Toward a new methodology for modeling highway accident frequency

Previous traffic accident models (regression, exponential, and survival) showed a strong association between accident frequency and highway characteristics. Using connected highway segments, these models relate accident frequency of any segment to traffic and geometric characteristics of the segment. That is, the effect of gradual or sudden changes in the characteristics of previous segments is not considered. The objective of this paper is to develop a time-series model to investigate the cases for which this effect should be considered to produce a better prediction of accident frequency. To achieve this objective, we considered connected highway segments as a space-domain series (a time series where the space is considered as a time increment). First, we model accident frequency as a function of traffic and geometric characteristics using the traditional regression model. Second, we test the random error term of the regression model. If the error term is autocorrelated, we employ the ARIMA time-series approach using the same independent variables of the regression model. The proposed methodology was applied to a highway in Japan using different accident types. The results showed that for some cases the error term was autocorrelated indicating that the accident frequency of a given segment depends not only on the characteristics of that segment, but also on the characteristics of the preceding segments. Thus, the effect of gradual or sudden changes in traffic and geometric characteristics on accident frequency can be evaluated. The proposed time series model fitted the observed data better than the regression model