CHAPTER 6. SAFETY EVALUATION OF BICYCLE TREATMENTS

6.1 METHODOLOGICAL APPROACH

In true experimental studies, comparable sites of similar traffic volume and geometric features are randomly assigned to a treatment or a nontreatment group. The treatment is then implemented at the sites in the treatment group, and crash and traffic volume data are obtained before and after implementing the treatment. Although these true experimental studies can minimize threats to the validity of studies including regression to the mean bias, randomization is almost never possible in reality, and true experiments such as randomized control group experiments are not feasible in safety evaluations. Available data in safety evaluations of on-street bicycle designs in this project were also observational.

6.1.1 Study Design

Various methodological choices are possible in terms of both study design and analysis for the safety evaluation of different treatments. Two types of data are frequently encountered in safety studies: before-after (B/A) and cross-sectional (C/S). Although before and after data are preferred, many jurisdictions do not inventory installation dates, and finding this information for a large sample of data is not feasible. Cross-sectional study design is therefore implemented in the absence of before and after data. Cross-sectional studies use statistical modeling techniques that consider the crash experience of sites with and without a particular treatment of interest (a bikeway facility in this case). In cross-sectional studies, analysts develop CMFs using the crash frequency of the treated and the comparison sites. The main drawback of cross-sectional analysis is the selection bias. The treated sites tend to experience a high number of crashes compared to the comparison sites. This implies that, even if the crash frequency at a treated site is lower after the treatment, the number of crashes will continue being higher compared to the number of crashes at the comparison sites. Therefore, comparing the crash frequency at the two sites might yield biased results.

One of the methods used to overcome selection bias in cross-sectional studies is the propensity score matching and weighting method. A propensity score represents the conditional probability of a facility receiving a treatment given the covariates and outcomes. It shows the relationship between treatment status (1=treated, 0=control) and covariates (i.e., variables that completely or partially account for the apparent association between an outcome and risk factor). Using this approach, the probability of treated and control sites being selected for treatment is nearly the same, and both treated and control sites have similar characteristics. The approach mimics

random selection and helps to avoid selection bias in the safety database. Using this method, the sites with treatment are matched to the sites without treatment (control sites) based on the propensity score. The propensity score, which imitates the random selection principle and indicates the probability of a site being selected, is formulated as follows:

Where:

• P(T_i|X_i) is the propensity score denoting the probability of the site i receiving the treatment T.
• T_i is the treatment status of the site i, which takes binary values {0, 1}.
• X_i is a vector of covariates that vary with the treatment.
• α_i is the vector of coefficients through the binary logistic regression.

The propensity scores of the treatment and control sites are selected such that both groups have similar key characteristics. Propensity scores can be estimated using several parametric and nonparametric tools. In this project, the research team used logistic regression models to select the covariates and estimate propensity scores. The response variable was a binary variable with two potential values: 1=treated (i.e., bike lanes, buffered bike lane, and separated bike lane), and 0=control. Data matching was performed by randomly selecting the control sites. Additionally, site replacements were allowed to prevent the same control site from being selected multiple times, allowing for better matching. After matching the data based on the selected covariates, we conducted a quality check of the matched dataset and implement negative binomial models to evaluate the safety effectiveness of bicycle lanes.

6.1.2 Sample Size and Power Estimations

The sample size required to achieve significant results depends on the study design and the statistical model employed, as well as the anticipated effect of the estimate (the CMF, in this study). Table 56 shows the required sample sizes (number of midblock sites) to implement a cross-sectional design and apply generalized linear models (e.g., Poisson, overdispersed Poisson). The table shows the required sample sizes to ensure that the estimated CMF is statistically significant. For example, if the estimated CMF for the baseline crash rate of 1 (crashes/bicyclist volume) is found to be equal to 0.4 (60 percent decrease), then we would require 17 sites to ensure that the estimated CMF is significant (i.e., high quality). If the data are overdispersed (i.e., the variance of observed crash data is significantly higher than the mean), the required sample size would be 34. However, one of the

challenges is that bicyclist crashes are extremely rare events; therefore, the baseline crash rate will most likely be below 0.02. As observed, the number of midblock sites must range between 800–12,000 to ensure that the CMF developed for this baseline crash rate is significant enough. Obviously, identifying this many sites per CMF is very challenging. One of the potential approaches would be to identify enough control sites to increase the overall sample size of the available data. Moreover, statistical models such as negative binomial, zero-inflated Poisson, and Tobit models provided other potential approaches to ensure that the CMFs developed in this study were statistically significant.

6.2 BIKE FACILITY AND CRASH DATA ASSESSMENT

The research team performed both cross-sectional (with PSM) and before-after analyses of the crash data based on the facility type. The B/A analysis conducted in this study did not utilize the empirical Bayes method, due to the lack of relevant SPFs. To perform the safety analysis, the team first conducted exploratory analyses of the crash data to assess the suitability of available data for

each analysis type. Figure 28 depicts the number of total and KABC (fatal and injury) bicyclist crashes that occurred in 2017–2020 (the evaluation period) per study site and facility type, together with the number of sites.

The number of sites indicates segments with each bikeway type. As discussed in Chapter 5, the research team used the segments from the roadway network data and conflated them with the bicycle network, where a segment was defined according to the roadway network database and stretched from one intersection to the next (see Figure 2 for the definition of a midblock segment). Because these segments were mainly located in dense urban areas, the segments were relatively short and had similar lengths in many cases. We used spatial join in ArcGIS to map each crash to the correct segment and reviewed the street names included in both the roadway network and crash database to ensure that each crash was assigned to the correct segment.

In 2017–2020, 54 bicyclist crashes occurred in Arlington County, Virginia, of which 38 occurred at sites with no bike lanes, 13 occurred in bike lanes, 1 occurred in buffered lanes, and 2 occurred in separated bike lanes. During this same period, Austin, Texas, experienced 905 bicyclist crashes, of which 704 occurred at sites with no bike lanes, 66 occurred in bike lanes, 9 occurred in buffered bike lanes, and 26 occurred in separated bike lanes. More than half of the SBL crashes in Austin occurred along two segments: Rio Grande and Guadalupe Streets. In Boston, Massachusetts, 79 bicyclist crashes occurred during the evaluation period; 73 of which occurred at sites with no bike lanes, while 4 occurred in bike lanes and 2 occurred in buffered bike lanes. In Dallas-Fort Worth, Texas, 821 bicyclist crashes occurred; of which 810 occurred at sites with no bike lanes, 5 occurred in bike lanes, and 6 occurred in buffered bike lanes. Of the 155 bicyclist crashes in Minneapolis, Minnesota, 87 occurred at sites with no bike lanes, 51 occurred in bike lanes, 7 occurred in buffered bike lanes, and 10 occurred in separated bike lanes. Of the 342 bicyclist crashes in Philadelphia, Pennsylvania, 213 occurred at sites with no bike lanes, 112 occurred in bike lanes, 14 occurred in buffered bike lanes, and 3 occurred in separated bike lanes. Finally, in Seattle, Washington, 510 bicyclist crashes occurred between 2017 and 2020, of which 323 occurred at sites with no bike lanes, 87 occurred in bike lanes, 10 occurred in buffered bike lanes, and 90 occurred in separated bike lanes. Most of the SBL crashes in Seattle (32 of the 90) occurred along three corridors: 2^nd Street, Roosevelt Way, and Pine Street.

As observed, most of the bicyclist crashes occurred at sites with no bike lanes, however, the sample size of these sites was considerably larger because it included almost the entire roadway network. At nearly all study sites, most of the on-road bike facility crashes occurred along conventional bike lanes, while buffered and separated bike lanes experienced very few crashes. Based on this assessment, the research team used crash data from all study locations except for Boston when estimating the safety effectiveness of conventional bike lanes but used crash data from only Austin, DFW, and Philadelphia to evaluate the safety effectiveness of installing buffered bicycle lanes. Because the bike and buffered bike lanes were already in place during the evaluation period, the research team conducted cross-sectional analyses using the propensity score matching method to assess their safety effectiveness.

To evaluate the safety effectiveness of SBLs, the research team collected the date of installation and separation type manually using Google Earth. According to this assessment, 135 SBLs were implemented before the evaluation period (i.e., before 2017): 9 in Arlington County, 7 in Austin, 3 in Boston, 12 in Minneapolis, 5 in Philadelphia, and 99 in Seattle.(Figure 29).

Figure 30 depicts the previous facility type where the SBLs were installed, together with the separation element. Among all study sites, Seattle had a reasonable sample size for conducting analysis. Therefore, the research team used the Seattle data for conducting a cross-sectional analysis. Conducting analyses for all SBLs across all study sites was not feasible due to inconsistencies in the data, particularly the roadway data, which was not collected consistently across all study sites.

Remaining SBLs were installed after the evaluation period (2017–2023). Overall, 170 SBLs were installed during this period across select study sites: 4 in Arlington County, 72 in Boston, 11 in Minneapolis, 2 in Philadelphia, and 81 in Seattle. As observed in Table 57, 1 crash occurred in Arlington County and Minneapolis before installing SBLs, and 1 crash occurred both before and after installing SBLs in Boston and Philadelphia. 28 crashes occurred in Seattle after installing SBLs, compared to 8 crashes before SBL installation. Note that the crash data for the before period may be incomplete, as the research team did not use the crash data before 2017. Based on this assessment, the research team used Seattle data for the before and after (B/A) analysis of SBLs. To ensure that there is adequate duration of crash data for both before and after periods, only the sites that were installed in 2018 and 2019 were considered. When considering the sites that were installed in 2018 and 2019, the average crashes at SBLs were 0.0833 in the before period and 0.0885 in the after period.

6.3 EXPOSURE ESTIMATION

As discussed earlier in Chapter 5, bicycle count data were obtained from several study sites; however, the sample sizes of available bicyclist count data (collected from on-road bicycle facilities) were not sufficient to estimate the bicyclist counts per study site. Using count data from shared use paths (where most of the counters were located) could introduce bias. Shared use path users have a limited exposure to risk as they are physically separated from the traffic. Therefore, using these counts estimate the exposure for midblock sites may introduce bias. To improve the sample size, the research team also used the estimated bicyclist counts obtained from Philadelphia and Texas cities. Figure 31 presents the number of sites used for developing exposure models together with the bikeway type and average daily bicycles.

An exposure model was developed using roadway (lane width, number of lanes); exposure (AADT, speed limit); land use (civic, commercial, residential, industrial); transit (transit routes, bus stops); sociodemographic (race, ethnicity); housing (households, housing units); and economic (median income, labor force) factors. To improve model accuracy, data from all study sites were combined, and one model was developed using the entire dataset without differentiation between study sites. Since each study site had a different data collection period, we assume that the seasonality is somewhat is accounted for in aggregated data, however this needs to be confirmed through more rigorous exposure modeling approaches. There is also possibility that the COVID-19 pandemic may have had an impact on observed counts. Findings of the literature review study by Li et al. (2022) indicate that recreational bicycling had increased during the pandemic while bike share trips had dropped significantly. Because most of the recreational bicycling trips occur at trails and shared use paths, it is not clear how pandemic had affected the bicycling trips on bikeways (that are mostly used for commuting). The exposure model estimation results are presented in Table 58.

Estimation results indicated that increased AADT and speed limits decreased bicyclist counts, while these counts increased as surface width (calculated as the number of lanes × lane width) increased. This result might be counterintuitive since bicyclist do not prefer cycling in wider streets since number of lanes is correlated with the increasing AADT. The ADB increased in civic land use areas but decreased in commercial land use areas. Increased total population (males and females) of a block group decreased bike counts. This might be due to the fact that most cyclists

prefer riding in urban areas where the area size of block groups is smaller (hence less population), however population density was not significant in this model which would account both for population and size of the block groups. Block groups with a higher percentage of White population had higher bike counts, while block groups with higher percentages of Black, Hispanic, and Asian populations had a lower number of bicyclists. Finally, a higher number of households in a block group resulted in higher bike counts, which again may be indicative of highly dense housing in urban areas where there is more biking activity.

The following equations can be used to estimate the bicyclist counts at separated, buffered, conventional, and no bike lanes:

for other variables indicate the effect of the variable on total and KABC bicyclist crashes and should not be interpreted as a CMF.

The base condition of this model included no bike lane in the before period. Overall, the number of bicyclist crashes decreased slightly after SBL installation (intercept=0.002). After accounting for the increase in ADB (9.4 percent), estimation results indicated a nearly 99 percent reduction in bicyclist crashes (denoted by the crash reduction factor [CRF]) after the installation of SBLs at sites with no bike lanes (although this was not statistically significant). Other interesting observations in this model included the relative reduction in crashes based on the previous facility type. For example, installing SBLs along former bike lanes and along former BBLs could potentially decrease crashes by 90 and 97 percent, respectively, compared to having no bike facility. Both of these findings were estimated to be statistically significant. Although a reduction in crashes was observed after installing SBLs, the magnitude of the reduction was much higher for sites with no bike lanes. Additionally, with respect to separation barrier type, installing flexi-posts may be more beneficial than concrete barriers because this factor was associated with an 88 percent reduction in crashes. Concrete barriers are usually installed at high-speed locations where the exposure to safety risk might be higher. They are also installed at sites with inherently high safety concerns. Most of the model estimates were found to be statistically significant. However, this model used only 1–2 years of before and after crash data; more years of data, as well as more sites, are needed to confirm the model findings.

Table 59. SBL Effectiveness, Before and After Analysis

Parameter		CMF (CRF)	Estimate	Std. Err.	p-value
(Intercept)		0.055	-2.907	1.326	0.028
Period (0 = Before, No Bike Lane, 1 = After, SBL)		0.007 (99%)	-5.017	4.546	0.270
Average Daily Bicycles		1.094	0.090	0.077	0.243
Segment Length		1.004	0.004	0.003	0.113
Previous Bikeway Type	Bicycle Lane	0.100 (90% )	-2.301	0.988	< 0.005
	Buffered Bicycle Lane	0.030 (97% )	-3.522	1.832	<0.001
Separation Type (0 = Concrete, 1= Flexi Post)		0.117 (88% )	-2.144	0.709	<0.005
Overdispersion		0.575
AIC		92

6.4.2 Cross-Sectional Analysis

For this analysis, we considered facilities that were installed before the evaluation period (i.e., before 2017) to ensure that the control and treatment (SBL) sites were already in place during 2017–2020. For the purposes of these analysis, we considered all SBLs in Seattle, regardless of their previous condition, and compared them with similar sites having (1) NBLs, (2) BLs, and (3) BBLs. Overall, 99 SBLs were in place before 2017 and were considered in this cross-sectional analysis.

6.4.2.1 Separated vs. No Bike Lanes

To compare SBL and NBL sites, we first conducted logistic regression to estimate propensity scores (i.e., the probability of a site with no bike lane to be selected for an SBL installation). According to the results of the logistic regression, the decision to install SBLs at sites with no existing bikeway was significantly associated with speed limit, segment length, surface width, land use type, average number of bus stops, and sociodemographic and economic factors (Table 60).

After identifying these variables, the research team developed a dataset for cross-sectional analysis, where the sites with and without treatment were selected. Sites with and without treatment were expected to be identical to ensure that any changes in crashes are mostly due to the presence of the treatment. Additionally, the number of control sites was expected to be 2 to 4 times larger than the treatment sites to achieve meaningful results. We allowed replacement in the data (i.e., when one site was selected more than once). The caliper width was 0.2 (Wang et al., 2013), while the ratio of treated to control sites was 1:2 (i.e., one treated site was matched with two control sites).

A total of 258 sites were selected for the negative binomial regression analysis for estimating the safety effectiveness of SBLs compared to no bike lanes (Table 61). Because of the repetition, the number of matched treatments sites was higher than the actual (raw) number of sites.

Finally, we developed safety models to estimate the impact of SBLs on total and KABC bicyclist crashes at sites with no previous bike facility. For the safety analysis, we developed negative binomial regression models. These models included treatment type and ADB, as well as the list of variables that were selected for estimating the propensity scores in the logistic regression. After removing highly insignificant variables, the following contextual factors were retained in the model: segment length, surface width, land use, and percentage of White population. Table 62 present the estimation result for total and KABC crashes.

As observed, SBLs were associated with significantly less bicyclist crashes. Compared to sites with significantly similar characteristics, sites with SBLs were found to have 98 percent and almost 99 percent less total and KABC bicyclist crashes, with statistical significance in each case. The ADB at these sites was observed to increase bicyclist crashes by approximately 10 percent when the ADB is increased by 100 percent (10.3 percent and 9.6 percent for total and KABC crashes, respectively) and have a significant effect on crash frequency. In addition to the bike facility type and ADB, bicyclist crashes were also affected by the speed limit, segment length, surface width, commercial land use, residential land use, and percentage of White population in a block group. As observed, the effect of the speed limit on crashes was negative, which may seem counterintuitive but is likely due to the confounding effects of the roadway characteristics (i.e., SBLs may have been installed along high-speed roads). Commercial land use had a significantly increasing effect on bicyclist crashes, while bicyclist crashes were estimated to decrease by 0.86 (14 percent) in residential areas.

Table 62. Safety Effectiveness of SBLs vs. NBLs.

Parameter	Total				KABC
	CMF (CRF)	Estimate	SE	P-value	CMF (CRF)	Estimate	SE	P-value
(Intercept)	0.217	-1.530	1.555	0.325	0.213	-1.545	1.586	0.330
Treated (0=NBL, 1=SBL)	0.016 (98%)	-4.115	3.238	<0.01	0.001 (~99%â)	-6.583	3.170	<0.05
ADB	1.103	0.098	0.045	<0.05	1.096	0.091	0.044	<0.05

6.4.2.2 Separated vs. Conventional Bike Lanes

Table 63 presents the list of contextual factors significantly associated with the probability of installing separated bicycle lanes along bike lanes, including surface width and various land use, transit, and sociodemographic factors. Based on these findings, 240 sites were selected for the safety analysis (Table 64).

The estimation results of the safety models are presented in Table 65. As observed, converting bike lanes to SBLs reduced total crashes by 99 percent and KABC crashes by 98 percent. Both CMFs were statistically significant. Additionally, the ADB at sites with SBLs had a significant increasing effect on bicyclist crashes (by almost 10 percent). Among other covariates, residential land

use decreased bicyclist crashes, indicating that converting bike lanes to SBLs in residential areas may improve bicyclist safety significantly.

Table 65. Safety Effectiveness of SBLs vs. BLs.

Parameter	Total				KABC
	CMF (CRF)	Estimate	SE	P-value	CMF (CRF)	Estimate	SE	P-value
(Intercept)	0.022	-3.796	0.513	<0.0001	0.019	-3.946	0.531	<0.0001
Treated (0=BL, 1=SBL)	0.002 (99%)	-6.068	3.298	<0.05	0.004 (99%â)	-5.627	3.329	<0.05
ADB	1.103	0.098	0.046	<0.01	1.098	0.093	0.047	<0.01
Segment length	1.004	0.004	0.001	<0.0001	1.004	0.004	0.001	<0.0001
Residential land use	0.556	-0.587	0.186	<0.001	0.583	-0.540	0.187	<0.001
Overdispersion	0.503				0.508
AIC	221				209

6.4.2.3 Separated vs. Buffered Bike Lanes

Finally, we compared SBLs with buffered bike lanes. Table 66 presents the list of contextual factors significantly associated with the probability of installing SBLs at former BBL sites. These factors include surface width and civic and commercial land use.

Table 67 depicts the sample size of matched segments. There were more SBLs than buffered bike lanes in Seattle. Therefore, some of the treatment sites were dropped from the sample, and a total of 112 sites were selected for the safety assessment (reverse matching).

Table 68 presents the estimation results. All other covariates were dropped from the model because they were not significant and introduced bias. As observed, SBLs were associated with 84 percent and 85 percent less total and KABC crashes, respectively, compared to BBLs. Although this effect was not statistically significant, the p-values were relatively small (p-value=~0.15), indicating that 85 percent of the time, converting a BBL to an SBL will result in an over 80 percent reduction

in crashes. The ADB again had a significant increasing effect on crashes; increasing ADB was estimated to increase total and KABC crashes by 8 percent.

Table 68. Safety Effectiveness of SBLs vs. BBLs.

Parameter	Total				KABC
	CMF (CRF)	Estimate	SE	P-value	CMF (CRF)	Estimate	SE	P-value
(Intercept)	0.005	-5.224	1.290	<0.0001	0.003	-5.699	1.334	<0.0001
Treated (0=BBL, 1=SBL)	0.156 (84%)	-1.855	1.276	0.1459	0.143 (85%â)	-1.942	1.318	0.141
ADB	1.083	0.080	0.032	<0.01	1.090	0.087	0.032	<0.0001
Overdispersion	0.146				0.145
AIC	415				391

6.5 BUFFERED BICYCLE LANES

To assess the safety effectiveness of buffered bike lanes, the research team used cross-sectional analysis to compare BBLs to: (1) no bike lanes and (2) bike lanes. As indicated earlier, these analyses were conducted using the crash data from three study sites: Austin, DFW, and Philadelphia. Because the data from Texas cities were collected consistently, Austin and DFW data were combined when performing this safety analysis. In Philadelphia, almost all bicyclist crashes resulted in fatalities or injuries. Therefore, we did not develop separate models for total and KABC crashes.

6.5.1 Buffered vs. No Bike Lanes

In this section, the research team estimated the safety effectiveness of installing buffered bicycle lanes compared to sites with no bike lanes.

6.5.1.1 Austin and Dallas-Fort Worth, Texas

Table 69 lists the significant factors (covariates) affecting the probability of installing buffered bike lanes at sites with no bike lanes in Austin and DFW. These factors included AADT, speed limit, lane width, and commercial land use. Increased AADT and lane width increased the probability of installing bike lanes, while increased speed limits and commercial land use decreased this probability.

Based on the selected covariates, treated sites (BBLs) were then matched with the control sites (no bike lanes). The sample size of the selected segments with significantly similar characteristics was 158 (Table 70).

Table 71 presents the estimation results. Installing buffered bike lanes at sites with no bike facility significantly reduced total crashes by 54 percent and KABC crashes by 52 percent. Other factors affecting bicyclist safety at buffered bike lanes included ADB, AADT, and speed limit; however, their effects were not statistically significant.

Table 71. Safety Effectiveness of BBLs vs. NBLs (Austin and Dallas-Fort Worth, Texas).

Parameter	Total				KABC
	CMF (CRF)	Estimate	SE	P-value	CMF (CRF)	Estimat e	SE	P-value
(Intercept)	0.049	-3.024	1.058	<0.001	0.047	-3.050	1.087	<0.001
Treated (0=NBL, 1=BBL)	0.451 (54%)	-0.796	0.311	<0.01	0.478 (52%â)	-0.737	0.311	<0.01
ADB	0.995	-0.005	0.012	0.673	0.996	-0.004	0.010	0.709
Log (AADT)	1.110	0.105	0.103	0.312	1.111	0.106	0.106	0.320
Speed limit	1.053	0.051	0.032	0.103	1.051	0.050	0.032	0.123
Overdispersion	0.180				0.186
AIC	295.000				298.000

6.5.1.2 Philadelphia, Pennsylvania

Table 72 shows lists the significant covariates affecting the probability of installing BBLs at sites with no bike lanes in Philadelphia. These factors included AADT, number of through lanes, lane width, residential land use, roadway class, daily transit, and population characteristics of the site.

Based on these covariates, 961 sites were selected for analysis; 433 sites had buffered bike lanes, while 523 sites had no bike lanes (Table 73).

The estimation results of the safety analysis (Table 74) again show that installing buffered bike lanes at sites with no bike lanes was associated with a 77 percent significant reduction in bicyclist crashes. The ADB had an increasing effect on bicyclist crashes; however, this effect was highly insignificant, indicating that despite increasing exposure, crashes can be reduced by installing BBLs.

Table 74. Safety Effectiveness of BBLs vs. NBLs (Philadelphia, Pennsylvania).

Parameter	CMF (CRF)	Estimate	SD	P-value
(Intercept)	0.096	-2.342	0.153	<0.0001
Treated (0=NBL, 1=BBL)	0.228 (77%)	-1.479	0.404	<0.0001
ADB	1.002	0.002	0.006	0.758
AADT	1.000	0.000	0.000	0.416
Through lanes	0.939	-0.063	0.134	0.639
Average daily transit	1.078	0.076	0.013	<0.0001
Overdispersion	0.112
AIC	542

6.5.2 Buffered vs. Bike Lanes

In this section, the research team estimated the safety effectiveness of installing buffered bicycle lanes compared to sites with conventional bike lanes.

6.5.2.1 Austin and Dallas-Fort Worth, Texas

After identifying the covariates (Table 75) that significantly impacted the probability of installing buffered bike lanes along conventional bike lanes in Austin and DFW, the sample size of the matched dataset was determined (Table 76). A total of 128 sites were selected; 54 had a buffered bike lane, while 74 had a conventional bike lane. These sites had significantly similar characteristics in terms of AADT, number of lanes, land use type, and percentage of White and Hispanic populations.

This imbalance indicated that bike lanes and buffered bike lanes were not usually installed at similar sites in Philadelphia; therefore, comparing them may not yield meaningful results.

Results of the safety analysis indicated a relatively small decrease in crashes (4.8 percent) when comparing BBLs to BLs; however, this estimate was highly insignificant. As expected, this model had a weaker performance (overdispersion>1).

Table 80. Safety Effectiveness of BBLs vs. BLs (Philadelphia, Pennsylvania).

Parameter	CMF (CRF)	Estimate	SE	P-value
(Intercept)	0.041	-3.199	0.458	<0.0001
Treated (0=BL, 1=BBL)	0.952(4.8%)	-0.049	0.400	0.792
ADB	1.002	0.002	0.006	0.755
AADT	1.000	0.000	0.000	0.181
Through lanes	0.721	-0.327	0.318	0.304
Overdispersion	1.060
AIC	298.000

6.6 CONVENTIONAL BICYCLE LANES

Finally, the research team conducted safety analyses for installing conventional bike lanes at sites with no bike facilities. As indicated earlier, crash data from the following study sites were used to estimate the safety effect of conventional bike lanes: Arlington County, Austin and DFW (combined), Minneapolis, Philadelphia, and Seattle.

6.6.1 Arlington County, Virginia

Table 81 summarizes the logistic regression results indicating the probability of a site being selected for a conventional bike lane treatment. The selection of sites for installing bicycle lanes was significantly dependent on street operation (one-way or two-way), AADT, speed limit, segment length, and land use type.

bicycle lane was significantly dependent on AADT, speed limit, lane width, number of lanes, land use type, number of bus stops, and population density.

Table 85 lists the number of treatment, control, and total sites after implementing the PSM. A total of 336 bike lanes were matched with 472 control sites, resulting in 808 segments for analysis.

Results of the crash models are presented in Table 86. Installing bike lanes was associated with a 38 percent significant reduction in total crashes and a 36 percent reduction in KABC crashes in Texas cities. The ADB had an increasing effect; however, the magnitude of this effect was very small (1 percent). Other factors affecting bicyclist safety at bike lanes included AADT, speed limit, lane width, civic land use, and number of bus stops on the bike lane route, all of which had an increasing effect on bicyclist crashes.

Table 86. Safety Effectiveness of BLs vs. NBLs (Austin and Dallas-Fort Worth, Texas).

Parameter	Total				KABC
	CMF (CRF)	Estimate	SE	P-value	CMF (CRF)	Estimate	SE	P-value
(Intercept)	0.070	-2.655	0.529	<0.0001	0.070	-2.664	0.542	<0.0001
Treatment (0=NBL, 1=BL)	0.617 (38%)	-0.483	0.121	<0.0001	0.634 (36%â)	-0.455	0.124	<0.0001
ADB	1.001	0.001	0.002	0.407	1.001	0.001	0.002	0.375
Log (AADT)	1.098	0.093	0.059	0.115	1.096	0.092	0.060	0.127
Speed limit	1.022	0.022	0.011	*	1.021	0.021	0.011	.
Lane width	1.015	0.015	0.012	0.232	1.015	0.015	0.013	0.234
Civic land use	1.080	0.077	0.014	<0.0001	1.080	0.077	0.015	<0.0001
Number of bus stops	1.070	0.068	0.019	<0.0001	1.068	0.066	0.020	<0.0001
Overdispersion	0.586				0.620
AIC	1703				1669

6.6.3 Minneapolis, Minnesota

Table 87 summarizes the logistic regression results indicating the contextual factors that affect the installation of conventional bicycle lanes in Minneapolis. The installation of bike lanes was significantly related to the number of through lanes, lane width, left shoulder width, land use type, and number of bus stops.

After identifying the covariates associated with the probability of installing bike lanes, we selected the sites for safety analysis. Of the 218 segments with a bicycle lane, 207 were matched with 217 control sites (with no facility), resulting in 424 segments for safety analysis (Table 88).

The results of the safety analysis indicated that the installation of bike lanes was associated with a relatively small reduction (5–6 percent) in crashes in Minneapolis; however, this reduction was not statistically significant. The ADB, lane width, commercial land use type, and number of bus stops had an increasing effect on crashes.

Table 89. Safety Effectiveness of BLs vs. NBLs (Minneapolis, Minnesota).

Parameter	Total				KABC
	CMF (CRF)	Estimate	SE	P-value	CMF (CRF)	Estimate	SE	P-value
(Intercept)	0.008	-4.883	0.761	<0.0001	0.006	-5.145	0.656	<0.000 1
Treatment (0=NBL, 1=BL)	0.949 (5%)	-0.052	0.547	0.2945	0.945 (6%â)	-0.057	0.525	0.194
ADB	1.031	0.031	0.024	0.1954	1.033	0.032	0.021	0.132
Lane width	1.046	0.045	0.014	<0.001	1.042	0.042	0.014	<0.001
Commercial land use	1.726	0.546	0.235	*	1.653	0.503	0.237	*

6.6.4 Philadelphia, Pennsylvania

Table 90 summarizes the logistic regression results indicating the contextual factors that affect the installation of conventional bicycle lanes in Philadelphia. In Philadelphia, the AADT, number of through lanes, and lane width, as well as various land use, transit, sociodemographic, and economic factors, were significantly associated with the probability of installing bicycle lanes.

After identifying these variables, the research team developed a dataset for cross-sectional analysis, where the sites with and without treatment were selected (Table 91). A total of 9,018 segments were selected for safety analysis.

Table 92 presents the estimation results for evaluating the safety effectiveness of bike lanes in Philadelphia. As mentioned earlier, almost all bicyclist crashes resulted in fatalities and injuries; we therefore developed only one model for this study site. According to the estimation results, bike lanes were associated with a 67 percent reduction in bicyclist crashes. This finding was statistically significant and very important for a city where bicyclist crashes tend to result in fatalities and

injuries. The ADB was associated with a significant increase in crashes; however, the magnitude of this increase was very small (2 percent). Other factors affecting bicyclist safety were lane width; speed limit; civic, commercial, and industrial land use; number of bus stops; average daily transit; and population density. Each of these factors, except lane width, was associated with a significant increase in bicyclist crashes.

• No bicycle facilities can potentially result in an over 98 percent reduction in bicyclist crashes.
• Bicycle lanes can potentially lead to a 90 percent reduction in total crashes and a near 100 percent significant reduction in KABC crashes.
• Buffered bicycle lanes can result in an 84–97 percent significant reduction in total crashes and an 86 percent reduction in KABC crashes.
• Installing buffered bicycle lanes at sites with:
  • No bicycle facilities can potentially result in a 55–77 percent reduction in total crashes and a 52–77 percent significant reduction in KABC crashes.
  • Bicycle lanes can potentially reduce the total and KABC bicyclist crashes by 5–25 percent. This CMF was not statistically significant. Overall, this CMF was weak due to the very small number of treatment and control sites selected for analysis, as well as the very small number of crashes at the selected sites.
• Installing conventional bicycle lanes at sites with no facilities can result in a 38–68 percent reduction in total crashes and a 37–68 percent reduction in KABC crashes. In Minneapolis, this reduction was only 5 percent; however, the p-value associated with this CMF was not statistically significant.

6.7.1 Bicycle Counts

Exposure had a considerable impact on bicyclist crashes. Moreover, the exposure changed based on the facility type. Sites with no bicycle lanes generally had a smaller number of bicyclists (9 per site). Bicyclist counts increased gradually at sites with conventional and buffered bicycle lanes (32 and 37 per site, respectively). However, the bicyclist counts increased substantially along separated bicycle lanes due to the comfort and perceived safety of these facilities (135 per site). Findings from the safety analyses indicated that the average daily bicyclists can potentially increase the exposure to safety concerns, hence increasing the crash frequency and severity. Generally, the ADB was associated with a 1–10 percent increase in bicyclist crashes based on the facility type; however, this effect was not always statistically significant due to limitations in the ADB estimation.

6.7.2 Other Contextual and Operational Factors

In addition to bikeway types and counts, the safety models included many contextual factors. The results of the cross-sectional safety analysis can be interpreted in two ways:

1. Factors influencing the probability of installing bicycle lane types.
2. Factors associated with the changes in safety outcomes.

When considering all variables across all the models, we observed that the decision to install a particular bikeway type was mainly affected by the AADT and speed limit, as well as various land use, transit, socioeconomic, and demographic factors. We then used all these important factors to assemble the dataset for analysis; thus, despite not being included in the individual safety models, all these factors were accounted for in the data that was used for the safety analysis.

Overall, the decision to install a bikeway and the consequential effects of a bikeway on bicyclist safety were affected by the following contextual factors:

• Roadway operations and design factors such as speed limit, number of lanes, and lane width were found to have a significant effect on both the selection of sites for installing bicycle lanes and bicyclist crashes. More lanes were found to be associated with less crashes. However, lane and surface width were found to have an increasing effect on bicyclist crashes in all cases (except in Philadelphia). Speed limit and AADT were also generally found to have a negative impact on bicyclist safety, but AADT was not a significant factor in most of the safety models. Although other factors such as median width, shoulder width, and parking type were considered in some of the models and were found to be significant factors in selecting sites for installing bicycle facilities, they were not found to be significant in the safety models and were thus removed. One explanation for this may be that in matched data (in cross-sectional analyses), sites with and without treatment are assumed to have very similar characteristics; therefore, the covariates affecting the site selection probability (i.e., propensity scores) are not significant because the data already accounts for their effect.
• Land use types such as civic (i.e., public and privately owned buildings for meeting and conventions), commercial, and residential had a significant effect in determining both the probability of a site being selected for a bike lane installation and safety outcomes. Overall, sites on commercial and civic lands (including gas stations, businesses, and

• parking lots) were associated with an increased number of crashes, while residential areas were mostly associated with a decreased number of bicyclist crashes.
• Transit variables such as transit routes and bus stops were also found to have a significant effect on the decision to install a bikeway. The findings from the safety analyses showed that the sites with higher numbers of bus stops had a negative impact on bicyclist safety.

Demographic and socioeconomic factors such as race, ethnicity, population density, median income, labor force, and housing units were also considered during modeling. Among these factors, the population density, percentage of different racial and ethnic groups, and median income were found to have a significant effect when selecting sites for installing bicycle facilities. However, these factors were ultimately removed from the safety models because they did not yield significant results.