category: literaturenote citekey: cottrillevaluatingpedestriancrashesareas2010 title: Evaluating pedestrian crashes in areas with high low-income or minority populations authors: "Cottrill, Caitlin D.; Thakuriah, Piyushimita (Vonu)" year: 2010 date: 2010-00-00 2010 doi: 10.1016/j.aap.2010.04.012 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2010.04.012" zotero_key: PFMRW59R zotero_storage: S997TRUT collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Cottrill, Caitlin D."

status: converted

Contents lists available at ScienceDirect

Accident Analysis and Prevention

journal homepage: www.elsevier.com/locate/aap

Evaluating pedestrian crashes in areas with high low-income or minority populations

Caitlin D. Cottrill ^a,∗, Piyushimita (Vonu) Thakuriah^b,¹

^a Urban Transportation Center, University of Illinois, Chicago, 412 South Peoria Street, Suite 340, Chicago, IL 60607-7036, United States
^b Department of Urban Planning and Policy, University of Illinois, Chicago, 412 South Peoria Street, Suite 340, Chicago, IL 60607-7036, United States

article info

Article history: Received 14 May 2009 Received in revised form 26 April 2010 Accepted 29 April 2010

Keywords: Pedestrian Crash Income Minority Crime Transit

abstract

In this paper, we present an analysis of the relationship between pedestrian–vehicle crashes and characteristics of areas with high low-income and minority populations in the Chicago metropolitan area (also called environmental justice or EJ areas in the United States). While related research has indicated that pedestrian crashes occur more frequently in these areas than in non-EJ areas, this paper attempts to relate the incidence to environmental characteristics and behavioral factors through a better understanding of the contributing factors present in crash occurrences in EJ versus non-EJ areas. Specially constructed small-area factors from a Spatial Decision Support System (SDSS) are used to explain pedestrian–vehicle crashes.

Using a Poisson model that corrects for underreporting, we find that pedestrian crash incidents in EJ areas are related to variables of exposure (including the suitability of the area for walking and transit accessibility), crime rates, transit availability, and general population demographics such as income and presence of children. Results suggest that it may be necessary to better incorporate a safety perspective or measures of safety improvements in pedestrian and transit improvements and expansion programs within EJ areas.

1. Introduction

In 2005, motor-vehicle crashes killed 4881 pedestrians in traffic in the United States and sent 120,815 pedestrians to emergency medical departments (National Highway Traffic Safety Administration [NHTSA], 2006). While the contributing causal factors to these crashes vary widely, previous studies have indicated that certain factors relating to the socio-demographics and the physical environment of neighborhoods may heighten the risk of pedestrian–vehicle crashes. These initial indications are particularly relevant in regards to environmental justice (EJ) guidelines which broadly require each U.S. federal agency to make the alleviation of inequitable environmental burden a part of its mission by identifying and addressing, as appropriate, disproportionately high and adverse human health or environmental effects of its programs, policies, and activities on minority and low-income populations (Environmental Protection Agency, EPA, Executive Order, 1994). This paper is intended to specifically address pedestrian–vehicle crashes within the framework of EJ guidance to determine if pedestrian safety is an issue that should be specifically addressed within EJ activities.

Several authors have noted that there is an increased incidence of pedestrian–vehicle crashes in EJ areas (Mid-America Regional Council, 2007; Murtha, 2005). EJ areas are those with high proportions of minority and low-income households (the specific definition used in this paper is given in Section 3.1). While numbers gathered from these evaluations are of interest, what is notable is that occurrences of crashes are often cited while paying little attention to the question of general exposure to risk, including such factors as mixture of land uses, the number of potential pedestrians in an area, or the overall character of the surrounding environment and whether it is inviting to and safe for pedestrian travel.

The current study more extensively explores some geographic correlates of pedestrian–vehicle crashes, through an analysis of data from the Chicago metropolitan area in the State of Illinois, USA, with particular emphasis on spatial variations on these events by EJ versus non-EJ area. We examine 2005 Chicago vehicle–pedestrian crash data, gathered from the Illinois Department of Transportation (IDOT) as a case study in order to hold factors beyond socio-demographics constant and to provide information relating to risk levels in environmental justice areas. Statistical evidence, as presented below, shows that an elevated risk is

^∗ Corresponding author. Tel.: +1 312 996 4820; fax: +1 312 413 0006.

E-mail addresses: ccottr2@uic.edu(C.D. Cottrill), vonu-pt@uic.edu(P. Thakuriah).

¹ Tel.: +1 312 996 4820; fax: +1 312 413 0006.

present, perhaps indicating that policies and programs addressing pedestrian–vehicle crashes in these areas should be targeted to those factors that are likely to account for such increases in risk.

The objective of the paper is to report on a comparative analysis of motor-vehicle–pedestrian crashes in EJ versus non-EJ areas and to understand which factors contribute to variations in pedestrian crashes in those areas. We have taken the view that two groups of geographic factors can potentially affect observed differences – those relating to the physical environment (called environmental factors) of the areas and those relating to social behaviors, cognitive and demographics (called behavioral factors) of the residents of the areas. These measures are a part of an ongoing research project to develop a Spatial Decision Support System on the Chicago metro area for planning purposes by the authors' research team. The construction and potential use of these measures in accident analysis are described. The importance of these factors in explaining EJ versus non-EJ differences in pedestrian crashes are demonstrated using exploratory methods as well as regression. An issue of concern that is addressed in the paper is underreporting of pedestrian crashes – we have adopted a statistical approach to address potential underreporting in the Chicago vehicle–pedestrian crash data.

The paper is organized as follows: prior literature on environmental justice and pedestrian crashes is reviewed in Section 2. The study area, the method by which EJ areas were determined and the environmental and behavioral factors considered are presented in Section 3. Section 4 presents an exploratory analysis of crash trends in EJ versus non-EJ areas and reports summary statistics on the factors and on crash patterns in the two types of areas. Statistical models of pedestrian crashes are given in Section 5. Summary and conclusions are drawn in Section 6.

2. Background and prior literature

In 1994, President Clinton signed Executive Order 12898, "Federal Actions to Address Environmental Justice in Minority Populations and Low-Income Populations". As part of this order, it was required that, "each Federal agency shall make achieving environmental justice part of its mission by identifying and addressing, as appropriate, disproportionately high and adverse human health or environmental effects of its programs, policies, and activities on minority populations and low-income populations..." (1994). Partially in response to this order, and partially in response to growing awareness by residents and activists in environmental justice communities, pedestrian issues in low-income and minority areas have of late been gaining increased visibility and dialogue.

Several authors have examined geographic variations in pedestrian crash rates at both a macro (state or regional) level as well as a micro (intra-regional) level. For example, Paulozzi (2006) found that the highest pedestrian fatality rates concentrate in Sunbelt states that have experienced rapid population growth in the past 50 years. This pattern may result from at least three features of these states: (a) a high percentage of urban vehicle miles traveled; (b) urban sprawl; and (c) a high prevalence of alcohol use—especially heavy use—among Sunbelt pedestrians. Beck et al. (2007), in a comparison of the Atlanta metro region with the rest of the U.S., reported that from 2000 to 2004, pedestrian fatality rates declined in the United States, but the Atlanta metropolitan statistical area did not experience the same decline. Pedestrian fatality rates for males, Hispanics, and the 15–34 and 35–54 year age groups were higher in Atlanta than in the United States overall.

Of direct relevance to this paper is a stream of literature on intra-regional variations in pedestrian crash rates. Dougherty et al. (1990) found that for Montreal, neighborhoods with income levels in the lowest quintile had death and accident rates that were at least four times the rates for children living in the least disadvantaged neighborhoods. Several authors have written about the complex interaction of factors that put pedestrians in poor neighborhoods, especially child pedestrians, at greater risk from moving vehicles (Padilla et al., 1976; Brown and Davidson, 1978; Bijur et al., 1986; Lyons et al., 2008; Guyer et al., 1985) including "the lack of secure play space; dwellings adjacent to busy traffic flows; hazardous journeys to school and other facilities; less developed cognitive skills, which would imply risk-taking; and greater stress (physiological, social, and psychological)." (Bagley, 1992) LaScala et al. (2004), using data from four California communities, found that annual numbers of pedestrian injuries were greater in areas with higher youth population densities, more unemployment, fewer high-income households, and greater traffic flow.

Loukaitou-Sideris and Liggett (2007) evaluated pedestrian accidents in Los Angeles in the context of socio-demographic and land use characteristics, eventually concluding that, "The coefficients on these variables [percent of the population below poverty and percent Latino] in each model were positive, supporting the assumption that pedestrian accidents are more likely to occur in low-income, minority neighborhoods once other aspects of risk are controlled for." Similar results were found in Nassau County, New York (pertaining to risk of hospitalization for pedestrian injuries). Additionally, the Surface Transportation Policy Project's (STPP) report, Mean Streets 2002 (STPP, 2002) stated that,

Nationwide, non-Hispanic whites comprise 69 percent of the population. However, only 60 percent of pedestrian deaths for which ethnicity is known are non-Hispanic whites. In contrast, African-Americans make up more than one in five (20 percent) pedestrian deaths, though they represent only 12 percent of the total U.S. population. Likewise, Latino pedestrians comprise 13.5 percent of deaths, but only 12.5 percent of the total U.S. population. However, although the current literature includes studies of individual and environmental characteristics related to injury rates, there exist only a handful of neighborhood studies that introduce specific geographic components with general applicability across communities (LaScala et al., 2000, 2001, 2004).

While it is not necessarily the case that these pedestrian deaths took place near the location of residence, it is critical to evaluate the potential that the socio-demographics and the nature of the built environment of the home location may play a part in the relative imbalance of the ethnicity of crash victims.

Several international studies have also indicated a relationship between pedestrian crashes and areas of high low-income and/or minority populations. For example, in a scan of literature addressing the issue from a number of countries, including the US, the United Kingdom, and Australia, among others, Laflamme and Diderichsen (2000) found that for most types of traffic injuries, including pedestrian injuries, "...mortality and morbidity are often higher among children from lower social positions and in more deprived socio-economic areas." In a study of English and Welsh families, Edwards et al. (2006) found that the pedestrian death rate of children of parents classified as long-term unemployed or never employed was 20.6 times higher than that of children of parents in higher managerial or professional occupations. Finally, a study of traffic crashes involving Maoris in New Zealand conducted by Sargent et al. (2004) found a higher proportion of traffic casualties and injuries for all crashes (including pedestrian) for those Maoris living in areas of higher relative deprivation. These findings indicate that potential linkages between socio-economic variables such as income and race and pedestrian crashes are a worldwide phenomena and require additional study.

3. Study area

In 2005, the Illinois Department of Transportation (IDOT) reported 4886 vehicle crashes in the Chicago region in which a pedestrian was involved. Of those crashes, 120 (or about 3%) resulted in the death of the pedestrian. Fig. 1 shows the distribution of vehicle–pedestrian crashes in the six-county Chicago metro area (consisting of Cook, McHenry, Kane, Will, Lake, and DuPage counties; Cook county includes the City of Chicago). Overall, pedestrian–vehicle crashes accounted for about 1% of all 2005 Illinois crashes and about 2% of all crashes in the Chicago region. Fig. 1 shows that the distribution of pedestrian–vehicle crashes are not random throughout the region; rather, it appears that crashes are concentrated in the downtown business district (the "Loop") and areas in south and west Chicago. The goal of this paper is to utilize the Chicago region as a case study to analyze pedestrian–vehicle crashes, focusing on those factors that are considered under environmental justice regulations and to determine if issues beyond pedestrian exposure play a part in influencing pedestrian–vehicle crash rates.

3.1. Determination of EJ areas

EJ and non-EJ areas are defined at the level of census tracts. There are several pros and cons in using census tracts as the spatial unit for equity studies. Census tracts have a relatively homogeneous population of about 4000 and are delimited by local persons and thus "reflect the structure of the metropolis as viewed by those most familiar with it" (Bogue, 1985). They are small enough to reduce the possibility of "aggregation errors" and "ecological fallacies" (Anderton et al., 1994) but large enough to potentially avoid impacts that spill over beyond smaller spatial units such as block or block groups (Been, 1994), although other authors have noted that impact areas may run well beyond the boundaries of individual tracts in metropolitan areas (Mohai, 1995). Further, it is the most commonly used geographic unit of analysis (Anderton et al., 1994) – in our case, the spatial level corresponds to the unit of analysis in the Chicago area Spatial Decision Support System and it is deemed to be a reasonable approximation of the concept of a neighborhood (Denton and Massey, 1991). Although there are several studies that compare outcomes related to mobility and transportation conditions in EJ versus non-EJ areas, there is no consistency among studies about how EJ areas are defined. In safety research, Wier et al. (2009) recently used census tracts in the San Francisco area to determine the relationship between street, land use, and population characteristics, including commute behaviors and census tract-level pedestrian–vehicle injury collisions.

The use of census tracts as the level of analysis also introduces the need to examine the way in which population exposure is evaluated. Because, as noted above, census tracts are fairly homogenous in population size, for purposes of this analysis we have chosen to use population density as a measure of exposure. While the use of actual pedestrian counts would be useful, such information is not generally available on an area-wide basis. The use of density in this calculation is consistent with methods proposed by such entities as the FHWA (1999), which suggests population density as a method of evaluating potential travel behavior from a contextoriented perspective, and such researchers as Wedagama et al. (2006) and Graham and Glaister (2003), who use population density as an explanatory variable in models of crash frequencies and who suggest that increased urban density may be both an indicator of deprivation, which has been shown as influencing pedestrian crashes, and as an influence on traffic volumes and speed, which may also have an influence. Raford and Ragland (2003), additionally state that, "Physical factors such as residential population density, mixed land use, street connectivity, and adequate pedestrian facilities have been identified as key physical variables that influence the number and types of walking trips." By using density, we are able to better represent the relative distribution of pedestrian crashes across the Chicagoland region. As the region includes a large variety of census tract types (ranging from rural to highly urban), the spatial area of each tract also varies greatly. It may be misleading to directly compare incidents of pedestrian crashes in highly rural tracts with those in highly urban tracts, as the spatial density will vary greatly. By using population density, we are better able to perform comparative analyses.

The definition of EJ areas adopted here replicates the rule given in Sriraj et al. (2003). EJ neighborhood criteria are determined in this paper by computing demographic and economic information for Cook County, and basing the threshold analysis for EJ census tracts on the results. Using this method, the following rules are determined to define EJ tracts: (a) Black – according to the 2000 Census, 26% of all residents of Cook County are Black. All census tracts with a Black population greater than 26% qualify as EJ neighborhoods; (b) Hispanic – according to the 2000 Census, 20% of all residents of Cook County are Hispanic. All census tracts with a Hispanic population greater than 20% qualify as EJ neighborhoods; (c) low-income – all households that are within 150% of the poverty threshold are considered low-income. Using the reference area technique, this population is 21.5% of the total population in Cook County. Accordingly, census tracts that have more than 21.5% of their population within the low-income threshold, qualify as EJ neighborhoods.

3.2. Environmental and behavioral factors relating to pedestrian crashes

The Chicago area Spatial Decision Support System (SDSS) integrates a Multi-Criteria Decision-Making (MCDM) environment with small-area spatial data to integrate transportation decision-making with housing, community development, economic development and physical planning. Using a variety of data sources and transportation modeling outputs, it consists of a number of indicators on traffic and transportation, accessibility measures, regional employment opportunities, small-area employment estimates based on forecasted job openings and actual jobs, affordable housing, school quality, crime, health, land use and the built environment.

The factors within the SDSS, which were used in this study, can be categorized into two groups: the first group of factors relate to the physical environment of the area including the level of vehicular traffic flow, quality and complexity of the walking environment, pedestrian crowding possibilities (at transit stations or commercial developments), road crossing opportunities which present possibilities of conflict with vehicular traffic, population density and the extent of crime (called environmental factors). The second group of factors relate to social behaviors, cognitive and demographics (called behavioral factors), which lead to variations in the ability to cope effectively with these hazards, including percent of population with little or no English-speaking ability as indicators of unfamiliarity with the environment, percent of school-going children and percent of carless households.

Table 1 presents the indicators used with a brief explanation of each indicator. Indicators were chosen mostly on the basis of prior, published research as well as local knowledge derived from professionals in the Chicago metro area, who work in the area of transportation, safety and public health. Using a methodologically rigorous approach to ensure that all possible factors that account for systematic differences between EJ and non-EJ tracts and omitted variable bias, is left for future research.

Fig. 1. Map showing 2005 pedestrian crashes in the Chicago metro area.

4. Exploratory analysis of pedestrian crashes in EJ versus non-EJ areas

Tables 2 and 3 show summary statistics on variables that are of importance in the analysis. Table 2 shows that the number of pedestrian crashes in EJ areas are well over twice that in non-EJ areas and that the pedestrian crash rate per 1000 population is 9.66 in EJ areas compared to 3.37 in non-EJ areas. Large differences do not exist in the proportion of injuries of different types between EJ and non-EJ areas. Injuries are classified as follows (IDOT Crash Report Data Dictionary, 2005):

- A Injuries: incapacitating but not fatal injuries, including those which prevent the injured party from walking, driving, or normally continuing the activities her or she was capable of performing before the injury occurred.
- B Injuries: nonincapacitating and non-fatal injuries, evident to observers at the crash scene.
C Injuries: possible injuries, which are not classified as type A or B above.

Fatal crashes are 2.2% of all crashes in EJ areas compared to 3% in non-EJ areas.

Pedestrian crash victims tend to be younger in EJ areas (close to 28% of pedestrians involved in crashes in EJ areas are less than 16 years of age, compared to about 10% of pedestrians in non-EJ areas) and where the gender of the crash victim is known, involve a slightly larger share of males (56% compared to 55%). Table 3 shows that the median household income in EJ areas is \$36,274.79 compared to \$61,414 in non-EJ areas. A larger proportion of EJ area residents have "low" levels of English-speaking ability and also a greater share of school-age children (30% compared to 26% in non-EJ areas). About 24% of households in EJ areas have no cars, compared to only 8% of households in non-EJ areas.

Table 3 shows that the mean pedestrian crash rate in EJ tracts is 3.26 and that in non-EJ tracts is 1.94, with a difference of 1.32.

Table 1 Environmental and behavioral Indicators used in the study.

Indicator	Description
Environmental Indicators at census tract-level
Transit availability index (TAI)	Composite index giving the extent to which residents have access to transit (bus and rail); based on three input measures – frequency (person-minutes served), hours of service (number of hours) and service coverage (percentage of census tract area covered). See Minocha et al. (2008) for more details
Pedestrian accessibility index (PED)	Composite index ranking tract suitability for pedestrian travel; based on input values of population, income, number of households, amount of commercial and residential land uses as a percentage of census tracts, weighted trip origins and destinations, and Pedestrian Environment Factor (PEF) values, where PEF's are the average number of blocks for the quarter section within each census tract and the eight adjacent quarter sections. See Cottrill and Thakuriah (2008) for more details
Sum of annual average daily traffic (SUM AADT)	Total annual average daily traffic on links of all highway functional classes within census tract; output from regional traffic assignment model and GIS
Total miles of roads (SUM LENGTH)	Total miles of roads of all functional classifications within census tract; output from GIS
Total number of schools (NO SCHOOLS)	Count of all schools within census tract
Crime rate (CRIME)	Total violent and non-violent crimes in census tracts
Behavioral indicators at census tract-level
Median household income (MEDINC)	Median household income; based on census 2000
Population density (POP SQMILE)	Population per square mile in census tract (census 2000 population)
Percent with no cars (PERNOCAR)	Percentage of population without access to a vehicle (census 2000 data)
Percent commercial (PER COMM)	Percentage of census tract used for commercial purposes (based on CATS land use 1990 file)
Percent children (PERCHILDREN)	Percentage of population under the age of 16 (census 2000 data)
Percent who speak limited or no English (PERLOWENGLISH)	Percentage of persons who speak no English or limited English (Census 2000 data)

Our objective in Section 5 is to examine if and how much, ceteris paribus, environmental and behavioral factors explain the crash rate difference between EJ and non-EJ areas.

The built environment in EJ areas differs from that in non-EJ areas in significant ways. EJ areas are of much higher density (13,910 population per square mile versus 7111 per square mile in non-EJ areas) but have fewer schools per census tract (1.89 on the average, compared to 2.68 in non-EJ areas). EJ tracts are estimated to have lower Average Annual Daily Traffic (with a SUM AADT value of 229,599 miles) compared to 340,739 miles in non-EJ areas. EJ areas also have lower highway length (with a SUM LENGTH value about 3.5 miles on the average per census tract) compared to non-EJ areas (with an average of 9.84 miles).

The average pedestrian accessibility rank (PED) is slightly higher in EJ areas (the t-test of difference between PED in EJ versus non-EJ areas is significant only at the .10 level). The transit availability

Table 2 Comparison of EJ and non-EJ tracts on pedestrian crash victim socio-demographics and injury severity.

	EJ tract		Non-EJ tract		All tracts
	Number	Percent	Number	Percent	Total	Percent
Crash totals	3367.00	100.00%	1501.00	100.00%	4868.00	100.00%
Crash rates
Pop 2001 Census Est.	3,486,832		4,458,285		4,458,285
Count of Ped crashes	3367		1501		4868
Crash rate	0.000966		0.000336677		0.000337
Crashes/10,000	9.66		3.37		6.13
Injury severity
A Injury crash*	664	19.72%	331	22.05%	995	20.44%
B Injury crash**	1717	50.99%	746	49.70%	2463	50.60%
C Injury crash***	875	25.99%	371	24.72%	1246	25.60%
Fatal crash	74	2.20%	45	3.00%	119	2.44%
No injuries	37	1.10%	8	0.53%	45	0.92%
Hit and run
No	2276	67.60%	1190	79.28%	3466	71.20%
Yes	1091	32.40%	311	20.72%	1402	28.80%
Age (in years) of victims
0–5	132	3.92%	26	1.75%	158	3.26%
6–16	824	24.47%	247	16.61%	1071	22.06%
17–35	899	26.70%	485	32.62%	1384	28.51%
36–50	666	19.78%	313	21.05%	979	20.17%
51–70	446	13.25%	262	17.62%	708	14.59%
71 and more	400	11.88%	154	10.36%	554	11.41%
Gender
Female	1314	39.03%	624	41.57%	1938	39.81%
Male	1709	50.76%	774	51.57%	2483	51.01%
Unknown	344	10.22%	103	6.86%	447	9.18%

^* A Injuries: Incapacitating but not fatal injuries, including those which prevent the injured party from walking, driving, or normally continuing the activities her or she was capable of performing before the injury occurred.

^** B Injuries: Nonincapacitating and non-fatal injuries, evident to observers at the crash scene.

^*** C Injuries: Possible injuries, which are not classified as type A or B above.

Table 3 Variables utilized in the analysis, tests of significance, effect size and correlations.

Variable	EJ tract			Test of significance and effect size
	"No" d ≤ .20: small; .20 < d ≤ .50: medium; .50 < d ≤ .80: large	"Yes" (N = 946)	t value*	Cohen's d**
	(N = 886)
	Mean	Mean
PED CRASH	1.94	3.26	−7.87	0.37
Environmental factors
TAI	0.61	0.67	−10.86	0.51	0.20206
PED LOW (binary)	0.26	0.25	0.47	−0.02	−0.03654
SUM AADT	340,739.10	229,599.87	5.44	−0.26	0.03028
SUM LENGTH	9.84	3.49	11.7	−0.56	−0.12138
NOOFSCHOOLS	2.68	1.89	6.83	−0.32	0.29065
POP00 SQMI	7,111.12	13,910.09	−14.26	0.67	0.29913
PERCRIME	3.03	4.91	−22.16	0.74	0.27328
PED	35.31	36.23	−1.8	0.68	0.31226
Behavioral factors
WLKTOWRK	79.25	52.59	3.24	−0.15	0.29943
MEDHHINC99	67,414.06	36,274.79	32.85	−0.78	−0.20477
PERNOCAR	0.08	0.24	−24.05	1.12	0.31196
PER COMM	0.04	0.09	−9.08	0.42	0.21792
PERCHILDREN	0.26	0.30	−10.19	0.48	0.00721
PERLOWENGLISH	0.25	0.43	−12.41	0.57	0.10207

^* Italicized and bold estimates are significant at the .01 level.

index (TAI) is significantly larger in EJ areas, indicating that the built environment, land uses and transit access levels may be more conducive to walking in the EJ areas. Figs. 2 and 3 show Trellis plots of variations in pedestrian crashes by EJ and non-EJ census tract, for different levels of TAI and PED. The top panel of the Trellis plots in Fig. 2 shows variations in pedestrian crashes by EJ versus non-EJ census tracts for "High" levels of TAI (a TAI value of "High" indicates census tracts with the top 25% of values of transit availability, "Medium High" indicates census tracts with the second highest 25% of TAI values and so on). Fig. 3 displays similar information on PED. The plots also show the mean pedestrian crash rates for each TAI/PED and EJ condition.

Overall, the mean pedestrian crash rates for both EJ and non-EJ tracts decrease with decrease in TAI levels. In the case of non-EJ

Fig. 2. Trellis plot of variations in count of crashes by census tract for combinations of EJ and transit availability index (TAI).

^** The interpretation of effect sizes given by Cohen's d for difference in means with pooled variance is generally difficult due to "the low level of consciousness of the magnitude of phenomenon" (Cohen, 1992), but may be interpreted generally as follows: ^|d| ≤ .20: small; .20 < ^|d| ≤ .50: medium; .50 < ^|d| ≤ .80: large. # Correlation with PED CRASH.

Fig. 3. Trellis plots of variations in counts of crashes by census tracts for combinations of pedestrian accessibility (PED) and EJ status of census tract.

tracts with High TAI levels (on the left hand side of the top panel), the mean crash rate is 3.75, decreasing to 1.37 in the case of tracts with Low TAI levels. The mean crash rates for each TAI condition is higher for EJ tracts compared to non-EJ tracts; but in the case of EJ tracts too, the mean pedestrian crash rate decreases from 3.94 in the case of tracts with HIGH TAI levels to a mean of 2.56 in the case of EJ tracts with Low TAI levels. The difference between the mean crash rates of EJ and non-EJ tracts is substantially greater for tracts with Low–Medium and Low levels of TAI compared to tracts with High and High–Medium TAI levels.

Further, the percent of both non-EJ and EJ tracts with zero crashes increases with decrease in their TAI level. The percent of non-EJ tracts that have zero pedestrian crashes increases from 15% of all non-EJ tracts that have High TAI levels, to about 26% for tracts with High–Medium TAI levels, to approximately 42% for tracts with Low–Medium TAI levels and leveling off at about 40% of non-EJ tracts with Low levels of TAI with zero crashes. The percent of EJ tracts (on the right hand side of the top panel) that have zero pedestrian crashes also share a similar relationship to TAI levels; the percent of EJ tracts with High TAI levels that have zero recorded pedestrian crashes is close to 15% and increases to about 24% for EJ tracts with Low TAI levels. The percent of non-EJ tracts with low levels of pedestrian crashes (1–3 crashes) remains constant (at about 32–45% of all non-EJ tracts) for each TAI level; the same trend can be noticed for EJ tracts (between 32 and 42% of EJ census tracts have 1–3 crashes, irrespective of TAI levels). However, the percent of both EJ and non-EJ tracts with High or Very High crash rates (over 9) occur only where TAI levels are High or High–Medium. For each TAI level, the percent of EJ tracts with Low crashes is lower than in non-EJ tracts while the percent of EJ tracts with High–Medium crashes is greater than in non-EJ tracts.

Table 3 also shows the effect sizes of the environmental and behavioral factors, as measured by Cohen's d index (Cohen, 1992). Based on this measure, the following environmental factors – TAI, SUM LENGTH, POP00 SQMI, PERCRIME and PED – have large differences between pedestrian crashes in EJ and non-EJ tracts. Among behavioral factors, MEDHHINC99 and PERNOCAR have large differences between EJ and non-EJ areas.

The table also shows that the environmental measures TAI, PER-CRIME and PED with large effect sizes are positively correlated with PED CRASH. Finally, MEDHHINC99 is negatively correlated, whereas PERNOCAR is positively correlated with PED CRASH.

The relationship of the pedestrian accessibility index (PED) to pedestrian crashes is similar to that of TAIto crashes and can be seen in the Fig. 3. The mean pedestrian crash rates for both EJ and non-EJ tracts decrease with decrease in PED levels. In the case of non-EJ tracts with High PED levels (on the left hand side of the bottom panel), the mean crash rate is 3.60, decreasing to 1.21 in the case of tracts with Low PED levels. The mean crash rates for each PED condition is higher for EJ tracts compared to non-EJ tracts; but in the case of EJ tracts too, the mean pedestrian crash rate decreases from 5.13 in the case of tracts with HIGH PED levels to a mean of 1.84 in the case of EJ tracts with Low PED levels. Overall, in both EJ and non-EJ tracts, there is an increase in the percent of census tracts with zero recorded crashes as PED levels decrease. However, in contrast to the case of TAI, the percent of both EJ and non-EJ tracts with High or Very High crash rates (over 9) mostly occur only where PED levels are High, with just a minimal percentage of tracts in the High–Medium and Low PED categories, in which such crash rates were experienced.

In summary, the mean crash rate of EJ tracts is estimated to be higher than in non-EJ tracts by 1.32, however, large differences do not exist between the two types of areas in the proportion of injuries of different types and in the proportion of fatalities. The mean crash rate and the percent of both non-EJ and EJ tracts with zero pedestrian crashes increase with decrease in their TAI and PED levels, and are highest in cases of tracts with High levels of TAI and PED. Further, the mean crash rates for each TAI and PED condition is higher for EJ tracts compared to non-EJ tracts. For each TAI and PED level, the percent of EJ tracts with Low crashes is lower than in comparable non-EJ tracts while the percent of EJ tracts with High–Medium crashes is greater than in non-EJ tracts.

5. Statistical models of pedestrian crash counts

The basic form of the statistical models considered is PED CRASHj = f(EJ TRACT, environmental factors, behavioral factors)ⁱ where PED CRASHj is the count of motor-vehicle–pedestrian crashes in the jth tract and EJ TRACTj is a binary variable taking the value of 1 if the jth tract is considered to be an EJ tract based on the rules given in Section 3.1 and the environmental and behavioral factors are as described in Section 3.2.

The crash data considered in this paper poses several methodological challenges including (i) the possibility of variations in the estimated effects of models arising from the choice of census tracts as the spatial unit, (ii) the presence of spatial autocorrelation in safety data, (iii) presence of overdispersion and heterogeneity and (iv) underreporting of crashes. Several different models were estimated in an attempt to address these different aspects of the data. The final results presented in Table 4 are from a Poisson regression with heterogeneity and a Poisson regression with heterogeneity which accounts for exogenous underreporting that are most relevant in explaining the distribution of PED CRASH with census tract-level environmental and behavioral factors.

Variations in the estimated model effects can arise from the choice of census tracts as the areal unit rather than zip code or some other spatial unit; this is referred to as the modifiable areal unit problem (MAUP) (Openshaw, 1979; Holt et al., 1996; Anselin et al., 2000). Aside from the reasons given in Section 3.1, our preference behind using census tracts in this analysis is that most EJ analysis have used this unit (Mid-America Regional Council, 2007; US EPA, 2009) and our results extend safety research into the emerging area of environmental justice analysis. The extent to which the model parameters are sensitive to the choice of this analysis is left for future research.

The presence of spatial autocorrelation can violate the independence assumption made by many statistical models (LeSage and Pace, 2004). A Local Indicators of Spatial Association (LISA) significance map of PED CRASH locations (as shown in Fig. 4) provides an indication of the existence of clusters in the spatial arrangement of pedestrian crashes. The similarity of the maps after both 1 and 199 permutations indicates that the sensitivity is fairly low, and supports the existence of spatial clusters. A visual review of the residual versus predicted PED CRASH values from the OLS model also supports the presence of spatial autocorrelation, indicating that crashes tend to be overpredicted in some suburban locations, and underpredicted in the Chicago downtown business district (the "Loop") and in southern areas of the city of Chicago. The calculated Moran's I value of 0.3984 was also significant, indicating that

Table 4 Alternative models, parameter estimates and measures of fit.

| Variable | Model I |-----------------------|-------- | | Intercept | −4.21*** | EJ TRACT (1 = "Yes") | 0.65*** | Environmental factors | | TAI2 | 2.01*** | PED LOW | 1.61*** | SUM AADT2 | 0.48e−06*** | SUM LENGTH2 | −0.28e−03*** | NO SCHOOLS | 0.19** | POP SQMILE | 0.91e−05 | PERCRIME | 0.24** | PED | 0.09*** | Behavioral factors | | WLKTOWRK | 0.0008*** | MEDHHINC99 | −0.22e−06 | PERNOCAR | 2.60** | PER COMM | 1.37 | PERCHILDREN | −2.09 | PERLOWENGLISH | 0.21 | | | Intercept | | COUNTY (1 = "Cook") | | R2 | 0.58# | Log–likelihood | −1763.25 | 2/df | 136.8 | Vuong's statistic | – | | | – | | | | ------------------------------------------|-------------------------------------------------------------|--|--| | Poisson with
heterogeneity
Marginal effect | Poisson with exogenous
underreporting
Marginal effect | | | | −2.13*** | | | | 0.33*** | | | | | | | | 1.01*** | | | | 0.59*** | | | | 0.24e−06*** | | | | −0.14e−03*** | | | | 0.09** | | | | 0.46e−03 | | | | 0.12** | | | | 0.05*** | | | | | | | | 0.0009 | | | | −0.11e−05 | | | | 1.31** | | | | 0.69 | | | | −1.05 | | | | −0.10 | | | | Probit reporting equation | | | | 0.54e−07 | | | | 0.018** | | | | 0.61# | | | | −1511.36 | | | | 93.76 | | | | −60.75 | | | | 0.13 (p < 0.0001) | 0.18 (p < 0.517) | | | | 0.00 | | |

^** Significant at 0.05. *** Significant at 0.01.

there is spatial autocorrelation in the residuals. While accounting for such spatial autocorrelation is important for this problem and a spatial autoregressive model was presented in an earlier version of this paper (Cottrill and Thakuriah, 2008), that model was limited in that the distribution of PED CRASH was assumed to be normal and further work is necessary to develop a spatial model with the appropriate distributional assumption for PED CRASH. Since there is, simultaneously, evidence of spatial autocorrelation, underreporting and heterogeneity in the count data under consideration in this paper, in the most ideal case, our requirement is to use a count data model that explicitly takes into account the spatial structure, underreporting and evidence of heterogeneity. In the model con-

Fig. 4. LISA cluster map: pedestrian crashes.

^# Pearson's R² given as R² ^p = 1 − -J j=1 yj ln(yj/j)/ J j=1 yj ln(yj/y¯)

sidered here, we account for three out of the four requirements – and leave the explicit introduction of spatial dependence in count models of PED CRASH with heterogeneity and underreporting for future research.

Table 4 reports the marginal effects of a Poisson regression that accounts for heterogeneity and a Poisson regression that accounts for both heterogeneity and potential underreporting of crashes. Associated statistics and measures of fit are also presented. Diagnosis of a preliminary OLS model showed there was no evidence of multicollinearity, which may be due to combining many different individual factors into the composite indices, in contrast to introducing them as separate effects. Partial residual plots suggested that several variables need transformation; in particular, TAI, and SUM AADT were squared, leading to improvement in model fit; SUM LENGTH, NO SCHOOLS and RANKTOT were also candidates for transformation, but we chose not to, as the parameter estimates and model fit did not substantially change with the transformations.

Poisson regression is typically used to model crash counts. However, the presence of overdispersion in the data led us to estimate a Negative Binomial model. In the incremental model development process, we estimated Poisson and Negative Binomial models with heterogeneity, of which, on the basis of three measures: Pearson's R2, log-likelihood and 2/df-statistic, we keep the Poisson Model with heterogeneity in Table 4 (under Model 1) for further consideration. The Poisson model with heterogeneity has a mean of exp(xˇ + ε) where ε∼N(0, -2) where ˇ's are parameters to be estimated, x's are exploratory variables explaining PED CRASHES and ε's are unobserved heterogeneity that are normally distributed across the individual data with mean 0 and standard deviation -The estimate of is 0.13 and is statistically significant at the .01 level. However, an additional issue, that of underreporting, was taken into account in the final model development step, which leads to Model II in Table 4.

Underreporting in low-income neighborhoods has been noted to be important in several studies. For example, Sciortino et al. (2005) found that the California Statewide Integrated Traffic Reporting System (SWITRS) under-reported San Francisco pedestrian injuries by 21% (using San Francisco General Hospital medical records as a gold standard), with African-Americans and males less likely to have a SWITRS-reported injury; Stutts et al. (1998) reported on the basis of specially-collected bicycle and pedestrian injury data from emergency hospitals that overall, 70% of the reported bicycle injury events and 64% of the reported pedestrian injury events did not involve a motor-vehicle.

From a modeling perspective, count data models that take heterogeneity and underreporting into account have been given by Winkelmann and Zimmermann (1993) and applied to safety data by Kumara and Chin (2005). Suchmodels acknowledge that only a subset of the actual number of crashes that occurred are reported and that the count of reported accidents is smaller than the total number of events. Let y* be the unobserved number of accidents and let y denote the recorded number of accidents. Let y*|ε have a Poisson distribution with mean E(y∗|x, ε) = exp(x ˇ + ε), where, as before, ˇ's are parameters to be estimated, x's are exploratory variables explaining PED CRASHES and ε's are unobserved heterogeneity that are normally distributed across the individual data with mean 0 and standard deviation -. Further, let ci denote a binary variable which is 1 if the net utility from reporting the ith crash is positive; ci is determined by a reporting process such that c^∗ ⁱ = zi + ui and ci = 1 iff c^∗ ⁱ > 0. Hence the number of reported crashes in census tract ^j is yj ⁼ ^y^∗ ⁱ=¹ci.

The vector z represents exploratory variables that determine the reporting mechanism; in the current case, the feasibility of several variables was examined, based on the literature. Ultimately, COUNTY, a binary variable, which is 1 in the case of Cook County, the most populous county, which includes the City of Chicago and includes over 88% of the EJ tracts, was included in z. The vector are parameters to be estimated and with a normality assumption for the ui's, we have a probit model for the selection mechanism. The Poisson probabilities in this model (Poisson with endogenous reporting), with the modified mean ¯ , are given by:

$$\begin{split} P(Y_j = y_j | \varepsilon_j) &= \frac{\exp^{-\tilde{\lambda}_j} \tilde{\lambda}_j^{y_j}}{y_j!} \ \text{where} \quad \tilde{\lambda}_j &= \exp(x_j \beta + \varepsilon_j) \times \varPhi\left(\frac{z_j \gamma + (\rho | \sigma) \varepsilon_j}{\sqrt{1 - \rho^2}}\right) \end{split}$$

where ˚ is the CDF of the standard normal distribution, is the correlation between ε and u, i.e. is the correlation between the variables relating to crash occurrence and reporting. If = 0, i.e., if reporting is independent from crash occurrence, we have a case of a Poisson model with exogenous underreporting with the conditional mean function:

$$\bar{\lambda}_j = \exp(x_j \beta + \varepsilon_j) \times \Phi(z_j \gamma)$$

The Poisson model with endogenous underreporting led to an estimate of which is small and not significant at any reasonable level of significance. On this basis, we estimated a Poisson model with exogenous underreporting, the results of which are shown as Model II in Table 4. In deciding whether to favor Model II or III, we use Vuong's test, along with three other measures of fit, including Pearson's R2, log-likelihood and 2/df. The Vuong's test statistic (V) is distributed as a standard normal variate and hence its value is comparable to the critical value of the standard normal distribution. Consequently, V > 1.96 favors the Poisson underreporting models, whereas V ≤ 1.96 favors the parent Poisson models. In the case of Model II, V = −60.75, which leads to the conclusion that Model II should be favored. Nevertheless, on the basis of the other measures: we explore the results of Model III further, since it leads to useful insights regarding underreporting and otherwise is a better fit compared to Model I (on the basis of the log-likelihood value at convergence, the Pearson's R² and the2/df value).

The reporting equation estimates that crashes in non-Cook County are likely to be under-reported by 1.8% points more than in Cook county. Since most EJ tracts are in Cook county, this result is contradictory with previous literature, which reports pedestrian crashes are more likely to be under-reported by minority and lowincome populations. We believe that this contradictory finding has to do with differences in the nature of the crashes and the severity of the crashes in Cook versus the suburban counties (here referred to as "collar counties"). Whereas crashes per capita are higher in Cook County, compared to collar counties, minor crashes (those without fatalities, incapacitating injuries or with other notable injuries) as a percent of all crashes is higher in collar counties compared to Cook. Minor crashes as a percent of all crashes are higher in EJ tracts than non-EJ tracts in collar counties. This is in contrast to Cook County, where this quantity is lower for EJ tracts than for non-EJ tracts. It is possible that minor crashes that are reported in EJ tracts in collar counties are only a subset of the total number of such crashes occurring there.

It may be recalled that the difference between mean PED CRASHES in EJ tracts and mean PED CRASHES in non-EJ tracts is 1.32. Controlling for environmental and behavioral factors, the OLS model estimates that this difference in average PED CRASH rates drops to 0.65. Across almost all the different models tested, the coefficient for EJ TRACT remains remarkably constant – between 0.62 to 0.67 – and statistically significant, indicating that holding the environmental and behavioral factors constant reduces the difference in PED CRASH between EJ and non-EJ tracts by about half.

With Model II, however, which corrects for underreporting, the estimate of the differences drops to 0.33. The discussion above, which attempts to relate differential reporting in EJ tracts between Cook and collar county, partly answers this reduction – once underreporting in non-Cook county tracts are accounted for, the difference in PED CRASH between EJ and non-EJ tracts reduces further.

The coefficients for TAI merit further discussion. The TAI estimates are significant in the three models shown in Table 4. In keeping with the Trellis plots in Figs. 2 and 3, as the value of these indicators increase, PED CRASH is estimated to increase. This finding implies that TAI is a potentially useful proxy measure to direct measurements on pedestrian exposure by tracts. PERCRIME is also an important predictor of PED CRASH; pedestrian crashes increase with recorded crime in the census tract. The exact mechanism of why this is so merits further research. It would also be interesting to find out whether some of the pedestrians involved in crashes in areas which are unsafe from a personal security perspective were in the process of avoiding crime being committed, or became careless with respect to the traffic environment, in attempts to deal with perceived unsafe situation that theymay have suddenly faced. This would require researchers to access detailed crime data from police department. The coefficient for SUM AADT² is as expected; however, the coefficient for SUM LENGTH² is negative, possibly because of the presence of freeway links in some of the census tracts, where there little possibility of walking. Whereas the coefficient of NO SCHOOLS is positive, as expected, that of PERCHIL-DREN is negative, indicating that child pedestrians may be more vulnerable to crashes in areas that are not within the immediate vicinity of their residential locations. WLKTOWRK is not significant at any level of significance. It is also likely that the effect of the density variable, POP00 SQMI, was explained by the variable EJ TRACT.

6. Summary and conclusions

This paper evaluates the incidence of pedestrian–vehicle crashes in high low-income and minority population (EJ) census tracts as compared to pedestrian–vehicle crashes in non-EJ tracts in the Chicago metropolitan area. The summary statistics presented above, primarily the overall incidence, are generally supportive of the finding that pedestrian–vehicle crashes are more prevalent in EJ census tracts, as indicated in previous research. By taking a two-pronged approach to the data set and addressing the issue from the standpoints of environmental characteristics and behavioral factors, these general findings may be expanded upon and more targeted methods of addressing this issue may be developed. While earlier studies have indicated an increased number of pedestrian–vehicle crashes in areas of high minority or low-income populations in comparison to those areas that do not meet these criteria, less attention has been paid to ascertaining contributing factors to this disparity. This paper has attempted to test both environmental and behavioral factors for their contributions to the perceived heightened incidence of these crashes in order to provide better guidance for potential needs for interventions and further study.

In general, the summary statistics indicate that the pattern of pedestrian–vehicle crashes in EJ and non-EJ tracts as evaluated in relation to environmental conditions are broadly similar, with comparable patterns obtained when evaluating these crashes against the transit availability index (TAI) and the pedestrian accessibility index (PED). These findings indicate that the general geographic characteristics of the pedestrian environment have roughly similar impacts on the likelihood of pedestrian–vehicle crashes in terms of rate, if not incidence. The relationship between TAI and pedestrian crashes may be especially significant, as it would suggest that expansions of and improvements to pedestrian-accessed transit stops should be planned and implemented in conjunction with pedestrian safety improvements in order to reduce the risk of incidents. Additionally, the relationship between PED and crashes may indicate the need to more thoroughly address the safety impacts of pedestrian improvements, particularly in the relationship between potential pedestrian and vehicle interactions, in order to more safely encourage walking. These findings are especially relevant in the domain of EJ regulations, as the environmental analysis indicates that EJ tracts tend to have higher levels of both TAI and PED, indicating the potential for more crashes.

The statistical analysis presented reveals a number of interesting relationships. First, the connection between crash underreporting and injury category within the collar counties surrounding the primary urban county in the metropolitan area suggests that there may be differences in these crashes unaccounted for in the model. Potential variables that may impact this underreporting and should be explored further include distance to police or hospital services and general insurance rates and coverage. It is possible, for example, that minor crashes in collar counties are less likely to impact travel flow or be observed by law enforcement, and so are not reported at the time of occurrence. Additionally, if either the driver or pedestrian is uninsured, it may be to their advantage to take care of the crash impacts outside of legal boundaries if the injuries are minor, or if no injury occurred. These and other potentially related variables should be explored further.

A second relationship of note is related to proxies for exposure rates. As noted above, high TAI and PED scores were related to an increase in crash incidence, which may indicate that they reveal areas of increased pedestrianism, thus increased exposure. Two other variables that may also be related to exposure rates are RANKTOTAL and PERCRIME, increases of which were also related to an increased number of crashes. For RANKTOTAL, the relationship is fairly clear, as this variable represents a ranking of census tracts by those factors that are typically associated with high pedestrian rates. PERCRIME, on the other hand, presents a less clear relationship, as noted in Section 5. Further analysis of pedestrian and driver behavior in areas of high crime levels should be conducted to more clearly evaluate how increased levels of crime impact pedestrian–vehicle crashes. Finally, the general spatial analysis presented here reveals the presence of spatial autocorrelation, indicating that a more thorough spatial analysis of the data should be conducted. Results from the initial spatial error model improved overall model performance, indicating that controlling spatial dependence would likely result in better overall model fit.

The implications of the findings of the analysis are fairly wideranging. As noted, further examination of a number of relationships is needed; however, it is evident that a number of factors related to pedestrian–vehicle crashes in EJ areas should be addressed. First, pedestrian and transit improvements and expansions should better incorporate safety considerations at locations of potential pedestrian–vehicle interaction. The raised incidence of pedestrian crashes in areas that rank highly for pedestrian amenities indicates that further analysis of safety requirements is needed, beyond the current practices. Secondly, further analysis of pedestrian and driver behavior in areas of high crime rates should be conducted in order to encourage more safe behavior in these areas. Finally, environmental justice analyses should include a review of how pedestrian–vehicle crash rates may be impacted by projects or programs developed for these areas.

References

Anderton, D.L., et al., 1994. Environmental equity: the demographics of dumping. Demography 31.2, 229–248, http://www.jstor.org/stable/2061884. Anselin, L., Cohen, J., Cook, D., Gorr, W., Tita., G., 2000. Spatial analyses of crimes. Criminal Justice 4.

Bagley, C., 1992. The urban setting of juvenile pedestrian injuries: a study of behavioural ecology and social disadvantage. Accident Analysis and Prevention 24 (6), 673–678.
Beck, L.F., Leonard, J.P., Stephen, C.D., 2007. Pedestrian fatalities, Atlanta metropolitan statistical area and United States, 2000–2004. Journal of Safety Research 38 (6), 613–616.
Been, V., 1994. Locally undesirable land uses in minority neighborhoods: disproportionate sitting or market dynamics? Yale Law Review 103.
Bijur, P.E., Stewart-Brown, S., Butler, N., 1986. Child behavior and accidental injury in 11,966 preschool children. American Journal of Diseases of Children (1960) 140.5, 487–492.
Bogue, D.J., 1985. The Population of the United States: Historical Trends and Future Projections, 2nd ed. The Free Press, New York.
Brown, G.W., Davidson, S., 1978. Social class, psychiatric disorder of mother, and accidents to children. Lancet 1.8060, 378–381.
Cohen, J., 1992. A power primer. Psychological Bulletin 112 (1), 155–159.
Cottrill, C., Thakuriah, P., 2008. Evaluating pedestrian risk in environmental justice areas. In: Proceedings of the 87th Transportation Research Board Annual Conference.
Denton, N.A., Massey, D.S., 1991. Patterns of neighborhood transition in a multiethnic world: U.S. Metropolitan Areas, 1970–1980. Demography 28.1, 41–63, http://www.jstor.org/stable/2061335.
Dougherty, G., Pless, I.B., Wilkins, R., 1990. Social class and the occurrence of traffic injuries and deaths in urban children. Canadian Journal of Public Health. Revue Canadienne de Sante Publique 81 (3), 204–209.
Edwards, P., Green, J., Roberts, I., Lutchmun, S., 2006. Deaths from injury in children and employment status in family: analysis of trends in class specific death rates. British Medical Journal 333 (July), 119–121.
EPA, 1994. Executive Order 12898 of February 11, 1994. Federal Register Environmental Documents, http://www.epa.gov/fedrgstr/eo/eo12898.htm.
FHWA, 1999. Guidebook on Methods to Estimate Non-Motorized Travel: Overview of Methods, p. 22.
Graham, D.J., Glaister, S., 2003. Spatial variation in road pedestrian casualties: the role of urban scale, density and land-use mix. Urban Studies 40 (July (8)), 1591–1607.
Guyer, B., Talbot, A.M., Pless, I.B., 1985. Pedestrian injuries to children and youth. Pediatric Clinics of North America 32.1, 163–174.
Holt, D., Steel, D.G., Tranmer,M.,Wrigley, N., 1996. Aggregation and ecological effects in geographically based data. Geographical Analysis 28, 224–262.
Illinois Department of Transportation, 2005. IDOT Crash Report Data Dictionary. Kumara, S.S.P., Chin., H.C., 2005. Application of Poisson underreporting model to examine crash frequencies at signalized three-legged intersections. Transportation Research Record 1908, 46–50.
Laflamme, L., Diderichsen, F., 2000. Social differences in the traffic injury risks in childhood and youth—a literature review and a research agenda. Injury Prevention 6, 293–298.
LaScala, E.A., Gruenewald, P.J., Johnson, F.W., 2004. An ecological study of the locations of schools and child pedestrian injury collisions. Accident Analysis and Prevention 36.4, 569–576.
LaScala, E.A., Gerber, D., Gruenewald, P.J., 2000. Demographic and environmental correlates of pedestrian injury collisions: a spatial analysis. Accident Analysis and Prevention 32, 651–658.
LaScala, E.A., Johnson, F.W., Gruenewald, P.J., 2001. Neighborhood characteristics of alcohol related pedestrian injury collisions: a geostatistical analysis. Prevention Science 2 (2), 123–134.
LeSage, J.P., Pace, R.K., 2004.Models for spatially dependentmissing data. The Journal of Real Estate Finance and Economics 29 (2), 233–254.
Loukaitou-Sideris, A., Liggett, R., 2007. Death on the crosswalk: a study of pedestrian accidents in Los Angeles. Funded by a UCTC year 16 research grant. Journal of Planning Education and Research.
Lyons, R.A., Towner, E., Christie, N., Kendrick, D., Jones, S.J., Hayes, M., Kimberlee, R., Sarvotham, T., Macey, S., Brussoni, M., Sleney, J., Coupland, C., Phillips, C., 2008. The advocacy in action study: a cluster randomized controlled trial to reduce pedestrian injuries in deprived communities. Injury Prevention 14 (2).
Mid-America Regional Council, 2007. Transportation Improvement Program 2008-2012: Environmental Justice Analysis., http://www.marc.org/ transportation/tip/TIP08-12.htm.
Minocha, I., Sriraj, P.S., Metaxatos, P., Thakuriah, P., 2008. Analysis of transit quality of service and employment accessibility for the greater Chicago, Illinois, region. Transportation Research Record: the Journal of the Transportation Research Board 2048, 20–29.
Mohai, P., 1995. Methodological Issues: The Demographics of Dumping Revisited: Examining the Impact of Alternate Methodologies in Environmental Justice Research. Copyright (c) 1995 Virginia Environmental Law Journal Association Virginia Environmental Law Journal.
Murtha, T., October, 2005. Pedestrian and bicycle safety data in Northeastern Illinois. In: Presentation to Association of Pedestrian and Bicycle Professionals, http://www.walkinginfo.org/pp/ppt/Murthanotes.ppt.
National Highway Traffic Safety Administration [NHTSA], 2006. Pedestrians [DOT HS 810 624]. Washington, DC.
Openshaw, S., 1979. A million or so correlation coefficients: three experiments on the modifiable areal unit problem. In: Wrigley, N. (Ed.), Statistical Methods in the Spatial Sciences. Pion, London.
Padilla, E.R., Rohsenow, D.J., Bergman, A.B., 1976. Predicting accident frequency in children. Pediatrics 58.2, 223–226.
Paulozzi, L.J., 2006. Is it safe to walk in the Sunbelt? Geographic variation among pedestrian fatalities in the United States, 1999–2003. Journal of Safety Research 37, 453–459.
Raford, N., Ragland, D.R., 2003. Space Syntax: An Innovative Pedestrian VolumeModeling Tool for Pedestrian Safety. Safe Transportation Research & Education Center, UC Berkeley, retrieved from: http://escholarship.org/uc/item/50m064zp.
Sargent, M., Begg, D., Broughton, J., Stephenson, S., Wright, C., Baxter, J., 2004. Motor vehicle traffic crashes involving Maori. Journal of the New Zealand Medical Association. 117 (January (1188)).
Sciortino, S., et al., 2005. San Francisco pedestrian injury surveillance: mapping, under-reporting, and injury severity in police and hospital records. Accident Analysis and Prevention 37.6, 1102–1113.
Sriraj, P.S., Fruin, G., McNeil., S., 2003. Analysis of the environmental justice compliance of the Chicago Transit Authority. In: Proceedings of the Mid-Continent Transportation Research Symposium.
Stutts, J.C., Stewart, J.R., Martell, C., 1998. Cognitive test performance and crash risk in an older driver population. Accident Analysis and Prevention 30.3, 337–346.
Surface Transportation Policy Project (STPP), 2002. Mean Streets 2002., http://www.transact.org/PDFs/ms2002/MeanStreets2002.pdf. US EPA, 2009. Environmental Justice Strategic Enforcement Assessment Tool
(EJSEAT)., http://www.epa.gov/compliance/resources/policies/ej/ej-seat.html. Wedagama, D.M.P., Bird, R.N., Metcalfe, A.V., 2006. The influence of urban land-use on non-motorised transport casualties. Accident Analysis and Prevention 38,
1049–1057. Wier, M., et al., 2009. An area-level model of vehicle–pedestrian injury collisions with implications for land use and transportation planning. Accident Analysis
and Prevention 41.1, 137–145. Winkelmann, R., Zimmermann, K.F., 1993. Poisson-Logistic Regression. Department of Economics, University of Munich, Working Paper No. 93-18.

cottrillevaluatingpedestriancrashesareas2010.md 80 KB History Raw