Exploring traffic safety climate with driving condition and driving behaviour: a random parameter structural equation model approach

This study aimed to explore traffic safety climate by quantifying driving conditions and driving behaviour. To achieve the objective, the random parameter structural equation model was proposed so that driver action and driving condition can address the safety climate by integrating crash features, vehicle profiles, roadway conditions and environment conditions. The geo-localized crash open data of Las Vegas metropolitan area were collected from 2014 to 2016, including 27 arterials with 16 827 injury samples. By quantifying the driving conditions and driving actions, the random parameter structural equation model was built up with measurement variables and latent variables. Results revealed that the random parameter structural equation model can address traffic safety climate quantitatively, while driving conditions and driving actions were quantified and reflected by vehicles, road environment and crash features correspondingly. The findings provide potential insights for practitioners and policy makers to improve the driving environment and traffic safety culture.


Introduction
As reported by World Health Organization (WHO), each year about 1.24 million people are killed in traffic collisions across the globe, and if the current crash trend is followed, traffic death will become the top fifth reason for deaths by 2030 [1]. Thus, improving traffic safety has been the uppermost priority.
Among the current safety improvement approaches, countermeasures and strategies, all have been tried and utilized in different ways from the perspectives of human factors, vehicles, roadways and environment, reducing traffic fatalities and improving the safety level to some extent. However, the extent of improvement has slowed down, implying that the impact of regular approaches on safety may be gradually reduced [1]. Therefore, it is urgent to import some new approaches to prevent the declining trend in traffic safety improvement.
Safety culture in organizations, as "the product of individual and group values, attitudes, competencies, and patterns of behaviour that determine the commitment to, and the style and proficiency of an organization's health and safety programs" [2,3], has been accepted by various enterprises and institutions. And organizational safety culture can be reflected by the safety climate within certain limited regions or areas. Safety climate, as a manifestation of safety culture, is more easily quantified by the participants' perception [3][4][5][6], especially their attitudes and behaviour.
Traffic safety culture, a relatively new concept in the safety research field, has been recently introduced from organizational safety culture. Because the traffic safety culture within certain regions or areas, such as a university campus or some community, CBD area, strip area in Las Vegas, etc., may cause road users to follow similar attitudes and behaviours, which is totally consistent with the requirements of organizational safety culture, and it can be transferred to the traffic safety field. Similarly, traffic safety culture can be revealed by the safety climate; hence, how to estimate and quantify the traffic safety climate is the key issue.
According to Zhang et al. [3] and Aburumman et al. [4], traffic safety climate (TSC) concentrates not only on "road users' (e.g. drivers, bicyclists and pedestrians) attitudes and perceptions", but also on vehicles, roadways and the environment, which makes it possible to be estimated and quantified. Currently, there are a variety of studies focusing on traffic safety climate; however, few have actually presented concrete results from the perspective of attitudes and behaviour. Therefore, the objective of this study is to explore the TSC and present a structural equation model so that the safety climate can be addressed by considering driving conditions and driving actions simultaneously.

Literature review
Initially, the basic concept and framework of TSC was introduced from organizational safety culture. Naevestad and Bjørnskau [7] examined how the safety culture idea can be transferred to road traffic from organizational safety culture, and the results concluded that for targeted traffic safety interventions, the peer group alternative seemed to be new and promising. Edwards et al. [8] constructed a framework for conceptualizing TSC.
Three key issues were revealed: the difference from organization safety culture, the components or factors that traffic safety culture was comprised of, and the extent that TSC was changed. Then Luria et al. [9] conceptualized and measured road safety climate from the community perspective. A qualitative interview-based study was conducted to investigate community impact on road safety and build a safety climate scale. The findings showed that road safety climate was concerned with geographical communities. On the other side, Coogan et al. [10] examined the differences among different groups in behaviour and attitude, and investigated ten separate factors hypothesized to explore the propensity so as to conduct risky driving behaviour, and establish a multistructural equation model. The results verified that drivers' attitudes and behaviours can be beneficial for the safety culture. Innovatively, Qu et al. [11] extended TSC to assess public receptivity for autonomous vehicles.
Extensively, traffic safety culture has been addressed in different countries from various perspectives. Gehlert et al. [12] validated traffic safety climate attitudes in Germany. A selfreported Traffic Climate Scale (TCS) questionnaire was proposed, and it was found that external affective demands were consistently related to the perception of the driving/riding style, while internal requirements were consistently concerned with one's own driving/riding style. The results pointed out that a positive traffic safety climate was related to more secondary tasks during driving and traffic violations. Naevestad et al. [13] reported the traffic safety culture study from bicyclists in Norway. Results indicated that respondents' TSC were associated with those Downloaded from https://academic.oup.com/tse/article/3/3/tdab015/6358800 by guest on 03 September 2021 that they attribute to their peers, and the safety behaviours can forecast their bicycle accident risk. Nevertheless, Grytnes et al. [14] focused on the nontraffic-related work safety with drivers of heavy freight vehicles in Denmark. Qualitative interviews were performed with drivers, and the results found that the organizational structure of the company shaped their individual attitudes towards safety, as well as being dependent on relationships with their colleagues and friends. From the quantitative perspective, Islam et al. [1] measured TSC among distracted driving, impaired driving and speeding in Canada. A telephone survey was conducted, and multivariate confirmatory factor analysis was performed with structural equation modelling. The results demonstrated that perceived threat to personal safety had an important impact on self-reported behaviour, while various sociodemographic characteristics had a critical impact on the perceived threat of traffic behaviour to personal safety. Continuously, Stringer [15] explored traffic safety culture and drunk driving in the USA. Multilevel growth curve models were proposed to investigate the relationship with panel data. It was found that an increase in the relevant anti-alcohol community standards were related to a decrease in alcohol-related crashes at the county level. From the system resilience perspective, Tümer et al. [16] investigated traffic system resilience and driving skills related to the traffic safety climate in Turkey. A questionnaire was conducted, and hierarchical regression analysis was employed. The results revealed that traffic system resilience and driving skills were important influencers for traffic safety climate, and safety skills in particular were found to be critical for traffic climate. Due to a different culture from Western countries, Xu et al. [17] and Zhang et al. [3] focused on the traffic climate in China from the pedestrian and driving behaviour perspectives, respectively. The Chinese version of TCS was validated. Exploratory factor analysis and confirmatory factor analysis were adopted, and the results showed that the relationship between the inconvenience to pedestrians and pedestrians' transgressive behaviour was fully alleviated by traffic safety climate.
Furthermore, some comparisons of TSCs have been made among different countries. Nordfjaern et al. [18] examined traffic safety and driver behaviour in Turkey and Iran, due to differences in road traffic culture, personal traits, risk perception and attitudes. A questionnaire survey was conducted in a sample of road users. The findings revealed that normlessness was the significant predictor in driver behaviour. Atchley et al. [19] compared traffic safety culture from a cultural foundations perspective in China, Japan and the United States. By comparing traffic culture and safety attitudes, three countries were explored, and the unique cultural influences on traffic safety were verified. Similarly, Nordfjaern et al. [20] checked country cluster differences and attitudes to traffic safety and driver behaviour from eight countries. The results revealed that Norwegians provided overall safer attitudes towards traffic safety and driver behaviour cluster while individuals in Africa experienced the highest risk perception. The latest study by Kacan et al. [21] examined the traffic safety climate relationship with driver behaviours from five countries. The results showed that internal requirements and self-transcendence had positive relation in all countries, while external affective demands and conservation revealed a partial positive relation.
Structural equation models (SEMs) have been applied in various fields [22][23][24][25]. In the transportation safety area, Ye et al. [25] employed a simultaneous equation model to analyse crash frequency by collision type for rural intersections, and Xu et al. [26] extended SEM models with panel data to determine the influencing factors of arterial mid-blocks.
To sum up, from the import of organizational safety culture to TSC, variety of studies in different countries or regions have been conducted to reveal the relationship between TSC and various relevant factors, in which most of the studies adopt qualitative analysis and only a few employ quantitative approaches. Questionnaire surveys have been widely used because culture, to some extent, is hard to quantify. Additionally, most studies concentrate on one aspect or a specified perspective to discuss TSC, while few focus on comprehensive interaction, human-vehicle-roadwayenvironment integration. Therefore, the purpose of this study is to quantify the traffic safety climate from attitudes and behaviour simultaneously. The quantitative SEM will be proposed so that driver personality and behaviour will be reflected from crash features, vehicle profiles, roadway conditions and environment conditions to comprehensively address the safety climate within a certain area.

Methodology
Different from traditional statistical analysis methods, structural equation modelling is a type of statistical procedure that integrates regression analysis of linear model with factor analysis, which can realize the model identification, estimation and validation of various causal models. Therefore, it has been widely applied in behavioural and social science, and plays a significant role in quantification research [1].
The general SEM is composed of two basic components, a measurement model and a structural model. Correspondingly, the former establishes the relationship between observed variables and latent variables, and can employ confirmatory factor analysis (CFA) to test, while the latter constructs the relationship between latent variables, and specifies the casual relationship among latent variables. Therefore, the measurement models can be expressed as Equations (1) and (2) x where x represents the exogenous measurement variable, x denotes the factor loading of x, ξ is the exogenous latent variable, and δ is the error term of the exogenous measurement variable explained by the exogenous latent variable, which is unrelated to ξ ; y represents the endogenous measurement variable, y denotes the factor loading of y, η is the endogenous latent variable, and ε is the error term of the endogenous measurement variable explained by the endogenous latent variable, which is unrelated to η. Specifically, measurement variables are vehicle 1 & 2 driver behaviour (age, action and condition), and latent variable is injury severity. The SEM can be described as Equation (3) where B denotes the related coefficient matrix between endogenous latent variables, denotes the regression coefficient matrix between exogenous latent variables and endogenous latent variables and ζ is error term of unpredicted or unaccounted in structural equation model. Specifically, B includes driver behaviour and driving condition, and involves other influencing factors.
In order to account for heterogeneity due to unobservable factors, a random parameter can be employed, and the estimable parameters can be expressed as where β represents estimated parameters, and ϕ i is a randomly distributed term.
For the model identification, it is completed according to rule t. If Equation (4) is allowed, the model can be identified exactly, and vice versa.
where df is the number of degrees of freedom in the model, and p and q are the number of exogenous and endogenous measurement variables, respectively. There are various approaches to estimate the parameters in SEM analysis, of which instrumental variables, two-stage least squares, maximum likelihood and generalized least squares are widely used. In this study, the maximum likelihood approach is employed to assess the random parameters of the SEM. Similar to Xu et al. [26], the estimation of two equations can be converted to one, which can be evaluated by maximum likelihood method. In this way, the random parameter model can address the heterogeneity issue due to unobserved factors while the SEM can accommodate the intrinsic endogeneity issue between dependent variables and independent variables.
After the model parameters are estimated, model modification is required for the goodnessof-fit between the hypothesized and observed values. This can be achieved by adding or removing some parameters in the model so that the goodness-of-fit is improved. In the following sections the estimation and validation steps will be illustrated.

Data description
The geographical information system (GIS) open data site maintained by Nevada Department of Transportation (NDOT) was adopted to collect the data from 2014 to 2016. And the target employed in this study was the metropolitan Las Vegas area, which is formed by major and minor arterials, involving City of Las Vegas, City of North Las Vegas, City of Henderson and Clark County. Literally, there were 27 major and minor arterials included, and 25 030 injuries involved as shown in Fig. 1. After some invalid data were removed, 16 827 injuries were kept. The main factors were extracted from four aspects: the road users' features (including drivers, pedestrians and cyclists), vehicle profiles, roadway conditions and environment conditions.
In order to quantify the traffic safety climate, the TCS presented by Gehlert et al. [12] and Zhang et al. [3] is considered as the index, which is composed of external demands, internal requirements and functionality. In order to match with the three aspects, the data were collected mainly focusing on the road users' features, including personal status, emotions, cognition and behaviour. Therefore, the road users can be considered as the dependent variables in the proposed model, while road users here include drivers, pedestrians, pedal cyclists and motorcyclists. Due to the limitations of the dataset, the road users' attributes, e.g. pedestrians, bicylists and motorcyclists, are not collcted except that binary attributes of crash injury are provided, i.e. yes or no. Thus the road users in this study mainly concentrate on the drivers' attributes.
Moreover, in the data set the collision includes either two vehicles or more than two vehicles, in which the at-fault vehicles are regarded as vehicle 1 and those not-at-fault are vehicle 2. In the same way, the drivers' features are separated into vehicle 1 and vehicle 2. Thus, according to this division, the variables can be listed as vehicle 1 driver age, vehicle 1 driver action, vehicle 2 driver age, vehicle 2 driver action, etc. Here, driver actions are the representatives of driver behaviour in this data set; thus, in the following modelling process, driver actions are considered to reflect driver behaviour.
In accordance with the vehicles involved, the vehicle profiles contain the total vehicle, vehicle types, vehicle direction and vehicle conditions (e.g. hit-and-run, mechanical defects, driving too fast, etc.).
In terms of roadway characteristics, the number of vehicle lanes, roadway conditions (e.g. dry, wet, ice, snow, etc.) and highway factor (work zone or not) are collected, whereas for the environment condition, the weather, lighting conditions and first harm (e.g. median, fence, pedestrian, etc.) are extracted.
In Nevada, the injury severity is typically classified as property damage only (PDO), injury and fatality. In the chosen sample, there is only about 0.5% fatality. Because the injury and fatality were severe, merging both did not potentially affect the inference. Reflected from ArcGIS, the injury within 100 ft (30.48m) of arterials was buffered; thus, the observed injury severity can be considered as a latent variable. Therefore, the latent variable injury severity was considered to be binary, in which PDO and injury together with fatality were named as 0 and 1, respectively.
To assess the proposed models in AMOS software, all the variables are categorized and digitalized as listed and summarized in Table 1. In the upper part, dependent, latent and categorical variables are listed, while the descriptive statistics of the continuous/indicator variables are in the lower part.

Results
According to the requirements of SEM, variables in the model are divided into latent variables and observational variables. And the interaction between variables is divided into the influence of exogenous variables on latent variables, and the influence of endogenous variables on latent variables. In terms of the objective, the state and behaviour of drivers can reflect the TCS of the local traffic safety climate. Therefore, in order to correspond to the external emotional needs, internal needs and functions of the TCS, the variables  reflecting the characteristics of drivers are considered as dependent variables of the model, namely, driving condition and driving action (driver age is firstly included, but the variance of residue is negative, thus it cannot be adopted in the AMOS software). The SEM is established by reflecting the explanatory variables with the crash features, road characteristics, environmental conditions and vehicle conditions as independent variables. On the basis of 16 827 sample data, AMOS software was employed to calculate the influencing path, and model estimation between variables was conducted by Stata. Fig. 2 gives the influencing path of all variables, and Table 2 shows the fitting results of the proposed model.

Fig. 2. Influencing path of all variables
As shown in Table 2, GFI, AGFI and PGFI all meet the test standard, especially GFI and AGFI larger than 0.9; the chi-square degree of freedom ratio is much higher than the judgement criteria, but RMR and RMSEA are larger than 0.05, denoting that the overall fit of the model is not so good.
In order to enhance the fitting degree of the proposed model, a modification is required, which includes two directions. One is to simplify the model by deleting or constraining some influencing paths; the other is to modify the paths, i.e. releasing some path restrictions and improving the fitting degree of the model. Obviously, the two cannot be realized at the same time. But for either direction, the ultimate goal is to obtain a model that is both concise and consistent with practical significance. Correction in these two directions is mainly based on the change in the size of the modification index (MI) and the critical ratio. Here, the degree of model fitting is improved mainly by MI. The following principles should be followed when revising the index.
(i) First, add or delete the direct path of the influence between potential variables in the structural model, because the analysis model of potential variable path focuses on the SEM instead of the measurement model. (ii) Then, the covariance between the error terms of observation variables in the same measurement model can be defined in priority. (iii) The next step is to define the covariant relationship between the error terms of two observation variables that are both potential external factors, and the covariant relationship  between the error terms of observation variables of two different measurement models, implying that there is a certain relationship between the dimensions of these two constructs reflecting different potential factors.
In the SEM of this study, the variables Crash type, Severity and First harm have covariant relationships among each other, which can explain each other to a certain extent. And there are causal relations between these three variables and Vehicle 1. There are also causal relationships between Crash type and Vehicle 2 action, Severity and Vehicle 1 driver. Therefore, the release of the covariance parameter can be performed on the path between the variables above. The modified influencing path of selected variables is shown in Fig. 3. Fig. 3 gives the modified influencing path of selected variables, and Table 3 shows the fitting results of the modified random parameter SEM. It can be found that in Table 3, GFI, AGFI, RMR, RMSEA and PGFI all meet the test standard, in which GFI and AGFI are larger than 0.90, and RMR and RMSEA are smaller than 0.05. Moreover, the chi-square degree of freedom ratio is a little higher than the judgement criterion, but also close to the standard value, implying that the SEM has a good fitting effect on the whole. From Fig. 3, it can be surmised that driving action and driving conditions can be reflected from a series of variables. Table 4 gives the estimation coefficients of all the variables by Stata. For standardized parameters, the closer the parameter value is to 1 or -1, the greater the weight or importance of the surface variable dimension will be. A positive value provides a positive effect, a negative value gives a negative effect, and p value less than 0.1 indicates a significance level of 10%. In the model, road environment (i.e. first harm, weather and highway factor) has a negative influence on driving condition (i.e. vehicle 1 and 2 driver condition), and vehicle 1 condition has a positive impact, while other variables are not significant. This implies that the driver conditions of vehicles 1 and 2 are influenced by first harm, weather and highway factor negatively, and by the at-fault vehicle condition positively, implying that the driver conditions are easily affected by vehicles, work zone or not and weather factors.
In the same manner, environment condition has a positive impact on driving behaviour (including vehicles 1 and 2 action), while vehicle 1 Note: * denotes significance at the level of P < i < 0.1; * * means significance at the level of P < 0.01. condition, crash type and accident (including total number and severity) have negative influence on driving behaviour. This indicates that whether the drivers make actions depends not only on road environment and at-fault vehicle condition, but also on the crash features. Reliability and validity are the two most important and fundamental features in the evaluation of criteria [27]. In the reliability analysis, Cronbach's alpha coefficient is the most widely used. However, it is supposed that the load values of potential variables to all index items are equal, which is obviously inconsistent with the reality. After the emergence of structural model mode, combination reliability (CR) and average variance extraction (AVE) are preferred to be employed. CR value is the integration of all measurement variables, and represents the internal consistency of the construct pointer, which is similar to Cronbach's alpha. The higher the CR value, the higher the internal consistency of the construct is, and 0.7 is an acceptable threshold. AVE is the explanatory power to calculate the variance of the measured variables of potential variables. The higher the AVE is indicates that the reliability and convergent validity of the construct is higher. The ideal upper standard value must be greater than 0.5, and 0.36-0.5 is an acceptable threshold range.
Validity refers to the degree to which the content of index data can correctly reflect the traits to be measured, including content validity, criterion validity and structure validity. Among them, structure validity can be monitored by quantitative means, mainly by measuring KMO (Kaiser-Meyer-Olkin) value, and conducting a Bartlett spherical test. The standard is that the KMO value is greater than 0.6, and the Bartlett spherical test is less than 0.01 for p value.
The test results of data reliability and validity are shown in Table 5. The CR value reaches the standard, and the AVE value lies between the standard values, amounting to the acceptable threshold. Because of the large number of samples, the KMO value greater than 0.6 can be regarded as meeting the reliability test standard. And the Bartlett spherical test is significant, so it can pass the reliability and validity test, and the proposed SEM can be verified.

Discussions
TSC, as a relatively new concept in the safety research field, is introduced from organizational safety culture, and so far there have been a number of studies to explore it from various perspectives. Because TSC is difficult to quantify, most of the current studies employ qualitative analysis, e.g. questionnaire survey, and only a few quantitative approaches have been used. In this study, the quantitative SEM is proposed to explore and quantify the TSC with the driver condition and driver behaviour, which fills the gap of qualitative analysis. More importantly, this study investigates the TSC from driver condition and driver behaviour simultaneously, which can address the safety climate within certain area comprehensively.
According to the results of the proposed model, driving condition and driving actions are two important aspects to be emphasized for TSC, and corresponding countermeasures can be made for practitioners and policy makers. First, whether the at-fault or not-at-fault drivers should pay attention to the vehicles, work zones and weather condition. Specifically, the vehicles should be maintained and examined regularly to maintain good condition, especially for those at-fault vehicles; the drivers should be concerned with the motor vehicle in transport to avoid first harm where possible; when the drivers arrive at work zones, more concentration should be taken-for instance, reducing speed to lessen the fluctuation, and education should be programmed if necessary; under bad weather conditions, especially rainy or snowy days, driving vision is damaged seriously, and although those days only account for a small proportion of injuries, the injury generated is more severe than that under clear conditions, thus the drivers should make cautious moves under those conditions in order to travel satisfactorily.
Second, the driving actions rely on vehicle and road environment as well as the crash features. Specifically, when the drivers make actions, attention should be paid to vehicles around to avoid first harm and more attention should be paid to work zones; when drivers pass through a location with high numbers of crashes, or a segment with severe crashes, alert should be taken by drivers so that proper actions can be made, for example, minding sharp bends and adapting to the speed appropriately.
Empirically, in accordance with the results obtained, if the TSC is constructed within the community or the city, driving condition and driving actions are the main aspects to be considered besides the general human-vehicle-roadwayenvironment factors. First, the vehicles' examination standard should be built up uniformly, while motor vehicles should be restricted in certain areas to prevent first harm. As for work zones, the length, lane closed and speed limit should stick to strict standards, and information should be issued through the Internet, cell phone and social media regularly so that the public are kept informed. More importantly, the drivers' "SAFETY FIRST" consciousness should be educated from the beginning of driving licence testing to everyday life, e.g. no speeding, pedestrian first, no road anger and civilized driving. Last but not least, road planning and design should follow standard procedures and consider the safety from the beginning, and set up necessary safety labels, message signs and boards at appropriate locations as well as maintaining the lighting condition properly and the environment comfortably.

Conclusions
In this study the traffic safety climate was investigated with the driving condition and driving actions by considering the crash-vehicleroadway-environment integration factor. To quantify the TSC, a random parameter structural equation model was proposed so that driver condition and actions can address the safety climate with crash features, vehicle profiles, roadway conditions and driving environment comprehensively within a certain area. The results showed that driving condition and driving actions can address traffic safety climate quantitatively and simultaneously, while driving condition and driving actions were influenced by vehicles, road environment and crash features correspondingly.
Two main findings can be obtained from the results. First, a random parameter structural equation model can accommodate the endogeneity and heterogeneity issues simultaneously and address the traffic safety climate quantitatively. To our knowledge, this is the first attempt to apply the proposed model in a traffic safety climate. Second, driving condition and driving actions can be quantified simultaneously and reflected by the traffic safety culture in Las Vegas, which provides potential insights for traffic policy.
Some drawbacks are still present in this study. One point is that the driver variables only include age, condition and action, and in the future more variables reflecting the drivers' personality should be included. Additionally, the data set from Las Vegas gives one example of the proposed model, and more data sources should be tried to confirm the findings and generality in further work. Future study may consider the impact of spatial or temporal features on TSC, which may address the TSC more accurately.

Supplementary data
Supplementary data are available at Transportation Safety and Environment online.