Transferability of safety performance functions and hotspot identification for freeways of the United States and China

https://doi.org/10.1016/j.aap.2020.105493Get rights and content

Highlights

  • The transferability of freeway SPFs between the U.S. and China was investigated.

  • The capability of transferred SPFs for hotspot identification was further analyzed.

  • SPFs were transferable between Texas/New York and Shanghai/Suzhou, but not between Florida and Shanghai/Suzhou.

  • The poor transferability was attributable to the considerable difference in traffic volume.

Abstract

Safety performance function (SPF) has been a vital tool in traffic safety evaluation including finding contributing factors to crashes, identifying hotspots, and assessing safety effects of countermeasures. In the United States (U.S.), the Highway Safety Manual provides a number of SPFs for a variety of road facilities. Due to the limited availability of traffic data in many regions, the transferability of SPFs has been an important topic in traffic safety analysis and has been evaluated by several studies. Nevertheless, the international transferability of freeway SPFs and the applicability of transferred SPFs on hotspot identification has been rarely investigated. Based on data from two Chinese cities, Shanghai and Suzhou, and three U.S. states, Texas, New York and Florida, this study analyzes the transferability of freeway SPFs between Chinese and U.S. regions. These SPFs are then transferred to the other country and their performance on hotspot identification is investigated. SPFs were developed in the frameworks of Poisson, Poisson-lognormal and negative binomial regressions for the five localities separately, and were calibrated using the calibration functions before being transferred. Without calibration, the poor model transferability was found between the two countries, while after calibration, the transferred SPFs between Shanghai/Suzhou and Texas/New York showed satisfactory performance on both model fitting and hotspot identification. However, the transferability of SPFs between Florida and the Chinese cities turned out to be unsatisfactory regardless of whether being calibrated or not, which was attributable to the considerable difference in traffic flow. The findings of this study are expected to be a good reference for researchers and practitioners who want to understand the transferability and applicability of SPFs in the international context.

Introduction

The rapid expansion of the freeway system in China has increased the exposure to crash occurrence and made freeway safety management one of the top priorities. To improve safety, accurate freeway hotspots should be identified to provide effective countermeasures to the locations with a priority. Hotspot identification is to find the most hazardous sites from massive road networks. A simple ranking using the number of observed crashes has been widely used to identify hotspots, but the nature of random fluctuations of crash data limits the reliability of the simple ranking method. Crash prediction model based methods can ameliorate random fluctuations but they require safety performance functions (SPFs) that areas with data limitations are difficult to develop. Such areas, however, may sometimes be able to apply SPFs developed for other areas. Therefore, the transferability of SPFs becomes an important research topic for the areas without sufficient data.

An SPF, a regression model that relates crash frequency to roadway geometric and traffic data, has been used for traffic safety analysis with various purposes. In the United States (U.S.), the Highway Safety Manual (HSM), published by the American Association of State Highway and Transportation Officials (AASHTO), and SafetyAnalyst®, a set of software tools developed by the Federal Highway Administration (FHWA), have provided systematic SPFs for different road entities such as urban arterials, intersections and freeways (Bonneson, 2010; Harwood et al., 2010). U.S. cities with no available local SPFs may adopt, or apply, the HSM SPFs, which can be modified and improved for local conditions by using a calibration factor. SPFs have been modified from HSM to local conditions in various states (Srinivasan et al., 2011; Xie et al., 2011; Shin et al., 2015; Troyer et al., 2015), and investigations of state-to-state transferable SPF characteristics (Farid et al., 2016, 2018b) have been conducted in the U.S. Considering China’s lack of safety modeling for road facilities, and of a systematic road safety manual like HSM, the study of SPF transfer between China and the U.S. has become imperative.

The transferability of SPF is assessed by measuring its goodness-of-fit, or the difference of log-likelihood between local and transferred SPFs. The transferability assessments are made on a one-to-one level, e.g., crash prediction of one segment is compared with the crash observation of the same segment. In general, however, more than one segment of a road is identified as a hotspot, which relaxes the precision requirement of one-to-one crash prediction. Consequently, indices for the transferability assessment do not always reflect the effectiveness of SPFs on hotspot identification. This study, therefore, has two main objectives: first, to evaluate the transferability of freeway SPFs between China and the U.S., and provide insights into the reasons for transferability or non-transferability. Second, to investigate the applicability of transferred SPFs for hotspot identification.

In China, sufficient data is available for Shanghai to develop SPFs. In 2006, Shanghai Traffic Police established an electronic database of crash information, and in 2009, the Shanghai Road Traffic Crash Analysis and Warning System was developed by Shanghai’s Tongji University. Since then, Shanghai Police has used this system to conduct temporal and spatial analyses of crashes and to investigate the causes of crash occurrence. In 2012, Shanghai implemented the Traffic Crash and Violation Location Description Standard, which locates crashes by using a five-element (road name, side of the road, referencing point, direction, and distance) recording method based on a linear reference system, and have improved crash location accuracy to 98 % (Li et al., 2018). Since 2010, local SPFs for urban arterials (Wang et al., 2015; Li and Wang, 2017), signalized intersections of urban arterials (Xie et al., 2013; Wang et al., 2014; Xie et al., 2014), suburban arterials (Wang et al., 2013, 2018c), and freeways (Wang et al., 2016, 2018a) have been developed, and an SPF transfer study has been conducted between Shanghai and Guangzhou (Wang et al., 2018b). Subsequently, Suzhou has implemented the same standard to locate crashes from 2016 in collaboration with Tongji University, and has improved its freeway crash location accuracy to 95 % by 2018. Based on the improved crash data, Shanghai and Suzhou can be used to study the international transferability of SPFs.

Hence, freeways in the two Chinese cities, Shanghai and Suzhou, and three U.S. states, Texas, New York, and Florida were investigated in this study. To obtain more reliable results, three widely used model structures were selected to develop SPFs, namely Poisson Regression, Poisson-Lognormal (PLN) Regression, and Negative Binomial (NB) regression. Similarly, two commonly used hotspot identification methods were adopted, i.e., the Empirical Bayes Method and the Potential for Safety Improvement. When transferring the SPFs, the calibration function was utilized to improve the SPFs for local conditions. Several measures including the transfer index, the cumulative residual plot, mean absolute deviance and root mean square error were used to evaluate the transferability of SPFs. And the capability of transferred SPFs for hotspot identification was assessed by testing the number of consistent hotspots identified by transferred and local SPFs.

Section snippets

Literature review

The HSM’s SPFs have been transferred to various states in the U.S., such as Oregon, North Carolina, Alabama, Florida, and so on (Srinivasan and Carter, 2011; Xie et al., 2011; Mehta and Lou, 2013; Lu et al., 2014; Claros et al., 2018). As revealed by previous studies, models, generally, are not transferable directly and calibration of the transferred model using local data appears to be necessary (Hadayeghi et al., 2006; Sawalha and Sayed, 2006; Venkataraman et al., 2016). The HSM provides a

Data preparation

This study is concentrated on freeways. Data was collected from two Chinese cities, Shanghai and Suzhou, and three U.S. states, Texas, New York, and Florida. The freeway G15 in Shanghai, G2, G1521 and G1522 in Suzhou, TX-45 and TX-130 in Texas, NY-440 in New York, SR408 and I-4 in Florida were investigated. Descriptions of these roads are shown in Table 1. Based on alignment features, these roads were divided into homogenous segments in traffic and geometric characteristics.

China’s crash data

Safety performance function

Three commonly used model structures for developing SPFs were investigated in this study. They are Poisson regression, PLN regression, and NB regression. The Poisson model has become the basic model framework for crash data because of the non-negative feature of crash count data, which cannot be handled by linear regression. However, one limitation of the Poisson model, the disability to account for the overdispersion existing in crash data, has prompted the derivation and wide use of PLN and

Modeling results

Freeway SPFs were developed for Shanghai, Suzhou, Texas, New York, and Florida, respectively. The modeling results are presented in Table 3. All variables in these SPFs were significant at the 95 % confidence interval. Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) were provided to show the model performance. For the five regions, AIC and BIC of the Poisson model were much larger than those of PLN and NB models, indicating that PLN and NB models performed better

Poor transferability

Different from the reasonably good inter-transferability of SPFs between Shanghai/Suzhou and Texas/New York, Florida’s SPFs showed inferior transferability than those of Texas and New York when applied to the Chinese cities, and its freeway data were also less receptive to the calibrated Chinese SPFs. In Table 5, MADs and RMSEs of calibrated Florida’s SPFs, when fitting the Chinese data, were always slightly higher than those of Texas’s and New York’s calibrated SPFs; in Table 6, the TIs of

Conclusion

This study has achieved two objectives. First, it has evaluated the transferability of freeway SPFs between China and the U.S., and provided insights into the reason for poor transferability. Second, it has investigated the applicability of transferred SPFs for hotspot identification. Data from two Chinese cities: Shanghai and Suzhou, and three U.S. states: Texas, New York, and Florida, was collected to develop freeway SPFs. When developing SPFs, three model structures were used: Poisson, PLN,

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This study was sponsored by the International Science & Technology Cooperation Program of China (2017YFE0134500).

References (48)

  • J. Lu et al.

    Developing local safety performance functions versus calculating calibration factors for safetyanalyst applications: a Florida case study

    Saf. Sci.

    (2014)
  • F.L. Mannering et al.

    Unobserved heterogeneity and the statistical analysis of highway accident data

    Anal. Methods Accid. Res.

    (2016)
  • A. Montella

    A comparative analysis of hotspot identification methods

    Accid. Anal. Prev.

    (2010)
  • Z. Sawalha et al.

    Transferability of accident prediction models

    Saf. Sci.

    (2006)
  • X. Wang et al.

    Safety modeling of suburban arterials in Shanghai, China

    Accid. Anal. Prevent.

    (2015)
  • X. Wang et al.

    Investigating the safety impact of roadway network features of suburban arterials in Shanghai

    Accid. Anal. Prev.

    (2018)
  • K. Xie et al.

    Corridor-level signalized intersection safety analysis in shanghai, China using Bayesian hierarchical models

    Accid. Anal. Prev.

    (2013)
  • K. Xie et al.

    Crash frequency modeling for signalized intersections in a high-density urban road network

    Anal. Methods Accid. Res.

    (2014)
  • K. Al Kaaf et al.

    Transferability and calibration of highway safety manual performance functions and development of new models for urban four-lane divided roads in Riyadh, Saudi Arabia

    Transp. Res. Rec. J. Transp. Res. Board

    (2015)
  • J.A. Bonneson

    Highway Safety Manual

    (2010)
  • S. Cafiso et al.

    Performance of safety indicators in identification of black spots on two-lane rural roads

    Transp. Res. Rec. J. Transp. Res. Board

    (2011)
  • Y. Chen et al.

    Improving transferability of safety performance functions by Bayesian model averaging

    Transp. Res. Rec. J. Transp. Res. Board

    (2012)
  • W. Cheng et al.

    New criteria for evaluating methods of identifying hot spots

    Transp. Res. Rec. J. Transp. Res. Board

    (2008)
  • B. Claros et al.

    Hsm calibration factor, calibration function, or jurisdiction-specific safety model–a comparative analysis

    J. Transp. Saf. Secur.

    (2018)
  • Cited by (15)

    • Transferability of the highway safety manual freeway model to EU countries

      2022, Accident Analysis and Prevention
      Citation Excerpt :

      The empirical results showed that calibration functions fitted better compared to calibrated HSM-default SPFs. Feng et al. (2020) evaluated the transferability of freeway SPFs between Chinese and the U.S. regions and their applicability for hotspot identification. The results revealed that after calibration the transferred SPFs between the two countries showed satisfactory performance on both model fit and hotspot identification.

    • Transfer learning-based highway crash risk evaluation considering manifold characteristics of traffic flow

      2022, Accident Analysis and Prevention
      Citation Excerpt :

      Nevertheless, the model transferability between Florida and Chinese cities is not expected. The reason may be that the model fails to identify traffic flow characteristics in Florida (Feng et al., 2020). Tang et al., believe that existing calibration methods require a large amount of data in the target area, ignoring the feasibility of data collection in real life.

    • Transfer learning for spatio-temporal transferability of real-time crash prediction models

      2022, Accident Analysis and Prevention
      Citation Excerpt :

      Regarding transferability in quantitative research, transferability referred to the ability of the model or the findings of one sample to be applied to another sample (Polit and Beck, 2010). For safety studies, a number of studies focused on transferability between safety performance functions (Feng et al., 2020; Venkataraman et al., 2016) and crash prediction models (Sawalha and Sayed, 2006). Yet, limited studies were conducted on the topic of model transferability in real-time crash prediction models.

    • Applicability of boosting techniques in calibrating safety performance functions for freeways

      2021, Accident Analysis and Prevention
      Citation Excerpt :

      The study of Li et al. (2018) has demonstrated that the crash frequency was more sensitive to the traffic volume in the U.S. more than it was in China. Also, as interpreted by Feng et al. (2020), the significant difference in traffic volume between Florida state and the Chinese cities, Shanghai and Suzhou, was a reason for the inferior transferability of SPFs. Generally, it can be concluded the superiority of the various boosting calibration techniques to transfer the crash prediction models using the limited available data in the international context.

    View all citing articles on Scopus

    This paper has been handled by associate editor Tony Sze

    View full text