Acasă Explorează Ajutor
Autentificare
ardo
/
Obsidian
1
0
Bifurcare 0
Fisiere Probleme 0 Trageți solicitările 0 Wiki
Arbore: a2f8402268
Ramuri Etichete
Obsidian
/
Zotero
/
001_artiklid
/
Road Safety Indicators
/
choucomparativemodelingriskfactors2024.md

choucomparativemodelingriskfactors2024.md 137 KB
Istoric Crud

category: literaturenote citekey: choucomparativemodelingriskfactors2024 title: Comparative modeling of risk factors for near-crashes from crowdsourced bicycle airbag helmet data and crashes from conventional police data authors: "Chou, Kuan-Yeh; Paulsen, Mads; Jensen, Anders Fjendbo; Rasmussen, Thomas Kjær; Nielsen, Otto Anker" year: 2024 date: 2024-00-00 2024 doi: 10.1016/j.jsr.2024.10.003 publication: Journal of Safety Research url: "https://doi.org/10.1016/j.jsr.2024.10.003" zotero_key: 75XKTJVR zotero_storage: 55QHPKFE collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Chou, Kuan-Yeh"

status: converted

Contents lists available at ScienceDirect

Journal of Safety Research

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

Comparative modeling of risk factors for near-crashes from crowdsourced bicycle airbag helmet data and crashes from conventional police data

Kuan-Yeh Chou ∗ , Mads Paulsen, Anders Fjendbo Jensen, Thomas Kjær Rasmussen, Otto Anker Nielsen

Technical University of Denmark, Department of Technology, Management and Economics, Akademivej, Building 358, 2800 Kgs., Lyngby, Denmark

A R T I C L E I N F O

Keywords: Bicycle crash Bicycle near-crash Crowdsourced cycling data Crash-rate model Bicycle infrastructure

A B S T R A C T

Introduction: Conventional cycling crash data is valuable for shaping safe cycling environments but has limitations due to the rarity and under-reporting of cycling crashes. However, recent technological developments can provide information from near-crashes. the subheads should be italic, not bf. Also in the Abstract, there shouldn't be hard return between subheads, the whole section should all run together, so run up any text between subheads. Method: With Metropolitan Copenhagen as a case, this study uses a very large crowdsourced near-crash dataset from Hövding bicycle airbag helmet users and conventional police crash data to model and identify differences in the infrastructure factors influencing rates of crashes and near-crashes in these datasets. Results: In contrast to existing literature, our results show considerable differences in the factors influencing the frequency of crashes and near-crashes. The risk of crashes increases predominantly at intersections and roundabouts, whereas near-crashes are also associated with infrastructure types shared with pedestrians. Conclusion: When used complementarily, crowdsourced near-crash data can enrich the data foundation and help increase the awareness of near-crash-prone infrastructure types necessary for shaping more comprehensive cycling safety policies. Practical Applications: The findings of the study advocate for a broader perspective on cyclist safety, incorporating currently undisclosed near-crash-prone infrastructure types, such as paths shared by cyclists and pedestrians.

1. Introduction

Cycling is increasingly gaining attention as a sustainable transport alternative, due to a growing interest in public health, environmental awareness, and efficient urban land use (Bagloee et al., 2016; Cole-Hunter et al., 2015; Raustorp & Koglin, 2019). Bicycle use has generally increased over the past two decades, mainly in Europe and the Americas, especially during the COVID-19 pandemic (Pucher & Buehler, 2017; Schwarz et al., 2022). In relation to this, bicycle safety is receiving more attention worldwide, and many cities seek to provide safe urban cycling environments through infrastructure and traffic regulations (e.g. Buehler & Pucher, 2022; Dozza et al., 2017; Li et al., 2022; Nikitas et al., 2021; Peden et al., 2022).

To effectively reduce bicycle crashes, credible data sources should be used to identify and analyze problematic locations and potential causes of bicycle crashes (Salmon et al., 2022). Most studies use crash data recorded by the police or hospitals to detect crash hotspots (Das et al., 2021; Hamann & Peek-Asa, 2017; Ringel et al., 2023). However, such records suffer from under-reporting and rare occurrences (Hosseinpour et al., 2021; Ringel et al., 2023; Springer et al., 2020) and typically involve motorized vehicles and severe injuries that are likely to occurin specific types of crashes. Consequently, it has been suggested to explore different types of incidents that might be related to risk for cyclists but occur much more frequently than crashes recorded by police or hospitals (Ibrahim et al., 2021; Puchades et al., 2018; Springer et al., 2020). For potential collisions between cars and cyclists, there is even evidence that the difference between a crash and a near-crash to a large extent can be attributed to the reaction time of the drivers and the time to collision when the cyclist enters the driver's field of vision (Ito et al., 2018).

Existing research related to bicycle near-crashes relies largely on survey data that has a limited number of observations and is subject to self-report bias (Aldred, 2016; Aldred & Goodman, 2018; Poulos et al., 2017). Thus, compared to actual bicycle crashes, there is a lack of data sources and understanding of the characteristics of bicycle nearcrashes (Aldred & Goodman, 2018; Karakaya et al., 2020; Terum &

E-mail addresses: kuach@dtu.dk (K.-Y. Chou), madsp@dtu.dk (M. Paulsen), afjje@dtu.dk (A.F. Jensen), tkra@dtu.dk (T.K. Rasmussen), oani@dtu.dk (O.A. Nielsen).

https://doi.org/10.1016/j.jsr.2024.10.003

∗ Corresponding author.

Svartdal, 2019). Consequently, Ibrahim et al. (2021) calls for more studies on the infrastructure, perceptional, and behavioral characteristics of bicycle near-crashes. Therefore, determining factors causing near-crashes becomes crucial since it enables us to identify potential causes and certain types of risky locations that cannot be identified solely by conventional crash data from the police or hospitals. In the present study, locations of near-crashes are identified by using a largescale crowdsourced data source that on top of recording trajectory data also classifies abnormal movements based on accelerometer data.

Due to emerging technology, sensors embedded in bicycles or bicycle devices can accurately detect not only crashes but also other abnormal movements, often referred to as near-crashes (Hamann & Peek-Asa, 2017; Ringel et al., 2023), near misses (Aldred, 2016; Aldred & Crosweller, 2015; Aldred & Goodman, 2018; Branion-Calles et al., 2017; Ibrahim et al., 2021; Karakaya et al., 2020; Matsui et al., 2016; Poulos et al., 2017; Puchades et al., 2018; Sanders, 2015; Sanders & Nelson, 2023), near accidents (Mesimäki & Luoma, 2021), safetycritical events (Das et al., 2021; Lindqvist & Roos, 2020; Schleinitz et al., 2015; Springer et al., 2020; Werneke et al., 2015), or dangerous decelerations (Strauss et al., 2017). Research indicates that crowdsourced bicycle crashes and near-crash data collected by sensors enable a more comprehensive analysis in the research domain of bicycle safety (Nelson et al., 2021). Indeed, some studies have shown that infrastructure types leading to near-crashes are similar to those leading to actual crashes, indicating that near-crash data can be a relevant supplement when analyzing crash risks (Guo et al., 2010; Puchades et al., 2018; Strauss et al., 2017; Wright & van der Schaaf, 2004).

The contributions of the study are as follows. Firstly, it is the first study to model the occurrence rate of bicycle near-crashes using revealed near-crash data. We do so through a crowdsourced dataset consisting of bicycle accelerometer data and trajectory data as well as detailed infrastructure characteristics. Secondly, we estimate a complementary network-based crash rate model from police crash data, enabling us to comparatively investigate the differences and similarities between infrastructure characteristics that, according to the estimated models, are more likely to cause bicycle near-crashes of Hövding users and crashes registered in police crash data, respectively. This comparative analysis method between the near-crash rate and crash rate models contributes to discovering near-crash-prone infrastructure characteristics that cannot be identified solely by conventional crash data. Thirdly, for the near-crash rate model, we utilize the uniqueness of our crowdsourced data source which allows us to use exact exposure directly obtained from the trajectories of users across all their trips, completely eliminating uncertainty from the exposure measures.

The remainder of the paper is structured as follows. Section 2 describes the literature related to bicycle near-crashes. Section 3 introduces the near-crash, crash, exposure, and network data and elaborates how these datasets were processed to be usable in the analysis. Section 4 presents the methodology for modeling near-crash and crash rates. Section 5 provides the specific findings and main results. Section 6 discusses the results. Section 7 elaborate limitations and future research of the study. Section 8 concludes the paper.

2. Literature review

In this section, we review studies investigating bicycle near-crashes to emphasize relevant research gaps.

Existing studies regarding bicycle near-crashes are organized in Tables 1 and 2, which also indicate differences in research objectives, study areas, data sources, and methods. We also refer readers to the review by Ibrahim et al. (2021).

Most studies shown in the table have explored the factors affecting bicycle near-crashes and investigated the similarities and differences between occurrence patterns of bicycle near-crashes and crashes. Generally, they found that factors causing near-crashes and crashes are similar (Branion-Calles et al., 2017; Lindqvist & Roos, 2020; Matsui et al., 2016; Mesimäki & Luoma, 2021; Poulos et al., 2017); thus, suggesting that near-crashes can represent or supplement crashes when analyzing and identifying crash-prone locations. One exception is Ringel et al. (2023) who found that relative to crashes, near-crashes are more often related to overtaking or lane changing, whereas solo near-crashes without collisions are rare. However, they only studied near-crashes through descriptive analysis of survey data and did not explore the underlying contributing factors.

Another takeaway from the previous literature (Table 5) is that large-scale crowdsourced bicycle data is an emerging data source, obtained through devices such as smartphones using sensors to record cyclists' movements in great detail. This includes precise location and accelerometer data, which allow for the detection of potentially risky situations for cyclists such as acute braking (Strauss et al., 2017), strong head movements (Lindqvist & Roos, 2020), close passes (Karakaya et al., 2020), and left/right hooks (Karakaya et al., 2020). According to Ibrahim et al. (2021), Karakaya et al. (2020), Puchades et al. (2018), a sufficient number of such near-crashes and their precise locations can identify crash-prone hotspots and thus serve as an early warning before the actual crashes occur. However, among the 20 relevant studies shown in Tables 1 and 2, only five studies used crowdsourced data, and only three used large-scale GPS trajectory and sensor data (i.e., more than 10,000 bicycle trips) (Karakaya et al., 2020; Lindqvist & Roos, 2020; Strauss et al., 2017). The rest of the studies are based on data from surveys, videos, and field observation, which suffer from small sample sizes and self-reporting bias.

Statistical modeling is a solid and widely recognized analysis method to investigate how different factors affect the occurrence of bicycle crashes (Poulos et al., 2017). Most studies based on large-scale crowdsourced data did not model near-crash rates (Karakaya et al., 2020; Lindqvist & Roos, 2020; Strauss et al., 2017). Indeed, only four of the reviewed studies modeled the frequency or rate of bicycle near-crashes (Aldred & Crosweller, 2015; Aldred & Goodman, 2018; Branion-Calles et al., 2017; Poulos et al., 2017). Based on survey data, Aldred and Goodman (2018) modeled factors influencing the frequency and near-crash rate (i.e., the number of near-crashes per mile/hour traveled). Aldred and Crosweller (2015) modeled how gender, age, speed, peak hour riding time, and weekday vs. weekend contribute to occurrence rates and scariness. Poulos et al. (2017) modeled how gender, age, trip purpose, and cycling experience contribute to near-crash rates (i.e., the number of near-crashes per 1000 h traveled). Finally, Branion-Calles et al. (2017) utilized crowdsourced data to model how demographic factors, trip features, and infrastructure contribute to the frequency of both near-crashes and crashes and compared similarities and differences across several factors between near-crashes and crashes.

The present study uses an unprecedentedly large crowdsourced dataset of bicycle trajectories and near-crashes based on very precise accelerometer data with a 200 Hz sampling rate from dedicated sensors, comprising 129,744 near-crashes occurring across 365,813 trips made by 10,049 cyclists in a 20-month observation period. By comparison, Lindqvist and Roos (2020) and Karakaya et al. (2020) obtained 2721 and 7173 near-crashes from 16,760 and over 10,000 trips, respectively, over much shorter study periods. Although Strauss et al. (2017) obtained a larger number of near-crashes than the present study, their data was collected from far fewer bicycle trips and only explored near-crashes that involved urgent deceleration.1 Their detection of near-crashes was based on smartphone GPS data of low granularity (i.e., less precise deceleration rates with a one-second frequency), which adds considerable uncertainty to the detection of urgent

1 The actual number is not stated in the paper, but based on numbers in tables the number of incidents can be estimated to be approximately to be just over 400,000. They furthermore report that ''over 10,000 trips were recorded by almost 1000 cyclists''.

K.-Y. Chou et al.

Table 1 Summary of studies related to observed bicycle near-crashes.

Study Study area
(Period)		 Objective Aspects explored Data source Number of observations Method Main finding
The present
study		 Metropolitan
Copenhagen,
Denmark
(20 months)		 1. Explore causes
of (near-)crashes
2. Compare
near-crashes
with crashes
3. Investigate
(near-)crash
hotspots		 Infrastructure Crowdsourced
(Bicycle airbag
helmet)		 Bicycle trips: 365,813
Crashes: 97
Near-crashes: 129,744
Crash from police data:
4,220		 Negative
binomial
regression model		 Some factors
causing
near-crashes and
crashes are
similar, but
others are
different
Lindqvist and
Roos (2020)		 Malmo, Sweden
(6 months)		 Bicycle trips: 16,760
Crash: 33
Near-crashes: 2721
Crash from
police/emergency data:
646		 Density-based
spatial clustering		 Hotspots of
near-crashes and
crashes are
similar
Strauss et al.
(2017)		 Montreal,
Canada (4.5
months)		 1. Explore causes
of near-crashesa
2. Compare
near-crashes
with crashes		 Infrastructure
Behavior		 Crowdsourced
(Smartphone)		 Bicycle trips: Over
10,000
Near-crashes: >400,000
Crashes from historic
data: 4,900 injuries		 Correlation test The frequency of
near-crashes is
significantly
correlated with
the frequency of
crashes
Matsui et al.
(2016)		 Japan
(5 years)		 1. Explore causes
of near-crashes
2. Compare
near-crashes
with crashes
(Specifically
between cyclist
and vehicle)		 Behavior Car video Participants: 463
Near-crashes: 229
Crashes from national
record: 2818		 Analyzed time to
collision		 Factors causing
near-crashes and
crashes are
similar
Ito et al.
(2018)		 Crashes: 47
Near-crashes: 532		 Crashes caused
by low time to
collision when a
cyclist appears
or late braking
Karakaya
et al. (2020)		 Germany: Berlin,
Augsburg,
Bochum,
Pforzheim,
Stuttgart.
Switzerland:
Bern
(7 months)		 1. Explore causes
of near-crashes
2. Investigate
near-crash
hotspots		 Infrastructure
Safety perception
Behavior		 Crowdsourcedb
(Smartphone)		 Bicycle trips: Over
10,000
Near-crashes: 7,173		 Score of
dangerous level		 Near-crashes can
be used to
identify
previously
unknown risky
hotspots
Springer
et al. (2020)		 Chemnitz,
Germany
(1 week)		 Explore causes
Behavior
of near-crashes		 Field observation Participants: 1,202
Near-crashes: 17		 Naturalistic
study		 Near-crashes can
be used to
identify
unrevealed risky
hotspots
Dozza and
Werneke
(2014)		 Gothenburg,
Sweden
(2 weeks)		 Infrastructure 1. GPS-enabled
video
2. Brake force
sensor		 Participants: 16
Bicycle trips: 332
Near-crashes: 63
Dozza (2020) Compare
near-crashes and
crashes		 Temporal
Spatial		 Near-crashes: 30
Police crashes: 481
Hamann and
Peek-Asa
(2017)		 Iowa City,
Johnson County,
US
(1 week)		 Explore causes
of (near-)crashes
among children
and adults		 Infrastructure
Behavior		 GPS-enabled
video		 Participants: 20
Bicycle trips: 261
Crashes: 2
Near-crashes: 10		 Cyclist behavior
during
near-crash and
crash between
children and
adults are
similar
Schleinitz
et al. (2015)		 Chemnitz,
Germany
(4 weeks)		 Explore different
types of
near-crashes in
different types of
infrastructure		 1. Video
2. Speed sensor		 Participants: 31
Bicycle trips: 1667
Near-crashes: 77		 Over half of
near-crashes
involve other
cyclists and
pedestrians, and
they occurred
frequently on
shared road

a Only dangerous decelerations were considered.

b After near-crashes are collected, those data will be labeled and annotated by participants to elaborate on which type of near-crash (in Karakaya et al. (2023) the labeling process of near-crash is partially automated) and how dangerous every participant perceived each recorded near-crash.

Table 2 Summary of studies related to self-reported bicycle near-crashes.

Study Study area (Period) Objective Aspects explored Data source Number of
observations		 Method Main finding
Branion
Calles et al.
(2017)		 Victoria and
Vancouver, Canada
(2 years)		 1. Explore causes of
(near-)crashes
2. Compare
near-crashes with
crashes		 Infrastructure
Demographicsa
Weather		 Crowdsourcedb
(Open information
platform)		 Crashes: 97
Near-crashes: 307
Crashes from
insurance company:
599		 Logistic regression
model		 Factors causing
near-crashes and
crashes are similar
Poulos et al.
c
(2017)		 New South Wales,
Australia
(9 months)		 Infrastructure
Behavior
Demographics
Temporal		 Survey
(Cycling diary)		 Participants: 2038
Crashes: 198
Near-crashes: 3437		 Poisson regression
model
Aldred
(2016)		 UK
(2 weeks)		 Infrastructure
Safety perception
Behavior		 Participants: 1525
Near-crashes: 3994
Crashes from police
data: 2565		 Binary logistic
regression model		 Factors causing
near-crashes and
minor-injury crashes
are similar
Ringel et al.
(2023)		 Zurich, Switzerland
(3 months)		 Safety perception
Demographic
Behavior		 Crashes: 416
Near-crashes: 1452
Crashes from police
data: 4,305		 Descriptive analysis Near-crashes can be
used to identify
unrevealed
Mesimäki
and Luoma
(2021)		 Cities in Finland
(3 years)		 1. Explore
near-crash frequency
2. Compare
near-crashes with
crashes (Specifically
between cyclist and
pedestrian)		 Infrastructure.
Safety perception.
Behavior		 Participants: 1046
Crashes: 21
Near-crashes: 1001		 Factors causing
near-crashes and
crashes are similar
Sanders
(2015)		 San Francisco Bay
Area, US
(1 month)		 1. Explore
near-crash frequency
2. Compare
near-crashes with
crashes (Specifically
between cyclist and
vehicle)		 Safety perception.
Behavior		 Survey Participants: 463
Crashes: 82
Near-crashes: 917		 1. Kruskal Wallis
test
2
2. 𝜒
test
3. Logistic
regression model		 Near-crashes has
more influence on
cyclists'
perceived safety
than crash
Das et al.
(2021)		 Phoenix Area, AZ,
US
(7 years)		 1. Explore causes of
(near-)crashes
2. Compare
near-crashes with
crashes
3. Examine changes
in perceived safety
after experiencing
(near-)crashes		 Infrastructure
Safety perception
Demographics
Temporal
Weather		 Crowdsourcedb
(Open information
platform)		 Crashes: 108
Near-crashes: 118		 Taxicab
correspondence
analysis		 Factors causing
near-crashes and
crashes are similar
Aldred and
Goodman
(2018)		 UK
(2 years)		 1. Explore causes of
near-crashes
2. Examine how
perceived safety
affects the judgment
of near-crashes		 Safety perception
Temporal		 Survey
(Cycling diary)		 Participants: 1061
Near-crashes: 1875		 Poisson regression
model		 Patterns and
scariness of
near-crashes are
similar to ones
recorded after a
year
Aldred and
Crosweller
(2015)		 UK
(2 weeks)		 1. Explore causes of
near-crashes
2. Examine changes
in perceived safety
after experiencing
(near-)crashes		 Safety perception
Temporal
Demographics		 Participants: 1525
Near-crashes: 3994		 1. Cross-tabulation
2. Regression model		 Cyclist-vehicle
near-crashes make
cyclists feel higher
levels of scariness
Sanders and
Nelson
(2023)		 Tempe, AZ, US
(1 day)		 Safety perception
Behavior		 Survey Participants: 1256
Near-crashes: 54
Near-crashes: 138		 1.Kruskal Wallis test
2
2.𝜒
test
3.Z test		 Cyclists-vehicle
near-crashes
occurred most
frequently
Puchades
et al. (2018)		 Italy
(3 months)		 Explore causes of
near-crashes		 Participants: 298
Near-crashes: 393		 Bayesian path
model		 Near-crashes can be
used to identify
previously unknown
risky hotspots
Werneke
et al. (2015)		 Gothenburg, Sweden
(2 weeks)		 Infrastructure
Safety perception
Behavior		 1. Survey (Cycling
diary)
2. Interview		 Participants: 16
Bicycle trips: 332
Near-crashes: 63		 Naturalistic study Near-crashes mainly
involve other road
users
and due to
unmaintained
infrastructure

a This contains also the trip features of participants.

b They used the data from Nelson et al. (2015).

c Detailed study protocols were adopted from Poulos et al. (2012).

decelerations. As mentioned previously, neither of the three studies used large-scale GPS data to model the occurrence rate for near-crashes and crashes.

According to Kaplan and Prato (2015), there is a need for highresolution network-based near-crash models that provide deeper insights into how comprehensive infrastructure such as various road types (e.g., protected bicycle tracks, bicycle lanes, or shared paths) and surface types (e.g., asphalt or cobblestones) influence near-crash rates. Nevertheless, none of the studies modeled bicycle near-crash rates with detailed infrastructure features. Only Branion-Calles et al. (2017) considered infrastructure features, and they did so through less detailed information, such as whether or not the road contains bicycle facilities like local street bikeways, painted lanes, or separated tracks. Therefore, only a superficial understanding of the impact of infrastructures on near-crash rates could be obtained.

Bicycle (near-)crash rates are calculated as (near-)crash frequency divided by bicycle exposure (the traffic at the location), which is a more accurate indicator for identifying crash-prone locations than using only (near-)crash frequency (Myhrmann & Mabit, 2023; Saad et al., 2019; Vanparijs et al., 2015). Thus, to accurately model (near- )crash rates, it is essential to include bicycle exposure, i.e. the annual number of kilometers traveled on each link of the network. (Near-)crash rates inferred from external exposure data sources are approximated through, for instance, traffic models or surveys which is less accurate than directly observed exposure. Among studies modeling near-crash rates, only Poulos et al. (2017) and Aldred and Crosweller (2015) included bicycle exposure obtained from surveys (cycling diaries) in their models. Exposure was in both studies calculated from temporal perspectives (i.e., number of near-crashes per distance and hour traveled at each time of day) without spatial aspects, thus making it difficult to identify crash-prone hotspots or risky infrastructure types at the network level.

Owing to the wide availability of smart-sensing devices in recent years, (near-)crashes and cyclists' trajectories can be collected simultaneously from devices such as smartphone applications using sensors and GPS (Dhakal et al., 2018; Ibrahim et al., 2021; Karakaya et al., 2023; Saad et al., 2019; Strauss et al., 2017). This allows the relevant exposure to be inferred directly from the trajectory data and thus allows (near-)crash rates to be calculated more accurately. Furthermore, matching the trajectories with high-resolution network data can make locations of exposure more accurate (Kaplan & Prato, 2015).

In summary, the present study addresses several research gaps identified in the research shown in Tables 1 and 2. First, no studies have identified the difference between factors causing near-crashes and crashes involving cyclists. All studies conducting comparative analysis between bicycle near-crashes and crashes focused on similar contributing factors and concluded that the two types of crashes were influenced by similar factors. In the present study, we find that the factors differ considerably. Second, no studies have modeled bicycle near-crash rates by using large-scale near-crash and exposure data. Aldred and Crosweller (2015), Poulos et al. (2017) modeled bicycle near-crash rates with exposure but did not use large-scale data, and although Karakaya et al. (2020), Lindqvist and Roos (2020), Strauss et al. (2017) used large-scale bicycle near-crash and trajectory data, they did not model near-crash rates nor frequencies. This gap is also highlighted in the review by Ibrahim et al. (2021), who address the necessity of sufficient near-crash data that allows estimating statistically significant risk factors. Finally, no studies have modeled bicycle near-crashes at high-resolution network levels containing detailed infrastructure features.

To fill these gaps, the present study models network-based bicycle near-crash and crash rates by using large-scale crowdsourced data, including exposure. Both models contain detailed infrastructure features, including 16 road types, four surfaces, five point-based infrastructures, and three other infrastructures. Furthermore, similarities and differences between infrastructure factors causing crowdsourced nearcrashes and conventional crashes (i.e., crashes reported to police) are investigated. Consequently, compared to existing research, the present study identifies crash-prone hotspots more robustly and reliably and finds that infrastructure types may influence the risk of near-crashes and crashes differently.

3. Data

This section introduces data sources and data processing methods used in the study. Sections 3.1 and 3.2 deal with the near-crash and crash data, respectively, whereas Section 3.3 is concerned with the associated exposure of both data types. Section 3.4 describes the infrastructure variables used for modeling (near-)crash rates.

3.1. Hövding data

The Hövding airbag2 is a head protection device for cyclists, worn around the neck as a collar. An accelerometer in the device tracks the movement and acceleration of cyclists and continuously determines the current safety state. Those safety states include a reference state where no abnormal movement is detected (i.e., mode 03 ) and three elevated safety states (modes 1, 2, and 3) representing different levels of abnormal movement. Modes 1 and 2 describe increasing levels of abnormal movement under the threshold for deployment whereas Mode 3 is where the movement is so abnormal that the device deploys (i.e., an actual crash occurs). The safety states are meticulously distinguished by accelerometers with a support vector machine (SVM) algorithm trained on data from 2000 h of normal cycling and around 3000 bicycle crashes carried out by stunt performers (Hövding, 2023). The cyclist's safety state is recorded 200 times per second, and when connected to an app on the user's phone (which is not a requirement for using the airbag), the cyclist's GPS location is also recorded every two seconds, allowing for determining the locations of elevated safety states. Overall, our data include 365,813 trips conducted by 10,049 Hövding users within Copenhagen Metropolitan area from September 16, 2019, to May 31, 2021. The data is protected under the General Data Protection Regulation (GDPR), and the privacy of Hövding users is fully secured. The gender and age distribution of Hövding users has a similar distribution of the general population in Copenhagen (Fosgerau et al., 2023).

Table 3 shows the frequency of each safety state after the data processing described in the next section. Mode 3 is when the helmet is deployed. Due to the low number of Mode 3 occurrences (Hövding data only includes 97 observations where the airbag was deployed), the present study only considers near-crashes in the Hövding data and uses external data from the police when considering actual crashes. Mode 1 is not categorized as a near-crash since it describes the lowest level of abnormal movement which happens so frequently that events could be due to uneven surfaces such as cobblestones and curbs (Lindqvist & Roos, 2020). We use the same classification for a near crash as in Lindqvist and Roos (2020) (originally classified as medium and high level of abnormal movement) and refer to this safety state as Mode 2 in this paper. The spatial distribution of Hövding near-crashes is shown in 2(a).

To extract valid potential crash and near-crash locations from Hövding safety states, we carried out the following four processing steps, which are depicted in Fig. 1:

  1. Select observations with elevated safety states (Modes 1 and 2) from the bicycle trajectories.

2 https://hovding.com

3 We have recoded the original value codings for a more intuitive representation in ascending order of severity. For a detailed classification of the original safety states please refer to Appendix A.

Fig. 1. Data processing example describing a hypothetical trip where the cyclists encounter both a near-crash (acute brake) and a crash (fall).

Table 3

| Descriptive analysis of principal safety states from Hövding data. | | | | | | | | | |--------------------------------------------------------------------|----------------|---------|-----------------------------|--|--|--|--|--| | Principal safety state | Classification | Count | Frequency
per
1000 km | | | | | | | Mode 3 | Crash | 97 | 0.08 | | | | | | | Mode 2 | Near-crash | 129,744 | 106.78 | | | | | | | Mode 1 | Non near-crash | 723,652 | 604.01 | | | | | |

Table 4 Buffer of point-based infrastructure variables.

Point-based Buffer radius
infrastructure
variables
Traffic light 18 m
Crossing 2.5 m
Bus stop 7.5 m
Roundabout 10 m
  • 2. Cluster elevated safety states from the same trip for consecutive elevated safety states within a five-second time interval. That interval is chosen as the clustering criterion, since over 90% of the safety states have additional elevated safety states occurring within five seconds. Therefore, clearly, each event is not an independent near-crash unit, and we use clusters as our units of analysis.
  • 3. For each cluster, assign the first occurrence of the maximum safety state within the cluster as the principal safety state; it represents the safety state and the exact location where the near-crash occurred.
  • 4. Filter out clusters with average speeds over 30 km/hr or with instantaneous speeds of the principal safety state over 50 km/h to avoid clusters including inaccurate GPS points. These thresholds were selected based on Lindqvist and Roos (2020). Additionally, clusters that occurred within the first five seconds of a trip are removed to avoid false-positive detections. With this filtering process, 914,784 clusters were reduced to 853,493 (i.e., a 6.7% reduction).

3.2. Police crash data

As the number of actual crashes in our Hövding dataset is low, we rely on external data for such events. Data describing crashes involving cyclists reported to the police (referred to as ''police crash data'') is obtained from the Danish Road Directorate (Vejman, 2022) with the same data collection period as the Hövding dataset (i.e., from September, 2019 to May, 2021). Fig. 2(b) shows the spatial distribution of police crash data, clearly indicating that the police crash data has a much lower coverage than the near-crash data shown in Fig. 2(a).

3.3. Exposure

To obtain how much bicycle traffic was traveled annually on each road segment by the relevant populations, the present study calculated exposure of Hövding and police crash data. Exposure of Hövding near-crashes were determined through map-matching the more than 360,000 trips in our Hövding dataset to their corresponding road segments in the OpenStreetMap (OSM) network (OpenStreetMap, 2022), and counting the occurrence of each segment. Since we only consider near-crashes of Hövding users in the near-crash data, it is natural to consider those users as our population. Thus, in contrast to common practice where exposure is based on modeled flows from transport models, the procedure provides the exact relevant exposure with minimal uncertainty. Multiplying the number of passing trips by the length of that segment yields the total distance traveled on that network segment.

For the police crash data, we rely on modeled traffic flows from the regional transport model COMPASS (Paag et al., 2019) that provides annual daily bicycle traffic on each road segment. Because of the reliance on modeled flows, exposure is restricted to the granularity of the network used in the COMPASS model, which is somewhat lower than in the richer OSM network used for near-crash exposure.

3.4. Infrastructure variables

To analyze how infrastructure features contribute to crashes and near-crashes, we use the same road type and surface variables as in Łukawska et al. (2023b), derived directly from the OSM network. All variables are listed in Appendix B. The binary bridge and tunnel variables are also network-based (i.e., link-based), and derived from the OSM network. Cycle superhighway is also a binary variable, but is based on an external source created for Hallberg et al. (2021).

The remaining variables (for traffic lights, crossing, bus stops, and roundabouts) cannot be directly assigned to the OSM network, as they are point-based features in the OSM point-of-interest layer. Instead, we use the following procedure based on buffers to account for the point-based attributes on each of the nearby links. This is done to identify the proportion of each network link covered by the respective buffers of point-based buffers. This results in a number between 0 and 1 for each point-based infrastructure type for each network link, which can then be used as a link-level attribute representing a point-based infrastructure type. The sizes of the buffers are based on the size of the infrastructure as shown in Table 4. An example of this is seen in Fig. 3. If 50% of a network link is covered by a bus stop buffer, the corresponding bus stop attribute for that link is 0.5.

Fig. 2. Spatial distribution of observed Hövding near-crashes and Police crashes.

Fig. 3. Explanation of point-based infrastructure calculation based on buffer. In this specific case, the buffer of the bus stop covers half the link, why the corresponding bus stop attribute of the link will be 0.5.

Since the exposure of police crash data was simulated in the COM-PASS network which contains less detailed information about infrastructure variables on each road segment, infrastructure variables of the OSM network were assigned and transferred to the COMPASS network. This is possible because the COMPASS network is a subset of the full OSM network.

4. Crash-rate model for near-crashes and crashes

This section describes the near-crash and crash rate models used to analyze how the infrastructure variables affect the occurrence rate of near-crashes and crashes, respectively. Separate models are estimated for near-crashes and crashes, but we will use the same type of model in both instances.

We find the negative binomial regression model to be a suitable model for our study for several reasons. Firstly, our (near-)crashes showed overdispersion issues when estimated through Poisson models,4 why a model capable of handling overdispersion is needed. Both the Negative Binomial model and the quasi-Poisson model can deal with

4 We elaborate on this issue later in this section.

overdispersion in a flexible manner. We deem both models to be suitable for our problem, but decide to use the Negative Binomial model since it is the most frequently used model in the crash modeling literature according to Lord and Mannering (2010) and Mannering and Bhat (2014). The probability distribution of the negative binomial distribution is shown in Eq. (1):

$$P(y_i) = \frac{\Gamma(\frac{1}{\alpha} + y_i!)}{\Gamma(\frac{1}{\alpha})y_i!} \left(\frac{\frac{1}{\alpha}}{\frac{1}{\alpha} + \mu_i}\right)^{\frac{1}{\alpha}} \left(\frac{\mu_i}{\frac{1}{\alpha} + \mu_i}\right)^{y_i}, \tag{1}$$

where is the dependent variable denoting the annual frequency of (near-)crashes at road segment , ( ) is the probability of (near- )crashes occurring at road segment annually, is the gamma function, is the annual expected number of (near-)crashes, and is the over-dispersion parameter.

The inclusion of serves to relax the mean–variance equality assumption that applies to the Poisson distribution. As opposed to the Poisson model where ( ) = , the variance of the negative binomial model is given by

$$Var(y_i) = \mu_i + \alpha \mu_i^2. \tag{2}$$

The over-dispersion parameter allows the variance to exceed the mean. If the over-dispersion parameter is close to zero, then the mean is equal to the variance, and there would be no issues in using Poisson regression. On the contrary, if is significantly larger than 0, the Poisson model would suffer from over-dispersion, and an alternative model type with more flexible variance (such as the negative binomial model) would have to be used. We calculated the ratio of the residual deviance to the degrees of freedom. This ratio exceeded 1 (1.35), indicating over-dispersion. In the present study, we find of the nearcrash model and crash model to be 2.158 and 8.134, respectively, and highly significantly larger than zero with a -value of 24.1 and 2.37, respectively, clearly evidencing that a Poisson regression model would not be suitable in this case.

Therefore, we use the negative binomial regression model, assuming a log-linear relationship between the expected number of (near-)crashes and explanatory variables, and use the specification in Eq. (3) for estimating (near-)crash rates:

$$\log(\mu_i) = \beta_0 + \beta_1 \cdot X_i + \log(ABT_i \times I_i). \tag{3}$$

Here , denotes column vectors containing explanatory variables of infrastructure factors, denotes the intercept, and denotes row vectors containing including coefficients of explanatory variables corresponding to each infrastructure attribute, and and denote the annual bicycle traffic and road length, respectively, at road segment . The product of the two is treated as exposure. It is important to clarify that while this study does model the near-crash and crash rates at the level of individual road segments, it does so only through (exposure and) infrastructural variables that are common across many links. We do not intend to directly compare the predicted number of nearcrashes and crashes on individual links. Instead, we assess the broader patterns from both models to gain insight into the overall similarities and differences in the infrastructural characteristics of near-crashes and crashes.

Finally, the logarithm of the exposures and is offset to the left-hand side with , as shown in Eq. (4):

$$\log(Rate_i) = \log\left(\frac{\mu_i}{ABT_i \times I_i}\right) = \beta_0 + \beta_1 \cdot X_i, \tag{4}$$

where denotes the (near-)crash rates at road segment . In this study, the unit for (near-)crash rates is frequency 100,000 bicyclekilometers traveled on road segment .

Prior to the model estimation, a Pearson correlation test was carried out to avoid high multicollinearity among explanatory variables. There were no signs of serious multicollinearity, i.e., no correlation coefficients over | ± 0*.*5|.

After the models were estimated, the incidence rate ratios (IRRs) of each explanatory variable were calculated by taking the exponential of estimated coefficients (i.e., ). IRRs are intuitive indicators showing the magnitude to which (near-)crash rates will be influenced by adjusting certain explanatory variables, when compared to their corresponding base-case variables.

5. Results

We now turn to investigating the types of infrastructure features that affect near-crash and crash rates. First, in Section 5.1 we provide descriptive statistics of the (near-)crash occurrence on the various infrastructure types. In Section 5.2 we estimate negative binomial nearcrash and crash-rate models and interpret the results. All analysis are based on a total of 129,744 near-crashes and 4220 crashes described in Section 3.

5.1. Descriptive analysis

Table 5 shows the proportion of each infrastructure variable for crashes and near-crashes and exposure in each category weighted by the length of the road segment. Several tendencies can be observed from Table 5 and are elaborated in the following paragraphs.

In general, road types with protected bicycle tracks have a lower occurrence of both near-crashes and police crashes compared to the corresponding exposure.

No bicycle infrastructure on medium roads has a disproportionately high number of police crashes compared to the amount of COMPASS exposure on those roads while Hövding near-crashes on these roads are proportionate to the level of Hövding exposure. On the other hand, No bicycle infrastructure on residential roads, cycleways, footways, and shared paths all exhibit a higher rate of Hövding near-crashes compared to the amount of Hövding exposure while they have a lower rate of police crashes compared to the amount of COMPASS exposure.

Regarding surface types, asphalt accounts for a lower percentage of Hövding near-crashes compared to their exposure, while cobblestones and gravel account for a higher percentage. All surface variables have the opposite trend for police crashes compared to near-crashes.

Regarding other infrastructure variables, Cycle superhighway has a lower share of Hövding near-crashes and police crashes compared to their corresponding exposures. This indicates that the focus on better mobility for cyclists on these routes also has an improved effect on cyclist safety.

5.2. Estimation results

Table 6 shows the estimation results of the negative binomial regression model presented with IRRs of the near-crash model from Hövding using the OSM network, and the crash model from police data using the COMPASS network. Besides, we also estimated models shown in Appendix C where we have removed all insignificant variables. For these reduced models, only very minor changes in IRRs of significant variables and models' goodness-of-fit were observed, indicating that the results are stable.

5.2.1. Hövding near-crash model

For the Hövding near-crash model, all variables affect near-crashes at a 5% significance level, except painted bicycle lanes on residential roads, bridges, and bus stops. For road type variables, only footway, living Street, shared path, pedestrian way, and steps have IRRs greater than one, indicating that these roads tend to have higher near-crash rates than No bicycle infrastructure on residential road (reference case), holding other variables unchanged. The two lowest IRRs, at 0.33 and 0.31, are obtained from protected bicycle tracks on medium roads and protected bicycle tracks on large roads, respectively, which may be due to cyclists

Table 5 Percentages of infrastructure variables for Hövding near-crashes, police crashes, exposures, and networks for each variable category.

Category Hövding Hövding OSM Police COMPASS COMPASS
Road type Near-Crash Exposure Network Crash Exposure Network
Protected bicycle track on large road 9.44 15.85 2.76 9.08 12.59 3.73
Protected bicycle track on medium road 18.34 34.98 7.23 26.34 28.65 10.38
Protected bicycle track on residential road 1.39 2.24 0.46 1.55 2.60 0.59
Painted bicycle lane on large road 0.64 0.91 0.16 0.43 0.86 0.24
Painted bicycle lane on medium road 2.25 3.08 2.08 4.96 2.86 3.30
Painted bicycle lane on residential road 0.66 0.51 0.18 0.36 0.43 0.22
No bicycle infrastructure on large road 1.44 1.64 1.64 1.66 1.70 0.85
No bicycle infrastructure on medium road 4.33 4.04 14.26 19.26 6.71 27.40
No bicycle infrastructure on residential road 27.34 13.68 25.06 15.74 16.14 21.99
Cycleway 15.58 12.57 15.51 16.18 17.27 19.75
Footway 5.11 3.38 8.26 0.57 2.17 1.76
Living street 0.91 0.35 0.56 0.19 0.39 0.30
Shared path 11.45 6.19 21.27 3.43 7.20 9.26
Pedestrian way 1.10 0.53 0.43 0.25 0.37 0.18
Step 0.03 0.04 0.15 0.01 0.07 0.07
Surface
Asphalt 86.27 91.39 83.81 93.42 91.2 88.43
Cobblestones 9.27 5.34 4.02 2.98 4.5 1.98
Gravel 4.12 3.01 11.5 3.42 4.03 9.17
Other 0.34 0.26 0.66 0.18 0.28 0.41
Point-based infrastructurea
Traffic light 0.06 0.10 0.02 0.06 0.08 0.02
Crossing 0.01 0.02 0.00 0.01 0.01 0.00
Bus stop 0.01 0.02 0.00 0.01 0.01 0.00
Roundabout 0.00 0.00 0.00 0.01 0.00 0.01
Other infrastructurea
Bridge 1.60 2.12 0.51 0.66 1.92 0.53
Tunnel 0.20 0.21 0.33 0.05 0.31 0.28
Cycle superhighway 14.77 20.83 5.39 11.54 17.65 4.65

a The share of each variable is calculated internally (e.g., the share of roads with and without traffic light sums to 100%).

having fewer interactions with motorized traffic at these particular road types.

For surface types, cobblestones has the highest IRR of 1.96, meaning that the near-crash rate for a cyclist riding on cobblestones is twice as high as for a cyclist riding on asphalt. For point-based infrastructure types, traffic lights, crossings, and roundabouts have IRRs around 1.8 to 2.4. For other infrastructure types, Cycle superhighways is the only variable with an IRR lower than one.

5.2.2. Police crash model

In general, road types with protected bicycle tracks or painted bicycle lanes have IRRs below 1, although the parameter for Painted bicycle lane on large road is strongly insignificant, and the parameters for Protected bicycle track on medium road and Protected bicycle track on residential road are only nearly significant (p-values at 0.08 and 0.11 respectively). An exception is on medium roads where crash rates are increasing both in the presence of painted bicycle lanes and when there is no bicycle infrastructure. Variables associated with different kinds of paths, including cycleways, pedestrian ways, and shared paths, have IRRs lower than one. For surface type, cobblestones has an IRR lower than one.

For point-based infrastructure types, traffic lights and roundabouts have IRRs of 15 and 23, respectively, indicating that signalized intersections and roundabouts generally have higher police crash rates. By contrast, crossing has an IRR of only 0.1, indicating Police crashes are far less likely to occur at crossings.

5.2.3. Comparison of estimation results

Below, we compare estimation results between the Hövding nearcrash model and the police crash model to find similar and different trends in the explanatory variables. For road-type variables, protected bicycle tracks on large roads and cycleways are significantly different from zero in both the Hövding near-crash and police crash models and they all have IRRs lower than one. This indicates that protected bicycle facilities tend to have a lower risk of both near-crashes and crashes. protected bicycle facilities and cycleways separate cyclists from motorized traffic, which is why it is reasonable that this greatly reduces the risk of collisions. It is also worth noting that protected bicycle tracks reduce the risk of near-crashes (IRRs around 0.3 to 0.6) by a factor of four to seven more than the risk of crashes (IRRs around 0.9)

By contrast, no bicycle infrastructure on medium roads, footways, and shared paths are significant in both models but have opposite IRR tendencies. Pedestrian-oriented ways and paths have a higher risk of near-crashes but a lower risk of crashes. This could be because cyclists riding on such pedestrian-oriented infrastructure have a high risk of interaction with pedestrians, resulting in evasive movements such as strong head or lateral movements. Such interactions rarely result in crashes that need to be reported to the police since the speeds of both cyclists and pedestrians are low.

For surface variables, cobblestones are significant in both models but have opposite IRR trends, implying that cobblestones have higher nearcrash rates than asphalt roads, while crash rates are lower. Crashes on cobble roads are rare due to low levels of motorized traffic and low speeds, while the increased near-crash rates might be related to uneven surfaces that can lead to near-crashes or crashes that are not severe

Table 6 Estimated incidence rate ratios (IRRs) from the Hövding near-crash and police crash models.

| | | Hövding Near-Crash | | | | Police Crash | | | | | |-----------------------------------------------|-------|--------------------------|---------|--------|--------|--------------------------|---------|--------|--|--| | | IRR | SE | Z-val | p-val | IRR | SE | Z-val | p-val | | | | Road type | | | | | | | | | | | | No bicycle infrastructure on residential road | | IRR = 1 (Reference case) | | | | IRR = 1 (Reference case) | | | | | | Protected bicycle track on large road | 0.313 | 0.009 | −39.000 | <0.001 | 0.630 | 0.054 | −5.397 | <0.001 | | | | Protected bicycle track on medium road | 0.334 | 0.007 | −50.998 | <0.001 | 0.901 | 0.060 | −1.579 | 0.114 | | | | Protected bicycle track on residential road | 0.436 | 0.027 | −13.486 | <0.001 | 0.724 | 0.133 | −1.761 | 0.078 | | | | Painted bicycle lane on large road | 0.436 | 0.044 | −8.216 | <0.001 | 0.834 | 0.175 | −0.864 | 0.387 | | | | Painted bicycle lane on medium road | 0.389 | 0.018 | −20.389 | <0.001 | 1.267 | 0.145 | 2.065 | 0.039 | | | | Painted bicycle lane on residential road | 0.844 | 0.084 | −1.707 | 0.088 | 0.814 | 0.214 | −0.781 | 0.435 | | | | No bicycle infrastructure on large road | 0.413 | 0.022 | −16.980 | <0.001 | 0.673 | 0.123 | −2.168 | 0.030 | | | | No bicycle infrastructure on medium road | 0.521 | 0.015 | −23.176 | <0.001 | 1.566 | 0.121 | 5.792 | <0.001 | | | | Cycleway | 0.502 | 0.012 | −28.979 | <0.001 | 0.391 | 0.034 | −10.953 | <0.001 | | | | Footway | 1.184 | 0.031 | 6.443 | <0.001 | 0.353 | 0.079 | −4.648 | <0.001 | | | | Living street | 1.319 | 0.101 | 3.603 | <0.001 | 0.568 | 0.233 | −1.378 | 0.168 | | | | Shared path | 1.312 | 0.028 | 12.663 | <0.001 | 0.377 | 0.053 | −6.889 | <0.001 | | | | Pedestrian way | 1.308 | 0.089 | 3.939 | <0.001 | 0.988 | 0.422 | −0.029 | 0.977 | | | | Step | 2.278 | 0.227 | 8.245 | <0.001 | 0.003 | 0.015 | −1.074 | 0.283 | | | | Surface | | | | | | | | | | | | Asphalt | | IRR = 1 (Reference case) | | | | IRR = 1 (Reference case) | | | | | | Cobblestones | 1.956 | 0.051 | 25.595 | <0.001 | 0.606 | 0.091 | −3.349 | 0.001 | | | | Gravel | 0.757 | 0.025 | −8.511 | <0.001 | 0.329 | 0.070 | −5.215 | <0.001 | | | | Other | 1.346 | 0.132 | 3.033 | 0.002 | 0.480 | 0.356 | −0.990 | 0.322 | | | | Point-based infrastructurea | | | | | | | | | | | | Traffic light | 1.787 | 0.057 | 18.346 | <0.001 | 14.932 | 1.211 | 33.324 | <0.001 | | | | Crossing | 2.125 | 0.224 | 7.153 | <0.001 | 0.099 | 0.056 | −4.095 | <0.001 | | | | Bus stop | 1.231 | 0.166 | 1.541 | 0.123 | 0.907 | 0.485 | −0.182 | 0.856 | | | | Roundabout | 2.415 | 0.224 | 9.503 | <0.001 | 23.833 | 2.895 | 26.104 | <0.001 | | | | Other infrastructurea | | | | | | | | | | | | Bridge | 1.114 | 0.073 | 1.645 | 0.100 | 0.151 | 0.058 | −4.933 | <0.001 | | | | Tunnel | 1.790 | 0.143 | 7.302 | <0.001 | 0.403 | 0.238 | −1.540 | 0.124 | | | | Cycle superhighway | 0.794 | 0.019 | −9.871 | <0.001 | 0.937 | 0.066 | −0.928 | 0.353 | | |

a For point-based infrastructure and other infrastructure, the reference case is corresponding road segments that are not labeled as these types of infrastructures (i.e., where the value is 0).

enough to be reported to the police. Gravel roads are also significant but have IRRs lower than one in both models, implying near-crash and crash rates are lower on gravel roads than on asphalt roads. This could be due to the average speeds being lower on gravel than on asphalt while the surface is smooth enough to provide a stable ride.

For point-based infrastructure variables, traffic lights and roundabouts are significant with IRRs greater than 1 in both models, implying most signalized intersections and roundabouts contribute to higher near-crash and crash rates than road segments without those infrastructure features. It is also worth noting that traffic lights and roundabouts increase the risk of crashes (IRRs around 15 and 24, respectively) by a factor of almost ten times more than the risk of near-crashes (IRRs around 1.8 and 2.4, respectively).

6. Discussion

We used large-scale crowdsourced near-crash data and conventional crash data to explore and compare how different infrastructure factors contribute to the occurrence rates of bicycle near-crashes and crashes. Contrary to most existing literature on the subject (Branion-Calles et al., 2017; Lindqvist & Roos, 2020; Matsui et al., 2016; Poulos et al., 2017), we have found multiple differences in terms of how infrastructure types contribute to near-crashes and crashes.

One such difference is that near-crashes tend to occur frequently on pedestrian-oriented roads such as footways, living streets, shared paths, and pedestrian ways, while crashes registered with the police do not. This difference has not previously been identified in the literature, although some indications exist. Pedestrian ways, footways and shared paths usually have many interactions with pedestrians and, according to Dozza and Werneke (2014), Mesimäki and Luoma (2021), Poulos et al. (2017), such infrastructure types are characterized by unpredictable maneuvers of pedestrians who may abruptly walk into cyclists' trajectories and by cyclists' speeding or lack of attention. Previous research has also found through surveys that such infrastructure types tend to have frequent collisions and near-crash situations (Mesimäki & Luoma, 2021). On the other hand, interactions with motor vehicles will likely be few, and most cyclists-pedestrian crashes are less severe and thus less likely to be reported to the police (Medury et al., 2019). To ensure the safety and mobility of both cyclists and pedestrians on such types of infrastructure, feasible solutions could be designing proper width and equipping dedicated space to cyclists and pedestrians according to cycling and pedestrian volume (Gkekas et al., 2020; Mesimäki & Luoma, 2021; Skoczyński, 2021).

Another identified difference is that cobblestones have relatively higher near-crash rates but lower crash rates. The results for cobblestones are supported by Lindqvist and Roos (2020), and this might be due to the slippery or uneven surface of the cobblestones. Moreover, cyclists typically travel at slower speeds on this type of surface compared to asphalt roads, leading to fewer severe crashes but increasing the likelihood of instability, resulting in more near-crashes. However, we acknowledge that it could also be due to the false detection of near-crashes caused by vibration on this type of infrastructure.

In line with existing literature on near-crashes (Mesimäki & Luoma, 2021) and crashes (DiGioia et al., 2017; Kaplan & Prato, 2015; Morrison et al., 2019; Wall et al., 2016), we found that both near-crashes and crashes are in general less likely to occur on roads with bicycle infrastructure such as protected bicycle tracks and bicycle lanes than on roads without bicycle infrastructure. This supports that creating dedicated infrastructure for cyclists likely provides safer conditions due to fewer interactions with other types of road users.

Furthermore, we found that both near-crashes and crashes are more likely to occur at signalized intersections and roundabouts. This result is in line with both near-crash research (Aldred, 2016; Dozza & Werneke, 2014; Strauss et al., 2017) and crash studies (Kaplan & Prato, 2015; Poudel & Singleton, 2021; Reynolds et al., 2009; Schröter et al., 2023). This indicates that these locations should receive significant attention.

7. Limitations and future research

The sample of users of the relatively expensive Hövding equipment may be biased towards high-income and safety-aware frequent cyclists (Łukawska et al., 2024). However, the gender and age composition of Hövding users closely resembles that of the wider population in Copenhagen (Fosgerau et al., 2023), and we find it to be the most suitable dataset to provide valuable insights about near-crashes. Police data suffers from under-reporting, as the crashes recorded by the police are mainly limited to bicycle crashes involving other road users, particularly motorized vehicles (Hosseinpour et al., 2021; Ringel et al., 2023; Springer et al., 2020). Other types of crashes, such as single bicycle crashes where it is not necessary to dispatch police and severe crashes that are only recorded by hospitals are underrepresented in the crash rate model.

Another limitation in this study relates to the Hövding near-crashes which are defined as abnormal movements detected by the sensors of the helmet. The classification of safety states made by accelerometers does not reveal anything about the types of near-crashes. Therefore, we cannot guarantee that all of them are actual near-crashes. Some might be false detections that do not occur under near-crash circumstances. On the other hand, some actual near-crashes might go unnoticed in the data leading to false-negative issues. As crash causation mechanisms differ across different types of (potential) crash situations (e.g. Ito et al., 2018) it is likely that our near-crash data has a bias towards certain types of near-crashes, namely those that cause an abnormal movement of the cyclist. Other types of near-crashes such as rearend near-collisions by vehicles and situations that primarily cause emotional stress may be underrepresented or completely missing. Still, our data provide a large amount of valuable information that yields meaningful results when used in a near-crash-rate model. Future research could use naturalistic studies to gather additional information and validate the near-crash types found in crowd-sourced datasets triggered by abnormal movements like the Hövding dataset used in this study.

It is also relevant to mention that the GPS locations of near-crash occurrences might be subject to minor geographical misplacements due to GPS inaccuracies. This may have led to some near-crash locations being assigned to a nearby road segment rather than the true one.

However, the data smoothing, filtering, and map-matching are based on existing research described in Łukawska et al. (2023a) and we do not expect the overall model results to be significantly affected by this issue.

Another limitation of this study is the unavailability of accurate exposure measures for police data crashes, which instead had to rely on modeled flows from COMPASS (Paag et al., 2019). This adds uncertainty to the estimated crash rate model, and therefore also to the comparisons between the near-crash and crash rate models. Obtaining more accurate flows to better represent the true exposure of the police data crashes would lead to model estimates with lower uncertainty, similar to what we see for the near-crash rate model where the uncertainty of the exposure is completely eliminated by utilizing the direct exposure from the trajectories of the Hövding dataset.

Finally, this study focuses on exploring infrastructural factors and we have thus included detailed network-based infrastructure information. However, other factors such as behavioral factors (e.g., speed, riding style, interaction with other road users, etc.), spatiotemporal factors (e.g., temporary construction work), and weather factors could also influence the number of (near-)crashes. While this is true, infrastructure will also affect the aforementioned factors, for instance, regarding behavioral factors, a shared path might induce more interaction between cyclists and pedestrians and likely make cyclists reduce their speed, while fewer interactions and higher speeds are expected at a protected bicycle track. Future research could include adding such factors to the (near-)crash rate models.

8. Conclusion

Through investigating potential factors causing bicycle near-crashes of Hövding users and crashes registered in police crash data, this study identifies the near-crash-prone locations that cannot be identified solely by conventional crash data and also identifies potential underlying infrastructural causes of near-crashes in such locations. With a modeling approach based on large-scale data, we have analyzed how the risk of near-crashes and crashes is affected by different types of infrastructure in Metropolitan Copenhagen. For instance, although much research investigates factors contributing to bicycle crashes, few studies have explored significant factors at such a detailed network level, and none have compared factors leading to near-crashes and crashes within the same area and modeling framework. Our main contribution is modeling of bicycle near-crash rates based on crowdsourced data from bicycle accelerometers and GPS trajectories, which additionally allow us to model the influence of various detailed infrastructure types while taking the exact exposure into account in the near-crash rate model. Moreover, we estimate an equivalent networkbased crash rate model from police data to compare how infrastructure factors contribute to both near-crashes and crashes. By investigating potential factors causing bicycle near-crashes and crashes, this study identifies near-crash-prone infrastructure types that, are primarily associated with pedestrian-cyclist interactions, that cannot be identified solely by conventional crash data. In conclusion, we determined that bicycle-crowdsourced near-crash data is a valuable source, and its complementarity with conventional crash data that usually suffer from under-reporting issues (Janstrup et al., 2016; Møller et al., 2021) enables a richer perspective when pinpointing risky locations.

Having an isolated focus on severe or fatal crashes may lead to neglecting relevant understanding of near-crashes that may influence the perceived safety or the willingness to cycle. As further elaborated in the discussion, we found that certain infrastructure factors, such as pedestrian-oriented ways and paths, tend to induce more frequent nearcrashes but less frequent crashes. These findings can be used to guide the development of urban space to reduce the risk of near-crashes and foster safer cycling environments. Furthermore, this study raises the discussion on the possible reasons behind the frequent occurrence of near-crashes at these locations and also suggests the opportunity for further comprehensive research exploring various aspects beyond just infrastructural factors.

9. Practical applications

The practical implications of this study lie in the transformation of how we identify risky locations for cyclists. Traditional crash data, primarily obtained from sources like police or hospitals, is no longer the only source of information. Instead, by combining large-scale crowdsourced bicycle near-crash data with traditional crash data, we can gain a more nuanced and comprehensive understanding of risky hotspots and infrastructure types for cyclists. This approach can at a disaggregated level help identify previously undisclosed specific geographic locations for cyclists, allowing more effective precautions to be taken.

As an example, our results show that shared paths tend to have near-crash rates that are between two to three times higher than that of dedicated cycleways. Similar results are seen for other pedestrianoriented infrastructure types. Thus, a very specific recommendation based on our study is to encourage stronger separation between pedestrians and cyclists. This illustrates how the outcomes of this study can be used to guide city planners how to design safer infrastructure.

CRediT authorship contribution statement

Kuan-Yeh Chou: Writing – original draft, Visualization, Software, Methodology, Investigation, Formal analysis, Conceptualization. Mads Paulsen: Writing – review & editing, Supervision, Software, Investigation, Conceptualization. Anders Fjendbo Jensen: Writing – review & editing, Supervision, Investigation, Conceptualization. Thomas Kjær Rasmussen: Writing – review & editing, Supervision, Conceptualization. Otto Anker Nielsen: Writing – review & editing, Supervision, Conceptualization.

Declaration of competing interest

None.

Acknowledgments

The study was funded by an alliance PhD grant from the Technical University of Denmark and the Technical University of Munich.

Appendix A. List of Hövding safety states

Original
safety state		 Classification Safety state
in this study		 Classification
in this study
Mode 0 Bluetooth
inactive		 NA NA
Mode 1 Bluetooth
active		 Mode 0 Reference
state
Mode 2 Deployed Mode 3 Crash
Mode 4 SVM Low Mode 1 Non
near-crash
Mode 5 SVM Medium Mode 2 Near-crash
Mode 6 SVM High Mode 2 Near-crash
Mode 7 SVM High NA NA
Mode 8 SVM High NA NA
Mode 9 SVM High NA NA

Table B.7

Variables [Unit] OSM tag
Road type [Binary]
Protected bicycle tracksa
-Medium road
-Large road Medium road
-Residential road (medium) highway: 'primary', 'secondary', 'tertiary', 'unclassified'
Painted bicycle lanes
-Medium road		 & at most one car lane per direction
Large road
-Large road highway: 'primary', 'secondary', 'tertiary', 'unclassified'
-Residential road (medium) & at least two car lanes in at least one direction
No bicycle infrastructure Residential road
-Medium road highway: 'residential'
-Large road
-Residential road (medium)
-Cycleway highway: 'cycleway'
-Footway highway: 'footway'
-Living street highway: 'living_street'
-Shared path highway: 'path', 'track', 'service'
-Pedestrian way highway: 'pedestrian'
-Steps highway: 'steps'
Surface [Binary]
-Asphalt surface: 'asphalt', 'paved', 'concrete', 'concrete:plates','concrete:lanes'
-Cobble surface: 'cobblestone', 'unhewn_cobblestone', 'sett','stone', 'paving_stones', 'pebblestone', 'rock'
-Gravel surface: 'gravel', 'dirt', 'compacted', 'unpaved','fine_gravel', 'ground'
-Other surface: 'wood', 'metal', 'granite', 'sand', 'earth', 'grass','grass_paver', 'rubbercrumb', 'mud', 'woodchips', 'plastic'
Point-based infrastructure [Proportion]
-Traffic light highway: 'traffic_signals' (18m-radius buffer)
-Crossing highway: 'crossing' (2.5m-radius buffer)
-Bus stop highway: 'bus_stop' (7.5m-radius buffer)
-Roundabout highway: 'roundabout (10m-radius buffer)
-Mini roundabout highway: 'mini_roundabout' (10m-radius buffer)
Other infrastructure [Binary]
-Bridge bridge
-Tunnel tunnel
-Cycle Superhighway Created through digital maps from the Office forCycle Superhighways (SUPER CYKELSTIER, 2022)

a Bicycle infrastructure type is defined based on OSM tag cycleway. cycleway='track' for protected bicycle tracks, cycleway='lane' or cycleway='crossing' for painted bicycle lanes, and cycleway='no' or missing value for no bicycle infrastructure.

Table C.8

Hövding Near-Crash Police crash
IRR SE Z-val p-val IRR SE Z-val p-val
Road type
No bicycle infrastructure on residential road IRR = 1 (Reference case) IRR = 1 (Reference case)
Protected bicycle track on large road 0.315 0.009 −38.924 0.000 0.673 0.051 −5.191 0.000
Protected bicycle track on medium road 0.336 0.007 −51.100 0.000
Protected bicycle track on residential road 0.436 0.027 −13.456 0.000
Painted bicycle lane on large road 0.439 0.044 −8.165 0.000
Painted bicycle lane on medium road 0.390 0.018 −20.318 0.000 1.367 0.147 2.898 0.004
Painted bicycle lane on residential road 0.845 0.084 −1.696 0.090
No bicycle infrastructure on large road 0.415 0.022 −16.917 0.000 0.727 0.130 −1.788 0.074
No bicycle infrastructure on medium road 0.523 0.015 −23.081 0.000 1.684 0.117 7.498 0.000
Cycleway 0.504 0.012 −28.892 0.000 0.412 0.032 −11.424 0.000
Footway 1.186 0.031 6.533 0.000 0.369 0.081 −4.522 0.000
Living street 1.319 0.101 3.601 0.000
Shared path 1.313 0.028 12.707 0.000 0.394 0.054 −6.790 0.000
Pedestrian way 1.310 0.089 3.963 0.000
Step 2.286 0.228 8.284 0.000
Surface
Asphalt IRR = 1 (Reference case) IRR = 1 (Reference case)
Cobblestones 1.957 0.051 25.602 0.000 0.601 0.088 −3.487 0.001
Gravel 0.756 0.025 −8.550 0.000 0.331 0.071 −5.170 0.000
Other 1.378 0.134 3.298 0.001
Point-based infrastructurea
Traffic light 1.782 0.056 18.281 0.000 14.449 1.1411 33.812 0.000
Crossing 2.120 0.223 7.135 0.000 0.091 0.052 −4.239 0.000
Bus stop
Roundabout 2.408 0.223 9.475 0.000 23.378 2.753 26.763 0.000
Other infrastructurea
Bridge 0.147 0.056 −5.005 0.000
Tunnel 1.786 0.142 7.271 0.000
Cycle superhighway 0.795 0.019 −9.821 0.000
Log likelihoodb −104250 −13037
Akaike information criterion (AIC)b 208548 26105

a For point-based infrastructure and other infrastructure, the reference case is corresponding road segments that are not labeled as these types of infrastructures (i.e., where the value is 0).

Appendix B. Road infrastructure variables

See Table B.7.

Appendix C. Estimated incidence rate ratios (IRRs) from the Hövding near-crash and police crash models with only significant variables

See Table C.8.

References

Aldred, R. (2016). Cycling near misses: Their frequency, impact, and prevention. Transportation Research Part A: Policy and Practice, 90, 69–83. http://dx. doi.org/10.1016/j.tra.2016.04.016, URL https://www.sciencedirect.com/science/ article/pii/S0965856416303639.

Aldred, R., & Crosweller, S. (2015). Investigating the rates and impacts of near misses and related incidents among UK cyclists. Journal of Transport & Health, 2(3), 379– 393. http://dx.doi.org/10.1016/j.jth.2015.05.006, URL https://www.sciencedirect. com/science/article/pii/S2214140515002236.

Aldred, R., & Goodman, A. (2018). Predictors of the frequency and subjective experience of cycling near misses: Findings from the first two years of the UK near miss project. Accident Analysis and Prevention, 110, 161–170. http://dx. doi.org/10.1016/j.aap.2017.09.015, URL https://www.sciencedirect.com/science/ article/pii/S0001457517303366.

Bagloee, S. A., Sarvi, M., & Wallace, M. (2016). Bicycle lane priority: Promoting bicycle as a green mode even in congested urban area. Transportation Research Part A: Policy and Practice, 87, 102–121. http://dx.doi.org/10.1016/j.tra.2016.03.004, URL https://www.sciencedirect.com/science/article/pii/S0965856416000495.

Branion-Calles, M., Nelson, T., & Winters, M. (2017). Comparing crowdsourced near-miss and collision cycling data and official bike safety reporting. Transportation Research Record: Journal of the Transportation Research Board, 2662(1), 1–11. http://dx.doi.org/10.3141/2662-01, URL http://journals.sagepub.com/doi/ 10.3141/2662-01.

Buehler, R., & Pucher, J. (2022). Cycling through the COVID-19 pandemic to a more sustainable transport future: Evidence from case studies of 14 large bicycle-friendly cities in Europe and north america. Sustainability, 14(12), 7293. http://dx.doi.org/ 10.3390/su14127293, URL https://www.mdpi.com/2071-1050/14/12/7293.

Cole-Hunter, T., Donaire-Gonzalez, D., Curto, A., Ambros, A., Valentin, A., Garcia-Aymerich, J., Martínez, D., Braun, L. M., Mendez, M., Jerrett, M., Rodriguez, D., de Nazelle, A., & Nieuwenhuijsen, M. (2015). Objective correlates and determinants of bicycle commuting propensity in an urban environment. Transportation Research Part D: Transport and Environment, 40, 132–143. http://dx.doi.org/10.

b The goodness-of-fits are not comparable across the two models as they are estimated on different datasets.

Thomas Kjær Rasmussen is Associate Professor and head of the Transportation Systems Modeling section in the Transportation Science Division at the Department of Technology, Management and Economics. His research interests include route choice modeling, traffic assignment, traveler behavior and systems design. His work is quantitative and spans across multiple modes of transport, with a recent focus on bicycle traffic and road (pricing).

Otto Anker Nielsen is Professor at the Technical University of Denmark and head of the Transportation Science Division at the Department of Technology, Management and Economics. His research interest includes route choice models and traffic assignment, transport demand modeling and transport behavior, Road pricing and GPS-based data, and integrated public transport optimization and planning.