Spatial patterns and temporal dynamics of urban bicycle crashes—A case study from Salzburg (Austria)

Spatial patterns and temporal dynamics of urban bicycle crashes—A case study from Salzburg (Austria)

Journal of Transport Geography 52 (2016) 38–50 Contents lists available at ScienceDirect Journal of Transport Geography journal homepage: www.elsevi...

4MB Sizes 0 Downloads 17 Views

Journal of Transport Geography 52 (2016) 38–50

Contents lists available at ScienceDirect

Journal of Transport Geography journal homepage: www.elsevier.com/locate/jtrangeo

Spatial patterns and temporal dynamics of urban bicycle crashes—A case study from Salzburg (Austria) Martin Loidl ⁎, Christoph Traun, Gudrun Wallentin Department of Geoinformatics, University of Salzburg, Hellbrunnerstraße 34, 5020 Salzburg, (Austria)

a r t i c l e

i n f o

Article history: Received 21 September 2015 Received in revised form 8 December 2015 Accepted 26 February 2016 Available online xxxx Keywords: Bicycle crashes Exploratory analysis Spatial and temporal dynamics

a b s t r a c t Most bicycle crash analyses are designed as explanatory studies. They aim to identify contributing risk factors and calculate risk rates based on – most of the time – highly aggregated statistical data. In contrast to such explanatory study designs, the presented study follows an exploratory approach, focusing on the absolute number of crashes. The aim is to reveal and describe patterns and dynamics of urban bicycle crashes on various spatial scale levels and temporal resolutions through a multi-stage workflow. Spatial units are delineated in the network space and serve as initial units of aggregation. In order to facilitate comparisons among regions and quantify temporal dynamics, a reference value of crash frequency is simulated for each unit of the respective spatial scale level and temporal resolution. For the presented case study, over 3000 geo-coded bicycle crashes in the city of Salzburg (Austria) were analyzed. The data set covers 10 years and comprises all bicycle crashes reported by the police. Distinct spatial and temporal patterns with clusters, seasonal variations, and regional particularities could be revealed. These insights are indicators for urban dynamics in the transport system and allow for further, targeted in-depth analyses and subsequent counter measures. Moreover, the results prove the applicability of the proposed multi-stage workflow and demonstrate the added value of analyses of small aggregates on various scale levels, down to single crashes, and temporal resolutions. © 2016 Elsevier Ltd. All rights reserved.

1. Introduction Cycling as an active mode of transport is healthy (Götschi et al., 2015; Holm et al., 2012) but dangerous relative to the distance traveled (Beck et al., 2007). Compared to driving a car, the risk for becoming involved in a crash is between 5.5 (ITF, 2012) and 12 (Delmelle and Thill, 2008) times higher for bicyclists—depending on the country and environment. Thus, it is not surprising that safety concerns are among the main reasons for rejecting the bicycle as a utilitarian mode of transport (Sanders, 2015; Winters et al., 2011). Generally, there is a broad consensus on the fact that road safety is one of the key factors for an increasing share of cyclists (Heinen et al., 2010; Thomas and DeRobertis, 2013; WHO, 2013). In order to face the safety issue, authorities, planners, and researchers have put much effort into building and improving bicycle infrastructure (Pucher et al., 2010; Rietveld and Daniel, 2004) and providing information about safe routes (Loidl and Zagel, 2014). It is thus of crucial importance to better understand the patterns of crashes in terms of their spatial (where?) and temporal (when?) occurrence on a city-scale level. This degree of detail allows for further, ⁎ Corresponding author. E-mail addresses: [email protected] (M. Loidl), [email protected] (C. Traun), [email protected] (G. Wallentin).

http://dx.doi.org/10.1016/j.jtrangeo.2016.02.008 0966-6923/© 2016 Elsevier Ltd. All rights reserved.

targeted in-depth analyses and subsequent countermeasures. Although the prevalent safety concerns are widely acknowledged, crashes are rarely investigated on multiple temporal intervals and spatial scale levels, ranging from the city-scale to single crash locations. Many studies on bicycle crashes follow an epidemiological approach. This means that aggregated crash data are related to aggregated statistics which serve as exposure variables, such as annual travel distance. This way, an average risk exposure can be calculated for spatial reference units of different scales. Vandenbulcke et al. (2009) and Yiannakoulias et al. (2012) provide examples for risk calculations on a regional level, while Beck et al. (2007) performed risk calculations on a national level. A complementary approach relies on systematic in-depth analysis of individual crash samples to draw general conclusions on causalities, e.g. concerning bicycle facilities, road design, environmental conditions, and sociodemographic variables (e.g. Chen and Fuller, 2014; Harris et al., 2013; Lovelace et al., 2015; Teschke et al., 2012). Both approaches allow for a better understanding of certain aspects of bicycle safety, especially of contributing risk factors. Nevertheless, at least two issues still remain: First, the calculation of the risk exposure is often based on weak (in terms of spatial and temporal resolution, scale and accuracy) data, as extensively noted in the latest OECD report on bicycle safety (OECD, 2013). Additionally, large aggregates don't account for the dynamics within the reference units. Second, although crashes are rarely random but tend to be spatially dependent within a certain

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

area (Anderson, 2009; Xie and Yan, 2013), the spatial and temporal context of the data is mostly not considered explicitly in crash data analysis (Vandenbulcke-Plasschaert, 2011). With the approach proposed in this paper, we reveal the patterns of crash occurrences on multiple spatial scale levels and temporal resolutions. For this, we make use of the spatial and temporal information attached to the crash reports and establish a multi-stage workflow for a seamless, interactive exploration of the crash data.

2. Spatial and temporal analysis of bicycle crashes: Background and objectives The body of scientific literature on bicycle safety has been growing constantly over the last two decades (Gerike and Parkin, 2015). Since the topic is tackled from various domain-specific and methodological perspectives, the intention here is to give a brief overview of a few major research avenues that are each further illuminated in the referenced literature. A large number of studies deals with the relation of bicycle crashes and the physical environment (De Rome et al., 2013; Dozza and Werneke, 2014; Harris et al., 2013; Lusk et al., 2011; Reynolds et al., 2009; Teschke et al., 2012; Thomas and DeRobertis, 2013). Teschke et al. (2012) identified physical route characteristics that contribute to the risk of being involved in a crash. Lusk et al. (2011) compared the risk of on-road cycling to cycle tracks. While the focus of many risk estimations lies on the road segment, Harris et al. (2013) investigated risk factors at intersections. Both studies suggest that the separation of cyclists from motor vehicles and a low traffic speed increase the overall safety for bicyclists. Studies focusing on the spatial distribution of risk highlight the spatial variability of crash occurrences in terms of geography, type of environment or road class. Obvious spatial variabilities can be detected on a national scale, with the “bikeability” of the environment and bicycle safety politics as major explanatory variables (Vandenbulcke et al., 2009). De Geus et al. (2012) found that the majority of crashes occurred in built-up areas, when cyclists were driving on the main lane. De Rome et al. (2013) revealed a relatively high risk for being involved in a bicycle crash cycling on road and on shared paths for a study site in Australia. Although there is evidence for gender-specific cycling behaviors (Beecham and Wood, 2013; Garrard et al., 2008), the gender-specific risk tends to be insignificant according to several studies (e.g. De Geus et al., 2012; De Rome et al., 2013). Nevertheless, Martínez-Ruiz et al. (2015) found higher death rates for male cyclists. Lovelace et al. (2015) were not able to estimate the gender-specific risk but point to the fact, that the ratio of crashes involving female and male bicyclists was 1:8. Concerning the age, Kröyer (2015) found that the fatality risk increases with the age of the bicyclist who is involved in a bicycle–motorized vehicle crash. This is in line with findings by Degraeuwe et al. (2015). Less research has been done so far on the temporal variability of bicycle crashes and the influence of seasonal dynamics, mainly associated with weather and road conditions. Doherty and Aultmann-Hall (2000) investigated the temporal variability of bicycle crash occurrences (not risk) and found significant differences between seasons but also between different urban environments. Aldred and Crosweller (2015) reconstructed incidents (crashes and near-misses) from travel diaries and correlated these events with the total travel time, binned to hours of a day. Their result shows distinct temporal patterns over the course of a day with peaks in the morning and evening rush hour. These findings are in line with Lovelace et al. (2015). Wanvik (2009) and Chen and Fuller (2014) linked road and light conditions with crash risk and found evidence for a larger prevalence of bicycle crashes in the darkness and in case of wet and icy road surfaces. From a methodological perspective, the general study design, the level of aggregation and scale, the definition of adequate spatial references and existing approaches in point pattern analysis, both in space and time, are relevant.

39

Independent from the variable that is tested for, explanatory study designs (Lord and Mannering, 2010) require an exposure variable (typically distance traveled or number of trips). The type of exposure variable and the quality of the data directly impact the explanatory power of the results (OECD, 2013; Schepers, 2013). For bicycle-related investigations of risk patterns, sound exposure variables are hardly ever available, especially on a micro-scale. Unlike to explanatory approaches, exploratory analyses focus on the absolute number of incidents (crash frequencies instead of rates) and their respective spatial and/or temporal distribution as well as on their characteristics (Dai, 2012). Exploratory approaches do not necessarily require any exposure variables. In turn, they don't allow for definite conclusions on underlying processes. Instead, they serve as hypothesis generator and initial point for further explanatory investigations. As explanatory analyses are very data intensive, a preceding exploratory analysis that identifies potential causalities, associated with specific locations and time intervals, contributes to cut down the amount of required data considerably. This aspect is even more important when bicycle crash locations are investigated on a single-crash level where adequate exposure variables are hardly ever available for the particular location and point of time. As mentioned in the introduction, many bicycle crash analyses follow an epidemiological approach where crash occurrences are related to spatially aggregated, statistical variables. Typically, the prevalence of crashes or the estimated risk are investigated on the level of administrative or statistical units, such as countries, states, municipalities, census districts, or regular grids (Anderson, 2009; Beck et al., 2007; Delmelle and Thill, 2008; Lovelace et al., 2015; Vandenbulcke et al., 2014). For spatial dynamics of bicycle crashes or risk exposure, especially on a micro-scale, one has to consider the potentially high variation within reference units (census districts tend to be too large) and furthermore account for the network-bound character of bicycle crashes. Thus, the Euclidean distance is inappropriate for the delineation of reference units and needs to be substituted by network-based alternatives (Okabe and Sugihara, 2012). As alternative to planar reference units for the analysis of bicycle crashes, several authors exclusively focus on junctions or segments respectively (Vandenbulcke-Plasschaert, 2011), while others slice the road network into equally long segments and use them as reference units. For the latter method, no matter which algorithm is applied, it has to be noted that fragments in peripheral areas always remain when the network is decomposed (Shiode, 2008). Roughly generalized, the delineation of reference units in networks can be approached from two sides. First, spatial, temporal, or spatio-temporal cluster detection algorithms can be applied in order to define regions that are similar in terms of crash occurrence or risk exposure (Eckley and Curtin, 2013; Shiode and Shiode, 2013; Steenberghen et al., 2010). Results of such data-driven approaches are hard to compare over several time intervals, due the variability of the spatial configuration. Second, spatial reference units are predefined, either using existing administrative units or applying delineation methods such as the quadrat method (Shiode, 2008) or Voronoi diagrams (Okabe and Sugihara, 2012). For the multi-stage workflow described in this paper, the latter approach was applied. This makes it easier to compare regions over time and reveal temporal dynamics. On the other hand, the MAUP (Wong, 2009) potentially remains an issue which needs to be taken into account when the results are interpreted. As soon as bicycle crashes are investigated on a single crash level, a large variety of well-established methods and tools from many domains can be applied (Cressie, 1993; Diggle, 2013). In the context of crash analysis, the detection and characterization of black-spots or blackzones (clusters), the determination of significance, the degree of spatial dependency, and the variability over time are among the most eminent concerns. However, it needs to be considered that the majority of available algorithms were originally designed for the planar space. Thus, an adaption for the network space is required (Okabe and Sugihara, 2012). This leads to more reliable results, for example, in

40

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

cluster analysis (Steenberghen et al., 2010) or density estimation (Borruso, 2008; Okabe et al., 2009). Besides, there is evidence from crime analysis that the focus on the micro-scale in networks reveals hotspots that might not be detectable on the macro-scale or in planar space (Braga et al., 2010). Although there are plenty of examples for network-based analysis of point patterns, the body of literature on spatio-temporal dynamics of point incidents in a network, especially in the context of crash analysis, is thin (Eckley and Curtin, 2013). Against this background, we follow an exploratory approach and develop a multi-stage workflow that allows for a seamless investigation of bicycle crashes on multiple spatial scale levels and temporal resolutions. Throughout the paper, we refer to “crashes,” indicating that we do not consider any qualitative judgment on the incident (liability, chance to prevent the crash, involvement of other parties etc.). The main objective of the presented study is to reveal spatial and temporal dynamics of bicycle crashes. Furthermore the following objectives will be dealt with: • develop an exploration environment for revealing and characterizing the spatial and temporal dynamics of crash occurrences in the absence of exposure variables; • delineate appropriate, network-based spatial units that serve as reference for the temporal variations; • evaluate the relation between pattern detection on the one hand and spatial scale and temporal resolution on the other hand; • reveal particularities in a case study and reflect the findings on the basis of existing literature.

To our current knowledge, there are no comparable studies which investigate bicycle crashes on multiple spatial scales and temporal resolutions, following a purely exploratory approach. 3. Methods: A multi-stage workflow For the spatial and temporal analysis of bicycle crash data, we integrated spatial analysis, descriptive statistics, and visual data exploration into a multi-stage workflow. The general idea of the proposed workflow is the following (details are provided in the following sections): comparable network-based units are delineated in order to serve as fixed spatial reference for further analyses. For each of these reference units, the number of crashes is simulated, based on the assumption of randomly distributed crashes per road category. Using the simulated mean and standard deviation as baseline, spatial and temporal variabilities of reported bicycle crash occurrences are quantified by a z-score. Finally, revealed patterns and dynamics are further investigated in an exploratory, visual environment, including descriptive statistics.

In this section, a conceptual description of the workflow is provided before it is employed in a case study. Here, the focus is put on the delineation of the network-based, spatial reference units (Fig. 1, stage 2), the calculation of z-scores which serve as proxy for spatial and temporal dynamics (Fig. 1, stages 2 and 3), and the analysis environment (Fig. 1, stage 3). The first three stages of the workflow in Fig. 1 are implemented in a GIS (Geographical Information System) environment. Bicycle crashes need to be geographically referenced (geo-coded) and a digital representation (graph) of the street network is required. In a first preparatory stage, the geo-coded crash reports are related to the graph by a simple geometric map-matching algorithm (Quddus et al., 2007). Next, the graph is decomposed into network-based, spatial reference units. 3.1. Definition of spatial reference units In Section 2 it became evident, that a disaggregated bicycle crash analysis on a city-scale level requires a network-based approach. As the focus of this study is not necessarily on spatial clusters at one specific point of time, but on the spatial and temporal dynamics over a longer period, we define spatially fixed, network-based reference units which serve as initial level of aggregation. The construction of the spatial reference units, which equal Voronoi diagrams as described by Okabe and Satoh (2009), works as follows (ref. Fig. 2): • Overlay the network with a regular grid. In the case study, presented in the next section, the entire study area was overlain with a 5 × 5 cells grid (equals a cell size of 2.3 × 2.7 km). • Calculate the network length ratio for each cell: Di/Dtotal, where Di is the network length per grid cell and Dtotal the network length of the whole road network. • Define the number of reference units. • Generate x seed points along the network for each cell, applying a buffered random sampling process (Kingsley et al., 2004), where x is the number of intended reference units multiplied with the ratio Di/Dtotal. In contrast to a completely random point generation, the applied restriction ensures randomness of the point locations while keeping a minimum distance buffer between each point. Since the points serve as seed points for the definition of comparable reference units, the generation of very small units is prevented. • Calculate network-based Voronoi diagrams (Okabe and Satoh, 2009) for the whole study area based on the seed points. The Voronoi diagrams define the spatial reference units in the network space for further analyses. With this approach, the network is split into relatively homogeneous units in terms of network density. Due to the bicycle's relatively small share in the modal split and immense underreporting (OECD, 2013;

Fig. 1. Multi-stage workflow for the spatio-temporal analysis of bicycle crashes.

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

41

Fig. 2. Definition of network-based, spatial reference units corresponding with network density.

Tin Tin et al., 2013; Wegman et al., 2012), (reported) bicycle crashes are comparatively rare phenomena (Nordback et al., 2014). Thus, the number of reference units is a trade-off between size (and connected to this the number of reported crashes) and homogeneity or between the statistical robustness of the results and their spatial and temporal resolution respectively: Larger reference units (or temporal bins) lead to larger sample sizes and a higher confidence about the results, but in turn might hide potentially relevant spatial or temporal patterns at a finer scale. 3.2. Simulation of random crash locations For the detection and analysis of spatial and temporal dynamics of crash occurrences, a reference value is calculated, against which the actual crash frequency can be compared. To do so, the same number of crash locations which are available from the crash reports is randomly distributed in the network space. For further refinement, additional constraints can be implemented, such as the proportional distribution of crashes for different road categories. For the simulation of crash locations, a random point generator is run for a sufficient number of realizations. The resulting normal distribution of simulated point numbers for each reference unit allows for the calculation of a mean value and the corresponding standard deviation. The simulated mean frequency of crash occurrences per reference unit neither serves as a basis for risk calculation nor is it the primary intent of this study to test the hypothesis of randomly distributed crashes. Instead, it is used as a (hypothetical) reference level with the number of standard deviations (z-score) as proxy for significant spatial patterns and temporal dynamics. Crash occurrences between ±2 standard deviations from the simulated mean are regarded as random and thus not considered in further analyses. In turn, very high or low z-scores indicate a non-random accumulation or variability of bicycle crashes. A similar approach can be found in Xie and Yan (2013), where a Monte-Carlo simulation is used as a “pseudo-significance” value for the determination of spatial autocorrelation in network clusters.

A high z-score indicates a larger population under risk and connected to this a higher number of crashes or points to other potentially underlying phenomena that might explain the accumulation or dynamic of crashes. Besides the most probable non-linear relation between traffic load and crash occurrence (“safety in numbers” phenomenon (Jacobsen, 2003)), reasons for crash accumulations can be temporal construction sites, bad street design, or road conditions, just to name a few. The result of the first analysis step, namely, the calculation of the z-score for each reference unit, is mapped spatially. This allows for the visual detection of regions with, relative to the simulated reference level, crash numbers significantly above or below the simulated reference level. With the help of temporal decomposition and aggregation, temporal variations in space can be visually observed (e.g. by animation) and statistically described (Fig. 1, stage 3). In order to further increase the amount of information, the results of stage 3 are implemented into an interactive environment for visual exploration (Fig. 1, stage 4). Here, additional variables from the police reports can be related to the spatial reference and the associated crashes, respectively. The visual exploration and statistical description of different spatial scale levels, levels of attributive aggregation, and temporal resolutions supports the generation of hypothesis for further, explanatory investigations. 4. Case study Following an exploratory approach, the presented case study was designed in absence of any statistical population under risk (exposure) data. Instead of risk estimation on a highly aggregated level, the single crashes were spatially and temporally binned and compared to simulated reference data. Changes in the deviation from the simulated mean were used as proxies for underlying dynamics. The proposed multi-stage workflow was applied to a case study in the city of Salzburg (Austria), where we analyzed a bicycle crash database, covering a period of 10 years. Before results are presented in Section 5, we introduce the study area, describe the data, and provide the specifications of the analysis environment.

3.3. Spatial and temporal analyses 4.1. Study area In order to detect and analyze spatial and temporal patterns and dynamics, the z-score is used as significance measure relative to random crash occurrences. In the spatial dimension, the analysis helps to identify network regions with crash occurrences considerably above or below the simulated random distribution. Here, the z-score is used for comparison purposes among different regions. For the temporal analysis, the simulated distribution serves as reference level for change over time.

The city of Salzburg has roughly 150,000 inhabitants, with approximately the same number of inhabitants in the larger agglomeration.1 Thus, the urban catchment of the city of Salzburg within bicycle distance 1 Figures for 2012 according to the Austrian federal bureau of statistics (Statistik Austria) and the Bavarian bureau of statistics (GENESIS online database).

42

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

Fig. 3. Urban catchment of the city of Salzburg (Austria) with approximately 305,000 inhabitants within 15 km from the city center (a). Road network in the city of Salzburg with bicycle facilities in green (b). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

covers slightly over 300,000 inhabitants in the Austrian–German border region (Fig. 3). The road network of the city of Salzburg which is legally relevant for bicycle traffic (highways and highway links are excluded) has a total length of 1045 km. About one third of these roads, 320 km, are equipped with some kind of bicycle facility, ranging from physically separated cycle ways to opened one-way roads with painted on-road cycle lanes. There are 180 km of dedicated cycle ways. According to official reports, the city of Salzburg has a bicycle share of 20% in the modal split, which is the highest share among the five largest cities in Austria (TEMS, 2011 cont.). The availability of bicycles per household is high: 81% of all households in the federal state of Salzburg own at least one bicycle (Illek and Mayer, 2013); the numbers tend to be even higher in the city of Salzburg. Although the number of all traffic crashes in the federal state of Salzburg significantly decreased over the last 15 years, the total number of bicycle crashes has steadily increased over the same time interval (Hemetsberger et al., 2013). 4.2. Data The two data sets, on which this study is based, are police reports of bicycle crashes and the digital road network graph. The police reports are collected by the federal bureau of statistics (Statistik Austria), which is responsible for regular, national crash reports. The processed data are then reported back to the responsible administrative units for further analysis and targeted countermeasures. In the case of Salzburg, all crash records are stored in a spatial database which is maintained by the city's department of transport and planning. The data of this study were extracted from this authoritative crash database. Each crash report contains the following attributive information: date, time, location, crash type (according to official categories), number of involved vehicles, weather condition, road surface condition and for all involved vehicles, the vehicle type, gender, and age of the driver, information about intoxication (alcohol) and injury severity in four classes (uninjured, slightly injured, badly injured, fatal). Information about liability is not contained in the data. The location information is mandatory and can be either stored as address location, linear referenced location, relative location from a prominent point of interest or by geographical coordinates. For this study, only crash reports with at

least one involved bicyclist were considered, resulting in a data set with 3096 reported incidents between January 2002 and December 2011. 48 reports were excluded due to invalid location information. In sum, 3048 crashes were objects of the analyses. In addition to the crash reports, the digital representation (graph) of the road network was used in the analyses. The road graph, managed by the city's department of transport and planning, contains attributive information about the physical and legal characteristics of the road network. This digital representation of the road network is used as network space for all further analyses. 4.3. Analysis environment The multi-stage workflow, as conceptually described in Section 3, was set up by coupling three software environments, via file exchange formats. All spatial analyses and mapping tasks have been performed using ESRI ArcGIS 10.1 with the non-commercial add-in SANET (Okabe and Sugihara, 2012). Tableau 8.1 has been used for visual data exploration. Statistics have been calculated with Python in ESRI ArcGIS 10.1. The temporal decomposition was operationalized with the “stl” algorithm in the “R” software package (Cleveland et al., 1990). For the case study, different numbers and sizes of reference units were tested. As elaborated on earlier, the number of reference units is a trade-off between number of cases (statistical robustness) and granularity (variability within reference units). For the present case study, we chose 100 reference units, with an average of 31.7 crashes per reference unit (standard deviation 47.5). Less reference units resulted in severe smoothing effects (x = 59.0 and σ = 87.2 for 50 reference units), whereas the average number of crashes per reference unit would have fallen below 30 by using more than 100 reference units. The seed points for the construction of the Voronoi diagrams were located randomly with a minimum distance buffer constraint, while considering the network density of the study area (Fig. 4). In order to calculate the mean number of simulated crash occurrence for each reference unit, random points were distributed along the network with 100 realizations. Because the crash locations are unevenly distributed and not proportionally related to the network length (Fig. 5), the road category was used as constraint for the random point generator. For each of the 100 conditional realizations the number of

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

43

spatial scales and temporal resolutions are elaborated on. Each result is discussed in the light of existing literature. 5.1. Spatial patterns

Fig. 4. Spatial reference units; colored and numbered reference units are relevant in the results section (ref. Figs. 9 and 10).

points per reference unit was counted. The resulting normal distribution of simulated point numbers for each reference unit allows for the calculation of mean and standard deviation. The z-score was calculated for each reference unit. These spatial units serve as initial, spatial level of aggregation. Further investigations within these units were done down to the single-crash-level. For the temporal dimension, the aggregate of the whole observation period, the single years, months, seasons, weekdays, and hour of the day were each considered as temporal subgroups. Years were analyzed chronologically, whereas the other temporal references referred to 10-year aggregates. 5. Results and discussion In the following, the major results of the analysis and visual exploration are presented for the case study. The section is structured as follows: first, the revealed spatial and temporal patterns are briefly presented, before spatio-temporal patterns, dynamics, and particularities on various

The 3048 reported crashes occurred at 1865 different locations (identical geographic coordinates derived from police reports): at 1379 locations, only one crash occurred, two to five crashes were reported at 423 locations, six to ten at 52, eleven to fifteen at 9, and more than that at 2 locations. Within the test area, the crashes are not evenly distributed but, as indicated in Fig. 6, strongly clustered. In approximately two thirds of all reference units, crashes are reported. Only ten reference units have more than 100 reported incidents within the 10 years of observation. Over the course of the investigated time period, most crashes occurred in reference unit 16. This reference unit is located in the city center (ref. Fig. 4) and has a comparable large share of high-quality bicycle infrastructure (48% of all segments are bicycle ways or have some kind of bicycle infrastructure). The statistical and visual exploration (steps 3 and 4 in the workflow described in Fig. 1) reveals that the characteristics of the crashes (weather and road conditions, severity of injuries, intoxication, age distribution, gender ratio, etc.) in this respective area are not significantly different from the rest of the city. Only the share of bicycle–bicycle and bicycle–pedestrian crashes is slightly higher than the overall average and the absence of fatalities is evident (Fig. 7). The reference units with the next highest numbers of crashes are adjacent to reference unit 16. They are located along the Salzach river to the north and the south and along a major corridor to the west. The five reference units with the highest number of crashes account for 30% of all incidents, although they comprise only 8.6% of the total network length. Again, there is no general tendency in the data, indicating any specific crash characteristics that would be different from the overall pattern. Aggregated to the highest temporal level (10 years of observation), a spatial clustering of crash locations along the Salzach river and the radial corridors becomes evident. The high spatial concentration of bicycle crashes is in line with Delmelle and Thill (2008) and Lovelace et al. (2015) who both found significant spatial concentrations (clusters) of bicycle crashes on the city-scale. Whereas the latter focused on single incidents, the first used census districts as aggregation level. In both studies, a planar kernel density estimation (KDE) was employed for cluster detection. Although the main focus of Vandenbulcke et al. (2009) is on crash risk and the spatial aggregation level and scale is different, they observe

Fig. 5. Number of bicycle crashes per road category.

44

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

Fig. 6. Spatial distribution of bicycles crashes: reference units and locations with multiple crash occurrences.

significant spatial heterogeneity as well. Generally, there are only a few studies that focus on the spatial clustering of bicycle crashes or risky spots on a local scale, down to single-crash locations (Steenberghen

et al. (2004) and (Xie and Yan, 2013) for road incidents in general, Lovelace et al. (2015) and Vandenbulcke et al. (2014) with a focus on bicycle crashes). However, for evidence-based interventions and planning

Fig. 7. Example for how visual exploration environments are employed. Reference unit 16 (in green) is plotted against the rest of the study area (in light grey) in this parallel coordinate plot (PCP).

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

purposes, an aggregated approach does not sufficiently account for the spatial variability on a micro-scale. As it becomes evident in this case study, a first exploratory investigation helps to focus on network areas where further data collection (for an exposure variable) and subsequent investigations are potentially beneficial. 5.2. Temporal patterns To explore long-run temporal patterns in the occurrence of crashes, the data set was decomposed into three components: a periodic seasonality, a long-term trend and the remaining noise. First, the seasonal pattern was determined by computing the mean of crash occurrences for each month. Second, a moving window with a time span of 2 years smoothed the remaining data to extract the long-term trend component. Third, the remaining noise was determined by removing trend and seasonal oscillation. The temporal decomposition of the crash occurrences shows a distinct periodicity, that is a variation of crash occurrences depending on the season, and a trend for an increasing number of incidents. Fig. 8 shows the observed number of crashes at a weekly granularity, where the absolute numbers range between 0 and 20 crashes per week. The seasonality (2nd row, Fig. 8) is expressed in terms of standard deviations from the mean. The trend curve shows the smoothed (2-year-moving window) result after the subtraction of the seasonal oscillation. What remains (bottom row, Fig. 8) is the weekly noise. In the winter season, the number of crashes decreases to five incidents per week below the mean, whereas the number peaks in the summer season at six crashes above the mean (Fig. 8). In addition to recurrent seasonal oscillations, the overall trend increases over time by 2.5 additional crashes per week. The trend line reached a maximum in 2007/08 with almost three additional crashes per week. The remaining random variation exhibits irregular behavior that ranges between −9 and +12 crashes per week. Similar oscillations were found as well at other temporal scales, including characteristic daily and weekly patterns. Daily means over the entire observation period range between 0.4 crashes on Sundays and an almost three times higher mean of 1.1 crashes on Wednesdays and Thursdays. The daily variation at an hourly resolution even amounts to a ratio of 1/37. The average temporal patterns in the course of a day are similar to findings from the UK (Aldred and Crosweller, 2015; Lovelace et al., 2015) and Canada (Doherty and Aultmann-Hall, 2000). The seasonal effect, with a distinct peak in the summer months, was also observed in Toronto and Ottawa. Interestingly, Doherty and Aultmann-Hall

45

(2000) found in this context that severity of collisions remains the same over the year. The overall increase of bicycle crashes in the same or similar time period is also reported for different levels of aggregation by DfT (2014) for the UK, Lovelace et al. (2015) for West Yorkshire or Hemetsberger et al. (2013) for the federal state of Salzburg. The sharp increase in 2007/08, the following relaxation and an increase after 2010 (see trend line in Fig. 8) is also visible in the nationwide statistics (Illek and Mayer, 2013). Thus, it can be concluded that the long-run temporal pattern can be explained by external variables, such as the relation of traffic volume and weather condition (Spencer et al., 2013). 5.3. Spatio-temporal patterns and crash characteristics Taken together, the spatial and temporal dimension of crashes allows for a further spatio-temporal investigation of these events. For this, temporal increments have been defined related to the respective years, weekdays, hour of day, and seasons (each aggregated over the ten years of observation). The predefined network units (Section 4.3 and Fig. 4) served as spatial reference. Fig. 9 indicates the seasonality of spatial patterns. Whereas most crashes (in terms of quantity and significance) occur in reference unit 16 during spring, summer, and fall season, the overall pattern obviously changes in winter. The number of crashes in general is lower and – different to the spring or summer seasons – there is no reference unit which stands out in terms of crash occurrences. From a spatial perspective, it becomes evident, that the reference units with the highest amount of crashes and the highest number of deviations from the simulated mean are not necessarily in the city center. Reference unit 8, for example, covers two major corridors that connect cycle ways from the western and southwestern periphery to the main cycle way along the Salzach river. Investigating the spatio-temporal pattern of bicycle crashes over the 10 years of observation reveals certain dynamics (see Fig. 10), although a general pattern remains stable. Reference unit 16 is the network region with most reported incidents in most years. Exceptions from this can be observed in 2002, 2006, and 2011. Nevertheless, even in these years, reference unit 16 is among the regions with the highest crash prevalence. Interestingly, reference units with a high number of crashes in 1 year might “disappear” in another one. In reference unit 17, for example, 17 crashes occurred in 2002, what was the highest amount that year. In the last year of observation, 2011, the same number of crashes occurred in this reference unit, but with less deviation from the simulated mean (2002/2011 with 19/9 standard deviations from mean). Reference unit 8 is another example for an area with a

Fig. 8. Weekly aggregated crash occurrences (1st row, “data”) are temporally decomposed into three parts: seasonal oscillations (2nd row, “seasonal”), a long-term trend (3rd row, “trend”), and the remaining noise (4th row, “remainder”).

46

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

Fig. 9. Seasonality of spatial pattern (blue and red represent the respective negative or positive deviation from simulated mean). Labels indicate the reference unit's ID. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

Fig. 10. Spatio-temporal dynamics over 10 years of observation. Labels indicate the reference unit's ID.

47

48

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

discontinuous development. In 2011, this was the area with the most reported crashes (28 incidents), whereas the year before “only” 16 crashes occurred in this particular reference unit. Such observations and the related deviation from the simulated mean might point to temporal, peculiar events, such as construction sites, in the respective areas. Apart from the observed dynamics in the annual comparison and the evident periodicity over the year, recurring temporal patterns can also be observed in the course of a week and a day, with remarkable outliers in the spatial dimension. In total, most crashes occur on Wednesdays (561 crashes = 18.41% of total). Slightly fewer crashes are reported on Thursdays. 15.52% of all crashes occur during the weekend. On the level of the reference units, the pattern are less homogenous. Above all, reference unit 16, the area with the highest total number of reported incidents, stands out. Here by far most crashes (62 or 24.41%) occur on Thursdays, whereas the day with the second most crashes is Wednesday with 44 incidents. Excluding reference unit 16 from the overall statistics reveals a rather constant number of crashes from Monday to Thursday with an average of 479 crashes per day (σ = 22.5) for the entire period of observation. The temporal deviation of reference unit 16 probably points to an underlying, recurring event, which would be worth further investigation: every Thursday, the city's largest greenmarket takes place in reference unit 16. This market is a major attractor and highly frequented by bicyclists. Although correlations between this event and the crash frequency cannot be directly deduced from the data, the analyses of the spatial and temporal dynamics suggest particularities in this area for a certain point of time. However, this example clearly demonstrates the effect of a disaggregated approach and demonstrates how an exploratory analysis can act as filter in order to focus on relevant areas and time intervals and draw more specific conclusions from statistics. Furthermore, the relation between spatial configuration (in this case reference unit 16 differs from the rest) and temporal resolution becomes evident. Thus the effect of spatial aggregation and temporal resolution is crucial for subsequent analyses; an aspect which is hardly ever addressed in epidemiological studies with aggregated statistics as input variables. Over the day, two peaks with different magnitudes can be observed. In total, most crashes occur between 5 and 6 pm (295 crashes = 9.68% of total). Starting from 2 pm, the number of crashes increases until this peak and sharply decreases between 6 and 7 pm. The second, but much more moderate, peak occurs in the late morning with a relative high between 10 am and 12 pm. The number of crashes is comparatively low in the morning hours where most commuters are usually on the road (182 crashes between 7 and 8 am). Again, on the level of reference units, the patterns are not homogenous at all. Crashes in reference unit 16 peak between 11 am and 12 pm and account for 11.64% of the crashes in this time interval. This peak stands out of the general pattern of small peaks in the morning and high peaks in the late afternoon. A very discontinuous pattern can be observed in reference unit 36 – which is located between the main station and the Salzach river, north of the city center – with four distinct peaks between 8 and 9 am, 11 am and 12 pm, 2 and 3 pm, and 5 and 6 pm. Again, the cause of this temporal pattern in reference unit 36 cannot be determined from the data. But the concentration of schools (with high numbers of people arriving and leaving at school start and end) in this area could serve as starting point for further in-depth investigations. Again, this example demonstrates the contribution of a spatio-temporal disaggregation of crash data for more targeted in-depth investigations. Obviously, driving factors for an increase of bicycle crashes can be very different, depending on spatial and temporal configurations. Although relations between crashes and external factors, such as weather and road conditions, can be observed on an aggregated level, the disaggregated analysis of the present crash data suggests additional, specific cofactors. Dividing the dataset into two subsets according to the gender of the crash victims does not reveal any spatial particularities. Only the

temporal variation over the course of a day exhibits a few remarkable characteristics. Female bicyclists were more often involved in a crash in the morning hours than during the rest of the day regarding the absolute numbers. Between 10 and 11 am is the only hour of the day where more female than male bicyclists have a crash. The statistical overhang of male victims is largest in the late afternoon hours and in the morning between 7 and 8 am. Whereas the predominance of male crash victims is expected from previous studies (Aldred and Crosweller, 2015; Lovelace et al., 2015) and can be explained by a general higher share of male bicyclists (Ogilvie and Goodman, 2012) and a higher incline to take risk (Cobey et al., 2013), the inverted ratio between 10 and 11 am calls for closer investigation. The segmentation of the data set into two groups shows how different patterns and dynamics can be revealed for subgroups. The exploration of other subgroups, for example, age groups, was not possible because of the comparable low number of crashes for the respective subgroups. 6. Conclusion In the presented case study, crashes are considered as explicitly spatial and temporal phenomena. This means that the location and point of time of bicycle crashes are regarded as fundamental characteristics. The proposed multi-staged workflow turned out to be performant and supportive in better understanding when and where bicycle crashes happen on a city-scale level. On the conceptual level, the workflow is transferable to other cities and application domains where point incidents are spatially and temporarily analyzed in a network space. Here, the delineation of network-based reference units as initial aggregate units and the conditional simulation of a random distribution for the calculation of a z-score which serves as proxy for variability are extensions or alternatives to existing approaches. The case study underpins the extensive variability of crash occurrences in the spatial and temporal dimension, which of course cannot be considered in more aggregated approaches. Characteristic patterns and dynamics emerge depending on the spatial level of aggregation and temporal resolution. These patterns and dynamics are crucial for evidencebased interventions and planning decisions within cities. The workflow proposed here can therefore serve as initial filter before specific data are captured and the respective areas are investigated in detail. From a methodological perspective, the workflow, illustrated in Fig. 1, merges and extends several existing approaches, from the transportation, spatial analysis, and visualization communities. In contrast to studies which analyze crash locations in planar space, the present approach accounts for the network-bound character of bicycle crashes. Consequently, all analysis steps, starting from the delineation of the reference units, are adapted to the network space. The proposed analysis approach starts from initially defined spatial and temporal aggregate units, following an interactive exploration down to the level of single crashes. The design of the implemented environment for visual exploration allows for a smooth transition between scales and resolutions. Because of this flexible scaling, additional information can be retrieved from the data and more efficient in-depth analysis can be done in a consecutive step. Although the findings in the case study are specific to the study site, general conclusions can be drawn from the results: • The spatial heterogeneity within typical spatial aggregates, be it census districts or the whole city, is high. Crash locations are highly clustered, primarily along major bicycle corridors with several radial connections (ref. Alrutz et al., 2015). Thus, it was shown that an exploration of bicycle crashes on a finer scale does generate valuable, additional information which is relevant for cutting down consecutive data demands (e.g. optimal location for counting stations) considerably. • The temporal variation of bicycle crash occurrences is evident on several time axes. Whereas the overall patterns and trends on the aggregation level of the whole city are perfectly in line with existing literature, the

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

case study revealed significant particularities on the local scale. Hence, through the exploration of the bicycle crash data on several levels of aggregation and scale additional insights were gained. Up to now, only a few studies have investigated bicycle crash occurrences and bicycle risk patterns on the city-scale level. This is mainly due to a shortage of available data which meet the requirements in terms of detail and quality. Implementing explorative studies such as this one prior to the in-depth investigation is expected to be an efficient way to cut down the data demand and the effort for the development of models and analyses. However, further research questions in the context of explorative studies which focus on spatial and temporal patterns and dynamics remain. With bicycle crash research and spatio-temporal point pattern analysis in general in mind, we thus suggest to address the following issues: • While the regionalization with network-based Voronoi diagrams from conditional random point distributions (delineation of spatial reference units) proved to be suitable, it is yet unclear to which degree the analysis results depend on the network-based regionalization approach. We defined the reference units considering a minimum number of incidents while aiming for a maximum of compactness. Spatial effects on the outcome, such as the MAUP (Wong, 2009), were not systematically investigated. • Any attempt to tackle the problem of massive underreporting helps to improve the results of bicycle crash analyses. This, of course, does not only hold true for the present study, but for any investigation of crash data. Complementary to additional data sources, such as hospital records, crowd sourcing initiatives (Nelson et al., 2015) and the implementation of active and passive (technical) human sensors (Griffin and Jiao, 2015; Resch et al., 2015) might be promising in this respect. • Although spatial and temporal dynamics became evident, it is not clear, which parameters exactly drive these dynamics on the microscale. Dependencies and causalities remain unclear. Further statistical analyses and qualitative investigations (e.g. focus group interviews) will help to gain more specific insights and judge on the generated hypotheses. • Given that the sample size is large enough, steps four and three in the workflow could be exchanged: from a visual exploration, potentially interesting subsets could be derived and subsequently investigated for distinct patterns in the spatial and/or temporal dimension. This would help to design further investigation accordingly in order to gain sub-group-specific insights (e.g. elderly people, children, female etc.).

The conditional random distribution of simulated crashes allows for the calculation of the z-score (standard deviation) based on a simulated mean, but cannot be used for any risk calculation. For this, traffic flows (or any alternative statistical population on the level of single road segments) need to be estimated. Such traffic models require highquality input data with a fine spatial (and ideally temporal) resolution for parametrization, calibration, and validation. Whereas transportation models on the macro-scale are not suitable for local risk estimations, aggregated flows derived from individual agents (Wallentin and Loidl, 2015) are regarded to be a promising alternative. Financial support None. Conflict of interest The authors declare that they have no conflict of interest.

49

References Aldred, R., Crosweller, S., 2015. Investigating the rates and impacts of near misses and related incidents among UK cyclists. J. Transp. Health 2, 379–393. Alrutz, D., Bohle, W., Maier, R., Enke, M., Pohle, M., Zimmermann, F., Ortlepp, J., Schreiber, M., 2015. Einfluss von Radverksaufkommen und Radverkehrsinfrastruktur auf das Unfallgeschehen. In: GDV (Ed.), Unfallforschung der Versicherer. Gesamtverband der Deutschen Versicherungswirtschaft, Berlin, p. 190. Anderson, T.K., 2009. Kernel density estimation and K-means clustering to profile road accident hotspots. Accid. Anal. Prev. 41, 359–364. Beck, L.F., Dellinger, A.M., O'Neil, M.E., 2007. Motor vehicle crash injury rates by mode of travel, United States: using exposure-based methods to quantify differences. Am. J. Epidemiol. 166, 212–218. Beecham, R., Wood, J., 2013. Exploring gendered cycling behaviours within a large-scale behavioural data-set. Transp. Plan. Technol. 37, 83–97. Borruso, G., 2008. Network density estimation: A GIS approach for analysing point patterns in a network space. Trans. GIS 12, 377–402. Braga, A.A., Hureau, D.M., Papachristos, A.V., 2010. The relevance of micro places to citywide robbery trends: A longitudinal analysis of robbery incidents at street corners and block faces in Boston. J. Res. Crime Delinq. 48, 7–32. Chen, D., Fuller, D., 2014. Analyzing road surface conditions, collision time, and road structural factors associated with bicycle collisions from 2000 to 2010 in Saskatoon, Saskatchewan. J. Transp. Health 1, 40–44. Cleveland, R.B., Cleveland, W.S., Mcrae, J.E., Terpenning, I., 1990. A seasonal-trend decomposition procedure based on loess. J. Off. Stat. 6, 3–73. Cobey, K.D., Stulp, G., Laan, F., Buunk, A.P., Pollet, T.V., 2013. Sex differences in risk taking behavior among dutch cyclists. Evol. Psychol. 11, 350–364. Cressie, N.A.C., 1993. Statistics for Spatial Data. John Wiley & Sons, New York. Dai, D., 2012. Identifying clusters and risk factors of injuries in pedestrian–vehicle crashes in a GIS environment. J. Transp. Geogr. 24, 206–214. De Geus, B., Vandenbulcke, G., Int Panis, L., Thomas, I., Degraeuwe, B., Cumps, E., Aertsens, J., Torfs, R., Meeusen, R., 2012. A prospective cohort study on minor accidents involving commuter cyclists in Belgium. Accid. Anal. Prev. 45, 683–693. De Rome, L., Boufous, S., Georgeson, T., Senserrick, T., Richardson, D., Ivers, R., 2013. Bicycle crashes in different riding environments in the Australian Capital territory. Traffic Inj. Prev. 15, 81–88. Degraeuwe, B., de Geus, B., Thomas, I., Vandenbulcke, G., Meeusen, R., Panis, L.I., 2015. Cycling behaviour and accident risk of utilitarian cyclists in Belgium. In: Gerike, R., Parkin, J. (Eds.), Cycling Futures—From Research into Practice. Ashgate, Surrey, pp. 33–51. Delmelle, E.C., Thill, J.-C., 2008. Urban bicyclists: Spatial analysis of adult and youth traffic hazard intensity. Transp. Res. Rec. J. Transp. Res. Board 2074, 31–39. DfT, 2014. Reported Road Casualties in Great Britain: Main Results 2013, In: Department for Transport (Ed.), London. Diggle, P., 2013. Statistical Analysis of Spatial and Spatio-Temporal Point Patterns. third ed. CRC Press, Boca Raton. Doherty, S.T., Aultmann-Hall, L., 2000. Commuter cyclist accident patterns in Toronto and Ottawa. J. Transp. Eng. 126, 21–26. Dozza, M., Werneke, J., 2014. Introducing naturalistic cycling data: what factors influence bicyclists' safety in the real world? Transport. Res. F: Traffic Psychol. Behav. 24, 83–91. Eckley, D.C., Curtin, K.M., 2013. Evaluating the spatiotemporal clustering of traffic incidents. Comput. Environ. Urban. Syst. 37, 70–81. Garrard, J., Rose, G., Lo, S.K., 2008. Promoting transportation cycling for women: the role of bicycle infrastructure. Prev. Med. 46, 55–59. Gerike, R., Parkin, J., 2015. Cycling Futures: From Research into Practice. Ashgate, Farnham, Surrey, p. 322. Götschi, T., Garrard, J., Giles-Corti, B., 2015. Cycling as a part of daily life: A review of health perspectives. Transp. Rev. 1–27. Griffin, G.P., Jiao, J., 2015. Where does bicycling for health happen? Analysing volunteered geographic information through place and plexus. J. Transp. Health 2, 238–247. Harris, M.A., Reynolds, C.C.O., Winters, M., Cripton, P.A., Shen, H., Chipman, M.L., Cusimano, M.D., Babul, S., Brubacher, J.R., Friedman, S.M., Hunte, G., Monro, M., Vernich, L., Teschke, K., 2013. Comparing the effects of infrastructure on bicycling injury at intersections and non-intersections using a case–crossover design. Inj. Prev. Heinen, E., van Wee, B., Maat, K., 2010. Commuting by bicycle: an overview of the literature. Transp. Rev. 30, 59–96. Hemetsberger, U., Kolator, R., Rettenbacher, P., 2013. Gib8! - Verkehrssicherheitsprogramm 2013–2016. Amt der Salzburger Landesregierung, Salzburg, p. 26. Holm, A.L., Glümer, C., Diderichsen, F., 2012. Health impact assessment of increased cycling to place of work or education in Copenhagen. BMJ Open 2. Illek, G., Mayer, I., 2013. Radverkehr in Zahlen - Daten, Fakten und Stimmungen. BMVIT, Wien. ITF, 2012. Cycling Safety: Key Messages. In: Tørsløv, N. (Ed.), International Transport Forum - Working Group on Cycling Safety. OECD, Paris. Jacobsen, P.L., 2003. Safety in numbers: more walkers and bicyclists, safer walking and bicycling. Inj. Prev. 9, 205–209. Kingsley, M., Kanneworff, P., Carlsson, D., 2004. Buffered random sampling: a sequential inhibited spatial point process applied to sampling in a trawl survey for northern shrimp Pandalus borealis in West Greenland waters. ICES J. Mar. Sci. J. Conseil. 61, 12–24. Kröyer, H.R.G., 2015. The relation between speed environment, age and injury outcome for bicyclists struck by a motorized vehicle—a comparison with pedestrians. Accid. Anal. Prev. 76, 57–63. Loidl, M., Zagel, B., 2014. Assessing bicycle safety in multiple networks with different data models. In: Vogler, R., Car, A., Strobl, J., Griesebner, G. (Eds.), GI-Forum. Wichmann, Salzburg, pp. 144–154.

50

M. Loidl et al. / Journal of Transport Geography 52 (2016) 38–50

Lord, D., Mannering, F., 2010. The statistical analysis of crash-frequency data: a review and assessment of methodological alternatives. Transp. Res. A Policy Pract. 44, 291–305. Lovelace, R., Roberts, H., Kellar, I., 2015. Who, where, when: the demographic and geographic distribution of bicycle crashes in West Yorkshire. Transport. Res. F: Traffic Psychol. Behav. (pending). Lusk, A.C., Furth, P.G., Morency, P., Miranda-Moreno, L.F., Willett, W.C., Dennerlein, J.T., 2011. Risk of injury for bicycling on cycle tracks versus in the street. Inj. Prev. 17, 131–135. Martínez-Ruiz, V., Jiménez-Mejías, E., Amezcua-Prieto, C., Olmedo-Requena, R., Luna-delCastillo, J.d.D., Lardelli-Claret, P., 2015. Contribution of exposure, risk of crash and fatality to explain age- and sex-related differences in traffic-related cyclist mortality rates. Accid. Anal. Prev. 76, 152–158. Nelson, T.A., Denouden, T., Jestico, B., Laberee, K., Winters, M., 2015. BikeMaps.org: a global tool for collision and near miss mapping. Front. Public Health 3, 53. Nordback, K., Marshall, W.E., Janson, B.N., 2014. Bicyclist safety performance functions for a U.S. city. Accid. Anal. Prev. 65, 114–122. OECD, 2013. Cycling, health and safety. In: Tørsløv, N. (Ed.), ITF-OECD Working Group on Cycling Safety, Paris, p. 249. Ogilvie, F., Goodman, A., 2012. Inequalities in usage of a public bicycle sharing scheme: socio-demographic predictors of uptake and usage of the London (UK) cycle hire scheme. Prev. Med. 55, 40–45. Okabe, A., Satoh, T., 2009. Spatial Analysis on a Network. In: Fotheringham, A.S., Rogerson, P.A. (Eds.), The SAGE Handbook of Spatial Analysis. Sage, London, pp. 443–464. Okabe, A., Sugihara, K., 2012. Spatial Analysis along Networks—Statistical and Computational Methods. John Wiley & Sons, Chichester. Okabe, A., Satoh, T., Sugihara, K., 2009. A kernel density estimation method for networks, its computational method and a GIS-based tool. Int. J. Geogr. Inf. Sci. 23, 7–32. Pucher, J., Dill, J., Handy, S., 2010. Infrastructure, programs, and policies to increase bicycling: an international review. Prev. Med. 50, 106–125. Quddus, M.A., Ochieng, W.Y., Noland, R.B., 2007. Current map-matching algorithms for transport applications: state-of-the art and future research directions. Transp. Res. C Emerg. Technol. 15, 312–328. Resch, B., Summa, A., Sagl, G., Zeile, P., Exner, J.-P., 2015. Urban emotions—geo-semantic emotion extraction from technical sensors, human sensors and crowdsourced data. In: Gartner, G., Huang, H. (Eds.), Progress in Location-Based Services 2014. Springer International Publishing, pp. 199–212. Reynolds, C.C., Harris, M.A., Teschke, K., Cripton, P.A., Winters, M., 2009. The impact of transportation infrastructure on bicycling injuries and crashes: a review of the literature. Environ. Heal. 8, 47. Rietveld, P., Daniel, V., 2004. Determinants of bicycle use: do municipal policies matter? Transp. Res. A Policy Pract. 38, 531–550. Sanders, R.L., 2015. Perceived traffic risk for cyclists: the impact of near miss and collision experiences. Accid. Anal. Prev. 75, 26–34. Schepers, P., 2013. A safer road environment for cyclists. Transport and Planning. TU Delft, Delft, p. 205. Shiode, S., 2008. Analysis of a distribution of point events using the network-based quadrat method. Geogr. Anal. 40, 380–400.

Shiode, S., Shiode, N., 2013. Network-based space–time search-window technique for hotspot detection of street-level crime incidents. Int. J. Geogr. Inf. Sci. 27, 866–882. Spencer, P., Watts, R., Vivanco, L., Flynn, B., 2013. The effect of environmental factors on bicycle commuters in Vermont: influences of a northern climate. J. Transp. Geogr. 31, 11–17. Steenberghen, T., Aerts, K., Thomas, I., 2010. Spatial clustering of events on a network. J. Transp. Geogr. 18, 411–418. Steenberghen, T., Dufays, T., Thomas, I., Flahaut, B., 2004. Intra-urban location and clustering of road accidents using GIS: a Belgian example. Int. J. Geogr. Inf. Sci. 18, 169–181. TEMS, 2011. Cont. City modal split database. European Platform on Mobility Management, Brussels. Teschke, K., Harris, M.A., Reynolds, C.C.O., Winters, M., Babul, S., Chipman, M., Cusimano, M.D., Brubacher, J.R., Hunte, G., Friedman, S.M., Monro, M., Shen, H., Vernich, L., Cripton, P.A., 2012. Route infrastructure and the risk of injuries to bicyclists: a casecrossover study. Am. J. Public Health 102, 2336–2343. Thomas, B., DeRobertis, M., 2013. The safety of urban cycle tracks: a review of the literature. Accid. Anal. Prev. 52, 219–227. Tin Tin, S., Woodward, A., Ameratunga, S., 2013. Completeness and accuracy of crash outcome data in a cohort of cyclists: a validation study. BMC Public Health 13. Vandenbulcke, G., Thomas, I., de Geus, B., Degraeuwe, B., Torfs, R., Meeusen, R., Int Panis, L., 2009. Mapping bicycle use and the risk of accidents for commuters who cycle to work in Belgium. Transp. Policy 16, 77–87. Vandenbulcke, G., Thomas, I., Int Panis, L., 2014. Predicting cycling accident risk in Brussels: a spatial case–control approach. Accid. Anal. Prev. 62, 341–357. Vandenbulcke-Plasschaert, G., 2011. Spatial Analysis of Bicycle Use and Accident Risks for Cyclists. Université catholique de Louvain, Leuven, p. 333. Wallentin, G., Loidl, M., 2015. Agent-based bicycle traffic model for Salzburg City. GI_ Forum – Journal for Geographic Information Science 2015, pp. 558–566. Wanvik, P.O., 2009. Effects of road lighting: an analysis based on Dutch accident statistics 1987–2006. Accid. Anal. Prev. 41, 123–128. Wegman, F., Zhang, F., Dijkstra, A., 2012. How to make more cycling good for road safety? Accid. Anal. Prev. 44, 19–29. WHO, 2013. Global Status Report on Road Safety 2013—Supporting a Decade of Action. World Health Organization, Geneva, p. 318. Winters, M., Davidson, G., Kao, D., Teschke, K., 2011. Motivators and deterrents of bicycling: comparing influences on decisions to ride. Transportation 38, 153–168. Wong, D., 2009. The Modifiable Areal Unit Problem (MAUP). In: Fotheringham, S., Rogerson, P.A. (Eds.), The SAGE Handbook of Spatial Analysis. SAGE Publications Ltd., London, pp. 105–123. Xie, Z., Yan, J., 2013. Detecting traffic accident clusters with network kernel density estimation and local spatial statistics: an integrated approach. J. Transp. Geogr. 31, 64–71. Yiannakoulias, N., Bennet, S.A., Scott, D.M., 2012. Mapping commuter cycling risk in urban areas. Accid. Anal. Prev. 45, 164–172.