Hourly associations between weather factors and traffic crashes: Non-linear and lag effects

•Examined the non-linear and lag effects of weather on crash risk.•A novel case-crossover design with distributed lag non-linear model was proposed.•The associations by different population characteristics were examined.•It could serve to inform decision makers in the context of climate change. Weather is well recognized as a significant environmental factor contributing to higher risk of road crashes. In the conventional road safety studies, weather effects had been set out either based on the instant weather conditions recorded by the police officer attained or the average of meteorological observations over a relatively long time period, such as daily, weekly or even monthly, etc. To the best of our knowledge, it is rare that the lag effect of weather in the preceding period on the crash risk in the current period was attempted. With the use of high-resolution meteorological data in very short time interval, it is possible to evaluate the role of lagged weather effect on safety. In this study, we propose a novel distributed lag non-linear model (DLNM), integrated with case-crossover design, to evaluate the lag effect of weather on crash incidence. The proposed modelling framework could describe the non-linear relationship between weather and crash and the lag effects. Also, the possible over-dispersion and autocorrelation of the time-series weather and crash data can be controlled for. The model was estimated using an integrated meteorological, traffic and crash dataset in Hong Kong. For instances, high resolution data on temperature, humidity, rain intensity and wind speed in 1-hour interval was available. The bi-dimensional exposure-lag-response surfaces are established to visualize the varying effects of possible weather factors on crash risk, with respect to the lag size. Such relationship between effect size and lag size is often overlooked in the literatures. Results indicate that model with 4 degrees of freedom for both weather condition (knots at equal spaces) and lag time (knots at equal intervals) best fit with the observations, in accordance to Quasi-likelihood Akaike information criterion (Q-AIC). Then, stratified analyses are conducted to evaluate the difference in the association among different clusters. Findings should shed light on the modelling of non-linear exposure-response relationship and lag effects in traffic safety time series analysis.