Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Journal of Transport & Health 27 (2022) 101529

Contents lists available at ScienceDirect

Journal of Transport & Health


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

Investigation of pedestrian and bicyclist safety in public


transportation systems
Md Tanvir Ashraf, Kakan Dey *, Dimitra Pyrialakou
Department of Civil and Environmental Engineering, West Virginia University, Morgantown, WV, 26505, USA

A R T I C L E I N F O A B S T R A C T

Keywords: Introduction: Understanding the crash risk factors associated with pedestrians and bicyclists is
Sustainable modes vital to promoting pedestrian- and bicycle-friendly communities and ensuring the sustainability of
Transit station the transportation system. Transit stations perform as important mode transfer locations for bi­
Traffic safety
cyclists and pedestrians. However, past studies on contributing factors to pedestrian and bicycle
Active transportation
Spatial dependency
crash in transit station service areas are limited. This study develops a model to guide the pri­
oritization of areas for bicyclist and pedestrian safety improvement.
Method: This study analyzes pedestrian and bicyclist safety in subway station service areas using
publicly available crash data from New York City (NYC) area. For the model development,
various subway station and built-environment characteristics related variables were aggregated
to the subway station service area level. Spatial regression and random parameter models were
developed to account for spatial heterogeneity and spatial dependency in the crash data. Per­
formance measures of the developed models showed that the Semi-Parametric Geographically
Weighted Poisson Regression (S-GWPR) model better explained the variation in the crash
frequency.
Results: The results showed that subway stations in NYC with high ridership were positively
associated with pedestrian and bicycle crashes. The number of bus stops in a subway station
service area was also positively associated with pedestrian and bicycle crashes. This study also
found that additional bike lanes can reduce bicycle-involved crashes.
Conclusion: Areas surrounding transit stations, where pedestrian and bicyclist activities are usu­
ally high, require special attention to deploy appropriate safety countermeasures. The findings of
this study suggest that cities can install more bike lanes to promote the safety of bicyclists. Using
the results of this study, traffic safety planners can prioritize areas for the implementation of
safety countermeasures in NYC.

1. Introduction

Transportation agencies have been implementing initiatives to promote active transportation modes (e.g., walking, bicycling) and
sustainable transportation modes (e.g., public transportation) to reduce the environmental impacts of transportation systems and
chronic congestion problems (Peng et al., 2021). Active transportation modes such as walking and cycling promote a healthy lifestyle
while being sustainable and environmentally friendly. However, pedestrians and bicyclists are among the most vulnerable road users,

* Corresponding author.
E-mail addresses: ma00001@mix.wvu.edu (M.T. Ashraf), kakan.dey@mail.wvu.edu (K. Dey), dimitra.pyrialakou@mail.wvu.edu (D. Pyrialakou).

https://doi.org/10.1016/j.jth.2022.101529
Received 4 January 2022; Received in revised form 26 October 2022; Accepted 26 October 2022
Available online 8 November 2022
2214-1405/© 2022 Elsevier Ltd. All rights reserved.
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

as they are the least protected in an automobile-involved crash. In 2019, the United States experienced 6,205 pedestrian and 846
bicyclist fatalities which is almost 17% of all traffic-related fatalities and the highest number since 1990 (NHTSA, 2019). Urban
planners and decision-makers have been focusing on pedestrian and bicycle-friendly roadways in recent years to ensure a safer and
more convenient road traffic environment for all modes of transportation. Nowadays, more people walk and bike to work or use
bike-sharing services to perform short utilitarian trips within the urban core (Noland et al., 2016). Ensuring the safety of pedestrians
and bicyclists is a challenging endeavor because of the lack of protective structure and the difference in body mass compared to motor
vehicles. Complete street design can be adopted to ensure pedestrian and bicycle safety (Smith et al., 2010). A complete street design
includes elements such as sidewalks, bicycle lanes (or wide, paved shoulders), shared-use paths, designated bus lanes, safe and
accessible transit stops, and frequent and safe crossings for pedestrians, including median islands, accessible pedestrian signals, and
curb extensions (Smith et al., 2010).
Because of its unique position as an intermodal interface, transit stops/stations play an important role in ensuring pedestrian and
bicyclist safety among the complete street design elements. A transit stop serves as a hub for various public and active modes of
transportation. Transit users need first-mile and last-mile connectivity to access transit services. For first- and last-mile connections,
travelers usually walk or use bicycles (Zuo et al., 2020). As a result, both pedestrian and bicyclist densities are high near transit
stations. The increased presence of bicyclists and pedestrians near transit stations might contribute to a higher number of pedestrian
and bicycle crashes. However, literature has not explored crashes involving pedestrians or bicyclists, exclusively focusing on the transit
stations and surrounding areas. This study uses spatial regression techniques to investigate the effects of transit station,
socio-demographics and other built environment characteristics on pedestrian and bicycle crashes surrounding transit stations. New
York City (NYC) subway network was chosen for the case study analysis as it serves more passengers than any other citywide public
transit networks in the US and plays a vital role in promoting a sustainable transportation system (Noland et al., 2016).
The rest of this paper is structured as follows: the literature review section provides a comprehensive overview of the past studies
focusing on pedestrian and bicycle crashes, followed by a discussion on the research gap in the existing literature. The following four
sections describe the pedestrian and bicycle-related crash data, the adopted research method for analysis, the results of the statistical
models, and the implications of the findings. Finally, the contributions of this study are discussed, and the last section presents
concluding remarks.

2. Literature review

Analyzing traffic crashes to understand the crash contributing factors has been a popular research area for the last few decades.
Previous studies related to traffic safety analysis can be categorized into two distinct groups. One group of studies focuses on specific
crash types and crash severities based on roadway elements such as intersections, roadway segments, corridors, ramps, and freeways.
Another group of studies focuses on a macro-level analysis of traffic crashes by aggregating them into different geographical units, such
as census tracts, Traffic Analysis Zones (TAZs), and census blocks (Siddiqui et al., 2012). Several studies have analyzed macro-level
pedestrian and bicycle crashes using different statistical models (Carter and Council, 2007; Kim et al., 2007; Siddiqui et al., 2012).
Siddiqui et al. (2012) investigated the roadway characteristics and various demographic and socio-economic factors related to
pedestrian and bicycle crashes using Bayesian spatial regression techniques by aggregating the pedestrian and bicycle crashes at the
TAZ level. The study found that the TAZ-level pedestrian and bicycle crashes were spatially autocorrelated, and these types of crashes
were more frequent at roadway intersections. Pedestrian density and employment density were positively associated with pedestrian
crashes. Moreover, long-term parking cost was positively associated with pedestrian crashes. The authors suggested that high parking
costs can discourage people from using a personal car and encourage public transit use, leading to more pedestrian-involved crashes.
Carter and Council (2007) explored the difference between contributing factors for pedestrian and bicycle crashes in rural and
urban areas in North Carolina. The study identified critical areas on rural highways where mitigation measures were needed to
improve safety (i.e., hot spots). Pedestrian crashes were more frequent in areas where pedestrians failed to yield at the midblock and
intersections. Bicyclists turning or merging into the path of the auto drivers at midblock, overtaking at the midblock, and failing to
yield at intersections were the main factors in bicyclist-involved crashes (Carter and Council, 2007). Saha et al. (2018) used bicycle
crash data at the census block level from the state of Florida and found that the bicycle crash counts were spatially autocorrelated. The
authors used smartphone-based bicycle ridership data and applied the Conditional Autoregressive (CAR) model, one of the popular
spatial analysis methods within the Bayesian framework. Population density, daily vehicle miles traveled (DVMT), age group,
households with no vehicles, urban road density selected by functional class, miles of bicycle trips, and intensity of bicycle trips were
positively associated with the total number of bicyclist crashes. Abdel Aty et al. (2011) examined the impact of road characteristics and
different trip productions and attraction characteristics on the total number of bicycle and pedestrian crashes at the TAZ level.
In addition to examining roadway characteristics and demographic characteristics, researchers have also analyzed built envi­
ronmental factors. For example, Chen (2015) investigated the effects of the built environment in Seattle, Washington, on TAZ-level
bicycle crashes. The study used several variables as surrogate measures of bicyclist exposure to traffic volume. The variables used
to measure bicycle exposure consist of bicycle trip volume, the total number of trips, and the modal share of bikes. Only the total
number of trip variables was significant among these exposure variables. The most widely used geographical unit for macro-level
pedestrian and bicycle crash analysis have been census blocks, block groups, tracts, and TAZs. TAZs have been well-established
because the study results/findings can be easily integrated into the TAZ-based transportation planning process (Ziakopoulos and
Yannis, 2020). Roadway segments or intersections have also been used to aggregate pedestrian and bicycle crashes (Huang and
Abdel-Aty, 2010; Pulugurtha and Sambhara, 2011; Ni et al., 2016; Lee et al., 2017).
The safety of active transportation modes around transit stops and the safety of transit access trips have recently gained attention (e.

2
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

g., González et al., 2019; Pulugurtha and Srirangam, 2021; Ulak et al., 2021). González et al. (2019) explored the effect of commercial
and residential gentrification of rail stations on pedestrian and bicyclist safety. They analyzed 213 rail stations in Los Angeles and San
Francisco to examine whether the pedestrian and bicyclist crash patterns differ between gentrified and non-gentrified areas. Com­
mercial gentrification due to new rail stations resulted in a chance of crashes two times higher in areas where more and more diverse
groups of pedestrian and bicyclist activities were present. Pulugurtha and Srirangam (2021) studied pedestrian safety at intersections
near light rail transit stations (LRT) before and after the LRT stations were opened in Charlotte, North Carolina. The results showed that
the number of bus stops and mixed-use, office, and industrial areas was negatively associated with pedestrian safety at intersections
near an LRT station. Ulak et al. (2021) developed a bus stop safety index to address pedestrian safety around bus stops and found a
significant positive relationship between bus stop locations and pedestrian-involved crashes.
For first- and last-mile connections to transit services, travelers often walk or use bicycles (Zuo et al., 2020). Previous studies have
reported that people usually walk to more distant transit stops motivated by health benefits or other benefits, such as avoiding crowded
bus stops or not wanting to wait unnecessarily at a bus stop (Ragaini et al., 2020; van Soest et al., 2020). On the other hand, people
living outside the transit coverage area tend to use park-and-ride (P&R) facilities to access the transit stations. Previous studies re­
ported that P&R users also usually walk more when the areas near subway stations are pedestrian-friendly (Cao and Duncan, 2019).
Furthermore, after the advent of bike-sharing systems, people are also becoming increasingly reliant on bike trips for both first- and
last-mile connectivity trips (Fan et al., 2019; Zuo et al., 2020; Baek et al., 2021). In light of the above, pedestrian and bicyclist densities
are high around transit stations. Due to increased pedestrian and bicyclist activity around transit stations, the risk of pedestrian and
bicyclist-involved crashes are expected to be higher in the public transit station service area. Past studies have not explored the as­
sociations between transit usage and pedestrian and bicycle crashes around transit station service areas. This study models the

Fig. 1. Subway Stations of NYC in 2019 with calculated service areas, based on Thiessen polygons truncated to ¼ mile.

3
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

pedestrian and bicycle crashes within the subway station service area to identify and quantify the effects of subway usage and area
characteristics (i.e., demographic and built environment) to address these research gaps. Further, as the following section presents, this
study aims to identify a well-performing method that can be used to prioritize high-risk areas to improve pedestrian and bicycle safety
and support the identification of appropriate countermeasures by comparing several methodologies.

3. Methods

3.1. Study area

This study used pedestrian and bicycle crash data from New York City area. New York City is the most densely populated large city
in the US. There are five boroughs in NYC: Brooklyn, Bronx, Manhattan, Queens, and Staten Island. Subway stations are located in four
of the five boroughs of NYC, and Staten Island is not connected to the subway network. Therefore, only the remaining four boroughs
were considered in the analysis. The subway system of NYC is the biggest in North America; every day, approximately 5.6 million
people use the subway as their daily mode of transport (Metropolitan Transportation Authority, 2020). The map of the study area and
the locations of the subway stations are shown in Fig. 1. The public and shared transportation network of the area is multimodal and
includes, in addition to the subway, buses and bike-sharing services.

3.2. Data collection and summary statistics

Pedestrian and bicycle crash data for the year 2019 were collected in this study. In addition, socio-demographic characteristics,
built environment, bike-sharing, and subway ridership data were collected to the same spatial and temporal extent. The pedestrian and
bicycle crash data and other variables were aggregated to the subway station service area defined in section 3.7 using ArcGIS Pro. The
summary statistics of the collected dependent and independent variables are shown in Table 1. Additional details about the dependent
and independent variables are provided in the following subsections.

3.3. Pedestrian and bicycle crash characteristics

For this study, the two dependent variables, modeled separately, are the pedestrian crash counts and bicycle crash counts at NYC
subway station service areas. To create the crash count variables, all pedestrian and all bicycle crashes, including fatal, injury, and non-
injury crashes, were aggregated to the subway station service area level. From January 2019 to December 2019, in NYC, there were a
total of 15,124 pedestrian and bicycle-involved fatal and injury crashes, among which 10,711 involved pedestrians and 4,413 involved
bicyclists. During this period, there were 97 fatal pedestrian and 10 fatal bicyclist crashes. Among the five boroughs, Queens expe­
rienced almost 42% of the total pedestrian fatalities in NYC, the highest number among the boroughs. Brooklyn, Bronx, and Queens
experienced two cyclist fatalities each in 2019, whereas there were three cyclist fatalities in Manhattan in the same year. In terms of
injury crashes, Brooklyn experienced the highest number of pedestrian and bicycle injury crashes in 2019 (3,591 and 1,773,
respectively).

3.4. Subway station and ridership data

One of the objectives of this study was to analyze the relationship between subway ridership, and pedestrian and bicyclist crashes
close to subway stations. Subway ridership is used in this analysis as a measure of exposure for pedestrians and bicyclists in the vicinity
of the subway station location. It is expected that higher subway ridership will correspond to more pedestrians and bicyclists traveling
in and out of the subway stations. The high pedestrian and bicyclist activity increases the exposure to roadway traffic and thus

Table 1
Summary Statistics of the selected Explanatory variables (N = 424).
Variable Mean Std. Dev Min Max

Dependent Variables
Bicycle Involved Crash Count 6.53 5.21 0 24.00
Pedestrian Involved Crash Count 12.22 7.84 0 52.00
Independent Variables
Number of Bus Stops 11.69 7.89 0 40.00
% of Households Without a Vehicle 60.25 16.77 0 90.50
Employment Density (#/per acre) 51.01 46.52 0 577.04
Subway Ridership (thousands/per day/station) 12.96 18.79 0 243.36
Intersection Density 107.88 58.59 0 584.63
Park Area (Acre) 6.43 11.98 0 74.77
Number of Schools 2.26 2.89 0 14.00
Citi Bike End Trips 200.27 118.23 0 680.88
Length of Bike Lanes (100 ft) 41.10 31.90 0 164.38
Number of Citi Bike Station 1.42 1.88 0 7.00
Number of DOT Bike Racks 42.35 43.39 0 283.00

4
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

increases the risk of crashes. The shapefile containing the NYC subway station’s geographic locations was collected from the NYC Open
Data portal, and the station-wise weekday ridership data from the Metropolitan Transportation Authority (MTA) website. There are
424 subway stations in NYC when stations connected by transfers are counted as single stations. The average daily ridership per
subway station was 12.95 thousand in 2019 (Table 1).

3.5. Socio-demographic data

Urban landscape and roadway characteristics in a metropolitan area largely depend on the nearby area’s socio-demographic
characteristics, affecting pedestrian and bicycle crashes. Socio-demographic characteristics-employment density and vehicle owner­
ship were collected from the Census Bureau website. These data were collected at the census tract level and then aggregated to the
subway station service area level using the “summarize within” tool in ArcGIS. When a subway station service area overlaps two census
tracts, the socio-demographic characteristics for that subway station service area were assigned as the average of the socio-
demographic characteristics of the overlapping census tracts. A summary description of these data is provided in Table 1.

3.6. Bike infrastructure data

Bike lanes can be considered as a safety proxy for bicyclists. Bike lanes provide a safe driving environment for bikers, and people are
more likely to use personal or bike-sharing services if the system provides a safe environment (NYCDOT, 2017). On the other hand,
bike racks and Citibike stations can be used as a proxy for bicyclists’ exposure to traffic. The presence of more bike racks and Citibike
stations indicates more bike users due to demand and supply issues and, therefore, traffic exposure. Three variables related to bike
infrastructure (i.e., length of bike lanes, number of bike racks, and number of Citibike station count) were considered in the model
development to investigate the effects of these factors on bicyclist safety. Summary statistics of the collected three bike
infrastructure-related variables are shown in Table 1. Approximately 85% of subway stations have bike lanes within their service areas,
and the average length of the bike lanes in a subway station service area is 4,110 ft.

3.7. Definition of subway service area

The subway station service/catchment area was defined as the area around a subway station from where any subway user can walk
to that station comfortably (Zhang et al., 2019). Areas beyond the catchment area of a subway station were not considered to affect the
ridership of that station. Past studies have used several distances to define the catchment area of a public transit station (Noland et al.,
2016; Faghih-Imani and Eluru, 2015; Ma et al., 2015). This study used a 400 m (or a quarter mile) radius of Thiessen polygon or
Voronoi cells from the station or the equivalent of a 5-min walking distance to define the catchment area, following literature practices
(Noland et al., 2016; Ma et al., 2015; Ashraf et al., 2021). The subway station catchment areas are shown in Fig. 1.

3.8. Model selection

The pedestrian and bicycle crash count data are non-negative integers for which count models such as Poisson models are better
suited (Xu and Huang, 2015). Moreover, the pedestrian and bicyclist crash count data were overdispersed (i.e., the variance exceeds
the mean of the crash counts) (Lord and Mannering, 2010; Xu and Huang, 2015), as confirmed by an overdispersion test on NYC data
based on Gelman and Hill (2006). Due to overdispersion, Poisson models may produce biased and inconsistent parameter estimations,
and Negative Binomial (NB) models are used to address this issue (Ashraf and Dey, 2022). In addition, traffic crash frequency count in a
geographic area tends to exhibit spatial dependency (Aguero-Valverde and Jovanis, 2006; Huang and Abdel-Aty, 2010; Xu and Huang,
2015). Based on the observed characteristics of NYC data and recommendations of past studies, the following three types of models
were developed and tested to accommodate the structure of the NYC data: (i) Bayesian Negative Binomial Conditional Autoregressive
(NB CAR) model, (ii) Bayesian Random Parameter Negative Binomial (RPNB) model, and (iii) Semi-Parametric Geographically
Weighted Poisson Regression (S-GWPR) model.

3.8.1. Bayesian Negative Binomial Conditional Autoregressive model (NB CAR)


This study first developed an NB CAR model under the Bayesian framework to incorporate correlation heterogeneity or spatial
dependency into the modeling structure. In this model, an extra error term φi is added in the traditional NB model equation that
accounts for the spatial autocorrelation. The NB model can be expressed as follows, where crash count (i.e., pedestrian and bicycle
crashes) at the service area level was modeled as a Poisson distribution:
yi ∼ Poisson (θi ) (1)
In equation (1), yi is the observed number of traffic crashes at the subway station i’s service area, and θi is the expected Poisson rate
for station i. The Poisson rate can be expressed in the logarithm format as follows:

log(θi ) = β0 + βk Xik + εi (2)
k

Where, β0 is the intercept of the model, βk is the vector of estimated parameters, Xik is the vector of independent variables, and EXP(εi )

5
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

is the gamma-distributed error term with mean 1, and variance α (Xu and Huang, 2015). The functional form of the NB CAR model then
can be written as follows:

log(θi ) = β0 + βk Xik + εi + φi (3)
k

Where, εi , β0 , βik , and Xik bear the same meaning as described in equation (2), and φi is the spatial random effect error term for each
subway station i. The NB CAR model was developed under Bayesian inference using the Markov Chain Monte Carlo (MCMC) estimation
procedure in Crimestat V4.02 developed by Levine (2015). Non-informative priors for the Bayesian estimation are used for the model
parameters. Based on the correlogram test and neighboring structure, the negative exponential decay of − 1 was used for the Bayesian
NB CAR models (Levine, 2015).

3.8.2. Bayesian Random Parameter Negative Binomial model (RPNB)


Although the NB CAR model described above contains a spatial dependency structure in the modeling framework, the parameter
estimates from this model still represent global variation. The parameter estimates from the NB model can vary from one subway
station service area to another to account for local variations in the data, which leads to the RPNB model. The following equation can
represent the RPNB model:

log(θi ) = βi0 + βik Xik + εi (4)
k

βik = βk + φik (5)

Where, βik is the parameter estimates of the k-th explanatory variable for subway station service area i, and φik is a randomly
distributed term with a mean 0 and variance σ 2k . The RPNB model was developed using the package INLA in R (Rue et al., 2009).
Non-informative priors are used for model development.

3.8.3. Semi-Parametric Geographically Weighted Poisson Regression (S-GWPR)


Another approach to account for spatial heterogeneity in the traffic crash count data is the use of the Semi-Parametric
Geographically Weighted Poisson Regression (S-GWPR) model, which allows for varying the co-efficient estimates over the study
area. The S-GWPR model can be presented as follows:

l ∑
K
ln(θi ) = β0 (ui , vi ) + βj Xij + βk (ui , vi )Xik (6)
j=1 k=l+1

Where, βj is the jth global coefficient and θi is the expected number of traffic crashes at subway station i, (ui , vi ) is the location of a
subway station, and βk (ui , vi ) is a function of coordinates of the center of the subway station service area i.
Finally, the overall complexity and goodness-of-fit for the models were assessed by measuring the Akaike Information Criteria
(AICc) and the Bayesian Information Criterion (BIC) values. The S-GWPR model was developed using the software GWR4.0 which was
developed by Nakaya et al. (2012). Details of three modeling techniques used in this study can be found in Xu and Huang (2015).

4. Results and discussion

Following the proposed methods, two sets of three models, (i) a CAR model, (ii) an RPNB model, and (iii) an S-GWPR model, were
developed separately for pedestrian and bicycle crashes. The following sections present the comparison of the models in terms of
goodness-of-fit measures and a discussion of the parameter estimates from the best-performing model.

4.1. Model comparison

The CAR models were developed under the Bayesian framework and implemented using a Markov Chain Monte Carlo (MCMC)
estimation procedure. For the estimation, the number of iterations was 25,000, and the number of burn-in samples for each iteration

Table 2
Measures of Model goodness-of-fit.
Pedestrian Crash Model Bicycle Crash Model

AICc BIC MAD BIC AICc MAD

Bayesian CAR 3468 3496 4.70 3148.1 3108.2 3.03


Bayesian RPNB 2851 2879 4.92 2539.6 2499.7 3.00
S-GWPR 897.8 1222.3 3.23 1025.5 882.7 3.02

N 424 424

AICc = Corrected Akaike Information Criteria; BIC=Bayesian Information Criteria; MAD = Mean Absolute Deviance.

6
M.T. Ashraf et al.
Table 3
Parameter estimates from models with pedestrian crash frequency as the dependent variable.
Pedestrian Crashes Semi-Parametric Geographically Weighted Poisson Regression Model Bayesian Conditional Autoregressive Model Bayesian Random Parameter Model

Variables Coef. Std. Dev Min Max Coef. Std.Dev 95% Credible Interval Coef. Std.Dev 95% Credible Interval

2.50% 97.50% 2.50% 97.50%

Intercept 0.920 0.98 − 1.483 2.760 1.634 0.118 1.402 1.866 1.543 0.117 1.312 1.771
Number of Bus Stops 0.023 0.004a 0.023 0.003 0.017 0.030 0.025 0.004 0.017 0.032
Subway Ridership 0.027 0.023 − 0.013 0.119 0.006 0.002 0.003 0.009 0.008 0.002 0.004 0.012
% of household without vehicle 0.020 0.014 − 0.004 0.055 0.011 0.002 0.007 0.016 0.012 0.002 0.007 0.016
Employment Density − 0.005 0.007 − 0.024 0.025 − 0.005 0.001 − 0.007 − 0.003 − 0.005 0.001 − 0.007 − 0.004
Number of School − 0.004 0.04 − 0.119 0.093 0.012 0.009 − 0.007 0.030 0.011 0.009 − 0.007 0.030
7

Total Park Area − 0.004 0.022 − 0.089 0.060 − 0.007 0.002 − 0.012 − 0.003 − 0.007 0.002 − 0.012 − 0.003
Intersection Density 0.0006 0.003 − 0.0059 0.0090 0.0003 0.0005 − 0.0006 0.001 0.0001 0.0005 − 0.0008 0.0011
Spatial Component (phi) 0.0069 0.013 − 0.017 0.035
Unobserved Effects

Intercept 5.09 1.43


% of household without vehicle 1.09 0.308
Number of Bus Stops 1.11 0.312
Subway Ridership 1.09 0.307
Number of School 1.1 0.309

Note: Italicized bold coefficients are significant at 95% significance level.


a
Geographically fixed variable.

Journal of Transport & Health 27 (2022) 101529


M.T. Ashraf et al.
Table 4
Parameter estimates from models with bicycle crash frequency as the dependent variable.
Bicycle Crashes Semi Parametric Geographically Weighted Poisson Regression Model Bayesian Conditional Autoregressive Model Bayesian Random Parameter Model

Independent Variables Coef. Std. dev Min Max Coef. Std.dev 95% Credible Interval Coef. Std.dev 95% Credible Interval

2.5% 97.5% 2.5% 97.5%

Intercept 0.6774 0.541 − 1.3056 1.8603 0.5792 0.16 0.2580 0.8979 0.5009 0.1652 0.1736 0.8225
Number of Bus Stops 0.0252 0.010 − 0.0084 0.0458 0.0216 0.004 0.0131 0.0302 0.0235 0.0045 0.0147 0.0323
Subway Ridership 0.0026 0.007 − 0.0069 0.0192 0.0021 0.001 − 0.0016 0.0060 0.0031 0.0022 − 0.0011 0.0076
% of HH without vehicle 0.0168 0.005 0.0011 0.0282 0.0138 0.003 0.0084 0.0193 0.0131 0.0028 0.0076 0.0185
Employment Density − 0.0045a 0.0010 − 0.0019 0.001 − 0.0040 0.0002 − 0.0017 0.0011 − 0.0039 0.0004
Number of School 0.000002 0.0000010 0 0.000007 − 0.0081 0.012 − 0.0318 0.0156 − 0.0077 0.0122 − 0.0318 0.0162
Length of Bike lanes (100’s of ft) − 0.0004a 0.0008 − 0.0002 0.001 − 0.0025 0.0020 − 0.0001 0.0012 − 0.0024 0.0022
8

Intersection Density 0.0003 0.001 − 0.0031 0.0055 0.0002 0.0005 − 0.0008 0.0013 0.0001 0.0005 − 0.0010 0.0012
Citi Bike trips − 0.0010 0.0009 − 0.0022 0.0031 − 0.0010 0.0004 − 0.0019 − 0.0002 − 0.0011 0.0004 − 0.0019 − 0.0003
Citi bike Station Count 0.0448 0.0408 − 0.0455 0.1316 0.0842 0.0224 0.0404 0.1284 0.0842 0.0224 0.0402 0.1281
DOT bike racks 0.0039 0.001 0.0012 0.0081 0.0056 0.0009 0.0039 0.0073 0.0055 0.0009 0.0038 0.0072
Spatial Component (phi) − 0.0052 0.0109 − 0.0364 0.0076
Unobserved Effects

Intercept 1.84 0.53


% of household without vehicle 1.12 0.32
Number of Bus Stops 1.13 0.33
Subway Ridership 1.1 0.31

Note: Italicized bold coefficients are significant at 95% significance level and the bold variables are significant at 90% significance level.

Journal of Transport & Health 27 (2022) 101529


a
Geographically fixed variable.
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

was 5000. Non-informative priors were used for the prior specification, and a negative exponential decay of − 1 miles, as presented in
section 3.8.1was used to control for spatial dependency among the subway stations. The convergence of the models was checked using
the Gelman-Rubin (G-R) test statistics. The G-R statistic for each coefficient was below 1.2, indicating the convergence of the models.
The spatial autocorrelation parameter was not statistically significant in both pedestrian and bicycle crash CAR models, which suggests
that the spatial autocorrelation was adequately controlled in our models (Noland et al., 2016). The RPNB model was developed under
the Bayesian framework. In the model development, non-informative priors were used for the parameter distribution. Finally, fixed
and adaptive kernels were applied to develop a weight matrix for the S-GWPR models. For both pedestrian and bicycle crash models,
the adaptive kernel models produced lower AICc values. The optimal bandwidth of the adaptive kernel was found to be 64.
The goodness-of-fit measures of the developed models are provided in Table 2. A common rule of thumb in the use of AICc is that a
difference of more than 2 between two comparing models indicates a substantial difference in performance, and the model with a
lower AICc value is better (Fabozzi et al., 2014; Burnham and Anderson, 2004). Similarly, a difference of 10 in BIC value indicates a

Fig. 2. Distribution of the parameter estimates from GWPR model using Pedestrian Crash as the dependent variable (a) Employment Density; (b)
Subway Ridership; (c) % of Households without Vehicle; (d) Total Park Area.

9
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

substantial difference in performance, and the model with the lower BIC value is better (Fabozzi et al., 2014; Burnham and Anderson,
2004). As shown in Table 2, AICc, BIC, and MAD values indicate that the S-GWPR model outperforms the other two models sub­
stantially for pedestrian crash data. For the bicycle crashes, all three models performed similarly in explaining the variations in the
bicycle crash data, which can be seen from the MAD values in Table 2. However, the AICc and BIC values were substantially lower for
the S-GWPR model than the other two models, indicating a better fit.

4.2. Parameter estimates

Parameter estimates from the developed models are summarized in Tables 3 and 4. The Bayesian CAR model only shows the fixed
effects of the parameter estimates. In the RPNB model, fixed and unobserved effects from varying coefficients are also shown. On the
other hand, in the S-GWPR model, the mean value of the spatially varying coefficients and the range of the parameter estimates are

Fig. 3. Distribution of the parameter estimates from GWPR model using Bicycle Crash as the dependent variable (a) Subway Ridership; (b) Citibike
Station Count; (c)% of Households without Vehicle; (d) Number of Bus Stops.

10
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

shown.
For the pedestrian crashes, the same five variables were found statistically significant in both the Bayesian CAR and Bayesian RPNB
models, and the parameter estimates were comparable. More specifically, the number of bus stops, the percentage of households without a
vehicle, employment density, total park area, and the subway ridership were significant. In the final RPNB pedestrian crash model, the
percentage of households without a vehicle, number of bus stops, subway ridership, and number of school variables, including the intercept,
were allowed to vary from one subway station service area to another. On the other hand, for the S-GWPR pedestrian crash model, all
variables except the number of bus stops variable were found to be geographically varying.
In the bicycle Bayesian CAR model, six variables were found significant. These variables were the number of bus stops, the subway
ridership, the percentage of households without a vehicle, the number of Citi Bike trips, Citi Bike station counts, and DOT bike racks. All but one
(i.e., subway ridership) of these variables were also found significant in the Bayesian RPNB model. In the final RPNB bicycle crash
model, the variables of the percent of households without a vehicle, subway ridership, the number of bus stops, and the intercept were treated
as random parameters. In the S-GWPR bicycle crash model, all variables except employment density and length of bike lanes were found
geographically varied. The spatial variations of the parameters within the boroughs of the New York City area are shown in Figs. 2 and
3.
As discussed in the model comparison section, the S-GWPR models performed better in model goodness-of-fit and overall model
complexity measures for pedestrian and bicycle crash data models. In light of this result, the discussion on the estimated parameters
presented in the next section is based on the S-GWPR models.

4.2.1. S-GWPR model for pedestrian crashes


As shown in Table 3, only the number of bus stops was geographically fixed over the study area, which indicates that keeping this
variable global and all other variables local would give the best fit. This variable was positively associated with the pedestrian crash
frequency in a subway station service area with a coefficient estimate of 0.023. Previously, several studies found a positive association
between pedestrian crash risk and bus stop counts (Su et al., 2021; Almasi and Behnood, 2022). However, these studies focused on
overall pedestrian crash risks and did not account for the increased risk of crash near the public transit stops and service area. The
estimated mean coefficient value of the variable indicates that one additional bus stop within a subway station service area was
associated with a 2.3% increase in the expected number of pedestrian crash frequencies. The positive relationship between the number
of bus stops and pedestrian crashes can result from the expected increased pedestrian activity at bus stops and their exposure to
surrounding traffic, as transit users often walk for the first- and last-mile connection to bus stops. Furthermore, it is likely that many
people use buses to connect to subway stations and walk in between the bus stops and the closest subway station, which can also lead to
an increased number of pedestrian crashes. Additional safety measures such as reduced speed zone, enhanced pedestrian crossing, and
speed humps can be implemented to improve pedestrian safety at bus stops.
Other variables in the S-GWPR pedestrian crash model showed significant variability over the study area. The variability of the
estimated coefficients for subway ridership, employment density, percent of households without a vehicle, and total park area variables is
shown in Fig. 2. The mean value of the spatially varying coefficients indicates that subway ridership, percent of households without a
vehicle, and intersection density were on average positively associated with pedestrian crash frequency in the subway station service
area. In addition, employment density, number of schools, and total park area were on average negatively associated with pedestrian crash
frequency.
The subway is the most used public transportation mode in NYC, and subway stations work as hubs for public transit users and
other public transportation modes (e.g., bus, taxi). An increase in subway ridership creates more pedestrian activities in subway service
areas. As expected, subway station ridership was positively associated with higher pedestrian crashes. The mean value of the
geographically varying coefficients for the subway ridership was 0.027 with a standard deviation (s.d) of 0.023. A higher ridership was
associated with an increase in the pedestrian crash frequency at 95% of the subway stations and a decrease in the pedestrian crash
frequency at the remaining 5% of the subway stations (Fig. 2b). The mean value of the local coefficients for this variable indicated that,
on average, there was a 2.7% increase in the expected number of pedestrian crash frequencies for a one thousand increase in daily
subway ridership. The spatial distribution of the local coefficients (presented in Fig. 2b) showed that subway stations located in the
Bronx, Brooklyn, and Queens had higher values of the coefficients compared to Manhattan. This finding suggests that the effect of
subway ridership on pedestrian crashes is low in Manhattan.
Households that do not have access to a personal vehicle depend on public transportation for their daily movement. Thus, a higher
percentage of households in a subway station service area without access to a vehicle could increase the pedestrian crash frequency.
The spatial distribution of the local coefficient estimates of this variable showed that the impact was positive in almost all subway
stations, whereas only 2% of stations had a negative relationship (Fig. 2c). The mean value of the estimated local coefficients for all
subway stations was 0.020 with a s.d of 0.014, which means that for a 1% increase in the average percentage of households without a
vehicle there was, on average, a 2% increase in the pedestrian crash frequency within subway station service areas. The positive
relationship between vehicle ownership and pedestrian crash risk is similar to the previous study by Siddiqui et al. (2012). The co­
efficient values were relatively higher in Brooklyn (denoted by the dark red color in Fig. 2c) compared to the other three boroughs. This
result means that subway station service areas in Brooklyn yielded a less safe environment for pedestrians who did not own a vehicle.
The total park area in a subway station service area had both positive and negative impacts on pedestrian crash frequency, although
it had a negative impact on the global model. The mean value of this coefficient was − 0.004, with an s.d of 0.022. This result suggests
that a higher value of the total park area was associated with a decrease in the pedestrian crash frequency for 58% of the subway
station service areas and an increase in the pedestrian crash frequency for 42% of the subway station service areas (Fig. 2d). As seen
from Fig. 2d, the total park area mainly had a negative impact on the pedestrian crash frequency in Queens. The local coefficients had

11
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

mixed effects (i.e., positive and negative) in Brooklyn, Bronx, and Manhattan (Fig. 2d). The mixed relationship between total park area
and pedestrian crash frequency can be explained by the fact that more park areas for recreational activities in a highly built envi­
ronment like Manhattan can absorb pedestrian exposure to traffic, reducing the chance of pedestrian crashes. On the other hand, as
parks can induce more pedestrian activities, lack of infrastructure (e.g., pedestrian crossing, pedestrian signals, median) near some of
the parks could lead to more pedestrian-involved crashes. The subway station service areas where the total park area had a positive
association with pedestrian crashes require more detailed analysis to find deficiencies in existing infrastructures.
Finally, the employment density variable primarily had a negative association with pedestrian crash frequency. The mean value for
this coefficient was − 0.005, with an s.d of 0.007. The result suggests that for 80% of the station service areas, a higher employment
density was associated with a decrease in the pedestrian crash frequency, and for 20% of them, a higher employment density was
associated with an increase in the pedestrian crash frequency (Fig. 2a). On average, a one-unit increase in employment density resulted
in a 0.50% decrease in pedestrian crash frequency. Previously, few studies reported a positive association between employment
density and pedestrian crash counts, where high employment density led to high pedestrian exposures (Cai et al., 2016; Bernhardt and
Kockelman, 2021). However, the result from the current study indicates that pedestrian crash risk had a negative association with
employment density in most subway station service areas. This suggests that higher employment density within several subway station
service areas increased the frequency of pedestrian crashes. In Queens, 50% of the subway station service areas were positively
associated with employment density and pedestrian crash frequency (Fig. 2a). Fig. 2a shows a positive association between pedestrian
crash frequency and employment density in other boroughs in a few stations. For all subway stations in Manhattan, the estimated
coefficients for employment density were found negative, denoted by different shades of green for varying intensity in Fig. 2a.

4.2.2. S-GWPR model for bicycle crashes


The S-GWPR bicycle crash model showed that the employment density and length of bike lanes variables were not significantly varying
over the study area. These results imply that keeping these two variables global and all other variables local gave the best fit model.
Different aspects of bike infrastructure can work as an exposure proxy and others as a safety proxy variable. As expected, therefore, the
result from our bicycle crash model showed mixed associations. Specifically, bicycle crashes were negatively associated with the length
of bike lanes but positively associated with Citibike Station count and the number of DOT bike racks. This result is intuitive as bike
lanes indicate a safe environment for bicyclists. More Citibike stops and DOT bike racks indicate higher bicyclist volume. This was
likely the case of our model, as it was found that both of these variables positively affect the number of bicycle crashes, which means
these variables worked as an exposure proxy variable. The coefficient estimate for employment density and length of bike lanes was
− 0.0045 and − 0.0004, respectively. For a one-unit increase in employment density and a 100 ft increase in the length of bike lanes, there
was a 0.45% and 0.040% decrease in the expected bicycle crash frequency, respectively.
Several variables specific to bicycle use were included in the bicycle crash model. The choropleth map containing five intervals was
plotted to investigate the borough-wise variation of the geographically varying coefficients, which are presented in Fig. 3. Similar to
the pedestrian crash frequency model, subway ridership was positively associated with bicycle crash frequency in the majority of the
subway stations. The mean value of the local coefficient for this variable was 0.0026 with an s.d of 0.007. A higher ridership was
associated with an increase in the bicycle crash frequency in 56% of the station service areas (Fig. 3a) and a decrease in bicycle crashes
in 44% of the areas. The spatial distribution of the local coefficient estimates (Fig. 3a) revealed that the coefficients had much higher
values in the Bronx than in the other three boroughs. The comparatively high values of the coefficient estimate for the subway ridership
variable in the Bronx might be because they implemented fewer safety countermeasures within subway station service areas compared
to Manhattan, Brooklyn, and Queens subway stations. Most of the subway stations in Brooklyn, Manhattan, and Queens showed a
negative association between bicycle crashes and subway ridership. The mean value of the local coefficient estimates for this variable
indicates that, on average, one thousand increase in average subway ridership was associated with a 0.26% increase in the expected
bicycle crash frequency. Subway stations in the Bronx with low average subway ridership showed a high marginal effect. Additional
safety countermeasures, such as bike priority areas, speed humps, bike lanes, and turn traffic calming, could improve bicycle safety in
the Bronx.
Coefficient estimates for three bike infrastructure-related variables were found to be significantly varying geographically. The DOT
bike racks variable had a positive relationship at all the subway stations in all four boroughs, a finding in line with results from the
literature. For example, Xie et al. (2021) found a positive association between bicycle crash risks and bike rack density in NYC. The
mean coefficient value of the DOT bike racks variable was 0.0039 (s.d was 0.001). There was a one-unit increase in the average number
of DOT bike racks within the subway station service area for a 0.39% increase in the bicycle crash frequency. The mean value of the
local coefficients for Citibike ridership was found negative (mean value of − 0.001 and s.d. of 0.0009), although the distribution showed
that positive relationships with bicycle crash frequency were in some cases found (9% had a positive association). Spatial distribution
showed that the Citibike Station count variable was positively associated with bicycle crash frequency in most subway stations in all
four boroughs. Only 10% had a negative association (Fig. 3b). The mean value of the Citibike Station count variable was 0.0448 with an
s.d of 0.0408, which means that for a one-unit increase in the average number of Citibike Station count within the subway station
service area, there was a 4.58% increase in the bicycle crash frequency. Bronx’s local coefficient estimates were much higher than the
other three Boroughs, which is denoted by the dark red color in Fig. 3b. Higher Citibike ridership could increase bicycle crash exposure,
which could increase bicycle crash frequency. However, higher Citibike ridership can also be a surrogate measure of bike-friendly
communities where people feel safer using bicycles for their trips.
The percentage of households without a vehicle had a positive association with bicycle crash frequency. The mean value of this
coefficient for this variable was 0.0168 (s.d was 0.005), which indicates that for a 1% increase in the percentage of households without
a vehicle, there was a 1.69% increase in the bicycle crash frequency in the subway station service area on average. The distribution of

12
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

the estimated local coefficients showed that all subway stations had a positive association with this variable. The spatial distribution of
the local coefficients for the households without a vehicle is shown in Fig. 3c, which shows that almost all subway stations were
positively associated with household vehicle availability. However, several subway stations in Brooklyn had a high value of the local
coefficients, meaning those subway station service areas yielded a less safe environment for households that did not own a vehicle. The
number of bus stops within the subway station service area was positively associated with bicycle crash frequency in all subway stations
except five stations in Queens. The mean value of the local coefficients was 0.025 (s.d = 0.010), which means that, on average, one
additional bus stop near a subway station could increase the bicycle crash frequency by 2.53%. The distribution of the local coefficients
for this variable showed a similar pattern as observed for the percentage of households without a vehicle variable (Fig. 3d). Several
subway stations in Brooklyn and Queens had a high magnitude of the estimated coefficients compared to the other two boroughs,
denoted dark red in Fig. 3d.

5. Case study and policy implications

5.1. Case study implications

The geographically varying coefficient estimates of pedestrian and bicycle crash models showed local variations in magnitude and
direction, presenting positive or negative correlations between explanatory variables and crash counts. Subway stations with positive
correlations were separated using the geographically varying coefficient estimates. For these subway stations, all the explanatory

Fig. 4. Prioritized areas for safety improvement.

13
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

variables contribute to an increase in crash counts and are prioritized for safety improvements. Thirty-seven subway stations (19
stations in the pedestrian crash model, 16 stations in the bicycle crash model, and 2 in common) were selected in the prioritization list
for safety improvements. A map of the prioritized subway stations is shown in Fig. 4. Most of the selected subway stations (33 out of 37)
are located in Brooklyn and Bronx. Only one is located in Manhattan, and three are located in Queens. In Fig. 4, safety initiatives taken
by NYCDOT, such as bike priority areas, street improvement projects, and enhanced crossing implementation location frequencies, are
shown. Fifty percent of the prioritized subway stations are located within the bike priority area (green polygon in Fig. 4). NYC DOT
defined bike priority areas as the neighborhoods with high numbers of bicyclists KSI (Killed or Seriously Injured) and limited dedicated
bicycle facilities. NYC DOT used these areas for bicycle network expansion projects (NYCDOT, 2017). Furthermore, Fig. 4 shows that
some of the street safety improvement initiatives by engineering intervention/enforcement (e.g., street improvement and enhanced
crossing in Fig. 4) were not implemented in the prioritized areas (i.e., Green Polygon in Fig. 4), which means the prioritization projects
need to be considered in these locations for timely implementations.

5.2. Policy implications

Transportation safety is one of the highest priorities when designing urban landscapes. The growing popularity of active trans­
portation modes increases users’ vulnerability to traffic crashes. Thus, it is essential to identify associated crash contributing factors to
develop policies for targeted safety countermeasures.
Ridership volume was positively associated with pedestrian and bicycle crashes, indicating the lack of appropriate facilities at
locations with high pedestrian and bicycle volume. The area around subway stations with high ridership should implement more safety
countermeasures. For example, this study found that increased bike lane length was associated with a reduction in bicycle crashes.
NYCDOT has expanded bicycle lanes substantially over the last ten years and reported fewer bicycle crashes (NYCDOT, 2017). Based
on this observed effectiveness, NYCDOT can focus on implementing additional bike lanes, especially near subway stations. Park and
recreational areas within the subway station service area were found to reduce pedestrian exposure to traffic and the number of
crashes. In this regard, enhanced pedestrian crossing, lower speed limit, sidewalks, and other pedestrian facilities can be provided in
and out of the park and recreational areas. Similarly, the number of schools in transit station service area was negatively associated
with pedestrian crashes, which could be attributed to the better traffic control and road safety education that induced safety awareness
among all roadway users (Su et al., 2021). Thus, localized better traffic control design and management around transit station service
area can improve the active transportation mode users’ safety.
The findings of this study also provide insights into the effectiveness of recent innovative transportation infrastructure design
policies, such as complete streets and vision zero. Several variables considered in this study (e.g., bike lanes, number of bus stops, park
area, intersection density) are integral components of the complete street design strategies. Pedestrian and bicyclist safety can be
improved by implementing complete street design strategies (Kim, 2019). Safety countermeasures such as narrower lanes, protected
signal phasing for pedestrian crossing, speed humps, left turn traffic calming, and signal retiming can be implemented on road net­
works near public transit stops to improve pedestrian and bicyclist safety (Congiu et al., 2019; Kim, 2019). Other than engineering
countermeasures, educational programs and media campaigns can be implemented to promote more responsible driving and walking
behavior among multimodal roadway users (Congiu et al., 2019; Morris et al., 2019).
In addition, all stakeholders (including legislators, planners, health agencies, law enforcement agencies, community organizations,
transportation service providers, and road users) should be involved in the planning process in safety planning for active transportation
modes (Congiu et al., 2019). A comprehensive pedestrian and bicyclist safety improvement plan should combine appropriate laws and
regulations, stringent law enforcement for the rider and driver’s risky behaviors, engineering countermeasures, and safety education
campaigns.

6. Research contributions

The contribution of this research is threefold. First, this paper adopted an innovative subway service area concept as the
geographical unit in developing an in-depth understanding of pedestrian and bicycle safety near transit stops/hubs. Widely used
geographical units in traffic safety research, such as Census tracts or TAZs, are suitable for integrating results into the overall trans­
portation planning process. However, geographical units such as subway service areas are more practical to focus safety improvements
on a high-risk area of a region and efficiently use the limited resources. Thus, this study contributed to the existing literature by
analyzing pedestrian and bicycle crashes at the subway station service area level. Second, this study developed three statistical models
incorporating spatial heterogeneity and spatial dependency among the subway station service area level crash frequencies and
compared the result. The results showed that the Semi-Parametric Geographically Weighted Poisson Regression (S-GWPR) model
performed better in explaining the spatial component in the subway station service area level pedestrian and bicyclist crash data. In
addition, this method was able to facilitate the identification of hotspots. Third, this study contributes to the current practices by
identifying high-risk subway station service areas. In recent years, public transit ridership has been decreasing in NYC. In addition to
increasing the quality of sustainable public transportation services, safety improvement near transit stations is necessary to promote
safe access to public transportation modes. This study prioritized areas for immediate safety improvements and thus contributes to
maintaining the overall sustainability of the public transportation system.

14
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

7. Conclusions

Walking and bicycling are two active modes of transportation that can facilitate a sustainable and environmentally friendly
transportation system. This study investigated factors associated with the subway network, socio-demographic, and built environment
characteristics in subway station service areas in NYC and their impacts on the number of pedestrian and bicycle crashes. Three types
of models were developed to estimate pedestrian and bicycle crash frequency separately: (i) Bayesian Conditional Autoregressive
(CAR), (ii) Semi-Parametric Geographically Weighted Poisson Regression (S-GWPR) model, to account for potential spatial autocor­
relation and spatial heterogeneity, and (iii) non-spatial Bayesian Random Parameter model (RPNB). The developed models were
compared using the three goodness-of-fit measures: AICc, BIC, and MAD. Based on the model goodness-of-fit measures, the S-GWPR
model performed best for pedestrian and bicycle crash frequency.
The coefficient estimates from all three models showed that subway ridership was positively associated with pedestrian and bicycle
crash frequency. The subway stations where the subway ridership had a strong positive relationship with pedestrian and bicycle crash
frequency need particular attention as the adjacent areas of these subway stations were less safe for pedestrians and bicyclists.
Transportation planners can prioritize these subway station service areas to implement safety countermeasures. The number of bus
stops in the subway station service area was also positively associated with the pedestrian and bicycle crash frequency.
Despite the high number of subway stations and high subway ridership in Manhattan, borough-wise variation of the local co­
efficients revealed that Manhattan’s stations were relatively safe compared to the other three boroughs. The geographically varying
coefficients can help prioritize areas for improvement. Subway station service areas with a high positive correlation between crash
frequency and crash-related factors can be prioritized, as these areas were less safe for pedestrians and bicyclists. Additionally, these
subway stations might be chosen to conduct corridor-level traffic safety studies to uncover more microscopic factors contributing to
pedestrian and bicycle crashes. The result from this study stresses the need of implementing transportation policies that focuses on
incorporating complete street design elements as an essential part of the transportation infrastructure development process. Moreover,
result suggest that complete street design policies can complement the vision zero goal and reduce the active tranasportation related
crashes.

8. Limitations and future research directions

This paper analyzed how transit station utilization and built environment factors might contribute to pedestrian and bicycle crashes
within subway station service areas. The focus was on examining relationships in the proximity of the subway stations, and thus, data
only within the service area were included in the analysis. Many worthwhile related research questions were, therefore, out of the
scope of this study. For example, an extension of this study could compare crashes around transit stations with places further away
from the transit stations. Such an extension could shed light on how the existence of a transit station might affect crashes. Additionally,
such an analysis might provide insights into potential differences in the impacts of specific built environment factors close and further
away from a station. Comparing pedestrian and bicycle crashes within and out of the service area can be explored in future research by
using several buffer distances and measuring the dissipating influence of factors.
Furthermore, despite an effort to include spatial autocorrelation and spatial heterogeneity in the modeling approach, the pedestrian
and bicycle crashes were modeled separately, and thus, potential common unobserved characteristics between the two dependent
variables were not considered. Nevertheless, there are several factors (both observed and unobserved) that can impact crashes
involving pedestrians and bicyclists (for example, how well-lit an area is). In future research, both crash types can be modeled as a
system to account for common unobserved characteristics and identify common crash contributing factors. Also, more microscopic
models (e.g., focusing on the road segments within subway service areas) can be developed in future studies to incorporate specific
local variables (e.g., speed distribution, geometrical properties of roadway) in the modeling process.
Finally, this study only considered variables related to the built environment and transit system. However, the impacts of
pedestrian and bicyclist characteristics and their behavior (e.g., age, use of cell phones and other distractions, jaywalking, risk per­
ceptions) need to be analyzed to develop a holistic understanding of safety around transit stations. Future studies could focus on
integrating behavioral, vehicular, and roadway characteristics data to develop comprehensive insights and identify associated
improvement strategies and policies. Moreover, future research can perform evidence-based effectiveness analysis to identify and
promote the most effective countermeasures.

Author statement

Md Tanvir Ashraf: Conceptualization, Methodology, Data Curation, Formal analysis, Writing- Original draft preparation. Kakan
Dey: Conceptualization, Methodology, Writing - Review & Editing, Supervision. V. Dimitra Pyrialakou: Conceptualization, Method­
ology, Writing - Review & Editing.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

15
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

Financial disclosure

The Authors did not receive any specific funding for this work.

Data availability

Data used for the research were described in the article.

References

Abdel-Aty, M., Siddiqui, C., Huang, H., Wang, X., 2011. Integrating trip and roadway characteristics to manage safety in traffic analysis zones. Transport. Res. Rec.
2213 (1), 20–28.
Aguero-Valverde, J., Jovanis, P.P., 2006. Spatial analysis of fatal and injury crashes in Pennsylvania. Accid. Anal. Prev. 38 (3), 618–625.
Almasi, S.A., Behnood, H.R., 2022. Exposure based geographic analysis mode for estimating the expected pedestrian crash frequency in urban traffic zones; case study
of Tehran. Accid. Anal. Prev. 168, 106576.
Ashraf, M.T., Dey, K., 2022. Application of Bayesian space-time interaction models for deer-vehicle crash hotspot identification. Accid. Anal. Prev. 171, 106646.
Ashraf, M.T., Hossen, M.A., Dey, K., El-Dabaja, S., Aljeri, M., Naik, B., 2021. Impacts of Bike Sharing Program on Subway Ridership in New York City. Transportation
Research Record, 03611981211004980.
Baek, K., Lee, H., Chung, J.H., Kim, J., 2021. Electric scooter sharing: how do people value it as a last-mile transportation mode? Transport. Res. Transport Environ.
90, 102642.
Bernhardt, M., Kockelman, K., 2021. An analysis of pedestrian crash trends and contributing factors in Texas. J. Transport Health 22, 101090.
Burnham, K.P., Anderson, D.R., 2004. Multimodel inference: understanding AIC and BIC in model selection. Socio. Methods Res. 33 (2), 261–304.
Cai, Q., Lee, J., Eluru, N., Abdel-Aty, M., 2016. Macro-level pedestrian and bicycle crash analysis: incorporating spatial spillover effects in dual state count models.
Accid. Anal. Prev. 93, 14–22.
Cao, J., Duncan, M., 2019. Associations among distance, quality, and safety when walking from a park-and-ride facility to the transit station in the Twin Cities.
J. Plann. Educ. Res. 39 (4), 496–507.
Carter, D.L., Council, F.M., 2007. Factors contributing to pedestrian and bicycle crashes on rural highways. In: Transportation Research Board 86th Annual Meeting
Paper, 07-2457.
Chen, P., 2015. Built environment factors in explaining the automobile-involved bicycle crash frequencies: a spatial statistic approach. Saf. Sci. 79, 336–343.
Congiu, T., Sotgiu, G., Castiglia, P., Azara, A., Piana, A., Saderi, L., Dettori, M., 2019. Built environment features and pedestrian accidents: an Italian retrospective
study. Sustainability 11 (4), 1064.
Fabozzi, F.J., Focardi, S.M., Rachev, S.T., Arshanapalli, B.G., 2014. The Basics of Financial Econometrics: Tools, Concepts, and Asset Management Applications. John
Wiley & Sons.
Faghih-Imani, A., Eluru, N., 2015. Analysing bicycle-sharing system user destination choice preferences: Chicago’s Divvy system. J. Transport Geogr. 44, 53–64.
Fan, A., Chen, X., Wan, T., 2019. How have travelers changed mode choices for first/last mile trips after the introduction of bicycle-sharing systems: an empirical
study in Beijing, China. J. Adv. Transport. 2019.
Gelman, A., Hill, J., 2006. Data Analysis Using Regression and Multilevel/hierarchical Models. Cambridge university press.
González, S.R., Loukaitou-Sideris, A., Chapple, K., 2019. Transit neighborhoods, commercial gentrification, and traffic crashes: exploring the linkages in Los Angeles
and the Bay Area. J. Transport Geogr. 77, 79–89.
Huang, H., Abdel-Aty, M., 2010. Multilevel data and Bayesian analysis in traffic safety. Accid. Anal. Prev. 42 (6), 1556–1565.
Kim, D., 2019. The transportation safety of elderly pedestrians: modeling contributing factors to elderly pedestrian collisions. Accid. Anal. Prev. 131, 268–274.
Kim, J.K., Kim, S., Ulfarsson, G.F., Porrello, L.A., 2007. Bicyclist injury severities in bicycle-motor vehicle accidents. Accid. Anal. Prev. 39 (2), 238–251.
Lee, J., Abdel-Aty, M., Cai, Q., 2017. Intersection crash prediction modeling with macro-level data from various geographic units. Accid. Anal. Prev. 102, 213–226.
Levine, N., 2015. A Spatial Statistics Program for the Analysis of Crime Incident Locations.(V 4.02). Ned Levine & Associates. Houston-National Institute of Justice,
Washington, DC.
Lord, D., Mannering, F., 2010. The statistical analysis of crash-frequency data: a review and assessment of methodological alternatives. Transport. Res. Pol. Pract. 44
(5), 291–305.
Ma, T., Liu, C., Erdoğan, S., 2015. Bicycle sharing and public transit: does Capital Bikeshare affect Metrorail ridership in Washington, DC? Transport. Res. Rec. 2534
(1), 1–9.
Metropolitan Transportation Authority, 2020. Subway and bus ridership for 2019. https://new.mta.info/agency/new-york-city-transit/subway-bus-ridership-2019.
(Accessed 20 February 2022).
Morris, N.L., Craig, C.M., Houten, R.V., 2019. Evaluation of Sustained Enforcement, Education, and Engineering Measures on Pedestrian Crossings.
Nakaya, T., Charlton, M., Lewis, P., Fortheringham, S., Brunsdon, C., 2012. Windows Application for Geographically Weighted Regression Modelling. Ritsumeikan
University, Kyoto, Japan.
National Highway Traffic Safety Administration, 2019. Overview of motor vehicle crashes. https://crashstats.nhtsa.dot.gov/Api/Public/Publication/813060.
(Accessed 20 February 2021).
Ni, Y., Wang, M., Sun, J., Li, K., 2016. Evaluation of pedestrian safety at intersections: a theoretical framework based on pedestrian-vehicle interaction patterns. Accid.
Anal. Prev. 96, 118–129.
Noland, R.B., Smart, M.J., Guo, Z., 2016. Bikeshare trip generation in New York city. Transport. Res. Pol. Pract. 94, 164–181.
NYCDOT. Safer cycling: bicycle ridership and safety in New York city. Available at. https://www1.nyc.gov/html/dot/downloads/pdf/bike-safety-study-
fullreport2017.pdf. 2017.
Peng, B., Erdoğan, S., Nasri, A.A., Zou, Z., 2021. Towards a health-conscious transportation planning: a framework for estimating health impacts of active
transportation at local level. J. Transport Health 22, 101231.
Pulugurtha, S.S., Sambhara, V.R., 2011. Pedestrian crash estimation models for signalized intersections. Accid. Anal. Prev. 43 (1), 439–446.
Pulugurtha, S.S., Srirangam, L.P., 2021. Pedestrian safety at intersections near light rail transit stations. Public Transport 1–26.
Ragaini, B.S., Sharman, M.J., Lyth, A., Jose, K.A., Blizzard, L., Peterson, C., Johnston, F.H., Palmer, A., Aryal, J., Williams, J., Marshall, E.A., 2020. A mixed-methods
study of the demographic and behavioural correlates of walking to a more distant bus stop. Transp. Res. Interdiscip. Perspect. 6, 100164.
Rue, H., Martino, S., Chopin, N., 2009. Approximate Bayesian inference for latent Gaussian models using integrated nested Laplace approximations (with discussion).
J. Roy. Stat. Soc. B 71 (2), 319–392.
Saha, D., Alluri, P., Gan, A., Wu, W., 2018. Spatial analysis of macro-level bicycle crashes using the class of conditional autoregressive models. Accid. Anal. Prev. 118,
166–177.
Siddiqui, C., Abdel-Aty, M., Choi, K., 2012. Macroscopic spatial analysis of pedestrian and bicycle crashes. Accid. Anal. Prev. 45, 382–391.
Smith, R., Reed, S., Baker, S., 2010. Street design: Part 1. Complete streets. Public Roads 74 (1), 12–17.
Su, J., Sze, N.N., Bai, L., 2021. A joint probability model for pedestrian crashes at macroscopic level: roles of environment, traffic, and population characteristics.
Accid. Anal. Prev. 150, 105898.
Ulak, M.B., Kocatepe, A., Yazici, A., Ozguven, E.E., Kumar, A., 2021. A stop safety index to address pedestrian safety around bus stops. Saf. Sci. 133, 105017.

16
M.T. Ashraf et al. Journal of Transport & Health 27 (2022) 101529

van Soest, D., Tight, M.R., Rogers, C.D., 2020. Exploring the distances people walk to access public transport. Transport Rev. 40 (2), 160–182.
Xie, K., Ozbay, K., Yang, D., Xu, C., Yang, H., 2021. Modeling bicycle crash costs using big data: a grid-cell-based Tobit model with random parameters. J. Transport
Geogr. 91, 102953.
Xu, P., Huang, H., 2015. Modeling crash spatial heterogeneity: random parameter versus geographically weighting. Accid. Anal. Prev. 75, 16–25.
Zhang, G., Yang, H., Li, S., Wen, Y., Li, Y., Liu, F., 2019. What is the best catchment area of bike share station? A study based on Divvy system in Chicago, USA. In:
In2019 5th International Conference on Transportation Information and Safety (ICTIS). IEEE, pp. 1226–1232.
Ziakopoulos, A., Yannis, G., 2020. A review of spatial approaches in road safety. Accid. Anal. Prev. 135, 105323.
Zuo, T., Wei, H., Chen, N., 2020. Promote transit via hardening first-and-last-mile accessibility: learned from modeling commuters’ transit use. Transport. Res.
Transport Environ. 86, 102446.

17

You might also like