The differences in traffic accident severity between urban and rural areas have been widely studied, but conclusions are still limited. To explore the factors influencing the occurrence of roadside accidents in urban and rural areas, 3735 roadside traffic accidents from 2017 to 2019 were analyzed. Fourteen variables from the aspects of driver, vehicle, driving environment, and other influencing factors were selected to establish a Bayesian binary logit model of roadside crashes. The deviance information criterion and receiver operating characteristic curve were used to test the goodness of fit for the traffic crash model. The results show that: (1) the Bayesian binary logit regression model well fits the traffic crash data, and the goodness of fit for sub-models based on separate urban and rural data is better than when based on all data (urban and rural); (2) 10 variables have a significant influence on the extent of roadside crash severity in the two areas, with different impacts; (3) drunk driving increases the probabilities of fatal traffic accidents by 10.8% and 16.4% in urban and rural areas, respectively; (4) the probabilities of fatality caused by traffic accidents involving trucks are 4.6% and 9.8% higher than those without trucks in urban and rural areas, respectively. The findings of this study can provide a theoretical foundation for traffic safety administration to formulate relevant policies or strategies in order to reduce the severity of roadside accidents.
Roadside crashes are accidents in which a vehicle deviates from the traffic lane and enters the road shoulder or the area outside the shoulder. The vehicle collides with guardrails, vehicles, pedestrians, or other obstacles. They also involve other accidents such as rollovers and fallings. According to the statistics, 31.7% of all vehicle traffic accidents were roadside accidents in 2017. Injuries and fatalities in accounted for 32.65% and 33.16% of the respective totals for all accidents [
Scholars have made great progress identifying influencing factors and their effects on traffic accidents [
There are significant differences between urban and rural areas in regards to highway mileage, road grade, car ownership, traffic volume, and number of car accidents [
Numerous studies have used real-world accident data in urban or rural areas to investigate factors affecting road safety [
The impacts of factors on accident severity between urban and rural areas has also been explored. For example, Al-Bdairi and Hernandez [
However, due to sample sizes, differences in driving styles [
To summarize, by not considering the differences between the two regions, many accident analysis models achieve low accuracy in a wide range of applications. More attention should be paid to the differences of influence factors on roadside crash severity between urban and rural regions. This study uses traffic accident data in urban and rural areas of Shandong, China, to identify the differences of influence factors on roadside traffic accident severity in the two regions and develop appropriate accident analysis models.
The occurrence and severity of traffic accidents are closely related to the factors of humans, vehicles, roads, and the environment [
Traffic accident data for this study were from Shandong province in 2017–2019, as provided by the Shandong Department of Transportation (SDOT). This study explores the influencing factors on the probability of roadside accidents in urban and rural areas, and the differences of the probability of fatal roadside accidents between urban and rural areas under different factors. Thus crashes occurring in urban and rural areas were reserved, and others, such as in suburbs, were removed by manual screening, to obtain a total of 2664 crashes in urban areas and 1071 in rural areas. All the model variables were encoded and transformed to categorical variables. For example, according to the worst injury severity in an accident, its severity was transformed to a dichotomous variable, coded as 1 if someone was killed (on-the-spot or within seven days), and 0 otherwise. The injury severity of the
Considering that two-wheeled motorcycles and electric bicycles have similar characteristics in traffic accidents (e.g., they cannot provide necessary protection for passengers), they are both classified as two-wheeled vehicles; if one of these is involved in an accident, the code (which is shown in
Factors | Code (Percentage of urban/rural accidents) |
---|---|
Injury severity ( |
1 = Fatality (29.4/35.6); 0 = No fatality (70.6/64.4) |
Gender ( |
0 = Male (90.4/92.5); 1 = Female (9.6/7.5) |
Age ( |
0 = Mid-Age (25–50)(79.4/81.9); |
1 = Young (<25)(14.2/11.8); 2 = Old (>50)(6.3/6.2) | |
Drunk driving ( |
0 = No (94.2/92.1); 1 = Yes (5.8/7.9) |
Use of seat belt/helmet ( |
0 = Use (79.5/39.9); 1 = No use (20.5/60.1) |
Fatigue driving ( |
0 = No (91.7/85.5); 1 = Yes (8.3/14.5) |
Accident type ( |
0 = Single-vehicle accident (14.6/27.9); |
1 = Multiple-vehicle accident (85.4/72.1) | |
Truck involved ( |
0 = No (78.1/62.6); 1 = Yes (21.9/37.4) |
Two-wheeled vehicle involved ( |
0 = No (56.3/47.5); 1 = Yes (43.7/52.5) |
Cause of accident ( |
0 = Vehicle (94.6/90.0); 1 = Others (5.4/10.0) |
Weather conditions ( |
0 = Clear weather (87.4/91.2); 1 = Non-clear weather (12.6/8.8) |
Road surface ( |
0 = Dry (89.8/90.5); 1 = Non-dry (10.2/9.5) |
Visibility ( |
0 = ≥200 m (59.2/47.7); 1 = 100–200 m (18.3/18.5) |
2 = 50–100 m (12.4/17.3); 3 = ≤50 m (10.1/16.5) | |
Time of accident ( |
0 = 7:00–09:59 (16.7/12.7); 1 = 10:00–16:59 (32.2/38.2) |
2 = 17:00–19:59 (35.6/23.6); 3 = 20:00–6:59 (15.6/25.5) | |
Date of accident ( |
0 = Work day (69.9/67.5); 1 = Holiday (30.1/32.5) |
As shown in
The proportion of single-vehicle roadside accidents is higher in rural areas (27.9%) than in urban areas (14.6%). This may be due to the poor road conditions in rural areas, resulting in roadside crashes due to an object falling from the roadside, hitting a fixed object alongside the road, or poor vision. The proportions of roadside crashes involving trucks and two-wheeled vehicles are 37.4% and 52.5%, respectively, in rural areas, which are higher than in urban areas. The proportions of roadside crashes in non-clear weather and with non-dry road surface conditions are 12.6% and 10.2%, respectively, in urban areas, which are higher than in rural areas. People in urban regions might be less likely to change their trip plans in bad weather (e.g., rainfall, snow, and wind) due to choices of transport modes [
The proportion of accidents is lower on holidays than on work days. The percentage of accidents on holidays is higher in rural districts than in urban districts. The proportion of accidents in the morning and evening is higher in urban areas than in rural areas, but the proportion of accidents through the night is higher in rural areas. This is probably because of poor lighting on many rural roads [
The logit model is considered a generalized linear regression model. The response variables are converted to the logit function, which improves the applicability and fitting performance of the model. Suppose that response variable
where
Bayesian analysis defines the unknown parameters as random variables. The posterior distribution of parameters is obtained as [
where
The prior information is expressed by a prior distribution of parameters in Bayesian analysis. The prior distribution includes the informative prior distributions and uniform priors or non-informative priors. And the posterior distribution of parameters is directly affected by the prior distribution. This research uses uniform priors following a normal distribution. The initial values of each parameter are estimated by maximum likelihood. The standard deviation of the normal prior distribution for the coefficients is
The joint density function of
Then the posterior distribution of parameter
Considering the complexity of the posterior distribution for parameter
The receiver operating characteristic (ROC) curve is often used to test the fitting performance of a binary response model (event = 1, non-event = 0) [
The ROC curve shows the relationship between sensitivity and 1-specificity under different thresholds. The area of the graph formed by the curve, (1, 0) and (1, 1) on the X-axis is represented by the area under curve (AUC) value. The value of AUC ranges from 0 to 1. AUC can also directly evaluate model fitting. The larger the value the better the model fits. Model fitting accuracy is higher as AUC approaches 1.
The deviance information criterion (DIC) comprehensively evaluates the goodness of fit and complexity of a Bayesian model, and is used to compare the advantages of different models [
Where
If the difference of the DIC values of two models is more than 5, then it can be considered that the model with smaller DIC has better fitting performance. If the difference is less than 5, then the DIC value cannot be used as the only index to judge the fitting performance of the model.
We calculate the average marginal effect of factors that have a significant influence on the severity of accidents. This is used for quantitative analysis of the change of a dependent variable when an independent variable changes. Since all independent variables in the model have been converted to dummy variables, the marginal effect is determined as
where
Model parameters were estimated using the Bayesian algorithm in Stata 16.0 software. Considering the complexity of the model, three MCMC chains were constructed for Bayesian inference. Each chain was iterated 20,000 times, and the historical graph and Brooks-Gelman-Rubin (BGR) values of the iterated chain for all parameters were used to check for reasonable convergence of the model. When the iterative historical graph of a model parameter shows no strong periodicity and the BGR is less than 1.2, model convergence occurs. The autocorrelation function of model parameters was tested to ensure that they were independent and identically distributed (IID). When the autocorrelation curve of the parameters quickly decreases to 0 and remains stable, the parameters can be considered to meet IID requirements. The length of the chain was detected by the Raftery–Lewis diagnostic based on the specific quantile of the parameter distribution [
Variables | Urban | Rural | ||
---|---|---|---|---|
Coef. (S.D.) | 95% BCI | Coef. (S.D.) | 95% BCI | |
Female | –0.538 (0.180) | –0.901, –0.198 | –0.369 (0.325) | –1.036, -0.075 |
Young (<25) | –0.517 (0.222) | –0.960, –0.100 | – | – |
Old (>50) | 0.051 (0.172) | 0.028, 0.401 | 0.786 (0.243) | 0.309, 1.265 |
Drunk driving | 2.437 (0.013) | 2.051, 3.825 | 3.714 (0.019) | 1.190, 5.252 |
Seat belt/helmet | 1.034 (0.105) | 0.194, 3.052 | 1.133 (0.067) | 0.008, 1.607 |
Fatigue driving | 1.106 (0.120) | 0. 801, 1.292 | 1.534 (0.129) | 0.471, 2.825 |
Multiple vehicles | 0.437 (0.196) | 0.051, 0.825 | –1.214 (0.207) | –1.788, –0.640 |
Truck | 1.619 (0.214) | 0.501, 2.796 | 3.249 (0.278) | 0.803, 5.953 |
Two-wheel vehicle | 0.905 (0.174) | 0.035, 1.638 | 0.734 (0.326) | 0.053, 1.574 |
Non-clear weather | 0.591 (0.052) | 0.408, 0.778 | – | – |
Non-dry surface | – | – | 0.142 (0.067) | 0.015, 1.184 |
Visibility (50 m–100 m) | – | – | 0.162 (0.092) | 0.015, 0.345 |
<50 m | 0.724 (0.021) | 0.590, 0.859 | 0.533 (0.094) | 0.348, 0.720 |
17:00–19:59 | – | – | 1.016 (0.190) | 0.111, 2.695 |
20:00–06:59 | 0.476 (0.156) | 0.018, 3.657 | 1.240 (0.164) | 0.255, 2.539 |
Holiday | – | – | 0.288 (0.107) | 0.068, 0.734 |
Model Intercept | –3.288 (0.199) | –5.080, –1.312 | –1.896 (0.015) | –4.214, –0.231 |
It should be noted that the model coefficients can indicate positive or negative correlation between independent and dependent variables, but they cannot quantify the influence of factors on accident levels. Therefore, the marginal effects of all independent variables with significant influence were calculated to directly show the influence of factors on the severity of roadside accidents. The marginal effects of model parameters are shown in
Variables | Urban | Rural |
---|---|---|
Marginal effects (S.D.) | Marginal effects (S.D.) | |
Female | –0.103 (0.034) | –0.069 (0.064) |
Young (<25) | –0.092 (0.035) | – |
Old (>50) | 0.091 (0.033) | 0.180 (0.058) |
Drunk driving | 0.108 (0.034) | 0.164 (0.092) |
Seat belt/helmet | 0.031 (0.109) | 0.042 (0.034) |
Fatigue driving | 0.039 (0.040) | 0.054 (0.030) |
Multiple vehicles | 0.045 (0.068) | –0.051 (0.068) |
Truck | 0.046 (0.017) | 0.098 (0.029) |
Two-wheel vehicle | 0.074 (0.005) | 0.062 (0.031) |
Non-clear weather | 0.014 (0.010) | – |
Non-dry surface | – | 0.006 (0.052) |
50 m–100 m | – | 0.016 (0.045) |
<50 m | 0.006 (0.030) | 0.004 (–0.035) |
17:00–19:59 | – | 0.059 (0.042) |
20:00–06:59 | 0.030 (0.043) | 0.089 (0.026) |
Holiday | – | 0.003 (0.028) |
Note: S.D. means standard deviation.
The following can be seen from
The goodness of fit tests of the roadside crash analysis model for urban and rural traffic accidents are shown in
Fitting test | Urban | Rural | Total |
---|---|---|---|
ROC | 0.798 | 0.797 | 0.654 |
DIC | 3248.076 | 1466.691 | 4772.946 |
Log marginal-likelihood | –1895.590 | –985.244 | –2941.520 |
According to the results of the Bayesian binary logit model, it was found that factors have different influences, and the same factor has different influences in urban and rural areas on the probability of a fatal roadside crash.
Twelve factors are significantly correlated with the accident severity of the roadside accident analysis model for urban areas, including drunk driving, the use of seatbelt/helmet, fatigue driving, number of vehicles involved, truck involved, two-wheeled vehicle involved, weather, visibility (less than 50 m), and time of occurrence (20:00–6:59). Fourteen factors are significantly correlated with the severity of roadside crashes in the model for rural areas, including driver gender, driver age (old), drunk driving, use of seat belts/helmet, fatigue driving, number of vehicles involved, truck involved, two-wheeled vehicle involved, road surface conditions, visibility (50 m–100 m, less than 50 m), time of occurrence, and holiday. Ten variables have an influence on the severity of roadside accidents in both rural and urban areas. The variables of road surface condition, visibility distance (50 m–100 m), time (20:00–6:59), and holiday have a great impact on roadside crashes only in rural areas, and the variables of weather and driver age (young) has a large effect on the severity of roadside crashes only in urban areas. We analyze the model factors from four aspects.
(1) Driver characteristics
Gender has a significant impact on the severity of urban and rural roadside accidents. Female drivers are less likely than males to be involved in fatal roadside accidents, with probabilities 10.3% less on urban roads and 6.9% on rural roads. Gender has a significant impact on the severity of various types of accidents. Wu et al. [
The probability of fatal injuries for young drivers (less than 25 years old) is 9.2% less than that of drivers 25–50 years of age in roadside accidents in urban areas, and the probabilities of urban and rural fatal injuries of old drivers (50 years old or older) are respectively 9.1% and 18% greater. This is consistent with the findings of previous studies [
The results show that drunk driving increases the probability of fatal roadside crashes in urban and rural areas by 10.8% and 16.4%, respectively. This is consistent with other research [
The research results show that non-use of seat belts/helmets increases the probability of fatal roadside crashes in urban and rural areas by 3.1% and 4.2%, respectively. It is noticed that increasing trends of urban and rural regions are approximately the same. This result is reasonable because it is widely accepted that the use of seat belts can significantly reduce the degree of injury to passengers [
Fatigue driving is not uncommon, especially among those who transport goods or passengers over long distances. They can easily feel exhausted after driving a long time, with serious consequences [
By comparing driver factors that have a significant influence on serious roadside traffic accidents on urban and rural roads, it was found that the factors of gender, age (old drivers), drunk driving, use of seat belt/helmet, and fatigue driving have a strong influence on the severity of roadside crashes in both regions. This indicates that driver factors are significantly correlated with the severity of roadside traffic collisions. The marginal effects of old drivers, drunk driving, use of seatbelt/helmet, and fatigue driving are greater in rural areas, which suggests that with the influence of the four factors, fatal roadside crashes are more likely to occur in rural areas. This may be due to poor road conditions, less organized traffic environments, lack of road safety awareness, and less effectual police enforcement.
(2) Vehicle characteristics
The probability of death in multiple-vehicle roadside crashes is 4.5% higher than in single-vehicle roadside accidents in urban areas. In rural areas, the probability of death in multiple-vehicle roadside accidents is 5.1% lower than in single-vehicle roadside accidents. Serious collisions are more likely to occur in multiple-vehicle accidents than in single-vehicle accidents in urban areas, and it is the opposite in rural areas. This is because surface conditions and auxiliary facilities of urban roads are pretty good, and most single-vehicle roadside accidents involve minor scratches and no injuries. Many rural roads have pavement in poor condition, and vehicles are prone to be involved in serious roadside accidents, such as by running into a ditch or sliding into a river. Thus, death is more likely in single-vehicle roadside accidents than in multiple-vehicle roadside accidents in rural areas.
The involvement of a truck often leads to fatalities in roadside accidents. The death rates in urban and rural regions are respectively 4.6% and 9.8% higher in truck-involved roadside accidents than without trucks. This suggests that vehicle crash deaths are more likely to occur in roadside traffic accidents involving trucks, and the casualty probabilities of roadside crashes involving trucks are higher in rural areas than in urban areas. It could therefore be said that the factor of truck involvement has a significant impact on the severity of roadside accidents in both areas. This is consistent with previous research [
The death rates of urban and rural regions are 7.4% and 6.3% higher in accidents involving two-wheeled vehicles than in those not involving two-wheeled vehicles. Motorcycles and bicycles increase the probability of fatal traffic accidents by 5–10 times [
(3) Driving environment characteristics
The regression coefficient of the weather factor in urban areas is positive, which shows that serious roadside traffic accidents in urban areas are more likely to occur in adverse weather than in sunny weather (the marginal effect is 1.4%). In addition, the impact of severe weather on the severity of the accident is nonlinear, thanks to the comprehensive effect of various factors. For example, the impact of rainfall on the severity of single-vehicle accidents is related to the rainfall intensity, wind speed, use of seat belt, driver’s gender, road type, and other factors [
The influence of road surface condition on the severity of roadside crashes is not significant in urban areas; however it is statistically significant in rural areas. Islam et al. [
Visibility of 50 m–100 m has no significant effect on the occurrence of fatal roadside crashes in urban areas, but the significant effect appears in rural areas. Compared with visibility above 200 m, the probability of fatal roadside crashes increases by 1.6% with the visibility distance of 50 m–100 m in rural areas, and by 0.6% and 0.4% with the visibility distance of less than 50 m in urban and rural areas, respectively. This shows that the worse the visibility the higher the probability of serious roadside crashes. The reason for this may be that drivers need long times to respond in case of emergency in low-visibility environments. This result is consistent with the findings of previous studies [
(4) Other influencing factors
Using the morning (7:00–09:59) as the reference period, the evening period (17:00–19:59) has no significant impact on the occurrence of serious roadside crashes in urban areas, but a significant influence on rural accidents, where fatal accidents are 5.9% more likely to occur. Compared to the reference period, the probabilities of serious roadside traffic crashes in urban and rural areas increase by 3.0% and 8.9%, respectively, in the period from night to early morning (20:00–09:59), which is consistent with Champahom et al. [
Taking non-holiday as a reference, holiday has no significant impact on serious roadside crashes in urban areas. The probability of a fatal roadside accident is 0.3% higher on a holiday in a village. This is because many people in rural communities return home for holidays, which increases traffic volumes.
Through analyzing the probability models of fatal roadside crashes in urban and rural areas, it can be seen that there is a difference between them, and the influence of different explanatory variables on the probability of fatal roadside crashes is heterogeneous. For example, factors with no significant influence on the occurrence of fatal roadside crashes in urban areas (e.g., dry road, holiday, visibility of 50 m–100 m, and time period 17:00–19:59), have a great effect on those in the countryside. The variable of severe weather has no clear influence on the probability of fatal roadside traffic accidents in rural areas, but it does in urban areas. Therefore, different measures should be adopted to accommodate the differences between urban and rural areas to effectively reduce fatal accidents.
(1) From the four aspects of driver, motor vehicle, road, and driving environment, 14 factors were selected to analyze the influence of variables on the severity of roadside crashes in urban and rural areas. Based on real-world traffic accidents in urban and rural areas of Shandong province, the binary logit regression model was established using Bayesian parameter estimation. The results show that the factors of driver age, visibility, time period, road surface condition, holiday, and weather have different effects on fatal crashes in rural and urban areas, and nine factors have different effects on serious traffic accidents.
The influence of young drivers and non-clear weather on the severity of roadside crashes is not significant in rural areas, but is statistically significant for roadside traffic collisions in urban areas. Some factors, e.g., non-dry road surface, visibility of 50 m–100 m, the evening period (17:00–19:59), and holiday, have no significant impact on the occurrence of serious roadside crashes in urban areas, but they have a significant impact on the severity of rural roadside accidents. The influence of different factors on the severity of roadside crashes shows diversity between urban and rural areas, and effective strategies should be developed based on these differences. In addition, 10 variables have a significant impact on the severity of roadside crashes in both urban and rural areas.
(2) Comparing the influence on the severity of roadside crashes in urban areas, nine factors can be expressed in descending order, as follows: drunk driving, female drivers, young drivers, old drivers, two-wheeled vehicles involved, trucks involved, multiple vehicles, fatigue driving, the use of seat belts/helmets, the time period from night to early morning (20:00–06:59), non-clear weather, two-wheel vehicles involved, visibility less than 50 m, weather conditions, the time period from night to early morning (20:00–06:59), the number of vehicles involved in the accident. Comparing the influence on the severity of traffic accidents in rural areas, 14 factors can be expressed in descending order as follows: old drivers, drunk driving, trucks involved, the time period from night to early morning (20:00–06:59), female drivers, two-wheeled vehicles involved, fatigue driving, use of safety belt/helmet, number of vehicles involved, the evening time period (17:00–19:59), visibility of 50 m–100 m, road surface condition, two-wheeled vehicles involved, number of vehicles involved, visibility less than 50 m, holidays, visibility of 50 m–100 m, road surface condition. Among them, the number of vehicles involved in an accident has a different impact in urban and rural areas. The percentage of severe injuries and death in urban areas is higher in multi-vehicle accidents than in single-vehicle accidents, while the opposite is true in rural areas.
(3) More broadly, this study recognizes the differences between urban and rural roadside accidents to obtain a deeper understanding of the influencing factors for road traffic crashes. The findings can help the Department of Transportation of China to take differentiated and targeted measures to effectively prevent and reduce urban and rural traffic accidents, especially those that are serious or fatal. A high-accuracy predictive model commonly needs large amounts of information. Hence a promising direction for future work would be to obtain a dataset with detailed information such as specific types of two-wheeled vehicles (e.g., motorcycle, motorized cycle, pedal cycle), drivers’ socioeconomic attributes, and environmental characteristics.
This study was supported by SDOT (Shandong Department of Transportation). We thank LetPub for its linguistic assistance during the preparation of this manuscript.