Temporal shifts in safety states through the COVID-19 pandemic: Insights from hidden semi-Markov models.

Abstract

The COVID-19 pandemic significantly impacted transportation safety, with an increase in risky driving behaviors observed during the initial lockdown period, leading to a higher likelihood of severe crashes. However, there is limited research on the post-pandemic effects on driving behaviors and safety. This study addresses this gap by analyzing open data from the state of Virginia to examine shifts in safety states from 2016 to 2024, covering the pre-, during-, and post-pandemic periods. Structural equation modeling (SEM) was utilized to measure latent variables representing aggressive and inattentive driving behaviors and to model their impacts on crash severity. Additionally, hidden semi-Markov models (HSMMs) were applied to infer shifts in safety states associated with these risky driving behaviors and the proportion of severe crashes. The strength of HSMM models lies in the ability to distinguish meaningful pattern changes from random noise. Compared with hidden Markov models (HMMs), HSMMs provide greater flexibility by accommodating arbitrary state duration distributions, contributing to better model performance and more reliable inferences. The HSMMs with four hidden states were utilized to reveal shifts in safety states over the eight-year analysis period in Virginia. Results suggested that safety states related to risky driving behaviors and the proportion of severe crashes were at lower-risk levels pre-pandemic from 2016 to 2019, then escalated to the highest-risk levels during the pandemic in 2020 and remained at higher-risk levels in 2021, 2022 and 2023. By 2024, safety states have returned to lower-risk levels similar to those inferred in the pre-pandemic period. A seasonal pattern was also identified in safety states, with lower-or-lowest-risk levels occurring in winter near the holiday season.