choose to bicycle on all facilities, there is a potential for omitted bias in the crash data. In other words, less experienced or risk adverse bicyclists may avoid higher speed roadways while a wider array of people may choose to bicycle on lower speed facilities, and this bias may impact safety analysis results in ways that are difficult to measure.

Video-Based Conflict Analysis

Conflicts, or interactions between vehicles, occur much more frequently than crashes. Using automated video-based conflict analysis gathers robust detail about each bicycle-motor vehicle interaction including speeds, trajectories and information about all vehicles’ paths. There is a reasonable body of research that correlates conflicts and other surrogate measures of safety with actual crashes (i.e., that observing conflicts is useful in understanding the actual reported crash performance). The research, however, is less developed for bicycle-motor vehicle crashes. It also is reasonable to assume that fewer conflicts would be associated with higher levels of comfort for road users, including people on bikes, and it is useful to know the rate and severity of conflicts, though this has not yet been validated by research. Lastly, the comment on omitted bias also applies to the conflict analysis.

Human Factors Study (Simulator)

Both the crash and video-based conflict studies are observational –what has been designed and built and how existing traffic and drivers use the roadway can only be observed. In the driving simulator, the experimental environment can be completely controlled and the variation in outcome is in the driver’s performance within the same situation. In addition, significant and detailed information about each driver’s performance can be collected, however, there is a limitation on the number of people that can be included in the experiment (typical experimental cohorts are between 30 and 40 people. While many measures have been validated to real world conditions and is directly related to safety (e.g., driver speed selection), not all data maps directly to safety outcomes.

Site Selection

This section describes the process used to select sites (i.e., intersection approaches) for the micro crash analysis and video-based conflict analysis. The selection process for the micro crash analysis involved consolidating data from several sources (data provided by cities, readily available data from Open Streets Map and detailed desktop data collected via Google Street View). The sites selected for the conflict analysis are a subset of the sites for the micro analysis. Sites were selected from four cities:

• Austin, TX
• Minneapolis, MN
• New York City, NY
• Seattle, WA

Intersection treatment Types

Based on the results of State of the Practice review, the site selection focused on five different intersection treatment types. Each of the treatments is listed below with a definition. Definitions for these treatments have been sourced primarily from the FHWA Separated Bike Lane Planning and Design Guide (FHWA 2015) and the upcoming AASHTO Bike Guide (AASHTO, forthcoming).

treatment, presence of bicycle signal). Starting with this output from candidate intersection locations, the project team manually conducted a desktop audit of sites to identify individual bicycle approaches, remove locations that do not have appropriate designs for the study and to collect information on key design and contextual data for each approach. The following factors were considered:

• Each bicycle approach that matched one of the design types was recorded as a candidate approach.
• If an intersection had no bikeway facilities on any approach it was removed.
• Candidate approaches might be excluded for other reasons, including if the approach bikeway for the bicycle is a shared lane (i.e., not a bike lane or separated bike lane), or a two-way separated bikeway.

In addition to the intersection design type, other information collected in the desktop audit included information on the intersection characteristics, such as the number of legs and angle of approach. Most information was specific to each approach (and progression through the intersection), including:

• The approach direction
• Presence of bike lane markings through the intersection
• Type of bikeway on approach
• Presence of a bus stops
• Bike signal
• Opposing left turns lanes (and their signal type)

Data was also collected on the number and types of lanes on approach, and the number and direction of lanes on the cross street. Finally, information using the time capsule feature on Street View to identify the month and year of the earliest captured Street View with the current facility layout, and the month and year of the latest captured Street View with the prior layout were also collected. Each intersection and approach received a unique identifier. Each approach was geo-coded to a specific OSM segment. Data collected in the desktop audit are shown in Appendix C.

Refinements and Exclusions

Once the initial desktop data collection was complete, the entire list of sites was reviewed more closely. The objective was to develop a list of candidate sites for the micro-crash and conflict analysis that did not have unusual features that would confound the safety analysis. Thus, intersections with the following criteria were excluded:

• With five or more legs
• Locations with channelized turn lanes
• Approaches that cross nine or more lanes
• Locations where the current design was implemented January 2019 or later were excluded because of the unavailability of after crash data at the time data were collected

Locations with large skews were generally excluded for the conventional bike lane and pocket bike lane locations due to the availability of sufficient non-skew locations for these intersection treatment types. Given a relatively small sample size for mixing zones, separated bike lanes and offset / protected locations, the team reviewed locations that were skewed and determined, that there were few that could be retained. Finally, filters were created to more easily identify the following characteristics:

• Offset / protected intersections were coded based on whether they had at least 6 feet of offset or not.
• Mixing zones were coded to indicate whether they had full mixing zone markings (turn arrow, bike stencil markings and / or shared turn lane signage) and were distinguished from those where the mixing zone markings was not explicit.
• Pocket lane designs were coded to distinguish between those wherein the turn lane was wide enough to fully fit a motor vehicle and in which roadway markings indicate that bikes should cross over prior to the turn lane rather than designs that were not explicit.