Macro and micro models for zonal crash prediction with application in hot zones identification
Introduction
Crash prediction model (CPM) is an essential tool in traffic safety analysis. Numerous studies have been conducted to evaluate the safety level of various types of road entities, to identify hotspots or sites with promise, and to find appropriate countermeasures. Recently, an increasing research effort is being focused on a higher aggregated level of crash analysis, which could be referred to as zonal CPM. Traffic crashes are typically aggregated at a certain spatial scale and researchers usually seek to relate safety to zone-level covariates. These macro-level CPMs may aid transportation agencies in more proactively incorporating safety consideration into long term transportation planning process (Washington et al., 2006).
Last decade has witnessed fast growing scope of scientific research to investigate crash propensity on macroscopic levels. Different area-wide characteristics were considered, including road characteristics such as intersections density (Huang et al., 2010, Xu and Huang, 2015), road length with different speed limit (Abdel-Aty et al., 2011, Siddiqui et al., 2012), road length with different functional classification (Quddus, 2008, Hadayeghi et al., 2010), junctions and roundabouts (Quddus, 2008); traffic patterns such as traffic flow and vehicle speed (Quddus, 2008, Hadayeghi et al., 2010); trip generation and distribution (Abdel-Aty et al., 2011, Dong et al., 2014, Dong et al., 2015); environment conditions such as total precipitation/snowfall, and number of rainy/snowy days per year (Aguero-Valverde and Jovanis, 2006); land use (Pulugurha et al., 2013); and socioeconomic factors such as population density (Huang et al., 2010, Siddiqui et al., 2012), age cohorts (Aguero-Valverde and Jovanis, 2006, Dong et al., 2015, Hadayeghi et al., 2010), household incomes (Xu and Huang, 2015) and employment (Quddus, 2008, Hadayeghi et al., 2010).
A wide array of spatial units have been employed, such as regions (Washington et al., 1999), counties (Miaou et al., 2003, Aguero-Valverde and Jovanis, 2006, Huang et al., 2010, Li et al., 2013), districts (Haynes et al., 2007), wards (Quddus, 2008), zip codes (Girasek and Taylor, 2010, Lee et al., 2014a), census tracts (Ukkusuri et al., 2011, Ukkusuri et al., 2012, Wang and Kockelman, 2013), block groups (Levine et al., 1995), and traffic analysis zones (i.e. TAZs1; Hadayeghi et al., 2010, Abdel-Aty et al., 2011, Siddiqui et al., 2012, Pulugurha et al., 2013, Wang et al., 2013, Dong et al., 2014, Dong et al., 2015, Xu and Huang, 2015). Among them, TAZs are now the only traffic-related zone system and are superior in being easily integrated with the transportation planning process, thus having been widely adopted.
However, the efficiency of macro-level traffic safety analysis may be subject to the well-known modifiable areal unit problem (Abdel-Aty et al., 2013, Lee et al., 2014b, Xu et al., 2014) and boundary issue (Siddiqui and Abdel-Aty, 2012, Cui et al., 2015). From another perspective, the safety problem is anyhow a microscopic problem and the direct contributing factors could be related to micro-level factors for a specific road segment or intersection, or the driver-vehicle units involved. In additional to macro level CPMs, an alternatively potential solution estimating zonal safety situation is to sum up crash predictions of all entities (i.e. road segments and intersections) located within the zones of interest, which could be regarded as micro level CPMs.
As road entities located in close proximity may share confounding factors, spatial correlation (or spatial dependency) tends to be a major concern. Research demonstrated that the consideration of spatial effect of adjacent road segments in crash prediction contributes to an unbiased parameter estimation, and significantly improves model predictive performance (Wang et al., 2009, Aguero-Valverde and Jovanis, 2010, Xie et al., 2013). Nevertheless, previous studies are mostly limited to individual types of road entities, i.e. either intersection or segments. Undoubtedly, spatial correlation exists not only between adjacent road segments or between adjacent intersections, but also, even more importantly, between road segments and their connected intersections. To this end, Zeng and Huang (2014) proposed a Bayesian spatial joint approach to simultaneously model crash frequency of intersections and the feeding road segments. Results revealed that the spatial correlation between segments and the connected intersections are more significant than those solely between segments or between intersections.
To our knowledge, there is no research comparing macro-level and micro-level models in predicting zonal safety levels. A comparative analysis could be interesting and beneficial to reveal the associations and differences between those two methods, as well as to provide an explicit template towards the application of either technique appropriately.
This study intends to empirically compare two types of zonal CPMs by evaluating model fitting and predictive performance, as well as identifying crash hot zones. Two state-of-art methods, i.e. the macro level Bayesian spatial model with conditional autoregressive (CAR) prior and micro level Bayesian spatial joint model are developed and empirically evaluated, respectively. The analysis is based on an urban road network with 346 segments and 198 intersections of 155 TAZs in Hillsborough County, Florida, U.S.
Section snippets
Bayesian spatial model with CAR prior
Traditional CPMs such as Poisson lognormal model and negative binomial model have largely ignored the issue of possible spatial correlation of traffic crashes among adjacent zones, which would be misleading as this cannot reflect the true underlying data generating process (Huang and Abdel-Aty, 2010). For this reason, by incorporating an error term followed by the CAR prior into the link function, the Bayesian spatial model with CAR prior has been widely applied in current macro-level crash
Data preparation
Dataset were elaborately collected based on 155 TAZs in Hillsborough County, Florida, U.S. These TAZs constituted a road network of 346 segments and 198 intersections, as shown in Fig. 1.
The crash data in a three-year period (2005–2007) were obtained from the Crash Analysis Reporting (C.A.R.) system of Florida. Data for road and traffic characteristics, trip generation, and demographic and socioeconomic factors were derived from the Florida Department of Transportation roadway characteristics
Results and discussion
Before the final sets of variables were determined, a multi-collinearity test was conducted to ensure the inclusion of no highly correlated independent variables. Results indicated a high correlation between total population and number of minority population. Similarly, the total trip productions and total trip attractions; the proportion of population aged under 15 years old and the population between 16 and 64 years old; the roadway function class and presence of median were all highly
Conclusion
This study mainly compared the performance of two different types of zonal CPMs, namely macro-level CAR model and micro-level spatial joint model. Results revealed that the micro-level spatial joint model has a better overall fit and predictive performance, provides better insights about the microscopic factors (e.g. the geometric design of segments and intersections) that closely contribute to crashes, and leads to more direct countermeasures for improving traffic safety.
Meanwhile, the
Acknowledgements
This work was jointly supported by the Natural Science Foundation of China (Project No. 71371192), the Joint Research Scheme of National Natural Science Foundation of China/Research Grants Council of Hong Kong (Project No. 71561167001 & N_HKU707/15), the Research Fund for the Fok Ying Tong Education Foundation of Hong Kong (142005), and the Fundamental Research Funds for the Central Universities of Central South University.
References (53)
- et al.
Geographical unit based analysis in the context of transportation safety planning
Transp. Res. A
(2013) Multivariate spatial models of excess crash frequency at area level: case of Costa Rica
Accid. Anal. Prev.
(2013)- et al.
Spatial analysis of fatal and injury crashes in Pennsylvania
Accid. Anal. Prev.
(2006) - et al.
A framework of boundary collision data aggregation into neighbourhoods
Accid. Anal. Prev.
(2015) - et al.
Support vector machine in crash prediction at the level of traffic analysis zones: assessing the spatial proximity effects
Accid. Anal. Prev.
(2015) - et al.
Macroscopic hotspots identification: a Bayesian spatio-temporal interaction approach
Accid. Anal. Prev.
(2016) - et al.
Development of planning level transportation safety tools using geographically weighted Poisson regression
Accid. Anal. Prev.
(2010) - et al.
Multilevel data and Bayesian analysis in traffic safety
Accid. Anal. Prev.
(2010) - et al.
Predicting crash frequency using an optimized radial basis function neural network model
Transportmetrica A
(2016) - et al.
Analysis of residence characteristics of at-fault drivers in traffic crashes
Saf. Sci.
(2014)