category: literaturenote citekey: leeassociationbetweenintersectioncharacteristics2016 title: Association between intersection characteristics and perceived crash risk among school-aged children authors: "Lee, Gain; Park, Yuna; Kim, Jeongseob; Cho, Gi-Hyoug" year: 2016 date: 2016-00-00 2016 doi: 10.1016/j.aap.2016.09.001 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2016.09.001" zotero_key: SFFCCV9S zotero_storage: X9WX5FQT collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Lee, Gain"

status: converted

Contents lists available at ScienceDirect

Accident Analysis and Prevention

j ournal homepage: www.elsevier.com/locate/aap

Association between intersection characteristics and perceived crash risk among school-aged children

Gain Lee, Yuna Park, Jeongseob Kim, Gi-Hyoug Cho^∗

Ulsan National Institute of Science and Technology School of Urban and Environmental Engineering, Korea

a r t i c l e i n f o

Article history: Received 18 February 2016 Received in revised form 7 July 2016 Accepted 2 September 2016 Available online 7 September 2016

Keywords: Perceived crash risk Pedestrian Intersection Safety

a b s t r a c t

This research examined how environmental attributes near intersections influence the perceived crash risk among school-aged children, which provides information on the potential risks of pedestrian crashes that can guide the development of proactive countermeasures. In a sample of 799 children aged 10–12 years old in Korea, the environmental attributes of intersections perceived as having a high risk of producing crashes near elementary schools were investigated using standard negative binomial and zero-inflated negative binomial models.The results showed that a higher number of student crossings, a wider road width, the presence of crosswalks, student-friendly facilities at the intersection, and fourway intersections were significant and positively associated with perceived crash risk among school-aged children. The findings related to building characteristics indicated that a higher number of entrances at an intersection increased the perceived crash risk while higher visibility at the intersection reduced the perception of risk. Associations with traffic-calming measures were weak,suggesting that the measures used in the study areas were not effective in reducing the perceived crash risk. The results of a policereported crash model showed that school-aged children have a relatively accurate perception of crash risk and that the perceived crash risk of school-aged children may provide valuable information on the intersection characteristics in need of attention near school sites.

1. Introduction

Road traffic injuries are one of the leading causes of childhood death and disability worldwide (World Health Organization [WHO], 2007; Peden, 2008). Children are particularly vulnerable to road traffic crashes (Stoker et al., 2015). In low-income and middleincome countries, children account for 30–40% of all road traffic deaths (WHO, 2007). In Korea, the annual number of child fatalities resulting from traffic accidents is decreasing, but the rate is still approximately 50% higher than the fatality rate of other developed countries (Lee and Lee, 2014). In particular, 65% of children killed by traffic accidents are victims of a pedestrian crash.

One of the interventions for reducing pedestrian crashes is the modification of the road environment (Retting et al., 2003; WHO, 2007). The existing literature has focused on identifying the built environment or road design characteristics that potentially increase the hazards for pedestrians. Researchers have reported

E-mail address: gicho@unist.ac.kr (G.-H. Cho).

that a higher population density (Clifton and Kreamer-Fults, 2007; Dissanayake et al., 2009; Elias and Shiftan, 2014; Rothman et al., 2014), commercial use (Dai et al., 2010; Yu and Zhu, 2015), a greater number of transit stops (Ukkusuri et al., 2012; Yu and Zhu, 2015), a higher number of lanes (Ukkusuri et al., 2012), and the absence of traffic signals (Moudon et al., 2011) or traffic-calming devices (Rothman et al., 2014) were all associated with a higher number of pedestrian crashes. These studies have used police-reported pedestrian crash data as their main source of information on pedestrian crash risk.

In contrast to police-reported pedestrian crashes, risk perception is a subjective assessment of the probability of experiencing a negative event(Lund and Rundmo, 2009). In the field of road safety, risk-perception techniques have been used to examine associations betweenthe behavior of roadusers andperceived risk and tounderstand how road users identify risk factors (Diógenes et al., 2010). The main advantage of using perceived crash risk in research is that it provides the possibility of assessing road safety even when crashes have not occurred (Diógenes et al., 2010). Since pedestrian crashes are relatively rare events, perceived crash risk can provide predictive information on the potential risk of pedestrian crashes and guide the development of proactive countermeasures to reduce

^∗ Corresponding author at: School of Urban and Environmental Engineering Building 106, 801-7 Ulsan, Korea.

the possibility of such crashes occurring (Schneider et al., 2004). Previous literature has shown that perceived crash risk tends to be positively correlated with police-reported crash risk data (Elvik and Bjørnskau, 2005), but sometimes perceived crash risk can provide information that cannot be addressed only with actual crash data (Cho et al., 2009). Studies on perceived crash risk can provide a better understanding of pedestrian behavior in different road environments (Hine, 1996) and potentially mitigate pedestrian crashes (Diógenes et al., 2010).

The present study examined the road environment factors at intersections associated with perceived crash risk among children by focusing on the school–home journey, which is a point of considerable traffic exposure and risk for children (Stoker et al., 2015). First, we intended to identify locations perceived as having a high risk of crashes near elementary schools and the roadway intersection characteristics associated with perceived crash risk. Then, through a comparison with the intersection characteristics associated with higher police-reported crash risk, this study examines the implications and potential usefulness of perceived crash risk in understanding pedestrian safety. Ultimately, this study aims to inform policy for creating safe routes to schoolthat reduce the crash risk to children in school zones based on the study's findings.

2. Background

Modifying the road environment is one of the main interventions for reducing the likelihood of children and youth being involved in road traffic crashes (WHO, 2007). Retting et al. (2003) suggested that engineering modifications to reduce pedestrian–vehicle crashes may generally be classified into three approaches: managing vehicle speeds, separating pedestrians and vehicles, and increasing pedestrian visibility. The evidence indicates that higher speed limits (Tiwari et al., 2007), a greater number of transit stops (Pulgurtha and Sambhara, 2011; Ukkusuri et al., 2012), an uncontrolled left-turn lane (Chin and Quddus, 2003), and wide lanes (Noland and Oh, 2004 Goldstein, 2006; Ukkusuri et al., 2012) are all associated with a greater crash risk. Conversely, proper roadway lighting design (Goldstein, 2006), street segments with sidewalks (Berhanu, 2004), and the installation of traffic-calming measures (Ewing, 2001) reduce the rates of vehicle–pedestrian crashes. Some studies have focused on the relationship between neighborhood characteristics and pedestrian safety, reporting that a higher population density (Pulugurtha and Sambhara, 2011), mixed land use (Wier et al., 2009; Dai et al., 2010; Ukkusuri et al., 2012), and retail uses (Wedagama et al., 2006) are associated with a greater frequency of pedestrian crashes.

A relatively small number of studies have examined the environmental attributes associated with pedestrian crashes involving children in general or crashes near schools in particular. Consistent with the results of studies targeting the general population, those studies conducted near schools have reported that a higher population density, mixed land use, and a higher number of transit stops are all associated with an increase of pedestrian crashes involving children (Clifton and Kreamer-Fults, 2007; Dissanayake et al., 2009; Elias and Shiftan, 2014; Yu and Zhu, 2015) while traffic-calming measures (Jones et al., 2005; Rothman et al., 2014) and more sidewalk coverage (Yu and Zhu, 2015) were found to be associated with lower rates of pedestrian crashes.

Whereas a significant body of research has focused on the environmental attributes associated with police-reported pedestrian crashes, the perceived risk from the child's perspective has not received much attention. Unlike police-reported crash data, perceived crash risk can be understood as a multifaceted concept because it involves the subjective interpretation of risks in various transport situations (Deery, 2000). Perceived crash risk is frequently defined as a level of awareness, with proper awareness toward a prospective risk, allowing individuals to avoid accident risk (Slovic, 1987). The literature examining the perceived risk of different modes of transport has shown that risk perception among the public is broadly correct (Elvik and Bjørnskau, 2005). However, children's ability to accurately identify the potential risk of pedestrian crashes is not always satisfactory (Meir et al., 2015). Often, young children do not correctly understand complex traffic situations in the same way as adults do, and their inaccurate awareness of impending danger increases their risk of being involved in road traffic crashes (WHO, 2007). Understanding the ways in which children perceive potential risks in road environments could provide some important information on children's behavior and this could be used to inform policy, thus improving safety.

Furthermore, the perceived crash risk provides information on potentially hazardous situations prior to their actualization (Meir et al., 2015). In studies on pedestrian crashes, it is often difficult to identify the risk factors associated with crashes using statistical methods since pedestrian crashes are rare events (Diógenes et al., 2010). In such cases, perception information can be used to identify locations where an accident is waiting to happen because a high level of perceived crash risk can imply that there is the potential danger of a crash even at locations where no crashes have yet occurred (Schneider et al., 2004). Cho et al. (2009), for example, examined how perceived and actual crash risk are related to the built environment characteristics in Montogomery County, MD in the Washington DC metropolitan area. Their study found that a low population density and non-mixed land use increase the perception of crash risk, although these characteristics reduce the actual crash risk in the neighborhood. Their results indicated that policereported crash risk data does not provide complete information on traffic safety because people change their behaviors to avoid exposure to high perceived-risk areas; thus, proactive interventions are desirable where the perceived crash risk is high. Schneider et al. (2004) analyzed perceived crash risk on the campus of the University of North Carolina at Chapel Hill and found that pedestrians reported high perceived risk at locations near the main campus where the actual crash risk was low. The study concluded that high perceived risk implies the need for proactive interventions to prevent future pedestrian crashes.

In sum, the perceived crash risk among children at school sites conveys information thatis importantfor enhancing the road safety of children by addressing the level of awareness for pedestrian crashes and indicating the potential crash risk to children at specific locations. In addition, children's perceived crash risk has a mediating role in achieving the successful implementation of a safe route to school program that encourages walking and bicycling (McMillan, 2005). The literature has shown that some of the factors influencing children's walking to and from school are safety concerns (Timperio et al., 2006; Zhou et al., 2010; Trapp et al., 2012).

However, the relationship between locations perceived as having a high risk of pedestrian crashes and the environmental attributes of school sites has rarely been examined. In particular, few studies have attempted to identify the perceived crash risk for school-aged children. The purpose of this paper, therefore, is to identify locations that school-aged children perceive as dangerous and what environmental factors are associated with their risk perception. Hypothetically, the patterns of perceived crash risk for school-aged children would not be identical to the patterns found in the police-reported crash risk data. Young children may not correctly understand the risk of traffic crashes or it would be difficult to identify the environmental attributes of crash locations using the data from the limited number of police-reported crashes. This study aims to determine the implications of perceived crash risk for pedestrian safety through a comparison with the patterns found in the data for police-reported crashes.

The present study addresses some gaps in the previous studies on perceived crash risk. Few studies, for example, have assessed the perceived crash risk among school-aged children in an analytical manner. Parents' perceptions of the environmental conditions along the route to school have been documented (Carver et al., 2010; Kerr et al., 2006), but perceptions of environments may differ by age (Nasar, 2008), which suggests that parents and children may have different risk perceptions of the same environments near schools (Napier et al., 2011). In particular, the perceived crash risk among school-aged children themselves is important in the context of the study site because the majority of children enrolled in elementary school in Korea walk to school without their parents. Second, we have attempted to include the attributes of buildings near intersections as risk factors that undermine pedestrian safety. The existing literature has heavily focused on the characteristics of roadway design and neighborhood characteristics that may raise or minimize the probability of pedestrian injury. However, in dense urban settings, the attributes of buildings near intersections may affect the visibility of pedestrians and pedestrian/traffic volumes. By measuring visibility and the number of entrances into buildings, this study additionally accounts for the characteristics of buildings in explaining perceived crash risk.

3. Methodology

This study used five complementary data sources for the analysis: 1) survey items asking school-aged children about perceived crash risk, 2) locations perceived as having a high risk of pedestrian crashes, 3) locations of police-reported pedestrian crashes, 4) walking routes to schools, and 5) the environmental attributes of intersections. Information about the first two sources and the fourth source regarding walking routes to schools was collected from a survey conducted in July 2015. To compare the results of perceived crash risk with the actual crash risk, geo-coded policereported crashes that occurred between 2007 and 2014 were retrieved from the Traffic Safety Information Management Complex SysteminKorea (TMACS;tmacs.ts2020.kr). The environmental attributes of intersections were collected through GIS measurements of local data and street audits using a Naver street-view system (map.naver.com). This section details the recruitment of the study's participants, the collection of data, and the statistical methods applied to the analysis.

3.1. Study area and participants

The study location is Ulsan, which is the largest industrial city in Korea. Its total population is approximately 1.2 million (the seventh largest city in Korea) (Park et al., 2016). In 2013, the median household income for Ulsan was \$44,971, which is significantly higher than the national average (\$37,014).

The selected school sites were eight elementary school zones in the Dong district, where heavy industries and shipyards are located (Fig. 1). The school zones are located in a relatively small geographical area (3.59 km2) and the distance from the downtown city area is approximately 10 km. The number of enrolled students in the eight elementary schools was 4130 in 2015. The land use at the selected sites was mainly residential but included mixed commercial-industrial areas. Thenumber ofpolice-reportedpedestrian crashes in this area between 2010 and 2014 was 792, which is the highest level of crash occurrence in Ulsan.

The study participants were fifth- and sixth-grade students (10–12 years old) enrolled in eight elementary schools. We excluded students below the fifth grade because they may have difficulty in identifying the potential hazards affecting pedestrian crashes (Meir et al., 2015). The participants were asked to fill out a survey form in the classroom. The survey comprised of a series of questions regarding demographic features, the mode of travel to school, and the child's perception of crash risk. A total of 1263 students returned the survey. Excluding 221 students who failed to complete the questionnaire, 1042 students were initially chosen as valid samples. The dominant mode of travel to school was walking (91.7%). Only 6.3% of students used motorized travel modes, either a car or bus. Among the 956 students who commute to schools by walking, 799 adhered to the protocol of the survey and successfully filled out all the responses for locations perceived as having a high risk of pedestrian crashes and walking routes to schools. The final sample used for the analysis included 799 students who commuted to school by walking.

3.2. Outcomes: perceived and actual crash risk at intersections

In addition to questions asking about gender, age, and experience of traffic crashes, transport risk perception was measured using thirteen items on a five-point Likert scale. The response categories for each question ranged from "strongly disagree" to "strongly agree," with higher scores indicating a higher risk perception. To evaluate the risk perceptions for sidewalk conditions, visibility, crossing aids, and traffic concerns, three questions were included for each dimension, and one question was included to evaluate street lighting. In order to derive a single indicator for each feature, we measured the average values for the three questions for each dimension: sidewalk conditions, visibility, crossing aids, and traffic concerns. Using the Student's t-test, responses were compared between grades, gender, and experience of pedestrian crashes.

The risk-perception survey included a map of school zones that illustrated the locations of the main buildings and schools, the crosswalks, and the bus stops. Participants were instructed to mark a couple of locations with red-colored stickers that they believed had the highest risk of pedestrian crashes and, with yellow-colored stickers, an additional couple of locations were marked having a moderate-high risk of pedestrian crashes. Excluding invalid responses, 799 students reported 2621 locations in the study area. We digitized those locations they perceived as having a high risk of pedestrian crashes using ArcGIS 10.2.

Unlike previous studies examining associations between perceived crash risk and demographic features (Lund and Rundmo, 2009; Nordfjærn and Rundmo, 2010; Chaurand and Delhomme, 2013), we were interested in identifying the characteristics of roadway intersections related to perceived crash risk. Thus, the locations perceived as having a high crash risk were required to be spatially matched to a corresponding intersection. Consistent with the definition of a road intersection in Schneider et al.'s (2010) study, we defined a road intersection area as a 15-m (50-ft) buffered zone from the crossing point of two road centerlines. There were 546 intersections located in the study area. The number of risk points located at each intersection was used to quantify the level of perceived crash risk for an intersection. Occasionally, the points perceived as having a high risk were located within multiple intersections. In such cases, the points were assigned to the closest intersection.

Locations of police-reported pedestrian crashes are downloadable from TMACS. TMACS provides information regarding crash type, injury severity, the dates of crashes, and the locations where two ormore crashes have occurred in a year.We selected the pedestrian crashes that occurred within a 15-m buffered zone from the 546 intersections and geo-coded them using ArcGIS 10.2. Overall, 217 pedestrian crashes occurred within the study intersections between 2007 and 2014.

Fig. 1. Location of the study area.

3.3. Independent variables: environmental attributes

Intersection audits included measures in five categories: 1) risk exposure, 2) familiarity, 3) traffic-calming devices and traffic signs, 4) road infrastructure, and 5) buildings at intersections. Besides familiarity, these features were selected based on existing studies (Lee and Abdel-Aty, 2005; Abdel-Aty et al., 2007; Dai et al., 2010; Schneider et al., 2010; Moudon et al., 2011; Aziz et al., 2013; Yu and Zhu, 2015) that examined the environmental correlates of police-reported pedestrian crashes. Risk exposure is one of the most important factors for understanding pedestrian safety. Ideally, pedestrian counts for each intersection would be an accurate measure of pedestrian exposure to risk (Clifton and Kreamer-Fults, 2007). However, no archived pedestrian count data are available for the study areas and manual counts at each intersection are not feasible given the number of intersections. Assuming that students choose regular walking routes to school, we collected data on their walking paths from home to school in order to estimate the number of student crossings at each intersection. Participants were asked to represent their regular walking routes to school on the provided map and we recoded the series of segments and intersections chosen for their routes. By counting the selected number of times the participants crossed intersections on their walking routes to school, a proxy measure to represent the pedestrian crossing count for each intersection was generated. In addition to the number of student crossings, risk exposure includes factors such as population density, street density, and student density. We obtained the population for the study area and the street centerlines in GIS from the Ulsan Statistical Information Service (2010) and Statistical Geographic Information Service (SGIS) in Korea. When an intersection was located at the boundary of multiple census blocks, we used the average values for those blocks.

The dependent variable in this study, perceived crash risk, has unique characteristics because it is a subjectively evaluated level of risk perceived by children. Identifying reliable perceived risk information from the survey of school-aged children is a difficult task since children are very likely to perceive the intersections they frequently visit and are therefore familiar with as locations with a high crash risk. The inclusion of the estimated number of student crossings at the intersections in the model may partly account for the effect of "familiarity." In addition, we defined two more variables to representfamiliarity in order to determine the true effect of intersection attributes when familiarity of the intersection to the children is accounted for. Elementary school-aged children show five types of behaviors when they come to school and go back to home; walking, crossing, waiting, playing, and consuming (Kim and Kwak, 2003). The presence of parks and playgrounds is a feature that attracts children to engage in more outdoor play activities, and the presence of stationery stores, snack bars, and private educational institutes (with student-friendly facilities) is related to more consuming and shopping activities. We audited for the presence of those facilities on the roads connected to the intersections, and then counted the number of roads with parks and playgrounds and student-friendly facilities at each intersection to define the familiarity of the intersection.

In the study area, three types of traffic-calming measures and signs were most consistently used: red-colored pavements at the intersections to enhance awareness, school zone signs to remind drivers to control their speeding impulses, and speed humps. Road infrastructure includes the number of marked crosswalks and sidewalks at an intersection, the maximum width of the road approaching an intersection, the width difference between two intersecting roads, and the number of legs at intersections.

Buildings located at intersections create interactions between the road and the pedestrian. We defined these interactions with two measurements:the number of building entrances atthe pedestrian level and visibility at the adjacent intersection. We assumed that a higher number of entrances might increase accessibility to the locations and imply more interactions between indoor and outdoor spaces in the environment surrounding the intersection. Increasing pedestrian visibility is one of the important engineering interventions for crash prevention (Retting et al., 2003). The built environment in the study area is dense, with residential areas and mixed retail storefronts, and both left and right turns are allowed at most intersections. Therefore, vehicles approaching an intersection are frequently not visible to pedestrians because the buildings located near the intersections block them from view. To measure the level of visibility at each intersection, we calculated the proportion of all pedestrian approaches from which pedestrians were able to observe incoming vehicles at an intersection. The previous literature has consistently reported that commercial or retail uses increase the likelihood of pedestrian crashes (Wedagama et al., 2006; Dai et al., 2010; Ukkusuri et al., 2012; Yu and Zhu, 2015). Building uses include binary variables, indicating the presence of commercial use, residential use, and public use. Table 1 provides definitions and descriptive statistics about the intersection variables used in our analysis.

3.4. Statistical analysis

In this research, two count data regression models were used to examine the association between perceived crash risk and the built environment. The count data for perceived crash risk indicates the number of locations perceived as having a high risk of crashes at each intersection. First, we used the negative binomial (NB) regression model. The standard NB model, with fixed dispersion parameters, is the most widely applied in traffic safety studies since the model is capable of accounting for over-dispersion and the between-location heterogeneity of crash variations. Moreover, model calibration is simple (Bowman et al., 1995; Ukkusuri et al., 2012). The standard NB model can be represented as,

$$Y_i | \theta \sim Poisson(\theta_i)$$

$$\theta_i = \mu_i \exp(\varepsilon_i)$$

where the perceived crash risk at an intersection i is denoted by Yi, which is a Poisson distribution with a mean of ⁱ. The multiplicative random effect of the model following a Gamma distribution is denoted by exp (εi), and i = f xi;ˇ represents a function of area-specific attributes and a measure of risk exposure, where xi is a vector of the environmental attributes of intersections, and ˇ is a vector of the regression parameters estimated from the data (Ukkusuri et al., 2012).

Second, we used a zero-inflated negative binomial(ZINB) model as an alternative count data model to handle problems related to excess zero counts. Our data showed that the proportion of zeroperceived risk intersections was approximately 40% of the sample. The ZINB model assumes that two different processes are at work for zero count data. One is the zero-risk state where the intersection is virtually safe. The other has the intersections at a non-zero count for perceived crash risk (Lee and Mannering, 2002). The ZINB regression model can be expressed as follows:

Yi∼0, with probability pi

$$Y_i | \theta \sim Poisson (\theta_i)$$ with probability $(1 - p_i)$

where pi is a parameter that represents the proportion of zeros added to the NB distribution (Ukkusuri et al., 2012). The ZINB model may be more flexible compared to the NB model since it can handle both over-dispersion due to unobserved heterogeneity and an excess of zero counts (Usman et al., 2010). STATA software (version 13.0) was used to estimate the NB and ZINB regression models.

With regard to the actual occurrence of pedestrian crashes, we ran both the standard NB and ZINB models to identify intersection characteristics that have a statistically significant relationship with the actual crash risk, but the ZINB model for the actual occurrence of pedestrian crashes did not create a model that resulted in convergence. In the results section, this study presents the standard NB model for actual crash risk so as to compare the findings with the results from the perceived crash risk analysis.

4. Results

4.1. Descriptive analysis

Among the 799 students surveyed who commuted to school by walking, slightly more than half of the participants were boys (53.9%). The sixth-grade students accounted for 53.4% ofthe sample and the rest were fifth-grade students. Of the sample, 6.6% of the students had experienced a pedestrian crash in a school zone and over a half of the students (52.7%) answered that they had experienced a "near miss" collision with a car. Of the surveyed students, 98.7% commuted to school without their parents or guardians.

The students' most pressing concern about road safety was the high traffic volume and speed (2.61/5.0) near school sites. They reported a moderate level of concern for sidewalk conditions (2.44/5.0) and crosswalks (2.43/5.0). Compared to girls, the boys consistently reported a lower level of risk perception for sidewalk conditions (t(797) = 2.54, p < 0.01), visibility (t(797) = 4.54, p < 0.001), crosswalks (t(797) = 4.31, p < 0.001), traffic conditions (t(797) = 3.19, p < 0.001), and road lighting (t(797) = 5.20, p < 0.001) (Fig. 2(a)). Previous studies have shown that males tend to underestimate the risks caused by traffic situations compared to females (DeJoy, 1992; Rhodes and Pivik, 2011) because males have a higher perceived sense of control over dangerous situations. For school-aged children, girls perceive the risk of environmental dangers to be higher compared to boys, although there is no significant difference in their appraisals of the relative risk of situations (Hillier and Morrongiello, 1998). Not surprisingly, those who answered that they had almost been involved in a car accident reported higher levels of safety concerns for sidewalk conditions (t(797) = 6.95, p < 0.001), visibility (t(797) = 7.29, p < 0.001), cross-

Table 1 Descriptive statistics of model variables.

Variable Name	Variable Description	Methodology	Mean	Std. Dev.	Min	Max
Outcomes
Perceived Crash Risk	Sum of perceived risk point at the intersection	Summing up the perceived risk points located within 15 m buffered area from the crossing point of two road	3.09	6.31	0.0	52.0
Pedestrian Accident	Number of pedestrian accidents at the intersection	centerlines. Counting the number of pedestrian accidents occurred within 15 m buffered area from the crossing point of two road centerlines (Based on the pedestrian accident black spot (2007–2014) data provided by TMACS)	0.40	1.44	0.0	12.0
Exposure Crossing	Estimated number of daily student crossings at the intersection	Using the collected walking routes to schools, walking route and the series of intersections chosen for the route was recoded in ArcGIS. Then the selected number of times of intersections for walking route to schools was	4.78	9.09	0.0	66.0
PopulationDensity	Population density of 2010	aggregated by the intersection. Population density of the census block where the intersection is located at. When an intersection is located at the boundary of multiple census blocks, the average of	18.59	13.58	0.1	100.2
StreetDensity	Street density of census block	those block was used. Total length of street centerline in each census block divided by the size of census block	55.53	27.33	0.4	125.5
Familiarity s
Parks and playgrounds	Number of pedestrian approaches with parks and playground	Audit the presence of parks and playgrounds on the roads connected to the intersection, and then count the number	0.19	0.51	0.0	3.0
Student-friendly facilities	Number of pedestrian approaches with student-friendly facilities	of roads with parks and playgrounds at the intersection. Define student-friendly facilities are stationery stores, snack bars and private educational institute. Audit the presence of student-friendly facilities on the roads connected to the intersection, and then count the number		1.00	0.0	4.0
Road infrastructure		of roads with student-friendly uses at the intersection.
Sidewalk Crosswalk	Presence of sidewalks Presence of crosswalks	Number of sidewalks on all sides of the intersection Number of marked crosswalks on all sides of the	1.95 0.60	2.31 1.09	0.0 0.0	8.0 4.0
MaxWidth	Maximum width of the road	intersection The maximum width of the road approaching to the	12.28	7.06	4.0	32.4
WidthDifference	approaching to the intersection Width difference between two	intersection. Subtract the minimum road width from maximum road		5.87	0.0	26.5
Approach	intersecting roads Total number of roads connected to the intersection	width	3.360	0.500	3.0	5.0
Traffic calming measures and traffic signs
Pavement	Presence of red-colored pavement at the intersection	Yes = 1, no = 0	0.20	0.40	0.0	1.0
Speedhump	Presence of speed hump at the		0.16	0.37	0.0	1.0
Schoolzone	intersection Presence of school zone sign at the intersection		0.07	0.26	0.0	1.0
Building at intersections Visibility	Degree of clear visibility at the intersections of streets	Proportion of pedestrian approaches from which pedestrians are able to observe incoming vehicles at the intersection	0.21	0.35	0.0	1.0
Entrance	Total number of building entrances at the intersections	Count number of building entrances at the pedestrian level	2.26	1.16	0.0	4.0
Building use
Residential	Presence of residential use	Yes = 1, no = 0	0.66	0.47	0.0	1.0
Public	Presence of public use		0.10	0.30	0.0	1.0
Commercial	Presence of commercial use		0.76	0.43	0.0	1.0

walks (t(797) = 8.42, p < 0.001), traffic conditions (t(797) = 7.81, p < 0.001), and road lighting (t(797) = 5.04, p < 0.001) (Fig. 2(b)). No significant difference between the sixth- and fifth-grade students was found. Regardless of their subgroup, the students' major concern about road safety was the traffic conditions (speed and volume) at the school site.

4.2. Estimated student crossings and perceived crash risk and pedestrian exposure at the intersections

Fig. 3(a)illustrates the selected number oftimes the participants crossed intersections on their walking routes to school. The total number of intersections located within the study areas was 546. On average,the intersections in the study areas were selected 4.8 times (SD = 9.0) and 215 intersections were not selected for the routes. The school-aged children reported 2621 locations that were perceived as having a high risk of crashes near elementary schools and 1673 (65%) of those locations are located within the 15-m buffered area of the intersections (Fig. 3(b)), which implies that school-aged childrenperceive intersections to bedangerous locations compared with segments. The mean value for the perceived crash risk atintersections was 3.09 (SD = 6.3). Of the 546 intersections located in the study areas, 252 were zero-risk intersections. Not surprisingly, the intersections with a higher number of student crossings tend to

Fig. 2. The mean perceived risk of boy and girl (a), and those who had experienced near a collision with a car and those who had not (b).

have a higher perceived crash risk. A higher than moderate level of correlation (r = 0.85) exists between the two.

4.3. Intersection characteristics associated with perceived and actual crash risk

Table 2 presents the parameter estimates and z values of the NB and ZINB models for perceived crash risk. In the NB model, the over-dispersion parameter estimate was found to be significantly higher than zero (95% CI (0.51, 0.83)), which implies that the NB distribution was more appropriate than the Poisson distribution. However, the Vuong test suggests that the ZINB model is a significant improvement over the standard NB model (z = 6.37, p < 0.001). The standard NBmodel showed thattenintersectioncharacteristics have statistically significant relationships with perceived crash risk at the intersections. The zero-inflated model includes both count and logit models. The count model showed six characteristics that have a statistically significant relationship with risk perception and the logit model revealed three significant characteristics.

The variables representing pedestrian exposure were all associated with perceived crash risk. In the NB model, a higher number of student crossings and a higher population density and more street networks were associated with a higher level of perceived crash risk. In the ZINB model, the number of student crossings increases the positive count of the perceived crash risk as well as the probability of the intersection being non-zero risk. With regard to the familiarity variables, the number of street legs with studentfriendly facilities was associated with perceived crash risk.

The proportion of marked crosswalk legs also increased the level of perceived crash risk in the NB model and the probability of the intersection being non-zero risk in the ZINB model. Whether a marked crosswalk reduces pedestrian crashes remains unclear, but the segments with a marked crosswalk alone seemed to be most positively associated with a pedestrian crash rate (Koepsell et al., 2002; Zegeer et al., 2005). Students may perceive a higher crash risk at those intersections with a marked crosswalk because a marked crosswalk indicates a place where pedestrian and vehicle flows are concentrated. The volumes at pedestrian crossings were three to four times higher at marked crosswalks than at equivalent unmarked crosswalks (Zegeer et al., 2005).

Meanwhile, the students showed a lower risk perception at three-way intersections than at four-way intersections in the ZINB model. The literature has consistently reported thatfour-way intersections have higher crash rates than three-way intersections do (Lovegrove and Sayed, 2006; Ewing and Dumbaugh, 2009). A fourway intersection with two-lane roads creates sixteen points of conflict between vehicles and pedestrians while there are only twelve conflict points at a three-way intersection (AASHTO, 2004), which implies that a higher level of awareness is required to walk across a four-way intersection to avoid crashes. Students perceived a higher crash risk at locations where the maximum width of the road was wider, while the width difference between the two roads reduces risk perception in both the NB and ZINB models. We spec-

Fig. 3. Estimated student crossings (a) and perceived crash risk (b) at the intersections.

Table 2 Results of the standard NB and ZINB models for perceived crash risk.

		Negative binomial model		Zero-inflated negative binomial model
	Variable				Non zero-risk state			Zero-risk state
		Coeff.	z value		Coeff.	z value		Coeff.	z value
Exposure	Crossing	0.077	11.96	***	0.059	11.86	***	−2.217	−2.42	***
	PopulationDensity	0.009	1.89	*	0.004	1.03		−0.043	−1.3
	StreetDensity	0.011	3.65	***	0.002	0.74		−0.015	−0.96
Familiarity	Park	0.133	1.38		0.086	1.02		0.347	0.63
	StudentFacility	0.226	3.89	***	0.132	2.63	***	−0.312	−0.79
Road infrastructure	Sidewalk	0.062	1.28		0.053	1.17		−0.151	−0.83
	Crosswalk	0.308	3.83	***	0.195	2.66	**	−0.483	−1.63
	MaxWidth	0.038	1.9	*	0.013	0.68		−0.137	−1.82	*
	WidthDifference	−0.038	−2.03	**	−0.016	−0.9		0.126	1.75	*
	Approach	0.154	1.25		0.203	1.7	*	0.734	0.93
Traffic calming measures and traffic	Pavement	0.193	1.39		0.14	1.15		−0.303	−0.45
signs	Speedhump	0.163	1.16		0.045	0.36		−0.03	−0.03
	Schoolzone	0.638	3.48	***	0.404	2.7	**	−2.555	−1.31
Building at intersections	Visibility	−0.532	−2.2	**	−0.43	−1.88	*	−1.583	−1.44
	Entrance	0.146	2.35	**	0.088	1.51		−0.361	−1.03
Building use	Residential	0.154	1.17		0.021	0.17		−0.161	−0.26
	Public	0.156	0.98		0.13	0.93		−0.25	−0.3
	Commercial	−0.216	−1.38		−0.227	−1.53		−0.851	−1.23
Number of observations			546						546
Dispersion parameter			0.653						0.35
AIC			1835.3						1734.5
Log likelihood			−897.7						−828.3
Vuong statistic									6.37

^* Statistically significant at 10%.

ulate that the alteration of the road width at an intersection may reduce the speed of the vehicles approaching the intersection and, as a result, decrease the perceived crash risk among pedestrians.

The results further showed that the number of building entrances at the pedestrian level and the visibility at an intersection were associated with a perceived crash risk. A higher number

^** Statistically significant at 5%.

^*** Statistically significant at 1%.

of entrances at an intersection increased the positive count for the perceived crash risk while higher visibility at the intersection reduced that risk perception. However, the attributes of building use were not significantly associated with risk perception in the NB model. The number of building entrances at the pedestrian level were likely to be associated with higher levels of pedestrian demand.

With regard to traffic-calming measures and signage, only the presence of a school zone sign was associated with a higher perceived crash risk.

Table 3 shows the results of the NB model for police-reported crash risk. Four independent variables were statistically significant. A higher population density, higher number of marked crosswalks, a wider width for the main road, and a higher number of building entrances at the intersection increase the occurrence of pedestrian crashes at the intersection. These associations are in line with the findings reported in existing studies on the relationship between the built environment and pedestrian crashes (Koepsell et al., 2002; Zegeer et al., 2005; Clifton and Kreamer-Fults, 2007; Wier et al., 2009; Ukkusuri et al., 2012). More importantly, intersection characteristics associated with actual crashes are broadly consistent with the relationship between perceived crash risk and intersection characteristics, which implies that risk perception among schoolaged children provides relatively accurate information with which to identify the potential risk of pedestrian crashes. Meanwhile, several intersection characteristics that have a significant association with perceived crash risk were not associated with actual crash risk. These variables are mainly relevant to pedestrian exposure, familiarity, and visibility at the intersection. A further discussion to compare these results will follow in the next section.

5. Discussion and conclusions

It is important to note that we generally understand the perceived crash risk among students as an indicator for potential crash risk and as a tool with which to provide guidelines for proactive interventions (Schneider et al., 2004; Cho et al., 2009; Diógenes et al., 2010). In that sense, the findings of this study confirm those from the body of the literature examining the environmental correlates of pedestrian crashes. A higher number of student crossings, a higher population density, a wider road width, the presence of a crosswalk, and four-way intersections were all significant and positively associated with a higher perceived crash risk or potential crash risk at the intersections.

The results from the actual crash risk model show similar findings to the model results for perceived crash risk. However, several variables associated with perceived crash risk were not statistically significant in explaining actual crash risk. It seems manifest that some ofthose variables are highly related to pedestrian exposure or familiarity of the environment. A higher number of student crossings, the presence of student-friendly facilities, and the presence of school zone signs increase the perceived crash risk, not because those characteristics are potential risk factors for pedestrians, but because school-aged childrenmay frequently visitthe intersections with those characteristics, and thus they are more likely to report the locations they frequently visit as having a higher crash risk. In contrast, some ofthe variables may provide useful information that cannot be addressed only with the crash data. Often, identifying the risk factors associated with actual crashes does not fully depict potentially hazardous situations prior to their actualization, in particular, when the study site covers relatively small areas (Diógenes et al., 2010). Although intersections with lower visibility and two large roads were not shown to have a significant association with actual crash risk,the higher perceived crash risk atthe intersections may imply that there is the potential danger of a crash or a change in children's commuting behavior. Our findings are inconclusive as of yet. More concrete hypotheses and theoretical frameworks need to be developed and tested in future studies.

One interesting finding is that the features of buildings near intersections were associated with perceived crash risk. The existing literature focusing on the design of roadway intersections has not given much attention to the characteristics of buildings at the intersections. A higher number of building entrances increased the perception of the crash risk among students and the occurrence of pedestrian crashes. We suggest that more entrances to buildings, which may indicate accessibility to the building, also generate more conflicts between indoor and outdoor pedestrian movement and traffic flow. Given that this feature can produce a more diverse and vital urban environment, suggesting design interventions to reduce building entrances at intersections may not be a desirable solution. Indeed, road safety studies have shown that environmental attributes such as mixed land use or street connectivity tend to increase crash risk, although these attributes are supposed to support walking or increase the vitality of a city. Intersections with those characteristics are more likely to have a concentration of pedestrians and a greater risk of crashes (Loukaitou-Sideris et al., 2007; Clifton et al., 2009). The literature has also reported that crash risk per capita tends to decrease as pedestrian volume increases (Leden, 2002; Jacobsen, 2003; Elvik, 2009). Examining the practice of "safety in numbers" with respectto perceived crash risk would be a promising topic for future research.Apparently, improving pedestrian visibility at an intersection is importantfor enhancing feelings of safety. Typically, in dense urban areas, intersection corners are frequently entrances to buildings that impair visibility in terms of oncoming traffic. To ensure safety for school-aged children, particular attention should be given to pedestrians through the provision of an adequate sight distance. Imposing a building setback regulation near school zones or creating open spaces near intersections may increase the visibility for both pedestrians and drivers and also reduce potential crash risks for children.

The literature on pedestrian crashes has commonly reported that traffic-calming measures can reduce crash rates (Retting et al., 2003; Ewing and Dumbaugh, 2009), but our results showed that these measures were not effective in reducing either perceived or actual crash risk. Perhaps the influence of traffic-calming measures on crash risk is minimal because they are mainly intended to change the behavior of drivers rather than of pedestrians. However, a more plausible explanation is that the treatments applied to the site are not effective at reducing crash risk. The commonly used measures were red-colored pavements atintersections, speed humps, and traffic signs to reduce vehicle speed. Other treatments that physically force lower speeds, such as narrowing traffic lanes or creating a horizontal deflection for vehicles, were not used in the study areas. Ewing and Dumbaugh (2009) suggested that the safety impacts of traffic calming could be diverse according to the context of the study areas, and they recommended installing more intensive traffic-calming treatments that were well integrated with the surroundings to improve the safety effect of the measurements.

Methodological limitations to this study should be noted. The first issue is the countermeasures to pedestrian exposure and the locations perceived as having a high crash risk. Given that students in elementary school tend to choose regular walking routes to school and have relatively limited information on the surrounding environment near schools compared with adults, students are very likely to report risk locations along the routes they choose. In estimating the number of student crossings, we used walking routes from home to school in the morning as data but did not collect behavioral information after school, which presumably would show more diverse patterns of walking near the schools. Lack of data on pedestrian exposure on a citywide scale, however, is a common difficulty in pedestrian safety studies (Loukaitou-Sideris

Table 3 Results of the standard NB model for actual crash risk.

		Negative binomial model
	Variable Crossing	Coeff.	z value
Exposure		−0.025	−1.04
	PopulationDensity	0.039	2.63	***
	StreetDensity	0.009	0.98
Familiarity	Park	−0.567	−1.43
	StudentFacility	0.12	0.62
Road infrastructure	Sidewalk	0.184	1.45
	Crosswalk	0.43	2.32	**
	MaxWidth	0.123	2.43	**
	WidthDifference	0.041	0.96
	Approach	−0.384	−0.93
Traffic calming measures and traffic signs	Pavement	−0.194	−0.48
	Speedhump	0.116	0.21
	Schoolzone	−0.035	−0.04
Building at intersections	Visibility	−0.079	−0.12
	Entrance	0.619	3.22	***
Building use	Residential	−0.164	−0.45
	Public	0.07	0.14
	Commercial	0.505	0.85
Number of observations			546
Dispersion parameter			3.06
AIC			561.6
Log likelihood			−260.8

^** Statistically significant at 5%.

et al., 2007; Clifton et al., 2009), including this one, and research uncovering more refined measures of pedestrian exposure are still required.

Most ofthe environmental attributes around intersections were collected by audit using street views without conducting a formal test of reliability and validity. Rundle et al. (2011) indicated that auditing neighborhood environments using street-view imagery had relatively high reliability and validity for measuring pedestrian safety, motorized traffic, and the infrastructure for active travel, whereas physical disorder and aesthetics had a lower concordance. In addition, the approximately two-year time gap between streetview imagery and present environmental conditions may reduce the validity of intersection audits.

These shortcomings aside,the findings ofthis study suggestthat the proposed method for identifying perceived crash risk among school-aged children can provide valuable information about locations in need of attention and can be used to suggest proactive interventions to reduce the potential crash risk at intersections. The major contribution of this paper is in identifying the potentially hazardous environmental attributes of intersections based on perceived crash risk as reported by school-aged children. To our knowledge, no study has investigated this subject with the method used herein. With our empirical analysis, we have not proven whether the identified potential risk factors were likely to materialize as pedestrian crashes in the near future. Future work, therefore, should incorporate more in-depth monitoring, comparison, and investigation of police-reported pedestrian crashes along with data on the perceived crash risk to understand more definitively the role of perceived crash risk among school-aged children in enhancing pedestrian safety.

Acknowledgements

This work was supported by TechnologyAdvancement Research Program (15CTAP-C078783-02) funded by Ministry of Land, Infrastructure and Transport of Korean government.

References

American Association of State Highway and Transportation Officials, 2004. Guide for the Planning, Design and Operation of Pedestrian Facilities. AASHTO, Washington, DC.

Abdel-Aty, M., Chundi, S.S., Lee, C., 2007. Geo-spatial and log-linear analysis of pedestrian and bicyclist crashes involving school-aged children. J. Saf. Res. 38 (5), 571–579.

Aziz, H.A., Ukkusuri, S.V., Hasan, S., 2013. Exploring the determinants of pedestrian–vehicle crash severity in New York City. Accid. Anal. Prev. 50, 1298–1309.

Berhanu, G., 2004. Models relating traffic safety with road environment and traffic flows on arterial roads in Addis Ababa. Accid. Anal. Prev. 36 (5), 697–704. Bowman, B.L., Vecellio, R.L., Miao, J., 1995. Vehicle and pedestrian accident models for median locations. J. Transp. Eng. 121 (6), 531–537.

Carver, A., Timperio, A., Hesketh, K., Crawford, D., 2010. Are children and adolescents less active if parents restrict their physical activity and active transport due to perceived risk? Social Sci. Med. 70 (11), 1799–1805.

Chaurand, N., Delhomme, P., 2013. Cyclists and drivers in road interactions: a comparison of perceived crash risk. Accid. Anal. Prev. 50, 1176–1184.

Chin, H.C., Quddus, M.A., 2003. Applying the random effect negative binomial model to examine traffic accident occurrence at signalized intersections. Accid. Anal. Prev. 35 (2), 253–259.

Cho, G., Rodríguez, D.A., Khattak, A.J., 2009. The role of the built environment in explaining relationships between perceived and actual pedestrian and bicyclist safety. Accid. Anal. Prev. 41 (4), 692–702.

Clifton, K.J., Kreamer-Fults, K., 2007. An examination of the environmental attributes associated with pedestrian–vehicular crashes near public schools. Accid. Anal. Prev. 39 (4), 708–715.

Clifton, K.J., Burnier, C.V., Akar, G., 2009. Severity of injury resulting from pedestrian–vehicle crashes: what can we learn from examining the built environment? Transp. Res. Part D 14 (6), 425–436.

Dai, D., Taquechel, E.P., Steward, J., Strasser, S., 2010. The impact of built environment on pedestrian crashes and the identification of crash clusters on an urban university campus. West. J. Emergency Med. 11 (3).

DeJoy, D.M., 1992. An examination of gender differences in traffic accident risk perception. Accid. Anal. Prev. 24 (3), 237–246.

Deery, H.A., 2000. Hazard and risk perception among young novice drivers. J. Saf. Res. 30 (4), 225–236.

Diógenes, M.C., Lindau, L.A., de Transportes–LASTRAN, L.D.S., 2010. Risk perception of pedestrians at midblock crossing in Brazil. In: Proceedings of the 12th World Conference on Transport Research (WCTR), Lisbon, Portugal, pp. 1–19.

Dissanayake, D., Aryaija, J., Wedagama, D.P., 2009. Modelling the effects of land use and temporal factors on child pedestrian casualties. Accid. Anal. Prev. 41 (5), 1016–1024.

Elias, W., Shiftan, Y., 2014. Analyzing and modeling risk exposure of pedestrian

children to involvement in car crashes. Accid. Anal. Prev. 62, 397–405. Elvik, R., Bjørnskau, T., 2005. How accurately does the public perceive differences in transport risks?: An exploratory analysis of scales representing perceived risk. Accid. Anal. Prev. 37 (6), 1005–1011.