Stability of regression coefficients corresponding to modeling factors for taxi speeding likelihood

Objective This study investigates the stability of regression coefficients corresponding to modeling factors for taxi speeding likelihood across different time scales, identifying stable factors that can support taxi safety management. Method GPS data were collected from 5 719 taxis in Chengdu City from 1st to 28th of November, 2016. Then, this study identified the speeding behaviors and measured the percentage of speeding distance with three time scales as the speeding likelihood, i.e., weeks, weekday & non-weekend, and workday. The operational factors were simultaneously extracted from three time scales. Bayesian random coefficient model was applied to model speeding likelihood in each time unit. The stability of regression coefficients was subsequently examined through the likelihood ratio test. Result The coefficients of various operational factors vary significantly across three time scales. The coefficient of driving distance ratio during peak hours is the most unstable, as it yields significant variation across time scales. The regression coefficient of daily driving distance is unstable neither. In contrast, regression coefficients of factors are relatively stable across all time scales, e.g., ratio of driving at night, ratio of driving on low-speed-limit roads, average hourly passenger capacity, average one-way passenger distance, and average no-load distance. The elasticity analysis further indicates that the pursuit of higher revenue is the primary reason for increasing taxi speeding likelihood. Conclusion Among contributing factors of taxi speeding likelihood, the partial operational factors' regression coefficients are stable across time scales, which require common countermeasures to tackle the speeding behaviors. However, for the factors with unstable regression coefficients, they require more tailored and flexible countermeasures to improve the effectiveness of safety management strategies.

研究目标：本研究旨在探究出租车超速概率建模因子对应的回归系数在不同时间尺度下的稳定性，识别可支撑出租车安全管理的稳定影响因子。
研究方法：本研究采集了2016年11月1日至28日成都市5719辆出租车的全球定位系统（Global Positioning System, GPS）数据。随后，本研究识别超速行为，并以三种时间尺度下的超速距离占比作为超速概率指标，分别为周度、工作日与非周末时段、工作日。同时从三种时间尺度中提取运行相关因子。采用贝叶斯随机系数模型（Bayesian random coefficient model）对各时间单元下的超速概率进行建模，并通过似然比检验（likelihood ratio test）考察回归系数的稳定性。
研究结果：各类运行因子的回归系数在三种时间尺度下存在显著差异。高峰时段行驶里程占比的系数稳定性最差，其在不同时间尺度间存在显著波动；日均行驶里程的回归系数同样不稳定。与之相对，部分因子的回归系数在所有时间尺度下均保持相对稳定，例如夜间行驶占比、低限速道路行驶占比、平均每小时载客量、平均单程载客里程及平均空载里程。弹性分析进一步表明，追求更高营收是出租车超速概率升高的主要诱因。
研究结论：在出租车超速概率的影响因子中，部分运行因子的回归系数在不同时间尺度下保持稳定，对此需采用通用管控措施以治理超速行为；而对于回归系数不稳定的因子，则需制定更具针对性与灵活性的管控策略，以提升安全管理举措的实施效果。