Spatial variation in bicycling risk based on crowdsourced safety data

Bicycling-related injury data are difficult to obtain from official reports, which capture only about 20% of crashes and often lack coordinates, outcomes, narratives needed for understanding where why incidents occurred. Crowdsourced on bicycling safety provides new opportunities the study risk. Our goal was quantify factors that influence spatial variation in unsafe across a city, based self-reports incidents. To meet this goal, we leveraged BikeMaps.org, global tool reporting incidents, drawing Metro Vancouver. We summarized incident conditions led injury, developed model identify predictors using random forest regression, mapped hot spots. results demonstrate injuries associated with older younger bicyclists, downhill slopes, parked cars, recreation weekend rides, falls, single bicycle infrastructure, roads, railroads. The broad range reported BikeMaps.org allows us add evidence falls collisions major causes injury. Also, value attributing spots contextual details infrastructure interventions can reduce bicyclists. Les données sur les blessures reliées à la pratique du vélo sont difficiles obtenir des rapports officiels qui ne recensent qu'environ collisions. De plus, il manque souvent dans ces coordonnées lieux de collision, nature et le détail faits requis pour comprendre où pourquoi se produits. type « crowdsourcing » sécurité cyclisme offrent nouvelles possibilités l'étude risques reliés vélo. L'objectif cette recherche est quantifier facteurs influencent spatiale associés une ville selon propres déclarations personnes impliquées Pour atteindre cet objectif, nous avons misé un outil reconnu signaler vélo, en focalisant grand Dans optique, compilé caractéristiques puis développé modèle arbre décision identifier variables explicatives et, finalement, cartographié points chauds Nos résultats démontrent que touchent davantage plus jeunes vieux cyclistes, elles localisent pentes descendantes, impliquent voitures stationnées, font lors balades récréatives fin semaine, aux chutes, infrastructures ainsi qu'aux routes voies ferrées. Le large éventail d'incidents signalé permet d'ajouter probantes fait chutes individuelles cyclistes majeures blessures. Au final, démontrons valeur l'identification incluant contexte l'incident visant routières peuvent réduire cyclistes. Bicycling is shown improve health relieve urban congestion (Oja al. 2011; Götschi 2016; Zahabi 2016). Cities around world promoting as healthy sustainable form travel (Buehler Pucher 2012; Fishman Buehler 2017; Fan 2019; Lee Pojani Rosas-Satizábal Rodriguez-Valencia Soliz 2021). However, concerns pose barriers, especially people riding bicycles share road driving vehicles (Heinen 2010; 2010). overcome increase bicycling, cities investing bicycling-specific separating bicyclists drivers has been real perceived (Sanders 2015). Lack data, including challenge when aiming implement pro-bicycling policy (Winters Branion-Calles 2017). Traditional include police insurance claims, hospital records, biased toward account ~20% all These sources under-sample crashes, do not involve vehicles, multiuse paths, near misses. datasets filling gaps provide enhanced ridership, perceptions attitudes (Nelson, Ferster, have sampling high ridership areas (e.g., paths), typically source miss (Mannering Bhat 2014; Nelson As such, crowdsourced be useful comparisons help prioritize investments will safety. effective use also requires it important present relative exposure, case number at particular place time. New approaches collection enabled by smartphone technology expanded possibilities much larger sample wider variety Websites mobile apps forum individuals report locations safety-related Crowdsourcing minor early warning preventable awareness risks lead serious (Aldred Furthermore, they negatively impact people's hampering efforts viable mode transport population Crosweller supplement traditional datasets, example providing more granular increasing amount available research. city drew upon (Nelson 2015) Vancouver, British Columbia. First, proportion those characteristics places incident; second, regression; third, risk kernel density estimation labelled narrative descriptions self-reports. set Columbia (Figure 1). Vancouver federation 21 municipalities (including City Vancouver), one Treaty First Nation (Tsawwassen Nation), electoral area diverse transportation (Metro promoted itself living, an emphasis fitness active outdoor lifestyle. Its temperate climate (average 4–18°C (Environment Canada 2021)) conducive year-round bicycling. Across journey-to-work above average Canada, overall rate 2.3% compared national 1.4% (Statistics 2019). municipalities, rates <0.5% less central sprawling 6.1% Bicycle nearly doubled 3.3% 1996 2016 Ridership doubling reflects positive outcomes support recent decades. In 2020, Vancouver's total route network 316 km, comprised local street bikeways (shared roadway along streets, traffic-calmed, 57%); cycle tracks (roadway lane exclusively physically separated motor sidewalk, 10%); paths (two-way paved path shared pedestrians, 20%); painted bike lanes (painted busy roadway, 13%) (City 2021; Firth densest downtown adjacent neighbourhoods, corresponds neighbourhoods highest densities region. used dataset 1,300 (collisions, misses) over six-year period (2014–2020, Figure Launched 2014, website app map For each incident, reporters location online complete questions dropdown options three categories: details, conditions, personal details. Extensive template elsewhere brief, Table 1 lists posed users. Mandatory “Were you injured?” answer “yes” or “no,” well related occurred). Reporters may detailed (up 300–500 characters) open-ended text section. reviewed ensure were correctly classified. considered 18 explanatory (Table majority (n = 17) collected derived included additional variable representing daily volumes locations. represent volumes, quantifying exposure potential (Ferster 2021), Strava generated popular tracking app. track activities provided free charge partner decision making. gained momentum stand-in pervasive count programs monitoring activities, control (Lee Sener app, although researchers strong correlations between levels (Jestico Conrow 2018; bias while benefiting continuous categorially define five categories ridership: very low (<6), (6–55), medium (56–230), (231–4,305), (>4,305). planning purposes classes appropriate given precision Roy, reports within 2014 2020 calculated gender, age, frequency trip purpose, occurred, object involved incident. understand what had greater prevalence category resulted trained balanced classifier predict injuries. Random supervised machine-learning algorithm explain importance (Breiman 2001). It ensemble trees, such tree grown values vector sampled independently same distribution trees forests classification procedure determines relevant splitting into list level “importance” classification. Variable measures accuracy, top ranking causing greatest decrease accuracy removed (Branion-Calles Fischer 2020). Variables negative thus scores final model. partial dependence plots visualize relationship Plots show probability variable, holding other constant. When developing priorities investments, consider density, “hot spots.” converted (KDE). KDE visualizing variability point works creating smoothed spot surface events. smoothness defined function bandwidth, literature suggests testing surfaces bandwidths accepted method selecting best represents (Atkinson Unwin 2002; Boots 2008). tested 100 m 500 50-m increments selected bandwidth 400 because distinct represented area. Following (2008), 10% values. better context spots, examined after mapping completed. grouped spot, Then inductively coded (Thomas 2006) common aspects spot. Summary statistics 2 information relate below. absolute counts vary widely, measure direct comparison. Of 1,297 quarter (26.5%, n 343) Incidents most men (44.3%), aged 30–49 (45.0%), who least once week (70.1%). 2, characteristic. Results lowest (19.9–20.9%). Older adults (50+) under 30 (32.1% 100.0%, respectively). Falls (83.3% vs. 74.2%). well, recreational (39.3%) (35.1%) proportions types trips weekday rides (see 2). occurred result crashing (curb, divider, pothole, sign/post), train (78.6–83.3%). regression. 92%. order prediction, were: (collision, fall, miss); terrain bicyclist (uphill, flat, downhill); age bicyclist; whether there cars Other minimal predictive power. Partial four 3) detail specific relationships prediction. stationary objects both slightly higher resulting than moving objects. Injury likely predicted terrain, (>49 years) (<30 streets roadside. identified 14 evaluated crash 4; 3). (64.0%) vehicles. Report open-text revealed some site-specific Four (E, G, K, M) involving either turning failing yield. At M, conflicts traffic circle. Driver failure stop yield 50% two (D J). Conflict (A, L). Only (C, H, N) did conflict these, H N tracks. biggest (B, F, I, L) diversity types. characterized pedestrian, bicyclist, car reflect inter-modal conflict. mechanism misses, injuries, bike-only falls). Another benefit ability happened during While qualitative analysis draw our discussion below contextualizing quantitative quotes study, found People 49 likelihood ending summary modelling, despite comprising lower reports. corroborate similar research findings associations (Vanparijs 2015; Prati Chen Shen Liu 2020; Meuleners example, Australian experience injured (Poulos could vulnerable due general aging-associated physiology. There young people; however, burden underestimated through mechanisms police, insurance) (Watson does uncover explanations being injured, but underrepresented studies view fill opportunities. topography, notably road. Downhill slopes linked increased severity several (Teschke Harris 2013; 2019), speed going downhill. Vehicles roads obstruct sightlines doorings, substantive (Ferenchak Marshall indicate Single This adds under-recognized significant cause Myhrmann part, made (Foley Utriainen outline how transitions off problematic. highlights curbs awkward shallow angles approach hazard: “decided enter gas station driveway east so path. 3?4 cm lip curb I entered angle ‘grabbed’ my tire me [fall]” (report #8765, regular cyclist, trip, no treatment, woman). Bollards, posts, meant manage flow instance, fixed bollards intended prevent onto multi-use hazardous “There pillars, visible bigger distance without any horizontal warnings signs. checking gear shifting few seconds suddenly hit them, falling completely uncontrolled” #4992, commute ER visit, man). Not surprisingly, showed consistent many previous Teschke Cripton Beck Several transient temporary surface. rocks, chunks concrete, gravel, slippery caused black ice leaves (Myhrmann Roadworks mentioned lost uneven slipped construction site materials: “I forced there's excessive works. My front metal plates laid ditches dug there. signs, tapes cones avoid hazards” #4302, treatment). illuminate often-overlooked experiences finding. One these commonly experienced (but rarely reported) misses Near BikeMaps.org. Past 2017), alarming Without doubt, captures outcome beyond Ideally, combined widespread analysis. But access, record consistency, need protect privacy make attributed outcomes. Through spatially, able contextualize clusters narratives. circles roundabouts approximately 6% vehicle-related entirely circle conflicts. describes struck rounding circle: cuts, scrapes strained wrist. driver vehicle alleged she couldn't see obstacles line sight” #367, family doctor Traffic residential traffic-calming strategy; cyclists (Harris 2013). Similarly, L, opening door (dooring) alley: “Struck oncoming Toyota Corolla attempted turn left alley Heather 10th Ave & Broadway. No explanation me” #4657, A strength leverage studying safety—self-reports. plentiful assess Linking helpful identification leads Further, prioritizing issues planners practitioners streamline upgrades interventions. limitation compare patterns exhibit selection available, demonstrates sources—insurance admissions—underestimate Evaluating depth circumstances aside occur Importantly, equity—either among users provision. Earlier indicates file tend per Likewise, investment concentrated densifying marked affluence (Firth paper occur. highlight characteristics, considering Self-reporting enhances professionals future investments.