Investigating the association between neighbourhood characteristics and e-scooter safety

Abstract

The uptake of e-scooters as an alternative mode of travel has risen sharply in recent years; however, their safety is less-understood compared to other modes of travel. For the first time in the extant literature, we explore the association between neighbourhood characteristics and e-scooter safety in Greater London, UK. We found that, over the study period, the expected e-scooter crash frequency was the highest in the City of London, followed by the West End, and then St. James's―both wards located in the borough of Westminster in central London. We found that e-scooter crash frequencies increase with an increase in area-level walking and cycling activities. Similarly, we found that the number of schools is positively associated with the expected e-scooter crash frequency. In contrast, the results indicated that as the proportion of ward-level greenspace increases, the number of crashes involving e-scooters decreases. The results also highlighted social inequalities in this context, with higher e-scooter crash frequencies in areas with larger Black, Asian and Minority Ethnic population, those with higher crime rates, and those with a higher population of children in out of work households. This research provides practical recommendations to prioritise areas for safety interventions and for selecting suitable safety improvement programmes.

Full text

Sustainable Cities and Society 83 (2022) 103982
Available online 3 June 2022
2210-6707/© 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
Investigating the association between neighbourhood characteristics and
e-scooter safety
Shahram Heydari
*
, Michael Forrest, John Preston
Transportation Research Group, Department of Civil, Maritime and Environmental Engineering, University of Southampton, Southampton, United Kingdom
ABSTRACT
The uptake of e-scooters as an alternative mode of travel has risen sharply in recent years; however, their safety is less-understood compared to other modes of travel.
For the rst time in the extant literature, we explore the association between neighbourhood characteristics and e-scooter safety in Greater London, UK. We found
that, over the study period, the expected e-scooter crash frequency was the highest in the City of London, followed by the West End, and then St. James’s–both wards
located in the borough of Westminster in central London. We found that e-scooter crash frequencies increase with an increase in area-level walking and cycling
activities. Similarly, we found that the number of schools is positively associated with the expected e-scooter crash frequency. In contrast, the results indicated that as
the proportion of ward-level greenspace increases, the number of crashes involving e-scooters decreases. The results also highlighted social inequalities in this
context, with higher e-scooter crash frequencies in areas with larger Black, Asian and Minority Ethnic population, those with higher crime rates, and those with a
higher population of children in out of work households. This research provides practical recommendations to prioritise areas for safety interventions and for
selecting suitable safety improvement programmes.
1. Introduction
Electric scooters (e-scooters) are becoming a popular form of
micromobility around the world with the sales of private devices ever
climbing and hire schemes appearing in an increasing number of loca-
tions. The rise of e-scooters is partly due to the recent pandemic that has
had a positive impact on micromobility use in urban areas; see, for
example, Wang & Noland (2021), and Heydari et al. (2021). E-scooters
have been praised for offering an alternative to short car and public
transit trips (Shaheen & Cohen, 2019) and may have a relatively small
carbon footprint as compared to other powered vehicles
1
as well as
taking up less room on roads, thereby easing congestion, which in turn
leads to improved air quality in urban areas. Therefore, e-scooters can
potentially contribute to sustainability in urban areas. On the other
hand, due to their small wheels, their ability to travel at relatively fast
speeds, and a lack of legislation and accountability for their use, safety
concerns arise. Due to the novelty of e-scooters and limited real-world
crash data being available, published literature on their operation and
safety is limited compared to other modes of transport. In fact, the safety
of e-scooters is less-understood.
Although e-scooters have been around for several decades, there has
been an increase in their popularity over the past few years with more
than 360,000 private e-scooters being purchased in the UK in 2020
(Winchcomb, 2021). In actuality, according to the UK Department for
Transport (DfT), it is not legal to ride private e-scooters on footpaths,
public roads or cycle lanes except in specially designated trial areas
(DfT, 2021). Despite the unlawful nature of private e-scooter travel, they
are still used as a mode of transport in the UK and repercussions for their
use seem to be rare. Hiring under an approved rental scheme is the only
way to ride an e-scooter legally in the UK (DfT, 2021). In fact, several
e-scooter rental schemes have been recently launched in different urban
areas across the country, but all are still subject to a trial period (DfT,
2020). Specically, the Transport for London (TfL) launched the London
e-scooter trial in June 2021 in ten boroughs: Camden, City of London,
Ealing, Hammersmith & Fulham, Kensington and Chelsea, Lambeth,
Southwark, Richmond upon Thames, Tower Hamlets, and Westminster
(TfL, 2021a).
1.1. Previous research
A number of previous studies have recently examined different as-
pects of emerging e-scooters. These include, for example, identifying
optimal locations for battery swapping stations (Torkayesh & Deveci,
2021), exposure to air pollution while travelling by e-scooters (Tran
et al., 2021), and factors associated with e-scooter usage (Hosseinzadeh
et al., 2021; Noland, 2021). With respect to safety, a relatively limited
* Corresponding author.
E-mail address: s.heydari@soton.ac.uk (S. Heydari).
1
Although this is dependent on the lifespan of the vehicles and the servicing arrangements – see, for example, de Bortoli & Christoforou (2020).
Contents lists available at ScienceDirect
Sustainable Cities and Society
journal homepage: www.elsevier.com/locate/scs
https://doi.org/10.1016/j.scs.2022.103982
Received 13 March 2022; Received in revised form 1 June 2022; Accepted 1 June 2022

Sustainable Cities and Society 83 (2022) 103982
2
number of previous studies examined various safety-related issues
relating to e-scooters (see; e.g., Kobayashi et al. (2019), Rix et al. (2021),
and Bodansky et al. (2022)). For example, illegal riding of e-scooters is
an issue which is addressed in some previous research. Behaviours
which constitute illegal riding differ between countries and even cities,
but some common themes are an inappropriate age of the rider, riding
under the inuence, riding while distracted, and riding in inappropriate
locations (Gioldasis et al., 2021; Haworth et al., 2021a, 2021b). A
scarcity of e-scooter crash data means that many researchers have
focused on the mechanics surrounding individual crashes. Findings such
as a tendency for crashes to be an isolated fall and not involving another
person or vehicle (Brownson et al., 2019) and a higher incidence of
crashes at weekends (Stigson et al., 2021) have been identied. In a
comparison study from Nashville (Tennessee), Shah et al. (2021) found
the majority of e-scooter crashes can be explained with just two crash
typologies: conicts with a motor vehicle either turning right or going
straight at an intersection. However, a greater number of crash typol-
ogies were found for cyclists, suggesting different collision mechanisms
between the modes.
To try and gain a more complete understanding of e-scooter crashes
as a big picture, a range of data sources have been utilised by re-
searchers. Questionnaires have been asked of people who visited hos-
pitals after a crash (Cicchino et al., 2021), of people at targeted locations
in the public realm (Gioldasis et al., 2021; Siebert et al., 2021) and of
people recruited as a random sample for their perception of the safety of
e-scooters (Gibson et al., 2021). A data mining technique with news-
paper reports is used by Yang et al. (2020) to gather information on
common rider demographics as well as crash types and locations. Given
the lack of available data on crash statistics from authorities, other data
sets such as insurance reports were sometimes used to supplement this
(Stigson et al., 2021). While these studies all have merits, research on
the association between e-scooter crashes and built and natural envi-
ronment, trafc, and socio-demographic characteristics are rare. To our
knowledge, only one study conducted by Azimian & Jiao (2022)
investigated the impact of built environment and sociodemographic
factors on e-scooter safety, considering dockless e-scooter injury acci-
dents in Austin, Texas. A signicant association was found between
crashes and population age/gender ratios, median household income,
the ratio of public transport users to private transport users, the land use
entropy index, and the percentage of restaurants and educational
centres.
Studies of other active travel modes, walking and cycling, are
numerous, providing valuable macro-level (area-level) transportation
safety and planning insights (Cai et al., 2016; Osama & Sayed, 2016;
Wang et al., 2016). These studies exemplify the importance of con-
ducting big picture analysis at a geographic area level and drawing
conclusions about a whole area. In Wang et al. (2016), for example, their
analysis is shown to have two implications: rst, areas with
higher-than-expected crashes can be identied and addressed; second,
any future-increase in expected pedestrian crashes due to land use
development can be predicted and minimised.
Tuli et al. (2021) utilised a macro-level approach with a range of
geographical characteristics to analyse e-scooter usage patterns in Chi-
cago, Illinois. This study used a random-effects negative binomial model
which was found to effectively model the origin-destination count of
e-scooter trips. The research revealed that more trips are generated by
more densely populated areas, those with more parks and open space,
and those with a higher number of zero-car households. Another de-
mographic study of how low-income areas affect shared e-scooter usage
was carried out by Frias-Martinez et al. (2021) who found that lower
income areas in four major US cities engaged in fewer trips. The
above-mentioned studies indicate that area-level characteristics have a
bearing on e-scooter mobility patterns and exposure. It is therefore
interesting to investigate how such characteristics affect e-scooter
safety.
1.2. The current research
This research contributes to the existing e-scooter safety literature by
investigating (to our knowledge, for the rst time) the association be-
tween various neighbourhood characteristics and zonal level e-scooter
safety. The present study differs principally from Azimian & Jiao (2022)
as crashes mostly involving private e-scooters are considered in our
research.
2
Additionally, the random parameters multilevel modelling
approach adopted in the present paper further considers the spatial
dependencies in the data while accounting for unobserved heterogeneity
more fully (Dupont et al., 2013; Heydari et al., 2018; Mannering et al.,
2016). In this paper, we utilise e-scooter crash data in the Greater
London area from the beginning of 2020 to the end of June 2021. Using
this crash data and other extensive data on built and natural environ-
ment characteristics, exposure measures, and socio-demographics, we
carry out a ward-level study to identify various area-level factors that
are associated with the propensity of e-scooters crashes in Greater
London. This research improves our understanding of e-scooter safety
and can lend itself to safety policy. The paper provides useful insights for
local authority decision making with the aim of promoting micro-
mobility in urban areas. To this end, we discuss practical implications of
the study based on our ndings.
2. Data description
2.1. Crash data
The crash data utilised in this study are obtained from STATS19
databases which are recorded by UK police forces and kept by the DfT.
The accidents are recorded with details of the local authority (borough
in Greater London) in which they occurred. The ward is not recorded by
the police; however, each crash is geo-tagged so using GIS software, we
identied crash counts in each ward. At the time of writing, the
STATS20 form, which is completed by police ofcers when recording a
crash, does not feature an option to record an e-scooter as the vehicle
type. Instead, an option of ‘other vehicle’ is selected and ‘electric
scooter’ is added in a free text box if a crash involves an e-scooter. The
data used here are of all crashes where at least one vehicle was tagged as
e-scooter. Specically, the crash data included 534 accidents of which
around 4% were falls (no other parties were involved), around 17% were
between e-scooters and pedestrians, and the remaining 79% involved
other parties including motorised vehicles.
The period of study includes 18 months of data from January 2020 to
June 2021, which is the period for which the crash data is available. The
2021 crash records, at the time of writing, are still provisional. The
decision to include the latter was taken to consider a longer time period,
which is known to be less affected by random uctuations in crash data
and thus resulting in a richer analysis (Hauer, 1997). While the
six-month period data in 2021 is provisional, we have judged that any
future changes in the 2021 data would be relatively minor and would
not have a signicant effect on the results of the study.
2.2. Explanatory variables
In addition to the crash data, we collected data on a broad range of
variables covering various land use, demographics, exposure measures,
and built and natural environment characteristics. Inspiration and
guidance were taken from other active travel safety literature to select
variables that had previously been found as determinants of active
2
Note that rental e-scooters were available in selected London boroughs only
during June 2021 as mentioned in Introduction; therefore, most crashes re-
ported over our study period involve private e-scooters. At the time of writing,
it was not possible to distinguish between rental and private e-scooters in
STATS19 databases.
S. Heydari et al.

Sustainable Cities and Society 83 (2022) 103982
3
modes safety, along with the knowledge and experience of the re-
searchers. The data were principally obtained from two sources that are
collected by the Greater London Authority (GLA): London ward proles
(GLA, 2015a) and London borough proles (GLA, 2015b). Owing to the
range of sources which make up these datasets, such as census data,
survey data, etc., and limitations with how often certain data are
collected, there is some disparity between the years from which the data
have provenance. As well as a temporal range, the variables are also split
between two types of spatial unit, some being recorded at a borough
level and others at a ward level. It is not always possible or practical for
authorities to record data granulated to a ward level.
No direct exposure measure (e.g., trip counts) for e-scooters was
available; therefore, proxy exposure measures, especially those relating
to walking and cycling, were considered. The proportion of adults who
cycle to work according to the 2011 census was used as a measure of
cycling activity. To capture walking activity, we used the 2018 data on
the proportion of population who walk at least one, three or ve times
per week for at least ten minutes and for any purpose. The walking and
cycling data were obtained from the DfT. London Tube (subway) entry
and exit counts were obtained from Transport for London (TfL). Tube
activity is correlated with active travel as underground trips often
include at least one stage of walking or cycling to begin or complete the
journey (TfL, 2021b). The London bike share scheme (Santander cycles)
has become a popular mode of travel across many parts of London
(Lovelace et al., 2020). Therefore, we obtained the ward-level numbers
of docking stations, which can offer another proxy exposure measure for
active travel, from TfL. Also, we considered trafc ow, which was
available in the form of vehicle kilometres travelled at borough level.
Note that, over the study period, around 75% of e-scooter crashes
involved a motorised vehicle. The rental e-scooter trial was launched in
10 London boroughs in June 2021 (TfL, 2021a), the end of our study
period; therefore, this was not relevant in this research.
Several built and natural environment characteristics were consid-
ered as well since these are known to have a bearing on area-level safety
for various modes of travel, particularly active modes. These included
land use with domestic garden, greenspace, water, non-domestic
buildings, and domestic buildings. We obtained the number of schools
and school enrolment in each ward from the Ordnance Survey
Fig. 1. Schematic view of the data.
Table 1
Summary statistics of the data.
Variable types Variables Spatial unit Mean SD Minimum Maximum
Crash counts
e-scooter Crashes Ward 0.843 1.158 0.000 9.000
Exposure
Population cycling to work (%)
1
Ward 4.011 3.395 0.242 19.092
Population walking at least once a week (%)
2
Borough 73.156 5.408 62.800 86.510
Population walking at least three times a week (%)
2
Borough 49.591 5.945 40.071 64.922
Population walking at least ve times a week (%)
2
Borough 38.223 5.224 30.090 51.830
Tube entries and exits (00 millions of travellers) Borough 0.831 1.242 0.000 6.553
Trafc ow (Billion vehicle kilometres travelled) Borough 0.936 0.422 0.150 2.147
Land use, built and natural environment
Area of ward (km
2
) Ward 2.720 2.760 0.350 29.04
Number of schools Ward 5.146 2.791 0.000 24.000
Land use with greenspace (%) Ward 26.20 16.60 15.30 90.00
Land use with domestic gardens (%) Ward 26.208 12.092 0.120 59.161
Land use with domestic buildings (%) Ward 11.904 5.154 0.850 30.540
Land use with non-domestic buildings (%) Ward 6.582 5.701 0.390 42.180
Land use with water (%) Ward 2.149 6.304 0.000 74.240
Cycle network density (total length per borough area) Borough 2.373 0.769 0.585 4.125
Density of pubs (pubs per km
2
) Borough 3.650 4.718 0.529 49.712
Road network (km) Borough 448.933 184.833 55.522 902.679
Santander docking stations (00 s) Borough 0.223 0.402 0.000 1.660
Socio-demographic
Number of cars per household Ward 0.840 0.327 0.233 1.705
Population (000 s) Ward 14.132 3.083 4.622 32.046
Population per square kilometre (0000 s) Ward 0.875 0.520 0.019 2.766
Child population (%) Ward 19.805 3.761 6.473 32.695
School enrolment (number of children) Ward 2287.986 1228.846 0.000 7388.000
Number of children in out of work households Ward 607.770 372.183 10.000 1940.000
BAME population (%)
3
Ward 0.389 0.189 0.041 0.937
Population with level 4 qualications and above (%) Ward 37.678 12.840 12.500 68.700
Crime rate (crimes committed per ward population) Ward 0.088 0.073 0.026 0.894
Children in poverty (%) Borough 19.270 6.020 8.8000 32.500
Lone parents without employment (%) Borough 46.110 8.540 20.820 73.580
Yearly Expenditure on alcohol (£00 millions) Borough 0.525 0.244 0.029 1.125
Average weekly earnings (£s) Borough 561.680 70.385 462.367 902.000
Unemployed population (%) Borough 5.322 0.999 3.867 19.633
1 Cycling measures the proportion of full time workers (adults) who cycle to work
2 Walking measures any continuous walk for at least 10 min for any purpose
3 BAME population refers to Black, Asian and Minority Ethnic population
S. Heydari et al.

Sustainable Cities and Society 83 (2022) 103982
4
Topography Layer for the year 2020. Other built environment variables
included the density of pubs (the number of pubs divided by the area of
the borough), the density of the cycle network (total length of cycle
lanes per area of borough), and the length of road network at borough
level.
Socio-demographic variables from a range of sources were consid-
ered. The number of cars per household was obtained from the 2011
census as was the percentage of people who hold a qualication at or
above level four. According to the UK Government’s qualication levels
in England, Wales and Northern Ireland, this level refers to the difculty
of obtaining the qualication, with higher levels being more difcult;
level four dictates that a person has achieved beyond A-level or equiv-
alent. Weekly earnings and unemployment rates representing the eco-
nomic characteristics of the wards were obtained from the Department
for Work and Pensions. Besides weekly earnings and unemployment
rate, economic status is represented by expenditure on alcohol and the
number of cars per household. Also, we considered data relating to child
poverty, education levels, and crime rates, provided by the London Data
Store and Metropolitan Police. The latter variables are markers of
deprivation and are often associated with areas of higher crash incidence
(Graham & Stephens, 2008; Graham et al., 2013; Li et al., 2017). Various
population related variables, including population, child population,
BAME (Black, Asian and Minority Ethnic) population, and population
density were obtained from the London Data Store. For example, pop-
ulation density was used as a proxy measure for active modes’ exposure
in previous literature (see e.g., Cottrill & Thakuriah, 2010). Fig. 1 dis-
plays a schematic view of the various components of the data compiled
in this study. Summary statistics of the data are reported in Table 1.
2.3. Statistical approach
Given the nature of the data being ward-level crash counts nested
within boroughs, we adopted a Bayesian multilevel random parameters
(slopes) Poisson lognormal regression approach (El-Basyouny & Sayed,
2009; Heydari et al., 2018). Random parameters models can address
unobserved heterogeneity more fully compared to the conventional
models (Mannering et al., 2016). To investigate how the t improves, we
also developed a simple Poisson lognormal model and a random in-
tercepts multilevel Poisson lognormal model. Note that multilevel
models, which are extensively used in the crash literature, can capture
spatially and non-spatially related unobserved factors effectively by
accommodating the hierarchical structure of the data (Dupont et al.,
2013; Heydari et al., 2016; Huang & Abdel-Aty, 2010; Islam & El-Ba-
syouny, 2015). We also investigated spatial autocorrelation in the data,
developing Bayesian conditional autoregressive models, but we did not
nd any evidence for such spatial dependency. We therefore discuss
only our nal model here.
2.4. Multilevel random parameters Poisson lognormal model
A multilevel random parameters Poisson lognormal model can be
specied as follows. Let y
i
and γ
i
be, respectively, observed and expected
crash frequencies for ward i. Let X and
α
be explanatory variables and
their respective regression coefcients. Let β
0j
represent the borough
effects (varying intercepts) that follow a normal distribution with the
mean
μ
β0
and the variance v
β0
, where j stands for borough. Let Z be
explanatory variables, the effects of which vary across different bor-
oughs, with their corresponding regression coefcients β. Let
ε
i
be a
normally distributed ward-level error term which has a mean of 0 and a
variance v
ε
, accounting for extra variability in the data. We can then
write:
yij ∼Poisson γij(1)
logγij=β0j+Zij βj+Xij
α
+
ε
ij
β0j∼normal
μ
β0,
ν
β0
βj∼normal
μ
β,vβ
ε
ij ∼normal(0,
ν
ε
)
We specied non-informative priors for various model parameters
and implemented the models in the Nimble package in R (de Valpine
et al., 2017). See Heydari et al. (2018) for further details on the model as
applied to at-grade crossings where grade crossings were nested within
various provinces.
3. Results and discussions
This section presents the model results and gives context to the
ndings, followed by the interpretation of the results, and nally some
thoughts on policy implications. The best performing model was the
multilevel random parameters Poisson lognormal model, providing the
best t to the data; therefore, our discussions will focus on the results of
Table 2
Estimation results of the regression coefcients.
Poisson lognormal model
Mean SD 95% Credible
Intervals
ln(Population walking at least three
times a week for any purpose)
1.338 0.521 0.328 2.361
ln(Population cycling to work) 0.159 0.073 0.015 0.302
ln(Crime rate) 0.741 0.086 0.573 0.910
ln(number of children in out of work
HH
1
)
0.258 0.078 0.108 0.415
Land use with greenspace −0.866 0.378 −1.600 −0.128
BAME population 0.926 0.342 0.250 1.592
Number of schools 0.033 0.015 0.002 0.062
Constant −0.455 0.061 −0.580 −0.338
Variance obs. level error term 0.077 0.056 0.006 0.206
Model t (WAIC) 1419 – – –
Multilevel random intercepts Poisson lognormal model
Mean SD 95% Credible
Intervals
ln(Population walking at least three
times a week for any purpose)
1.317 0.647 0.056 2.593
ln(Population cycling to work) 0.184 0.087 0.015 0.356
ln(Crime rate) 0.764 0.087 0.594 0.935
ln(number of children in out of work
HH
1
)
0.241 0.083 0.079 0.405
Land use with greenspace −0.796 0.386 −1.566 −0.053
BAME population 0.925 0.366 0.205 1.649
Number of schools 0.030 0.016 0.004 0.055
Borough effect −0.458 0.071 −0.604 −0.324
Variance Borough effect 0.050 0.036 0.002 0.139
Variance obs. level error term 0.040 0.042 0.001 0.152
Model t (WAIC) 1412 – – –
Multilevel random parameters Poisson lognormal model
Mean SD 95% Credible
Intervals
ln(Population walking at least three
times a week for any purpose)
1.232 0.631 0.009 2.467
ln(Population cycling to work) 0.202 0.099 0.009 0.401
Variance ln(Population cycling to work) 0.065 0.054 0.002 0.200
ln(Crime rate) 0.766 0.088 0.594 0.937
ln(number of children in out of work
HH
1
)
0.238 0.084 0.077 0.402
Land use with greenspace −0.773 0.388 −1.551 −0.016
BAME population 1.030 0.380 0.297 1.787
Number of schools 0.028 0.016 0.002 0.054
Borough effect −0.467 0.069 −0.608 −0.334
Variance Borough effect 0.022 0.027 0.001 0.095
Variance obs. level error term 0.046 0.043 0.002 0.159
Model t (WAIC) 1407 – – –
1 HH: household
S. Heydari et al.

Sustainable Cities and Society 83 (2022) 103982
5
this model. However, the results of the other two models are provided
for context and comparison.
3.1. Estimation results
The parameter estimates for the variables that were found to be
statistically signicant (their 95% credible intervals not containing zero)
are reported in Table 2. Note that, since we conducted the analysis using
Bayesian statistics, we obtained credible intervals, which are analogue
to frequentist condence intervals. Credible intervals have a more
intuitive interpretation compared to their classical counterparts; that is,
a 95% credible interval indicates that there is 95% chance that an
estimated coefcient happens to be in the range of that interval
(Daziano et al., 2013). However, this probability is either zero or one
when considering condence intervals. While we discuss the interpre-
tation of the estimated regression coefcients in Section 4.2.2, as it can
be seen in Table 2, all the models provide less or more similar regression
coefcient estimates. The results indicate that variables walking,
cycling, crime, children in out of work households, BAME population,
and the number of schools are positively associated with ward level
e-scooter crash frequencies in Greater London. However, land use with
greenspace is negatively associated with these crashes. The random
parameters model, which can accommodate the varying effect of the
explanatory variables, revealed that the effect of the variable cycling
varies across boroughs. This interesting nding indicates that perhaps
other unknown factors inuence the impact of cycling on e-scooter crash
counts. To investigate this further, we attempted to explain the het-
erogeneity in the mean and the variance of this random parameter (see,
for example, Seraneeprakarn et al., 2017); however, no variable in the
data was found to be able to explain its variability.
We also considered alternative distributional assumptions (e.g.,
lognormal) for our random coefcients; however, this did not result in
any improvement. Allowing for the varying effect of cycling improved
the model t; that is, a lower Watanabe-Akaike information criterion
(WAIC) value compared to the other models (see Table 2). Note that
WAIC is among the most valid Bayesian model tting criteria (Gelman
et al, 2013; Watanabe, 2010). The varying borough effects indicate that
there is a difference between various boroughs due to unobservables
that have a bearing on safety. Through the multilevel model we can
indirectly capture such differences. The observation level error term
accounts for extra variation, which is not accounted for by the explan-
atory variables and the hierarchical component of the model.
Based on our results, an increase in both walking and cycling levels
leads to an increase in e-scooter crash frequency. Walking and cycling
are generally found to have a positive association with pedestrian and
cyclist crash frequencies (et al., 2016; Heydari, Fu, Miranda-Moreno &
Joseph, 2017). It may also be the case that walking exposure correlates
with cycle casualties and vice versa and this may extend to other similar
travel modes such as e-scooters. Certainly, this appears to be the case in
the current research, and this stands to reason as areas where people
tend to walk and cycle more and drive less will likely be the same areas
where more people have taken to using e-scooters as a mode of travel
and thus more crashes are likely due to a higher exposure.
Our results showed that as the number of crimes per population
increased, e-scooter crash frequencies increase as well. One possible
explanation for this nding is that, since e-scooter use was largely illegal
during the study period (as discussed in Section 1), e-scooter crash fre-
quencies may be related to crime rates and those with a propensity for
deviant behaviour. Overall, this nding is in accordance with previous
research. For example, crime rate is noted to be associated with an in-
crease in pedestrian crashes by Cottrill & Thakuriah (2010) though the
mechanisms behind this are unclear. Crime is however one of the indices
of multiple deprivation (IMD) (see Graham & Stephens (2008) for a full
specication of the indices) and research has shown that areas with
higher IMD scores usually correlate with higher active travel crash fre-
quencies (Graham et al., 2005; Green, Muir & Maher, 2011; Li et al.,
2017). Similarly, we found that as the number of children in out of work
households increases, e-scooter crash frequencies increase–deteriorat-
ing trafc safety. Similar to crime rate, child poverty variables are a
deprivation measure; and therefore, a positive association with crash
frequency is unsurprising. Also, the results indicated that an increase in
the proportion of a ward’s population who identify as BAME is
Fig. 2. Spatial distribution of expected e-scooter crash frequency across Greater London wards.
S. Heydari et al.

Sustainable Cities and Society 83 (2022) 103982
6
associated with an increase in ward-level crash frequencies involving
e-scooters. Previous studies have also found that the ethnic de-
mographics of an area are associated with crash frequency for cyclists
(Ding et al., 2020) and pedestrians (Su et al., 2021).
We found a negative association between the proportion of land use
with greenspace and e-scooter crash frequencies. This is an area of in-
terest as greenspace has been shown to be positively associated with the
number of crashes for cyclists in London (Ding et al., 2020). In that study
it was posited that over the summer months when leisure cycling levels
are higher, areas with green space attract these cyclists and thus expe-
rience a higher number of cyclist crashes. The same could potentially be
true of e-scooters; further study into seasonal instability of e-scooter
usage and crashes would be needed. However, the result of the current
paper may be explained by the fact that riding an e-scooter in a park, for
example, means e-scooter users are less exposed to road trafc, leading
to a reduced crash risk.
Also, we found that, as the number of schools increases, e-scooter
crash frequency increases. This is in accordance with previous literature.
For example, an increase in crash frequency was shown for pedestrians
with a higher number of schools (Zhan et al., 2015) as they attract more
trips from children who possess innate cognitive, physical and behav-
ioural traits which make them more vulnerable to road accidents
(Gitelman et al., 2019). Bhat et al. (2017) found an increase in pedes-
trian injuries in census tracts in New York as the number of schools in-
creases, though with the caveat that these injuries were less likely to be
incapacitating due to lower speeds and heightened driver awareness.
Also, with respect to school as a risk factor, Heydari et al. (2020) found
that pedestrian injury frequencies increased at intersections in proximity
to schools in Montreal, Quebec. Nevertheless, the latter requires further
investigation in the context of e-scooter safety.
4. Policy analysis
4.1. Practical area level inferences
Figs. 2 and 3 display the spatial distribution of expected e-scooter
crash frequencies over the study period across different Greater London
wards and boroughs, respectively. A darker colour indicates a higher
expected crash frequency. Figs. 2 and 3 can be used to identify overall
spatial patterns in terms of e-scooter safety in the Greater London area
and to detect high crash wards and boroughs. These could be prioritised
for safety improvement programmes that can follow from our study,
considering the most important area-level variables that are associated
with e-scooter safety (see Section 4.2.2.).
It can be seen in Fig. 2 that the propensity of e-scooter crashes is, in
general, higher in inner London wards. Specically, we found that the
City of London (if we consider it as a ward given its relatively small size),
followed by the West End, and then St. James’s–both wards located in
the borough of Westminster in central London–had the highest expected
e-scooter crash frequencies. It can be inferred from Fig. 3 that the
Fig. 3. Spatial distribution of expected e-scooter crash frequency across Greater London boroughs.
S. Heydari et al.

Sustainable Cities and Society 83 (2022) 103982
7
boroughs of Westminster, Tower Hamlets, and Lambeth had the highest
expected e-scooter crash frequency over the study period, followed by
Hackney, Camden, Southwark, and Wandsworth. It can be seen in Fig. 3
that Croydon has the highest expected e-scooter crash frequency among
the outer London boroughs.
4.2. Elasticities and marginal effects
To interpret the magnitude of the association between each explan-
atory variable and ward-level e-scooter crash frequencies over the study
period, Table 3 reports average elasticities and marginal effects. These
are based on the results of the multilevel random parameters count
model, which provided the best t to the data. For the log transformed
variables, elasticities can be readily obtained based on their respective
estimated regression coefcients (see Table 3). For the rest of the vari-
ables, we estimated average marginal effects (Washington et al., 2020).
The estimation of the elasticities and marginal effects provides a clear
understanding of the effect of the explanatory variables on ward level
e-scooter crash frequencies, allowing us to identify the most impactful
contributory factors. For example, our results revealed that walking has
a much higher impact (having a bigger elasticity value) on e-scooter
crash frequencies compared to cycling. This highlights the need for
tailored safety interventions in areas with high levels of walking to
mitigate e-scooter crash risk propensity.
Our results indicate that a 10% increase in the proportion of the
population who walk at least three times a week for any purpose will
increase the expected e-scooter crash frequency (per 18-month period; i.
e., the study period) by 12.32%. The same increase in the proportion of
people who cycle to work (i.e., those commuting by bike) will increase
the expected e-scooter crash frequency by an average of 2.02%. Simi-
larly, a 10% increase in the ward crime rate will result in an average
increase of 7.66% in the expected e-scooter crash frequency, and the
same percentage increase in the number of children in out of work
households will yield a 2.38% increase in the frequency of crashes
involving e-scooters. The marginal effects show that a unit increase in
land use with greenspace will reduce the expected crash frequency by an
average of 0.652 crashes over an 18-mounth period (the study period).
One unit increase in BAME population, on average, will lead to 0.869
additional crashes over an 18-month period, and an additional school in
a ward will result in an average increase of 0.024 in the expected e-
scooter involved crash count over an 18-month period.
To be able to compare the explanatory variables, which are in the
model, in terms of their impacts on ward-level e-scooter crash fre-
quency, we computed marginal effects for the log transformed variables
as well. These are ordered in terms of their impacts as follows (from the
highest to the lowest): BAME population, green space, crime rate,
walking, cycling, children in out of work households, and the number of
schools.
5. Summary and conclusions
Following the proliferation of e-scooters in many urban settings
worldwide in recent years, the intent of this study was to provide an
improved understanding of zonal level e-scooter safety (measured in
terms of crash frequency). Specically, we investigated the association
between various neighbourhood characteristics–including the built/
natural environment and socio-demographics–and ward-level e-scooter
safety in Greater London. We used a Bayesian multi-level random pa-
rameters count model to effectively account for unobserved heteroge-
neity and the hierarchical structure of the data (wards nested within
boroughs).
We found that walking and cycling activities are important correlates
of area level e-scooter safety. In fact, this research shows that, in the
absence of reliable e-scooter trip data that can constitute a direct
exposure measure for this novel mode of travel, walking and cycling can
act as suitable proxy exposure measures for investigating e-scooter
safety. Note that trafc ow was not found to have a statistically
important effect on ward-level e-scooter safety. This is perhaps due to
the fact that trafc ow was available at borough level in our study. Had
ward- level trafc ow been available, this variable would have been
appeared in the model as a statistically signicant variable. Also, the
effect of trafc volume is partly captured through other variables that
are in model. The built and natural environment attributes such as the
number of schools and greenspace were found to be associated with the
number of crashes involving e-scooters. Specically, we found that e-
scooter crash counts are slightly higher in wards with higher numbers of
schools. This indicates that a particular attention should be given to
these wards when it comes to increasing trafc safety. The results also
indicated that ward-level greenspace is benecial to trafc safety: the
larger the proportion of greenspace, the lower the expected e-scooter
crash frequency. Different types of variables that were investigated in
this research appear to align with literature on other active travel
modes, such as walking and cycling. In other words, factors affecting the
area level safety of pedestrians and cyclists, in general, would appear to
affect the safety of e-scooter users in a similar fashion. In the absence of
extensive real-world e-scooter crash data, this could provide valuable
information on the way in which e-scooter use and regulation, and large
scale transportation planning and safety policies are viewed. This will in
turn help improve road safety as the uptake of this specic form of
micromobility increases.
In addition, this research revealed important socioeconomic and
ethnic background differences in e-scooter related road crashes in Lon-
don. Specically, we found higher e-scooter crash frequencies among
areas with larger BAME population, those with higher crime rates, and
those with higher numbers of children in out of work households. In this
regard, the planning of interventions, which aim at increasing trafc
safety, should consider the latter factors to reduce inequalities relating
to e-scooter safety in the Greater London area. Note that this does not
necessarily imply that the risk (in its epidemiological term) of getting
involved in an e-scooter crash is higher, for example, among BAME
population. As discussed by Noland & Laham (2018) caution must be
taken in drawing conclusive conclusions in this regard based on
ecological (spatial) studies. In fact, a crash-level analysis, with
individual-level characteristics, would provide more detailed insights in
this regard.
Addressing inequalities in this context would be particularly
important at this stage in which e-scooter riding is a relatively novel
mode of travel. This can help contain the gap between people from
diverse socioeconomic and ethnic backgrounds in a timely manner. This
calls (i) for research to better understand the reason for such in-
equalities, and (ii) for policies that consider social inequality in decision-
making processes and that consider suitable remedies to address
inequity.
Our ndings, being based on a rigorous statistical analysis, can be
utilised by authorities to identify high-crash wards and boroughs; and
Table 3
Average elasticities and marginal effects based on the random parameters count
model.
Log transformed variables Elasticities
1
Population walking at least three time a week 12.32%
Population cycling to work 2.02%
Crime rate 7.66%
Number of children in out of work households 2.38%
Other Variables Marginal effects
Land use with greenspace −0.652
BAME population 0.869
Number of schools 0.024
1 Elasticities are given based on a 10% increase in the variable in question
S. Heydari et al.

Sustainable Cities and Society 83 (2022) 103982
8
consequently, decide where to allocate safety improvement funds and
projects. Also, this research can help authorities in selecting and
implementing a variety of countermeasures: (i) educational programmes
and publicity campaigns that aim at raising awareness among the pop-
ulation about the safety implications of e-scooters on the road network
(e.g., in more deprived areas or where the number of schools is higher);
(ii) stricter enforcement in high e-scooter crash locations (e.g., areas
where walking is more prevalent); and (iii) trafc engineering in-
terventions. With respect to the latter; for example, the inclusion of e-
scooters on trafc signage, at least, in high e-scooter crash wards, with
the aim of warning other road users (e.g., pedestrians and car drivers) of
their presence could be considered. Another engineering intervention
may consider more segregated “e-scooter ways” which would allow e-
scooters, like bicycles, to take advantage of being physically shielded
from motorised trafc. This could happen in the form of designated e-
scooter facilities that separate e-scooter riders from other road users
where the prevalence of e-scooter riding is comparatively high or
through sharing cycleways with cyclists. Finally, the use of dedicated
trafc lights for e-scooter users, similar to those existing for pedestrians,
at junctions with relatively high prevalence of e-scooter users, and “e-
scooter boxes” (similar to bike boxes) at junctions in high-crash loca-
tions/wards would help improve e-scooter safety; and consequently,
trafc safety.
Note that our research did not investigate the impact of the above-
mentioned proposed safety interventions as data on such interventions
were not available. However, these are recommended based on our
ndings, domain knowledge, and the study of previous research con-
ducted on pedestrian and cyclist safety. As more data become available,
estimating the effectiveness of countermeasures in the context of e-
scooter safety would be an important direction for future research.
Table 4 reports some exemplars of relevant stakeholders for each safety
intervention category discussed above.
Our results might have been affected by the recent pandemic to some
extent; however, the model can be readily updated as more recent e-
scooter crash data (e.g., post-pandemic crash data) become available. In
future research considering differing injury severity levels (e.g., slight,
serious, and fatal injuries) would provide more detailed insights into
understanding e-scooter safety. As new evidence will appear in this
context in the future, our ndings can be used to improve road safety,
proactively, not only for e-scooter riders but also for other road users,
particularly pedestrians.
Declaration of Competing Interest
The authors declare no conict of interest.
References
Amoh-Gyimah, R., Saberi, M., & Sarvi, M. (2016). Macroscopic modeling of pedestrian
and bicycle crashes: A cross-comparison of estimation methods. Accident Analysis and
Prevention, 93, 147–159.
Azimian, A., & Jiao, J. (2022). Modeling factors contributing to dockless e-scooter injury
accidents in Austin. Texas. Trafc Injury Prevention, 23(2), 107–111.
Bhat, C. R., Astroza, S., & Lavieri, P. S. (2017). A new spatial and exible multivariate
random-coefcients model for the analysis of pedestrian injury counts by severity
level. Analytic Methods in Accident Research, 16, 1–22.
Bodansky, D., Gach, M., Grant, M., Solari, M., Nebhani, N., Crouch-Smith, H., et al.
(2022). Legalisation of e-scooters in the UK: The injury rate and pattern is similar to
those of bicycles in an inner city metropolitan area. Public Health, 206, 15–19.
Brownson, A. B., Fagan, P. V., Dickson, S., & Civil, I. D. (2019). Electric scooter injuries at
Auckland City Hospital. The New Zealand Medical Journal (Online), 132(1505),
62–72. Available: https://journal.nzma.org.nz/journal-articles/electric-scooter-injur
ies-at-auckland-city-hospital. Accessed online 14/12/2021.
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.
Accident Analysis and Prevention, 93, 14–22.
Cicchino, J. B., Kulie, P. E., & McCarthy, M. L. (2021). Severity of e-scooter rider injuries
associated with trip characteristics. Journal of Safety Research, 76, 256–261.
Cottrill, C. D., & Thakuriah, P. V. (2010). Evaluating pedestrian crashes in areas with
high low-income or minority populations. Accident Analysis and Prevention, 42(6),
1718–1728.
Daziano, R. A., Miranda-Moreno, L., & Heydari, S. (2013). Computational Bayesian
statistics in transportation modeling: From road safety analysis to discrete choice.
Transport Reviews, 33(5), 570–592.
de Bortoli, A., & Christoforou, Z. (2020). Consequential LCA for territorial and
multimodal transportation policies: Method and application to the free-oating e-
scooter disruption in Paris. Journal of Cleaner Production, 273, Article 122898.
de Valpine, P., Turek, D., Paciorek, C. J., Anderson-Bergman, C., Lang, D. T., & Bodik, R.
(2017). Programming with models: Writing statistical algorithms for general model
structures with NIMBLE. Journal of Computational and Graphical Statistics, 26(2),
403–413.
Department for Transport (DfT) (2020). E-scooter trials: Guidance for local areas and
rental operators. Available: https://www.gov.uk/government/publications/e-scoote
r-trials-guidance-for-local-areas-and-rental-operators/e-scooter-trials-guidance-for-
local-areas-and-rental-operators. Accessed online 15/12/2021.
Department for Transport (DfT) (2021). Perceptions of current and future e-scooter use
in the UK. https://www.gov.uk/government/publications/e-scooters-public-percept
ions. Accessed online 15/12/2021.
Ding, H., Sze, N. N., Li, H., & Gup, Y. (2020). Roles of infrastructure and land use in
bicycle crash exposure and frequency: A case study using Greater London bike
sharing data. Accident Analysis and Prevention, 144, Article 105652.
Dupont, E., Papadimitriou, E., Martensen, H., & Yannis, G. (2013). Multilevel analysis in
road safety research. Accident Analysis and Prevention, 60, 402–411.
El-Basyouny, K., & Sayed, T. (2009). Accident prediction models with random corridor
parameters. Accident Analysis and Prevention, 41(5), 1118–1123.
Frias-Martinez, V., Sloate, E., Manglunia, H., & Wu, J. (2021). Causal effect of low-
income areas on shared dockless e-scooter use. Transportation Research Part D:
Transport and Environment, 100, Article 103038.
Gelman, A., Hwang, J., & Vehtari, A. (2013). Understanding predictive information
criteria for Bayesian models. Statistics and Computing, 24, 997–1016.
Gibson, H., Curl, A., & Thompson, L. (2021). Blurred boundaries: E-scooter riders’ and
pedestrians’ experiences of sharing space. Mobilities, 69–84. https://doi.org/
10.1080/17450101.2021.1967097
Gioldasis, C., Christoforou, Z., & Seidowsky, R. (2021). Risk-taking behaviors of e-scooter
users: A survey in Paris. Accident Analysis and Prevention, 163, Article 106427.
Gitelman, V., Levi, S., Carmel, R., Korchatov, A., & Hakkert, S. (2019). Exploring patterns
of child pedestrian behaviors at urban intersections. Accident Analysis and Prevention,
122, 36–47.
Graham, D., Glaister, S., & Anderson, R. (2005). The effects of area deprivation on the
incidence of child and adult pedestrian casualties in England. Accident Analysis and
Prevention, 37, 125–135.
Graham, D., Mccoy, E., & Stephens, D. (2013). Quantifying the effect of area deprivation
on child pedestrian casualties by using longitudinal mixed models to adjust for
confounding, interference and spatial dependence. Journal of the Royal Statistical
Society. Series A: Statistics in Society, 176(4), 931–950.
Graham, D., & Stephens, D. (2008). Decomposing the impact of deprivation on child
pedestrian casualties in England. Accident Analysis and Prevention, 40, 1351–1364.
Green, J., Muir, H., & Maher, M. (2011). Child pedestrian casualties and deprivation.
Accident Analysis and Prevention, 43(3), 714–723.
Greater London Authority (GLA), (2015a). London Ward Proles and Atlas. Available:
https://data.london.gov.uk/dataset/ward-proles-and-atlas. Accessed online 11/
11/21.
Greater London Authority (GLA) (2015b). London Borough Proles and Atlas. Available:
https://data.london.gov.uk/dataset/london-borough-proles. Accessed online 11/
11/21.
Hauer, E. (1997). Observational before/after studies in road safety: Estimating the effect of
highway and trafc engineering measures on road safety (1st ed.). Oxford: Pergamon
Press.
Haworth, N., Schramm, A., & Twisk, D. (2021a). Comparing the risky behaviours of
shared and private e-scooter and bicycle riders in downtown Brisbane, Australia.
Accident Analysis and Prevention, 152, Article 105981.
Haworth, N., Schramm, A., & Twisk, D. (2021b). Changes in shared and private e-scooter
use in Brisbane, Australia and their safety implications. Accident Analysis and
Prevention, 163, Article 106451.
Heydari, S., Fu, L., Lord, D., & Mallick, B. K. (2016). Multilevel Dirichlet process mixture
analysis of railway grade crossing crash data. Analytic Methods in Accident Research,
9, 27–43.
Table 4
Categories of safety interventions against potential stakeholders
1
.
Safety intervention category Potential respective stakeholders
Educational programmes and
publicity campaigns
Micromobility service providers, Local
Authorities, Transport for London, Department for
Transport, The Royal Society for the Prevention of
Accidents
Enforcement Police forces, Driver and Vehicle Licensing
Agency, and micromobility service providers, to
some extent, particularly over parking, where
Local Authorities may also have a role
Trafc engineering
interventions
Local Authorities, Transport for London,
Department for Transport
1 Note that this is not an exhaustive list and in many cases the roles may overlap.
S. Heydari et al.

Sustainable Cities and Society 83 (2022) 103982
9
Heydari, S., Fu, L., Miranda-Moreno, L., & Joseph, L. (2017). Using a exible multivariate
latent class approach to model correlated outcomes: A joint analysis of pedestrian
and cyclist injuries. Analytic Methods in Accident Research, 13, 16–27.
Heydari, S., Fu, L., Thakali, L., & Joseph, L. (2018). Benchmarking regions using a
heteroskedastic grouped random parameters model with heterogeneity in mean and
variance: Applications to grade crossing safety analysis. Analytic Methods in Accident
Research, 19, 33–48.
Heydari, S., Konstantinoudis, G., & Behsoodi, A. (2021). Effect of the COVID-19
pandemic on bike-sharing demand and hire time: Evidence from Santander cycles in
London. PloS one, 16(12), Article E0260969.
Heydari, S., Miranda-Moreno, L., & Hickford, A. J. (2020). On the causal effect of
proximity to school on pedestrian safety at signalized intersections: A heterogeneous
endogenous econometric model. Analytic Methods in Accident Research, 26, Article
100115.
Hosseinzadeh, A., Algomaiah, M., Kluger, R., & Li, Z. (2021). E-scooters and
sustainability: Investigating the relationship between the density of E-scooter trips
and characteristics of sustainable urban development. Sustainable Cities and Society,
66, 10262.
Huang, H., & Abdel-Aty, M. (2010). Multilevel data and Bayesian analysis in trafc
safety. Accident Analysis and Prevention, 42(6), 1556–1565.
Islam, M. T., & El-Basyouny, K. (2015). Multilevel models to analyze before and after
speed data. Analytic Methodsin Acc. Res., 8, 33–44.
Kobayashi, L., Williams, E., Brown, C., et al. (2019). The e-merging e-pidemic of e-
scooters. Trauma Surgery & Acute Care Open;, 4, Article E000337.
Li, H., Graham, D. J., & Liu, P. (2017). Safety effects of the London cycle superhighways
on cycle collisions. Accident Analysis and Prevention, 99, 90–101.
Lovelace, R., Beecham, R., Heinen, E., Tortosa, E. V., Yang, Y., Slade, C., et al. (2020). Is
the London cycle hire scheme becoming more inclusive? An evaluation of the
shifting spatial distribution of uptake based on 70 million trips. Transportation
Research Part A: Policy and Practice, 140, 1–15.
Mannering, F. L., Shankar, V., & Bhat, C. R. (2016). Unobserved heterogeneity and the
statistical analysis of highway accident data. Analytic Methods in Accident Research,
11, 1–16.
Noland, R. B. (2021). Scootin’ in the rain: Does weather affect micromobility?
Transportation Research Part A: Policy and Practice, 149, 114–123.
Noland, R. B., & Laham, M. L. (2018). Are low income and minority households more
likely to die from trafc-related crashes? Accident Analysis and Prevention, 120,
233–238.
Osama, A., & Sayed, T. (2016). Evaluating the impact of bike network indicators on
cyclist safety using macro-level collision prediction models. Accident Analysis and
Prevention, 97, 28–37.
Rix, K., Demchur, N., Zane, D., & Brown, L. (2021). Injury rates per mile of travel for
electric scooters versus motor vehicles. The American Journal of Emergency Medicine,
40, 166–168.
Seraneeprakarn, P., Huang, S., Shankar, V., Mannering, F., Venkataraman, N., &
Milton, J. (2017). Occupant injury severities in hybrid-vehicle involved crashes: A
random parameters approach with heterogeneity in means and variances. Analytic
Methods in Accident Research, 15, 41–55.
Shah, N. R., Aryal, S., Wen, Y., & Cherry, C. R. (2021). Comparison of motor vehicle-
involved e-scooter and bicycle crashes using standardized crash typology. Journal of
Safety Research, 77, 217–228.
Shaheen, S., & Cohen, A. (2019). Shared micromoblity policy toolkit: Docked and
dockless bike and scooter sharing. Available: https://doi.org/10.7922/G2TH8JW7.
Accessed online 14/01/2022.
Siebert, F. W., Ringhand, M., Englert, F., Hoffknecht, M., Edwards, T., & R¨
otting, M.
(2021). Braking bad - ergonomic design and implications for the safe use of shared E-
scooters. Safety Science, 140, Article 105294.
Stigson, H., Malakuti, I., & Klingegård, M. (2021). Electric scooters accidents: Analyses of
two Swedish accident data sets. Accident Analysis and Prevention, 163, Article
106466.
Su, J., Sze, N. N., & Bai, L. (2021). A joint probability model for pedestrian crashes at
macroscopic level: Roles of environment, trafc, and population characteristics.
Accident Analysis and Prevention, 150, Article 105898.
Torkayesh, A. E., & Deveci, M. (2021). A multi-normalization multi-distance assessment
(TRUST) approach for locating a battery swapping station for electric scooters.
Sustainable Cities and Society, 74, Article 103243.
Tran, P. T. M., Adam, M. G., Tham, K. W., Schiavon, S., Pantelic, J., Linden, P., et al.
(2021). Assessment and mitigation of personal exposure to particulate air pollution
in cities: An exploratory study. Sustainable Cities and Society, 72, Article 103052.
Transport for London (TfL), (2021a). Electric scooters. Available: https://t.gov.uk/
modes/driving/electric-scooter-rental-trial. Accessed online 22/01/2022.
Transport for London (TfL), (2021b). Travel in London report 14. Available: https://t.
gov.uk/corporate/publications-and-reports/travel-in-london-reports. Accessed
online 17/01/2022.
Tuli, F., Mitra, S., & Crews, M. (2021). Factors inuencing the usage of shared E-scooters
in Chicago. Transportation Research Part A: Policy and Practice, 154, 164–185.
Wang, H., & Noland, R. (2021). Bikeshare and subway ridership changes during the
COVID-19 pandemic in New York City. Transport Policy, 106, 262–270.
Wang, X., Yang, J., Lee, C., Ji, Z., & You, S. (2016). Macro-level safety analysis of
pedestrian crashes in Shanghai, China. Accident Analysis & Prevention, 96, 12–21.
Washington, S., Karlaftis, M. G., Mannering, F., & Anastasopoulos, P. (2020). Statistical
and econometric methods in transportation. Chapman and Hall/CRC. https://doi.
org/10.1201/9780429244018
Watanabe, S. (2010). Asymptotic equivalence of Bayes cross validation and widely
applicable information criterion in singular learning theory. Journal of Machine
Learning Research, 11, 3571–3594.
Winchcomb, M. (2021). The safety of private e-scooters in the uk interim report. PACTS.
Available: https://www.pacts.org.uk/the-safety-of-private-e-scooters-interim-report
-from-pacts/. Accessed online 16/12/2021.
Yang, H., Ma, Q., Wang, Z., Cai, Q., Xie, K., & Yang, D. (2020). Safety of micro-mobility:
Analysis of E-Scooter crashes by mining news reports. Accident Analysis and
Prevention, 143, Article 105608.
Zhan, X., Abdul Aziz, H. M., & Ukkusuri, S. V. (2015). An efcient parallel sampling
technique for Multivariate Poisson-Lognormal model: Analysis with two crash count
datasets. Analytic Methods in Accident Research, 8, 45–60.
S. Heydari et al.