Obsidian/Zotero/001_artiklid/Road Safety Indicators/liexploringimpactbuiltenvironment2025.md

---

category: literaturenote citekey: liexploringimpactbuiltenvironment2025 title: Exploring the impact of built environment on crash risks at transportation hubs authors: "Li, Chuanyao; Chen, Li" year: 2025 date: 2025-00-00 2025 doi: 10.1016/j.aap.2025.108079 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2025.108079" zotero_key: C5P7NFTK zotero_storage: E2W6TAR9 collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Li, Chuanyao"

status: converted

Contents lists available at ScienceDirect

Accident Analysis and Prevention

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

Exploring the impact of built environment on crash risks at transportation hubs

Chuanyao Li * , Li Chen

School of Traffic and Transportation Engineering, Central South University, Changsha 410075, China

ARTICLE INFO

Keywords: Transportation hub Built environment Crash risk Geographically weighted regression

ABSTRACT

This study investigates the impact mechanism of the built environment surrounding transportation hubs on crash risks (CR). Three buffer zones (300 m, 500 m, and 800 m) are defined as the spatial analysis units, and Ordinary Least Squares (OLS), Geographically Weighted Regression (GWR), and Multiscale Geographically Weighted Regression (MGWR) are utilized in this study. The results reveals that the 800 m buffer zone provides deeper insights into the factors affecting CR related to the built environment surrounding transportation hubs. Additionally, MGWR demonstrates superior performance in explaining the built environment's impact on CR compared to the other two methods, with an explanation rate of 83.7 %. To reduce CR near transportation hubs, rationally planning the surrounding land use layout and reducing population density per unit area are recommended. Moreover, the density of road networks surrounding airports and railway stations should be kept at a lower level to reduce CR. The findings of this study contribute to a deeper understanding of the relationship between the built environment surrounding transportation hubs and crashes, providing planning guidance and creating a friendly environment surrounding transportation hubs.

1. Introduction

The transportation hubs, including airports, railway stations, and bus stations, serve as vital nodes within the urban transportation system. They facilitate the connection between various modes of transport and the movement of passengers and goods (Zhu et al. 2021, Yin et al. 2023). Nevertheless, integrating diverse transport modes and complex land uses surrounding transportation hubs leads to heightened traffic density and conflicts among heterogeneous traffic flows. Consequently, this increases the likelihood of vehicle crashes (Yasmin et al. 2018, Ali et al. 2023). In 2017–––2018, 6042 crashes in Maryland occurred 300 m near transportation hubs (MVA 2023). These findings prompt consideration of the impact of the built environment surrounding transportation hubs on crash occurrences, in addition to traditional factors that scholars often focus on, such as road users, road design, and vehicle safety (Jin et al. 2024, Zou et al. 2024).

Scholars have extensively studied the influence of the built environment, denoting the overall surroundings of a developed city or area, on road safety (Xin et al. 2017, Lyu et al. 2023). Particularly, focused attention has been given to how the built environment affects crash risk, including the impact of factors like intersection density, building density, and mixed land use on bicycle accidents and overall road network safety (Asadi et al. 2022, Xiao et al. 2024). Interestingly, scholarly investigations on transportation hub safety have mainly concentrated on broader indoor safety aspects, encompassing fire safety and pedestrian evacuation procedures, among other considerations (Salarian et al. 2020, Shiwakoti et al. 2020). Nevertheless, the impact of the external built environment at transportation hubs on the frequency and severity of traffic accidents still needs to be explored, revealing significant research gaps. Understanding how the external built environment surrounding transportation hubs influences crash risk is crucial, as this knowledge can empower authorities to devise evidencebased interventions and policies to reduce crash probability and enhance external traffic safety at these hubs.

Given the existing research limitations, this study aims to investigate how the built environment influences crash risk (CR) at transportation hubs. The Maryland open crash dataset and additional open datasets delineated various buffer zones surrounding transportation hubs. Accident frequencies within these zones and built environment variables were analyzed. Three models—Ordinary Least Squares (OLS), Geographically Weighted Regression (GWR), and Multiscale Geographically Weighted Regression (MGWR)—were employed and

E-mail addresses: chuanyaoli@csu.edu.cn (C. Li), chenlijt@csu.edu.cn (L. Chen).

^* Corresponding author.

compared across the buffer zones. The discussion focuses on the MGWR regression coefficients, offering insights from diverse spatial perspectives and comparing different hub types. The primary objective of this research is to pinpoint specific built environment factors contributing to crash risk at transportation hubs. By comprehensively examining how the built environment of transportation hubs influences peripheral crash risk, this study provides valuable insights for urban planners, policymakers, and traffic management authorities.

This study is outlined as follows. Section 2 presents the literature review, acknowledging the relationship between the built environment and traffic safety, along with the common methodologies employed to establish this connection. Section 3 encompasses the methodology, detailing the study region, data sources, data processing methods, and an in-depth explanation of both global and local regression models utilized. Section 4 consists of the results, which incorporates spatial analysis of crash risk and the outcomes from the modeling. Section 5 is discussion, including the contributions, policy implications, and limitations of this study. Section 6 is conclusion,summarizes the main findings of the study and suggests directions for future research.

2. Literature review

2.1. The impact of built environments on traffic safety

Numerous studies have established strong correlations between various built environments and traffic safety (Merlin et al. 2020, Asadi et al. 2022). These correlations involve bicycle crashes and the built environment of streets, student commuting safety and its built surroundings, and roadway network safety and the built environment. Specifically, Costa et al. (2024) demonstrated the connection between bicycle accidents and the type of built environment using a latent class discrete outcome model. Their findings highlighted that busy intersections and a small percentage of road surface area increase the likelihood of crashes and influence crash severity; Wu et al. (2023) delved into the interrelation between the built environment and crash risks for school-commuting children. Their research emphasized the importance of factors like lower road network densities, rational educational facility layouts, reduced bus routes, and increased on-street parking spaces in mitigating crash risks; Xiao et al. (2024) explored how the built environment and traffic flow impact crash frequency on urban road networks. Their study revealed that variables such as employment density, road density, land use diversity, and public transport accessibility significantly impact crash occurrences. Moreover, attention has been directed towards examining the correlation between transportation safety and the built environment for older people (Lee et al. 2020) and investigating the impact of the built environment on children's traffic security (Macedo et al. 2023).

While existing studies have extensively explored the correlation between various types of built environments and crash risk, there needs to be more research investigating the link between the built environment surrounding transportation hubs and crashes. Most prominently, research on transportation hub safety has predominantly focused on the indoor aspects of the hub, encompassing passenger safety (Xie et al. 2022), safety facility design (Xie et al. 2023), and emergency evacuation and escape of crowds (Alam et al. 2023). Therefore, there remains a substantial knowledge gap concerning the external safety of transport hubs, specifically regarding how the surrounding built environment influences crash risk.

2.2. Modeling approaches to link the built environments and traffic safety

There are two primary modelling categories in exploring the link between the built environment and traffic safety: machine learning methods and traditional regression modelling. Machine learning methods leverage extensive datasets for training models to comprehend the influence of built environment factors on traffic safety. This method offers the advantage of handling the intricate, non-linear connections and uncovering nuanced patterns between the built environment and traffic safety (Bi et al. 2023, Galvez-P ´ ´erez et al. 2023). These models can forecast the impact of changes in the built environment on traffic safety to inform decision-making processes (Chang et al. 2022, Costa et al. 2024). However, machine learning methods necessitate substantial quantities of high-quality data for effective model training, and the absence of inadequacy of data may compromise model performance (Zhu et al. 2023). Furthermore, specific deep learning models exhibit a profound black-box nature that hinders a comprehensive understanding of the correlation between the built environment and traffic safety, including the spatial relationships between the built environment and crash risk.

Another category is traditional regression modeling. The first is a global model that does not take scale effects into account, including OLS, Generalized Linear Models (GLM), and binary logit models (Umair et al. 2022, AlKheder 2023, Soltani and Roohani Qadikolaei 2024). Although these models can identify potential control factors impacting traffic safety due to the built environment, they fail to address spatial dependencies. To address this limitation, GWR models are recognized for their spatial analysis capability, effectively illustrating the spatial relationships between dependent and independent variables (Gao et al. 2024, Liu et al. 2024). However, standard GWR models assume uniform spatial interactions at the same scale, applying a single bandwidth across all variables. This uniformity restricts the model's ability to adequately capture variations across multiple scales. Therefore, Fotheringham et al. (2017) introduced MGWR based on a generalized additive model. MGWR allows for individual spatial smoothing levels for each covariate, utilizing optimal bandwidths for each independent variable. This customization of spatial scales optimizes the modeling of each spatial process, enhancing the model's explanatory power and realism (Duan et al. 2024). In MGWR, the model employs the golden search algorithm to determine the most suitable bandwidths for diverse variables in regression modeling. This approach minimizes discrepancies between bandwidths and variable action scales, thereby reducing noise in parameter estimation (Zhao et al. 2024). MGWR has found applications in interdisciplinary studies encompassing environmental protection, urban landscapes, and public transport (Tang et al. 2023, Wang et al. 2023, Ma et al. 2024).

2.3. Research summary

In conclusion, while existing studies have extensively examined the relationship between built environments and traffic safety, there has been limited focus on the external safety of transportation hubs, particularly regarding how the surrounding built environment influences crash risk. Furthermore, most prior research has relied on either local or global spatial models, with insufficient attention to multi-scale spatial interactions. This study's contribution is to addresses these gaps by employing Multi-Scale Geographically Weighted Regression (MGWR) to analyze the Maryland crash dataset and other open-source datasets. By investigating the built environment's multi-scale spatial effects on crashes around transportation hubs, this study not only advances the theoretical understanding of the spatial dynamics between built environments and traffic safety but also provides practical recommendations for optimizing the external built environment of transportation hubs. These findings contribute to the existing body of knowledge and offer decision-makers actionable insights to enhance transportation safety in critical urban areas.

3. Methodology

3.1. Study area and data source

The State of Maryland is in the Mid-Atlantic region of the United States and is also known as the "Free State". The state includes 23 counties and one independent city, with the largest city being Baltimore and the capital being Annapolis. According to the U.S. Census Bureau's 2022 data, the state has 6, 165, 000 residents, making it one of the smallest states in the U.S. in terms of land area and one of the most densely populated.¹ Maryland's economy is predominantly tertiary, with transportation, centered on the Port of Baltimore being the most critical component of Maryland's economy.2 As a result, Maryland is not only home to world-class ports such as the Port of Baltimore and a welldeveloped network of roadways but also includes the Baltimore-Washington International Thurgood Marshall Airport (BWI), as well as significant air and rail carriers such as Amtrak, the Maryland Area Regional Commuter Railroad (MARC), and Washington, D.C.'s Metropolitan Area Rapid Transit (MART) system.³ Overall, Maryland is very well connected.

This study uses transportation hub data, administrative district data, and built environment data from Open Street Map (OSM).4 OSM is an open and accessible geographic information dataset from volunteers worldwide and is widely used in many studies. The data at the national level of the United States was acquired from OSM and subsequently subjected to data processing using ArcGIS 10.5 software. Finally, the dataset about the state of Maryland was exported. This includes multiple layers such as buildings, land use, POIs, transport, etc. We extracted geographic data from the transport layer for 208 transportation hubs, specifically 110 railway stations, 69 airports, 24 ferry terminals, and 5 bus stations. Their geographic distribution in Maryland is shown in Fig. 1.

In addition, a substantial dataset was obtained from the Open Data Portal of the Maryland State Government,5 which provides official and publicly accessible data for developing and studying traffic crashes. This dataset comprises 187, 405 crash records from January 1, 2017, to December 31, 2018. The original dataset consisted of 38 fields, including the location (down to the lane level), type of crash, and time of day each accident occurred. For this study, we selected the necessary fields, which include five key variables: date of the crash, type of crash, and geographic coordinates (including latitude and longitude), and assigned a unique identifier to each data record. The locations of the crash points are presented in Fig. 2.

3.2. Spatial scale

This study investigated how the built environment surrounding a transportation hub influences its CR. Therefore, we used the buffer zone as the basic spatial unit of analysis. Specifically, a buffer zone is established by designating a specific radius centered on the transportation hub. However, there is an important issue that needs to be considered: the modifiable areal unit problem (MAUP), which affects the interpretation of the results by the model, as different spatial division methods and scale choices have a significant impact on the results of the analyses (Clark and Scott 2014). Equally important, another concern is that the choice of multiple spatial scales can provide information on the built environment at different levels and scales. This helps to gain a comprehensive understanding of the factors influencing the occurrence of traffic accidents (Nigg et al. 2022). The study by Jun et al. (2015) states that the regular grid scale division ranges from 300 m to 1000 m, reducing the impact of MAUP. Therefore, to reduce the effect of MAUP and to capture multi-level information, we refer to the studies of Su et al. (2022) and Liu et al. (2023) and choose 300 m, 500 m, and 800 m as the spatial analysis units for this study.

3.3. Measuring the variables

3.3.1. Measurement of the dependent variable: crash risk of transportation hubs

Risk (R) is defined as the likelihood and potential impact of hazardous events, which can be understood as the product of the likelihood (L) of an event occurring and the consequences (C), denoted as R = L × C (Youli et al. 2018). Therefore, based on the risk theory, in this study, the crash risk (CR) of the buffer zone around the transportation hub is defined by the crash likelihood (CL) and crash severity (CS), which can be expressed as CR = CL × CS (Merlin et al. 2020, Wu et al. 2023). Further, considering the accessibility of the crash dataset, we adjusted the use of the absolute number of crash occurrences to denote the likelihood of occurrence (i.e., the higher the number of crashes occurring, the higher the likelihood of a crash occurring as well), for model development and result analysis. We have considered three different crash types and propose the quantitative crash risk assessment model for the buffer zone of a transportation hub, which was integrated based on the number of crashes occurring around transportation hubs. The equation is shown below:

$$CR_i = \alpha P_i + \beta I_i + \gamma F_i \tag{1}$$

In Eq. (1), CRi denotes the crash risk equivalent of the buffer i. Pi is the number of "Property Damage Crash" crash types in buffer i; Ii is the number of "Injury Crash" accident types in buffer i; and Fi is the number of "Fatal Crash" accident types in buffer i. α, β and γ are severity weights for property damage crash, injury crash, and fatal crash, respectively. Due to the limitations of our crash dataset, which did not include information on property damage (in US dollars), injuries, and fatalities per crash, we relied on the study conducted by Wu et al. (2023) as a reference. They utilized crash data from Shenzhen and employed structural equation modeling to develop a measurement model that establishes the relationship between school commuting crash risk and various measures, including the number of fatalities, serious injuries, slight injuries, and direct property damages. We acknowledge that the severity weights determined by Wu et al. (2023) differ from this study in the exact coefficient values. Nevertheless, Wu et al. (2023) confirmed the coefficients of fatal crashes were higher than those for injury crashes, followed by property damage crashes. Therefore, considering their findings, we adopted the estimated coefficients for α = 0*.22, γ* = 0*.*52 from Wu et al. (2023) in this study. It should be noted that they differentiate between injury crash types, including slight and serious injuries, and we averaged the coefficients for both, so β = 0*.*33.

3.3.2. Measurement of independent variables: metrics of the built environment of transportation hubs

The most commonly employed framework for describing the urban built environment is the 5D conceptual model proposed by Ewing and Cervero (2010), specifically density, diversity, design, destination accessibility, and distance. Based on this framework, the availability of the dataset and the characteristics of the POIs data were combined to select 13 metrics describing the built environment surrounding the transportation hub from three dimensions: land use, transport system, and urban design in this study. The specific details of these metrics are presented in Table 1, which includes the names, symbols, and measurement methods.

Due to space constraints, we highlight the descriptive statistics of the built environment of the total sample (N = 208) of transportation hubs at three spatial scales: 300 m, 500 m, and 800 m, as shown in Tables 2-4, respectively. The differences in the built environment of transportation hubs at different spatial scales can be observed by comparison. As the buffer ring expands, the indicator values for many density types decrease. However, the BSD did not change within the different buffer zones, which is because the expansion of the buffer zone did not change the accessibility measurements. Moreover, multicollinearity diagnostics

¹ Data from: https://www.census.gov/.

² Data from: https://zh.wikipedia.org/wiki/Maryland.

³ Data from: https://mdot.maryland.gov/pages/home.aspx#.

⁴ OSM website: https://www.openstreetmap.org/.

⁵ https://opendata.maryland.gov/.

Fig. 1. Geographic distribution of transportation hubs.

Fig. 2. Geographic distribution of crash data.

were performed by calculating the Variance Inflation Factor (VIF). The results showed no multicollinearity (Tolerances > 0.1, VIF < 5) was observed in the selected 13 variables, allowing us to proceed to the following analysis step. Itshould be noted that large orders of magnitude between different variables can lead to intrinsic anisotropy in nature, further blurring the relationship between the independent and the dependent variables. In this regard, we performed the next step of the model analysis with logarithmic changes in 2 high-level variables: population density and road intersection density.

3.4. Measuring the method

The three models were used for the modeling analysis include OLS, GWR, and MGWR in this study.

OLS is a global model that does not consider spatial effects, and its mathematical expression is shown in Eq. (2):

$$y_i = \alpha_0 + \sum_m \alpham x{im} + \varepsilon_i \tag{2}$$

In Eq. (2), yi is the dependent variable, which is the CR of transportation hub i(i = 1, 2, ⋯, 208); xim denotes the independent variable is the built environment variable m for transportation hub i, which represents the built environment variables of transportation hub i; *α0, α^m* are the regression coefficients; εⁱ is the random error in Eq. (2). In this study, OLS serves a dual purpose. It not only captures the global impact of the respective variables on the dependent variable, but also plays a crucial role in assessing the presence of multicollinearity in these variables.

In this regard, we introduced GWR, whose regression coefficients vary with geographic location, allowing for a more spatial interpretation, and whose mathematical expression is shown in Eq. (3):

$$y_i = \beta_0(\sigma_i, \taui) + \sum{r=1} \beta_r(\sigma_i, \taui) x{ir} + \epsilon_i$$ (3)

In Eq. (3), yi has the same meaning as Eq. (2); (σi, τi) is the longitude and latitude of transportation hub i; *β*0(σi, τi) is the constant term of transportation hub i;βr(σi, τi) is the coefficient of the built environment variable r for transportation hub i; xir is the built environment variable r for transportation hub i; ∊ⁱ represents uncaptured variation in random error and spatial effects.

As we have illustrated in the literature review, in comparison to the GWR, the MGWR regresses each independent variable using its optimal bandwidth, personalizing the spatial scale at which each spatial process operates, resulting in a spatial process model that is closer to the real one and has more explanatory power. The mathematical formulation of the

Table 1 Definitions and measurements of variables used to describe the built environment.

Variable (Symbol)	Measurement (Unit)
Enterprise office POI density (EO)	Number of enterprise office POIs in the buffer zone /area (#/km2 )
Leisure and entertainment POI density (LE)	Number of leisure and entertainment POIs in the buffer zone/area (#/km2 )
Tourism POI density (TOUR)	Number of tourism POIs in the buffer zone/area (#/km2 )
Life service POI density (LS)	Number of life service POIs in the buffer zone/ area (#/km2 )
Shopping center POI density (SC)	Number of shopping center POIs in the buffer zone/area (#/km2 )
Dining POI density (DIN)	Number of dining POIs in the buffer zone/area (#/km2 )
Bus stop accessibility (BSA)	Distance from the hubs to the nearest bus stop in the buffer zone (km)
Bus stop density (BSD)	Number of bus stops in the buffer zone /area (#/km2 )
Traffic signal density (TSD)	Number of traffic signals in the buffer zone/area (#/km2 )
Road intersection density (RID)	Number of road intersections in the buffer zone/ area (#/km2 )
Road network density (RND)	Sum of road network lengths in the buffer zone/ area (km/km2 )
Land using mixing degree (LUM)	∑N 2 (Xij/Xi) j=1
Population density (PD)	Number of residents in the buffer zone / areas (#/km2 )

Note: 1. The "area" refers to the size of different spatial unit; 2. The Herfindahl-Hirschman Index (HHI) is used for "land use mixing degree"; the smaller the HHI value, the higher the mixing, and vice versa. Xi is the number of total POIs in the buffer i; Xij is the number of total POIs in category j in the buffer i.

MGWR model is shown in Eq. (4):

$$y_i = \beta_0(\sigma_i, \taui) + \sum{j=1}^n \beta_{bwj}(\sigma_i, \taui)x{ij} + \epsilon_i$$ (4)

In Eq. (4), bwj represents the bandwidth used for the regression coefficients of the variable j; βbwj(σi, τi) denotes the regression coefficient of the variable j of transportation hub i under the bandwidth bwj; xij is the built environment variable j of transportation hub i. Refer to Eq. (3) for the meaning of the other terms.

4. Results

4.1. Spatial dependence and heterogeneity analysis of CR in transportation hubs

According to Eq. (1), the CR values of each transportation hub were calculated under various buffers. The spatial distribution of these values is displayed in Fig. 3. The transportation hubs with higher CR are predominantly located in downtown Baltimore, Montgomery, and Prince George's areas immediately adjacent to Washington, DC. And some of the more remote transportation hubs have lower CR. However, we combine this with Fig. 1 to show that the CR of railway station hubs located in urban areas is higher than that of other hubs. Note that we only show the spatial distribution of CR in this section, which has no effect on the results of CR as a dependent variable.

To explore the spatial dependence of CR at transportation hubs, we calculated the Moran's index under different buffers, as shown in Table 5. The 300 m buffer zone was better at catching the spatial dependence of CR of transportation hubs when hub diversity was not considered (all transportation hubs, N = 208). Whereas, as the buffer zone expands, the value of Moran's index becomes smaller, implying that the spatial dependence of the CR of the transportation hub decreases, and all statistical indicators pass the significance test (P < 0*.*050, Z < 1*.*923). The total sample demonstrated good spatial dependence within any buffer zone. This further shows that it is reasonable to utilize 300 m, 500 m, and 800 m buffer zones as the spatial unit of study to investigate the impact mechanisms of CR at transportation hubs.

In addition, we calculated Moran's index separately for railway station hubs and airport hubs under different buffer ranges. The Moran's index for ferry and bus stations could not be calculated due to insufficient sample size (N < 30). Moran's index of CR for train stations and airports is positive in any buffer and passes the significance test (P < 0*.*050, Z < 1*.*378). It shows the favorable spatial dependence of different transportation hubs CR within the 300 m, 500 m, and 800 m buffer zones. The variability of factors influencing the CR of different transportation hubs can be further explored in subsequent studies.

Lastly, the Theil Index was introduced to materialize the spatial heterogeneity of the CR. Lower values of the Theil Index indicate minimal heterogeneity among buffers. Fig. 4 depicts the results of the Theil index, which illustrates the spatial heterogeneity of CR. The airport CR exhibits the highest Theil index among buffers of the same size, while the Theil index of the total sample of CR is the lowest, suggesting that CR heterogeneity in space is most obvious in airport hubs. Comparing buffers across different spatial scales, transportation hub CR within the 300 m buffer demonstrates a higher Theil index, indicating more significant heterogeneity than the other buffers.

Table 2 Results of descriptive statistics and multicollinearity diagnostics for built environment variables in 300 m buffer zone.

Variables	Obs	Mean	Std.Dev.	Min	Max	Multi-collinearity diagnosis
						Tolerances	VIF
PD (#/km2 )	208	1355.483	2119.259	0	11840.010	0.508	1.967
LUM (no unit)	208	0.765	0.308	0.200	1	0.391	2.556
RND (km/km2 )	208	20.284	18.465	0	91.282	0.263	3.809
TSD (#/km2 )	208	8.605	15.741	0	74.283	0.382	2.620
RID (#/km2 )	208	87.403	102.398	0	291.703	0.256	3.902
BSD (#/km2 )	208	4.685	5.644	0	18.061	0.339	2.949
BSA (km/km2 )	208	7.028	10.989	0	62.872	0.538	1.857
EO (#/km2 )	208	0.901	2.299	0	17.687	0.600	1.667
LE (#/km2 )	208	8.980	21.817	0	134.418	0.493	2.030
TOUR (#/km2 )	208	2.329	7.909	0	95.507	0.778	1.285
LS (#/km2 )	208	8.571	18.593	0	99.045	0.288	3.467
SC (#/km2 )	208	3.639	11.373	0	106.119	0.405	2.470
DIN (#/km2 )	208	7.363	20.988	0	187.478	0.358	2.793

Table 3 Results of descriptive statistics and multicollinearity diagnostics for built environment variables in 500 m buffer zone.

Variables	Obs	Mean	Std.Dev.	Min	Max	Multi-collinearity diagnosis
						Tolerances	VIF
PD (#/km2 )	208	963.298	1465.740	0	10112.711	0.301	3.325
LUM (no unit)	208	0.703	0.318	0.220	1	0.390	2.565
RND (km/km2 )	208	13.360	9.370	0	51.697	0.532	1.878
TSD (#/km2 )	208	3.097	7.012	0	39.471	0.230	4.339
RID (#/km2 )	208	61.739	80.926	0	449.738	0.518	1.930
BSD (#/km2 )	208	4.083	8.721	0	64.936	0.333	3.005
BSA (km/km2 )	208	7.028	10.989	0	62.872	0.675	1.483
EO (#/km2 )	208	0.710	1.274	0	7.639	0.471	2.121
LE (#/km2 )	208	5.447	12.283	0	71.301	0.367	2.726
TOUR (#/km2 )	208	1.138	3.461	0	40.743	0.714	1.401
LS (#/km2 )	208	5.754	13.015	0	104.406	0.242	4.131
SC (#/km2 )	208	2.062	5.165	0	35.650	0.302	3.306
DIN (#/km2 )	208	4.688	12.544	0	87.853	0.233	4.292

Table 4 Results of descriptive statistics and multicollinearity diagnostics for built environment variables in 800 m buffer zone.

Variables	Obs	Mean	Std.Dev.	Min	Max	Multi-collinearity diagnosis
						Tolerances	VIF
PD (#/km2 )	208	864.7242	1213.952	0	5918.940	0.324	3.807
LUM (no unit)	208	0.664	0.326	0.189	1	0.374	2.674
RND (km/km2 )	208	6.401	4.744	0	31.592	0.584	1.712
TSD (#/km2 )	208	0.774	1.830	0	10.504	0.348	2.875
RID (#/km2 )	208	49.618	67.324	0	264.337	0.803	1.245
BSD (#/km2 )	208	1.020	2.128	0	12.414	0.479	2.089
BSA (km/km2 )	208	7.027	10.989	0	62.872	0.733	1.364
EO (#/km2 )	208	0.511	0.822	0	5.411	0.481	2.078
LE (#/km2 )	208	2.561	6.103	0	39.470	0.247	4.048
TOUR (#/km2 )	208	0.515	1.335	0	12.095	0.534	1.873
LS (#/km2 )	208	2.561	6.104	0	39.470	0.158	6.342
SC (#/km2 )	208	0.809	1.918	0	13.687	0.270	3.699
DIN (#/km2 )	208	2.049	4.904	0	31.831	0.245	4.080

4.2. Modeling analysis

This study designed two types, totaling 21 modeling scenarios (1 × 3 × 3 + 2 × 3 × 3), as shown in Fig. 5. Type I does not differentiate between transportation hubs, and the results are based on total sample regression. Type II distinguishes between transportation hubs, and as previously elaborated, bus stations and ferries have too few sample sizes to be used for developing separate models. Therefore, the Type II scenarios were used to compare the differences in the built environment of different transportation hubs regarding their CR impacts.

Nine scenarios (1 × 3 × 3) are constructed to investigate the impact of the built environment on the CR of transportation hubs, and the scenarios are designed as shown in Fig. 5, Type I. We compared the evaluation indicators of the 9 model scenarios, as shown in Table 6, to find the optimal explanatory model.

The AIC and AICc metrics favor lower values, indicating superior model performance. For the R2 and Adj R2 metrics, values closer to 1 indicate better model fit. Within the 300-meter buffer zone, the AIC for OLS is 376.915, the GWR is 365.005, and the MGWR is 335. The MGWR model demonstrates the lowest AIC value, signifying its superior performance in terms of explanatory variable selection and model fitting within the 300 m buffer zone. Similarly, within the 500 m and 800 m buffer zones, OLS and GWR underperformed MGWR based on the AIC metrics, consistently attaining the lowest values. Moreover, for both R² and Adj R2 metrics, the MGWR model has higher values relative to the other models in any buffer. This indicates that the MGWR model possesses advantages in terms of explanatory variable selection and model fitting across different buffer zone ranges and is better able to account for the correlation between the built environment around transportation hubs and CR at any spatial scale.

Further, we compare the bandwidth and standard deviation (STD) of

two local models, GWR and MGWR. For the MGWR model, we sorted the bandwidths of the independent variables in ascending order. The bandwidths and STD of the estimated parameters for the GWR and MGWR models for the three buffers of 300 m, 500 m, and 800 m are shown in Figs. 6 - 8. Generally, we observe that the STD in the MGWR model decreases as the bandwidth increases. This is likely because as the bandwidth increases, the number of neighboring observations the model considers increases, which can provide more spatial information to estimate the parameters, reducing the estimation uncertainty due to sample finiteness and thus reducing the standard error. However, the STD of the GWR model varies irregularly because GWR uses a fixed bandwidth for estimation, which may result in an uneven density and distribution of observations at different locations. Hence, the standard errors may also vary irregularly.

4.3. Mechanisms of CR impacts from the built environment of transportation hubs

From the analyses in Section 4.2.1, MGWR provides optimal results under different buffers compared to the other two methods. Further, we compare the estimated coefficients of the external built environment of transportation hubs on CR based on MGWR under different buffers, as shown in Table 7.

The findings reveal considerable variability in the key variables across different buffer distances, and their impact on CR exhibits significant variations in magnitude and significance. Within the 300 m buffer, we identified 5 key variables: population density, traffic signal density, bus stop density, enterprise office POI density, and dining POI density. As we expanded the buffer to 500 m, the number of critical variables increased to 6, encompassing population density, land use mix, road network density, traffic signal density, bus stop density, and dining

Fig. 3. Spatial distribution of CR under different buffers zone.

POI density. Further expanding the buffer to 800 m resulted in identifying 7 key variables: population density, traffic signal density, road intersection density, bus stop density, enterprise office POI density, life service POI density, and dining POI density.

A spatial presentation comprises these four variables: population density, traffic signal density, bus stop density, and enterprise office POI density. All four variables have significant survival correlations and a large degree of influence on the CR of the transportation hub, which offsets the effect of the MAUP. Population density increases external CR in the hubs of Baltimore's urban center and the northeastern Maryland region, and the same trend is observed for the other three variables, as shown in Fig. 9.

The result of the optimal model based on the 800 m buffer is analyzed as an example. Specifically, four key variables—population density (Cofe = 0*.418,* p < 0*.01), traffic signal density (Cofe = 0.250,* p < 0*.01), bus stop density (Cofe = 0.348,* p < 0*.*01), and Enterprise

office POI density (Cofe = 0*.170,* p < 0*.*01) have a significant positive effect on the CR of the transportation hubs at the 1 % confidence level. This association may be attributed to the notion that higher residential population density, traffic signal density, bus stop density, and enterprise office POI density indicate a bustling transportation hub area, signifying increased traffic flow, elevated vehicle densities, and heightened points of conflict among diverse traffic flow, thereby amplifying crash risks.

Conversely, three key variables—road intersection density (Cofe = − 0*.083,* p < 0*.05), life service POI density (Cofe = − 0.163,* p < 0*.1), and dining POI density (Cofe = − 0.114,* p < 0*.*1), exhibit a negative influence on the CR of transportation hubs at the 5 % to 10 % confidence levels, albeit with a moderated impact. Greater road intersection densities typically indicate a more resilient road network in the vicinity, enhancing evacuation efficiency and reducing dependence on individual intersections, hence bolstering network reliability. Higher life service

Fig. 3. (continued).

Table 5 Spatial dependence of CR of transportation hubs.

Categories	Area	Global Moran's I	P value	Z score	S.D.
All transportation hubs	300	0.852	0.000	5.359	0.026
(N = 208)	m
	500	0.664	0.000	2.429	0.025
	m
	800	0.347	0.003	1.923	0.026
	m
Railway station (N =	300	0.575	0.002	2.777	0.038
110)	m
	500	0.351	0.030	3.723	0.039
	m
	800	0.281	0.048	1.387	0.038
	m
Airport (N = 69)	300	0.314	0.001	2.184	0.033
	m
	500	0.223	0.047	2.039	0.012
	m
	800	0.186	0.004	2.248	0.057
	m

POI densities and Dining POI densities may reflect a more balanced urban planning approach, diminishing the necessity for extended travel distances to access services, thereby alleviating traffic congestion and unnecessary vehicular movements, consequently lowering crash probabilities.

Table 8 presents the estimated bandwidth results of the MGWR model within the 800 m buffer, revealing intriguing insights. The analysis highlights a substantial variance in the optimal bandwidths across the seven key variables. Notably, the impact scale of traffic signal density is 69, representing only 33.17 % of the total sample size, indicating a relatively lower effect than the other variables. Moreover, the effect scales for the remaining five crucial factors—residential population density, bus stop density, corporate office Point of Interest (POI) density, life service POI density, and food service POI density—are relatively close, constituting 61.06 %, 59.62 %, 60.58 %, 61.54 %, and 62.02 % of the total sample size, respectively. Noteworthy is the impact scale of road intersections at 117, accounting for 56.25 % of the total sample, underscoring its significance in the model. Remarkably, the bandwidth of the intercept is observed to be 207, nearing the global scale, which is a crucial observation in the context of the analysis.

4.4. Analysis of impact mechanisms of different types of transportation hubs

Similarly, this subsection presents the results for "Railway Station − MGWR-800 m" and the results for "Airport − MGWR-800 m". The CR impacts of different types of transportation hub's built environments are also compared. The results indicate substantial variations in the CR influence mechanisms among different types of transportation hubs, as detailed in Table 9 and 10.

At the land use level, a noteworthy positive effect of population density surrounding railway stations on CR was observed at the 1 % significance level (Cofe = 0*.396,*p < 0*.01), contrasting with the lack of significant impact of population density near airports on CR. Regarding the transport system level, the road network density encircling airports significantly positively affected CR (Cofe = 0.114,*p < 0*.01), indicating a direct relationship between higher road network density and increased crash likelihood. Conversely, the road network density surrounding railway stations did not display a significant effect on CR; instead, the traffic signal density near railway stations (Cofe = 0.401,*p < 0*.01) and road intersection density (Cofe = − 0.158,*p < 0*.05) demonstrated significant impacts on CR. Notably, these effects were not observed for airports. In terms of urban design, the life service POI density surround airports exhibited a significant positive effect on CR (Cofe = 0.400,* p < 0*.*05). At the same time, the same factor surrounding railway stations did not yield a significant impact on CR. This disparity may arise from the urban setting differences between airport cities, usually situated in remote or suburban areas, and railway station, typically located in city centers, leading to distinct urban design features impacting CR. These observations are further exemplified in Fig. 10, illustrating how higher bus stops and corporate office densities elevate CR surround railway stations predominantly positioned in Baltimore City's center. Conversely, bus stops exert a more pronounced influence on airfield in Maryland's southern region.

Moreover, the effect scales of key variables with the same impact effect for neighboring CR were also compared. Notably, the effect scale of bus stop density in railway station is 98, representing 89.09 % of the total sample, whereas the effect scale of bus stop density in airport is 35, accounting for 50.72 % of the total sample. Similarly, both enterprise office POI density and dining POI density in railway station are recorded at 108, closely aligning with the global scale of 110, indicating relatively stable coefficients across space. Conversely, the enterprise office POI

Fig. 4. Theil index in different buffer zones at different transportation hubs.

Fig. 5. Model development scenarios.

density and dining POI density in airport stand at 30 and 68, respectively, showcasing significant spatial heterogeneity. Furthermore, the effect scale of the intercept for the railway station is calculated at 106, nearing the full sample, whereas the intercept effect scale for the airport is 54, equivalent to 78.26 % of the total dataset.

4.5. Robustness analysis

we conducted robustness analyses of the core findings. We tested the robustness of the results by performing independent regression analyses for each of the three types of crashes (property damage crashes, injury crashes, and fatal crashes), using the number of occurrences within the 800 m buffer zone as the dependent variable, and comparing the results with those of the benchmark group model. The regression results are shown in Table 11.

The results indicate that the benchmark models are generally consistent with the direction and significance of the independent variables in the other groups. For example, Ln (PD) and TSD are positive and significant in all models and are indicated as significant at the 5 % level. Although there is a notable difference in the magnitude of significance, this does not change the core findings. This result certainly validates the robustness of the findings of this study. Therefore, the weighted combination of this study with reference to Wu et al. (2023) can reflection of the general trend of the impact of the built environment around transportation hubs on crash risk.

5. Discussion and policy recommendations

This study reveals the relationship between the external built environment of transportation hubs and CR (Crash Risk). Despite the variability in CR performance across different buffer zones, common spatial patterns emerge. Specifically, hubs with higher CR are predominantly in the Baltimore urban center and two cities adjacent to Washington, DC. This phenomenon may be attributed to increased traffic flow in urban centers and the presence of diverse transportation users (Asadi et al. 2022, Yoon and Lee, 2021). Analysis of model evaluation metrics revealed that the 800 m scale captured more spatial dependencies than the 300 m and 500 m scales, likely due to the broader geographical coverage of larger buffers, yielding more informative data points (Liu

Table 6 Indicators of optimal model estimation results of CR impact mechanisms under different buffer zones.

| Model indicators | | 300-meters buffer | | | 500-meters buffer | | | 800-meters buffer | | | |------------------|---------|-------------------|---------|---------|-------------------|---------|---------|-------------------|---------|--| | | OLS | GWR | MGWR | OLS | GWR | MGWR | OLS | GWR | MGWR | | | AIC | 376.915 | 365.005 | 335.000 | 237.107 | 240.583 | 229.062 | 306.136 | 229.613 | 218.426 | | | AICc | 381.415 | 376.789 | 340.952 | 241.607 | 245.741 | 234.294 | 310.636 | 261.496 | 232.339 | | | R2 | 0.687 | 0.750 | 0.765 | 0.840 | 0.850 | 0.857 | 0.777 | 0.890 | 0.880 | | | Adj R2 | 0.666 | 0.707 | 0.737 | 0.829 | 0.833 | 0.841 | 0.762 | 0.857 | 0.857 | |

Fig. 6. Bandwidth and STD for GWR and MGWR model parameter estimation in 300 m buffer.

Fig. 7. Bandwidth and STD for GWR and MGWR model parameter estimation in 500 m buffer.

Fig. 8. Bandwidth and STD for GWR and MGWR model parameter estimation in 800 m buffer.

Table 7 Estimated coefficients of external built environment of transportation hubs on CR under different buffers based on MGWR.

Variables	300-meters buffer		500-meters buffer		800-meters buffer
	Cofe.	S.E.	Cofe.	S.E.	Cofe.	S.E.
Ln (PD)	0.095**	0.056	0.155***	0.052	0.418***	0.059
LUM	− 0.030	0.064	− 0.103**	0.046	− 0.005	0.056
RND	0.086	0.078	0.075*	0.039	0.056	0.045
TSD	0.644***	0.065	0.433***	0.060	0.250***	0.058
Ln (RID)	− 0.032	0.079	0.029	0.040	− 0.083**	0.038
BSD	0.133**	0.069	0.319***	0.050	0.348***	0.049
BSA	− 0.002	0.055	0.008	0.035	− 0.019	0.039
EO	0.096*	0.052	0.022	0.042	0.170***	0.049
LE	0.003	0.057	− 0.055	0.047	0.091	0.068
TOUR	0.020	0.046	− 0.002	0.034	0.011	0.046
LS	− 0.105	0.075	− 0.063	0.058	− 0.163*	0.085
SC	0.045	0.063	− 0.053	0.052	− 0.003	0.065
DIN	0.129*	0.067	0.104*	0.059	− 0.114*	0.069
Intercept	− 0.019**	0.040	− 0.011**	0.029	− 0.071**	0.034

Note: 1. Cofe. stands for coefficient, S.E. stands for standard error; 2. ***p < 0.01; **p < 0.05; *p < 0.1.

et al. 2023). This contrasts with the findings of Wu et al. (2023) exploring the school perimeter built environment and CR, in which they found that small buffer zones as a spatial unit of analysis were more explanatory of the school perimeter built environment and CR. Moreover, regarding method performance comparison, it is evident that Multiscale Geographically Weighted Regression (MGWR) exhibits superior explanatory power, consistent with previous research findings (An et al. 2022, Kuo et al. 2024).

This study provides insights into spatial dependence and CR heterogeneity within transportation hubs. Based on these findings, specific and practicable policy recommendations are proposed for different spatial buffer scale (300 m, 500 m, and 800 m) and hub types (railway stations and airports). These recommendations emphasize adjustments to peripheral land use, infrastructure improvements, and urban planning strategies tailored to the unique characteristics of each spatial scale and hub types.

The 300 m buffer zone presents the strongest spatial dependence on

CR in densely populated urban areas near the transportation hubs. This study found a significant positive correlation between the density of traffic signalization equipment and the reduction in CR at this scale. Therefore, increasing the traffic signal density in and around railway stations should be considered on a case-by-case basis to regulate traffic flow and minimize conflict points between vehicles and pedestrians. Furthermore, the study revealed that areas with a poor land use mix have a higher risk of crashes, leading to the recommendation that highrisk land uses (e.g., heavy commercial or industrial areas) within 300 m should be restricted and replaced with pedestrian-friendly mixed-use developments. Regarding pedestrian infrastructure, new developments in the vicinity of hubs should incorporate traffic calming measures, such as raised crossings, speed humps, and clear signage, to reduce vehicle speeds and enhance pedestrian safety.

The 500 m buffer zone demonstrates a balanced combination of CR determinants, with significant contributions from factorslike population density, road network density, and bus stop density. It was observed that higher road network density is associated with lower CR. Consequently, the road network density within 500 m of the hub was increased to enhance connectivity and reduce congestion. Specific measures include providing additional secondary roads and alternate routes to alleviate congestion at major intersections. To improve transit accessibility, bus stop density in high pedestrian areas traffic should be increased while ensuring that the design minimizes CR by relocating bus stops away from intersections and incorporating shelters, lighting, and marked crossing areas. Regarding land use development, a mix of residential and commercial land uses will be encouraged within this buffer zone to shorten commuting distances and alleviate traffic congestion.

The 800 m buffer zone illustrates a broader spatial pattern, emphasizing the significance of built environment factors, such as population density, corporate office density, and road intersection density. Given that corporate office POI density positively affects CR, it is recommended that corporate office developments be centrally located in areas with well-developed transport networks to minimize reliance on private vehicles. For instance, zoning corporate parks in locations with good access to public transport can achieve this goal. Regarding amenity facilities, increasing the density of life service POI (e.g., grocery stores,

Fig. 9. Spatial distribution of estimated coefficients for four variables.

Table 8 Results of MGWR model estimation of bandwidth in 800 m buffers.

Variables	Bandwidth	Cofe.	STD	Min	Median	Max
Ln (PD)	127.000	0.418***	0.100	0.248	0.356	0.514
LUM	204.000	− 0.005	0.007	0.020	0.038	0.049
RND	121.000	0.056	0.111	− 0.082	0.190	0.367
TSD	69.000	0.250***	0.137	− 0.170	0.124	0.417
Ln (RID)	117.000	− 0.083**	0.099	− 0.289	0.089	0.035
BSD	124.000	0.348***	0.093	0.006	0.226	0.287
BSA	207.000	− 0.019	0.007	− 0.073	− 0.057	− 0.043
EO	126.000	0.170***	0.085	0.026	0.171	0.256
LE	107.000	0.091	0.181	− 0.218	− 0.104	0.206
TOUR	191.000	0.011	0.014	0.030	0.077	0.099
LS	128.000	− 0.163*	0.090	− 0.030	0.164	0.227
SC	207.000	− 0.003	0.001	− 0.007	− 0.003	0.000
DIN	129.000	− 0.114*	0.062	− 0.230	− 0.147	− 0.057
Intercept	207.000	− 0.071**	0.003	− 0.088	− 0.070	− 0.066

Note: 1. Cofe. stands for coefficient, STD stands for standard deviation; 2. ***p < 0.01; **p < 0.05; *p < 0.1.

Table 9 Estimation results for railway station-MGWR-800 m.

Variables	Bandwidth	Cofe.	STD	Min	Median	Max
Ln (PD)	82.000	0.396***	0.322	0.163	0.822	0.850
LUM	80.000	− 0.026	0.086	− 0.158	0.016	0.147
RND	92.000	0.071	0.105	− 0.058	0.142	0.356
TSD	107.000	0.401***	0.005	0.004	0.011	0.040
Ln (RID)	108.000	− 0.158**	0.011	− 0.179	− 0.179	− 0.111
BSD	83.000	0.245***	0.101	− 0.081	0.134	0.317
BSA	98.000	− 0.075	0.373	− 1.019	− 0.916	− 0.112
EO	108.000	0.159**	0.008	0.021	0.059	0.065
LE	74.000	0.145	0.167	− 0.111	0.213	0.275
TOUR	108.000	0.046	0.008	0.043	0.074	0.081
LS	96.000	− 0.143	0.062	− 0.101	0.036	0.103
SC	108.000	− 0.063	0.003	− 0.061	− 0.044	− 0.043
DIN	108.000	− 0.204*	0.003	0.088	0.106	0.109
Intercept	106.000	0.010**	0.043	− 0.242	0.022	0.027

Note: 1. Cofe. stands for coefficient, STD stands for standard deviation; 2. ***p < 0.01; **p < 0.05; *p < 0.1.

Table 10 Estimation results for airport-MGWR-800 m.

Variables	Bandwidth	Cofe.	STD	Min	Median	Max
Ln (PD)	35.000	0.001	0.137	− 0.268	− 0.202	0.152
LUM	53.000	0.078	0.026	0.000	0.072	0.087
RND	56.000	0.114***	0.060	0.039	0.068	0.213
TSD	53.000	− 0.011	0.039	− 0.055	0.090	0.095
Ln (RID)	50.000	0.026	0.025	− 0.059	− 0.015	0.039
BSD	54.000	0.463***	0.025	0.076	0.173	0.190
BSA	32.000	− 0.025	0.065	− 0.253	− 0.076	0.046
EO	30.000	0.231***	0.059	0.047	0.140	0.291
LE	68.000	0.116	0.001	0.074	0.077	0.077
TOUR	39.000	0.086	0.192	− 0.053	0.451	0.459
LS	68.000	0.400**	0.001	0.811	0.813	0.814
SC	53.000	0.094	0.027	0.137	0.139	0.237
DIN	68.000	− 0.328**	0.001	− 0.649	− 0.647	− 0.646
Intercept	54.000	− 0.016**	0.014	− 0.031	− 0.020	0.023

Note: 1. Cofe. stands for coefficient, STD stands for standard deviation; 2. ***p < 0.01; **p < 0.05; *p < 0.1.

and medical centers) can help decrease unnecessary trips and traffic congestion, thereby reducing CR. Urban design policies should focus on encouraging the development of essential services in proximity to residential areas.

The study identified significant differences in CR determinants between railway stations and airports, highlighting the need for tailored recommendations for these two transportation hubs. For railway stations, it is recommended to increase the bus stops density in the vicinity to enhance "last mile" connectivity and create an integrated transportation system that connects bus stops directly to station entrances, as this significantly reduces CR. Additionally, improving traffic signal density is essential to manage the high volume of pedestrian and vehicular traffic around the railway stations. For airports, increasing the road network density surrounding airports can reduce congestion and improve evacuation efficiency. This can be achieved by constructing additional access roads and providing high-capacity roundabouts at key intersections. Furthermore, to address the impact of amenity service density in the vicinity of airports on CR, a zoning policy should be implemented to support balanced urban planning and prevent overconcentration of service facilities around airports.

The limitations of this study should be acknowledged. Firstly, severity weights for crash types were adopted from Wu et al. (2023) without localized calibration. While these coefficients are consistent with established crash risk theories and have been validated in previous

Fig. 10. Spatial distribution of estimated coefficients for BSD and EO: railway stations and airports.

Table 11 Comparison of results for different groups with core explanatory variables using MGWR model within 800 m buffer.

Groups	Ln (PD)	TSD	Ln (RID)	BSD	EO	LS	DIN
Benchmark group	0.418***	0.250***	− 0.083**	0.348***	0.170***	− 0.163*	− 0.114*
Property Damage Crash	0.383***	0.216***	− 0.064**	0.369***	0.176***	− 0.176**	− 0.086
Injury Crash	0.463***	0.293***	− 0.042*	0.314***	0.155**	− 0.098**	− 0.155*
Fatal Crash	0.059**	0.185***	0.048	0.165**	0.364*	− 0.056	− 0.076**

^***p < 0.01; **p < 0.05; *p < 0.1.

studies, the application of these coefficients in different geographic environments may result in nuances. In addition, the absolute number of crashes was used as a proxy for crash risk without considering exposure factors such as traffic volume, trips. While absolute crash counts offer valuable insights into the spatial relationship between the built environment and crash occurrences, they may not fully capture the underlying risk dynamics. Despite these constraints, this study undeniably contributes to the theoretical understanding of the interplay between the built environment and traffic safety, shedding light on strategies to enhance the external safety of transportation hubs through optimized built environment planning.

6. Conclusion

This study reveals the relationship between the built environment around transportation hubs and crash risk, with the following main findings: (1) The 800 m buffer spatial scale demonstrates a stronger explanatory power in capturing various features of the built environment that influence crash risk around transportation hubs. (2) In the explanatory model, MGWR considers a flexible bandwidth compared to the other two methods, making it more suitable for analyzing the spatial heterogeneity of the dependent variable. (3) There are heterogeneous effects in the built environment of different types of transportation hubs regarding their impacts on surrounding crash risk, but it is not appropriate to attract too high a population density and build too high a road density around all transportation hubs.

Several potential directions for future research can be identified: (1)

When data become available in the future, localized weights for severity should be calibrated as well as exposure indicators incorporated to provide a more accurate assessment of crash risk. (2) While this study examined multiple spatial scales, exploring different time windows could offer valuable insights. (3) Employing appropriate machine learning techniques to explore the relationship between the external traffic environment of transportation hubs and crash risk may yield intriguing findings.

CRediT authorship contribution statement

Chuanyao Li: Writing – original draft, Supervision, Project administration, Methodology, Investigation, Funding acquisition. Li Chen: Writing – original draft, Software, Methodology, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This study was supported by grants from the National Natural Science Foundation of China (72271248), the Natural Science Foundation of Hunan Province, China (2024JJ5449), the Scientific Research Program of the Education Department of Hunan Province (24B0002), the Chinese Academy of Engineering Major Projects of Academy-Local Cooperation (2024-DFZD-57), the National Natural Science Foundation of China (72288101), the Higher-end Think-Tank Project of Central South University (2022znzk07), and the Fundamental Research Funds for the Central Universities of Central South University (Grant No.1053320240405).

Data availability

The authors do not have permission to share data.

References

• Alam, M.D.J., Pinchemel, A., Habib, M.A., Caetano, M., 2023. Airport evacuation under panic conditions: a microsimulation modeling applied at ottawa international airport. Simulation-Transactions of the Society for Modeling and Simulation International 99 (12), 1213–1224.
• Ali, Y., Washington, S., Haque, M.M., 2023. Estimating real-time crash risk at signalized
• intersections: a bayesian generalized extreme value approach. Saf. Sci. 164. Alkheder, S., 2023. Studying the effect of built environment on traffic safety with random
• parameter and generalized ordered logit models. Urban Clim. 47, 101388. An, R., Wu, Z., Tong, Z., Qin, S., Zhu, Y., Liu, Y., 2022. How the built environment promotes public transportation in wuhan: a multiscale geographically weighted
• regression analysis. Travel Behav. Soc. 29, 186–199. Asadi, M., Ulak, M.B., Geurs, K.T., Weijermars, W., Schepers, P., 2022. A comprehensive analysis of the relationships between the built environment and traffic safety in the dutch urban areas. Accid. Anal. Prev. 172, 106683.
• Bi, H., Li, A., Zhu, H., Ye, Z., 2023. Bicycle safety outside the crosswalks: Investigating cyclists' risky street-crossing behavior and its relationship with built environment. J. Transp. Geogr. 108, 103551.
• Chang, I., Park, H., Hong, E., Lee, J., Kwon, N., 2022. Predicting effects of built environment on fatal pedestrian accidents at location-specific level: application of xgboost and shap. Accid. Anal. Prev. 166, 106545.
• Clark, A., Scott, D., 2014. Understanding the impact of the modifiable areal unit problem on the relationship between active travel and the built environment. Urban Stud. 51 (2), 284–299.
• Costa, M., Lima Azevedo, C., Siebert, F.W., Marques, M., Moura, F., 2024. Unraveling the relation between cycling accidents and built environment typologies: capturing spatial heterogeneity through a latent class discrete outcome model. Accid. Anal. Prev. 200, 107533.
• Duan, H., Cao, Q., Wang, L., Gu, X., Ashrafi, K., 2024. Exploring the relationships between 3d urban landscape patterns and pm2.5 pollution using the multiscale geographic weighted regression model. Progress in Physical Geography-Earth and
• Ewing, R., Cervero, R., 2010. Travel and the built environment. J. Am. Plann. Assoc. 76 (3), 265–294.
• Fotheringham, A.S., Yang, W., Kang, W., 2017. Multiscale geographically weighted regression (mgwr). Ann. Am. Assoc. Geogr. 107 (6), 1247–1265.
• G´ alvez-P´erez, D., Guirao, B., Ortuno, ˜ A., 2023. Analysis of the elderly pedestrian injury severity in urban traffic accidents in spain using machine learning techniques. Transp. Res. Procedia 71, 6–13.
• Gao, C., Li, S., Sun, M., Zhao, X., Liu, D., 2024. Exploring the relationship between urban vibrancy and built environment using multi-source data: Case study in munich. Remote Sens. (Basel) 16 (6).
• Jin, J., Huang, H., Zhou, R., Chen, J., Liu, P., 2024. Analysis of dynamic determinants of vehicles involved in crash affecting severity based on in-depth crash data. Traffic Inj. Prev.
• Jun, M.-J., Choi, K., Jeong, J.-E., Kwon, K.-H., Kim, H.-J., 2015. Land use characteristics of subway catchment areas and their influence on subway ridership in seoul. J. Transp. Geogr. 48, 30–40.
• Kuo, P.-F., Hsu, W.-T., Putra, I.G.B., Sulistyah, U.D., 2024. The proposed model for analyzing off-street parking dynamics: a case study of taipei city. Transp. Res. A Policy Pract. 180, 103965.
• Lee, S., Yoon, J., Woo, A., 2020. Does elderly safety matter? Associations between built environments and pedestrian crashes in seoul, korea. Accid. Anal. Prev. 144,
• Liu, W., Li, D., Meng, Y., Guo, C., 2024. The relationship between emotional perception and high-density built environment based on social media data: evidence from spatial analyses in wuhan. Land 13 (3).

• Liu, X., Chen, X., Tian, M., De Vos, J., 2023. Effects of buffer size on associations between the built environment and metro ridership: a machine learning-based sensitive analysis. J. Transp. Geogr. 113, 103730.

• Lyu, T., Wang, Y., Ji, S., Feng, T., Wu, Z., 2023. A multiscale spatial analysis of taxi ridership. J. Transp. Geogr. 113, 103718.

• Ma, R., Huang, A., Cui, H., Yu, R., Peng, X., 2024. Spatial heterogeneity analysis on distribution of intra-city public electric vehicle charging points based on multi-scale geographically weighted regression. Travel Behaviour and Society 35.

• Macedo, F.E., Raaphorst, K.M.C., Bevelander, K.E., Van Der Krabben, E., 2023. The influence of the built environment on active school travel in the netherlands: a mode choice analysis. Multimodal Transportation 2 (4), 100103.

• Merlin, L.A., Guerra, E., Dumbaugh, E., 2020. Crash risk, crash exposure, and the built environment: a conceptual review. Accid. Anal. Prev. 134, 105244.

• Mva, M., 2023. Crash data dashboard-maryland takes a data-driven approach to reach our goal of zero roadway fatalities and serious injuries.

• Nigg, C., Niessner, C., Burchartz, A., Woll, A., Schipperijn, J., 2022. The geospatial and conceptual configuration of the natural environment impacts the association with health outcomes and behavior in children and adolescents. Int. J. Health Geogr. 21

• Salarian, A.H., Mashhadizadeh, A., Bagheri, M., 2020. Simulating passenger evacuation in railway station under fire emergency using safe zone approach. Transp. Res. Rec. 2674 (11), 806–812.

• Shiwakoti, N., Wang, H., Jiang, H., Wang, L., 2020. A 'role-rule' model to examine passengers' likely behaviour and their perceived ability to evacuate safely from airport in an emergency evacuation. Saf. Sci. 124.

• Soltani, A., Roohani Qadikolaei, M., 2024. Space-time analysis of accident frequency and the role of built environment in mitigation. Transp. Policy 150, 189–205.

• Su, L., Zhou, S., Kwan, M.-P., Chai, Y., Zhang, X., 2022. The impact of immediate urban

• environments on people's momentary happiness. Urban Stud. 59 (1), 140–160. Tang, X., Bi, R., Wang, Z., 2023. Spatial analysis of moving-vehicle crashes and fixedobject crashes based on multi-scale geographically weighted regression. Accid. Anal. Prev. 189.

• Umair, M., Rana, I.A., Lodhi, R.H., 2022. The impact of urban design and the built environment on road traffic crashes: a case study of rawalpindi, pakistan. Case Studies on Transport Policy 10 (1), 417–426.

• Wang, Z., Gong, X., Zhang, Y., Liu, S., Chen, N., 2023. Multi-scale geographically weighted elasticity regression model to explore the elastic effects of the built environment on ride-hailing ridership. Sustainability 15 (6).

• Wu, Y., Hu, X., Ji, X., Wu, K., 2023. Exploring associations between built environment and crash risk of children in school commuting. Accid. Anal. Prev. 193, 107287.

• Xiao, D., Ding, H., Sze, N.N., Zheng, N., 2024. Investigating built environment and traffic flow impact on crash frequency in urban road networks. Accid. Anal. Prev. 201, 107561.

• Xie, C.-Z., Tang, T.-Q., Hu, P.-C., Huang, H.-J., 2022. A civil aircraft passenger deplaning model considering patients with severe acute airborne disease. Journal of Transportation Safety & Security 14 (6), 1063–1084.

• Xie, W., Liu, W., Xiao, Y., Dai, Y., Dong, J., Liu, R., 2023. Mrsa-net: Anomaly detection of metro stations based on residual attention network with path aggregation and efficient representation. IEEE Sens. J. 23 (22), 28340–28354.

• Xin, C., Guo, R., Wang, Z., Lu, Q., Lin, P.-S., 2017. The effects of neighborhood characteristics and the built environment on pedestrian injury severity: a random parameters generalized ordered probability model with heterogeneity in means and variances. Anal. Methods Accid. Res 16, 117–132.

• Yasmin, S., Eluru, N., Wang, L., Abdel-Aty, M.A., 2018. A joint framework for static and real-time crash risk analysis. Anal. Methods Accid. Res 18, 45–56.

• Yin, C., Zhang, Z., Zhang, X., Chen, J., Tao, X., Yang, L., 2023. Hub seaport multimodal freight transport network design: perspective of regional integration development. Ocean Coast. Manag. 242.

• Yoon, J., Lee, S., 2021. Spatio-temporal patterns in pedestrian crashes and their determining factors: application of a space-time cube analysis model. Accid. Anal. Prev. 161, 106291.

• Youli, Y., Yingjian, P., Xiaoxia, L., Airan, Z., 2018. Research on safety risk management of civil construction projects based on risk matrix method. IOP Conf. Ser.: Mater. Sci. Eng. 392 (6), 062080.

• Zhao, J., Guo, F., Zhang, H., Dong, J., 2024. Mechanisms of non-stationary influence of urban form on the diurnal thermal environment based on machine learning and mgwr analysis. Sustain. Cities Soc. 101.

• Zhu, J., Cao, J., Saxena, D., Jiang, S., Ferradi, H., 2023. Blockchain-empowered federated learning: challenges, solutions, and future directions. ACM Comput. Surv. 55 (11).

• Zhu, Y., Diao, M., Li, J.J., 2021. Examining indoor air pollution in a large-scale integrated transportation hub in shanghai. Transportation Research Part D-Transport & Environment 97.

• Zou, R., Yu, H., Yang, H., Chen, C., Zhang, G., 2024. Severity analyses of urban twovehicle-involved rear-end crashes characterized by different configurations using mixed logit models. Journal of Transportation Engineering Part A-Systems 150 (2).