Dissertation Title: "Building and Evaluating a Surveillance System for Bicycle Crashes and Injuries"
Chapter 1: For cities aiming to create a useful surveillance system for bicycle injuries, a common challenge is that city crash reporting is scattered, faulty, or non-existent. We document some of the lessons learned in helping the City of Boston to: 1) create a prototype for a comprehensive police crash database, 2) produce the city’s first Cyclist Safety report, 3) make crash data available to the public, and 4) generate policy recommendations for both specific roadside improvements and for sustainable changes to the police department’s crash reporting database. Some of the lessons include finding and using committed champions, prioritizing the use of existing data, creating opportunities to bridge divisions between stakeholders, partnering with local universities for assistance with advanced analytics, and using deliverables, such as a Cyclist Safety Report, to advocate for sustainability. After the release of the Report, the City implemented immediate interventions for injury prevention, including 1) furnishing over 1,800 taxis with stickers to prevent “dooring,” 2) adding pavement markings at trolley tracks to decrease the likelihood that cyclists would fall from getting their wheels lodged in the tracks, 3) conducting targeted enforcement of traffic laws, and 4) working directly with state and federal agencies to fund a more comprehensive surveillance system. As of February of 2016, nearly three years after its public release, the crash database has been viewed over 2,500 times and cited repeatedly by the media. Our hope is that the lessons we learned can help other cities do even better.
Chapter 2: Police reports are the main (and sometimes the only) source of crash data that injury prevention professionals (e.g., road engineers, urban planners, and public health practitioners) use to inform decisions about how to prevent bicycle crashes and injuries. Police report templates generally include a limited number of forced-choice fields and an open-ended narrative text field where the officer can describe the circumstances of the crash. There are currently no standardized guidelines for completing these text fields, especially with regards to reporting auto-related crashes. Several studies have demonstrated meaningful data could be extracted from accident narratives in large databases. The rapid development of text-mining tools has created an impetus to extract elements of interest that are not available as a forced-choice field (e.g., vehicle or bicycle turning maneuvers). However, the question is whether the text data are rich enough to generate information that is useful for designing and implementing appropriate injury prevention interventions. Using the Boston Police Bicycle Crash Database, the aim of this study was to determine situations (e.g., injury status of the bicyclist, time of day, type of road, etc.) under which the content of a police narrative text is more and less informative. I used a part-of-speech tagger in Python's Natural Language Toolkit (NLTK) to generate outcome variables that are considered linguistically rich in content. I then fit a linear regression of each outcome on the explanatory variables and found that when an involved party filed a report either in-person or by phone, the narrative contained significantly fewer words than when a police officer was present at the scene. Report length also was lower when crashes occurred mid-block vs. at an intersection, when crashes did not involve dooring, and when the crashes occurred at night. To further examine the value of mining the text narratives, a random sample of narratives were coded against a government-recommended bicycle crash analysis tool to determine the quality and utility of the data. I provide recommendations for improvements to the Boston Police Department's bicycle crash surveillance system.
Chapter 3: Understanding bicycle-vehicle collisions that result in hit-and-run (HAR) behavior is an important concern for law enforcement, public health, and affected individuals. If bicyclists are injured, this issue has implications for expedient access to medical care and for protection from the financial burden of associated medical costs. This study aimed to identify significant predictors of vehicle-bicycle HARs, the results of which can potentially inform preventive interventions for this type of injury and crime. Method: Data were collected from Boston Police Department bicycle crash reports for 2009-2012. The data identified whether a crash was a hit-and-run and other predictor variables including road and bicyclist characteristics. The probability of a HAR was fit to selected variables through logistic regression models. Effects of the predictors were reported as odds ratios. Results: Of the 1646 bike-vehicle collisions, 6% (n=93) resulted in a HAR and 80% (n=1309) involved an injury to the bicyclist. Controlling for all other variables, the odds of a HAR were not greater when the bicyclist was injured versus not injured or when the bicyclist was male versus female. The odds of a HAR were 2.40 (95% CI: 1.31, 4.23) times as likely when the vehicle was a taxi. The odds of a HAR were 1.65 (95% CI: 1.08, 2.54) times as likely during night as during daylight hours and 1.74 (95%: 1.07, 2.66) times as likely during the weekend versus during the week. The odds were 3.18 (95%: 1.29, 11.51) times as likely when the incident did not involve a dooring. The interactions of male-by-injured, taxi-by-injured, and night-by-weekend were nonsignificant. Conclusion: The probability of a hit-and-run partially depends on time, day of the week, and whether the vehicle type was a taxi. We discuss implications for policies and interventions aimed at preventing this type of collision and crime.