Analysis of Influencing Factors for Nighttime Pedestrian-vehicle Crash Injury Severity Considering Temporal Instability

Nighttime pedestrian-vehicle crashes exhibit significantly higher injury severity than daytime crashes due to visibility limitations and other factors. To accurately identify influencing factors, this study develops a hybrid model integrating a random parameters Logit model with heterogeneity in means and variances and a random forest (RF) algorithm based on SHapley Additive exPlanation (SHAP), i.e. RF-SHAP, using crash data from 2017 to 2022. The log-likelihood ratio test confirms temporal instability in the dataset, necessitating separate models for 2017—2019, 2020, 2021, and 2022 with calculated average marginal effects for significant variables. Results demonstrate that random effects exist for drinking pedestrians (2017—2019), ambulance required (2020), local street crashes (2021), and 48—56 km/h speed limits (2022), with their mean/variance influenced by traffic control and road classification. Drinking pedestrians, pedestrians aged over 45 to 60 years, driver injuries, vehicle types (pickup trucks and trucks), divided roadways, speed limits (32—40, 48—56, 64—72 km/h), weekends, and winter conditions have begun to exhibit statistically significant effects on nighttime pedestrian-vehicle crashes in recent years. In addition, the RF-SHAP algorithm quantifies heterogeneous contributions of all sub-variables within four random parameters to crash severity. Policy implications highlight three priorities: addressing pedestrian drinking behavior, enhancing nighttime crash prevention on expressways and arterial routes, and establishing appropriate speed limits while avoiding excessively high or low values.