Probabilistic models of freeway safety performance using traffic flow data as predictors

In this paper we lay the groundwork for gauging the level of safety of any type of traffic flow on a freeway, based on data from single loop detectors; the procedure can be implemented wherever such data are monitored or simulated. Our analyses are based on loop detector data for each of the freeway lanes for a short period of time preceding teach of over 1700 accidents in our case study. This case study covers the six major freeways in Orange County, California, for a six-month period in 2001.Recognizing that loop detector data at a specific time and place cannot be converted to speed, because it is not possible to know effective vehicle length at such a detailed level (that is, the mix of long and short vehicles is unknown at a specific place for a short period of time), we avoid using any direct speed or density measures among the parameters. Rather, we employ explanatory parameters that include not only central tendencies (means and medians), but variations, and measures of systematic and synchronized traits that capture patterns in short period of loop detector data. Such patterns include breakdown from free flow to congested operations or recovery back to free flow, and differences in traffic conditions across lanes. In the analysis, we uncover an extensive set of statistical parameters that capture those aspects of traffic flow that are strongly related to accident potential. We demonstrate that the parameters can account for speed and density, even though these are not used directly. Moreover, the parameters account for important differences among the types of accidents that occur under different types of traffic flow.