Improving crash predictability of the Highway Safety Manual through optimizing local calibration process

The predictive method of the Highway Safety Manual (HSM) estimates crash frequency by applying an uncalibrated safety performance function (SPF) and a set of uncalibrated crash modification factors (CMFs) to each location individually; then the predicted crashes must be adjusted by a local calibration factor (LCF) at the aggregate level for at least 30-50 sites per SPF. Although this calibration procedure assures total predicted crashes will be localized, still the prediction of crashes for individual locations suffers from the aggregate localization process. An alternative approach of locally calibrating the HSM predictive method is proposed to improve prediction quality at individual locations while maintaining equality of total observed and total predicted crashes. The methodology incorporates multiple calibration factors for different components of the predictive method (SPF parameters and CMFs) rather than a single calibration factor as recommended by the HSM that only calibrates at the aggregate level. In the proposed method, the application of calibration factors expressed in both weight and power function better reflects the local conditions while still ensuring calibration at the aggregate level. The parameters are estimated through an optimization process of five different methods. Rural two-lane, two-way roads (R2U) data was used from the states of Maryland, Illinois, and Washington. A tool named "Roadway Safety Data Integrator (RSDI)" was developed for data preparation. Different Goodness-of-Fit measures along with CURE plots indicated that the proposed method performed significantly better than the HSM calibration method, calibration function (that will most likely be calibration process in the HSM 2^nd edition), calibrated Washington models (for the case of Washington data), and some alternative calibration methods suggested by past studies. Moreover, the results indicated that the additional parameters for CMFs could improve the prediction significantly; a previous study did not find this to be so due to data limitations, but we have improved the methodology and are not so limited. Application of the proposed approach can lead to more accurate identification of hot-spots and site-specific strategies. Considering the limitations of this study, some avenues for further research are discussed.