xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam 2200385Ka 4500
controlfield tag 001 002063950
007 cr mnu|||uuuuu
008 100318s2009 flu s 000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0003033
Air toxics and equity :
b a geographic analysis of environmental health risks in Florida
h [electronic resource] /
by Angela Gilbert.
[Tampa, Fla] :
University of South Florida,
Title from PDF of title page.
Document formatted into pages; contains 101 pages.
Thesis (M.A.)--University of South Florida, 2009.
Includes bibliographical references.
Text (Electronic thesis) in PDF format.
ABSTRACT: A large number of quantitative studies have examined social inequities in the geographic distribution of air pollution. Although previous research has made strides towards understanding the nature and extent of inequities, they have been limited methodologically in three ways. First, the presence of pollutants have been rarely linked to their adverse health effects, with many studies using proximity to sources as a proxy for risk. Second, there has been a tendency to study a single pollution source instead of assessing multiple types of sources. Finally, conventional statistical methods such as multivariate regression have been limited by their inability to discern spatial variations in the relationships between dependent and explanatory variables. This thesis addresses these gaps in environmental justice analysis of air pollution by using data from U.S. Environmental Protection Agency's 1999 National-Scale Air Toxics Assessment in combination with 2000 U.S. Census data to evaluate inequities in the geography of cancer risks from hazardous air pollutants in Florida.The objective is to determine if there are racial/ethnic inequities in the distribution of estimated cancer risks from outdoor exposure to point and mobile sources of air pollutants, after controlling for well-documented contextual variables. The first phase of the study utilizes traditional correlation and regression techniques to reveal that cancer risk from most air pollution sources are distributed inequitably with respect to race, ethnicity, and socioeconomic state. In the second phase, geographically weighted regression is used along with choropleth mapping to explore the spatial nonstationarity of regression model parameters and geographic variations in the statistical association between cancer risks and various explanatory variables.Results indicate that while Black and Hispanic proportions remain consistent indicators of cancer risk from most pollution sources, these relationships vary across space within Florida. This thesis contributes to environmental justice analysis by demonstrating that conventional multivariate regression can hide important local variations in the relationships between environmental risk and explanatory variables such as race, ethnicity, and socioeconomic status. Since this spatial nonstationarity can be significant within an entire region or a single urban area, understanding its nature and extent is imperative to advancing environmental justice goals.
Mode of access: World Wide Web.
System requirements: World Wide Web browser and PDF reader.
Advisor: Jayajit Chakraborty, Ph.D.
Geographically weighted regression
t USF Electronic Theses and Dissertations.
Air Toxics and Equity: A Geographic Analysis of Environm ental Health Risks in Florida by Angela Gilbert A thesis submitted in partial fulfillment of the requirements for the degree of Master of Arts Department of Geography College of Arts and Sciences University of South Florida Major Professor: Jaya jit Chakraborty, Ph.D. Pratyusha Basu, Ph.D. Steven Reader, Ph.D. Date of Defense: April 30, 2009 Keywords: environmental justice, air pollution, NATA, geographically weighted regression, public health Copyright 2009, Angela Gilbert
ii ACKNOWLEDGEMENTS This thesis would not be possible without th e guidance, clarity, and confidence of Jay Chakraborty. Thank you. Much gratitude goes to my family for their kindness, love, and generosity. Finally, IÂ’d like to acknowle dge Wilco, whose music provid ed sanity and perspective during the long, late hours of reading, writing, and analyzing.
i TABLE OF CONTENTS List of Tables ................................................................................................................ List of Figures ............................................................................................................... Abstract ...................................................................................................................... ... Chapter 1. Introduction ................................................................................................. Chapter 2. Background and Literature Review ............................................................. 2.1 Measuring Exposure to Air Pollution ......................................................... 2.1.1 From Individual Source Analysis to a Cumulative Approach ..... 2.1.2 Measuring the Extent of Exposure to Air Pollution ..................... 2.2 Statistical Methods ...................................................................................... 2.3 Summary ..................................................................................................... Chapter 3. Study Area and Methodology ..................................................................... 3.1 Study Area .................................................................................................. 3.2 Data Sources and Variables ........................................................................ 3.2.1 Dependent Variables .................................................................... 3.2.2 Explanatory Variables .................................................................. 3.3 Methods....................................................................................................... Chapter 4. Conventional Environmental Ju stice Analysis of Estimated Cancer Risks ........................................................................................................................ 4.1 Descriptive Choropleth Mapping and Statistics ......................................... 4.2 Bivariate Correlation Analysis .................................................................... 4.3 Traditional Multiple Regression Analysis .................................................. Chapter 5. Geographically Weighted Regre ssion Analysis of Estimated Cancer Risks ........................................................................................................................ 5.1 Cumulative Cancer Risk from All Sources ................................................. 5.2 Lifetime Cancer Risk from Major Point Sources ....................................... 5.3 Lifetime Cancer Risk from Other Point Sources ........................................ 5.4 Lifetime Cancer Risk from On-Road Mobile Sources ............................... 5.5 Lifetime Cancer Risk from Non-Road Mobile Sources ............................. iii iv vii 1 6 7 7 11 15 18 20 20 21 23 25 28 32 32 40 42 46 48 57 62 69 72
ii 5.6 Summary of GWR Results ......................................................................... Chapter 6. Conclusions ................................................................................................. References .................................................................................................................... Appendices .................................................................................................................... Appendix A ....................................................................................................... 78 81 88 95 96
iii LIST OF TABLES Table 3.1 Definitions of Explanatory Variables and Expected Statistical Association with Cancer Risk ............................................................................ Table 4.1 Descriptive Statistic s for Variables Analyzed ......................................... Table 4.2 PearsonÂ’s Correlation Coefficien ts of Lifetime Cancer Risk by Source and Explanatory Variables ................................................................................. Table 4.3 Multivariate Regression Results: Cumulative Cancer Risk from All Sources ............................................................................................................... Table 4.4 Multivariate Regression Results: Cancer Risk from Four Known Sources ............................................................................................................... Table 5.1 GWR Results for Cumulative Cancer Risk from All Sources ................. Table 5.2 GWR Results for Cancer Ri sk from Major Point Sources ...................... Table 5.3 GWR Results for Cancer Ri sk from Minor Point Sources ...................... Table 5.4 GWR Results for Cancer Risk from On-Road Mobile Sources .............. Table 5.5 GWR Results for Cancer Ri sk from Non-Mobile Sources ...................... 28 40 41 43 44 49 58 63 70 73
iv LIST OF FIGURES Figure 4.1 Estimated Lifetime Cancer Risk From All Sources by Census Tract, 1999....................................................................................................................... Figure 4.2 Estimated Lifetime Cancer Risk From Major Point Sources by Census Tract, 1999 ............................................................................................................ Figure 4.3 Estimated Lifetime Cancer Ri sks From Minor Point Sources, 1999 ........ Figure 4.4 Estimated Lifetime Cancer Risk From On-Road Mobile Sources, 1999 .. Figure 4.5 Estimated Lifetime Cancer Risk From Non-Road Mobile Sources, 1999....................................................................................................................... Figure 4.6 Proportion Non-Hispanic Black by Census Tract, 2000 ........................... Figure 4.7 Proportion Hispanic by Census Tract, 2000 .............................................. Figure 4.8 Proportion Asian by Census Tract, 2000 ................................................... Figure 4.9 Homeowner Occupied Housi ng Units by Census Tracts, 2000 ................ Figure 4.10 Proportion of Population Belo w the Federal Poverty Rate, 2000 ........... Figure 4.11 Population Density by Census Tract, 2000 ............................................. Figure 5.1 Local R-Squared from GWR Model of Cancer Risk from All Sources .... Figure 5.2 Distribution of t-values for Pr oportion Black and Cancer Risk from All Sources .................................................................................................................. Figure 5.3 Distribution of t-values for Pr oportion Hispanic and Cancer Risk from All Sources ............................................................................................................ Figure 5.4 Distribution of t-values for Pr oportion Asian and Cancer Risk from All Sources .................................................................................................................. Figure 5.5 Distribution of t-values for Ho me Ownership Rate and Cancer Risk from All Sources ................................................................................................... Figure 5.6 Distribution of t-values for P overty Rate and Cancer Risk from All Sources .................................................................................................................. Figure 5.7 Distribution of t-values for P opulation Density and Cancer Risk from All Sources ............................................................................................................ Figure 5.8 Local R-Squared of Cancer Risk from Major Point Sources .................... Figure 5.9 Distribution of t-values for Proportion Black and Cancer Risk from Major Point Sources .............................................................................................. 33 34 34 35 35 36 37 37 38 38 39 52 53 53 54 54 55 55 59 59
v Figure 5.10 Distribution of t-values for Proportion Hispanic and Cancer Risk from Major Point Sources ..................................................................................... Figure 5.11 Distribution of t-values for Pr oportion Asian and Cancer Risk from Major Point Sources .............................................................................................. Figure 5.12 Distribution of t-values for Ho me Ownership Rate and Cancer Risk from Major Point Sources ..................................................................................... Figure 5.13 Distribution of t-values for Pove rty Rate and Cancer Risk from Major Point Sources ........................................................................................................ Figure 5.14 Local R-Squared of Cancer Risk from Other Point Sources ................... Figure 5.15 Distribution of t-values for Pr oportion Black and Cancer Risk from Other Point Sources .............................................................................................. Figure 5.16 Distribution of t-values for Proportion Hispanic and Cancer Risk from Other Point Sources...................................................................................... Figure 5.17 Distribution of t-values for Pr oportion Asian and Cancer Risk from Other Point Sources .............................................................................................. Figure 5.18 Distribution of t-values for Ho me Ownership Rate and Cancer Risk from Other Point Sources...................................................................................... Figure 5.19 Distribution of t-values for Pove rty Rate and Cancer Risk from Other Point Sources ........................................................................................................ Figure 5.20 Distribution of t-values for P opulation Density and Cancer Risk from Other Point Sources .............................................................................................. Figure 5.21 Local R-Squared of Cancer Ri sk from On-Road Mobile Sources .......... Figure 5.22 Distribution of t-values for P overty Rate and Cancer Risk from OnRoad Mobile Sources ............................................................................................ Figure 5.23 Local R-Squared of Cancer Ri sk from Non-Road Mobile Sources ........ Figure 5.24 Distribution of t-values for Pr oportion Black and Cancer Risk from Non-Road Mobile Sources .................................................................................... Figure 5.25 Distribution of t-values for Proportion Hispanic and Cancer Risk from Non-Road Mobile Sources ........................................................................... Figure 5.26 Distribution of t-values for Pr oportion Asian and Cancer Risk from Non-Road Mobile Sources .................................................................................... Figure 5.27 Distribution of t-values for Ho me Ownership Rate and Cancer Risk from Non-Road Mobile Sources ........................................................................... Figure 5.28 Distribution of t-values for P overty Rate and Cancer Risk from NonRoad Mobile Sources ............................................................................................ Figure 5.29 Distribution of t-values for P opulation Density and Cancer Risk from Non-Road Mobile Sources .................................................................................... 60 60 61 61 65 65 66 66 67 67 68 71 71 74 74 75 75 76 76 77
vi Figure A-1 Frequency Histogr am of Cumulative Cancer Risk per Million People ... Figure A-2 Frequency Histogram of Majo r Point Sources of Cancer Risk per Million People ....................................................................................................... Figure A-3 Frequency Histogr am of Other Point Sources of Cancer Risk per Million People ....................................................................................................... Figure A-4 Frequency Histogram of On-Road Mobile Sources of Cancer Risk per Million People ....................................................................................................... Figure A-5 Frequency Histogr am of Non-Road Mobile Sources of Cancer Risk per Million People ................................................................................................. Figure A-6 Frequency Histogram of Proportion Black by Census Tract, 2000 ......... Figure A-7 Frequency Histogr am of Proportion Hispanic by Census Tract, 2000 .... Figure A-8 Frequency Histogram of Proportion Asian by Census Tract, 2000 ......... Figure A-9 Frequency Hist ogram of Poverty Rate by Census Tract, 2000 ................ Figure A-10 Frequency Hist ogram of Home Ownership Rate by Census Tract, 2000....................................................................................................................... Figure A-11 Frequency Hist ogram of Population Density by Census Tract, 2000 .... Figure A-12 Frequency Histogram of th e Natural Log of Population Density by Census Tract, 2000................................................................................................ 96 96 97 97 98 98 99 99 100 100 101 101
vii AIR TOXICS AND EQUITY: A GEOGRAPHIC ANALYSIS OF ENVIRONMENTA L HEALTH RISKS IN FLORIDA Angela Gilbert ABSTRACT A large number of quantitative studies have examined social inequities in the geographic distribution of ai r pollution. Although previous re search has made strides towards understanding the nature and extent of inequities, they have been limited methodologically in three ways. First, the pres ence of pollutants have been rarely linked to their adverse health effects, with many st udies using proximity to sources as a proxy for risk. Second, there has been a tendency to study a single polluti on source instead of assessing multiple types of sources. Finally, conventional statistical methods such as multivariate regression have been limited by thei r inability to discern spatial variations in the relationships between depende nt and explanatory variables. This thesis addresses these gaps in envi ronmental justice analysis of air pollution by using data from U.S. Environmental Protection AgencyÂ’s 1999 National-Scale Air Toxics Assessment in combination with 2000 U.S. Census data to evaluate inequities in the geography of cancer risks from hazardous air pollutants in Florida. The objective is to determine if there are racial/ethnic inequities in the distribution of estimated cancer risks from outdoor exposure to point and mobile sour ces of air pollutants, after controlling for well-documented contextual variables. The fi rst phase of the study utilizes traditional correlation and regression techni ques to reveal that cancer ri sk from most air pollution
viii sources are distributed inequitably with re spect to race, ethnicity, and socioeconomic state. In the second phase, geographically we ighted regression is used along with choropleth mapping to explore the spatial nonsta tionarity of regression model parameters and geographic variations in the statistical association between can cer risks and various explanatory variables. Results indicate that while Black and Hispanic proportions remain consistent indicators of cancer risk from most pollution source s, these relationships vary across space within Florida. This thesis cont ributes to environmental justice analysis by demonstrating that conventional multivariate regression can hide important local variations in the relationship s between environmental risk an d explanatory variables such as race, ethnicity, and socioeconomic status Since this spatial nonstationarity can be significant within an entire region or a single urban area, understa nding its nature and extent is imperative to advanc ing environmenta l justice goals.
1 CHAPTER 1: INTRODUCTION Increasingly in the last few decades, ge ographers have been concerned with the unequal spatial distribution of both the bene fits and negative byproducts of modern society. Of particular intere st is the distribution of th e adverse effects of various technological and natural hazards (Walke r and Bulkeley 2006). By documenting and analyzing the characteristics of communities that are exposed to the externalities of industrial and commercial growth such as air pollution and it s adverse health outcomes, geographers hope to uncover dist ributional inequities of envi ronmental risks and advance the aims of social justice. Environmental justice is concerned with spatial and social inequities in the distribution of environmental pollution and adverse health consequences of industrial activities and environmental policies. Envi ronmental justice began as a movement stemming from an incident in Warren County, North Carolina. In 1982, the state of North Carolina selected a predominately African-Ame rican and low-income area to locate a toxic waste landfill for the disposal of polychlorinated biphenyls (PCBs) dumped illegally in other areas of the state. The site chosen was declared by experts to be unsafe and unsuitable. The community, believing they had been targeted due to their poverty and racial composition, engaged in legal action and a campaign of civil disobedience to keep the landfill out of their neighborhood. Despite their protests and more than 500 arrests, the landfill was placed in the planned location in 1983 (McGurty 2000). Although the residents of Warren County failed to keep th e hazardous waste out of their community,
2 they succeeded in attracting national attention to the issue of environmental justice. As a result, the U.S. General Accounting Office ( 1983) launched an investigation into the distribution of hazardous facili ties in the South and found that Black residents comprised the majority of the population in three of the four commun ities that contained landfills. This report was followed by a more comprehensive national study conducted by the United Church of Christ (UCC) Commissi on for Racial Justic e (1987). This study indicated that race was the most significan t factor in determining the location of commercial hazardous waste facili ties and that three out of every five Black and Hispanic individuals in the U.S. lived in communitie s containing uncontrolled waste sites. The 1987 UCC report set the wheels in motion for va rious quantitative and qualitative studies that attempted to seek empirical evidence for the claims made by environmental justice activists. Quantitative environmental justice resear ch has sought to provide statistical evidence of environmental inequity, or the di sproportionate distribution of environmental burdens on people and places. These studies have led to the implementation of policies at the national, state, and local level, incl uding President ClintonÂ’s 1994 Executive Order that requires federal agencies to include e nvironmental justice considerations in their plans, programs, and all aspects of operati on. The U.S. Environmental Protection Agency (EPA) has recently expanded the definition of environmental justice to the Â“the fair treatment and meaningful involvement of all people regardless of race, color, national origin, or income with respect to the deve lopment, implementation, and enforcement of environmental laws, regulations, and policies. Â” (EPA 2008a). Even w ith these policies in place, an update of the UCC study on the stat e of environmental ju stice shows that race
3 remains the most statistically significant in dicator of the presence of hazardous waste sites in the U.S. (UCC 2007). While various types of hazardous facilitie s and undesirable land uses have been investigated, the distributional impacts of ai r pollution remain a persistent public health and social concern. Although previous empiri cal studies have made important strides towards understanding the causes and consequences of the inequities in the geography of air pollution, they have been limited methodologically in three critical ways. First, there has been a lack of connect ion between the presence of emissions and the adverse health risks they engender. Many studies have ut ilized proximity to pollution sources as a proxy for risk (e.g., Pollock and Vitt as 1995; Cutter et al. 1996; Perlin et al. 2001; Pastor et al. 2004), but this approach has several drawbacks. An exclusive focus on proximity ignores the quantity, toxicity, and environmental fate of released chemicals. Local meteorological conditions and other factors that heavily influence the direction and distance traveled by pollutants are also overl ooked. While specific studies have focused on modeling exposure to toxic pollution (e.g., Chakraborty and Armstrong 1997; Bevc et al. 2007), few have attempted to examine wh ether unequal exposure patterns lead to disproportionate health risks among mi nority and low-income communities. A second pitfall of environmental justice research on air pollution is the tendency to study a single pollution source, particularly in dustrial facilities, inst ead of cumulatively assessing exposure to multiple chemicals and emission sources. It is equally important to consider mobile emission sources, smaller emitters, and preexisting background concentrations of naturally occurring and pe rsistent air pollution. While other overlooked polluters such as dry-cleaning facilities, au to-body shops, and off-road mobile sources
4 may release fewer quantities of air polluta nts than their large counterparts, they potentially and cumulatively contribute to adve rse health risks in residential areas (Fitos and Chakraborty 2003). Finally, statistical methods such as multiv ariate regression have been used to examine the inequity hypothesis by evaluati ng the association be tween magnitude of pollution and well-documented socioeconomic and demographic variables. Conventional regression techniques, however, are limited by their inability to discern local variations in the relationships between th e dependent and independent variables. When assessing environmental injustice across a large area such as a nation, re gion, or state, the lack of geographic specificity can obfuscate underlying patterns of inequity. This thesis seeks to address these th ree methodological gaps in quantitative environmental justice analysis of air pollution through a case study that examines adverse health risks from multiple types of pollution s ources in the state of Florida. The research utilizes an EPA database of ambient air emission information and a regression methodology that allows the evaluation of geogra phic variation in anal ytical results. The specific objective is to determine if there ar e racial/ethnic and soci oeconomic inequities in the distribution of estimated cancer risks fr om outdoor exposure to both stationary and mobile sources of air polluta nts, after controlling for well-documented contextual variables. The specific research questions investigated in Florida are as follows: (a) Is cancer risk from outdoor exposure to hazardous air pollutants from various known sources distributed inequitably w ith respect to r ace/ethnicity and socioeconomic status?
5 (b) How does the strength and significance of the statistical relationships between cancer risks from various known sources and race/ethnicity or socioeconomic status vary across the state? The key data source is the EPAÂ’s 1999 Na tional-Scale Air Toxics Assessment (NATA), which integrates information from the local, state and federal levels in order to produce health risk estimates at the census trac t level from four sour ces of air pollution: major stationary sources, other stationary s ources, on-road mobile sources, and non-road mobile sources. A pertinen t set of variables desc ribing population and housing characteristics at the census tract level fr om U.S. Census 2000 is used to analyze inequities in modeled health risks. The firs t phase of the analysis utilizes traditional techniques such as bivariate linear correla tion and multivariate regression analysis to examine the relationship between estimated can cer risk from the four source categories and relevant racial/ethnic, demographic, a nd socioeconomic characteristics. The second phase uses a local spatial statistical te chnique known as geographically weighted regression (GWR) in conjunction with chor opleth mapping to address the problems inherent in conventional regr ession and investigate the spat ial nonstationarity of model parameters and model performance. GWR can be used to examine geographic variations in estimated regression parameters and provides localized coefficients for analytical units in a given study area (Fot heringham et al. 2002).
6 CHAPTER 2: BACKGROUND A ND LITERATURE REVIEW The research literature on the inequitabl e distribution of the adverse effects of technological hazards, specifically air pollu tion, can be generally classified into two distinct categories: (a) studi es that document the histor ical processes leading to environmental inequities and grassroots move ments to achieve environmental justice, and (b) studies that attempt to determine if curre nt patterns of environmental risk resulting from past decisions are inequitable with resp ect to race, ethnicity, or income. Research concerned with the historical production of sp atial and social inequities has relied on qualitative methods such as interviews, pa rticipant observation, and archival document analysis (e.g., Pulido 2000, Boone 2002, Grineski et al. 2005). Quantitative methods are typically used to describe th e nature and extent of the st atistical association between a measure of the presence or magnitude of envi ronmental pollution and the racial/ethnic or socioeconomic characteristics of th e potentially exposed population. Undesirable land uses and hazards that have been studied in environmental justice literature include landfills (e.g., Been 1994, Liu 1997), industrial manufacturing facilities releasing toxic chemicals (e.g., Mitchell et al. 1999, Pastor et al. 2004), hazardous waste treatment, storage, and disposal facilities (e.g., Anderton et al. 1994, Pastor et al. 2001), facilities storing chemicals classified as extremely hazardous substances (e.g., Chakraborty 2001), nuclear power plants (e.g., Knezevic and Chakraborty 2004), and noise pollution (e.g., Most et al. 2004). Desp ite the abundance of hazards that could potentially impact human health and welfare, air pollution was one of the first public
7 concerns to be directly addre ssed by legislation in the U.S. with the passage of the Clean Air Act in 1970. Inequities in the spatial distribu tion of air pollution have been studied since the 1970s (Freeman 1972, Asch and Se neca 1978, and Gianessi et al. 1979), even before the emergence of the environmental ju stice movement and research agenda in the 1980s. In addition to often being colorless and odorless, air pollution can be difficult to ascertain due to its quick dispersion, leavi ng those impacted by it pot entially unaware of its presence. Analyzing the environmental justice imp lications of exposure to air pollution using quantitative methods requires: (a) the se lection of the measurement used to detect the presence of emissions, and (b) methods for estimating the magnitude of exposure to the pollution. Both decisions come with potentia l problems that will be explored in this literature review. The first se ction will review research c oncerning the selection of the source of exposure, then the magnitude of the exposure to air pollution. The second section will provide an overview of common statistical methods used in environmental justice research to measure relationships between variables. 2.1 Measuring Exposure to Air Pollution 2.1.1 From Individual Source Analysis to a Cumulative Approach Air pollution is the amount of contaminants in the air or the presence of one or more contaminants at concentrations high e nough to cause adverse he alth effects (Godish 2004). Air pollution is produced by many different types of sources, and while large factories that release ominous plumes may a ttract more policy and media attention, other sources of air pollution that ha ve a negative impact on public h ealth are just as important to assess. In addition to point sources such as factories and power plants, there are less
8 conspicuous point sources like auto body and paint shops and dry cleaning facilities. There are also mobile sources, including car s, trucks, and motorcycles, and off road mobile sources such as airplanes, all-terr ain vehicles, and snow mobiles. Additionally, naturally occurring and lingering background con centrations of chemicals in the air can contribute to pollution levels Environmental justice research, however, has analyzed major point sources more than any other type of source due to the eas e of availability of data and their conspicuous presence. Facilities listed in the EPAÂ’s annual T oxic Release Inventory (TRI) represent one specific point source that has been studied extensively in the environmental justice research literature. This database includes industrial manufacturing fa cilities that employ more than ten people and either manufacture or process more than 25,000 lbs of chemicals. TRI sites and emissions data ar e particularly well suited for environmental justice analysis. Facilities that meet certai n threshold requirements must report annually to the EPA the quantities of more than 600 toxi c chemicals that they release directly to air, water, or land, and/or tran sport to offsite facilities. This means that a central agency can collect and manage all data nationwide. TR I data is also publicly accessible and fairly straightforward to download and use. Since the 1990s, the TRI has emerged as the most comprehensive data source on industrial toxic emissions in the U.S. (Chakraborty 2004). The environmental justice impacts of TRI facilities and releases have been studied in Florida (Pollock and Vittas 1995), South Carolina (Cutter et al. 1996, Mitchell et al. 1999), California (Pastor et al. 2004), New Je rsey (Mennis and Jordan 2005), Texas (Tiefenbacher and Hagelman 1999), Des Mo ines, Iowa (Chakraborty and Armstrong 1997), Minneapolis, Minnesota (Sheppard et al. 1999), U.S. counties (Daniels and
9 Friedman 1999), U.S. states (Chakraborty 2004), and U.S. metropol itan areas (Ash and Fetter 2004). The scope and resolu tion of the studies vary, bu t TRI facilities remain an important part of environmental justice assessment. An exclusive focus on TRI facilities or ot her large industrial sources, however, is likely to bias the assessment of environmen tal risk burdens within a community, because emissions from automobiles, smaller industries, and various other sour ces that contribute to air pollution are ignored. Smaller point s ource emitters have received less attention from environmental justice researchers. Fitos and Chakrabor ty (2003) examined inequities in the spatial di stribution of dry cleaning facilities in Hillsborough County, Florida. They found that drycleaners are le ss likely to locate inside minority and impoverished neighborhoods but more likely to be found in densely populated areas near these neighborhoods. The author s pointed out that while drycleaner emissions may be less apparent and less regulated than those from more cons picuous industrial facilities, small point source polluters can be dangerous due to their tendency to be located near residential neighborhoods, increasing the risk of exposure. Also, small point sources do not fall under as strict EPA guidelines and m onitoring may be nonexistent. Other studies that analyze non-TRI point sour ces include a study in Orlando Florida that utilized the EPAÂ’s aerometric information retrieval system (AIRS) in order to assess the environmental justice implica tions of childrenÂ’s exposure to air pollution (Chakraborty and Zandbergen 2007). AIRS contains data for automobile mechanics, dry cleaning facilities, hospitals, and manu facturing facilities too small to be regulated as a TRI facility. By including small and large point sources, the researchers provided a more comprehensive profile of the present risks.
10 Point sources, large or small, are easier to analyze because they are stationary. The presence of air pollution from mobile sources, however, is difficult to assess, therefore distance to roadways is commonly us ed as an indication of risk. In a case study of New York City, Jacobson et al. (2005) used proximity to highways as an indicator of exposure, recognizing that prox imity to a highway can be both a benefit and a health hazard. Using advanced simulated pollution mo dels, Kingham et al. (2007) analyzed the environmental justice implications of traffi c related air pollution in Christchurch, New Zealand. They found evidence of distinct socioeconomic inequ ity in the distribution of vehicle emissions, with the areas that had the highest vehicle ownership demonstrating lower air pollution concentrations. Wu and Batterman (2006) assessed the environmental justice implications of the locations of school s near high traffic roadways in the Detroit area. The study found that Black and Hispan ic children were more likely to attend schools near these high traffic roads, which can increase the chances of various health problems such as asthma. As with point s ources, the mobile sources analyzed in the reviewed literature required simplistic assump tions about the nature and environmental fate of emitted air pollutants. From this overview of source-by-source stud ies, it is evident that environmental justice analysis could benefit from a more cu mulative approach that incorporates multiple pollution sources. Recent studies have acknowle dged the need to include multiple types of sources and emissions that can be found in a given study area. Mo rello-Frosch et al. (2001) modeled estimated lifetime cancer risk s derived from TRI sites, small point sources, and automobile emission data released from the EPAÂ’s Cumulative Exposure Project (CEP). According to these author s, small point and mobile sources are
11 problematic to regulate due to their disper sion and diversity, but the lack of guidelines makes their health impacts all the more impor tant to consider. Con cern over the absence of detailed information on multiple sources led several studies to use a new data set released by the EPA that incorporates both m obile and point sources (Pastor et al. 2005, Apelberg et al. 2005, Linder et al. 2008). However, these studies compared 1996 emission data with Census 2000 socio-demogra phic data a temporal mismatch that can be avoided by using more recent information on air pollution. In light of these points, this thesis uses a data set from 1999 that in corporates multiple sources of outdoor air pollution in order to create a pr ofile of resulting adverse heal th risks within the state of Florida. 2.1.2 Measuring the Extent of Exposure to Air Pollution Environmental justice research on air pollution distribution began by measuring the adverse risk from air pollution as a func tion of distance to the pollution source. While many studies disclaimed their findings with th e caveat that proximity to pollution did not necessarily reflect the actual health risks imposed on the population, this proxy for risk was widely used. The earliest studies used proximity to hazardous facilities as an indicator for the risks minority populations encountered, regardless of the ways these pollutants and facilities may adversely imp act public health (e.g., Asch and Seneca 1978, Gianessi et al.1979). These studies asserted that the mere pres ence of these facilities had a detrimental effect on the community econom ically and socially by implying that the neighborhood is not a desirabl e place in which to live. The study of air pollution with regards to its impact on the well being of the surrounding community is different because immediate dispersion of releases makes
12 emissions harder to contain and keep track of. Ambient pollution is not intended to remain at the location in which they are re leased. Nevertheless, Pollock and Vittas (1995) used proximity to facilities as an indicator of risk in their study of toxic release inventory (TRI) sites in Florida. The an alysis explored the environmen tal justice implications of potential pollution exposure to TRI emissi ons, measuring exposure as a non-linear function of distance to the f acility. While the authors di d acknowledge that risk of exposure is not fully explained by distance from the source, no possible avenues for improvement were provided. Several researchers have used GIS-based circular buffers around point sources of pollution to obtain a more valid or consistent spatial definition of the affected area, instead of relying on administ rative units such as census tract or zip code boundaries. Glickman (1994) explored how a GIS can be ut ilized to incorporate ac tual risk into EJ analysis, not just proximity to the source. Pa rt of the analysis uses a proximity-based measurement in the form of buffers placed around each source to create a circle of potential risk and also includes risk-based measurements derived from information from extremely hazardous substances facilities. Perl in et al.Â’s 1999 study also used circular buffers around TRI facilities in block gr oups, studying the Kanawha Valley in West Virginia, the Baton Rouge to New Orleans co rridor in Louisiana, and the Baltimore, Maryland metropolitan area. Buffers of radii 0.5, 1.0, 1.5, and 2.0 miles were placed around the sources in order to assess the im pacts these arbitrary distances had on the results. Although racial/ethnic and economic inequities were found in all these areas, these authors acknowledged that proximity to TRI facilities does not necessarily mean exposure to pollution is the result.
13 Keeping with the trend of moving towards the buffer analysis method in order to accurately predict the location of impacted areas and populations, Sheppard et al. (1999) experimented with buffer sizes in order to assess the inequity associated with TRI locations in Minneapolis and to ascertain the statistical outcomes of different buffer distances on the result s. Buffers in all these studies ar e used to determine the exposed area around a facility, moving away from munici pal boundaries such as census tracts that may not accurately reflect the range of the air pollution plume. Chakraborty and Armstrong (1997) focus on the environmental fate and transport of pollution after its release a nd translating that informati on to spatially define areas potentially exposed to adverse health risks. These authors a sserted the circular buffers around air pollution emitters used in other studi es may be inadequate because chemical dispersion plumes are not released in a pe rfect circle, but are affected by local wind patterns. This is important because while some people may be in closer proximity to the source, it does not mean they are in more danger when circumstances of release such as wind speed and direction is taken into accoun t. Also, buffer sizes are based on arbitrary distances that may not reflect the nature and quantity of substances emitted (Chakraborty and Armstrong 2004). Research concerning the most effective wa y to approximate areas exposed to air emissions contributed to exposure assessm ent by moving beyond mere proximity based analysis towards more sophisticated methods that incorporate the behavior of air pollutants and their known health effects. E nvironmental justice researchers have begun to incorporate more accurate estimates of h ealth risks that accompany exposure. In 2001, Morello-Frosch et al. confronted this ch allenge of making the connection between the
14 presence of a polluting facility and the resul ting health risks in S outhern California. Air toxics concentration estimates were gathered from the U.S. EPAÂ’s Cumulative Exposure Project, which estimates the annual averag e outdoor concentrati ons of a class of pollutants. The researchers then analyzed lif etime cancer risks derive d from data obtained from the U.S. EPA and the California Air Pollution Control Officers Association. This article makes an important step towards understanding why the pr esence of pollution alone is not enough to assess risk and health ri sk estimates is a better measure of potential inequities. Apelberg et al. (2005) anal yzed health risks derived from EPA estimates at the census tract level in Maryland. The risk stud ied was lifetime cancer risk for 29 hazardous air pollutants. This database incorporated fate and transport of the chemicals, the resulting ambient exposure, and the associated he alth risks for each census tract. Pastor et al. (2005) utilized the same database for a study on California and Linder et al. (2008) worked with an updated and expanded version of the data to asse ss cancer risks in the Houston area. These three studie s have utilized the most co mplete profile of the actual health risks associated with air toxics fo r environmental justice analysis. While these studies incorporated health risk estimates, they did not examine the environmental justice implications of health risks fr om different emissions sources. This thesis project uses the most recent version of the database used by Apelberg et al. (2005) and Pastor et al. (2005) and the one used by Linder et al. (2008). As described further in Chapter 3, this advanced database of pollution emissions endeavors to improve upon past methodologies and provide a more precise portrai t of adverse health risk from air pollution by incorporating lo cal meteorology, chemical dispersion patterns,
15 and circumstances of release. This informa tion will be used to derive human exposure concentrations and estimated cancer risks fo r four different sources of air toxics. 2.2 Statistical Methods Once the source of environmental pollution and the method for approximating exposure to the risk is selected in a quantita tive environmental just ice study, a statistical method can be used to analyze the collected data and to draw conclusions about the inequities in the demographic and socio economic distribution of exposure. These methods can range from straightforward descri ptive and correlation statistics, to multiple regression models, and newly emerging sta tistical methods that account for spatial processes and effects. Correlation analysis can be used to explore the linear relationship between individual variables, such as the associati on between the percentage of Black residents and the quantity of toxic air releases. Co rrelation can provide evidence of positive or negative associations between two variables, which can be useful in order to seek evidence to support an inequity hypothesis. Bi variate correlation tests have been used by Bowen et al. (1995) to analyze TRI locations and emissions in Ohio, by Cutter et al. (1996) to evaluate the location of TRI facilities, toxic stor age and disposal sites, and inactive hazardous waste sites in South Ca rolina, and by Tiefenbacher and Hagelman (1999) to assess acute and chroni c toxic air releases in urban counties of Texas. However, in situations where variables may overlap with each other, bivariat e correlation statistics do not account for how the effect of an inde pendent variable (e.g., minority percentage and poverty rate) may change in the presence of another. These corr elation tests can be
16 used for initial exploration of data, but is prim arily used today in variable identification in order to improve subseque nt statistical analysis. Multivariate regression methods expre ss how the relationship between the dependent variable and an explanatory variab le changes in the presence of several other explanatory variables in the same model. Environmental justice research encompasses social, economic, and demographic explanator y variables that are often related to each other and difficult to isolate. Prior studies ha ve used the multiple regression approach to analyze the equity implications of hazardous waste sites (e.g., Boer et al. 1997), TRI sites (e.g., Bowen et al. 1995, Daniels and Frie dman 1999, Mennis 2002, Pastor et al. 2004) Superfund sites (e.g., Bevc et al. 2005) and tr eatment, storage, and disposal facilities (e.g., Mennis 2002, Zahran et al. 2008). Multivaria te regression analysis also allows for the creation of multiple combinations of expl anatory variables that can provide additional depth to the exploration of th e relationship between the nature or extent of pollution and relevant explanatory factors. A problem arises, however, when conven tional linear regression is applied to spatial data. Standard regression produces a global model where a single equation is provided to explain all variation in the relationships between the dependent and explanatory variables. The processes being exam ined are thus assumed to be constant or stationary over space. In th e context of environmental ju stice analysis, there are two reasons to consider spatial va riability in the relationships between variables (Mennis and Jordan 2005). First, differences may occur due to misspecification of the model because of missing variables such as metropolitan designation. However, the attempt to account for such variation can lead to oversimp lification of regional differences and the
17 assumption that it is possible to categori ze complex local characteristics. Second, there may be nothing wrong with the model; the vari ation can be attributed to the unique characteristics of different locations. There has been a very recent trend toward s the application of spatially oriented statistical methods. Geographi cally weighted regression (GWR) is a local spatial statistical technique used to analyze spatial nonstationarity, or when the measurement of relationships among variables differs from location to location (F otheringham et al. 2002). The analytical utility of GWR has been demonstrated in geographic studies of disaster declarations (Schmi dlein et al. 2008), poverty in urban areas (Longley and Tobon 2004), patterns of violent crime (Cahill and Mulligan 2007), and local politics and voting (Calvo and Escolar 2003). These studies utiliz ed GWR because of the inherent spatial nature of the data and the re lationships being analyzed. It is important to consider that enviro nmental justice is an explicitly spatial problem, concerned with the geographic dist ribution of exposure to pollution and its relationship to explanatory va riables that are rarely dist ributed uniformly across any study area. However, only one published study has used GWR to examine the environmental inequity hypothesis (Mennis and Jordan 2005). These authors applied GWR in order to assess the relationship between the location of TRI facilities and race/ethnicity and other socioeconomic va riables in the state of New Jersey. GWR allowed Mennis and Jordan to map model para meters that represented the statistical relationship between the dependent and releva nt independent variables. The resulting maps showed significant racial/ethnic inequity in the distribution of TRI facilities in urban areas of the state, while simultaneously finding a lack of inequity for Blacks in
18 Newark and evidence of disparity for Hispan ics in rural Vineland. The results provided different explanations for the presence of thes e facilities in each part of the state. This study was limited by its use of a single pollu tion source (TRI sites) and proximity as a surrogate for exposure and advers e health risk-limita tions that will be addressed in this proposed thesis. 2.3 Summary This literature review has traced the methodological history of quantitative environmental justice research in order to elucidate three problems encountered in the search for answers to questions of environmenta l inequity. First, ther e has been a lack of cumulative source assessment, with many studies focusing on human exposure to only one pollution source, despite th e potential oversight of equa lly dangerous sources. Past empirical studies have focused primarily on in equities associated with major stationary sources of pollution, thus ignoring mobile emi ssion sources and smaller emitters that also pollute the local environment. The failure to consider emissions from automobiles, smaller industries, and other less conspicuous sources that contribute to air pollution is likely to distort the assessment of inequities in the distri bution of environmental risk burdens. Several recent environmental justice studies have emphasized the need for going beyond locational inequities and a focus on sta tionary emission sources towards a more cumulative exposure approach th at considers the health risk s that a community may face from various types of pollutants and emissi ons sources. A risk m odeling approach that considers multiple sources of pollution is also consistent with the emerging policy focus on cumulative exposure assessment. Second, estimating exposure to air pollution has been problematic due to the difficulty in obt aining the abundance of data necessary to
19 model emissions and the resul ting health risks. A large num ber of studies have used proximity to hazardous facilities or pollution sources as a proxy for potential human exposure, instead of assessing the nature and quantity of pollutants emitted, local meteorological conditions, and other factors that influence exposure. While specific studies focused on modeling exposure to toxi c pollution, few have attempted to examine whether unequal exposure patterns lead to di sproportionate health risks among minority and low-income communities. Finally, comm only used statistical methods produce a single model to explain all variation in the da ta thus ignoring the f act that statistical associations between environmental health ri sks and explanatory f actors can vary from place to place within a study ar ea. Geographically weighted regression, a technique that examines local variation in model paramete r estimates, has rarely been utilized in environmental justice studies to investig ate the spatial non-sta tionarity of model parameters and understand local and regional differences. These three gaps in environmental jus tice methodology present an opportunity for further research and improvement. The follo wing chapter outlines the data sources and methodology used in a case study that focuse s on evaluating the geographic distribution of estimated health risks from air pollution in the state of Florida.
20 CHAPTER 3: STUDY AREA AND METHODOLOGY This section describes th e study area, data sources used, and the methodology followed in order to assess th e equitable distribution of va rious known sources of cancer risk in Florida. First, the study area is introduced and the source of the data and the process used to derive the ke y variables are outlined. Next, th e variables used in the case study are defined and described, along with their data sources Finally, the methods that were chosen to address the limitations of previous studies are explained. 3.1 Study Area The geographic scope of this analysis includes the entire state of Florida. Florida provides a suitable setting for a study concerni ng the environmental justice implications of air pollution, in part, because of its rapidl y growing population. It is the fourth largest state in the U.S. in terms of total population and one of the fastest growing states in the nation. Since 1970, nine million people have moved to Florida and an average of 800 individuals relocate to th e state daily (Chapin 2006). This state experienced a 23.5 percent increase in population from 1990 to 2000 accompanied by swift economic growth in the past few decades that has led to unch ecked commercial, residential, industrial, and infrastructure development in some cases. As Florida continues its growth, it is important to ensure that growth for some reside nts of Florida does not mean adverse and disproportionate health risks for others. Florida is also a diverse state in terms of race/ethnicity, with the thr ee largest minority groups, Hisp anic, non-Hispanic Black, and Asian collectively comprising over a third of the entire population. As the populations of
21 these minority groups continue to grow, the urgency to assess the environmental justice implications of the health risks from air pollution becomes even more pronounced. Systematic quantitative research on the environmental justice implications of exposure to air pollution has been limited at the state level. Past environmental justice studies in the state of Florida include Pollock and Vittas (1995), who focused on analyzing distance to TRI facilities in the state, based on demographic and socioeconomic characteristics of census bl ock groups. Racial and ethnic minority subpopulations, particularly African American hous eholds, were found to reside closer to polluting sources. Stretesky and Hogan (1998) examined the spatial relationship between Superfund sites and socio-demographic charact eristics of census tr acts surrounding these sites in a longitudinal study of Florida. They found that Blacks and Hispanics were more likely to live near these sites and this a ssociation increased with time. While these previous studies have investig ated inequities associated with major point sources, more research is necessary to analyze exposure to other types of pollution sources in Florida. By utilizing the 1999 NATA, this thesis will a ttempt to improve the information available by including estimated health risks and in corporating several different sources of emissions. 3.2 Data Sources and Variables The Clean Air Act of 1990 separated air po llutants into two di stinct categories, criteria air pollutants and air toxics. Criteri a air pollutants are a na rrow classification of common pollutants including ozone, sulfur dioxide, carbon monoxide, nitrogen dioxide, particulate matter, and lead. The EPA has set cr iteria, or threshold levels, at which these substances become potentially harmful to publ ic health. Air toxics, however, do not have
22 set levels at which the EPA deem them dange rous; instead, any exposure to air toxics is considered unsafe. Air toxics are a group of 188 air pollutants that are known to cause or are suspected of causing cancer and other serious health prob lems such as respiratory, reproductive, and neurological damage. Sensit ive populations such as the elderly and children are particularly vulnerable to health risks caused by air toxics. Air toxics include metals such as cadmium, mercury, and chromium, as well as chemicals such as asbestos, benzene, and dioxin. While the environmental justice implications of criteria air poll utants have been extensively studied (Wernette and Nieves 1992; Jerrett et al. 2001; Kingham et al. 2007), air toxics have received less attentio n. Beginning in 2002, the EPA developed a nationwide database describing th e release of and health risks associated with air toxics. The first release of the National-Scale Air Toxics Assessment (NATA) profiled emissions from 1996. This assessment included 33 air pollutants identified as the most dangerous to public health in the larges t urban areas (EPA 2008b). Because the EPA updates its emission databases every three year s, the next assessment featured 1999 data. The 1999 NATA increased the number of air toxi cs included in the da tabase to 177, with health risk estimates for 133 of these polluta nts. The NATA estimates health risks at the census tract level providing a geographically de tailed resolution that can be matched with population and housing data from Census 2000. The 1999 NATA utilizes a four-step process in order to obtain and calculate health risk estimates at the census tract leve l. First, the EPA gathers emission data and estimations directly from the sources of po llution as well as state and tribal agencies responsible for tracking outdoor emissions For some sources, the EPA estimates
23 emissions using models and measurements. This information is compiled into the National Emissions Inventory (NEI). The NEI draws its data from four sources: major point sources, such as industr ial facilities and waste inci nerators, minor point sources, including dry cleaning faciliti es and auto-body shops, mobile on-road sources like cars and trucks, and mobile off-road sources, in cluding boats and all-te rrain vehicles. Next, ambient air concentrations ar e estimated using a Gaussian dispersion equation known as the Assessment System for Population Expos ure Nationwide (ASPEN) model. Using the 1999 NEI as the input data, the ASPEN computer model estimates the concentrations in every census tract in the U.S. by factoring in lo cation of release, height of release, local meteorological conditions, and the behavior of the particular chemical once it is released, including the conversion of one chemical into another and the settling out of chemicals in the atmosphere. Background concentrations, or levels of pollutants th at preexist in the environment either due to natural formation or lingering presence from previous release, are also included. Third, inhalation exposur e is estimated utilizing the Hazardous Air Pollutant Exposure Model, Version 5 (HAP EM5). This model incorporates ASPEN ambient air pollution estimates, average i ndoor/outdoor activity patt erns, climate data, and census data to derive estimated inhalation exposure. Finally, the EPA estimates specific types of health risks associated with inhalation of air toxics according to current data on health effects, EPA risk assessme nt characterization, and exposure estimates (EPA 2008c). 3.2.1 Dependent Variables The result of the EPAÂ’s four-step proce ss is a database containing tract-level estimates for three types of public health risks: cancer risk, respiratory risk, and
24 neurological risk. Each risk is further divided by emissi on source category: major point sources, minor point sources, on-road mobile sources, off-road mobile sources, and background concentrations. The NATA also pr ovides cumulative exposure estimates for each census tract that collectivel y includes all sources of toxics. This analysis will be focusing on estimated lifetime cancer risks. Cancer risks are expressed as individual lifetime excess risk from a lifetime of exposure to one unit of a pollutant. The NATA estimates cancer risk on the basis of the inhalation unit risk (IUR) factor, a measur e of the cancer-causing potential of each air toxic. Lifetime cancer risk for each pollutant in each census tract is calculated by: R ij C ij I URj Where R ij = estimate of individual lifetime cancer risk from air pollutant j in census tract i Cij = Concentration in micrograms of pollutant per cubic meter of air of air pollutant j in census tract i I UR= inhalation unit risk estimate for pollutant j. Cancer risks are assumed to be additive a nd lifetime cancer risk from all air toxics present in a tract are summed to obtain the total estimated lifetime cancer risk for the tract. The estimated lifetime cancer risk variab le is expressed as N in one million people, or the number of people expected out of one million to contract cance r if exposed to the 1999 concentrations estimated for that census tract for 24 hours a day over seventy years, which is the EPAÂ’s estimate of an average li fe span. These estimates are in addition to cancer cases unrelated to air toxic ex posure, not including them (EPA 2008d).
25 In order to explore spatial inequities in cancer risk from outdoor air pollution in detail, this thesis analyzes cumulative lifetime cancer risk from all sources in Florida, as well as risk from the four following indi vidual sources of ai r toxic emissions: Major point sources: These are defined by the Clean Air Act as facilities that have the potential to emit 10 tons of one air toxic or 25 tons of a combina tion of air toxics. Other point sources: These ar e also known as area/other point sources and include facilities below the Â‘majorÂ’ threshold levels such as auto body shops and dry-cleaning facilities and also include s ources such as wildfires. On-road mobile sources include motorized vehicles that normally operate on public roadways and comprises passenger cars, motorc ycles, minivans, sport-utility vehicles, light-duty trucks, heavyduty trucks, and buses. Non-road mobile sources are mobile sour ces not found on roads and highways, such as trains, airplanes, lawn mowers, constr uction vehicles, and farm machinery (EPA 2008e). 3.2.2 Explanatory Variables All explanatory variables for this stu dy include well-documented variables from previous environmental justice studies extr acted from Census 2000 data at the census tract level. These are defined in Table 3.1, along with their ex pected statistical relationship with the dependent variables. Th e explanatory variables can be divided into three categories: racial/ethni c variables, socio-demographi c variables, and density. In light of the primary concern of race and ethni city in environmental justice research, the three largest racial/ethnic minority groups in the state of Florida will be included. Persons self-identifying as non-Hispanic Black, Hispan ic/Latino of any race, and Asian together
26 make up about 37 percent of the population in Florida (U.S. Census 2006) and their inclusion in this study is cr ucial to exploring equity. R acial and ethnic minorities are generally expected to have in creased exposure to health ri sks when compared to their white counterparts due to disc rimination and institutional racism in the housing market (Pulido 2000, Pastor et al. 2001). Urban areas have a legacy of residential segregation in the U.S. and Florida that invariably contribute to the current loca tion pattern of the Black population at the metropolitan scale. Lingering racial disc rimination in the housing market and employment opportunities, along with various educational and economic disparities, can lead to limited housing choices and restrict their locations to areas disproportionately exposed to adverse health risks from t echnological hazards (Stretesky and Hogan 1998, Pastor et al. 2001). While Hispan ic populations do not have as deeply ingrained legacies of urban segregation, these communities can experience similar outcomes and exposure to toxic pollution (Pulido 2001). Particularly in Florida, migrant Hispanic workers may be at increased risk of hous ing discrimination, possibly cont ributing to increased health risks (Pastor et al. 2005 ). While Asians have been included in the minority category in previous environmental justice research, fe w studies have separately analyzed the relationship between Asian populations and e nvironmental health risks (Pastor et al. 2001). This study assumes similar racial disc riminatory practices are at work and examines whether Asian individual are di sproportionately exposed to environmental health risks. Several socioeconomic variables will be included to provide other avenues through which inequity can occur and to ensu re that any unjust di stribution found among
27 minorities is not due to differences in econom ic status across the various racial/ethnic groups. The proportion of residents below the federal poverty level indicates financial resources, which can become a limiting factor not only with regards to the areas available in which to live, but the ability to move out of tracts that may become more risky over time. The proportion of homes owned by their occupants is another economic indicator that gives an impression of immediate financ ial resources within a tract and residentsÂ’ wealth and power. Some researchers use hom e ownership as an indicator of political participation (Pastor et al. 2005), believing that home ownership leads to increased investment in local decision-making proce sses. The environmental justice literature indicates that tracts with re latively lower levels of home ownership and higher poverty rates will be expected to have a higher risk of exposure to the adverse health effects caused by air toxics (Boer et al. 1997, MorelloFrosch et al. 2001, Pa stor et al. 2005). Population density is a commonly used cont rol variable in en vironmental justice studies because densely populated areas are more likely to co ntain polluting facilities and activities. While population de nsity is commonly measured as the number of people per square mile, the natural logari thm of this value will be take n in order to account for the diminishing effect of higher numbers, as s uggested by Mennis (2002) and Pastor et al. (2005). The explanatory variables (proportio n non-Hispanic Black, proportion Black, proportion Asian, proportion below poverty, proportion owner occupied, and population density) have been obtained from the 2000 Census at the tract level, matching up with the resolution and time frame of th e 1999 NATA dependent variables.
28 Table 3.1 Definitions of Explanatory Variables and Expected Statistical Association with Cancer Risk. Variable Definition Expected Statistical Relationship Proportion NonHispanic Black Ethnically non-Hispanic individuals identifying themselves as Black expressed as a proportion of the census tractÂ’s total population. Positive Proportion Hispanic/Latino Individuals identifying themselves as Hispanic/Latino (of any race) expressed as a proportion of the census tractÂ’s total population. Positive Proportion Asian Ethnically non-Hispanic individuals identifying themselves as Asian expressed as a proportion of the census tractÂ’s total population. Positive Proportion Below Poverty Proportion of the population within a census tract with an annual family income below the federal poverty level. Positive Proportion Owner Occupied Proportion of occupied housing units that are owner-occupied within a census tract. Negative Population Density Number of peopl e per square mile in a census tract. Positive 3.3 Methods In order to assess the dist ribution of estimated cancer risks from air toxics in Florida, this thesis project impl ements three stages of statisti cal analysis. First, the linear association between the dependent and each explanatory variable is measured using PearsonÂ’s correlation coefficient at the census tract level. Bivariate correlations provide an initial indication of the relationships that exist between the health risks and the various racial/ethnic, and socioeconomic variables ut ilized in this study. Conventional multiple regression is used to develop five sets of models, one for each source category of estimated cancer risk and one for cumulative cancer risk from all sources, based on the ordinary least squares method. These models express the relations hip all independent
29 variables possess with each source of cancer ri sk simultaneously, at th e census tract level. The multivariate regression models can be summarized by the following equation: y 0 1 x 1 2 x 2 ... k x k Where y is the dependent variable, x1, x2,Â… xk are the independent variables, and 0, 1, 2,Â… k are the model parameters. The model parameters indicate the nature and strength of the association between the particular independent variable and the de pendent variable, negative or positive, when all other variables are also taken into account. Multivariate regression allows the influence each independent variable exerts ov er the dependent variab le to be examined, after controlling for the effects of the other independent variable. While multivariate regression has been used extensively in environmental justice studies, this methodology does not account for geographic variations in signs and coefficients of model parameters with a gi ven study area, as described previously. In order to detect trends within the state of Florida, the second phase of analysis uses geographically weighted regression (GWR), a local spatial statistical technique used to analyze spatial nonstationarity (Fotheringham et al. 2002). Instead of calibrating a single or global regression equation, GWR generates a separate regression equation for each observation. Each equation is calibrated usi ng a different weighti ng of the observations contained in the data set. Each GWR equation may be expressed as yi 0 ( u i v i ) 1 ( u i v i ) x i 1 2 ( u i v i ) x i 2 ... k ( u i v i ) x i k i Where (ui, vi) is the location of i. The assumption is that near by observations have greater influences on one anotherÂ’s parameter esti mates than observations farther apart. The weight assigned to each observation is based on a distance decay function that is centered
30 on observation i. In the case of areal data (e.g., ce nsus tracts), the distance between observations is calculated as the distance between polygon centroids. The distance decay function, which may take multiple forms, is modified by a bandwidth setting at which distance the weight approaches zero. GWR software version 3.0.18 provides the user with three options to assist with calibration of the model to achieve optimal results. Generally, all observations are weighed by a function that decreas es with distance from point i. This weighted window is referred to as a kernel. Wij exp[ (dij/b)2 Where W ij is the weight of data point j at regression point i, dij is the Euclidean distance between i and j, and b is the bandwidth. The farther away data point j is from regression point i, the lower its influence will be. Choosi ng the appropriate bandw idths is essential to producing a dataset that will not oversimp lify regional differences by being too large, or be too detailed to provide a clear pattern by being too small. The bandwidth can be manually set by the user after experimenti ng with previous models, or it can be automatically selected by cross-validati on or minimizing the Akaike Information Criterion (AIC), which is the method used in this study (Fotheringham et al. 2002). AIC 2 n loge( Âˆ ) n loge(2) n n tr ( S ) n 2 tr ( S ) Where n = the sample size, Âˆ the estimated standard deviation of the error term, and t r ( S ) the trace of the hat matrix which is a function of the bandwidth. The lower the AIC, the better the fit of the model. The AI C method has the advantage of considering the
31 fact that the degrees of freedom may vary among regression models centered on different observations. In the application of GWR, the user may choose a fixed bandwidth that is used for every observation (e.g., Mennis and Jordan 2005) or a variable bandwidth that expands in areas of sparse observations and shrinks in areas of dense observa tions (e.g., Calvo and Escolar 2003). This analysis uses a variable kernel bandwidth that adapts for the density of data in each location due to the va riability in the sizes of census tracts. Because the regression equation is calibra ted independently for each analytical unit, a separate parameter estimate, t-valu e, and goodness-of-fit is calculated for each census tract in Florida. These values are mapped using GIS softwa re, thus providing a way to visually interpret the geographic dist ribution of the nature and strength of the relationships between explanat ory and dependent variables. Indeed, in many cases the GWR output would be inscrutable without the ab ility to map the result s. Fotheringham et al. (2002) describe several different methods for displa ying the data, including point symbols, area symbols, contour plots, and pseudo3D displays. This study utilizes choropleth mapping techniques to clearly display significance of local regression coefficients and is also suitable for comp aring descriptive chor opleth maps of the dependent variables. These maps provide the ba sis for local and region al analysis of the relationship between cancer risks from vari ous sources of air toxics and various explanatory variables, as well as compar isons between the analyzed results of conventional regression and GWR.
32 CHAPTER 4: CONVENTIONAL ENVIRO NMENTAL JUSTICE ANALYSIS OF ESTIMATED CANCER RISKS This chapter focuses on the use of traditional statistical techniques to explore the environmental justice implicati ons of estimated cancer risk from inhalation exposure to air toxics in Florida. Firs t, descriptive choropleth mappi ng and summary statistics are used to explore the spatial distribution of th e dependent and explanatory variables at the census tract level. Second, bi variate correlations are derive d to examine the statistical association between each explan atory variable and the five dependent variables: lifetime estimated cancer risk from all sources and the four known sources of ambient air toxics: major point sources, other point sources, on -road mobile sources, and non-road mobile sources. Finally, conventional multivariate regression analysis is employed to estimate each dependent variable as a function of the racial/ethnic and socioeconomic variables in a single model. By analyzing cumulative cancer risk and risks from four different sources of air toxics in Florida, it is possible to assess how sta tistical associations between relevant explanatory variables and magnitude of cancer risk varies by emission source. 4.1 Descriptive Choropleth Mapping and Statistics The five dependent variables are displaye d as classified choropleth maps in Figures 4.1 to 4.5. Census tracts in Flor ida are grouped into fo ur quartiles based on estimated values of lifetime cancer risk. Lifetim e cancer risk from all sources of air toxics is concentrated primarily in the largest metropo litan areas of Florida (Figure 4.1) such as Jacksonville, Orlando, Tampa, and Miami. B ecause most pollution-generating activities
33 take place in urban areas, this pattern is not surprising. However, when cancer risks are separated by release source, different patterns emerge. Can cer risk from major sources alone (Figure 4.2) is concentrat ed in both urban centers such as Jacksonville and in rural areas. Lifetime cancer risk from minor point so urces (Figure 4.3) is comparably lower in the central southern portion of the state. Life time estimated cancer risk from both on-road mobile sources (Figure 4. 4) and from non-ro ad mobile sources (Figure 4.5) are highly concentrated in the metropolitan areas of Florida, reflecting areas of dense vehicular traffic and other modes of transportati on such as airports and railroads. Figure 4.1 Estimated Lifetime Cancer Risk Fr om All Sources by Census Tract, 1999
34Figure 4.2 Estimated Lifetime Cancer Risk From Major Point Sources by Census Tract, 1999 Figure 4.3 Estimated Lifetime Cancer Risk s From Minor Point Sources, 1999
35Figure 4.4 Estimated Lifetime Cancer Risk From On-Road Mobile Sources, 1999 Figure 4.5 Estimated Lifetime Cancer Risk From Non-Road Mobile Sources, 1999
36 The geographic distributions of the six e xplanatory variables are summarized in Figures 4.6 to 4.11. These choropleth maps cl assify each variable into four categories based on the quartile method, and suggest subs tantial spatial variation for each variable. The Black population (Figure 4.6) is concentrated mostly in FloridaÂ’s panhandle region and major metropolitan areas. Additionally, some rural tracts near Lake Okeechobee show comparatively higher pr oportions of this variable. Th e Hispanic population (Figure 4.7) is concentrated in the southern portion of the state, both in urban and rural tracts. Higher proportions of Asian populations ex ist in and around major cities like Tampa, Orlando, and Jacksonville (Figure 4.8). Figure 4.6 Proportion Non-Hispanic Bl ack by Census Tract, 2000
37Figure 4.7 Proportion Hispanic by Census Tract, 2000 Figure 4.8 Proportion Asian by Census Tract, 2000
38Figure 4.9 Homeowner Occupied Housing Units by Census Tracts, 2000 Figure 4.10 Proportion of Population Below th e Federal Poverty Rate, 2000
39Figure 4.11 Population Density by Census Tract, 2000 Places with comparatively higher rates of home ownership are mostly in rural and suburban areas of Florida (Figure 4.9). Hi gher proportions of persons living below the poverty line can be observed in rural tracts as well as major urban centers (Figure 4.10). As expected, areas of high populat ion density coincide with th e largest cities in Florida (Figure 4.11). Summary statistics for all va riables used in this study are provided in Table 4.1, with frequency histograms included in Appe ndix A. The mean values of the dependent variables suggest that major po int sources of air toxics pose the lowest cancer risk, while on-road mobile sources are responsible for th e highest. Most of the explanatory variables suggest substantial variability in their values across census tr acts in Florida. The average proportion of Blacks and Hispanics approximate ly equal 0.15, but these values range from 0 to 0.99 and 0.95, respectively. The Asian proportion has a much lower mean of only 0.02 and a considerably smaller range The mean proportion of owner occupied
40 housing units (home ownership rate) is 0.63, while proportion below poverty has a mean of 0.13 among census tracts in the state of Flor ida. Population density appears to have the most skewed distribution, with a mean value of 3,223 people per square mile that is lower than its standard deviation and almost te n times smaller than its maximum value. Table 4.1 Descriptive Statistics for Variables Analyzed N Mean Min Max Std Dev. Dependent Variables (estimated lifetime cancer risk in persons per million): All Sources 3,154 37.295 0.000 108.133 11.998 Major Point Sources 3,154 0.463 0.000 32.397 1.335 Other Point Sources 3,154 5.818 0.000 54.77 3.728 On-Road Mobile Sources 3,154 9.295 0.000 81.995 5.628 Non-Road Mobile Sources 3,154 2.083 0.000 52.503 1.862 Independent Variables: Proportion non-Hispanic Black 3,151 0.154 0.000 0.990 0.227 Proportion Hispanic 3,151 0.146 0.000 0.954 0.196 Proportion Asian 3,151 0.016 0.000 0.202 0.016 Proportion Owner Occupied 3,151 0.628 0.000 1.000 0.236 Proportion Below Poverty 3,151 0.130 0.000 0.768 0.105 Population Density 3,151 3,223 0.000 34,289 3,500 4.2 Bivariate Correlation Analysis Bivariate parametric correla tions are used to assess the linear relationship between each explanatory variable and the five depende nt variables, at the census tract level. PearsonÂ’s correlation coefficients ( r -values), presented in Table 4.2, indicate the strength and significance of each indivi dual exploratory variableÂ’s statistical association with estimated lifetime cancer risk, and do not accoun t for the presence of other variables. The
41 table indicates that cumulative cancer risk is significantly and negatively correlated with the proportion of owner-occupied homes and positively correlated with all the other independent variables, which is consistent with the theoretical expectations of inequity outline in Chapter 3. Table 4.2 PearsonÂ’s Correlation Coefficients of Lifetime Cancer Risk by Source and Explanatory Variables (N=3,154). All Sources Major Point Sources Other Point Sources On Road Mobile Sources Non-Road Mobile Sources Proportion nonHispanic Black 0.253** 0.054** 0.280** 0.232** 0.089** Proportion Hispanic 0.467** 0.084** 0.240** 0.401** 0.342** Proportion Asian 0.195** -0.028 0.087** 0.148** 0.072** Proportion Owner Occupied -0.251** -0.042* -0.186** -0.257** -0.231** Proportion Below Poverty 0.242** 0.096** 0.278** 0.245** 0.143** Population Density (natural log) 0.618** -0.012 0.276** 0.547** 0.370** *p<0.05; **p<0.01 The strength and direction of correlations differ when individual sources of estimated cancer risk are analyzed. Cancer ri sk from major point sour ces is positively and significantly correlated with the proportion of Black and Hispan ic residents, as well as with poverty rate, which corresponds with the exp ected behavior of this variable in light of the environmental justice framework. However, the Asian proportion and population density are not considered significantly corr elated. Home ownershi p rate behaves as expected and is negatively and significantly correlated with estimated cancer risks from major point sources. Statistical associations between lifetime cancer risk from other point sources and the proportion of Blacks, Hispan ics, Asians, poverty rate, and population
42 density remain positive matching the findings for cancer risk from all point sources. Additionally, home ownership rate remains negatively and significantly correlated with estimated cancer risk from other point s ources. For both on-road and non-road mobile sources, home ownership rate demonstrates a significant and negative relationship, while all other variables indicate a significant a nd positive correlation with estimated cancer risk. In summary, the explanatory variables disp lay the expected sta tistical correlations not only with cumulative cancer risk from all sources of air toxics, but also with those from both types of mobile sources and minor poi nt sources. Major point sources differ in that the proportion of Asians and populat ion density do not exhi bit a signification relationship with cancer risk. In the next se ction, multivariate regression is used to explore the relationship between each of the five dependent variables and the combined presence of all six explanatory variables. 4.3 Traditional Multiple Regression Analysis In order to assess the simultaneous effects of all explanatory variables on each of the five dependent variables, conventional multivariate regression based on the ordinary least squares (OLS) method is utilized. These regression results are summarized in Tables 4.3 and 4.4. To check for multicollinearity, the condition index was calculated for each multiple regression model. None of the five models indicated a condition index greater than 30, suggesting the absence of serious collinearity problems. Table 4.3 provides the regression results for cumulative cancer risk from all sources of air toxics. The ANOVA F -test indicates statistical significance for the overall model ( p <.001) and the value of the adjusted multiple R -squared exceeds 50 percent,
43 suggesting a reasonably high goodnessof-fit. After controlling fo r the effects of the other independent variables, population density and all racial and ethnic variables are significantly and positively associated with cancer risk, while home ownership and poverty rates are not significantl y related to cancer risk from all sources of air toxics. These results match the bivariate correlations reported previously, with the exception of the non-significance exhibited by the socioeconomic variables. Table 4.3 Multivariate Regression Results: Cumulative Cancer Risk From All Sources Variables Coefficientt-statistic Proportion non-Hispanic Black 12.5113.13** Proportion Hispanic 21.2625.02** Proportion Asian 82.868.61** Proportion Owner Occupied -1.26-1.30 Proportion Below Poverty -3.15-1.29 *** Population Density (natural log) 3.4131.75** Adjusted R-squared 0.52 F-statistic 566.87** N 3145 *p<0.05; **p<0.01 The multiple regression models associated w ith the different sources of air toxics indicate substantial variation in model perf ormance and significance of model parameters (Table 4.4). Although the ANOVA F -test indicates overall si gnificance for all these OLS models, the adjusted R-squared values are generally smaller compared to the multiple regression for cumulative cancer risk. In presence of the other explanat ory variables, only the Hispanic proportion and poverty rate show an expected and significantly positive effect on cancer risk from major point sources. The only other signifi cant variable is the
44 natural log of population density, which exhibi ts a negative association with cancer risk. All other explanatory va riables are not significant in the regression model for major point sources. Additionally, this model yields the lowest R-squared value, indicating unsatisfactory goodness-of-fit. Table 4.4 Multivariate Regression Results: Cancer Risk from Four Known Sources Major Point Sources Minor Point Sources On-Road Mobile Sources Non-Road Mobile Sources Coeff t-stat. Coeff. t-stat. Coeff. t-stat. Coeff. t-stat. Proportion Black 0.11 0.70 ** 3.64 9.37** 4.44 8.87** 0.56 2.93** Proportion Hispanic 0.57 4.18** 3.73 10.75** 8.12 18.17** 2.47 14.58** Proportion Asian -0.62 -0.40 ** 21.94 5.59** 26.12 5.16** 0.23 0.12 ** Proportion Owner Occupied 0.10 0.66 ** 0.44 1.10 ** -1.27 -2.49* -1.39 -7.17** Proportion Below Poverty 1.01 2.59** 3.85 3.85** 0.63 0.47 ** -1.24 -2.53* Population Density (ln) -0.04 -2.04* 0.36 8.25** 1.44 25.57** 0.30 13.95** Adjusted R2 0.01 0.18 0.40 0.21 F-statistic 7.90** 118.48** 350.20** 141.71** N 3145 3145 3145 3145 *p<0.05; **p<0.01 The regression model for estimated lifetim e cancer risk from minor point sources indicates positive and significant effects for a ll three racial/ ethnic va riables, poverty rate, and population density, as indicated by the biva riate correlations. Th e R-squared value of 0.18 is higher than model for major point sour ces, but is considerably smaller than the model for cumulative cancer risk from all sources. The multiple regression model for onroad mobile sources shows a significant and positive relationship be tween cancer risk and proportion of Black, Hispanic, and Asian, resi dents, as well as population density. Home ownership rate behaves as expected, demonstr ating a significant and negative relationship
45 with cancer risk from on-road mobile sources of air toxics. This m odel yields a relatively high R-squared value of 0.40, indicating a more satisfactory fit than those based on point sources. The multivariate model for non-road mobile sources of cancer risk provides similar results to the one derived from on-road sources, with two exceptions. The proportion of Asians is not significant and proportion below poverty is negatively significant, which is incongruous with the expectations of environmental inequity. In summary, the presence of Black and Hispanic residents and higher population density represent the most consistent and si gnificant indicators of lifetime cancer risk from all emission source categories, after acc ounting for the independe nt effects of other explanatory variables. The so cioeconomic variables suggest mixed results based on the source of air toxics. Cancer ri sk from point source emissions is significantly greater in areas of high poverty rate, while cancer risk fr om mobile sources is greater in areas of low home ownership. It is important to consider, however, that these conventional regression models assume spa tial stationarity in the relationships between estimated cancer risk and each explanatory variable. Give n the spatial variabil ity observed in the choropleth maps for the dependent and indepe ndent variables, it is to necessary to examine if the statistical relationships indicated by traditional regression (i.e., the five global models) remain consistent across the state of Florida. The next chapter thus focuses on the application of geographically weighted regression to analyze spatial variation in the OLS regression mode l performance and model parameters.
46 CHAPTER 5: GEOGRAPHICALLY WEIGH TED REGRESSION ANALYSIS OF ESTIMATED CANCER RISKS This chapter focuses on exploring the statistical relationship between the dependent variables representing cancer risk and explanatory variab les on the basis of geographically weighted regression (GWR), a method that allows a unique local regression model to be produced for every i ndividual census tract in Florida. Again, cumulative cancer risk from all sources is ev aluated, followed by can cer risks associated with the four known sources of ambient ai r toxic emissions. The local estimates of regression parameters and regression model performance provided by GWR are compared to the results of traditional multiple regression analysis that were reported in the previous chapter. In order to determine the appropriate bandwidth for GWR analysis, an adaptive kernel method is used for this study, instea d of a fixed bandwidth setting. Previous studies indicate that choosi ng a fixed bandwidth is difficult when census enumeration units are used, because these vary in size a nd shape according to the population density of the area (Mennis and Jordan 2005; Mennis 2006; Kavanagh et al. 2006). A spatially adaptive kernel selects an optimal number of neighboring units for the analysis, relying upon contiguity rather than distance to s eek a specific number of nearest neighbors to ensure a constant size of local samples (Zuhuang 2006). Where census tract centroids are farther apart (e.g., rural areas), the bandwidth becomes larger, and when tracts are closer together (e.g., urban areas), the bandwidth becomes smaller. The optimal bandwidth for
47 this study is determined through an iterative process to minimize the Akaike information criterion (AIC), following previous applica tions of GWR (e.g., L ongley 2004; Cahill and Mulligan 2007; Ali et al. 2007). In this method, the software calculates the AIC for several different local sample sizes, finally selecting the minimum value. The AIC is considered to be the most appropriate me thod for implementing the adaptive kernel method due to its ability to take goodness of fit and degrees of freedom into account (Fotheringham et al. 2002). In order to maintain a consistent spatia l definition of neares t neighbors for census tracts near the boundary of the study area, tracts from the neighboring states of Alabama and Georgia that are sufficiently close to Fl orida are also included in the GWR analyses. The data set used to estimate the GWR models for this study thus encompasses all tracts in Florida (n=3,151) and a dditional tracts from neighbor ing states (n=359) whose centroid falls within a 50-mile radius of the centroid of a Florida tract located at the state border. While the GWR analyses utilize 3,510 total observations, the results for only the census tracts within Florida are presented he re to allow comparison with the respective global regression models. The output for GWR includes model parameter estimates and t-values for each explanatory variable, along w ith local R-squared values to indicate model performance. While this is the same output provided by a global regression model, GWR produces this set of diagnostics for each data point, or in this case, a census tract. This makes it possible to map the regression results and visualize how relationships vary across space, rather than relying upon one set of parameters to express the relationship for an entire study area. In this chapter, the GW R results associated with each dependent variable (cancer
48 risk by source) are organized and summarized in the form of a table, a map depicting the spatial variability in model performance, and a series of maps that show the distribution of relevant regression model parameters across Florida. 5.1 Cumulative Cancer Ri sk from All Sources The numerical results associated with the GWR model for cancer risk from all sources are summarized in Table 5.1. This ta ble allows the global regression parameters presented in the previous chapter to be comp ared with the range of results produced by GWR. For example, traditional multiple regression indicated that cancer risk from all sources is positively and significantly asso ciated with proportion Black among census tracts in Florida. While the global regression model estimated the parameter at 12.51 for the entire state (Table 4.3), the GWR model s hows that it ranges from 0.09 to 32.97, with a median of 11.34 (Table 5.1). This variability in the model coefficient suggests that the relationship between cancer risk from all sour ces of air toxics and the proportion of Black residents is not static across the state.
49 Table 5.1 GWR Results for Cumulative Cancer Risk from All Sources Global GWR Coefficients Statistical Significance of t-values Across Census Tracts in Florida Regression Coefficient Min Median Max Negative 90% level Not Significant Positive 90% level Proportion Black 12.51** 0.0911.3432.970.00% 7.20% 92.80% Proportion Hispanic 21.26** -28.2316.4668.036.35% 27.20% 66.45% Proportion Asian 82.86** -88.29104.21289.992.05% 29.67% 68.28% Proportion Owner Occupied -1.26 ** -23.99-4.829.0255.56% 44.44% 0.00% Proportion Below Poverty -3.15 ** -63.100.0117.2130.12% 59.88% 10.00% Population Density (Ln) 3.41** 0.000.474.0850.21% 2.19% 47.60% Adjusted R squared 0.52 ** 0.140.470.78 Akaike Information Criterion (AIC) 26186 ** 23795 p<0.05 **p<0.01 For each explanatory variable, the last th ree columns of Table 5.1 categorize the tvalues produced by GWR for each Florida tract by the level of statistical significance (p<.10) and direction of the association. Because of the large sample size (n=3,151), standard t-values of -1.645 and 1.645 are used to represent the 90 percent le vel threshold for negative and positive signifi cance, respectively. In this example, proportion Black is positively and significantly related to cancer risk from all sources in 93 percent of tracts and is not significant in 7 percent of tracts. Although the Hispanic and Asian proporti ons were positively and significantly related to cancer risk from all sources in th e global regression model, the GWR results in Table 5.1 indicate that these variable coefficients include both negative and positive values. For both variables, the statistical a ssociation with cumulative cancer risk is not significant in more than one-fourth of census tracts in Florida. Across a small percentage
50 of tracts, the t-values are negatively significant the opposit e relationship than what was indicated by the global regression model. For proportion below poverty, 60 percent of tracts indicate non-significant t-values, ma tching the global regression model result. However, the global model does not account for the fact that cancer risk has a positively significant relationship with povert y rate in more than ten percent of Florida tracts, which is consistent with the environmental injust ice expectations. GWR model parameters for population density are divided almost equally between posit ive and negative significance, while the global regression model produced a positively significant coefficient for the entire study area. These varia tions in not only the signifi cance (t-values), but also direction (coefficient signs), point to the need to assess statistical relationships for environmental justice analysis locally, rath er than globally. Th e Akaike Information Criterion (AIC) is the recommended measure for comparing a global regression model with a GWR model (Fotheringham et al. 2002). As shown in Table 5.1, the GWR model yields a lower AIC score than the global mode l, indicating a superior fit or improvement in overall model performance. To find evidence of spatial nonstationarity in the relationshi p between cumulative cancer risk and each explanator y variable, the interquartile range of an individual GWR parameter can be compared to twice the sta ndard error of the globa l regression parameter (Fotheringham et al. 2002). For cumulative cancer risk from all sources, the interquartile ranges for all six explanatory variables from the GWR model were found to be larger than their respective standard error obtained from conventio nal regression, thus providing an informal confirmation of the presence of spatial nonstationarity for their estimated parameters across Florida.
51 To assess spatial differences in overall model fit between the GWR and global regression models, estimates of local model pe rformance for each census tract in the form of R-squared values are depicted as a chor opleth map in Figure 5.1. The red line in the map legend denotes the adjusted R squared valu e from the original global model, with two classes above and below this value. This map can be used to visualize and identify locations within Florida where GWR produces an improvement or decline in overall model fit. The local R-square d values improve upon the global R-squared value of 0.52 in a majority of tracts in Florida, with the GWR models performing the best in northern Florida and parts of central and southern Florida. The final method of assessing spatially vary ing relationships ar e choropleth maps for each explanatory variable that represent the geographic distribution of their t-values across Florida, with respect to cumulative can cer risk (Figures 5.2 to 5.7). The choropleth classification is based on a manual distributi on that separates both positive and negative t-values at the 90 percent and 95 percent le vel of significance. Because of the large sample size (n=3,151), standard t-values of + 1.96 and + 1.645 are used to denote 95 percent and 90 percent levels of significance, respectively. On all these maps, the red color is used to depict positive significance, the blue color is used to display negative significance, and the grey co lor is used to represent non-significant t-values.
52Figure 5.1 Local R-Squared from GWR Model of Cancer Risk from All Sources
53Figure 5.2 Distribution of t-values for Proportion Black and Cancer Risk from All Sources Figure 5.3 Distribution of t-values for Proportion Hispanic and Cancer Risk from All Sources
54Figure 5.4 Distribution of t-values for Proportion Asian and Cancer Risk from All Sources Figure 5.5 Distribution of t-values for Home Ownership Rate and Cancer Risk from All Sources
55Figure 5.6 Distribution of t-values for Poverty Ra te and Cancer Risk from All Sources Figure 5.7 Distribution of t-values for Population Density an d Cancer Risk from All Sources
56 Proportion Black demonstrates a positive re lationship with cancer risk from all sources in almost 93 percent of tracts th roughout Florida, with the exception of the panhandle (Figure 5.2). This corresponds to the global regression result. However, a number of tracts show a negative relationship between the Hispanic and Asian proportions and cumulative cancer risk from a ll sources, contrary to the global regression results. While proportion Hispanic is positively related to cancer risk across much of the state, it is negatively associat ed with risk in the northeas tern portion of Florida (Figure 5.3). Proportion Asian is also positively associated with can cer risk from all sources across most of the state, but is negatively significant near the Miami metropolitan area (Figure 5.4). Home ownership rate was not significant in the global model, yet is negatively associated with cancer risk from all sources in tracts located in north, central, and southwest Florida (Figure 5.5). The poverty ra te variable behaves inconsistent with the expectations of environmental injustice across most of the state, exhibiting a negative relationship in most tracts. However, th e Tampa Bay metropolitan area is the only location in Florida where census tracts have an expected positive relationship with cancer risk (Figure 5.6). Tracts are sharply split between having a significantly positive and negative relationship between cancer risk and population density, with positive associations located in the panhandle and through Tampa Bay and ru ral central Florida and negative associations existing in north-c entral and south Flor ida (Figure 5.7). The results associated with GWR analysis of cancer risk from four known sources of air toxics are summarized and described in the remaining sections of this chapter.
575.2 Lifetime Cancer Risk from Major Point Sources When cancer risks from major point sources of air toxics such as manufacturing facilities and power plants are examined, the GWR parameter estimates for all explanatory variables do not show substantia l variation when compared to the global regression model parameters (Table 5.2). Th e GWR results for the Black proportion, for example, are reasonably consistent with th e global regression mode l with 93 percent of tracts yielding non-significant t-values. The Hi spanic proportion was positively related to cancer risk in the global mode l, but was found to be positiv ely significant in less than 50 percent of Florida tracts according to GW R. The Asian proportion was not deemed significant in the global model, but is found to be negatively significant in 36 percent of tracts based on the GWR output. In terms of socioeconomic characteristic s, home ownership rate was also not significant in the multiple regression model, but GWR results indicate negative significance in 18 percent of tracts and positiv e significance in 34 percent of tracts. Most tracts mirrored the global coefficient for poverty rate, while in the case of population density, the global estimate does not fall with in the GWR range and no tracts were found to be significant. The AIC score for the GWR model is marginally lower than the same statistic for the global regression model, indi cating a slight improvement in overall model performance. For cancer risk from major po int sources of air toxics, the comparison between the interquartile ranges from GWR m odel parameters and standard errors from the global regression model parameters indi cate that the Black proportion and home ownership rate are the only variables sugge sting spatial nonstationarity based on this measure.
58Table 5.2 GWR Results For Cancer Risk From Major Point Sources Global GWR Coefficients Statistical Significance of t-values Across Census Tracts in Florida Regression Coefficient Min Med Max Negative 90% level Not Significant Positive 90% level Proportion Black 0.11 ** -0.80-0.020.257.36% 92.64% 0.00% Proportion Hispanic 0.57** -0.530.550.880.00% 51.38% 48.62% Proportion Asian -0.62 ** -7.88-3.461.7335.52% 64.48% 0.00% Proportion Owner Occupied 0.10 ** -1.25-0.160.6117.60% 48.40% 34.00% Proportion Below Poverty 1.01** -1.111.323.920.43% 23.57% 76.00% Population Density (Ln) -0.04* 0.000.000.000.00% 100.00% 0.00% Adjusted R squared 0.01 ** 0.010.030.06 Akaike Information Criterion (AIC) 12602 ** 12554 p<0.05 **p<0.01 Local R-squared values associated with the regression model for cancer risks from major point sources are very low with a median value of only 0.03 and a maximum of 0.06 (Table 5.2). However, the global m odelÂ’s R-squared value of 0.01is exceeded in most census tracts in Florida. The spatial di stribution of R-squared values show slightly better model performance at the extreme north and south portions of the state with the poorest performance in central Florida (Fi gure 5.8). Choropleth maps depicting the geographic distribution of expl anatory variables whose signifi cant t-values vary across space are provided in Figures 5.9 to 5.13. On all these maps, the red color is used to depict positive significance, the blue color is used to display negative significance, and the grey color is used to re present non-significant t-values.
59Figure 5.8 Local R-Squared of Cancer Ri sk from Major Point Sources Figure 5.9 Distribution of t-values for Proportion Black and Cancer Risk from Major Point Sources
60Figure 5.10 Distribution of t-values for Proportion Hispanic and Cancer Risk from Major Point Sources Figure 5.11 Distribution of t-values for Proportion Asian and Cancer Risk from Major Point Sources
61Figure 5.12 Distribution of t-values for Home Ow nership Rate and Cancer Risk from Major Point Sources Figure 5.13 Distribution of t-values for Poverty Ra te and Cancer Risk from Major Point Sources
62 Although no census tracts show a positive re lationship between proportion Black and cancer risk from major point sources, th e 7 percent of tracts that are negatively associated with this type of risk are located in the east-central porti on of the state (Figure 5.9) where the Black proportion is rela tively smaller. Tracts showing a positive association between cancer risk from major point sources a nd Hispanic proportion can be found mainly in the southern Florida (Figur e 5.10). While the Asian proportion is not positively associated with cancer risk from major point sources in any tract, 36 percent of tracts indicate a negative relationship. Thes e tracts are located in the northeast and east central areas of Florida (Figure 5.11). Although a negative association was expect ed between home ownership rate and cancer risk from major point sources, only 18 pe rcent of tracts exhibit this relationship. These tracts are located in northern Florida (Figure 5.12), whereas those with a positive relationship (34 percent) are lo cated in southern Florida. A ccording to the environmental justice framework, the proportion below pove rty is anticipated to have a positive relationship with risk, and i ndeed 76 percent of tracts conf irm this association. These tracts are located throughout central and southern Florid a (Figure 5.13). A few tracts showing a negative relationship between poverty rate and can cer risk from major point sources are located north of Jacksonville on the Georgia border. 5.3 Lifetime Cancer Risk from Other Point Sources When cancer risks from other point sour ces or small emitters such as auto body shops and dry cleaning facilities are exam ined, the GWR parameter estimates for the explanatory variables exhibit some varia tion compared to those from the global regression model (Table 5.3) The GWR coefficients for the Black proportion indicate
63 that less than three-fourth of tracts in Flor ida have a significantly positive relationship with cancer risk which matches the global regr ession result, but the re maining tracts yield non-significant t-values. The coefficient for the Hispanic proportion also demonstrates the same relationship as the global regression parameter in a majority of tracts, but the same variable is either non-significant or ne gatively associated with cancer risk from this source in the remaining 30 percent of tracts in the state. For the pr oportion of Asians, 60 percent of tracts are positively associated with cancer risk from major point sources, also matching the global regression coefficient. Table 5.3 GWR Results For Cancer Risk From Minor Point Sources Global GWR Coefficients Statistical Significance of t-values Across Census Tracts in Florida Regression Coefficient Min Med Max Negative 90% level Not Significant Positive 90% level Proportion Black 3.64** -1.083.3810.900.00% 27.11% 72.89% Proportion Hispanic 3.73** -19.855.2826.1115.40% 14.97% 69.63% Proportion Asian 21.94** -21.6721.1873.330.00% 40.39% 59.61% Proportion Owner Occupied 0.44 ** -7.28-0.483.8926.06% 68.61% 5.33% Proportion Below Poverty 3.85** -14.360.4710.6212.43% 59.26% 28.31% Population Density (Ln) 0.36** 0.000.221.4241.12% 8.75% 50.13% Adjusted R squared 0.18 ** 0.070.380.62 Akaike Information Criterion (AIC) 18586 ** 16948 p<0.05 **p<0.01 Tracts exhibiting a negative relationship between cancer risk from other point sources and home ownership rate comprise 26 percent, while 5 pe rcent show a positive relationship. The global model, however, did not find home ownership to have a significant effect on cancer risk from this s ource. Reflecting the global coefficient, 28 percent of tracts in Florida indicate a positiv e relationship between risk and poverty rate,
64 while 12 percent show the opposite relationshi p. While the global regression parameter found population density to be significant acro ss Florida, GWR finds 41 percent of tracts to have the opposite relationship. As show n in Table 5.3, the GWR model yields a substantially lower AIC score than the global model, indicating a considerable improvement in overall model performance. Fo r cancer risk from other point sources of air toxics, the comparison between the in terquartile ranges from the GWR model parameters and standard errors from the global regression model parameters indicated spatial nonstationarity for al l explanatory variables. Local R-squared values associated with the multivariate regression model for cancer risks from other point sources are generally higher and suggest a much better model performance compared to cancer risk fr om major point sources, based on a median value of 0.38 and a maximum of 0.62 (Table 5.3). The global modelÂ’s R-squared value of 0.18 is exceeded in a majority of census trac ts in Florida. The sp atial distribution of Rsquared values shows that tracts with the best model performance are located in Jacksonville, Tampa, and West Palm Beach (Figure 5.14). Choropleth maps depicting the geographic di stribution of significant t-values for explanatory variables are provi ded in Figures 5.15 to 5.20. On all these maps, the red color is used to depict positive significance, the blue color is used to display negative significance, and the grey co lor is used to represent non-significant t-values.
65Figure 5.14 Local R-Squared of Cancer Ri sk from Other Point Sources Figure 5.15 Distribution of t-values for Proportion Black and Cancer Risk from Other Point Sources
66Figure 5.16 Distribution of t-values for Proportion Hispanic and Cancer Risk from Other Point Sources Figure 5.17 Distribution of t-values for Proportion Asian and Cancer Risk from Other Point Sources
67Figure 5.18 Distribution of t-values for Home Ow nership Rate and Cancer Risk from Other Point Sources Figure 5.19 Distribution of t-values for Poverty Ra te and Cancer Risk from Other Point Sources
68Figure 5.20 Distribution of t-values for Population Density and Cancer Risk from Other Point Sources The Black proportion is positively related to cancer risk from other point sources throughout northeastern Florida, south Florida, and the west-central coast (Figure 5.15). Tracts that are positively associ ated with proportion Hispanic are located in south Florida and west Florida, while tracts negatively a ssociated with cancer risk are located in northeast Florida (Figure 5.16). Cancer risk is positively related to the Asian proportion in northeast Florida, on the east coast, a nd around the Tampa Bay metropolitan area (Figure 5.17). The socioeconomic variables show both positive and negative associations with cancer risk from other sources. Home ownership rate is nega tively significant with tracts in the northeast and north-central portio ns of the state, spreading down to the Tampa Bay metropolitan area (Figure 5.18). A handful of tracts show a positive relationship in areas of West Palm Beach a nd the Florida Keys. While poverty rate has a
69 significantly positive effect on cancer risk among tracts located in Tampa Bay and Daytona Beach (Figure 5.19), tracts in southw est Florida and north-cen tral Florida depict a negative relationship between poverty and ca ncer risk the opposite relationship with respect to the global model parameter. Popula tion density is expected to have a positive association with cancer risk fr om other point sources and trac ts located in the panhandle and central to south Florida reflects this relationship (Figure 5.20) However, tracts in north-central and south Florid a show the opposite relationship. 5.4 Lifetime Cancer Risk from On-Road Mobile Sources When cancer risks from on-road mobile so urces such as cars, trucks, and other vehicles found on public roadways are examined, the GWR parameter estimates for the explanatory variables suggest minimal spatial nonstationarity (Table 5.4). The t-values generated by the GWR model for all three racial/ethnic variables are positively related to cancer risk in all Florida tracts, an exact match with the global regression result. The tvalues for home ownership rate also reflect the global coefficient, with 100 percent of tracts classified as negatively significant. The only variable indi cating any variation in results is poverty rate. While the global regres sion model did not find this variable to be significant, its coefficient is negatively signifi cant in 97 percent of tracts according to the GWR model. Population density also showed no spatial variation in the significance of its t-values, with all tracts indicating a significan tly positive relationship that is identical to the global regression findings. The lack of variation across th e study area in model parameters could point to a problem in th e adaptive bandwidth selection for this particular dependent variable. The AIC scor e for the GWR model associated with cancer risk from on-road mobile sources is margin ally lower than the same statistic for the
70 global model, indicating only a minor improve ment in model performance. As can be expected based on the GWR results summarized in Table 5.4, the comparison between the interquartile ranges from the GWR model parameters and standard errors from the global regression model did not indicate spatia l nonstationarity for any of the explanatory variables. Table 5.4 GWR Results For Cancer Risk From On-Road Mobile Sources Global GWR Coefficients Statistical Significance of t-values Across Census Tracts in Florida Regression Coefficient Min Med Max Negative 90% level Not Significant Positive 90% level Proportion Black 4.44** 6.267.998.990.00% 0.00% 100.00% Proportion Hispanic 8.12** 6.4511.1418.320.00% 0.00% 100.00% Proportion Asian 26.12** 56.9870.1682.280.00% 0.00% 100.00% Proportion Owner Occupied -1.27* -6.06-4.09-2.88100.00% 0.00% 0.00% Proportion Below Poverty 0.63 ** -11.76-5.98-1.6096.83% 3.17% 0.00% Population Density (Ln) 1.44** 0.000.000.000.00% 0.00% 100.00% Adjusted R squared 0.40 ** 0.190.280.33 Akaike Information Criterion (AIC) 20797 ** 20577 p<0.05 **p<0.01 The local R-squared values indicate a me dian value of 0.28 and a maximum of 0.33, with no tracts improving upon the global model performance of 0.40 (Figure 5.21) for on-road mobile sources of ambient air t oxics. The GWR model performed the best in the Florida panhandle and South Florida (Figur e 5.21). Because of the lack of spatial variation in the significant tvalues, only the poverty rate choropleth map is provided (Figure 5.22). Unlike the rest of the state, tr acts in the southern portion of Florida do not show a significant relationship between poverty rate and cancer risk from on-road mobile sources. On all these maps, the red color is us ed to depict positive significance, the blue
71 color is used to display negative significance, and the grey color is used to represent nonsignificant t-values. Figure 5.21 Local R-Squared of Cancer Risk from On-Road Mobile Sources Figure 5.22 Distribution of t-values for Poverty Rate and Cancer Risk from On-Road Mobile Sources
72 5.5 Lifetime Cancer Risk from Non-Road Mobile Sources When cancer risks from non-road mobile so urces of air toxics (e.g., airplanes, lawnmowers, and ships) are examined, the GWR parameter estimates for the explanatory variables suggest a greater amount of spatial nonstationarity (Table 5.4) compared to the GWR model for on-road sources. This model provides fewer tracts with significant tvalues than those from the othe r dependent variables, with a ll six variables indicating that at least 50 percent tracts are classified as non-significant. The t-values for the Black proportion show that this variab le is not significant in 87 percent of tracts, yet has a significantly positive relationship in the global model. For this variable, only 3 percent of tracts are positively associated with risk, while 10 percent are negatively related. The Hispanic variable behaves similarly, with 19 percent of tracts re flecting the positively significant global coefficient, while 15 percen t of tracts yielding a negative association with cancer risk. The proportion of Asians was not a significan t indicator of risk in the global model, but 23 percent of tracts in th e GWR model indicate a significantly negative relationship between this variable and can cer risk from non-road mobile sources. Home ownership rate was negatively rela ted to cancer risk in the global model, and maintained this association in 41 percen t of tracts in the GWR model. Poverty rate only matched the negative global coefficient in 7 percent of tracts, while 13 percent of tracts exhibited the opposite re lationship. Population density was split between negatively (20 percent) and positively (30 percent) significant t-values in the GWR model, while the global model was positively related to cancer ri sk from non-road mobile sources. As seen in Table 5.5, the GWR model produces a subs tantially lower AIC score than the global model, indicating a considerable improveme nt in overall model performance over the
73 global model for cancer risks from non-road mobile sources. The comparison between the interquartile ranges for all GWR model pa rameters and the standard error of the global regression parameters provided strong evidence of no nstationarity for all six explanatory variables. Table 5.5 GWR Results For Cancer Risk Fr om Non-Road Mobile Sources Global GWR Coefficients Statistical Significance of t-values Across Census Tracts in Florida Regression Coefficient Min Med Max Negative 90% level Not Significant Positive 90% level Proportion Black 0.56** -3.82-0.014.7910.46% 87.75% 2.79% Proportion Hispanic 2.47** -16.06-0.317.4514.49% 66.49% 19.02% Proportion Asian 0.23 ** -41.13-2.1340.6523.59% 73.65% 2.76% Proportion Owner Occupied -1.39** -5.45-1.170.7740.55% 59.45% 0.00% Proportion Below Poverty -1.24* -6.620.118.496.91% 80.09% 13.00% Population Density (Ln) 0.30** -0.910.110.4819.94% 50.06% 30.00% Adjusted R squared 0.21 ** 0.060.390.81 Akaike Information Criterion (AIC) 13508 ** 12739 p<0.05 **p<0.01 The local R-squared values generated by the GWR model for cancer risks from non-road mobile sources are generally highe r than global regression modelÂ’s R-squared value of 0.21, with a median value of 0.39 a nd a maximum value of 0.81. The tracts with the best model performance are located thr ough the panhandle and centr al Florida (Figure 5.23). The geographic distribution of significant t-values for all explanatory variables are shown in Figures 5.23 to 5.29. On all these ma ps, the red color is used to depict positive significance, the blue color is used to displa y negative significance, and the grey color is used to represent non-significant t-values.
74Figure 5.23 Local R-Squared of Cancer Risk from Non-Road Mobile Sources Figure 5.24 Distribution of t-values for Proportion Black and Cancer Risk from NonRoad Mobile Sources
75Figure 5.25 Distribution of t-values for Proportion Hispanic a nd Cancer Risk from NonRoad Mobile Sources Figure 5.26 Distribution of t-values for Proportion Asian and Cancer Risk from NonRoad Mobile Sources
76Figure 5.27 Distribution of t-values for Home Ow nership Rate and Cancer Risk from Non-Road Mobile Sources Figure 5.28 Distribution of t-values for Poverty Rate and Cancer Risk from Non-Road Mobile Sources
77Figure 5.29 Distribution of t-values for Population Density an d Cancer Risk from NonRoad Mobile Sources The Black proportion has a negative relati onship with cancer risk from non-road mobile sources in tracts lo cated in the Miami metropolitan area and central Florida (Figure 5.24), while 3 percent of tracts have a positive relationship with this type of risk in southwest Florida, the Tampa Bay metropolitan area, Fort Lauderdale, and central Florida. The Hispanic proportion demonstr ates a negative relationship with this dependent variable in tracts located in the Miami and Or lando metropolitan areas as well as along the Alabama border (Figure 5.25), while a positive association is observed around the Tampa Bay and Palm Bay metropol itan areas and in south Florida. In Orlando, Miami, and West Palm Beach, tracts indicate a negative relationship between cancer risk from non-road mobile sources a nd the Asian proportion (Figure 5.26), while
78 tracts throughout central Flor ida and southwest Florida dem onstrate the expected positive relationship with risk. Tracts that show a significant and nega tive association between home ownership rate and cancer risk from non-road mobile sources are located along the Alabama border, through the Jacksonville, Orlando, Tampa, Fort Myers/ Cape Coral, Port St. Lucie, and Miami metropolitan areas, along with all of southwest Florid a (Figure 5.27). The expected positive relationship exists between this dependent variable and poverty rate among tracts in the Miami, West Palm B each, Jacksonville, Tampa, and Palm Bay metropolitan areas, while a negative relationshi p is found in south Florida (Figure 5.28). Population density is positively associated with cancer risk from non-sources of air toxics throughout north Florida, parts of central Florida, and th roughout southwest Florida. Pockets of positive relationships can be seen around Orlando and throughout south Florida (Figure 5.29). 5.6 Summary of GWR Results Satisfactory model performance among the five sources of risk varies across Florida, but certain areas of the state cons istently demonstrate better model fit when compared with surrounding tracts. GWR model performance improves upon traditional multivariate regression in north FloridaÂ’s panhandle area and north of the Tampa Bay region for most sources of cancer risk. However, it becomes difficult to compare performance between models due to the method used to determine the optimal number of census tracts necessary to conduct GWR analys is. The criteria for bandwidth selection were consistent for all models but the adaptive kernel me thod (as described in section 3.3) results in variation in the sample size produced by minimization of the AIC. Out of a
79 total xxx census tracts within and adjacent to Florida, 719 tracts were used to determine local parameters in the GWR models for both cumulative cancer risk and minor point sources. Larger sample sizes of 2,470 and 2,966 tracts were utilized in the GWR models for major point sources and on-road mobile sources, respectively. These sample sizes explain the relative homogeneity in the c horopleth maps for local model parameters associated with these two variables. The GWR model for nonÂ–road mobile sources used the smallest sample size of only 276 tracts This low number is reflected in the comparatively nuanced choropleth maps that show greater spatial variability in model parameters compared to the other GWR models. Even with this variation in sample size, certain geographic patterns are consistently displayed in the relationshi p between the dependent and explanatory variables. Proportion Black indicates a posit ive relationship with multiple sources of cancer risk throughout peninsular Florida. Proportion Hispanic has a positive association with cancer risk from various sources thr oughout south Florida, a nd also exhibits a negative relationship in many models in north Florida. This uniformity of patterns in both the expected positive relationships between racial variables and risk as well as the unanticipated negative a ssociations can also be seen am ong the distributions of the Asian GWR results. For the most part, relationships remain positive across much of the state, but negative associations can be seen in the Miami metropolitan area and in central Florida, in multiple models. Home ownershi p rate has pockets of unexpected positive relationships in south Florida, while th e poverty rate variable shows the opposite expected associations with risk throughout the east coast and Tampa Bay. Population density demonstrates a positive relationship w ith multiple sources of cancer risk in the
80 panhandle and central Florida, while the opposit e is observed in north central and south Florida.
81 CHAPTER 6: CONCLUSIONS Human geography has been concerned w ith how the industrial and commercial developments of the last 150 years have advers ely impacted different sectors of society, particularly the under-privileged and under-represented. The emergence of the environmental justice movement in the 1980s led to an increased focus on how racial/ethnic minorities and socioeconomically disadvantaged residents were impacted by human-induced hazards and related health risks. Geographers became interested in environmental justiceÂ’s inherently spatial issues of distributional equi ty and social justice (Walker and Bulkeley 2006). A variety of qua ntitative methodologies have been utilized in the environmental justice research literatur e to evaluate whether technological hazards and risks are distributed e quitably with respect to di fferent racial, ethnic, and socioeconomic groups. Previous quantitative re search on environmental justice, however, has been impeded by three specific limitations in the types of data utilized and methodologies employed. First, a large number of studies have measured dispropor tionate burdens based on distance to pollution sources or the quantit y or toxicity of emitted pollutants, thus failing to account for the adverse health imp acts of exposure to these chemicals. Second, past studies have focused prim arily on inequities associated w ith major stationary sources of pollution such as industri al manufacturing facilities, i gnoring mobile emission sources and smaller emitters that also pollute the local environment and contribute substantially to environmental health risks. Finally, almo st all previous stud ies using conventional
82 statistical methods such as multivariate re gression have assumed that relationships between dependent and explanatory variables ar e stationary across sp ace in a given study area. Since traditional regre ssion analysis cannot uncover lo cal variations, it leads to broad generalizations about the entire study region, thus igno ring the notion that specific places within the same study area might differ from each other with regard to the nature and extent of environmental injustice. This thesis extends quantitative research on environmental justice by addressing these three problems, through a case study of estimated cancer risks from inhalation exposure to air toxics in Florida. Th e 1999 National-Scale Air Toxics Assessment (NATA) data used for this study provides modeled estimates of cancer and non-cancer risks from air toxics at the census tract le vel. This is an improvement over previous studies because the risk assessment met hodology incorporates chemical fate and transport, weather patterns, and human inhalation data. The NATA also provides estimates of adverse health risks for multiple types of air toxics emission sources, allowing for less readily acknowledged sour ces of pollution to be accounted for and incorporated into environmental justice anal yses. In order to addr ess the limitations of conventional regression, this study uses a spa tially sensitive statis tical technique that allows for a unique regression model to be produced at each data point. Instead of generating one set of model parameters a nd diagnostics for the entire study area, geographically weighted regression gives more influence to data points closer to the origin point for which the model is being built, thus allowing local variations in statistical relationships to be examined.
83 The first stage of this study utilizes trad itional statistical methods such as linear correlation and multiple regression to examine the environmental justice implications of lifetime cancer risk for entire study area. Biva riate correlation analysis provides strong evidence of racial/ethnic and socioeconomic in equities in the distribution of cancer risks from all emissions sources. The multivariate regression results provides evidence that race and ethnicity, particularly Black and Hisp anic proportions, are consistent predictors of the presence of cumulative cancer risk, as we ll as cancer risk from the four individual sources of air toxics, even after contro lling for socioeconomic characteristics and population density. Increased Asia n proportions are indicative of greater cancer risk from other point sources and on-road mobile sour ces. The socioeconomic variables indicate mixed results, with lower home ownership ra te being related to increased cancer risk from only mobile sources of air toxics. An increased poverty rate is associated with greater cancer risk for point sources, but not mobile sources of air toxics. Higher population density was a positive indicator of cancer risk for both types of mobile sources, minor point sources, and cumula tive cancer risk from all sources. In contrast to the conventional correla tion and multivariate regression analysis, the geographically weighted regression (GWR ) analysis indicated that many of the observed statistical associations between can cer risk and specific explanatory variables are not uniform across Florida, but are more significant in some areas and not evident in other locations. Despite the differences in regression methodology, the GWR results consistently indicate the perv asive effect of Black and Hisp anic proportions in explaining the geographic distribution of cancer risks fr om air toxics in Florida, even when controlling for other explanatory factors. A higher proportion of Bl ack residents in a
84 census tract suggests greater cancer risk from all sources throughout much of the state. More specifically, the Black proportion is positively related to cancer risk from other point sources among tracts in Jacksonville and other areas in north Florida, around Tampa Bay, and throughout south Florida, whil e cancer risk from mobile sources display the same relationship northwest of Orlando a nd Tampa, southeast of Naples, in Miami Beach, and West Palm Beach. Although cancer risks are also distributed in equitably with respect to the Hispanic population, the GWR analysis reveals that locations which sh ow a significant and positive association for the Hispanic proportion are different from those which depict a similar relationship for the Black proportion. Th e presence of Hispanic residents indicates greater cancer risks from all sources th roughout the Tampa Bay metropolitan area and along the east coast of Flor ida. More specifically, highe r proportions of Hispanic residents suggest increased cancer risk from point sources across south Florida, and from non-road mobile sources in Tampa Bay, Cape Canaveral, Miami, and Fort Lauderdale. While GWR confirms where these expected relationships exist, it also highlights areas where the opposite is observed. While the Hispanic proportion is positively associated with cancer risk across much of Flor ida, it seems to negatively relate to cancer risk from other point sources around Jacks onville. Higher home ownership rates indicate increased risk from other point sources in West Palm Beach and the Florida Keys. Lower poverty rates point to increased cumulative can cer risk from all sour ces in most Florida tracts, but the exception occurs in the Ta mpa Bay metropolitan ar ea where cancer risks are disproportionately di stributed with respect to people in poverty. These local variations in the global positive relationship between can cer risk and ethnicity or socioeconomic
85 status are obscured by conventional multivariate regression, but are revealed by the use of GWR in conjunction with chor opleth mapping. The results of this study clearly suggest the need for researchers to recognize the us efulness of GWR as an exploratory data analysis tool for environmental justice assessment. It is important to discuss, however the caveats that bound the improvements made by the data and methodology used for this research. There are specific limitations associated with EPAÂ’s NATA data set used for this research. The 1999 NATA estimates cancer risk only from inhalation exposure to air toxics and does not account for exposure through other pathways such as ingestion a nd skin contact. The NATA information is also not a substitute for actual health outcomes data, and only represent modeled estimates of cancer risk based on EPAÂ’ s risk assessment guidelines. The 1999 NATA also assesses the cancer risk from 133 out of 188 air toxics, and t hus fail to provide a complete assessment of all toxic air polluta nts of concern (EPA 2008d). Also, this study is a cross-sectional analysis of adverse health risks at a spec ific point in time (1999), and should not be used to elucidate causal rela tionships between race/ethnicity and exposure to pollution or deduce the chronol ogical order of events leadi ng to the current disparities. Geographically weighted regression also pr esents challenges in its application. A relatively new method of analysis, GWR is ripe for critique and assessment of its capabilities. The choice of adaptive or fixe d bandwidth for the selection of neighboring census tracts is a critical decision that could greatly influence the results. The adaptive or variable bandwidth approach was utilized in this study to account fo r the spatial variation in the size of the tracts or the density of tract centroids. However, a fixed kernel bandwidth which uses distance rather than a specific numbe r of neighbors to determine
86 the local sample size could be a more appropr iate method for analysis in certain cases where resolution is more uniform (Mennis a nd Jordan 2005). It is also important to consider that GWR also does not always im prove the overall performance or fit of a multivariate regression model. As seen in the results for cancer risk from major point sources, both the global regres sion and GWR models resulted in low R-squared values and point to a potential pr oblem with the choice of explanatory variables. This thesis has several important implica tions for public policy. The results of the study indicate that environmenta l injustice is occurring in spec ific places within Florida and also identify the air pollution sources re sponsible for the disp roportionate health impacts. This information can assist local advocacy groups, grassroo ts organizations, and minority populations in their attempts to obtai n increased oversight of hazardous facilities or highways near areas populated by minority or low-income residents, in addition to providing evidence of unequal exposure to ha zards for possible litigation. Since this study uses health risk assessments from the EP A to evaluate spatial inequities, the results can lend legitimacy to the environmental health burdens imposed on community members and thus advance the aims of envir onmental justice. GWR can be particularly useful because it has the capacity to provide sp ecific information about locations that are disproportionately impacted by hazardous air pol lutants. These locational differences in the unique relationships between the explan atory variables and different types of pollution sources can assist policy formati on and regulation by high lighting the unique issues faced by a city or region and represen t a starting point for qua litative analysis of the various economic, historical, and political pr ocesses at work in the study area. Several variables such as the proportion of Black a nd Hispanic populations are consistently and
87 positively associated with cancer risk from a ll sources in all parts of the state. This inequitable pattern points to the need to asse ss the environmental justice implications of current and future zoning and siting regulat ions in the state of Florida. Along with enforcement of environmental regulations, disproportionately exposed minority and lowincome residents must be included in the decision-making process in order to combat unintended discrimination and inequitable exposure to the negative by-products of modern society. In conclusion, this thesis provides strong evidence of racial/ethnic and socioeconomic disparities in the adverse health effects of exposure to outdoor air toxics from both point and mobile air pollution sources in Florida. The results demonstrate that conventional multivariate regression for e nvironmental justice analysis can hide important local variations in the relati onships between cancer risk and relevant explanatory variables such as race, ethnicity and socioeconomic status. Further research is needed to characterize the various socio economic, political, and sp atial processes that are causing cancer risk inequities in specific areas of Florida. Meanwhile, these results raise new challenges for both policy makers a nd environmental justice advocates in terms of developing regulatory and po llution prevention strategies that address both stationary and mobile emission sources.
88 REFERENCES Anderton, Douglas L., Andy B. Anderson, J ohn M. Oaks, and Michael R. Fraser. 1994. Environmental equity: The demographics of dumping. Demography 31:229-48. Ali, Kamar, Mark D. Partridge, and M. Ro se Olfert. 2007. Can geographically weighted regressions improve regional analysis and policy making? International Regional Science Review 30: 300-29. Apelburg, Benjamin J., Timothy J. Buckley, and Ronald H. White. 2005. Socioeconomic and racial disparities in cancer risk from air toxics in Maryland. Environmental Health Perspectives 113:693-99. Asch, Peter and Joseph J. Seneca. 1978. Some evidence on the distribution of air quality. Land Economics 54: 278-97. Ash, Michael and T. Robert Fetter. 2 004. Who lives on the wrong side of the environmental tracks? Evidence from the EPAÂ’s Risk-Screening Environmental Indicators Model. Social Science Quarterly 85: 441-62. Been, Vicki. 1994. Locally undesirable land uses in minority neighborhoods: Disproportionate siting or market dynamics? The Yale Law Journal 103:13831406. Bevc, Christine A., Brent K. Marshall, and J. Steven Picou. 2007. Environmental justice and toxic exposure: Toward a spatial model of physical health and psychological well-being. Social Science Research 36:48-67. Boer, J. Thomas, Manuel Pastor, Jim L. Sadd, and Lori D. Snyder. 1997. Is there environmental racism? The demographics of hazardous waste in Los Angeles County. Social Science Quarterly 78:793-810. Boone, Christopher G. 2002. An assessment and explanation of environmental inequity in Baltimore. Urban Geography 23:581-95. Bowen, William M., Mark J. Salling, Kindsley E. Haynes, Ellen J. Cyran. 1995. Towards environmental justice: Spatial equity in Ohio and Cleveland. Annals of the Association of American Geographers 85: 641-63.
89 Cahill, Meagan and Gordon Mulligan. 2007. Using geographically weighted regression to explore local crime patterns. Social Science Computer Review 25: 174-93. Calvo, Ernesto and Marcelo Escolar. 2003. The local voter: A geographically weighted approach to ecological inference. American Journal of Political Science 47: 189204. Chakraborty, Jayajit. 2001. Acute exposure to extremely hazardous substances: An analysis of environmental equity. Risk Analysis 21:883-95 Chakraborty, Jayajit. 2004. The geographi c distribution of potential risks posed by industrial toxic emi ssions in the U.S. Journal of Environmental Science and HealthÂ—Part A 39: 559-75. Chakraborty, Jayajit and Marc P. Armstrong. 1997. Exploring the use of buffer analysis for the identification of impacted area s in environmental equity assessment. Cartography and Geographic Information Systems 24: 145-57. Chakraborty, Jayajit and Marc P. Armstrong. 2004. Thinking outside the circle: Using geographical knowledge to focus environm ental risk assessmen t investigations. WorldMinds: Geographical Perspectives on 100 Problems ed. Donald G. Janelle, Barney Warf, and Kathy Hansen, 435-42. Dordecht, The Netherlands: Kluwer Academic Publishers. Chakraborty, Jayajit and Paul Zandberg en. 2007. Children at risk: Measuring racial/ethnic disparities in potential e xposure to air pollution at school and home. Journal of Epidemiology and Community Health 61:1074-79. Chapin, Tim. 2006. Still booming after all thes e years: Looking at FloridaÂ’s projected population and employment in 2015. www.cba.ufl.edu/purc/docs/presentation_2006Chapin.pdf (Last accessed November 1, 2008). Clinton, William J. 1994. Federal actions to address environmental justice in minority populations and low-income populations: Executive orde r 12898 of February 11, 1994. Federal Register 59, 7629-7633. Cutter, Susan L., Danika Holm, and Lloyd Clar k. 1996. The role of geographic scale in monitoring environmental justice. Risk Analysis 16:517-26. Daniels, Glynis and Samantha Friedman. 1999. Spatial inequality a nd the distribution of industrial toxic re leases: Evidence from the 1990 TRI. Social Science Quarterly 80: 244-62. EPA (U.S. Environmental Protection Agency). 2008a. Basic information. http://www.epa.gov/compliance/basics/ej.html (Last accessed October 20, 2008).
90 EPA (U.S. Environmental Protection Agency). 2008b. About air toxics. http://www.epa.gov/air/toxicair/newtoxics.html (Last accessed October 20, 2008). EPA (U.S. Environmental Protection Agency). 2008c. The four steps of the assessment. http://www.epa.gov/ttn/at w/nata1999/4steps99.html (Last accessed October 20, 2008). EPA (U.S. Environmental Protection Agency). 2008d. Frequent questions. http://www.epa.gov/ttn/atw/nata1999/nata 99faq.html (Last accessed October 20, 2008). EPA (U.S. Environmental Protection Agen cy). 2008e. Glossary of key terms. http://www.epa.gov/ttn/atw/nata1999/glo ss1.html (Last accessed March 7, 2009). Fitos, Athanasia, and Jayajit Chakrabort y. 2003. Who lives near drycleaners? A geographic analysis of environmental in equity on Hillsborough County, Florida. The Florida Geographer 34:44-64. Fotheringham, Stewart A., Chris Bruns don, and Martin E. Charlton. 2002. Geographically Weighted Regression: Th e Analysis of Spatially Varying Relationships Chichester, U.K.: Wiley. Freeman, A.M. 1972. Distribution of environmental equity. Environmental Quality Analysis ed. A. Kneese and B. Bower, 163-76. Baltimore, MD: Johns Hopkins Press. GAO (U.S. General Accounting Office) 1983. Siting of hazardous waste landfills and their correlations with th e racial and socio-economi c status of surrounding communities. Washington D.C.: General Accounting Office. Gianessi, Leonard P., Henry M. Peskin, a nd Edward Wolff. 1979. The distributional effects of uniform air pollution policy in the United States. Quarterly Journal of Economics 93:281-301. Glickman Theodore S. 1994. Measuring environmental equity with Geographic Information Systems. Renewable Resources Journal 12: 17-21. Godish, Thad. 2004. Air Quality, 4th Edition Boca Raton, Florid a: Lewis Publishers. Grineski, Sara E., Bob Bolin, and Victor Agadjanian. 2005. Tuberculosis and urban growth: Class, race, and disease in early Phoenix, Arizona, USA. Health and Place 12: 603-16.
91 Jacobson, Jerry O., Nicolas W. Hengartner and Thomas A. Louis. 2005. Inequity measures for evaluations of environmental justice: a case study of close proximity to highways in New York City. Environment and Planning 37: 21-43. Jerrett, Michael, Michael Buzzelli, Richard T. Burnett, and Patrick F. DeLuca. 2001. Particulate air pollution, so cial confounders, and mortalit y in small areas of an industrial city. Social Science and Medicine 60: 2845-63. Kavanagh, Adrian, Richard Sinnott, Stewart Fo theringham, and Martin Charlton. 2006. A geographically weighted regression analys is of general elec tion turnout in the Republic of Ireland. Papers of the Political Studi es Association of Ireland Conference October 20, 2006. Kingham, Simon, Jamie Pearce, and Peyman Zawar-Reza. 2007. Driven to injustice? Environmental justice and vehicle pollution in Chri stchurch, New Zealand. Transportation Research Part D 12: 254Â–63. Knezevic, Zlatko and Jayajit Chakraborty. 2004. Exploring the environmental justice consequences of nuclear power pl ant siting in the US Southeast. Papers of the Applied Geography Conferences 27: 332-41. Linder, Stephen H., Dritana Marko, and Ke n Sexton. 2008. Cumulativ e cancer risk from air pollution in Houston: Disparities in risk burden and social disadvantage. Environmental Science & Technology 42: 4312-22. Liu, Feng. 1997. Dynamics and causation of environmental equity, locally unwanted land uses, and neighborhood changes. Environmental Management 21: 643-56. Longley, Paul A. and Carolina Tobon. 2004. Spatial dependence and heterogeneity in patterns of hardship: An intra-urban analysis. Annals of the Association of American Geographers 94: 503-19. McGurty, Eileen. 2000. Warren County, NC, and the emergence of the environmental justice movement: Unlikely coalitions a nd shared meanings in local collective action. Society & Natural Resources 13: 373-87. Mennis, Jeremy. 2002. Using geographic info rmation systems to create and analyze statistical surfaces of population and risk for environmental justice analysis. Social Science Quarterly 83: 281-97. Mennis, Jeremy. 2006. Mapping the results of geographically weighted regression. The Cartographic Journal 43: 171-79.
92 Mennis, Jeremy L. and Lisa Jordan. 2005. The distribution of environmental equity: Exploring spatial nonstationar ity in multivariate models of air toxic releases. Annals of the Associatio n of American Geographers 95: 249-68. Mitchell, Jerry, Deborah S. K. Thomas, a nd Susan L. Cutter. 1999. Dumping in dixie revisited: The evolution of environm ental injustices in South Carolina. Social Science Quarterly 80: 229-43. Mohai, Paul, and Bunyan Bryant. 1992. Envi ronmental Racism: Reviewing the evidence. Race and the incidence of environmental hazards ed. Bunyan Bryant and Paul Mohai, 163-76. Boulder, CO: Westview Press. Morello-Frosch, Rachel, Manuel Pastor, a nd James Sadd. 2001. Environmental justice and Southern CaliforniaÂ’s Â“riskscapeÂ”: Th e distribution of air toxics exposures and health risks among diverse communities. Urban Affairs Review 36:551-78. Most, Michael, Raja Sengupta, Michael Bu rgener. (2004) Spatia l scale and population assignment choices in envir onmental justice analyses. The Professional Geographer 56: 574-86. Pastor, Manuel, Rachel More llo-Frosch, and James L. Sa dd. 2001. Which came first? Toxic facilities, minority move-i n and environmental justice. Journal of Urban Affairs 23:1-21. Pastor, Manuel, James L. Sadd, and Rachel Morello-Frosch. 2004. Reading, writing, and toxics: ChildrenÂ’s health, academic performance, and environmental justice in Los Angeles. Environment and Planning C: Government and Policy 22:271-90. Pastor, Manuel, Rachel Morell o-Frosch, and James L. Sadd. 2005. The air is always cleaner on the other side: Race, space, and ambient air toxics exposures in California. Journal of Urban Affairs 27:127-48. Perlin, Susan A., Ken Sexton and David W. S. Wong. 1999. An examination of race and poverty for populations living near i ndustrial sources of air pollution. Journal of Exposure Analysis and Environmental Epidemiology 9: 29-48. Perlin, Susan A., David Wong, and Ken Sexton. 2001. Residential proximity to industrial sources of air pollution: Interrelations hips between race, poverty, and age. Journal of the Air &Waste Management Association 51:406-21. Pollock III, Philip H., and M. Elliot Vittas. 1995. Who bears the burdens of environmental pollution? Race, ethnicity, and environmental equity in Florida. Social Science Quarterly 76: 294-310.
93 Pulido, Laura. 2000. Rethinking environmen tal racism: White privilege and urban development in Southern California. Annals of the Associ ation of American Geographers 90:12-40. Schmidtlein, Mathew C., Christina Finch, and Susan L. Cutter. 2008. Disaster declarations and major hazard oc currences in the United States. The Professional Geographer 60: 1-14. Sheppard, Eric, Helga Leitner, Robert B. McMaster, and Hongguo Tian. 1999. GIS-based measures of environmental equity: Explor ing their sensitivity and significance. Journal of Exposure Analysis and Environmental Epidemiology 9:18-28. Stretesky, Paul, and Michael J. Hogan. 1998. An analysis of Superfund sites in Florida. Social Problems 45:268-87. Tiefenbacher, John P. and Ronald R. Hagelm an. 1999. Environmental equity in urban Texas: Race, income, and patterns of acu te and chronic toxic air releases in metropolitan counties. Urban Geography 19: 516-33. United Church of ChristÂ’s Commi ssion for Racial Justice. 1987. Toxic Wastes and Race in the United States: A National Re port on the Racial and Socio-Economic Characteristics of Communities with Hazardous Waste Sites New York: Public Data Access. United Church of ChristÂ’s Commi ssion for Racial Justice. 2007. Toxic Wastes and Race at Twenty: 1987-2007. Cleveland, OH: United Church of Christ. United States Census Bureau. 2006. 2006 American Community Survey. http://factfinder.census.gov/servlet/Da tasetMainPageServlet?_program=ACS&_s ubmenuId=&_lang=en&_ts = (Last accessed February 9, 2009). Walker, Gordon and Bulkeley, Harriet. 2006. Geographies of environmental justice. GeoForum 37: 655-59. Wernette, D.R. and L.A. Nieves. 1992. Breathing polluted air: Minorities are disproportionately exposed. EPA Journal 18:16-17. Wu, Yi-Chen and Stuart A. Batterman. 2006. Pr oximity of schools in Detroit, Michigan to automobile and truck traffic. Journal of Exposure Science and Environmental Epidemiology 16: 457-70. Zahran, Sammy, Donald W. Hastings, and Samuel D. Brody. 2008. Rationality, inequity, and civic vitality: The distri bution of treatment, storage, and disposal facilities in the Southeast. Society & Natural Resources 21: 179-96.
94 Zhuang, Duan. 2006. Spatial dependence and neighborhood effects in mortgage lending: A geographically weighted regression a pproach. LUSK Center for Real Estate, USC.
96Appendix A: Fre quency Histograms Figure A-1 Frequency Histogram of Cumulative Cancer Risk per Million People Figure A-2 Frequency Histogram of Major Point S ources of Cancer Risk per Million People
97Appendix A (continued) Figure A-3 Frequency Histogram of Other Point S ources of Cancer Risk per Million People Figure A-4 Frequency Histogram of On-Road Mob ile Sources of Cancer Risk per Million People
98Appendix A (continued) Figure A-5 Frequency Histogram of Non-Road M obile Sources of Cancer Risk per Million People Figure A-6 Frequency Histogram of Propor tion Black by Census Tract, 2000
99 Appendix A (continued) Figure A-7 Frequency Histogram of Proportion Hispanic by Census Tract, 2000 Figure A-8 Frequency Histogram of Propor tion Asian by Census Tract, 2000
100 Appendix A (continued) Figure A-9 Frequency Histogram of Povert y Rate by Census Tract, 2000 Figure A-10 Frequency Histogram of Home Ownership Rate by Census Tract, 2000
101Appendix A (continued) Figure A-11 Frequency Histogram of Population Density by Census Tract, 2000 Figure A-12 Frequency Histogram of the Natural Log of Population Density by Census Tract, 2000