ABSTRACT：Safety performance functions (SPFs) are critical for traffic safety management. They have been applied for identifying significant risk factors, estimating crash frequencies, and screening potentially hazardous locations. Since SPFs proposed by Highway Safety Manual (HSM) are developed based on certain states in the United States, regions without jurisdiction-specific SPFs need model calibrations for the localization of SPFs. The main objective of this study is to compare the typical calibration methods that used in the literature and identify the appropriate ones. Random effects Negative Binomial (NB) models were established for urban arterials in Shanghai and Guangzhou during peak hours and off-peak hours separately. Four calibration methods, including the calibration factor, empirical Bayes (EB) method, K Nearest Neighbor (KNN) regression method, and pooled data, were applied. The performance in improving model transferability was measured by transfer index and the adaptability to insufficient data was assessed by necessary data collected for each method. Based on the modeling results, pooled data approach that composed of the entire Shanghai dataset and 50% proportion of the Guangzhou dataset provides the best performance. And EB method and KNN regression method are preferable to the calibration factor. Furthermore, modeling and calibrating for different time periods should be considered when average speed is incorporated in the model.
Xuesong Wang*, Dongjie Tang, Saijun Pei. Comparison of Calibration Methods for Improving the Transferability of Safety Performance Functions. Transportation Research Board 98th Annual Meeting, Washington D.C., USA, 2019. 1.13-17.