CHAPTER 7

Video-Based Conflict Analysis

The video-based surrogate safety analysis, or conflict analysis, examines conflicts via surrogate safety measures at a further targeted scale to provide additional insight into the dynamics between bicycle safety and known risk factors like turning volumes, speed, and design features. This chapter presents the findings from the conflict analysis, including site selection, video data collection, conflict extraction and results.

Site Selection

The goal of the conflict analysis site selection process was to identify 25-30 sites, with: five to six locations per treatment type, a mix of traffic intensities (medium, high) and locations with higher bicycle volumes and crash history to increase the likelihood of recordable conflicts.

For conflict analysis site selection, the primary consideration was for traffic intensity and recorded or estimated AADT. Recorded AADT came from local jurisdiction volume data for specific segments immediately entering or exiting the intersection on the approach street. Estimated AADT were derived by averaging across the entire city AADT dataset, assigned by functional class of the roadway. AADTs were categorized⁵ as:

• High - above 15,000
• Medium - 9,000 to 15,000
• Low - below 9,000

To select the candidate locations for video-based conflict analysis, the entire set of sites selected for the micro crash analysis were reviewed and filtered by type of treatment type (e.g., bike lane, pocket bike lane, mixing zone, separated bike lane, protected intersection). The available list of sites were reviewed and sites with high-quality designs and feasible camera mounting locations were prioritized for selection. At each location, the cameras were mounted at a certain height and angle in consultation with the video processing vendor to facilitate good views for optimal data extraction and minimize occlusion. Several locations in each city were included to incorporate geographic diversity. The team intentionally did not select sites with known low bicycle volumes due to the concern about not being able to record enough motor vehicle - bicycle interactions for analysis.

Additional sites were added to the list of sites that were selected for crash analysis to ensure that we have enough locations with high-quality design for each design treatment type. For example, most of the sites initially identified for offset /protected intersection locations were in New York, and generally were on the

CONVENTIONAL BIKE LANES

POCKET OR KEYHOLE BICYCLE LANE

SEPARATED BIKE LANE

MIXING ZONES

OFFSET / PROTECTED

Video Data Collection

Once the locations were selected for the conflict analysis, they were shared with two different vendors; one contracted to assist with the data collection and the other for the automated conflict data extraction process. The data collection vendor coordinated with the data extraction vendor to optimize the camera placement at each of the sites. A map of the locations was generated, and the research team identified the conflict areas for each site to be recorded. Potential camera placements and mounting specifications were identified using Google Street View and satellite images and further refined to maximize the conflict area view at each site.

The goal of the data collection process was to collect at least five days of video, with 12 hours of video per day (i.e., 7 am - 7 pm), resulting in approximately 60 hours of video collected per site. However, due to camera malfunctions at many sites, video data was recollected resulting in more than 60 hours of video data being available. This was particularly relevant at the Austin and Seattle locations. Only one location (Yesler Way and 12^th Ave) in Seattle had 28 hours of usable video data. Table 41 shows the dates and time of the video data collection.

Table 40. Conflict analysis sites and key characteristics

(* AADT is estimated)

Table 41. Video data collection hours

Conflict Identification

After the videos were collected at each site, they were processed using automated video processing methods by a vendor selected for this task. For each selected site, the research team worked with the vendor to identify the locations where the design required motor vehicles and bicycles to interact or cross paths. In general, the locations for the conflict areas are shown in Figure 12.

At each site, the vendor identified conflicts from the video using algorithms that track each user’s trajectory in the time-space plane. The measure used to identify conflicts was the PET, a surrogate safety metric. PET was originally introduced by Allen et al. (Allen et al., 1978) and is defined as “the time between the moment when the first road user leaves the path of the second and the moment when the second reaches the path of the first” (Johnsson et al., 2018). Thus, PET essentially measures how close in time the road users are to each other when occupying the same space (Russo et al., 2022). Lower values of PET indicate the potential for more severe conflicts.

PET was chosen as the surrogate safety metric of choice as opposed to TTC because the road users do not have to be on a collision course to calculate PET and it is easier to measure by defining the conflict point and recording time stamps for the different road users.

The vendors algorithm identified all interactions with PETs less than or equal to 10 seconds for the movement of interest at each location. The raw data provided by the vendor included two files. The safety events file included:

• the observed PETs,
• the date and time of each observed conflict,
• type of road users,
• which road user arrived first,
• median and conflict speed of the two road users involved in the conflict, and
• the type of conflict.

The road user’s data file included all observed road users (including those that were not involved in conflicts) along with their median speed and location. In addition to these files, videos for all conflicts with PETs less than or equal to 2 seconds were provided on the web portal for review. Figure 13 shows a sample visual out from the software mapping the trajectories onto world space. For each site in the video data analysis, similar maps are shown for all locations in the Appendix E: Conflict Analysis Site by Site Graphics.

Data Cleaning and Preparation

To prepare the data for modeling and analysis, filters were applied to the dataset. Figure 14 presents a flowchart of the data cleaning process. As mentioned, the software identified all PETs <= 10 secs. A total of 47,112 interactions were included. First, all interactions with PETs > 5 seconds were filtered out, leaving 33,524 interactions with PETs >= 5 seconds. Next, observations were removed from analysis as follows:

• 1,158 observations due to violations (bicycles traveling the wrong way) and vehicle turning movements that were not of relevance to the study.
• 79 pedestrian conflicts and vulnerable road user (VRU), which were classified as VRU’s when they were not able to be classified as one of the other user types by the automated software) and motorized vehicle observations.
• 61 bicycle-to-bicycle conflicts.
• 10 erroneous speed observations were removed, including a set of bicycle speeds of 186.41 mi/hr, a vehicle speed of 186.41 mi/h, and a set of bicycle speeds of 80 mi/h.
• 16,109 duplicate observations; specifically, if Road User I ID, Road User II ID and the timestamp were all the same.

This filtering and removal left 16,107 observations for analysis. Of those, there were 15,561 conflicts in the same direction (i.e., right- or left-turning vehicle on the approach conflicting with a through bike on the same approach), and 546 conflicts were observed in the opposing direction (i.e., a right- or left-turning vehicle from the opposing direction conflicting with the bicycle on the approach of interest).

Conflict Thresholds and Severity

PET is the time gap between two intersecting trajectories crossing the same space. However, it is helpful to further refine and categorize the PETs to better describe the interactions. One approach is to define thresholds for the value of PET. Previous research for bicycle-motor vehicle interactions (Zangenehpour et al., 2016, Russo et al., 2020, 2022) has suggested threshold levels are classified as PET ≤ 1.5 sec, 1.5 < PET ≤ 3.0 and 3.0 < PET ≤ 5.0. We have used these thresholds in the descriptive analyses and in filtering the conflicts for modeling purposes. The lower PETs values generally indicate a closer interaction between the two users.

However, PET alone can be insufficient to describe the interaction as a conflict. In the State of the Practice literature review, alterative surrogate safety measures were reviewed that include some characterization or use of speed (e.g. TTC). The vendor’s algorithm is only trained to produce PET observations so any of these other measures are not available. Intuitively, the potential for a conflict measured by PET threshold can be further informed by the vehicle speed in the conflict area. We can classify the conflicts based on vehicle speed and PETs as a proxy for severity potential. Previous research suggests that a threshold value of 9 mi/h is appropriate to classify high and low vehicle speeds with respect to bicycle-motor vehicle crashes/conflicts (Jurewicz, Sobhani, Woolley, Dutschke, & Corben, 2016; Russo et al., 2022,).

In the remainder of this chapter, we apply the severity label for observations with the following combination of PET and vehicles speed in the conflict area:

• High Severity (PET ≤ 1.5 sec and vehicle speed at conflict greater than 9 mi/h).
• Medium Severity (1.5 sec < PET ≤ 3.5 sec and vehicle speed at conflict greater than 9 mi/h).
• Low Severity (3.5 < PET ≤ 5 sec and vehicle speed at conflict greater than 9 mi/h).

Conflict Analysis Methods

To further explore the observed conflicts, conflict frequency and CS models were estimated. These are further described in the following sections. Table 42 summarizes which CS was were used in the two modeling approaches.

Count-Data Models for Conflict Frequency

Using the safety events files and the road user movement files provided via Transoft, a series of conflict frequency and CS models were fit. Each row in the road user movement file represented one road user. Other variables that were included in the road user movement file included:

• date,
• ID,
• entry and exit time of the road user in the conflict area,
• road user type,
• median speed in mph, and
• direction of movement.

For conflict frequency, the road user movement files were used to determine hourly volumes during the data collection period, where the number of observations observed for a given hour was used as the hourly volume. Hourly volumes were determined for bicycle and approach turn movements.

Each row in the safety event file represented a conflict between two road users. Other variables in this file included:

• observed PET value,
• date and time of the conflict,
• which road user arrived first at the conflict area,
• details of each road user (ID, movement direction and type),
• conflict and median speed of each road user in miles per hour, and
• the movement scenario (i.e., movement direction of the two conflicting road users).

The safety events files were used to determine the number of conflicts for a given hour. The number of conflicts in an hour is the conflict frequency that was analyzed. Many modeling frameworks and data aggregations were explored. These data were joined to create a single file based on location, date and hour for modeling purposes. For CS, only the safety events files were used and no aggregation was required.

The number of conflicts in a given hour is represented by a non-negative integer value, or a count; as such, the appropriate class of models are count-data models. One challenge with using PET with bicycle-motor vehicle conflicts is that some conflicts that are identified only by low PET times are more similar to the expected user interactions and yielding than a pre-cursor to a crash. Initially, models of all conflicts were explored but the results were difficult to interpret due to the large numbers of observations with PETs < 5 secs. There was greatest interest in the most crash-potential events from a safety perspective. Thus, speed of the vehicle and low PET times were combined and only severe conflicts PET less than or equal to 1.5 sec and vehicle speed greater than 9 mi/h were modeled. The most common count-data model is the Poisson model. In the Poisson model, y_i is drawn from a Poisson population with parameter λ_i, where the model is represented as (Greene, 2018):

where P(y_i) is the probability of hour i having y_i conflicts and λ_i is the Poisson parameter for hour i. The Poisson parameter is also equal to the expected number of conflicts, E[y_i], in hour i. The Poisson model is estimated by specifying the Poisson parameter as a function of explanatory variables (e.g., bicycle volume, approach turning volume), which then takes the form of a log-linear model:

or equally:

where X_i is a vector of explanatory variables and β is vector of parameters to be estimated. The expected number of conflicts in hour i conditional on X_i is then given by:

Despite the Poisson model being a non-linear regression model, it is easier to estimate parameters using maximum likelihood, where parameters are estimated as:

The Poisson model, however, suffers from a distinct limitation: it is unable to handle over- or under-dispersion. The Poisson model, and the Poisson distribution, operate under the assumption that the mean and variance are equal, E[y_i] = Var [y_i]. If this equality, or assumption, is found to be untrue, the data being analyzed is considered under-dispersed (E[y_i] > Var[y_i]) or over-dispersed (E[y_i] < Var[y_i]). If under-or over-dispersion is present and actions not taken to correct for it, parameter estimates will no longer be unbiased and standard errors will be incorrect. The determination of dispersion can be computed manually but is also a common output from models that account and correct for dispersion, such as the negative binomial model, through the estimation of a dispersion parameter. If the estimated dispersion parameter is statistically significant, the Poisson model is not appropriate. It was determined that for each of the conflict frequency models, the dispersion parameter was not statistically significant (i.e., no significant dispersion); therefore, Poisson models were used.

Discrete Choice Models for Conflict Severity

Two severity analyses were conducted, the first of which considers three severity levels:

• high severity (PET ≤ 1.5 sec and vehicle speed greater than 9 mi/h)
• medium severity (1.5 sec < PET ≤ 3.5 sec and vehicle speed greater than 9 mi/h)
• low severity (3.5 < PET ≤ 5 sec and vehicle speed greater than 9 mi/h).

The second consists of analyzing high severity conflicts alone, such that 1 if high severity conflict (PET ≤ 1.5 sec and vehicle speed greater than 9 mi/h), and 0 otherwise. As defined previously, in addition to considering these PET ranges, vehicle speed was also considered. To summarize, this resulted in the following severity categories:

• High Severity (PET ≤ 1.5 sec and vehicle speed at conflict greater than 9 mi/h).
• Medium Severity (1.5 sec < PET ≤ 3.5 sec and vehicle speed at conflict greater than 9 mi/h).
• Low Severity (3.5 < PET ≤ 5 sec and vehicle speed at conflict greater than 9 mi/h).

Ordered Probit Models

Based on the ordered nature of severity (low severity to high severity), an ordered probit model was chosen to model CS when considering all three severities. In the ordered probit model, an unobservable variable y* is defined as a linear function (Greene, 2018):

where X is a vector of explanatory variables (e.g., vehicle category, vehicle speed, user arrival), β is a vector of parameters to be estimated, and ε is an independently randomly distributed disturbance term with mean 0 and variance 1. Utilizing Eq. (10), conflict severities can be represented as observable (Greene, 2018):

where µ are the thresholds used to define the ordered conflict severities. During model estimation, µ is estimated simultaneously with β, where integer ordering is used such that J is the highest integer order (i.e., high severity conflict). To determine the probabilities of J, ε is assumed to be normally distributed (mean 0 and variance 1) such that the ordered probabilities are (Greene, 2018):

with:

where y = 2 is high severity, y = 1 is medium severity, y = 0 is low severity, and Eq. (9) shows the required relationship for the threshold parameters to ensure all probabilities are positive.

To assess the impact of explanatory variable X on severity probability, marginal effects are computed. Marginal effects represent the change in probability due to a one-unit increase in explanatory variable X. For continuous explanatory variables, marginal effects are determined as (Greene, 2018):

while for indicator variables:

where ${\ddot{X}}_{X_k}$ is the mean of all other variables (i.e., variables are held constant) while indicator X_k changes from zero to one. Ordered probit marginal effects have an inherent difficulty in interpreting interior orders (in this study, y = 1). This results in the analyst not knowing for certain what impact a positive or negative parameter has on this probability, only that it is positive or negative (Washington, Karlaftis, Mannering, & Anastasopoulos, 2020). Therefore, marginal effects are interpreted at the outer orders (high severity and low severity).

Binary Logit Models

For modeling the likelihood of a severe conflict (1 if high severity conflict, 0 otherwise), a binary logit model framework was applied. The binary logit framework follows the form of the logit probability where one of the two outcomes is normalized to zero (Greene, 2018):

where P_n(i) is the probability of observation n having outcome i (high severity conflict), β is a vector of parameters to be estimated, X is a vector of explanatory variables, and ε_in is a Type I Extreme Value distributed error term. The model is interpreted via marginal effects, where for continuous variables:

where $\frac{\partial P_n(i)}{\partial X_{ink}}$ is the derivative of the probability of observation n having CS i. For indicator variables, marginal effects are computed as shown in Eq. (11).

Parameter Transferability

In this work, disaggregate data models by treatment type were applied; that is, a separate conflict frequency and CS model was fit for each treatment type. To determine if separate models by treatment type are appropriate, the following log-likelihood ratio test is applied (Washington et al., 2020):

where LL $\left({\beta }_{t_{b_{t_a}}}\right)$ is the log-likelihood at converge of a model using parameters from the treatment b model on the data for treatment α, LL(β_α) is the log-likelihood at convergence of the treatment α model, and χ² is a chi-square statistics with degrees of freedom equal to the number of estimated parameters in ${\beta }_{t_{b_{t_a}}}$. If χ² results in statistical significance, the null hypothesis that all treatments should be modeled together in a single model is rejected in favor for the alternative hypothesis that each treatment type should consist of its own model.

Conflict Analysis Results

This section presents the descriptive statistics, conflict frequency and conflict analysis models.

Descriptive Statistics

Table 43 shows the frequency and proportion of conflicts by treatment type where the PET <=5 sec, first presenting it regardless of arrival and then by if the vehicle or bicycle arrived first at the conflict location.

Overall, the highest number of conflicts (10,151) were observed at the protected intersection locations (63% of the total), followed by mixing zone locations (3,201 and 19.9%). This is not surprising, as six of the eight protected intersection sites and three out of five mixing zone sites were in New York City, where the bicycle and vehicle volumes are higher compared to other locations. The lowest number of observed conflicts were observed at the pocket bike lane locations (584, 3.6% of total).

Tabel 44 shows the frequency and proportion of high severity conflicts (PET ≤ 1.5 sec) by treatment type. The proportion among conventional bike lanes remains consistent, while other treatment types have

Table 45 shows the frequency of conflicts and exposure by treatment type, location and exposure (i.e., number of through bicycles, number of approach turns and number of oncoming turns). The lowest frequency of conflicts was observed at the pocket or keyhole bike lane locations (n=584), while the highest frequency was observed at the offset/protected intersection locations (n=10,151). The offset/protected intersection sites were primarily located in New York City where observed bicycle volumes were much higher compared to locations at other treatment types as seen in Table 45.

Table 45. Total Frequency of Conflicts and Exposure Measures by treatment type and location

Table 46 shows the frequency of conflicts by city and treatment type. The highest number of conflicts were observed in New York City (79.4%) at the protected intersection and mixing zone locations at least partly due to higher volumes. Seattle had the second highest proportion of conflicts (9.5%), followed by Minneapolis (5.7%) and Austin (5.4%).

Table 47 shows a distribution of conflicts based on PET thresholds (PET ≤ 1.5 sec, 1.5 < PET ≤ 3.0, and 3.0 < PET ≤ 5.0. Mean PETs by treatment type are also shown in Table 47 with the values ranging between 2.66 sec for mixing zones and 3.17 sec for separated bike lane locations. Of all the conflicts observed across the treatment types, the highest proportion were conflicts where PET was between 1.5 and 3.0 sec (47.0%), followed by conflicts where PET was between 3.0 and 5.0 sec (34.5%). Overall, 18.5% of all conflicts had PETs less than or equal to 1.5 sec.

Within each treatment type also, higher proportions of conflicts were observed with PETs greater than 1.5 sec and less than or equal to 5 sec. For conflicts with PETs less than or equal to 1.5 sec, the largest proportion was observed at the mixing zone locations (24.4%) followed by pocket or keyhole bike lane locations (18.3%), with the separated bike locations showing the lowest proportion (9.1%). Figure 15 shows a plot of the mean PET by intersection treatment type, with separated bike lane facilities having the highest and mixing zones the lowest observed mean PET.

Table 48 shows the distribution of conflicts by location, city, treatment type and PET. Across all treatment types, the highest proportion of conflicts were those with PETs greater than 1.5 sec, which have lower propensity for crashes. The highest proportion of conflicts with PETs less than or equal to 1.5 sec were observed at the mixing zone locations (24.4%), followed by pocket/keyhole bike lane locations (18.3%).

In addition to PETs, vehicle speeds were also extracted during the video processing. Vehicle speed is a key contributor to crash severity. Table 49 shows the vehicle speeds descriptive statistics at conflict areas by treatment type and location. Across cities, average turning vehicle speeds at the pocket/keyhole bike

Table 48. Number of Conflicts by Location, City, Treatment Type and PET

Table 49. Vehicle Speed at Conflict Point by Treatment Type and Location (mi/h)

Table 50 shows the distribution of conflicts by treatment type, severity and speed. Conflicts with PETs < 1.5 sec and high vehicle speeds are likely to be the most severe and within this category, protected intersection locations had the lowest proportion (4.9%), followed by the separated bike lane locations (5.5%). Figure 19 shows the observed proportions of conflicts by severity and treatment type. For each treatment type, the figure shows the percentage of each PET and vehicle speed category that is shown in the bottom half of Table 50. The stacked bar figure allows visual comparison of the distribution of the PET and vehicle speed combinations.

Parameter Transferability Results

Table 51 and Table 52 show the parameter transferability results for the conflict frequency models and the CS models. For the severity models, the binary logit specifications were used for the transferability test. The tables show the chi-square statistic and degrees of freedom, each which indicate with at least 90% confidence that separate treatment type models are the appropriate approach; that is, the null hypothesis that parameters are not different across model types is rejected. Based on the results of this test, conflict frequency and CS results are presented and discussed by treatment type model.

Conflict Frequency Analysis Results

As a reminder, only severe conflicts with PET less than or equal to 1.5 sec and vehicle speed greater than 9 mi/h were modeled in the frequency count models.

Conflict frequency models by treatment type are shown in Table 53 to Table 56, and CURE plots for each model are shown in Figure 20. Both Poisson and negative binomial models were fit for each model, where no significant dispersion was found; therefore, all treatment type models are Poisson. The CURE plots show the model residuals bounded by 2 standard deviations and is a typical way crash and conflict frequency models are assessed for fit. The residual line should track within the bounds of the standard deviations, and, overall, indicate good fit. The plots do show that the model fits are not as good as the exposure measures increase.

The individual estimates for each treatment type shown in the model specification tables are the estimates for hourly through bicycle volume, hourly vehicle turn volume on approach and site indicators. Also shown in the model specification tables are the incidence rate ratios, or IRRs. An IRR represents the expected change in conflict frequency as a multiplicative factor and is determined by exponentiating the parameter estimate; that is, e^β. For example, the IRR for Yesler Way and 12^th Ave indicate that the expected conflict frequency at this location is expected to increase by a factor of 4.085. Alternately, IRRs can be interpreted as a percent increase in the expected number of conflicts as e^β — 1. In each model, all parameters for hourly through bicycle volume and hourly vehicle turn volume on approach are significant with at least 90% confidence, while parameter significance for site indicators varies. The site indicators, regardless of significance, were included in final model specifications to control for site-level differences.

Figure 21 summarizes the IRR for the exposure measures (bicycle volume and same direction motor vehicle turning volumes). For all five treatment types, the contribution of the same direction motor vehicle

Conflict Severity Analysis Results

Ordered Probit Model for High, Low and Medium Severity Conflicts

Ordered probit model specifications for CS, and corresponding marginal effects, are shown in Table 57 to Table 66.

To provide an example of how to interpret the marginal effects, refer to Table 62. The marginal effect for continuous explanatory variable vehicle speed (mi/h) is 0.0045, indicating that for a one-unit increase in speed (1 mi/h), there is a 0.0045 higher probability of the conflict being severe. The marginal effect for the indicator variable for if the vehicle arrived first is 0.0927, indicating if the vehicle arrived first there was a 0.0927 higher probability of a high severity conflict.

As part of the modeling efforts, we explored all data that were available. We attempted to include geometric and other variables in the modeling process but none of these variables were significant. This is primarily because the variation between sites on these variables is minimal due to the intentional site selection process. As a proxy, we included site level indicator variables, some of which were significant. In the text we explain the characteristics of these sites for context, but these should not be interpreted as the cause or reason for the severity differences in the model estimates.

Conventional Bike Lanes

Three parameters were found to be statistically significant in the bike lane severity model (see Table 57 and Table 58). The first of these was for the indicator for articulated truck, box truck, or bus, where marginal effects indicate a 0.1263 higher probability of a severe conflict if the conflict was with an articulated truck, box truck, or bus. Also significant were two parameters for site indicators. Marginal effects suggest that if a conflict occurred at Yesler Way and 12^th Ave (Seattle) there was a 0.1448 higher probability of a severe conflict. There are no bike lane markings at this location, three motor vehicle cross street lanes, two-way motor vehicle traffic, no bus stop, no left-turn lane or signal and a speed limit of 25 mi/h. Also with a significant parameter, marginal effects indicate that if a conflict occurred at 28^th St E and Park Ave S (Minneapolis-St. Paul) there was a 0.0787 higher probability of a severe conflict. This site also has no bike lane markings, three motor vehicle cross street lanes, no bus stop, no left-turn lane or signal and a speed limit of 25 mi/h. Conflicts that occurred at Dean Keeton St and Guadalupe St (Austin) had a similar probability of resulting in a severe conflict; however, this site has full-green bike lane markings, an exclusive left-turn lane with a 5-head signal and turn arrow and a speed limit of 35 mi/h – all other characteristics were the same.

Pocket/Keyhole Bike Lanes

In the pocket bike lane model, just one parameter was statistically significant, as well as the constant. Marginal effects suggest that if a conflict occurred with an articulated truck, box truck, or bus, there was a 0.0887 higher probably of a severe conflict (Table 59 and Table 60). Although all sites in the pocket/keyhole bike lane model have a negative coefficient and decrease the probability of a severe conflict (relative to the base case of 7th Ave S and Dearborn St in Austin), the parameters for these indicators were not significant.

Separated Bike Lanes

There were six significant parameters in the separated bike lane model. Unlike the previous two models, marginal effects suggest that if a conflict occurred with an articulated truck, box truck, or bus, there was a 0.0432 decrease in probability of a severe conflict (see Table 61 and Table 62). If the conflict occurred with the bicycle and vehicle traveling in the same direction (i.e., vehicle was turning on approach), marginal effects indicate a 0.0383 higher probability of a severe conflict. The parameter for if the vehicle arrived first was highly significant, where marginal effects suggest a 0.0927 higher probability of a severe conflict if the vehicle arrived first. Lastly, despite all four sites having similar effects on probability, three had significant parameters. Marginal effects indicate that if a conflict occurred at Bouldin Ave and Barton Springs Rd (Austin), there was a 0.0399 lower probability of a severe conflict. This location has no bike lane markings, one motor vehicle cross street lane, two-way motor vehicle traffic, no bus stops, a left-turn lane with a 3-head ball and a speed limit of 45 mi/h. If a conflict occurred at 7th Ave S and Stewart St (Seattle), marginal effects suggest a 0.0391 lower probability of the conflict being severe. This location differs from the previous, as there are full-green bike lane markings, four motor vehicle cross street lanes, one-way motor vehicle traffic, no bus stop, no left-turn lane or signal and a speed limit of 25 mi/h. The final location with a significant parameter was Congress Ave and Woodward St (Austin), where if a conflict occurred here marginal effects indicate a 0.0450 lower probability of a severe conflict. This location has full-white bike lane markings, four motor vehicle cross street lanes, two-way motor vehicle traffic, a far-side bus stop, an exclusive left-turn lane with a 5-head signal and turn arrow and a speed limit of 45 mi/h.

Mixing Zones

The mixing zone model had five significant parameters. Once more, the indicator for articulated truck, box truck, or bus had a significant parameter. According to marginal effects, if the conflict occurred with an articulated truck, box truck, or bus, there was a 0.0663 higher probability of a severe conflict (see Table 63 and Table 64). If the vehicle arrived first, there was a 0.0658 higher probability of a severe conflict, based on marginal effects. The first site with a significant parameter was 6^th Ave and W 21^st St (New York), where marginal effects indicate a 0.1122 higher probability of a severe conflict. Although estimates indicate an increase in probability of a severe conflict, this site has full-white bike lane markings, one motor vehicle cross street lane, one-way motor vehicle traffic, a near-side bus stop, no left-turn lane or signal, and a speed limit of 25 mi/h. Also significant was the parameter for E 37^th St and 1^st Ave (New York), which has the same characteristics and nearly the same effects on probability of a severe conflict. The final site with a significant parameter was Linden Blvd and Fountain Ave (New York), where marginal effects suggest that if a conflict occurred here there was a 0.2044 higher probability of the conflict being severe. This site has full-white bike lane markings, 10 motor vehicle cross street lanes, two-way motor vehicle traffic, no bus stop, a left-turn lane with a 3-head ball signal and a speed limit of 25 mi/h. This site was also the only mixing zone site in which the mixing zone marking was determined to be minimal.

Protected Intersections

Three variables had significant parameters in the protected intersection model. Once more, the parameter for articulated truck, box truck, or bus was found to be significant (see Table 65 and Table 60). If the conflict occurred with an articulated truck, box truck, or bus, marginal effects indicate a 0.0326 higher probability of a severe conflict. Also observed once more is the significance of the parameter for if the vehicle arrived first, where marginal effects suggest a 0.0975 higher probability of a severe conflict if the vehicle arrived first. Just one site had a significant parameter, 1st Ave and E 11th St (New York), in which marginal effects indicate a 0.0790 higher probability of a severe conflict if a conflict occurred at this location. This site has full-green bike lane markings, one motor vehicle cross street lane, one-way motor vehicle traffic, no bus stops, no left-turn lane or signal and a speed limit of 25 mi/h.

Binary Logit Model for High Severity Conflicts

To model only high severity conflicts (PET ≤ 1.5 sec and vehicle speed greater than 9 mi/h, 0 otherwise), a series of binary logit models were generated. Despite similarities, there were differences observed as well (compared to the ordered probit models). The binary logit models were a bit more sensitive to variables in which a large proportion of severe conflicts, or otherwise, were present.

Conventional Bike Lanes

Two parameters were significant in the bike lane model, while a third was just inside the fail to reject region (i.e., test statistic of 1.64, see Table 67). The indicator for the vehicle arriving first had a significant parameter, where marginal effects suggest a 0.100 higher probability of the conflict being severe if the vehicle arrived first. Yesler Way and 12th Ave (Seattle), also significant in the ordered probit framework, had a significant parameter and marginal effects that indicate a 0.189 higher probability of the conflict being severe if it occurred at this location. There are no bike lane markings at this location, three motor vehicle cross street lanes, two-way motor vehicle traffic, no bus stop, no left-turn lane or signal, and a speed limit of 25 mi/h. Just inside the fail to reject region was the indicator for Dean Keeton St and Guadalupe St (Austin), which was found to be significant through the ordered probit framework. In the current