基于变量选择的交通事故数据建模研究

Research on Traffic Accident Data Modeling Based on Variable Selection

摘要: 道路交通事故数据作为经典计数数据，其建模在复杂交通事故分析应用中具有重要意义。本研究基于英国交通部的道路交通事故统计系统(STATS19)中2022~2023年的道路交通事故记录数据，构建涵盖时间效应、自回归效应及交互项的Poisson回归模型，系统对比了逐步回归与五种正则化方法(Ridge回归，LASSO，Elastic Net，SCAD，Adaptive LASSO)的性能。实证表明，SCAD与Adaptive LASSO均显著优于传统逐步回归，其中SCAD在预测精度与模型简洁性之间取得了最优平衡，Adaptive LASSO展现出优异的变量选择一致性与稳定性。本研究为计数数据在实际中的应用提供了兼具精度与稳定性的建模工具。

Abstract: Road traffic accident data, as a classic form of count data, plays a pivotal role in complex traffic accident analysis. This study utilizes road accident records from the UK Department for Transport’s STATS19 system (2022~2023) to construct a Poisson regression model incorporating temporal effects, autoregressive terms, and interaction effects. We systematically compare the performance of stepwise regression against five regularization methods: Ridge, LASSO, Elastic Net, SCAD, and Adaptive LASSO. Empirical results demonstrate that both SCAD and Adaptive LASSO significantly outperform traditional stepwise regression. Specifically, SCAD achieves an optimal balance between predictive accuracy and model parsimony, while Adaptive LASSO exhibits superior consistency and stability in variable selection. This study provides robust modeling tools that combine precision and stability for practical applications involving count data.