USFDC Home  USF Electronic Theses and Dissertations   RSS 
Material Information
Subjects
Notes
Record Information

Full Text 
xml version 1.0 encoding UTF8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchemainstance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd leader nam 2200397Ka 4500 controlfield tag 001 002046419 005 20100106154242.0 007 cr mnuuuuuu 008 100106s2008 flu s 000 0 eng d datafield ind1 8 ind2 024 subfield code a E14SFE0002724 035 (OCoLC)496021936 040 FHM c FHM 049 FHMM 090 G116 (Online) 1 100 Williams, Marilyn Marie. 0 245 Linking health hazards and environmental justice : b a case study in Houston, Texas h [electronic resource] / by Marilyn Marie Williams. 260 [Tampa, Fla] : University of South Florida, 2008. 500 Title from PDF of title page. Document formatted into pages; contains 264 pages. Includes vita. 502 Dissertation (Ph.D.)University of South Florida, 2008. 504 Includes bibliographical references. 516 Text (Electronic dissertation) in PDF format. 520 ABSTRACT: This dissertation seeks to extend quantitative research on environmental justice and address methodological limitations of previous studies by: (a) using new indicators of exposure to air pollution and contemporary risk modeling techniques; (b) assessing disparities in human health risks, instead of focusing only on potential exposure or proximity to pollution sources; and (c) using multivariate regression models that consider the effects of spatial dependence. The case study examines racial/ethnic and socioeconomic disparities in the geographic distribution of exposure to airborne toxic emissions from industrial point sources in the HoustonGalvestonBrazoria metropolitan statistical area. Industrial pollution sources for this study comprise facilities listed in the US EPA's Toxic Release Inventory (TRI).The RiskScreening Environmental Indicator (RSEI) model is used to estimate potential human health risks from air pollutants based on data on toxicity and dispersion of chemical releases from TRI facilities. The analyses utilize four indicators of potential exposure to industrial pollution: (a) presence or absence of air emissions, (b) total quantities (pounds) of air emissions, (c) toxicityweighted quantities of emissions and (d) modeled risk scores based on the cumulative health risk posed by air emissions. Traditional linear regression and spatial autoregressive techniques based on several neighborhood configurations are used to model the occurrence and magnitude of these four indicators, using relevant explanatory variables from the 2000 census, at the census tract and block groups levels of aggregation. Results indicate a disproportionate pattern of health risks from TRI facilities in the HGBMSA, with the Hispanic population facing the highest exposure.The locations and magnitude of toxic pollution are significantly statistical effected by the presence of minority residents and population density. Additionally, key differences in the significance of explanatory variables between the spatial and conventional regression models demonstrate the importance of correcting for spatial dependence in environmental justice analysis. The analytical results for several variables are also sensitive to the choice of data resolution (tract or block group). Overall, this study indicates that more flexible spatial analytic techniques are required to improve the identification of environmental injustice and past studies utilizing conventional statistical methods should be revisited to explicitly account for spatial effects. 538 Mode of access: World Wide Web. System requirements: World Wide Web browser and PDF reader. 590 Advisor: Jayajit Chakraborty, Ph.D. 653 Environmental justice Pollution GIS Spatial statistics Health disparities 690 Dissertations, Academic z USF x Geography Doctoral. 773 t USF Electronic Theses and Dissertations. 4 856 u http://digital.lib.usf.edu/?e14.2724 PAGE 1 Linking Health Hazards and Environmental Justice: A Case Study in Houston, Texas by Marilyn Marie Williams A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy Department of Geography Co llege of Arts and Sciences University of South Florida Major Professor: Jayajit Chakraborty, Ph.D M. Martin Bosman, Ph.D Thomas Mason, Ph.D Steven Reader, Ph .D Graham A. Tobin, Ph.D Date of Approval: November 19, 2008 Keywords: Environmental Justice, Po llution, GIS Spatial Statistics, Health Disparities Copyright 2008, Marilyn M. Williams PAGE 2 i Table of Contents List of Tables iii Lis t of Figures x List of Terms xi Linking Health Hazards and Environmental Justice: A Case Study in Houston, Texas xii 1. Introduction 1 1.1 Types of Exposure 2 1.2 Environmental Justice and Hea lth Disparities Research: Trends and Limitations 4 1.3 Research Objectives 6 2. Literature Review 8 2.1 Theoretical Explanations for Spatial and Environmental Justice 8 2.1.1 Theories of Justice 9 2.1.2 Alternative Explanations 11 2.1.3 Contemporary Explanations of Environmental Justice 13 2.1.4 Geographical Principles of Environmental Justice 18 2.2 Methodology for Environmental Justice Assessment 20 2.2.1 Geographical Unit of Analysis 20 2.2.2 Geographic Problems and Spatial Effects 24 2.2.3 Spatial Analytical Approaches to Identifying Impacted Areas 30 2.3 Environmental Health Disparities 35 2.3.1 Influential Variables on Health Disparities 37 2.3.2 Measures of Health Disparities 40 2.3.3 Effects of Environmental Health Disparities 44 3. Study Area 48 3.1 Racial, Economic, and Industrial Context 50 3.2 Previous Environmental Justice Studies in the Houston Area 53 3.3 Suitability of the Study Area for Proposed Research 54 4. Data Sources, Models, and Methods 56 4.1 Data Sources for Pollution Exposure and Health Risk Assessment 56 4.1.1 Chronic Exposure 57 4.2 Populations at Risk 61 4.3 Methodology for Statistical Analyses 64 4.6 Presentation of Statistical Analysis Results 70 5. Locational Analysis of Industrial Pollution Sources 72 5.1 Census Tract Level Analysis 76 5.2 Block Group Level Analysis 78 5.3 Discussion and Interpretation 80 6. Analyzing the Quantity of Emissions from Industrial Pollution Sources 93 6.1 Census Tract Level Analysis 101 6.2 Block Group Level Analysis 106 6.3 Discussion and Interpretation 111 7. Analyzing the Toxicity of Emissions from Industrial Pollution Sources 139 7.1 Census Tract Level Analysis 148 PAGE 3 ii 7.2 Block Group Level Analysis 153 7.3 Discussion and Interpretation 158 8. Analyzing the Health Risks of Emissions from Industrial Pollution Sources 188 8.1 Census Tract Level Analysis 197 8.2 Block Group Level Analysis 202 8.3 Discussion and Interpretation 207 9. Conclusions 239 9.1 Evaluating Inequities for the Four Pollution Indicators 241 9.2 Effect of Regression Methodology and Analytical Unit Selection 243 9.3 Geographies of Environmental Justice 247 9.4 Public Health Implications 249 9.5 Limitations and Future Research 250 9.6 Policy Recommendations 252 Literature Ci ted 254 About the Author End Page PAGE 4 iii LIST OF TABLES Table 1: Government definitions of hea lth disparities 36 Table 2: Census Year 2000 HoustonGalveston Brazoria, TX CMSA Sociodemographic and Socioeconomic Status 49 Table 3: Industries in the HGB in 2000 51 Table 4: Selected Studies Examining Aspects of E nvironmental Justice in Houston, Texas 54 Table 5: RSEI Indicator Element mode 59 Table 6: Descriptive Statistics for Variables at the Census Tract Level 83 Table 7: Descriptive Statistics for Variables at the Block Group Level 84 Table 8: Pearson Correlation Coefficients at the Census Tract Level Associated with the Square Root of TRI Frequencies 85 Table 9: Pearson Cor relation Coefficients at the Block Group Level Associated with the Square Root of TRI Frequencies 86 Tables 10: Logistic Regression Analysis of TRI Location at Zero Mile from Census Tract Boundary 87 Tables 11: Logistic Regression Analysis of TRI Locatio n at Half Mile from Census Tract Boundary 88 Tables 12: Logistic Regression Analysis of TRI Location at One Mile from Census Tract Boundary 89 Tables 13: Logistic Regression Analysis of TRI Location at Zero Mile from Block Group Boundary 90 Tables 14: Log istic Regression Analysis of TRI Location at Half Mile from Block Group Boundary 91 Tables 15: Logistic Regression Analysis of TRI Location at One Mile from Block Group Boundary 92 Table 16: Descriptive Statistics for Variables at the Census Tract Level 114 Table 17: Descriptive Statistics for Variables at the Block Group Level 115 PAGE 5 iv Table 18: Pearson Correlation Coefficients Associated with the Density of Pounds (Natural Log) Released at the Census Tract Level 116 Table 19: Pearson Correlation Coefficients A ssociated with the Density of Pounds (Natural Log) Released at the Block Group Level 117 Table 20: Least Squares Regression Results for Density of Pounds Released (Natural Log) at Zero Miles from Census Tract Boundary 118 Table 21: Least Squares Regression Results for Density of Pounds Released (Natural Log) at Half Mile from the Census Tract Boundary 119 Table 22: Least Squares Regression Results for Density of Pounds Released (Natural Log) at One Mile from the Census Tr act Boundary 120 Table 23: Spatial R egression Results for Density of Pounds Released (Natural Log) at Zero Mile of Census Tract Boundary Using Que en Contiguity 121 Table 24: Spatial Regression Results for Density of Pounds Released (Natural Log) at Zero Mile of Census Tract Boundary Using the R ook Contiguity 122 Table 25: Spatial Regression Results for Density of Pounds Released (Natural Log) at Half Mile of Census Tract Boundary Using the Queen C ontiguity 123 Table 26: Spatial Regression Results for Density of Pounds Released (Natural Log) at Half Mile of Census Tract Boundary Using the R ook Contiguity 124 Table 27: Spatial Regression Results for Density of Pounds Released (Natural Log) at One Mile of Census Tract Boundary Using the Queen Contiguity 125 Table 28: Spatial Regression Results for Density of Pounds Released (Natural Log) at One Mile of Census Tract Boundary Using the Rook Contiguity 126 Table 29: Spatial Regression Results for Density of Pounds Released (Natural Log) at One Mile of Census Tract Boundary Using Distan ce 127 Table 30: Least Squares Regression Results for Density of Pounds Released (Natural Log) at Zero Mile from Block Group Boundary 128 Table 31: Least Squares Regression Results for Density of Pounds Released (Natural Log) at Half Mile from Block Group Boundary 129 Table 32: Least Squares Regression Results for Density of Pounds Released (Natural Log) at One Mile from Block Group Boundary 130 PAGE 6 v Table 33: Spatial Regression Results for Density of Pounds Released (Natural Log) at Zero Miles from Block Group Boundary Usin g t he Queen Contiguity 131 Table 34: Spatial Regression Results for Density of Pounds Released (Natural Log) at Zero Miles from Block Group Boundary Using the Rook Contiguity 132 Table 35: Spatial Regression Results for Density of Pounds Released (Natural Log) at Half Mile from Block Group Boundary Using th e Queen Contiguity 133 Table 36: Spatial Regression Results for Density of Pounds Released (Natural Log) at Half Mile from Block Group Boundary Using the Rook Contiguity 134 Table 37: Spatial Regression R esults for Density of Pounds Released (Natural Log) at Half Mile from Block Group Boundary Using Dist ance 135 Table 38: Spatial Regression Results for Density of Pounds Released (Natural Log) at One Mile from Block Group Boundary Using th e Queen Contiguity 136 Table 39: Spatial Regression Results for Density of Pounds Released (Natural Log) at One Mile from Block Group Boundary Using th e Rook Contiguity 137 Table 40: Spatial Regression Results for Density of Pounds Released (Natural Log) at One Mile from Block Group Boundary Using Dis tance 137 Table 41: Descriptive Statistics for Variables at the Census Tract Level Associated with Hazard S cores 162 Table 42: Descriptive Statistics for Variables at the Block Group Level Associated with Hazard Scores 163 Tab le 43: Pearson Correlation Coefficients Associated with the Density of Hazard Scores (Natural Log) at the Census Tract Level 164 Table 44: Pearson Correlation Coefficients Associated with the Density of Hazard Scores (N atural Log) at the Block Group L evel 165 Table 45: Least Squares Regression Results for Density of Hazard Scores (Natural Log) at Zero Miles of Census Tract Boundary 166 Table 46: Least Squares Regression Results for Density of Hazard Scores (Natural Log) at Half Mile from the Census Tra ct Bo undary 167 Table 47: Least Squares Regression Results for Density of Hazard Scores (Natural Log) at One Mile from the Census Tract Boundary 168 PAGE 7 vi Table 48: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Zero Mile from Census Tract B oundary U sing Queen Contiguity 169 Table 49: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Zero Mile from Census Tract Boundary Using Rook Contiguity 170 Table 50: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Half Mile from Census Tract Boundary Using Queen Contiguity 171 Table 51: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Half Mile from Census Tract Boundary Using Rook Contiguity 172 Table 52: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Half Mile from Census Tract Boundary Using Dist ance Band 173 Table 53: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Census Tract Boundary Related U sing Queen Contigui ty 174 Table 54: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Census Tract Boundary Using Rook Contiguity 175 Table 55: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Census Tract Boundary Using Dist ance Band 176 Table 56: Least Squares Regression Results for Density of Hazard Scores (Natural Log) at Zero Mile from the Block Gro up Boundary 177 Table 57: Least Squares Regression Results for Density of Hazard Scores (Natural Log ) at Half Mile from the Block Group Boundary 178 Table 58: Least Squares Regression Results for Density of Hazard Scores (Natural Log) at One Mile from the Block Gro up Boundary 179 Table 59: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Zero Miles from Block Group Boundary Using Queen Contiguity 180 Table 60: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Zero Miles from Block Group Boundary Using Rook Contiguity 181 Table 61: Spatial Regression Results f or Density of Hazard Scores (Natural Log) at Half Mile from Block Group Boundary Using Queen Contiguity 182 Table 62: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Half Mile from Block Group Boundary Using Rook Contiguity 183 PAGE 8 vii Tab le 63: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Half Mile from Block Group Boundary Using Dista nce Band 184 Table 64: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Block Group Bounda ry Using Queen Contiguity 185 Table 65: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Block Group Boundary Using Rook Contiguity 186 Table 66: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Block Group Boundary Using Distance Band 187 Table 67: Descriptive Statistics for Variables at the Census Tract Level 211 Table 68: Descriptive Statistics for Variables at the Block Group Level 212 Table 69: Pearson Correlation Coefficients Associated with the Density of Risk Scores (Natural Log) Released at the Census Tract Level 213 Table 70: Pearson Correlation Coefficients Associated with the Density of Risk Scores (Natural Log) Released at the Block Group Level 214 Table 71: Least Squa res Regression Results for Density of Risk Scores (Natural Log) at Zero Mile from Census Tract Boundary 215 Table 72: Least Squares Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Census Tract Boundary 216 Table 73: Least Squa res Regression Results for Density of Risk Scores (Natural Log) at OneMile from Census Tract Boundary 217 Table 74: Spatial Regression Results for Density of Risk Scores (Natural Log) at Zero Mile from Census Tract Boundary Using Queen Contiguity 218 Tabl e 75: Spatial Regression Results for Density of Risk Scores (Natural Log) at Zero Mile from Census Tract Boundary Using Rook Contiguit y 219 Table 76: Spatial Regression Results for Density of Risk Scores (Natural Log) at Zero from Census Tract Boundary Usi ng Distance B and 220 Table 77: Spatial Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Census Tract Boundary Using Queen Contiguity 221 Table 78: Spatial Regression Results for Density of Risk Scores (Natural Log) at Half Mi le from Census Tract Boundary Using Rook Contiguity 222 PAGE 9 viii Table 79: Spatial Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Census Tract B oundary Using Distance Band 223 Table 80: Spatial Regression Results for Density of Risk Scores (Natural Log) at One Mile from Census Tract Boundary Using Q ueen Contiguity 224 Table 81: Spatial Regression Results for Density of Risk Scores (Natural Log) at One Mile from Census Tract Boundary Using Rook Contiguity 225 Table 82: Spatial Regressi on Results for Density of Risk Scores (Natural Log) at One Mile from Census Tract Boundary Using Distance Band 226 Table 83: Least Squares Regression Results for Density of Risk Scores (Natural Log) at Zero Mile from Block Group Boundary 227 Table 84: Leas t Squares Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Block Group Boundary 228 Table 85: Least Squares Regression Results for Density of Risk Scores (Natural Log) at One Mile from Block Group Boundary 229 Table 86: Spatial Regression Results for Density of Risk Scores (Natural Log) at Zero Miles from Block Group Boundary Using Queen Contiguity 230 Table 87: Spatial Regression Results for Density of Risk Scores (Natural Log) at Zero Miles from Block Group Boundary Using Rook Contiguity 231 Table 88: Spatial Regression Results for Density of Risk Scores (Natural Log) at Zero Miles from Block Group Boundary Using Distance Band 232 Table 89: Spatial Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Block Group Boundary Using Queen Contig uity 233 Table 90: Spatial Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Block Group Boundary Using R ook Contiguity 234 Table 91: Spatial Regression Results for Density of Risk Scores (Na tural Log) at Half Mile from Block Group Boundary Using Distance Band 235 Table 92: Spatial Regression Results for Density of Risk Scores (Natural Log) at One Mile from Block Group Boundary Using Q ueen Contiguity 236 Table 93: Spatial Regression Results for Density of Risk Scores (Natural Log) at One Mile from Block Group Boundary Using R ook Contiguity 237 PAGE 10 ix Table 94: Spatial Regression Results for Density of Risk Scores (Natural Log) at One Mile from Block Group Boundary Using Distance B and 238 PAGE 11 x LIST OF FI GURES Figure 1: Geographical Hierarchy for the D ecennial Census 21 Figure 2: Heavy gas dispersion from gravity 33 Figure 3: Gaussian distribution and spread of chemical from source 34 Figure 4: Schematic of exposure and eff ects (outcomes) 45 Figure 5: Ho uston Galveston Brazoria Consolidated Met ropolitan Statistical Area 48 Figure 6: Locations of Toxic Release Inventory F acilities in the HGB MSA 51 Figure 7: Organizational Char t of Study 69 Figure 8 : C ens us tract connectivity histograms used for pounds em itted 98 Figure 9: B lock group connectivity histograms used for pounds emitted 99 Figure 10: Census tract connectivity histograms used for hazard scores 145 Figure 11: Block group connectivity histograms used for hazard scores 146 Figure 12 : C ensus tra ct connectivity histograms used for risk scores 194 Figure 13 : Block group connectivity histograms used for risk scores 195 PAGE 12 xi LIST OF TERMS Agency for Toxic Substances and Disease Registry (ATSDR) Areal Locations of Hazardous Atmospheres (ALOHA) Centers for Disease Control and Prevention (CDC) Emergency Planning and Community Right to Know Act (EPCRA) Federal Highway Administration (FHWA) Hazardous Air Pollutants (HAPs) Hazardous Materials Information System (HMIS) Hazardous Substances Emergency Events S urveillance (HSEES) Health Effects Assessment Summary Tables (HEAST) Integrated Risk Information System (IRIS) National Environmental Protect Act (NEPA) National Health and Nutrition Examination Survey (NHANES) National Oceanic and Atmospheric Administration (NOAA) National Response Center (NRC) National Scale Air Toxics Assessment (NATA) Pipeline and Hazardous Materials Safety Administration (PHMSA) Resource Conservation and Recovery Act (RCRA) Risk Screening Environmental Indicators for Chronic Human He alth Model (RSEI) Toxic Release Inventory Site (TRI) U.S. Coast Guard's National Response Center (NRC) U.S. Department of Transportation (DOT) U.S. Environmental Protection Agency (USEPA PAGE 13 xii LINKING HEALTH HAZAR DS AND ENVIRONMENTAL JUSTICE: A CASE STUDY IN HOUSTON, TEXAS Marilyn M. Williams ABSTRACT This dissertation seeks to extend quantitative research on environmental justice and address methodological limitations of previous studies by: (a) using new indicators of exposure to air pollution and contempor ary risk modeling techniques; (b) assessing disparities in human health risks, instead of focusing only on potential exposure or proximity to pollution sources; and (c) using multivariate regression models that consider the effects of spatial dependence. T h e case study examines racial/ethnic and socioeconomic disparities in the geographic distribution of exposure to airborne toxic emissions from industrial point sources in the HoustonGalveston Brazoria metropolitan statistical area Industrial pollution s ources for this study comprise facilities listed in the US EPAs Toxic Release Inventory (TRI). The Risk Screening Environmental Indicator (RSEI) model is used to estimate potential human health risks from air pollutants based on data on toxicity and dispersion of chemical releases from TRI facilities The analys es utilize four indicators of potential exposure to industrial pollution : (a) presence or absence of air emissions, (b) total quantities ( pounds) of air emissions, (c) toxicity weighted quantit ies of emissions and (d) modeled risk score s based on the cumulative healt h risk posed by air emissions. Traditional linear regression and spatial autoregressive techniques based on several neighborhood configurations are used to model the occurrence and magn itude of the se four indicators, using relevant explanatory PAGE 14 xiii variables from the 2000 census at the census tract and block groups levels of aggregation. Results indicate a disproportionate pattern of health risks from TRI facilities in the HGB MSA, with the Hispanic population facing the highest exposure. T he locations and magnitude of toxic pollution are significantly statistical effect e d by the presence of minority residents and population density Additionally key differences in the significance of explanatory variables between the spatial and conventional regression models demonstrate the importance of correcting for spatial dependence in environmental justice analysis. The analytical results for several variables are also sensitive t o the choice of dat a resolution (tract or block group) Overall this study indicates that more flexible spatial analytic techniques are required to improve the identification of environmental injustice and past studies utilizing conventional statistical methods should be r evisited to explicitly account for spatial effects . PAGE 15 1 1. INTRODUCTION The use and production of industrial chemicals has increased significantly in the United States since the Industrial Revolution, resulting in an increase of potential exposure and health r isks to the population. Racial and socioeconomic inequities associated with the distribution of pollution burdens and related health impacts sparked the environmental justice movement in the 1980s. The goal of environmental justice or environmental equit y is to ensure that all people, regardless of race, national origin or income, are protected from disproportionate impacts of environmental hazards (Clinton 1994) Spatial inequities in the distribution of toxic pollution in the US was first reported by a study conducted by the General Accounting Office (GAO) in 1983, which found the location pattern of hazardous waste landfills to be inequitable with respect to race and income in eight southeastern states (US General Accounting Office 1983) This study followed on the heels of resident protests in Warren County, North Carolina, against the location of a proposed landfill for polychlorinated biphenyl (a known carcinogen) in a predominantly Black neighborhood. These protests attracted substantial media and congressional attention and became a pivotal event for the environmental justice movement. Following the Warren County protests, the United Church of Christ (UCC) conducted the first comprehensive national scale stu dy on environmental justice (United Church of Christ Commission for Racial Justice 1987) The UCC report was significant because it found race to be the most potent variable in predicting where commercial hazardous waste facilities were located in the US. In addition to galvanizing PAGE 16 2 the attention of the government and the citizens to the occurrence of environmental injustice, the UCC report triggered a series of empirical case studies that investigated the association between environmental risks and potentiall y affected populations, based on race/ethnicity and economic status. There is a growing literature which suggests that industries that store, use, manufacture, and release toxic chemicals are disproportionately located in greater densities in urban and m inority spaces as compared to suburban and nonminority areas (Mohai and Bryant 1992; Liu 2001) This location pattern may result in greater exposure of industrial chemicals to surrounding residents and consequently imposes an uneven health risk upon the affected population. This increased exposure has created a new environmental health burden on society where communities are exposed to a multitude of toxic chemicals on a continuous basis. Studies have found that the impact of sustained atmospheric, water and land exposure may re sult in adverse health effects. These health effects can cause increased morbidity and mortality including cancers, renal and liver diseases (Gee and PayneSturges 2004) as well as low birth weight babies (Dolk and Vrijheid 2003) 1.1 Types of Exposure There are two distinct types of chemical exposures identified in toxicological studies, namely, acute and chronic. Chronic exposures occur as a result of long term exposures to toxic chemicals and routine emissions from facilities such as industrial manufacturing plants, toxic waste disp osal sites, and landfills. Chronic exposures also can result in lifetime adverse health effects which have been known to increase the incidence of low birth weight, birth defects and asthma (Gilboa and Fixler 2005; Morello PAGE 17 3 Frosch and Shenassa 2006) As the siting of industries has progressed in residential neighborhoods and as neighborhoods have encroached on the locations of some of the most toxic spa ces in the US, an increase of chronic exposure and adverse health effects have occurred. Adverse health outcomes can be exacerbated by prolonged exposure to a single chemical and interaction with multiple chemical exposures potentially resulting in greate r morbidity and mortality (Klassen 2001) Acute pollution events are nonroutine, accidental releases of hazardous substances that have immediate public health consequences for the affected population (Chakraborty 2001) Unlike chronic hazards, acute hazards are short term, unintentional spills that frequently lead to injury, fatalities, and adverse health outcomes which are known to range from minor respiratory irritation to protra cted and progressively debilitating diseases, such as cancers (Gee and Payne Sturges 2004) Acute exposures can result from unexpected releases from railway cars, fuel truck, and fixed facilities. Increased morbidity and mort ality rates have been known to result from these releases. These exposures can be additive and synergistic to chronic exposures and as such there is a difficulty collecting data and determining independent effects on the affected population. Measurement o f adverse health outcomes requires the implementation of longitudinal epidemiological studies. The lack of sufficient data collection supports the need of adequate databases that contain information collected for extended periods of time (Payne Sturges and Gee 2006) Data gathering of acute health outcomes requires an established record keeping system which tracks the type of chemical exposure as well as demographics and resultant health outcomes. Collection of data on chronic heal th PAGE 18 4 outcomes differs from those on acute events in that it requires many years for the onset of the disease or death to occur. Therefore the modeled risk to the population is used to examine potential health impacts of chronic exposure while acute health outcomes data are collected over a short period of time. These two types of health outcomes have been observed to occur disproportionately in minority and low income neighborhoods resulting in environmental health disparities (Pastor Jr et al. 2005; MorelloFrosch and Lopez 2006) 1.2 Environmental Justice and Health Disparities Research: Trends and Limitations Environm ental health disparities are defined as racial/ethnic and socioeconomic inequities in illness, exposure and outcomes which are influenced by the biophysical, social and built environments (Payne Sturges and Gee 2006) Such disparitie s have arisen as a result of racial/ethnic minorities and persons of low er socioeconomic status living in spaces exposed to environmental pollution and related hazards (Morello Frosch and Lopez 2006) The study of the distribution and extent of environmental health disparities has recently emerged as a discipline which extends environmental justice research (Institute of Medicine 1999; Woodruff et al. 2003; Payne Sturges and Gee 2006) Studies have begun to explore the association between the social construction of health disparities and neighborhood locations, with respect to minority and/or low income residents in urban areas. Adverse health outcomes and patterns of population locations are now being evaluated through geospatial tools that not only provide spatial representations of affected populations but also can be used to identify the social dynamics which influence racial/ethnic and economic disparities. PAGE 19 5 Previous empirical research on environmental justice and health disparities has been plagued by several methodological limitations. First, a majority of studies have evaluated racial and economic inequities in the distribution of pollution sources such as hazardous waste facilities and industrial manufacturing plants based solely on their locations (United Church of Christ Commission for Racial Justice 1987; Mohai and Bryant 1992; Cutter and Clark 1996; Perlin and Wong 1999; Mohai and Saha 2006) thus ignoring the nature quantity, or properties of the emitted pollutants and other factors that influence exposure. Second, although some studies have incorporated data on quantity and toxicity of released pollutants and local meteorological conditions to assess potential exp osure to toxic emissions (Chakraborty and Armstrong 2004; Dolinoy and Miranda 2004) few studies have attempted to l ink disproportionate environmental exposure to their adverse health effects or risks (Maantay 2002) Third, m ost studies have used conventional statistical methods such as linear regress ion analysis to determine th e effect of racial/ethnic and socioeconomic influences on response variables that represent exposure to toxic pollution. Previous research on environmental justice has thus failed to adequately account for spatial processes and effects that are an inheren t feature of geographically referenced data sets In order to address these methodological limitations, this dissertation focuses on assessing chronic expos ure to toxic atmospheric pollution sources within a single US metropolitan area on the basis of ne wly emerging risk modeling and spatial statistical techniques The study area is the HoustonGalveston Brazoria (HGB) consolidated metropolitan statistical area, where several prominent studies on racial inequities PAGE 20 6 associated with hazardous waste facility siting have been conducted (Bullard 1983; Bullard 1990) 1.3 Research Objectives The broader aims of this dissertation are to extend quantitative research on environmental justice and address several meth odological limitations of previous studies, by : 1) using new indicators of exposure to air pollution and contemporary risk modeling technique s ; 2) assessing disparities in human health risks instead of focusing only on potential exposure or proximity to pollution sources ; and 3) i mplementing multivariate r egression methods that consider the effect of spatial dependence and are thus more suitable for analyzing geographic data The data sets used in the dissertation project for estimating exposure and health risks have not previously been applied in environmental justice analysis and require d new methodological avenues to be explored. Fixed facility point sources of i ndustrial toxic emissions in the study area were used to estimate chronic or long term exposure to outdoor air pollution in this dissertation Specific research questions posed to examine the geographic distribution of chronic industrial atmospheric pollution sources and health disparities in the study area are summarized as follows: a. What is the association between locations of industrial atmospheric pollution sources and the racial/ethnic and socioeconomic status of the population? b. What is the association between quantity (pounds) of pollutants released from industrial pollution sources and t he racial/ethnic and socioeconomic status of the population? PAGE 21 7 c. What is the association between toxicity weighted quantit ies (hazard scores ) of emissions from industrial pollution sources and the racial/ethnic and socioeconomic status of the population? d. What is the association between health risks (modeled risk scores) from chronic exposure to industrial pollutants and the racial/ethnic and socioeconomic status of the population? Th is dissertation is organized into nine chapters. The Introduction is followed by a comprehensive review of the pertinent literature in chapter 2, which includes the theoretical foundations for the research questions and an overview of geospatial techniques and methodologies used in this study. Chapter 3 provides background informat ion on the study area and chapter 4 describes the data sources, models, and statistical technique s used in this research. The empirical findings associated with each research question are presented in the following four chapters (chapters 5 to 8), while c hapter 9 summarizes the conclusions and broader significance of this work. PAGE 22 8 2. LITERATURE REVIEW This chapter provides a review of the environmental justice and health disparities research literature. It encompasses relevant theories as well as methodolog ies used in prior studies to explain, document, and analyze disparit ies o n the spatial distribution and impact of toxic pollution. The first section of this chapter outlines a selection of common and contemporary theories used to examine the causes of environmental injustice and inequity. The second section describes analytical tools and methodologies that have been used in environmental justice research, while the third section explores the analysis and measurement of health disparities. The fo urth section explains how disease clusters are defined and outlines several statistical tools that can be used to analyze the spatial distribution of disease health events to determine the presence of statistically significant disease clusters. 2.1 Theoretical Exp lanations for Spatial and Environmental Justice A large variety of theories have been proposed to explain and substantiate the occurrence of injust ices. Sections one and two outlines the common theories used to address social justice, as well as several alternativ e theories. Section three offers more contemporary theories and explanations of environmental justice. Section four define s how and why these t heories and explanations were used in relation to geographical principles of environmental justice. PAGE 23 9 Th eories developed to interpret the causes and consequences of environmental injustice and health disparities include social justice, economic residential and location theories, as well as theories of neighborhood change. Mainstream theories of justice and equity have been identified as utilitarianism, contractarianism, egalitarianism, and libertarianism. These theories can be expanded upon by the principles of Marxism and feminism in order to dispel the rigidity brought to the concept of justice through the mainstream theories. Additionally, specific environmental justice explanations such as the neighborhood life cycle model, has brought to light new ideas that can be used to examine environmental justice issues. 2.1.1 Theories of Justice The four mainstre am theories of justice and equity have been identified as utilitarianism, contractarianism, egalitarianism and libertarianism. These theories are representative of distributive justice through use of goods and services. Egalitarianism is the equal distri bution of justice and treatment for individuals (Smith 1994) This theory hypothesizes that people should receive equal inputs (e.g., education, resources) and have equal outputs (e.g., monies, social welfare), regardless of their geographical location. There are several problems with strict egalitarian principles in that: 1) although people may receive the same education they may not receive the same jobs or monies once they graduate; 2) geographical locations inherently have different resources available; and 3) redistribution of goods from the wealthiest group of people may not be feasible or practical. Another theory of distributive justice is utilitarianism. Utilitarianism states that goods should be produced and distributed to maximize the total welfare or aggregate PAGE 24 10 social utility (Smith 1994; Rawls 1999; Liu 2001) Implementation of utilitarian policy ignores individual inequity and would distribute justice over the aggregate, thereby ignoring the effe cts on individuals. This theory is not concerned with equity and distributive justice since the goal is the aggregate effect and not individuals who are in greatest need. In essence, utilitarianism could increase the existing inequity which is the polar opposite of the goal of egalitarianism. Contractarianism uses a social contract where individuals would distribute goods equally unless an unequal distribution of any or all goods is to the advantage of the least favored (Smi th 1994; Rawls 1999) This theory of justice as represented by Rawls states that 1) each person is to have an equal right to the most extensive basic liberty compatible with a similar liberty for others; and 2) social and economic inequalities are to be arranged so that they are both: a) to the greatest benefit of the least advantaged, and b) attached to offices and positions open to all under conditions of fair equality of opportunity (Rawls 1980, 1999) This theory, like the egalitarian theory, strives to make the least well off in society as well off as possible, but unlike the egalitarian approach it does not strive to eliminate the inequality. Therefo re, Rawlsian theory is seen as the principle theory of fairness and not equality. Libertarian ism stresses the freedom of the individual to perform whatever task he/she desire as long as he/she respects the rights of others to do the same. The state, (i. e., government), is only allowed to interfere if there is aggression or fraud. Libertarianism emphasizes free choice, free market economy, and property rights of the individual, but it does not provide avenues of conflict resolution when dealing with poll ution or other justice issues that involve large groups of people (Liu 2001) PAGE 25 11 Therefore, this theory would have limited ability to resolve environmental justice issues since these issues inherently involve large groups of people who expe rience varying levels of injustice. 2.1.2 Alternative Explanations Common social justice theories are limited in their ability to help explain the intricacies of why environmental injustices may occur. Further explanations have been suggested and have be come alternative pathways used to broaden justice theory. Two additional ideas to the more established theories are Marxism and feminism, which are summarized below. Marxism : The Marxist view of justice is anticapitalist, and as such it expresses opposit e views of what influences justice in most developed countries. Karl Marx focused on two particular avenues of justice, namely alienation and exploitation. He used these ideals as explanations of how production in a capitalist economy (i.e., means and mo de of production), contribute to injustices and inequalities among workers. Marx identifies alienation as the ability of the worker in a capitalist society to be alienated from their most important capacity thereby denying the worker the ability to realize his/her own potential. The alienation denies the worker any creativity and thus inhibits true fulfillment. Although fulfillment may be a goal of a worker as viewed by this theory, it is not necessarily unjust if a person does not feel that they are unf ulfilled. Smith (1994) recognizes that a person can perform a job which is monotonous, but provides significant monetary compensation therefore allowing them to be fulfilled in other aspects of their lives outside of their job. Consequently, the evaluation of exploitation may be a better PAGE 26 12 choice as a representation in Marxism as it pertains to explanations of environmental justice. Marxists view capitalist exploitation as the ability of the persons who possess ownership to receive a greater value from the production of the workers than is given to the workers. In essence, the product produced by the worker is of greater value than the worker is compensated and therefore the owner receives the surplus as profit. A more contempora ry view of this Marxist principle is the unfair advantage within broader distribution arrangements which include uneven access to resources in general and to the means of production in particular (Smith 1994) This defi nition of exploitation as part of justice issues, particularly environmental justice, helps to explain the importance of equal access to resources and uneven distribution of externalities from toxic pollution. For example, by having locally unwanted land uses (LULUs) concentrated in impoverished or minority neighborhoods, the pollution caused by these LULUs denies the residents of these neighborhoods access to clean air as a resource and consequently denies them equal access to good health, thus representi ng this principle of Marxist theory. Feminism : Feminism explores the use of gender identity as an influence on justice. It explores the inequalities experienced by women due to the paternalistic view that shapes society. Understanding the impact of how societal structure influences where people live, the jobs they receive and how much they earn are important in identifying the basic principles of justice. Young (1990) proposes that distributive justice, whic h includes distribution of benefits and burdens among society, limits social justice because it does not explore the institutional or social context where these burdens were created. The PAGE 27 13 contemporary framework of theories of justice, as represented by Raw ls (contractarian), utilitarian, egalitarian and libertarian principles, as well as Marxism, are based on principles of distributive justice. These theories emphasize the importance of distribution of goods and worker related production, but not decision making processes or exploration of the social and institutional context which determines the distribution of these goods. Marxism, for example, views exploitation in the sense of the worker only but does not examine exploitation of women or individuals by race (Young 1990; Smith 1994) In exploring siting of LULUs, Young (1983) argues that the prevailing approaches to reasoning about justice cannot treat justice used for decision making of where pollution sources should be sited. Justice does not only include the distribution of facilities but participation in the process of choosing where to place such facilities. Young states that the utilitarian argument for siting pollution sources, which consists of placing the facilities in one area, would provide the greatest benefit to most people, but does not take into account fairness as it effects distributive justice. Although Rawl sian theory of justice is viewed as representing fairness as justice and distributive justice, in particular, it does not outline the decision making processes. This theory neglects the rights of the potentially affected population to participate in the s iting process, which denies them power, thereby denying them justice. 2.1.3 Contemporary Explanations of Environmental Justice While mainstream social, moral, ethical and political theories were outlined in the previous section, more specific theories he lp to identify and explain the concentration of pollution sources in minority and low income neighborhoods. PAGE 28 14 2.1.3.1 Theories of Economics and Location Economics can play an important role in determining how justice is distributed. Pareto efficiency is a s tate where no one can be made better off without making someone worse off, and therefore to achieve efficiency there should be no unexploited gains to trade, no unexploited way of increasing output with the same level of inputs and no mix of products that do not reflect the preferences of the consumer (Liu 2001) Environmental problems such as environmental inequity and environmental injustice are produced as externalities of processes of industrial pollution and therefore do not achieve efficiency. Industrial pollution can result in externalities that require a higher cost to be paid by communities which are in vicinity to the industries. These negative externalities resulting from air, land, and water pollution can lead to adverse and disproportionate health effects as referenced in EO 12898 (Clinton 1994) 2.1.3.1.1 Industrial Location Location theory for industries in the US can be explored using a combination of frameworks. Least cost is identified as a theory where industrial location decisions are made based upon labor and transport costs and Hotelling (1990) outlined that location of industries was influenced by spatial competition and oligopolistic (small number of sellers of a single product, e.g., airlines, personal computers) (Liu 2001) In essence, industries locate near areas which minimize the cost of transport, whether the cost is from direct access to raw goods or waste disposal sites. These theories are represented by the agglomeration of a variety of industries and businesses in close proximity to each other and close proximity to transportation routes. The choice of industry locations has also been found to include land prices which are known to vary dependent upon the distance PAGE 29 15 from the city center. Additionally, researchers have stressed that i ndustry locations follow the path of least resistance and are thus deliberately placed in areas where land prices are low, labor is inexpensive, and residents have little political influence to resist siting (Mohai and Bryant 1992) 2.1.3.1.2 Residential Loca tion Von Thunen developed the theory of agricultural landuse where land is distributed according to an individuals willingness to bid. Initially house location theory proposed that locations were determined by accessibility and space. Later it was expand ed to include environmental amenities to the residential population (Fujita 1989) Kanemoto (1987) found that land rent increases at the boundary between industrial and residential zones. In essence, if pollution zones expand or encroach on neighborhoods then the following items hold true: 1) The residential land and/or rent will decrease allowing for an increase in industrial purchasing of inexpensive land. This can result in an increase in negative impact on the population in the neighborhood and can force people to relocate. 2) More low income individuals can afford to live in these areas increasing the exposure to a larger population. 3) If there are people of middle income it may drive them to other areas away from the industry. 4) Externalities among neighbors can occur whe n one group has racial and/or economic prejudice against another group (Fujita 1989) This can result in segregation in the neighborhood as well as W hite flight where certain PAGE 30 16 race/ethnic gr oups (e.g., Hispanics) move into a majority White neighborhoods resulting in the population relocating elsewhere (Pulido 2000) 5) Externalities resulting from transportation can be from high congestion of traffic alone, locations of freeways near neighborhoods and air pollution resulting from traffic. These externalities can result in property values declining and increased health concerns from exposure to car exhaust. 2.1.3.2 Explanations of Neighborhood Change The theoretical models outlined below provide a temporal component to environmental justice explanations. They ask the question how and whether environmental risks change over time, by understanding the siting process of local unwanted land uses (LULUs), the spatial associations of potential exposures to the impacted populations, and the factors influencing the construction and degradation of neighborhoods. 2.1.3.2.1 Classical InvasionSuccession Model This model originated through ecological theory and can be metaphorically represented by an example where an indigenous group (White population) living in a neighborhood is invaded by an outsider (Black person) resulting in the neighborhood initially having some resistance. As the Black person becomes more acclimated and begins to change their space others wit h their similar background and socioeconomic status begin to move into the neighborhood resulting in an increased resistance. If the indigenous group fails in its resistance then they withdraw resulting in a racial/ethnic change in the neighborhood. PAGE 31 17 Thi s model shows the importance of understanding the construction of the neighborhood and the temporal aspects of its space. It is important to understand the history not only of the neighborhoods succession process but of the people (e.g., Black migration) who have inhabited the neighborhood in the past This model stresses that once the succession has occurred it is expected for density of the new residents to increase thereby causing the boundary areas inhabited by Whites to decline in value and potential ly selling to industries resulting in an increase in pollution. Industries could be sited before the succession began (pre siting) or after the succession was complete (post siting). Therefore, it is important to understand when industries were sited in order to determine if an environmental injustice has occurred. 2.1.3.2.2 Neighborhood Life Cycle Model The importance of recognizing the temporal element of neighborhoods can also be seen in the neighborhood life cycle model which also originates from an ecological theory. This model states that a neighborhood has a natural life cycle and as it ages and declines its residents become members of successively lower socioeconomic groups. The model has five cycles which are: residential development, transiti on, downgrading, thinning out and renewal. The life cycle model is exclusive of LULUs or other external forces. Therefore it is anticipated that each neighborhood naturally will physically age and decline as well as socially and economically decline. Additionally, post siting of LULUs in this neighborhood may exacerbate the decline but is not the cause. This model seems intuitively misguided since it does recognize that pre siting of industries influences the degradation of the neighborhood but it str esses that people can relocate to another neighborhood at a higher cycle of development thereby gaining a benefit. Additionally, it PAGE 32 18 ignores the factors which may influence why the neighborhood changes overtime but embraces that the change occurs. 2.1.3.2.3 PushPull Model The last model to be addressed is the pushpull model. It is represented by many social, cultural, and economic impacts. This model relates why people move or remain in certain spaces. The push factors are represented as undesirable factors that influence people to move while pull factors are positive factors which attract people to an area (Been 1994) Push factors are identified in neighborhoods as LULUs, fear, violence, lack of employment, aging infrastructure, prejudice, traffic congestion etc. Some pull factors may include job opportunities, better housing, safety, and lack of pollution. These factors work independently but tend to have some interconnectivity. 2.1.4 Geographical Principles of Environmental Justice The theoretical bases for environmental justi ce and inequity have been addressed in the previous section. The foundation of this topic has been laid in economic, political and social theories and has progressed into urban and location (i.e., residential and industry) theories. Additionally, a more practical component of environmental justice theories are the temporal theories and explanations as previously indicated. The challenge is to understand the interdisciplinary nature of environmental justice and to encourage a variety of thought be applied to solutions of inequities. In blending of these theories with geographical principles, a broader perspective of the environment and of justice can be achieved. The basic tenets of geography are space, place, scale and time and they all have numerous and varied definitions (Castree 2005) The Explanations of Neighborhood Change as previously defined, stress the PAGE 33 19 importance of relating time to space (spatio temporal relationships). Space can be perceived as undergoing continual construction exact ly through the agency of things encountering each other more or less organized circulations ; in essence this is a relational view of space (Thrift 2005) In environmental justice these spaces are created by the relationship of neighborhoods (not just a house but the demographics and socioeconomics of the neighborhood), industries, pollution and adverse health effects. In recognizing that space is recreated throughout time, it is also necessary to recognize that a definition for space is the creation of place. Place once again has several definitions. It can be defined as: 1) a location on earth, 2) a sense of place which can be a subjective feeling a person has about a place and 3) a setting and scale for peoples daily actions and interactions (Castree 20 05) It is understood that places are interconnected and exert certain influences upon each other but it is also recognized that places exist independently. In relating environmental justice to neighborhoods, it is important to understand not only the history of the neighborhood, i.e., the period each group inhabited the space, but also the social issues of the time (e.g., segregation, desegregation, and migration patterns of populations into areas), the economic structure of the city, and different type s and length of time of the industries inhabited by the space, as well as the size of their footprint on the landscape. A place has a distinct individuality. The idea of place is one of the most important aspects of understanding why environmental injust ices occur. The nexus of certain conditions e.g., economic social, pervasive institutional racism has lead to these injustices and health disparities occurring in specific and distinct locations. PAGE 34 20 As a geographic principle, the understanding of scale f or environmental justice research must also be stressed. The spatial scale is determined by the resolution which indicates the level at which we focus on a particular item of interest (Herod 2005) Numerous environmental justice studies have been conducted at different spatia l scales in order to assess different outcomes. Scales include national, state, county, and local, and can be represented by different units of analysis such as census unit boundaries or zip codes. 2.2 Methodology for Environmental Justice Assessment Th is section describes geographi c analytical units and approaches utilized to identify potentially exposed areas and populations in environmental justice research. I t provides the evolution of how geographical analytical units have been previously selected and used in environmental justice research, their contemporary usage and some of the pi tfalls and problems of selecting geographical analytical units. 2.2.1 Geographical Unit of Analysis The selection of the geographical unit of analysis for environment al justice research has been subject to an ongoing debate in determining the proper scale of selection and representation of neighborhoods (Willia ms 1999; Liu 2001) The spatial definition of a neighborhood has been routinely left to the researcher s and their research agenda and thus these studies have varying results based on the choice of analytical unit Several geographical units of analysis such as zip codes, census tracts, counties, and census blocks/block groups have been used in the research literature. Additionally, nonstandard geographical analytical units are used by the US census bureau namely PAGE 35 21 statistical areas e.g., consolidated metr opolitan statistical areas for year 2000 census or micropolitan statistical areas (Figure 1). Additionally, with the advent of grids a new way of structuring census data has been established. Figure 1 depicts the variety of units of analysis and their co mponents as used by the US Census with the most commonly used units in environmental justice studies explained in the following section. Figure 1:Geographical Hierarchy for the Decennial Census, (US Census Bureau 2005) PAGE 36 22 2.2.1.1 Zip Codes In the infancy of environmental justice research, zip codes were the primary geographic unit of analysis. The pioneering United Church of Christ (1987) study was the first on a national scale to use five digit zip codes as units to conclude that race was the most influential variable in the location of hazardous waste facilities. Zip codes are arbitrary five digit numbers used by the US Postal Service for ease of distribut ing mail to neighborhoods not for analyzing populations. These numeric codes represent the aggregation of households and businesses into small areas and thus facilitate a configuration of neighborhoods throughout the US. One concern with using the zip code as an analytical unit is that not all businesses or industries receive mail at the site location, therefore certain important sources of pollution could be missed. Additionally, large neighborhoods can be comprised of several zip codes and as such the se zip codes must be aggregated to obtain a true representation of the affected locations (Liu 2001) Furthermore, since zip codes are not units defined by the US Census, population data are not collected specifically for their configuration, and the sizes and shapes of the zip codes do not remain consistent from year to year. Therefore, zip codes are an unstable unit if performing inequity studies. Although the use of zip codes as analytical units have facilitated the discovery of envi ronmental inequity in the past (Brooks and Sethi 1997) their limitations called for a more rigorous unit to be developed and us ed to improve robustness of equity studies. 2.2.1.2 Census Tracts Census units including tracts (Downey 2006) block groups (Grineski et al. 2007) and blocks have emerged as the most commonly used geographical units of analyses used PAGE 37 23 in environmental justice research. Census tracts are subdivisions of counties that were established to provide a small statistical area un it with comparable boundaries between decennial censuses (US Census Bureau 1997) For the 2000 US Census, a census tract must 1) contain between 1,500 and 8,000 people (4000 person average) (Williams 1999) 2) meet boundary feature criteria, 3) be reasonably compact land area, 4) be accessible by roads, 5) have the same boundary as the county, and 4) cover the entire land and inland water area of each county. Census tract data are routinely used in environmental justice studies because there is a wealth of data variables collected on a decennial basis by the US Census Bureau on various population and housing characteristics. A second reason to use census tract data is that on average they comprise a smaller area of land than zip codes and therefore are more representative of neighborhoods (Anderton et al. 1994) Additionally, because of the inherent criteria for establishing a census tract, i.e., the population size (Figure 1), there is reas onable comparison between census tracts versus zip codes where there is no population lim it attributed to spatial size. 2.2.1.3 Census Blocks/Block Groups Census block groups are statistical subdivisions of census tracts and are comprised of census blocks. For the 2000 US Census, a census block group must 1) contain between 600 and 3,000 people, 2) meet boundary feature criteria, be reasonably compact, contiguous land area, all parts must be accessible by roads, 3) be reasonably compact contiguous land a rea, 4) be accessible by roads and 5) have the same boundary as the census tract and cover the entire land and inland water area of each county. Additionally, each census tract must contain a minimum of one block group and a PAGE 38 24 maximum of nine block groups. A block group within a Native American reservation is allowed to extend across state lines. The 1990 US Census allowed for the first time the land mass of the US to be numbered using census blocks as the unit. The census block is the smallest unit of ana lysis collected by the US Census (US Census Bureau 1997) The census block must fulfill the s ame boundary criteria as the census block group and unlike census block groups its delineation is not primarily based on the population. The census block can contain zero or thousands of people (up to the 3,000 person limit of the census block group as it is the larger of the two boundaries). The primary reason for using the census block and/or block groups is to allow for analysis of the population at a finer resolution. The census block group is more useful than the census block for environmental justi ce research because it contains detailed data on population, housing, income, poverty, education and other many other characteristics In order to preserve the privacy of individuals and households only a limited number of variables describing basic popul ation and housing characteristics are published at the block level. 2.2.2 Geographic Problems and Spatial Effects In environmental justice research, it is important to understand the strengths and weaknesses of each unit of analysis in order to limit the errors associated with geospatial analysis. This section introduces the major limitations of selection criteria for geographical units of analysis and stresses the importance of appropriate selection. Furthermore it outlines the necessity for spatial an alyses that help to mitigate the effects of spatial dependence in the data PAGE 39 25 2.2.2.1 Modifiable Areal Unit Problem There are several common problems related to the selection of proper geographical unit of analysis. The modifiable areal unit problem (MAUP) is comprised of the scale effect and the zoning effect (Fotheringham and Wong 1997) The scale effect occurs when different statistical findings are obtained from the same data set when the information is grouped at different levels of resolution e.g., census tracts, blocks and counties) (Wrigley et al. 1996) The zoning effect occurs when different statistical findings are obtained with a modifiable areal unit as a function of various ways these units can be grouped at a give scale, and not as a result of the variation in the size of those areas i.e., the difference in results which follows from merely altering the boundaries or configurations of the zones a t a given scale of analysis Results from both effects have been found to result in varying levels of goodness of fit for a model, different regression coefficient estimates and t values, as well as an increase in correlations between variables in zoning alone. A more direct issue stemming from scale effect is des cribed as the ecological fallacy. Anderton (1994) writes that an ecological fallacy is the reaching of conclusions from a larger unit of analysis that does not hold true in the analyses of smaller, more refined units. This reflects a sensitivity of results to the aggregation of data. This was observed in the Glickman et al., (1995) study, where racial inequities were not observed at the census block group and census tract levels, but detected only when the impacted area was defined by a circular buffer of radius 0.5 mile or 1 mile. PAGE 40 26 2.2.2.2 Edge Effect A common problem identified a s a result of geographic analytical unit selection and/or analytical technique selection is edge effect. Edge effect (or border effect) occurs as a result of pollution sources under study being located near the edge of two or more census units where pollution is counted as belonging solely to the host census unit of the facility. Methods suggested to mitigate the edge effect (Liu 2001) include: 1) aggregating the census units that are along the edge into one by determining the distance of impact (exposure ) on the affected neighborhood; and 2) not aggregating the units but treating all units as potentially exposed. Both of these methods have been used to rectify the problem, but they work with varying degrees of results (Downey 2006; Mohai and Saha 2006) Since pollution from a facility is not limited by the boundaries of the analytical unit hosting the facility, if the pollution is not allocated to all areas where it travels then there can be an overestimation in the host analytical unit and an underestimation in adjacent nonhost analytical units that are also likely to be impacted 2.2.2.3 Spatial Autocorrelation The first law of geography as expl ained by Waldo Tobler (1970) states that everything is related to everything else, but near things are more related to each other A variable is considered to be spatially autocorrelated if the geographic distribution of a variable indicates a systematic pattern and spatially independent i f no pattern ca n be discerned. Positive spatial autocorrelation is often caused by geographic clustering when values at neighboring locations are more similar than would be expected of a random distribution (Kissling and Carl 2008) Most conventional statistical tests, however, are based on the assumption that the values of observations in each sample are independent PAGE 41 27 of one another. Although the ordinary least square s (OLS) method for multiva riate regression has been used extensively in quantitative studies on environmental justice and health disparities spatial autocorrelation potentially leads to spatial dependence of regression model residuals, thus violating the OLS assumption of independent observations (uncorrelated errors) and biasing the regression coefficients associated with the explanatory variables When analyzing spatial data, therefore, appropriate statistical modeling techniques must be used to account for the effects of spatial dependence Spatial regression, or the s patial autoregressive (SAR) model s address the spatial autocorrelation problem by augmenting the standard OLS regression model ( X Y ) to include the effects of spatial dependence (Anselin and Bera 1998) The SAR model is in effect, a combination of the OLS model that includes the response variable )( Y and the term for predictors and errors ) ( X along with an additional term which accounts for the spatial autocorrelation structure ( W ) of a given data set Th is additional term is implemented with a spatial weights matrix ( W ) where the neighborhood of location and the weight of each neighbor need to be defined (Kissling and Carl 2008) The SAR model requires the development of an appropriate set of neighbor relationships, or spatial weights matrix. The spatial weights matrix accounts for variation in the dependent variable explained by values at nearby locations instead of the explanatory variables. The spatial weights matrix is constructed by identifying the neighbo rhood structure of each location or areal unit, based on a contiguity based or distance based technique. The queen contiguity includes all common points (boundaries and vertices) of the areal unit to determine the neighbors while the rook continuity uses the common boundaries alone to determine the neighbors. Additionally, the distance band selects all PAGE 42 28 areal units whose centroids fall within a specified distance from the centroid of the areal unit under consideration. Since any chosen distance can be us ed to define the spatial extent of neighborhood effects, it is important to select the minimal distance from the boundary as the distance band which also negates spatial autocorrelation. The opensource spatial statistics application GeoDA (Anselin 2004 ) can be used to develop OLS and SAR models detect spatial autocorrelation, and generate weight matrices for this study. The SAR models operationalized by GeoDA utilize both contiguity based spatial weights and distancebased spatial weights for a specif ied spatial analytical unit to identify the neighborhood configuration (Anselin 2006) S patial weight matrices based on queen contiguity, rook contiguity or distance band s can be calculated to represent neighbor relationship s between areal units Two types of SAR models can be used to improve the OLS model and account for spatial dependence in the data: spatial error and spatial lag. The spatial error regression model is implemented when the assumptio n of uncorrelated error terms is violated. This model assumes that the autoregressive process occurs only in the error term and not in the pre dictor or explanatory variables (Dormann et al. 2007; Kissling and Carl 2008) The spatial error model is specified as: W X Y (Dormann et al. 2007; Kissling and Carl 2008) (1) where ( W ) represents the spatial structure ( W ) in the spatially dependent error term ( ). The s patial lag model on the other hand, assumes that spatial autocorrelation occurs only in the response variable This model thus includes a term for the spatial autocorrelation in PAGE 43 29 the d ependent variable and also the standard error term that is used in the OLS regression model It is specified as: X WY Y (Dormann et al. 2007; Kissling and Carl 2008) (2) where ( W ) is the term for spatial autocorrelation in the response variable ( Y ). The presence of spatial autocorrelation for regression residuals (error terms) is determined by the global Morans I in this study. The Morans I is a g lobal test that has an associated z score which indicates that the point pattern could be a result of random chance if the z score is near 0 and non significant. The Morans I measures the spatial autocorrelation by m j ij n i m j j i ij n iw s x x x x w I1 1 2 1 1 (Chang 2006; Torrens 2008) (3) where ix is the value at point i jx is the value at point i s neighbor j ijw is the coefficient for measuring adjacency between two points, n is the number of points, and s2 is the variance of x values with a mean of x A positive I indicates that points tend to be located to similar high or low values, while a negative I suggests that high value s tend to locate near low values in the study area. Although multivariate regression has been used extensively in environmental justice research (Bowen and Cyran 1995; Pollock and Vittas 1995; Daniels and Friedman 1999; Bouwes et al. 2001; MorelloFrosch et al. 2001; Ash and Fetter 2004; Sicotte and Swanson 2007) spatial auto regressive (SAR) models have rarely been implemented as a PAGE 44 30 means to account for spatial dependence in the data U tilizing SAR for multiple regression with spatially referenced data could potentially lead to more statistically valid conclusions about the si gnificance of explanatory variables such as race/ethnicity or poverty, in quantitative analys e s of environmental justice. 2.2.3 Spatial Analytical Approaches to Identifying Impacted Areas Environmental justice studies have used several approaches used to determine the geographic boundaries of areas exposed to environmental pollution and related health risks and to estimate the population characteristics of these potentially impacted areas. These are described in detail below. 2.2.3.1 Common Analytical T echniques Downey (2003) and Mohai and Saha (2006) outlined the unit hazard coincidence method, which assumes that ambient pollution is confined to and evenly distributed within the host unit of analysis. This method sums the number of hazards located in the study area, sums the pounds of pollution emitted in the analysis unit or creates a dummy variable that indicates whether or not an analysis unit contains a hazard. This method does not ac count for edge effect s ambient pollution spreading outside the pre set boundaries, or the lack of even distribution of pollution which can occur within the host unit. A large number of environmental justice studies have used this approach to compare the p opulation characteristics of host areas with areas that do not host polluting facilities (e.g., United Church of Christ Commission for Racial Justice 1987; Anderton et al. 1994) PAGE 45 31 Circular buffer analysis is an alternative method used to establish impacted areas and populations from pollution exposure, typically on the basis of GIS software. The idea of buffering or uniform buffering is to identify a pollution source and then choose circular areas of a specified rad ius (e.g., 0.5, 1, 2, 5 miles from the center) with the pollution source representing the center of the circle. While circular areas address some of the limitations of the unithazard coincidence method, there are a few problems associated with its applic ation. The first problem is the arbitrary choice of radius to define circular areas which is not generally based on the quantity or toxicity of emitted pollutants (Chakraborty and Armstrong 1997) Secondly, uniform buffering does not account for varying toxicity levels of exposure and assumes that all areas within the buffer are exposed to equal ris ks. The third problem is that it assumes that a population is only exposed to the one source instead of multiple sources. These issues can be mitigated by several more rigorous methods of buffering and modeling techniques. The circular buffer approach h as been utilized in many prior studies of environmental injustice. Glickman (1994) for example, used buffer analysis in Allegany County, Pennsylvania to determine the proximity of nonWhite residents to industrial manufacturing facilities reporting toxic emissions. He found that nonWhites bear less of a burden than Whites and populations above the poverty line. This finding was due to the aggregate effect of combining all hazard sites and based on the large size (2 mile radi us ) of the buffers. At a smaller scale (0.5 mil e and 1 mile radius ) the residents closer to the sites were predominantly nonWhite and impoverished but as the radius of the buffer expanded it encompassed the White and wealthier population. The updated version of the national scale Toxic Waste and Race study (United Church of Christ Commission for PAGE 46 32 Racial Justice 2007) used 1.8 mile radius buffers to examine populations in proximity to hazardous waste facilities. This study found that minorities were the predominant population who lived within the analyzed distance of hazardous waste sites in the US. Although studies relying on the circular buffer approach have reported disproportionate risk burdens imposed on the population, uniform buffers lack the ability to analyze dispersion patterns of emitted chemi cals based on their physiochemical properties and related meteorological factors. Additionally, the buffers fail to determine the impact of individual and cumulative effects of toxicity related to the point sources of pollution and thus limits the discove ry of health outcomes and the resultant health disparities on the affected populations 2.2.3.2 Contemporary Analytical Techniques New methods are being developed to address exposure risks and outcomes as a result of hazardous pollutant emissions. In the past, limitations have been placed on environmental justice studies because of the lack of data pertaining to exposures from chemicals. Currently, methods are being constructed to take into account the influence of varying levels of toxicity and multiple chemical exposures on a population. Dispersion Models Atmospheric dispersion modeling is a predictive method used to estimate concentrations of pollutants in the atmosphere resulting from point or nonpoint atmospheric emissions (Dent and Henriques 2000) Advances in fate and transport techniques have introduced dispersion models, allowing them to be interwoven with census derived socioeconomic param eters and demographics using GIS to go beyond basic identification of the location of industries and the populations. The models can PAGE 47 33 now provide assistance in exploring the potential health impacts as a result of the pollution. Several dispersion models are available for modeling of fate and transport of industrial releases into the atmosphere (e.g., DEGADIS and ISCLT ). These models are routinely used by the US EPA, NOAA and other private and governmental bodies as methods of discerning fate and transpor t of chemicals through the atmosphere. Specifically, DEGADIS and ISCLT are routinely used in conjunction with other models to identify different types of industrial releases (short and longterm releases). Short Term Release Model DEGADIS is a mathematical dispersion model that can be used as a refined modeling approach to estimate short term ambient concentrations and the expected area of exposure to concentrations above specified threshold values for toxic chemical releases. Chemicals with an atmospher ic density greater than air, i.e., a heavy gas, is modeled using DEGADIS which accounts for the sinking of the gas cloud to the ground and as it dissipates with distance from the source then the chemicals behavior imitates a chemical with physiochemical pr operties more similar to air which is signified when the concentration drops below 1% or 10,000 parts per million (ppm) of air (Figure 2). PAGE 48 34 Figure 2: Heavy gas dispersion from gravity (US Environmental Protection Agency and National Oceanic and Atmospheric Administration 2007) Limitations of the DEGADIS model include the fact that they cannot be used in complex terrain ( e.g., valleys ) and during high sea breeze. In spite of these limitations this nonGaussian model is particularly important for analyzing chemicals which are heavier than air and do not easily disperse in the environment Scenna and Santa Cruz (2005) used DEGADIS to explore the risk of chlorine gas dispersion from mobile sources in Argentina. They found that the model was able to interpret that there was an increased risk to persons in cars and the resident population if a release w ere to occur. Long Term Release Model The Industrial Source Complex (ISCLT versions 1, 2 or 3) is a Gaussian model used for long term releases into the atmosphere. It provides good spatial resolution near sources but cannot predict fate and transport greater than 50km from the source (US Environmental Protection Agency 1996b) Chemicals with an atmospheric density similar to air (1.1 kg m3) are processed in the Gaussian model which is represented by a bell curve that becomes flat as the concentration of the chemical declines as it diffuses from the source of the release (Figure 3). Figure 3. Gaussian distribution and spread of chemical from source. (US Environmental Protection Agency and Nati onal Oceanic and Atmospheric Administration 2007) PAGE 49 35 Both these models provide significant contribution s to interpreting the complexity of how toxic chemicals are dispersed throughout the environment. They lend important guidance to modelers and risk as sessors to the identification of areas potentially exposure to chemical emissions and adverse health effects Furthermore they provide necessary information in pinpointing where environmental inequitie s can occur as a means of outlining the potential exp osure to the population. 2.3 Environmental Health Disparities Numerous definitions have been provided to explain a health disparity (Table 1). Keppel (2005) writes that health disparities are not simply about differences in health. The term difference in itself has many meanings including inequity, inequality and injustice (Krieger 2005) The first definition developed by the US government was prepared by the National Institutes of Health in 1999 as part of a development plan on reducing health dispari ties (Table 1). A year later this definition was improved with a broader definition provided by the National Center for Minority Health Disparities. This new definition included a proportional comparison to the general public not just among population g roups (US Congress 2000; Institute of Medicine 2006) PAGE 50 36 Table 1: Government definitions of health disparities. Government Source Definitions 1 National Institutes of Health (1999) Health disparities are differences in the incidence, prevalence, mortality, and burden of diseases and other adverse health conditions that exist among specific population groups in the United States. 2Minority Health and Health Disparities Research and Education Act (2000) A population is a health disparity population if there is a significant disparity in the overall rate of disease incidence, prevalence, morbidity, mortality or surviva l rates in the population as compared to the health status of the general population. 3 National Cancer Institute Differences in the incidence, prevalence, mortality, and burden of cancer and related adverse health conditions that exist among specific populations groups in the United States. date unknown 4 Healthy People 2010 (2000) The second goal of Healthy People 2010 is to eliminate health disparities among different segments of the population, including differences that occur by gender, race or ethnicity, education or income, disability, geographic location, or sexual orientation. 1(Institute of Medicine 2006) 2(US Congress 2000) 3(National Cancer Institute 2007) 4(US Department of Health and Human Services 2000) The National Cancer Institute (NCI) and Healthy People 2010 goals provide more topically focused definitions. The NCI specifically focused on cancer disparities while Healthy People 2010 goals are geared independently on all affected demographic groups (US Department of Health and Human Services 2000; National Cancer Institute 2007) He althy People 2010 is a set of goals established by the US Department of Health and Human Services and includes this definition (Table 1) as part of its mission not only to decrease but eliminate health disparities in the US (US Department of Health and Human PAGE 51 37 Services 2000) This far reaching definition encompasses all demographics of the population thereby extending the goal to all residents. The following section outlines the causes of health disparities, how they are measured, and their health outcomes. 2.3.1 Influential Variables on Health Dispariti es There is not a single cause of a health disparity. Health disparities result from the interaction of social, cultural, racial and economic factors, as well as geographical locations. Socioeconomic status (SES) as measured by education attained, income levels, and occupation along with race/ethnicity are listed as the most common influences for health disparities (National Research Council 2004) Additional factors, including personal behavi ors, culture (language and acculturation), genetics, gender, and physical environment (where you live, work, types of exposures), can also contribute to health disparities. 2.3.1.1 Race/Ethnicity Numerous studies have determined that race/ethnicity is an influential factor of health disparities. Studies have found that minorities have higher risk to exposures because Whites and minorities do not work, live or play in the same places (Lee 2002) This idea is promoted through the persistent role of residential segregation still in effect post desegregation and the related policies (Williams and Jackson 2005; MorelloFrosch an d Lopez 2006) AcevedoGarcia (2000) stressed that segregation con centrates poverty, overcrowded and dilapidated housing, and social disintegrati on in minority neighborhoods. Maantay (2005) identified that in the Bronx, New York, 13 percent of the population living near the hazardous facilities hospitalized for asthma were more likely to be minorities At a different scale, Bai et al. (2007) found that Black children in PAGE 52 38 Pennsylvania were twice as likely to have more severe asthma symptoms after controlling for SES. Although anti discrimination policies have been in effect in the US for over 40 years (US Department of J ustice 1964) institutionalized racism is pervasive in society and has a powerful impact on where people live and work and their economic status According to a study on infant mortality rates (IMR) conducted by Laveist (2002) the US White (IMR=11) population closely resembles more developed countries while the US Black population (IMR=20) reflects mortality rates in less developed countries when viewed separately on an international scale. 2.3.1.2 Socioeconomic Status Socioeconomic status (SES), independent of other factors, has been found to influence health disparities. When looking at the impact of SES on health disparities Maantay (2005) found that 30 percent of the population living near hazardous facilities in Bronx were hospitalized for asthma were more likely to be below the poverty level. Several other studies have observed that wealth is a contributing factor to health outcomes. Chen (2006) found that persons living in census tracts with high poverty levels have an increased incidence of premature death. Low income areas are fraught with environmental pollution and poor infrastructure. Additionally, persons of low income tend to have limited access to health care and proper nutrition further contributing to the process of health disparities in this group (O'neill et al. 2003) 2.3.1.3 Other Factors and Susceptible Populations Gender and genetics play a distinct role in health disparities. Health disparities have a link to genetics as they have been found to increase the susceptibility of certain PAGE 53 39 diseases on individuals carrying the gene (e .g. breast cancer, BRCA1,2 genes or Colon cancer, MSH2 gene) (Edlin and Golanty 2004) Furthermore, gender (as a biological role and a social role) can influence whether a health disparity occurs (Coreil and Henderson 2001) Human males and females not only have genetic variability, but also have a physiologically different metabolism thereby allowing certain people to be susceptible to various exposures. For example, men and women have different breathing patterns on average and thus their inhalation of chemicals would be different (Hinds 1999) The social role of gender can be expressed through cigarette smoking. Smoking is known to be highly correlated with lung cancer. In the US, men smoke at a higher percentage (23.4 percent) than women (18.5 percent) (Ce nters for Disease Control and Prevention 2004) This lifestyle behavior provides evidence that if it was adapted then the prevalence of lung cancer in the US populations would decrease. Additionally, other lifestyle behaviors like nutrition, alcohol, and occupation could contribute to health disparities and should be evaluated for causation. Additional susceptible populations such as persons with compromised immune systems, young children, elderly and fetuses (maternal exposure) can play a role in outcom es of health disparities. Young children are susceptible to lead poisoning and other neurologic damage as they grow into adulthood. The elderly population may have compromised immune system from slower metabolisms and difficulty recovering from injury. L astly, fetuses are a susceptible population because exposure in the womb may result in the increased opportunity for preterm deliveries, deformities, neurotoxicity, decreased mental acuity and death (Carpenter et al. 2002; Woodruff et al. 2003) PAGE 54 40 2.3.2 Measures of Health Disparities The choice of subpopulations as a method of measurement for environmental inequity studies is a critical issue for researchers. Several subpopulations are customarily studied as variables used to represent environmental inequities namely, race/ethnicity, and income. Additionally, there are specific susceptible subpopulations identified through evidence obtained from environmental health disparity and adverse health effect studies e.g., the elderly, children, fetuses (prenatal exposures) and persons with depressed immune systems (HIV and chronic illnesses (kidney, liver diseases). 2.3.2.1 Race/Ethnicity Many studies use race/ethnicity as variables of choice for assessing health disparities but there are a multitude of challenges and potential inaccuracies that can, will, and do result from the use of race/ethnicity variables. In the US, race/ethnicity is primarily self reported on the US Census and health care documents and thus it carries an inherent bias (positive or negative). The limitations of the US Census surveys as well as the multitude of mixed race individuals confound analyses in social and geographic research. Furthermore, the change in the racial/ethnic categories between decennial surveys results in a lack of consistency between the collection years. For example, over the last 30 years on the US Census, the category of Black race has been altered three times from Black in the 1900s, to Negro in the 1950s, to Black/African American in the year 2000 (National Research Council 2004) Additionally, race/ethnicity when not self reported is assumed by the perception of individuals, entering more bias into the data stream. For example, persons from many Latin American countries may have African ancestry but tend to id entify themselves as being Latin o or Hispanic These individuals PAGE 55 41 could be perceived as Black if visually surveyed, but if verbally questioned they would identify themselves as being of Hispanic origin. This is further exacerbated by the introduction of self reported multiple race categories on the 2000 census. Environmental justice studies typically use variables describing racial/ethnic and economic characteristics of the populations for studying the disproportionate impacts of toxic exposures. Most studies examine Black and White populations as the primary variables for assessing environmental injustice, but neglect the myriad of diverse races/ethnicities that exist in the US (Bullard 1990; Perlin and Wong 1999) This dichotomous relationship is slowly being expanded by the inclusion of Hispanic and Asian populations but other races/ethnicities such as Native Americans that potentially experience environmental injustice and environmental health disparities are neglected (Ash and Fetter 2004) 2.3.2.2 Socioeconomic Status and Income Socioeco nomic status encompasses income, occupation and education. Income itself can be parsed into current income, wealth, and lifetime wealth (National Research Council 2004) Current income variables provide data for annual cash flows and are presented by the most common used measures of income namely, median household, poverty levels and percent poverty. When using current income as a variable it is important to realize that, like race/ethnicity data, income is also self reported. Additionally, many individuals are unwilling to disclose their income level and will therefore leave the item blank or report false data. Annual household income levels have been used to measure economic status in man y environmental justice analyses Studies use the income distribution of each PAGE 56 42 individual race/ethnicity, household income (median or mean), poverty thresholds and/or poverty guidelines. Median household income (MHI) is used in a particular area and can a ccount for income status over space and time (US Census Bureau 2000; Liu 2001) The MHI reports the income of all residents of t he household over 15 years of age. There are several problems with the median household income variable. First, it is not available at finer geographical scales. Second, the decennial US census reports MHI to the block group level while in outside of the decennial census year the county level is reported. Third, it is not possible to count low income individuals. Poverty thresholds are also used as tools for measuring income. The federal poverty threshold is based on the definition of poverty establishe d by the US Social Security Administration (SSA) in 1964. The poverty measure is based on the recognition that families of three or more people across all income levels spend 1/3 of their income on food; therefore the SSA multiplied the Economy Food Plan established by the US Department of Agriculture by a factor of three to obtain the poverty threshold (US Census Bureau 2000; Liu 2001) The poverty threshold is revised annually to reflect the cost of living of United States in totem, but does not account for variation in regional cost of living. Additionally, household poverty levels only measure bloodrelated family members in h ouseholds wher e a family member is the head. This measurement can result in an underestimation of the number of people in poverty since people who are unrelated may live together to pool their economic resources. 2.3.2.3 Wealth, Education, and Occupation Wealth which is comprised of the accumulation of assets (e.g., stock and home ownership) and lifetime wealth are used in longitudinal s tudies as a measure of lifetime PAGE 57 43 income. Wealth and lifetime wealth can be difficult to measure, but are useful in that t hey can provide observations into how persons with high wealth can have access to health care thereby preventing a disparity and persons with low wealth could be mired in the welfare health care system resulting in higher morbidity and mortality rates. Edu cational levels are also used to determine SES. This variable typically is based on level of school completed; it gives a representation of intellectual, behavioral and financial ability. It can also be used as a proxy for the ability of a person to obta in and comprehend health information and diagnoses (National Research Council 2004) This measure is relatively stable over a lifetime and correlates well with occupation, income and social s tatus. The inadequacy in the measure is that it is not tracked geographically on an individual level; therefore the quality of education cannot be determined. Ash and Fetter (2004) and Sicotte and Swanson (2007) included residents with less than high school education as a variable in their studies and found that block groups with a larger amount of adults without diplomas tend to live in more polluted areas. Anderton et al. (1994) used the mean value of housing stock as a proxy for ne ighborhood wealth and found significantly lower values in tracts containing hazardous waste facilities. More recently, Pastor et al. (2005) have argued that home ownership is not only a measure of the wealth possessed by a community, but also a potential indicator of political involvement based on the assumption t hat home ownership implies leads to higher concern for the local environment due to invested capital. Their study indicated that the census tracts with higher levels of home ownership are less likely to be exposed to the adverse health risks from air pollution in California PAGE 58 44 Occupation is used in exposure studies and can be an indicator of outcomes. Although the occupation variable may be less difficult to collect than the income variable and is a proxy for income, it suffers from a lack of economic info rmation 2.3.3 Effects of Environmental Health Disparities Routes of exposure can be dermal, respiratory and oral pathways (Figure 4). These exposures can result from moderate to terminal adverse health outcomes (effects). Outcomes are defined as the aff ect of exposures on a given population. Some outcomes of environmental health have been identified as increased incidence of asthma in children in the inner city, cancer clusters and congenital birth defects (Oleckno 2002) Chemical exposures to populations can either be acute or chronic and can result in numerous outcomes to the exposed population. PAGE 59 45 Figure 4: Schematic of exposure and effects (outcomes). (US Environmental P rotection Agency 1992) 2.3.3.1 Acute Exposure and Health Effects According to the US Environmental Protection Agency (US Environmental Protection Agency 2006a) acute exposures are defined as single, nonrepetitive exposures between 10 minutes and 8 hours. These exposures can cause adverse health outcomes ranging from mild eye and respiratory discomfort, irritation, prolonged multi organ systemic effects to death. The effects of acute exposure are ba sed on the interactions of susceptible populations the toxicity levels and the route of exposure. An acute exposure of mercury has been found to cause coughing, burning in the lungs, gastrointestinal PAGE 60 46 problems, increased blood pressure, eye irritation and in high doses kidney problems and development issues with fetuses and death (Zeitz and Kaye 2002) Although mercury can enter the body via dermal absorption or ingestion its most effective route of entry is by inhalation therefore atmospheric exposure during spills can result in significant exposure and concentrations to the general population. 2.3.3.2 Chronic Exposure and Health Effects Vulnerable populations, namely the elderly, the chronic ill, people with compr omised immune systems and children, have been recognized as special populations when evaluating health outcomes. These populations tend to have a greater risk from environmental exposures for several reasons. The elderly as they age may suffer debilitating diseases (decreased liver and kidney functions) and have increased poverty rates resulting in poor health care and subsequently resulting in poor health. The chronic ill with compromised immune systems (e.g., people with AIDS, cancers, liver and kidney disease) tend to have depressed immune systems which make them highly susceptible to environmental exposures. Children are listed as a special group because environmental exposures to the young could negatively impact them for a lifetime. Lead (Pb) exposure during the mental development phase of young children has been found to cause neurological damage and result in slow mental cognition and bra in damage (Lanphear and Roberts 2005; Jain and Hu 2006) Additionally, numerous cohort studies have identified the increased prevalence of asthma in the US of 75 percent between 1980 and 1994 (Brown and McCormick 2004) Although a single cause of asthma has not been identified most studies agree that air pollution specifically PM2.5 is the primary outdoor pollutant that influence asthma attacks. PAGE 61 47 Additionally fertility issues concerning exposure and disparities have arisen in the general population. Studies have found that air pollution during pregnancy increased the risk of poor birth outcomes in minorities and low income populations (Bobak and Leon 1999; Woodruff et al. 2003) Specifically, low birth weight, premature births and infant mortality have been found to be greater in areas where the mothers are of low socioeconomic status (as measured by education level attained) and minority status. Berry and Bove, (1997) found that full term births of parents living closest to landfills in Gloucester County, New Jersey were 5 times more likely (OR=5.1) than the general population to have a low birth weight baby. Wilhelm and Ritz, ( 2003) also found that in addition to air pollution negatively impacting the birth outcomes residential proximity to traffic could also affect birth outcomes to the exposed populations. Specifically, they found that women who were exposed to traffic pollution during the winter months (more stagnant air) received greater exposure and therefore negative birth outcomes could be increased. PAGE 62 48 3. STUDY AREA The study area for the proposed dissertation research is the Houston Galveston Brazoria Consolidated Metropolitan Statistical Area (HGB CMSA) which is comprised of eight counties (Brazoria, Chambers, Fort Bend, Galveston, Harris, Liberty, Montgomery, and Waller counties) as defined by the 2000 US Census (Figure 5). Over the last 50 years, this area has experienced substantial growth in population and land area. In 1950, the incorporated city of Houston comprised 167 square miles which increased to 617 square miles by the year 2000 (Melosi and Pratt 2007) In the same time period (19502000), the population in the metropolitan statistical area more than quadrupled from 908, 822 to 4,715,407, and the land area of the MSA increased to approximately 2,000 square miles. This increase of land area was driven in part by annexation, uncontrolled urban sprawl and lack of zoning laws (Melosi and Pratt 2007) PAGE 63 49 Figur e 5: Counties of the Houston Galveston Brazoria Consolidated Metropolitan Statistical Area. PAGE 64 50 3.1 Racial, Economic, and Industrial Context Table 2 outlines the racial/ethnic and poverty characteristics of the population in this metropolitan area. Over 90 percent of the population in 2000 lived in the urban areas of the MSA. According to the 2000 census, total population of the HGB CMSA was about 4.7 million with nearly 15 percent below the federal poverty level. In the year 2000, racial/ethnic minoritie s comprised approximately half of the total population with the Hispanic population being the largest group. Table 2: Census Year 2000 HoustonGalveston Brazoria, TX CMSA Sociodemographic and Socioeconomic Status (Green Medi a Toolshed 2005) Variable Values Total Population 4,669,571 Urban: 4,294,571 Rural: 375,000 Percent Hispanic or Latino 28.9 Percent White alone 47.9 Percent Black or African American alone 16.6 Percent American Indian and Alaska Native alone 0.3 P ercent Asian alone 4.8 Percent Native Hawaiian and Other Pacific Islander alone 0.0 Percent Some other race alone 0.1 Percent Two or more races 1.4 Percent Below Poverty Level 13.5 Median household income in 1999 ($) 44,761 The increase of the population from about 120,000 in 1900 to todays count of 5.2 million (ca 2007) along with an increase in automobile usage has lead to the area becoming a nexus for widespread toxic pollution where oil production, oil refineries and automobile transportation c oincide. Historically, this area has morphed from primarily farming to manufacturing industry. The discovery of crude oil at Spindletop (Beaumont, Texas) in 1901 provided ample employment and migration of populations into the Texas PAGE 65 51 Gulf Coast (US Census Bureau 2002) As the extraction of crude oil increased, related facilities (i.e., refineries, polymerization plants and facilities) were established to synthesize chemicals with the majority of industries related to oil processing. As a result, an increase in the population surrounding these facilities occurred along with an increase in ground transportation (e.g., trucking routes and railways), in addition to the existing shipping corridors via the Gulf of Mexico. Table 3 classifies polluting industrial facilities located in the study area listed in the USEPAs Toxic Release Inventory database, based on their manufacturing activities as defined by Standard Industrial Classification (SIC) codes. As indicated in the table, a large majority of these polluting industries are involved in the production of organic chemicals and paint related chemicals as well as industries related to oil and gas refineries. The location s of the TRI facilities are provided in Figure 6. Ta ble 3: Industries in the HGB in 2000 (US Environmental Prot ection Agency 1996a) SIC Code Long Names Number of Facilities Percent Industrial organic chemicals, n.e.c. 87 13.0 Plastics materials, synthetic resins, and non vulcanizable elastomers 29 4.3 Oil and gas field machinery and equipment 27 4.0 Coating, engraving and allied services, n.e.c. 26 3.9 Chemicals and allied products, n.e.c. (wholesale trade) 22 3.3 Industrial inorganic chemicals, n.e.c. 22 3.3 Chemicals and chemical preparations, n.e.c. 21 3.1 Paints, varnishes, lacquers, enamels, and allie d products 21 3.1 Industrial gases 19 2.8 Petroleum bulk stations and terminals 19 2.8 Plastics products, n.e.c. 18 2.7 Ship building and repairing 13 1.9 Petroleum refining 12 1.8 Other Industries 332 49.7 Total 668 100.0 PAGE 66 52 Figure 6: Locations of Toxic Release Inventory Facilities in the HGB MSA, 2000 PAGE 67 53 3.2 Previous Environmental Justice Studies in the Houston Area Several prior studies have examined environmental justice issues in the Houston metropolitan area. Table 4 provides a summary of empi rical studies conducted in the Houston area, including the type of hazard examined, analytical unit used, and their research findings. Bullards seminal study, initially conducted in 1983 and published in his book Dumping in Dixie (1990) found solid waste facilities in Houston to be disproportionately located in Black neighborhoods. Been (1994) and Liu (1997) used longitudinal analyses to extend Bullards study, in order to assess pre siting and post siting changes in neighborhoods hosting these waste facilities. The obj ective of these studies was to determine whether polluting facilities were disproportionately sited in minority neighborhoods, or if minority residents moved to host neighborhoods subsequent to siting. Although Beens (1994) choice of analytical unit was different from t hose used in the Bullards study (census tracts versus neighborhoods), Beens study did not provide rigorous statistical comparisons. However, Been found that the percentage of African Americans and low income individuals in census tracts containing landfi lls had increased over time. Liu (1994) on the other hand, used quantitative statistical techniques and did not find any evidence of solid waste facilities playing an important role in the locations of Black or impoverished populations in Houston. Additionally, an updated version of the pi oneering Toxics Wastes and Race study conducted by the United Church of Christ (2007) found that Houston had the second largest percentage of minorities living in neighborhoods that contain hazardous waste facilities. Although these studies played an important role in exploring inequities in the siting process of waste facilities, they used simplistic methods to identify populations at risk and did not PAGE 68 54 assess disproportionate exposure to atmospheric pollution or related health effects associ ated with polluting facilities. Table 4: Selected Studies Examining Aspects of Environmental Justice in Houston, Texas Author/Year Hazard Unit of Analysis Analytic Method Findings Bullard (1983) solid waste facilities City of Houston neighborhoods Descri ptive and historical analysis Sites were more likely to be found near Black neighborhoods and Black schools Been (1994) solid waste facilities City of Houston census tracts Descriptive and historical analysis a)LULUs had a disproportionate affect on Bla cks and poor b)areas where LULUs were located drove down housing prices c)neighborhoods became increas ing populated with Blacks Liu (1997) solid waste facilities City of Houston census tracts Difference of means parametric and nonparametric tests No evidence provided to support the hypothesis that the presence of LULUs made the neighborhoods home to more Blacks and poor people UCC (2007) hazardous waste facilities City of Houston neighborhoods Descriptive and historical analysis Second largest % of people of color living in host neighborhoods 3.3 Suitability of the Study Area for Proposed Research The Houston Galveston Brazoria Consolidated Metropolitan Statistical Area (HGB CMSA) area is particularly well suited for studying environmental justice and health disparities in the distribution of toxic pollution for several reasons. First, its location al ong the mid Texas Gulf Coast has historically included a variety of industrial, petrochemical, and manufacturing companies. This agglomeration of industries was situated in proximity to shipping routes for ease of transporting crude oil throughout the glo bal oil economy and as such the area is prone to exposure from toxic pollutants associated with petrochemical related emissions and runoff. Second, the area has a racially and ethnically diverse population as outlined in Table 3, with the White PAGE 69 55 population comprising less than 50 percent of the total population. Third, several national environmental databases suggest that counties in the Houston metropolitan area are among the most polluted in the US. In 2002, four of the eight counties in the CMSA (Brazo ria, Fort Bend, Galveston, and Harris) were ranked in the top 10 percent of all US counties in terms of total chemical releases and waste generation (Green Media Toolshed 2005) Furthermore, three out of eight counties (Harris, Brazoria and Galveston) were ranked in the top five of counties in Texas with the greatest reported releases of recognized carcinogens to air, with Harris County being number one. These pollutionbased rankings further support the selection of this area as appropria te for this dissertation research. The rankings underscore the severity of exposure on the population and how these exposures can potentially result in adverse health outcomes. PAGE 70 56 4 DATA SOURCES MODELS, AND METHODS A variety of methodologies and inform ation sources have been used for environmental justice and environmental health disparities research, over the last two decades. Data sources are typically provided by government entities such as the US Environmental Protection Agency (USEPA), Centers for Disease Control and Prevention (CDC) and the National Cancer Institute (NCI). These organizations provide information on sources of toxic pollution and their health effects including morbidity, mortality and disease incidence and prevalence in a given po pulation at the national, state and county levels. Recent advances in computer databases, surveillance systems, environmental monitoring and pollution modeling allow an improved and more rigorous analysis of environmental health risks and associated dispa rities. This chapter explores the use of data sources and risk models for assessi ng chronic exposure that were utilized in this research and provides an overview of the statistical techniques used to analyze racial/ethnic and socioeconomic inequities in th e distribution of exposure Demographic and socioeconomic data that were used to identify the population at risk and measure health disparities are also described. 4.1 Data Sources for Pollution Exposure and Health Risk Assessment Chronic exposures wer e analyzed using the USEPAs Toxic Release Inventory (TRI) database in conjunction with the Risk Screening Environmental Indicator (RSEI) PAGE 71 57 model. These data sources and models are thoroughly examined in the following sections to provide guidance on their usage in this research. 4.1.1 Chronic Exposure 4.1.1.1 The Toxic Release Inventory The USEPAs Toxic Release Inventory (TRI) database is the most widely used resource for assessing toxic pollution in research on environmental justice and health dispar ities since the early 1990s. The TRI provides information on the annual amounts of toxic chemicals released into the environment from industrial manufacturing activities, metals and coal mining, electric utilities, commercial hazardous waste treatment and government facilities. The TRI began in 1988 as a response to the Emergency Planning and Community Right to Know Act (EPCRA) and currently reports releases of nearly 650 different hazardous chemicals annually (US Environmental Protection Agency 2006b) For each facility, the TRI database inclu des its name and address, locational coordinates (degrees of latitude and longitude), the type of industry or manufacturing activity, federal information processing standards (FIPS) codes, and the annual quantity of each toxic pollutant released. A facil ity must report to the database if it meets the following conditions: 1) is involved in manufacturing, metal mining, coal mining, electrical utilities that combust coal and/or oil for the purpose of generating power for distribution in commerce; 2) is part of t he Resource Conservation and Recovery Act (RCRA) Subtitle C hazardous waste treatment and disposal facilities, chemical wholesaler, petroleum PAGE 72 58 terminals and bulk stations, solvent recovery services with certain Standard Industrial Classification Codes (SIC) a federal facility in any SIC code; 3) employs 10 or more full time equivalent employees; and 4) manufactures or processes more than 25,000 pounds or otherwise uses more than 10,000 pounds of any listed toxic chemical during the calendar year (US Environmental Protection Agency 2004c) In spite of providing a comprehensive inventory of annual releases, there are several limitations of the TRI for measuring adverse health effects. First, it reports only releases of chemicals and not exposure or toxicity data. Second, TRI data are self reported and not monitored by the USEPA, which can lead to under and/ or overestimation of actual chemical releases. Finally, not all chemical releases are reported and only those which reach the established thresholds are collected. Additionally, it is important to note that small quantities of highly toxic chemicals may p ersist longer in the environment and be more toxic to the exposed population than certain chemicals that are in larger quantities and less toxic. Therefore, it is important to include fate and transport data, exposure and toxicity levels in order to have an effective tool for assessing risks to a given population. 4.1.1.2 Risk Screening Environmental Indicators for Chronic Human Health Model The Risk Screening Environmental Indicators (RSEI) for Chronic Human Health Model is a screening tool that eval uates the exposure of toxic chemicals released to the environment from industrial sources. It was developed by the USEPA for assessing potential risk to human health based on physiochemical properties of ambient pollutants and related fate and transport in the environment, as opposed to using only the quantity PAGE 73 59 (pounds) of annual releases that are reported by the Toxic Release Inventory (TRI) database. The RSEI is capable of extending environmental health disparities and justice research by adding a much needed health risk assessment component to the TRI database. The RSEI combines fate and transport, toxicity, exposure, population (US Census) with TRI data on a 1 km by 1 km tessellated grid system and models the risk from individual facilities at local, state and national levels (US Environmental Protection Agency 2004a) It combines the amount of chemicals released, the facility location, the toxicity of the chemical, fate and transport through the environment, the route and extent of human exposure, and the number of people affected for up to 44 miles from e ach industrial release source (US Environmental Protection Agency 2004b) The model itself can be adapted for use on other chemicals n ot included in the TRI, thereby expanding the number of substances that can be explored for human health and environment impacts. The RSEI model begins with an Indicator Element which is the combination of the chemical release from the TRI facility an d exposure pathway (e.g., emissions from powerplants). Each Indicator Element is associated with a set of results, which consist of three types of scores: (a) pounds based, (b) hazard based, and (c) riskbased, as summarized in Table 5. Table 5 : RSEI In dicator Element mode Source: (US Environmental Protection Agency 2004d) Results Model Type Risk based results Surrogate Dose Toxicity Weight Population Hazard based results Pounds Toxicity Weight Pounds based results TRI Pounds released PAGE 74 60 Pounds based results are provided for Indicator Elements as the number of pounds associated with each toxic chemica l released annually from each facility reported by TRI. The model also calculates the pounds along with the chemical specific toxicity weights. All of these Indicator Elements can be numerically combined to assess modeled risk of pollution to chemicals, facilities and by exposure pathways on local, state and national levels. Hazard based results are calculated by multiplying pounds released by the chemical specific toxicity weight for the given exposure route (oral or inhalation). It also calculates Mo deled Hazard which is the chemical specific toxicity weighted multiplied by the Modeled Pounds without fate and transport modeling (Table 5). The risk based results score is a unitless value proportional to the potential riskrelated impact of each e lement. Risk based results are obtained by modeling the surrogate dose by using fate and transport and population exposure factors (exposure pathway). The model calculates Indicator Elements for the entire population, children under 10, children 10 to 17, males and females age 18 to 44 and adults over 65. It also calculates Modeled Pounds which is the number of pounds that can be modeled before fate and transport and exposure assumptions are applied (Table 5). The scores for each hazard (i.e., pounds, hazard and risk related results) are based on the aggregation of all chemicals released from each facility and their related toxicity. For each type of pollution indicator (pounds, hazard and modeled risk), a single number is computed for each released chemical and aggregated to represent the total atmospheric emissions from that source (TRI facility). Specifically, for hazard and modeled risk scores, the scores are calculated by determining the toxicity of each chemical and aggregating the resultant score to the TRI facility thereby creating a single score per pollution source. PAGE 75 61 The limitations for the RSEI are that the modeled scores are arbitrary in nature meaning that they are related to each other and toxicity weights may not directly correlate with cr iteria used for listing chemicals in the TRI. Therefore, the RSEI must be used as a screening tool only. Additionally, the Chronic Human Health model is for chronic effects associated with long term exposure and not acute toxicity to humans or environment Other limitations are that the toxicity weights used reflect the single most sensitive chronic human health effect dependent upon inhalation and oral pathways (fish consumption only), and not the most severe health effect or the additive or synergistic effects of exposure to multiple chemicals. These issues can contribute to biases in toxicity calculations by positively or negatively skewing the results and by not accounting for the additive effects of multiple chemicals. For this dissertation research a detailed dataset requested from th e developers of the RSEI was used to a llocate levels of modeled risk to each grid cell (1 km by 1 km). The model calculates concentrations of a 101 km by 101 km grid around the TRI facilities which is then separated i nto 1 km by1 km grids. The USEPA provides crosswalk data which contains coordinates of each grid and the proportion of each census block which is contained in each grid. These data allow for ease of transfer of population data from the census block to the grid level. 4.2 Populations at Risk The census t opologically i ntegrated g eographic e ncoding and r eferencing ( TIGER) database was developed by the US Census to establish official cartographic boundaries for the decennial censuses. This data source c ontains street centerlines, census unit boundaries, and landmarks in digital form. Additionally, data from the 2000 PAGE 76 62 US Census from the summary file 3 (SF3) on the racial/ethnic and economic characteristics of the population was linked with the TIGER/Line files at both the census tract and block group levels of aggregation. These data can be used with GIS tools, specifically ArcGIS, to sp atially relate socioeconomic data to the cartographic boundary files. Presidential Executive Order (EO) 12898 (Clinton 1994) states that each Federal agency shall make achieving environmental justice part of its mission by identifying and addressing, as appropriate, disproportionately high and adverse human health or environmental effects of its programs, policies, and activities on minority populations and low income populations. The literature on e nvironmental justice and health disparities research has used a large variety of variables to represent minority and low income populations as outlined in EO 12898 and reviewed by (Liu 2001) and (McMaster et al. 1997) Minority variables in previous studies have included individuals of Hispanic origin, American Indian/Alaska Natives, Asian, and Blacks or AfricanAmericans. Variables used to measure low income populations and economic status have included the population with a family income bel ow the federal poverty level, median household income, median housing values, home ownership rate, and level of education. This dissertation research includes all racial and ethnic subgroups outlined in the year 2000 US census namely, White alone, Hispan ic alone, American Indian or Alaska Native alone, Asian alone, Black or African American alone, Native Hawaiian or Other Pacific Islander alone, and Other racealone. Racial subcategories under the nonHispanic ethnic classification were used in this rese arch to avoid double counting or overlap of racial and ethnic categories defined by the US Census. Additionally, the PAGE 77 63 overall minority percent was included as a general indicator of the racial/ethnic status of the population; this variable was computed by subtracting the percentage of the Non Hispanic White population from the t otal population. Explanator y variables describing the socioeconomi c characteristics of the population included the percentage of the population below the poverty line labor force, owner occupied houses, and median household income Population density was used a s a control variable based on the assumption that densely populated areas are more likely to contain air polluting facilities and activities. While population density is commonly measured as the number of people per square mile, the natural logarithm (LN) of this value was included to account for the skewed distribution of the variable as suggested by Mennis and Jordan (2005) and Pastor et al. (2005) The percentage of individuals with an annual family income below the federal poverty line (poverty rate) and median household inco me w ere included to measure economic status following num erous prior studies on environmental justice The percentage of occupied housing units that are owner occupied, also known as home ownership rate, was also included as an explanatory variable i n an attempt to distinguish between wealth and income. Wealth measures family assets which can be used in case of economic hardship while income is a measurement of disposable cash (Pastor Jr et al. 2004) Previous studies suggest that an increase in the percent age of owner occupied housing is associated with lower levels of toxic pollution and facility occurrence (Pastor Jr et al. 2004; Pastor Jr et al. 2005) The civilian labor force was also included in this study as an explanatory variable. Pulido (1996) found that census tracts with industrial facilities tend to have a higher percent of their population that are in the labor force PAGE 78 64 4.3 M ethodology for Stat istical Analyses This section focuses on the implementation of methodological and statistical tech niques utilized in this study. Th e initial phase of this research encompasses analytical methods utilized in previous studies of environmental just ice, such as host/nonhost techniques (bivariate analy s is) and ordinary least squares (OLS) regression ( multivariate analysis ) to examine the effects of variables that influence the distribution of pollution sour ces and their adverse impacts. The subsequent phases of this study aim to enhance statistical analysis of environmental justice through the use of spatial regression models that control for the effects of spatial dependence in the data. Each of the four research questions in this study emphasize a n unique quantitative indicator of exposure to air pollution from facilities listed in the USEPAs Toxic Release Inventory (TRI) for 2001 : 1) presence/absence of TRI facilities reporting toxic emissions; 2) the total quantity (in pounds ) of toxic chemicals released into the environment ; 3) the toxicity weighted hazard score associated with released chemicals; and 4) the modeled risk score based on the quantity and physiochemical properties of released chemicals, in conjunction with fate and transport modeling The four research questions thus focus on deter mining if each of these variables are related to relevant explanatory factors that describe the race/ethni c ity and socioeco nomic characteristics of the population in the Houston MSA Linear regression, specifically ordinary least squares (OLS) regression, was used as the statistical technique for exploring the associations between each dependent variable (i.e., facility presence/absence, pounds emitted, hazard scores, and modeled risk scores ) and the set of PAGE 79 65 explanatory variables. To measure each dependent variable, t hree different search radii were implemented for each analytical unit (census tract and block group) to assess exposure to the adverse health impacts of toxic chemical emission s and consider rele ases from f acilities that are situated at or near the boundar y of the unit (edge effect) The se search radii included: 0 mile (at the boundary of the analytical unit) 0.5 mile (0.8 km), and 1 mile (1.6 km). These distances are derived from buffer sizes used in previous studies (Anderton et al. 1994; Glickman 1994; Chakraborty and Armstrong 1997; United Church of Christ Commission for Racial Justice 1987) and recent research which suggests that atmospheric pollution rarely extend s beyond two miles from its source (United Church of Christ Commission for Racial Justice 2007) Racial /ethnic and socioeconomic explanatory variables for multivariate regression were finalized after an initial univariate analysis. Based on the univariate analysis and literature, White alone was removed based on high multicollinearity with Black alone, and median household income was removed based on its high multicollinearity with the other socioeconomic variables such as the percent belo w poverty and owner occupied housing The percentage of Native American s was removed because of its extremely low values across all census units in this study area Furthermore, each dependent outcome variable and the explanatory variables were assessed for the common statistical assumptions of uncorrelated errors, normality and constant variance. Following this assessment, the quantity or pounds of released chemicals toxicity weighted hazard scores and the modeled risk scores were transformed to thei r natural logarithm (LN), while percentages of relevant racial / ethnic categories individuals below the poverty line, and owner occupied housing units were determined. PAGE 80 66 Four final multiple regression models were f ormulated using various combinations of t he individual explanatory variables to explore the statistical effects of different configurations of race/ethnicity and socioeconomic status on the dependent variables (Figure 7 ) All four models contain population density (natural log of the number of people per square mile), one of the strongest predictors of economic activity and industrial pollution (Ash and Fetter 2004; Pastor Jr et al. 2005) Model 1 and 3 co mprise only the race/ethnicity variables with mo del 1 containing the overall percentage of minorities and model 3 containing separate variables that represent the percentage of Blacks, Asians, and Hispanics, respectively. Models 2 and 4 combine the race/ethnicity variables along with the socioeconomic variables and population density. These variables provide a more complete picture in order to help differentiate between whether race/ethnicity is the most influential variable on environmental justice or health disparity issues or if economic variables h ave a contribution. Given that the first wave of environmental justice studies focused primarily on location inequity in the distribution of pollution sources (United Church of Christ Commission for Racial Justice 1987, 2007) the first research question in this dissertation attempts to determine if the presence or absence of TRI facilities are influenced by different racial/ethnic and socioeconomic variables. Binary logistic regression analysis was used to answer this question, with the dependent variable coded as 1 if the analytical unit (cens us tract or block group) contained at least one TRI facilit y and as 0 if it did not contain such a facility based on the three distance definitions around each unit (0, 0.5, and 1 mile ) as mentioned previously For the remaining research questions that focus on pounds emitted, hazard scores, and modeled risk scores, respectively, ordinary least PAGE 81 67 squares (OLS) regression was initially used to determine the statistical association s between aggregated density of the se dependent variables (based on the three di stance definitions) and race/ethnici ty and socioeconomic status. Since the assumptions of OLS regression may not be valid if the residuals (error terms) are spatially autocorrelated further analysis is necessary to assess the effects of spatial dependenc e in the data. It should be noted that the use of buffers (0.5 mile and 1.0 mile) around census units to compute these quantities (pounds emitted, hazard scores, and modeled risk scores) can be expected to introduce an additional amount of spatial clust ering or autocorrelation in the dependent variable that potentially increases the extent of spatial autocorrelation in the regression residuals. While it is not possible to separate these effects or sources of autocorrelation, the results associated with s patial regression when the dependent variable is based on values obtained from neighboring census units must be interpreted with caution. For this study, t he SAR models use a spatial weights matrix which is specified by three configurat ions of neighbors : queen based contiguity, rook based contiguity, and distancebased bands The spatial weights ma trix for queen contiguity was created by selecting neighbor ing areal units that touch the boundary and the vertices of the analytical unit being examined. The r ook contiguity identifies areal units that only touch the boundary of a given unit The identification of neighbors in the distance based spatial weights matrix is based on the inclusion of areal units whose centroids lie within a minimum distance from th e centroid of the analytical unit in question, as computed in GeoDA. In order to create spatial weights matrices using distance bands, several different distance values (3 to 5) were examined to define the set of neighboring areal PAGE 82 68 units for assessing and minimizing spatial autocorrelation. Because census tracts and block groups vary widely in shape and size within a single urban area, a fixed distance band is more likely to ensure a consistent spatial definition of neighbors for detecting and addressing t he effect of spatial dependence. After specifying the spatial weight matrix on the basis of three neighborhood configurations (rook based, queenbased, and distance based) the initial diagnoses for spatial autocorrelation were conducted for the OLS regr ession models First, Morans I was used to determine if spatial autocorrelation existed in the residuals. T he residuals were found to be spatially dependent if they yielded a high coefficient for Morans I that was significantly different from 0 (p<0.05). To incorporate the effect of spatial a utocorrelation, an appropriate spatial autoregressive model ( spatial lag or spatial error ) was utilized. Four statistical tests were used t o determine if the spatial lag approach or spatial error approach would pr ovide the most appropriate correction for each OLS regression model (Anselin 2006) The Lagrange Multiplier (error) was used to assess error dependence and the Lagrange Multiplier (lag) was used to determine if there was a missing spatially lagged variable. Additionally, a robust form of t he Lagrange Multiplier (error) was used to test for error dependence in the possible presence of a missing lagged de pendent variable and the Robust Lagrange Multiplier ( lag ) wa s used to test for a lag ged variable in the presence of a missing error dependent variable Following the selection of the spatial autoregressive term (lag or error), the residuals from the spatial regression model were estimated and Moran s I was again calculated. The residuals were found to be spatially dependent if they still indicated a high coefficient for Morans I that was significantly different from 0, based on an PAGE 83 69 inferential test that uses 999 permutations. On ce the term was corrected for spati al lag or error, minimization of spatial dependence was confirmed by a lower value of Morans I that was not significantly different from zero The best fitted spatial autoregressive ( SAR ) model was first chosen based on the nonsig nificant Morans I with a p value closest to 0. A SAR model with a nonsignificant Morans I improves the OLS model by negating spatial dependence exhibited by the residuals obtained from conventional regression analysis of census data. The maximum likelihood method was used t o fit the SAR models as represented by the Akaike Information Criterion (AIC). The AIC as outlined in (Allison 1999; Grove et al. 2006) helps to select the order of the likelihood of a set of nested or nonnested models. The AIC as represented by ), 2 ) ( log 2 k M L AIC (4) accounts for the fit of the model by using the maximized log likelihood term (log L(M)) as well as the num ber of parameters in the model ( k ). The benefit of this test is that it accounts for the model fit as well as the complexity of the model, thereby providing an effective tool in determining the best fitted model. This test has not previously been used in environmental justice research and can potentially contribute to an improved selection of explanatory variables. T he AIC tests for the fit of the regression model with an enhancement of complexity of the model (number of variables) The AIC alone cannot be used for likelihood examination, but each measure must be compared to other AICs. The model yielding the smallest AIC is considered to be the best fitted model. PAGE 84 70 Lastly, the modifiable areal unit problem (MAUP) was examined for this study by runni ng all regression models at both the censu s tract and block group levels, which allow us to assess the sensitivity of analytical results to the geographic resolution of the data. This issue is critically important in a large metropolitan area such as Hous ton, where variations in size of the geographic analytical units as well as their number (886 census tracts and 2 ,705 block groups ) could potentially lead to different findings regarding inequitable environmental outcomes 4.6 Presentation of Statistical A nalysis Results T he results of environmental justice analys es associated with the four dependent variables (Figure 7 ) are discussed in the following four chapters (chapters 5 to 8). Each chapter begins with a brief introduction of the research question f ollowed by specific methodologies and data used to answer each q uestion, and the results associated with summary statistics bivariate correlation, conventional multiple regression, and spatial autoregressive models, at both the census tract and block group levels. The conclusion chapter combines all the discussions and allows for a comprehensive interpretation of the results PAGE 85 71 Figure 7 : Organizational Chart of Study PAGE 86 72 5. LOCATIONAL ANALYSIS OF INDUSTRIAL POLLUT ION SOURCES This chapter focuses on the anal ysis of environmental justice with regard to the location of facilities with toxic air releases listed in the USEPAs Toxic Release Inventory (TRI). The goal is to examine the association between locations of these industrial pollution sources in the Hous ton MSA and selected demographic and socioeconomic variables from the Census 2000, in order to determine the presence of inequity with respect to the racial/ethnic and poverty status of the population. This analysis begins with an examination of bivariate parametric correlations between the frequency of TRI facilities at three different search radii (0, 0.5, and 1 mile) from each analytical unit (census tract and block group) and explanatory variables described in Chapter 4. Because of the skewed distribu tion of TRI frequency, the square root (SR ) of the number of TRI facilities in each census unit was used for the statistical description and analyses presented in Tables 6 to 9. Summary statistics for all variables analyzed are provided in Tables 6 and 7, respectively, based on census tract and block group level data. Tables 6 and 7 show that census tracts contain a greater number of TRI facilities than block groups and the average number of facilities hosted increase as the search radius increases from 0 to 1 mile from the boundary Pearsons correlation coefficients associated with TRI frequency within the three distances of each analytical unit are presented in Table 8 for census tract level and Table 9 for block group level analyses. The bivariate cor relation analysis at the census tract level (Table 8) shows that at 0 miles from the boundary, percent Hispanic has the highest PAGE 87 73 positive correlation with the frequency of TRIs, followed by the overall minority percentage and percent below the poverty line. Additionally, population density and labor force indicate the greatest negative correlation followed by percent Asian and percent owner occupied housing. Percent Black, however, does not show a significant relationship with the frequency of TRIs at this level. At 0.5 miles from the census tract boundary, percent Hispanic continues to indicate the highest positive correlation in relation to TRI frequency, followed by percent below poverty and percent minority. Additionally, percent Asian has the greates t negative correlation followed by labor force and percent owner occupied housing. Population density and percent Black does not show any significant association with the frequency of TRIs at this level. At 1 mile from the census tract boundary, percent H ispanic continues to indicate highest positive correlation in relation to TRI frequency, followed by percent minority and percent below poverty line. Percent Asian again indicates the greatest negative correlation, followed by labor force and percent owne r occupied housing. Population density and percent Black does not show a significant relationship with the frequency of TRIs at this level. At the block group level (Table 9), the correlational analysis indicates that Hispanic percentage has the highest positive correlation with the frequency of TRIs at 0 miles from the boundary followed by the overall minority percentage and percent below poverty. Additionally, population density has the greatest negative correlation followed by labor force, percent Asi an, and percent owner occupied housing. Percent Black does not show a significant relationship with the frequency of TRIs at this level. PAGE 88 74 At 0.5 miles from the block group boundary, percent Hispanic has the highest positive correlation with TRI frequency f ollowed by the overall percent minority and percent below poverty. Additionally, labor force has the greatest negative correlation followed by percent owner occupied housing and percent Asian. Population density and percent Black do not show a significant association with the frequency of TRIs at this level. At 1 mile from the block group boundary, percent Hispanic has the highest positive correlation with TRI frequency, followed by percent minority and percent below poverty. Additionally, percent Asian had the greatest negative correlation, followed by labor force and percent owner occupied housing. Population density and percent Black does not show a significant relationship with the frequency of TRIs at this level. In summary, the correlation analysi s indicates that the percentage of Hispanics has the strongest and most significant positive linear association with the presence of TRI facilities, at both the census tract and block levels. The environmental inequity hypothesis is supported by three var iables (overall minority percentage, percent Hispanic, and percent below poverty) which consistently exhibit a significant and positive relationship with TRI frequency at all search radii and at both the census tract and block group levels. The percentage of Asians, percent owner occupied housing, and labor force suggest a negative linear relationship at all distances and levels of aggregation. Population density, on the other hand, is negatively correlated with the number of TRI facilities only at the bou ndaries (0 mile search radius) of census tracts and the block groups. Although percent Black is not significantly related to TRI frequency at any level or unit, the overall minority percentage which encompasses percent Black as well as the PAGE 89 75 remaining racia l/ethnic minority groups indicates a significant and positive correlation. This finding underscores the necessity to evaluate race/ethnicity variables both jointly and separately when analyzing the social distribution of environmental pollution sources. While correlation analysis provides a preliminary view of the statistical association between each demographic and socioeconomic variable and TRI frequency, multivariate regression is used to analyze the simultaneous statistical effect s of the explanatory variables on the location of TRI facilities at the census tract and block group level, respectively. The binary logistic regression approach was utilized to model the presence or absence of TRI facilities within each analytical unit in the study ar ea, based on Census 2000 data. Although the number of facilities could have been used as a quantitative dependent variable, there are two reasons why a dichotomous measure (presence/absence) was implemented First, few census tracts and block groups in the stu dy area contained more than one or two TRI facilities. Second, binary logistic regression is a standard approach that has been used to examine inequities in the spatial distribution of pollution sources in previous studies (Daniels and Friedman 1999; Pastor Jr et al. 2004; United Church of Christ Commission for Racial Justice 2007) The dependent variable in this analysis is thus a binary measure of TRI location which was coded as 1 if one or more TRI facilities were present in a analytical unit, and 0 if there were no such facilities in any unit. The logistic regression models were estimated on the basis of the maximum likelihood method. The results associated with each type of analytical unit are described below. PAGE 90 76 5.1 Census Tract Level Analysis The results of the logistic regression analysis at the census tract level, for each of the three distance definitions (0 mile, 0.5 mile, and 1 mile from each tract boundary), are summarized in Tables 10 to 12, respectively. These tables provide the logistic regression coefficients and the odds ratios associated with each independent variable, for the four logistic models that were fitted to predict TRI locations at the tract level. This ratio can be interpreted as a multiplier of the odds of hosting a TRI facility within a census tract. When this odds ratio exceeds one, the probability of hosting a TRI facility increases that many times for each unit of increase in the explanatory variable in the model. An odds ratios smaller than one, in contrast, implies that the probability of hosting a TRI facility decreases with an increase in the value of the explanatory variable. Table 10 summarizes the logistic regression analysis results for TRI location within the boundary of cens us tracts (0 mile). Models 1 and 2 both indicate that the odds of hosting a TRI facility increases significantly with an increase in the overall percentage of minorities, even after controlling for the statistical effect s of socioeconomic variables (model 2). When the minority subgroups are analyzed separately (models 3 and 4), the odds of the hosting a TRI within the census tract increases by 4 and 5 percent, respectively, with every one percent increase in the Hispanic population. This suggests that the minority population in this study area probably comprises a large proportion of Hispanics. Additionally, it stresses the importance of analyzing each racial/ethnic subgroup individually and collectively to gain clarity on their environmental justice imp lications. For models 2 and 4, the odds of TRI facilities being hosted within the census tract significantly increases as the size of the labor force increases. In all four PAGE 91 77 models, the odds of TRI location significantly decreases with an increase in population density. The inclusion of socioeconomic variables in models 1 and 3 do not indicate a significant change in the coefficients or odds ratios of the race/ethnicity variables. At 0.5 mile from the boundary of the census tracts (Table 11), the resul ts for the racial/ethnic variables are similar to those observed at 0 mile from the tracts. For models 1 and 2, the odds of hosting a TRI facility increase significantly as the overall percentage of minority residents increases. Furthermore, for models 3 and 4, the odds of TRI location significantly increase with an increase in the Hispanic percentage. Model 3 is the only regression model where the odds of hosting a TRI facility at 0.5 mile from a census tract increase significantly with an increase in t he percentage of Black residents. Percent Black term becomes non significant, however, when socioeconomic variables are included in model 4. This could be a possible consequence of a strong and positive linear correlation between the Black percentage and poverty rate ( r = 0.52). For both models 2 and 4, the odds of hosting a TRI increases significantly by 4 and 3 percent respectively, with every one percent increase of the percent below poverty. For all models, at 0.5 mile from the tract boundary, the odds of hosting a facility again significantly decrease as the population density increases. With the exception of percent Black in model 4, the addition of socioeconomic variables does not cause a significant change in the coefficients or odds ratios of the race/ethnicity variables. At 1 mile from the boundary of the census tracts (Table 12), models 1 and 2 indicate a positive statistical effect of the overall minority percentage and percent Hispanic (models 3 and 4) on the odds of hosting a TRI facili ty that is consistent with the results at 0 and 0.5 miles from tract boundaries. As observed at a distance of 0.5 mile PAGE 92 78 from the tract boundary, the odds of hosting a TRI facility increases significantly when percent Black increases. For model 2 only, the odds of hosting a TRI facility significantly increases by 3 percent with every one percent increase in the percent below poverty. For all models at 1 mile from the boundary, the odds of hosting a TRI facility significantly decreases as the population density increases. With the exception of percent Black in model 4, the inclusion of socioeconomic variables does not cause a significant change in the coefficients or odds ratios of the race/ethnicity variables. 5.2 Block Group Level Analysis The result s of the logistic regression analysis at the block group level, for each of the three distance definitions (0 mile, 0.5 mile, and 1 mile from each block group boundary), are summarized in Tables 13 to 15, respectively. These tables provide the regression c oefficients and the odds ratios associated with each independent variable, for the four logistic models that were fitted to predict TRI location at the block group level. Table 13 summarizes the logistic regression analysis results for TRI location with in the boundary of block groups (0 mile). At the boundary of the block group, for both models 1 and 2, the odds of hosting a TRI facility significantly increases with an increase in the minority percentage. When the minority subgroups are analyzed separ ately for both models 3 and 4, the odds of hosting a TRI facility significantly increases by 3 and 4 percent respectively, as the percent Hispanic increases by one percent. As with the census tracts, this suggests that the minority population is mostly co mprised of Hispanics in this study area. However, for model 3 alone, the odds of hosting a TRI facility at the block group boundary significantly increases as the Black percentage increases. The Black term becomes non significant when socioeconomic PAGE 93 79 variables are added in model, possibly because of its strong association with lower socioeconomic status. Additionally, it stresses the importance of modeling racial/ethnic terms collectively and individually to gain clarity of their impacts on the dependent variable. For all models, the odds of hosting a facility within the block group boundary significantly decreases as the population density increases. The addition of socioeconomic variables did not suggest a significant change in the coefficients or odds ratios of race/ethnicity variables, except for percent Black in model 4. At 0.5 miles from the block group boundary, for both models 1 and 2 the odds of hosting a TRI facility increases significantly with the overall minority and Hispanic percentages, similarly to what was observed at 0 miles from the block group (Table 14). Additionally, the sign and significance of the Black percentage for model 3 matches the results at 0 mile from the boundary the odds of TRI location increase with percent Black. For models 2 and 4, the odds of hosting a TRI facility decreases with an increase of the percent owner occupied housing. The odds of TRI location significantly decrease by 3.1 and 2.5 percent, respectively as labor force increases, in models 2 and 4. F or all models at 0.5 miles from the block group boundary, the odds of hosting a facility significantly decreases as the population density increases. After the addition of socioeconomic variables to the models there were no significant changes in the coef ficients or odds ratios of race/ethnicity variables except for the percent Black in model 4. At 1 mile from the boundary of block groups (Table 15), the regression results are similar to those observed at 0 and 0.5 miles; the odds of hosting a TRI fac ility significantly increases as the overall minority percent and Hispanic percent increases. PAGE 94 80 Percent Black shows a significant statistical effect in model 3, but not in the presence of other socioeconomic variables in model 4. The odds of hosting a facil ity significantly decrease again with increases in labor force and population density. A notable exception is the percent below poverty which shows a significant and positive statistical effect on TRI location in both models 2 and 4. 5.3 Discussion and Interpretation The results reveal that location of TRI facilities in this MSA is significantly related to population density, the overall minority percentage, and the Hispanic percentage at the census tract level. While densely populated tracts are less likely to contain TRI facilities, the presence of racial/ethnic minorities and Hispanics, in particular, indicate a positive statistical effect on TRI location at all distances, even after controlling for the statistical effect s of other explanatory variab les. The Black percent shows a significant and positive statistical effect but only when socioeconomic variables are not included in the model. This implies that for the Black subgroup, economic status provides a better statistical explanation for the p resence of TRI sites at the tract level. Additionally, the poverty rate and labor force are the only two socioeconomic variables to show significance at the census tract level. Labor force shows a positive statistical effect on the odds of hosting a TRI site only at the boundary which suggests that the presence of a readily available workforce plays a key role in the location of these industrial facilities. Furthermore, the statistical effect of people in poverty is more pronounced only at half mile from the boundary. As observed at the census tract level, population density, minority percent, and percent Hispanic, are again the most influential variables in explaining the location of PAGE 95 81 TRI facilities at the block group level. TRI facilities are signific antly more likely to locate in block groups containing a higher proportion of racial/ethnic minorities and Hispanic populations, in all models and distance definitions. Furthermore, the Black percent has a significant and positive statistical effect on TR I location when socioeconomic variables are not included. As for the other socioeconomic variables, the presence of owner occupied housing significantly increases the odds of hosting a TRI facility at only 0.5 miles, while labor force suggests a negative association at 0.5 and 1 mile. Poverty rate shows a significant and positive statistical effect at the block group level only at 1mile. In summary, the results suggest that statistical effects of race/ethnicity on TRI location become stronger and more si gnificant as the search radius increases at both the census tract and block levels of aggregation. An exception is the Hispanic percentage which retains a significant statistical effect at all three distances and aggregation levels. The significant statistical effect of population density also does not diminish with distance and as the search radius distance increases for both units of analysis, the significance of poverty increases. These findings suggest that although the proportion of Black and below poverty residents has less significance in determining the location of facilities than the Hispanic proportion and population density, the shape and size of the census tract and block group units of analyses may not capture all of the potential effects as sociated with these populations. Consequently, a large number of Blacks and those in poverty probably reside in census tracts and block groups that are adjacent to those hosting TRI sites. Thus the necessity of implementing a more spatially sensitive ana lysis which is not limited by arbitrary boundary configurations is necessary to incorporate the PAGE 96 82 impact of industrial pollution sources which are located at the edge of the boundaries of analytical units. While the analyses presented in this chapter provi de an assessment of inequities in the location of TRI facilities, a more detailed spatial exploration of the quantities of emitted pollutants is necessary to obtain a better understanding of the disproportionate risk burdens imposed by such industrial faci lities in the Houston MSA. PAGE 97 83 Table 6: Descriptive Statistics for Variables at the Census Tract Level. n =886 Simple Statistics Mean Std Dev Min Max Variables TRI frequency (SR) : 0 Miles 0.105 0.306 0.000 2.223 TRI frequency(SR): 0.5 Mile s 0.167 0.293 0.000 1.448 TRI frequency(SR): 1 Mile 0.219 0.284 0.000 1.278 % Black 18.140 24.290 0.000 98.160 % Asian 4.140 5.890 0.000 37.710 % Hispanic 28.380 23.310 0.000 100.000 % Minority 52.700 29.830 0.000 100.000 Population Density (LN) 7.550 8.020 0.000 10.390 % Below Poverty 14.210 10.920 0.000 66.690 % Owner Occupied 61.010 25.440 0.000 100.00 Labor Force (SR) 60.270 16.530 0.000 115.450 PAGE 98 84 Table 7: Descriptive Statistics for Variables at the Block Group Level. n=2 705 Simple Statistics Mean Std Dev Min Max Variables TRI frequency(SR) : 0 Miles 0.064 0.305 0.000 3.794 TRI frequency(SR): 0.5 Miles 0.134 0.301 0.000 1.805 TRI frequency(SR): 1 Mile 0.189 0.295 0.000 1.556 % Black 17.560 25.760 0.000 100.000 % Asian 3.760 6.660 0.000 68.160 % Hispanic 28.860 25.590 0.000 100.000 % Minority 52.210 31.630 0.000 100.000 Population Density (LN) 7.710 1.580 0.000 11.150 % Below Poverty 14.490 12.190 0.000 73.360 % Owner Occupied 63.110 27.430 0.000 100.000 Labor Force (SR) 34.140 10.650 0.000 83.610 PAGE 99 85 Table 8: Pearson Correlation Coefficients at the Census Tract Level Associated with the Square Root of TRI Frequenc y n=886 p<0.05 0 Mile 0.5 Mile 1 Mile Variables r p Value r p Value r p Value % Black 0.050 0.130 0.010 0.800 0.010 0.790 % Asian 0.100 0.004 0.180 <0.001 0.180 <0.001 % Hispanic 0.270 <0.001 0.360 <0.001 0.410 <0.001 % Minority 0.170 <0.001 0.250 <0.001 0.310 <0.001 Population Density (LN) 0.140 <0.001 0.050 0.120 0.010 0.680 % Below Poverty 0.160 <0.001 0.270 <0.001 0.300 <0.001 % Owner Occupied 0.090 0.010 0.110 0.001 0.140 <0.001 Labor Force (SR) 0.140 <0.001 0.150 <0.001 0.170 <0.001 PAGE 100 86 Table 9: Pearson Correl ation Coefficients at the Block Group Level Associated with the Square Root of TRI Frequenc y n=2705 p<0.05 0 Mile 0.5 Mile 1 Mile Variables r p Value r p Value r p Value % Black 0.030 0.160 0.020 0.380 0.004 0.820 % Asian 0.070 0.001 0.110 <0.001 0.150 <0.001 % Hispanic 0.150 <0.001 0.320 <0.001 0.380 <0.001 % Minority 0.100 <0.001 0.230 <0.001 0.290 <0.001 Population Density (LN) 0.130 <0.001 0.002 0.920 0.020 0.240 % Below Poverty 0.090 <0.001 0.220 <0.001 0.280 <0.001 % Owner Occupied 0.070 0.001 0.120 <0.001 0.140 <0.001 Labor Force (SR) 0.120 <0.001 0.130 <0.001 0.140 <0.001 PAGE 101 87 Tables 10: Logistic Regression Analysis of TRI Location at Zero Mile from Census Tract Boundary. n=886 *p<0.05 Independent Variables Model 1 Model 2 Model 3 Model 4 Coefficient OR Coefficient OR Coefficient OR Coefficient OR %Black 1.000 0.011 1.010 0.001 %Asian 0.0210 1.020 0.0270 1.030 %Hispanic 0.043* 1.040 0.046* 1.050 %Minority 0.024* 1.020 0.024* 1.030 Population Density(LN) 0.6 67* 0.513 0.773* 0.462 0.756* 0.470 0.868* 0.420 %Below Poverty 0.018 1.020 0.0156 1.020 %Owner Occupied 0.003 1.000 0.061 1.010 Labor Force(SR) 0.021* 1.020 0.021* 1.020 Akaike Info Criterion 653.394 647.009 619.965 613.141 Log lik elihood 647.394 635.009 609.965 597.141 PAGE 102 88 Tables 11: Logistic Regression Analysis of TRI Location at Half Mile from Census Tract Boundary. n=886 *p<0.05 Independent Variables Model 1 Model 2 Model 3 Model 4 Coefficient OR C oefficient OR Coefficient OR Coefficient OR %Black 0.009* 1.01 0.005 1.010 %Asian 0.033 0.968 0.024 0.977 %Hispanic 0.031* 1.030 0.028* 1.030 %Minority 0.019* 1.020 0.012* 1.010 Population Density(LN) 0.506* 0.603 0.544* 0.580 0.5 06* 0.603 0.569* 0.566 %Below Poverty 0.038* 1.040 0.027* 1.030 %Owner Occupied 0.005 1.010 0.006 1.010 Labor Force(SR) 0.010 1.010 0.009 1.010 Akaike Info Criterion 1037.513 1028.198 993.377 990.811 Log likelihood 1031.513 1016 .198 983.377 974.811 PAGE 103 89 Tables 12: Logistic Regression Analysis of TRI Location at One Mile from Census Tract Boundary. n=886 *p<0.05 Independent Variables Model 1 Model 2 Model 3 Model 4 Coefficient OR Coefficient OR Coefficient OR Coefficient OR %Black 0.009* 1.010 0.007 1.010 %Asian 0.017 0.983 0.012 0.988 %Hispanic 0.033* 1.030 0.032* 1.030 %Minority 0.019* 1.020 0.014* 1.010 Population Density(LN) 0.535* 0.586 0.504* 0.604 0.555* 0.574 0 .541* 0.582 %Below Poverty 0.027* 1.030 0.013 1.010 %Owner Occupied 0.006 1.010 0.007 1.010 Labor Force(SR) 0.001 1.000 0.0001 1.000 Akaike Info Criterion 1152.150 1150.415 1106.742 1108.614 Log likelihood 1146.150 1138.415 10 96.742 1092.614 PAGE 104 90 Tables 13: Logistic Regression Analysis of TRI Location at Zero Mile from Block Group Boundary. n=2705 *p<0.05 Independent Variables Model 1 Model 2 Model 3 Model 4 Coefficient OR Coefficient OR Coefficient OR Coefficient OR %Bla ck 1.010* 1.010 0.008 1.010 %Asian 0.010 1.010 0.004 1.000 %Hispanic 0.030* 1.040 0.035* 1.040 %Minority 0.022* 1.020 0.020* 1.020 Population Density(LN) 0.672* 0.511 0.692* 0.501 0.722* 0.486 0.756* 0.470 %Below Poverty 0.014 1. 010 0.012 1.010 %Owner Occupied 0.001 1.000 0.002 1.000 Labor Force(SR) 0.007 1.010 0.010 1.010 Akaike Info Criterion 1054.772 1057.858 1017.335 1020.365 Log likelihood 1048.772 1045.858 1007.335 1004.365 PAGE 105 91 Tables 14: Logistic Regression Analysis of TRI Location at Half Mile from Block Group Boundary. n=2705 *p<0.05 Independent Variables Model 1 Model 2 Model 3 Model 4 Coefficient OR Coefficient OR Coefficient OR Coefficient OR %Black 0.014* 1.010 0 .005 1.010 %Asian 0.019 0.982 0.020 0.980 %Hispanic 0.042* 1.040 0.033* 1.030 %Minority 0.029* 1.030 0.0198* 1.020 Population Density(LN) 0.445* 0.641 0.242* 0.785 0.507* 0.602 0.339* 0.712 %Below Poverty 0.011 1.010 0.014 1.010 %Owner Occupied 0.009* 0.991 0.009* 0.991 Labor Force(SR) 0.032* 0.969 0.026* 0.975 Akaike Info Criterion 1152.667 1134.017 1079.170 1067.573 Log likelihood 1146.667 1122.017 1069.170 1051.573 PAGE 106 92 Tables 15: Logistic Regression Analysis of TRI Location at One Mile from Block Group Boundary. n=2705 *p<0.05 Independent Variables Model 1 Model 2 Model 3 Model 4 Coefficient OR Coefficient OR Coefficient OR Coefficient OR %Black 0.012* 1.010 0.004 1.000 %Asian 0.017 0.983 0.015 0.985 %Hispanic 0.035* 1.030 0.027* 1.030 %Minority 0.028* 1.020 0.014* 1.010 Population Density(LN) 0.279* 0.756 0.076* 0.986 0.314* 0.574 0.037* 0.964 %Below Poverty 0.016* 1.080 0.014* 1.010 %Owner Occupied 0.003 0.99 7 0.002 0.998 Labor Force(SR) 0.041* 0.960 0.034* 0.966 Akaike Info Criterion 2485.853 2441.003 2348.188 2323.025 Log likelihood 2479.853 2429.003 2338.188 2307.025 PAGE 107 93 6. ANALYZING THE QUANTI TY OF EMISS IONS FROM INDUSTRIAL POLLUTION SOURCES This chapter focuses on the use of pounds of toxic chemicals emitted by facilities listed in the USEPAs Toxic Release Inventory (TRI) as an indicator of potential exposure to chronic air pollution in the analysis of environmental justice in the HoustonGalveston Brazoria metropolitan statistical area (HBG MSA). The goal is to examine the association between the total quantity (pounds) of pollutants released from TRI facilities and selected demographic and socioeconom ic variables from Census 2000, in order to determine the presence of inequity with respect to the racial/ethnic and poverty status of the population. The analysis begins with an examination of bivariate parametric correlations between the density (pounds per square mile) of atmospheric pollutants released at three different search radii (0, 0.5, and 1 mile) and relevant explanatory variables at the census tract and block group levels of aggregation. The density of atmospheric pollution (all emitted chemic als) from TRI facilities were calculated for each analytical unit following the technique for modeling point density described previously in Chapter 4. For each census tract or block group, the areal density of pounds released is measured by dividing the total quantity of pounds emitted from their host facilities by the area enclosed, at three different distances from the tract or block group boundary (0, 0.5, and 1 mile). Because of the skewed distribution of this variable, the natural logarithm (LN) of the density of pounds emitted was used for all statistical analyses presented in this chapter. The use of the natural log reduces outliers and yields a normal style distribution of the dependent variable that is more consistent with the requirements of re gression modeling. Summary PAGE 108 94 statistics for all variables analyzed are provided in Tables 16 and 17 for census tract and block group level data, respectively. The descriptive statistics show that census tracts contain a larger density of pounds of emitted pollutants than block groups. In addition, the density of pounds emitted increase as the search radius increases from 0 to 1 mile from the boundary due to the possible inclusion of more TRI facilities at larger distances from each unit. Pearsons correl ation coefficients associated with the density of pounds emitted within the three distances of each analytical unit are presented in Table 18 for census tract level and Table 19 for block group level analyses. The correlation analysis at the census tract l evel shows that at 0 miles from the boundary, percent Hispanic has the highest positive correlation with the density of pounds emitted followed by the percent below poverty. Additionally, population density has the greatest negative correlation followed b y percent Asian, labor force, and percent Black. Percent owner occupied housing and the overall minority percent does not indicate a significant correlation with the density of pounds emitted at this level. At 0.5 mile from the census tract boundary, perc ent Hispanic again has the highest positive correlation in relation to pounds emitted followed by percent below poverty. Additionally, population density has the greatest negative correlation followed by percent Asian, percent Black and labor force. Percent owner occupied housing and overall minority percent again do not show a significant correlation with pounds emitted at the tract level. At 1 mile from the census tract boundary, percent Hispanic has the highest positive correlation in relation to pounds emitted followed by percent below poverty and PAGE 109 95 overall percent minority. Additionally, percent Asian has the greatest negative correlation followed by population density and labor force. Percent owner occupied housing and percent Black do not show a sig nificant correlation with pounds emitted at this level. At the block group level, the correlation analysis indicates that at 0 miles from the boundary percent Hispanic has the highest positive correlation with the pounds emitted. Additionally, population density has the greatest negative correlation followed by labor force, percent Asian, and percent Black. Percent owner occupied housing, percent below poverty and percent minority do not show a significant correlation with pounds emitted at this level. A t 0.5 mile from the block group boundary, percent Hispanic has the highest positive correlation in relation to pounds emitted followed by percent below poverty and percent minority. Additionally, population density has the greatest negative correlation f ollowed by percent Asian, labor force and percent owner occupied housing. Percent Black does not show a significant correlation with pounds emitted at this level. At 1 mile from the block group boundary, percent Hispanic has the highest positive correlation in relation to pounds emitted followed by percent minority and percent below poverty. Additionally, percent Asian has the greatest negative correlation followed by labor force, population density and percent owner occupied housing. Percent Black does not show a significant correlation with pounds emitted at this level. In summary, the correlation analysis indicates that the percentage of Hispanics has the strongest and most significant positive linear association with the pounds of chemicals released by TRI facilities into the atmosphere, at both the census tract and PAGE 110 96 block levels. The environmental inequity hypothesis is supported by the overall percent Hispanic variable which consistently exhibits a significant and positive relationship with density of pounds emitted at all search radii and at both the census tract and block group levels. The percentage of Asians, population density, and labor force suggest a negative linear relationship at all distances and levels of aggregation. Blacks, on the other hand, have a negative relationship with density of pounds emitted at 0 and 0.5 mile for census tracts and at 0 miles only for block groups. Additionally, the presence of minorities becomes significant as the search radii approach 1 mile (1 mile at bloc k group; and 0.5 and 1 mile at census tract). This suggests that once again the Hispanic populati on has the greatest association with the minority percentage since the coefficient overall for the minority percentage is positive while all of the remaining minority subgroups yield negative coefficients. Additionally, the negative correlation of owner occupied housing at the block group level only reinforces the need to address environmental justice variables at different resolutions. While correlation a nalysis provides a preliminary view of the statistical association between each demographic and socioeconomic variable and density of pounds emitted, it does not clarify the underlying pattern in sufficient detail. Multivariate regression is therefore use d to analyze the simultaneous statistical effect of the explanatory variables on the density of pounds emitted by TRI facilities at the census tract and block group level, respectively. The ordinary least squares (OLS) regression approach is first utiliz ed to model the association of explanatory variables with the density of pounds released from TRI facilities within each analytical unit in the study area, based on the three search radii (0, PAGE 111 97 0.5 and 1 mile). The results from the OLS regression models for the four different combinations of explanatory variables are summarized in the first se t o f tables (Tables 2022 and 3032). This set of tables also provides results from an analysis of spatial dependence using both the contiguity based and distance base d methodologies for identifying spatial neighbors. The results in the OLS tables contain regression coefficients that are bolded, z scores which are in parentheses and p values which are k. Additionally these tables contain the values of Morans I for each model which is used for testing the presence of spatial autocorrelation. The second phase of the analysis uses the spatial autoregressive models (SAR) that include an autoregressive c oefficient to account for spatial autocorrelation. The appropriate SAR function (spatial lag or spatial error) was determined using the Lagrange Multiplier test statistic results (Anselin 2006) obtained while running the corresponding OLS model. As mentioned previously, the SAR models use a spatial weights matrix specified by three configurations of neighboring census units: queenbased contiguity, rookbased contiguity, and distance based bands. Separate queen and rook contiguity analysis was therefore performed in hopes of finding the best fitted model that minimizes the effects of spatial dependence. Additionally, three to five distance bands were constructed to identify neighboring analytical units in order to identify the most appropriate distance from the search radius which yields the best fit model and minimizes the effects of spatial dependence. After the completion of the analyses, the best model was selected based upon the achievement of nonsignifica nt (p<0.05) Morans I which signifies no spatial autocorrelation and the lowest Akaike Information Criterion PAGE 112 98 (AIC) which is used to determine the overall fit of the model. The census tract re sults are listed in Tables 23 24 a nd the block groups in Tables 3340. The results in the tables contain regression coeffic ients which are bolded, z scores which are in parentheses, and pThe histograms in Figure 8 summarize the connectivity distribution for census tracts, based on the num ber of neighbors in each configuration. For queen contiguity Figure 8 (a) shows that the majority of the census tracts have 4 to 6 neighbors and peaks at 6 neighbors For the rook contiguity in Figure 8(b) the number of neighbors peaks at 4 but the cens us tracts noticeably had a positive skew. The histogram for the distance based configuration in Figure 8(c) provides the connectivity distribution at 0.50 mile from the centroid of each unit and indicates that a majority of census tracts have no neighbors at this distance The connectivity distributions for census block groups are summarized in Figure 9 The histogram for queen contiguity in Figure 9 ( a ) shows that the majority of the block groups have 5 to 7 neighbors and peaks at 6 neighbors For the r ook contiguity in Figure 9 ( b) the number of neighbors peaks at 5 and noticeably ha s a positive skew. The histogram for the distance based configuration in Figure 9( c ) provides the connectivity distribution at 0.19 mile from the centroid of each unit and indicates that the number of block groups with no neighbors at this distance is considerably higher than the number of census tracts with no neighbors in Figure 9(c) PAGE 113 99 Figure 8 : C ensus t ract connectivity histograms used for modeling pounds emitt ed a. Queen Contiguity b. Rook Contiguity c. Distance Based Configuration (0.50 miles) PAGE 114 100 Figure 9: Block group connectivity histograms used for modeling pounds emitted d. Queen Contiguity e. Rook Contiguity f. Distance Based Configuration (0.19 miles) PAGE 115 101 6.1 Census Tract Level Analysis Ordinary Least Squares Regression This section focuses on multivariate regression analysis of the areal density of pounds emitted a t the census tract level, at zero (0), half (0.5) and one (1) mile search radii, based on the application of ordinary least squares (OLS) and spatial autoregressive (SAR) techniques. Table 20 summarizes the results of OLS regression analysis at the bounda ry of the census tracts. Models 1 and 2 both indicate that the overall minority percent shows a significant and positive statistical effect on the density of pounds released. Additionally, when the minority subgroups are analyzed separately for models 3 and 4, the Hispanic percent indicates a significant and positive association with the density of pounds emitted. This suggests that the minority population of tracts with larger quantities of TRI emissions probably comprises a large proportion of Hispani cs. Additionally, it stresses the importance of analyzing each racial/ethnic sub group individually and collectively to gain clarity on their environmental justice implications. Percent Black term for model 3 is significant and becomes nonsignificant wh en socioeconomic variables are included in model 4. As mentioned previously, this could be a possible consequence of a strong and positive linear correlation between the Black percentage and poverty rate ( r = 0.52) at the tract level. In all four models, the density of pounds emitted significantly increases with a decrease in population density. At 0.5 mile from the boundary of the census tracts (Table 21), the results for the racial/ethnic variables are similar to those observed at 0 mile from the tracts. For models 1 and 2, the density of pounds emitted increases significantly as the overall percentage of minority residents increases. Furthermore, for models 3 and 4, the Hispanic percentage PAGE 116 102 has a significant and positive statistical effect on the densit y of pounds released. For both models 2 and 4, the density of pounds emitted increase significantly as the overall labor force increases. For all models, at 0.5 mile from the tract boundary, the density of pounds emitted significantly increases with a de crease in population density. At 1 mile from the boundary of the census tracts (Table 22), models 1 and 2 indicate a positive statistical effect of the overall minority percentage and percent Hispanic (models 3 and 4) on the density of pounds emitted which is consistent with the results at 0 and 0.5 miles from tract boundaries. As observed at a distance of 0 mile from the tract boundary, percentage Black becomes nonsignificant when socioeconomic variables are included (model 4). For model 2 only, the density of pounds emi tted is positively associated with percent below poverty. As with the 0 and 0.5 mile from the tract boundary, the density of pounds emitted significantly increases with a decrease in population density. Morans I for the regression models based on queen and rook contiguities show significance at 0, 0.5, and 1 miles search radii indicating the presence of significant spatial autocorrelation in the data and justifying the need to account for spatial dependence by using a SAR model. However, at 0.5 mile from the boundary when using the distance band for both 0 and 0.5 miles search radii, the Morans I was not statistically significant. Consequently, spatial dependence was not detected and the spatial autoregressive function was not implemented. However, at the 1 mile search radii for the 0.5 mile distance band, models 1 and 2 only were found to be significant indicating the presence of significant spatial autocorrelation in the data and the need to use the SAR model. As a result of the statistically significant Morans I, SAR was run to account for PAGE 117 103 spatial dependence. This distance band configuration was chosen based on the best fit using AIC and the achievement of nonsignificant spatial autocorrelation based on Morans I. Following the assessment for spatial dependence, the SAR model was thus implemented for all OLS models in which significant spatial autocorrelation was detected. Spatial Regression At the boundary of the census tract level for the queen contiguity (Table 23), model s 1 and 2 indicate a significant and positive association between the density of pounds at 0 miles and the overall minority percent. Additionally, for models 3 and 4, the Asian and Hispanic populations positively affect the density of pounds released. Labor force however, has a positive statistical effect on density of pounds emitted only for model 2 which ceases to be significant when the minority subgroups are separated. For all models 0 miles from the tract boundary, density of pounds emitted significantly increases when population density decreases. The queen contiguity SAR models differs from OLS in that Black percentage becomes nonsignificant while percent Asian and labor have a positive statistical effect on the density of pounds released. At the boundary of the census tract level for the rook contiguity (Table 24), all SAR models exhibit the same statistical associations as observed in the model for queen contiguity and had similar differences from the corresponding OLS models. Model s for th e distance band at 0.5 mile were not run because Morans I from the OLS regression were not significantly different from zero when determining spatial autocorrelation. At 0.5 mile from the boundary of census tracts (Table 25) for the queen contiguity mode ls 1 and 2, percent minority has a positive association with the density of pounds PAGE 118 104 emitted. Additionally, for models 3 and 4 the percent of Asians and Hispanics independently have a significantly positive statistical effect on the density of pounds emitte d, as well as labor force for models 2 and 4. For all models at the boundary, population density is a negative predictor of the density of pounds emitted. At the queen contiguity SAR models differ from OLS in that Asian percentage for models 3 and 4 are s ignificant and have a positive statistical effect on the density of pounds emitted unlike the OLS model. For the rook contiguity models at 0.5 mile from the boundary of census tracts (Table 26), percent minority, percent Asian, and labor force exhibit the same positive statistical effect as the queen contiguity, while population density has the same negative prediction for density pounds emitted. Percent Hispanic, however, has a positive statistical effect on the density of pounds emitted at model 3, but not in model 4 when the socioeconomic variables are introduced. This suggests that economic factors have a greater influence on the density of TRI emissions than the presence of Hispanics when the minority subgroups are examined separately. The rook co ntiguity SAR models differ from OLS in that Asian percentage for models 3 and 4 are significant and have a positive statistical effect on the density of pounds emitted unlike the OLS model and that percent Hispanic becomes nonsignificant for model 4 in th e SAR model. Models for the distance band at 0.5 mile were not run be cause the distance band at 0.5 mile in the OLS regression was found to be nonsignificant for spatial autocorrelation using Morans I. At 1 mile from the boundary of census tracts (Table 27) for queen contiguity, percent minority positively associated with the density of pounds emitted, in model 1. PAGE 119 105 Additionally, for models 3 and 4, the Hispanic population has a positive statistical effect on the density pounds emitted. For all models at 1 mile from the boundary, population density is a negative predictor of the density of pounds emitted. The queen contiguity SAR models differ from the OLS model in that Black percentage is nonsignificant. In model 2 only, percent minority and percent be low poverty become non significant. At 1 mile from the boundary of the census tracts (Table 28) for the rook contiguity, percent minority and percent Hispanic exhibits the same positive statistical effect as the queen contiguity, while population density has the same negative statistical effect on pounds emitted. The SAR model exhibits the same difference for the rook contiguity at 1 mile as for the queen contiguity. A distance band of 0.37 mile was constructed for models 1 and 2 only beginning at 1 mile fr om each census tract boundary to identify neighboring tracts for assessing spatial dependence (Table 29). This distance was chosen based on a selection criteria that comprised the minimization of the likelihood term (AIC) and significance of spatial autoc orrelation using Morans I. Percent Hispanic in models 3 and 4 and percent minority in model 1 exhibit the same positive statistical effects on density of pounds emitted as observed in the rookbased contiguity model, while population density has the same negative statistical effect. Additionally, percent below poverty in model 2 and percent Black in model 3 is positively associated with the density of pounds emitted. The 0.37 mile distance band did not show any difference for SAR as compared to the OLS models. PAGE 120 106 6.2 Block Group Level Analysis Ordinary Least Squares Regression This section focuses on multivariate regression analysis of the areal density of pounds emitted at the block group level, at zero (0), half (0.5) and one (1) mile search radii base d on the application of ordinary least squared regression and spatial autoregressive techniques. Table 30 summarizes the results of OLS regression analysis for at the boundary of the block groups. Models 1 and 2 both indicate that the overall percent min ority shows a significant and positive statistical effect on the density of pounds released. Additionally, when the minority subgroups are analyzed separately for models 3 and 4, the Hispanic percent indicates a significant and positive association with the density of pounds emitted. This suggests that the minority population at the block group level with larger quantities emissions probably comprises a large proportion of Hispanics. Additionally, it stresses the importance of analyzing each racial/ethn ic sub group individually and collectively to gain clarity on their environmental justice implications. Percent Black term becomes non significant, however, when socioeconomic variables are included in model 4. This could be a possible consequence of a st rong and positive linear correlation between the Black percentage and poverty rate ( r = 0.45) at the block group level. In all four models, the density of pounds emitted significantly increases with a decrease in population density. At 0.5 mile from the b oundary of the block groups (Table 31), the results for the racial/ethnic variables are similar to those observed at 0 mile from the block groups. For models 1 and 2, the effect on density of pounds emitted increases significantly as the overall percentag e of minority residents increases. Furthermore, for models 3 and 4, the Hispanic percentage has a significant and positive statistical effect on the density of PAGE 121 107 pounds emitted while in model 3 percent Black retains a positive statistical effect on density of pounds. For both models 2 and 4, the density of pounds emitted increases significantly as the overall percent below poverty and labor force increase. For all models, at 0.5 mile from the block group boundary, the density of pounds emitted increases si gnificantly with a decrease in population density. At 1 mile from the boundary of the block groups (Table 32), models 1 and 2 indicate a positive statistical effect of the overall minority percentage, along with percent Black at model 3 and percent Hispanic (models 3 and 4) on the density of pounds emitted that is consistent with the results at 0 and 0.5 miles from block group boundaries. As with 0.5 mile the density of pounds emitted is positively affected by percent below poverty. For all models, at 1 mile from the block group boundary, the density of pounds emitted significantly increases with a decrease in population density. Morans I for the regression models based on queenbased and rookbased contiguities show significance at 0.5 mile and 1 m iles search radii indicating the presence of significant spatial autocorrelation in the data and justifying the need to account for spatial dependence by using the SAR model. However, at 0.19 mile from the boundary when using the distance band for 0 miles search radii only, the distance band was found to yield a nonsignificant Morans I. Consequently, spatial dependence was not detected and the spatial autoregressive function was therefore not implemented. This distance band configuration was chosen base d on the best fit using AIC and the achievement of nonsignificance for spatial autocorrelation using Morans I. Following the assessment for spatial dependence, the SAR model was thus implemented for all OLS models in which significant spatial autocorrelation was detected. PAGE 122 108 Spatial Regression At the boundary of the block group level (Table 33) for the queen contiguity in models 1 and 2, a significant and positive association is observed between the density of pounds at 0 miles and the overall minority perc ent. Additionally, in models 3 and 4 the Hispanic population positively predicts the amount of pounds released. Percent Black however, positively affects density of pounds emitted at model 3, but not when the economic terms were introduced. For all models, the density of pounds emitted significantly increases with a decrease in population density. The queen contiguity SAR models differs from OLS in that owner occupied housing is nonsignificant for models 2 and 4 when the SAR term is introduced. At the boundary of the block group level (Table 34) for the rook contiguity, all SAR models exhibit the same statistical associations as observed in the model for queen contiguity and had similar differences from the corresponding OLS models. The rook contiguit y differed from the OLS model the same as the queen contiguity where owner occupied housing is nonsignificant for models 2 and 4 when SAR is introduced. Models for the distance band at 0 miles was not run be cause the distance band at 0.19 mile for all m odels in the OLS regression was found to be nonsignificant when testing for spatial autocorrelation using Morans I. At 0.5 mile from the boundary of the block group level (Table 35) for the queen contiguity, for models 1 and 2 percent minority positivel y predict the density of pounds emitted. Additionally, models 3 and 4, the Hispanic population and labor force (models 2 and 4) independently and significantly has a positive statistical effect on density of pounds released. Percent Asian however, positi vely predicts density of pounds emitted at model 3, but is not a predictor in model 4 when the economic term was introduced. PAGE 123 109 Percent below poverty positively affects the density of pounds emitted at model 4, but not in model 3 when there was not an economic term. For all models, the density of pounds emitted significantly increases with a decrease in population density. The queen contiguity SAR models differ from OLS in that in percent Black in model 3 and percent below poverty in model 2 are no longer significant, but percent Asian gains significance and has a positive statistical effect on density of pounds released. At 0.5 mile from the boundary of block groups (Table 36) in the rook contiguity model, percent minority, percent Asian, percent Hispanic and labor force exhibit the same positive statistical effect as the queen contiguity, while population density has the same negative prediction on density of pounds emitted. However, percent below poverty and percent owner occupied housing for both models 2 and 4 have positive statistical effect s on density of pounds emitted. The rook contiguity SAR models differ from OLS in that the percent Black in model 3 is no longer significant. However, percent Asian gains significance in model 3 without the economic terms and percent owner occupied for both model significant and have a positive statistical effect on density of pounds released. A distance band of 0.19 mile (Table 37) was constructed beginning from 0.5 mile from each block group boundary to ident ify neighboring tracts for assessing spatial dependence. Percent minority, percent Hispanic, and percent below poverty exhibit the same positive statistical effect of density of pounds emitted as the rookbased contiguity model, while population density has the same negative association reported earlier. Percent Black, however, positively predicts pounds emitted in model 3, but is not a significant predictor in model 4 when the socioeconomic terms are introduced. PAGE 124 110 Additionally, labor force positively pre dicts density of pounds emitted in model 4, but is not a predictor in model 3 when socioeconomic variables are included. The SAR models for the distance based model at 0.19 mile from the boundary differ from OLS in that labor force in exhibits a positive and significant statistical effect in model 4, but ceases to be significant in model 3. At 1 mile from the boundary (Table 38) of block groups for the queen contiguity models 1 and 2, percent minority has similar statistical effect as in the SAR models f or 0 and 0.5 miles. Additionally, in models 3 and 4 the Hispanic population has a positive association with the density of pounds released while labor force has a positive association in models 2 and 4. For all models, the density of pounds emitted sign ificantly increases with a decrease in population density. The queen contiguity SAR models differ from OLS in that in percent Black in model 3 no longer significant. Additionally, percent Asian (models 3 and 4) and percent below poverty (models 2 and 4) l ose significance and no longer contribute to density of pounds released. Labor force on the other hand shows a significant statistical effect on emission density for models 2 and 4. At 1 mile from the boundary of block groups for the rook contiguity, perc ent minority, percent Hispanic, and labor force exhibit the same positive statistical effect on the density of pounds emitted as the queen contiguity, while population density has the same negative statistical effect on the density of pounds emitted. The rook contiguity SAR models exhibit the same difference from OLS as the queen contiguity A distance band of 0.19 mile (Table 40) was constructed beginning from 1 mile from the block group boundary. Percent minority and percent Hispanic show the same PAGE 125 111 posi tive prediction as the rook contiguity, while population density had the same negative statistical effect on the density of pounds emitted. Additionally, percent below poverty is positively associated with the density of pounds emitted at model 2 and 4 and percent Black at model 3. Percent Asian however, is negatively associated with the density of pounds emitted at models 3 and 4. The distance based SAR models do not exhibit any significant difference from the OLS models. 6.3 Discussion and Interpret ation The analyses presented on this chapter focuses on the geographic distribution of pollution burdens measured in terms of the quantity of emissions from industrial point sources. The results reveal that the areal density of pounds released by TRI facil ities in the HoustonMSA are significantly associated with population density, the overall proportion of minorities, and the Hispanic proportion at the census tract level, regardless of whether conventional or spatial regression was utilized. While dense ly populated tracts are less likely to contain TRI facilities and associated emission quantities, the presence of racial/ethnic minorities and Hispanics, in particular, indicate a positive statistical effect on the density of pounds released at all distanc es, even after controlling for the statistical effect s of other explanatory variables. The percentage of Asians living within or near census tract boundaries are only significant when spatial autocorrelation is removed. The Black percentage shows a signi ficant and positive statistical effect in two separate cases. The first case occurs only when socioeconomic variables are not included in the model, near the tract boundary and when ordinary regression modeling is applied. This implies that for the Black sub group, economic status provides a better statistical explanation for the presence of pounds at the tract level. The second case where Black PAGE 126 112 percentage is significant is when the density of emissions is measured at one mile from the tract boundary, for both ordinary and spatial regression. This suggests that less affluent Blacks and those in poverty could be residing closer to TRI facilities with higher quantities of emissions than more affluent Blacks. Furthermore, labor force and poverty rate are t he only two socioeconomic variables to show significance at the census tract level. Labor force shows a positive statistical effect on the density of pounds only when spatial autocorrelation is accounted for. This suggests that the presence of a readily available workforce plays a key role in the location of these industrial facilities. As observed at the census tract level, population density, minority percent, and percent of Hispanics, are again the most influential variables related to the density of pounds emitted at the block group level. There tends to be a significantly greater density of pounds in block groups containing a higher proportion of racial/ethnic minorities and Hispanic populations, in all models and distance definitions. Furthermore, unlike the census tract level, the Black population is significantly associated with emission density at all distances with and without the adjustment for spatial autocorrelation. This suggests that a higher concentration of both Black and Hispanic resi dents can be found near TRI facilities with greater density of pounds, at the block group level. Again poverty rate and labor force are the only two socioeconomic variables to show a significant statistical effect on emission density at distances of 0.5 a nd 1 mile from the block group boundaries, for both conventional and spatial regression techniques. In summary, the results suggest that statistical effects of race/ethnicity on density of pounds released are influential at all search radii and at both the census tract and block levels of aggregation. More specifically, the overall minority and Hispanic population PAGE 127 113 and the number of people per square mile are the most significant explanatory factors in determining the density of pounds released at all se arch radii, all aggregations, and both regression modeling techniques. Additionally, the proportion of Asians is only significant when spatial autocorrelation is removed and the significance of labor force increases when spatial dependence is accounted fo r as well. This finding indicates how spatial autocorrelation can mask the effects of specific variables (e.g., percent Asian and labor force) when determining their statistical association with the magnitude of pollution and justifies the need to control for spatial dependence in regression analysis of environmental justice. The proportion of Blacks was found to be significant when emission density was measured at the boundaries of census tracts, but not at the block group boundaries. Additionally, the Black sub group was significant only at the farthest distance (1 mile) away from tract boundaries and at all three distances from the block group boundaries, for both regression modeling techniques. This confirms the importance of analyzing data at more t han one spatial resolution because the nuances between the distances were not evident at a single analytical unit. While the analyses presented in this chapter provide an assessment of inequities in the distribution of TRI emission amounts, it is important to consider that the density of pounds released into the atmosphere does not account for how ambient chemicals metabolically impact humans. The introduction of toxicity weighted hazard scores, therefore, is necessary to address this limitation and provide a more refined assessment of disproportionate risk burdens imposed on the population in the Houston MSA. PAGE 128 114 Table 16: Descriptive Statistics for Variables at the Census Tract Level n =886 Simple Statistics Mean Std Dev Min Max Variables D ensity of Pounds(LN): 0 Miles 1.180 4.590 0.000 20.710 Density of Pounds(LN): 0.5 Miles 1.690 4.300 0.000 19.620 Density of Pounds(LN): 1 Mile 5.410 6.360 0.000 20.310 % Black 18.140 24.290 0.000 98.160 % Asian 4.140 5.890 0.000 37.710 % Hispanic 28.380 23.310 0.000 100.000 % Minority 52.700 29.830 0.000 100.000 Population Density (LN) 7.550 1.590 0.000 10.390 % Below Poverty 14.210 10.920 0.000 66.690 % Owner Occupied 61.010 25.440 0.000 100.000 Labor Force (SR) 60.270 16.530 0.000 115.450 PAGE 129 115 Tab le 17: Descriptive Statistics for Variables at the Block Group Level n=2705 Simple Statistics Mean Std Dev Min Max Variables Density of Pounds(LN): 0 Miles 0.770 3.200 0.000 23.200 Density of Pounds(LN): 0.5 Miles 2.290 5.000 0.000 21.990 Density of Pounds(LN): 1 Mile 3.990 5.990 0.000 21.060 % Black 17.560 25.760 0.000 100.000 % Asian 3.760 6.660 0.000 68.160 % Hispanic 28.860 25.590 0.000 100.000 % Minority 52.210 31.630 0.000 100.000 Population Density (LN) 7.710 1.580 0.000 11.150 % Below Poverty 14.490 12.190 0.000 73.360 % Owner Occupied 63.110 27.430 0.000 100.000 Labor Force (SR) 34.140 10.650 0.000 83.610 PAGE 130 116 Table 18: Pearson Correlation Coefficients Associated with the Density of Pounds (Na tural Log) Released at the Census Tract Level. n =886 p<0.05 0 Mile 0.5 Mile 1 Mile Variables r p Value r p Value r p Value % Black 0.080 0.020 0.080 0.020 0.020 0.530 % Asian 0.110 0.002 0.110 0.002 0.170 <0.001 % Hispanic 0.160 <0.001 0.150 <0.001 0.260 <0.001 % Minority 0.060 0.110 0.050 0.170 0.170 <0.001 Population Density (LN) 0.300 <0.001 0.310 <0.001 0.160 <0.001 % Below Poverty 0.080 0.020 0.070 0.040 0.170 <0.001 % Owner Occupied 0.020 0.520 0.030 0.400 0.040 0.280 Labor Force (SR) 0.100 0.003 0.100 0.003 0.130 0.001 PAGE 131 117 Table 19: Pearson Correlation Coefficients Associated with the Density of Pounds (Natural Log) Released at the Block Group Le vel n =2705 p<0.05 0 Mile 0.5 Mile 1 Mile Variables r p Value r p Value r p Value % Black 0.050 0.010 0.040 0.060 0.030 0.170 % Asian 0.070 0.001 0.110 <0.001 0.130 <0.001 % Hispanic 0.080 <0.001 0.220 <0.001 0.270 <0.001 % Minority 0.030 0.170 0.140 <0.001 0.210 <0.001 Population Density (LN) 0.280 <0.001 0.160 < 0.001 0.110 <0.001 % Below Poverty 0.080 0.060 0.150 <0.001 0.180 <0.001 % Owner Occupied 0.010 0.530 0.060 0.003 0.060 0.001 Labor Force (SR) 0.120 <0.001 0.110 <0.001 0.120 <0.001 PAGE 132 118 Table 20: Least Squares Regression Results for Density of Pounds Released (Natural Log) at Zero Miles from Census Tract Boundary Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.0085* (1.380) 0.003 (0.374) % Asian 0.037 0.041 (1.440) (1.55) % Hispanic 0.061* 0.055* (8.920) (6.090) %Minority 0.032* 0.025* (5.960) (3.210) Population Density(LN) 1.130* 1.170* 1.220* 1.240* ( 11.290) ( 10.620) ( 11.890) ( 11.290) % Below Poverty 0.027 0.026 (1.240) (1.120) % Owner Occupied 0.107 0.004 (0.107) (0.546) Labor Force (SR) 0.011 0.007 (1.220) (0.762) R 2 0.128 0.130 0.170 0.171 F Statistic 64.600* 26.370* 45.000* 25.940* Akaike Info Criterion 5097.290 5100.570 5057.580 5061.730 Log likelihood 2545.640 2544. 280 2523.790 2522.870 Morans I queen 0.185* 0.188* 0.153* 0.156* Morans I rook 0.179* 0.183* 0.144* 0.148* Morans I 0.50 mile 0.003 0.002 0.002 0.004 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 133 119 Table 21: Least Squar es Regression Results for Density of Pounds Released (Natural Log) at Half Mile from the Census Tract Boundary a. All race variables represent a single rac e. b. *p<0.05, **p<0.10 Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.004 (0.682) 0.004 (0.458) % Asian 0.032 0.037 (1.290) (1.490) % Hispanic 0.0520* 0.051* (8.050) (5.920) % Minority 0.025* 0.002* (5.090) (3.170) Population Density(LN) 1.290* 1.490* 1.390* 1.580* ( 10.650) ( 10.360) ( 11.180) ( 10.990) % Below Poverty 0.023 0.022 (1.140) (1.020) % Owner O ccupied 0.005 0.007 (0.733) (1.180) Labor Force (SR) 0.030* 0.027* (3.020) (2.780) R 2 0.115 0.125 0.150 0.164 F Statistic 57.230* 25.240* 40.020* 24.550* Akaike Info Criterion 4997.450 4992.680 4961.510 4957.070 Log likeli hood 2495.730 2490.340 2475.750 2470.540 Morans I queen 0.186* 0.192* 0.156* 0.163* Morans I rook 0.179* 0.187* 0.146* 0.155* Morans I 0.50 mile 0.005 0.003 0.001 0.002 PAGE 134 120 Table 22: Least Squares Regression Results for Density of Pounds Released (Natural Log) at One Mile from the Census Tract Boundary Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.014** (1.680) 0.014 (1.1 60) % Asian 0.048 0.040 ( 1.290) ( 1.060) % Hispanic 0.091* 0.092* (9.610) (7.060) % Minority 0.050* 0.038* (6.750) (3.330) Population Density(LN) 1.430* 1.460* 1.450* 1.500* ( 6.930) ( 5.520) ( 6.910) ( 5.850) % Below Pover ty 0.056* 0.015 (1.810) (0.476) % Owner Occupied 0.010 0.010 (1.060) (1.120) Labor Force(SR) 0.010 0.008 (0.612) (0.510) R 2 0.072 0.076 0.132 0.134 F Statistic 34.390* 14.560* 33.570* 19.400* Akaike Info Criterion 57 31.710 5733.730 5676.450 5680.740 Log likelihood 2862.860 2860.870 2833.230 2832.370 Morans I queen 0.536* 0.537* 0.502* 0.505* Morans I rook 0.570* 0.571* 0.539* 0.541* Morans I 0.50 mile 0.012* 0.013* 0.006 0.006 a. All race variables represe nt a single race. b. *p<0.05, **p<0.10 PAGE 135 121 Table 23: Spatial Regression Results for Density of Pounds Released (Natural Log) at Zero Mile of Census Tract Boundary Using Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.012 0.005 ( 1.530) (1.560) % Asian 0.057** 0.061* (1.890) (2.000) % Hispanic 0.060* 0.052* (7.220) (4.920) % Minority 0.003* 0.027* (4.950) (2.910) Population Density(LN) 1.150* 1.250* 1.220* 1.300* ( 9.960) ( 9.68 0) ( 10.680) ( 10.310) % Below Poverty 0.033 0.038 (1.420) (1.610) % Owner Occupied 0.005 0.915 (0.717) (0.360) Labor Force(SR) 0.018** 0.014 (1.880) (1.480) Spatial Autoregressive Coefficient 0.387* 0.397* 0.342* 0.354* Akaike Info Criterion 5029.910 5030.330 5009.110 5010.090 Log likelihood queen 2511.956 2509.167 2499.550 2497.040 Moran's I queen 0.009 0.009 0.006 0.007 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 136 122 Tab le 24: Spatial Regression Results for Density of Pounds Released (Natural Log) at Zero Mile of Census Tract Boundary Using the Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.008 (1.380) 0.004 (0.440) % Asian 0.05 1* 0.062* (2.010) (2.050) % Hispanic 0.051* 0.052* (7.420) (5.050) % Minority 0.033* 0.027* (5.030) (2.960) Population Density(LN) 1.100* 1.200* 1.050* 1.260* ( 9.710) ( 9.510) ( 10.110) ( 10.240) % Below Poverty 0.0310 0.037 (1.330) (1.540) % Owner Occupied 0.004 0.005 (0.556) (0.780) Labor Force(SR) 0.019* 0.0150 (2.000) (1.560) Spatial Autoregressive Coefficient 0.323* 0.333* 0.268* 0.289* Akaike Info Criterion 5043.770 5044.140 5050.190 5022.380 Log likelihood rook 2518.890 2516.070 2504.730 2503.190 Moran's I rook 0.014 0.014 0.015 0.011 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 137 123 Table 25: Spatial Regression Results for Density of Pounds Released (N atural Log) at Half Mile of Census Tract Boundary Using the Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.004 (0.783) 0.005 (0.505) % Asian 0.046** 0.057* (1.950) (1.970) % Hispanic 0.041* 0.047* (6 .550) (4.720) % Minority 0.027* 0.025* (4.250) (2.850) Population Density(LN) 1.220* 1.580* 1.190* 1.650* ( 9.140) ( 9.400) ( 9.610) ( 9.980) % Below Poverty 0.031 0.036 (1.430) (1.620) % Owner Occupied 0.008 0.090 (1.250) (1. 470) Labor Force(SR) 0.036* 0.033* (3.520) (3.320) Spatial Autoregressive Coefficient 0.390* 0.402* 0.331* 0.363* Akaike Info Criterion 4929.080 4919.690 4910.900 4901.780 Log likelihood queen 2461.540 2453.850 2449.240 2442.890 Moran' s I queen 0.008 0.009 0.001 0.006 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 138 124 Table 26: Spatial Regression Results for Density of Pounds Released (Natural Log) at Half Mile of Census Tract Boundary Using the Rook Contigui ty Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.004 (0.797) 0.003 (0.338) % Asian 0.046** 0.050* (1.940) (2.070) % Hispanic 0.042* 0.041 (6.690) (4.790) % Minority 0.026* 0.025* (4.260) (2.880) Population De nsity(LN) 1.170* 1.510* 1.190* 1.350* ( 8.830) ( 9.190) ( 9.640) ( 9.300) % Below Poverty 0.029 0.022 (1.330) (1.070 % Owner Occupied 0.007 0.004 (1.080) (0.744) Labor Force(SR) 0.036* 0.022* (3.520) (2.350) Spatial Autoregre ssive Coefficient 0.327* 0.337* 0.290* 0.280* Akaike Info Criterion 4943.310 4934.060 4918.240 4917.070 Log likelihood rook 2468.650 2461.030 2453.120 2449.540 Moran's I rook 0.014 0.114 0.016 0.003 a. All race variables represent a singl e race. b. *p<0.05, **p<0.10 PAGE 139 125 Table 27: Spatial Regression Results for Density of Pounds Released (Natural Log) at One Mile of Census Tract Boundary Using the Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.003 0.00 1 ( 0.517) ( 0.181) % Asian 0.013 0.014 (0.511) (0.557) % Hispanic 0.026* 0.028* (4.030) (3.140) % Minority 0.009** 0.009 (1.850) (1.250) Population Density (LN) 0.628* 0.732* 0.718* 0.814* ( 4.560) ( 4.130) ( 4.990) ( 4.520 ) % Below Poverty 0.004 0.001 (0.187) (0.056) % Owner Occupied 0.001 0.001 ( 0.060) (0.227) Labor Force(SR) 0.011 0.010 (1.020) (0.914) Spatial Autoregressive Coefficient 0.829* 0.829* 0.814* 0.814* Akaike Info Criterion 5129.5 30 5134.470 5118.730 5123.750 Log likelihood queen 2560.770 2560.230 2553.370 2552.880 Moran's I queen 0.004 0.001 0.003 0.001 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 140 126 Table 28: Spatial Regression Results for Den sity of Pounds Released (Natural Log) at One Mile of Census Tract Boundary Using the Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.002 0.001 ( 0.426) (0.067) % Asian 0.012 0.012 (0.488) (0.490) % Hispanic 0.026* 0.029* (4.150) (3.390) % Minority 0.009* 0.011 (2.000) (1.510) Population Density (LN) 0.608* 0.719* 0.694* 0.793* ( 4.600) ( 4.230) ( 5.020) ( 4.590) % Below Poverty 0.001 0.004 ( 0.041) ( 0.210) % Owner Occupied 0.0 01 0.001 ( 0.146) (0.118) Labor Force(SR) 0.011 0.010 (0.287) (0.964) Spatial Autoregressive Coefficient 0.824* 0.825* 0.810* 0.810* Akaike Info Criterion 5084.700 5089.540 5073.570 5078.450 Log likelihood rook 2538.350 2537.770 2530 .790 2530.220 Moran's I rook 0.002 0.002 0.001 0.002 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 141 127 Table 29: Spatial Regression Results for Density of Pounds Released (Natural Log) at One Mile of Census Tract Boundary Using Distance Independent Variables Model 1 0.37 mile Model 2 0.37 mile Model 3 0.37 mile Model 4 0.37 mile % Black 0.014** (1.680) 0.014 (1.160) % Asian 0.048 0.040 ( 1.290) ( 1.060) % Hispanic 0.091* 0.092* (9.610) (7.060) % Minority 0. 050* 0.039* (6.790) (3.460) Population Density (LN) 1.400* 1.430* 1.450* 1.500* ( 6.860) ( 5.450) ( 6.910) ( 5.850) % Below Poverty 0.055* 0.015 (1.810) (0.476) % Owner Occupied 0.012 0.010 (1.340) (1.120) Labor Force(SR) 0.00 9 0.008 (0.586) (0.510) Spatial Autoregressive Coefficient 0.490* 0.502* X X Akaike Info Coefficient 5726.120 5727.720 5676.450 5680.740 Log likelihood 2860.060 2857.860 2833.230 2832.370 Moran's I distance 0.019 0.002 0.006 0.00 6 a. All race variables represent a single race. b. *p<0.05, **p<0.10 c. X represents data provided in related OLS table PAGE 142 128 Table 30: Least Squares Regression Results for Density of Pounds Released (Natural Log) at Zero Mile from Block Group Bounda ry a. All race variables represent a single race. b. *p<0.05, **p<0.10 Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.005* 0.003 (2.110) (1.260) % Asian 0.011 0.011 (1.210) (1.220) % Hispanic 0.025* 0.023* (9.800) (7.360) % Minority 0.015* 0.013* (7.250) (4.670) Population Density (LN) 0.689* 0.681* 0.718* 0.712* ( 17.070) ( 15.99) ( 17.500) ( 16.620) % Below Poverty 0.001 0.002 ( 1.740) ( 0.216) % Owner Occupied 0.006* 0.005* ( 2.320) ( 2.020) Labor Force (SR) 0.009 0.008 ( 1.460) ( 1.330) R 2 0.097 0.100 0.112 0.133 F Statistic 145.757* 60.023 85.200* 49.623* Akaike Info Criterion 13695.000 13693.000 13654.600 13654.500 Log likelihood 6844.520 6840.490 6822.320 6819.240 Morans I queen 0.141* 0.137* 0.125* 0.122* Morans I rook 0.153* 0.148* 0.137* 0.132* Morans I 0.19 mile 0.001 0.001 0.001 0.001 PAGE 143 129 Table 31: Least Squares Regression Result s for Density of Pounds Released (Natural Log) at Half Mile from Block Group Boundary Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.009* 0.003 (2.380) (0.728) % Asian 0.007 0.008 ( 0.504) ( 0.547) % Hispanic 0.058* 0.0 53* (14.740) (10.570) % Minority 0.032* 0.023* (10.290) (5.350) Population Density (LN) 1.800* 1.140* 1.130* 1.260* ( 11.660) ( 10.130 ( 12.090) ( 11.270) % Below Poverty 0.037* 0.028* (3.240) (2.390) % Owner Occupied <0.001 0.002 (0.016) (0.390) Labor Force (SR) 0.018** 0.026* (1.640) (2.430) R 2 0.063 0.067 0.104 0.108 F Statistic 90.310* 38.870* 78.660* 46.590* Akaike Info Criterion 16217.700 16210.600 16098.500 16093.900 Log likelihood 8105.830 8099.290 8044.260 8038.950 Morans I queen 0.515* 0.722* 0.483* 0.488* Morans I rook 0.534* 0.537* 0.503* 0.508* Morans I 0.19 mile 0.011* 0.011* 0.009* 0.009* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 144 130 Table 32: Least Squares Re gression Results for Density of Pounds Released (Natural Log) at One Mile from Block Group Boundary Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.011* 0.005 (2.540) (3.660) % Asian 0.034** 0.001** ( 1.950) ( 0.135) % Hisp anic 0.075* 0.016* (16.270) (11.830) % Minority 0.042* 0.003* (11.33) (5.830) Population Density (LN) 1.190* 1.140* 1.210* 1.330* ( 9.290) ( 6.550) ( 9.400) ( 7.780) % Below Poverty 0.049* 0.032* (3.560) (2.320) % Owner Occup ied 0.005 0.007 (1.060) (1.450) Labor Force (SR) 0.004 0.020 (0.290) (1.400) R 2 0.058 0.062 0.112 0.115 F Statistic 82.770* 35.770* 85.240* 49.820* Akaike Info Criterion 17213.400 17206.700 17056.600 17055.300 Log likelihood 8603.710 8597.350 8523.290 8519.660 Morans I queen 0.672* 0.672* 0.639* 0.642* Morans I rook 0.692* 0.692* 0.658* 0.663* Morans I 0.19 mile 0.018* 0.018* 0.017* 0.016* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 145 131 Table 33: Spatia l Regression Results for Density of Pounds Released (Natural Log) at Zero Miles from Block Group Boundary Using the Queen Contiguity a. All race variables represent a single race. b. *p<0.05, **p<0.10 Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.005** 0.003 (1.840) (0.944) % Asian 0.010 0.009 (1.020) (0.898) % Hispanic 0.023* 0.020* (7.480) (5.600) %Minority 0.014* 0.011* (5.650) (3.660) Population Density (LN) 0.732* 0.733* 0.736* 0.748* ( 15.620) ( 14.700) ( 15.960) ( 15.150) % Below Poverty 0.004 0.004 (0.58 6) (0.468) % Owner Occupied 0.004 0.004 ( 1.460) ( 1.430) Labor Force(SR) 0.0003 0.0003 (0.051) (0.056) Spatial Autoregressive Coefficient 0.305* 0.303* 0.281* 0.280* Akaike Info Criterion 13574.100 13576.800 13556.200 13559.300 Log likelihood queen 6784.030 6782.380 6773.110 6771.640 Moran's I queen 0.006 0.005 0.004 0.004 PAGE 146 132 Table 34: Spatial Regression Results for Density of Pounds Released (Natural Lo g) at Zero Miles from Block Group Boundary Using the Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.005** 0.003 (1.860) (0.958) % Asian 0.011 0.010 (1.070) (0.964) % Hispanic 0.023* 0.020* (7.510) (5.62 0) % Minority 0.014* 0.011* (5.690) (3.680) Population Density (LN) 0.716* 0.717* 0.733* 0.735* ( 15.350) ( 14.450) ( 15.730) ( 14.930) % Below Poverty 0.005 0.004 (0.679) (0.555) % Owner Occupied 0.003 0.003 ( 1.280) ( 1.24 0) Labor Force(SR) 0.001 0.001 (0.080) (0.086) Spatial Autoregressive Coefficient 0.297* 0.295* 0.275* 0.273* Akaike Info Criterion 13567.300 13570.400 13549.400 13552.900 Log likelihood rook 6780.650 6779.210 6769.690 6768.430 Moran's I rook 0.002 0.002 0.001 0.001 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 147 133 Table 35: Spatial Regression Results for Density of Pounds Released (Natural Log) at Half Mile from Block Group Boundary Using the Queen Contiguit y Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.003 0.0003 (1.130) (0.088) % Asian 0.003** 0.015 (1.130) (1.430) % Hispanic 0.019* 0.017* (6.690) (4.600) % Minority 0.010* 0.008* (4.680) (2.530) Population Den sity (LN) 0.512* 0.642* 0.566* 0.699* ( 7.680) ( 7.890) ( 8.250) (8.500) % Below Poverty 0.013 0.014** (1.570) (1.700) % Owner Occupied 0.003 0.002 ( 1.060) ( 0.697) Labor Force(SR) 0.024* 0.025* (3.150) (3.270) Spatial Aut oregressive Coefficient 0.783* 0.783* 0.771* 0.771* Akaike Info Criterion 14679.400 14671.700 14660.800 14652.600 Log likelihood queen 7335.720 7328.840 7324.420 7317.280 Moran's I queen 0.013 0.016 0.013 0.017 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 148 134 Table 36: Spatial Regression Results for Density of Pounds Released (Natural Log) at Half Mile from Block Group Boundary Using the Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.003 (1.230) 0.008 (1.410 % Asian 0.018** 0.017 (1.730) (1.240) % Hispanic 0.020* 0.023* (6.880) (4.240) % Minority 0.011* 0.016* (4.850) (3.450) Population Density (LN) 0.501* 1.200* 0.555* 1.210* ( 7.650) ( 10.130) ( 8.230) ( 10.190) % Below Poverty 0.027* 0.026* (3.080) (2.960) % Owner Occupied 0.009* 0.008* (2.710) (2.590) Labor Force (SR) 0.066* 0.067* (7.050) (7.100) Spatial Autoregressive Coefficient 0.770* 0.787* 0.759* 0.781* Akaike Info Crite rion 14641.300 14601.100 14621.800 14595.500 Log likelihood queen 7316.660 7294.530 7304.890 7291.250 Moran's I queen 0.012 0.006 0.008 0.006 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 149 135 Table 37: Spatial Regression R esults for Density of Pounds Released (Natural Log) at Half Mile from Bloc k Group Boundary Using Distance Independent Variables Model 1 0.19 mile Model 2 0.19 mile Model 3 0.19 mile Model 4 0.19 mile % Black 0.009* 0.004 (2.460) (0.764) % Asian 0.007 0.008 ( 0.487) ( 0.539) % Hispanic 0.058* 0.052* (14.700) (10.530) % Minority 0.032* 0.023* (10.290) (5.330 Population Density (LN) 1.080* 1.130 1.130 1.250* ( 11.700) ( 10.100) ( 12.120) ( 11.240) % Below Poverty 0.036 0.027* (3.180) (2.360) % Owner Occupied 0.001 0.001 ( 0.197) (0.244) Labor Force(SR) 0.016 0.024* (1.510) (2.310) Spatial Autoregressive Coefficient 0.290* 0.286* 0.261* 0.252* Akaike Info Criterion 16208.900 16202.000 16091.70 0 16087.600 Log likelihood 8101.430 8095.020 8040.850 8035.800 Moran's I distance 0.003 0.003 0.003 0.003 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 150 136 Table 38: Spatial Regression Results for Density of Pounds Released (N atural Log) at One Mile from Block Group Boundary Using the Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.001 0.0004 (0.599) (0.134) % Asian 0.010 0.011 (1.040) (1.060) % Hispanic 0.016* 0.015* (5.910 ) (4.510) % Minority 0.008* 0.007* (3.870) (2.390) Population Density (LN) 0.449* 0.605* 0.491* 0.670* ( 6.230) ( 6.150) ( 6.650) ( 6.730) % Below Poverty 0.011 0.012 (1.510) (1.520) % Owner Occupied 0.001 0.002 (0.458) (0.786 ) Labor Force(SR) 0.021* 0.024* (2.690) (2.940) Spatial Autoregressive Coefficient 0.876* 0.876* 0.868* 0.868* Akaike Info Criterion 14517.100 14514.400 14500.800 14496.500 Log likelihood queen 7254.540 7250.210 7244.420 7239.240 Moran' s I queen 0.008 0.004 0.007 0.003 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 151 137 Table 39: Spatial Regression Results for Density of Pounds Released (Natural Log) at One Mile from Block Group Boundary Using the Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.002 0.001 (0.865) (0.339) % Asian 0.009 0.010 (0.996) (1.000) % Hispanic 0.016* 0.016* (6.140) (4.690) % Minority 0.008* 0.007* (4.160) (2.550) Population Density (LN) 0.450* 0.610* 0.493* 0.673* ( 6.400) ( 6.350) ( 6.800) ( 6.930) % Below Poverty 0.012 0.012 (1.600) (1.550) % Owner Occupied 0.001 0.002 (0.425) (0.744) Labor Force(SR) 0.022* 0.024* (2.810) (3.080) Spatial Autoregressiv e Coefficient 0.866* 0.866* 0.858* 0.858* Akaike Info Criterion 14449.300 14446.200 14433.200 14428.000 Log likelihood rook 7220.830 7216.080 7210.610 7205.020 Moran's I rook 0.017 0.012 0.015 0.010 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 152 138 Table 40: Spatial Regression Results for Density of Pounds Released (Natural Log) at One Mile from Block Group Boundary Using Distance Independent Variables Model 1 0.19 mile Model 2 0.19 mile Model 3 0.19 mile Model 4 0.19 mi le % Black 0.012* 0.006 (2.750) (1.020) % Asian 0.035* 0.033** ( 2.060) ( 1.880) % Hispanic 0.075* 0.069* (16.230) (11.680) % Minority 0.042* 0.030* (11.380) (5.720) Population Density (LN) 1.180* 1.110* 1.200* 1.290* ( 9.280) ( 6.410) ( 9.360) ( 7.610 % Below Poverty 0.050* 0.033* (3.670) (2.430) % Owner Occupied 0.004 0.006 (0.830) (1.240) Labor Force(SR) 0.001 0.017 (0.070) (1.180) Spatial Autoregressive Coefficient 0.401* 0.408* 0.385* 0.38 4* Akaike Info Coefficient 17193.000 17185.200 17038.300 17037.300 Log likelihood 8593.480 8586.600 8514.170 8510.670 Moran's I distance 0.005 0.005 0.003 0.005 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 153 139 7. ANALYZ ING THE TOXICITY OF EMISSIONS FROM INDUSTRIAL POLLUTION SOURCES This chapter focuses on the use of hazard scores calculated from points sources listed in the USEPAs Toxic Release Inventory (TRI) as an indicator of exposure to chronic air pollution in the analysis of environmental justice in the HoustonGalveston Brazoria metropolitan statistical area (HGB MSA) The goal is to explore the association between toxicity weighted quantities of emissions (hazard score) from chronic industrial air pollution sources and selected demographic and socioeconomic variables from Census 2000, in order to determine the presence of inequity with respect to the racial/ethnicity and poverty status of the population. The analysis begins with an examination of bivariate par ametric correlations between the toxicity weighted quantities of emissions (hazard scores) at three search radii (0, 0.5, and 1 mile) and relevant explanatory variables at the census tract and block group level of aggregation. The hazard scores of atmospheric pollution (all emitted chemicals) from TRI facilities were calculated for each geographic analytical unit following the technique for modeling point density described previously in Chapter 4. For each census tract or block group, the areal density of hazard scores is measured by dividing the total toxicity weighted hazard scores from their host facilities by the area enclosed, at three different distances from the tract or block group boundary (0, 0.5, and 1 mile). Because of the skewed distribution of this variable, the natural logarithm (LN) of the density of hazard scores was used for all statistical analyses presented in this chapter. The use of PAGE 154 140 the natural log reduces outliers and yields a normal style distribution of the dependent variable that is more consistent with the requirements of regression modeling. Summary statistics for all variables analyzed are provided in Tables 41 and 42, respectively, for census tract and block group level data. The descriptive statistics show that census tract s contain a larger density of hazard scores from pollutants than block groups. In addition, the density of hazard scores increase as the search radius increases from 0 to 1 mile from the boundary due to the possible inclusion of more TRI facilities at lar ger distances from each unit. Pearsons correlation coefficients associated with the density of hazard scores within the three distances of each analytical unit are presented in Table 43 for census tract level and Table 44 for block group level analyses The correlation analysis at the census tract level show that at 0 miles from the boundary, percent Hispanic has the highest positive correlation with the density of hazard scores followed by percent below poverty and percent minority. Additionally, population density has the greatest negative correlation followed by percent Asian, labor force, and percent Black. Percent owner occupied housing does not indicate significant correlation with density of hazard scores at this level. At 0.5 mile from the cen sus tract boundary, percent Hispanic again has the highest positive correlation in relation to density of hazard scores followed by percent below poverty and percent minority. Additionally, percent Asian has the greatest negative correlation followed by population density, and labor force. Percent owner occupied housing and percent Black does not indicate a significant correlation with density of hazard scores at this level. PAGE 155 141 At 1 mile from the census tract boundary, percent Hispanic has the highest positi ve correlation in relation to density of hazard scores followed by percent minority and percent below poverty. Additionally, percent Asian has the greatest negative correlation followed by labor force and population density. Percent owner occupied housing and percent Black do not indicate a significant correlation with density of hazard scores at this level. At the block group level, the correlation analysis show that at 0 miles from the boundary percent Hispanic has the highest positive correlation to t he density of hazard scores followed by percent below poverty and percent minority. Additionally, population density has the greatest negative correlation followed by labor force, percent Asian, and percent Black. Percent owner occupied housing does not indicate a significant correlation with density of hazard scores at this level. At 0.5 mile from the block group boundary, percent Hispanic has the highest positive correlation in relation to density of hazard scores followed by percent below poverty and percent minority. Additionally, population density has the greatest negative correlation followed by labor force, percent Asian, and percent owner occupied housing. Percent Black does not indicate a significant correlation with density of hazard scores at this level. At 1 mile from the block group boundary, percent Hispanic has the highest positive correlation in relation to hazard scores followed by percent minority and percent below poverty. Additionally, percent Asian along with labor force has the gre atest negative correlation followed by population density and percent owner occupied housing. PAGE 156 142 Percent Black does not show a significant correlation with density of hazard scores at this level. In summary, the correlation analysis indicates that the perce nt Hispanic has the strongest positive correlation and most significant positive linear association with the toxicity weighted quantities of chemicals released by TRI facilities into the atmosphere, at both the census tract and block levels. The environme ntal inequity hypothesis is supported by three variables (overall percent Hispanic, percent minority, and percent below poverty) which consistently exhibit a significant and positive relationship with density of hazard scores at all search radii and at bot h the census tract and block group levels. The percentage of Asians, population density, and labor force suggest a negative linear relationship at all distances and levels of aggregation. Blacks, on the other hand, have a negative relationship with densi ty of hazard scores at 0 miles only for both census tracts and block groups. Additionally, the presence of minorities becomes significant as the search radii approach 1 mile (1 mile at block group; and 0.5 and 1 mile at census tract). This suggests once again that the Hispanic population has the greatest statistical effect on the minority proportion since the coefficient for overall minority percentage is positive while all of the remaining minority sub groups are negative. Additionally, the negative cor relation of owner occupied housing at the block group level only reinforces the need to address environmental justice variables at different resolutions. The negative correlation for the percent Asian on the other hand becomes stronger as the search radius increases from the boundary to 1 mile for both analytical units, while population density became less influential. This suggests that more Asians could reside at the edge or outside of analytical units characterized by higher hazard scores. PAGE 157 143 While cor relation analysis provides a preliminary view of the statistical association between each demographic and socioeconomic variable and the density of hazard scores, it does not clarify the underlying pattern in sufficient detail. Multivariate regression is therefore used to analyze the simultaneous statistical effects of the explanatory variables on the density of hazard scores derived on the basis of the toxicity of chemicals released from TRI facilities at the census tract and block group level, respective ly. The ordinary least squares (OLS) regression approach is first utilized to model the statistical effect of explanatory variables on the density of hazard scores within each analytical unit in the study area, based on the three search radii (0, 0.5 and 1 mile). The results from the OLS regression models for the four different combinations of explanatory variables are summarized in the first set of tabl es (Tables 45 47 and 5658). Following Chapter 6, this set of tables also provides results from an an alysis of spatial dependence using both the contiguity based and distancebased methodologies. The results in the OLS tables contain regression coefficients that are bolded, z scores which are in parentheses and p values which are identified utilizing the testing for the existence of spatial autocorrelation. The second phase of the analysis uses the spatial autoregressive function (SAR) by utiliz ing an autoregressive coefficient to account for spatial autocorrelation. The appropriate SAR function (spatial lag or spatial error) was determined using the Lagrange Multiplier test statistic results (Anselin 2006) obtained while running the corresponding OLS model. As mentioned previously, the SAR models use a spatial PAGE 158 144 weights matrix specified by three configurations of neighboring census units: queenbased contiguity, rookbased contiguity, and distance based bands. Separate queen and rook contiguity analysis was therefore performed in hopes of finding the best fitted model that minimizes the effects of spatial dependence. Additionally, separate analysis for distance bands were constructed selecting between 3 and 5 distances which were used as adjustments to eliminate spatial autocorrelation. After the completion of the analyses, the best model was selected based upon the achievement of nonsignificant (p<0.05) Morans I which signifies no spatial autocorrela tion and the lowest Akaike Information Criterion (AIC) which is used to determine the fit of the model. The census tract results are listed in Tables 4855 and the block groups in Tables 5966. The results in the tables contain regression coefficients wh ich are bolded, z scores which are in parentheses, and p value s The histograms in Figure 10 summarize the connectivity distribution for census tracts, based on the number of neighbors in each configuration. For queen contiguity, Fi gure 10(a) shows that the majority of the census tracts have 5 to 8 neighbors and peaks at 7 neighbors. For the rook contiguity in Figure 10(b), the number of neighbors peaks at 4 but the census tracts noticeably had a positive skew. The histogram for the distancebased configuration in Figure 10(c) provides the connectivity distribution at 0.50 and (d) 0.38 mile from the centroid of each unit and indicates that a majority of census tracts have no neighbors at this distance. The connectivity distributions for census block groups are summarized in Figure 11. The histogram for queen contiguity in Figure 11(a) shows that the majority of the block groups have 5 to 7 neighbors and peaks at 6 neighbors. For the rook contiguity in Figure 11(b), the number of n eighbors peaks at 5 PAGE 159 145 and noticeably has a positive skew. The histogram for the distance based configuration in Figure 11(c) provides the connectivity distribution at 0.19 mile from the centroid of each unit and indicates that the number of block groups wit h no neighbors at this distance is considerably higher than the number of census tracts with no neighbors in Figure 11(c). PAGE 160 146 Figure 10 : C ensus t ract connectivity histograms used for hazard scores a. Queen Contiguit y b. Rook Contiguity c. Distance Based Configuration (0.50 miles) d. Distance Based Configuration (0.38 miles) PAGE 161 147 Figure 11 : B lock group connectivity his tograms used for hazard scores a. Queen Contiguity b. Rook Contiguity c. Distance Based Configuration (0.19 miles) PAGE 162 148 7.1 Census Tract Level Analysis Ordinary Least Squares Regression This section compares the density of hazard scores as computed from toxicity weighted quantities of emissions at the census tract level, at zero (0), half (0.5 ) and one (1) mile search radii, based on the application of ordinary least squares (OLS) and spatial autoregressive (SAR) techniques. Table 45 summarizes the results of OLS regression analysis at the boundary of the census tracts. Models 1 and 2 both indicate that the overall percent minority shows a significant and positive statistical effect on the density of hazard scores. Additionally, when the minority subgroups are analyzed separately for models 3 and 4, the Hispanic percent indicates a significa nt and positive association with the density of hazard scores. This suggests that the minority population in tracts exposed to the most toxic emissions probably comprises a large proportion of Hispanics. Additionally, it stresses the importance of analyz ing each racial/ethnic sub group individually and collectively to gain clarity on their environmental justice implications. In all four models, the density of hazard scores significantly increases with a decrease in population density. At 0.5 mile from th e tract boundary (Table 46), the results for the racial/ethnic variables are similar to those observed at 0 mile from the tracts. For models 1 and 2, the density of hazard scores increases significantly as the overall percentage of minority residents incr eases. Furthermore, for models 3 and 4, the Hispanic percentage has a significant and positive statistical effect on the density of hazard scores. Percent Black however, has a positive and significant statistical effect on hazard scores at model 3, but not in model 4 when the socioeconomic terms are introduced. Furthermore, percent PAGE 163 149 below poverty increases significantly for models 2 and 4. For all models at the boundary, population density is a negative predictor of hazard score density. At 1 mile fro m the boundary of census tracts (Table 47) models 1 and 2 indicate a positive statistical effect of the overall minority percentage and percent Hispanic (models 3 and 4) on the density of hazard scores, which is consistent with the results at 0 and 0.5 miles from tract boundaries. Percentage Black is again positive and significant in model 3. Percent below poverty on the other hand is significant in model 2, but becomes nonsignificant when the individual minority subgroups are introduced. In all four m odels, the density of hazard scores significantly increases with a decrease in population density. Morans I for the regression models based on queen and rook contiguities shows significance at 0, 0.5, and 1 mile search radii indicating the presence of sig nificant spatial autocorrelation in the data and justifying the need to account for spatial dependence by using a SAR model. However, at 0.5 mile from the 0 mile boundary when using the distance band at all search radii the distance band was found to be nonsignificant. Consequently, spatial dependence was not detected and the spatial autoregressive function was therefore not implemented. The distance band was also found to be nonsignificant at 0.37 mile at 0.5 and 1 mile search radii for models 3 and 4 only. Consequently, spatial dependence was not detected and the spatial autoregressive function was implemented only for models 1 and 2. Following the assessment for spatial dependence, the SAR model was thus implemented for all OLS models in which signi ficant spatial autocorrelation was detected. Spatial Regression At the boundary of the census tract level for the queen contiguity (Table 48), models 1 and 2 indicate a significant and positive association between the PAGE 164 150 density of hazard scores at 0 miles an d the overall minority percent. Additionally, for models 3 and 4, the Asian and Hispanic populations positively affect the density of hazard scores. Labor force however, has a positive statistical effect on density of hazard scores only for model 2, but is not significant when the minority subgroups are separated. For all models at 0 mile from the tract boundary, density of hazard scores significantly increases when population density decreases. The queen contiguity SAR models differ from OLS in that A sian percentage and labor force becomes significant and has a positive statistical effect on the density of hazard scores. At the boundary of the census tract level for the rook contiguity (Table 49), all SAR models exhibit the same statistical associations as observed in the model for queen contiguity and had similar differences from the corresponding OLS models. Model s fo r the distance band at 0.5 mile were not run because the Morans I from the OLS regression were not significantly different from zer o when determining spatial autocorrelation. At 0.5 mile from the boundary of census tracts (Table 50) for the queen contiguity, percent minority is positively associated with the density of hazard scores (model 1) only in the absence of socioeconomic vari ables. Additionally, for models 3 and 4, percent Hispanic has a significantly positive statistical effect on the density of hazard scores, along with labor force and percent below poverty for models 2 and 4. For model 2, the density of hazard scores incr eases significantly only when percent owner occupied increases, but not when the minority sub groups are separated. For all models at the boundary of tracts, population density is a negative predictor of the density of hazard scores. The queen contiguity SAR models differ from OLS in that the Black PAGE 165 151 percent becomes non significant for model 3 and percent minority becomes nonsignificant for model 2 only. Additionally, percent owner occupied housing and labor force have a positive statistical effect and ar e significant and for models 2 only on the density of hazard scores unlike the OLS model. At 0.5 mile from the boundary of the census tracts for the rook contiguity (Table 51), explanatory variables in all SAR models exhibit the same statistical associations as observed in the models for queen contiguity except percent occupied housing and labor which were both nonsignificant in this contiguity. Additionally, rook contiguity achieved the same significance and non significance as the corresponding OLS model. A distance band of 0.37 mile was constructed for models 1 and 2 only beginning 0.5 mile from each census tract boundary to identify neighboring tracts for assessing spatial dependence (Table 52). This distance was chosen based on a selection criteria that comprised the minimization of the likelihood term (AIC) and significance of spatial autocorrelation using Morans I. Percent Hispanic for models 3 and 4 has a significantly positive statistical effect on the density of hazard scores, as well as per cent minority for models 1 and 2, and percent below poverty for models 2 and 4. The distance band did not show any difference for SAR as compared to the OLS models. At 1 mile from the boundary of census tracts (Table 53) for queen contiguity, percent mino rity positively statistical effect s the density of hazard scores in model 1. Additionally, for models 3 and 4, the Hispanic population has a positive statistical effect on the density of hazard scores. For all models at 1 mile from the boundary, populati on density is a negative predictor of the density of hazard scores. The queen contiguity SAR PAGE 166 152 models differ from OLS in that Black percentage is not significant. In model 2 only, percent minority and percent below poverty become nonsignificant. At 1 mile from the boundary of the census tracts (Table 54) for the rook contiguity model, percent Hispanic exhibits the same positive prediction as in the queen contiguity model, while population density has the same negative statistical effect on the density of h azard scores. Percent minority is significant and a positive predictor of density of hazard scores for models with and without the socioeconomic variables. The SAR model exhibits the same difference for the rook contiguity at 1 mile as for the queen cont iguity, except the minority percent for model 2 is significant. A distance band of 0.37 mile was constructed for models 1 and 2 only at 1 mile from each census tract boundary to identify neighboring tracts for assessing spatial dependence (Table 55). This distance was chosen based on selection criteria that comprised the minimization of the likelihood term (AIC) and significance of spatial autocorrelation using Morans I. Percent Hispanic and percent minority exhibit the same positive statistical effect s on density of hazard scores as in the rook contiguity model, while population density has the same negative statistical effect. Additionally, percent below poverty and percent owner occupied in model 2, and labor force and percent Black in model 3 show a positive and significant statistical effect on the density of hazard scores. The distancebased model did not show any difference for SAR as compared to the OLS model, except that percent owner occupied housing and labor force became significant while using the distancebased SAR model. PAGE 167 153 7.2 Block Group Level Analysis Ordinary Least Squares Regression This section focuses on multivariate regression analysis of the areal density of hazard scores at the block group level, at zero (0), half (0.5), and one (1) mile search radii based on the application of ordinary least squared regression and spatial autoregressive techniques. Table 56 summarizes the results of OLS regression analysis at the boundary of the block groups. Models 1 and 2 both indicate that the overall percent minority shows a significant and positive statistical effect on the density of hazard scores. Additionally, when the minority subgroups are analyzed separately in models 3 and 4, the Hispanic percent and Black percent indicates a signific ant and positive association with the density of hazard scores. Additionally, percent owner occupied is a negative predictor in models 3 and 4, while the density of hazard scores significantly increases with a decrease in population density in all four mo dels. At 0.5 mile from the boundary of the block groups (Table 57), the results for the racial/ethnic variables are similar to those observed at 0 mile from the block groups. For models 1 and 2, the effect on density of hazard scores increase significantl y as the overall percentage of minority residents increases. Furthermore, for models 3 and 4, the Hispanic percentage has a significant and positive statistical effect on the density of hazard scores while in model 3 percent Black retains a positive statistical effect on density hazard scores. For both models 2 and 4, the density of hazard scores increases significantly as the percent below poverty increases. For all models, at 0.5 mile from the block group boundary, the density of hazard scores signific antly increases with a decrease in population density. PAGE 168 154 At 1 mile from the boundary of block groups (Table 58), models 1 and 2 indicate a positive statistical effect of the overall minority percentage, along with percent Black and percent Hispanic (model s 3 and 4) on the density of hazard scores that is consistent with the results at 0 miles from block group boundaries. As with 0.5 mile density of hazard scores is positively associated by percent below poverty. Additionally, the hazard scores increases significantly as percent Asian decreases for model 3 alone and for all models with a decrease in population density. Morans I for the regression models based on the queen and rook contiguities show significance at 0.5 and 1 miles search radii indicating the presence of significant spatial autocorrelation in the data, therefore justifying the need to account for spatial dependence by using the SAR model. However, at 0.19 mile from the boundary when using the 0 mile search radii only, the distance band wa s found to be nonsignificant. Consequently, spatial dependence was not detected and the spatial autoregressive function was therefore not implemented. This distance band configuration was chosen based on the best fit using AIC and the achievement of nonsignificance for spatial autocorrelation using Morans I. Following the assessment for spatial dependence the SAR model was thus implemented for all OLS models in which significant spatial autocorrelation was detected. Spatial Regression At the boundary of block groups (Table 59) for the queen contiguity in models 1 and 2, the overall minority percent indicates a significant and positive association with the density of hazard scores. Additionally, in models 3 and 4, the Hispanic percent has a positive sta tistical effect on the density of hazard scores. Percent Black however, positively influences density of hazard scores in model 3, but not when PAGE 169 155 the socioeconomic variables are introduced. For all SAR models, the density of hazard scores significantly inc reases with a decrease in population density. Hazard scores also increase when the presence of percent owner occupied housing decreases in models 2 and 4. The queen contiguity SAR models differs from OLS in that the Black percent becomes nonsignificant in model 4, while it remains significant in model 3 when socioeconomic terms are absent. At the boundary of the block group level (Table 60) for the rook contiguity, all SAR models exhibit the same statistical associations as observed in the model for quee n contiguity and similar differences from the corresponding OLS models. The rook contiguity differed from the OLS model the same as the queen contiguity where owner occupied housing is nonsignificant for models 2 and 4 when SAR is introduced and labor fo rce became significant in model 4. Models for the distance band at 0 miles was not run because the distance band at 0.19 mile for all models in the OLS regression was found to be nonsignificant when testing for spatial autocorrelation using Morans I. At 0.5 mile from the block group boundary (Table 61) for the queen contiguity, percent minority shows a positive statistical effect on the density of hazard scores in models 1 and 2. Additionally, in models 3 and 4, the Hispanic population has significant ly a positive statistical effect on density of hazard scores. In models 2 and 4, percent below poverty, percent owner occupied, and labor force independently and significantly has a positive statistical effect on density of hazard scores. Percent Black however, positively predicts density of hazard scores at model 3, but is not a predictor in model 4 when the socioeconomic terms are introduced. For all models, the density of PAGE 170 156 hazard scores significantly increases with a decrease in population density. T he queen contiguity SAR models differ from OLS in that the percent owner occupied housing and labor force have a positive statistical effect on density of hazard scores. At 0.5 mile from the boundary at the block group level (Table 62) for rook contiguit y, percent minority and percent Hispanic exhibit the same positive statistical effect as the queen contiguity, while population density has the same negative prediction on density of hazard scores. Additionally, percent Asian has a positive statistical ef fect on density of hazard scores for models 3 and 4. However, percent below poverty (model 2) and labor force (model 4) both have positive statistical effects on density of hazard scores. The rook contiguity SAR models differ from OLS in that the percent Black in model 3 is no longer significant. However, percent Asian gains significance in models 3 and 4 and labor force gains significance in the model with separate minority subgroups. The Black percentage on the other hand loses significance for model 3 while below poverty percentage loses significance when minority sub groups are included. A distance band of 0.19 mile (Table 63) was cons tructed beginning from 0.5 mile from each block group boundary to identify neighboring tracts for assessing spatial dependence. Percent minority and percent Hispanic exhibit the same positive statistical effect of density of hazard scores as the rook contiguity, while population density has the same negative prediction. Additionally, percent below poverty is a positi ve predictor for both models. Percent Black however, positively predicts hazard scores at model 3, but is not a predictor in model 4 when the economic terms were introduced. Labor force is positively associated with the density of hazard scores in model 4, but is not a predictor in model 3 when the minorities are classified into subgroups. Lastly, percent owner PAGE 171 157 occupied is a significant and negative predictor of density of hazard scores. The SAR models for the distance band model at 0.19 mile from the boundary differ from OLS in that percent owner occupied and labor force in model 3 are both significant and exhibits a positive statistical effect in model 4 on density of hazard scores. At 1 mile from the boundary of block groups (Table 64) for queen c ontiguity, percent minority and percent Hispanic has similar positive statistical effect in t he SAR models as 0 and 0.5 mile while population has a similar negative statistical effect on density of hazard scores. Hazard scores also increase as labor force increases for models 2 and 4. The queen contiguity SAR models differ from OLS in that in percent Black in models 3 and 4 are no longer significant. Additionally, percent Asian (model 3) and percent below poverty (models 2 and 4) lose significance and no longer contribute to density of hazard scores. Labor force on the other hand gains significance for models 2 and 4. At 1 mile from the boundary of block groups (Table 65) for rook contiguity, percent minority, percent Hispanic, and labor force exhibit t he same positive statistical effect on the density of hazard scores as the queen contiguity, while population density has the same negative statistical effect on the density of hazard scores. The rook contiguity SAR models exhibit the same difference from OLS as the queen contiguity models. A distance band of 0.19 mile (Table 66) was constructed beginning from 1 mile from the block group boundary. Percent minority and percent Hispanic show the same positive prediction as the rook contiguity, while populat ion density had the same negative statistical effect on of density of hazard scores. Additionally, percent below poverty has PAGE 172 158 a positive association with the density of hazard scores at model 2 and 4 and percent Black at models 3and 4. Pe rcent Asian howe ver, has a negative association with the density of hazard scores at model 3. The distance based SAR models do not exhibit any significant differences from the OLS models. 7.3 Discussion and Interpretation The analyses presented in this chapter focuses on the geographic distribution of pollution burdens measured in terms of the toxicity weighted quantities of emissions (hazard scores) from industrial point sources. The results reveal that the areal density of hazard scores from TRI facilities in the Hous tonMSA are significantly associated with population density, the overall proportion of minorities, and the Hispanic proportion at the census tract level, regardless of whether conventional or spatial regression was utilized. While densely populated tract s are less likely to be exposed to the most toxic emissions, the presence of racial/ethnic minorities and Hispanics, in particular, indicate a positive statistical effect on the density of hazard scores at all distances, even after controlling for the effe cts of other explanatory variables. Additionally, the percentage of Asians living within the census tract boundaries are only significant when spatial autocorrelation is removed. The Black percentage shows a significant and positive statistical effect wh en the density of hazard scores is measured at 0.5 mile from the tract boundary, for both ordinary regression and spatial regression based on distance based contiguity ( 0.37 mile ). The poverty rate also shows a positive relationship with hazard scores at 0.5 mile Additionally, the percentage of owner occupied housing at distances (0.5 and 1 mile) away from the census tract boundaries are only positive and significant when spatial autocorrelation is removed. Labor force shows a positive statistical effec t PAGE 173 159 on the density of hazard scores near the boundary only when spatial autocorrelation is taken into account. This suggests that the presence of a readily available workforce plays a key role in the location of these industrial facilities. As observed at the census tract level, population density, minority percent, and percent of Hispanics, are again the most influential variables related to the density of hazard scores at the block group level. Hazard scores are found to be significantly greater in block groups containing a higher proportion of racial/ethnic minorities and Hispanic populations, in all models and distance definitions. Furthermore, unlike the census tract level, the Black population is significantly associated with hazard score density at all distances with and without the adjustment for spatial autocorrelation. This suggests that a higher concentration of both Black and Hispanic residents can be found near TRI facilities releasing chemicals of greater toxicity, at the block group level. It also again suggests that the minorities facing the greatest exposure to the most toxic TRI chemicals are overwhelmingly Hispanics. Again the Asian population is significant only when spatial dependence is removed. As for poverty rate, a positive and significant statistical effect is found only when hazard scores are measured at distances that extend beyond block group boundaries (0.5 mile ; and 1 mile 0.19 mile distance based), with and without adjusting for spatial autocorrelation. Additionally, the percentage of owner occupied housing at distances (0 and 0.5 mile) away from the block group boundaries are negative and significant using both ordinary and spatial regression techniques, while labor force is positive and significant at almost all distances In summary, the results suggest that the toxicity weighted quantities of TRI emissions increase significantly with an increase in the proportion of racial/ethnic PAGE 174 160 minorities at all search radii and at both the census tract and block levels of aggregation. More specifically, the overall minority, Hispanic population, and population density are the most significant explanatory factors in determining the density of hazard scores at all search radii, all data resolutions, and for both regression modeling te chniques. Additionally, the proportion of Asians is only significant when spatial autocorrelation is removed and the significance of labor force increases when spatial dependence is accounted for as well. This finding indicates how spatial autocorrelation can mask the effects of specific variables (e.g., percent Asian and labor force) when determining their statistical association with the toxicity weighted magnitude of emissions and justifies the need to control for spatial dependence in regression analys is of environmental justice. The Black population was found to be positive and significant at the boundary of the tract level for the OLS models only. Additionally, this variable was found to be significant at the 0.5 and 1 mile search radii and the 0.5 search radii using the 0 distance based SAR models that did not include socioeconomic variables. However, at the block group level, Black population was significant at all distances and in majority of the models tested with and without the socioeconomic variables. It should be noted that owner occupied housing had a positive statistical effect on the hazard scores at the census tract level but exhibited a negative statistical effect at the block group level. This confirms the importance of analyzing dat a at more than one spatial resolution because the nuances between the distances were not evident at a single analytical unit. While the analyses presented in this chapter provide an assessment of inequities based on toxicity of chemicals emitted from TRI f acilities, it is important to consider that the toxicity weighted quantities (hazard scores) do not account for the environmental fate PAGE 175 161 and transport of released chemicals. Thus, the human health risks of exposure to toxic emissions within and outside the a nalytical unit containing the TRI facilities are not considered. The application of modeled risk scores, therefore, is necessary to address this limitation and provide a more exact assessment of disproportionate risk burdens imposed on the population in the Houston MSA. PAGE 176 162 Table 41: Descriptive Statistics for Variables at the Census Tract Level Associated with Hazard Scores n=886 Hazard Scores Mean Std Dev Min Max Variables Density of Hazard Scores (LN): 0 Miles 2.580 6.340 0.000 26.480 Density of Hazard Scores(LN): 0.5 Mile 5.380 8.250 0.000 26.580 Density of Hazard Scores(LN): 1 Mile 8.060 8.870 0.000 26.010 % Black 18.140 24.290 0.000 98.160 % Asian 4.140 5.890 0.000 37.710 % Hispanic 28.380 23.310 0.000 100.000 % Minority 52.700 29.830 0.000 100.000 Population Density (LN) 7.550 8.020 0.000 10.390 % Below Poverty 14.210 10.920 0.000 66.690 % Owner Occupied 61.010 25.440 0.000 100.000 Labor Force (SR) 60.270 16.530 0.000 115.450 PAGE 177 163 Table 42: Descriptive Statistics for Variables a t the Block Group Level Associated with Hazard Scores n=2705 Hazard Scores Mean Std Dev Min Max Variables Density of Hazard Scores (LN): 0 Miles 1.080 4.370 0.000 27.340 Density of Hazard Scores(LN): 0.5 Mile 3.340 7.000 0.000 27.920 De nsity of Hazard Scores(LN): 1 Mile 5.960 8.450 0.000 26.970 % Black 17.560 25.760 0.000 100.000 % Asian 3.760 6.660 0.000 68.160 % Hispanic 28.860 25.590 0.000 100.000 % Minority 52.210 31.630 0.000 100.000 Population Density (LN) 7.710 1.580 0.000 11.150 % Below Poverty 14.490 12.190 0.000 73.360 % Owner Occupied 63.110 27.430 0.000 100.00 Labor Force (SR) 34.140 10.650 0.000 83.610 PAGE 178 164 Table 43: Pearson Correlation Coefficients Associated with the Density of Hazard Scores (Natural Log) at the Census Tract Level. n =886 p<0.05 0 Mile 0.5 Mile 1 Mile Variables r p Value r p Value r p Value % Black 0.071 0.035 0.013 0.693 0.006 0.862 % Asian 0.110 0.001 0.185 <0.001 0.171 <0.001 % Hispanic 0.174 <0.001 0.255 <0.001 0.283 <0.001 % Minority 0.068 0.044 0.162 <0.001 0.204 <0.001 Population Density (LN) 0.297 <0.001 0.180 <0.001 0.140 <0.001 % Below Poverty 0.080 0.017 0.195 <0.001 0.203 <0.001 % Owner Occupied 0.013 0.699 0.034 0.307 0.040 0.233 Labor Force (SR) 0.095 0.005 0.124 0.001 0.145 <0.001 PAGE 179 165 Table 44: Pearson Correlation Coefficients Associated with the Density of Hazard Scores (Natural Log) at the Block Group Level n=2705 p<0.05 0 Mile 0.5 Mile 1 Mile Variables r p Value r p Value r p Value % Black 0.044 0.022 0.026 0.183 0.006 0.746 % Asian 0.073 0.001 0.107 <0.001 0.130 <0.001 % Hispanic 0.086 <0.001 0.240 <0.001 0.292 <0.001 % Minority 0.039 0.043 0.161 <0.001 0.212 <0.001 Population Density (LN) 0.283 <0.001 0.156 0.004 0.102 <0.001 % Below Poverty 0.042 0.030 0.166 <0.001 0.204 <0.001 % Owner Occupied 0.019 0.331 0.074 0.001 0.076 <0.001 Labor Force (SR) 0.124 <0.001 0.116 <0.001 0.130 <0.001 PAGE 180 166 Table 45: Least Squares Regression Results for Density of Hazard Scores (Natural Log) at Zero Miles of Cen sus Tract Boundary Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.013 0.010 (1.580) (0.807) % Asian 0.050 0.052 (1.400) (1.420) % Hispanic 0.088* 0.083* (9.320) (6.640) %Minority 0.046* 0.040* (6.310) (3 .650) Population Density(LN) 1.560* 1.620* 1.680* 1.720* ( 11.240) ( 10.680) ( 11.860) ( 11.330) % Below Poverty 0.027 0.022 (0.919) (0.686) % Owner Occupied 0.001 0.002 (0.129) (0.267) Labor Force (SR) 0.0180 0.012 (1.3 50) (0.916) R 2 0.128 0.130 0.172 0.173 F Statistic 64.630* 26.330* 45.810* 26.310* Akaike Info Criterion 5671.740 5675.280 5629.360 5634.070 Log likelihood 2832.870 2831.640 2809.680 2809.040 Morans I queen 0.176* 0.179* 0.141* 0.144* Morans I rook 0.168* 0.172* 0.130* 0.135* Morans I 0.5 mile 0.003 0.002 0.003 0.004 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 181 167 Table 46: Least Squares Regression Results for Density of Hazard Scores (Natural Log) at H alf Mile from the Census Tract Boundary a. All race variables represent a single race. b. *p<0.05, **p<0.10 Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.025* (2.220) 0.009 (0.542) % Asian 0.075 0.057 ( 1.580) ( 1.18) % Hispanic 0.121* 0.105* (9.800) (6.310) % Minority 0.069* 0. 039* (7.230) (2.650) Population Density(LN) 1.870* 1.860* 1.860* 1.920* ( 7.950) ( 6.610) ( 7.770) ( 6.900) % Below Poverty 0.128* 0.079** (3.240) (1.890) % Owner Occupied 0.015 0.015 (1.220) (1.280) Labor Force (SR) 0.015 0.012 (0.800) (0.645) R 2 0.086 0.098 0.146 0.150 F Statistic 41.740* 19.030* 37.570* 22.150* Akaike Info Criterion 6178.330 6173.400 6122.810 6124.240 Log likelihood 3086.170 3080.700 3056.410 3054.120 Morans I queen 0.433* 0.4 36* 0.387* 0.393* Morans I rook 0.461* 0.465* 0.419* 0.425* Morans I 0.37 mile 0.013* 0.014* 0.006 0.007 PAGE 182 168 Table 47: Least Squares Regression Results for Density of Hazard Scores (Natural Log) at OneMile from the Census Tract Boundary Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.032* (2.670) 0.026 (1.550) % Asian 0.063 0.047 ( 1.240) ( 0.897) % Hispanic 0.137* 0.1 34* (10.450) (7.440) % Minority 0.080* 0.059* (7.870) (3.740) Population Density(LN) 2.030* 1.920* 2.050* 2.010* ( 7.110) ( 5.240) ( 7.04) ( 5.580) % Below Poverty 0.096* 0.040 (2.520) (0.883) % Owner Occupied 0.021 0.021 (1.620) (1.690) Labor Force(SR) 0.001 0.002 (0.018) ( 0.098) R 2 0.086 0.092 0.146 0.149 F Statistic 41.480* 17.870* 37.690* 21.960* Akaike Info Criterion 6307.650 6307.510 6251.260 6254.260 Log likelihood 3150.820 3147.760 312 0.630 3119.130 Morans I queen 0.530* 0.528* 0.496* 0.497* Morans I rook 0.561* 0.559* 0.531* 0.521* Morans I 0.37 mile 0.012* 0.013* 0.007 0.008 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 183 169 Table 48: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Zero Mile from Census Tract Boundary Using Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.016 0.010 ( 1.570) (0.740) % Asian 0.085** 0.078** (1.790) (1.840) % Hispanic 0.085* 0.078* (7.510) (5.360) % Minority 0.047* 0.041* (5.080) (3.190) Population Density(LN) 1.560* 1.710* 1.670* 1.780* ( 9.010) ( 9.620) ( 10.580) ( 10.290) % Below Poverty 0.033 0.037 (1.020) (1.130) % Owner Oc cupied 0.004 0.022 (0.368) (1.630) Labor Force(SR) 0.027* 0.022 (2.030) (1.630) Spatial Autoregressive Coefficient 0.372* 0.381* 0.322* 0.333* Akaike Info Criterion 5610.470 5611.420 5587.530 5589.570 Log likelihood queen 2802.230 279 9.710 2788.760 2786.790 Moran's I queen 0.008 0.009 0.005 0.006 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 184 170 Table 49: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Zero Mil e from Census Tract Boundary Using Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.012 (1.490) 0.007 (0.637) % Asian 0.066** 0.068** (1.910) (1.900) % Hispanic 0.073* 0.068* (7.730) (5.440) % Minority 0.047* 0.041* (5.210) (3.240) Population Density (LN) 1.500* 1.650* 1.450* 1.490* ( 9.590) ( 9.480) ( 10.080) (9.650) % Below Poverty 0.033 0.023 (1.030) (0.730) % Owner Occupied 0.003 0.001 (0.302) ( 0.063) Labor Force(SR) 0.029* 0.010 (2.160) (0.795) Spatial Autoregressive Coefficient 0.309* 0.319* 0.252* 0.252* Akaike Info Criterion 5623.860 5624.380 5598.410 5603.280 Log likelihood rook 2808.930 2806.190 2793.200 2792.640 Moran's I rook 0.013 0.013 0 .006 0.002 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 185 171 Table 50: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Half Mile from Census Tract Boundary Using Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.010 (1.110) 0.005 ( 0.396) % Asian 0.006 0.016 (0.166) (4.240) % Hispanic 0.051* 0.360* (5.090) (2.710) % Minority 0.027* 0.023 (3.580) (1.380) Population Density(LN) 1.010* 1.670* 1.09 0* 1.160* ( 5.440) ( 5.390) ( 5.640) ( 5.130) % Below Poverty 0.111* 0.062** (3.020) (1.860) % Owner Occupied 0.019** 0.003 (1.720) (0.337) Labor Force(SR) 0.040* 0.011* (2.320) (0.747) Spatial Autoregressive Coefficient 0.717* 0.735* 0.692* 0.691* Akaike Info Criterion 5812.130 5805.870 5802.380 5804.430 Log likelihood queen 2902.060 2896.930 2895.190 2893.220 Moran's I queen 0.005 0.000 0.004 0.009 a. All race variables represent a single race. b. *p<0.05, **p< 0.10 PAGE 186 172 Table 51: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Half Mile from Census Tract Boundary Using Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.010 (1.210) 0.004 ( 0.313) % Asi an 0.006 0.016 (0.161) (0.437) % Hispanic 0.051* 0.037* (5.240) (2.850) % Minority 0.027* 0.011 (3.750) (1.040) Population Density (LN) 0.951* 0.997* 1.030* 1.100* ( 5.280) ( 4.630) ( 5.480) ( 5.000) % Below Poverty 0.064* 0. 061** (2.140) (1.890) % Owner Occupied 0.002 0.004 (0.202) (0.470) Labor Force(SR) 0.013 0.011 (0.928) (0.773) Spatial Autoregressive Coefficient 0.715* 0.711* 0.691* 0.691* Akaike Info Criterion 5783.210 5783.920 5772.850 5774.810 Log likelihood rook 2887.600 2884.960 2880.420 2878.400 Moran's I rook 0.006 0.013 0.006 0.011 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 187 173 Table 52: Spatial Regression Results for Density of Hazard Scores (Natural Log) a t Half Mile from Census Tract Boundary Using Distance Band Independent Variables Model 1 0.37 mile Model 2 0.37 mile Model 3 0.37 mile Model 4 0.37 mile % Black 0.025* (2.220) 0.009 (0.542) % Asian 0.075 0.057 % Hispanic ( 1.580) ( 1.180) 0.121* 0.105* % Minority 0.038* (9.800) (6.310) 0.068* (7.150) (2.600) Population Density (LN) 1.840* 1.820* 1.860* 1.920* ( 7.850) ( 6.550) ( 7.770) ( 6.900) % Below Poverty 0.131* 0.079** (3.330) (1.890) % Owner Occupied 0.0170 0.015 (1.410) (1.280) Labor Force(SR) 0.016 0.012 (0.864) (0.645) Spatial Autoregressive Coefficient 0.417 0.438 X X Akaike Info Criterion 6173.950 6168.310 6122.810 6124.240 Log likelihood 3082.990 3077.150 3056.410 3054 .120 Moran's I distance 0.003 0.003 0.006 0.007 a. All race variables represent a single race. b. *p<0.05, **p<0.10 c. X represents data provided in related OLS table PAGE 188 174 Table 53: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Census Tract Boundary Related Using Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.001 0.001 ( 0.135) ( 0.022) % Asian 0.014 0.018 (0.394) (0.500) % Hispanic 0.040* 0.041* (4.380) (3.33 0) % Minority 0.016* 0.015 (2.310) (1.400) Population Density (LN) 0.900* 1.010* 1.020* 1.120* ( 4.670) ( 4.080) ( 5.050) ( 4.460) % Below Poverty 0.012 0.007 (0.432) (0.242) % Owner Occupied 0.002 0.005 (0.268) (0.546) Labor Force(SR) 0.020 0.011 (0.831) (0.715) Spatial Autoregressive Coefficient 0.823* 0.823* 0.807* 0.807* Akaike Info Criterion 5718.480 5723.560 5707.190 5712.250 Log likelihood queen 2855.240 2854.780 2847.600 2847.130 Moran's I queen 0.0 14 0.010 0.012 0.008 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 189 175 Table 54: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Census Tract Boundary Using Rook Contiguity Independent Variab les Model 1 Model 2 Model 3 Model 4 % Black 0.001 0.002 ( 0.109) (0.215) % Asian 0.009 0.011 (0.270) (0.333) % Hispanic 0.041* 0.044* (4.570) (3.670) % Minority 0.016* 0.017** (2.450) (1.670) Population Density (LN) 0.868* 0 .987* 0.977* 1.090* ( 4.670) ( 4.130) ( 5.030) ( 4.480) % Below Poverty 0.006 0.002 (0.208) ( 0.065) % Owner Occupied 0.002 0.004 (0.224) (0.470) Labor Force(SR) 0.012 0.011 (0.866) (0.752) Spatial Autoregressive Coefficient 0 .816* 0.816* 0.800* 0.800* Akaike Info Criterion 5681.920 5687.050 5669.280 5674.280 Log likelihood rook 2836.960 2836.520 2828.640 2828.140 Moran's I rook 0.013 0.010 0.011 0.007 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 190 176 Table 55: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Census Tract Boundary Using Distance Band Independent Variables (m) Model 1 0.37 mile Model 2 0.37 mile Model 3 0.37 mile Model 4 0.37 mile % B lack 0.032* 0.026 (2.670) (1.55) % Asian 0.063 0.047 ( 1.240) ( 0.897) % Hispanic 0.137* 0.134* (10.450) (7.440) % Minority 0.080* 0.060* (7.910) (3.870) Population Density (LN) 2.000* 1.880* 2.050* 2.010* ( 7.090) ( 5 .200) ( 7.040) ( 5.580) % Below Poverty 0.094* 0.040 (2.250) (0.883) % Owner Occupied 0.024** 0.021 (0.012) (1.690) Labor Force(SR) 0.001* 0.002 ( 0.002) ( 0.098) Spatial Autoregressive Coefficient 0.501* 0.517* X X Akaike Info Coefficient 6301.750 6300.590 6251.260 6254.260 Log likelihood 3147.870 3144.440 3120.630 3119.130 Moran's I distance 0.003 0.003 0.007 0.008 a. All race variables represent a single race. b. *p<0.05, **p<0.10 c. X represents data provided in related OLS table PAGE 191 177 Table 56: Least Squares Regression Results for Density of Hazard Scores (Natural Log) at Zero Mile from the Block Group Boundary a. All race variables represent a single race. b. *p<0.05, **p<0.10 Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.009* 0.008** (2.640) (1.850) % Asian 0.016 0.017 (1.300) (1.300) % Hispanic 0.026* 0.035* (10.480) (8.050) % Minority 0.022* 0.020* (7.970) (5.350) Population Density(LN) 0.958* 0.948* 0.999* 0.811* ( 17.410) ( 16.320) ( 17.850) ( 16.970) % Below Poverty 0.005 0.006 ( 0.478) ( 0.584) % Owner Occupied 0.009* 0.008* ( 2.580) ( 2.290) Labor Force (SR) 0.013 0.012 ( 1.630) ( 1.480) R2 0.101 0.104 0.117 0.119 F Statistic 152.155* 62.890* 89.090* 52.045* Akaike Info Criterion 15374.000 15370.500 15331.300 15329.900 Log likelihood 7683.990 7679.270 7660.650 7656.960 Morans I queen 0.139* 0.134* 0.123* 0.118* Morans I rook 0.151* 0.145* 0.134* 0.129* Morans I 0.19 mile 0.001 0.001 0.001 0.001 PAGE 192 178 Table 57: Least Squares Regression Results for Density of Hazard Scores (Natural Log) at Half Mile from the Block Group Boundary Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.017* 0.010 (3.280) (1.460) % Asian 0.002 0.004 ( 0.102) ( 0.173) % Hispanic 0.087* 0.080* (15.960) (11.510) % Minority 0.050* 0.038* (11.540) (6.230) Population Density (L N) 1.540* 1.590* 1.620* 1.770* ( 11.950) ( 10.190) ( 12.460) ( 11.400) % Below Poverty 0.047* 0.034* (2.950) (2.090) % Owner Occupied 0.003 0.001 ( 0.508) ( 0.088) Labor Force (SR) 0.018 0.030 (1.230) (2.020) R 2 0.071 0. 075 0.114 0.117 F Statistic 102.640* 43.520* 86.990* 50.990* Akaike Info Criterion 18009.600 18003.090 17883.700 17881.400 Log likelihood 9001.780 8995.940 8936.850 8932.710 Morans I queen 0.505* 0.506* 0.469* 0.474* Morans I rook 0.524* 0.526* 0.488* 0.494* Morans I 0.19 mile 0.010* 0.011* 0.008* 0.008* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 193 179 Table 58: Least Squares Regression Results for Density of Hazard Scores (Natural Log) at One Mile from the Block Group Boundary Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.024* 0.016* (3.920) (1.990) % Asian 0.041** 0.036 ( 1.690) ( 1.470) % Hispanic 0.114* 0.011* (17.690) (12.800) % Minority 0.067* 0.050* (13.000) (6.850) Population Density (LN) 1.650* 1.480* 1.680* 1.750* ( 9.220) ( 6.070) ( 9.300) ( 7.320) % Below Poverty 0.069* 0.045* (3.620) (2.320) % Owner Occupied 0.007 0.009 (1.080) (2.320) Labor Force (SR) 0.007 0.015 ( 0.339) (0.77 6) R 2 0.069 0.073 0.124 0.126 F Statistic 99.550* 42.730* 95.460* 55.500* Akaike Info Criterion 19032.20 19024.40 18870.70 18870.50 Log likelihood 9513.10 8597.35 9430.35 9427.26 Morans I queen 0.663* 0.661* 0.628* 0.630* Morans I rook 0.682* 0.680* 0.648* 0.650* Morans I 0.19 mile 0.020* 0.021* 0.019* 0.018* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 194 180 Table 59: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Zero Miles from Blo ck Group Boundary Using Queen Contiguity a. All race variables represent a single race. b. *p<0.05, **p<0.10 Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.009* 0.006 (2.260) (1.370) % Asian 0.015 0.014 (1.090) (0.966) % Hispanic 0.033* 0.030* (8.110) (6.190) %Minority 0.010* 0.020* (6.260) (4.190) Population Density (LN) 1.020* 1.020* 1.040* 1.040* ( 15.990) ( 15.020) ( 16.350) ( 15.500) % Below Poverty 0.004 0.002 (0.387) (0.223) % Owner Occupied 0.006** 0.006** ( 1.680) ( 1.650) Labor Force(SR) 0.0 01 0.001 ( 0.118) ( 0.101) Spatial Autoregressive Coefficient 0.301* 0.297* 0.275* 0.273* Akaike Info Criterion 15256.900 15259.200 15237.100 15239.800 Log likelihood queen 7625.470 7623.620 7613.570 7611.920 Moran's I queen 0.006 0.00 5 0.004 0.004 PAGE 195 181 Table 60: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Zero Miles from Block Group Boundary Using Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.009* 0.006 (2.270) (1.330) % Asian 0.016 0.014 (1.120) (1.010) % Hispanic 0.033* 0.031* (8.120) (6.300) % Minority 0.021* 0.018* (6.280) (4.200) Population Density (LN) 0.999* 0.997* 1.02 0* 0.961* ( 15.730) ( 14.770) ( 16.130) ( 13.730) % Below Poverty 0.005 0.008 (0.497) (0.854) % Owner Occupied 0.005 0.002 ( 1.460) (2.690) Labor Force(SR) 0.001 0.089* ( 0.105) ( 2.610) Spatial Autoregressive Coefficient 0.2 93* 0.289* 0.269* 0.261* Akaike Info Criterion 15250.100 15253.000 15230.400 15228.600 Log likelihood rook 7622.070 7620.510 7610.190 7606.310 Moran's I rook 0.003 0.003 0.002 0.001 a. All race variables represent a single race. b. *p<0.0 5, **p<0.10 PAGE 196 182 Table 61: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Half Mile from Block Group Boundary Using Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.013** 0.010 (1.740) (1.260) % Asian 0.028 0.018 (1.360) (0.906) % Hispanic 0.033* 0.030* (4.770) (3.890) % Minority 0.024* 0.021* (4.120) (3.070) Population Density (LN) 1.010* 1.810* 1.010* 1.820* ( 8.250) ( 10.720) ( 8.280) ( 10.750) % Below Poverty 0.03 6* 0.035* (2.860) (2.730) % Owner Occupied 0.009** 0.008** (1.870) (1.760) Labor Force(SR) 0.091* 0.092* (6.820) (6.850) Spatial Autoregressive Coefficient 0.786* 0.793* 0.780* 0.788* Akaike Info Criterion 16539.000 16495.500 165 37.600 16493.600 Log likelihood queen 8266.600 8241.710 8263.800 8238.800 Moran's I queen 0.002 0.001 0.003 0.002 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 197 183 Table 62: Spatial Regression Results for Density of Hazard S cores (Natural Log) at Half Mile from Block Group Boundary Using Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.006 (1.590) 0.002 (0.495) % Asian 0.028** 0.024** (1.940) (1.670) % Hispanic 0.030* 0.030* (7.350) (5.150) % Minority 0.016* 0.013* (5.300) (3.000) Population Density (LN) 0.737* 0.900* 0.819* 0.987* ( 7.970) ( 7.990) ( 8.600) ( 8.650) % Below Poverty 0.015* 0.017 (1.380) (1.490) % Owner Occupied 0.005 0.003 ( 1. 270) ( 0.893) Labor Force(SR) 0.030 0.031* (2.810) (2.940) Spatial Autoregressive Coefficient 0.763* 0.763* 0.750* 0.751* Akaike Info Criterion 16492.500 16486.500 16471.600 16465.200 Log likelihood queen 8242.230 8236.240 8229.670 8223 .590 Moran's I queen 0.012 0.015 0.012 0.016 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 198 184 Table 63: Spatial Regression Results for Density of Hazard Scores (Natural Log) at Half Mile from Block Group B oundary Using Distance Ban d Independent Variables Model 1 0.19 mile Model 2 0.19 mile Model 3 0.19 mile Model 4 0.19 mile % Black 0.017* 0.010 (3.370) (1.500) % Asian 0.002 0.004 ( 0.095) ( 0.174) % Hispanic 0.087* 0.079* (15.890) (11.450) % Minority 0.050* 0.037* (11.530) (6.210) Population Density (LN) 1.540* 1.580* 1.620* 1.760* ( 11.970) ( 10.150) ( 12.470) ( 11.360) % Below Poverty 0.046* 0.033* (2.900) (2.080) % Owner Occupied 0.004 0.001* ( 0.674) ( 0.218) Labor Force(SR ) 0.016 0.028** (1.100) (1.910) Spatial Autoregressive Coefficient 0.279* 0.279* 0.233* 0.229* Akaike Info Coefficient 18001.300 17995.600 17878.400 17876.300 Log likelihood 8997.670 8991.810 8934.190 8930.160 Moran's I distance 0.003 0 .003 0.008 0.002 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 199 185 Table 64: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Block Group Boundary Using Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.003 0.002 (0.944) (0.430) % Asian 0.015 0.015 (1.080) (1.090) % Hispanic 0.024* 0.023* (6.300) (4.810) % Minority 0.013* 0.011* (4.270) (2.650) Population Density (LN) 0.637* 0.807* 0.702 0.901* ( 6.240) ( 5.810) ( 6.660) ( 6.410) % Below Poverty 0.014 0.014 (1.300) (1.270) % Owner Occupied 0.001 0.002 (0.279) (0.608) Labor Force(SR) 0.024* 0.027* (2.100) (2.360) Spatial Autoregressive Coefficient 0.873* 0.873* 0.864* 0.863* Akaike Info Criterion 16400.000 16400.500 16382.600 16381.900 Log likelihood queen 8196.000 8193.240 8185.320 8181.950 Moran's I queen 0.009 0.006 0.008 0.005 a. All race variables represent a single race. b. *p<0.05, **p< 0.10 PAGE 200 186 Table 65: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Block Group Boundary Using Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.004 0.003 (1.250) (0.630) % Asian 0.014 0.014 (0.986) (0.985) % Hispanic 0.024* 0.024* (6.520) (4.940) % Minority 0.013* 0.011* (4.570) (2.780) Population Density(LN) 0.630* 0.806* 0.690* 0.898* ( 6.310) ( 5.950) ( 6.700) ( 6.530) % Below Poverty 0.015 0.015 (1.450) (1.340) % Owner Occupied 0.001 0.002 (0.213) (0.527) Labor Force(SR) 0.025* 0.028* (2.240) (2.520) Spatial Autoregressive Coefficient 0.863* 0.863* 0.854* 0.854* Akaike Info Criterion 16339.500 16339.000 16322.300 16320.600 L og likelihood rook 8165.750 8162.510 8155.140 8151.320 Moran's I rook 0.017 0.013 0.013 0.012 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 201 187 Table 66: Spatial Regression Results for Density of Hazard Scores (Natural Log) at One Mile from Block Group Boundary Using Distance Band Independent Variables Model 1 0.19 mile Model 2 0.19 mile Model 3 0.19 mile Model 4 0.19 mile % Black 0.025* 0.016* (4.120) (2.030) % Asian 0.044** 0.040 ( 1.840) ( 1.630) % Hispan ic 0.114* 0.105* (17.590) (12.650) % Minority 0.067* 0.049* (13.000) (6.750) Population Density (LN) 1.640* 1.430* 1.660* 1.700* ( 9.200) ( 5.920) ( 9.270) ( 7.130) % Below Poverty 0.070* 0.460* (3.700) (2.400) % Owner Occupied 0.006 0.008 (0.861) (1.260) Labor Force(SR) 0.011 0.011 ( 0.553) (0.566) Spatial Autoregressive Coefficient 0.421* 0.427* 0.405* 0.405* Akaike Info Coefficient 19008.100 18999.300 18849.300 18849.200 Log likelihood 9501.070 9493.66 0 9419.630 9416.590 Moran's I distance 0.005 0.005 0.004 0.004 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 202 188 8. ANALYZING THE HEALTH RISKS OF EMISSIONS FROM INDUSTRIAL POLLUTION SOURCES This chapter focuses on the use of hum an health risks from points sources listed in the USEPAs Toxic Release Inventory (TRI) as an indicator of exposure to chronic air pollution in the analysis of environmental justice in the HoustonGalveston Brazoria metropolitan statistical area (HGB MSA) The goal is to explore the association between health risks (modeled risk scores) from chronic exposure to industrial air pollution sources and selected demographic and socioeconomic variables from Census 2000, in order to determine the presence of inequity with respect to the racial/ethnicity and poverty status of the population. Risk scores from the Risk Screening Environmental Indicators (RSEI) model are used to assess health risks from chronic exposure of industrial pollution from TRI facilities. T o estimate the relative risks to chronic human health posed by toxic chemical releases, the model integrates toxicity weights for individual chemicals and chemical categories and exposure estimates, based upon pathway specific reporting of releases to air, water, and land, and the size of the potentially exposed residential population. The toxicity weights are assigned separately for the oral and inhalation exposure pathways, and include both cancer and noncancer effects. Chemical release data from the TR I and pathway specific fate and transport models that consider factors such as wind patterns and stream flow are used to calculate the doses to which people may be exposed (Bouwes et al. 2001) PAGE 203 189 The RSEI model combines toxicity weights with the surrogate dose delivered by each release to obtain a part ial score for each square kilometer grid cell that represents the toxicity adjusted potential human health effects from chronic exposure. The partial scores resulting from TRI emissions at different facilities are summed to obtain an aggregated risk score for each 1km by 1km grid cell for the HoustonMSA. Areal interpolation based on proportional allocation was used to assign risk scores from the 1km by 1km grid cells to the boundaries of census tracts and block groups in the study area, resulting in each analytical unit containing a nonzero value representing health risks throughout the MSA. Although the 1km by 1km grids provide a finer spatial resolution than census tracts or block groups the grids were not used as analytical units to maintain consisten cy with the equity analyses associated with the other pollution indicators that are based on census units. The final risk score of atmospheric pollution (all emitted chemicals) from TRI facilities were then calculated for each geographic analytical unit following the technique for modeling point density described previously in Chapter 4. For each census tract or block group, the areal density of risk scores is measured by dividing the total modeled risk scores from their host facilities by the area enclo sed, at three different distances from the tract or block group boundary (0, 0.5, and 1 mile). Because of the skewed distribution of this variable, the natural logarithm (LN) of the density of risk scores was used for all statistical analyses presented in this chapter. The use of the natural log reduces outliers and yields a normal style distribution of the dependent variable that is more consistent with the requirements of regression modeling. PAGE 204 190 Summary statistics for all variables analyzed are provided i n Tables 67 and 68, respectively, for census tract and block group level data. The descriptive statistics show that census tracts contain a larger density of risk scores than block groups. In addition, the density of risk scores increase as the search rad ius increases from 0 to 1 mile from the boundary due to the possible inclusion of more pollution emissions at larger distances from each unit. The analysis begins with an examination of bivariate parametric correlations between the risk score density at three search radii (0, 0.5, and 1 mile) and relevant explanatory variables at the census tract and block group level of aggregation. Pearsons correlation coefficients associated with the density of risk scores at the three distances from each analytical unit are presented in Table 69 for census tract level and Table 70 for block group level analyses. The correlation analysis at the census tract level shows that at 0 miles from the boundary, population density has the highest positive correlation with t he density of risk scores followed by percent minority, percent Hispanic, percent below poverty, and percent Black. Additionally, percent owner occupied housing has the greatest negative correlation. Labor force and percent Asian do not indicate signific ant correlation with density of risk scores at this level. At 0.5 mile from the census tract boundary population density again has the highest positive correlation with the density of risk scores followed by percent minority, percent Hispanic, percent belo w poverty, and percent Black. Additionally, owner occupied housing has the greatest negative correlation followed by labor force. Percent Asian does not indicate a significant correlation with density of risk scores at this level. PAGE 205 191 The correlations at 1 mile from the tract boundary are similar to those observed at 0.5 mile for all variables. At the block group level, the correlation analysis shows that at 0 miles from the boundary population density has the highest positive correlation with the density of risk scores followed by percent minority, percent Hispanic, percent below poverty, and percent Black. Additionally, owner occupied housing has the greatest negative correlation followed by labor force. Percent Asian does not indicate a significant c orrelation with density of risk scores at this level. At both 0.5 mile and 1 mile from the block group boundary, the correlations for all explanatory variables are similar to those observed at 0 mile. In summary, the correlation analysis consistently indicates that the population density has the strongest positive correlation and most significant linear association with the risk scores at both the census tract and block levels. The premise of environmental inequity is supported by three variables (overal l population density, percent minority, percent Hispanic, percent below poverty, and percent Black) which consistently exhibit a significant and positive relationship with density of risk scores at all search radii and at both the census tract and block gr oup levels. The percentage of owner occupied housing units suggests a negative linear relationship at all distances and levels of aggregation along with labor force except at the boundary of the tract level. Additionally, the Asian percentage does not ex hibit a significance association with the risk scores for any level or aggregation. While correlation analysis provides a preliminary view of the statistical association between each demographic and socioeconomic variable and the density of PAGE 206 192 risk scores, it does not clarify the underlying pattern in sufficient detail. Multivariate regression is therefore used to analyze the simultaneous statistical effects of the explanatory variables on the density of risk scores as related to the health effects of toxic c hemicals released from TRI facilities at the census tract and block group level, respectively. The ordinary least squares (OLS) regression approach is first utilized to model the statistical effect of explanatory variables on the density of risk scores w ithin each analytical unit in the study area, based on the three search radii (0, 0.5, and 1 mile). The results from the OLS regression models for the four different combinations of explanatory variables are summarized in the first set of tables (Tables 7 173 and 7779). This set of tables also provides results from an analysis of spatial dependence using both the contiguity based and distance based methodologies. The results in the OLS tables contain regression coefficients that are bolded, z scores whi ch are in parentheses and p Additionally, these tables contain Morans I which is used for testing for the existence of spatial autocorrelation. The second phase of the analysis uses the spatial autoregressive function (SAR) by utilizing an autoregressive coefficient to account for spatial autocorrelation. The appropriate SAR function (spatial lag or spatial error) was determined using the Lagrange Multiplier test statistic results obtained while running the corresponding OLS model (Anselin 2006) As mentioned previously, the SAR models use a spatial weights matrix specified by three configurations of neighboring census uni ts: queen based contiguity, rookbased contiguity, and distance based bands. Separate queen and rook PAGE 207 193 contiguity analysis was therefore performed in hopes of finding the best fitted model that minimizes the effects of spatial dependence. Additionally, s eparate analysis for distance bands were constructed selecting between 3 and 5 distances which were used as adjustments to eliminate spatial autocorrelation. After the completion of the analyses, the best model was selected based upon the achievement of nonsignificant (p<0.05) Morans I which signifies no spatial autocorrelation and the lowest Akaike Information Criterion (AIC) which is used to determine the fit of the model. The census tract re sults are listed in Tables 7482 and the block g roups are li sted in Tables 8694. The results in the tables contain regression coefficients which are bolded, z scores which are in parentheses, and pvalues which The histograms in Figure 12 summarize the connectivity distribution for census tracts, based on the number of neighbors in each configuration. For queen contiguity, Figure 12(a ) shows that the majority of the census tracts have 5 to 8 neighbors and peaks at 7 neighbors. For the rook contiguity in Figure 12(b), the number of neighbors peaks at 4 but the census tracts noticeably had a positive skew. The histogram for the distanc e based configuration on Figure 12 (c) at 0.50 provides t he connectivity distribution at from the centroid of each unit and indicates that a majority of census tracts have no neighbors at this distance. The connectivity distributions for census block groups are summarized in Figure 13. The histogram for queen contiguity in Figure 13(a) shows that the majority of the block groups have 4 to 8 neighbors and peaks at 6 neighbors. For the rook contiguity in Figure 13 (b), the number of neighbors peaks at 5 and noticeably has a positive skew. The histogram for the distance based configuration in Figure 13(c) provides the connectivity distribution at 0.19 mile from the centroid of each unit and PAGE 208 194 indicates that the number of block groups with no neighbors at this distance is considerably higher than the number of census tracts with no neighbors in Figure s 12. PAGE 209 195 Figure 12 : C ensus tract connectivity histograms used for modeling risk scores a. Queen Contiguity b. Rook Contiguity c. Distance Based Conf iguration (0.50 mile) PAGE 210 196 Figure 13 : B lock group connectivity histograms used for modeling risk scores a. Queen Contiguity b. Rook Contiguity c. Distance Based Configuration (0.19 mile) PAGE 211 197 8.1 Census Tract Level Analysis Ordinary Least Squares Regression This section compares the density of risk scores as computed from the quantity and physiochemical properties of released chemicals, in conjunction with fate and transport modeling at the census tract level, at zero (0), half (0.5) and one (1) mile search radii, based on the application of ordinary least squares (OLS) and spatial autoregressive (SAR) techniques. Table 71 summarizes the results from OLS regression analysis a t the boundary of the census tracts. Models 1 and 2 both indicate that the overall percent minority shows a significant and positive statistical effect on the density of risk scores. When the minority subgroups are analyzed separately in models 3 and 4, the Hispanic percent indicates a significant and positive association with the density of risk scores while percent Asian has a significant and negative association. Percent Black however, has a positive and significant statistical effect on risk scores in model 3, but not in model 4 when the socioeconomic terms are introduced. This stresses the importance of analyzing each racial/ethnic sub group individually and collectively to gain clarity on their environmental justice implications. Furthermore, bo th percent owner occupied housing and labor force are significant and have a negative association with density of risk scores in models 2 and 4. In all four models, the density of risk scores significantly increases with an increase in population density. At 0.5 from the boundary of the census tract (Table 72), the results for the racial/ethnic variables are similar to those observed at 0 mile from the tracts. For models 1 and 2, the density of risk scores increases significantly as the overall percentage of minority residents increases. Furthermore, for models 3 and 4, the Hispanic percentage has a significant and positive statistical effect on the density of risk scores while percent PAGE 212 198 Asian has a significant and negative association for both models. Per cent Black however, has a positive and significant statistical effect on risk scores at model 3, but not in model 4 when the socioeconomic terms are introduced. Furthermore, percent owner occupied has a negative and significant statistical effect on risk scores at model 4, but not in model 2 when the minority subgroups are combined. Again labor force is significant and has a negative association with density of risk scores for models 2 and 4. In all four models, the density of risk scores significantly increases with an increase in population density. At 1 mile from the boundary of the census tracts (Table 73) models 1 and 2 indicate a positive statistical effect of the overall minority percentage and percent Hispanic (models 3 and 4) on the density of r isk scores, which is consistent with the results at 0 and 0.5 miles from the tract boundary. Percentage Black is again positive and significant in model 3, while percent Asian has a negative association at model 3 only. Percent owner occupied housing and labor force are both consistent with 0 mile, where they both have a negative and significant association with density of risk scores. In all four models, the density of risk scores significantly increases with an increase in population density. Morans I for the queen and rook contiguities as well as the distance based ( 0.5 mile ) configuration shows significance at 0, 0.5, and 1 miles search radii, indicating the presence of significant spatial autocorrelation in the data and justifying the need to accoun t for spatial dependence by using a SAR model. Following the assessment for spatial dependence the SAR model was thus implemented for all OLS models where spatial autocorrelation remained. PAGE 213 199 Spatial Regression At the boundary of the census tract level for t he queen contiguity (Table 74), models 1 and 2 indicate a significant and negative association between the density of risk scores at 0 miles and the overall minority percent. Additionally, for models 3 and 4, the Black and Asian percentages have a negative statistical effect on the density of risk scores. Percent owner occupied housing, however, has a negative statistical effect on density of risk scores for model 4 only, but is not influential when the minorities are classified into sub groups. For all models at 0 mile from the tract boundary, the density of risk scores significantly increases when population density increases. Labor force, however, has a positive statistical effect on density of risk scores only for model 2, but is not influential whe n the minority subgroups are included. The queen contiguity SAR models differ from OLS in that the minority and Black percentages indicate negative associations while the Hispanic percentage and labor force becomes nonsignificant. At the boundary of the census tract level for the rook contiguity (Table 75), all SAR models exhibit the same statistical associations as observed in the model for queen contiguity and have similar differences from the corresponding OLS models, except for percent owner occup ied housing which is negative and significant in model 2. Several distance bands were constructed based on selection criteria resulting in minimization of the likelihood term (AIC) and significance of spatial autocorrelation using Morans I (Table 76). Distance bands for models 1 and 3 were constructed at 0.93 mile while models 2 and 4 were constructed at 1.24 miles Percent Black in model 1 has a positive and significant association with density of risk scores. Percent Hispanic for models 3 and 4 has a significantly positive statistical effect on the density of risk scores, PAGE 214 200 as well as percent minority for model 1 only. All socioeconomic variables (percent below poverty, percent owner occupied, and labor force) have a negative and significant associat ion with density of risk scores for models 2 and 4. The only difference between the SAR and the OLS models is that percent below poverty became significant and percent Asian became non significant when spatial dependence was taken into account. At 0.5 mile from the boundary of census tracts for the queen contiguity (Table 77), models 1 and 2 indicate a significant and negative association between the density of risk scores at 0.5 mile and the overall minority percent. Additionally, for models 3 and 4, th e Black and Asian proportions negatively affect the density of risk scores. Percent owner occupied housing and labor force, however, independently have a negative statistical effect on density of risk scores in both models 2 and 4. For all models at 0.5 mile from the tract boundary, density of risk scores significantly increases when population density increases. The queen contiguity SAR models differ from OLS in that the minority and Black percentages show a negative association while the Hispanic percentage for both models becomes nonsignificant. Additionally, percent owner occupied housing for model 2 becomes significant while maintaining the same negative sign. At 0.5 mile from the boundary of census tracts for the rook contiguity (Table 78), all SAR models exhibit the same statistical associations as observed in the model for queen contiguity and had similar differences from the corresponding OLS models. Several distance bands were constructed based on selection criteria resulting in minimization of the likelihood term (AIC) and significance of spatial autocorrelation using Morans I (Table 79). Model 1 was constructed at 0.37 mile model 2 at 0.62 mile PAGE 215 201 and models 3 and 4 at 0.5 mile Percent Black in model 1 has a positive and significant ass ociation with density of risk scores. Percent Hispanic for models 3 and 4 has a significantly positive statistical effect on the density of risk scores, as well as percent minority for models 1 and 2. Percent Asian on the other hand, has a negative and s ignificant association with density of risk scores. Percent owner occupied and labor force indicates a negative and significant association with density of risk scores for models 2 and 4. The only difference between the SAR as compared to the OLS models is that percent owner occupied housing in model 2 became significant when spatial dependence was negated. At 1 mile from the boundary of the census tract level for the queen contiguity (Table 80), model 2 indicates a significant and negative association between the density of risk scores at 1 mile and the overall minority percent. Additionally, percent Black in model 4 has a significant and negative association with density of risk scores. Percent Asian is significant and negative for both models with and without the socioeconomic variables. Percent owner occupied housing and labor force however, independently have a negative statistical effect on density of risk scores on both models 2 and 4. For all models 1 mile from the tract boundary, the densit y of risk scores increases significantly when population density increases. The queen contiguity SAR models differ from OLS in that the overall minority percent indicates a negative association in model 2 and minority percent becomes nonsignificant in mo del 1. Black percentages also show a negative association and become significant in model 4. The Hispanic percentage, on the other hand, becomes non significant in both models. Additionally, the Asian percentage becomes significant in both models while maintaining the same negative sign. PAGE 216 202 At 1 from the boundary of the census tract level for the rook contiguity (Table 81), all SAR models exhibit the same statistical associations as observed in the model for queen contiguity and had similar differences from the corresponding OLS models. Several distance bands at 1 mile were constructed based on selection criteria resulting in minimization of the likelihood term (AIC) and significance of spatial autocorrelation using Morans I (Table 82). Models 1 and 3 were constructed at 0.5 mile model 2 at 0.62 mile and model 4 at 0.37 mile Percent Black in model 1 has a positive and significant association with density of risk scores. Percent Hispanic for models 3 and 4 has a significantly positive statistical effect on the density of risk scores, as well as percent minority for models 1 and 2. Percent owner occupied and labor force had a negative and significant association with density of risk scores for models 2 and 4. The only difference between the SAR as compared to the OLS models is that percent owner occupied housing in model 2 became significant when spatial dependence was negated and percent Asian for both models became non significant. 8.2 Block Group Level Analysis Ordinary Least Squares Regressio n This section compares the density of risk scores as computed from the quantity and physiochemical properties of released chemicals, in conjunction with fate and transport modeling at the block group level, at zero (0), half (0.5) and one (1) mile search radii, based on the application of ordinary least squares (OLS) and spatial autoregressive (SAR) techniques. Table 83 summarizes the OLS regression analysis at the boundary of the block groups. Models 1 and 2 both indicate that the overall percent minori ty shows a significant and positive statistical effect on the density of risk scores. Percent Asian has a significant and negative association in model PAGE 217 203 3, but is nonsignificant when the minority subgroups are included separately for model 4. Percent Bl ack and percent Hispanic have a significant and positive association with density of risk scores. All socioeconomic variables (percent below poverty, percent owner occupied, and labor force) show a negative and significant association with density of risk scores for models 2 and 4. In all four models, the density of risk scores significantly increases with an increase in population density. At 0.5 mile from the boundary of the block group (Table 84), the results for the racial/ethnic variables are similar to those observed at 0 mile from the block groups. At 1 mile from the boundary of the block group (Table 85), the results for the racial/ethnic variables are also similar to those observed at 0 mile from the block groups. The only difference is that the Asian percentage is non significant at 1 mile from the boundary. Morans I for the queen and rook contiguities as well as the distance based ( 0.5 mile ) configuration shows significance at 0, 0.5, and 1 mile search radii indicating the presence of signifi cant spatial autocorrelation in the data and justifying the need to account for spatial dependence by using a SAR model. Following the assessment for spatial dependence, the SAR model was thus implemented for all OLS models where spatial autocorrelation w as detected. Spatial Regression At the boundary of block groups for the queen contiguity (Table 86), models 1 and 2 indicate a significant and negative association between the density of risk scores at 0 miles and the overall minority percent. Additionall y, for models 3 and 4, the Black and Asian populations negatively affect the density of risk scores. For all models 0 miles from the block group boundary, density of risk scores significantly increases when PAGE 218 204 population density increases. Percent Hispanic and all socioeconomic variables are non significant and have no overall statistical effect on the density of risk scores. The queen contiguity SAR models differ from OLS in that the Asian percentage became significant for model 4 and all the socioeconomic variables yield nonsignificant coefficients. At the boundary of block groups for the rook contiguity (Table 87), all SAR models exhibit the same statistical associations as observed in the model for queen contiguity and had similar differences from the corresponding OLS models. A distance band of 0.19 mile (Table 88) was constructed beginning from 0.5 mile from each block group boundary to identify neighboring block groups for assessing spatial dependence. The distance bands were constructed based on s election criteria resulting in minimization of the likelihood term (AIC) and significance of spatial autocorrelation using Morans I. Models 1 and 2 both indicate that the overall percent minority shows a significant and positive statistical effect on the density of risk scores. Percent Asian has a significant and negative association in model 3, but is nonsignificant when the economic terms fare introduced in model 4. Percent Black and percent Hispanic have a significant and positive association with density of risk scores. Percent owner occupied and labor force have a negative and significant association with density of risk scores for models 2 and 4, while percent below poverty is negative and significant only in model 4. In all four models, the de nsity of risk scores significantly increases with an increase in population density. The distance SAR models differ from OLS in that percent below poverty in model 3 becomes nonsignificant. At 0.5 mile from the boundary of block groups for the queen cont iguity (Table 89), models 1 and 2 indicate a significant and negative association between the density of PAGE 219 205 risk scores at 0.5 mile and the overall minority percent. Additionally, for models 3 and 4, the Black and Asian populations negatively affect the dens ity of risk scores. Percent Hispanic, however, is significant only in the model that contains the socioeconomic terms. Labor force has a negative statistical effect on density of risk scores on both models 2 and 4, while Percent owner occupied housing ha s a significant and negative statistical effect in model 4 only. For all models at 0.5 mile from the block group boundary, density of risk scores significantly increases when population density increases. The queen contiguity SAR models differ from OLS in that the percent Asian loses significance when the socioeconomic terms are introduced. Hispanic percentage also becomes nonsignificant for model 3. Additionally, percent owner occupied housing for model 2 loses significance while maintaining the same signage and percent below poverty loses significance for both models. At 0.5 mile from the boundary of block groups for the rook contiguity (Table 90), all SAR models exhibit the same statistical associations as observed in the model for queen contigui ty and had similar differences from the corresponding OLS models. A distance band of 0.19 mile (Table 91) was constructed beginning from 0.5 mile from each block group boundary to identify neighboring block groups for assessing spatial dependence. The distance bands were constructed based on selection criteria resulting in minimization of the likelihood term (AIC) and significance of spatial autocorrelation using Morans I. There are no statistical differences between the SAR models as compared to the OLS models for the 0.5 mile distance band. At 1 mile from the boundary of block groups for the queen contiguity (Table 92), model 2 indicates a significant and negative association between the density of risk PAGE 220 206 scores at 1 mile and the overall minority perc ent. Additionally, percent Black and percent Hispanic in model 4 have a significant and negative association with density of risk scores. Percent Asian is significant and negative for both models with and without the economic variables. Percent owner occupied housing and labor force however, independently have a negative statistical effect on density of risk scores in both models 2 and 4. For all models at 1 mile from the block group boundary, density of risk scores significantly increases when populat ion density increases. The queen contiguity SAR models differ from OLS in that the percent minority gains a negative association and model 1 becomes nonsignificant. Additionally, Black and Hispanic percentages gain negative association and lose signific ance for model 3. Asian percentage for both models also becomes significant while maintaining the same signage. Lastly, percent below poverty loses significance. At 1 mile from the boundary of block groups for the rook contiguity (Table 93), all SAR mod els exhibit the same statistical associations as observed in the model for queen contiguity and had similar differences from the corresponding OLS models, except percent owner occupied housing became nonsignificant in the model with separate minority sub groups. A distance band of 0.19 mile (Table 94) was constructed beginning from 0.5 mile from each block group boundary to identify neighboring block groups for assessing spatial dependence. The distance bands were constructed based on selection criteri a resulting in minimization of the likelihood term (AIC) and significance of spatial autocorrelation using Morans I. Percent Black and percent Hispanic in models 3 and 4 indicate a positive and significant association with the density of risk scores, as well as PAGE 221 207 percent minority for models 1 and 2. All socioeconomic variables are significant and have a positive statistical effect on the density of risk scores. The only difference between the SAR as compared to the OLS models is that percent Asian for model 4 became significant. 8.3 Discussion and Interpretation The analyses presented in this chapter focuses on the geographic distribution of pollution burdens measured in terms of human health risks from industrial point sources. The results reveal that the areal density of modeled risk scores in the HoustonMSA is significantly associated with population density, the overall proportion of minorities, and the Hispanic percentage at the census tract level, regardless of whether conventional or spatial regr ession was utilized. The findings from the analysis of risk scores reveals distinct differences between the OLS, contiguity based SAR, and distance based SAR techniques. The OLS and distance based methodologies provide similar statistical results at each level of aggregation and distance from the analytical unit. The presence of racial/ethnic minorities and Hispanics, in particular, indicate a positive statistical effect on the density of risk scores at all distances, even after controlling for the effec ts of other explanatory variables, in OLS and distance based SAR models. Additionally, the percentage of Asians living within the census tract boundaries indicates a significantly negative statistical effect at all search radii in the OLS models and when spatial autocorrelation is removed at the 0.5 mile distance band ( 0.5 mile ). The Black percentage shows a significant and positive statistical effect when the density of risk scores is measured at all search radii from the tract boundary, for both ordinar y regression and when using the spatial regression at all distance bands. The poverty rate PAGE 222 208 also shows a negative relationship with the density of risk scores at the 0 mile search radius while using the distance band only. Additionally, the percentages of owner occupied housing units are negative and significant only when spatial autocorrelation is removed using the distance bands. Labor force shows a negative statistical effect on the density of risk scores using OLS and when spatial autocorrelation is t aken into account using the distance bands. Lastly, population density has a positive association with the risk scores at all aggregation, search radii and distances using all methodological techniques. The contiguity based SAR models exhibit results that are similar to those from the OLS and distancebased techniques, except that the Black and minority percentages become negative and more significant. This indicates how the definition of neighborhood for measuring spatial dependence affects the sign and s ignificance of a key explanatory variable (i.e., Black percentage). Additionally, the Asian percentage is negative and significant at all levels and aggregations using the queen and rook contiguities, except when using the distance band at 0 and 1 mile se arch radii. This suggests that a large proportion of the Asian population in this MSA is located outside and away from census tracts facing the greatest risks from industrial pollution sources. Furthermore, the positive significance of Hispanics at all levels using the distance based SAR suggests that a large proportion of the Hispanic population resides in tracts that are adjacent to those facing greater health risks from toxic emissions. As observed at the census tract level, population density, minorit y percent, and percent of Hispanics, are again the most influential variables related to the density of risk scores at the block group level. Again, the OLS and distance based SAR models indicate PAGE 223 209 similar statistical results at each aggregation and distanc e. The presence of racial/ethnic minorities and Hispanics (OLS and distance bands only), in particular, indicate a positive statistical effect on the density of risk scores at all distances, even after controlling for the effects of other explanatory vari ables. Additionally, the percentage of Asians living within the block group boundaries shows negative significance using OLS at all search radii and when spatial autocorrelation is removed. The Black percentage shows a significant and positive statistica l effect when the density of risk scores is measured at all search radii from the tract boundary, for both ordinary regression and when using the spatial regression at all distance bands for the models. The poverty rate also shows a negative relationship to risk scores overall using OLS and distance bands at all search radii. Additionally, the percentage of owner occupied housing and labor force are negative and significant using OLS and when spatial autocorrelation is removed using the distance bands. L astly population density ha s a positive association with the risk scores at all aggregation, search radii, and distances for both regression techniques. In summary, the results suggest that that the density of risk scores from industrial toxic emissions increase significantly with an increase in the proportion of racial/ethnic minorities at all search radii and at both the census tract and block levels of aggregation. More specifically, the overall minority, Black, and Hispanic proportions and population density are the most significant explanatory factors in determining the density of risk scores at all search radii, all aggregations, and both regression modeling techniques. Blacks and Asians were found to reside outside of areas with the highest health r isk. Therefore, it can be surmised that minorities facing the highest health risks are primarily comprised of the Hispanic population, particularly when using the contiguity based PAGE 224 210 spatial regression models. Furthermore, the distance bands provide evidenc e that a large percentage of the Hispanic population resides along the edges of analytical units and are exposed to higher health risks. Additionally, risk scores decrease consistently across all search radii as home ownership rates decline and the signif icance of the negative association between the Asian proportion and risk scores increases at the block group resolution. Lastly, the analytical units with higher population density were found to be exposed to greater health risks. The combination of the aforementioned findings suggests that the Hispanic residents comprise the largest component of the potentially exposed population. The locations of this minority subgroup coincide spatially with areas facing the greatest health risks from industrial air pollution in the HoustonGalvestonBrazoria MSA. PAGE 225 211 Table 67: Descriptive Statistics for Variables at the Census Tract Level n =886 Simple Statistics Mean Std Dev Min Max Variables Density of Risk Scores(LN): 0 Miles 5.110 1.880 0.010 9.660 Density of Risk Scores(LN): 0.5 Miles 5.190 1.830 0.010 9.050 Density of Risk Sco res(LN): 1 Mile 5.210 1.800 0.010 8.800 % Black 18.140 24.290 0.000 98.160 % Asian 4.140 5.890 0.000 37.710 % Hispanic 28.380 23.310 0.000 100.000 % Minority 52.700 29.830 0.000 100.000 Population Density (LN) 7.550 8.020 0.000 10.390 % Below Poverty 14.210 10.920 0.000 66.690 % Owner Occupied 61.010 25.440 0.000 100.000 Labor Force (SR) 60.270 16.530 0.000 115.450 PAGE 226 212 Table 68: Descriptive Statistics for Variables at the Block Group Level n=2705 Simple Statistics Mean Std Dev Min Max Variables Density of Risk Scores(LN): 0 Miles 5.160 1.920 0.000 9.930 Density of Risk Scores(LN): 0.5 Miles 5.270 1.850 0.010 9.350 Density of Risk Scores(LN): 1 Mile 5.290 1.810 0.010 9.210 % Black 17.560 25.760 0.000 100.000 % Asian 3.760 6.660 0.000 68.160 % Hispanic 28.860 25.590 0.000 100.000 % Minority 52.210 31.630 0.000 100.000 Population Density (LN) 7.710 1.580 0.000 11.150 % Below Poverty 14.490 12.190 0.000 73.360 % Owner Occupied 63.110 27.430 0.000 100.000 Labor Force (SR) 34.140 10.650 0.000 83.610 PAGE 227 213 Table 69: Pearson Correlation Coefficients Associated with the Density of Risk Scores (Natural Log) Released at t he Census Tract Level. n=886 p<0.05 0 Mile 0.5 Mile 1 Mile Variables r p Value r p Value r p Value % Black 0.122 0.001 0.130 0.001 0.133 <0.001 % Asian 0.028 0.404 0.026 0.439 0.024 0.477 % Hispanic 0.418 <0.001 0.432 <0.001 0.437 <0.001 % Minority 0.429 <0 .001 0.446 <0.001 0.454 <0.001 Population Density (LN) 0.648 <0.001 0.577 <0.001 0.510 <0.001 % Below Poverty 0.332 <0.001 0.354 <0.001 0.360 <0.001 % Owner Occupied 0.349 <0.001 0.361 <0.001 0.362 <0.001 Labor Force (SR) 0.017 0.616 0.046 0.017 0.068 0.045 PAGE 228 214 Table 70: Pearson Correlation Coefficients Associated with the Density of Risk Scores (Natural Log) Released at the Block Gro up Level n=2705 p<0.05 0 Mile 0.5 Mile 1 Mile Variables r p Value r p Value r p Value % Black 0.094 <0.001 0.105 <0.001 0.111 <0.001 % Asian 0.021 0.275 0.018 0.359 0.019 0.331 % Hispanic 0.380 <0.001 0.406 <0.001 0.409 <0.001 % Minority 0.382 <0.001 0.414 <0.001 0.423 <0.001 Population Density (LN) 0.625 <0.001 0.538 <0.001 0.450 <0.001 % Below Poverty 0.261 < 0.001 0.285 <0.001 0.289 <0.001 % Owner Occupied 0.274 <0.001 0.294 <0.001 0.299 <0.001 Labor Force (SR) 0.046 <0.001 0.071 0.001 0.086 <0.001 PAGE 229 215 Table 71: Least Squares Regression Results for Density of Risk Scores (Natural Log) at Zero Mile from Census Tract Boundary Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.005* 0.003 (2.390) (1.290) % Asian 0.024* 0.024* ( 2.820) ( 2.870) % Hispanic 0.019* 0.017* (8.420) (6.120) %Minority 0.011* 0.006* (6.540) (2.390) Population Density (LN) 0.676* 0.757* 0.690* 0.713* (20.630) (22.130) (20.760) (21.420) % Below Poverty 0.002 0.012 (0.275) ( 1.590) % Owner Occupied 0.008* 0.010* ( 4.040) ( 5.080) Labor Force (SR) 0.023* 0.001* ( 7.710) ( 5.690) R 2 0.446 0.499 0.479 0.514 F Statistic 355.720* 175.634* 202.820* 132.850* Akaike Info Criterion 3115.890 3032.270 3065.120 3044.490 Log likelihood 1554.940 1510.130 1527.560 1496.760 Morans I queen 0.643* 0. 643* 0.602* 0.612* Morans I rook 0.657* 0.656* 0.616* 0.624* Morans I 0.5 mile 0.060* 0.071* 0.048* 0.055* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 230 216 Table 72: Least Squares Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Census Tract Boundary a. All race variables represent a single race. b. *p<0.05, **p<0.10 Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.009* 0.002 (4.220) (0.680) % Asian 0.015** 0.020* ( 1.750) ( 2.470) % Hispanic 0.023* 0.017* (10.410) (6.120) % Minority 0.015* 0.005* (9.280) (2.220) Population Density (LN) 0.715* 1.010* 0.728* 1.010* (16.760) (21.950) (16.700) (22.170) % Below Poverty 0.006 0.006 (0.861) ( 0.920) % Owner Occupied 0.010 0.010* ( 5.160) ( 5.4 10) Labor Force (SR) 0.038* 0.038* ( 12.130) ( 12.570) R 2 0.388 0.507 0.424 0.538 F Statistic 280.398* 180.823* 161.823* 145.904* Akaike Info Criterion 3152.400 2967.860 3104.000 2914.400 Log likelihood 1573.200 1477.930 154 7.000 1449.200 Morans I queen 0.718* 0.691* 0.681* 0.631* Morans I rook 0.729* 0.704* 0.692* 0.675* Morans I 0.5 mile 0.049* 0.049* 0.039* 0.036* PAGE 231 217 Ta ble 73: Least Squares Regression Results for Density of Risk Scores (Natural Log) at One Mile from Census Tract Boundary Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.011* 0.002 (5.240) (0.874) % Asian 0.006 0.015** ( 0. 681) ( 1.880) % Hispanic 0.027* 0.018* (11.640) (6.400) % Minority 0.018* 0.006* (10.720) (2.570) Population Density (LN) 0.671* 1.150* 0.674* 1.140* (13.580) (20.550) (13.300) (20.620) % Below Poverty 0.005 0.005 (0.801) ( 0.80 0) % Owner Occupied 0.011* 0.011* ( 5.540) ( 5.700) Labor Force(SR) 0.046* 0.047* ( 13.700) ( 14.180) R 2 0.337 0.488 0.373 0.519 F Statistic 224.695* 167.553* 130.881* 135.274* Akaike Info Criterion 3196.270 2974.160 315 1.540 2922.530 Log likelihood 1595.130 1481.080 1570.770 1453.270 Morans I queen 0.750* 0.701* 0.689* 0.672* Morans I rook 0.759* 0.711* 0.727* 0.683* Morans I 0.5 mile 0.050* 0.046* 0.042* 0.033* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 232 218 Table 74: Spatial Regression Results for Density of Risk Scores (Natural Log) at Zero Mile from Census Tract Boundary Using Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.003* 0.002* ( 2.8 30) ( 1.560) % Asian 0.016* 0.018* ( 4.180) ( 4.460) % Hispanic 0.001 0.001 ( 0.570) ( 0.037) % Minority 0.002* 0.003* ( 2.150) ( 2.460) Population Density (LN) 0.230* 0.233* 0.246* 0.247* (13.180) (11.950) (13.940) (12.620) % Below Poverty 0.002 0.004 (0.466) ( 1.270) % Owner Occupied 0.002 0.002* ( 1.830) ( 2.36) Labor Force(SR) 0.001** 0.001 ( 0.462) ( 0.394) Spatial Autoregressive Coefficient 0.878* 0.871* 0.869* 0.862* Akaike Info Criterion 1868.550 1840.650 1854.030 1853.540 Log likelihood queen 930.274 927.614 921.013 917.280 Moran's I queen 0.033 0.030 0.024 0.038* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 233 219 Table 75: Spatial Regres sion Results for Density of Risk Scores (Natural Log) at Zero Mile from Census Tract Boundary Using Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.003* 0.002** ( 2.770) ( 1.570) % Asian 0.016* 0.017* ( 4.2 00) ( 4.510) % Hispanic 0.001 0.001 ( 0.670) ( 0.230) % Minority 0.002* 0.003* ( 2.160) ( 2.310) Population Density (LN) 0.227* 0.231* 0.242* 0.239* (13.270) (12.100) (13.030) (13.130) % Below Poverty 0.002 0.004 (0.582) ( 1.0 60) % Owner Occupied 0.002** 0.002* ( 1.680) ( 2.280) Labor Force(SR) 0.001 0.001 ( 0.566) (0.360) Spatial Autoregressive Coefficient 0.870* 0.863* 0.861* 0.857* Akaike Info Criterion 1855.940 1856.870 1841.720 1840.130 Log likeliho od rook 923.971 921.437 914.861 912.043 Moran's I rook 0.029 0.039* 0.018 0.029 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 234 220 Table 76: Spatial Regression Results for Density of Risk Scores (Natural Log) at Zero from Cens us Tract Boundary Using Distance Band Independent Variables Model 1 0.93 mile Model 2 1.24 miles Model 3 0.93 mile Model 4 1.24 miles % Black 0.005* 0.001 (2.290) (0.277) % Asian 0.001 0.001 (0.041) (0.030) % Hispanic 0.016* 0.008* (6.560) (2.880) % Minority 0.010* 0.004 (5.290) (1.460) Population Density (LN) 0.596* 0.648* 0.598* 0.648* (18.580) (19.430) (18.180) (19.350) % Below Poverty 0.160* 0.018* ( 2.790) ( 2.980) % Owner Occupied 0.011* 0.011* ( 6 .290) ( 6.410) Labor Force(SR) 0.017* 0.018* ( 7.060) ( 7.270) Spatial Autoregressive Coefficient 0.519* 0.696* 0.489* 0.696* Akaike Info Criterion 3098.620 2728.060 2942.320 2724.570 Log likelihood 1443.990 1357.030 1466.160 1354.28 0 Moran's I distance 0.034 0.031 0.033 0.030 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 235 221 Table 77: Spatial Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Census Tract Boundary Using Queen Cont iguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.002* 0.002* ( 2.610) ( 2.600) % Asian 0.011* 0.012* ( 4.020) ( 4.350) % Hispanic 0.001 0.001 ( 0.984) ( 1.240) % Minority 0.001* 0.007* ( 2.260) ( 5.45 0) Population Density (LN) 0.205* 0.332* 0.218* 0.257* (14.180) (14.800) (14.900) (13.920) % Below Poverty 0.003 0.001 ( 1.020) ( 0.443) % Owner Occupied 0.005* 0.002* ( 6.760) ( 2.940) Labor Force(SR) 0.006* 0.004* ( 5.3 20) ( 3.640) Spatial Autoregressive Coefficient 0.945* 0.982* 0.940* 0.922* Akaike Info Criterion 1236.210 1221.290 1224.890 1205.410 Log likelihood queen 614.100 604.650 606.440 593.710 Moran's I queen 0.102* 0.060* 0.091* 0.120* a. All ra ce variables represent a single race. b. *p<0.05, **p<0.10 PAGE 236 222 Table 78: Spatial Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Census Tract Boundary Using Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.002* 0.002* ( 2.640) ( 2.760) % Asian 0.011* 0.011* ( 4.170) ( 4.440) % Hispanic 0.001 0.001 ( 1.220) ( 1.500) % Minority 0.016* 0.003* ( 2.410) ( 3.590) Population Density (LN) 0.199* 0.241* 0.212* 0.251* 14.350 13.630 (15.120) (14.190) % Below Poverty 0.003 0.001 (1.340) ( 0.239) % Owner Occupied 0.001* 0.002* (2.170) ( 2.660) Labor Force(SR) 0.004* 0.004* ( 3.810) ( 3.860) Spatial Autoregressive Coefficient 0.941* 0.924* 0.9 37* 0.919* Akaike Info Criterion 1192.440 1171.140 1180.670 1160.880 Log likelihood rook 592.218 578.569 584.336 571.438 Moran's I rook 0.110* 0.114* 0.968* 0.126* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 237 223 Table 79 : Spatial Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Census Tract Boundary Using Distance Band Independent Variables Model 1 0.37 mile Model 2 0.62 mile Model 3 0.5 mile Model 4 0.5 mile % Black 0.005* 0.003 ( 2.480) (1.080) % Asian 0.023* 0.022* ( 2.760) ( 2.730) % Hispanic 0.018* 0.016* (8.110) (5.850) % Minority 0.002* 0.005* (2.150) (2.210) Population Density (LN) 0.230* 0.101* 0.693* 0.761* (13.180) (21.970) (20.700) (22.620) % Be low Poverty 0.003 0.011 (0.501) ( 1.550) % Owner Occupied 0.010* 0.009* ( 5.140) ( 4.280) Labor Force(SR) 0.039* 0.024* ( 12.480) ( 8.240) Spatial Autoregressive Coefficient 0.227* 0.290* 0.259* 0.295* Akaike Info Criterion 3137.960 2944.410 3053.870 2965.250 Log likelihood 1564.980 1466.200 1521.930 1457.620 Moran's I distance 0.006 0.017 0.010 0.014 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 238 224 Table 80: Spatial Regression Results for Density of Risk Scores (Natural Log) at One Mile from Census Tract Boundary Using Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.001 0.002* ( 1.470) ( 2.340) % Asian 0.006* 0.007* ( 2.630) ( 3.21) % Hispan ic 0.001 0.001 ( 0.276) ( 1.150) % Minority 0.001 0.002* ( 1.140) ( 2.880) Population Density (LN) 0.155* 0.219* 0.163* 0.227* (12.270) (12.600) (12.590) (12.920) % Below Poverty 0.002 0.001 (0.849) ( 0.292) % Owner Occupied 0.002* 0.002* ( 2.870) ( 3.200) Labor Force(SR) 0.005* 0.005* ( 5.390) ( 5.480) Spatial Autoregressive Coefficient 0.974* 0.956* 0.972* 0.952* Akaike Info Criterion 902.225 862.775 898.860 859.156 Log likelihood queen 447.110 424.3 87 443.430 420.578 Moran's I queen 0.112* 0.145* 0.111* 0.140* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 239 225 Table 81: Spatial Regression Results for Density of Risk Scores (Natural Log) at One Mile from Census Tract Bounda ry Using Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.001 0.002* ( 1.550) ( 2.060) % Asian 0.006* 0.007* ( 2.830) ( 3.360) % Hispanic 0.001 0.001 ( 0.518) ( 1.380) % Minority 0.001 0.002* ( 1.330) ( 3.130) Population Density (LN) 0.149* 0.214* 0.157* 0.222* (12.270) (12.820) (12.660) (13.170) % Below Poverty 0.002 0.001 (1.050) ( 0.136) % Owner Occupied 0.001* 0.002* ( 2.530) ( 2.880) Labor Force(SR) 0.005* 0. 005* ( 5.760) ( 5.840) Spatial Autoregressive Coefficient 0.971* 0.952* 0.968* 0.949* Akaike Info Criterion 862.448 820.155 858.575 816.490 Log likelihood rook 427.224 403.078 423.288 399.245 Moran's I rook 0.125* 0.152* 0.116* 0.147* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 240 226 Table 82: Spatial Regression Results for Density of Risk Scores (Natural Log) at One Mile from Census Tract Boundary Using Distance Band Independent Variables Model 1 0.5 mile Model 2 0.62 mile Model 3 0.5 mile Model 4 0.37 mile % Black 0.010* 0.002 (4.800) (0.676) % Asian 0.008 0.020 ( 0.971) ( 1.920) % Hispanic 0.025* 0.017* (11.380) (6.290) % Minority 0.017* 0.006* (9.820) (2.580) Population Density (LN) 0.667* 1.140* 0.672* 1.140* (13.770) (20.620) (13.560) (20.970) % Below Poverty 0.003 0.004 (0.444) ( 0.660) % Owner Occupied 0.011* 0.010* ( 5.520) ( 4.280) Labor Force(SR) 0.047* 0.024* ( 14.010) ( 5.520) Spatial Auto regressive Coefficient 0.160* 0.175* 0.162* 0.212* Akaike Info Coefficient 3168.430 2929.720 3151.540 2912.010 Log likelihood 1564.980 1466.200 1521.930 1447.010 Moran's I distance 0.019 0.005 0.008 0.005 a. All race variables represent a sing le race. b. *p<0.05, **p<0.10 PAGE 241 227 Table 83: Least Squares Regression Results for Density of Risk Scores (Natural Log) at Zero Mile from Block Group Boundary a. All race variables represent a single race. b. *p<0.05, **p<0.10 Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.004* 0.003* ( 3.410 ) (2.080) % Asian 0.013* 0.006 ( 2.910) ( 1.280) % Hispanic 0.016* 0.014* ( 12.720) ( 9.650) % Minority 0.010* 0.007* (9.790) (5.780) Population Density(LN) 0.682* 0.754* 0.681* 0.741* (34.880) (38.000) (34.540) (37.40 0) % Below Poverty 0.006** 0.009* ( 1.740) ( 2.600) % Owner Occupied 0.006* 0.006* ( 5.330) ( 5.240) Labor Force (SR) 0.039* 0.006* ( 14.340) ( 13.770) R 2 0.412 0.459 0.430 0.475 F Statistic 946.950* 458.330 516.290 348.5590* Akaike Info Criterion 9780.500 9560.600 9684.650 9543.390 Log likelihood 4887.250 4774.300 4937.300 4734.230 Morans I queen 0.682* 0.670* 0.661* 0.654* Morans I rook 0.692* 0.680* 0.670* 0.665* Morans I 0.19 mile 0.033* 0.04 8* 0.033* 0.046* PAGE 242 228 Table 84: Least Squares Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Block Group Boundary Ind ependent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.008* 0.004* (7.270) (2.780) % Asian 0.086** 0.004 ( 1.93) ( 0.980) % Hispanic 0.022* 0.015* (18.120) (11.150) % Minority 0.015* 0.008* (15.570) (6.930) Population Density (LN) 0.772* 1.210* 0.777* 1.191* (27.100) (39.830) (26.980) (39.230) % Below Poverty 0.007* 0.010* ( 2.270) ( 3.160) % Owner Occupied 0.009* 0.009* ( 8.480) ( 8.3600) Labor Force (SR) 0.078* 0.075* ( 26.650) ( 26. 140) R 2 0.347 0.495 0.377 0.513 F Statistic 719.690* 529.370* 409.040* 405.528* Akaike Info Criterion 9843.390 9155.830 9721.070 9063.430 Log likelihood 4918.700 4571.910 4855.540 4523.710 Morans I queen 0.777* 0.722* 0.754* 0.706* Morans I roo k 0.784* 0.729* 0.761* 0.714* Morans I 0.19 mile 0.011* 0.021* 0.010* 0.020* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 243 229 Table 85: Least Squares Regression Results for Density of Risk Scores (Natural Log) at One Mile from Bloc k Group Boundary Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.009* 0.005* (7.270) (3.660) % Asian 0.006 0.001 ( 1.160) ( 0.135) % Hispanic 0.023* 0.016* (17.680) ( 11.830) % Minority 0.018* 0.009* (18.740) (7 .870) Population Density (LN) 0.699* 1.470* 0.823* 1.436* (20.080) (37.600) (22.630) (36.830) % Below Poverty 0.009* 0.012* ( 2.960) ( 3.710) % Owner Occupied 0.020* 0.010* ( 9.700) ( 9.530) Labor Force (SR) 0.097* 0.094* ( 29.920) ( 29.420) R 2 0.290 0.480 0.321 0.497 F Statistic 551.670* 498.600* 318.730* 380.913* Akaike Info Criterion 9977.040 9139.550 9861.100 9053.700 Log likelihood 4985.520 4563.780 5097.330 4518.850 Morans I queen 0.816* 0.728* 0.796* 0.7 14* Morans I rook 0.821* 0.734* 0.722* 0.722* Morans I 0.19 mile 0.009* 0.018* 0.008* 0.017* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 244 230 Table 86: Spatial Regression Results for Density of Risk Scores (Natural L og) at Zero Miles from Block Group Boundary Using Queen Contiguity a. All race variables represent a single race. b. *p<0.05, **p<0.10 Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.002* 0.002* ( 3.240) ( 3.150) % Asian 0.007* 0.007* ( 3.240) ( 3.660) % Hispanic 0.001 0.001 ( 1.610 ) ( 1.850) %Minority 0.001* 0.002* ( 2.930) ( 3.270) Population Density (LN) 0.149* 0.153* 0.154* 0.155* (16.460) (15.910) (16.920) (16.160) % Below Poverty 0.002 0.001 (1.370) (0.701) % Owner Occupied 0.001 0.0002 ( 0.187) ( 0.355) Labor Force(SR) 0.001 0.001 ( 0.910) ( 0.030) Spatial Autoregressive Coefficient 0.912* 0.909* 0.909* 0.907* Akaike Info Criterion 5320.860 5323.080 5311.560 5317.900 Log likelihood queen 2656.430 2654.540 2649.780 2649.260 Mor an's I queen 0.019 0.019 0.017 0.017 PAGE 245 231 Table 87: Spatial Regression Results for Density of Risk Scores (Natural Log) at Zero Miles from Block Group Boundary Using Rook Contiguity Indepe ndent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.001* 0.002* ( 3.040) ( 2.880) % Asian 0.006* 0.006* ( 3.650) ( 3.500) % Hispanic 0.001 0.001 ( 1.290) ( 1.500) % Minority 0.001* 0.002* ( 2.620) ( 2.910) Populat ion Density (LN) 0.147* 0.151* 0.152* 0.154* (16.700) (16.100) (17.120) (16.340) % Below Poverty 0.002 0.001 (1.190) (0.543) % Owner Occupied 0.001 0.0002 ( 0.211) ( 0.372) Labor Force(SR) 0.0009 0.0005 (0.824) ( 0.470) Spati al Autoregressive Coefficient 0.907* 0.905* 0.904* 0.903* Akaike Info Criterion 5284.500 5287.510 5275.730 5280.970 Log likelihood rook 2638.250 2636.760 2631.860 2631.390 Moran's I rook 0.010 0.010 0.008 0.008 a. All race variables represen t a single race. b. *p<0.05, **p<0.10 PAGE 246 232 Table 88: Spatial Regression Results for Density of Risk Scores (Natural Log) at Zero Miles from Block Group Boundary Using Distance Band Independent Variables Model 1 0.19 mile Model 2 0.19 mile Model 3 0.19 mile Mo del 4 0.19 mile % Black 0.004* 0.002* (3.670) (1.970) % Asian 0.013* 0.006 ( 3.020) ( 1.310) % Hispanic 0.016* 0.014* (12.860) (9.490) % Minority 0.010* 0.097* (9.790) (5.610) Population Density (LN) 0.682* 0.767* 0.686* 0.75 5* (34.880) (39.250) (35.160) (38.670) % Below Poverty 0.005 0.008* ( 1.440) ( 2.280) % Owner Occupied 0.007* 0.006* ( 5.770) ( 5.660) Labor Force(SR) 0.042* 0.040* ( 15.440) ( 14.870) Spatial Autoregressive Coefficient 0.5 01* 0.579* 0.507* 0.581* Akaike Info Coefficient 9732.900 9480.030 9636.070 9404.940 Log likelihood 4863.450 4734.010 4813.030 4694.470 Moran's I distance 0.002 0.002 0.002 0.003 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 247 233 Table 89: Spatial Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Block Group Boundary Using Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.001* 0.0010* ( 2.620) ( 3.860) % As ian 0.005* 0.005* ( 5.210) ( 4.630) % Hispanic 0.001 0.001* ( 1.62) ( 2.85) % Minority 0.001* 0.001* ( 2.470) ( 4.230) Population Density (LN) 0.142* 0.196* 0.150* 0.200* (20.000) (21.510) (20.960) (21.870) % Below Poverty 0. 001 0.0004 (1.270) (0.494) % Owner Occupied 0.001 0.001** ( 1.640) ( 1.900) Labor Force(SR) 0.007* 0.007* ( 9.480) ( 9.150) Spatial Autoregressive Coefficient 0.966* 0.951* 0.965* 0.950* Akaike Info Criterion 2339.850 2247.500 2319.100 2236.020 Log likelihood queen 1165.930 1116.750 1153.550 1108.010 Moran's I queen 0.121* 0.121* 0.113* 0.118* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 248 234 Table 90: Spatial Regression Results for Density of Ri sk Scores (Natural Log) at Half Mile from Block Group Boundary Using Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.001* 0.001* ( 2.530) ( 3.650) % Asian 0.005* 0.004* ( 5.040) ( 4.470) % Hispanic 0.00 1 0.001* ( 1.530) ( 2.650) % Minority 0.0005* 0.001* ( 2.360) ( 3.980) Population Density (LN) 0.139* 0.191* 0.146* 0.194* (20.220) (21.570) (21.130) (21.930) % Below Poverty 0.001 0.001 (1.060) (0.318) % Owner Occupied 0.004 0.001** ( 1.600) ( 1.860) Labor Force(SR) 0.007* 0.007* ( 9.400) ( 9.080) Spatial Autoregressive Coefficient 0.964* 0.949* 0.963* 0.948* Akaike Info Criterion 2259.630 2170.290 2240.330 2159.540 Log likelihood queen 1125.810 1078.140 1114.170 1070.770 Moran's I queen 0.114* 0.112* 0.109* 0.109* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 249 235 Table 91: Spatial Regression Results for Density of Risk Scores (Natural Log) at Half Mile from Block Group Bounda ry Using Distance Band Independent Variables Model 1 0.19 mile Model 2 0.19 mile Model 3 0.19 mile Model 4 0.19 mile % Black 0.008* 0.004* (7.320) (2.740) % Asian 0.009** 0.004** ( 1.920) ( 0.950) % Hispanic 0.022* 0.015* (18.07 0) (11.040) % Minority 0.015* 0.008* (15.520) (6.840) Population Density (LN) 0.771* 1.220* 0.776* 1.200* (27.140) (40.270) (27.020) (39.660) % Below Poverty 0.007* 0.009* ( 2.140) ( 3.000) % Owner Occupied 0.009* 0.009* ( 8.72 0) ( 8.580) Labor Force(SR) 0.079* 0.076* ( 27.110) ( 26.610) Spatial Autoregressive Coefficient 0.287* 0.424* 0.268* 0.416* Akaike Info Coefficient 9834.870 9480.030 9713.940 9040.300 Log likelihood 4914.430 4559.330 4815.970 4512.16 0 Moran's I distance 0.003 0.004 0.003 0.004 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 250 236 Table 92: Spatial Regression Results for Density of Risk Scores (Natural Log) at One Mile from Block Group Boundary Using Queen Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0.001* 0.001* ( 0.943) ( 3.270) % Asian 0.003* 0.003* ( 3.800) ( 3.590) % Hispanic 0.001 0.001* ( 0.439) ( 2.81) % Minority 0.001 0.008* ( 0.755) ( 3.780) Po pulation Density (LN) 0.104* 0.192* 0.110* 0.195* (17.220) (21.950) (17.920) (22.170) % Below Poverty 0.003 0.005 (0.488) ( 0.078) % Owner Occupied 0.007* 0.008* ( 3.410) ( 3.610) Labor Force(SR) 0.009* 0.009* ( 14.010) ( 13. 790) Spatial Autoregressive Coefficient 0.985* 0.967* 0.984* 0.966* Akaike Info Criterion 1022.090 821.093 1011.450 817.101 Log likelihood queen 507.047 403.547 499.726 399.55 Moran's I queen 0.130* 0.111* 0.117* 0.110* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 251 237 Table 93: Spatial Regression Results for Density of Risk Scores (Natural Log) at One Mile from Block Group Boundary Using Rook Contiguity Independent Variables Model 1 Model 2 Model 3 Model 4 % Black 0. 001 0.001* ( 0.867) ( 3.060) % Asian 0.003* 0.003* ( 3.660) ( 3.430) % Hispanic 0.001 0.001* ( 0.437) ( 2.640) % Minority 0.001 0.001* ( 0.714) ( 3.550) Population Density (LN) 0.102* 0.189* 0.107* 0.189* (17.400) (21.860 ) (18.060) (22.070) % Below Poverty 0.001 0.001 (0.357) ( 0.186) % Owner Occupied 0.001* 0.001 ( 3.200) ( 3.410) Labor Force(SR) 0.009* 0.009* ( 13.790) ( 13.570) Spatial Autoregressive Coefficient 0.983* 0.965* 0.982* 0.965* Akaike Info Criterion 960.780 767.392 951.258 764.040 Log likelihood rook 476.390 376.696 469.629 373.020 Moran's I rook 0.110* 0.962* 0.104* 0.930* a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 252 238 Table 94: Spatial Regressi on Results for Density of Risk Scores (Natural Log) at One Mile from Block Group Boundary Using Distance Band Independent Variables Model 1 0.19 mile Model 2 0.19 mile Model 3 0.19 mile Model 4 0.19 mile % Black 0.011* 0.005* (9.120) (3.620) % Asian 0.001 0.001** ( 0.183) ( 0.088) % Hispanic 0.025* 0.016* (20.390) (11.690) % Minority 0.018* 0.009* (18.130) (7.770) Population Density (LN) 0.699* 1.480* 0.698* 1.440* (20.840) (37.980) (20.500) (37.200) % Below Poverty 0.009* 0.011* ( 2.850) ( 3.570) % Owner Occupied 0.010* 0.010* ( 9.920) ( 9.720) Labor Force(SR) 0.098* 0.095* ( 30.320) ( 29.810) Spatial Autoregressive Coefficient 0.274* 0.409* 0.248* 0.394* Akaike Info Coefficient 9969.93 0 9118.220 9855.630 9034.850 Log likelihood 4981.970 4559.110 4922.810 4506.430 Moran's I distance 0.005 0.005 0.003 0.005 a. All race variables represent a single race. b. *p<0.05, **p<0.10 PAGE 253 239 9. CONCLUSIONS Issues of fairness and equity in the distribution of environmental pollution sources such as industrial manufacturing facilities, landfills, and hazardous waste disposal sites sparked the beginning of the environmental justice movement. The disproportionate placement of polluting sites in p redominantly low income and minority neighborhoods has resulted in an uneven health burden that requires researchers to move beyond locational analyses of facilities to more rigorous methodological approaches that make a connection between exposure to toxi c chemical releases and adverse health risks. The siting patterns of industrial facilities and the exploration of their relationship with proximate populations are grounded in geographic principles and spatial techniques are thus necessary to analyze asso ciations between environmental pollution and the characteristics of potentially exposed residents While n umerous environmental justice studies have examined the location pattern of polluting facilities, recent research has emphasized the need to analyze disproportionate exposure to pollutants and related health risks on the basis of chemical specific emission data and atmospheric modeling techniques The goals of this dissertation were to extend quantitative research on environmental justice and address s everal methodological limitations of previous studies by: 1) using new indicators of exposure to air pollution and contemporary risk modeling techniques; 2) assessing disparities in human health risks, instead of focusing only on potential exposure or pro ximity to pollution sources; and 3) implementing multivariate PAGE 254 240 regression models that consider the effect of spatial autocorrelation and are thus more suitable for analyzing variables derived from census data. These goals were specifically addressed by usin g the Risk Screening Environmental Indicators model (RSEI) as provided by the US Environmental Protection Agency. This model provides information on toxicity weighted emissions (hazard scores) and human health risks (modeled risk scores) from air pollutan ts by including detailed data on the quantity, nature, and dispersion of chemical releases from facilities in the Toxic Release Inventory (TRI). Compared to the measures used in previous studies on the environmental justice implications of TRI facilities, the RSEI data provides two distinct advantages. First, the detailed information on chemical toxicity allows an improved measure of the potential human health effects caused by pollution. Second, the data used are based on a realistic representation of exposure, incorporating chemical fate and transport, as well as stack heights and exit gas velocities. The application of this new data set thus allowed for an extension of the traditional approach for measuring the magnitude of pollution in quantitative analy sis of environmental justice. Additionally, the implementation of advanced multivariate and spatial regression techniques instead of traditional bivariate statistical measures and regression models allowed for the indicators of environmental justice to be examined on the basis of differ ent analytical units and distances. Furthermore, the assessment and adjustment for spatial dependence resulted in the use of multivariate models that are less likely to violate the assumptions associated with regression res iduals or bias the regression coefficients. Lastly, the use of least likelihood statistics combined with the adjustments for spatial dependence lends to an improved PAGE 255 241 selection of variables for explaining the geographic distribution of health risks and pollution sources. Newly emerging themes in the environmental justice research literature (Gee and Payne Sturges 2004; Maantay 2005; Payne Sturges and Gee 2006) such as inequities in the levels of pollution and public health risks were explored in the HoustonMSA by determining the association between various pollution indicators and the racial/ethnic and socioeconomic characteristics of the population. Four quantitative measures of pollution were used for these assess ments. First, locations of industrial toxic emissions were examined by assessing the presence or absence of TRI facilities across census units in the study area. Second, the quantity of chemicals (pounds) emitted into the atmosphere was used as a method of exposure assessment on the population. Third, the quantity of pounds emitted was extended by assessing the contribution of toxicity of each chemical as part of the exposure assessment (hazard scores). Modeling the atmospheric dispersion of chemicals i n conjunction with their toxic impacts on the population was the fourth and final avenue of risk assessment. 9.1 Evaluating Inequities for the Four Pollution Indicators The findings from the analyses of the four pollution sources overwhelmingly indicate th at the overall minority percentage, the Hispanic percentage, and population density provi de the most consistent association s on the magnitude of pollution at all search radii, all aggregations, and both standard and spatial regression modeling techniques. The nuances of these analytical results lend significant findings to this study. Locational analysis of TRI facilities provides evidence that the effect of race/ethnicity on TRI location becomes stronger and more significant as the search radius increases, at PAGE 256 242 both the census tract and block group levels of aggregation. Thus, choosing both the correct analytical unit and the correct scale for measuring pollution is essential. Although higher Hispanic proportions suggested a stronger statistical relations hip with TRI pollution than higher Black p roportions and percent below poverty the results associated with the Black population and poverty rates provide necessary evide nce that not all statistical effect s from pollution sources can be captured by using a single analytical unit. Thus it becomes important to utilize more flexible analytical techniques in order to assess whether minority subgroups (e.g., Blacks) reside inside or on the edge of boundaries of census tracts or block groups. The analysis of r acial/ethnic inequity for the density of pounds and density of hazard scores suggested similar results. As observed in the locational analysis, levels of pounds and hazard scores are higher in census units containing a greater percentage of Hispanics. Additionally, a higher percentage of Asians showed a negative statistical effect while Blacks and labor force indicated a positive statistical effect on levels of pounds and hazard scores. The human health risks from toxic emissions, as measured by modeled risk scores, are also impacted by the proportion of minorities, Hispanics, and population density. Although census units associated with TRI facilities, higher emission quantities, and higher hazard scores indicated lower population density, modeled risk scores are greatest in areas of higher population density. Since the modeled risk scores are interpolated data which provide consistent scores for all analytical units, it presents an improved picture of the magnitude of health risks on the studied popul ation. Thus, the findings of the positive statistical effect of population density on adverse health risks PAGE 257 243 from air pollution are consistent with other environmental justice and health disparities studies (Krieger a nd Subramanian 2003; Gee and Payne Sturges 2004; Maantay 2005; Payne Sturges and Gee 2006; Krieger 2005; Chen et al. 2006) The Hispanic population represents the minority subgroup exposed to the highest levels of health risk in the study area, while the Asian population is facing the lowest health risks The Black population, on the other hand, experiences an increase in health risks only when neighbor relationships in the analysis of spatial dependence are based on the distance based approach Most of the model results suggest that the Black and Hispanic populations within the boundaries of both the analytical units have the greatest health risks. While these two minority subgroups share the same spaces in certain parts of the study area, the history of a greater density of Blacks residing within the inner core and Hispanics being more dispersed throughout the HoustonMSA may influence these findings. The findings for owner occupied housing are also consistent with the results of previous environment al justice studies that found fewer toxic facilities and lower levels of pollution in areas with smaller home ownership rate (Morello Frosch et al. 2001; Ash and Fetter 2004; Pastor Jr et al. 2005) 9.2 Effect of Regression Methodology and Analytical Unit Selection Ordinary Least Squares Regression vs. Spatial Regression An important goal of this research was to implement multivariate regression models that considered the effect of spatial dependence and do not violate the assumptions of ordinary least squares regression (OLS) regression, in order to obtain a clearer picture of the disproportionate impacts of industrial pollution in the HoustonMSA. OLS regression was the traditional modeling technique utilized to examine the statistical effects of various racial/ethnic and PAGE 258 244 socioeconomic explanatory variables. The inherent geographic distribution of these population variables along with pollution sources suggested a need to identify if spatial autocorrelation exist s and the independence assumptions of OLS are satisfied. SAR models were then used to improve the OLS models and account for spatial autocorrelation in the data. OLS and SAR techniques were not used for the traditional host/nonhost analysis of census units that was based on bivariate analysis and logistic regression of TRI locations. Therefore, three quantitative pollution indicators were analyzed to supply evidence that accounting for spatial dependence provides a significant improvement over using OLS regression alone. Additionally, multiple neighborhood configurations were used to identify the set of neighbors which provide the best fit for each regression model while minimizing spatial dependence. In assessing the influences of racial/ethnic and socioeconomic variables on pounds emitted and toxicity weighted pounds (hazard scores), the Asian percentage and labor force was found to be significant only when spatial dependence was taken into account. Consequently, these relationships with health risks w ould not be noticed if spatial autocorrelation existed in the data. Additionally, when looking at the health risks, the Black and minority populations who reside within the boundary of the analytical units experience increased health risks as compared to the same groups outside the boundary. This would not be obvious if the distance band and contiguity based approaches were not implemented and compared to each other. Thus, neighbors in the unit boundaries have a greater influence on health risks than nei ghbors at a greater distance and when spatial dependence is not accounted for when using OLS. The findings clearly demonstrate that PAGE 259 245 the choice of technique for constructing neighbor relationships on the basis of census units (contiguity based or distancebased) affects the results of the SAR models. It is important to consider the limitations associated with the use of buffers (0.5 mile and 1.0 mile) to calculate certain dependent variables (pounds emitted, hazard scores, and modeled risk scores) Thes e buffer distances potentially introduce additional spatial autocorrelation in the dependent variable that could be partially responsible for causing OLS regression residuals to be spatially autocorrelated. The regression results obtained from the analyse s of dependent variables at the boundaries of census units (0 mile buffer) are thus more reliable because they are not based on values derived from neighboring census units. The findings associated with the 0.5 mile and 1.0 mile buffer s should therefore be interpreted with caution. Although the exact amount of spatial autocorrelation in regression errors introduced by the dependent variable itself cannot be currently estimated, the results indicate that the value of Morans I from the OLS regressions incr ease as the buffer size around census units increases. Future research should focus on the systematic separation of these spatial effects and determining the extent to which spatial autocorrelation in regression residuals is influenced by spatial autocorr elation in the dependent variable Census tract vs. Block group analysis Another goal of this research was to assess the effects of the modifiable areal unit problem (Anderton et al. 1994; Glickman et al. 1995) on the findings of environmental justice analysis. More specifically, the study attempted to examine whether conducting the same set of analyses at two different spatial resolutions (census tract or block group) lead to different results, when analyzing the PAGE 260 246 influence of racial/ethnic and socioeconomic explanatory variables on the pollution indicators. For locational analysis of TRI facilities, it was found that the likelihood of TRI location is related to population density, overall minority percentage, and t he Hispanic percentage at both the census tract and block group levels. While racial/ethnic subgroups were found to be related to the quantity of pollutants emitted from industrial sources, the Black percentage was found to be significant only when emissi on density was measured at the boundaries of census tracts, but not at the block group boundaries. Additionally, the Black sub group was significant only at the farthest distance (1 mile) away from tract boundaries and at all three distances from the bloc k group boundaries, for both regression modeling techniques. This confirms the importance of analyzing data at more than one spatial resolution because the nuances between the distances were not evident at a single analytical unit. As for the hazard score s, the Black percentage at the block group level showed significance at all distances and in majority of the models tested with and without the socioeconomic variables. Furthermore, home ownership rate had a positive statistical effect on the hazard score s at the census tract level but exhibited a negative statistical effect at the block group level. Regarding the health risks from toxic emissions, the findings in this study suggest that race/ethnicity and socioeconomic variables are influential at both t he census tract and block levels of aggregation. Overall minority, Black, and Hispanic populations, along with population density, are influential at both analytical units. Although the Asian percentage showed a negative association with health risks, the significance of this variable increases when the block group unit of analysis is used. These findings suggest that a quantitative study based on one particular PAGE 261 247 areal unit is unlikely to produce reliable empirical support for environmental injustice, because the statistical results are affected by the spatial resolution of the data. 9.3 Geographies of Environmental Justice A variety of theories have been proposed to understand and explain the occurrence of spatial and en vironmental injustices. These we re outlined in Chapter 2 of this dissertation and include theories of social justice, theories of neighborhood change, and economic r esidential, and location theories The results suggest that principles of contractarianism, egalitarianism, and libertaria nism are violated because environmental health risks are not evenly distributed across all racial/ethnic and socioeconomic groups in the Houston MSA. Industrial pollution sources were found to be predominantly located in areas containing a higher proporti on of minorities, and Hispanic s, and to a lesser e xtent Blacks, in particular Additionally, the magnitude of pollution was found to be primarily in the same communities with a disproportionate risk burden being placed on these groups. Although equal acc ess to clean air and low health risks is not evident in this research, basic principles of utilitarianism are observed. By placing industrial facilities in confined areas of the MSA it limits the overall exposure to the majority of the population thus supporting a basic utilitarian principle. The contemporary view of the Marxist exploitation principle as outlined by (Smith 1994) is supported by the findings of this research. This idea states that unfair advantage within broader distribution arrangements which include uneven access to resources in general and to the means of production in particular Since neighborhoods populated with minorities contain the majority of pollution sources, greater exposures and health r isks, this group does not receive equal access to clean air. Empirical support for theories of neighborhood change PAGE 262 248 cannot be established through this research because longitudinal analysis is necessary to understand the process through which the socially disadvantaged groups and pollution sources came to occupy the same spaces. Examples of the tenets of geographic principles associated with place, scale, and space in relation to environmental justice were addressed in this research The over representatio n of the minority population in the most polluted parts of the study area exhibits the importance of place. The fact that White and Asian populations are loca ted primarily in areas without industrial pollution sources and low health risks show that place matters in the MSA when addressing environmental justice concerns Furthermore, scale was found to be of substantial importance with respect to the resolution of spatial data ( census tracts to block groups ) The use of block groups or a finer resolution provided evidence of an increase in the significan ce of specific explanatory variables and contributed a more detailed understanding of how race/ethnicity and socioeconomic factors relates to pollution sources and the magnitude of pollution. Thus, it is e ssential to examine the role of geographic scale when analyzing environmental justice and health disparities. Lastly, the effects of space were recognized explicitly in the analysis and findings of this research. In previous environmental justice research, traditional regression modeling has been conducted without taking into account the spatial clustering of pollution and populations. This research demonstrated the importance of assessing and adjusting for spatial dependence in the data Spatial autoco rrelation was found to exist in all analyses based on traditional regression modeling. Furthermore, several variables including the Asian percentage are only significant when spatial dependence is negated. Therefore, understanding the importanc e of spati al relationships and PAGE 263 249 geographically referenced data is essential when analyzing environmental injustices and health disparities. 9.4 Public Health Implications The World Health Organization (WHO) (2007) states that the determinant s of health include : the social and economic environment, the physical environment, and the persons individual characteristics and the behaviors. This research assesse d specific aspects of these determinants by evaluating relationships between race/ethnicity and socioeconomic status in relation to pollution sources and indicators of exposure in the Houston MSA. Evidence of disparities related to excess exposure of toxi c chemicals was observed using the Risk Screening Environmental Indicators (RSEI) model Specifically, minorities and the Hispanic population, in particular, were found to be di sproportionately exposed to health risks compared to the remainder of the studied population. These results beg for future questions to be posed. From this research systematic connections between levels of exposure and measureable outcomes such as cancers and asthma rates can be addressed. These results will be provided to the Houston MSA public health community so that they can use these findings as a guide to address the needs of populations that are potentially facing excess toxic chemical exposure. The importance of conducting public health research in a spatial context ve rsus traditional analytical methods alone was also evident from this study The research provides evidence that the data and variables commonly selected for health disparities research (e.g., race/ ethnicity, income) are inherently spatial C onsequently analyses conducted with spatial tools and methodologies could provide an improved assessment of relationships. While this study focused on the general population, future research will PAGE 264 250 utilize a more detailed exploration of specific sub populations that mo re susceptible to the adverse health effects of toxic chemical exposure. The se groups will include young children, fetuses to address prenatal exposure, the elderly and persons with compromised immune systems (e.g., persons with HIV/AIDS and liver and ki dney disease ) 9.5 Limitations and Future Research This study indicated that certain racial/ethnic or socioeconomic groups are restricted to specific locations within the Houston metropolitan area and are disproportionately exposed to the adverse health risks from industrial pollution These groups reside in these spaces for numerous reasons, such as economical limitations that restrict housing choices employment (dis)opportunities, lack of political representation, racial/ethnic discrimination or soci al and cultural links to the local community (Been 1994; Liu 2001) The methodologies implemented in this study capture the pattern of inequity associated with the geography of pollution sources, racial/ethnic and socioeconomic factors in the Houston MSA. Although this study provides evidence for environm ental inequity, there are several limitations that can be addressed by future research. First, the Risk Screening Environmental Indicators (RSEI) modeling data, although a great improvement on traditional locational information on polluting facilities, su ffer from three limitations. The RSEI data are arbitrary in nature meaning that the emissions data can only be compared to each other and toxicity weights may not directly correlate with criteria used for listing chemicals in the TRI. Therefore, the RSEI should be use d only as a screening tool Environmental analysis with RSEI data could be enhanced by conducting a longitudinal study on locations of facilitie s and nearby PAGE 265 251 populations over a period of several decades using consistent census data to deter mine the population and facilities succes sion of the area. Additionally, the RSEI model analyzes chronic effects associated with long term exposure and not acute toxicity to humans or the environment. This limitation could be rectified by performing a si milar analysis using data on accidental releases of toxic chemicals in conjunction with explanatory variables from Census 2000. This would provide a more comprehensive picture of the combined effects of chronic and acute human health risk on the residents in this study area. Other limitations are that the toxicity weights used reflect the single most sensitive chronic human health effect dependent upon inhalation and oral pathways (fish consumption only), and not the most severe health effect or the syner gistic effects of exposure to multiple chemicals. These issues can contribute to biases in toxicity calculations by positively or negatively skewing the results and by not accounting for the additive effects of simultaneous releases of toxic chemicals. B ecause the RSEI data provides information for the person with the highest sensitivity to a chemical instead of only a severe health effect, an improvement in understanding the additive and synergistic effects of the chemicals may be the most important exte nsion that can be made to this study. A second limitation to this study is that no qualitative research was conducted to complement or support the quantitative analysis of environmental justice. Although the RSEI database provides an im provement over envir onmental data sets used previously to examine industrial air pollution from point sources, qualitative analysis could better help ascertain population exposure to other types of pollution sources ( stationary and mobile) Future research should focus on de termining occupational exposures of the residents and PAGE 266 252 exposure from other pollution sources in regards to locations of dwellings. This would allow first hand historical information concerning possible exposure sources to be obtained and related to the hea lth risks. Additionally, the length of time and other locations where residents have previously resided need to be considered. Research on health disparities would benefit this study if government officials provided historical information of where and how long the polluting industries have existed in the studied areas and any notable patterns of previously detected health effects on the impacted populations. Finally there is a need to link data from the RSEI model to measureable health outcomes. The m odeled risk scores in the RSEI database are based on toxicity of chemicals and their dispersal in the environment, which cannot be used to determine the actual health outcomes such as the occurrence of a type of cancer or respiratory illness. Studies of e nvironmental justice can be improved substantially by combining the spatial modeling benefits of the RSEI model with data on the spatial distribution of individual cancer cases and other adverse health outcomes to assess potential cause and effect for expo sure to toxic emissions. 9.6 Policy R ecommendations Although there are several additional questions which could be asked to improve our understanding of environmental injustice the findings of this study provide several a venues for public policy recommendations Regardless of the methodology utilized, Hispanics and Blacks represent the two minority subgroups that are disproportionately exposed to health risks from industrial pollution in the MSA although Hispa nics overwhelmingly received the greatest ex posures This information can be used by local PAGE 267 253 environmental regulatory agencies to advance environmental justice through the restriction of new industrial facilities near neighborhoods populated by minority residents. The identification of t he health di sparities reported in this dissertation could provide the local government with the ability to formulate a grassroots effort to assess the health and wellbeing of minority residents In addition, this study can provide local community groups and environme ntal activists empirical evidence of unequal exposure and health risks which can help them achieve improved health surveillance in this metropolitan area or support possible litigation against specific polluti ng industries. In conclusion, this research gi ves credence to the disparate impact s of industrial pollution and the resultant health effects on people and places It also provides evidence that although lower socioeconomic status is associated with the extent of pollution, race/ethnicity remains one of the strongest predictors of environmental health risks. The results substantiate that environmental injustices and health disparities exist in the Houston MSA, even after controlling for population density and various economic factors that are known to influence industrial location. PAGE 268 254 LITERATURE CITED Acevedo Garcia, D. 2000. Residential segregation and the epidemiology of infectious diseases. Social Science and Medicine 51:11431161. Allison, P. D. 1999. Logistic Regression Using SA S: Theory and Apllication. Cary: SAS Institute Inc. Anderton, D. L., A. B. Anderson, J. M. Oakes, and M. R. Fraser. 1994. Environmental equity: the demographics of dumping. Demography 31 (2):229248. Anselin, L. 2006. Spatial Regression. Urbana: Spatial Analysis Laboratory. Anselin, L., and A. K. Bera. 1998. Spatial Dependence in Linear Regression Models with and Introduction to Spatial Econometrics. In Handbook of Applied Economic Statistics eds. A. Ullah and D. E. A. Giles, 237289. New York: Marcel Dekker. Ash, M., and T. R. Fetter. 2004. Who Lives on the Wrong Side of the Environmental Tracks? Evidence from the EPA's Risk Screening Environmental Indicators Model. Social Science Quarterly (Blackwell Publishing Limited) 85 (2):441462. Bai, M., M. M. Hill emeier, and E. J. Lengerich. 2007. Racial/ethnic disparities on symptom severity among children hospitalized with asthma. Journal of Health Care for the Poor and Underserved 18:5461. Been, V. 1994. Locally undesirable land uses in minority neighborhoods: disproportionately siting or market dynamics? The Yale Law Journal 103 (6):13831422. Berry, M., and P. Bove. 1997. Birth weight reduction associated with residents nears hazardous waste landfill Environmental Health Perspectives. 105 (8):856861. Bobak, M ., and D. A. Leon. 1999. Pregnancy outcomes and air pollution: an ecological study in the districts of the Czech Republic 19861988. Occupational and Environmental Medicine 56 (8):539543. PAGE 269 255 Bouwes, N., S. M. Hassur, and M. D. Shapiro. 2001. Empowerment through risk related information: EPA's Risk Screening Environmental Indicators Project, 131. Amherst: University of Massachusetts Amherst. Bowen, W. M., Salling, M.J., Haynes, K.E., and E. J. Cyran. 1995. Toward environmental jusitice:spatial equity in Ohio and Cleveland. Annals of the Association of American Geographers 85 (4):641663. Brooks, N., and R. Sethi. 1997. The distribution of pollution: community characteristics and exposure to air toxics. Journal of environmental economics and management 32:233250. Brown, P., Mayer, B., Zavestoski, S., Luebke, T. Mandelbaum, T and S. McCormick. 2004. Clearing the air and breathing freely: the health politics of air pollution and asthma. International Journal of Health Services 34 (1):3963. Bullard, R. D. 1983. Solid waste sites and the black Houston community. Sociological Inquiry 53 (23). . 1990. Dumping in Dixie: Race, Class, and Environmental Quality 3rd ed. Boulder Westview Carpenter, D. O., R. J. Hussain, D. F. Berger, J. P. Lombardo, and H. Y. Park. 2002. Electrophysiologic and behavioral effects of perinatal and acute exposure of rats to lead and polychlorinated biphenyls. Environmental Health Perspectives 110 (Supplemental 3):377386. Castree, N. 2005. Place:connections and boundaries in an interde pendent world. In Key Concepts in Geography eds. S. L. Holloway, Rice, S.P. and G. Valentine, 165185. London: Sage Publications. Centers for Disease Control and Prevention. 2004. Smoking and Tobacco Use, ed. Health and Human Services. http://0 www.cdc.gov.mill1.sjlibrary.org/tobacco/data_statistics/tables/adult/table_2.htm (Accessed September 21, 2007). Chakraborty, J. 2001. Acute exposure to extrem ely hazardous substances: an analysis of environmental equity. Risk Analysis 21 (5). PAGE 270 256 Chakraborty, J., and M. P. Armstrong. 1997. Exploring the use of buffer analysis for the identification of impacted areas in environmetnal equity assessment. Cartography and Geographic Information Systems 24 (3):145157. . 2004. Thinking Outside the Circle: Using Geographical Knowledge to Focus Environmental Risk Assessment Investigation. In WorldMinds: Geographical Perspectives on 100 Problems eds. D. Janelle, Warf, B. and K. Hansen, 435442. Dordrecht, The Netherlands: Kluwer Academic Publications. Chang, K. 2006. Introduction to Geographic Information Systems New York: McGrawHill. Chen, J. T., D. H. Rehkopf, Waterman, P.D., Subramanian, S.V., Coull, B.A., Cohen, B., Ostrem, M., and N. Krieger. 2006. Mapping and measuring social disparities in premature mortality: The impact of census tract poverty within and across Boston neighborhoods, 19992001. Journal of urban health 83 (6):10631084. Clinton, W. J. 1994. Executi ve Order 12898: Federal actions to address environmental justice in minority populations and low income populations. The White House. Coreil, J., Bryant, C.A., and J. N. Henderson. 2001. Social and Behavioral Foundations of Public Health Thousand Oaks: S age Publications Inc. Cutter, S. L., Holm, Danika, and L. Clark. 1996. The role of geographic scale in monitoring environmental justice. Risk Analysis 16 (4):517122. Daniels, G., and S. Friedman. 1999. Spatial inequality and distribution of industrial tox ic releases: evidence from the 1990 TRI. Social Science Quarterly (Blackwell Publishing Limited) 80 (2):244261. Dent, A. L., Fowler, D.A., Kapla, B.M., Zarus, G.M., and W. D. Henriques. 2000. Using GIS to study the health impact of air emissions. Drug and Chemical Toxicology 23 (1):161178. Dolinoy, D. C., and M. L. Miranda. 2004. GIS modeling of air toxics releases from TRI reporting and nonTRI reporting facilities: impacts for environmental justice. Environmental Health Perspectives 112 (17). PAGE 271 257 Dolk, H., and M. Vrijheid. 2003. The impact of environmental pollution on congenital anomalies. British Medical Bulletin 68:2545. Dormann, F., M. Carsten, J. McPherson, M. B. Arajo, R. Bivand, J. Bolliger, G. Carl, R. G. Davies, A. Hirzel, W. Jetz, W. Daniel Kis sling, I. Khn, R. Ohlemller, P. R. Peres Neto, B. Reineking, B. Schrder, F. M. Schurr, and R. Wilson. 2007. Methods to account for spatial autocorrelation in the analysis of species distributional data: a review. Ecography 30, no. 5. Downey, L. 2003. Spatial measurement, geography, and urban racial inequality. Social Forces 81:937954. . 2006. Environmental racial inequality in Detroit. Social Forces 85 (2):771796. Edlin, G., and E. Golanty. 2004. Health and Wellness Boston: Jones and Bartlett Publi shers. Fotheringham, S. A., and W. S. Wong. 1997. The modifiable areal unit problem in multivariate statistical analysis. Environment and Planning A 23:1025144. Fujita, M. 1989. Urban Economic Theory: Land Use and City Size Cambridge, England: Cambridge University Press. Gee, G. C., and D. C. Payne Sturges. 2004. Environmental health disparities: a framework integrating psychosocial and environmental concepts. Environmental Health Perspectives. 112 (17):16451949. Gilboa, S. M., Mendola, P., Olshan, A.F., Langlois, P.H., Savitz, D.A., Loomus, A.H., and D. E. Fixler. 2005. Relation between ambient air quality and selected birth defects, seven county, Texas, 19972000. American Journal of Epidemiology 162 (3):238252. Glickman, T. S. 1994. Measuring environm ental equity with geographic information systems. Resources :2 6. Glickman, T. S., D. Golding, and R. Hersh. 1995. GIS based environmental equity analysis:A case study of TRI facilities in the Pittsburg area. In Computer Supported Risk Management. PAGE 272 258 Green Media Toolshed. 2005. Pollution report card: www.scorecard.org Grineski, S., B. Bolin, and C. Boone. 2007. Criteria air pollution and marginalized population: environmental inequity in metropolitan Phoenix, Arizona. Social Science Quarterly (Blackwell Publishing Limited) 88 (2):535554. Grove, J., A. Troy, J. O'Neil Dunne, W. R. Burch, M. Cadenasso, and S. Pickett. 2006. Characterization of Households and its Implications for the Vegetation of Urban Ecosystems. Ecosystems 9, no. 4. Herod, A. 2005. Scale: the local and the global In Key Concepts in Geography eds. S. L. Holloway, Rice, S.P., and G. Valentine, 229247. London: Sage Publications. Hinds, W. C. 1999. Aerosol Technology: Properties, Behavior, and Measureme nt of Airborne Particles New York: John Wiley & Sons. Hotelling, H., and A. C. Darnell eds. 1990. The collected economics articles of Harold Hotelling Springer Verlag: New York. Institute of Medicine. 1999. Toward Environmental Justice: Research, Education and Health Policy Needs Washington, D.C. : National Academy Press. . 2006. Examining the health disparities research plan of the national institutes of health. Washington, D.C.: The National Academies Press. Jain, N. B., and H. Hu. 2006. Childhood correlates of blood lead levels in Mumbai and Delhi. Environmental Health Disparities 111 (3):466470. Kanemoto, Y. ed. 1987. Externalities in space. In Urban dynamics and urban externalities Switzerland: Hardwood Academics. Keppel, K., E. Pamuk, and J. Lynch. 2005. Methodological issues in measuring health disparities, 2005. Vital Health Statistics 2 (141):116. Kissling, W. D., and G. Carl. 2008. RESEARCH PAPER: Spatial autocorrelation and the selection of simultaneous autoregressive models. Global Ecology & Biogeography 17, no. 1. PAGE 273 259 Klassen, C. D. 2001. Casarett & Doull's Toxicology: The Basic Science of Poisons : McGraw Hill. Krieger, N. 2005. Defining and investigating social disparities in cancer:critical issues. Cancer causes and control 16:514. Krieger, N., Chen, J.T., Waterman, Pamela, D., Rehkopf, D.H., and S. V. Subramanian. 2003. Race/ethnicity, gender, and monitoring socioeconomic gradients in health: a comparison of area based socioeconomic measures the public health disparities geocoding projec t. American Journal of Public Health 93 (10):16551671. Lanphear, B. P., Hornung, R., Khoury, J., Yolton, K., Baghurst, P., Vellinger, D.C., Canfield, R.L., Dietrich, K.N., Bornschein, R., Greene, T., Rothenberg, S.J., Needleman, H.L., Schnaas, L., Wasserm an, G., Graziano, J. and R. Roberts. 2005. Low level environmental lead exposure and childrens intellectual function: an international pooled analysis. Environmental Health Perspectives 113 (7):894946. Laveist, T. A. ed. 2002. Race, Ethnicity and Healt h. San Francisco: Jossey Bass. Lee, C. 2002. Environmental justice: building a unified vision of health and the environment. Environmental Health Disparities 110 (Supplemental 2):141144. Liu, F. 1997. Dynamics and causation of environmental equity, locall y unwanted land uses, and neighborhood changes. Environmental Management 21 (5):643656. . 2001. Environmental Justice Analysis Boca Raton: CRC Press. Maantay, J. 2002. Mapping Environmental Injustices: Pitfalls and Potential of Geographic Information Systems in Assessing Environmental Health and Equity. Environmental Health Perspectives 110 (Supplement 2):161171. . 2005. Asthma and air pollution in the Bronx: methodological and data considerations in using GIS for environmental justice and health r esearch. Health and Place 13:3256. McMaster, R., H. Leitner, and E. Sheppard. 1997. GIS Based environmental equity and risk assessment: methodological problems and prospects. Cartography and Geographic Information Systems 24 (3):172189. PAGE 274 260 Melosi, M. V., and J. A. Pratt eds. 2007. Energy Metropolis Pittsburgh: University of Pittsburgh Press. Mennis, J. L., and L. Jordan. 2005. The distribution of environmental equity:exploring spatial nonstationarity in multivariate models of air toxic releases. Annals of t he Association of American Geographers 95 (2):249268. Mohai, P., and B. Bryant eds. 1992. Race and the Incidence of Environmental Hazards: A Time for Discourse Boulder: Westview. Mohai, P., and R. Saha. 2006. Reassessing racial and socioeconomic disparit ies in environmental justice research. Demography 43:383399. Morello Frosch, R., and R. Lopez. 2006. The riskscape and the color line: examining the role of segregation in environmental health disparities. Environmental Research 102:181196. Morello Frosc h, R., M. Pastor Jr, and J. L. Sadd. 2001. Environmental justice and southern Calfornia's "riskscape": The distribution of air toxics exposures and health risks among diverse communities. Urban Affairs Review 36 (4):551578. Morello Frosch, R., and E. D. S henassa. 2006. The environmental riskscape and social inequality: implications for explaining maternal and child health disparities. Environmental Health Perspectives 114 (8):11501153. National Cancer Institute. 2007. Health Disparities Defined, ed. Cente r to Reduce Cancer Health Disparities. http://crchd.cancer.gov/definitions/defined.html (Accessed September 21, 2007). National Research Council. 2004. Eliminating health disparities: measu rement and data needs, eds. M. Ver Ploeg and E. Perrin. Washington, D.C.: National Academies of Science. O'neill, M. S., M. Jerrett, and I. Kawachi. 2003. Health, wealth and air pollution:advancing theory and methods. Environmental Health Perspectives 111 (16):18611870. Oleckno, W. A. 2002. Essential Epidemiology: Principles and Applications : Waveland Press. PAGE 275 261 Pastor Jr, M., R. Morello Frosch, and J. L. Sadd. 2005. The air is always cleaner on the other side: race, space, and ambient air toxics exposure in California. Journal of urban affairs 27 (2):127148. Pastor Jr, M., J. L. Sadd, and R. MorelloFrosch. 2004. Waiting to Inhale: The Demographics of Toxic Air Release Facilities in 21stCentury California. Social Science Quarterly (Blackwell Publishing Limi ted) 85 (2):420440. Payne Sturges, D., and G. C. Gee. 2006. National environmental health measures for minority and low income populations: tracking social disparities in environmental health. Environmental Research 102:154171. Perlin, S. A., Sexton, K., and W. S. Wong. 1999. An examination of race and poverty for populations living near industrial sources of air pollution. Journal of Exposure Analysis an Environmental Epidemiology 9:2948. Pollock, P. H., and M. E. Vittas. 1995. Who bears the burdens of environmental pollution? Race, ethnicity, and environmental equity in Florida. Social Science Quarterly (Blackwell Publishing Limited) 76 (2):294310. Pulido, N. 1996. The role of geographic scale in monitoring environmental justice. Risk Analysis 16 (4):517526. . 2000. Rethinking environmental racism: white privilege and urban development. Annals of the Association of American Geographers 90 (1):1240. Rawls, J. 1980. John Rawls' Theory of Social Justice : Ohio University Press. . 1999. A Theory of J ustice, Revised Harvard: Harvard University Press. Scenna, N. J., and A. S. M. Santa Cruz. 2005. Road risk analysis due to the transportation of chlorine in Rosario city. Reliability Engineering & System Safety 90 (1):8390. Sicotte, D., and S. Swanson. 2007. Whose risk in Philadelphia? Proximity to unequally hazardous industrial facilities. Social Science Quarterly (Blackwell Publishing Limited) 88 (2):515534. Smith, D. M. 1994. Geography and Social Justice Oxford: Blackwell Publishers Inc. PAGE 276 262 Thrift, N. 2005. Space: the fundamental stuff of human geography. In Key Concepts in Geography eds. S. L. Holloway, Rice, S.P., and G. Valentine, 95107. London: Sage Publications. Tobler, W. R. 1970. A computer movie simulating urban growth in the Detroit region. Ec onomic Geography 46 (2):234240. Torrens, P. M. 2008. Wi fi geographies. Annals of the Association of American Geographers 98 (1):5984. United Church of Christ Commission for Racial Justice. 1987. Toxic wastes and race in the United States: A national rep ort on the racial and socioeconomic characteristics of communities with hazardous waste sites. New York: United Church of Christ. . 2007. Toxic wastes and race in the United States: A national report on the racial and socioeconomic characteristics of communities with hazardous waste sites New York: United Church of Christ. US Census Bureau. 1997. Participant statistical areas program guidelines: census tracts, block groups (BGs), census designated places (CDPs), census county divisions (CCDs) U.S. Cens us, Form D 1500. Washington, D.C. . 2000. Poverty Status in 1999. In American Fact Finder http://factfinder.census.gov . 2002. Census Basics. http://www.census.gov/mso/www/c2000basics/00Basics.pdf . 2005. Geographic Areas Reference Manual, ed. Georgraphy Division. Washington DC: Bureau US Census. US Congress. 2000. Minority Health and Health Disparities Research and Education Act In United States Public Law 106525, 24952511. Washington D.C.: 106th Congress. US Department of Health and Human Services. 2000. Healthy People 2010: Understanding and Improving Health (second edition, four color version): PAGE 277 263 http://www.healthypeople.gov/Document/tableofcontents.htm#under (Accessed September 21, 2007). US Department of Justice. 1964. Title VI Legal Manual 28 C.F.R. 42.101 et seq., ed. Civil Rights Divisio n. Washington DC: US Department of Justice. US Environmental Protection Agency. 1992. Guidelines of Exposure Assessment. Accessed February 3, 2007. . 1996a. Chemical accident prevention and the clean air act amendments of 1990 Fact sheet 550F 96 004, ed. Office of Solid Waste and Emergency Response (5101). . 1996b. An SAB report: the cumulative exposure project: review of the planning, policy, an evaluations cumulative exposure project phase I EPA SAB IHEC ADV96004. Washington, D.C.: Science Advisory Board Integrated Human Exposure Committee. . 2004a. EPAs RiskScreening Environmental Indicators (RSEI) Chronic Human Health Methodology.RSEI Version 2.1.2. http://www.epa.g ov/opptintr/rsei/pubs/method_aug2004.pdf (Accessed December 1, 2006). . 2004b. RSEI Basic Information. http://www.epa.gov/opptintr/rsei/pubs/basic_information.html (Accessed December 1, 2006). . 2004c. The toxics release inventory (TRI) and factors to consider when using TRI data. http://www.epa.gov/tri/FactorsToConPDF.pdf (Accessed December 1, 2006). . 2004d. Users Manual for RSEI Version 2.1: Economic, Exposure and Technology Division, Office of Pollution Prevention and Toxics, August 2004. . 2006a. Acute Exposure Guidelines: (Accessed February 3, 2007). . 2006b. TRI Public Data Release eReport Summa ry of Key Findings, ed. A. 2006. http://www.epa.gov/tri/tridata/tri04/ereport/KeyFind.pdf (Accessed December 1, 2006). PAGE 278 264 US Environmental Protection Agency, and National Oceanic and A tmospheric Administration. 2007. ALOHA Users Manual, February. US General Accounting Office. 1983. Siting of hazardous waste landfills and their correlation with the racial and socioeconomic status of surrounding communities. Washington D.C.: General Accounting Office. Wilhelm, M., and B. Ritz. 2003. Residential proximity to traffic and adverse birth outcomes in Los Angeles, California, 19941996. Environmental Health Perspectives 111 (2):207 216. Williams, D. R., and P. B. Jackson. 2005. Social sources of racial disparities in health. Health Affairs 24 (2). Williams, R. W. 1999. The contested terrain of environmental justice research: community as unit of analysis. Social Science Journal 36 (2):313328. Woodruff, T. J., J. D. Parker, A. D. Kyle, and K. C. Schoendorf. 2003. Disparities in exposure to air pollution during pregnancy. Environmental Health Perspectives 111 (7):942946. World Health Organization. 2007. The determinants of health, ed. http://www.who.int/hia/evidence/doh/en/print.html Wrigley, N., T. Holt, D. Steel, and M. Tranmer eds. 1996. Analysis, modelling and resolving the ecological fallacy New York: J. Wiley. Young, I. M. 1990. Justice and the Politics of Difference Prince ton: Princeton University Press. ed. 1983. Justice and Hazardous Waste Bowling Green: The Applied Philosophy Program Bowling Green State University. Zeitz, P., Orr, M.F., and W. E. Kaye. 2002. Public health consequences of mercury spills: hazardous substances emergency events surveillance system, 1993 1998. Environmental Health Perspectives 110 (2):129132. PAGE 279 ABOUT THE AUTHOR Marilyn M. Williams began her collegiate journey at the University of Colorado at Boulder where she received her Bachelor s degree in Environmental Studies. She then continued her education at Ball State University where she received a Master of Science in Natural Resources and Environmental Management. She then relocated to Chicago, IL where she was employed as an Environm ental Scientist for several years Th e author returned to college where she received both a Masters of Public Health in Environmental Health and went on to obtain a Ph.D. at the University of South Florida in Geography and Environmental Science and Policy While pursuing these degrees she was a team member on several research projects and presented at numerous local and national conferences including the meetings of the American Public Health Association and Association of American Geographers. Today, Ms Williams is a Lecturer at the University of Denver and is seeking a tenure track position with a focus on Health Geography. 