Accounting for seasonal effects on cyclist-vehicle crashes

Abstract Crash data is usually aggregated over time where temporal correlation contributes to the unobserved heterogeneity. Since crashes that occur in temporal proximity share some unobserved characteristics, ignoring these temporal correlations in safety modeling may lead to biased estimates and a loss of model power. Seasonality has several effects on cyclists’ travel behavior (e.g., the distribution of holidays, school schedules, weather variations) and consequently cyclist-vehicle crash risk. This study aims to account for the effect of seasonality on cyclist-vehicle crashes by employing two groups of models. The first group, seasonal cyclist-vehicle crash frequency, employs four vectors of the dependent variables for each season. The second group, rainfall involved cyclist-vehicle crash frequency, employs two vectors of the dependent variables for crashes that occurred on rainy days and non-rainy days. The two model groups were investigated using three modeling techniques: Full Bayes crash prediction model with spatial effects (base model), varying intercept and slope model, and First-Order Random Walk model with a spatial–temporal interaction term. Crash and volume data for 134 traffic analysis zones (TAZ’s) in the City of Vancouver were used. The results showed that the First-Order Random Walk model with spatial–temporal interaction outperformed the other developed models. Some covariates have different associations with crashes depending on the season and rainfall conditions. For example, the seasonal estimates for the bus stop density are significantly higher for the summer and spring seasons than for the winter and autumn seasons. Also, the intersection density estimate for a rainy day is significantly higher than a non-rainy day. This indicates that on a rainy day each intersection to the network adds more risk to cyclists compared to a non-rainy day.