CHAPTER 2

Data Collection and Analysis

Our analysis plan for this research was based on gaps identified in the literature review and the results of the state of the practice review. Our plan consists of several interrelated analyses beginning with a macro-level analysis of national crash data, followed by a more detailed, micro-level crash analysis to assess trends and contributing factors to pedestrian safety in darkness in several case cities. We then held a series of focus groups to understand pedestrian and/or driver expectations and behavior at night. The focus groups were complemented by experiments in a driving simulator to provide objective data on driver behavior relative to important factors of interest. Lastly, we conducted interviews with agencies to validate our findings.

This approach allowed us to consider an extensive array of contributing factors, including roadway characteristics, pedestrian and driver behavior, and factors that manifest in sociodemographic trends, across a number of urban environments. In keeping with the Literature Review and the State of the Practice, the emphasis was on urban environments (as defined by FARS) because of the highly disproportionate number of pedestrian fatalities in urban and suburban areas. Additionally, we intentionally included a focus on the safety of Black and Brown pedestrians in darkness in recognition of the disproportionate harm those communities suffer from traffic fatalities.

2.1 Macro-Level Analysis

Background

Our macro-analysis evaluates high-level trends and characteristics related to pedestrian safety in darkness. The primary purpose of our analysis is to determine factors that are associated with pedestrian fatalities occurring at night versus occurring during the daytime. This summary presents the descriptive statistics for key variables used in our analysis followed by our multivariate analysis. We summarize these descriptive statistics on a yearly basis to explore trends over time. For our multivariate analysis, we used binomial logistic regression to quantify the association between pedestrian fatality variables and pedestrian fatalities occurring during darkness. Specifically, we model whether or not a pedestrian fatality occurred in darkness, given that a pedestrian fatality occurred. In other words, our model results do not necessarily suggest that nighttime pedestrian fatalities are more likely to occur when significant variables are present. The significant variables in our models indicate that pedestrian fatalities with those specific characteristics are more likely to have occurred in the dark.

We consider several categories of factors associated with the Safe System Approach, including safe roadways, safe speeds, safe vehicles, and safe road users. We also examine contextual factors such as temporal (e.g., time of week, month of year) and weather conditions associated with pedestrian fatalities. Our analysis categorizes these factors more specifically into temporal (e.g., time of week, month of year), environmental (e.g., weather), location (e.g., intersection, midblock, sidewalk, crosswalk), roadway (e.g., number of lanes, speed limit), vehicle (e.g., car, truck, sport utility vehicle), movement (e.g., left turn, right turn), behavior (e.g., intoxication, speeding), and demographic (e.g., race/ethnicity, sex, age) variables.

We base our analysis on the eleven years of Fatality Analysis Reporting System (FARS) data from 2010 to 2020. FARS is a census of all fatalities on all public trafficways in the United States, including people who die from their injuries at the scene of the crash or up to 30 days after the crash (NHTSA 2022). It is considered the most reliable source of traffic crash data in the U.S. Our analyses build on recent multivariate research by members of the research team (Sanders et al. 2022) and others (Ferenchak et al. 2022) and allows us to further investigate the descriptive analysis of trends in pedestrian fatalities as reported in Schneider (2020) and Tefft et al. (2021).

Our analyses of characteristics associated with darkness compare darkness against daylight, dawn, or dusk. For this analysis, we define darkness using the FARS lighting condition variable, which is derived from police reports. Our definition of darkness includes “darkness with streetlights,” “darkness without streetlights,” and “darkness with unknown street lighting,” but it does not include “dawn” or “dusk.” For comparison, we also derived a variable for darkness based on the number of degrees the sun was below the horizon at the given time of day and latitude of the crash (“civil twilight” ends when the sun reaches six degrees below the horizon). Our empirically-derived darkness variable had a high correlation with the police-reported darkness variable, so we chose to use the police-reported variable to maintain consistency with the FARS database and make our results easier to replicate. The judgment of police on the scene might also better represent the influence of cloud cover on darkness.

Descriptive Statistics

Our dataset shows that the total number of U.S. pedestrian fatalities increased by 51% between 2010 and 2020 (4,302 to 6,516 fatalities per year) (Figure 3). During the same period, pedestrian fatalities during darkness (not including dawn or dusk) increased by 63% (3,030 to 4,951 fatalities per year). Therefore, the proportion of pedestrian fatalities that occurred during darkness increased from 71% to 76% between 2010 and 2020.

The exposure context for these pedestrian fatalities is important. Currently, the best available national-level data on both pedestrian and motor vehicle travel at different times of day comes from the National Household Travel Survey (NHTS) (FHWA 2017). This survey provides the start times for trips made between pairs of origins and destinations by different types of transportation (it also includes trips made for exercise or recreational purposes). While these trips are not coded by daylight versus darkness, we analyzed the proportion of trips starting between 6:00 pm and 5:59 am as a rough approximation of nighttime trips. We examined pedestrian trip and vehicle miles traveled data from 2009 and 2017 to approximate exposure by time of day. We analyzed pedestrian trips rather than pedestrian miles traveled because the 2017 NHTS pedestrian miles traveled data includes some large outlier values and we believe that the pedestrian trip data are more accurate.

The 2009 and 2017 NHTS show that there was between three and four times more pedestrian and vehicle travel during the hours of 6:00 am and 5:59 pm than between 6:00 pm and 5:59 am (Figure 4). There may have been a slight shift from nighttime to daytime travel between the 2009 and 2017 surveys. Given the high proportion of pedestrian fatalities corresponding with lower levels of pedestrian and vehicle activity at night, the relative risk of pedestrian crashes (per pedestrian trip and per vehicle mile traveled) occurring at night is much higher than during the day.

Additional contextual information about exposure is included in Appendix A.

The following subsections summarize the overall prevalence of specific characteristics in pedestrian fatalities as a whole and the proportion of pedestrian fatalities with these characteristics that occurred during darkness (Table 3 through Table 6). More detailed information about the descriptive statistics, including additional data tables, are provided in Appendix B. We also discuss the influence of exposure and other possible reasons for these specific results in this appendix.

Overall, the fourth quarter of the year has more pedestrian fatalities than any other quarter, as October, November, and December each account for more than 10% of total pedestrian fatalities (Table 3). The highest percentages of fatalities during darkness correspond with the darkest months of the year in the Northern Hemisphere (October through March). Comparing extremes, 81% of December pedestrian fatalities occur in darkness compared to 67% of June pedestrian fatalities.

Fridays overnight (noon on Friday to 11:59 am on Saturday) and Saturdays overnight (noon on Saturday to 11:59 am on Sunday) have more pedestrian fatalities than any other days of the week, each with more than 16% of total pedestrian fatalities (Table 3). These two days also have the highest proportion of pedestrian fatalities during darkness (79% of Friday fatalities and 83% of Saturday fatalities are during darkness).

Most pedestrian fatalities (75%) occur in clear conditions (Table 3). Of the fatalities that occur under clear conditions, 73% are during darkness. In contrast, 9% of pedestrian fatalities occur during rainy conditions yet 83% of rainy fatalities are in darkness. Therefore, rain may be associated with increased pedestrian fatality risk during darkness.

Most pedestrian fatalities occur at non-intersection locations (81%) (Table 4). Locations away from intersections are also overrepresented in fatalities during darkness (76%). Looking in more detail at intersection locations, pedestrian fatalities are more likely to occur on the far side of intersections (58%) than the near side (42%). However, pedestrian fatalities on the near side of intersections are more likely to occur in darkness (74% on the near side versus 61% on the far side). Marked crosswalk locations have far fewer pedestrian fatalities than other locations, as marked crosswalks account for less than 11% of all fatalities. Marked crosswalk fatalities are also less likely to be in darkness than other fatalities (56% versus 76%, respectively). We do not know the amount of pedestrian activity at marked crosswalks versus other locations, so we cannot compare the risk of these different crossing types.

Most pedestrian fatalities occur in the roadway (89%) rather than on driveways, sidewalks, and other non-roadway locations (Table 4). In-road pedestrian fatalities are much more likely to occur during darkness (77%) than pedestrian fatalities that happen in other locations (45%). This could be due to in-road locations being associated with higher vehicle approach speeds (and less reaction time) than driveway or other fatality locations. It could also be associated with roadway lighting being focused near intersections, driveways, or at the sides of roads rather than in the middle of the road. However, these are theories that require more detailed assessment.

Overall, pedestrian fatalities are most common along arterial roadways (69% of fatalities), roadways with four or more lanes (61% of fatalities), and roadways with 35 to 50 mph speed limits (53% of fatalities) (Table 4). These roadway characteristics are often found simultaneously on major thoroughfares, many of which are part of the national and state highway networks and are under the jurisdiction of state DOTs. These common roadway characteristics are also overrepresented in darkness-related fatalities. Pedestrian fatalities on roadways with higher speed limits are more likely to occur in darkness, ranging from 55% in darkness on roadways with speed limits of 30 mph or lower to 81% on roadways with speed limits of 55 mph or higher. Approximately 67% of pedestrian fatalities on roadways with three or fewer lanes occur in darkness, but this proportion increases to 80% in darkness on roadways with four or more lanes. Local roadways have less than 60% of their pedestrian fatalities occur at night, a much lower percentage than arterial roadways (76% in darkness) and freeways (81% in darkness).

Passenger cars (46%) and pickup trucks, vans, and SUVs (46%) are involved in similar proportions of overall pedestrian fatalities (Table 5). For context, passenger cars made up slightly more of U.S. household vehicles (50% in 2009 and 2017) than pickup trucks, vans, and SUVs (45% in 2009 and 46% in 2017) (McGuckin and Fucci 2018). These figures do not include commercial vehicles. Pedestrian fatalities involving passenger cars are more likely to occur during darkness (77%) than pickup trucks, vans, and SUVs (71%) or commercial trucks and other heavy vehicles (57%). This may reflect less nighttime use of commercial vehicles in areas with high levels of pedestrian activity. The 2017 NHTS shows that 21% of passenger car miles traveled are between 6 pm and 6 am compared to 19% for pickup trucks, 19% for vans, and 18% for SUVs (FHWA 2017). This finding could also be related to geographic differences in vehicle characteristics (e.g., urban versus rural), where areas with more pedestrian activity at night might have more passenger cars and areas with less pedestrian activity at night might have more large trucks and SUVs. Additional research could help clarify these relationships.

Most pedestrian fatalities occur when vehicles are traveling straight (84%). Straight-traveling vehicle pedestrian fatalities are also overrepresented in darkness (79% of fatalities in darkness). Examining the movements of pedestrians relative to vehicles, most pedestrian fatalities involve a pedestrian traveling perpendicular to the vehicle’s line of travel. More detailed intersection data indicate that pedestrians struck while traveling from the driver’s left to right account for 43% of fatalities, and pedestrians traveling from the driver’s right to left account for 39% of fatalities. Both of these types of interactions are also overrepresented in darkness, with 78% of fatalities involving pedestrians traveling from the driver’s left to right occurring in darkness.

Approximately 35% of pedestrian fatalities are flagged for pedestrian alcohol involvement and 11% are flagged for driver alcohol involvement. Alcohol use by pedestrians is legal but alcohol use by drivers over a given blood alcohol concentration is illegal in all states. This finding is based on analysis of the “DRINKING” variable in FARS (NHTSA 2022), which indicates police-reported alcohol involvement in the fatal crash and therefore reflects the judgement of law enforcement rather than the result of a blood test. We used this variable in the analysis because it is more complete than blood test data. Research has found that driver alcohol usage is positively associated with hit-and-run crashes (described below) (Benson et al. 2017), suggesting that driver alcohol impairment may be underestimated in this sample. Approximately 19% of pedestrian fatalities are flagged for pedestrian drug involvement and 7% are flagged for driver drug involvement. Of these types of pedestrian fatalities, pedestrian alcohol involvement (92% occur in darkness), driver alcohol involvement (82%), and pedestrian drug involvement (88%) are overrepresented at night. Sixty-two percent of pedestrian fatalities that involve a driver with drugs occur in darkness. This does not mean that driver drug involvement is safer for pedestrians at night. Better data on how driver drug use is distributed throughout the daytime versus nighttime would help clarify risk related to this variable.

Nearly 20% of pedestrian fatalities involve hit-and-run drivers (Table 5). These hit-and-run fatalities are overrepresented in darkness: approximately 85% occur in darkness. Reckless driving is cited in approximately 2% of pedestrian fatalities (Table 5). Fifty-five percent of reckless driving pedestrian fatalities occur in darkness.

Table 5. Overall Prevalence of Specific Characteristics in Pedestrian Fatalities and Proportion of Fatalities with Each Characteristic that Occurred in Darkness, 2010 to 2020

1) The total number of fatalities listed for each characteristic is the number of fatalities with known values for lighting condition and known values for the given characteristic. Fatalities with an unknown value for either variable are excluded from this table.

2) Passenger Car is classified as BODY_TYP = 1, 2, 3, 4, 5, 6, 7, 8, 9, and 17 from the FARS database.

3) Pickup Truck/Van/Sport Utility Vehicle is classified as BODY_TYP = 10, 14, 15, 16, 19, 20, 21, 22, 28, 29, 30, 31, 32, 34, 39, 40, 42, 45, 48, 49, and 67 from the FARS database.

4) Commercial Truck or Other Heavy Vehicle is classified as BODY_TYP = 12, 50, 51, 52, 55, 58, 59, 60, 61, 62, 63, 64, 66, 71, 72, 78, and 79 from the FARS database.

5) Other Vehicle Type is classified as BODY_TYP = 11, 13, 33, 41, 65, 68, 73, 80-89, and 90-97 from the FARS database.

6) Fatal crashes involving the pedestrian and vehicle traveling in either the same initial or opposite initial direction occurred in the far side crosswalk when a vehicle turned left or right. For example, if a northbound pedestrian was struck by an initially northbound vehicle that was turning left to go west, it was classified as “Pedestrian and Vehicle Traveling Same Direction.” If a southbound pedestrian was struck by an initially northbound vehicle that was turning left to go west, it was classified as “Pedestrian and Vehicle Traveling in Opposite Direction.”

7) Fatal crashes involving the pedestrian intersecting the vehicle path from left to right or right to left occurred in either the near side crosswalk for vehicles going straight, left, or right or the far side crosswalk for vehicles going straight.

Seventy-two percent of pedestrian fatalities involve people between age 20 and 64 (Table 6). The highest proportions of fatalities in darkness involve pedestrians between age 20 and 44, with more than 80% of those pedestrian fatalities occurring in darkness. Child pedestrians (younger than age 15) and older adults (older than age 64) are underrepresented in fatalities during darkness. Drivers aged 20 to 34 account for the largest shares of pedestrian fatalities (13% for ages 20 to 24 and 22% for ages 25 to 34). Drivers in the 20 to 24 and 25 to 34 age groups also have the highest proportions of pedestrian fatalities in darkness (76% and 75%, respectively). As described earlier, there is very likely an exposure element associated with age groups that are and are not more likely to be walking or driving at night.

Male pedestrians and male drivers account for more than twice as many pedestrian fatalities as their female counterparts (Table 6). Male pedestrians are also more likely to be killed in darkness (76%) than female pedestrians (68%). Some of this disparity could be explained by exposure since the 2017 NHTS shows that 24% of male pedestrian trips are between 6 pm and 6 am compared to 21% of female pedestrian trips (FHWA 2017). Overall, male drivers are more likely than female drivers to kill pedestrians (male drivers account for 71% of pedestrian fatalities). The disparity between male driver and female driver involvement in nighttime pedestrian fatalities is smaller, with 73% of male and 71% of female driver-involved pedestrian fatalities occurring in darkness. One possible reason for this result could be that male drivers are riskier at night around pedestrians. Another could be that males generally drive more at night. According to the 2017 NHTS, men did a greater proportion of their daily vehicle travel between 6 pm and 6 am (22%) than women (17%) (Federal Highway Administration 2017). More research would help clarify the degree to which these differences between men and women are due to behavior, exposure, or other factors.

More pedestrian fatalities involve pedestrians who are White (53%), Black (20%), and Hispanic/Latino (20%) than other racial and ethnic groups (Table 6). Other studies show that American Indian/Alaska Native, Black, and Hispanic/Latino pedestrians are overrepresented while White pedestrians are underrepresented in pedestrian fatalities on a per capita basis (Sanders and Schneider 2022). Pedestrian fatalities among American Indian/Alaska Native (86% in darkness), Black (80% in darkness), and Hispanic/Latino (75% in darkness) groups are all overrepresented in darkness.

There are likely exposure-related factors for differences in sex, race, and ethnicity that we are unable to tease out from the FARS data. However, it is unlikely that exposure alone explains these disparities (Sanders and Schneider 2022).

Table 6. Overall Prevalence of Specific Characteristics in Pedestrian Fatalities and Proportion o Fatalities with Each Characteristic that Occurred in Darkness, 2010 to 2020

1) The total number of fatalities listed for each characteristic is the number of fatalities with known values for lighting condition and known values for the given characteristic. Fatalities with an unknown value for either variable are excluded from this table.

Method: Multivariate Statistical Analysis

We used multivariate statistical modeling to identify specific characteristics that had a significant association with pedestrian fatalities occurring in darkness versus daylight. Our analysis also quantified the odds of a pedestrian fatality with a specific characteristic occurring during darkness rather than daylight, all else equal. This modeling approach involved two steps: random forest analysis and binomial logistic regression modeling.

Random Forest Analysis

First, we prepared a database of more than 100 explanatory variables that could potentially be related to pedestrian fatalities occurring at night. We were initially concerned that this large number of variables might be impractical to work with and could have resulted in some variables showing statistical significance by random chance. Additionally, some of these variables were correlated, so we wanted to avoid including them in the same model. We used random forest analysis as a screening step to identify which of the FARS database variables might be the most important to include in our multivariate models. This was an early exploratory step, and we ultimately tested combinations of all variables in the binomial logistic regression modeling process.

Important variables identified by the random forest analysis included crashes occurring in the roadway, pedestrian and driver drinking and drug use, driver distraction, hit-and-run crash, vehicle traveling straight, speed limit less than 25 mph, pedestrian age (under age 16 and over age 64), day of week (Friday overnight and Saturday overnight), and season of year (October through December). Random forest analysis identifies the potential importance of a variable for predicting whether or not a pedestrian fatality occurs in darkness; it does not indicate the direction of the relationship (i.e., some of these variables could have positive associations with darkness while others could have negative associations with darkness).

Binomial Logistic Regression Modeling

Second, we used binomial logistic regression models to identify the significance and magnitude of relationships between specific variables and fatal pedestrian crashes occurring during darkness rather than daylight. Each multivariate model controls for the simultaneous influence of the full range of variables that are included in the model.

We estimated four sets of models and present final versions of each model in the results section (Table 7). Each model is shown as a main column in the table. Model 1 (n = 59,780) uses the entire dataset of pedestrian fatalities reported between 2010 and 2020. Model 2 is the subset of pedestrian fatalities that occurred at non-intersection locations (n = 47,889), and Model 3 is the subset of pedestrian fatalities that occurred at intersections (n = 10,870). Model 4 is estimated from a database of crashes that occurred at intersections from 2014 to 2020 (n = 7,224). This seven-year dataset incorporates pedestrian circumstance variables that were not available in the FARS database until 2014.

The total number of fatal pedestrian crashes used in each model (n) and an overall model Akaike information criterion (AIC) are listed at the top of each model column. The odds ratio for each variable (derived by taking e to the power of the estimated parameter), 95% confidence interval for the odds ratio, and statistical significance of the parameter estimate (Z-value) are given in the rows of the table. Asterisks indicate statistical significance at the 99.9% level (***), 99% level (**), and 95% level (*). Lower-script dots indicate parameter estimates that are significant at the 90% level (.).

The FARS database has missing or unknown values for several variables. We controlled for these missing values by incorporating missing value indicator (dummy) variables into the models. Not controlling for missing values appropriately could lead to systematic bias in the estimated effects of particular variables. For example, many hit-and-run crashes do not include information about driver characteristics, such as the driver drinking. Treating an unknown driver drinking value as zero (indicating “not drinking”) would likely bias the results to show that driver drinking at night is less prevalent than it is in reality. Controlling for missing values provides more accurate estimates of the association between known driver drinking and pedestrian fatalities at night.

We examined correlations between explanatory variables to avoid model estimation problems due to multicollinearity. Several pairs of missing value indicator variables were highly correlated (|p| > 0.7). These pairs were among the variables representing missing pedestrian drinking, missing pedestrian drug use, missing driver drinking, missing driver drug use, missing driver age, missing driver sex, and missing vehicle type. (Missing values for driver age, driver sex, driver drinking, driver drug use, and vehicle type also had high correlations with hit-and-run crashes, but they were below the |p| > 0.7 threshold.) However, the estimated coefficients for these variables were not central to the results, so we kept each of these variables to provide the most complete set of controls for missing values. The only other variables that were highly correlated were the interactions of arterial functional classification and traffic control (arterial × signal, arterial × stop sign, arterial × other, arterial × missing) and the corresponding traffic control variables (signal, stop sign, arterial, and other). We accepted this high correlation in order to better understand how different types of traffic control along arterials were associated with pedestrian fatalities during darkness.

Finally, we developed a separate model comparing factors associated with pedestrian fatalities that occurred between 2015 and 2019 with those that occurred between 2010 and 2014. We used darkness as an explanatory variable in these models. This allowed us to test whether or not darkness had a significant relationship with the increase in pedestrian fatalities during the 2010s, after controlling for a large set of other explanatory variables. Appendix C shows that darkness had a significant positive relationship with the increase in pedestrian fatalities over the last decade.

Findings: Variables Associated with Pedestrian Fatalities During Darkness

Many of the explanatory variables in our four binomial logistic regression models had statistically significant relationships with pedestrian fatalities occurring during darkness rather than daylight, dawn, or dusk (Table 7). Overall, the following variables had the strongest relationships with nighttime pedestrian fatalities (odds ratios generally greater than two): October through December time period, rainy weather, crashes occurring in the roadway at non-intersection locations, roadways with higher posted speed limits, drivers going straight (rather than turning), pedestrians and drivers drinking, and pedestrians who were age 16 to 64 (rather than children or older adults).

Several of these variables likely reflect the influence of exposure, as discussed earlier. For example, October through December have more hours of darkness than daylight in the United States. People are probably more likely to drink alcohol in the evening hours, which correspond with darkness. Children and older adults are probably less likely to be walking at night than people who are aged 16 to 64. However, reliable national data on how much pedestrian and driving activity occurs during different time periods, is associated with drinking, and is done by people of different ages, is unavailable.

Within the context of this limitation, we describe statistically significant model results for temporal, weather, location, roadway, vehicle, movement, behavior, and demographic variables in more detail below.

Table 7. Binomial Logistic Regression Model Results: Variables Associated with Pedestrian Fatalities During Darkness

1) Days of the week correspond with the 24-hour period from noon for the day listed to 11:59 am the following day (e.g., Friday represents noon on Friday until 11:59 am on Saturday). This is done to ca

2) Missing values for all location-related variables are controlled by the missing location variable. By definition, there are no FARS fatality records with “other location” or “missing location” values i

3) There are only 10 missing values for work zones in the entire FARS pedestrian fatality dataset, so the estimates for this parameter are very small or undefined in Model 3 and Model 4.

4) Missing values for one-way streets are identical to missing values for the number of lanes, so a missing value indicator variable was not included.

5) The model samples exclude FARS pedestrian fatality records that had unknown values for lighting condition.

Temporal Characteristics

After controlling for all other variables, the odds of a pedestrian fatality occurring in darkness increased over time. The odds of a pedestrian fatality being in darkness increased steeply in the mid-2010s, and highest odds of a pedestrian fatality being in darkness were during 2018 (28% higher than 2010 in the model with all pedestrian fatalities). The odds ratios for 2019 and 2020 decreased slightly but were still higher than 2010 in all three models that covered the full study period.

These year-specific variables allow us to organize the otherwise unexplained variation in our models temporally. There are a number of important factors that could affect the proportion of pedestrian fatalities that occur in darkness, but we do not have data to represent. Importantly, these unmeasured factors could have shifted in the last 10 years. For example, the distribution of pedestrian volume during night vs. day could have increased (due to changing employment shift hours), traffic volumes at night vs. day could have increased (due to increased nighttime TNC use), or the prevalence of driver distraction or blinding from in-vehicle screens could have increased more at night (due to vehicle entertainment system changes). We were not able to measure these factors directly, but finding that the likelihood of nighttime pedestrian fatalities became higher during particular years when some of these changes were probably happening provides insight into what further research could explore. For example, additional studies could try to see if any of the factors mentioned in this paragraph (or others) increased fairly steeply between about 2014-2018.

October through December had the highest odds of pedestrian fatalities at night, with more than twice the likelihood of being at night than April through June. The higher odds of nighttime fatalities are likely due to more hours of darkness during October through December than other times of year.

Fridays overnight (noon on Friday to 11:59 am on Saturday) and Saturdays overnight (noon on Saturday to 11:59 am on Sunday) are more likely to have pedestrian fatalities in darkness than other days of the week. Since the models control for behaviors such as drinking that may be more likely on Friday and Saturday nights, these results likely reflect higher pedestrian and motor vehicle activity during these two nights.

Weather Characteristics

Rain was consistently associated with a higher likelihood of pedestrian fatalities during darkness across all four models (nearly twice as high as clear weather at non-intersection locations and three times as high at intersection locations). A possible explanation for this result is that rain increases roadway glare, and it might make it more difficult for drivers to see through their windshields, especially at night. Rain also decreases roadway friction, increasing the time and space needed for drivers to stop. “Other” conditions associated with pedestrian fatalities at night include snow and sleet, which may have similar effects as rain.

The intersection models showed that cloudy conditions were also significant, but these parameter estimates are less precise and do not have a clear theoretical connection with fatalities at night.

Location Characteristics

The non-intersection model showed that crashes occurring within the roadway were over four times more likely to occur at night versus day than non-roadway locations (such as shoulder, parking lane, driveway, or sidewalk). Nearly all crashes at intersections occur in roadways (regardless of day or night), so this finding did not translate to intersection locations.

Pedestrian fatalities at crosswalks had a significantly lower likelihood of occurring at night than during the day across all four models. This could indicate that crosswalk locations are less risky at night (potentially

Vehicle Characteristics

Pedestrian fatalities involving sedans (i.e., smaller cars) had significantly higher odds of occurring during darkness than larger vehicles. Large commercial trucks are probably more likely to be on roadways during the day than at night. Pickup trucks, vans, and SUVs could possibly be more likely than sedans to be used for daytime work purposes, so they may also have lower odds of pedestrian fatalities at night because of lower exposure. The 2017 NHTS shows that 21% of passenger car miles traveled are between 6 pm and 6 am compared to 19% for pickup trucks, 19% for vans, and 18% for SUVs (FHWA 2017). However, we did not examine commercial versus personal use of vehicles in this analysis, so this hypothesis requires further research. It might also be possible that these larger personal vehicles could have some type of design advantage (e.g., headlights that are higher or at a different angle than smaller cars) that reduces their risk to pedestrians at night. While the possibility of a marginal nighttime safety benefit associated with pickup trucks, vans, and SUVs could be explored through further research, research shows that these heavier vehicles with higher front ends are associated with more severe and fatal injuries to pedestrians overall (Lefler and Gabler 2004). Additionally, SUVs were overrepresented in the increase in pedestrian fatalities that occurred at night when comparing 2010-2017 with 2002-2009 (Ferenchak and Abadi 2021).

Movement Characteristics

Pedestrian fatalities that involve drivers going straight are at least three times more likely than fatalities that involve drivers turning left or right to occur during darkness. This likely reflects the importance of vehicle speed for drivers being able to detect and react to pedestrians in the dark. Since left- and right-turning vehicles tend to travel at slower speeds than those traveling straight, drivers are more likely to have adequate time to see pedestrians at night when they are turning.

We did not find statistically significant differences in the odds of nighttime pedestrian fatalities between near side and far side crosswalks at intersections.

Pedestrian fatalities involving pedestrians and drivers traveling in the same direction prior to the crash were significantly more likely to be at night than during the day. This is likely due to drivers having a harder time seeing pedestrians ahead of them on the roadway during darkness. Given this finding, further human factors research could explore the challenge of visually detecting longitudinal movement versus lateral movement at night.

Pedestrians crossing from left to right in front of a driver were more likely than pedestrians crossing from right to left to be associated with nighttime pedestrian fatalities. While this difference was not statistically significant, this result could add support to the theory that pedestrian visibility at night is diminished by opposing traffic headlights.

Behavior Characteristics

Pedestrian fatalities that involve pedestrians or drivers who were drinking are significantly more likely to be at night. Most models showed that the odds of nighttime pedestrian fatalities were at least three times higher for drinking pedestrians and two times higher for drinking drivers, which is likely related to greater alcohol consumption at night.

Pedestrian drug use is positively associated with pedestrian fatalities at night across all four models. Yet, driver drug use is negatively associated with pedestrian fatalities at night (significant association in the total pedestrian fatality model and non-intersection model but non-significant association in the intersection models). We are not aware of a theoretical explanation for these opposite results for pedestrians and drivers, but further research could help identify possible explanations.

Hit-and-run pedestrian fatalities are significantly more likely to occur during darkness. This may be due to drivers not realizing that they hit a pedestrian (and the difficulty of detecting them in the dark in the first place). It could also relate to some drivers perceiving that they can leave the scene of a crash more easily at night without being identified. As stated earlier, leaving the scene may also be associated with driver impairment, although this information is not directly available in FARS data.

Pedestrian fatalities due to speeding and reckless driving were significantly less likely to be reported during darkness (reckless driving was only significant in the models with all pedestrian fatalities and non-intersection pedestrian fatalities). We think that speeding and reckless driving behaviors would make it more difficult to detect and react to pedestrians at night. However, there may be fewer witnesses observing these behaviors at night, and law enforcement officers may be less likely to record these behaviors due to less clear visual evidence, so these behaviors may be underreported for nighttime pedestrian fatalities. Further research could help to explore this finding.

Demographic Characteristics

After controlling for all other variables, several demographic characteristics of pedestrians and drivers are still associated with higher odds of pedestrian fatalities occurring at night.

Pedestrians younger than age 16 were less than one-third as likely and pedestrians older than age 64 were less than half as likely as pedestrians aged 16 to 64 to be killed at night rather than during the day. This likely reflects that children and older adults are less likely to be walking along and across roadways at night. Male pedestrians were more likely to be killed at night than female pedestrians, especially at intersections. We do not know why intersections seemed to have a stronger relationship with nighttime pedestrian fatalities for males.

Pedestrians who were Black or Pacific Islander (most commonly Native Hawaiian) were significantly more likely to be killed at night in all four models. Hispanic/Latino and American Indian/Alaska Native pedestrians were significantly associated with nighttime fatalities in the models with all fatalities and non-intersection fatalities. The overrepresentation of these racial and ethnic groups may have to do with exposure, such as walking at night for work in service jobs. For instance, higher proportions of Black workers work at night than White workers (Population Reference Bureau 2008; Lieberman et al. 2020). People who are Black also have lower motor vehicle ownership and access (Brumbaugh 2020), which research suggests stems from historic exclusion from many kinds of work (Adkins-Jackson et al. 2022). So, to access third-shift jobs, Black workers may be more likely than White workers to rely on walking or taking the bus in darkness. However, this association may also reflect a more general exposure to less safe roadway conditions due to historical factors like redlining (Roll 2021; Sanders and Schneider 2022). There may also be a relationship between pedestrian safety, lighting effectiveness, and skin tone that additional research could help illuminate.

Pedestrian fatalities involving drivers aged 15 to 19 had significantly higher odds of being at night than those involving drivers older than age 64. This result is unlikely to be due to older drivers being able to see pedestrians better at night. More likely, it could be due to exposure, especially for older drivers who may try to avoid driving at night. The higher likelihood of young drivers being involved in pedestrian fatalities at night could also reflect their inexperience dealing with dark conditions.

Female drivers had 20% lower odds of being involved in a pedestrian fatality during darkness than male drivers. This could indicate that female drivers are less risky at night around pedestrians or that females

2.2 Micro-Level Analysis

Background

There is no travel mode more theoretically accessible than walking. It is free, requires no or relatively minimal equipment compared to other modes, and accrues mental, emotional, and physical health benefits. Yet many people and places in the U.S. experience less accessibility, due in part to a lack of safety from traffic, particularly in dark conditions. Despite decades of progress in reducing pedestrian fatalities between 1980 and 2010 (Schneider 2020), a confluence of factors since that time has resulted in increasing pedestrian fatalities, culminating in a 40-year high in pedestrian fatalities in 2021 – of which 76% occurred in dark conditions (Petraglia and Macek 2023). Myriad studies in the past twenty years focused on understanding factors that are associated with fatal and injurious outcomes of motorist-pedestrian crashes, routinely finding darkness as a significant correlate (e.g., Siddiqui et al. 2006; Kim et al. 2010). However, few studies until recently have examined pedestrian safety explicitly in darkness, and those that did (e.g., Sullivan and Flanagan 2002) often used simpler, bivariate analysis to investigate the problem. For example, in their comprehensive review of factors, Tefft et al. (2021) found that 73% of the increase in pedestrian fatalities between 2009 and 2018 occurred in darkness, and that roadway classification (particularly arterials) and posted speed were strongly correlated with the increase in pedestrian fatalities, but the authors did not control for lighting condition when examining other factors.

Other recent studies have sought to understand the significant rise in pedestrian fatalities since 2010 by focusing specifically on dark conditions and using multivariate models to provide insight into the relative importance of certain factors while controlling for others. Several studies have found significant associations between pedestrian fatalities in dark conditions and roadway design and operations (e.g., Sanders et al. 2022; Ferenchak and Abadi 2021), sociodemographic characteristics (e.g., Long and Ferenchak 2021; Sanders and Schneider 2022; Dumbaugh et al. 2023), and land uses such as liquor stores, restaurants, and shopping centers (Long and Ferenchak 2021; Dumbaugh et al. 2023). In other words, while these factors may be associated with pedestrian injuries and fatalities in both daytime and dark conditions, their relationship to injury severity appears to be particularly related to the lighting condition. For example, a pedestrian hit at 45 mph is highly unlikely to survive at any time of day. But if pedestrian fatalities are significantly more likely to occur on 45-mph roadways in dark conditions (Sanders et al. 2022), it would be helpful to understand the root causes of this phenomenon.

This section presents the results of a case-control analysis that builds on recent efforts to examine pedestrian safety specifically in darkness (Sanders et al. 2022; Ferenchak and Abadi 2021; Kumfer et al. 2019). Through this work, we intentionally controlled for the known risk factor of functional street classification to try to further isolate and identify the factors associated with an increased likelihood of a pedestrian fatality or severe injury in dark conditions.

Methodology and Data

Methodology

This study uses a case-control methodology to evaluate the likelihood of certain variables being significantly associated with a fatal or severe pedestrian injury in darkness. Case-control studies originate in the field of public health, where they were developed to analyze rare events (and correspondingly small sample sizes) with additional rigor. In this case, while the overall number of pedestrians who are killed or severely injured in the U.S. has steadily increased in recent years, the number dying or being severely injured in any one year in any one city is relatively small from a statistical perspective, which makes the case-control methodology an attractive option for deeper analysis of the many potential factors associated with these fatalities and severe injuries. This study focuses on motorist-pedestrian crashes from the years 2015-2019 to reflect recent trends, including the steep rise in pedestrian fatalities in the last decade, while precluding pandemic-era effects from clouding the findings. In seeking a balance between statistical rigor and the study timeline and budget, we elected from the outset to gather a sample of 100 cases and 100 controls for each case city.

A key component of case-control studies is selecting an appropriate matching characteristic for the comparison. Cases and controls are selected from datasets matched on certain criteria to ensure that they are sufficiently comparable to provide insights about other variables of interest once they are evaluated in comparison to one another. In addition to being a key aspect of the case-control methodology, this matching helps to prevent the crediting of conclusions to certain variables when they are more substantially related to other, often underlying variables. For example, controlling for functional class while allowing variation in number of lanes and posted speed allows greater insight into the role of the two latter variables when functional class might otherwise dominate in a regression model.

Case-control studies can also provide deeper investigation into known risk factors by controlling for those factors themselves. In this case, research has established that most pedestrian fatalities at night occur along urban arterials, with a disproportionate number at midblock locations. But urban arterials comprise dozens if not hundreds of roadway miles in most urban and suburban areas, and most of those locations do not experience a pedestrian fatality or serious injury, even when they appear to be equally high risk. Thus, this study sought to provide insight into factors associated with a fatal or severe pedestrian injury in darkness by controlling for functional class and then examining if and how various built environment and land use factors differ between locations that have experienced a fatal or severe pedestrian injury in darkness (cases) and matched locations that have not experienced a fatal or severe pedestrian injury in the same timeframe (controls). In this vein, this study used functional classification (primary and secondary arterials and major collectors) and location type (midblock and unsignalized intersections) as the matching criteria. Variables that were not controlled for through matching, including other elements of roadway design and exposure proxies, were then explicitly examined as potential contributors to the outcome of interest (fatal or severe pedestrian crash in darkness).

Case City Selection

Identifying potential case cities

In order to conduct this analysis, we first identified a list of potential case cities for subsequent analysis. To identify this list, we examined the number of pedestrian fatalities in cities across the country and selected the 30 cities with the highest numbers of pedestrian fatalities between 2015-2019 (shown in Table 8) as initial candidates.

Table 8. Top 30 Cities to Examine Pedestrian Safety in Darkness (2010-2020 FARS Data)

We sent the list to the project panel and received their concurrence to move forward with additional data collection and evaluation to further refine the list.

Evaluating available data

We then conducted a high-level investigation into the presence and type of data available via online search, using the following questions to frame our search:

• Are crash data available for public download?
  • – Are pedestrian crashes easily identifiable within the data? To determine this, one may need to download the state’s crash data manual.
  • – Are data in a flat file (e.g., one row per crash), or are there multiple files (e.g., for victim, collision, and party) that can be triangulated for more information?
• Are roadway design and operations data available for public download via the target city’s website?
  • – If not available for download, is there information on the website that indicates a contact person and willingness to share the data?
  • – What data are available, and in what format (e.g., integrated network file v. separate files for number of lanes, signalization, average annual daily traffic (AADT), sidewalks, etc.)?
  • – Is there a readily available data dictionary for each dataset?

At the end of this step, we narrowed our case city options from the original 30 to the 16 cities shown in Table 9.

Digging deeper to determine data quality

The third step in the process included a more detailed look at the presence and type of data available via online search and city contacts where necessary.

Key elements of this investigation included:

• What elements are included in the crash data?
  • – Fields for pedestrian pre-crash action/movement and location.
  • – Availability of crash severity levels within the data (i.e., are each of the KABCO [a commonly used injury severity scale in police crash reports, with the following values: K = Killed, A = Suspected serious injury, B = Suspected moderate injury, C = Suspected minor injury, O = Property damage only] values available, or are the data aggregated into fatality v. injury v. non-injury?)
• Are roadway design and operations data available for public download via the target city’s website?
  • – How are the data coded relative to spatial units?
  • – Are the data available network-wide, or are they available in patches throughout the network, e.g., isolated counts?
  • – In what format are the data (e.g., tabular, spatial, PDF)?
  • – Are there primary keys available for joining the data?
• Are pedestrian and/or bike exposure data (e.g., average annual daily pedestrians (AADP) and/or average annual daily bicyclists (AADB) or counts) available? [While pedestrian safety is the primary concern of this study, bicycle volumes were of interest as a potential proxy for pedestrian activity at this screening stage.]
  • – Actual pedestrian and/or bike counts and/or exposure models
  • – Pedestrian/bike-specific page with additional information, even if not counts?
  • – Is the city represented in the Ped/Bike Data Portal (https://bikeped.trec.pdx.edu/)?
• Other data of interest:
  • – Turning movement counts (e.g., as related to intersection timing updates)
  • – Vendor data subscriptions (e.g., Streetlight, Strava)
  • – Observed speed data
  • – Pedestrian-specific infrastructure: sidewalks, curb bulb-outs, median crossing islands, midblock crossings, etc.

We also examined the number of fatal and serious pedestrian injuries specifically in darkness, as well as the number that occurred midblock or at unsignalized locations along arterial or collector streets – the areas of primary interest for our study. Along the way, we encountered barriers to getting crash data in several locations, whether due to a lack of key information within the data or a lack of access. For this reason, several of the top 16 cities fell out of contention and the city of Portland, OR, was added to our finalized list. At the end of this step, we narrowed our case city options from 16 to the six cities listed in Table 10 and began to build the city-specific datasets.

Table 10. Final Six Cities to Examine Pedestrian Safety in Darkness (2014-2020 FARS Data)

Data and Quality Control

After finalizing the six case cities, we created city-specific datasets from a combination of police-reported crashes, manually collected data, city data, open-source data, and publicly available data, as described below.

Creating city-specific datasets

We created the city-specific datasets through a selection of locations that fit certain criteria. For cases, these criteria included:

• Location of a police-reported pedestrian fatality or serious injury during dark conditions between the years 2015-2019
• Location midblock or at unsignalized intersection
• Location along an arterial or collector roadway

For controls, the criteria included:

• Not having been the location of a police-reported pedestrian fatality or serious injury at any point between the years 2015-2019
• Location midblock or at unsignalized intersection
• Location along an arterial or collector roadway
• For midblock locations: block length of at least 200 feet in length

At the outset of the study, we intended to just focus on pedestrian fatalities on urban arterials at midblock locations, based on findings from Tefft et al. (2021) about the predominance of urban arterials and midblock locations for pedestrian fatalities at night. However, upon finding that there were not enough cases for these fairly restrictive criteria within our case cities, we expanded our matching criteria in the following ways:

• We added secondary arterials and major collectors, which have similar characteristics to primary arterials in many cities, and
• We added unsignalized intersection locations, which are often problematic locations for pedestrian safety along arterials due to the driver priority (similar to midblock locations) that is communicated by the lack of signal control for the intersecting street.

Control locations were randomly selected using only the segment that met the above requirements. The final balance between midblock and unsignalized intersections was a 50/50 split.

Each case and control location was treated as a point. For intersections, crashes are often geo-located to the intersection centroid, so that convention was kept for this study. For midblock locations, the midpoint of the segment was selected. When contextualizing the locations, buffers were used for the contextualization process, which varied by attribute.

We also added severe pedestrian injuries to the pedestrian fatalities when needed to increase the sample size. For some cities, the numbers were large enough that we used a random selection within the pool of cases that match the criteria, up to 100 cases. For other cities, we used all available cases to get as close to 100 cases as possible. Control locations were always randomly selected, then assigned a faux “crash year” in proportion to the distribution of years in the case set. Both groups were reviewed for quality control before proceeding with data collection. Our final analysis consisted of 539 cases and 595 controls, for a total of 1134 locations. The distribution of cases and controls by year and by city follows in Table 11 and Table 12. Table 13 shows the distribution of fatal and severe crashes by city.

Due to the need for an adequate size for the analysis, we used all of the cases that were available in most of our case cities. However, we conducted exploratory analysis of the crashes at the case locations to understand if and how the crash types at these locations differed, which could influence our interpretation of results. As Table 14 shows, the majority of the fatal and severe pedestrian injury crashes in our case dataset involved a pedestrian who was crossing the street (76.4%) and a motorist going straight (90%). This distribution was similar by location type (intersection v. midblock). These findings support the case-control approach and the ability to draw broader conclusions from this work.

Table 14. Distribution of Case Crashes by Location Type, Pedestrian Pre-Crash Movement, and Motorist Pre-Crash Movement (Crash Type)¹

¹ Data from Houston and Charlotte excluded from this table due to lack of usable information about pedestrian actions and driver movements.

Manual data collection

To collect the necessary contextual data for each case and control site, we developed a data collection protocol and used the Fulcrum™ app to help ensure consistency in data collection and input. Project team members were trained on the data collection app, which included up to 37 separate fields for data collection when all fields were applicable. The data collection protocol was designed to gather variables that previous research, including research reviewed through this project’s extensive literature review and our macro-level pedestrian fatal crash analysis, had found to be significantly or likely related to pedestrian crash risk. These variables included roadway design and operations variables that were difficult to find via other reliable data sources, such as number of through and turn lanes, posted speed limit, presence and completeness of sidewalks, presence and type of street lighting, and the presence and type of pedestrian countermeasures. Other variables of primary interest for manual data collection included land use variables related to pedestrian activity, such as the presence and type of housing and commercial uses such as grocery, convenience, and liquor stores. A selection of the data collected at each location is represented in Figure 5; additional data from the manual data collection effort are shown in Table 15 through Table 17.

The data collection team was given a set of 60 locations (case/control status was concealed) to pilot-test the data collection process using Google Maps™ and Google Streetview™. Streetview generally contains annual images of the sites, which allowed the data collection process to reflect site conditions at the time of the crash (or the faux crash year, in the case of controls). In the event that an image from the crash year was not available, the team was instructed to use the closest image, with priority given to past, rather than future, years. The team collected a sample of 60 sites, evaluated and slightly revised the data collection protocol, and then proceeded with the remainder of data collection. In the event of construction or other problems with a location, queries were sent to the data analysis team and replacements were generated when possible (i.e., for controls and when there were additional case options).

Quality control

Quality control was conducted in rounds. The first round occurred after the data collection form was piloted on 60 locations. After the data were gathered for the 60 locations, a member of the analysis team reviewed every 10th location to assess the data collection and determine whether the data collection was proceeding as planned. The review effort revealed a couple of fields for which the coding options did not sufficiently cover conditions on the ground. These fields were revised and the data collection team received additional instruction and proceeded with data collection for the remaining sites across the six cities.

The second round of quality control occurred once the data collection was finished. At that point, a member of the analysis team examined every 15th to 20th record in detail to look for discrepancies between the collected data and the data available online. Because each location could have information for up to 37 variables, there were occasional errors, usually in the range of 1 or 2 errors for every few records. When a discrepancy was identified, the project’s research director was brought in for review and made the final determination as to the accurate data collection entry, and the record was updated and noted in the team’s quality control records. Given the extensive nature of the manual data collection, some degree of error is to be expected. While we corrected what we could, it was beyond scope and budget to exhaustively re-review each data record, and variables that were found to be consistently less reliable (generally those being measured via online tools) have been excluded from the analysis in this report. We recorded the discrepancies and corrected records and estimate a manual data collection error rate of less than 3 percent for the full dataset.

Open-source and public data

After completing this QA/QC process, we matched the geocoded location for each record to contextual variables from larger open-source and publicly available datasets to allow for additional analysis. These data included infrastructure data from Open Street Maps (www.openstreetmap.org); data from the Environmental Protection Agency’s Smart Location Database (EPA SLD) (https://www.epa.gov/smartgrowth/smart-location-mapping, 2019 vintage), on housing density, road network density and characteristics, and jobs; sociodemographic data from the U.S. Census American Community Survey (ACS) (https://www.census.gov/, five-year estimates from 2015-2019); transit stop and service data from the General Transit Feed Specification (GTFS) dataset (https://gtfs.org/); and highway performance monitoring system (HPMS) data, which provides information on functional classification, roadway ownership, and AADT

(https://www.fhwa.dot.gov/policyinformation/hpms/fieldmanual/page05.cfm#toc249159691).

All census and EPA SLD values were calculated at the Census block group (CBG) level using an area-weighted proportional aggregation approach for the ½-mile radius around each case and control location. Using GTFS data, service level at each transit stop was estimated using a systematic sample of weekdays and weekends in October for each year of crash data. The number of times a transit vehicle stopped at each stop location per day was tabulated and averaged over all the days in the sample. These data were assigned to case and control locations by aggregation - the distance to nearest stop and the service level estimate for that stop, the count of stops and sum of service level estimates for all stops within 200 feet, and the count of stops and sum of service level estimates for all stops within ¼ mile.

Exposure data

We know that pedestrian exposure plays some role in pedestrian risk – if no pedestrians are in an area, no pedestrians will be hit, regardless of other present risk factors. However, no pedestrian exposure data were available in any of the case cities, so this study used proxies in the form of pedestrian activity generators (e.g., certain land uses, like convenience stores), census variables known to have a relationship with pedestrian activity, such as percentage of residents commuting to work by walking and total

Table 15. Summary Statistics for Continuous Variables of Interest in the Dataset

1 All ACS and EPA SLD variables calculated for ½-mile radius around the case and control locations.

2 EPA SLD documentation defines auto-oriented facilities as: “Any controlled access highway, tollway, highway ramp, or other facility on which automobiles are allowed but pedestrians are restricted; any link having a speed category value of 3 or lower (speeds are 55 mph or higher); any link having a speed category value of 4 (between 41 and 54 mph) where car travel is restricted to one-way traffic; or any link having four or more lanes of travel in a single direction (implied eight lanes bi-directional – turn lanes and other auxiliary lanes are not counted). For all of the above, ferries and parking lot roads were excluded.”

3 EPA SLD documentation defines multimodal facilities as: “Any link having a speed category of 4 (between 41 and 54 mph) where car travel is permitted in both directions; any link having a speed category of 5 (between 31 and 40mph); or any link having a speed category of 6 (between 21 and 30 mph) where car travel is restricted to one-way traffic. For all of the above, autos and pedestrians must be permitted on the link. For all of the above, controlled access highways, tollways, highway ramps, ferries, parking lot roads, tunnels, and facilities having four or more lanes of travel in a single direction (implied eight lanes bi-directional) are excluded.”

4 EPA SLD documentation defines pedestrian-oriented facilities as: Any link having a speed category of 6 (between 21 and 30 mph) where car travel is permitted in both directions; any link having a speed category of 7 or higher (less than 21mph); any link having a speed category of 6 (between 21 and 30mph); or any pathway or trail (pending data availability) on which automobile travel is not permitted (speed category 8). For all of the above, pedestrians must be permitted on the link. For all of the above, controlled access highways, tollways, highway ramps, ferries, parking lot roads, tunnels, and facilities having four or more lanes of travel in a single direction (implied eight lanes bi-directional) are excluded.

Findings

Bivariate Data Exploration

We conducted an exploratory analysis of the data to understand the degree to which the cases (locations with at least one fatal or severe pedestrian crash in darkness between 2015 and 2019) and controls (locations with no fatal or severe pedestrian crash in any lighting condition between 2015 and 2019) differed with regard to variables of interest such as roadway design and operational characteristics, nearby land uses, and the sociodemographic characteristics and travel behavior of the surrounding population. All p-values in this section reflect the distribution between the two variables, and thus the p-value is listed on the line of the variable name. For example, the posted speed limit distribution is significantly related to case-control status, with a p-value of 0.10. Relationships between case-control status and each value of the variable are further explored in the conditional logistic regression section.

Roadway Design and Operations

Table 16 shows significant bivariate relationships between contextual and design variables and the cases and controls. Immediately, we see that cases are significantly more likely to have occurred on a roadway with two or more through lanes, a posted speed limit of 35 mph, at least one turn lane, and that operate bidirectionally. These findings are likely influenced by pedestrian exposure to some degree, whether through pedestrian route choice or the location of pedestrian activity generators (explored further below). Additionally, despite selecting controls via random selection from within a control population matched to the functional class and location type of the case population, cases were significantly more likely to be located along primary roads (i.e., principal arterials) compared to secondary roads (minor arterials or major collectors).

Table 16. Roadway Design and Operations Variables by Case-Control Status

Significance indicated by bolding and the following: *** p ≤ 0.001, ** p ≤ 0.01, * p ≤ 0.05, ^#p ≤ 0.10

Countermeasures

Pedestrian countermeasures aim to improve pedestrian safety by influencing driver and pedestrian behavior, separating pedestrian and driver movements, and/or increasing pedestrian visibility. While countermeasures should not be needed to help pedestrians safely cross, given that all cities in this study are governed by laws or vehicle codes stating that drivers must yield to pedestrians trying to cross, countermeasures are often critical to encouraging that yielding. Yet less than 10 percent of the case and control locations had some type of crossing countermeasure, as shown in Table 17. Where countermeasures were present, their prevalence varied between the two groups. Control locations were significantly (p ≤ 0.05) more likely to have advanced stop lines, while case locations were marginally significantly (p ≤ 0.10) more likely to have high-visibility crosswalk markings. However, there was no significant difference between case and control locations with regard to the presence of any type of crosswalk, the quality of crosswalk markings, or the presence of at least one countermeasure of any type.

In contrast, the majority of locations had complete sidewalks on both sides and street lighting on at least one side, although these two countermeasures were significantly more likely to be at case locations. This association almost assuredly reflects that cases were also significantly more likely to occur near pedestrian attractors like stores and along larger roadways, rather than any risk from the countermeasure itself.

Table 17. Countermeasure Presence and Type by Case-Control Status

Significance indicated by bolding and the following: *** p ≤ 0.001, ** p ≤ 0.01, * p ≤ 0.05, ^#p ≤ 0.10

Land Use

Several land uses were highly significantly associated with case-control status (see Table 18). In particular, cases were significantly more likely to be co-located with a liquor store, convenience store, grocery store, or any type of commercial center, although only marginally more likely to have a drug or marijuana store on the segment or at the intersection. Cases were also significantly more likely to be co-located with high-density housing, while being significantly less likely than controls to be co-located with low-density housing. While the commercial and housing variables are significantly correlated with case-control status, high-density housing is not significantly correlated with either low- or high-density commercial, suggesting that they are not operating as proxies for one another.

Table 18. Land Use Presence and Type by Case-Control Status

Significance indicated by bolding and the following: *** p ≤ 0.001, ** p ≤ 0.01, * p ≤ 0.05, ^#p ≤ 0.10

Transit

Finally, there were also highly significant associations between case-control status and the presence of a transit stop, both within 200 feet and within one-quarter of a mile (see Table 19). This finding likely reflects pedestrian exposure, underscoring the importance of providing safe pedestrian crossings and facilities at transit stops.

Multivariate Analysis

The next step in the analysis were multivariate analyses that explored the many significant bivariate relationships to home in on the most significant factors associated with case versus control locations. Because of the large number of variables in the dataset, we used random forest analysis to help reduce the data under consideration for the subsequent regression modeling. After reviewing the results of the random forest analysis, we used conditional logistic regression to further explore the influence of various factors on the likelihood of being a case or control observation.

Random Forest

Random forest is a data reduction technique that can make subsequent modeling more efficient. Because of the high number of variables in our dataset and the likelihood of correlation between, particularly, some of the census and EPA SLD variables, using tools to help with data reduction is particularly important. The results of the random forest analysis, displayed in Figure 6, include two different measures – Mean Decrease in Accuracy and Mean Decrease in the Gini Coefficient – to depict the relative importance of each variable. The accuracy plot shows the negative impact of leaving out any one variable; the Gini model reflects the purity of the nodes at the end of the tree. The plots are not expected to match exactly. For both plots, a higher score reflects relative greater importance in predicting case-control status.

Conditional Logistic Regression

Based on the results of the bivariate significance testing, random forest analysis, and our understanding of past research and transportation and public health theory, we began to test varying combinations of variables with a significant bivariate relationship to the case-control status in a multivariate conditional logistic regression model. We used conditional logistic regression to control for the matching variable used in the design of the case-control sample. While we did not test every possible combination of these variables, we iterated on this model dozens of times, using a combination of both forward and backward stepwise regression and guided by theory, the random forest results, and each subsequent regression model’s results. No model included predictor variables that were correlated with one another above a Pearson’s score of 0.6, a generally accepted cut-off. Our preferred final model homes in on several land use and population characteristics of the half mile surrounding each case and control location that are significantly associated with case-control status, as shown in Table 20. Among the land use variables, we see a significant and positive association with cases and the presence of (separately) a convenience store (p ≤ 0.01), grocery store (p ≤ 0.001), liquor store (p ≤ 0.05), and other low-density commercial (p ≤ 0.05). We see a significant and negative association between cases and the presence of low-density housing (p ≤ 0.05).