This study aims to identify key factors affecting crash frequencies under various times of the day, so as to propose effective and time-specific countermeasures. Two approaches are proposed and compared. The clustering approach combines a crash count model to predict total number of crashes and a clustering model to divide segments into clusters according to their time-of-day distribution patterns of crash frequency. The multivariate approach treats the crash frequencies of various times of the day as target variables and accommodates potential correlation among them. Crash datasets of Taiwan Freeway No.1 are used to estimate and validate the models. Four times of the day, late-night/dawn (24-06), morning/noon (07-13), afternoon/evening (14-19), and night (20-23) are formed according to crash count distribution. In terms of Adj-MAPE and RMSE, the clustering approach performs better than the multivariate approach. According to the clustering results, segments in metropolitan areas have higher crash frequency in the afternoon/evening, while those in rural areas have higher crash frequency in late-night/dawn, suggesting the time-of-day distributions of crash frequency markedly differ among segments. Time-specific countermeasures are then proposed accordingly.
- Time-of-day crash frequency distribution
- multivariate modeling approach
- negative binomial regression