Spatiotemporal Heterogeneity Analysis of Provincial Road Traffic Accidents and Its Influencing Factors in China

Abstract

To objectively evaluate the road traffic safety levels across different provinces in China, this study investigated the spatiotemporal heterogeneity characteristics of macro factors influencing road traffic accidents. Panel data from 31 provinces in China from 2009 to 2021 were collected, and after data preprocessing, traffic accident data were selected as the dependent variables. Population size, economic level, motorization level, highway mileage, unemployment rate, and passenger volume were selected as explanatory variables. Based on the spatiotemporal non-stationarity testing of traffic accident data, three models, namely, ordinary least squares (OLS), geographically weighted regression (GWR), and geographically and temporally weighted regression (GTWR), were constructed for empirical research. The results showed that the spatiotemporal heterogeneity characterizing the macro factors of traffic accidents could not be ignored. In terms of impact effects, highway mileage, population size, motorization level and passenger volume had positive promoting effects on road traffic accidents, while economic level and unemployment rate mainly exhibited negative inhibitory effects. In terms of impact magnitude, highway mileage had the greatest impact on traffic accidents, followed by population size, motorization level, and passenger volume. Comparatively, the impact magnitude of economic level and unemployment rate was relatively small. The conclusions were aimed at contributing to the objective evaluation of road traffic safety levels in different provinces and providing a basis for the formulation of reasonable macro traffic safety planning and management decisions. The findings offer valuable insights that can be used to optimize regional traffic safety policies and strategies, thereby enhancing road safety.

1. Introduction

Road traffic accidents pose a great danger to the safety of people’s lives and property around the world. Globally, approximately 1.3 million people die in road traffic accidents annually [1]. According to data released by the Ministry of Transport in 2021, Germany experienced 2.3 million road traffic accidents, resulting in approximately 2562 deaths; Japan had a total of 305,000 traffic accidents, leading to 2636 deaths; and the United States saw around 43,000 fatalities due to traffic accidents. China, with its vast landmass of 9.6 million square kilometers and a population exceeding 1.4 billion, had 270,000 accidents and 62,218 deaths reported in 2021 [2].

Previous studies have shown that road traffic safety is influenced by a multitude of factors, including geographical, socio-economic, and cultural conditions [3]. In urban environments, road traffic safety not only involves the safety of vehicle drivers but also requires paying attention to all road users, especially pedestrians and cyclists [4]. The complex interactions between vehicles and these more vulnerable road users significantly increase the complexity of traffic safety issues. However, the impact efficacy of these influencing factors is not entirely consistent when examined across different spatial and temporal scales, which limits our understanding of the full complexity of traffic safety dynamics. In large and environmentally diverse regions like China, this complexity is further amplified, posing significant challenges in accurately assessing traffic safety across different provinces. The provinces of China vary widely in geography, economy, and social conditions, making traffic safety a particularly complex issue. Evaluating the traffic safety level of different provinces solely based on the number of road traffic accidents is neither sufficiently scientific nor rigorous. Therefore, selecting universally applicable and objective evaluation indicators to accurately assess the road traffic safety levels across provinces becomes a challenging research task.

The purpose of this study is to explore the traffic safety conditions in various regions of China and to provide a scientific and reliable basis for evaluating provincial traffic safety levels. By selecting a set of evaluation indicators, this study investigates their correlation with road traffic accidents and examines the influence patterns of these indicators across different provinces. The findings contribute to the objective evaluation of road traffic safety levels and inform sound macro-level traffic safety planning and management decisions.

The structure of this paper is as follows. Firstly, a literature review including macro factors and accident analysis methods is presented. Secondly, we describe the data sources and research methods including methods for spatial autocorrelation and spatiotemporal non-stationarity testing, regression model construction, and comparison indicators of model results. Thirdly, comparisons of the results of OLS, GWR, and GTWR models and a spatiotemporal analysis of the regression coefficients of the GTWR model are described. Finally, we present the conclusions of the empirical results.

2. Literature Review

2.1. Macro Factors

Existing research has extensively analyzed the relationship between various influence factors and road traffic accidents. Amoros et al. [5] pointed out that road traffic accidents often resulted from the combined effects of multiple factors. We summarize recent studies, focusing on year, author, location, and factors in these studies, as shown in

Table 1. Summary of macro factors.

Year	Author	Location	Factors
2003	Hadayeghi et al. [6]	Toronto	Resident population, Road kilometers
2010	Huang et al. [7]	Florida	Population density, Economic development
2014	Qiu Chenlu et al. [8]	China	GDP, Population, Motor vehicles, Highway mileage, Passenger volume, Freight volume
2017	Pirdavani et al. [9]	Flanders	Socio-economic characteristics
2019	Sun et al. [10]	China	Economic, Road mileage, Vehicles, Population
2020	Akinyemi et al. [11]	Nigeria	Per capita GDP, Unemployment rate
2021	Taruwere Yakubu et al. [12]	Nigeria	Unemployment rate
2023	Forrest et al. [13]	Greater London	Population, Average number of cars per household

. Hadayeghi et al. [6] found a significant positive correlation between resident population, road kilometers, and traffic accidents. Huang et al. [7] analyzed motor vehicle collision traffic accident data and found that population density had a positive impact on traffic accidents, while the level of economic development had a negative impact on the traffic accident rate. Qiu Chenlu et al. [8] used multiple linear regression analyses to analyze six influencing factors: GDP, population, number of motor vehicles, highway mileage, highway passenger volume, and highway freight volume. They found that GDP was negatively correlated with the number of accidents, whereas the population size was positively correlated, further confirming the significant impacts of GDP and population data on traffic accidents. Consistent with the results of Huang et al., Pirdavani et al. [9] found a negative correlation between socio-economic characteristics and the severity of traffic accidents, implying that higher socio-economic levels contributed to reducing casualties in traffic accidents. Sun et al. [10] discovered through research that economic development can enhance road traffic safety to a certain extent, whereas the increase in road mileage, new vehicles, and population growth contributes to a rise in road traffic casualties. Akinyemi et al. [11] conducted a correlation analysis of per capita gross domestic product (GDP), unemployment rate, and annual road traffic accident data. They found that in the long term, as per capita GDP increased, the number of accidents and fatalities decreased, while the number of injuries increased. However, there was no statistically significant short-term relationship among per capita GDP, unemployment rate, and traffic accidents. In contrast to the findings of Akinyemi, Taruwere Yakubu et al. [12] found that an increase in the unemployment rate increased the number of road traffic accidents in all geopolitical zones of Nigeria, demonstrating the correlation between the unemployment rate and traffic accidents. Forrest et al. [13] discovered that the increase in population was associated with pedestrian accidents, while the influence of the average number of cars per household on pedestrian accidents was relatively small. The unemployment rate was also considered an important factor influencing the number of traffic accidents.

2.2. Methods of Accident Analysis

Road traffic accidents exhibited randomness and uncertainty [14]. Quantitative models have been widely applied in the analysis of traffic accident data [15]. As shown in Table 2, various models have been used to explore the association between influencing factors and the occurrence of traffic accidents [16]. Current research suggests that the effects of explanatory variables are not fixed and may vary due to the unobserved factors. Many studies have proposed stochastic parameter models or logistic regression models to explain the unobserved heterogeneity (i.e., a variable having different parameters in different observations), thereby enhancing the predictive ability of the model [17]. Chen et al. [18] used a polynomial logit model to investigate and identify important influencing factors for the severity of rear-end collision injuries. Sun et al. [19] analyzed three collision accident datasets using the RP-logit model, identifying significant factors and heterogeneity across these datasets. Wu et al. [20] developed a mixed logit model to analyze the severity of driver injuries in single and multiple vehicle collisions on rural two-lane highways in New Mexico from 2010 to 2011. Yu et al. [21] used data collected in North Carolina and employed a random parameter logit method to investigate factors affecting the severity of rear-end collisions in work zones. Although stochastic parameter or mixed logit models were helpful in capturing some unobserved heterogeneity, they failed to capture the spatial characteristics of the data. Therefore, the geographically weighted regression (GWR) model, with spatial heterogeneity considered, had been widely applied.

The GWR model, proposed by Brunsdon et al. [22] in 1986, could generate a regression model describing local relationships at each location in the study area, thus explaining the local spatial relationships and spatial heterogeneity of explanatory variables. Accidents in a region often exhibited spatial heterogeneity and were unevenly distributed [23]. Wachnicka et al. [24] analyzed differences in road safety levels among regions based on the GWR model, finding that this model had advantages over the OLS model. Rahman et al. [25] used GWR analysis to reveal the spatial relationship between traffic accidents and population density, linking it to land use in each community, and found that traffic accidents mainly occurred near residential areas and public facilities. Al-Hasani et al. [26] applied the GWPR model to road traffic accident (RTA) data in Oman and found that the model could better capture spatial heterogeneity and influencing factors. Huang et al. [27] used the GWR model to explore the relationship between traffic accidents and built environment in the Detroit area of Michigan, finding that the relationship between built environment and accidents was non-stationary in space and varied in strength and direction with spatial changes.

However, the GWR model did not consider the impact of temporal changes on traffic accidents. In addition to spatial heterogeneity, traffic accidents also exhibited temporal heterogeneity [28]. Therefore, the geographically and temporally weighted regression (GTWR) model, with both spatial and temporal heterogeneity considered, were applied in traffic accident analysis.

Building upon the GWR method, Huang [29] proposed the GTWR model to capture spatial non-stationarity. In terms of traffic accident analysis, Mohammadnazar et al. [16] integrated data from multi-lane rural road segments from 2013 to 2017 and used the GTWR model to find significant differences in most variables in space and time. Liu et al. [30] used the GTWR model to explore pedestrian accident data in the metropolitan area of Hong Kong from 2008 to 2012, finding relationships and spatiotemporal heterogeneity between pedestrian injury severity and influencing factors. Cheng et al. [31] analyzed the influencing factors of traffic safety using OLS, GWR, and GTWR models, finding that the GTWR model had the best modeling effect on spatiotemporal heterogeneous data, and factors such as traffic condition information, road network structure, and existing environment affected the occurrence of traffic accidents significantly. Therefore, it was necessary to apply the GTWR model, which could explain both temporal and spatial heterogeneity, as the ignorance of these attributes may lead to unobserved temporal heterogeneity and lack of accuracy.

Table 2. Summary of methods of accident analysis.

Year	Author	Model
1986	Brunsdon et al. [22]	GWR model
2010	Huang [29]	GTWR model
2014	Wu et al. [20]	Mixed logit model
2015	Chen et al. [18]	Polynomial logit model
2018	Huang et al. [27]	GWR model
2020	Yu et al. [21]	Random parameter logit method
2020	Rahman et al. [25]	GWR model
2021	Wachnicka et al. [24]	GWR model
2021	Al-Hasani et al. [26]	GWPR model
2021	Mohammadnazar et al. [16]	GTWR model
2022	Sun et al. [19]	RP-logit model
2023	Cheng et al. [31]	OLS, GWR, GTWR models
2023	Liu et al. [30]	GTWR model

Table 2. Summary of methods of accident analysis.

Year	Author	Model
1986	Brunsdon et al. [22]	GWR model
2010	Huang [29]	GTWR model
2014	Wu et al. [20]	Mixed logit model
2015	Chen et al. [18]	Polynomial logit model
2018	Huang et al. [27]	GWR model
2020	Yu et al. [21]	Random parameter logit method
2020	Rahman et al. [25]	GWR model
2021	Wachnicka et al. [24]	GWR model
2021	Al-Hasani et al. [26]	GWPR model
2021	Mohammadnazar et al. [16]	GTWR model
2022	Sun et al. [19]	RP-logit model
2023	Cheng et al. [31]	OLS, GWR, GTWR models
2023	Liu et al. [30]	GTWR model

In summary, the academic community has conducted extensive and comprehensive research on the influencing factors of traffic accidents, revealing that factors such as socio-economic characteristics [7,8,9,10,11,13], population [6,7,8,10], and unemployment rate [11,12] have a significant impact on traffic accidents. However, it is worth noting that when the spatial and temporal scales of traffic accidents studies differ, the effects of different influencing factors on traffic accidents may not be entirely consistent. For example, while some studies have found a positive correlation between the unemployment rate and the number of road traffic accidents [12], others have pointed out that there is no statistically significant correlation between the unemployment rate and traffic accidents in the short term [11]. In addition, China has a vast land area and diverse geographical environments, and significant differences exist in population and area, as well as social, economic, and cultural factors among different provinces in China. The geographical and socio-economic differences among different provinces in China may lead to differences in traffic safety levels. However, existing research has mostly focused on independent cities, lacking research on the spatiotemporal heterogeneity of traffic accident influencing factors at the provincial level in China. Previous studies have shown that it is crucial to consider spatial and temporal heterogeneity in traffic accidents analysis [16,29,30,31]. Therefore, this study applies spatial econometric methods to identify key factors influencing traffic accidents, deeply analyze the spatiotemporal evolution characteristics of traffic accident influencing factors in different provinces of China, provide a scientific basis for provincial traffic safety level evaluation research, and provide important insights for government and decision-makers to improve traffic safety.

3. Methods and Data

3.1. Data Source and Preprocessing

3.1.1. Data Source

This study selected panel data from 31 provinces and municipalities in China (excluding Hong Kong, Macau, and Taiwan) from 2009 to 2021. The number of traffic accidents at the provincial level was chosen as the dependent variable. The causes of road traffic accidents are generally categorized into three groups: human factors, mechanical factors related to the vehicle, and environmental factors and road conditions [3,5,6,7,8,9,10,11,12,13]. Based on previous research findings [9,13], as shown in

Table 3. Description of macro factors influencing road traffic accidents at provincial level in China.

Indicator	Measurement Methods
Population size	Population/Provincial area
Economic level	GDP/Population
Motorization level	Number of motor vehicles/Population
Technological level	Number of patent applications granted
Highway mileage	Highway mileage
Government intervention	Local financial expenditure on transportation/GDP
Freight volume	Weight of goods transported by various means of transport during a certain period
Higher education level	Number of people with college education or above/Population
Unemployment rate	Urban registered unemployment rate
Passenger volume	Number of passengers transported by various means of transport during a certain period

, initially, 10 macro factors influencing traffic accidents were selected as explanatory variables. Among them, population data, gross domestic product (GDP), number of motor vehicles, number of patent applications granted, highway mileage, local financial expenditure on transportation, freight volume, passenger volume, urban registered unemployment rate, and the number of traffic accidents were sourced from the National Bureau of Statistics website. Provincial area data were obtained from the statistical yearbooks and bulletins of each province and municipality. The seven major geographic divisions of China are shown in Figure 1. All data were collected and preprocessed based on provincial boundaries as spatial units.

3.1.2. Data Preprocessing

First, we considered the presence of multicollinearity when two or more explanatory variables were highly correlated, which can lead to inaccurate estimation of the individual effects on the dependent variable [33]. Therefore, a multicollinearity test and stepwise regression were conducted using STATA software (Stata 17.0), and factors with a VIF greater than 10 were removed. Second, given the significant disparity in the magnitude of data for factors influencing traffic accidents, logarithmic transformation was applied to the data before empirical analysis [34] to mitigate the impact of outliers on the empirical results.

3.2. Research Methodology

3.2.1. Spatial Autocorrelation

Spatial autocorrelation analysis refers to the assessment of the similarity between a point in space and its neighboring points. Moran’s I is a commonly used method for spatial autocorrelation analysis, which can be employed to evaluate the spatial autocorrelation of explanatory variables [16,30]. The mathematical expression for Moran’s test is as follows:

$$\begin{array}{c}\hfill I={\displaystyle \frac{n}{{\displaystyle \sum _{p=1}^n}{\displaystyle \sum _{q=1}^n}{w}_{pq}}}{\displaystyle \frac{{\displaystyle \sum _{p=1}^n}{\displaystyle \sum _{q=1}^n}{w}_{pq}\left(x_p-\overline{x}\right)\left(x_q-\overline{x}\right)}{{\displaystyle \sum _{p=1}^n}{\left(x_p-\overline{x}\right)}^2}},\end{array}$$

(1)

where I represents the Moran’s index, $w_{pq}$ the spatial weight between positions p and q, n the total number of spatial elements, $x_p$ and $x_q$ the values at locations p and q, respectively, and $-x$ the mean of all observed values. The range of I is from greater than or equal to $-1$ to less than or equal to 1. A value greater than 0 implies spatial clustering, while a value less than 0 indicates spatial dispersion. Values approaching 0 suggest spatial randomness. Additionally, the Z-score is commonly used as a significance indicator for Moran’s test. The formula is calculated as follows:

$$\begin{array}{c}\hfill Z\left(I\right)={\displaystyle \frac{I-E\left(I\right)}{\sqrt{V\left(I\right)}}},\end{array}$$

(2)

where $E\left(I\right)$ is the expected value of Moran’s I index under the null hypothesis and $V\left(I\right)$ is the variance of Moran’s I index under the null hypothesis.

3.2.2. Spatiotemporal Non-Stationarity Test

The spatiotemporal non-stationarity test aims to take hypothesis testing for the changes of the model and its regression parameters over time and space, thus evaluating the non-stationarity of spatiotemporal data. The commonly used spatiotemporal non-stationarity test method is based on the results of Monte Carlo simulations, comparing the quartiles of the geographically and temporally weighted regression (GTWR) model with doubled standard error of the ordinary least squares (OLS) model [35]. If the quartiles of the GTWR model are greater than twice the standard error of the OLS model, it indicates that the model parameters may vary significantly at different time and spatial locations, suggesting significant spatiotemporal non-stationarity. Conversely, if the quartiles of the GTWR model are less than or equal to the doubled standard error of the OLS model, it indicates that the model parameters are relatively stable over time and space, with no significant spatiotemporal non-stationarity.

3.2.3. Build Models

Three models, namely OLS, GWR, and GTWR, were constructed for empirical analysis, and the fitting effects of the models were compared and verified [31].

(1)

OLS Model

The OLS model is a classical regression analysis method commonly used to estimate the parameters of linear regression models. The model expression is as follows:

$$\begin{array}{c}\hfill y_i={\beta }_0+\sum _{i=1}{\beta }_k{X}_{ik}+{\epsilon }_i,\end{array}$$

(3)

where i represents the ith sample point, $y_i$ represents the accident data value of the ith sample point, $X_{ik}$ represents the kth traffic accident influencing factor of the ith sample point, ${\epsilon }_i$ is the random error term, ${\beta }_0$ represents the regression constant of the regression equation, and ${\beta }_k$ represents the regression coefficient of the kth traffic accident influencing factor.

(2)

GWR Model

Compared to the OLS model, which does not reflect the spatial heterogeneity of influencing factors, the parameters of independent variables in the GWR model vary with geographical locations, reflecting spatial heterogeneity. The model expression is as follows:

$$\begin{array}{c}\hfill y_i={\beta }_0\left({\mu }_i,v_i\right)+\sum _k{\beta }_k\left({\mu }_i,v_i\right)X_{ik}+{\epsilon }_i,\end{array}$$

(4)

where $({\mu }_i,v_i)$ represents the latitude and longitude coordinates of the ith area, $y_i$ represents the accident data value of the ith sample point, $X_{ik}$ represents the kth traffic accident influencing factor of the ith sample point. ${\epsilon }_i$ is the model error term, ${\beta }_0({\mu }_i,v_i)$ represents the regression constant of the ith sample point, and ${\beta }_k({\mu }_i,v_i)$ represents the regression coefficient of the kth traffic accident influencing factor of the ith sample point.

(3)

GTWR Model

Compared to the GWR model, the parameters of independent variables in the GTWR model vary with both time and space, thus reflecting spatial heterogeneity while considering temporal heterogeneity [36]. China has a vast land area and diverse geographical environments, and there are significant differences in population, society, economy, culture, and area among different provinces in China. Therefore, at different times, the influencing factors of traffic accidents in different provinces may have temporal and spatial differences. The model expression is as follows:

$$\begin{array}{c}\hfill y_i={\beta }_0\left({\mu }_i,v_i,t_i\right)+\sum _k{\beta }_k\left({\mu }_i,v_i,t_i\right)X_{ik}+{\epsilon }_i,\end{array}$$

(5)

where $({\mu }_i,v_i)$ represents the latitude and longitude coordinates of the ith area, $t_i$ represents the observation time, which refers to the year in this context, $y_i$ represents the dependent variable value of the ith sample point, $X_{ik}$ represents the kth traffic accident influencing factor of the ith sample point, ${\epsilon }_i$ is the model error term, ${\beta }_0({\mu }_i,v_i,t_i)$ represents the regression constant of the ith sample point, and ${\beta }_k({\mu }_i,v_i,t_i)$ represents the regression coefficient of the kth traffic accident influencing factor of the ith sample point.

The model expression is as follows:

$$\begin{array}{c}\hfill \widehat{\beta }\left({\mu }_i,v_i,t_i\right)={\left[X^{T}W\left({\mu }_i,v_i,t_i\right)X\right]}^{-1}{X}^{T}W\left({\mu }_i,v_i,t_i\right)Y,\end{array}$$

(6)

where $W({\mu }_i,v_i,t_i)$ represents the weight of the spatiotemporal position i. The GTWR model determines the influence of other sample points’ values on the regression sample points by constructing a spatiotemporal weight matrix; thus, the spatiotemporal weight matrix plays a core role in the calculation process of the GTWR model, which takes the form of a diagonal matrix, denoted as $W\left({\mu }_i,v_i,t_i\right)=diag\left[W\left({\mu }_1,v_1,t_1\right),W\left({\mu }_2,v_2,t_2\right),\dots ,W\left({\mu }_i,v_i,t_i\right)\right]$. The expression for the space-time distance calculation is as follows:

$$\begin{array}{c}\hfill {\left(d^{ST}\right)}^2={\gamma \left(d^S\right)}^2+{u\left(d^T\right)}^2,\end{array}$$

(7)

where $d^{ST}$ represents spatiotemporal distance, $d^T$ represents time distance, and $d^S$ represents spatial distance.

The weight function is generally chosen as a Gaussian function or a bi-square function. After substitution, the weight function can be transformed, and the weight matrix can be calculated as follows:

$$\begin{array}{c}\hfill W_{ij}=exp\left\{-{\displaystyle \frac{\left[{\left(u_i-u_j\right)}^2+{\left(v_i-v_j\right)}^2\right]+{\tau \left(t_i-t_j\right)}^2}{{\left(h^{ST}\right)}^2}}\right\},\end{array}$$

(8)

$$\begin{array}{c}\hfill =exp\left\{-{\displaystyle \frac{\left[{\left(u_i-u_j\right)}^2+{\left(v_i-v_j\right)}^2\right]}{{\left(h^S\right)}^2}}-{\displaystyle \frac{{\left(t_i-t_j\right)}^2}{{\left(h^T\right)}^2}}\right\},\end{array}$$

(9)

where $h^S$ represents spatial bandwidth, $h^T$ represents time bandwidth, and $h^{ST}$ represents spatiotemporal bandwidth. $\tau$ represents ${\displaystyle \frac{u}{\gamma }}$. $h^{ST}$ is linearly related to $h^S$ and $h^T$, respectively, as shown below as follows:

$$\begin{array}{c}\hfill {\left(h^{ST}\right)}^2={\gamma \left(h^S\right)}^2,\end{array}$$

(10)

$$\begin{array}{c}\hfill {\left(h^{ST}\right)}^2={u\left(h^T\right)}^2.\end{array}$$

(11)

(4)

Model Comparison

The residual sum of squares (RSS), the corrected Akaike information criterion (AICc), and the coefficient of determination ($R^2$) are commonly used metrics in final model decision making. In statistics, the difference between data points and their corresponding positions on the regression line is called residual. RSS represents the influence of random errors. AICc measures the complexity and goodness of fit of the model. $R^2$ represents the ability of the response variable to be explained by the predictor variables in the regression model. The value of $R^2$ is between 0 and 1. The closer it is to 1, the better the goodness of fit.

Regarding diagnostic indicators, higher values of $R^2$, smaller values of AICc, and smaller values of RSS are generally considered indicators of better model fit [31].

4. Results and Discussion

4.1. Model Calculation and Related Tests

4.1.1. The Result of Moran’s I

Through multicollinearity testing [33], population size, economic level, motorization level, highway mileage, unemployment rate, and passenger volume were ultimately selected as the six influencing factors.

Table 4. Global Moran’s I index results.

Variable	Moran’s I	Z-Value	p-Value	VIF
Population size	0.611	5.802	0.001	3.870
Economic level	0.415	3.931	0.001	5.120
Motorization level	0.406	3.647	0.001	3.900
Highway mileage	0.191	2.001	0.050	3.350
Unemployment rate	0.011	0.390	0.594	1.260
Passenger volume	0.304	3.165	0.002	4.670

shows the global Moran’s I index results for each variable. Except for the unemployment rate, the other variables were statistically significant at the $1\%$ level, indicating that most variables had spatial autocorrelation. Therefore, conducting an analysis of traffic accident influencing factors from a spatial perspective is closer to the actual situation, thus avoiding biases that may result from ignoring spatial interaction relationships among variables.

4.1.2. The Results of the Non-Stationarity Test

The results of the non-stationarity test are shown in

Table 5. Spatiotemporal non-stationarity test for variables.

Variable	Interquartile (GTWR)	2SE (OLS)	Extra Local Variation
Population size	0.549	0.102	Yes
Economic level	0.287	0.117	Yes
Motorization level	0.493	0.102	Yes
Highway mileage	0.284	0.095	Yes
Unemployment rate	0.126	0.112	Yes
Passenger volume	0.401	0.052	Yes

. The second column represents the quartiles (i.e., the difference between the upper quartile and the lower quartile). The third column indicates two times the standard error of the OLS model. The fourth column indicates whether a variable has additional local variation. All variables exhibited additional variation, indicating the presence of non-stationarity in both time and space. This suggests that better yields can be achieved by using the GTWR model to analyze the spatiotemporal heterogeneity of factors influencing traffic accidents [35], allowing for a more comprehensive consideration of the evolving trends of the data.

4.1.3. Model Comparison Results

As shown in

Table 6. Comparative analysis results of models.

Evaluation Metrics	OLS	GWR	GTWR
R2	0.664	0.840	0.905
AICc	627.493	431.59	341.247
RSS	108.134	51.709	30.629

, the $R^2$ of the GTWR model is 0.065 higher than that of the GWR model and 0.241 higher than that of the OLS model. Additionally, both the AICc and RSS values of the GTWR model are smaller than those of the GWR and OLS models, indicating that the GTWR model can better explain the influence of independent variables on traffic accidents and account for the spatiotemporal heterogeneity of factors influencing traffic accidents.

4.1.4. Analysis of GTWR Model Results

The descriptive statistics of the GTWR model’s fitted coefficients are shown in

Table 7. Descriptive statistics of regression coefficients for impact factors.

Variable	MIN	LQ	UQ	MAX	AVG	Significance Level at 10%	Significance Level at 5%	Significance Level at 1%
Population size	−2.173	−0.172	0.377	3.276	0.217	92.060%	89.081%	84.864%
Economic level	−5.120	−0.025	0.262	1.156	−0.021	90.074%	88.089%	84.864%
Motorization level	−1.270	−0.134	0.359	2.721	0.167	87.841%	84.367%	78.908%
Highway mileage	−1.589	0.243	0.528	5.827	0.447	92.556%	92.060%	89.826%
Unemployment rate	−0.406	−0.098	0.028	0.400	−0.028	77.171%	75.186%	67.742%
Passenger volume	−0.425	0.085	0.486	0.871	0.284	85.608%	81.886%	78.164%

MIN = Minimum value of the variable; LQ = Lower quartile (25th percentile); UQ = Upper quartile (75th percentile); MAX = Maximum value of the variable; AVG = Average value of the variable.

, providing a more intuitive presentation of the temporal and spatial effects of various macroscopic influencing factors on traffic accidents from 2009 to 2021 across provinces. The descriptive statistics of the fitted coefficients adopted the methods proposed by Li Yingli [37] and Lv Yanqin et al. [38] to examine the significance of each explanatory variable. According to the data in the table, the proportions of significance at the $10\%$, $5\%$, and $1\%$ levels for each explanatory variable all exceed $50\%$, indicating that each explanatory variable had a strong explanatory power in the model under Gaussian weights.

Moreover, the signs and magnitudes of the fitted coefficients provided information about the direction and extent of the influence of the independent variables on the dependent variable. When the fitted coefficient of an independent variable is positive, it indicates a positive correlation between the increase in the independent variable and the increase in the dependent variable. The larger the absolute value of the fitted coefficient, the greater the impact of this positive correlation. Conversely, when the fitted coefficient of an independent variable is negative, it indicates a negative correlation between the increase in the independent variable and the decrease in the dependent variable. Observing the minimum and maximum values of the regression coefficients, there is significant variability in the coefficients of factors affecting traffic accidents across space and time, indicating significant spatiotemporal instability and heterogeneity.

Upon further examination of the impact effects of highway mileage, population size, motorization level, and passenger volume, it was found that there were positive promoting effects on the occurrence of road traffic accidents and negative inhibitory effects of economic level and unemployment rate. This finding differed slightly from the results of Akinyemi [11] and Taruwere Yakubu [12]. In terms of the degree of influence, highway mileage has the greatest impact on the occurrence of traffic accidents, followed by population size, motorization level, and passenger volume. In comparison, the influence of economic level and unemployment rate is relatively smaller. Therefore, different regions have different social, economic, and cultural environments, leading to significant differences in the correlation of factors influencing road traffic accidents.

4.2. Spatiotemporal Heterogeneity Analysis of Impact Factors

4.2.1. Temporal Heterogeneity Analysis of Impact Factors

Based on the regression results of the GTWR model, in order to further reveal the temporal heterogeneity of the factors influencing traffic accidents, this paper drew boxplots of the fitted coefficients of each influencing factor from 2009 to 2021, as shown in Figure 2. These boxplots illustrated the contribution coefficients of each factor to traffic accidents from 2009 to 2021.

(1) From 2009 to 2021, the regression coefficient of population size of decreasing initially and increasing gradually, the average regression coefficient first decreased from 0.195 in 2009 to 0.141 in 2015, then grew to 0.312 in 2021, indicating significant temporal variation. Additionally, there was considerable variability in the dispersion across different years, confirming the temporal heterogeneity of population size. However, overall, population size showed a positive effect on traffic accidents. This is because provinces with larger populations typically have more traffic activities, including vehicles, pedestrians, and other participants, thereby increasing the probability of traffic accidents.

(2) From 2009 to 2021, the regression coefficients of economic level were both positive and negative, experiencing the process of positive to negative and decreasing year by year, from 0.165 in 2009 to −0.169 in 2021 specifically, of which there was a positive correlation from 2009 to 2015 and a negative correlation from 2015 onwards, with obvious differences. Additionally, there was considerable variability in dispersion across the years from 2009 to 2021, confirming the temporal heterogeneity of economic level. The improvement of economic level is often associated with an increase in people’s education level and social quality, leading to greater adherence to traffic rules by drivers. Moreover, governments typically allocate more resources to improve transportation infrastructure, which may contribute to reducing the occurrence of traffic accidents.

(3) From 2009 to 2021, the average regression coefficient of motorization level underwent a process of being initially negative and then positive, first decreasing and then increasing, with the significant exhibition of temporal heterogeneity. The average regression coefficient slightly decreased from −0.017 to −0.049 from 2009 to 2011 and then increased from −0.027 to 0.526 from 2012 to 2021. Additionally, there was considerable variability in dispersion across the years from 2009 to 2021, confirming the temporal heterogeneity of motorization level. The increase in motorization level leads to an increase in traffic volume, highlighting issues such as traffic congestion, and possibly putting significant pressure on transportation infrastructure, thereby increasing the occurrence rate of traffic accidents.

(4) The impact of highway mileage on traffic accidents showed relatively stable changes from 2009 to 2021, with the average regression coefficient being positive. However, there was considerable variability in dispersion across different years, indicating the temporal heterogeneity of road mileage. An increase in highway mileage results in longer distances traveled by vehicles, thereby increasing the risk of accidents. It also encourages more vehicles to travel on roads, thereby increasing traffic volume and consequently, the probability of accidents.

(5) From 2009 to 2021, the average regression coefficient of the unemployment rate showed both positive and negative distributions, with a small decrease from 2009 to 2016 from 0.025 to −0.072, and an increase from 2017 to 2021 from −0.065 to 0.021, with temporal heterogeneity. Additionally, from the perspective of dispersion, it confirmed temporal heterogeneity. This is because an increase in the unemployment rate implies a decrease in commuting numbers and driving frequency, which may lead to a reduction in traffic accidents.

(6) Similar to highway mileage, the average regression coefficient for passenger trips was essentially positive over the study period, and the positive effect increased and then decreased over time, with the average regression coefficient increasing from 0.306 in 2009 to 0.433 in 2014 before decreasing to 0.080 in 2021, with significant temporal variability. Additionally, from the perspective of dispersion, the impact varies across provinces. An increase in passenger volume implied more passenger vehicles, passenger numbers and longer operating times, which may increase the incidence rate of traffic accidents. However, with the improvement of traffic safety awareness and traffic management level, some of the increased risks of traffic accidents resulting from the increase in passenger volume may be mitigated.

4.2.2. Spatial Heterogeneity Analysis of Influencing Factors

According to the natural breakpoint grading principle in ArcGIS 10.2, the average regression coefficients of each province from 2009 to 2021 were calculated. The spatial distribution map of the influencing factors is plotted in Figure 3.

(1) The regression coefficients of population size cover a wide range, distributed between −0.029 and 0.639, showing significant spatial heterogeneity. Provinces with negative regression coefficients are mainly located in the East China region. This may be due to stricter traffic management, better traffic facilities, and higher traffic civilization in densely populated areas, thereby reducing the occurrence of traffic accidents.

(2) The impact of economic level on traffic accidents did not exhibit consistent positive or negative correlations. The distribution of regression coefficients covers a wide range, distributed between −0.069 and 0.407, indicating spatial heterogeneity. Provinces with negative coefficients are mainly located in North China and Northwest China. North China has higher levels of traffic planning, road construction, and traffic law enforcement, which contribute to the reduction of traffic accidents and improved road safety. However, the traffic infrastructure and safety facilities in Northwest China, once economically lagging behind the other regions, are relatively backward. With economic development, investments in traffic infrastructure and safety facilities are increasing to alleviate traffic pressure and reduce the incidence of traffic accidents.

(3) The spatial heterogeneity of the regression coefficients of motorization level was evident, distributed between −0.281 and 0.267. There were relatively higher regression coefficients in provinces in North China and Northeast China, ranging from above 0.129 to below 0.267. In North China, although there is good infrastructure construction and traffic management, an increase in motorization levels still leads to increased traffic flow and pressure, thereby increasing the probability of traffic accidents. In Northeast China, adverse weather conditions (snow, sandstorms, and haze) often increase the risk of road travel, potentially increasing the probability of driving errors and hazardous driving actions, making accidents more likely to occur [39].

(4) Highway mileage had a positive impact on traffic accidents, especially in the Southwest, Central, and South China regions, with regression coefficients reaching above 0.524 and below 0.570. These southern regions have diverse geographical environments, with mountains in Southwest China, relatively flat terrain in Central China, and hills, plains, and coastal areas in South China. As highway mileage increases, the distance traveled by vehicles increases, leading to increased traffic flow and the risk of accidents. Particularly in mountainous and plateau areas, terrain and climate conditions limit drivers’ visibility and operations, further increasing the risk of traffic accidents.

(5) The regression coefficients of unemployment rates varied in different provinces, ranging from −0.032 to 0.111, with a high positive correlation observed in Central China and South China. In economically developed regions like South China, high unemployment rates often have a significant impact on people’s mental health [40], causing them to be more susceptible to fatigue, anxiety, distraction, and even errors in judgment while driving, which in turn significantly increases the risk of traffic accidents. In Central China, located in the convergence zone of the middle and lower reaches of the Yellow River and the middle reaches of the Yangtze River, with a relatively complete industrial system, an increase in unemployment rates may lead to more people seeking livelihoods, thereby increasing travel demands and the probability of traffic accidents.

(6) The regression coefficients of passenger volume in each province were positive, with high values mainly concentrated in Central China and East China, and relatively high values distributed in North China and South China. These regions are located within the area east of the “Hu Line”, which is densely populated and industrialized, with higher demands for passenger transportation. An increase in passenger volume leads to more passengers, vehicles, and longer operating times, thereby increasing the risk of traffic accidents. The widespread distribution of regression coefficients for passenger volume confirmed the significant spatial heterogeneity of this explanatory variable.

4.2.3. Managerial Implications

Given the vast geographical, social, economic, and cultural differences among provinces, a one-size-fits-all approach to traffic safety evaluation is insufficient. This study offers significant insights for policymakers in addressing the complex issue of traffic safety across China’s diverse provinces.

Firstly, the study reveals the temporal and spatial law of the impact of macro indicators such as highway mileage, population, motorization level, passenger volume, economic level, and unemployment rate on traffic accidents. For example, the effect of economic level on traffic accidents was positive from 2009 to 2015, but has gradually turned negative since 2015, indicating that its effect is not static. This finding provides a reference for the economic development of other countries or regions in different stages of development and helps to clarify the key weights of macro indicators on the impact of traffic accidents in different stages of economic development.

Secondly, in a country as vast as China, the complexity of traffic safety issues increases as a result of the different geographical, social, economic, and cultural environments in different provinces. Evaluating traffic safety solely on the basis of the number of road traffic accidents is insufficient and lacks scientific validity. Therefore, decision makers need to take a more nuanced approach and select evaluation indicators that are universally applicable and objective. It provides a basis for decision makers to develop area-specific strategies.

Finally, given the complex and varied factors affecting traffic accidents in the specific process of management and construction, we should comprehensively consider the relevance of macro indicators and take corresponding preventive measures. For example, when carrying out road construction, we should not only pay attention to the potential impact of road mileage on traffic accidents, but also combine factors such as population and economy, and improve public safety awareness through government propaganda and other means to reduce the possibility of accidents during construction. This predictive management method can effectively reduce potential risks and improve the overall level of traffic safety.

4.2.4. Limitations and Future Directions

Although some progress has been made in this study, its limitations should not be overlooked. Firstly, the current research has not elucidated the underlying mechanisms as to how influencing factors such as population size, economic level, motorization level, highway mileage, unemployment rate, and passenger volume interact with each other to affect traffic accidents. Therefore, future research will employ principal component analysis to extract more influencing factors and analyze the comprehensive effects among different factors. Secondly, the current study mainly focuses on the macro-scale traffic safety level of provincial regions, with insufficient attention to micro-level details. Future research could further refine the analysis of the spatial regions of cities or counties to reveal the differences in traffic safety among regions and their underlying reasons more specifically. Lastly, future research should expand to different analytical models, such as nonlinear models, to reveal the complex relationships between traffic accidents and influencing factors more deeply.

5. Conclusions

(1) This study utilized Moran’s I test to analyze six influencing factors: population size, economic level, motorization level, highway mileage, unemployment rate, and passenger volume. The results indicated that the spatial distribution of most variables exhibited significant spatial heterogeneity.

(2) While studying the influencing factors of road traffic accidents, the construction of the GTWR model had better fitting effects and accuracy compared with the OLS and GWR models. The GTWR model was more suitable for the analysis of provincial-level road traffic safety.

(3) The results of the GTWR model showed that the macroscopic influencing factors of traffic accidents exhibited significant spatiotemporal heterogeneity. Regarding the impact effects, highway mileage, population size, motorization level, and passenger volume showed positive promotion effects on road traffic accidents, while economic level and unemployment rate exhibited negative inhibitory effects. In terms of impact magnitude, highway mileage had the greatest impact on the occurrence of traffic accidents, followed by population size, motorization level, and passenger volume. In comparison, the impact magnitude of economic level and unemployment rate was relatively smaller.

(4) Observing from the dimensions of space and time, respectively, from 2009 to 2021, the influence of various factors on traffic accidents showed significant temporal heterogeneity. The regression coefficient of population size decreased first and then increased, indicating temporal differences. The influence of economic level on traffic accidents gradually changed from positive to negative and weakened year by year. The average regression coefficient of motorization level fluctuated, showing its volatility in influence. The dispersion of highway mileage varied greatly in different years, reflecting its temporal heterogeneity. The average regression coefficient of unemployment rate varied positively and negatively, while the average regression coefficient of passenger volume increased first and then decreased. The influence of each factor on traffic accidents exhibited significant spatial heterogeneity. The regression coefficient of population size fluctuated between −0.029 and 0.639, showing significant spatial differences. The influence of economic level also exhibited varying spatial characteristics. Motorization level had higher positive correlations in North China and Northeast China. The positive impact of highway mileage on traffic accidents was particularly significant in Southwest, Central, and South China. The influence of unemployment rate varied positively and negatively in different regions, with Central China and South China being particularly prominent. Passenger volume was positively correlated with traffic accidents, with high values mainly concentrated in Central and Eastern China, and also significant distribution in North China and South China.