category: literaturenote citekey: wangsystematicapproachmacrolevel2024 title: A systematic approach to macro-level safety assessment and contributing factors analysis considering traffic crashes and violations authors: "Wang, Xuesong; Zhang, Xueyu; Pei, Yingying" year: 2024 date: 2024-00-00 2024 doi: 10.1016/j.aap.2023.107323 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2023.107323" zotero_key: 7XG2ERRW zotero_storage: 54TXWUTP collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Wang, Xuesong"

status: converted

Contents lists available at ScienceDirect

Accident Analysis and Prevention

journal homepage: www.elsevier.com/locate/aap

A systematic approach to macro-level safety assessment and contributing factors analysis considering traffic crashes and violations

Xuesong Wang a,b,* , Xueyu Zhang a,b , Yingying Pei a,b

^a School of Transportation Engineering, Tongji University, Shanghai 201804, China
^b The Key Laboratory of Road and Traffic Engineering, Ministry of Education, Shanghai 201804, China

ARTICLE INFO

Keywords: Crash and violation Police enforcement Crash-prone and violation-prone identification Bivariate negative binomial spatial CAR model Potential for safety improvement Interpretable machine learning framework

ABSTRACT

During rapid urbanization and increase in motorization, it becomes particularly important to understand the relationships between traffic safety and risk factors in order to provide targeted improvements and policy recommendations. Violations and police enforcement are key variables, but the endogenous relationship between crashes and violations has made these variables unreliable and has limited their use. To manage this problem, this study developed a systematic approach for the joint modeling of crashes and violations to identify crash and violation hotspots and examine the mechanisms underlying macro-level contributing factors. Socio-economic, road network, public facility, traffic enforcement, and land use intensity data from 115 towns in Suzhou, China, were collected as independent variables. A bivariate negative binomial spatial conditional autoregressive model (BNB-CAR) and the potential for safety improvement (PSI) method were adopted to identify crash-prone and violation-prone areas, and an interpretable machine learning framework was applied to explore the factors' effects by area. Results showed that the proposed framework was able to accurately identify problem areas and quantify the impact of key factors, which, in Suzhou, were the number of traffic police and their daily patrol time. Considering such enforcement-related information provided important insights into reducing crash and violation frequency; for example, keeping the number of traffic police and daily patrol time under certain thresholds (number of police lower than 11 and patrol time lower than 2.3 h in this sample) was as effective as increasing these numbers for reducing the probability of high-crash and high-violation areas. The proposed approach can help traffic administrators identify the key contributing factors, especially enforcement factors, in crash-prone and violation-prone areas and provide guidelines for improvement.

1. Introduction

Booming urbanization and motorization in many cities have put enormous pressure on urban traffic operations, with the increase in traffic leading to frequent violations and crashes. When administrators designate safety improvements such as traffic enforcement to particular areas, taking active traffic safety management countermeasures that are generated by identifying crash-prone areas and analyzing contributing factors is an approach which is more reliable than relying solely on the experience of traffic police for optimizing the allocation of resources.

Many crash analysis models have taken the frequency of traffic violations as one of the explanatory variables (Zhang et al. 2013, Hos seinlou et al. 2018) and confirmed positive correlations between crashes and violations. Hong et al. (2019), however, found an endogenous relationship between violations and crashes and concluded that violations should therefore not be considered as an explanatory variable for crashes. Feng et al. (2020) also found underlying interactions between violations and crashes, observing that traffic enforcement factors such as the number of traffic police directly influenced violations and thus reduced the occurrence of crashes. That is, the problem with endogenous variables is that their internal influences complicate the interpretation of correlation and potential causality. For example, traffic and roadway factors can increase both crashes and violations, but increased enforcement may both increase reported violations and reduce crashes. The result of these inconsistencies is that traditional regional safety improvement models generally began to identify hazardous areas by crash frequency only, and neglect violation frequency as a potential influencing factor. There has also been some neglect of traffic violations as a dependent variable, and little research conducting joint analyses of traffic crashes and violations. Yet a two-dimensional

^* Corresponding author at: School of Transportation Engineering, Tongji University, Shanghai 201804, China. E-mail address: wangxs@tongji.edu.cn (X. Wang).

identification approach considering both crashes and violations is more capable of uncovering safety problems than the one-dimensional identification of crash-prone areas by crash frequency only. For example, if there are many crashes but few violations in a hazardous area, the low number of violations may suggest that the crash frequency may be due to either poor enforcement by police or poor engineering facilities. To accurately determine the problems of a region in order to optimize enforcement and other improvement strategies, it is necessary to consider both traffic crashes and violations together, as has been done in a few studies. Existing comprehensive analyses of traffic crashes and violations, however, have focused on the individual driver (Joo et al. 2022) or have been micro-level studies (Feng et al. 2020, Cheng et al., 2023), but there have been few regional, or macro-level, studies.

Macro-level safety analysis, compared to meso-level or micro-level safety analysis, examines the effects of zonal factors contributing to crashes or violations (Wang et al. 2016, Alver et al. 2014, Pei et al. 2021), such as socio-economic, road network, land use, traffic, and enforcement characteristics. In active safety management of a region, where traffic police must know which improvement countermeasures are most beneficial for which high-risk areas, it is essential to consider these zonal, or regional, characteristics in order to provide policy-level solutions. These regional factors influence both crashes and violations, but as enforcement can influence crashes and violations in opposing directions, omitting consideration of their endogenous relationship may produce biased parameter estimates and ineffective recommendations for safety improvements.

In a macro-level analysis, methods such as the potential for safety improvement (PSI) can be applied to identify hazardous areas, but most of the existing PSI studies only identify hazardous areas by crashes, not violations. If identification methods can take into account regional influencing factors such as traffic police enforcement information, both crash-prone and violation-prone areas can be used to guide targeted safety improvements. Traditional interpretation of risk factors in hazardous areas, however, relies on field research and manual causation analysis, which cannot address the complex mechanisms underlying the influence on crashes and violations of some of the key factors such as enforcement. Recently, machine learning (ML) methods have attracted increasing attention in the traffic safety research field (Wen et al. 2021, An et al. 2022). ML models can overcome traditional models' strict assumptions of differences between independent and dependent variables and are widely used for safety analysis modeling. The interpretable ML framework developed in recent years (Wen et al. 2021, Yang et al. 2021) has facilitated the handling of the ML models' black box problem, and offers a new data-driven approach to determining the factors contributing to hazardous areas and explaining their effects and influencing mechanisms. ML thus has the potential to enable modeling to consider the endogeneity of crashes and violations to avoid misinterpretation and

inaccurate estimation results. ML can provide both a theoretical basis and data support for safety improvement, but these models have not yet been used at the macro-level to regionally interpret the relationships between crashes and violations.

To produce a regionally categorized assessment of engineering and enforcement that can guide the proposal of targeted safety improvement countermeasures, this study proposes an approach that analyzes macrolevel contributing factors, with particular attention to the joint modeling of crashes and violations. The center panel of Fig. 1 shows the basic process of regional safety improvement, with the left and right panels comparing the traditional method with the proposed method, respectively. The traditional method is dominated by crash analysis only, modeling traffic crashes and identifying crash-prone areas. Traffic authorities then implement improvements in road facilities and traffic enforcement in these crash-prone areas. This method, however, makes it difficult to accurately identify which of the crash-prone areas have inadequate enforcement and which have inadequate road facilities, resulting in unclear analysis of the specific safety issues and limited targeting of improvements. The proposed method, on the other hand, identifies both crash-prone and violation-prone areas, enabling the comprehensive classification of areas with the potential for engineering and/or enforcement improvement. The first step in the proposed method is to determine appropriate analysis units and collect a variety of macro factors related to road networks and traffic enforcement. Second, to identify and explore these contributing factors' effects to provide focus for government and urban planning departments, a bivariate model is used instead of a univariate model to estimate crash and violation frequency data jointly rather than crash frequency only. Third, PSI values are obtained according to the modeling results and are employed to identify violation-prone areas as well as crash-prone areas and their distribution characteristics. Finally, to identify the key factors' influencing mechanisms, an interpretable ML model is adopted to analyze and assess the effects of the variables, particularly the effect of police enforcement on regional traffic safety. Based on the analysis results, pertinent improvement countermeasures are proposed in this paper for different area types (areas with different combinations of high and low crash and violation frequency) to provide guidance on traffic police enforcement.

2. Literature review

2.1. Macro-level contributing factors for crashes and violations

Previous studies have revealed that crashes and violations are related to many macro-level, or regional, contributing factors, which can be categorized into socio-economic, traffic, road network, land use, and police enforcement characteristics. Socio-economic indicators mainly

Fig. 1. Study Framework.

include age, gender, gross domestic product (GDP), and population density. For example, Zhang et al. (2013) found that male drivers were more likely to be involved in violations and fatal crashes than female drivers. Huang et al. (2018) showed that population density was positively associated with crash rates in most (about 80%) of their study's block groups.

The main road network characteristics are road network scale and pattern. Studies generally have shown that the greater the density of the road network and the longer the length of arterials in the region, the more crashes occur (Pei et al. 2021, Wang et al. 2019). Research has shown differing results on the effect of road type on crashes and violations. Alver et al. (2014) found that young drivers were more likely to violate traffic rules such as seat belt usage and speed limits on urban arterials than on rural roads, but a study by Mane and Pulugurtha (2018) showed that crash risk was higher on suburban arterials than on urban arterials.

Traffic and land use characteristics are also commonly assessed in macro-level safety studies. Wang et al. (2019) found that there was a significant positive correlation between trip generation and traffic crashes in a given region. Pei et al. (2021) found higher land use intensity showed a higher crash frequency and violation frequency. Traffic police enforcement factors, which also have important correlations with crash frequency, are discussed in the next section.

2.2. Correlations between crashes and violations

Yannis et al. (2008) discovered that increasing the number of patrols significantly reduced crashes caused by speeding and drunk driving. Holland and Conner (1996) found that an active police patrol effectively reduced the proportion of drivers exceeding the speed limit. Studies have shown that electronic policing devices have significant associations with violations and the risk of crashes (Wang et al. 2020, Martinez-Ruiz et al. 2019, Rothengatter 1982). Wang et al. (2020) showed that cameras recording red-light running and speeding reduced crash frequency. Similarly, the use of enforcement cameras were found by Martinez-Ruiz et al. (2019) to increase detected violations and decrease crash frequency. Other studies have used the number of traffic citations as a measure to examine the effect of enforcement on crashes. For example, a negative correlation between the number of citations and the number of crashes was found by Terrill et al. (2016) using ordinary least squares, suggesting that enforcement is a preventive measure for crashes.

Zhang et al. (2013) established a binary logit model to determine the risk factors correlated with traffic violations and crashes and found that violation behavior was one of the fundamental associations with crashes: as the frequency of traffic violations decreased, the frequency of injury and fatal crashes also decreased. Using a zero-truncated Poisson model, Hosseinlou et al. (2018) also found a positive correlation between violations and crashes. A study of drunk driving (Valen et al. 2019) found that most drunk drivers involved in fatal crashes were also engaged in violation behaviors such as speeding and neglecting seat belt use. These correlations, however, appear to contradict other findings mentioned above such as the negative correlation between the number of citations and the number of crashes found by Terrill et al. (2016). Researchers have thus warned against the use of traffic violation frequency as an explanatory variable due to the endogeneity problem (Hong et al. 2019, Feng et al. 2020) discussed in this paper's Introduction.

2.3. Methods for identifying crash-prone areas

Identification of crash-prone areas, or hotspots, provides a foundation for managing routine traffic safety (Wang et al. 2021). These areas were traditionally identified by crash frequency or density, but neither measure addressed the random fluctuation of crashes (Hauer 1997). The potential for safety improvement (PSI) method, proposed by Persaud et al. (1999), refers to the difference between the expected crash

frequency at a certain location and the average predicted crash frequency for similar locations. PSI has had ongoing widespread use in crash-prone area identification, having multiple advantages such as reducing the impact of random fluctuations and considering regional characteristics. As it provides the best reference for regional traffic safety management, this study used the PSI method to identify hazardous areas.

2.4. Application of machine learning methods in traffic safety

ML models, particularly models based on a decision tree, can recognize the complex relationships between independent variables and dependent variables and have high predictive accuracy. These models have gradually been adopted in traffic safety in addition to or instead of traditional statistical models. Rahman et al. (2019) were the first to develop decision tree regression macro-level traffic safety models for vulnerable road user crashes. Kus¸kapan et al. (2021) used ML to construct speed violation classification models for heavy vehicles on highways; they identified the highest speed violations provinces according to the naive Bayes classification algorithm and a density map of the speed violation rate. ML is especially useful for studies with multisource data. For example, An et al. (2022) examined non-linear effects on injurious traffic collisions using a gradient boosting decision tree (GBDT), and Chang et al. (2022) applied extreme gradient boosting (XGBoost) to explore the effects of built environment variables such as land use and road characteristics on fatal pedestrian accidents. Both studies concluded that ML methods were able to capture the nonlinearity of independent variables. Classification and regression tree (CART), adaptive boosting (AdaBoost), and light gradient boosting machine (LightGBM) (Wen et al. 2021, Yuan et al. 2022) are also widely used ML methods in current traffic safety modeling. The ML model results, however, can be difficult to interpret; that is, the model has a black box problem.

Proposed by Lundberg and Lee (n.d) to tackle the ML model's black box problem, the SHAP (SHapley Additive exPlanations) framework, based on cooperative game theory, has been utilized in many fields to interpret the output of ML models. SHAP values measure the marginal effects of individual contributing factors on predicted outcomes while controlling the average effects of other factors. An increasing number of traffic safety analyses employ SHAP. Yang et al. (2021) used XGBoost and SHAP to explore the built environment factors related to freight truck crashes. Wen et al. (2021) quantified and compared the impact of key risk factors on various crash types using LightGBM and SHAP. Yuan et al. (2022) compared multiple ML models to obtain a reliable and practical real-time safety model; finding that the random forest method achieved the best performance, the researchers added the SHAP framework to improve the model's interpretability. These studies suggest the SHAP technique can effectively enhance the ML model's transparency and interpretability.

3. Data preparation

3.1. Analysis units

Suzhou is one of China's megacities. It covers 8657.32 km2 , has a population of 12,848 million persons, and is experiencing increasingly rapid urbanization and a resultant increase in motorization. Over the last few years, the city's traffic safety level has increasingly become the main factor affecting and even restricting the development of the city. It has become urgent to assess traffic safety by analyzing crash contributing factors at the macro level to improve the active safety management of the city's urban traffic.

Like many large cities in China, Suzhou consists of multiple towns, which are the basic administrative districts in which Chinese traffic police conduct routine safety management activities. As shown in Fig. 2, Suzhou contains 115 towns. The towns' boundaries are consistent with

Fig. 2. Town Boundaries in Suzhou.

the jurisdictional boundaries of surface road management. Freeways and urban elevated expressways are excluded from the road networks as they are separately managed. The town is used as the analysis unit in the current study, providing a sample size that is comparable to other macro traffic safety studies (Kamel et al. 2019, Guo et al. 2018).

3.2. Description of dependent variables

Crash frequency and violation frequency were set as dependent variables. Crash and violation data from January 1 to June 30, 2020, were derived from the Suzhou Road Traffic Crash Analysis and Warning System. The system responds to statistical query and provides a spatial display function to access crash, violation, and police data. The crashes are divided into three categories, depending on crash severity, according to road traffic accident handling procedures in China: general processing procedure crashes, simple processing procedure crashes, and quick processing procedure crashes. General processing procedure crashes are serious injury or fatal crashes, simple processing procedure crashes are property damage or minor injury crashes, and quick processing procedure crashes are property damage only crashes. The three types of crashes differ in the structure of record tables and the quantity of crash information. This study used only simple and general processing procedure crash data.

The violation data were derived from the same Suzhou Road Traffic Crash Analysis and Warning System, and included violations detected by either enforcement cameras or police officers. All violations were considered in this study except crash violation. Examples of included violations were speeding, illegal turns, not stopping or yielding where

required, careless driving, drunk driving, and illegal parking.

3.3. Description of independent variables

The independent variables in this study were divided into five categories of characteristics: socio-economic, road network, public facility, traffic enforcement, and land use intensity. The population and regional GDP data were collected from the 2020 statistical yearbooks of each administrative district. Road network features, including the total length of major arterials, the total length of minor arterials, road density, and intersection density, were obtained manually based on a geographic information system (GIS) base map. Because the collection of regional traffic operation data was limited for some analysis units, public facility data, which included the density of restaurants, parks, and schools, was used as an alternative measure to reflect the mobility of people within the area (Jia et al. 2018). Enforcement information, provided by the Suzhou Traffic Police Department, included the number of traffic police officers, the daily patrol time of police officers, and the proportion of electronic policing. Land use data were obtained from the Google Earth 2019 historical image map. In line with previous research (Pei et al. 2021, Wang et al. 2019), land use intensity was classified into three levels according to the floor-area ratio (proportion of a building's total floor area to the area of the land it rests on): low (*<25%), medium (25% ~65%), and high (>*65%).

SPSS 22.0® software was used to check for multi-collinearity. Results showed that all variance inflation factors (VIF) of the variables were lower than 3, indicating they could be used for modeling. Statistical descriptions of the independent variables are summarized in Table 1.

Table 1 Statistical descriptions of independent variables.

Variables		Min	Max	Mean	Std. Dev.
Socio-economic characteristics	Population density (104 persons per km2	0.027	2.488	0.387	0.504
	) GDP (104 yuan per person)	1.991	66.667	19.961	13.692
Road network features	Total length of major arterials	0.000	66.133	17.160	15.440
	(km) Total length of minor arterials (km)	0.006	73.669	20.873	22.303
	Road density (km/ km2 )	0.582	14.735	3.154	2.508
	Intersection density (count/ km2 )	0.261	52.980	4.138	5.906
Public facility data	Restaurant density (count/km2 )	0.000	3.232	0.578	0.838
	Park density (count/km2 )	0.000	1.776	0.213	0.360
	School density (count/km2 )	0.000	2.454	0.304	0.477
Traffic enforcement	Daily patrol time of police officers (h)	0.000	9.679	1.731	2.020
information	Proportion of electronic policing	0.013	0.639	0.314	0.145
	enforcement Number of traffic police officers (count)	1	35	11	4.721
Variables		Number of		Percentage
Land use intensity	0: low 1: medium 2: high	observations 22 42 51		19.13% 36.52% 44.35%

4. Methodology

4.1. Crash and violation frequency joint modeling

Unlike existing macro-level safety analysis models (Wang et al. 2019, Lee et al. 2014), this study considered both crashes and violations in regional safety assessment. The negative binomial (NB) model was selected as the basic statistical modeling method to deal with the overdispersion problem in the crash data. To handle the endogeneity problem, the correlations and interactions between crashes and violations should be considered when modeling. Correlations among analysis units are also inevitable, producing unobserved spatial effects. To address the spatial correlations, this study adopted a conditional autoregressive (CAR) model to capture the unobserved spatially structured effects. Then, a bivariate model was adopted instead of a univariate model to estimate crash and violation frequency data jointly, to deal with the endogeneity problem, rather than crash frequency only. The resultant BNB-CAR model was defined as follows:

$$y{i1} Negbin(\theta{i1}, r_1)$$ (1)

$$y{i2} Negbin(\theta{i2}, r_2)$$ (2)

$$log\theta{i1} = \beta{01} + X_{i1}\beta1 + \varepsilon{i1} + \varphi_i$$ (3)

$$log\theta{i2} = \beta{02} + X_{i2}\beta2 + \varepsilon{i2} + \varphi_i \tag{4}$$

$$(\boldsymbol{\varepsilon}_1, \boldsymbol{\varepsilon}_2) \ \mathrm{N}{(0,0), (\sigma_1^2, \rho \sigma_1 \sigma_2, \sigma_2^2)}$$ (5)

where *yi*1 and *yi*2 are crash and violation counts of town i, respectively; *θi*1 and *θi*2 are expected crashes and expected violations, respectively; *Xi*1 and *Xi*2 are vectors of the independent variables; *β*1 and *β*2 are vectors of regression coefficients; *ε*1 and *ε*2 are measures of the correlations between crashes and violations; *φi*is the CAR effect term; and ρ is the parameter that measures the correlation effect between crashes and violations. The conditional distribution of *φi*and φ⁻ *ⁱ*follows a normal distribution with mean ^∑ wj,ⁱ wi⁺ *φj*and variance ¹ τcwi⁺ .

$$\varphii | \varphi{-i} N(\sum{j} \frac{w{j,i}}{w_{i+}} \varphi_j, \frac{1}{\tauc w{i+}})$$ (6)

where φ⁻ *ⁱ*is the set of φi, wi⁺ ⁼ ^∑^N ⁱ=1*wi,j*refers to the sum of weights in all towns near town i, j is the set of towns adjacent to town i, and *τc*is the quality coefficient.

The reciprocals of distance between towns were used in the BNB-CAR model. An example of the weight matrix is shown:

$$w{i,j} = \begin{bmatrix} 1/d{i,j} & i \neq j \ 0 & i = j \end{bmatrix} \tag{7}$$

where *di,j*is the distance between the centroids of town i and town j.

4.2. Crash-prone and violation-prone area identification

PSI was utilized as the performance measure to identify regional safety levels due to its ability to ameliorate random fluctuations in crash and violation counts. PSI, in this study, is the difference between the expected crash (or violation) frequency at a certain location and the predicted crash (or violation) frequency for similar locations. The expected crash and violation frequencies were estimated by the BNB-CAR model, which considered the problem of regression-to-the-mean (RTM) in crash and violation occurrence. PSI was calculated by the following equation:

$$PSI = N{expected} - N{predicted}$$

$$= exp(\beta{01} + X{i1}\beta1 + \varepsilon{i1} + \varphii) - exp(\beta{01} + X_{i1}\beta_1)$$

$$= exp(\beta{01} + X{i1}\beta1)(exp(\varepsilon{i1} + \varphi_i) - 1)$$ (8)

The essence of safety assessment is to measure the number of standard deviations from the mean. In this study, combining the distributions of PSI values for crashes and violations, towns with a PSI value for crashes (or violations) equal to or more than one standard deviation above the mean were classified as high-crash (or high-violation) towns, or areas, while others were low-crash (or low-violation) areas. The strategy of analyzing crash results first and violation results second was adopted to classify the towns into four types: high-crash and highviolation, high-crash and low-violation, low-crash and high-violation, and low-crash and low-violation.

4.3. Influencing mechanism analysis of regional contributing factors

LightGBM is a boosting framework proposed by Ke et al. (n.d) based on the GBDT model. GBDT is improved in the LightGBM by using a histogram-based algorithm, a leaf-wise growth strategy with depth limitation, a gradient-based one-side sampling (GOSS) algorithm, and an exclusive feature bundling (EFB) algorithm. These algorithms significantly speed up model training, effectively reduce system memory consumption, and improve model prediction.

In addition, GBDT typically performs well on small datasets (Ding et al. 2019, Shao et al. 2020). LightGBM, based on GBDT, cycles through samples during training iterations to approximate the true value and enhance the efficiency of sample utilization, which is not sensitive to

sample size. LightGBM can also recognize the complex relationships between independent variables and dependent variables. To investigate the mechanisms underlying the contributing factors in the different crash-prone and violation-prone areas, LightGBM was therefore used to construct a classification model. SHAP was then introduced to identify contributing factors and quantify their effects. SHAP (Lundberg and Lee, n.d), which can handle the relationships between expected and predicted values, can improve the interpretability of the ML model's prediction results. The SHAP framework calculates SHAP values for any feature in a sample that corresponds to the predicted value. A positive or negative SHAP value indicates the feature's positive or negative influence on the prediction results, and its magnitude indicates the feature's degree of influence.

Mathematically, the main idea of the SHAP framework is to fit the trainer F(x) for interpretation with an additive model G(x), as follows:

$$F(x) = G(x) = \phi{base} + \sum{i=1}^{n} \phi_{i}$$ (9)

where ϕbase is the model's prediction base for the sample; *ϕi*is the SHAP value of feature i in the sample x, calculated as follows:

$$\begin{cases} \phii = \sum{S} \frac{|S|!(n-|S|-1)}{n!} (F(S \cup {x_i}) - F(S)) \ S \subseteq {x_1, x_2, \dots, x_n} / x_i \end{cases}$$ (10)

where {*x1, x2,* ⋯, xn} is the set of all input features, and S is the subset of features that do not contain the feature xi. Based on Eq. (10), the marginal contributions of all features can be calculated to explain the prediction results of the ML model.

5. Results

5.1. Results of BNB-CAR model

The BNB-CAR model was estimated in WinBUGS® using the Markov Chain Monte Carlo (MCMC) algorithm. Two MCMC chains of 50,000 iterations were generated, with the first 5,000 samples discarded for numerical instability.

The parameter estimation results are shown in Table 2. The correlation coefficient of the BNB-CAR model was 0.955, indicating a significant correlation between crashes and violations.

Population density and GDP were positively correlated with crashes and violations, which is consistent with most macro-level safety studies (Pei et al. 2021, Lee et al. 2014). This is reasonable because heavily populated and economically developed towns are more likely to generate trips, which increases the risk of crashes and violations.

The road network features of total length of major and minor arterials, road density, and intersection density were also positively associated with crashes and violations. These results correspond to most studies (Wang et al. 2021, Wang et al. 2015), which have found that the total length of major and minor arterials significantly affects traffic safety. High traffic volume and speed on major arterials increase the interaction of vehicles, which negatively impacts traffic safety; on minor arterials, the interaction and conflicts between motor vehicles and pedestrians and non-motorized vehicles increase exposure to crashes and violations (Pei et al. 2021, Wang et al. 2019). Road density and intersection density were positively correlated in this study with crashes and violations, similar to the findings of Wang et al. (2020). Denser road networks tend to increase more turning traffic, leading to higher crash probability.

Restaurant density had a positive correlation with crashes and violations. This finding is consistent with the assumption in this study that higher restaurant density tends to generate and attract more trips. Additionally, the likelihood of illegal parking and drunk driving near restaurants may be higher than in other areas, resulting in more crashes

Table 2 BNB-CAR model results for crashes and violations.

| Variables | Crashes | | | | | | |------------------------------------------------|------------|-------------|-----------|------------|--|--| | | Mean | Std.
Dev | 2.50% BCI | 97.50% BCI | | | | Intercept | 5.760** | 0.189 | 5.374 | 6.111 | | | | Population density | 0.038* | 0.094 | − 0.056 | 0.104 | | | | GDP | 0.0005* | 0.002 | − 0.001 | 0.005 | | | | Total length of major
arterials | 0.014** | 0.003 | 0.009 | 0.020 | | | | Total length of minor
arterials | 0.011** | 0.002 | 0.007 | 0.015 | | | | Road density | 0.020* | 0.024 | − 0.025 | 0.069 | | | | Intersection density | 0.015* | 0.012 | − 0.008 | 0.038 | | | | Restaurant density | 0.061** | 0.065 | 0.004 | 0.148 | | | | Park density | 0.026 | 0.160 | − 0.279 | 0.339 | | | | School density | − 0.134 | 0.0088 | − 0.300 | 0.042 | | | | Daily patrol time of police
officers | − 0.051** | 0.019 | − 0.090 | − 0.014 | | | | Proportion of electronic
police enforcement | 0.610** | 0.296 | 0.044 | 1.205 | | | | Number of traffic police
officers | 0.066** | 0.009 | 0.048 | 0.085 | | | | Land use intensity | | | | | | | | medium vs. low | 0.426** | 0.092 | 0.239 | 0.602 | | | | high vs. low | 0.584** | 0.091 | 0.402 | 0.760 | | | | Variables | Violations | | | | | | | | Mean | Std.
Dev | 2.50% BCI | 97.50% BCI | | | | Intercept | 8.481** | 0.171 | 8.153 | 8.834 | | | | Population density | 0.076* | 0.085 | − 0.094 | 0.241 | | | | GDP | 0.0006* | 0.002 | − 0.003 | 0.005 | | | | Total length of major
arterials | 0.008** | 0.002 | 0.003 | 0.017 | | | | Total length of minor
arterials | 0.002* | 0.002 | − 0.0009 | 0.006 | | | | Road density | 0.068* | 0.020 | 0.028 | 0.110 | | | | Intersection density | 0.027** | 0.010 | 0.008 | 0.047 | | | | Restaurant density | 0.100* | 0.055 | − 0.005 | 0.209 | | | | Park density | 0.262 | 0.138 | − 0.008 | 0.535 | | | | School density | − 0.053* | 0.079 | − 0.206 | 0.102 | | | | Daily patrol time of police
officers | − 0.013* | 0.018 | − 0.048 | 0.021 | | | | Proportion of electronic
police enforcement | − 0.168* | 0.273 | − 0.361 | 0.072 | | | | Number of traffic police
officers | 0.052** | 0.008 | 0.036 | 0.069 | | | | Land use intensity | | | | | | | | medium vs. low | 0.313** | 0.081 | 0.153 | 0.470 | | | | high vs. low | 0.441** | 0.075 | 0.295 | 0.584 | | | | Correlation coefficient | 0.955 | | | | | | | CAR effects | 0.233 | 0.077 | | | | |

Notes: Mean values noted by ** are significant at the 0.05 level; mean values noted by * are significant at the 0.1 level; BCI stands for Bayesian credible interval.

and violations. In contrast, school density had a negative correlation with traffic violations. This association may be caused by the fact that the area around schools is a preferred spot for traffic police patrols. Drivers usually regulate their driving behavior in areas where police cars are present, resulting in a lower violation frequency (Pei et al. 2021).

The proportion of electronic enforcement had opposite effects on crashes and violations: in contrast to the positive relationship between crash frequency and electronic enforcement, there was a negative correlation between violation frequency and electronic enforcement. Results suggest that while increasing the proportion of electronic enforcement reduces violations, it may not reduce the risk of crashes. The negative coefficient of police officers' daily patrol time for both dependent variables implies that increasing patrol time helps reduce crash and violation frequency, which is consistent with Feng et al. (2020). The number of police officers, however, was positively associated with both crashes and violations. In theory, the increase in the

number of police should increase the frequency of citations, in turn reducing the risk of accidents, which is supported by Terrill et al. (2016), who found a negative correlation between the number of citations and the number of crashes. This counterintuitive effect will be discussed in Section 5.3.3.

The frequency of crashes and violations was higher in areas with medium and high land use intensity compared to low land use intensity. This accords with the results of Mane and Pulugurtha (2018), who conclude that a probable explanation is that areas with high land use intensity usually have a higher concentration of people, and the increased commercial activity in the surrounding area causes an increase in traffic flow, which increases exposure to crashes and violations.

5.2. Identification results

The PSI values for crashes and violations in the 115 towns were calculated based on the BNB-CAR model. The results are shown in Fig. 3. The number of high-crash and low-violation towns is 18 (orange dots in the lower right quadrant), accounting for 15.7% of the towns. Because violations can only be known by citations, these areas may have insufficient enforcement efforts or poor coordination between the times and locations of enforcement and the times and locations of crashes and violations. Conversely, there are 11 low-crash and high-violation towns (yellow dots in the upper left quadrant), accounting for 9.6% of the towns. These areas may have excessive enforcement and wasted police resources. In 86 towns (green and red dots, indicating low-crash and low-violation towns and high-crash and high-violation towns, respectively), the level of crashes and violations is roughly the same, indicating that the level of traffic enforcement is compatible with the safety status.

Low-crash and low-violation (green) areas are mainly in the city's central towns of Gusu, Wuzhong, Huqiu, Xiangcheng, and Suzhou Industrial Park Districts, and peripheral towns of the city. Towns in the three problem categories are mainly adjacent to each other, located mainly in and near Wujiang, Kunshan, Changshu, and Zhangjiagang, showing a spatial clustering tendency. Towns near high-crash and highviolation towns tend to suffer higher safety risk than those away from the hazardous towns, indicating the spatial correlations among towns. High-crash and low-violation towns are spatially aggregated in the westcentral Kunshan section of Huning Corridor.

Compared to traditional hot zone identification methods generated by crash frequency, the bivariate-generated PSIs identify the zones precisely, and thus more usefully, for their classification based on combinations of crash and violation frequency. Using the results of this

kind of comprehensive analysis, traffic safety management departments can more assuredly identify each area's safety issues as primarily due to either enforcement or infrastructure problems, and thereby propose more targeted and effective safety improvement countermeasures.

5.3. Influencing mechanisms of crash-prone and violation-prone areas

5.3.1. Evaluation of prediction accuracy results

A confusion matrix of the crash-prone and violation-prone towns, shown in Fig. 4, was used to determine the LightGBM models' prediction accuracy. The darker elements on the diagonal of the matrix represent the number of correctly classified towns for each of the true label crashviolation classifications, or area types, such as low-crash and lowviolation; the remaining elements are the number of incorrectly classified towns. The numbers within each element indicate the number of towns in that group, which is also reflected in each element's color: the numbers beside the colored bar on the right indicate the number of towns described by that color. The results show that only one town is misclassified, suggesting that the crash-prone and violation-prone models based on the LightGBM algorithm have accurate prediction capability.

5.3.2. Importance of contributing factors

Fig. 5 depicts the importance of each of the 13 contributing factors in the four area types according to the interpretable ML model. The contributing factors were ranked in descending order based on their mean absolute SHAP values, that is, their average impact on the magnitude of model output. Of all factors, the number of traffic police officers had the greatest contribution to predicting crash-prone and violation-prone areas. This suggests that the number of police officers plays an important role in these areas, which is plausible for at least two reasons. First, the number of police contributes to reporting more on-site traffic violations. Second, dense police presence serves as a warning, resulting in reducing crash risk. Officers' daily patrol time was the second most important independent variable, indicating that both crashprone and violation-prone areas should pay attention to patrol time. It is worth noting that park density, which was used as a measure of people's mobility (28), was the third most important independent variable. The total length of major arterials also showed important associations with traffic safety.

Table 3 presents the relative importance ranking of the 13

Fig. 3. PSIs for Crashes and Violations. Fig. 4. Confusion Matrix for LightGBM Crash and Violation models.

Fig. 5. Feature Importance Based on Mean Absolute SHAP Values.

Table 3 Ranking of Feature Importance Based on Mean Absolute SHAP Value.

Variable	High-crash and high violation	High-crash and low violation	Low-crash and high violation	Low-crash and low violation
Population density	2	7	2	8
GDP	5	3	6	9
Total length of major arterials	7	5	8	3
Total length of minor arterials	11	4	9	5
Road density	3	6	7	7
Intersection density	10	11	13	12
Restaurant density	6	10	12	11
Park density	9	8	1	4
School density	12	12	5	10
Daily patrol time of police officers	4	2	10	2
Proportion of electronic police enforcement	8	9	3	6
Number of traffic police officers	1	1	4	1
Land use intensity	13	13	11	13

contributing factors according to mean absolute SHAP values, separated by crash-violation classification area type. Among the characteristics, population density ranked higher in high-violation areas than in lowviolation areas (ranks of 2 & 2 vs. 7 & 8), whereas GDP's importance was more related to crashes: GDP was more important in high-crash areas than low-crash areas (ranks of 5 & 3 vs. 6 & 9). The total length of major and minor arterials ranked higher for low-violation areas (5 & 3 vs. 7 & 8 and 4 & 5 vs. 11 & 9, respectively). It was notable that the relative ranking of road density for high-crash areas was higher than for low-crash areas (3 & 6 vs. 7 & 7). Daily police patrol time had a higher relative ranking for low-violation areas (2 & 2 vs. 4 & 10), showing the importance of this factor in low-violation areas. With the exception of low-crash and high-violation areas, the number of traffic police officers ranked as number 1.

High-crash and low-violation areas were selected for further analysis. The traffic enforcement and road network features (specifically, the mean number of traffic police officers and daily patrol time, and the

mean total length of major and minor arterials) ranked higher importance for high-crash and low-violation areas in Table 3. Fig. 6 shows the mean number of traffic police officers and daily patrol time in the four area types. It can be observed that high-crash and low-violation areas had the highest number of police officers (15.56), but the lowest daily patrol time (0.32 h). The inadequate allocation of police resources resulted in the issuing of an insufficient number of on-site traffic violations, which ultimately led to the areas being crash-prone. Traffic police departments in these areas should consider adjusting police resources to increase the intensity of patrol time. Fig. 7 illustrates the mean total length of major and minor arterials. High-crash and low-violation areas had the highest lengths of both major and minor arterials, increasing the interaction of traffic participants, which negatively impacted traffic safety. Therefore, there is a need to control the scale of the arterial road network to reduce the crash risk.

5.3.3. Directional impact of contributing factors

The feature importance plot in Fig. 5 qualitatively describes the relative importance among contributing factors in the four area types, but to specifically quantify how contributing factors affect the crashprone and violation-prone areas and in which direction, SHAP provides partial dependence plots (PDPs), which illustrate both the impact of individual factors and the impact of interaction between two factors on the dependent variables.

Fig. 8 and Fig. 9 show the SHAP PDPs of the top two factors (per Fig. 5) in predicting crash-prone and violation-prone areas. The horizontal axis is the value of the primary contributing factor, which is the number of officers in Fig. 8 and daily patrol time in Fig. 9. The left vertical axis is the contributing factor's SHAP value, which represents the degree and direction of impact on crashes and violations. That is, in addition to showing the degree of contribution by low to high SHAP values, a dot with a positive SHAP value indicates a positive impact on the areas and a negative value indicates a negative impact. The color of the dots shows the value of an additional contributing factor, described in the right vertical bar. The additional factor in Fig. 8 is daily patrol time, and in Fig. 9 it is electronic enforcement. The color indicates the magnitude of the value, with red and blue dots indicating higher and lower values, respectively. The interaction of the two factors is thus illustrated: the placement of the dots accords with the horizontal and left vertical axis, and their color accords with the right vertical bar.

The placement of the dots in Fig. 8(a and b) show that in crash-prone areas there was an obvious increasing trend for SHAP values as the number of traffic police officers increased, regardless of the corresponding daily patrol time (indicated by red and blue colors). In

Fig. 6. Comparison of the mean number of traffic police officers and daily patrol time in the four area types.

Fig. 7. Comparison of the mean total length of major and minor arterials in the four area types.

predicting high-crash and high-violation areas (9a), when the number of traffic police was under 11, SHAP values were negative; that is, the police presence appeared to reduce crashes. Then, when the number of traffic police rose above 11, SHAP values became positive, which indicates that the number of traffic police had a positive impact on highcrash and high-violation areas; that is, crashes and violations increased when the number of police increased beyond 11. This finding explains the overall positive correlations found between the number of traffic police and crashes and violations (Section 5.1): the SHAP analysis suggests that on-site enforcement by traffic police can reduce the probability of crash and violation, but this influence of the number of traffic police levels out after a certain range (in Suzhou, according to this study, it was less than 11). When the number exceeds this range, the probability of violations may increase and may waste resources. The patterns in Fig. 8(b) are comparable to those in Fig. 8(a): SHAP values were negative when the number of traffic police was below 10, but the impact became positive when the number of police was above 10. One possible explanation for the patterns in both figures is that dense police coverage has a deterrent effect, but only up to a certain threshold.

Fig. 8(c and d) show that the thresholds for the number of traffic police were 13 and 10 in predicting low-crash and high-violation areas and low-crash and low-violation areas, respectively; but beyond the thresholds, police influence shifted from positive to negative. That is, whereas in the high-crash areas shown in 9 (a) and (b), more officers beyond the thresholds correlated with more violations and crashes, in the low-crash areas of 9 (c) and (d), more officers beyond the thresholds correlated with fewer violations and crashes. The red dots' predominant location on the left side of all four graphs in Fig. 8(a – d) shows that when the number of police was small, the daily patrol time was higher. Conversely, when the number of police was large, the daily patrol time was lower. This finding indicates daily management should pay attention to the correlation between the number of traffic police and the patrol time. Specifically, when the number of traffic police reaches a threshold, administrative departments can consider optimizing patrol time and law enforcement structure in traffic safety management, rather than simply increasing the number of police.

As shown in Fig. 9(a), regardless of the proportion of electronic police enforcement (right-side vertical bar), an increasing and then

Fig. 8. SHAP Partial Dependence Plots of the Number of Traffic Police Officers.

stabilizing trend of SHAP values was found for the daily patrol time of police officers in predicting high-crash and high-violation areas. Patrol times less than 0.9 h had a negative influence on crashes and violations. SHAP values sharply increased to 0, that is, the factor began to have little impact, once it reached 0.9 h, whereupon its influence fluctuated around 0 between 0.9 and 2.3 h of patrol time. Then when the time was above 2.3 h, SHAP values became positive. This finding suggests that the daily patrol time, when it is below the threshold (2.3 h), can curb the probability of an area being high-crash and high-violation, but additional patrol time beyond the threshold is a waste of police resources. On the other hand, Fig. 9(b) shows a decreasing and then stabilizing trend of SHAP values for daily patrol time in predicting high-crash and lowviolation areas. SHAP values were positive, that is, there were more crashes, when the daily patrol time was below 0.6 h, but the factor's impact became negative and stable when the daily patrol time passed the 0.6 h threshold, suggesting that the lower threshold for the daily patrol time should also be considered. Fig. 9(c) shows that the daily patrol time had a relatively stable effect on low-crash and high-violation areas except when the patrol time was low. This implies that the daily patrol time plays a small role in low-crash and high-violation areas, which is consistent with Table 3 (10th ranking). SHAP values also show an increasing and stabilizing trend for daily patrol time in predicting low-crash and low-violation areas in Fig. 9(d). SHAP values were negative when the daily patrol time was below 0.4 h, after which the impact became positive and stable. This result may be explained by the possibility that the low patrol time may not be enough to improve safety, but when the patrol time increases, the effect on safety improvement is significant. Taking into consideration the factor's interaction with the proportion of electronic police enforcement (right bar), the considerable number of red points concentrated in low patrol time in all four graphs indicates that when police resources are not adequately allocated, the risk of crashes and violations can be reduced through electronic

enforcement.

6. Summary and conclusions

Because endogeneity complicates the role of traffic violations as a crash contributing factor, this study proposed a regional safety analytical approach focusing on enforcement-related factors. A bivariate model was developed and the potential for safety improvement (PSI) method was applied to identify crash-prone and violation-prone areas, after which the effects of the contributing factors were explored.

The 115 towns in Suzhou, China, were selected as units of analysis in order to conform with the jurisdictional boundaries of routine traffic safety management. Five categories of independent variables (socioeconomic characteristics, road network features, public facility data, traffic enforcement information, and land use intensity) and two dependent variables (crashes and violations) were collected for the 115 towns.

To deal with the endogeneity between crashes and violations and their spatial correlation among towns, a BNB-CAR model was adopted to jointly model crashes and violations. The results showed that there was a significant correlation between crashes and violations. Most influencing factors had a positive correlation with both crashes and violations, but the daily patrol time of police officers had a negative correlation with both crashes and violations, which is consistent with Feng et al. (2020). The proportion of electronic police enforcement had opposite effects on crashes and violations: electronic policing correlated with an increase in crashes and decrease in violations.

To identify crash-prone and violation-prone areas, the PSIs for crashes and violations in the 115 towns were calculated based on the BNB-CAR model. The bivariate PSI identification, in contrast to traditional PSI hot zone identification by crash frequency, were able to precisely classify the areas into four types based on both dependent

Fig. 9. SHAP Partial Dependence Plots of the Daily Patrol Time of Police Officers.

variables: high-crash and high-violation, high-crash and-low violation, low-crash and high violation, and low-crash and low violation. The spatial distributions for the four area types were analyzed separately.

LightGBM and SHAP were employed to quantify, compare, and interpret the influencing mechanisms of the primary contributing factors in crash-prone and violation-prone areas. The results suggested the number of traffic police officers and their daily patrol time had the greatest contribution to predicting crash-prone and violation-prone areas. The SHAP analysis showed that the importance of the contributing factors differed, however, depending on whether the area was high-crash and high-violation, high-crash and-low violation, or lowcrash and high violation. These differences led us to recommend countermeasures. For example, keeping the number of traffic police and daily patrol time under certain thresholds (number of police lower than 11 and patrol time lower than 2.3 h in this sample) was enough to potentially reduce the probability of high-crash and high-violation areas while not wasting police resources. Previous studies (Pei et al. 2021, Holland and Conner 1996) have found that the number of traffic police can reduce crashes, but the approach taken in this study enabled it to demonstrate that the number of police can reduce crashes only within a certain range. This study also provided some insight into ways of handling the endogenous relationship between crashes and violations: by modeling them together, the important factor of traffic violations can be used more reliably to predict crashes.

The proposed approach, which involves a systematic analysis considering both traffic crashes and violations for regional safety, can be used to identify hazardous areas and key contributing factors of traffic crashes and violations, enabling a categorized regional assessment of engineering and enforcement as well as a reference for targeting new countermeasures. A limitation of this study for real-world application is the determination of appropriate analysis units. Spatial analysis faces the modifiable areal unit problem (MAUP), which is a statistical bias that significantly affects the variation of results between observed variables due to changes in aggregation levels of spatial data (scale effect) as well as changes in configuration, shape, and arrangement of the aggregation unit (zone effect). The town unit was selected as the analysis unit in this study because conclusions drawn from the town can be readily applied into Chinese traffic police routine management. Nevertheless, it is possible that the evaluation results and key factors obtained from the different analysis units may vary. For practical application, it is necessary to compare the results for multiple analysis units to obtain more reliable conclusions, and for future study, to delineate more appropriate analysis units.

In addition, while the conclusions of the existing analysis can provide macro-level policy guidance for government departments. It is still necessary to combine sub-region crash and violation characteristic analysis and field investigation to find out the main focus such as road safety and traffic enforcement for the improvement of specific problem areas. The field investigation can be carried out in detail from several perspectives: road network, road connections, land use, and public facilities distribution. Only then can the actual causes of regional safety problems be identified and pertinent countermeasures can be proposed to improve different crash-prone and violation-prone areas. The optimization of enforcement strategies cannot, of course, be considered only at the macro level, but the actual police numbers in each jurisdiction must be considered in order to reasonably allocate police resources.

This study has introduced a viable method for handling the endogenous relationship between crashes and violations, but several possible extensions to this research remain for future analysis. One recommended extension is that the types of crashes can be subdivided, such as crashes with other motor vehicles and crashes with non-motor vehicles. Other extensions include exploring the interactions between factors not considered in this study, such as the interaction between violations and road type, alignments, and conditions such as weather. Different types of

violations should also be considered separately, and more attention should be given to type of enforcement, that is, contrasting the effects of police presence and electronic enforcement. Also, this study's sample of 115 towns is moderate for a macro traffic safety study; for example, the sample size in (Kamel et al. 2019) and (Guo et al. 2018) is 134 traffic analysis zones. It would be beneficial, therefore, to test this method with a larger sample size. To generalize the method to other cities, due to variation in size, culture, and other characteristics, the key factors contributing to traffic crashes and violations would need to be deeply researched and scaled because different cities may have different regional and traffic characteristics, as well as different traffic management patterns and practices. Researchers in the field will be interested in discovering whether data from other countries and regions confirms the results of this study.

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.

Data availability

The authors do not have permission to share data.

Acknowledgment

This study was sponsored by the Chinese National Science Foundation (52372335). The authors are grateful to Barbara Rau Kyle for her helpful edit.

References

crash records. Accident Analysis and Prevention 176, 106790.

Ke, G.L., Meng, Q., Finley, T., Wang, T.F., Chen, W., Ma, W.D., Ye, Q.W., Liu, T.Y., Year. Lightgbm: A highly efficient gradient boosting decision tree. In: Proceedings of the 31st Annual Conference on Neural Information Processing Systems (NIPS), Long Beach, CA.
Kus¸kapan, E., Çodur, M.Y., Atalay, A., 2021. Speed violation analysis of heavy vehicles on highways using spatial analysis and machine learning algorithms. Accident Analysis and Prevention 155, 106098.
Lee, J., Abdel-Aty, M., Jiang, X.M., 2014. Development of zone system for macro-level traffic safety analysis. Journal of Transport Geography 38, 13–21.
Lundberg, S.M., Lee, S.I., Year. A unified approach to interpreting model predictions. In: Proceedings of the 31st Annual Conference on Neural Information Processing Systems (NIPS), Long Beach, CA.
Mane, A.S., Pulugurtha, S.S., 2018. Influence of on-network, traffic, signal, demographic, and land use characteristics by area type on red light violation crashes. Accident Analysis and Prevention 120, 101–113.
Martinez-Ruiz, D.M., Fandino-Losada, A., De Leon, A.P., Arango-Londono, D., Mateus, J. C., Jaramillo-Molina, C., Bonilla-Escobar, F.J., Vivase, H., Vanlaar, W., Gutierrez-Martinez, M.I., 2019. Impact evaluation of camera enforcement for traffic violations in cali, colombia, 2008–2014. Accident Analysis and Prevention 125, 267–274.
Pei, Y.Y., Wang, X.S., Feng, M.J., Zhu, Z.X., Liu, F., Y., Q.Z., Jovanis, P.P., 2021. Macrolevel safety analysis of crashes and violations: Influencing factors and crash hotspots. In: Proceedings of the 100th annual meeting of the transportation research board.
Persaud, B., Lyon, C., Nguyen, T., 1999. Empirical bayes procedure for ranking sites for safety investigation by potential for safety improvement. Transportation research record 1665 (1), 7–12.
Rahman, M.S., Abdel-Aty, M., Hasan, S., Cai, Q., 2019. Applying machine learning approaches to analyze the vulnerable road-users' crashes at statewide traffic analysis zones. Journal of Safety Research 70, 275–288.
Rothengatter, T., 1982. The effects of police surveillance and law-enforcement on road user behavior. Bulletin of the British Psychological Society 35 (DEC), A101-A101.
Shao, Q., Zhang, W., Cao, X., Yang, J., Yin, J., 2020. Threshold and moderating effects of land use on metro ridership in shenzhen: Implications for tod planning. Journal of Transport Geography 89, 102878.
Terrill, T.T., Mashhadi, M.M.R., Ksaibati, K., 2016. Developing a tool to help highway patrol in allocating resources to crashes. International Journal of Police Science & Management 18 (4), 231–241.
Valen, A., Bogstrand, S.T., Vindenes, V., Frost, J., Larsson, M., Holtan, A., Gjerde, H., 2019. Driver-related risk factors of fatal road traffic crashes associated with alcohol or drug impairment. Accident Analysis and Prevention 131, 191–199.
Wang, X.S., Fan, T.X., Chen, M., Deng, B., Wu, B., Tremont, P., 2015. Safety modeling of urban arterials in shanghai, china. Accident Analysis and Prevention 83, 57–66.
Wang, X., Pei, Y., Yang, M., Yuan, J., 2021. Meso-level hotspot identification for suburban arterials. Accident Analysis and Prevention 156, 106148.
Wang, C., Xu, C., Fan, P., 2020. Effects of traffic enforcement cameras on macro-level traffic safety: A spatial modeling analysis considering interactions with roadway and land use characteristics. Accident Analysis and Prevention 144, 105659.
Wang, X.S., Yang, J.G., Lee, C., Ji, Z.R., You, S.K., 2016. Macro-level safety analysis of pedestrian crashes in shanghai, china. Accident Analysis and Prevention 96, 12–21.
Wang, X.S., Zhou, Q.Y., Yang, J.G., You, S.K., Song, Y., Xue, M.G., 2019. Macro-level traffic safety analysis in shanghai, china. Accident Analysis and Prevention 125, 249–256.
Wen, X., Xie, Y., Wu, L., Jiang, L., 2021. Quantifying and comparing the effects of key risk factors on various types of roadway segment crashes with lightgbm and shap. Accident Analysis and Prevention 159, 106261.
Yang, C., Chen, M., Yuan, Q., 2021. The application of xgboost and shap to examining the factors in freight truck-related crashes: An exploratory analysis. Accident Analysis and Prevention 158, 106153.
Yannis, G., Papadimitriou, E., Antoniou, C., 2008. Impact of enforcement on traffic accidents and fatalities: A multivariate multilevel analysis. Safety Science 46 (5), 738–750.
Yuan, C., Li, Y.e., Huang, H., Wang, S., Sun, Z., Wang, H., 2022. Application of explainable machine learning for real-time safety analysis toward a connected vehicle environment. Accident Analysis and Prevention 171, 106681.
Zhang, G.N., Yau, K.K.W., Chen, G.H., 2013. Risk factors associated with traffic violations and accident severity in china. Accident Analysis and Prevention 59, 18–25.

wangsystematicapproachmacrolevel2024.md 80 KB Історія Запис