不同区划方案对宏观交通事故建模的影响

依托美国佛罗里达州Hillsborough县历史数据，分别提取人口普查单元组、交通分析小区、人口普查区、邮政投递区等4种区划方案的事故数据、路网交通特征数据和经济 社会 人口数据；基于贝叶斯方法构建负二项条件自回归模型，从模型拟合度、模型估计参数、小区事故黑面识别等3个方面定量评价不同区划方案对宏观交通安全分析结果的影响.研究表明：宏观交通安全分析结果会随着空间单元划分方式不同而产生显著差异；小区数目越少，事故预测越为准确；对比人口普查单元组、人口普查区和邮政投递区，基于交通分析小区的模型拟合度最低；变量中等家庭收入对分区规模最不敏感，其参数估计结果具有稳健性和可靠性.
Based on the historic data from Hillsborough County, Florida, U.S., the zone level factors including crashes counts, road network, traffic pattern, and various social economic factors were explicitly collected for four different zoning schemes, i.e. block groups, traffic analysis zones, census tracts, and zone improvement plan codes. Then, a Bayesian negative binomial model with conditional autoregressive prior was developed for each spatial units, respectively. The impacts of zonal variations on macro level safety modeling were investigated mainly from three aspects, i.e. model performance, model parameter estimates, as well as crash hotspots identification. Results revealed that statistical results based on different aggregation configurations could be significantly different. Zoning schemes with less number of zones tend to have higher crash prediction precision. Compared with block groups, census tracts, and zone improvement plan codes, traffic analysis zones level model preforms worst in terms of model goodness of fit. The variable of median household income shows consistently significant effects on crash frequency and is robust to variation in data aggregation